pandas: powerful Python data analysis

toolkit
Release 0.25.3

Wes McKinney& PyData Development Team

Nov 02, 2019

CONTENTS

i
ii
 pandas: powerful Python data analysis toolkit, Release 0.25.3

Date: Nov 02, 2019 Version: 0.25.3

Download documentation: PDF Version | Zipped HTML
Useful links: Binary Installers | Source Repository | Issues & Ideas | Q&A Support | Mailing List
pandas is an open source, BSD-licensed library providing high-performance, easy-to-use data structures and data
analysis tools for the Python programming language.
See the overview for more detail about whats in the library.

CONTENTS 1
pandas: powerful Python data analysis toolkit, Release 0.25.3

2 CONTENTS
 CHAPTER

ONE

WHATS NEW IN 0.25.2 (OCTOBER 15, 2019)

These are the changes in pandas 0.25.2. See release for a full changelog including other versions of pandas.

Note: Pandas 0.25.2 adds compatibility for Python 3.8 (GH28147).

1.1 Bug fixes

1.1.1 Indexing

• Fix regression in DataFrame.reindex() not following the limit argument (GH28631).

• Fix regression in RangeIndex.get_indexer() for decreasing RangeIndex where target values may be
improperly identified as missing/present (GH28678)

1.1.2 I/O

• Fix regression in notebook display where <th> tags were missing for DataFrame.index values (GH28204).
• Regression in to_csv() where writing a Series or DataFrame indexed by an IntervalIndex would
incorrectly raise a TypeError (GH28210)
• Fix to_csv() with ExtensionArray with list-like values (GH28840).

1.1.3 Groupby/resample/rolling

• Bug incorrectly raising an IndexError when passing a list of quantiles to pandas.core.groupby.

DataFrameGroupBy.quantile() (GH28113).
• Bug in pandas.core.groupby.GroupBy.shift(), pandas.core.groupby.GroupBy.
bfill() and pandas.core.groupby.GroupBy.ffill() where timezone information would be
dropped (GH19995, GH27992)

• Compatibility with Python 3.8 in DataFrame.query() (GH27261)

• Fix to ensure that tab-completion in an IPython console does not raise warnings for deprecated attributes
(GH27900).

3
pandas: powerful Python data analysis toolkit, Release 0.25.3

1.2 Contributors

A total of 8 people contributed patches to this release. People with a + by their names contributed a patch for the first
time.
• Felix Divo +
• Jeremy Schendel
• Joris Van den Bossche
• MeeseeksMachine
• Tom Augspurger
• Will Ayd
• William Ayd
• jbrockmendel
{{ header }}

4 Chapter 1. Whats new in 0.25.2 (October 15, 2019)

 CHAPTER

TWO

INSTALLATION

The easiest way to install pandas is to install it as part of the Anaconda distribution, a cross platform distribution for
data analysis and scientific computing. This is the recommended installation method for most users.
Instructions for installing from source, PyPI, ActivePython, various Linux distributions, or a development version are
also provided.

2.1 Python version support

Officially Python 3.5.3 and above, 3.6, 3.7, and 3.8.

2.2 Installing pandas

2.2.1 Installing with Anaconda

Installing pandas and the rest of the NumPy and SciPy stack can be a little difficult for inexperienced users.
The simplest way to install not only pandas, but Python and the most popular packages that make up the SciPy stack
(IPython, NumPy, Matplotlib, ) is with Anaconda, a cross-platform (Linux, Mac OS X, Windows) Python distribution
for data analytics and scientific computing.
After running the installer, the user will have access to pandas and the rest of the SciPy stack without needing to install
anything else, and without needing to wait for any software to be compiled.
Installation instructions for Anaconda can be found here.
A full list of the packages available as part of the Anaconda distribution can be found here.
Another advantage to installing Anaconda is that you dont need admin rights to install it. Anaconda can install in the
users home directory, which makes it trivial to delete Anaconda if you decide (just delete that folder).

2.2.2 Installing with Miniconda

The previous section outlined how to get pandas installed as part of the Anaconda distribution. However this approach
means you will install well over one hundred packages and involves downloading the installer which is a few hundred
megabytes in size.
If you want to have more control on which packages, or have a limited internet bandwidth, then installing pandas with
Miniconda may be a better solution.
Conda is the package manager that the Anaconda distribution is built upon. It is a package manager that is both
cross-platform and language agnostic (it can play a similar role to a pip and virtualenv combination).

5
pandas: powerful Python data analysis toolkit, Release 0.25.3

Miniconda allows you to create a minimal self contained Python installation, and then use the Conda command to
install additional packages.
First you will need Conda to be installed and downloading and running the Miniconda will do this for you. The
installer can be found here
The next step is to create a new conda environment. A conda environment is like a virtualenv that allows you to specify
a specific version of Python and set of libraries. Run the following commands from a terminal window:

conda create -n name_of_my_env python

This will create a minimal environment with only Python installed in it. To put your self inside this environment run:

source activate name_of_my_env

On Windows the command is:

activate name_of_my_env

The final step required is to install pandas. This can be done with the following command:

conda install pandas

To install a specific pandas version:

conda install pandas=0.20.3

To install other packages, IPython for example:

conda install ipython

To install the full Anaconda distribution:

conda install anaconda

If you need packages that are available to pip but not conda, then install pip, and then use pip to install those packages:

conda install pip

pip install django

2.2.3 Installing from PyPI

pandas can be installed via pip from PyPI.

pip install pandas

2.2.4 Installing with ActivePython

Installation instructions for ActivePython can be found here. Versions 2.7 and 3.5 include pandas.

2.2.5 Installing using your Linux distributions package manager.

The commands in this table will install pandas for Python 3 from your distribution. To install pandas for Python 2,
you may need to use the python-pandas package.

6 Chapter 2. Installation
 pandas: powerful Python data analysis toolkit, Release 0.25.3

Distribution Status Download / Reposi- Install method

tory Link
Debian stable official Debian reposi- sudo apt-get install python3-pandas
tory
Debian & unstable NeuroDebian sudo apt-get install python3-pandas
Ubuntu (latest
pack-
ages)
Ubuntu stable official Ubuntu reposi- sudo apt-get install python3-pandas
tory
OpenSuse stable OpenSuse Repository zypper in python3-pandas
Fedora stable official Fedora reposi- dnf install python3-pandas
tory
Centos/RHELstable EPEL repository yum install python3-pandas

However, the packages in the linux package managers are often a few versions behind, so to get the newest version of
pandas, its recommended to install using the pip or conda methods described above.

2.2.6 Installing from source

See the contributing guide for complete instructions on building from the git source tree. Further, see creating a
development environment if you wish to create a pandas development environment.

2.3 Running the test suite

pandas is equipped with an exhaustive set of unit tests, covering about 97% of the code base as of this writing. To
run it on your machine to verify that everything is working (and that you have all of the dependencies, soft and hard,
installed), make sure you have pytest >= 4.0.2 and Hypothesis >= 3.58, then run:

>>> pd.test()
running: pytest --skip-slow --skip-network C:\Users\TP\Anaconda3\envs\py36\lib\site-
,→packages\pandas

============================= test session starts =============================

platform win32 -- Python 3.6.2, pytest-3.6.0, py-1.4.34, pluggy-0.4.0
rootdir: C:\Users\TP\Documents\Python\pandasdev\pandas, inifile: setup.cfg
collected 12145 items / 3 skipped

..................................................................S......
........S................................................................
.........................................................................

==================== 12130 passed, 12 skipped in 368.339 seconds =====================

2.3. Running the test suite 7

pandas: powerful Python data analysis toolkit, Release 0.25.3

2.4 Dependencies

Package Minimum supported version

setuptools 24.2.0
NumPy 1.13.3
python-dateutil 2.6.1
pytz 2017.2

2.4.1 Recommended dependencies

• numexpr: for accelerating certain numerical operations. numexpr uses multiple cores as well as smart chunk-
ing and caching to achieve large speedups. If installed, must be Version 2.6.2 or higher.
• bottleneck: for accelerating certain types of nan evaluations. bottleneck uses specialized cython routines
to achieve large speedups. If installed, must be Version 1.2.1 or higher.

Note: You are highly encouraged to install these libraries, as they provide speed improvements, especially when
working with large data sets.

2.4.2 Optional dependencies

Pandas has many optional dependencies that are only used for specific methods. For example, pandas.
read_hdf() requires the pytables package. If the optional dependency is not installed, pandas will raise an
ImportError when the method requiring that dependency is called.

8 Chapter 2. Installation
 pandas: powerful Python data analysis toolkit, Release 0.25.3

Dependency Minimum Version Notes

BeautifulSoup4 4.6.0 HTML parser for read_html (see note)
Jinja2 Conditional formatting with DataFrame.style
PyQt4 Clipboard I/O
PyQt5 Clipboard I/O
PyTables 3.4.2 HDF5-based reading / writing
SQLAlchemy 1.1.4 SQL support for databases other than sqlite
SciPy 0.19.0 Miscellaneous statistical functions
XLsxWriter 0.9.8 Excel writing
blosc Compression for msgpack
fastparquet 0.2.1 Parquet reading / writing
gcsfs 0.2.2 Google Cloud Storage access
html5lib HTML parser for read_html (see note)
lxml 3.8.0 HTML parser for read_html (see note)
matplotlib 2.2.2 Visualization
openpyxl 2.4.8 Reading / writing for xlsx files
pandas-gbq 0.8.0 Google Big Query access
psycopg2 PostgreSQL engine for sqlalchemy
pyarrow 0.9.0 Parquet and feather reading / writing
pymysql 0.7.11 MySQL engine for sqlalchemy
pyreadstat SPSS files (.sav) reading
pytables 3.4.2 HDF5 reading / writing
qtpy Clipboard I/O
s3fs 0.0.8 Amazon S3 access
xarray 0.8.2 pandas-like API for N-dimensional data
xclip Clipboard I/O on linux
xlrd 1.1.0 Excel reading
xlwt 1.2.0 Excel writing
xsel Clipboard I/O on linux
zlib Compression for msgpack

Optional dependencies for parsing HTML

One of the following combinations of libraries is needed to use the top-level read_html() function:
Changed in version 0.23.0.
• BeautifulSoup4 and html5lib
• BeautifulSoup4 and lxml
• BeautifulSoup4 and html5lib and lxml
• Only lxml, although see HTML Table Parsing for reasons as to why you should probably not take this approach.

Warning:
• if you install BeautifulSoup4 you must install either lxml or html5lib or both. read_html() will not work
with only BeautifulSoup4 installed.
• You are highly encouraged to read HTML Table Parsing gotchas. It explains issues surrounding the installa-
tion and usage of the above three libraries.

2.4. Dependencies 9
pandas: powerful Python data analysis toolkit, Release 0.25.3

{{ header }}

10 Chapter 2. Installation
 CHAPTER

THREE

GETTING STARTED

{{ header }}

3.1 Package overview

pandas is a Python package providing fast, flexible, and expressive data structures designed to make working with
relational or labeled data both easy and intuitive. It aims to be the fundamental high-level building block for doing
practical, real world data analysis in Python. Additionally, it has the broader goal of becoming the most powerful
and flexible open source data analysis / manipulation tool available in any language. It is already well on its way
toward this goal.
pandas is well suited for many different kinds of data:
• Tabular data with heterogeneously-typed columns, as in an SQL table or Excel spreadsheet
• Ordered and unordered (not necessarily fixed-frequency) time series data.
• Arbitrary matrix data (homogeneously typed or heterogeneous) with row and column labels
• Any other form of observational / statistical data sets. The data actually need not be labeled at all to be placed
into a pandas data structure
The two primary data structures of pandas, Series (1-dimensional) and DataFrame (2-dimensional), handle the
vast majority of typical use cases in finance, statistics, social science, and many areas of engineering. For R users,
DataFrame provides everything that Rs data.frame provides and much more. pandas is built on top of NumPy
and is intended to integrate well within a scientific computing environment with many other 3rd party libraries.
Here are just a few of the things that pandas does well:
• Easy handling of missing data (represented as NaN) in floating point as well as non-floating point data
• Size mutability: columns can be inserted and deleted from DataFrame and higher dimensional objects
• Automatic and explicit data alignment: objects can be explicitly aligned to a set of labels, or the user can
simply ignore the labels and let Series, DataFrame, etc. automatically align the data for you in computations
• Powerful, flexible group by functionality to perform split-apply-combine operations on data sets, for both ag-
gregating and transforming data
• Make it easy to convert ragged, differently-indexed data in other Python and NumPy data structures into
DataFrame objects
• Intelligent label-based slicing, fancy indexing, and subsetting of large data sets
• Intuitive merging and joining data sets
• Flexible reshaping and pivoting of data sets

11
pandas: powerful Python data analysis toolkit, Release 0.25.3

• Hierarchical labeling of axes (possible to have multiple labels per tick)

• Robust IO tools for loading data from flat files (CSV and delimited), Excel files, databases, and saving / loading
data from the ultrafast HDF5 format
• Time series-specific functionality: date range generation and frequency conversion, moving window statistics,
moving window linear regressions, date shifting and lagging, etc.
Many of these principles are here to address the shortcomings frequently experienced using other languages / scientific
research environments. For data scientists, working with data is typically divided into multiple stages: munging and
cleaning data, analyzing / modeling it, then organizing the results of the analysis into a form suitable for plotting or
tabular display. pandas is the ideal tool for all of these tasks.
Some other notes
• pandas is fast. Many of the low-level algorithmic bits have been extensively tweaked in Cython code. However,
as with anything else generalization usually sacrifices performance. So if you focus on one feature for your
application you may be able to create a faster specialized tool.
• pandas is a dependency of statsmodels, making it an important part of the statistical computing ecosystem in
Python.
• pandas has been used extensively in production in financial applications.

3.1.1 Data structures

Dimensions Name Description

1 Series 1D labeled homogeneously-typed array
2 DataFrame General 2D labeled, size-mutable tabular structure with potentially
heterogeneously-typed column

Why more than one data structure?

The best way to think about the pandas data structures is as flexible containers for lower dimensional data. For
example, DataFrame is a container for Series, and Series is a container for scalars. We would like to be able to insert
and remove objects from these containers in a dictionary-like fashion.
Also, we would like sensible default behaviors for the common API functions which take into account the typical
orientation of time series and cross-sectional data sets. When using ndarrays to store 2- and 3-dimensional data, a
burden is placed on the user to consider the orientation of the data set when writing functions; axes are considered
more or less equivalent (except when C- or Fortran-contiguousness matters for performance). In pandas, the axes are
intended to lend more semantic meaning to the data; i.e., for a particular data set there is likely to be a right way to
orient the data. The goal, then, is to reduce the amount of mental effort required to code up data transformations in
downstream functions.
For example, with tabular data (DataFrame) it is more semantically helpful to think of the index (the rows) and the
columns rather than axis 0 and axis 1. Iterating through the columns of the DataFrame thus results in more readable
code:

for col in df.columns:

series = df[col]
# do something with series

12 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

3.1.2 Mutability and copying of data

All pandas data structures are value-mutable (the values they contain can be altered) but not always size-mutable. The
length of a Series cannot be changed, but, for example, columns can be inserted into a DataFrame. However, the vast
majority of methods produce new objects and leave the input data untouched. In general we like to favor immutability
where sensible.

3.1.3 Getting support

The first stop for pandas issues and ideas is the Github Issue Tracker. If you have a general question, pandas community
experts can answer through Stack Overflow.

3.1.4 Community

pandas is actively supported today by a community of like-minded individuals around the world who contribute their
valuable time and energy to help make open source pandas possible. Thanks to all of our contributors.
If youre interested in contributing, please visit the contributing guide.
pandas is a NumFOCUS sponsored project. This will help ensure the success of development of pandas as a world-
class open-source project, and makes it possible to donate to the project.

3.1.5 Project governance

The governance process that pandas project has used informally since its inception in 2008 is formalized in Project
Governance documents. The documents clarify how decisions are made and how the various elements of our commu-
nity interact, including the relationship between open source collaborative development and work that may be funded
by for-profit or non-profit entities.
Wes McKinney is the Benevolent Dictator for Life (BDFL).

3.1.6 Development team

The list of the Core Team members and more detailed information can be found on the peoples page of the governance
repo.

3.1.7 Institutional partners

The information about current institutional partners can be found on pandas website page.

3.1.8 License

BSD 3-Clause License

Copyright (c) 2008-2012, AQR Capital Management, LLC, Lambda Foundry, Inc. and PyData
,→Development Team

All rights reserved.

Redistribution and use in source and binary forms, with or without

modification, are permitted provided that the following conditions are met:
(continues on next page)

3.1. Package overview 13

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

* Redistributions of source code must retain the above copyright notice, this
list of conditions and the following disclaimer.

* Redistributions in binary form must reproduce the above copyright notice,

this list of conditions and the following disclaimer in the documentation
and/or other materials provided with the distribution.

* Neither the name of the copyright holder nor the names of its
contributors may be used to endorse or promote products derived from
this software without specific prior written permission.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

{{ header }}

3.2 10 minutes to pandas

This is a short introduction to pandas, geared mainly for new users. You can see more complex recipes in the Cookbook.
Customarily, we import as follows:

In [1]: import numpy as np

In [2]: import pandas as pd

3.2.1 Object creation

See the Data Structure Intro section.

Creating a Series by passing a list of values, letting pandas create a default integer index:

In [3]: s = pd.Series([1, 3, 5, np.nan, 6, 8])

In [4]: s
Out[4]:
0 1.0
1 3.0
2 5.0
3 NaN
4 6.0
5 8.0
dtype: float64

Creating a DataFrame by passing a NumPy array, with a datetime index and labeled columns:

14 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [5]: dates = pd.date_range('20130101', periods=6)

In [6]: dates
Out[6]:
DatetimeIndex(['2013-01-01', '2013-01-02', '2013-01-03', '2013-01-04',
'2013-01-05', '2013-01-06'],
dtype='datetime64[ns]', freq='D')

In [7]: df = pd.DataFrame(np.random.randn(6, 4), index=dates, columns=list('ABCD'))

In [8]: df
Out[8]:
A B C D
2013-01-01 1.832747 1.515386 1.793547 -0.360634
2013-01-02 -0.913436 0.035141 3.437482 -1.106914
2013-01-03 -1.323650 0.427355 0.835343 -0.000698
2013-01-04 0.509859 -2.769586 1.000521 -0.865748
2013-01-05 0.139488 -0.259328 1.082034 -0.902452
2013-01-06 -0.130327 -0.372906 1.072236 -0.424347

Creating a DataFrame by passing a dict of objects that can be converted to series-like.

In [9]: df2 = pd.DataFrame({'A': 1.,

...: 'B': pd.Timestamp('20130102'),
...: 'C': pd.Series(1, index=list(range(4)), dtype='float32'),
...: 'D': np.array([3] * 4, dtype='int32'),
...: 'E': pd.Categorical(["test", "train", "test", "train"]),
...: 'F': 'foo'})
...:

In [10]: df2
Out[10]:
A B C D E F
0 1.0 2013-01-02 1.0 3 test foo
1 1.0 2013-01-02 1.0 3 train foo
2 1.0 2013-01-02 1.0 3 test foo
3 1.0 2013-01-02 1.0 3 train foo

The columns of the resulting DataFrame have different dtypes.

In [11]: df2.dtypes
Out[11]:
A float64
B datetime64[ns]
C float32
D int32
E category
F object
dtype: object

If youre using IPython, tab completion for column names (as well as public attributes) is automatically enabled. Heres
a subset of the attributes that will be completed:

In [12]: df2.<TAB> # noqa: E225, E999

df2.A df2.bool
df2.abs df2.boxplot
df2.add df2.C
(continues on next page)

3.2. 10 minutes to pandas 15

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

df2.add_prefix df2.clip
df2.add_suffix df2.clip_lower
df2.align df2.clip_upper
df2.all df2.columns
df2.any df2.combine
df2.append df2.combine_first
df2.apply df2.compound
df2.applymap df2.consolidate
df2.D

As you can see, the columns A, B, C, and D are automatically tab completed. E is there as well; the rest of the attributes
have been truncated for brevity.

3.2.2 Viewing data

See the Basics section.

Here is how to view the top and bottom rows of the frame:
In [13]: df.head()
Out[13]:
A B C D
2013-01-01 1.832747 1.515386 1.793547 -0.360634
2013-01-02 -0.913436 0.035141 3.437482 -1.106914
2013-01-03 -1.323650 0.427355 0.835343 -0.000698
2013-01-04 0.509859 -2.769586 1.000521 -0.865748
2013-01-05 0.139488 -0.259328 1.082034 -0.902452

In [14]: df.tail(3)
Out[14]:
A B C D
2013-01-04 0.509859 -2.769586 1.000521 -0.865748
2013-01-05 0.139488 -0.259328 1.082034 -0.902452
2013-01-06 -0.130327 -0.372906 1.072236 -0.424347
Display the index, columns:
In [15]: df.index
Out[15]:
DatetimeIndex(['2013-01-01', '2013-01-02', '2013-01-03', '2013-01-04',
'2013-01-05', '2013-01-06'],
dtype='datetime64[ns]', freq='D')

In [16]: df.columns
Out[16]: Index(['A', 'B', 'C', 'D'], dtype='object')
DataFrame.to_numpy() gives a NumPy representation of the underlying data. Note that this can be an expensive
operation when your DataFrame has columns with different data types, which comes down to a fundamental differ-
ence between pandas and NumPy: NumPy arrays have one dtype for the entire array, while pandas DataFrames
have one dtype per column. When you call DataFrame.to_numpy(), pandas will find the NumPy dtype that
can hold all of the dtypes in the DataFrame. This may end up being object, which requires casting every value to a
Python object.
For df, our DataFrame of all floating-point values, DataFrame.to_numpy() is fast and doesnt require copying
data.

16 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [17]: df.to_numpy()
Out[17]:
array([[ 1.83274697e+00, 1.51538609e+00, 1.79354724e+00,
-3.60634458e-01],
[-9.13435768e-01, 3.51414290e-02, 3.43748191e+00,
-1.10691447e+00],
[-1.32365020e+00, 4.27354647e-01, 8.35343007e-01,
-6.97782589e-04],
[ 5.09859417e-01, -2.76958615e+00, 1.00052083e+00,
-8.65747849e-01],
[ 1.39488428e-01, -2.59327906e-01, 1.08203428e+00,
-9.02451725e-01],
[-1.30327388e-01, -3.72906082e-01, 1.07223611e+00,
-4.24346700e-01]])

For df2, the DataFrame with multiple dtypes, DataFrame.to_numpy() is relatively expensive.

In [18]: df2.to_numpy()
Out[18]:
array([[1.0, Timestamp('2013-01-02 00:00:00'), 1.0, 3, 'test', 'foo'],
[1.0, Timestamp('2013-01-02 00:00:00'), 1.0, 3, 'train', 'foo'],
[1.0, Timestamp('2013-01-02 00:00:00'), 1.0, 3, 'test', 'foo'],
[1.0, Timestamp('2013-01-02 00:00:00'), 1.0, 3, 'train', 'foo']],
dtype=object)

Note: DataFrame.to_numpy() does not include the index or column labels in the output.

describe() shows a quick statistic summary of your data:

In [19]: df.describe()
Out[19]:
A B C D
count 6.000000 6.000000 6.000000 6.000000
mean 0.019114 -0.237323 1.536861 -0.610132
std 1.117102 1.415574 0.988006 0.416115
min -1.323650 -2.769586 0.835343 -1.106914
25% -0.717659 -0.344512 1.018450 -0.893276
50% 0.004581 -0.112093 1.077135 -0.645047
75% 0.417267 0.329301 1.615669 -0.376563
max 1.832747 1.515386 3.437482 -0.000698

Transposing your data:

In [20]: df.T
Out[20]:
2013-01-01 2013-01-02 2013-01-03 2013-01-04 2013-01-05 2013-01-06
A 1.832747 -0.913436 -1.323650 0.509859 0.139488 -0.130327
B 1.515386 0.035141 0.427355 -2.769586 -0.259328 -0.372906
C 1.793547 3.437482 0.835343 1.000521 1.082034 1.072236
D -0.360634 -1.106914 -0.000698 -0.865748 -0.902452 -0.424347

Sorting by an axis:

In [21]: df.sort_index(axis=1, ascending=False)

Out[21]:
(continues on next page)

3.2. 10 minutes to pandas 17

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

D C B A
2013-01-01 -0.360634 1.793547 1.515386 1.832747
2013-01-02 -1.106914 3.437482 0.035141 -0.913436
2013-01-03 -0.000698 0.835343 0.427355 -1.323650
2013-01-04 -0.865748 1.000521 -2.769586 0.509859
2013-01-05 -0.902452 1.082034 -0.259328 0.139488
2013-01-06 -0.424347 1.072236 -0.372906 -0.130327

Sorting by values:

In [22]: df.sort_values(by='B')
Out[22]:
A B C D
2013-01-04 0.509859 -2.769586 1.000521 -0.865748
2013-01-06 -0.130327 -0.372906 1.072236 -0.424347
2013-01-05 0.139488 -0.259328 1.082034 -0.902452
2013-01-02 -0.913436 0.035141 3.437482 -1.106914
2013-01-03 -1.323650 0.427355 0.835343 -0.000698
2013-01-01 1.832747 1.515386 1.793547 -0.360634

3.2.3 Selection

Note: While standard Python / Numpy expressions for selecting and setting are intuitive and come in handy for
interactive work, for production code, we recommend the optimized pandas data access methods, .at, .iat, .loc
and .iloc.

See the indexing documentation Indexing and Selecting Data and MultiIndex / Advanced Indexing.

Getting

Selecting a single column, which yields a Series, equivalent to df.A:

In [23]: df['A']
Out[23]:
2013-01-01 1.832747
2013-01-02 -0.913436
2013-01-03 -1.323650
2013-01-04 0.509859
2013-01-05 0.139488
2013-01-06 -0.130327
Freq: D, Name: A, dtype: float64

Selecting via [], which slices the rows.

In [24]: df[0:3]
Out[24]:
A B C D
2013-01-01 1.832747 1.515386 1.793547 -0.360634
2013-01-02 -0.913436 0.035141 3.437482 -1.106914
2013-01-03 -1.323650 0.427355 0.835343 -0.000698

In [25]: df['20130102':'20130104']

18 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[25]:
A B C D
2013-01-02 -0.913436 0.035141 3.437482 -1.106914
2013-01-03 -1.323650 0.427355 0.835343 -0.000698
2013-01-04 0.509859 -2.769586 1.000521 -0.865748

Selection by label

See more in Selection by Label.

For getting a cross section using a label:

In [26]: df.loc[dates[0]]
Out[26]:
A 1.832747
B 1.515386
C 1.793547
D -0.360634
Name: 2013-01-01 00:00:00, dtype: float64

Selecting on a multi-axis by label:

In [27]: df.loc[:, ['A', 'B']]

Out[27]:
A B
2013-01-01 1.832747 1.515386
2013-01-02 -0.913436 0.035141
2013-01-03 -1.323650 0.427355
2013-01-04 0.509859 -2.769586
2013-01-05 0.139488 -0.259328
2013-01-06 -0.130327 -0.372906

Showing label slicing, both endpoints are included:

In [28]: df.loc['20130102':'20130104', ['A', 'B']]

Out[28]:
A B
2013-01-02 -0.913436 0.035141
2013-01-03 -1.323650 0.427355
2013-01-04 0.509859 -2.769586

Reduction in the dimensions of the returned object:

In [29]: df.loc['20130102', ['A', 'B']]

Out[29]:
A -0.913436
B 0.035141
Name: 2013-01-02 00:00:00, dtype: float64

For getting a scalar value:

In [30]: df.loc[dates[0], 'A']

Out[30]: 1.8327469709663295

For getting fast access to a scalar (equivalent to the prior method):

3.2. 10 minutes to pandas 19

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [31]: df.at[dates[0], 'A']

Out[31]: 1.8327469709663295

Selection by position

See more in Selection by Position.

Select via the position of the passed integers:

In [32]: df.iloc[3]
Out[32]:
A 0.509859
B -2.769586
C 1.000521
D -0.865748
Name: 2013-01-04 00:00:00, dtype: float64

By integer slices, acting similar to numpy/python:

In [33]: df.iloc[3:5, 0:2]

Out[33]:
A B
2013-01-04 0.509859 -2.769586
2013-01-05 0.139488 -0.259328

By lists of integer position locations, similar to the numpy/python style:

In [34]: df.iloc[[1, 2, 4], [0, 2]]

Out[34]:
A C
2013-01-02 -0.913436 3.437482
2013-01-03 -1.323650 0.835343
2013-01-05 0.139488 1.082034

For slicing rows explicitly:

In [35]: df.iloc[1:3, :]
Out[35]:
A B C D
2013-01-02 -0.913436 0.035141 3.437482 -1.106914
2013-01-03 -1.323650 0.427355 0.835343 -0.000698

For slicing columns explicitly:

In [36]: df.iloc[:, 1:3]

Out[36]:
B C
2013-01-01 1.515386 1.793547
2013-01-02 0.035141 3.437482
2013-01-03 0.427355 0.835343
2013-01-04 -2.769586 1.000521
2013-01-05 -0.259328 1.082034
2013-01-06 -0.372906 1.072236

For getting a value explicitly:

20 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [37]: df.iloc[1, 1]
Out[37]: 0.03514142900432859

For getting fast access to a scalar (equivalent to the prior method):

In [38]: df.iat[1, 1]
Out[38]: 0.03514142900432859

Boolean indexing

Using a single columns values to select data.

In [39]: df[df.A > 0]

Out[39]:
A B C D
2013-01-01 1.832747 1.515386 1.793547 -0.360634
2013-01-04 0.509859 -2.769586 1.000521 -0.865748
2013-01-05 0.139488 -0.259328 1.082034 -0.902452

Selecting values from a DataFrame where a boolean condition is met.

In [40]: df[df > 0]

Out[40]:
A B C D
2013-01-01 1.832747 1.515386 1.793547 NaN
2013-01-02 NaN 0.035141 3.437482 NaN
2013-01-03 NaN 0.427355 0.835343 NaN
2013-01-04 0.509859 NaN 1.000521 NaN
2013-01-05 0.139488 NaN 1.082034 NaN
2013-01-06 NaN NaN 1.072236 NaN

Using the isin() method for filtering:

In [41]: df2 = df.copy()

In [42]: df2['E'] = ['one', 'one', 'two', 'three', 'four', 'three']

In [43]: df2
Out[43]:
A B C D E
2013-01-01 1.832747 1.515386 1.793547 -0.360634 one
2013-01-02 -0.913436 0.035141 3.437482 -1.106914 one
2013-01-03 -1.323650 0.427355 0.835343 -0.000698 two
2013-01-04 0.509859 -2.769586 1.000521 -0.865748 three
2013-01-05 0.139488 -0.259328 1.082034 -0.902452 four
2013-01-06 -0.130327 -0.372906 1.072236 -0.424347 three

In [44]: df2[df2['E'].isin(['two', 'four'])]

Out[44]:
A B C D E
2013-01-03 -1.323650 0.427355 0.835343 -0.000698 two
2013-01-05 0.139488 -0.259328 1.082034 -0.902452 four

3.2. 10 minutes to pandas 21

pandas: powerful Python data analysis toolkit, Release 0.25.3

Setting

Setting a new column automatically aligns the data by the indexes.

In [45]: s1 = pd.Series([1, 2, 3, 4, 5, 6], index=pd.date_range('20130102',

,→periods=6))

In [46]: s1
Out[46]:
2013-01-02 1
2013-01-03 2
2013-01-04 3
2013-01-05 4
2013-01-06 5
2013-01-07 6
Freq: D, dtype: int64

In [47]: df['F'] = s1

Setting values by label:

In [48]: df.at[dates[0], 'A'] = 0

Setting values by position:

In [49]: df.iat[0, 1] = 0

Setting by assigning with a NumPy array:

In [50]: df.loc[:, 'D'] = np.array([5] * len(df))

The result of the prior setting operations.

In [51]: df
Out[51]:
A B C D F
2013-01-01 0.000000 0.000000 1.793547 5 NaN
2013-01-02 -0.913436 0.035141 3.437482 5 1.0
2013-01-03 -1.323650 0.427355 0.835343 5 2.0
2013-01-04 0.509859 -2.769586 1.000521 5 3.0
2013-01-05 0.139488 -0.259328 1.082034 5 4.0
2013-01-06 -0.130327 -0.372906 1.072236 5 5.0

A where operation with setting.

In [52]: df2 = df.copy()

In [53]: df2[df2 > 0] = -df2

In [54]: df2
Out[54]:
A B C D F
2013-01-01 0.000000 0.000000 -1.793547 -5 NaN
2013-01-02 -0.913436 -0.035141 -3.437482 -5 -1.0
2013-01-03 -1.323650 -0.427355 -0.835343 -5 -2.0
2013-01-04 -0.509859 -2.769586 -1.000521 -5 -3.0
2013-01-05 -0.139488 -0.259328 -1.082034 -5 -4.0
2013-01-06 -0.130327 -0.372906 -1.072236 -5 -5.0

22 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

3.2.4 Missing data

pandas primarily uses the value np.nan to represent missing data. It is by default not included in computations. See
the Missing Data section.
Reindexing allows you to change/add/delete the index on a specified axis. This returns a copy of the data.

In [55]: df1 = df.reindex(index=dates[0:4], columns=list(df.columns) + ['E'])

In [56]: df1.loc[dates[0]:dates[1], 'E'] = 1

In [57]: df1
Out[57]:
A B C D F E
2013-01-01 0.000000 0.000000 1.793547 5 NaN 1.0
2013-01-02 -0.913436 0.035141 3.437482 5 1.0 1.0
2013-01-03 -1.323650 0.427355 0.835343 5 2.0 NaN
2013-01-04 0.509859 -2.769586 1.000521 5 3.0 NaN

To drop any rows that have missing data.

In [58]: df1.dropna(how='any')
Out[58]:
A B C D F E
2013-01-02 -0.913436 0.035141 3.437482 5 1.0 1.0

Filling missing data.

In [59]: df1.fillna(value=5)
Out[59]:
A B C D F E
2013-01-01 0.000000 0.000000 1.793547 5 5.0 1.0
2013-01-02 -0.913436 0.035141 3.437482 5 1.0 1.0
2013-01-03 -1.323650 0.427355 0.835343 5 2.0 5.0
2013-01-04 0.509859 -2.769586 1.000521 5 3.0 5.0

To get the boolean mask where values are nan.

In [60]: pd.isna(df1)
Out[60]:
A B C D F E
2013-01-01 False False False False True False
2013-01-02 False False False False False False
2013-01-03 False False False False False True
2013-01-04 False False False False False True

3.2.5 Operations

See the Basic section on Binary Ops.

Stats

Operations in general exclude missing data.

Performing a descriptive statistic:

3.2. 10 minutes to pandas 23

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [61]: df.mean()
Out[61]:
A -0.286344
B -0.489887
C 1.536861
D 5.000000
F 3.000000
dtype: float64

Same operation on the other axis:

In [62]: df.mean(1)
Out[62]:
2013-01-01 1.698387
2013-01-02 1.711838
2013-01-03 1.387809
2013-01-04 1.348159
2013-01-05 1.992439
2013-01-06 2.113801
Freq: D, dtype: float64

Operating with objects that have different dimensionality and need alignment. In addition, pandas automatically
broadcasts along the specified dimension.
In [63]: s = pd.Series([1, 3, 5, np.nan, 6, 8], index=dates).shift(2)

In [64]: s
Out[64]:
2013-01-01 NaN
2013-01-02 NaN
2013-01-03 1.0
2013-01-04 3.0
2013-01-05 5.0
2013-01-06 NaN
Freq: D, dtype: float64

In [65]: df.sub(s, axis='index')

Out[65]:
A B C D F
2013-01-01 NaN NaN NaN NaN NaN
2013-01-02 NaN NaN NaN NaN NaN
2013-01-03 -2.323650 -0.572645 -0.164657 4.0 1.0
2013-01-04 -2.490141 -5.769586 -1.999479 2.0 0.0
2013-01-05 -4.860512 -5.259328 -3.917966 0.0 -1.0
2013-01-06 NaN NaN NaN NaN NaN

Apply

Applying functions to the data:

In [66]: df.apply(np.cumsum)
Out[66]:
A B C D F
2013-01-01 0.000000 0.000000 1.793547 5 NaN
2013-01-02 -0.913436 0.035141 5.231029 10 1.0

24 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

2013-01-03 -2.237086 0.462496 6.066372 15 3.0

2013-01-04 -1.727227 -2.307090 7.066893 20 6.0
2013-01-05 -1.587738 -2.566418 8.148927 25 10.0
2013-01-06 -1.718066 -2.939324 9.221163 30 15.0

In [67]: df.apply(lambda x: x.max() - x.min())

Out[67]:
A 1.833510
B 3.196941
C 2.602139
D 0.000000
F 4.000000
dtype: float64

Histogramming

See more at Histogramming and Discretization.

In [68]: s = pd.Series(np.random.randint(0, 7, size=10))

In [69]: s
Out[69]:
0 3
1 5
2 1
3 4
4 3
5 3
6 5
7 3
8 1
9 1
dtype: int64

In [70]: s.value_counts()
Out[70]:
3 4
1 3
5 2
4 1
dtype: int64

String Methods

Series is equipped with a set of string processing methods in the str attribute that make it easy to operate on each
element of the array, as in the code snippet below. Note that pattern-matching in str generally uses regular expressions
by default (and in some cases always uses them). See more at Vectorized String Methods.
In [71]: s = pd.Series(['A', 'B', 'C', 'Aaba', 'Baca', np.nan, 'CABA', 'dog', 'cat'])

In [72]: s.str.lower()
Out[72]:
0 a
(continues on next page)

3.2. 10 minutes to pandas 25

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

1 b
2 c
3 aaba
4 baca
5 NaN
6 caba
7 dog
8 cat
dtype: object

3.2.6 Merge

Concat

pandas provides various facilities for easily combining together Series and DataFrame objects with various kinds of
set logic for the indexes and relational algebra functionality in the case of join / merge-type operations.
See the Merging section.
Concatenating pandas objects together with concat():

In [73]: df = pd.DataFrame(np.random.randn(10, 4))

In [74]: df
Out[74]:
0 1 2 3
0 -0.060367 1.524865 0.131370 -1.258189
1 1.350664 0.281335 -1.137282 0.837561
2 -0.693000 0.319506 -0.629133 0.372674
3 0.044135 0.012920 0.369063 0.966930
4 2.401162 -0.910612 0.105657 -1.556792
5 0.633341 -1.540587 -1.379657 0.402392
6 -1.098101 2.017028 -0.309634 0.421307
7 0.098253 -1.233080 0.729444 -0.055611
8 -1.361441 -0.220021 -0.670728 -0.796977
9 -0.430638 -0.515560 1.171513 -0.172970

# break it into pieces

In [75]: pieces = [df[:3], df[3:7], df[7:]]

In [76]: pd.concat(pieces)
Out[76]:
0 1 2 3
0 -0.060367 1.524865 0.131370 -1.258189
1 1.350664 0.281335 -1.137282 0.837561
2 -0.693000 0.319506 -0.629133 0.372674
3 0.044135 0.012920 0.369063 0.966930
4 2.401162 -0.910612 0.105657 -1.556792
5 0.633341 -1.540587 -1.379657 0.402392
6 -1.098101 2.017028 -0.309634 0.421307
7 0.098253 -1.233080 0.729444 -0.055611
8 -1.361441 -0.220021 -0.670728 -0.796977
9 -0.430638 -0.515560 1.171513 -0.172970

26 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Join

SQL style merges. See the Database style joining section.

In [77]: left = pd.DataFrame({'key': ['foo', 'foo'], 'lval': [1, 2]})

In [78]: right = pd.DataFrame({'key': ['foo', 'foo'], 'rval': [4, 5]})

In [79]: left
Out[79]:
key lval
0 foo 1
1 foo 2

In [80]: right
Out[80]:
key rval
0 foo 4
1 foo 5

In [81]: pd.merge(left, right, on='key')

Out[81]:
key lval rval
0 foo 1 4
1 foo 1 5
2 foo 2 4
3 foo 2 5
Another example that can be given is:
In [82]: left = pd.DataFrame({'key': ['foo', 'bar'], 'lval': [1, 2]})

In [83]: right = pd.DataFrame({'key': ['foo', 'bar'], 'rval': [4, 5]})

In [84]: left
Out[84]:
key lval
0 foo 1
1 bar 2

In [85]: right
Out[85]:
key rval
0 foo 4
1 bar 5

In [86]: pd.merge(left, right, on='key')

Out[86]:
key lval rval
0 foo 1 4
1 bar 2 5

Append

Append rows to a dataframe. See the Appending section.

3.2. 10 minutes to pandas 27

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [87]: df = pd.DataFrame(np.random.randn(8, 4), columns=['A', 'B', 'C', 'D'])

In [88]: df
Out[88]:
A B C D
0 -0.588928 -0.915521 -1.117137 0.203822
1 -1.122360 0.304649 1.552535 0.002581
2 0.508278 -0.514235 1.658575 -0.641554
3 1.350919 1.331824 -0.207267 2.652396
4 -0.899836 2.550564 -2.093739 1.951529
5 -0.868126 -1.450904 0.513953 -1.340940
6 0.704139 -0.254053 -1.231656 0.412464
7 1.165165 -0.985036 -0.380646 0.322498

In [89]: s = df.iloc[3]

In [90]: df.append(s, ignore_index=True)

Out[90]:
A B C D
0 -0.588928 -0.915521 -1.117137 0.203822
1 -1.122360 0.304649 1.552535 0.002581
2 0.508278 -0.514235 1.658575 -0.641554
3 1.350919 1.331824 -0.207267 2.652396
4 -0.899836 2.550564 -2.093739 1.951529
5 -0.868126 -1.450904 0.513953 -1.340940
6 0.704139 -0.254053 -1.231656 0.412464
7 1.165165 -0.985036 -0.380646 0.322498
8 1.350919 1.331824 -0.207267 2.652396

3.2.7 Grouping

By group by we are referring to a process involving one or more of the following steps:
• Splitting the data into groups based on some criteria
• Applying a function to each group independently
• Combining the results into a data structure
See the Grouping section.

In [91]: df = pd.DataFrame({'A': ['foo', 'bar', 'foo', 'bar',

....: 'foo', 'bar', 'foo', 'foo'],
....: 'B': ['one', 'one', 'two', 'three',
....: 'two', 'two', 'one', 'three'],
....: 'C': np.random.randn(8),
....: 'D': np.random.randn(8)})
....:

In [92]: df
Out[92]:
A B C D
0 foo one -0.800850 -0.401946
1 bar one 0.574070 -0.239254
2 foo two 0.076080 -0.851860
3 bar three 0.658058 1.362899
4 foo two 0.014033 -0.724381
(continues on next page)

28 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

5 bar two -2.027835 -0.554717
6 foo one -0.871462 -0.654966
7 foo three 0.898298 0.280791

Grouping and then applying the sum() function to the resulting groups.

In [93]: df.groupby('A').sum()
Out[93]:
C D
A
bar -0.795706 0.568928
foo -0.683901 -2.352362

Grouping by multiple columns forms a hierarchical index, and again we can apply the sum function.

In [94]: df.groupby(['A', 'B']).sum()

Out[94]:
C D
A B
bar one 0.574070 -0.239254
three 0.658058 1.362899
two -2.027835 -0.554717
foo one -1.672312 -1.056912
three 0.898298 0.280791
two 0.090113 -1.576241

3.2.8 Reshaping

See the sections on Hierarchical Indexing and Reshaping.

Stack

In [95]: tuples = list(zip(*[['bar', 'bar', 'baz', 'baz',

....: 'foo', 'foo', 'qux', 'qux'],
....: ['one', 'two', 'one', 'two',
....: 'one', 'two', 'one', 'two']]))
....:

In [96]: index = pd.MultiIndex.from_tuples(tuples, names=['first', 'second'])

In [97]: df = pd.DataFrame(np.random.randn(8, 2), index=index, columns=['A', 'B'])

In [98]: df2 = df[:4]

In [99]: df2
Out[99]:
A B
first second
bar one 0.572341 -0.753047
two 2.561035 -2.172735
baz one -0.747480 -1.381895
two -1.271458 1.926008

The stack() method compresses a level in the DataFrames columns.

3.2. 10 minutes to pandas 29

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [100]: stacked = df2.stack()

In [101]: stacked
Out[101]:
first second
bar one A 0.572341
B -0.753047
two A 2.561035
B -2.172735
baz one A -0.747480
B -1.381895
two A -1.271458
B 1.926008
dtype: float64

With a stacked DataFrame or Series (having a MultiIndex as the index), the inverse operation of stack() is
unstack(), which by default unstacks the last level:
In [102]: stacked.unstack()
Out[102]:
A B
first second
bar one 0.572341 -0.753047
two 2.561035 -2.172735
baz one -0.747480 -1.381895
two -1.271458 1.926008

In [103]: stacked.unstack(1)
Out[103]:
second one two
first
bar A 0.572341 2.561035
B -0.753047 -2.172735
baz A -0.747480 -1.271458
B -1.381895 1.926008

In [104]: stacked.unstack(0)
Out[104]:
first bar baz
second
one A 0.572341 -0.747480
B -0.753047 -1.381895
two A 2.561035 -1.271458
B -2.172735 1.926008

Pivot tables

See the section on Pivot Tables.

In [105]: df = pd.DataFrame({'A': ['one', 'one', 'two', 'three'] * 3,

.....: 'B': ['A', 'B', 'C'] * 4,
.....: 'C': ['foo', 'foo', 'foo', 'bar', 'bar', 'bar'] * 2,
.....: 'D': np.random.randn(12),
.....: 'E': np.random.randn(12)})
(continues on next page)

30 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

.....:

In [106]: df
Out[106]:
A B C D E
0 one A foo -0.780671 -0.261755
1 one B foo -0.294512 0.557932
2 two C foo 0.214464 -0.014909
3 three A bar 0.176056 3.227353
4 one B bar -0.836311 -0.832515
5 one C bar 0.579569 0.072856
6 two A foo -0.746106 -0.620299
7 three B foo -0.214858 1.084295
8 one C foo 0.206084 0.244296
9 one A bar 0.505611 -0.589272
10 two B bar -0.181058 -1.117857
11 three C bar 0.527833 -1.061391

We can produce pivot tables from this data very easily:

In [107]: pd.pivot_table(df, values='D', index=['A', 'B'], columns=['C'])
Out[107]:
C bar foo
A B
one A 0.505611 -0.780671
B -0.836311 -0.294512
C 0.579569 0.206084
three A 0.176056 NaN
B NaN -0.214858
C 0.527833 NaN
two A NaN -0.746106
B -0.181058 NaN
C NaN 0.214464

3.2.9 Time series

pandas has simple, powerful, and efficient functionality for performing resampling operations during frequency con-
version (e.g., converting secondly data into 5-minutely data). This is extremely common in, but not limited to, financial
applications. See the Time Series section.
In [108]: rng = pd.date_range('1/1/2012', periods=100, freq='S')

In [109]: ts = pd.Series(np.random.randint(0, 500, len(rng)), index=rng)

In [110]: ts.resample('5Min').sum()
Out[110]:
2012-01-01 24418
Freq: 5T, dtype: int64

Time zone representation:

In [111]: rng = pd.date_range('3/6/2012 00:00', periods=5, freq='D')

In [112]: ts = pd.Series(np.random.randn(len(rng)), rng)

(continues on next page)

3.2. 10 minutes to pandas 31

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [113]: ts
Out[113]:
2012-03-06 0.198218
2012-03-07 -0.649384
2012-03-08 0.743513
2012-03-09 -0.541289
2012-03-10 -0.594013
Freq: D, dtype: float64

In [114]: ts_utc = ts.tz_localize('UTC')

In [115]: ts_utc
Out[115]:
2012-03-06 00:00:00+00:00 0.198218
2012-03-07 00:00:00+00:00 -0.649384
2012-03-08 00:00:00+00:00 0.743513
2012-03-09 00:00:00+00:00 -0.541289
2012-03-10 00:00:00+00:00 -0.594013
Freq: D, dtype: float64

Converting to another time zone:

In [116]: ts_utc.tz_convert('US/Eastern')
Out[116]:
2012-03-05 19:00:00-05:00 0.198218
2012-03-06 19:00:00-05:00 -0.649384
2012-03-07 19:00:00-05:00 0.743513
2012-03-08 19:00:00-05:00 -0.541289
2012-03-09 19:00:00-05:00 -0.594013
Freq: D, dtype: float64

Converting between time span representations:

In [117]: rng = pd.date_range('1/1/2012', periods=5, freq='M')

In [118]: ts = pd.Series(np.random.randn(len(rng)), index=rng)

In [119]: ts
Out[119]:
2012-01-31 -0.579681
2012-02-29 0.235443
2012-03-31 0.121223
2012-04-30 1.630841
2012-05-31 -0.795301
Freq: M, dtype: float64

In [120]: ps = ts.to_period()

In [121]: ps
Out[121]:
2012-01 -0.579681
2012-02 0.235443
2012-03 0.121223
2012-04 1.630841
2012-05 -0.795301
Freq: M, dtype: float64

32 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [122]: ps.to_timestamp()
Out[122]:
2012-01-01 -0.579681
2012-02-01 0.235443
2012-03-01 0.121223
2012-04-01 1.630841
2012-05-01 -0.795301
Freq: MS, dtype: float64
Converting between period and timestamp enables some convenient arithmetic functions to be used. In the following
example, we convert a quarterly frequency with year ending in November to 9am of the end of the month following
the quarter end:

In [123]: prng = pd.period_range('1990Q1', '2000Q4', freq='Q-NOV')

In [124]: ts = pd.Series(np.random.randn(len(prng)), prng)

In [125]: ts.index = (prng.asfreq('M', 'e') + 1).asfreq('H', 's') + 9

In [126]: ts.head()
Out[126]:
1990-03-01 09:00 0.456987
1990-06-01 09:00 -1.222265
1990-09-01 09:00 0.773708
1990-12-01 09:00 -1.138070
1991-03-01 09:00 -0.860873
Freq: H, dtype: float64

3.2.10 Categoricals

pandas can include categorical data in a DataFrame. For full docs, see the categorical introduction and the API
documentation.

In [127]: df = pd.DataFrame({"id": [1, 2, 3, 4, 5, 6],

.....: "raw_grade": ['a', 'b', 'b', 'a', 'a', 'e']})
.....:

Convert the raw grades to a categorical data type.

In [128]: df["grade"] = df["raw_grade"].astype("category")

In [129]: df["grade"]
Out[129]:
0 a
1 b
2 b
3 a
4 a
5 e
Name: grade, dtype: category
Categories (3, object): [a, b, e]

Rename the categories to more meaningful names (assigning to Series.cat.categories is inplace!).

3.2. 10 minutes to pandas 33

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [130]: df["grade"].cat.categories = ["very good", "good", "very bad"]

Reorder the categories and simultaneously add the missing categories (methods under Series .cat return a new
Series by default).

In [131]: df["grade"] = df["grade"].cat.set_categories(["very bad", "bad", "medium",

.....: "good", "very good"])
.....:

In [132]: df["grade"]
Out[132]:
0 very good
1 good
2 good
3 very good
4 very good
5 very bad
Name: grade, dtype: category
Categories (5, object): [very bad, bad, medium, good, very good]

Sorting is per order in the categories, not lexical order.

In [133]: df.sort_values(by="grade")
Out[133]:
id raw_grade grade
5 6 e very bad
1 2 b good
2 3 b good
0 1 a very good
3 4 a very good
4 5 a very good

Grouping by a categorical column also shows empty categories.

In [134]: df.groupby("grade").size()
Out[134]:
grade
very bad 1
bad 0
medium 0
good 2
very good 3
dtype: int64

3.2.11 Plotting

See the Plotting docs.

In [135]: ts = pd.Series(np.random.randn(1000),
.....: index=pd.date_range('1/1/2000', periods=1000))
.....:

In [136]: ts = ts.cumsum()

In [137]: ts.plot()
Out[137]: <matplotlib.axes._subplots.AxesSubplot at 0x124144910>

34 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

On a DataFrame, the plot() method is a convenience to plot all of the columns with labels:
In [138]: df = pd.DataFrame(np.random.randn(1000, 4), index=ts.index,
.....: columns=['A', 'B', 'C', 'D'])
.....:

In [139]: df = df.cumsum()

In [140]: plt.figure()
Out[140]: <Figure size 640x480 with 0 Axes>

In [141]: df.plot()
Out[141]: <matplotlib.axes._subplots.AxesSubplot at 0x129c58f90>

In [142]: plt.legend(loc='best')
Out[142]: <matplotlib.legend.Legend at 0x129c5df10>

3.2. 10 minutes to pandas 35

pandas: powerful Python data analysis toolkit, Release 0.25.3

3.2.12 Getting data in/out

CSV

Writing to a csv file.

In [143]: df.to_csv('foo.csv')

Reading from a csv file.

In [144]: pd.read_csv('foo.csv')
Out[144]:
Unnamed: 0 A B C D
0 2000-01-01 -0.327225 2.629468 0.152017 0.095606
1 2000-01-02 -0.924297 0.950886 0.363196 -1.492713
2 2000-01-03 -0.987506 0.909911 0.357891 -1.526351
3 2000-01-04 -2.537033 1.483865 0.264791 -2.119897
4 2000-01-05 -2.152314 2.492344 1.618494 -2.466438
.. ... ... ... ... ...
995 2002-09-22 11.523422 -30.110846 69.683435 68.775420
996 2002-09-23 11.002664 -29.664630 69.359104 67.976973
997 2002-09-24 11.327230 -28.225995 69.602334 66.540402
(continues on next page)

36 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

998 2002-09-25 12.654992 -28.103791 70.420846 66.271805
999 2002-09-26 13.310373 -27.917046 71.319203 67.595342

[1000 rows x 5 columns]

HDF5

Reading and writing to HDFStores.

Writing to a HDF5 Store.

In [145]: df.to_hdf('foo.h5', 'df')

Reading from a HDF5 Store.

In [146]: pd.read_hdf('foo.h5', 'df')

Out[146]:
A B C D
2000-01-01 -0.327225 2.629468 0.152017 0.095606
2000-01-02 -0.924297 0.950886 0.363196 -1.492713
2000-01-03 -0.987506 0.909911 0.357891 -1.526351
2000-01-04 -2.537033 1.483865 0.264791 -2.119897
2000-01-05 -2.152314 2.492344 1.618494 -2.466438
... ... ... ... ...
2002-09-22 11.523422 -30.110846 69.683435 68.775420
2002-09-23 11.002664 -29.664630 69.359104 67.976973
2002-09-24 11.327230 -28.225995 69.602334 66.540402
2002-09-25 12.654992 -28.103791 70.420846 66.271805
2002-09-26 13.310373 -27.917046 71.319203 67.595342

[1000 rows x 4 columns]

Excel

Reading and writing to MS Excel.

Writing to an excel file.

In [147]: df.to_excel('foo.xlsx', sheet_name='Sheet1')

Reading from an excel file.

In [148]: pd.read_excel('foo.xlsx', 'Sheet1', index_col=None, na_values=['NA'])

Out[148]:
Unnamed: 0 A B C D
0 2000-01-01 -0.327225 2.629468 0.152017 0.095606
1 2000-01-02 -0.924297 0.950886 0.363196 -1.492713
2 2000-01-03 -0.987506 0.909911 0.357891 -1.526351
3 2000-01-04 -2.537033 1.483865 0.264791 -2.119897
4 2000-01-05 -2.152314 2.492344 1.618494 -2.466438
.. ... ... ... ... ...
995 2002-09-22 11.523422 -30.110846 69.683435 68.775420
996 2002-09-23 11.002664 -29.664630 69.359104 67.976973
997 2002-09-24 11.327230 -28.225995 69.602334 66.540402
(continues on next page)

3.2. 10 minutes to pandas 37

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

998 2002-09-25 12.654992 -28.103791 70.420846 66.271805
999 2002-09-26 13.310373 -27.917046 71.319203 67.595342

[1000 rows x 5 columns]

3.2.13 Gotchas

If you are attempting to perform an operation you might see an exception like:

>>> if pd.Series([False, True, False]):

... print("I was true")
Traceback
...
ValueError: The truth value of an array is ambiguous. Use a.empty, a.any() or a.all().

See Comparisons for an explanation and what to do.

See Gotchas as well. {{ header }}

3.3 Essential basic functionality

Here we discuss a lot of the essential functionality common to the pandas data structures. Heres how to create some
of the objects used in the examples from the previous section:

In [1]: index = pd.date_range('1/1/2000', periods=8)

In [2]: s = pd.Series(np.random.randn(5), index=['a', 'b', 'c', 'd', 'e'])

In [3]: df = pd.DataFrame(np.random.randn(8, 3), index=index,

...: columns=['A', 'B', 'C'])
...:

3.3.1 Head and tail

To view a small sample of a Series or DataFrame object, use the head() and tail() methods. The default number
of elements to display is five, but you may pass a custom number.
In [4]: long_series = pd.Series(np.random.randn(1000))

In [5]: long_series.head()
Out[5]:
0 -0.183284
1 -1.084859
2 0.804220
3 -2.278089
4 0.736586
dtype: float64

In [6]: long_series.tail(3)
Out[6]:
997 -0.935665

38 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

998 0.865342
999 -0.798301
dtype: float64

3.3.2 Attributes and underlying data

pandas objects have a number of attributes enabling you to access the metadata
• shape: gives the axis dimensions of the object, consistent with ndarray
• Axis labels
– Series: index (only axis)
– DataFrame: index (rows) and columns
Note, these attributes can be safely assigned to!

In [7]: df[:2]
Out[7]:
A B C
2000-01-01 -0.216974 0.840135 -0.733559
2000-01-02 0.054693 -2.419175 -0.196602

In [8]: df.columns = [x.lower() for x in df.columns]

In [9]: df
Out[9]:
a b c
2000-01-01 -0.216974 0.840135 -0.733559
2000-01-02 0.054693 -2.419175 -0.196602
2000-01-03 0.088430 0.295848 -0.321244
2000-01-04 -0.405187 0.377840 -0.991461
2000-01-05 -0.204197 -0.949996 -0.892288
2000-01-06 -1.144473 0.336703 -1.526753
2000-01-07 -1.591884 -1.341427 -1.075799
2000-01-08 0.101217 0.253586 1.122798

Pandas objects (Index, Series, DataFrame) can be thought of as containers for arrays, which hold the actual
data and do the actual computation. For many types, the underlying array is a numpy.ndarray. However, pandas
and 3rd party libraries may extend NumPys type system to add support for custom arrays (see dtypes).
To get the actual data inside a Index or Series, use the .array property
In [10]: s.array
Out[10]:
<PandasArray>
[-0.0065828035649045415, -2.7167036660839705, 1.2287390078978135,
0.9324879342551502, 0.2940471792704379]
Length: 5, dtype: float64

In [11]: s.index.array
Out[11]:
<PandasArray>
['a', 'b', 'c', 'd', 'e']
Length: 5, dtype: object
array will always be an ExtensionArray. The exact details of what an ExtensionArray is and why pandas
uses them is a bit beyond the scope of this introduction. See dtypes for more.

3.3. Essential basic functionality 39

pandas: powerful Python data analysis toolkit, Release 0.25.3

If you know you need a NumPy array, use to_numpy() or numpy.asarray().

In [12]: s.to_numpy()
Out[12]: array([-0.0065828 , -2.71670367, 1.22873901, 0.93248793, 0.
,→29404718])

In [13]: np.asarray(s)
Out[13]: array([-0.0065828 , -2.71670367, 1.22873901, 0.93248793, 0.
,→29404718])

When the Series or Index is backed by an ExtensionArray, to_numpy() may involve copying data and coercing
values. See dtypes for more.
to_numpy() gives some control over the dtype of the resulting numpy.ndarray. For example, consider date-
times with timezones. NumPy doesnt have a dtype to represent timezone-aware datetimes, so there are two possibly
useful representations:
1. An object-dtype numpy.ndarray with Timestamp objects, each with the correct tz
2. A datetime64[ns] -dtype numpy.ndarray, where the values have been converted to UTC and the time-
zone discarded
Timezones may be preserved with dtype=object

In [14]: ser = pd.Series(pd.date_range('2000', periods=2, tz="CET"))

In [15]: ser.to_numpy(dtype=object)
Out[15]:
array([Timestamp('2000-01-01 00:00:00+0100', tz='CET', freq='D'),
Timestamp('2000-01-02 00:00:00+0100', tz='CET', freq='D')],
dtype=object)

Or thrown away with dtype='datetime64[ns]'

In [16]: ser.to_numpy(dtype="datetime64[ns]")
Out[16]:
array(['1999-12-31T23:00:00.000000000', '2000-01-01T23:00:00.000000000'],
dtype='datetime64[ns]')

Getting the raw data inside a DataFrame is possibly a bit more complex. When your DataFrame only has a single
data type for all the columns, DataFrame.to_numpy() will return the underlying data:

In [17]: df.to_numpy()
Out[17]:
array([[-0.21697434, 0.84013463, -0.73355917],
[ 0.05469281, -2.41917478, -0.19660224],
[ 0.08843011, 0.29584755, -0.32124436],
[-0.40518687, 0.37784039, -0.99146082],
[-0.20419686, -0.94999608, -0.89228838],
[-1.14447294, 0.33670262, -1.52675331],
[-1.59188382, -1.34142681, -1.07579903],
[ 0.1012167 , 0.2535859 , 1.12279785]])

If a DataFrame contains homogeneously-typed data, the ndarray can actually be modified in-place, and the changes
will be reflected in the data structure. For heterogeneous data (e.g. some of the DataFrames columns are not all the
same dtype), this will not be the case. The values attribute itself, unlike the axis labels, cannot be assigned to.

Note: When working with heterogeneous data, the dtype of the resulting ndarray will be chosen to accommodate all
of the data involved. For example, if strings are involved, the result will be of object dtype. If there are only floats and

40 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

integers, the resulting array will be of float dtype.

In the past, pandas recommended Series.values or DataFrame.values for extracting the data from a Series
or DataFrame. Youll still find references to these in old code bases and online. Going forward, we recommend
avoiding .values and using .array or .to_numpy(). .values has the following drawbacks:
1. When your Series contains an extension type, its unclear whether Series.values returns a NumPy array
or the extension array. Series.array will always return an ExtensionArray, and will never copy data.
Series.to_numpy() will always return a NumPy array, potentially at the cost of copying / coercing values.
2. When your DataFrame contains a mixture of data types, DataFrame.values may involve copying data and
coercing values to a common dtype, a relatively expensive operation. DataFrame.to_numpy(), being a
method, makes it clearer that the returned NumPy array may not be a view on the same data in the DataFrame.

3.3.3 Accelerated operations

pandas has support for accelerating certain types of binary numerical and boolean operations using the numexpr
library and the bottleneck libraries.
These libraries are especially useful when dealing with large data sets, and provide large speedups. numexpr uses
smart chunking, caching, and multiple cores. bottleneck is a set of specialized cython routines that are especially
fast when dealing with arrays that have nans.
Here is a sample (using 100 column x 100,000 row DataFrames):

Operation 0.11.0 (ms) Prior Version (ms) Ratio to Prior

df1 > df2 13.32 125.35 0.1063
df1 * df2 21.71 36.63 0.5928
df1 + df2 22.04 36.50 0.6039

You are highly encouraged to install both libraries. See the section Recommended Dependencies for more installation
info.
These are both enabled to be used by default, you can control this by setting the options:
New in version 0.20.0.

pd.set_option('compute.use_bottleneck', False)
pd.set_option('compute.use_numexpr', False)

3.3.4 Flexible binary operations

With binary operations between pandas data structures, there are two key points of interest:
• Broadcasting behavior between higher- (e.g. DataFrame) and lower-dimensional (e.g. Series) objects.
• Missing data in computations.
We will demonstrate how to manage these issues independently, though they can be handled simultaneously.

Matching / broadcasting behavior

DataFrame has the methods add(), sub(), mul(), div() and related functions radd(), rsub(), for carrying
out binary operations. For broadcasting behavior, Series input is of primary interest. Using these functions, you can
use to either match on the index or columns via the axis keyword:

3.3. Essential basic functionality 41

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [18]: df = pd.DataFrame({
....: 'one': pd.Series(np.random.randn(3), index=['a', 'b', 'c']),
....: 'two': pd.Series(np.random.randn(4), index=['a', 'b', 'c', 'd']),
....: 'three': pd.Series(np.random.randn(3), index=['b', 'c', 'd'])})
....:

In [19]: df
Out[19]:
one two three
a 0.280390 0.244345 NaN
b 1.169430 0.685821 1.063425
c 1.009711 -0.081301 -1.233237
d NaN 0.589941 0.853134

In [20]: row = df.iloc[1]

In [21]: column = df['two']

In [22]: df.sub(row, axis='columns')

Out[22]:
one two three
a -0.889039 -0.441476 NaN
b 0.000000 0.000000 0.000000
c -0.159719 -0.767122 -2.296662
d NaN -0.095880 -0.210291

In [23]: df.sub(row, axis=1)

Out[23]:
one two three
a -0.889039 -0.441476 NaN
b 0.000000 0.000000 0.000000
c -0.159719 -0.767122 -2.296662
d NaN -0.095880 -0.210291

In [24]: df.sub(column, axis='index')

Out[24]:
one two three
a 0.036045 0.0 NaN
b 0.483609 0.0 0.377604
c 1.091012 0.0 -1.151936
d NaN 0.0 0.263193

In [25]: df.sub(column, axis=0)

Out[25]:
one two three
a 0.036045 0.0 NaN
b 0.483609 0.0 0.377604
c 1.091012 0.0 -1.151936
d NaN 0.0 0.263193
Furthermore you can align a level of a MultiIndexed DataFrame with a Series.

In [26]: dfmi = df.copy()

(continues on next page)

42 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [27]: dfmi.index = pd.MultiIndex.from_tuples([(1, 'a'), (1, 'b'),
....: (1, 'c'), (2, 'a')],
....: names=['first', 'second'])
....:

In [28]: dfmi.sub(column, axis=0, level='second')

Out[28]:
one two three
first second
1 a 0.036045 0.000000 NaN
b 0.483609 0.000000 0.377604
c 1.091012 0.000000 -1.151936
2 a NaN 0.345596 0.608789

Series and Index also support the divmod() builtin. This function takes the floor division and modulo operation at
the same time returning a two-tuple of the same type as the left hand side. For example:
In [29]: s = pd.Series(np.arange(10))

In [30]: s
Out[30]:
0 0
1 1
2 2
3 3
4 4
5 5
6 6
7 7
8 8
9 9
dtype: int64

In [31]: div, rem = divmod(s, 3)

In [32]: div
Out[32]:
0 0
1 0
2 0
3 1
4 1
5 1
6 2
7 2
8 2
9 3
dtype: int64

In [33]: rem
Out[33]:
0 0
1 1
2 2

3.3. Essential basic functionality 43

pandas: powerful Python data analysis toolkit, Release 0.25.3

3 0
4 1
5 2
6 0
7 1
8 2
9 0
dtype: int64

In [34]: idx = pd.Index(np.arange(10))

In [35]: idx
Out[35]: Int64Index([0, 1, 2, 3, 4, 5, 6, 7, 8, 9], dtype='int64')

In [36]: div, rem = divmod(idx, 3)

In [37]: div
Out[37]: Int64Index([0, 0, 0, 1, 1, 1, 2, 2, 2, 3], dtype='int64')

In [38]: rem
Out[38]: Int64Index([0, 1, 2, 0, 1, 2, 0, 1, 2, 0], dtype='int64')
We can also do elementwise divmod():
In [39]: div, rem = divmod(s, [2, 2, 3, 3, 4, 4, 5, 5, 6, 6])

In [40]: div
Out[40]:
0 0
1 0
2 0
3 1
4 1
5 1
6 1
7 1
8 1
9 1
dtype: int64

In [41]: rem
Out[41]:
0 0
1 1
2 2
3 0
4 0
5 1
6 1
7 2
8 2
9 3
dtype: int64

44 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Missing data / operations with fill values

In Series and DataFrame, the arithmetic functions have the option of inputting a fill_value, namely a value to substitute
when at most one of the values at a location are missing. For example, when adding two DataFrame objects, you may
wish to treat NaN as 0 unless both DataFrames are missing that value, in which case the result will be NaN (you can
later replace NaN with some other value using fillna if you wish).
In [42]: df
Out[42]:
one two three
a 0.280390 0.244345 NaN
b 1.169430 0.685821 1.063425
c 1.009711 -0.081301 -1.233237
d NaN 0.589941 0.853134

In [43]: df2
Out[43]:
one two three
a 0.280390 0.244345 1.000000
b 1.169430 0.685821 1.063425
c 1.009711 -0.081301 -1.233237
d NaN 0.589941 0.853134

In [44]: df + df2
Out[44]:
one two three
a 0.560780 0.488689 NaN
b 2.338859 1.371642 2.126849
c 2.019421 -0.162603 -2.466474
d NaN 1.179882 1.706268

In [45]: df.add(df2, fill_value=0)

Out[45]:
one two three
a 0.560780 0.488689 1.000000
b 2.338859 1.371642 2.126849
c 2.019421 -0.162603 -2.466474
d NaN 1.179882 1.706268

Flexible comparisons

Series and DataFrame have the binary comparison methods eq, ne, lt, gt, le, and ge whose behavior is analogous
to the binary arithmetic operations described above:
In [46]: df.gt(df2)
Out[46]:
one two three
a False False False
b False False False
c False False False
d False False False

In [47]: df2.ne(df)
Out[47]:

3.3. Essential basic functionality 45

pandas: powerful Python data analysis toolkit, Release 0.25.3

one two three

a False False True
b False False False
c False False False
d True False False
These operations produce a pandas object of the same type as the left-hand-side input that is of dtype bool. These
boolean objects can be used in indexing operations, see the section on Boolean indexing.

Boolean reductions

You can apply the reductions: empty, any(), all(), and bool() to provide a way to summarize a boolean result.
In [48]: (df > 0).all()
Out[48]:
one False
two False
three False
dtype: bool

In [49]: (df > 0).any()

Out[49]:
one True
two True
three True
dtype: bool
You can reduce to a final boolean value.

In [50]: (df > 0).any().any()

Out[50]: True

You can test if a pandas object is empty, via the empty property.
In [51]: df.empty
Out[51]: False

In [52]: pd.DataFrame(columns=list('ABC')).empty
Out[52]: True
To evaluate single-element pandas objects in a boolean context, use the method bool():
In [53]: pd.Series([True]).bool()
Out[53]: True

In [54]: pd.Series([False]).bool()
Out[54]: False

In [55]: pd.DataFrame([[True]]).bool()
Out[55]: True

In [56]: pd.DataFrame([[False]]).bool()
Out[56]: False

46 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Warning: You might be tempted to do the following:

>>> if df:
... pass

Or
>>> df and df2

These will both raise errors, as you are trying to compare multiple values.:
ValueError: The truth value of an array is ambiguous. Use a.empty, a.any() or a.
,→all().

See gotchas for a more detailed discussion.

Comparing if objects are equivalent

Often you may find that there is more than one way to compute the same result. As a simple example, consider df
+ df and df * 2. To test that these two computations produce the same result, given the tools shown above, you
might imagine using (df + df == df * 2).all(). But in fact, this expression is False:
In [57]: df + df == df * 2
Out[57]:
one two three
a True True False
b True True True
c True True True
d False True True

In [58]: (df + df == df * 2).all()

Out[58]:
one False
two True
three False
dtype: bool
Notice that the boolean DataFrame df + df == df * 2 contains some False values! This is because NaNs do
not compare as equals:

In [59]: np.nan == np.nan

Out[59]: False

So, NDFrames (such as Series and DataFrames) have an equals() method for testing equality, with NaNs in corre-
sponding locations treated as equal.

In [60]: (df + df).equals(df * 2)

Out[60]: True

Note that the Series or DataFrame index needs to be in the same order for equality to be True:
In [61]: df1 = pd.DataFrame({'col': ['foo', 0, np.nan]})

In [62]: df2 = pd.DataFrame({'col': [np.nan, 0, 'foo']}, index=[2, 1, 0])

In [63]: df1.equals(df2)

3.3. Essential basic functionality 47

pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[63]: False

In [64]: df1.equals(df2.sort_index())
Out[64]: True

Comparing array-like objects

You can conveniently perform element-wise comparisons when comparing a pandas data structure with a scalar value:
In [65]: pd.Series(['foo', 'bar', 'baz']) == 'foo'
Out[65]:
0 True
1 False
2 False
dtype: bool

In [66]: pd.Index(['foo', 'bar', 'baz']) == 'foo'

Out[66]: array([ True, False, False])
Pandas also handles element-wise comparisons between different array-like objects of the same length:
In [67]: pd.Series(['foo', 'bar', 'baz']) == pd.Index(['foo', 'bar', 'qux'])
Out[67]:
0 True
1 True
2 False
dtype: bool

In [68]: pd.Series(['foo', 'bar', 'baz']) == np.array(['foo', 'bar', 'qux'])

Out[68]:
0 True
1 True
2 False
dtype: bool
Trying to compare Index or Series objects of different lengths will raise a ValueError:

In [55]: pd.Series(['foo', 'bar', 'baz']) == pd.Series(['foo', 'bar'])

ValueError: Series lengths must match to compare

In [56]: pd.Series(['foo', 'bar', 'baz']) == pd.Series(['foo'])

ValueError: Series lengths must match to compare

Note that this is different from the NumPy behavior where a comparison can be broadcast:

In [69]: np.array([1, 2, 3]) == np.array([2])

Out[69]: array([False, True, False])

or it can return False if broadcasting can not be done:

In [70]: np.array([1, 2, 3]) == np.array([1, 2])

Out[70]: False

48 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Combining overlapping data sets

A problem occasionally arising is the combination of two similar data sets where values in one are preferred over the
other. An example would be two data series representing a particular economic indicator where one is considered to
be of higher quality. However, the lower quality series might extend further back in history or have more complete
data coverage. As such, we would like to combine two DataFrame objects where missing values in one DataFrame
are conditionally filled with like-labeled values from the other DataFrame. The function implementing this operation
is combine_first(), which we illustrate:
In [71]: df1 = pd.DataFrame({'A': [1., np.nan, 3., 5., np.nan],
....: 'B': [np.nan, 2., 3., np.nan, 6.]})
....:

In [72]: df2 = pd.DataFrame({'A': [5., 2., 4., np.nan, 3., 7.],

....: 'B': [np.nan, np.nan, 3., 4., 6., 8.]})
....:

In [73]: df1
Out[73]:
A B
0 1.0 NaN
1 NaN 2.0
2 3.0 3.0
3 5.0 NaN
4 NaN 6.0

In [74]: df2
Out[74]:
A B
0 5.0 NaN
1 2.0 NaN
2 4.0 3.0
3 NaN 4.0
4 3.0 6.0
5 7.0 8.0

In [75]: df1.combine_first(df2)
Out[75]:
A B
0 1.0 NaN
1 2.0 2.0
2 3.0 3.0
3 5.0 4.0
4 3.0 6.0
5 7.0 8.0

General DataFrame combine

The combine_first() method above calls the more general DataFrame.combine(). This method takes
another DataFrame and a combiner function, aligns the input DataFrame and then passes the combiner function pairs
of Series (i.e., columns whose names are the same).
So, for instance, to reproduce combine_first() as above:

3.3. Essential basic functionality 49

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [76]: def combiner(x, y):

....: return np.where(pd.isna(x), y, x)
....:

3.3.5 Descriptive statistics

There exists a large number of methods for computing descriptive statistics and other related operations on Series,
DataFrame. Most of these are aggregations (hence producing a lower-dimensional result) like sum(), mean(), and
quantile(), but some of them, like cumsum() and cumprod(), produce an object of the same size. Generally
speaking, these methods take an axis argument, just like ndarray.{sum, std, }, but the axis can be specified by name or
integer:
• Series: no axis argument needed
• DataFrame: index (axis=0, default), columns (axis=1)
For example:
In [77]: df
Out[77]:
one two three
a 0.280390 0.244345 NaN
b 1.169430 0.685821 1.063425
c 1.009711 -0.081301 -1.233237
d NaN 0.589941 0.853134

In [78]: df.mean(0)
Out[78]:
one 0.819843
two 0.359701
three 0.227774
dtype: float64

In [79]: df.mean(1)
Out[79]:
a 0.262367
b 0.972892
c -0.101609
d 0.721537
dtype: float64
All such methods have a skipna option signaling whether to exclude missing data (True by default):
In [80]: df.sum(0, skipna=False)
Out[80]:
one NaN
two 1.438805
three NaN
dtype: float64

In [81]: df.sum(axis=1, skipna=True)

Out[81]:
a 0.524735
b 2.918675
c -0.304828

50 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

d 1.443075
dtype: float64
Combined with the broadcasting / arithmetic behavior, one can describe various statistical procedures, like standard-
ization (rendering data zero mean and standard deviation 1), very concisely:

In [82]: ts_stand = (df - df.mean()) / df.std()

In [83]: ts_stand.std()
Out[83]:
one 1.0
two 1.0
three 1.0
dtype: float64

In [84]: xs_stand = df.sub(df.mean(1), axis=0).div(df.std(1), axis=0)

In [85]: xs_stand.std(1)
Out[85]:
a 1.0
b 1.0
c 1.0
d 1.0
dtype: float64

Note that methods like cumsum() and cumprod() preserve the location of NaN values. This is somewhat different
from expanding() and rolling(). For more details please see this note.

In [86]: df.cumsum()
Out[86]:
one two three
a 0.28039 0.244345 NaN
b 1.44982 0.930166 1.063425
c 2.45953 0.848864 -0.169812
d NaN 1.438805 0.683321

Here is a quick reference summary table of common functions. Each also takes an optional level parameter which
applies only if the object has a hierarchical index.

3.3. Essential basic functionality 51

pandas: powerful Python data analysis toolkit, Release 0.25.3

Function Description
count Number of non-NA observations
sum Sum of values
mean Mean of values
mad Mean absolute deviation
median Arithmetic median of values
min Minimum
max Maximum
mode Mode
abs Absolute Value
prod Product of values
std Bessel-corrected sample standard deviation
var Unbiased variance
sem Standard error of the mean
skew Sample skewness (3rd moment)
kurt Sample kurtosis (4th moment)
quantile Sample quantile (value at %)
cumsum Cumulative sum
cumprod Cumulative product
cummax Cumulative maximum
cummin Cumulative minimum

Note that by chance some NumPy methods, like mean, std, and sum, will exclude NAs on Series input by default:
In [87]: np.mean(df['one'])
Out[87]: 0.8198434389015307

In [88]: np.mean(df['one'].to_numpy())
Out[88]: nan
Series.nunique() will return the number of unique non-NA values in a Series:

In [89]: series = pd.Series(np.random.randn(500))

In [90]: series[20:500] = np.nan

In [91]: series[10:20] = 5

In [92]: series.nunique()
Out[92]: 11

Summarizing data: describe

There is a convenient describe() function which computes a variety of summary statistics about a Series or the
columns of a DataFrame (excluding NAs of course):

In [93]: series = pd.Series(np.random.randn(1000))

In [94]: series[::2] = np.nan

In [95]: series.describe()
Out[95]:
count 500.000000
(continues on next page)

52 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

mean 0.002283
std 0.975484
min -2.857251
25% -0.633802
50% -0.017143
75% 0.693621
max 3.144426
dtype: float64

In [96]: frame = pd.DataFrame(np.random.randn(1000, 5),

....: columns=['a', 'b', 'c', 'd', 'e'])
....:

In [97]: frame.iloc[::2] = np.nan

In [98]: frame.describe()
Out[98]:
a b c d e
count 500.000000 500.000000 500.000000 500.000000 500.000000
mean -0.015788 -0.044768 0.002118 0.071407 -0.006535
std 1.032578 0.988722 1.024366 0.992153 1.024544
min -3.301982 -3.304828 -4.185081 -3.080779 -3.260973
25% -0.620001 -0.756503 -0.691177 -0.564064 -0.709754
50% -0.004659 -0.076543 -0.021568 0.018575 0.057651
75% 0.630194 0.673233 0.693157 0.744467 0.762700
max 3.210008 2.340774 3.014330 3.111480 2.846670

You can select specific percentiles to include in the output:

In [99]: series.describe(percentiles=[.05, .25, .75, .95])

Out[99]:
count 500.000000
mean 0.002283
std 0.975484
min -2.857251
5% -1.595479
25% -0.633802
50% -0.017143
75% 0.693621
95% 1.571126
max 3.144426
dtype: float64

By default, the median is always included.

For a non-numerical Series object, describe() will give a simple summary of the number of unique values and
most frequently occurring values:

In [100]: s = pd.Series(['a', 'a', 'b', 'b', 'a', 'a', np.nan, 'c', 'd', 'a'])

In [101]: s.describe()
Out[101]:
count 9
unique 4
top a
freq 5
dtype: object

3.3. Essential basic functionality 53

pandas: powerful Python data analysis toolkit, Release 0.25.3

Note that on a mixed-type DataFrame object, describe() will restrict the summary to include only numerical
columns or, if none are, only categorical columns:

In [102]: frame = pd.DataFrame({'a': ['Yes', 'Yes', 'No', 'No'], 'b': range(4)})

In [103]: frame.describe()
Out[103]:
b
count 4.000000
mean 1.500000
std 1.290994
min 0.000000
25% 0.750000
50% 1.500000
75% 2.250000
max 3.000000

This behavior can be controlled by providing a list of types as include/exclude arguments. The special value
all can also be used:
In [104]: frame.describe(include=['object'])
Out[104]:
a
count 4
unique 2
top No
freq 2

In [105]: frame.describe(include=['number'])
Out[105]:
b
count 4.000000
mean 1.500000
std 1.290994
min 0.000000
25% 0.750000
50% 1.500000
75% 2.250000
max 3.000000

In [106]: frame.describe(include='all')
Out[106]:
a b
count 4 4.000000
unique 2 NaN
top No NaN
freq 2 NaN
mean NaN 1.500000
std NaN 1.290994
min NaN 0.000000
25% NaN 0.750000
50% NaN 1.500000
75% NaN 2.250000
max NaN 3.000000
That feature relies on select_dtypes. Refer to there for details about accepted inputs.

54 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Index of min/max values

The idxmin() and idxmax() functions on Series and DataFrame compute the index labels with the minimum and
maximum corresponding values:
In [107]: s1 = pd.Series(np.random.randn(5))

In [108]: s1
Out[108]:
0 -0.203845
1 0.699335
2 1.576849
3 0.013191
4 0.222092
dtype: float64

In [109]: s1.idxmin(), s1.idxmax()

Out[109]: (0, 2)

In [110]: df1 = pd.DataFrame(np.random.randn(5, 3), columns=['A', 'B', 'C'])

In [111]: df1
Out[111]:
A B C
0 -0.432209 -0.068295 0.381522
1 0.414550 -0.113251 0.718740
2 1.841494 0.060288 -1.210947
3 0.119289 -0.496862 1.078065
4 -0.288879 0.362025 -3.061166

In [112]: df1.idxmin(axis=0)
Out[112]:
A 0
B 3
C 4
dtype: int64

In [113]: df1.idxmax(axis=1)
Out[113]:
0 C
1 C
2 A
3 C
4 B
dtype: object
When there are multiple rows (or columns) matching the minimum or maximum value, idxmin() and idxmax()
return the first matching index:
In [114]: df3 = pd.DataFrame([2, 1, 1, 3, np.nan], columns=['A'],
,→index=list('edcba'))

In [115]: df3
Out[115]:
A

3.3. Essential basic functionality 55

pandas: powerful Python data analysis toolkit, Release 0.25.3

e 2.0
d 1.0
c 1.0
b 3.0
a NaN

In [116]: df3['A'].idxmin()
Out[116]: 'd'

Note: idxmin and idxmax are called argmin and argmax in NumPy.

Value counts (histogramming) / mode

The value_counts() Series method and top-level function computes a histogram of a 1D array of values. It can
also be used as a function on regular arrays:
In [117]: data = np.random.randint(0, 7, size=50)

In [118]: data
Out[118]:
array([0, 0, 2, 3, 2, 5, 4, 0, 3, 4, 6, 4, 1, 5, 6, 5, 3, 1, 3, 6, 2, 0,
0, 0, 3, 5, 5, 0, 5, 1, 3, 3, 2, 5, 6, 1, 1, 1, 0, 1, 0, 1, 5, 6,
5, 3, 0, 0, 5, 2])

In [119]: s = pd.Series(data)

In [120]: s.value_counts()
Out[120]:
0 11
5 10
3 8
1 8
6 5
2 5
4 3
dtype: int64

In [121]: pd.value_counts(data)
Out[121]:
0 11
5 10
3 8
1 8
6 5
2 5
4 3
dtype: int64
Similarly, you can get the most frequently occurring value(s) (the mode) of the values in a Series or DataFrame:

In [122]: s5 = pd.Series([1, 1, 3, 3, 3, 5, 5, 7, 7, 7])

(continues on next page)

56 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [123]: s5.mode()
Out[123]:
0 3
1 7
dtype: int64

In [124]: df5 = pd.DataFrame({"A": np.random.randint(0, 7, size=50),

.....: "B": np.random.randint(-10, 15, size=50)})
.....:

In [125]: df5.mode()
Out[125]:
A B
0 3 -5
1 6 9

Discretization and quantiling

Continuous values can be discretized using the cut() (bins based on values) and qcut() (bins based on sample
quantiles) functions:

In [126]: arr = np.random.randn(20)

In [127]: factor = pd.cut(arr, 4)

In [128]: factor
Out[128]:
[(0.883, 1.671], (-1.483, -0.692], (-0.692, 0.0955], (0.0955, 0.883], (0.883, 1.671],
,→..., (0.0955, 0.883], (-1.483, -0.692], (0.883, 1.671], (-1.483, -0.692], (0.0955,

,→0.883]]

Length: 20
Categories (4, interval[float64]): [(-1.483, -0.692] < (-0.692, 0.0955] < (0.0955, 0.
,→883] <

(0.883, 1.671]]

In [129]: factor = pd.cut(arr, [-5, -1, 0, 1, 5])

In [130]: factor
Out[130]:
[(1, 5], (-5, -1], (-1, 0], (0, 1], (1, 5], ..., (0, 1], (-1, 0], (1, 5], (-1, 0], (0,
,→ 1]]

Length: 20
Categories (4, interval[int64]): [(-5, -1] < (-1, 0] < (0, 1] < (1, 5]]

qcut() computes sample quantiles. For example, we could slice up some normally distributed data into equal-size
quartiles like so:
In [131]: arr = np.random.randn(30)

In [132]: factor = pd.qcut(arr, [0, .25, .5, .75, 1])

In [133]: factor
Out[133]:
[(-0.32, 0.378], (0.378, 1.888], (-1.054, -0.32], (-2.8859999999999997, -1.
,→054], (-0.32, 0.378], ..., (-2.8859999999999997, -1.054], (-1.054, -0.32],

3.3. Essential basic functionality 57

pandas: powerful Python data analysis toolkit, Release 0.25.3

,→(-1.054, -0.32], (-0.32, 0.378], (-2.8859999999999997, -1.054]]

Length: 30
Categories (4, interval[float64]): [(-2.8859999999999997, -1.054] < (-1.054, -
,→0.32] < (-0.32, 0.378] <

(0.378, 1.888]]

In [134]: pd.value_counts(factor)
Out[134]:
(0.378, 1.888] 8
(-2.8859999999999997, -1.054] 8
(-0.32, 0.378] 7
(-1.054, -0.32] 7
dtype: int64
We can also pass infinite values to define the bins:

In [135]: arr = np.random.randn(20)

In [136]: factor = pd.cut(arr, [-np.inf, 0, np.inf])

In [137]: factor
Out[137]:
[(-inf, 0.0], (-inf, 0.0], (0.0, inf], (-inf, 0.0], (0.0, inf], ..., (-inf, 0.0], (0.
,→0, inf], (0.0, inf], (-inf, 0.0], (0.0, inf]]

Length: 20
Categories (2, interval[float64]): [(-inf, 0.0] < (0.0, inf]]

3.3.6 Function application

To apply your own or another librarys functions to pandas objects, you should be aware of the three methods below.
The appropriate method to use depends on whether your function expects to operate on an entire DataFrame or
Series, row- or column-wise, or elementwise.
1. Tablewise Function Application: pipe()
2. Row or Column-wise Function Application: apply()
3. Aggregation API: agg() and transform()
4. Applying Elementwise Functions: applymap()

Tablewise function application

DataFrames and Series can of course just be passed into functions. However, if the function needs to be called
in a chain, consider using the pipe() method. Compare the following

# f, g, and h are functions taking and returning ``DataFrames``

>>> f(g(h(df), arg1=1), arg2=2, arg3=3)

with the equivalent

>>> (df.pipe(h)
... .pipe(g, arg1=1)
... .pipe(f, arg2=2, arg3=3))

58 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Pandas encourages the second style, which is known as method chaining. pipe makes it easy to use your own or
another librarys functions in method chains, alongside pandas methods.
In the example above, the functions f, g, and h each expected the DataFrame as the first positional argument. What
if the function you wish to apply takes its data as, say, the second argument? In this case, provide pipe with a tuple
of (callable, data_keyword). .pipe will route the DataFrame to the argument specified in the tuple.
For example, we can fit a regression using statsmodels. Their API expects a formula first and a DataFrame as the
second argument, data. We pass in the function, keyword pair (sm.ols, 'data') to pipe:

In [138]: import statsmodels.formula.api as sm

In [139]: bb = pd.read_csv('data/baseball.csv', index_col='id')

In [140]: (bb.query('h > 0')

.....: .assign(ln_h=lambda df: np.log(df.h))
.....: .pipe((sm.ols, 'data'), 'hr ~ ln_h + year + g + C(lg)')
.....: .fit()
.....: .summary()
.....: )
.....:
Out[140]:
<class 'statsmodels.iolib.summary.Summary'>
"""
OLS Regression Results
==============================================================================
Dep. Variable: hr R-squared: 0.685
Model: OLS Adj. R-squared: 0.665
Method: Least Squares F-statistic: 34.28
Date: Sat, 02 Nov 2019 Prob (F-statistic): 3.48e-15
Time: 16:04:51 Log-Likelihood: -205.92
No. Observations: 68 AIC: 421.8
Df Residuals: 63 BIC: 432.9
Df Model: 4
Covariance Type: nonrobust
===============================================================================
coef std err t P>|t| [0.025 0.975]
-------------------------------------------------------------------------------
Intercept -8484.7720 4664.146 -1.819 0.074 -1.78e+04 835.780
C(lg)[T.NL] -2.2736 1.325 -1.716 0.091 -4.922 0.375
ln_h -1.3542 0.875 -1.547 0.127 -3.103 0.395
year 4.2277 2.324 1.819 0.074 -0.417 8.872
g 0.1841 0.029 6.258 0.000 0.125 0.243
==============================================================================
Omnibus: 10.875 Durbin-Watson: 1.999
Prob(Omnibus): 0.004 Jarque-Bera (JB): 17.298
Skew: 0.537 Prob(JB): 0.000175
Kurtosis: 5.225 Cond. No. 1.49e+07
==============================================================================

Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly
,→specified.

[2] The condition number is large, 1.49e+07. This might indicate that there are
strong multicollinearity or other numerical problems.
"""

The pipe method is inspired by unix pipes and more recently dplyr and magrittr, which have introduced the popular
(%>%) (read pipe) operator for R. The implementation of pipe here is quite clean and feels right at home in python.

3.3. Essential basic functionality 59

pandas: powerful Python data analysis toolkit, Release 0.25.3

We encourage you to view the source code of pipe().

Row or column-wise function application

Arbitrary functions can be applied along the axes of a DataFrame using the apply() method, which, like the de-
scriptive statistics methods, takes an optional axis argument:
In [141]: df.apply(np.mean)
Out[141]:
one 0.819843
two 0.359701
three 0.227774
dtype: float64

In [142]: df.apply(np.mean, axis=1)

Out[142]:
a 0.262367
b 0.972892
c -0.101609
d 0.721537
dtype: float64

In [143]: df.apply(lambda x: x.max() - x.min())

Out[143]:
one 0.889039
two 0.767122
three 2.296662
dtype: float64

In [144]: df.apply(np.cumsum)
Out[144]:
one two three
a 0.28039 0.244345 NaN
b 1.44982 0.930166 1.063425
c 2.45953 0.848864 -0.169812
d NaN 1.438805 0.683321

In [145]: df.apply(np.exp)
Out[145]:
one two three
a 1.323646 1.276784 NaN
b 3.220155 1.985401 2.896273
c 2.744807 0.921916 0.291348
d NaN 1.803882 2.346990
The apply() method will also dispatch on a string method name.
In [146]: df.apply('mean')
Out[146]:
one 0.819843
two 0.359701
three 0.227774
dtype: float64

In [147]: df.apply('mean', axis=1)

60 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[147]:
a 0.262367
b 0.972892
c -0.101609
d 0.721537
dtype: float64
The return type of the function passed to apply() affects the type of the final output from DataFrame.apply for
the default behaviour:
• If the applied function returns a Series, the final output is a DataFrame. The columns match the index of
the Series returned by the applied function.
• If the applied function returns any other type, the final output is a Series.
This default behaviour can be overridden using the result_type, which accepts three options: reduce,
broadcast, and expand. These will determine how list-likes return values expand (or not) to a DataFrame.
apply() combined with some cleverness can be used to answer many questions about a data set. For example,
suppose we wanted to extract the date where the maximum value for each column occurred:

In [148]: tsdf = pd.DataFrame(np.random.randn(1000, 3), columns=['A', 'B', 'C'],

.....: index=pd.date_range('1/1/2000', periods=1000))
.....:

In [149]: tsdf.apply(lambda x: x.idxmax())

Out[149]:
A 2000-02-28
B 2001-02-03
C 2002-04-25
dtype: datetime64[ns]

You may also pass additional arguments and keyword arguments to the apply() method. For instance, consider the
following function you would like to apply:

def subtract_and_divide(x, sub, divide=1):

return (x - sub) / divide

You may then apply this function as follows:

df.apply(subtract_and_divide, args=(5,), divide=3)

Another useful feature is the ability to pass Series methods to carry out some Series operation on each column or row:
In [150]: tsdf
Out[150]:
A B C
2000-01-01 -0.482395 0.259874 -0.377157
2000-01-02 1.775855 0.614308 0.014913
2000-01-03 -1.122091 0.798609 0.259841
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 -0.543235 -0.939397 -0.534679
2000-01-09 -0.857516 -0.277678 0.631568
2000-01-10 -0.169672 -0.135461 -1.163331

In [151]: tsdf.apply(pd.Series.interpolate)

3.3. Essential basic functionality 61

pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[151]:
A B C
2000-01-01 -0.482395 0.259874 -0.377157
2000-01-02 1.775855 0.614308 0.014913
2000-01-03 -1.122091 0.798609 0.259841
2000-01-04 -1.006319 0.451007 0.100937
2000-01-05 -0.890548 0.103406 -0.057967
2000-01-06 -0.774777 -0.244195 -0.216871
2000-01-07 -0.659006 -0.591796 -0.375775
2000-01-08 -0.543235 -0.939397 -0.534679
2000-01-09 -0.857516 -0.277678 0.631568
2000-01-10 -0.169672 -0.135461 -1.163331
Finally, apply() takes an argument raw which is False by default, which converts each row or column into a Series
before applying the function. When set to True, the passed function will instead receive an ndarray object, which has
positive performance implications if you do not need the indexing functionality.

Aggregation API

New in version 0.20.0.

The aggregation API allows one to express possibly multiple aggregation operations in a single concise way. This API
is similar across pandas objects, see groupby API, the window functions API, and the resample API. The entry point
for aggregation is DataFrame.aggregate(), or the alias DataFrame.agg().
We will use a similar starting frame from above:

In [152]: tsdf = pd.DataFrame(np.random.randn(10, 3), columns=['A', 'B', 'C'],

.....: index=pd.date_range('1/1/2000', periods=10))
.....:

In [153]: tsdf.iloc[3:7] = np.nan

In [154]: tsdf
Out[154]:
A B C
2000-01-01 -1.505167 0.208159 0.666916
2000-01-02 1.198501 0.575635 0.519301
2000-01-03 0.787026 -1.428255 -0.778372
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 -0.529666 0.296110 0.843091
2000-01-09 -2.084206 0.578940 0.672280
2000-01-10 1.161593 -1.926451 -0.851579

Using a single function is equivalent to apply(). You can also pass named methods as strings. These will return a
Series of the aggregated output:
In [155]: tsdf.agg(np.sum)
Out[155]:
A -0.971919
B -1.695863
C 1.071637
dtype: float64

62 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [156]: tsdf.agg('sum')
Out[156]:
A -0.971919
B -1.695863
C 1.071637
dtype: float64

# these are equivalent to a ``.sum()`` because we are aggregating

# on a single function
In [157]: tsdf.sum()
Out[157]:
A -0.971919
B -1.695863
C 1.071637
dtype: float64
Single aggregations on a Series this will return a scalar value:

In [158]: tsdf.A.agg('sum')
Out[158]: -0.9719194560759219

Aggregating with multiple functions

You can pass multiple aggregation arguments as a list. The results of each of the passed functions will be a row in the
resulting DataFrame. These are naturally named from the aggregation function.

In [159]: tsdf.agg(['sum'])
Out[159]:
A B C
sum -0.971919 -1.695863 1.071637

Multiple functions yield multiple rows:

In [160]: tsdf.agg(['sum', 'mean'])

Out[160]:
A B C
sum -0.971919 -1.695863 1.071637
mean -0.161987 -0.282644 0.178606

On a Series, multiple functions return a Series, indexed by the function names:

In [161]: tsdf.A.agg(['sum', 'mean'])

Out[161]:
sum -0.971919
mean -0.161987
Name: A, dtype: float64

Passing a lambda function will yield a <lambda> named row:

In [162]: tsdf.A.agg(['sum', lambda x: x.mean()])

Out[162]:
sum -0.971919
<lambda> -0.161987
Name: A, dtype: float64

Passing a named function will yield that name for the row:

3.3. Essential basic functionality 63

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [163]: def mymean(x):

.....: return x.mean()
.....:

In [164]: tsdf.A.agg(['sum', mymean])

Out[164]:
sum -0.971919
mymean -0.161987
Name: A, dtype: float64

Aggregating with a dict

Passing a dictionary of column names to a scalar or a list of scalars, to DataFrame.agg allows you to customize
which functions are applied to which columns. Note that the results are not in any particular order, you can use an
OrderedDict instead to guarantee ordering.
In [165]: tsdf.agg({'A': 'mean', 'B': 'sum'})
Out[165]:
A -0.161987
B -1.695863
dtype: float64

Passing a list-like will generate a DataFrame output. You will get a matrix-like output of all of the aggregators. The
output will consist of all unique functions. Those that are not noted for a particular column will be NaN:
In [166]: tsdf.agg({'A': ['mean', 'min'], 'B': 'sum'})
Out[166]:
A B
mean -0.161987 NaN
min -2.084206 NaN
sum NaN -1.695863

Mixed dtypes

When presented with mixed dtypes that cannot aggregate, .agg will only take the valid aggregations. This is similar
to how groupby .agg works.
In [167]: mdf = pd.DataFrame({'A': [1, 2, 3],
.....: 'B': [1., 2., 3.],
.....: 'C': ['foo', 'bar', 'baz'],
.....: 'D': pd.date_range('20130101', periods=3)})
.....:

In [168]: mdf.dtypes
Out[168]:
A int64
B float64
C object
D datetime64[ns]
dtype: object

In [169]: mdf.agg(['min', 'sum'])

Out[169]:
(continues on next page)

64 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

A B C D
min 1 1.0 bar 2013-01-01
sum 6 6.0 foobarbaz NaT

Custom describe

With .agg() is it possible to easily create a custom describe function, similar to the built in describe function.
In [170]: from functools import partial

In [171]: q_25 = partial(pd.Series.quantile, q=0.25)

In [172]: q_25.__name__ = '25%'

In [173]: q_75 = partial(pd.Series.quantile, q=0.75)

In [174]: q_75.__name__ = '75%'

In [175]: tsdf.agg(['count', 'mean', 'std', 'min', q_25, 'median', q_75, 'max'])

Out[175]:
A B C
count 6.000000 6.000000 6.000000
mean -0.161987 -0.282644 0.178606
std 1.423915 1.101757 0.776770
min -2.084206 -1.926451 -0.851579
25% -1.261292 -1.019151 -0.453954
median 0.128680 0.252134 0.593108
75% 1.067951 0.505754 0.670939
max 1.198501 0.578940 0.843091

Transform API

New in version 0.20.0.

The transform() method returns an object that is indexed the same (same size) as the original. This API allows
you to provide multiple operations at the same time rather than one-by-one. Its API is quite similar to the .agg API.
We create a frame similar to the one used in the above sections.
In [176]: tsdf = pd.DataFrame(np.random.randn(10, 3), columns=['A', 'B', 'C'],
.....: index=pd.date_range('1/1/2000', periods=10))
.....:

In [177]: tsdf.iloc[3:7] = np.nan

In [178]: tsdf
Out[178]:
A B C
2000-01-01 -0.279295 0.417231 0.537995
2000-01-02 0.807265 1.387190 -0.982545
2000-01-03 -0.538076 0.192555 -1.012581
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
(continues on next page)

3.3. Essential basic functionality 65

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2000-01-07 NaN NaN NaN
2000-01-08 -1.981527 0.274648 -0.747219
2000-01-09 1.473561 0.232948 0.192742
2000-01-10 0.263927 -0.153731 0.831473

Transform the entire frame. .transform() allows input functions as: a NumPy function, a string function name or
a user defined function.
In [179]: tsdf.transform(np.abs)
Out[179]:
A B C
2000-01-01 0.279295 0.417231 0.537995
2000-01-02 0.807265 1.387190 0.982545
2000-01-03 0.538076 0.192555 1.012581
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 1.981527 0.274648 0.747219
2000-01-09 1.473561 0.232948 0.192742
2000-01-10 0.263927 0.153731 0.831473

In [180]: tsdf.transform('abs')
Out[180]:
A B C
2000-01-01 0.279295 0.417231 0.537995
2000-01-02 0.807265 1.387190 0.982545
2000-01-03 0.538076 0.192555 1.012581
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 1.981527 0.274648 0.747219
2000-01-09 1.473561 0.232948 0.192742
2000-01-10 0.263927 0.153731 0.831473

In [181]: tsdf.transform(lambda x: x.abs())

Out[181]:
A B C
2000-01-01 0.279295 0.417231 0.537995
2000-01-02 0.807265 1.387190 0.982545
2000-01-03 0.538076 0.192555 1.012581
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 1.981527 0.274648 0.747219
2000-01-09 1.473561 0.232948 0.192742
2000-01-10 0.263927 0.153731 0.831473
Here transform() received a single function; this is equivalent to a ufunc application.
In [182]: np.abs(tsdf)
Out[182]:
(continues on next page)

66 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

A B C
2000-01-01 0.279295 0.417231 0.537995
2000-01-02 0.807265 1.387190 0.982545
2000-01-03 0.538076 0.192555 1.012581
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 1.981527 0.274648 0.747219
2000-01-09 1.473561 0.232948 0.192742
2000-01-10 0.263927 0.153731 0.831473

Passing a single function to .transform() with a Series will yield a single Series in return.
In [183]: tsdf.A.transform(np.abs)
Out[183]:
2000-01-01 0.279295
2000-01-02 0.807265
2000-01-03 0.538076
2000-01-04 NaN
2000-01-05 NaN
2000-01-06 NaN
2000-01-07 NaN
2000-01-08 1.981527
2000-01-09 1.473561
2000-01-10 0.263927
Freq: D, Name: A, dtype: float64

Transform with multiple functions

Passing multiple functions will yield a column MultiIndexed DataFrame. The first level will be the original frame
column names; the second level will be the names of the transforming functions.
In [184]: tsdf.transform([np.abs, lambda x: x + 1])
Out[184]:
A B C
absolute <lambda> absolute <lambda> absolute <lambda>
2000-01-01 0.279295 0.720705 0.417231 1.417231 0.537995 1.537995
2000-01-02 0.807265 1.807265 1.387190 2.387190 0.982545 0.017455
2000-01-03 0.538076 0.461924 0.192555 1.192555 1.012581 -0.012581
2000-01-04 NaN NaN NaN NaN NaN NaN
2000-01-05 NaN NaN NaN NaN NaN NaN
2000-01-06 NaN NaN NaN NaN NaN NaN
2000-01-07 NaN NaN NaN NaN NaN NaN
2000-01-08 1.981527 -0.981527 0.274648 1.274648 0.747219 0.252781
2000-01-09 1.473561 2.473561 0.232948 1.232948 0.192742 1.192742
2000-01-10 0.263927 1.263927 0.153731 0.846269 0.831473 1.831473

Passing multiple functions to a Series will yield a DataFrame. The resulting column names will be the transforming
functions.
In [185]: tsdf.A.transform([np.abs, lambda x: x + 1])
Out[185]:
absolute <lambda>
2000-01-01 0.279295 0.720705
(continues on next page)

3.3. Essential basic functionality 67

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2000-01-02 0.807265 1.807265
2000-01-03 0.538076 0.461924
2000-01-04 NaN NaN
2000-01-05 NaN NaN
2000-01-06 NaN NaN
2000-01-07 NaN NaN
2000-01-08 1.981527 -0.981527
2000-01-09 1.473561 2.473561
2000-01-10 0.263927 1.263927

Transforming with a dict

Passing a dict of functions will allow selective transforming per column.

In [186]: tsdf.transform({'A': np.abs, 'B': lambda x: x + 1})

Out[186]:
A B
2000-01-01 0.279295 1.417231
2000-01-02 0.807265 2.387190
2000-01-03 0.538076 1.192555
2000-01-04 NaN NaN
2000-01-05 NaN NaN
2000-01-06 NaN NaN
2000-01-07 NaN NaN
2000-01-08 1.981527 1.274648
2000-01-09 1.473561 1.232948
2000-01-10 0.263927 0.846269

Passing a dict of lists will generate a MultiIndexed DataFrame with these selective transforms.

In [187]: tsdf.transform({'A': np.abs, 'B': [lambda x: x + 1, 'sqrt']})

Out[187]:
A B
absolute <lambda> sqrt
2000-01-01 0.279295 1.417231 0.645935
2000-01-02 0.807265 2.387190 1.177790
2000-01-03 0.538076 1.192555 0.438811
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 1.981527 1.274648 0.524068
2000-01-09 1.473561 1.232948 0.482647
2000-01-10 0.263927 0.846269 NaN

Applying elementwise functions

Since not all functions can be vectorized (accept NumPy arrays and return another array or value), the methods
applymap() on DataFrame and analogously map() on Series accept any Python function taking a single value and
returning a single value. For example:
In [188]: df4
Out[188]:
one two three

68 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

a 0.280390 0.244345 NaN

b 1.169430 0.685821 1.063425
c 1.009711 -0.081301 -1.233237
d NaN 0.589941 0.853134

In [189]: def f(x):

.....: return len(str(x))
.....:

In [190]: df4['one'].map(f)
Out[190]:
a 19
b 18
c 18
d 3
Name: one, dtype: int64

In [191]: df4.applymap(f)
Out[191]:
one two three
a 19 19 3
b 18 18 18
c 18 20 19
d 3 18 18
Series.map() has an additional feature; it can be used to easily link or map values defined by a secondary series.
This is closely related to merging/joining functionality:
In [192]: s = pd.Series(['six', 'seven', 'six', 'seven', 'six'],
.....: index=['a', 'b', 'c', 'd', 'e'])
.....:

In [193]: t = pd.Series({'six': 6., 'seven': 7.})

In [194]: s
Out[194]:
a six
b seven
c six
d seven
e six
dtype: object

In [195]: s.map(t)
Out[195]:
a 6.0
b 7.0
c 6.0
d 7.0
e 6.0
dtype: float64

3.3. Essential basic functionality 69

pandas: powerful Python data analysis toolkit, Release 0.25.3

3.3.7 Reindexing and altering labels

reindex() is the fundamental data alignment method in pandas. It is used to implement nearly all other features
relying on label-alignment functionality. To reindex means to conform the data to match a given set of labels along a
particular axis. This accomplishes several things:
• Reorders the existing data to match a new set of labels
• Inserts missing value (NA) markers in label locations where no data for that label existed
• If specified, fill data for missing labels using logic (highly relevant to working with time series data)
Here is a simple example:
In [196]: s = pd.Series(np.random.randn(5), index=['a', 'b', 'c', 'd', 'e'])

In [197]: s
Out[197]:
a -1.577151
b 1.165881
c 0.408422
d 1.733967
e -1.179138
dtype: float64

In [198]: s.reindex(['e', 'b', 'f', 'd'])

Out[198]:
e -1.179138
b 1.165881
f NaN
d 1.733967
dtype: float64
Here, the f label was not contained in the Series and hence appears as NaN in the result.
With a DataFrame, you can simultaneously reindex the index and columns:
In [199]: df
Out[199]:
one two three
a 0.280390 0.244345 NaN
b 1.169430 0.685821 1.063425
c 1.009711 -0.081301 -1.233237
d NaN 0.589941 0.853134

In [200]: df.reindex(index=['c', 'f', 'b'], columns=['three', 'two', 'one'])

Out[200]:
three two one
c -1.233237 -0.081301 1.009711
f NaN NaN NaN
b 1.063425 0.685821 1.169430
You may also use reindex with an axis keyword:

In [201]: df.reindex(['c', 'f', 'b'], axis='index')

Out[201]:
one two three
c 1.009711 -0.081301 -1.233237
(continues on next page)

70 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

f NaN NaN NaN
b 1.169430 0.685821 1.063425

Note that the Index objects containing the actual axis labels can be shared between objects. So if we have a Series
and a DataFrame, the following can be done:
In [202]: rs = s.reindex(df.index)

In [203]: rs
Out[203]:
a -1.577151
b 1.165881
c 0.408422
d 1.733967
dtype: float64

In [204]: rs.index is df.index

Out[204]: True
This means that the reindexed Seriess index is the same Python object as the DataFrames index.
New in version 0.21.0.
DataFrame.reindex() also supports an axis-style calling convention, where you specify a single labels argu-
ment and the axis it applies to.
In [205]: df.reindex(['c', 'f', 'b'], axis='index')
Out[205]:
one two three
c 1.009711 -0.081301 -1.233237
f NaN NaN NaN
b 1.169430 0.685821 1.063425

In [206]: df.reindex(['three', 'two', 'one'], axis='columns')

Out[206]:
three two one
a NaN 0.244345 0.280390
b 1.063425 0.685821 1.169430
c -1.233237 -0.081301 1.009711
d 0.853134 0.589941 NaN
See also:
MultiIndex / Advanced Indexing is an even more concise way of doing reindexing.

Note: When writing performance-sensitive code, there is a good reason to spend some time becoming a reindexing
ninja: many operations are faster on pre-aligned data. Adding two unaligned DataFrames internally triggers a
reindexing step. For exploratory analysis you will hardly notice the difference (because reindex has been heavily
optimized), but when CPU cycles matter sprinkling a few explicit reindex calls here and there can have an impact.

Reindexing to align with another object

You may wish to take an object and reindex its axes to be labeled the same as another object. While the syntax for this
is straightforward albeit verbose, it is a common enough operation that the reindex_like() method is available

3.3. Essential basic functionality 71

pandas: powerful Python data analysis toolkit, Release 0.25.3

to make this simpler:

In [207]: df2
Out[207]:
one two
a 0.280390 0.244345
b 1.169430 0.685821
c 1.009711 -0.081301

In [208]: df3
Out[208]:
one two
a -0.539453 -0.038610
b 0.349586 0.402866
c 0.189867 -0.364256

In [209]: df.reindex_like(df2)
Out[209]:
one two
a 0.280390 0.244345
b 1.169430 0.685821
c 1.009711 -0.081301

Aligning objects with each other with align

The align() method is the fastest way to simultaneously align two objects. It supports a join argument (related to
joining and merging):
• join='outer': take the union of the indexes (default)
• join='left': use the calling objects index
• join='right': use the passed objects index
• join='inner': intersect the indexes
It returns a tuple with both of the reindexed Series:
In [210]: s = pd.Series(np.random.randn(5), index=['a', 'b', 'c', 'd', 'e'])

In [211]: s1 = s[:4]

In [212]: s2 = s[1:]

In [213]: s1.align(s2)
Out[213]:
(a -0.673094
b 1.111754
c 0.936573
d 1.666941
e NaN
dtype: float64, a NaN
b 1.111754
c 0.936573
d 1.666941
e -0.170962
dtype: float64)

72 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [214]: s1.align(s2, join='inner')

Out[214]:
(b 1.111754
c 0.936573
d 1.666941
dtype: float64, b 1.111754
c 0.936573
d 1.666941
dtype: float64)

In [215]: s1.align(s2, join='left')

Out[215]:
(a -0.673094
b 1.111754
c 0.936573
d 1.666941
dtype: float64, a NaN
b 1.111754
c 0.936573
d 1.666941
dtype: float64)
For DataFrames, the join method will be applied to both the index and the columns by default:

In [216]: df.align(df2, join='inner')

Out[216]:
( one two
a 0.280390 0.244345
b 1.169430 0.685821
c 1.009711 -0.081301, one two
a 0.280390 0.244345
b 1.169430 0.685821
c 1.009711 -0.081301)

You can also pass an axis option to only align on the specified axis:

In [217]: df.align(df2, join='inner', axis=0)

Out[217]:
( one two three
a 0.280390 0.244345 NaN
b 1.169430 0.685821 1.063425
c 1.009711 -0.081301 -1.233237, one two
a 0.280390 0.244345
b 1.169430 0.685821
c 1.009711 -0.081301)

If you pass a Series to DataFrame.align(), you can choose to align both objects either on the DataFrames index
or columns using the axis argument:

In [218]: df.align(df2.iloc[0], axis=1)

Out[218]:
( one three two
a 0.280390 NaN 0.244345
b 1.169430 1.063425 0.685821
c 1.009711 -1.233237 -0.081301
(continues on next page)

3.3. Essential basic functionality 73

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

d NaN 0.853134 0.589941, one 0.280390
three NaN
two 0.244345
Name: a, dtype: float64)

Filling while reindexing

reindex() takes an optional parameter method which is a filling method chosen from the following table:

Method Action
pad / ffill Fill values forward
bfill / backfill Fill values backward
nearest Fill from the nearest index value

We illustrate these fill methods on a simple Series:

In [219]: rng = pd.date_range('1/3/2000', periods=8)

In [220]: ts = pd.Series(np.random.randn(8), index=rng)

In [221]: ts2 = ts[[0, 3, 6]]

In [222]: ts
Out[222]:
2000-01-03 -1.352366
2000-01-04 -0.102184
2000-01-05 -0.992602
2000-01-06 -0.324007
2000-01-07 -0.589906
2000-01-08 -1.324784
2000-01-09 1.426685
2000-01-10 0.213322
Freq: D, dtype: float64

In [223]: ts2
Out[223]:
2000-01-03 -1.352366
2000-01-06 -0.324007
2000-01-09 1.426685
dtype: float64

In [224]: ts2.reindex(ts.index)
Out[224]:
2000-01-03 -1.352366
2000-01-04 NaN
2000-01-05 NaN
2000-01-06 -0.324007
2000-01-07 NaN
2000-01-08 NaN
2000-01-09 1.426685
2000-01-10 NaN
Freq: D, dtype: float64

74 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [225]: ts2.reindex(ts.index, method='ffill')

Out[225]:
2000-01-03 -1.352366
2000-01-04 -1.352366
2000-01-05 -1.352366
2000-01-06 -0.324007
2000-01-07 -0.324007
2000-01-08 -0.324007
2000-01-09 1.426685
2000-01-10 1.426685
Freq: D, dtype: float64

In [226]: ts2.reindex(ts.index, method='bfill')

Out[226]:
2000-01-03 -1.352366
2000-01-04 -0.324007
2000-01-05 -0.324007
2000-01-06 -0.324007
2000-01-07 1.426685
2000-01-08 1.426685
2000-01-09 1.426685
2000-01-10 NaN
Freq: D, dtype: float64

In [227]: ts2.reindex(ts.index, method='nearest')

Out[227]:
2000-01-03 -1.352366
2000-01-04 -1.352366
2000-01-05 -0.324007
2000-01-06 -0.324007
2000-01-07 -0.324007
2000-01-08 1.426685
2000-01-09 1.426685
2000-01-10 1.426685
Freq: D, dtype: float64
These methods require that the indexes are ordered increasing or decreasing.
Note that the same result could have been achieved using fillna (except for method='nearest') or interpolate:

In [228]: ts2.reindex(ts.index).fillna(method='ffill')
Out[228]:
2000-01-03 -1.352366
2000-01-04 -1.352366
2000-01-05 -1.352366
2000-01-06 -0.324007
2000-01-07 -0.324007
2000-01-08 -0.324007
2000-01-09 1.426685
2000-01-10 1.426685
Freq: D, dtype: float64

reindex() will raise a ValueError if the index is not monotonically increasing or decreasing. fillna() and
interpolate() will not perform any checks on the order of the index.

3.3. Essential basic functionality 75

pandas: powerful Python data analysis toolkit, Release 0.25.3

Limits on filling while reindexing

The limit and tolerance arguments provide additional control over filling while reindexing. Limit specifies the
maximum count of consecutive matches:

In [229]: ts2.reindex(ts.index, method='ffill', limit=1)

Out[229]:
2000-01-03 -1.352366
2000-01-04 -1.352366
2000-01-05 NaN
2000-01-06 -0.324007
2000-01-07 -0.324007
2000-01-08 NaN
2000-01-09 1.426685
2000-01-10 1.426685
Freq: D, dtype: float64

In contrast, tolerance specifies the maximum distance between the index and indexer values:

In [230]: ts2.reindex(ts.index, method='ffill', tolerance='1 day')

Out[230]:
2000-01-03 -1.352366
2000-01-04 -1.352366
2000-01-05 NaN
2000-01-06 -0.324007
2000-01-07 -0.324007
2000-01-08 NaN
2000-01-09 1.426685
2000-01-10 1.426685
Freq: D, dtype: float64

Notice that when used on a DatetimeIndex, TimedeltaIndex or PeriodIndex, tolerance will coerced
into a Timedelta if possible. This allows you to specify tolerance with appropriate strings.

Dropping labels from an axis

A method closely related to reindex is the drop() function. It removes a set of labels from an axis:
In [231]: df
Out[231]:
one two three
a 0.280390 0.244345 NaN
b 1.169430 0.685821 1.063425
c 1.009711 -0.081301 -1.233237
d NaN 0.589941 0.853134

In [232]: df.drop(['a', 'd'], axis=0)

Out[232]:
one two three
b 1.169430 0.685821 1.063425
c 1.009711 -0.081301 -1.233237

In [233]: df.drop(['one'], axis=1)

Out[233]:
two three
a 0.244345 NaN

76 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

b 0.685821 1.063425
c -0.081301 -1.233237
d 0.589941 0.853134
Note that the following also works, but is a bit less obvious / clean:

In [234]: df.reindex(df.index.difference(['a', 'd']))

Out[234]:
one two three
b 1.169430 0.685821 1.063425
c 1.009711 -0.081301 -1.233237

Renaming / mapping labels

The rename() method allows you to relabel an axis based on some mapping (a dict or Series) or an arbitrary function.
In [235]: s
Out[235]:
a -0.673094
b 1.111754
c 0.936573
d 1.666941
e -0.170962
dtype: float64

In [236]: s.rename(str.upper)
Out[236]:
A -0.673094
B 1.111754
C 0.936573
D 1.666941
E -0.170962
dtype: float64
If you pass a function, it must return a value when called with any of the labels (and must produce a set of unique
values). A dict or Series can also be used:

In [237]: df.rename(columns={'one': 'foo', 'two': 'bar'},

.....: index={'a': 'apple', 'b': 'banana', 'd': 'durian'})
.....:
Out[237]:
foo bar three
apple 0.280390 0.244345 NaN
banana 1.169430 0.685821 1.063425
c 1.009711 -0.081301 -1.233237
durian NaN 0.589941 0.853134

If the mapping doesnt include a column/index label, it isnt renamed. Note that extra labels in the mapping dont throw
an error.
New in version 0.21.0.
DataFrame.rename() also supports an axis-style calling convention, where you specify a single mapper and the
axis to apply that mapping to.
In [238]: df.rename({'one': 'foo', 'two': 'bar'}, axis='columns')
Out[238]:

3.3. Essential basic functionality 77

pandas: powerful Python data analysis toolkit, Release 0.25.3

foo bar three

a 0.280390 0.244345 NaN
b 1.169430 0.685821 1.063425
c 1.009711 -0.081301 -1.233237
d NaN 0.589941 0.853134

In [239]: df.rename({'a': 'apple', 'b': 'banana', 'd': 'durian'},

,→axis='index')

Out[239]:
one two three
apple 0.280390 0.244345 NaN
banana 1.169430 0.685821 1.063425
c 1.009711 -0.081301 -1.233237
durian NaN 0.589941 0.853134
The rename() method also provides an inplace named parameter that is by default False and copies the under-
lying data. Pass inplace=True to rename the data in place.
New in version 0.18.0.
Finally, rename() also accepts a scalar or list-like for altering the Series.name attribute.

In [240]: s.rename("scalar-name")
Out[240]:
a -0.673094
b 1.111754
c 0.936573
d 1.666941
e -0.170962
Name: scalar-name, dtype: float64

New in version 0.24.0.

The methods rename_axis() and rename_axis() allow specific names of a MultiIndex to be changed (as
opposed to the labels).
In [241]: df = pd.DataFrame({'x': [1, 2, 3, 4, 5, 6],
.....: 'y': [10, 20, 30, 40, 50, 60]},
.....: index=pd.MultiIndex.from_product([['a', 'b', 'c'],
,→ [1, 2]],

.....: names=['let', 'num']))

.....:

In [242]: df
Out[242]:
x y
let num
a 1 1 10
2 2 20
b 1 3 30
2 4 40
c 1 5 50
2 6 60

In [243]: df.rename_axis(index={'let': 'abc'})

Out[243]:
x y

78 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

abc num
a 1 1 10
2 2 20
b 1 3 30
2 4 40
c 1 5 50
2 6 60

In [244]: df.rename_axis(index=str.upper)
Out[244]:
x y
LET NUM
a 1 1 10
2 2 20
b 1 3 30
2 4 40
c 1 5 50
2 6 60

3.3.8 Iteration

The behavior of basic iteration over pandas objects depends on the type. When iterating over a Series, it is regarded
as array-like, and basic iteration produces the values. DataFrames follow the dict-like convention of iterating over the
keys of the objects.
In short, basic iteration (for i in object) produces:
• Series: values
• DataFrame: column labels
Thus, for example, iterating over a DataFrame gives you the column names:

In [245]: df = pd.DataFrame({'col1': np.random.randn(3),

.....: 'col2': np.random.randn(3)}, index=['a', 'b', 'c'])
.....:

In [246]: for col in df:

.....: print(col)
.....:
col1
col2

Pandas objects also have the dict-like items() method to iterate over the (key, value) pairs.
To iterate over the rows of a DataFrame, you can use the following methods:
• iterrows(): Iterate over the rows of a DataFrame as (index, Series) pairs. This converts the rows to Series
objects, which can change the dtypes and has some performance implications.
• itertuples(): Iterate over the rows of a DataFrame as namedtuples of the values. This is a lot faster than
iterrows(), and is in most cases preferable to use to iterate over the values of a DataFrame.

Warning: Iterating through pandas objects is generally slow. In many cases, iterating manually over the rows is
not needed and can be avoided with one of the following approaches:

3.3. Essential basic functionality 79

pandas: powerful Python data analysis toolkit, Release 0.25.3

• Look for a vectorized solution: many operations can be performed using built-in methods or NumPy func-
tions, (boolean) indexing,
• When you have a function that cannot work on the full DataFrame/Series at once, it is better to use apply()
instead of iterating over the values. See the docs on function application.
• If you need to do iterative manipulations on the values but performance is important, consider writing the in-
ner loop with cython or numba. See the enhancing performance section for some examples of this approach.

Warning: You should never modify something you are iterating over. This is not guaranteed to work in all cases.
Depending on the data types, the iterator returns a copy and not a view, and writing to it will have no effect!
For example, in the following case setting the value has no effect:
In [247]: df = pd.DataFrame({'a': [1, 2, 3], 'b': ['a', 'b', 'c']})

In [248]: for index, row in df.iterrows():

.....: row['a'] = 10
.....:

In [249]: df
Out[249]:
a b
0 1 a
1 2 b
2 3 c

items

Consistent with the dict-like interface, items() iterates through key-value pairs:
• Series: (index, scalar value) pairs
• DataFrame: (column, Series) pairs
For example:

In [250]: for label, ser in df.items():

.....: print(label)
.....: print(ser)
.....:
a
0 1
1 2
2 3
Name: a, dtype: int64
b
0 a
1 b
2 c
Name: b, dtype: object

80 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

iterrows

iterrows() allows you to iterate through the rows of a DataFrame as Series objects. It returns an iterator yielding
each index value along with a Series containing the data in each row:

In [251]: for row_index, row in df.iterrows():

.....: print(row_index, row, sep='\n')
.....:
0
a 1
b a
Name: 0, dtype: object
1
a 2
b b
Name: 1, dtype: object
2
a 3
b c
Name: 2, dtype: object

Note: Because iterrows() returns a Series for each row, it does not preserve dtypes across the rows (dtypes are
preserved across columns for DataFrames). For example,

In [252]: df_orig = pd.DataFrame([[1, 1.5]], columns=['int', 'float'])

In [253]: df_orig.dtypes
Out[253]:
int int64
float float64
dtype: object

In [254]: row = next(df_orig.iterrows())[1]

In [255]: row
Out[255]:
int 1.0
float 1.5
Name: 0, dtype: float64

All values in row, returned as a Series, are now upcasted to floats, also the original integer value in column x:
In [256]: row['int'].dtype
Out[256]: dtype('float64')

In [257]: df_orig['int'].dtype
Out[257]: dtype('int64')
To preserve dtypes while iterating over the rows, it is better to use itertuples() which returns namedtuples of the
values and which is generally much faster than iterrows().

For instance, a contrived way to transpose the DataFrame would be:

In [258]: df2 = pd.DataFrame({'x': [1, 2, 3], 'y': [4, 5, 6]})

In [259]: print(df2)
x y

3.3. Essential basic functionality 81

pandas: powerful Python data analysis toolkit, Release 0.25.3

0 1 4
1 2 5
2 3 6

In [260]: print(df2.T)
0 1 2
x 1 2 3
y 4 5 6

In [261]: df2_t = pd.DataFrame({idx: values for idx, values in df2.iterrows()}

,→)

In [262]: print(df2_t)
0 1 2
x 1 2 3
y 4 5 6

itertuples

The itertuples() method will return an iterator yielding a namedtuple for each row in the DataFrame. The first
element of the tuple will be the rows corresponding index value, while the remaining values are the row values.
For instance:

In [263]: for row in df.itertuples():

.....: print(row)
.....:
Pandas(Index=0, a=1, b='a')
Pandas(Index=1, a=2, b='b')
Pandas(Index=2, a=3, b='c')

This method does not convert the row to a Series object; it merely returns the values inside a namedtuple. Therefore,
itertuples() preserves the data type of the values and is generally faster as iterrows().

Note: The column names will be renamed to positional names if they are invalid Python identifiers, repeated, or start
with an underscore. With a large number of columns (>255), regular tuples are returned.

3.3.9 .dt accessor

Series has an accessor to succinctly return datetime like properties for the values of the Series, if it is a date-
time/period like Series. This will return a Series, indexed like the existing Series.
# datetime
In [264]: s = pd.Series(pd.date_range('20130101 09:10:12', periods=4))

In [265]: s
Out[265]:
0 2013-01-01 09:10:12
1 2013-01-02 09:10:12
2 2013-01-03 09:10:12
3 2013-01-04 09:10:12
dtype: datetime64[ns]

82 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [266]: s.dt.hour
Out[266]:
0 9
1 9
2 9
3 9
dtype: int64

In [267]: s.dt.second
Out[267]:
0 12
1 12
2 12
3 12
dtype: int64

In [268]: s.dt.day
Out[268]:
0 1
1 2
2 3
3 4
dtype: int64
This enables nice expressions like this:

In [269]: s[s.dt.day == 2]
Out[269]:
1 2013-01-02 09:10:12
dtype: datetime64[ns]

You can easily produces tz aware transformations:

In [270]: stz = s.dt.tz_localize('US/Eastern')

In [271]: stz
Out[271]:
0 2013-01-01 09:10:12-05:00
1 2013-01-02 09:10:12-05:00
2 2013-01-03 09:10:12-05:00
3 2013-01-04 09:10:12-05:00
dtype: datetime64[ns, US/Eastern]

In [272]: stz.dt.tz
Out[272]: <DstTzInfo 'US/Eastern' LMT-1 day, 19:04:00 STD>
You can also chain these types of operations:

In [273]: s.dt.tz_localize('UTC').dt.tz_convert('US/Eastern')
Out[273]:
0 2013-01-01 04:10:12-05:00
1 2013-01-02 04:10:12-05:00
2 2013-01-03 04:10:12-05:00
3 2013-01-04 04:10:12-05:00
dtype: datetime64[ns, US/Eastern]

You can also format datetime values as strings with Series.dt.strftime() which supports the same format as

3.3. Essential basic functionality 83

pandas: powerful Python data analysis toolkit, Release 0.25.3

the standard strftime().

# DatetimeIndex
In [274]: s = pd.Series(pd.date_range('20130101', periods=4))

In [275]: s
Out[275]:
0 2013-01-01
1 2013-01-02
2 2013-01-03
3 2013-01-04
dtype: datetime64[ns]

In [276]: s.dt.strftime('%Y/%m/%d')
Out[276]:
0 2013/01/01
1 2013/01/02
2 2013/01/03
3 2013/01/04
dtype: object
# PeriodIndex
In [277]: s = pd.Series(pd.period_range('20130101', periods=4))

In [278]: s
Out[278]:
0 2013-01-01
1 2013-01-02
2 2013-01-03
3 2013-01-04
dtype: period[D]

In [279]: s.dt.strftime('%Y/%m/%d')
Out[279]:
0 2013/01/01
1 2013/01/02
2 2013/01/03
3 2013/01/04
dtype: object
The .dt accessor works for period and timedelta dtypes.
# period
In [280]: s = pd.Series(pd.period_range('20130101', periods=4, freq='D'))

In [281]: s
Out[281]:
0 2013-01-01
1 2013-01-02
2 2013-01-03
3 2013-01-04
dtype: period[D]

In [282]: s.dt.year
Out[282]:
0 2013

84 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

1 2013
2 2013
3 2013
dtype: int64

In [283]: s.dt.day
Out[283]:
0 1
1 2
2 3
3 4
dtype: int64
# timedelta
In [284]: s = pd.Series(pd.timedelta_range('1 day 00:00:05', periods=4,
,→freq='s'))

In [285]: s
Out[285]:
0 1 days 00:00:05
1 1 days 00:00:06
2 1 days 00:00:07
3 1 days 00:00:08
dtype: timedelta64[ns]

In [286]: s.dt.days
Out[286]:
0 1
1 1
2 1
3 1
dtype: int64

In [287]: s.dt.seconds
Out[287]:
0 5
1 6
2 7
3 8
dtype: int64

In [288]: s.dt.components
Out[288]:
days hours minutes seconds milliseconds microseconds nanoseconds
0 1 0 0 5 0 0 0
1 1 0 0 6 0 0 0
2 1 0 0 7 0 0 0
3 1 0 0 8 0 0 0

Note: Series.dt will raise a TypeError if you access with a non-datetime-like values.

3.3. Essential basic functionality 85

pandas: powerful Python data analysis toolkit, Release 0.25.3

3.3.10 Vectorized string methods

Series is equipped with a set of string processing methods that make it easy to operate on each element of the array.
Perhaps most importantly, these methods exclude missing/NA values automatically. These are accessed via the Seriess
str attribute and generally have names matching the equivalent (scalar) built-in string methods. For example:

In [289]: s = pd.Series(['A', 'B', 'C', 'Aaba', 'Baca', np.nan, 'CABA', 'dog

,→', 'cat'])

In [290]: s.str.lower()
Out[290]:
0 a
1 b
2 c
3 aaba
4 baca
5 NaN
6 caba
7 dog
8 cat
dtype: object

Powerful pattern-matching methods are provided as well, but note that pattern-matching generally uses regular expres-
sions by default (and in some cases always uses them).
Please see Vectorized String Methods for a complete description.

3.3.11 Sorting

Pandas supports three kinds of sorting: sorting by index labels, sorting by column values, and sorting by a combination
of both.

By index

The Series.sort_index() and DataFrame.sort_index() methods are used to sort a pandas object by its
index levels.
In [291]: df = pd.DataFrame({
.....: 'one': pd.Series(np.random.randn(3), index=['a', 'b', 'c']),
.....: 'two': pd.Series(np.random.randn(4), index=['a', 'b', 'c',
,→'d']),

.....: 'three': pd.Series(np.random.randn(3), index=['b', 'c', 'd'])})

.....:

In [292]: unsorted_df = df.reindex(index=['a', 'd', 'c', 'b'],

.....: columns=['three', 'two', 'one'])
.....:

In [293]: unsorted_df
Out[293]:
three two one
a NaN -0.910346 1.695397
d 1.795269 -0.834629 NaN
c -0.228802 0.349813 -0.668278
b -0.907416 0.713418 0.349313

86 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

# DataFrame
In [294]: unsorted_df.sort_index()
Out[294]:
three two one
a NaN -0.910346 1.695397
b -0.907416 0.713418 0.349313
c -0.228802 0.349813 -0.668278
d 1.795269 -0.834629 NaN

In [295]: unsorted_df.sort_index(ascending=False)
Out[295]:
three two one
d 1.795269 -0.834629 NaN
c -0.228802 0.349813 -0.668278
b -0.907416 0.713418 0.349313
a NaN -0.910346 1.695397

In [296]: unsorted_df.sort_index(axis=1)
Out[296]:
one three two
a 1.695397 NaN -0.910346
d NaN 1.795269 -0.834629
c -0.668278 -0.228802 0.349813
b 0.349313 -0.907416 0.713418

# Series
In [297]: unsorted_df['three'].sort_index()
Out[297]:
a NaN
b -0.907416
c -0.228802
d 1.795269
Name: three, dtype: float64

By values

The Series.sort_values() method is used to sort a Series by its values. The DataFrame.sort_values()
method is used to sort a DataFrame by its column or row values. The optional by parameter to DataFrame.
sort_values() may used to specify one or more columns to use to determine the sorted order.
In [298]: df1 = pd.DataFrame({'one': [2, 1, 1, 1],
.....: 'two': [1, 3, 2, 4],
.....: 'three': [5, 4, 3, 2]})
.....:

In [299]: df1.sort_values(by='two')
Out[299]:
one two three
0 2 1 5
2 1 2 3
1 1 3 4
3 1 4 2

The by parameter can take a list of column names, e.g.:

3.3. Essential basic functionality 87

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [300]: df1[['one', 'two', 'three']].sort_values(by=['one', 'two'])

Out[300]:
one two three
2 1 2 3
1 1 3 4
3 1 4 2
0 2 1 5

These methods have special treatment of NA values via the na_position argument:
In [301]: s[2] = np.nan

In [302]: s.sort_values()
Out[302]:
0 A
3 Aaba
1 B
4 Baca
6 CABA
8 cat
7 dog
2 NaN
5 NaN
dtype: object

In [303]: s.sort_values(na_position='first')
Out[303]:
2 NaN
5 NaN
0 A
3 Aaba
1 B
4 Baca
6 CABA
8 cat
7 dog
dtype: object

By indexes and values

New in version 0.23.0.

Strings passed as the by parameter to DataFrame.sort_values() may refer to either columns or index level
names.

# Build MultiIndex
In [304]: idx = pd.MultiIndex.from_tuples([('a', 1), ('a', 2), ('a', 2),
.....: ('b', 2), ('b', 1), ('b', 1)])
.....:

In [305]: idx.names = ['first', 'second']

# Build DataFrame
In [306]: df_multi = pd.DataFrame({'A': np.arange(6, 0, -1)},
(continues on next page)

88 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

.....: index=idx)
.....:

In [307]: df_multi
Out[307]:
A
first second
a 1 6
2 5
2 4
b 2 3
1 2
1 1

Sort by second (index) and A (column)

In [308]: df_multi.sort_values(by=['second', 'A'])

Out[308]:
A
first second
b 1 1
1 2
a 1 6
b 2 3
a 2 4
2 5

Note: If a string matches both a column name and an index level name then a warning is issued and the column takes
precedence. This will result in an ambiguity error in a future version.

searchsorted

Series has the searchsorted() method, which works similarly to numpy.ndarray.searchsorted().

In [309]: ser = pd.Series([1, 2, 3])

In [310]: ser.searchsorted([0, 3])

Out[310]: array([0, 2])

In [311]: ser.searchsorted([0, 4])

Out[311]: array([0, 3])

In [312]: ser.searchsorted([1, 3], side='right')

Out[312]: array([1, 3])

In [313]: ser.searchsorted([1, 3], side='left')

Out[313]: array([0, 2])

In [314]: ser = pd.Series([3, 1, 2])

In [315]: ser.searchsorted([0, 3], sorter=np.argsort(ser))

Out[315]: array([0, 2])

3.3. Essential basic functionality 89

pandas: powerful Python data analysis toolkit, Release 0.25.3

smallest / largest values

Series has the nsmallest() and nlargest() methods which return the smallest or largest n values. For a
large Series this can be much faster than sorting the entire Series and calling head(n) on the result.
In [316]: s = pd.Series(np.random.permutation(10))

In [317]: s
Out[317]:
0 4
1 7
2 8
3 1
4 9
5 3
6 6
7 0
8 2
9 5
dtype: int64

In [318]: s.sort_values()
Out[318]:
7 0
3 1
8 2
5 3
0 4
9 5
6 6
1 7
2 8
4 9
dtype: int64

In [319]: s.nsmallest(3)
Out[319]:
7 0
3 1
8 2
dtype: int64

In [320]: s.nlargest(3)
Out[320]:
4 9
2 8
1 7
dtype: int64
DataFrame also has the nlargest and nsmallest methods.
In [321]: df = pd.DataFrame({'a': [-2, -1, 1, 10, 8, 11, -1],
.....: 'b': list('abdceff'),
.....: 'c': [1.0, 2.0, 4.0, 3.2, np.nan, 3.0, 4.0]})
.....:

90 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [322]: df.nlargest(3, 'a')

Out[322]:
a b c
5 11 f 3.0
3 10 c 3.2
4 8 e NaN

In [323]: df.nlargest(5, ['a', 'c'])

Out[323]:
a b c
5 11 f 3.0
3 10 c 3.2
4 8 e NaN
2 1 d 4.0
6 -1 f 4.0

In [324]: df.nsmallest(3, 'a')

Out[324]:
a b c
0 -2 a 1.0
1 -1 b 2.0
6 -1 f 4.0

In [325]: df.nsmallest(5, ['a', 'c'])

Out[325]:
a b c
0 -2 a 1.0
1 -1 b 2.0
6 -1 f 4.0
2 1 d 4.0
4 8 e NaN

Sorting by a MultiIndex column

You must be explicit about sorting when the column is a MultiIndex, and fully specify all levels to by.

In [326]: df1.columns = pd.MultiIndex.from_tuples([('a', 'one'),

.....: ('a', 'two'),
.....: ('b', 'three')])
.....:

In [327]: df1.sort_values(by=('a', 'two'))

Out[327]:
a b
one two three
0 2 1 5
2 1 2 3
1 1 3 4
3 1 4 2

3.3. Essential basic functionality 91

pandas: powerful Python data analysis toolkit, Release 0.25.3

3.3.12 Copying

The copy() method on pandas objects copies the underlying data (though not the axis indexes, since they are im-
mutable) and returns a new object. Note that it is seldom necessary to copy objects. For example, there are only a
handful of ways to alter a DataFrame in-place:
• Inserting, deleting, or modifying a column.
• Assigning to the index or columns attributes.
• For homogeneous data, directly modifying the values via the values attribute or advanced indexing.
To be clear, no pandas method has the side effect of modifying your data; almost every method returns a new object,
leaving the original object untouched. If the data is modified, it is because you did so explicitly.

3.3.13 dtypes

For the most part, pandas uses NumPy arrays and dtypes for Series or individual columns of a DataFrame. NumPy
provides support for float, int, bool, timedelta64[ns] and datetime64[ns] (note that NumPy does not
support timezone-aware datetimes).
Pandas and third-party libraries extend NumPys type system in a few places. This section describes the extensions
pandas has made internally. See Extension types for how to write your own extension that works with pandas. See
Extension data types for a list of third-party libraries that have implemented an extension.
The following table lists all of pandas extension types. See the respective documentation sections for more on each
type.

Kind of Data Data Type Scalar Array Documentation

tz-aware date- DatetimeTZDtype Timestamp arrays. Time zone handling
time DatetimeArray
Categorical CategoricalDtype (none) Categorical categorical
period (time PeriodDtype Period arrays.PeriodArray Time span representa-
spans) tion
sparse SparseDtype (none) arrays.SparseArray sparse
intervals IntervalDtype Interval arrays. IntervalIndex
IntervalArray
nullable integer Int64Dtype, (none) arrays. Nullable integer data
IntegerArray type

Pandas uses the object dtype for storing strings.

Finally, arbitrary objects may be stored using the object dtype, but should be avoided to the extent possible (for
performance and interoperability with other libraries and methods. See object conversion).
A convenient dtypes attribute for DataFrame returns a Series with the data type of each column.
In [328]: dft = pd.DataFrame({'A': np.random.rand(3),
.....: 'B': 1,
.....: 'C': 'foo',
.....: 'D': pd.Timestamp('20010102'),
.....: 'E': pd.Series([1.0] * 3).astype('float32'),
.....: 'F': False,
.....: 'G': pd.Series([1] * 3, dtype='int8')})
.....:

In [329]: dft

92 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[329]:
A B C D E F G
0 0.409005 1 foo 2001-01-02 1.0 False 1
1 0.620879 1 foo 2001-01-02 1.0 False 1
2 0.760605 1 foo 2001-01-02 1.0 False 1

In [330]: dft.dtypes
Out[330]:
A float64
B int64
C object
D datetime64[ns]
E float32
F bool
G int8
dtype: object
On a Series object, use the dtype attribute.
In [331]: dft['A'].dtype
Out[331]: dtype('float64')

If a pandas object contains data with multiple dtypes in a single column, the dtype of the column will be chosen to
accommodate all of the data types (object is the most general).
# these ints are coerced to floats
In [332]: pd.Series([1, 2, 3, 4, 5, 6.])
Out[332]:
0 1.0
1 2.0
2 3.0
3 4.0
4 5.0
5 6.0
dtype: float64

# string data forces an ``object`` dtype

In [333]: pd.Series([1, 2, 3, 6., 'foo'])
Out[333]:
0 1
1 2
2 3
3 6
4 foo
dtype: object
The number of columns of each type in a DataFrame can be found by calling DataFrame.dtypes.
value_counts().
In [334]: dft.dtypes.value_counts()
Out[334]:
datetime64[ns] 1
object 1
int8 1
bool 1
float64 1
(continues on next page)

3.3. Essential basic functionality 93

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

int64 1
float32 1
dtype: int64

Numeric dtypes will propagate and can coexist in DataFrames. If a dtype is passed (either directly via the dtype
keyword, a passed ndarray, or a passed Series, then it will be preserved in DataFrame operations. Furthermore,
different numeric dtypes will NOT be combined. The following example will give you a taste.
In [335]: df1 = pd.DataFrame(np.random.randn(8, 1), columns=['A'],
,→dtype='float32')

In [336]: df1
Out[336]:
A
0 1.289112
1 -0.073190
2 -0.302878
3 1.074794
4 0.771071
5 -0.806465
6 0.145373
7 -0.325009

In [337]: df1.dtypes
Out[337]:
A float32
dtype: object

In [338]: df2 = pd.DataFrame({'A': pd.Series(np.random.randn(8),

,→dtype='float16'),

.....: 'B': pd.Series(np.random.randn(8)),

.....: 'C': pd.Series(np.array(np.random.randn(8),
.....: dtype='uint8'))})
.....:

In [339]: df2
Out[339]:
A B C
0 -0.530273 1.006054 1
1 0.092712 -0.639761 0
2 0.259521 -0.577626 1
3 0.835449 -0.484128 0
4 0.478760 1.201647 0
5 -0.703613 -1.065766 0
6 0.287598 0.098521 0
7 0.812988 -1.754450 0

In [340]: df2.dtypes
Out[340]:
A float16
B float64
C uint8
dtype: object

94 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

defaults

By default integer types are int64 and float types are float64, regardless of platform (32-bit or 64-bit). The
following will all result in int64 dtypes.
In [341]: pd.DataFrame([1, 2], columns=['a']).dtypes
Out[341]:
a int64
dtype: object

In [342]: pd.DataFrame({'a': [1, 2]}).dtypes

Out[342]:
a int64
dtype: object

In [343]: pd.DataFrame({'a': 1}, index=list(range(2))).dtypes

Out[343]:
a int64
dtype: object
Note that Numpy will choose platform-dependent types when creating arrays. The following WILL result in int32
on 32-bit platform.

In [344]: frame = pd.DataFrame(np.array([1, 2]))

upcasting

Types can potentially be upcasted when combined with other types, meaning they are promoted from the current type
(e.g. int to float).
In [345]: df3 = df1.reindex_like(df2).fillna(value=0.0) + df2

In [346]: df3
Out[346]:
A B C
0 0.758839 1.006054 1.0
1 0.019523 -0.639761 0.0
2 -0.043356 -0.577626 1.0
3 1.910244 -0.484128 0.0
4 1.249831 1.201647 0.0
5 -1.510078 -1.065766 0.0
6 0.432970 0.098521 0.0
7 0.487979 -1.754450 0.0

In [347]: df3.dtypes
Out[347]:
A float32
B float64
C float64
dtype: object
DataFrame.to_numpy() will return the lower-common-denominator of the dtypes, meaning the dtype that can
accommodate ALL of the types in the resulting homogeneous dtyped NumPy array. This can force some upcasting.

3.3. Essential basic functionality 95

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [348]: df3.to_numpy().dtype
Out[348]: dtype('float64')

astype

You can use the astype() method to explicitly convert dtypes from one to another. These will by default return a
copy, even if the dtype was unchanged (pass copy=False to change this behavior). In addition, they will raise an
exception if the astype operation is invalid.
Upcasting is always according to the numpy rules. If two different dtypes are involved in an operation, then the more
general one will be used as the result of the operation.
In [349]: df3
Out[349]:
A B C
0 0.758839 1.006054 1.0
1 0.019523 -0.639761 0.0
2 -0.043356 -0.577626 1.0
3 1.910244 -0.484128 0.0
4 1.249831 1.201647 0.0
5 -1.510078 -1.065766 0.0
6 0.432970 0.098521 0.0
7 0.487979 -1.754450 0.0

In [350]: df3.dtypes
Out[350]:
A float32
B float64
C float64
dtype: object

# conversion of dtypes
In [351]: df3.astype('float32').dtypes
Out[351]:
A float32
B float32
C float32
dtype: object
Convert a subset of columns to a specified type using astype().
In [352]: dft = pd.DataFrame({'a': [1, 2, 3], 'b': [4, 5, 6], 'c': [7, 8, 9]})

In [353]: dft[['a', 'b']] = dft[['a', 'b']].astype(np.uint8)

In [354]: dft
Out[354]:
a b c
0 1 4 7
1 2 5 8
2 3 6 9

In [355]: dft.dtypes
Out[355]:

96 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

a uint8
b uint8
c int64
dtype: object
New in version 0.19.0.
Convert certain columns to a specific dtype by passing a dict to astype().
In [356]: dft1 = pd.DataFrame({'a': [1, 0, 1], 'b': [4, 5, 6], 'c': [7, 8, 9]}
,→)

In [357]: dft1 = dft1.astype({'a': np.bool, 'c': np.float64})

In [358]: dft1
Out[358]:
a b c
0 True 4 7.0
1 False 5 8.0
2 True 6 9.0

In [359]: dft1.dtypes
Out[359]:
a bool
b int64
c float64
dtype: object

Note: When trying to convert a subset of columns to a specified type using astype() and loc(), upcasting occurs.
loc() tries to fit in what we are assigning to the current dtypes, while [] will overwrite them taking the dtype from
the right hand side. Therefore the following piece of code produces the unintended result.

In [360]: dft = pd.DataFrame({'a': [1, 2, 3], 'b': [4, 5, 6], 'c': [7, 8, 9]})

In [361]: dft.loc[:, ['a', 'b']].astype(np.uint8).dtypes

Out[361]:
a uint8
b uint8
dtype: object

In [362]: dft.loc[:, ['a', 'b']] = dft.loc[:, ['a', 'b']].astype(np.uint8)

In [363]: dft.dtypes
Out[363]:
a int64
b int64
c int64
dtype: object

object conversion

pandas offers various functions to try to force conversion of types from the object dtype to other types. In cases
where the data is already of the correct type, but stored in an object array, the DataFrame.infer_objects()
and Series.infer_objects() methods can be used to soft convert to the correct type.

3.3. Essential basic functionality 97

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [364]: import datetime

In [365]: df = pd.DataFrame([[1, 2],

.....: ['a', 'b'],
.....: [datetime.datetime(2016, 3, 2),
.....: datetime.datetime(2016, 3, 2)]])
.....:

In [366]: df = df.T

In [367]: df
Out[367]:
0 1 2
0 1 a 2016-03-02
1 2 b 2016-03-02

In [368]: df.dtypes
Out[368]:
0 object
1 object
2 datetime64[ns]
dtype: object
Because the data was transposed the original inference stored all columns as object, which infer_objects will
correct.
In [369]: df.infer_objects().dtypes
Out[369]:
0 int64
1 object
2 datetime64[ns]
dtype: object

The following functions are available for one dimensional object arrays or scalars to perform hard conversion of objects
to a specified type:
• to_numeric() (conversion to numeric dtypes)
In [370]: m = ['1.1', 2, 3]

In [371]: pd.to_numeric(m)
Out[371]: array([1.1, 2. , 3. ])

• to_datetime() (conversion to datetime objects)

In [372]: import datetime

In [373]: m = ['2016-07-09', datetime.datetime(2016, 3, 2)]

In [374]: pd.to_datetime(m)
Out[374]: DatetimeIndex(['2016-07-09', '2016-03-02'], dtype='datetime64[ns]',
,→freq=None)

• to_timedelta() (conversion to timedelta objects)

In [375]: m = ['5us', pd.Timedelta('1day')]

(continues on next page)

98 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [376]: pd.to_timedelta(m)
Out[376]: TimedeltaIndex(['0 days 00:00:00.000005', '1 days 00:00:00'], dtype=
,→'timedelta64[ns]', freq=None)

To force a conversion, we can pass in an errors argument, which specifies how pandas should deal with elements
that cannot be converted to desired dtype or object. By default, errors='raise', meaning that any errors encoun-
tered will be raised during the conversion process. However, if errors='coerce', these errors will be ignored
and pandas will convert problematic elements to pd.NaT (for datetime and timedelta) or np.nan (for numeric).
This might be useful if you are reading in data which is mostly of the desired dtype (e.g. numeric, datetime), but
occasionally has non-conforming elements intermixed that you want to represent as missing:
In [377]: import datetime

In [378]: m = ['apple', datetime.datetime(2016, 3, 2)]

In [379]: pd.to_datetime(m, errors='coerce')

Out[379]: DatetimeIndex(['NaT', '2016-03-02'], dtype='datetime64[ns]', freq=None)

In [380]: m = ['apple', 2, 3]

In [381]: pd.to_numeric(m, errors='coerce')

Out[381]: array([nan, 2., 3.])

In [382]: m = ['apple', pd.Timedelta('1day')]

In [383]: pd.to_timedelta(m, errors='coerce')

Out[383]: TimedeltaIndex([NaT, '1 days'], dtype='timedelta64[ns]', freq=None)

The errors parameter has a third option of errors='ignore', which will simply return the passed in data if it
encounters any errors with the conversion to a desired data type:
In [384]: import datetime

In [385]: m = ['apple', datetime.datetime(2016, 3, 2)]

In [386]: pd.to_datetime(m, errors='ignore')

Out[386]: Index(['apple', 2016-03-02 00:00:00], dtype='object')

In [387]: m = ['apple', 2, 3]

In [388]: pd.to_numeric(m, errors='ignore')

Out[388]: array(['apple', 2, 3], dtype=object)

In [389]: m = ['apple', pd.Timedelta('1day')]

In [390]: pd.to_timedelta(m, errors='ignore')

Out[390]: array(['apple', Timedelta('1 days 00:00:00')], dtype=object)

In addition to object conversion, to_numeric() provides another argument downcast, which gives the option of
downcasting the newly (or already) numeric data to a smaller dtype, which can conserve memory:
In [391]: m = ['1', 2, 3]

In [392]: pd.to_numeric(m, downcast='integer') # smallest signed int dtype

Out[392]: array([1, 2, 3], dtype=int8)

In [393]: pd.to_numeric(m, downcast='signed') # same as 'integer'

3.3. Essential basic functionality 99

pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[393]: array([1, 2, 3], dtype=int8)

In [394]: pd.to_numeric(m, downcast='unsigned') # smallest unsigned int dtype

Out[394]: array([1, 2, 3], dtype=uint8)

In [395]: pd.to_numeric(m, downcast='float') # smallest float dtype

Out[395]: array([1., 2., 3.], dtype=float32)
As these methods apply only to one-dimensional arrays, lists or scalars; they cannot be used directly on multi-
dimensional objects such as DataFrames. However, with apply(), we can apply the function over each column
efficiently:
In [396]: import datetime

In [397]: df = pd.DataFrame([
.....: ['2016-07-09', datetime.datetime(2016, 3, 2)]] * 2, dtype='O')
.....:

In [398]: df
Out[398]:
0 1
0 2016-07-09 2016-03-02 00:00:00
1 2016-07-09 2016-03-02 00:00:00

In [399]: df.apply(pd.to_datetime)
Out[399]:
0 1
0 2016-07-09 2016-03-02
1 2016-07-09 2016-03-02

In [400]: df = pd.DataFrame([['1.1', 2, 3]] * 2, dtype='O')

In [401]: df
Out[401]:
0 1 2
0 1.1 2 3
1 1.1 2 3

In [402]: df.apply(pd.to_numeric)
Out[402]:
0 1 2
0 1.1 2 3
1 1.1 2 3

In [403]: df = pd.DataFrame([['5us', pd.Timedelta('1day')]] * 2, dtype='O')

In [404]: df
Out[404]:
0 1
0 5us 1 days 00:00:00
1 5us 1 days 00:00:00

In [405]: df.apply(pd.to_timedelta)
Out[405]:
0 1

100 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

0 00:00:00.000005 1 days
1 00:00:00.000005 1 days

gotchas

Performing selection operations on integer type data can easily upcast the data to floating. The dtype of the
input data will be preserved in cases where nans are not introduced. See also Support for integer NA.
In [406]: dfi = df3.astype('int32')

In [407]: dfi['E'] = 1

In [408]: dfi
Out[408]:
A B C E
0 0 1 1 1
1 0 0 0 1
2 0 0 1 1
3 1 0 0 1
4 1 1 0 1
5 -1 -1 0 1
6 0 0 0 1
7 0 -1 0 1

In [409]: dfi.dtypes
Out[409]:
A int32
B int32
C int32
E int64
dtype: object

In [410]: casted = dfi[dfi > 0]

In [411]: casted
Out[411]:
A B C E
0 NaN 1.0 1.0 1
1 NaN NaN NaN 1
2 NaN NaN 1.0 1
3 1.0 NaN NaN 1
4 1.0 1.0 NaN 1
5 NaN NaN NaN 1
6 NaN NaN NaN 1
7 NaN NaN NaN 1

In [412]: casted.dtypes
Out[412]:
A float64
B float64
C float64
E int64
dtype: object

3.3. Essential basic functionality 101

pandas: powerful Python data analysis toolkit, Release 0.25.3

While float dtypes are unchanged.

In [413]: dfa = df3.copy()

In [414]: dfa['A'] = dfa['A'].astype('float32')

In [415]: dfa.dtypes
Out[415]:
A float32
B float64
C float64
dtype: object

In [416]: casted = dfa[df2 > 0]

In [417]: casted
Out[417]:
A B C
0 NaN 1.006054 1.0
1 0.019523 NaN NaN
2 -0.043356 NaN 1.0
3 1.910244 NaN NaN
4 1.249831 1.201647 NaN
5 NaN NaN NaN
6 0.432970 0.098521 NaN
7 0.487979 NaN NaN

In [418]: casted.dtypes
Out[418]:
A float32
B float64
C float64
dtype: object

3.3.14 Selecting columns based on dtype

The select_dtypes() method implements subsetting of columns based on their dtype.

First, lets create a DataFrame with a slew of different dtypes:
In [419]: df = pd.DataFrame({'string': list('abc'),
.....: 'int64': list(range(1, 4)),
.....: 'uint8': np.arange(3, 6).astype('u1'),
.....: 'float64': np.arange(4.0, 7.0),
.....: 'bool1': [True, False, True],
.....: 'bool2': [False, True, False],
.....: 'dates': pd.date_range('now', periods=3),
.....: 'category': pd.Series(list("ABC")).astype('category')})
.....:

In [420]: df['tdeltas'] = df.dates.diff()

In [421]: df['uint64'] = np.arange(3, 6).astype('u8')

In [422]: df['other_dates'] = pd.date_range('20130101', periods=3)

(continues on next page)

102 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [423]: df['tz_aware_dates'] = pd.date_range('20130101', periods=3, tz='US/Eastern')

In [424]: df
Out[424]:
string int64 uint8 float64 bool1 ... category tdeltas uint64 other_dates
,→ tz_aware_dates
0 a 1 3 4.0 True ... A NaT 3 2013-01-01 2013-
,→01-01 00:00:00-05:00

1 b 2 4 5.0 False ... B 1 days 4 2013-01-02 2013-

,→01-02 00:00:00-05:00

2 c 3 5 6.0 True ... C 1 days 5 2013-01-03 2013-

,→01-03 00:00:00-05:00

[3 rows x 12 columns]

And the dtypes:

In [425]: df.dtypes
Out[425]:
string object
int64 int64
uint8 uint8
float64 float64
bool1 bool
bool2 bool
dates datetime64[ns]
category category
tdeltas timedelta64[ns]
uint64 uint64
other_dates datetime64[ns]
tz_aware_dates datetime64[ns, US/Eastern]
dtype: object

select_dtypes() has two parameters include and exclude that allow you to say give me the columns with
these dtypes (include) and/or give the columns without these dtypes (exclude).
For example, to select bool columns:

In [426]: df.select_dtypes(include=[bool])
Out[426]:
bool1 bool2
0 True False
1 False True
2 True False

You can also pass the name of a dtype in the NumPy dtype hierarchy:

In [427]: df.select_dtypes(include=['bool'])
Out[427]:
bool1 bool2
0 True False
1 False True
2 True False

select_dtypes() also works with generic dtypes as well.

For example, to select all numeric and boolean columns while excluding unsigned integers:

3.3. Essential basic functionality 103

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [428]: df.select_dtypes(include=['number', 'bool'], exclude=['unsignedinteger'])

Out[428]:
int64 float64 bool1 bool2 tdeltas
0 1 4.0 True False NaT
1 2 5.0 False True 1 days
2 3 6.0 True False 1 days

To select string columns you must use the object dtype:

In [429]: df.select_dtypes(include=['object'])
Out[429]:
string
0 a
1 b
2 c

To see all the child dtypes of a generic dtype like numpy.number you can define a function that returns a tree of
child dtypes:

In [430]: def subdtypes(dtype):

.....: subs = dtype.__subclasses__()
.....: if not subs:
.....: return dtype
.....: return [dtype, [subdtypes(dt) for dt in subs]]
.....:

All NumPy dtypes are subclasses of numpy.generic:

In [431]: subdtypes(np.generic)
Out[431]:
[numpy.generic,
[[numpy.number,
[[numpy.integer,
[[numpy.signedinteger,
[numpy.int8,
numpy.int16,
numpy.int32,
numpy.int64,
numpy.int64,
numpy.timedelta64]],
[numpy.unsignedinteger,
[numpy.uint8,
numpy.uint16,
numpy.uint32,
numpy.uint64,
numpy.uint64]]]],
[numpy.inexact,
[[numpy.floating,
[numpy.float16, numpy.float32, numpy.float64, numpy.float128]],
[numpy.complexfloating,
[numpy.complex64, numpy.complex128, numpy.complex256]]]]]],
[numpy.flexible,
[[numpy.character, [numpy.bytes_, numpy.str_]],
[numpy.void, [numpy.record]]]],
numpy.bool_,
numpy.datetime64,
numpy.object_]]

104 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Note: Pandas also defines the types category, and datetime64[ns, tz], which are not integrated into the
normal NumPy hierarchy and wont show up with the above function.

{{ header }}

3.4 Intro to data structures

Well start with a quick, non-comprehensive overview of the fundamental data structures in pandas to get you started.
The fundamental behavior about data types, indexing, and axis labeling / alignment apply across all of the objects. To
get started, import NumPy and load pandas into your namespace:

In [1]: import numpy as np

In [2]: import pandas as pd

Here is a basic tenet to keep in mind: data alignment is intrinsic. The link between labels and data will not be broken
unless done so explicitly by you.
Well give a brief intro to the data structures, then consider all of the broad categories of functionality and methods in
separate sections.

3.4.1 Series

Series is a one-dimensional labeled array capable of holding any data type (integers, strings, floating point numbers,
Python objects, etc.). The axis labels are collectively referred to as the index. The basic method to create a Series is
to call:

>>> s = pd.Series(data, index=index)

Here, data can be many different things:

• a Python dict
• an ndarray
• a scalar value (like 5)
The passed index is a list of axis labels. Thus, this separates into a few cases depending on what data is:
From ndarray
If data is an ndarray, index must be the same length as data. If no index is passed, one will be created having values
[0, ..., len(data) - 1].
In [3]: s = pd.Series(np.random.randn(5), index=['a', 'b', 'c', 'd', 'e'])

In [4]: s
Out[4]:
a 1.585681
b -0.737876
c -1.603758
d -0.668269
e -1.104086
dtype: float64

3.4. Intro to data structures 105

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [5]: s.index
Out[5]: Index(['a', 'b', 'c', 'd', 'e'], dtype='object')

In [6]: pd.Series(np.random.randn(5))
Out[6]:
0 0.873370
1 -1.516058
2 1.625341
3 0.561420
4 -0.009922
dtype: float64

Note: pandas supports non-unique index values. If an operation that does not support duplicate index values is
attempted, an exception will be raised at that time. The reason for being lazy is nearly all performance-based (there
are many instances in computations, like parts of GroupBy, where the index is not used).

From dict
Series can be instantiated from dicts:

In [7]: d = {'b': 1, 'a': 0, 'c': 2}

In [8]: pd.Series(d)
Out[8]:
b 1
a 0
c 2
dtype: int64

Note: When the data is a dict, and an index is not passed, the Series index will be ordered by the dicts insertion
order, if youre using Python version >= 3.6 and Pandas version >= 0.23.
If youre using Python < 3.6 or Pandas < 0.23, and an index is not passed, the Series index will be the lexically
ordered list of dict keys.

In the example above, if you were on a Python version lower than 3.6 or a Pandas version lower than 0.23, the Series
would be ordered by the lexical order of the dict keys (i.e. ['a', 'b', 'c'] rather than ['b', 'a', 'c']).
If an index is passed, the values in data corresponding to the labels in the index will be pulled out.
In [9]: d = {'a': 0., 'b': 1., 'c': 2.}

In [10]: pd.Series(d)
Out[10]:
a 0.0
b 1.0
c 2.0
dtype: float64

In [11]: pd.Series(d, index=['b', 'c', 'd', 'a'])

Out[11]:
b 1.0
c 2.0
d NaN

106 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

a 0.0
dtype: float64

Note: NaN (not a number) is the standard missing data marker used in pandas.

From scalar value

If data is a scalar value, an index must be provided. The value will be repeated to match the length of index.

In [12]: pd.Series(5., index=['a', 'b', 'c', 'd', 'e'])

Out[12]:
a 5.0
b 5.0
c 5.0
d 5.0
e 5.0
dtype: float64

Series is ndarray-like

Series acts very similarly to a ndarray, and is a valid argument to most NumPy functions. However, operations
such as slicing will also slice the index.
In [13]: s[0]
Out[13]: 1.5856810982636127

In [14]: s[:3]
Out[14]:
a 1.585681
b -0.737876
c -1.603758
dtype: float64

In [15]: s[s > s.median()]

Out[15]:
a 1.585681
d -0.668269
dtype: float64

In [16]: s[[4, 3, 1]]

Out[16]:
e -1.104086
d -0.668269
b -0.737876
dtype: float64

In [17]: np.exp(s)
Out[17]:
a 4.882616
b 0.478129
c 0.201139
d 0.512595
e 0.331514
dtype: float64

3.4. Intro to data structures 107

pandas: powerful Python data analysis toolkit, Release 0.25.3

Note: We will address array-based indexing like s[[4, 3, 1]] in section.

Like a NumPy array, a pandas Series has a dtype.

In [18]: s.dtype
Out[18]: dtype('float64')

This is often a NumPy dtype. However, pandas and 3rd-party libraries extend NumPys type system in a few places,
in which case the dtype would be a ExtensionDtype. Some examples within pandas are categorical and Nullable
integer data type. See dtypes for more.
If you need the actual array backing a Series, use Series.array.

In [19]: s.array
Out[19]:
<PandasArray>
[ 1.5856810982636127, -0.7378757380201878, -1.6037579905058967,
-0.668268701321749, -1.1040860124173795]
Length: 5, dtype: float64

Accessing the array can be useful when you need to do some operation without the index (to disable automatic
alignment, for example).
Series.array will always be an ExtensionArray. Briefly, an ExtensionArray is a thin wrapper around one or
more concrete arrays like a numpy.ndarray. Pandas knows how to take an ExtensionArray and store it in a
Series or a column of a DataFrame. See dtypes for more.
While Series is ndarray-like, if you need an actual ndarray, then use Series.to_numpy().

In [20]: s.to_numpy()
Out[20]: array([ 1.5856811 , -0.73787574, -1.60375799, -0.6682687 , -1.10408601])

Even if the Series is backed by a ExtensionArray, Series.to_numpy() will return a NumPy ndarray.

Series is dict-like

A Series is like a fixed-size dict in that you can get and set values by index label:
In [21]: s['a']
Out[21]: 1.5856810982636127

In [22]: s['e'] = 12.

In [23]: s
Out[23]:
a 1.585681
b -0.737876
c -1.603758
d -0.668269
e 12.000000
dtype: float64

In [24]: 'e' in s
Out[24]: True

108 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [25]: 'f' in s
Out[25]: False
If a label is not contained, an exception is raised:

>>> s['f']
KeyError: 'f'

Using the get method, a missing label will return None or specified default:

In [26]: s.get('f')

In [27]: s.get('f', np.nan)

Out[27]: nan

See also the section on attribute access.

Vectorized operations and label alignment with Series

When working with raw NumPy arrays, looping through value-by-value is usually not necessary. The same is true
when working with Series in pandas. Series can also be passed into most NumPy methods expecting an ndarray.
In [28]: s + s
Out[28]:
a 3.171362
b -1.475751
c -3.207516
d -1.336537
e 24.000000
dtype: float64

In [29]: s * 2
Out[29]:
a 3.171362
b -1.475751
c -3.207516
d -1.336537
e 24.000000
dtype: float64

In [30]: np.exp(s)
Out[30]:
a 4.882616
b 0.478129
c 0.201139
d 0.512595
e 162754.791419
dtype: float64
A key difference between Series and ndarray is that operations between Series automatically align the data based on
label. Thus, you can write computations without giving consideration to whether the Series involved have the same
labels.

In [31]: s[1:] + s[:-1]

Out[31]:
(continues on next page)

3.4. Intro to data structures 109

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

a NaN
b -1.475751
c -3.207516
d -1.336537
e NaN
dtype: float64

The result of an operation between unaligned Series will have the union of the indexes involved. If a label is not found
in one Series or the other, the result will be marked as missing NaN. Being able to write code without doing any explicit
data alignment grants immense freedom and flexibility in interactive data analysis and research. The integrated data
alignment features of the pandas data structures set pandas apart from the majority of related tools for working with
labeled data.

Note: In general, we chose to make the default result of operations between differently indexed objects yield the
union of the indexes in order to avoid loss of information. Having an index label, though the data is missing, is
typically important information as part of a computation. You of course have the option of dropping labels with
missing data via the dropna function.

Name attribute

Series can also have a name attribute:

In [32]: s = pd.Series(np.random.randn(5), name='something')

In [33]: s
Out[33]:
0 -0.987980
1 1.649534
2 -0.333885
3 0.303428
4 1.388549
Name: something, dtype: float64

In [34]: s.name
Out[34]: 'something'
The Series name will be assigned automatically in many cases, in particular when taking 1D slices of DataFrame as
you will see below.
New in version 0.18.0.
You can rename a Series with the pandas.Series.rename() method.

In [35]: s2 = s.rename("different")

In [36]: s2.name
Out[36]: 'different'

Note that s and s2 refer to different objects.

110 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

3.4.2 DataFrame

DataFrame is a 2-dimensional labeled data structure with columns of potentially different types. You can think of it
like a spreadsheet or SQL table, or a dict of Series objects. It is generally the most commonly used pandas object.
Like Series, DataFrame accepts many different kinds of input:
• Dict of 1D ndarrays, lists, dicts, or Series
• 2-D numpy.ndarray
• Structured or record ndarray
• A Series
• Another DataFrame
Along with the data, you can optionally pass index (row labels) and columns (column labels) arguments. If you pass
an index and / or columns, you are guaranteeing the index and / or columns of the resulting DataFrame. Thus, a dict
of Series plus a specific index will discard all data not matching up to the passed index.
If axis labels are not passed, they will be constructed from the input data based on common sense rules.

Note: When the data is a dict, and columns is not specified, the DataFrame columns will be ordered by the dicts
insertion order, if you are using Python version >= 3.6 and Pandas >= 0.23.
If you are using Python < 3.6 or Pandas < 0.23, and columns is not specified, the DataFrame columns will be the
lexically ordered list of dict keys.

From dict of Series or dicts

The resulting index will be the union of the indexes of the various Series. If there are any nested dicts, these will first
be converted to Series. If no columns are passed, the columns will be the ordered list of dict keys.
In [37]: d = {'one': pd.Series([1., 2., 3.], index=['a', 'b', 'c']),
....: 'two': pd.Series([1., 2., 3., 4.], index=['a', 'b', 'c', 'd'])}
....:

In [38]: df = pd.DataFrame(d)

In [39]: df
Out[39]:
one two
a 1.0 1.0
b 2.0 2.0
c 3.0 3.0
d NaN 4.0

In [40]: pd.DataFrame(d, index=['d', 'b', 'a'])

Out[40]:
one two
d NaN 4.0
b 2.0 2.0
a 1.0 1.0

In [41]: pd.DataFrame(d, index=['d', 'b', 'a'], columns=['two', 'three'])

Out[41]:

3.4. Intro to data structures 111

pandas: powerful Python data analysis toolkit, Release 0.25.3

two three
d 4.0 NaN
b 2.0 NaN
a 1.0 NaN
The row and column labels can be accessed respectively by accessing the index and columns attributes:

Note: When a particular set of columns is passed along with a dict of data, the passed columns override the keys in
the dict.

In [42]: df.index
Out[42]: Index(['a', 'b', 'c', 'd'], dtype='object')

In [43]: df.columns
Out[43]: Index(['one', 'two'], dtype='object')

From dict of ndarrays / lists

The ndarrays must all be the same length. If an index is passed, it must clearly also be the same length as the arrays.
If no index is passed, the result will be range(n), where n is the array length.
In [44]: d = {'one': [1., 2., 3., 4.],
....: 'two': [4., 3., 2., 1.]}
....:

In [45]: pd.DataFrame(d)
Out[45]:
one two
0 1.0 4.0
1 2.0 3.0
2 3.0 2.0
3 4.0 1.0

In [46]: pd.DataFrame(d, index=['a', 'b', 'c', 'd'])

Out[46]:
one two
a 1.0 4.0
b 2.0 3.0
c 3.0 2.0
d 4.0 1.0

From structured or record array

This case is handled identically to a dict of arrays.

In [47]: data = np.zeros((2, ), dtype=[('A', 'i4'), ('B', 'f4'), ('C',
,→'a10')])

In [48]: data[:] = [(1, 2., 'Hello'), (2, 3., "World")]

In [49]: pd.DataFrame(data)
Out[49]:
A B C

112 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

0 1 2.0 b'Hello'
1 2 3.0 b'World'

In [50]: pd.DataFrame(data, index=['first', 'second'])

Out[50]:
A B C
first 1 2.0 b'Hello'
second 2 3.0 b'World'

In [51]: pd.DataFrame(data, columns=['C', 'A', 'B'])

Out[51]:
C A B
0 b'Hello' 1 2.0
1 b'World' 2 3.0

Note: DataFrame is not intended to work exactly like a 2-dimensional NumPy ndarray.

From a list of dicts

In [52]: data2 = [{'a': 1, 'b': 2}, {'a': 5, 'b': 10, 'c': 20}]

In [53]: pd.DataFrame(data2)
Out[53]:
a b c
0 1 2 NaN
1 5 10 20.0

In [54]: pd.DataFrame(data2, index=['first', 'second'])

Out[54]:
a b c
first 1 2 NaN
second 5 10 20.0

In [55]: pd.DataFrame(data2, columns=['a', 'b'])

Out[55]:
a b
0 1 2
1 5 10

From a dict of tuples

You can automatically create a MultiIndexed frame by passing a tuples dictionary.

In [56]: pd.DataFrame({('a', 'b'): {('A', 'B'): 1, ('A', 'C'): 2},

....: ('a', 'a'): {('A', 'C'): 3, ('A', 'B'): 4},
....: ('a', 'c'): {('A', 'B'): 5, ('A', 'C'): 6},
....: ('b', 'a'): {('A', 'C'): 7, ('A', 'B'): 8},
....: ('b', 'b'): {('A', 'D'): 9, ('A', 'B'): 10}})
....:
Out[56]:
a b
(continues on next page)

3.4. Intro to data structures 113

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

b a c a b
A B 1.0 4.0 5.0 8.0 10.0
C 2.0 3.0 6.0 7.0 NaN
D NaN NaN NaN NaN 9.0

From a Series

The result will be a DataFrame with the same index as the input Series, and with one column whose name is the
original name of the Series (only if no other column name provided).
Missing data
Much more will be said on this topic in the Missing data section. To construct a DataFrame with missing data, we use
np.nan to represent missing values. Alternatively, you may pass a numpy.MaskedArray as the data argument to
the DataFrame constructor, and its masked entries will be considered missing.

Alternate constructors

DataFrame.from_dict
DataFrame.from_dict takes a dict of dicts or a dict of array-like sequences and returns a DataFrame. It operates
like the DataFrame constructor except for the orient parameter which is 'columns' by default, but which can
be set to 'index' in order to use the dict keys as row labels.

In [57]: pd.DataFrame.from_dict(dict([('A', [1, 2, 3]), ('B', [4, 5, 6])]))

Out[57]:
A B
0 1 4
1 2 5
2 3 6

If you pass orient='index', the keys will be the row labels. In this case, you can also pass the desired column
names:

In [58]: pd.DataFrame.from_dict(dict([('A', [1, 2, 3]), ('B', [4, 5, 6])]),

....: orient='index', columns=['one', 'two', 'three'])
....:
Out[58]:
one two three
A 1 2 3
B 4 5 6

DataFrame.from_records
DataFrame.from_records takes a list of tuples or an ndarray with structured dtype. It works analogously to the
normal DataFrame constructor, except that the resulting DataFrame index may be a specific field of the structured
dtype. For example:
In [59]: data
Out[59]:
array([(1, 2., b'Hello'), (2, 3., b'World')],
dtype=[('A', '<i4'), ('B', '<f4'), ('C', 'S10')])

In [60]: pd.DataFrame.from_records(data, index='C')

Out[60]:

114 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

A B
C
b'Hello' 1 2.0
b'World' 2 3.0

Column selection, addition, deletion

You can treat a DataFrame semantically like a dict of like-indexed Series objects. Getting, setting, and deleting
columns works with the same syntax as the analogous dict operations:

In [61]: df['one']
Out[61]:
a 1.0
b 2.0
c 3.0
d NaN
Name: one, dtype: float64

In [62]: df['three'] = df['one'] * df['two']

In [63]: df['flag'] = df['one'] > 2

In [64]: df
Out[64]:
one two three flag
a 1.0 1.0 1.0 False
b 2.0 2.0 4.0 False
c 3.0 3.0 9.0 True
d NaN 4.0 NaN False

Columns can be deleted or popped like with a dict:

In [65]: del df['two']

In [66]: three = df.pop('three')

In [67]: df
Out[67]:
one flag
a 1.0 False
b 2.0 False
c 3.0 True
d NaN False

When inserting a scalar value, it will naturally be propagated to fill the column:

In [68]: df['foo'] = 'bar'

In [69]: df
Out[69]:
one flag foo
a 1.0 False bar
b 2.0 False bar
c 3.0 True bar
d NaN False bar

3.4. Intro to data structures 115

pandas: powerful Python data analysis toolkit, Release 0.25.3

When inserting a Series that does not have the same index as the DataFrame, it will be conformed to the DataFrames
index:

In [70]: df['one_trunc'] = df['one'][:2]

In [71]: df
Out[71]:
one flag foo one_trunc
a 1.0 False bar 1.0
b 2.0 False bar 2.0
c 3.0 True bar NaN
d NaN False bar NaN

You can insert raw ndarrays but their length must match the length of the DataFrames index.
By default, columns get inserted at the end. The insert function is available to insert at a particular location in the
columns:

In [72]: df.insert(1, 'bar', df['one'])

In [73]: df
Out[73]:
one bar flag foo one_trunc
a 1.0 1.0 False bar 1.0
b 2.0 2.0 False bar 2.0
c 3.0 3.0 True bar NaN
d NaN NaN False bar NaN

Assigning new columns in method chains

Inspired by dplyrs mutate verb, DataFrame has an assign() method that allows you to easily create new columns
that are potentially derived from existing columns.
In [74]: iris = pd.read_csv('data/iris.data')

In [75]: iris.head()
Out[75]:
SepalLength SepalWidth PetalLength PetalWidth Name
0 5.1 3.5 1.4 0.2 Iris-setosa
1 4.9 3.0 1.4 0.2 Iris-setosa
2 4.7 3.2 1.3 0.2 Iris-setosa
3 4.6 3.1 1.5 0.2 Iris-setosa
4 5.0 3.6 1.4 0.2 Iris-setosa

In [76]: (iris.assign(sepal_ratio=iris['SepalWidth'] / iris['SepalLength'])

....: .head())
....:
Out[76]:
SepalLength SepalWidth PetalLength PetalWidth Name sepal_ratio
0 5.1 3.5 1.4 0.2 Iris-setosa 0.686275
1 4.9 3.0 1.4 0.2 Iris-setosa 0.612245
2 4.7 3.2 1.3 0.2 Iris-setosa 0.680851
3 4.6 3.1 1.5 0.2 Iris-setosa 0.673913
4 5.0 3.6 1.4 0.2 Iris-setosa 0.720000

116 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In the example above, we inserted a precomputed value. We can also pass in a function of one argument to be evaluated
on the DataFrame being assigned to.

In [77]: iris.assign(sepal_ratio=lambda x: (x['SepalWidth'] / x['SepalLength'])).

,→head()

Out[77]:
SepalLength SepalWidth PetalLength PetalWidth Name sepal_ratio
0 5.1 3.5 1.4 0.2 Iris-setosa 0.686275
1 4.9 3.0 1.4 0.2 Iris-setosa 0.612245
2 4.7 3.2 1.3 0.2 Iris-setosa 0.680851
3 4.6 3.1 1.5 0.2 Iris-setosa 0.673913
4 5.0 3.6 1.4 0.2 Iris-setosa 0.720000

assign always returns a copy of the data, leaving the original DataFrame untouched.
Passing a callable, as opposed to an actual value to be inserted, is useful when you dont have a reference to the
DataFrame at hand. This is common when using assign in a chain of operations. For example, we can limit the
DataFrame to just those observations with a Sepal Length greater than 5, calculate the ratio, and plot:

In [78]: (iris.query('SepalLength > 5')

....: .assign(SepalRatio=lambda x: x.SepalWidth / x.SepalLength,
....: PetalRatio=lambda x: x.PetalWidth / x.PetalLength)
....: .plot(kind='scatter', x='SepalRatio', y='PetalRatio'))
....:
Out[78]: <matplotlib.axes._subplots.AxesSubplot at 0x129542a90>

3.4. Intro to data structures 117

pandas: powerful Python data analysis toolkit, Release 0.25.3

Since a function is passed in, the function is computed on the DataFrame being assigned to. Importantly, this is the
DataFrame thats been filtered to those rows with sepal length greater than 5. The filtering happens first, and then the
ratio calculations. This is an example where we didnt have a reference to the filtered DataFrame available.
The function signature for assign is simply **kwargs. The keys are the column names for the new fields, and the
values are either a value to be inserted (for example, a Series or NumPy array), or a function of one argument to be
called on the DataFrame. A copy of the original DataFrame is returned, with the new values inserted.
Changed in version 0.23.0.
Starting with Python 3.6 the order of **kwargs is preserved. This allows for dependent assignment, where an
expression later in **kwargs can refer to a column created earlier in the same assign().
In [79]: dfa = pd.DataFrame({"A": [1, 2, 3],
....: "B": [4, 5, 6]})
....:

In [80]: dfa.assign(C=lambda x: x['A'] + x['B'],

....: D=lambda x: x['A'] + x['C'])
....:
Out[80]:
A B C D
0 1 4 5 6
1 2 5 7 9
2 3 6 9 12

In the second expression, x['C'] will refer to the newly created column, thats equal to dfa['A'] + dfa['B'].
To write code compatible with all versions of Python, split the assignment in two.
In [81]: dependent = pd.DataFrame({"A": [1, 1, 1]})

In [82]: (dependent.assign(A=lambda x: x['A'] + 1)

....: .assign(B=lambda x: x['A'] + 2))
....:
Out[82]:
A B
0 2 4
1 2 4
2 2 4

Warning: Dependent assignment may subtly change the behavior of your code between Python 3.6 and older
versions of Python.
If you wish to write code that supports versions of python before and after 3.6, youll need to take care when passing
assign expressions that
• Update an existing column
• Refer to the newly updated column in the same assign
For example, well update column A and then refer to it when creating B.
>>> dependent = pd.DataFrame({"A": [1, 1, 1]})
>>> dependent.assign(A=lambda x: x["A"] + 1, B=lambda x: x["A"] + 2)

For Python 3.5 and earlier the expression creating B refers to the old value of A, [1, 1, 1]. The output is then
A B
0 2 3
1 2 3
2 2 3

118 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

For Python 3.6 and later, the expression creating A refers to the new value of A, [2, 2, 2], which results in
A B
0 2 4
1 2 4
2 2 4

Indexing / selection

The basics of indexing are as follows:

Operation Syntax Result

Select column df[col] Series
Select row by label df.loc[label] Series
Select row by integer location df.iloc[loc] Series
Slice rows df[5:10] DataFrame
Select rows by boolean vector df[bool_vec] DataFrame

Row selection, for example, returns a Series whose index is the columns of the DataFrame:
In [83]: df.loc['b']
Out[83]:
one 2
bar 2
flag False
foo bar
one_trunc 2
Name: b, dtype: object

In [84]: df.iloc[2]
Out[84]:
one 3
bar 3
flag True
foo bar
one_trunc NaN
Name: c, dtype: object
For a more exhaustive treatment of sophisticated label-based indexing and slicing, see the section on indexing. We
will address the fundamentals of reindexing / conforming to new sets of labels in the section on reindexing.

Data alignment and arithmetic

Data alignment between DataFrame objects automatically align on both the columns and the index (row labels).
Again, the resulting object will have the union of the column and row labels.
In [85]: df = pd.DataFrame(np.random.randn(10, 4), columns=['A', 'B', 'C', 'D'])

In [86]: df2 = pd.DataFrame(np.random.randn(7, 3), columns=['A', 'B', 'C'])

In [87]: df + df2
Out[87]:
(continues on next page)

3.4. Intro to data structures 119

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

A B C D
0 -2.386114 -1.656789 0.925576 NaN
1 -0.504949 -0.468878 -1.876635 NaN
2 0.230231 -3.873474 2.610702 NaN
3 -0.564199 0.101636 0.369056 NaN
4 1.322597 0.820841 -1.877026 NaN
5 -0.311563 -0.185355 0.628811 NaN
6 1.943183 -0.098391 0.250144 NaN
7 NaN NaN NaN NaN
8 NaN NaN NaN NaN
9 NaN NaN NaN NaN

When doing an operation between DataFrame and Series, the default behavior is to align the Series index on the
DataFrame columns, thus broadcasting row-wise. For example:

In [88]: df - df.iloc[0]
Out[88]:
A B C D
0 0.000000 0.000000 0.000000 0.000000
1 1.167144 0.019752 -0.477114 -0.905114
2 1.506963 -2.859436 2.505428 1.842367
3 2.289411 0.669679 2.262291 0.641429
4 1.516210 1.316061 -1.161409 3.026566
5 1.333741 -0.282192 0.843573 0.798528
6 2.657332 1.228825 -0.475312 2.257556
7 1.965229 0.851194 2.054340 -0.250865
8 0.802379 0.151740 -0.152969 1.492131
9 0.620426 2.396284 -2.063297 -0.284884

In the special case of working with time series data, if the DataFrame index contains dates, the broadcasting will be
column-wise:
In [89]: index = pd.date_range('1/1/2000', periods=8)

In [90]: df = pd.DataFrame(np.random.randn(8, 3), index=index,

,→columns=list('ABC'))

In [91]: df
Out[91]:
A B C
2000-01-01 0.780948 0.587141 0.386433
2000-01-02 -1.150341 0.553978 1.047461
2000-01-03 0.021671 -1.720821 0.192664
2000-01-04 1.058264 0.507151 -0.622097
2000-01-05 0.722388 -0.072578 0.626614
2000-01-06 -0.899071 0.593542 1.557537
2000-01-07 0.446388 1.357955 0.979408
2000-01-08 1.434806 -1.811677 -0.556518

In [92]: type(df['A'])
Out[92]: pandas.core.series.Series

In [93]: df - df['A']
Out[93]:
2000-01-01 00:00:00 2000-01-02 00:00:00 2000-01-03 00:00:00 ...

120 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

,→ A B C
2000-01-01 NaN NaN NaN ...
,→ NaN NaN NaN
2000-01-02 NaN NaN NaN ...
,→ NaN NaN NaN
2000-01-03 NaN NaN NaN ...
,→ NaN NaN NaN
2000-01-04 NaN NaN NaN ...
,→ NaN NaN NaN
2000-01-05 NaN NaN NaN ...
,→ NaN NaN NaN
2000-01-06 NaN NaN NaN ...
,→ NaN NaN NaN
2000-01-07 NaN NaN NaN ...
,→ NaN NaN NaN
2000-01-08 NaN NaN NaN ...
,→ NaN NaN NaN

[8 rows x 11 columns]

Warning:
df - df['A']

is now deprecated and will be removed in a future release. The preferred way to replicate this behavior is
df.sub(df['A'], axis=0)

For explicit control over the matching and broadcasting behavior, see the section on flexible binary operations.
Operations with scalars are just as you would expect:
In [94]: df * 5 + 2
Out[94]:
A B C
2000-01-01 5.904738 4.935707 3.932165
2000-01-02 -3.751704 4.769892 7.237304
2000-01-03 2.108354 -6.604107 2.963318
2000-01-04 7.291318 4.535757 -1.110487
2000-01-05 5.611942 1.637111 5.133071
2000-01-06 -2.495355 4.967712 9.787685
2000-01-07 4.231938 8.789774 6.897039
2000-01-08 9.174028 -7.058386 -0.782592

In [95]: 1 / df
Out[95]:
A B C
2000-01-01 1.280496 1.703167 2.587770
2000-01-02 -0.869308 1.805125 0.954690
2000-01-03 46.144995 -0.581118 5.190396
2000-01-04 0.944944 1.971798 -1.607465
2000-01-05 1.384297 -13.778321 1.595878
2000-01-06 -1.112259 1.684799 0.642039
2000-01-07 2.240205 0.736402 1.021025
2000-01-08 0.696959 -0.551975 -1.796886

3.4. Intro to data structures 121

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [96]: df ** 4
Out[96]:
A B C
2000-01-01 3.719525e-01 0.118842 0.022300
2000-01-02 1.751080e+00 0.094183 1.203791
2000-01-03 2.205472e-07 8.768861 0.001378
2000-01-04 1.254225e+00 0.066153 0.149773
2000-01-05 2.723222e-01 0.000028 0.154170
2000-01-06 6.533950e-01 0.124110 5.885094
2000-01-07 3.970532e-02 3.400487 0.920140
2000-01-08 4.238110e+00 10.772667 0.095922
Boolean operators work as well:
In [97]: df1 = pd.DataFrame({'a': [1, 0, 1], 'b': [0, 1, 1]}, dtype=bool)

In [98]: df2 = pd.DataFrame({'a': [0, 1, 1], 'b': [1, 1, 0]}, dtype=bool)

In [99]: df1 & df2

Out[99]:
a b
0 False False
1 False True
2 True False

In [100]: df1 | df2

Out[100]:
a b
0 True True
1 True True
2 True True

In [101]: df1 ^ df2

Out[101]:
a b
0 True True
1 True False
2 False True

In [102]: -df1
Out[102]:
a b
0 False True
1 True False
2 False False

Transposing

To transpose, access the T attribute (also the transpose function), similar to an ndarray:

# only show the first 5 rows

In [103]: df[:5].T
Out[103]:
(continues on next page)

122 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2000-01-01 2000-01-02 2000-01-03 2000-01-04 2000-01-05
A 0.780948 -1.150341 0.021671 1.058264 0.722388
B 0.587141 0.553978 -1.720821 0.507151 -0.072578
C 0.386433 1.047461 0.192664 -0.622097 0.626614

DataFrame interoperability with NumPy functions

Elementwise NumPy ufuncs (log, exp, sqrt, ) and various other NumPy functions can be used with no issues on Series
and DataFrame, assuming the data within are numeric:
In [104]: np.exp(df)
Out[104]:
A B C
2000-01-01 2.183540 1.798839 1.471722
2000-01-02 0.316529 1.740162 2.850404
2000-01-03 1.021907 0.178919 1.212475
2000-01-04 2.881363 1.660554 0.536817
2000-01-05 2.059346 0.929993 1.871264
2000-01-06 0.406948 1.810390 4.747114
2000-01-07 1.562657 3.888233 2.662879
2000-01-08 4.198829 0.163380 0.573201

In [105]: np.asarray(df)
Out[105]:
array([[ 0.78094758, 0.58714139, 0.38643309],
[-1.15034079, 0.55397835, 1.04746083],
[ 0.02167082, -1.7208214 , 0.19266354],
[ 1.0582636 , 0.50715133, -0.62209742],
[ 0.72238838, -0.07257778, 0.62661419],
[-0.89907091, 0.59354246, 1.55753692],
[ 0.44638764, 1.35795473, 0.97940772],
[ 1.43480561, -1.81167717, -0.55651844]])
DataFrame is not intended to be a drop-in replacement for ndarray as its indexing semantics and data model are quite
different in places from an n-dimensional array.
Series implements __array_ufunc__, which allows it to work with NumPys universal functions.
The ufunc is applied to the underlying array in a Series.

In [106]: ser = pd.Series([1, 2, 3, 4])

In [107]: np.exp(ser)
Out[107]:
0 2.718282
1 7.389056
2 20.085537
3 54.598150
dtype: float64

Changed in version 0.25.0: When multiple Series are passed to a ufunc, they are aligned before performing the
operation.
Like other parts of the library, pandas will automatically align labeled inputs as part of a ufunc with multiple inputs.
For example, using numpy.remainder() on two Series with differently ordered labels will align before the

3.4. Intro to data structures 123

pandas: powerful Python data analysis toolkit, Release 0.25.3

operation.
In [108]: ser1 = pd.Series([1, 2, 3], index=['a', 'b', 'c'])

In [109]: ser2 = pd.Series([1, 3, 5], index=['b', 'a', 'c'])

In [110]: ser1
Out[110]:
a 1
b 2
c 3
dtype: int64

In [111]: ser2
Out[111]:
b 1
a 3
c 5
dtype: int64

In [112]: np.remainder(ser1, ser2)

Out[112]:
a 1
b 0
c 3
dtype: int64
As usual, the union of the two indices is taken, and non-overlapping values are filled with missing values.
In [113]: ser3 = pd.Series([2, 4, 6], index=['b', 'c', 'd'])

In [114]: ser3
Out[114]:
b 2
c 4
d 6
dtype: int64

In [115]: np.remainder(ser1, ser3)

Out[115]:
a NaN
b 0.0
c 3.0
d NaN
dtype: float64
When a binary ufunc is applied to a Series and Index, the Series implementation takes precedence and a Series is
returned.

In [116]: ser = pd.Series([1, 2, 3])

In [117]: idx = pd.Index([4, 5, 6])

In [118]: np.maximum(ser, idx)

Out[118]:
0 4
(continues on next page)

124 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

1 5
2 6
dtype: int64

NumPy ufuncs are safe to apply to Series backed by non-ndarray arrays, for example SparseArray (see Sparse
calculation). If possible, the ufunc is applied without converting the underlying data to an ndarray.

Console display

Very large DataFrames will be truncated to display them in the console. You can also get a summary using info().
(Here I am reading a CSV version of the baseball dataset from the plyr R package):
In [119]: baseball = pd.read_csv('data/baseball.csv')

In [120]: print(baseball)
id player year stint team lg g ab r h ... rbi sb
,→ cs bb so ibb hbp sh sf gidp
0 88641 womacto01 2006 2 CHN NL 19 50 6 14 ... 2.0 1.0
,→1.0 4 4.0 0.0 0.0 3.0 0.0 0.0
1 88643 schilcu01 2006 1 BOS AL 31 2 0 1 ... 0.0 0.0
,→0.0 0 1.0 0.0 0.0 0.0 0.0 0.0
.. ... ... ... ... ... .. .. ... .. ... ... ... ...
,→... .. ... ... ... ... ... ...
98 89533 aloumo01 2007 1 NYN NL 87 328 51 112 ... 49.0 3.0
,→0.0 27 30.0 5.0 2.0 0.0 3.0 13.0
99 89534 alomasa02 2007 1 NYN NL 8 22 1 3 ... 0.0 0.0
,→0.0 0 3.0 0.0 0.0 0.0 0.0 0.0

[100 rows x 23 columns]

In [121]: baseball.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 100 entries, 0 to 99
Data columns (total 23 columns):
id 100 non-null int64
player 100 non-null object
year 100 non-null int64
stint 100 non-null int64
team 100 non-null object
lg 100 non-null object
g 100 non-null int64
ab 100 non-null int64
r 100 non-null int64
h 100 non-null int64
X2b 100 non-null int64
X3b 100 non-null int64
hr 100 non-null int64
rbi 100 non-null float64
sb 100 non-null float64
cs 100 non-null float64
bb 100 non-null int64
so 100 non-null float64
ibb 100 non-null float64

3.4. Intro to data structures 125

pandas: powerful Python data analysis toolkit, Release 0.25.3

hbp 100 non-null float64

sh 100 non-null float64
sf 100 non-null float64
gidp 100 non-null float64
dtypes: float64(9), int64(11), object(3)
memory usage: 18.1+ KB
However, using to_string will return a string representation of the DataFrame in tabular form, though it wont
always fit the console width:
In [122]: print(baseball.iloc[-20:, :12].to_string())
id player year stint team lg g ab r h X2b X3b
80 89474 finlest01 2007 1 COL NL 43 94 9 17 3 0
81 89480 embreal01 2007 1 OAK AL 4 0 0 0 0 0
82 89481 edmonji01 2007 1 SLN NL 117 365 39 92 15 2
83 89482 easleda01 2007 1 NYN NL 76 193 24 54 6 0
84 89489 delgaca01 2007 1 NYN NL 139 538 71 139 30 0
85 89493 cormirh01 2007 1 CIN NL 6 0 0 0 0 0
86 89494 coninje01 2007 2 NYN NL 21 41 2 8 2 0
87 89495 coninje01 2007 1 CIN NL 80 215 23 57 11 1
88 89497 clemero02 2007 1 NYA AL 2 2 0 1 0 0
89 89498 claytro01 2007 2 BOS AL 8 6 1 0 0 0
90 89499 claytro01 2007 1 TOR AL 69 189 23 48 14 0
91 89501 cirilje01 2007 2 ARI NL 28 40 6 8 4 0
92 89502 cirilje01 2007 1 MIN AL 50 153 18 40 9 2
93 89521 bondsba01 2007 1 SFN NL 126 340 75 94 14 0
94 89523 biggicr01 2007 1 HOU NL 141 517 68 130 31 3
95 89525 benitar01 2007 2 FLO NL 34 0 0 0 0 0
96 89526 benitar01 2007 1 SFN NL 19 0 0 0 0 0
97 89530 ausmubr01 2007 1 HOU NL 117 349 38 82 16 3
98 89533 aloumo01 2007 1 NYN NL 87 328 51 112 19 1
99 89534 alomasa02 2007 1 NYN NL 8 22 1 3 1 0

Wide DataFrames will be printed across multiple rows by default:

In [123]: pd.DataFrame(np.random.randn(3, 12))
Out[123]:
0 1 2 3 4 ... 7 8 9
,→ 10 11
0 -1.533705 -0.176507 0.608559 0.978633 -1.593325 ... 0.648689 0.311338 1.
,→019847 -0.905475 0.979426
1 0.301441 1.043466 -0.585706 -0.300476 -0.631623 ... 1.778709 0.084805 0.
,→480836 0.003515 0.726600
2 0.261610 -0.446297 1.996565 0.344258 1.052674 ... -1.230991 -0.413491 -0.
,→145399 -0.301483 1.844836

[3 rows x 12 columns]

You can change how much to print on a single row by setting the display.width option:
In [124]: pd.set_option('display.width', 40) # default is 80

In [125]: pd.DataFrame(np.random.randn(3, 12))

Out[125]:
0 1 2 3 4 ... 7 8 9
,→ 10 11
0 0.821556 -0.392733 0.317696 -3.065966 0.216755 ... 0.883135 -1.962886 -1.
,→096080 -0.787760 -0.290642

(continues on next page)

126 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

1 -0.127228 1.021397 0.061536 0.340173 -0.447509 ... 0.268224 -1.339587 -0.
,→476597 0.180935 0.841153
2 2.699198 0.291861 -0.295606 -1.319987 -0.114294 ... 1.565075 -0.317159 -1.
,→399831 -0.254755 0.225071

[3 rows x 12 columns]

You can adjust the max width of the individual columns by setting display.max_colwidth

In [126]: datafile = {'filename': ['filename_01', 'filename_02'],

.....: 'path': ["media/user_name/storage/folder_01/filename_01",
.....: "media/user_name/storage/folder_02/filename_02"]}
.....:

In [127]: pd.set_option('display.max_colwidth', 30)

In [128]: pd.DataFrame(datafile)
Out[128]:
filename path
0 filename_01 media/user_name/storage/fo...
1 filename_02 media/user_name/storage/fo...

In [129]: pd.set_option('display.max_colwidth', 100)

In [130]: pd.DataFrame(datafile)
Out[130]:
filename path
0 filename_01 media/user_name/storage/folder_01/filename_01
1 filename_02 media/user_name/storage/folder_02/filename_02

You can also disable this feature via the expand_frame_repr option. This will print the table in one block.

DataFrame column attribute access and IPython completion

If a DataFrame column label is a valid Python variable name, the column can be accessed like an attribute:
In [131]: df = pd.DataFrame({'foo1': np.random.randn(5),
.....: 'foo2': np.random.randn(5)})
.....:

In [132]: df
Out[132]:
foo1 foo2
0 0.486038 -0.803225
1 -0.223749 -1.166967
2 0.128104 0.344985
3 -0.367409 -1.789576
4 -0.684883 -1.035161

In [133]: df.foo1
Out[133]:
0 0.486038
1 -0.223749
2 0.128104
3 -0.367409

3.4. Intro to data structures 127

pandas: powerful Python data analysis toolkit, Release 0.25.3

4 -0.684883
Name: foo1, dtype: float64
The columns are also connected to the IPython completion mechanism so they can be tab-completed:

In [5]: df.fo<TAB> # noqa: E225, E999

df.foo1 df.foo2

{{ header }}

3.5 Comparison with other tools

{{ header }}

3.5.1 Comparison with R / R libraries

Since pandas aims to provide a lot of the data manipulation and analysis functionality that people use R for, this
page was started to provide a more detailed look at the R language and its many third party libraries as they relate to
pandas. In comparisons with R and CRAN libraries, we care about the following things:
• Functionality / flexibility: what can/cannot be done with each tool
• Performance: how fast are operations. Hard numbers/benchmarks are preferable
• Ease-of-use: Is one tool easier/harder to use (you may have to be the judge of this, given side-by-side code
comparisons)
This page is also here to offer a bit of a translation guide for users of these R packages.
For transfer of DataFrame objects from pandas to R, one option is to use HDF5 files, see External compatibility
for an example.

Quick reference

Well start off with a quick reference guide pairing some common R operations using dplyr with pandas equivalents.

Querying, filtering, sampling

R pandas
dim(df) df.shape
head(df) df.head()
slice(df, 1:10) df.iloc[:9]
filter(df, col1 == 1, col2 == 1) df.query('col1 == 1 & col2 == 1')
df[df$col1 == 1 & df$col2 == 1,] df[(df.col1 == 1) & (df.col2 == 1)]
select(df, col1, col2) df[['col1', 'col2']]
select(df, col1:col3) df.loc[:, 'col1':'col3']
select(df, -(col1:col3)) df.drop(cols_to_drop, axis=1) but see1
distinct(select(df, col1)) df[['col1']].drop_duplicates()
distinct(select(df, col1, col2)) df[['col1', 'col2']].drop_duplicates()
sample_n(df, 10) df.sample(n=10)
sample_frac(df, 0.01) df.sample(frac=0.01)

128 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Sorting

R pandas
arrange(df, col1, col2) df.sort_values(['col1', 'col2'])
arrange(df, desc(col1)) df.sort_values('col1', ascending=False)

Transforming

R pandas
select(df, col_one = df.rename(columns={'col1': 'col_one'})['col_one']
col1)
rename(df, col_one = df.rename(columns={'col1': 'col_one'})
col1)
mutate(df, c=a-b) df.assign(c=df.a-df.b)

Grouping and summarizing

R pandas
summary(df) df.describe()
gdf <- group_by(df, col1) gdf = df.groupby('col1')
summarise(gdf, avg=mean(col1, na. df.groupby('col1').agg({'col1':
rm=TRUE)) 'mean'})
summarise(gdf, total=sum(col1)) df.groupby('col1').sum()

Base R

Slicing with Rs c

R makes it easy to access data.frame columns by name

df <- data.frame(a=rnorm(5), b=rnorm(5), c=rnorm(5), d=rnorm(5), e=rnorm(5))

df[, c("a", "c", "e")]

or by integer location

df <- data.frame(matrix(rnorm(1000), ncol=100))

df[, c(1:10, 25:30, 40, 50:100)]

Selecting multiple columns by name in pandas is straightforward

In [1]: df = pd.DataFrame(np.random.randn(10, 3), columns=list('abc'))

In [2]: df[['a', 'c']]

Out[2]:
a c
0 0.207136 -0.469146
1 Rs shorthand for a subrange of columns (select(df, col1:col3)) can be approached cleanly in pandas, if you have the list of columns,

for example df[cols[1:3]] or df.drop(cols[1:3]), but doing this by column name is a bit messy.

3.5. Comparison with other tools 129

pandas: powerful Python data analysis toolkit, Release 0.25.3

1 1.385538 -0.362643
2 -0.571230 1.414622
3 -0.860474 0.448659
4 0.198087 -0.138916
5 -0.236168 0.368541
6 0.077176 1.707149
7 1.212901 0.166563
8 1.116482 1.600544
9 -1.464652 -0.306999

In [3]: df.loc[:, ['a', 'c']]

Out[3]:
a c
0 0.207136 -0.469146
1 1.385538 -0.362643
2 -0.571230 1.414622
3 -0.860474 0.448659
4 0.198087 -0.138916
5 -0.236168 0.368541
6 0.077176 1.707149
7 1.212901 0.166563
8 1.116482 1.600544
9 -1.464652 -0.306999
Selecting multiple noncontiguous columns by integer location can be achieved with a combination of the iloc indexer
attribute and numpy.r_.
In [4]: named = list('abcdefg')

In [5]: n = 30

In [6]: columns = named + np.arange(len(named), n).tolist()

In [7]: df = pd.DataFrame(np.random.randn(n, n), columns=columns)

In [8]: df.iloc[:, np.r_[:10, 24:30]]

Out[8]:
a b c d e ... 25 26
,→27 28 29
0 -0.157020 -0.312100 2.101265 0.715737 0.161958 ... -0.592069 0.892628 -1.
,→426348 0.859143 -1.247781
1 -0.648453 0.427706 -1.167462 -0.693429 1.902080 ... 1.237770 0.661616 1.
,→144256 0.447000 0.198894
2 -0.574805 -1.204465 -0.201374 0.278877 -0.286763 ... -0.106527 -0.456767 0.
,→706188 0.133951 0.792075
3 -0.938013 -0.552114 -0.266028 -0.838281 -0.138009 ... -0.120739 -0.195936 0.
,→089482 -0.052919 -0.123138

4 0.348967 -1.399891 -1.166618 0.608277 -1.297820 ... 0.879326 1.234826 1.

,→130073 -0.820004 -1.120951

5 0.003281 -0.998084 1.300842 1.437468 -0.261703 ... -1.455117 0.004933 -1.

,→507294 -1.165402 -1.945610

6 -1.951386 1.894199 1.466576 0.731260 -0.068252 ... 1.449550 0.197532 0.

,→139688 -1.246171 -0.690641

7 2.119767 0.461675 1.116264 0.099048 1.471643 ... -0.444326 -1.369138 2.

,→162595 1.986876 -0.458651
8 -0.471204 -2.008981 0.609188 1.187271 0.178697 ... 1.538607 -0.038627 -0.
,→442633 -0.577955 -0.161513
(continues on next page)

130 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

9 -1.080629 -1.573208 0.240054 -1.182698 1.380183 ... -0.158802 0.767407 -0.
,→541999 0.275594 -0.193256
10 1.212011 0.021855 0.787764 -0.111502 0.978365 ... 1.190005 -0.341142 0.
,→019357 -0.954364 -0.667115

11 -0.372254 -1.419194 1.284916 -2.567632 -0.394702 ... -0.076347 0.540433 -0.

,→272025 -0.397703 0.119171
12 0.782784 -0.120299 1.073066 -0.462378 0.709914 ... -2.521942 0.150246 -0.
,→365755 -0.004324 0.019941
13 -1.955448 -0.152551 -1.394527 -1.838862 0.116834 ... 0.678305 -0.439020 -0.
,→004064 0.211563 0.366738
14 -1.345105 -0.939819 -0.126120 -0.325684 -1.346863 ... -0.004781 1.457464 -0.
,→859323 -0.615783 0.958581
15 1.661150 0.229221 -0.207492 1.324775 0.095197 ... 0.337167 -0.536268 -1.
,→166466 -0.873786 0.970503
16 -1.721043 0.234624 1.131312 0.187112 -0.916265 ... -0.666787 -0.460773 0.
,→189052 -0.540372 -0.873749

17 0.530914 0.092203 0.004254 -1.445988 -1.683797 ... -0.338648 -1.736859 0.

,→924184 -0.086912 -1.122609

18 1.272979 0.039344 -0.723751 -0.926619 0.982875 ... 0.369011 0.601631 1.

,→047098 0.168620 1.486580
19 1.019541 1.022316 -0.989192 1.331725 -1.333733 ... 0.259385 -0.265833 0.
,→927853 1.725549 0.617201
20 0.211400 0.788539 0.191757 1.249484 -1.506977 ... -2.191752 0.028034 -0.
,→118943 1.828197 -1.588506
21 0.435479 -1.408082 -0.518882 0.547208 0.792259 ... 0.315643 1.377493 0.
,→694965 -0.107745 -0.824049

22 -0.265984 1.632759 -1.072779 1.101620 0.169399 ... 1.635833 -0.175988 0.

,→993764 -0.049508 -0.738073

23 -1.593320 0.940353 -0.641346 1.229798 -1.588487 ... 1.768347 -1.377505 -0.

,→013970 0.495098 0.928113
24 -1.044262 0.476848 -0.559825 0.128314 1.010380 ... -0.368830 -1.917762 0.
,→879227 2.214419 -1.710267
25 -0.656052 0.911241 -0.698254 -0.304604 -0.144525 ... 0.188098 -0.079259 -0.
,→655610 -0.699509 1.270956
26 1.079778 0.928320 -0.079028 0.163994 1.242867 ... -0.032398 -0.614531 0.
,→159650 1.161948 -0.289595
27 0.104740 -0.179138 -1.081585 0.366706 0.072617 ... -0.014595 0.042685 -0.
,→211296 1.429080 -0.246837
28 -0.683389 1.678525 -0.869742 -0.459486 -4.081324 ... -0.663183 0.054473 1.
,→520833 -2.107380 -0.497143

29 -1.643519 0.094721 1.377571 -1.433719 0.744651 ... 0.611787 0.655887 0.

,→327674 0.144085 -1.006701

[30 rows x 16 columns]

aggregate

In R you may want to split data into subsets and compute the mean for each. Using a data.frame called df and splitting
it into groups by1 and by2:
df <- data.frame(
v1 = c(1,3,5,7,8,3,5,NA,4,5,7,9),
v2 = c(11,33,55,77,88,33,55,NA,44,55,77,99),
by1 = c("red", "blue", 1, 2, NA, "big", 1, 2, "red", 1, NA, 12),
by2 = c("wet", "dry", 99, 95, NA, "damp", 95, 99, "red", 99, NA, NA))
(continues on next page)

3.5. Comparison with other tools 131

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

aggregate(x=df[, c("v1", "v2")], by=list(mydf2$by1, mydf2$by2), FUN = mean)

The groupby() method is similar to base R aggregate function.

In [9]: df = pd.DataFrame(
...: {'v1': [1, 3, 5, 7, 8, 3, 5, np.nan, 4, 5, 7, 9],
...: 'v2': [11, 33, 55, 77, 88, 33, 55, np.nan, 44, 55, 77, 99],
...: 'by1': ["red", "blue", 1, 2, np.nan, "big", 1, 2, "red", 1, np.nan, 12],
...: 'by2': ["wet", "dry", 99, 95, np.nan, "damp", 95, 99, "red", 99, np.nan,
...: np.nan]})
...:

In [10]: g = df.groupby(['by1', 'by2'])

In [11]: g[['v1', 'v2']].mean()

Out[11]:
v1 v2
by1 by2
1 95 5.0 55.0
99 5.0 55.0
2 95 7.0 77.0
99 NaN NaN
big damp 3.0 33.0
blue dry 3.0 33.0
red red 4.0 44.0
wet 1.0 11.0

For more details and examples see the groupby documentation.

match / %in%

A common way to select data in R is using %in% which is defined using the function match. The operator %in% is
used to return a logical vector indicating if there is a match or not:
s <- 0:4
s %in% c(2,4)

The isin() method is similar to R %in% operator:

In [12]: s = pd.Series(np.arange(5), dtype=np.float32)

In [13]: s.isin([2, 4])

Out[13]:
0 False
1 False
2 True
3 False
4 True
dtype: bool

The match function returns a vector of the positions of matches of its first argument in its second:
s <- 0:4
match(s, c(2,4))

For more details and examples see the reshaping documentation.

132 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

tapply

tapply is similar to aggregate, but data can be in a ragged array, since the subclass sizes are possibly irregular.
Using a data.frame called baseball, and retrieving information based on the array team:

baseball <-
data.frame(team = gl(5, 5,
labels = paste("Team", LETTERS[1:5])),
player = sample(letters, 25),
batting.average = runif(25, .200, .400))

tapply(baseball$batting.average, baseball.example$team,
max)

In pandas we may use pivot_table() method to handle this:

In [14]: import random

In [15]: import string

In [16]: baseball = pd.DataFrame(

....: {'team': ["team %d" % (x + 1) for x in range(5)] * 5,
....: 'player': random.sample(list(string.ascii_lowercase), 25),
....: 'batting avg': np.random.uniform(.200, .400, 25)})
....:

In [17]: baseball.pivot_table(values='batting avg', columns='team', aggfunc=np.max)

Out[17]:
team team 1 team 2 team 3 team 4 team 5
batting avg 0.287166 0.392928 0.372269 0.398628 0.393112

For more details and examples see the reshaping documentation.

subset

The query() method is similar to the base R subset function. In R you might want to get the rows of a data.
frame where one columns values are less than another columns values:

df <- data.frame(a=rnorm(10), b=rnorm(10))

subset(df, a <= b)
df[df$a <= df$b,] # note the comma

In pandas, there are a few ways to perform subsetting. You can use query() or pass an expression as if it were an
index/slice as well as standard boolean indexing:
In [18]: df = pd.DataFrame({'a': np.random.randn(10), 'b': np.random.
,→randn(10)})

In [19]: df.query('a <= b')

Out[19]:
a b
0 -1.829965 0.318380
6 0.533221 0.686366
9 0.290722 0.761557

In [20]: df[df.a <= df.b]

3.5. Comparison with other tools 133

pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[20]:
a b
0 -1.829965 0.318380
6 0.533221 0.686366
9 0.290722 0.761557

In [21]: df.loc[df.a <= df.b]

Out[21]:
a b
0 -1.829965 0.318380
6 0.533221 0.686366
9 0.290722 0.761557
For more details and examples see the query documentation.

with

An expression using a data.frame called df in R with the columns a and b would be evaluated using with like so:

df <- data.frame(a=rnorm(10), b=rnorm(10))

with(df, a + b)
df$a + df$b # same as the previous expression

In pandas the equivalent expression, using the eval() method, would be:
In [22]: df = pd.DataFrame({'a': np.random.randn(10), 'b': np.random.
,→randn(10)})

In [23]: df.eval('a + b')

Out[23]:
0 1.134063
1 -1.663533
2 2.394069
3 2.476911
4 1.059003
5 1.941878
6 -0.499256
7 1.955427
8 1.029021
9 -1.430095
dtype: float64

In [24]: df.a + df.b # same as the previous expression

Out[24]:
0 1.134063
1 -1.663533
2 2.394069
3 2.476911
4 1.059003
5 1.941878
6 -0.499256
7 1.955427
8 1.029021
9 -1.430095

134 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

dtype: float64
In certain cases eval() will be much faster than evaluation in pure Python. For more details and examples see the
eval documentation.

plyr

plyr is an R library for the split-apply-combine strategy for data analysis. The functions revolve around three data
structures in R, a for arrays, l for lists, and d for data.frame. The table below shows how these data
structures could be mapped in Python.

R Python
array list
lists dictionary or list of objects
data.frame dataframe

ddply

An expression using a data.frame called df in R where you want to summarize x by month:

require(plyr)
df <- data.frame(
x = runif(120, 1, 168),
y = runif(120, 7, 334),
z = runif(120, 1.7, 20.7),
month = rep(c(5,6,7,8),30),
week = sample(1:4, 120, TRUE)
)

ddply(df, .(month, week), summarize,

mean = round(mean(x), 2),
sd = round(sd(x), 2))

In pandas the equivalent expression, using the groupby() method, would be:

In [25]: df = pd.DataFrame({'x': np.random.uniform(1., 168., 120),

....: 'y': np.random.uniform(7., 334., 120),
....: 'z': np.random.uniform(1.7, 20.7, 120),
....: 'month': [5, 6, 7, 8] * 30,
....: 'week': np.random.randint(1, 4, 120)})
....:

In [26]: grouped = df.groupby(['month', 'week'])

In [27]: grouped['x'].agg([np.mean, np.std])

Out[27]:
mean std
month week
5 1 54.610775 24.673729
2 79.400383 54.221977
3 87.966804 55.426962
6 1 75.458265 40.219732
2 63.014659 30.576905
3 82.579357 56.435221
(continues on next page)

3.5. Comparison with other tools 135

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

7 1 66.587862 40.926231
2 92.141920 45.262135
3 81.048190 48.547999
8 1 89.249868 50.984085
2 84.233718 51.690704
3 96.823561 43.465622

For more details and examples see the groupby documentation.

reshape / reshape2

melt.array

An expression using a 3 dimensional array called a in R where you want to melt it into a data.frame:

a <- array(c(1:23, NA), c(2,3,4))

data.frame(melt(a))

In Python, since a is a list, you can simply use list comprehension.

In [28]: a = np.array(list(range(1, 24)) + [np.NAN]).reshape(2, 3, 4)

In [29]: pd.DataFrame([tuple(list(x) + [val]) for x, val in np.ndenumerate(a)])

Out[29]:
0 1 2 3
0 0 0 0 1.0
1 0 0 1 2.0
2 0 0 2 3.0
3 0 0 3 4.0
4 0 1 0 5.0
5 0 1 1 6.0
6 0 1 2 7.0
7 0 1 3 8.0
8 0 2 0 9.0
9 0 2 1 10.0
10 0 2 2 11.0
11 0 2 3 12.0
12 1 0 0 13.0
13 1 0 1 14.0
14 1 0 2 15.0
15 1 0 3 16.0
16 1 1 0 17.0
17 1 1 1 18.0
18 1 1 2 19.0
19 1 1 3 20.0
20 1 2 0 21.0
21 1 2 1 22.0
22 1 2 2 23.0
23 1 2 3 NaN

melt.list

An expression using a list called a in R where you want to melt it into a data.frame:

136 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

a <- as.list(c(1:4, NA))

data.frame(melt(a))

In Python, this list would be a list of tuples, so DataFrame() method would convert it to a dataframe as required.

In [30]: a = list(enumerate(list(range(1, 5)) + [np.NAN]))

In [31]: pd.DataFrame(a)
Out[31]:
0 1
0 0 1.0
1 1 2.0
2 2 3.0
3 3 4.0
4 4 NaN

For more details and examples see the Into to Data Structures documentation.

melt.data.frame

An expression using a data.frame called cheese in R where you want to reshape the data.frame:

cheese <- data.frame(

first = c('John', 'Mary'),
last = c('Doe', 'Bo'),
height = c(5.5, 6.0),
weight = c(130, 150)
)
melt(cheese, id=c("first", "last"))

In Python, the melt() method is the R equivalent:

In [32]: cheese = pd.DataFrame({'first': ['John', 'Mary'],
....: 'last': ['Doe', 'Bo'],
....: 'height': [5.5, 6.0],
....: 'weight': [130, 150]})
....:

In [33]: pd.melt(cheese, id_vars=['first', 'last'])

Out[33]:
first last variable value
0 John Doe height 5.5
1 Mary Bo height 6.0
2 John Doe weight 130.0
3 Mary Bo weight 150.0

In [34]: cheese.set_index(['first', 'last']).stack() # alternative way

Out[34]:
first last
John Doe height 5.5
weight 130.0
Mary Bo height 6.0
weight 150.0
dtype: float64

3.5. Comparison with other tools 137

pandas: powerful Python data analysis toolkit, Release 0.25.3

For more details and examples see the reshaping documentation.

cast

In R acast is an expression using a data.frame called df in R to cast into a higher dimensional array:

df <- data.frame(
x = runif(12, 1, 168),
y = runif(12, 7, 334),
z = runif(12, 1.7, 20.7),
month = rep(c(5,6,7),4),
week = rep(c(1,2), 6)
)

mdf <- melt(df, id=c("month", "week"))

acast(mdf, week ~ month ~ variable, mean)

In Python the best way is to make use of pivot_table():

In [35]: df = pd.DataFrame({'x': np.random.uniform(1., 168., 12),

....: 'y': np.random.uniform(7., 334., 12),
....: 'z': np.random.uniform(1.7, 20.7, 12),
....: 'month': [5, 6, 7] * 4,
....: 'week': [1, 2] * 6})
....:

In [36]: mdf = pd.melt(df, id_vars=['month', 'week'])

In [37]: pd.pivot_table(mdf, values='value', index=['variable', 'week'],

....: columns=['month'], aggfunc=np.mean)
....:
Out[37]:
month 5 6 7
variable week
x 1 114.080088 85.737423 106.491489
2 51.936679 42.642634 112.288362
y 1 119.753362 196.773182 177.983139
2 289.007313 183.868930 217.038570
z 1 7.881457 5.557442 10.480275
2 10.927325 13.645086 7.487791

Similarly for dcast which uses a data.frame called df in R to aggregate information based on Animal and
FeedType:

df <- data.frame(
Animal = c('Animal1', 'Animal2', 'Animal3', 'Animal2', 'Animal1',
'Animal2', 'Animal3'),
FeedType = c('A', 'B', 'A', 'A', 'B', 'B', 'A'),
Amount = c(10, 7, 4, 2, 5, 6, 2)
)

dcast(df, Animal ~ FeedType, sum, fill=NaN)

# Alternative method using base R
with(df, tapply(Amount, list(Animal, FeedType), sum))

Python can approach this in two different ways. Firstly, similar to above using pivot_table():

138 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [38]: df = pd.DataFrame({
....: 'Animal': ['Animal1', 'Animal2', 'Animal3', 'Animal2', 'Animal1',
....: 'Animal2', 'Animal3'],
....: 'FeedType': ['A', 'B', 'A', 'A', 'B', 'B', 'A'],
....: 'Amount': [10, 7, 4, 2, 5, 6, 2],
....: })
....:

In [39]: df.pivot_table(values='Amount', index='Animal', columns='FeedType',

....: aggfunc='sum')
....:
Out[39]:
FeedType A B
Animal
Animal1 10.0 5.0
Animal2 2.0 13.0
Animal3 6.0 NaN

The second approach is to use the groupby() method:

In [40]: df.groupby(['Animal', 'FeedType'])['Amount'].sum()

Out[40]:
Animal FeedType
Animal1 A 10
B 5
Animal2 A 2
B 13
Animal3 A 6
Name: Amount, dtype: int64

For more details and examples see the reshaping documentation or the groupby documentation.

factor

pandas has a data type for categorical data.

cut(c(1,2,3,4,5,6), 3)
factor(c(1,2,3,2,2,3))

In pandas this is accomplished with pd.cut and astype("category"):

In [41]: pd.cut(pd.Series([1, 2, 3, 4, 5, 6]), 3)
Out[41]:
0 (0.995, 2.667]
1 (0.995, 2.667]
2 (2.667, 4.333]
3 (2.667, 4.333]
4 (4.333, 6.0]
5 (4.333, 6.0]
dtype: category
Categories (3, interval[float64]): [(0.995, 2.667] < (2.667, 4.333] < (4.333,
,→6.0]]

In [42]: pd.Series([1, 2, 3, 2, 2, 3]).astype("category")

Out[42]:
0 1

3.5. Comparison with other tools 139

pandas: powerful Python data analysis toolkit, Release 0.25.3

1 2
2 3
3 2
4 2
5 3
dtype: category
Categories (3, int64): [1, 2, 3]
For more details and examples see categorical introduction and the API documentation. There is also a documentation
regarding the differences to Rs factor. {{ header }}

3.5.2 Comparison with SQL

Since many potential pandas users have some familiarity with SQL, this page is meant to provide some examples of
how various SQL operations would be performed using pandas.
If youre new to pandas, you might want to first read through 10 Minutes to pandas to familiarize yourself with the
library.
As is customary, we import pandas and NumPy as follows:

In [1]: import pandas as pd

In [2]: import numpy as np

Most of the examples will utilize the tips dataset found within pandas tests. Well read the data into a DataFrame
called tips and assume we have a database table of the same name and structure.

In [3]: url = ('https://raw.github.com/pandas-dev'

...: '/pandas/master/pandas/tests/data/tips.csv')
...:

In [4]: tips = pd.read_csv(url)

In [5]: tips.head()
Out[5]:
total_bill tip sex smoker day time size
0 16.99 1.01 Female No Sun Dinner 2
1 10.34 1.66 Male No Sun Dinner 3
2 21.01 3.50 Male No Sun Dinner 3
3 23.68 3.31 Male No Sun Dinner 2
4 24.59 3.61 Female No Sun Dinner 4

SELECT

In SQL, selection is done using a comma-separated list of columns youd like to select (or a * to select all columns):

SELECT total_bill, tip, smoker, time

FROM tips
LIMIT 5;

With pandas, column selection is done by passing a list of column names to your DataFrame:

In [6]: tips[['total_bill', 'tip', 'smoker', 'time']].head(5)

Out[6]:
(continues on next page)

140 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

total_bill tip smoker time
0 16.99 1.01 No Dinner
1 10.34 1.66 No Dinner
2 21.01 3.50 No Dinner
3 23.68 3.31 No Dinner
4 24.59 3.61 No Dinner

Calling the DataFrame without the list of column names would display all columns (akin to SQLs *).

WHERE

Filtering in SQL is done via a WHERE clause.

SELECT *
FROM tips
WHERE time = 'Dinner'
LIMIT 5;

DataFrames can be filtered in multiple ways; the most intuitive of which is using boolean indexing.

In [7]: tips[tips['time'] == 'Dinner'].head(5)

Out[7]:
total_bill tip sex smoker day time size
0 16.99 1.01 Female No Sun Dinner 2
1 10.34 1.66 Male No Sun Dinner 3
2 21.01 3.50 Male No Sun Dinner 3
3 23.68 3.31 Male No Sun Dinner 2
4 24.59 3.61 Female No Sun Dinner 4

The above statement is simply passing a Series of True/False objects to the DataFrame, returning all rows with
True.
In [8]: is_dinner = tips['time'] == 'Dinner'

In [9]: is_dinner.value_counts()
Out[9]:
True 176
False 68
Name: time, dtype: int64

In [10]: tips[is_dinner].head(5)
Out[10]:
total_bill tip sex smoker day time size
0 16.99 1.01 Female No Sun Dinner 2
1 10.34 1.66 Male No Sun Dinner 3
2 21.01 3.50 Male No Sun Dinner 3
3 23.68 3.31 Male No Sun Dinner 2
4 24.59 3.61 Female No Sun Dinner 4
Just like SQLs OR and AND, multiple conditions can be passed to a DataFrame using | (OR) and & (AND).

-- tips of more than $5.00 at Dinner meals

SELECT *
FROM tips
WHERE time = 'Dinner' AND tip > 5.00;

3.5. Comparison with other tools 141

pandas: powerful Python data analysis toolkit, Release 0.25.3

# tips of more than $5.00 at Dinner meals

In [11]: tips[(tips['time'] == 'Dinner') & (tips['tip'] > 5.00)]
Out[11]:
total_bill tip sex smoker day time size
23 39.42 7.58 Male No Sat Dinner 4
44 30.40 5.60 Male No Sun Dinner 4
47 32.40 6.00 Male No Sun Dinner 4
52 34.81 5.20 Female No Sun Dinner 4
59 48.27 6.73 Male No Sat Dinner 4
116 29.93 5.07 Male No Sun Dinner 4
155 29.85 5.14 Female No Sun Dinner 5
170 50.81 10.00 Male Yes Sat Dinner 3
172 7.25 5.15 Male Yes Sun Dinner 2
181 23.33 5.65 Male Yes Sun Dinner 2
183 23.17 6.50 Male Yes Sun Dinner 4
211 25.89 5.16 Male Yes Sat Dinner 4
212 48.33 9.00 Male No Sat Dinner 4
214 28.17 6.50 Female Yes Sat Dinner 3
239 29.03 5.92 Male No Sat Dinner 3

-- tips by parties of at least 5 diners OR bill total was more than $45
SELECT *
FROM tips
WHERE size >= 5 OR total_bill > 45;

# tips by parties of at least 5 diners OR bill total was more than $45
In [12]: tips[(tips['size'] >= 5) | (tips['total_bill'] > 45)]
Out[12]:
total_bill tip sex smoker day time size
59 48.27 6.73 Male No Sat Dinner 4
125 29.80 4.20 Female No Thur Lunch 6
141 34.30 6.70 Male No Thur Lunch 6
142 41.19 5.00 Male No Thur Lunch 5
143 27.05 5.00 Female No Thur Lunch 6
155 29.85 5.14 Female No Sun Dinner 5
156 48.17 5.00 Male No Sun Dinner 6
170 50.81 10.00 Male Yes Sat Dinner 3
182 45.35 3.50 Male Yes Sun Dinner 3
185 20.69 5.00 Male No Sun Dinner 5
187 30.46 2.00 Male Yes Sun Dinner 5
212 48.33 9.00 Male No Sat Dinner 4
216 28.15 3.00 Male Yes Sat Dinner 5

NULL checking is done using the notna() and isna() methods.

In [13]: frame = pd.DataFrame({'col1': ['A', 'B', np.NaN, 'C', 'D'],

....: 'col2': ['F', np.NaN, 'G', 'H', 'I']})
....:

In [14]: frame
Out[14]:
col1 col2
0 A F
1 B NaN
2 NaN G
3 C H
4 D I

142 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Assume we have a table of the same structure as our DataFrame above. We can see only the records where col2 IS
NULL with the following query:

SELECT *
FROM frame
WHERE col2 IS NULL;

In [15]: frame[frame['col2'].isna()]
Out[15]:
col1 col2
1 B NaN

Getting items where col1 IS NOT NULL can be done with notna().

SELECT *
FROM frame
WHERE col1 IS NOT NULL;

In [16]: frame[frame['col1'].notna()]
Out[16]:
col1 col2
0 A F
1 B NaN
3 C H
4 D I

GROUP BY

In pandas, SQLs GROUP BY operations are performed using the similarly named groupby() method. groupby()
typically refers to a process where wed like to split a dataset into groups, apply some function (typically aggregation)
, and then combine the groups together.
A common SQL operation would be getting the count of records in each group throughout a dataset. For instance, a
query getting us the number of tips left by sex:

SELECT sex, count(*)

FROM tips
GROUP BY sex;
/*
Female 87
Male 157
*/

The pandas equivalent would be:

In [17]: tips.groupby('sex').size()
Out[17]:
sex
Female 87
Male 157
dtype: int64

Notice that in the pandas code we used size() and not count(). This is because count() applies the function
to each column, returning the number of not null records within each.

3.5. Comparison with other tools 143

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [18]: tips.groupby('sex').count()
Out[18]:
total_bill tip smoker day time size
sex
Female 87 87 87 87 87 87
Male 157 157 157 157 157 157

Alternatively, we could have applied the count() method to an individual column:

In [19]: tips.groupby('sex')['total_bill'].count()
Out[19]:
sex
Female 87
Male 157
Name: total_bill, dtype: int64

Multiple functions can also be applied at once. For instance, say wed like to see how tip amount differs by day of
the week - agg() allows you to pass a dictionary to your grouped DataFrame, indicating which functions to apply to
specific columns.

SELECT day, AVG(tip), COUNT(*)

FROM tips
GROUP BY day;
/*
Fri 2.734737 19
Sat 2.993103 87
Sun 3.255132 76
Thur 2.771452 62
*/

In [20]: tips.groupby('day').agg({'tip': np.mean, 'day': np.size})

Out[20]:
tip day
day
Fri 2.734737 19
Sat 2.993103 87
Sun 3.255132 76
Thur 2.771452 62

Grouping by more than one column is done by passing a list of columns to the groupby() method.

SELECT smoker, day, COUNT(*), AVG(tip)

FROM tips
GROUP BY smoker, day;
/*
smoker day
No Fri 4 2.812500
Sat 45 3.102889
Sun 57 3.167895
Thur 45 2.673778
Yes Fri 15 2.714000
Sat 42 2.875476
Sun 19 3.516842
Thur 17 3.030000
*/

144 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [21]: tips.groupby(['smoker', 'day']).agg({'tip': [np.size, np.mean]})

Out[21]:
tip
size mean
smoker day
No Fri 4.0 2.812500
Sat 45.0 3.102889
Sun 57.0 3.167895
Thur 45.0 2.673778
Yes Fri 15.0 2.714000
Sat 42.0 2.875476
Sun 19.0 3.516842
Thur 17.0 3.030000

JOIN

JOINs can be performed with join() or merge(). By default, join() will join the DataFrames on their indices.
Each method has parameters allowing you to specify the type of join to perform (LEFT, RIGHT, INNER, FULL) or
the columns to join on (column names or indices).
In [22]: df1 = pd.DataFrame({'key': ['A', 'B', 'C', 'D'],
....: 'value': np.random.randn(4)})
....:

In [23]: df2 = pd.DataFrame({'key': ['B', 'D', 'D', 'E'],

....: 'value': np.random.randn(4)})
....:

Assume we have two database tables of the same name and structure as our DataFrames.
Now lets go over the various types of JOINs.

INNER JOIN

SELECT *
FROM df1
INNER JOIN df2
ON df1.key = df2.key;

# merge performs an INNER JOIN by default

In [24]: pd.merge(df1, df2, on='key')
Out[24]:
key value_x value_y
0 B -1.152748 -0.878607
1 D -0.808546 -0.106710
2 D -0.808546 0.533295

merge() also offers parameters for cases when youd like to join one DataFrames column with another DataFrames
index.
In [25]: indexed_df2 = df2.set_index('key')

In [26]: pd.merge(df1, indexed_df2, left_on='key', right_index=True)

Out[26]:
(continues on next page)

3.5. Comparison with other tools 145

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

key value_x value_y
1 B -1.152748 -0.878607
3 D -0.808546 -0.106710
3 D -0.808546 0.533295

LEFT OUTER JOIN

-- show all records from df1

SELECT *
FROM df1
LEFT OUTER JOIN df2
ON df1.key = df2.key;

# show all records from df1

In [27]: pd.merge(df1, df2, on='key', how='left')
Out[27]:
key value_x value_y
0 A -1.421715 NaN
1 B -1.152748 -0.878607
2 C -0.129893 NaN
3 D -0.808546 -0.106710
4 D -0.808546 0.533295

RIGHT JOIN

-- show all records from df2

SELECT *
FROM df1
RIGHT OUTER JOIN df2
ON df1.key = df2.key;

# show all records from df2

In [28]: pd.merge(df1, df2, on='key', how='right')
Out[28]:
key value_x value_y
0 B -1.152748 -0.878607
1 D -0.808546 -0.106710
2 D -0.808546 0.533295
3 E NaN -1.408811

FULL JOIN

pandas also allows for FULL JOINs, which display both sides of the dataset, whether or not the joined columns find a
match. As of writing, FULL JOINs are not supported in all RDBMS (MySQL).

-- show all records from both tables

SELECT *
FROM df1
FULL OUTER JOIN df2
ON df1.key = df2.key;

146 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

# show all records from both frames

In [29]: pd.merge(df1, df2, on='key', how='outer')
Out[29]:
key value_x value_y
0 A -1.421715 NaN
1 B -1.152748 -0.878607
2 C -0.129893 NaN
3 D -0.808546 -0.106710
4 D -0.808546 0.533295
5 E NaN -1.408811

UNION

UNION ALL can be performed using concat().

In [30]: df1 = pd.DataFrame({'city': ['Chicago', 'San Francisco', 'New York City'],

....: 'rank': range(1, 4)})
....:

In [31]: df2 = pd.DataFrame({'city': ['Chicago', 'Boston', 'Los Angeles'],

....: 'rank': [1, 4, 5]})
....:

SELECT city, rank

FROM df1
UNION ALL
SELECT city, rank
FROM df2;
/*
city rank
Chicago 1
San Francisco 2
New York City 3
Chicago 1
Boston 4
Los Angeles 5
*/

In [32]: pd.concat([df1, df2])

Out[32]:
city rank
0 Chicago 1
1 San Francisco 2
2 New York City 3
0 Chicago 1
1 Boston 4
2 Los Angeles 5

SQLs UNION is similar to UNION ALL, however UNION will remove duplicate rows.

SELECT city, rank

FROM df1
UNION
SELECT city, rank
FROM df2;
(continues on next page)

3.5. Comparison with other tools 147

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

-- notice that there is only one Chicago record this time
/*
city rank
Chicago 1
San Francisco 2
New York City 3
Boston 4
Los Angeles 5
* /

In pandas, you can use concat() in conjunction with drop_duplicates().

In [33]: pd.concat([df1, df2]).drop_duplicates()

Out[33]:
city rank
0 Chicago 1
1 San Francisco 2
2 New York City 3
1 Boston 4
2 Los Angeles 5

Pandas equivalents for some SQL analytic and aggregate functions

Top N rows with offset

-- MySQL
SELECT * FROM tips
ORDER BY tip DESC
LIMIT 10 OFFSET 5;

In [34]: tips.nlargest(10 + 5, columns='tip').tail(10)

Out[34]:
total_bill tip sex smoker day time size
183 23.17 6.50 Male Yes Sun Dinner 4
214 28.17 6.50 Female Yes Sat Dinner 3
47 32.40 6.00 Male No Sun Dinner 4
239 29.03 5.92 Male No Sat Dinner 3
88 24.71 5.85 Male No Thur Lunch 2
181 23.33 5.65 Male Yes Sun Dinner 2
44 30.40 5.60 Male No Sun Dinner 4
52 34.81 5.20 Female No Sun Dinner 4
85 34.83 5.17 Female No Thur Lunch 4
211 25.89 5.16 Male Yes Sat Dinner 4

Top N rows per group

-- Oracle's ROW_NUMBER() analytic function

SELECT * FROM (
SELECT
t.*,
ROW_NUMBER() OVER(PARTITION BY day ORDER BY total_bill DESC) AS rn
(continues on next page)

148 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

FROM tips t
)
WHERE rn < 3
ORDER BY day, rn;

In [35]: (tips.assign(rn=tips.sort_values(['total_bill'], ascending=False)

....: .groupby(['day'])
....: .cumcount() + 1)
....: .query('rn < 3')
....: .sort_values(['day', 'rn']))
....:
Out[35]:
total_bill tip sex smoker day time size rn
95 40.17 4.73 Male Yes Fri Dinner 4 1
90 28.97 3.00 Male Yes Fri Dinner 2 2
170 50.81 10.00 Male Yes Sat Dinner 3 1
212 48.33 9.00 Male No Sat Dinner 4 2
156 48.17 5.00 Male No Sun Dinner 6 1
182 45.35 3.50 Male Yes Sun Dinner 3 2
197 43.11 5.00 Female Yes Thur Lunch 4 1
142 41.19 5.00 Male No Thur Lunch 5 2

the same using rank(method=first) function

In [36]: (tips.assign(rnk=tips.groupby(['day'])['total_bill']
....: .rank(method='first', ascending=False))
....: .query('rnk < 3')
....: .sort_values(['day', 'rnk']))
....:
Out[36]:
total_bill tip sex smoker day time size rnk
95 40.17 4.73 Male Yes Fri Dinner 4 1.0
90 28.97 3.00 Male Yes Fri Dinner 2 2.0
170 50.81 10.00 Male Yes Sat Dinner 3 1.0
212 48.33 9.00 Male No Sat Dinner 4 2.0
156 48.17 5.00 Male No Sun Dinner 6 1.0
182 45.35 3.50 Male Yes Sun Dinner 3 2.0
197 43.11 5.00 Female Yes Thur Lunch 4 1.0
142 41.19 5.00 Male No Thur Lunch 5 2.0

-- Oracle's RANK() analytic function

SELECT * FROM (
SELECT
t.*,
RANK() OVER(PARTITION BY sex ORDER BY tip) AS rnk
FROM tips t
WHERE tip < 2
)
WHERE rnk < 3
ORDER BY sex, rnk;

Lets find tips with (rank < 3) per gender group for (tips < 2). Notice that when using rank(method='min')
function rnk_min remains the same for the same tip (as Oracles RANK() function)
In [37]: (tips[tips['tip'] < 2]
....: .assign(rnk_min=tips.groupby(['sex'])['tip']
(continues on next page)

3.5. Comparison with other tools 149

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

....: .rank(method='min'))
....: .query('rnk_min < 3')
....: .sort_values(['sex', 'rnk_min']))
....:
Out[37]:
total_bill tip sex smoker day time size rnk_min
67 3.07 1.00 Female Yes Sat Dinner 1 1.0
92 5.75 1.00 Female Yes Fri Dinner 2 1.0
111 7.25 1.00 Female No Sat Dinner 1 1.0
236 12.60 1.00 Male Yes Sat Dinner 2 1.0
237 32.83 1.17 Male Yes Sat Dinner 2 2.0

UPDATE

UPDATE tips
SET tip = tip*2
WHERE tip < 2;

In [38]: tips.loc[tips['tip'] < 2, 'tip'] *= 2

DELETE

DELETE FROM tips

WHERE tip > 9;

In pandas we select the rows that should remain, instead of deleting them
In [39]: tips = tips.loc[tips['tip'] <= 9]

{{ header }}

3.5.3 Comparison with SAS

For potential users coming from SAS this page is meant to demonstrate how different SAS operations would be
performed in pandas.
If youre new to pandas, you might want to first read through 10 Minutes to pandas to familiarize yourself with the
library.
As is customary, we import pandas and NumPy as follows:
In [1]: import pandas as pd

In [2]: import numpy as np

Note: Throughout this tutorial, the pandas DataFrame will be displayed by calling df.head(), which displays
the first N (default 5) rows of the DataFrame. This is often used in interactive work (e.g. Jupyter notebook or
terminal) - the equivalent in SAS would be:
proc print data=df(obs=5);
run;

150 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Data structures

General terminology translation

pandas SAS
DataFrame data set
column variable
row observation
groupby BY-group
NaN .

DataFrame / Series

A DataFrame in pandas is analogous to a SAS data set - a two-dimensional data source with labeled columns that
can be of different types. As will be shown in this document, almost any operation that can be applied to a data set
using SASs DATA step, can also be accomplished in pandas.
A Series is the data structure that represents one column of a DataFrame. SAS doesnt have a separate data
structure for a single column, but in general, working with a Series is analogous to referencing a column in the
DATA step.

Index

Every DataFrame and Series has an Index - which are labels on the rows of the data. SAS does not have an
exactly analogous concept. A data sets rows are essentially unlabeled, other than an implicit integer index that can be
accessed during the DATA step (_N_).
In pandas, if no index is specified, an integer index is also used by default (first row = 0, second row = 1, and so on).
While using a labeled Index or MultiIndex can enable sophisticated analyses and is ultimately an important part
of pandas to understand, for this comparison we will essentially ignore the Index and just treat the DataFrame as
a collection of columns. Please see the indexing documentation for much more on how to use an Index effectively.

Data input / output

Constructing a DataFrame from values

A SAS data set can be built from specified values by placing the data after a datalines statement and specifying
the column names.

data df;
input x y;
datalines;
1 2
3 4
5 6
;
run;

3.5. Comparison with other tools 151

pandas: powerful Python data analysis toolkit, Release 0.25.3

A pandas DataFrame can be constructed in many different ways, but for a small number of values, it is often
convenient to specify it as a Python dictionary, where the keys are the column names and the values are the data.

In [3]: df = pd.DataFrame({'x': [1, 3, 5], 'y': [2, 4, 6]})

In [4]: df
Out[4]:
x y
0 1 2
1 3 4
2 5 6

Reading external data

Like SAS, pandas provides utilities for reading in data from many formats. The tips dataset, found within the pandas
tests (csv) will be used in many of the following examples.
SAS provides PROC IMPORT to read csv data into a data set.

proc import datafile='tips.csv' dbms=csv out=tips replace;

getnames=yes;
run;

The pandas method is read_csv(), which works similarly.

In [5]: url = ('https://raw.github.com/pandas-dev/'

...: 'pandas/master/pandas/tests/data/tips.csv')
...:

In [6]: tips = pd.read_csv(url)

In [7]: tips.head()
Out[7]:
total_bill tip sex smoker day time size
0 16.99 1.01 Female No Sun Dinner 2
1 10.34 1.66 Male No Sun Dinner 3
2 21.01 3.50 Male No Sun Dinner 3
3 23.68 3.31 Male No Sun Dinner 2
4 24.59 3.61 Female No Sun Dinner 4

Like PROC IMPORT, read_csv can take a number of parameters to specify how the data should be parsed. For
example, if the data was instead tab delimited, and did not have column names, the pandas command would be:

tips = pd.read_csv('tips.csv', sep='\t', header=None)

# alternatively, read_table is an alias to read_csv with tab delimiter

tips = pd.read_table('tips.csv', header=None)

In addition to text/csv, pandas supports a variety of other data formats such as Excel, HDF5, and SQL databases. These
are all read via a pd.read_* function. See the IO documentation for more details.

Exporting data

The inverse of PROC IMPORT in SAS is PROC EXPORT

152 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

proc export data=tips outfile='tips2.csv' dbms=csv;

run;

Similarly in pandas, the opposite of read_csv is to_csv(), and other data formats follow a similar api.

tips.to_csv('tips2.csv')

Data operations

Operations on columns

In the DATA step, arbitrary math expressions can be used on new or existing columns.

data tips;
set tips;
total_bill = total_bill - 2;
new_bill = total_bill / 2;
run;

pandas provides similar vectorized operations by specifying the individual Series in the DataFrame. New
columns can be assigned in the same way.

In [8]: tips['total_bill'] = tips['total_bill'] - 2

In [9]: tips['new_bill'] = tips['total_bill'] / 2.0

In [10]: tips.head()
Out[10]:
total_bill tip sex smoker day time size new_bill
0 14.99 1.01 Female No Sun Dinner 2 7.495
1 8.34 1.66 Male No Sun Dinner 3 4.170
2 19.01 3.50 Male No Sun Dinner 3 9.505
3 21.68 3.31 Male No Sun Dinner 2 10.840
4 22.59 3.61 Female No Sun Dinner 4 11.295

Filtering

Filtering in SAS is done with an if or where statement, on one or more columns.

data tips;
set tips;
if total_bill > 10;
run;

data tips;
set tips;
where total_bill > 10;
/* equivalent in this case - where happens before the
DATA step begins and can also be used in PROC statements */
run;

DataFrames can be filtered in multiple ways; the most intuitive of which is using boolean indexing

3.5. Comparison with other tools 153

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [11]: tips[tips['total_bill'] > 10].head()

Out[11]:
total_bill tip sex smoker day time size
0 14.99 1.01 Female No Sun Dinner 2
2 19.01 3.50 Male No Sun Dinner 3
3 21.68 3.31 Male No Sun Dinner 2
4 22.59 3.61 Female No Sun Dinner 4
5 23.29 4.71 Male No Sun Dinner 4

If/then logic

In SAS, if/then logic can be used to create new columns.

data tips;
set tips;
format bucket $4.;

if total_bill < 10 then bucket = 'low';

else bucket = 'high';
run;

The same operation in pandas can be accomplished using the where method from numpy.

In [12]: tips['bucket'] = np.where(tips['total_bill'] < 10, 'low', 'high')

In [13]: tips.head()
Out[13]:
total_bill tip sex smoker day time size bucket
0 14.99 1.01 Female No Sun Dinner 2 high
1 8.34 1.66 Male No Sun Dinner 3 low
2 19.01 3.50 Male No Sun Dinner 3 high
3 21.68 3.31 Male No Sun Dinner 2 high
4 22.59 3.61 Female No Sun Dinner 4 high

Date functionality

SAS provides a variety of functions to do operations on date/datetime columns.

data tips;
set tips;
format date1 date2 date1_plusmonth mmddyy10.;
date1 = mdy(1, 15, 2013);
date2 = mdy(2, 15, 2015);
date1_year = year(date1);
date2_month = month(date2);
* shift date to beginning of next interval;
date1_next = intnx('MONTH', date1, 1);
* count intervals between dates;
months_between = intck('MONTH', date1, date2);
run;

The equivalent pandas operations are shown below. In addition to these functions pandas supports other Time Series
features not available in Base SAS (such as resampling and custom offsets) - see the timeseries documentation for
more details.

154 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [14]: tips['date1'] = pd.Timestamp('2013-01-15')

In [15]: tips['date2'] = pd.Timestamp('2015-02-15')

In [16]: tips['date1_year'] = tips['date1'].dt.year

In [17]: tips['date2_month'] = tips['date2'].dt.month

In [18]: tips['date1_next'] = tips['date1'] + pd.offsets.MonthBegin()

In [19]: tips['months_between'] = (
....: tips['date2'].dt.to_period('M') - tips['date1'].dt.to_period('M'))
....:

In [20]: tips[['date1', 'date2', 'date1_year', 'date2_month',

....: 'date1_next', 'months_between']].head()
....:
Out[20]:
date1 date2 date1_year date2_month date1_next months_between
0 2013-01-15 2015-02-15 2013 2 2013-02-01 <25 * MonthEnds>
1 2013-01-15 2015-02-15 2013 2 2013-02-01 <25 * MonthEnds>
2 2013-01-15 2015-02-15 2013 2 2013-02-01 <25 * MonthEnds>
3 2013-01-15 2015-02-15 2013 2 2013-02-01 <25 * MonthEnds>
4 2013-01-15 2015-02-15 2013 2 2013-02-01 <25 * MonthEnds>

Selection of columns

SAS provides keywords in the DATA step to select, drop, and rename columns.

data tips;
set tips;
keep sex total_bill tip;
run;

data tips;
set tips;
drop sex;
run;

data tips;
set tips;
rename total_bill=total_bill_2;
run;

The same operations are expressed in pandas below.

# keep
In [21]: tips[['sex', 'total_bill', 'tip']].head()
Out[21]:
sex total_bill tip
0 Female 14.99 1.01
1 Male 8.34 1.66
2 Male 19.01 3.50
3 Male 21.68 3.31
4 Female 22.59 3.61

3.5. Comparison with other tools 155

pandas: powerful Python data analysis toolkit, Release 0.25.3

# drop
In [22]: tips.drop('sex', axis=1).head()
Out[22]:
total_bill tip smoker day time size
0 14.99 1.01 No Sun Dinner 2
1 8.34 1.66 No Sun Dinner 3
2 19.01 3.50 No Sun Dinner 3
3 21.68 3.31 No Sun Dinner 2
4 22.59 3.61 No Sun Dinner 4

# rename
In [23]: tips.rename(columns={'total_bill': 'total_bill_2'}).head()
Out[23]:
total_bill_2 tip sex smoker day time size
0 14.99 1.01 Female No Sun Dinner 2
1 8.34 1.66 Male No Sun Dinner 3
2 19.01 3.50 Male No Sun Dinner 3
3 21.68 3.31 Male No Sun Dinner 2
4 22.59 3.61 Female No Sun Dinner 4

Sorting by values

Sorting in SAS is accomplished via PROC SORT

proc sort data=tips;

by sex total_bill;
run;

pandas objects have a sort_values() method, which takes a list of columns to sort by.

In [24]: tips = tips.sort_values(['sex', 'total_bill'])

In [25]: tips.head()
Out[25]:
total_bill tip sex smoker day time size
67 1.07 1.00 Female Yes Sat Dinner 1
92 3.75 1.00 Female Yes Fri Dinner 2
111 5.25 1.00 Female No Sat Dinner 1
145 6.35 1.50 Female No Thur Lunch 2
135 6.51 1.25 Female No Thur Lunch 2

String processing

Length

SAS determines the length of a character string with the LENGTHN and LENGTHC functions. LENGTHN excludes
trailing blanks and LENGTHC includes trailing blanks.

data _null_;
set tips;
put(LENGTHN(time));
put(LENGTHC(time));
run;

156 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Python determines the length of a character string with the len function. len includes trailing blanks. Use len and
rstrip to exclude trailing blanks.
In [26]: tips['time'].str.len().head()
Out[26]:
67 6
92 6
111 6
145 5
135 5
Name: time, dtype: int64

In [27]: tips['time'].str.rstrip().str.len().head()
Out[27]:
67 6
92 6
111 6
145 5
135 5
Name: time, dtype: int64

Find

SAS determines the position of a character in a string with the FINDW function. FINDW takes the string defined by
the first argument and searches for the first position of the substring you supply as the second argument.

data _null_;
set tips;
put(FINDW(sex,'ale'));
run;

Python determines the position of a character in a string with the find function. find searches for the first position
of the substring. If the substring is found, the function returns its position. Keep in mind that Python indexes are
zero-based and the function will return -1 if it fails to find the substring.

In [28]: tips['sex'].str.find("ale").head()
Out[28]:
67 3
92 3
111 3
145 3
135 3
Name: sex, dtype: int64

Substring

SAS extracts a substring from a string based on its position with the SUBSTR function.

data _null_;
set tips;
put(substr(sex,1,1));
run;

3.5. Comparison with other tools 157

pandas: powerful Python data analysis toolkit, Release 0.25.3

With pandas you can use [] notation to extract a substring from a string by position locations. Keep in mind that
Python indexes are zero-based.

In [29]: tips['sex'].str[0:1].head()
Out[29]:
67 F
92 F
111 F
145 F
135 F
Name: sex, dtype: object

Scan

The SAS SCAN function returns the nth word from a string. The first argument is the string you want to parse and the
second argument specifies which word you want to extract.

data firstlast;
input String $60.;
First_Name = scan(string, 1);
Last_Name = scan(string, -1);
datalines2;
John Smith;
Jane Cook;
;;;
run;

Python extracts a substring from a string based on its text by using regular expressions. There are much more powerful
approaches, but this just shows a simple approach.

In [30]: firstlast = pd.DataFrame({'String': ['John Smith', 'Jane Cook']})

In [31]: firstlast['First_Name'] = firstlast['String'].str.split(" ", expand=True)[0]

In [32]: firstlast['Last_Name'] = firstlast['String'].str.rsplit(" ", expand=True)[0]

In [33]: firstlast
Out[33]:
String First_Name Last_Name
0 John Smith John John
1 Jane Cook Jane Jane

Upcase, lowcase, and propcase

The SAS UPCASE LOWCASE and PROPCASE functions change the case of the argument.

data firstlast;
input String $60.;
string_up = UPCASE(string);
string_low = LOWCASE(string);
string_prop = PROPCASE(string);
datalines2;
John Smith;
Jane Cook;
(continues on next page)

158 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

;;;
run;

The equivalent Python functions are upper, lower, and title.

In [34]: firstlast = pd.DataFrame({'String': ['John Smith', 'Jane Cook']})

In [35]: firstlast['string_up'] = firstlast['String'].str.upper()

In [36]: firstlast['string_low'] = firstlast['String'].str.lower()

In [37]: firstlast['string_prop'] = firstlast['String'].str.title()

In [38]: firstlast
Out[38]:
String string_up string_low string_prop
0 John Smith JOHN SMITH john smith John Smith
1 Jane Cook JANE COOK jane cook Jane Cook

Merging

The following tables will be used in the merge examples

In [39]: df1 = pd.DataFrame({'key': ['A', 'B', 'C', 'D'],
....: 'value': np.random.randn(4)})
....:

In [40]: df1
Out[40]:
key value
0 A -1.579390
1 B 0.454513
2 C 0.051246
3 D 2.043664

In [41]: df2 = pd.DataFrame({'key': ['B', 'D', 'D', 'E'],

....: 'value': np.random.randn(4)})
....:

In [42]: df2
Out[42]:
key value
0 B -0.016440
1 D -1.413628
2 D 0.117613
3 E -1.797024

In SAS, data must be explicitly sorted before merging. Different types of joins are accomplished using the in= dummy
variables to track whether a match was found in one or both input frames.
proc sort data=df1;
by key;
run;

proc sort data=df2;

(continues on next page)

3.5. Comparison with other tools 159

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

by key;
run;

data left_join inner_join right_join outer_join;

merge df1(in=a) df2(in=b);

if a and b then output inner_join;

if a then output left_join;
if b then output right_join;
if a or b then output outer_join;
run;

pandas DataFrames have a merge() method, which provides similar functionality. Note that the data does not have
to be sorted ahead of time, and different join types are accomplished via the how keyword.
In [43]: inner_join = df1.merge(df2, on=['key'], how='inner')

In [44]: inner_join
Out[44]:
key value_x value_y
0 B 0.454513 -0.016440
1 D 2.043664 -1.413628
2 D 2.043664 0.117613

In [45]: left_join = df1.merge(df2, on=['key'], how='left')

In [46]: left_join
Out[46]:
key value_x value_y
0 A -1.579390 NaN
1 B 0.454513 -0.016440
2 C 0.051246 NaN
3 D 2.043664 -1.413628
4 D 2.043664 0.117613

In [47]: right_join = df1.merge(df2, on=['key'], how='right')

In [48]: right_join
Out[48]:
key value_x value_y
0 B 0.454513 -0.016440
1 D 2.043664 -1.413628
2 D 2.043664 0.117613
3 E NaN -1.797024

In [49]: outer_join = df1.merge(df2, on=['key'], how='outer')

In [50]: outer_join
Out[50]:
key value_x value_y
0 A -1.579390 NaN
1 B 0.454513 -0.016440
2 C 0.051246 NaN
3 D 2.043664 -1.413628
4 D 2.043664 0.117613
5 E NaN -1.797024

160 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Missing data

Like SAS, pandas has a representation for missing data - which is the special float value NaN (not a number). Many
of the semantics are the same, for example missing data propagates through numeric operations, and is ignored by
default for aggregations.
In [51]: outer_join
Out[51]:
key value_x value_y
0 A -1.579390 NaN
1 B 0.454513 -0.016440
2 C 0.051246 NaN
3 D 2.043664 -1.413628
4 D 2.043664 0.117613
5 E NaN -1.797024

In [52]: outer_join['value_x'] + outer_join['value_y']

Out[52]:
0 NaN
1 0.438072
2 NaN
3 0.630036
4 2.161278
5 NaN
dtype: float64

In [53]: outer_join['value_x'].sum()
Out[53]: 3.0136973094137085
One difference is that missing data cannot be compared to its sentinel value. For example, in SAS you could do this
to filter missing values.

data outer_join_nulls;
set outer_join;
if value_x = .;
run;

data outer_join_no_nulls;
set outer_join;
if value_x ^= .;
run;

Which doesnt work in pandas. Instead, the pd.isna or pd.notna functions should be used for comparisons.
In [54]: outer_join[pd.isna(outer_join['value_x'])]
Out[54]:
key value_x value_y
5 E NaN -1.797024

In [55]: outer_join[pd.notna(outer_join['value_x'])]
Out[55]:
key value_x value_y
0 A -1.579390 NaN
1 B 0.454513 -0.016440
2 C 0.051246 NaN
3 D 2.043664 -1.413628

3.5. Comparison with other tools 161

pandas: powerful Python data analysis toolkit, Release 0.25.3

4 D 2.043664 0.117613
pandas also provides a variety of methods to work with missing data - some of which would be challenging to express
in SAS. For example, there are methods to drop all rows with any missing values, replacing missing values with a
specified value, like the mean, or forward filling from previous rows. See the missing data documentation for more.
In [56]: outer_join.dropna()
Out[56]:
key value_x value_y
1 B 0.454513 -0.016440
3 D 2.043664 -1.413628
4 D 2.043664 0.117613

In [57]: outer_join.fillna(method='ffill')
Out[57]:
key value_x value_y
0 A -1.579390 NaN
1 B 0.454513 -0.016440
2 C 0.051246 -0.016440
3 D 2.043664 -1.413628
4 D 2.043664 0.117613
5 E 2.043664 -1.797024

In [58]: outer_join['value_x'].fillna(outer_join['value_x'].mean())
Out[58]:
0 -1.579390
1 0.454513
2 0.051246
3 2.043664
4 2.043664
5 0.602739
Name: value_x, dtype: float64

GroupBy

Aggregation

SASs PROC SUMMARY can be used to group by one or more key variables and compute aggregations on numeric
columns.

proc summary data=tips nway;

class sex smoker;
var total_bill tip;
output out=tips_summed sum=;
run;

pandas provides a flexible groupby mechanism that allows similar aggregations. See the groupby documentation for
more details and examples.

In [59]: tips_summed = tips.groupby(['sex', 'smoker'])['total_bill', 'tip'].sum()

In [60]: tips_summed.head()
Out[60]:
total_bill tip
(continues on next page)

162 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

sex smoker
Female No 869.68 149.77
Yes 527.27 96.74
Male No 1725.75 302.00
Yes 1217.07 183.07

Transformation

In SAS, if the group aggregations need to be used with the original frame, it must be merged back together. For
example, to subtract the mean for each observation by smoker group.
proc summary data=tips missing nway;
class smoker;
var total_bill;
output out=smoker_means mean(total_bill)=group_bill;
run;

proc sort data=tips;

by smoker;
run;

data tips;
merge tips(in=a) smoker_means(in=b);
by smoker;
adj_total_bill = total_bill - group_bill;
if a and b;
run;

pandas groupby provides a transform mechanism that allows these type of operations to be succinctly expressed
in one operation.
In [61]: gb = tips.groupby('smoker')['total_bill']

In [62]: tips['adj_total_bill'] = tips['total_bill'] - gb.transform('mean')

In [63]: tips.head()
Out[63]:
total_bill tip sex smoker day time size adj_total_bill
67 1.07 1.00 Female Yes Sat Dinner 1 -17.686344
92 3.75 1.00 Female Yes Fri Dinner 2 -15.006344
111 5.25 1.00 Female No Sat Dinner 1 -11.938278
145 6.35 1.50 Female No Thur Lunch 2 -10.838278
135 6.51 1.25 Female No Thur Lunch 2 -10.678278

By group processing

In addition to aggregation, pandas groupby can be used to replicate most other by group processing from SAS. For
example, this DATA step reads the data by sex/smoker group and filters to the first entry for each.
proc sort data=tips;
by sex smoker;
run;

(continues on next page)

3.5. Comparison with other tools 163

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

data tips_first;
set tips;
by sex smoker;
if FIRST.sex or FIRST.smoker then output;
run;

In pandas this would be written as:

In [64]: tips.groupby(['sex', 'smoker']).first()
Out[64]:
total_bill tip day time size adj_total_bill
sex smoker
Female No 5.25 1.00 Sat Dinner 1 -11.938278
Yes 1.07 1.00 Sat Dinner 1 -17.686344
Male No 5.51 2.00 Thur Lunch 2 -11.678278
Yes 5.25 5.15 Sun Dinner 2 -13.506344

Other Considerations

Disk vs memory

pandas operates exclusively in memory, where a SAS data set exists on disk. This means that the size of data able to
be loaded in pandas is limited by your machines memory, but also that the operations on that data may be faster.
If out of core processing is needed, one possibility is the dask.dataframe library (currently in development) which
provides a subset of pandas functionality for an on-disk DataFrame

Data interop

pandas provides a read_sas() method that can read SAS data saved in the XPORT or SAS7BDAT binary format.
libname xportout xport 'transport-file.xpt';
data xportout.tips;
set tips(rename=(total_bill=tbill));
* xport variable names limited to 6 characters;
run;

df = pd.read_sas('transport-file.xpt')
df = pd.read_sas('binary-file.sas7bdat')

You can also specify the file format directly. By default, pandas will try to infer the file format based on its extension.
df = pd.read_sas('transport-file.xpt', format='xport')
df = pd.read_sas('binary-file.sas7bdat', format='sas7bdat')

XPORT is a relatively limited format and the parsing of it is not as optimized as some of the other pandas readers. An
alternative way to interop data between SAS and pandas is to serialize to csv.
# version 0.17, 10M rows

In [8]: %time df = pd.read_sas('big.xpt')

Wall time: 14.6 s

(continues on next page)

164 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [9]: %time df = pd.read_csv('big.csv')
Wall time: 4.86 s

{{ header }}

3.5.4 Comparison with Stata

For potential users coming from Stata this page is meant to demonstrate how different Stata operations would be
performed in pandas.
If youre new to pandas, you might want to first read through 10 Minutes to pandas to familiarize yourself with the
library.
As is customary, we import pandas and NumPy as follows. This means that we can refer to the libraries as pd and np,
respectively, for the rest of the document.

In [1]: import pandas as pd

In [2]: import numpy as np

Note: Throughout this tutorial, the pandas DataFrame will be displayed by calling df.head(), which displays
the first N (default 5) rows of the DataFrame. This is often used in interactive work (e.g. Jupyter notebook or
terminal) – the equivalent in Stata would be:

list in 1/5

Data structures

General terminology translation

pandas Stata
DataFrame data set
column variable
row observation
groupby bysort
NaN .

DataFrame / Series

A DataFrame in pandas is analogous to a Stata data set – a two-dimensional data source with labeled columns that
can be of different types. As will be shown in this document, almost any operation that can be applied to a data set in
Stata can also be accomplished in pandas.
A Series is the data structure that represents one column of a DataFrame. Stata doesnt have a separate data
structure for a single column, but in general, working with a Series is analogous to referencing a column of a data
set in Stata.

3.5. Comparison with other tools 165

pandas: powerful Python data analysis toolkit, Release 0.25.3

Index

Every DataFrame and Series has an Index – labels on the rows of the data. Stata does not have an exactly
analogous concept. In Stata, a data sets rows are essentially unlabeled, other than an implicit integer index that can be
accessed with _n.
In pandas, if no index is specified, an integer index is also used by default (first row = 0, second row = 1, and so on).
While using a labeled Index or MultiIndex can enable sophisticated analyses and is ultimately an important part
of pandas to understand, for this comparison we will essentially ignore the Index and just treat the DataFrame as
a collection of columns. Please see the indexing documentation for much more on how to use an Index effectively.

Data input / output

Constructing a DataFrame from values

A Stata data set can be built from specified values by placing the data after an input statement and specifying the
column names.
input x y
1 2
3 4
5 6
end

A pandas DataFrame can be constructed in many different ways, but for a small number of values, it is often
convenient to specify it as a Python dictionary, where the keys are the column names and the values are the data.
In [3]: df = pd.DataFrame({'x': [1, 3, 5], 'y': [2, 4, 6]})

In [4]: df
Out[4]:
x y
0 1 2
1 3 4
2 5 6

Reading external data

Like Stata, pandas provides utilities for reading in data from many formats. The tips data set, found within the
pandas tests (csv) will be used in many of the following examples.
Stata provides import delimited to read csv data into a data set in memory. If the tips.csv file is in the
current working directory, we can import it as follows.
import delimited tips.csv

The pandas method is read_csv(), which works similarly. Additionally, it will automatically download the data
set if presented with a url.
In [5]: url = ('https://raw.github.com/pandas-dev'
...: '/pandas/master/pandas/tests/data/tips.csv')
...:

In [6]: tips = pd.read_csv(url)

(continues on next page)

166 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [7]: tips.head()
Out[7]:
total_bill tip sex smoker day time size
0 16.99 1.01 Female No Sun Dinner 2
1 10.34 1.66 Male No Sun Dinner 3
2 21.01 3.50 Male No Sun Dinner 3
3 23.68 3.31 Male No Sun Dinner 2
4 24.59 3.61 Female No Sun Dinner 4

Like import delimited, read_csv() can take a number of parameters to specify how the data should be
parsed. For example, if the data were instead tab delimited, did not have column names, and existed in the current
working directory, the pandas command would be:

tips = pd.read_csv('tips.csv', sep='\t', header=None)

# alternatively, read_table is an alias to read_csv with tab delimiter

tips = pd.read_table('tips.csv', header=None)

Pandas can also read Stata data sets in .dta format with the read_stata() function.

df = pd.read_stata('data.dta')

In addition to text/csv and Stata files, pandas supports a variety of other data formats such as Excel, SAS, HDF5,
Parquet, and SQL databases. These are all read via a pd.read_* function. See the IO documentation for more
details.

Exporting data

The inverse of import delimited in Stata is export delimited

export delimited tips2.csv

Similarly in pandas, the opposite of read_csv is DataFrame.to_csv().

tips.to_csv('tips2.csv')

Pandas can also export to Stata file format with the DataFrame.to_stata() method.

tips.to_stata('tips2.dta')

Data operations

Operations on columns

In Stata, arbitrary math expressions can be used with the generate and replace commands on new or existing
columns. The drop command drops the column from the data set.

replace total_bill = total_bill - 2

generate new_bill = total_bill / 2
drop new_bill

3.5. Comparison with other tools 167

pandas: powerful Python data analysis toolkit, Release 0.25.3

pandas provides similar vectorized operations by specifying the individual Series in the DataFrame. New
columns can be assigned in the same way. The DataFrame.drop() method drops a column from the DataFrame.

In [8]: tips['total_bill'] = tips['total_bill'] - 2

In [9]: tips['new_bill'] = tips['total_bill'] / 2

In [10]: tips.head()
Out[10]:
total_bill tip sex smoker day time size new_bill
0 14.99 1.01 Female No Sun Dinner 2 7.495
1 8.34 1.66 Male No Sun Dinner 3 4.170
2 19.01 3.50 Male No Sun Dinner 3 9.505
3 21.68 3.31 Male No Sun Dinner 2 10.840
4 22.59 3.61 Female No Sun Dinner 4 11.295

In [11]: tips = tips.drop('new_bill', axis=1)

Filtering

Filtering in Stata is done with an if clause on one or more columns.

list if total_bill > 10

DataFrames can be filtered in multiple ways; the most intuitive of which is using boolean indexing.

In [12]: tips[tips['total_bill'] > 10].head()

Out[12]:
total_bill tip sex smoker day time size
0 14.99 1.01 Female No Sun Dinner 2
2 19.01 3.50 Male No Sun Dinner 3
3 21.68 3.31 Male No Sun Dinner 2
4 22.59 3.61 Female No Sun Dinner 4
5 23.29 4.71 Male No Sun Dinner 4

If/then logic

In Stata, an if clause can also be used to create new columns.

generate bucket = "low" if total_bill < 10

replace bucket = "high" if total_bill >= 10

The same operation in pandas can be accomplished using the where method from numpy.

In [13]: tips['bucket'] = np.where(tips['total_bill'] < 10, 'low', 'high')

In [14]: tips.head()
Out[14]:
total_bill tip sex smoker day time size bucket
0 14.99 1.01 Female No Sun Dinner 2 high
1 8.34 1.66 Male No Sun Dinner 3 low
2 19.01 3.50 Male No Sun Dinner 3 high
3 21.68 3.31 Male No Sun Dinner 2 high
4 22.59 3.61 Female No Sun Dinner 4 high

168 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Date functionality

Stata provides a variety of functions to do operations on date/datetime columns.

generate date1 = mdy(1, 15, 2013)

generate date2 = date("Feb152015", "MDY")

generate date1_year = year(date1)

generate date2_month = month(date2)

* shift date to beginning of next month

generate date1_next = mdy(month(date1) + 1, 1, year(date1)) if month(date1) != 12
replace date1_next = mdy(1, 1, year(date1) + 1) if month(date1) == 12
generate months_between = mofd(date2) - mofd(date1)

list date1 date2 date1_year date2_month date1_next months_between

The equivalent pandas operations are shown below. In addition to these functions, pandas supports other Time Series
features not available in Stata (such as time zone handling and custom offsets) – see the timeseries documentation for
more details.

In [15]: tips['date1'] = pd.Timestamp('2013-01-15')

In [16]: tips['date2'] = pd.Timestamp('2015-02-15')

In [17]: tips['date1_year'] = tips['date1'].dt.year

In [18]: tips['date2_month'] = tips['date2'].dt.month

In [19]: tips['date1_next'] = tips['date1'] + pd.offsets.MonthBegin()

In [20]: tips['months_between'] = (tips['date2'].dt.to_period('M')

....: - tips['date1'].dt.to_period('M'))
....:

In [21]: tips[['date1', 'date2', 'date1_year', 'date2_month', 'date1_next',

....: 'months_between']].head()
....:
Out[21]:
date1 date2 date1_year date2_month date1_next months_between
0 2013-01-15 2015-02-15 2013 2 2013-02-01 <25 * MonthEnds>
1 2013-01-15 2015-02-15 2013 2 2013-02-01 <25 * MonthEnds>
2 2013-01-15 2015-02-15 2013 2 2013-02-01 <25 * MonthEnds>
3 2013-01-15 2015-02-15 2013 2 2013-02-01 <25 * MonthEnds>
4 2013-01-15 2015-02-15 2013 2 2013-02-01 <25 * MonthEnds>

Selection of columns

Stata provides keywords to select, drop, and rename columns.

keep sex total_bill tip

drop sex

rename total_bill total_bill_2

3.5. Comparison with other tools 169

pandas: powerful Python data analysis toolkit, Release 0.25.3

The same operations are expressed in pandas below. Note that in contrast to Stata, these operations do not happen in
place. To make these changes persist, assign the operation back to a variable.
# keep
In [22]: tips[['sex', 'total_bill', 'tip']].head()
Out[22]:
sex total_bill tip
0 Female 14.99 1.01
1 Male 8.34 1.66
2 Male 19.01 3.50
3 Male 21.68 3.31
4 Female 22.59 3.61

# drop
In [23]: tips.drop('sex', axis=1).head()
Out[23]:
total_bill tip smoker day time size
0 14.99 1.01 No Sun Dinner 2
1 8.34 1.66 No Sun Dinner 3
2 19.01 3.50 No Sun Dinner 3
3 21.68 3.31 No Sun Dinner 2
4 22.59 3.61 No Sun Dinner 4

# rename
In [24]: tips.rename(columns={'total_bill': 'total_bill_2'}).head()
Out[24]:
total_bill_2 tip sex smoker day time size
0 14.99 1.01 Female No Sun Dinner 2
1 8.34 1.66 Male No Sun Dinner 3
2 19.01 3.50 Male No Sun Dinner 3
3 21.68 3.31 Male No Sun Dinner 2
4 22.59 3.61 Female No Sun Dinner 4

Sorting by values

Sorting in Stata is accomplished via sort

sort sex total_bill

pandas objects have a DataFrame.sort_values() method, which takes a list of columns to sort by.

In [25]: tips = tips.sort_values(['sex', 'total_bill'])

In [26]: tips.head()
Out[26]:
total_bill tip sex smoker day time size
67 1.07 1.00 Female Yes Sat Dinner 1
92 3.75 1.00 Female Yes Fri Dinner 2
111 5.25 1.00 Female No Sat Dinner 1
145 6.35 1.50 Female No Thur Lunch 2
135 6.51 1.25 Female No Thur Lunch 2

String processing

170 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Finding length of string

Stata determines the length of a character string with the strlen() and ustrlen() functions for ASCII and
Unicode strings, respectively.

generate strlen_time = strlen(time)

generate ustrlen_time = ustrlen(time)

Python determines the length of a character string with the len function. In Python 3, all strings are Unicode strings.
len includes trailing blanks. Use len and rstrip to exclude trailing blanks.
In [27]: tips['time'].str.len().head()
Out[27]:
67 6
92 6
111 6
145 5
135 5
Name: time, dtype: int64

In [28]: tips['time'].str.rstrip().str.len().head()
Out[28]:
67 6
92 6
111 6
145 5
135 5
Name: time, dtype: int64

Finding position of substring

Stata determines the position of a character in a string with the strpos() function. This takes the string defined by
the first argument and searches for the first position of the substring you supply as the second argument.

generate str_position = strpos(sex, "ale")

Python determines the position of a character in a string with the find() function. find searches for the first
position of the substring. If the substring is found, the function returns its position. Keep in mind that Python indexes
are zero-based and the function will return -1 if it fails to find the substring.

In [29]: tips['sex'].str.find("ale").head()
Out[29]:
67 3
92 3
111 3
145 3
135 3
Name: sex, dtype: int64

Extracting substring by position

Stata extracts a substring from a string based on its position with the substr() function.

3.5. Comparison with other tools 171

pandas: powerful Python data analysis toolkit, Release 0.25.3

generate short_sex = substr(sex, 1, 1)

With pandas you can use [] notation to extract a substring from a string by position locations. Keep in mind that
Python indexes are zero-based.

In [30]: tips['sex'].str[0:1].head()
Out[30]:
67 F
92 F
111 F
145 F
135 F
Name: sex, dtype: object

Extracting nth word

The Stata word() function returns the nth word from a string. The first argument is the string you want to parse and
the second argument specifies which word you want to extract.

clear
input str20 string
"John Smith"
"Jane Cook"
end

generate first_name = word(name, 1)

generate last_name = word(name, -1)

Python extracts a substring from a string based on its text by using regular expressions. There are much more powerful
approaches, but this just shows a simple approach.

In [31]: firstlast = pd.DataFrame({'string': ['John Smith', 'Jane Cook']})

In [32]: firstlast['First_Name'] = firstlast['string'].str.split(" ", expand=True)[0]

In [33]: firstlast['Last_Name'] = firstlast['string'].str.rsplit(" ", expand=True)[0]

In [34]: firstlast
Out[34]:
string First_Name Last_Name
0 John Smith John John
1 Jane Cook Jane Jane

Changing case

The Stata strupper(), strlower(), strproper(), ustrupper(), ustrlower(), and ustrtitle()

functions change the case of ASCII and Unicode strings, respectively.

clear
input str20 string
"John Smith"
"Jane Cook"
end
(continues on next page)

172 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

generate upper = strupper(string)

generate lower = strlower(string)
generate title = strproper(string)
list

The equivalent Python functions are upper, lower, and title.

In [35]: firstlast = pd.DataFrame({'string': ['John Smith', 'Jane Cook']})

In [36]: firstlast['upper'] = firstlast['string'].str.upper()

In [37]: firstlast['lower'] = firstlast['string'].str.lower()

In [38]: firstlast['title'] = firstlast['string'].str.title()

In [39]: firstlast
Out[39]:
string upper lower title
0 John Smith JOHN SMITH john smith John Smith
1 Jane Cook JANE COOK jane cook Jane Cook

Merging

The following tables will be used in the merge examples

In [40]: df1 = pd.DataFrame({'key': ['A', 'B', 'C', 'D'],

....: 'value': np.random.randn(4)})
....:

In [41]: df1
Out[41]:
key value
0 A -0.358468
1 B -0.697821
2 C 0.162113
3 D -0.076884

In [42]: df2 = pd.DataFrame({'key': ['B', 'D', 'D', 'E'],

....: 'value': np.random.randn(4)})
....:

In [43]: df2
Out[43]:
key value
0 B -0.235803
1 D 0.449752
2 D -2.220693
3 E 0.186546

In Stata, to perform a merge, one data set must be in memory and the other must be referenced as a file name on disk.
In contrast, Python must have both DataFrames already in memory.
By default, Stata performs an outer join, where all observations from both data sets are left in memory after the merge.
One can keep only observations from the initial data set, the merged data set, or the intersection of the two by using
the values created in the _merge variable.

3.5. Comparison with other tools 173

pandas: powerful Python data analysis toolkit, Release 0.25.3

* First create df2 and save to disk

clear
input str1 key
B
D
D
E
end
generate value = rnormal()
save df2.dta

* N ow create df1 in memory

clear
input str1 key
A
B
C
D
end
generate value = rnormal()

preserve

* Left join
merge 1:n key using df2.dta
keep if _merge == 1

* Right join
restore, preserve
merge 1:n key using df2.dta
keep if _merge == 2

* Inner join
restore, preserve
merge 1:n key using df2.dta
keep if _merge == 3

* Outer join
restore
merge 1:n key using df2.dta

pandas DataFrames have a DataFrame.merge() method, which provides similar functionality. Note that different
join types are accomplished via the how keyword.

In [44]: inner_join = df1.merge(df2, on=['key'], how='inner')

In [45]: inner_join
Out[45]:
key value_x value_y
0 B -0.697821 -0.235803
1 D -0.076884 0.449752
2 D -0.076884 -2.220693

In [46]: left_join = df1.merge(df2, on=['key'], how='left')

In [47]: left_join
Out[47]:
(continues on next page)

174 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

key value_x value_y
0 A -0.358468 NaN
1 B -0.697821 -0.235803
2 C 0.162113 NaN
3 D -0.076884 0.449752
4 D -0.076884 -2.220693

In [48]: right_join = df1.merge(df2, on=['key'], how='right')

In [49]: right_join
Out[49]:
key value_x value_y
0 B -0.697821 -0.235803
1 D -0.076884 0.449752
2 D -0.076884 -2.220693
3 E NaN 0.186546

In [50]: outer_join = df1.merge(df2, on=['key'], how='outer')

In [51]: outer_join
Out[51]:
key value_x value_y
0 A -0.358468 NaN
1 B -0.697821 -0.235803
2 C 0.162113 NaN
3 D -0.076884 0.449752
4 D -0.076884 -2.220693
5 E NaN 0.186546

Missing data

Like Stata, pandas has a representation for missing data – the special float value NaN (not a number). Many of the
semantics are the same; for example missing data propagates through numeric operations, and is ignored by default
for aggregations.
In [52]: outer_join
Out[52]:
key value_x value_y
0 A -0.358468 NaN
1 B -0.697821 -0.235803
2 C 0.162113 NaN
3 D -0.076884 0.449752
4 D -0.076884 -2.220693
5 E NaN 0.186546

In [53]: outer_join['value_x'] + outer_join['value_y']

Out[53]:
0 NaN
1 -0.933624
2 NaN
3 0.372867
4 -2.297577
5 NaN
dtype: float64

3.5. Comparison with other tools 175

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [54]: outer_join['value_x'].sum()
Out[54]: -1.0479449929596008
One difference is that missing data cannot be compared to its sentinel value. For example, in Stata you could do this
to filter missing values.

* Keep missing values

list if value_x == .
* Keep non-missing values
list if value_x != .

This doesnt work in pandas. Instead, the pd.isna() or pd.notna() functions should be used for comparisons.
In [55]: outer_join[pd.isna(outer_join['value_x'])]
Out[55]:
key value_x value_y
5 E NaN 0.186546

In [56]: outer_join[pd.notna(outer_join['value_x'])]
Out[56]:
key value_x value_y
0 A -0.358468 NaN
1 B -0.697821 -0.235803
2 C 0.162113 NaN
3 D -0.076884 0.449752
4 D -0.076884 -2.220693
Pandas also provides a variety of methods to work with missing data – some of which would be challenging to express
in Stata. For example, there are methods to drop all rows with any missing values, replacing missing values with a
specified value, like the mean, or forward filling from previous rows. See the missing data documentation for more.
# Drop rows with any missing value
In [57]: outer_join.dropna()
Out[57]:
key value_x value_y
1 B -0.697821 -0.235803
3 D -0.076884 0.449752
4 D -0.076884 -2.220693

# Fill forwards
In [58]: outer_join.fillna(method='ffill')
Out[58]:
key value_x value_y
0 A -0.358468 NaN
1 B -0.697821 -0.235803
2 C 0.162113 -0.235803
3 D -0.076884 0.449752
4 D -0.076884 -2.220693
5 E -0.076884 0.186546

# Impute missing values with the mean

In [59]: outer_join['value_x'].fillna(outer_join['value_x'].mean())
Out[59]:
0 -0.358468
1 -0.697821

176 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

2 0.162113
3 -0.076884
4 -0.076884
5 -0.209589
Name: value_x, dtype: float64

GroupBy

Aggregation

Statas collapse can be used to group by one or more key variables and compute aggregations on numeric columns.

collapse (sum) total_bill tip, by(sex smoker)

pandas provides a flexible groupby mechanism that allows similar aggregations. See the groupby documentation for
more details and examples.

In [60]: tips_summed = tips.groupby(['sex', 'smoker'])['total_bill', 'tip'].sum()

In [61]: tips_summed.head()
Out[61]:
total_bill tip
sex smoker
Female No 869.68 149.77
Yes 527.27 96.74
Male No 1725.75 302.00
Yes 1217.07 183.07

Transformation

In Stata, if the group aggregations need to be used with the original data set, one would usually use bysort with
egen(). For example, to subtract the mean for each observation by smoker group.

bysort sex smoker: egen group_bill = mean(total_bill)

generate adj_total_bill = total_bill - group_bill

pandas groupby provides a transform mechanism that allows these type of operations to be succinctly expressed
in one operation.

In [62]: gb = tips.groupby('smoker')['total_bill']

In [63]: tips['adj_total_bill'] = tips['total_bill'] - gb.transform('mean')

In [64]: tips.head()
Out[64]:
total_bill tip sex smoker day time size adj_total_bill
67 1.07 1.00 Female Yes Sat Dinner 1 -17.686344
92 3.75 1.00 Female Yes Fri Dinner 2 -15.006344
111 5.25 1.00 Female No Sat Dinner 1 -11.938278
145 6.35 1.50 Female No Thur Lunch 2 -10.838278
135 6.51 1.25 Female No Thur Lunch 2 -10.678278

3.5. Comparison with other tools 177

pandas: powerful Python data analysis toolkit, Release 0.25.3

By group processing

In addition to aggregation, pandas groupby can be used to replicate most other bysort processing from Stata. For
example, the following example lists the first observation in the current sort order by sex/smoker group.

bysort sex smoker: list if _n == 1

In pandas this would be written as:

In [65]: tips.groupby(['sex', 'smoker']).first()

Out[65]:
total_bill tip day time size adj_total_bill
sex smoker
Female No 5.25 1.00 Sat Dinner 1 -11.938278
Yes 1.07 1.00 Sat Dinner 1 -17.686344
Male No 5.51 2.00 Thur Lunch 2 -11.678278
Yes 5.25 5.15 Sun Dinner 2 -13.506344

Other considerations

Disk vs memory

Pandas and Stata both operate exclusively in memory. This means that the size of data able to be loaded in pandas is
limited by your machines memory. If out of core processing is needed, one possibility is the dask.dataframe library,
which provides a subset of pandas functionality for an on-disk DataFrame. {{ header }}

3.6 Tutorials

This is a guide to many pandas tutorials, geared mainly for new users.

3.6.1 Internal guides

pandas own 10 Minutes to pandas.

More complex recipes are in the Cookbook.
A handy pandas cheat sheet.

3.6.2 Community guides

pandas Cookbook by Julia Evans

The goal of this 2015 cookbook (by Julia Evans) is to give you some concrete examples for getting started with pandas.
These are examples with real-world data, and all the bugs and weirdness that entails. For the table of contents, see the
pandas-cookbook GitHub repository.

Learn Pandas by Hernan Rojas

A set of lesson for new pandas users: https://bitbucket.org/hrojas/learn-pandas

178 Chapter 3. Getting started

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Practical data analysis with Python

This guide is an introduction to the data analysis process using the Python data ecosystem and an interesting open
dataset. There are four sections covering selected topics as munging data, aggregating data, visualizing data and time
series.

Exercises for new users

Practice your skills with real data sets and exercises. For more resources, please visit the main repository.

Modern pandas

Tutorial series written in 2016 by Tom Augspurger. The source may be found in the GitHub repository
TomAugspurger/effective-pandas.
• Modern Pandas
• Method Chaining
• Indexes
• Performance
• Tidy Data
• Visualization
• Timeseries

Excel charts with pandas, vincent and xlsxwriter

• Using Pandas and XlsxWriter to create Excel charts

Video tutorials

• Pandas From The Ground Up (2015) (2:24) GitHub repo

• Introduction Into Pandas (2016) (1:28) GitHub repo
• Pandas: .head() to .tail() (2016) (1:26) GitHub repo
• Data analysis in Python with pandas (2016-2018) GitHub repo and Jupyter Notebook
• Best practices with pandas (2018) GitHub repo and Jupyter Notebook

Various tutorials

• Wes McKinneys (pandas BDFL) blog

• Statistical analysis made easy in Python with SciPy and pandas DataFrames, by Randal Olson
• Statistical Data Analysis in Python, tutorial videos, by Christopher Fonnesbeck from SciPy 2013
• Financial analysis in Python, by Thomas Wiecki
• Intro to pandas data structures, by Greg Reda
• Pandas and Python: Top 10, by Manish Amde

3.6. Tutorials 179

pandas: powerful Python data analysis toolkit, Release 0.25.3

• Pandas DataFrames Tutorial, by Karlijn Willems

• A concise tutorial with real life examples
{{ header }}

180 Chapter 3. Getting started

 CHAPTER

FOUR

USER GUIDE

The User Guide covers all of pandas by topic area. Each of the subsections introduces a topic (such as working with
missing data), and discusses how pandas approaches the problem, with many examples throughout.
Users brand-new to pandas should start with 10min.
Further information on any specific method can be obtained in the API reference. {{ header }}

4.1 IO tools (text, CSV, HDF5, )

The pandas I/O API is a set of top level reader functions accessed like pandas.read_csv() that generally
return a pandas object. The corresponding writer functions are object methods that are accessed like DataFrame.
to_csv(). Below is a table containing available readers and writers.

Format Data Description Reader Writer

Type
text CSV read_csv to_csv
text JSON read_json to_json
text HTML read_html to_html
text Local clipboard read_clipboard to_clipboard
binary MS Excel read_excel to_excel
binary OpenDocument read_excel
binary HDF5 Format read_hdf to_hdf
binary Feather Format read_feather to_feather
binary Parquet Format read_parquet to_parquet
binary Msgpack read_msgpack to_msgpack
binary Stata read_stata to_stata
binary SAS read_sas
binary Python Pickle Format read_pickle to_pickle
SQL SQL read_sql to_sql
SQL Google Big Query read_gbq to_gbq

Here is an informal performance comparison for some of these IO methods.

Note: For examples that use the StringIO class, make sure you import it according to your Python version, i.e.
from StringIO import StringIO for Python 2 and from io import StringIO for Python 3.

181
pandas: powerful Python data analysis toolkit, Release 0.25.3

4.1.1 CSV & text files

The workhorse function for reading text files (a.k.a. flat files) is read_csv(). See the cookbook for some advanced
strategies.

Parsing options

read_csv() accepts the following common arguments:

Basic

filepath_or_buffer [various] Either a path to a file (a str, pathlib.Path, or py._path.local.

LocalPath), URL (including http, ftp, and S3 locations), or any object with a read() method (such as
an open file or StringIO).
sep [str, defaults to ',' for read_csv(), \t for read_table()] Delimiter to use. If sep is None, the C engine
cannot automatically detect the separator, but the Python parsing engine can, meaning the latter will be used
and automatically detect the separator by Pythons builtin sniffer tool, csv.Sniffer. In addition, separators
longer than 1 character and different from '\s+' will be interpreted as regular expressions and will also force
the use of the Python parsing engine. Note that regex delimiters are prone to ignoring quoted data. Regex
example: '\\r\\t'.
delimiter [str, default None] Alternative argument name for sep.
delim_whitespace [boolean, default False] Specifies whether or not whitespace (e.g. ' ' or '\t') will be used as
the delimiter. Equivalent to setting sep='\s+'. If this option is set to True, nothing should be passed in for
the delimiter parameter.
New in version 0.18.1: support for the Python parser.

Column and index locations and names

header [int or list of ints, default 'infer'] Row number(s) to use as the column names, and the start of the data.
Default behavior is to infer the column names: if no names are passed the behavior is identical to header=0
and column names are inferred from the first line of the file, if column names are passed explicitly then the
behavior is identical to header=None. Explicitly pass header=0 to be able to replace existing names.
The header can be a list of ints that specify row locations for a MultiIndex on the columns e.g. [0,1,3].
Intervening rows that are not specified will be skipped (e.g. 2 in this example is skipped). Note that this
parameter ignores commented lines and empty lines if skip_blank_lines=True, so header=0 denotes the
first line of data rather than the first line of the file.
names [array-like, default None] List of column names to use. If file contains no header row, then you should
explicitly pass header=None. Duplicates in this list are not allowed.
index_col [int, str, sequence of int / str, or False, default None] Column(s) to use as the row labels of the
DataFrame, either given as string name or column index. If a sequence of int / str is given, a MultiIndex
is used.
Note: index_col=False can be used to force pandas to not use the first column as the index, e.g. when you
have a malformed file with delimiters at the end of each line.
usecols [list-like or callable, default None] Return a subset of the columns. If list-like, all elements must either be
positional (i.e. integer indices into the document columns) or strings that correspond to column names provided
either by the user in names or inferred from the document header row(s). For example, a valid list-like usecols
parameter would be [0, 1, 2] or ['foo', 'bar', 'baz'].

182 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Element order is ignored, so usecols=[0, 1] is the same as [1, 0]. To instantiate a

DataFrame from data with element order preserved use pd.read_csv(data, usecols=['foo',
'bar'])[['foo', 'bar']] for columns in ['foo', 'bar'] order or pd.read_csv(data,
usecols=['foo', 'bar'])[['bar', 'foo']] for ['bar', 'foo'] order.
If callable, the callable function will be evaluated against the column names, returning names where the callable
function evaluates to True:
In [1]: from io import StringIO, BytesIO

In [2]: data = ('col1,col2,col3\n'

...: 'a,b,1\n'
...: 'a,b,2\n'
...: 'c,d,3')
...:

In [3]: pd.read_csv(StringIO(data))
Out[3]:
col1 col2 col3
0 a b 1
1 a b 2
2 c d 3

In [4]: pd.read_csv(StringIO(data), usecols=lambda x: x.upper() in

,→['COL1', 'COL3'])

Out[4]:
col1 col3
0 a 1
1 a 2
2 c 3
Using this parameter results in much faster parsing time and lower memory usage.
squeeze [boolean, default False] If the parsed data only contains one column then return a Series.
prefix [str, default None] Prefix to add to column numbers when no header, e.g. X for X0, X1,
mangle_dupe_cols [boolean, default True] Duplicate columns will be specified as X, X.1X.N, rather than XX.
Passing in False will cause data to be overwritten if there are duplicate names in the columns.

General parsing configuration

dtype [Type name or dict of column -> type, default None] Data type for data or columns. E.g. {'a': np.
float64, 'b': np.int32} (unsupported with engine='python'). Use str or object together with
suitable na_values settings to preserve and not interpret dtype.
New in version 0.20.0: support for the Python parser.
engine [{'c', 'python'}] Parser engine to use. The C engine is faster while the Python engine is currently more
feature-complete.
converters [dict, default None] Dict of functions for converting values in certain columns. Keys can either be integers
or column labels.
true_values [list, default None] Values to consider as True.
false_values [list, default None] Values to consider as False.
skipinitialspace [boolean, default False] Skip spaces after delimiter.

4.1. IO tools (text, CSV, HDF5, ) 183

pandas: powerful Python data analysis toolkit, Release 0.25.3

skiprows [list-like or integer, default None] Line numbers to skip (0-indexed) or number of lines to skip (int) at the
start of the file.
If callable, the callable function will be evaluated against the row indices, returning True if the row should be
skipped and False otherwise:
In [5]: data = ('col1,col2,col3\n'
...: 'a,b,1\n'
...: 'a,b,2\n'
...: 'c,d,3')
...:

In [6]: pd.read_csv(StringIO(data))
Out[6]:
col1 col2 col3
0 a b 1
1 a b 2
2 c d 3

In [7]: pd.read_csv(StringIO(data), skiprows=lambda x: x % 2 != 0)

Out[7]:
col1 col2 col3
0 a b 2
skipfooter [int, default 0] Number of lines at bottom of file to skip (unsupported with engine=c).
nrows [int, default None] Number of rows of file to read. Useful for reading pieces of large files.
low_memory [boolean, default True] Internally process the file in chunks, resulting in lower memory use while
parsing, but possibly mixed type inference. To ensure no mixed types either set False, or specify the type with
the dtype parameter. Note that the entire file is read into a single DataFrame regardless, use the chunksize
or iterator parameter to return the data in chunks. (Only valid with C parser)
memory_map [boolean, default False] If a filepath is provided for filepath_or_buffer, map the file object
directly onto memory and access the data directly from there. Using this option can improve performance
because there is no longer any I/O overhead.

NA and missing data handling

na_values [scalar, str, list-like, or dict, default None] Additional strings to recognize as NA/NaN. If dict passed,
specific per-column NA values. See na values const below for a list of the values interpreted as NaN by default.
keep_default_na [boolean, default True] Whether or not to include the default NaN values when parsing the data.
Depending on whether na_values is passed in, the behavior is as follows:
• If keep_default_na is True, and na_values are specified, na_values is appended to the default NaN values
used for parsing.
• If keep_default_na is True, and na_values are not specified, only the default NaN values are used for
parsing.
• If keep_default_na is False, and na_values are specified, only the NaN values specified na_values are
used for parsing.
• If keep_default_na is False, and na_values are not specified, no strings will be parsed as NaN.
Note that if na_filter is passed in as False, the keep_default_na and na_values parameters will be ignored.

184 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

na_filter [boolean, default True] Detect missing value markers (empty strings and the value of na_values). In data
without any NAs, passing na_filter=False can improve the performance of reading a large file.
verbose [boolean, default False] Indicate number of NA values placed in non-numeric columns.
skip_blank_lines [boolean, default True] If True, skip over blank lines rather than interpreting as NaN values.

Datetime handling

parse_dates [boolean or list of ints or names or list of lists or dict, default False.]
• If True -> try parsing the index.
• If [1, 2, 3] -> try parsing columns 1, 2, 3 each as a separate date column.
• If [[1, 3]] -> combine columns 1 and 3 and parse as a single date column.
• If {'foo': [1, 3]} -> parse columns 1, 3 as date and call result foo. A fast-path exists for iso8601-
formatted dates.
infer_datetime_format [boolean, default False] If True and parse_dates is enabled for a column, attempt to infer
the datetime format to speed up the processing.
keep_date_col [boolean, default False] If True and parse_dates specifies combining multiple columns then keep
the original columns.
date_parser [function, default None] Function to use for converting a sequence of string columns to an array of
datetime instances. The default uses dateutil.parser.parser to do the conversion. pandas will try to
call date_parser in three different ways, advancing to the next if an exception occurs: 1) Pass one or more arrays
(as defined by parse_dates) as arguments; 2) concatenate (row-wise) the string values from the columns defined
by parse_dates into a single array and pass that; and 3) call date_parser once for each row using one or more
strings (corresponding to the columns defined by parse_dates) as arguments.
dayfirst [boolean, default False] DD/MM format dates, international and European format.
cache_dates [boolean, default True] If True, use a cache of unique, converted dates to apply the datetime conversion.
May produce significant speed-up when parsing duplicate date strings, especially ones with timezone offsets.
New in version 0.25.0.

Iteration

iterator [boolean, default False] Return TextFileReader object for iteration or getting chunks with get_chunk().
chunksize [int, default None] Return TextFileReader object for iteration. See iterating and chunking below.

Quoting, compression, and file format

compression [{'infer', 'gzip', 'bz2', 'zip', 'xz', None}, default 'infer'] For on-the-fly decompres-
sion of on-disk data. If infer, then use gzip, bz2, zip, or xz if filepath_or_buffer is a string ending in .gz, .bz2,
.zip, or .xz, respectively, and no decompression otherwise. If using zip, the ZIP file must contain only one data
file to be read in. Set to None for no decompression.
New in version 0.18.1: support for zip and xz compression.
Changed in version 0.24.0: infer option added and set to default.
thousands [str, default None] Thousands separator.

4.1. IO tools (text, CSV, HDF5, ) 185

pandas: powerful Python data analysis toolkit, Release 0.25.3

decimal [str, default '.'] Character to recognize as decimal point. E.g. use ',' for European data.
float_precision [string, default None] Specifies which converter the C engine should use for floating-point values.
The options are None for the ordinary converter, high for the high-precision converter, and round_trip for
the round-trip converter.
lineterminator [str (length 1), default None] Character to break file into lines. Only valid with C parser.
quotechar [str (length 1)] The character used to denote the start and end of a quoted item. Quoted items can include
the delimiter and it will be ignored.
quoting [int or csv.QUOTE_* instance, default 0] Control field quoting behavior per csv.QUOTE_* constants.
Use one of QUOTE_MINIMAL (0), QUOTE_ALL (1), QUOTE_NONNUMERIC (2) or QUOTE_NONE (3).
doublequote [boolean, default True] When quotechar is specified and quoting is not QUOTE_NONE, indi-
cate whether or not to interpret two consecutive quotechar elements inside a field as a single quotechar
element.
escapechar [str (length 1), default None] One-character string used to escape delimiter when quoting is
QUOTE_NONE.
comment [str, default None] Indicates remainder of line should not be parsed. If found at the beginning of a line,
the line will be ignored altogether. This parameter must be a single character. Like empty lines (as long
as skip_blank_lines=True), fully commented lines are ignored by the parameter header but not by
skiprows. For example, if comment='#', parsing #empty\na,b,c\n1,2,3 with header=0 will result in a,b,c
being treated as the header.
encoding [str, default None] Encoding to use for UTF when reading/writing (e.g. 'utf-8'). List of Python standard
encodings.
dialect [str or csv.Dialect instance, default None] If provided, this parameter will override values (default or
not) for the following parameters: delimiter, doublequote, escapechar, skipinitialspace, quotechar, and quoting.
If it is necessary to override values, a ParserWarning will be issued. See csv.Dialect documentation for
more details.

Error handling

error_bad_lines [boolean, default True] Lines with too many fields (e.g. a csv line with too many commas) will by
default cause an exception to be raised, and no DataFrame will be returned. If False, then these bad lines
will dropped from the DataFrame that is returned. See bad lines below.
warn_bad_lines [boolean, default True] If error_bad_lines is False, and warn_bad_lines is True, a warning for
each bad line will be output.

Specifying column data types

You can indicate the data type for the whole DataFrame or individual columns:
In [8]: data = ('a,b,c,d\n'
...: '1,2,3,4\n'
...: '5,6,7,8\n'
...: '9,10,11')
...:

In [9]: print(data)
a,b,c,d
1,2,3,4
5,6,7,8

186 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

9,10,11

In [10]: df = pd.read_csv(StringIO(data), dtype=object)

In [11]: df
Out[11]:
a b c d
0 1 2 3 4
1 5 6 7 8
2 9 10 11 NaN

In [12]: df['a'][0]
Out[12]: '1'

In [13]: df = pd.read_csv(StringIO(data),
....: dtype={'b': object, 'c': np.float64, 'd': 'Int64'})
....:

In [14]: df.dtypes
Out[14]:
a int64
b object
c float64
d Int64
dtype: object
Fortunately, pandas offers more than one way to ensure that your column(s) contain only one dtype. If youre
unfamiliar with these concepts, you can see here to learn more about dtypes, and here to learn more about object
conversion in pandas.
For instance, you can use the converters argument of read_csv():
In [15]: data = ("col_1\n"
....: "1\n"
....: "2\n"
....: "'A'\n"
....: "4.22")
....:

In [16]: df = pd.read_csv(StringIO(data), converters={'col_1': str})

In [17]: df
Out[17]:
col_1
0 1
1 2
2 'A'
3 4.22

In [18]: df['col_1'].apply(type).value_counts()
Out[18]:
<class 'str'> 4
Name: col_1, dtype: int64
Or you can use the to_numeric() function to coerce the dtypes after reading in the data,

4.1. IO tools (text, CSV, HDF5, ) 187

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [19]: df2 = pd.read_csv(StringIO(data))

In [20]: df2['col_1'] = pd.to_numeric(df2['col_1'], errors='coerce')

In [21]: df2
Out[21]:
col_1
0 1.00
1 2.00
2 NaN
3 4.22

In [22]: df2['col_1'].apply(type).value_counts()
Out[22]:
<class 'float'> 4
Name: col_1, dtype: int64
which will convert all valid parsing to floats, leaving the invalid parsing as NaN.
Ultimately, how you deal with reading in columns containing mixed dtypes depends on your specific needs. In the case
above, if you wanted to NaN out the data anomalies, then to_numeric() is probably your best option. However, if
you wanted for all the data to be coerced, no matter the type, then using the converters argument of read_csv()
would certainly be worth trying.
New in version 0.20.0: support for the Python parser.
The dtype option is supported by the python engine.

Note: In some cases, reading in abnormal data with columns containing mixed dtypes will result in an inconsistent
dataset. If you rely on pandas to infer the dtypes of your columns, the parsing engine will go and infer the dtypes for
different chunks of the data, rather than the whole dataset at once. Consequently, you can end up with column(s) with
mixed dtypes. For example,
In [23]: col_1 = list(range(500000)) + ['a', 'b'] + list(range(500000))

In [24]: df = pd.DataFrame({'col_1': col_1})

In [25]: df.to_csv('foo.csv')

In [26]: mixed_df = pd.read_csv('foo.csv')

In [27]: mixed_df['col_1'].apply(type).value_counts()
Out[27]:
<class 'int'> 737858
<class 'str'> 262144
Name: col_1, dtype: int64

In [28]: mixed_df['col_1'].dtype
Out[28]: dtype('O')
will result with mixed_df containing an int dtype for certain chunks of the column, and str for others due to the
mixed dtypes from the data that was read in. It is important to note that the overall column will be marked with a
dtype of object, which is used for columns with mixed dtypes.

188 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Specifying categorical dtype

New in version 0.19.0.

Categorical columns can be parsed directly by specifying dtype='category' or
dtype=CategoricalDtype(categories, ordered).
In [29]: data = ('col1,col2,col3\n'
....: 'a,b,1\n'
....: 'a,b,2\n'
....: 'c,d,3')
....:

In [30]: pd.read_csv(StringIO(data))
Out[30]:
col1 col2 col3
0 a b 1
1 a b 2
2 c d 3

In [31]: pd.read_csv(StringIO(data)).dtypes
Out[31]:
col1 object
col2 object
col3 int64
dtype: object

In [32]: pd.read_csv(StringIO(data), dtype='category').dtypes

Out[32]:
col1 category
col2 category
col3 category
dtype: object
Individual columns can be parsed as a Categorical using a dict specification:
In [33]: pd.read_csv(StringIO(data), dtype={'col1': 'category'}).dtypes
Out[33]:
col1 category
col2 object
col3 int64
dtype: object

New in version 0.21.0.

Specifying dtype='category' will result in an unordered Categorical whose categories are the unique
values observed in the data. For more control on the categories and order, create a CategoricalDtype ahead of
time, and pass that for that columns dtype.
In [34]: from pandas.api.types import CategoricalDtype

In [35]: dtype = CategoricalDtype(['d', 'c', 'b', 'a'], ordered=True)

In [36]: pd.read_csv(StringIO(data), dtype={'col1': dtype}).dtypes

Out[36]:
col1 category
col2 object
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 189

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

col3 int64
dtype: object

When using dtype=CategoricalDtype, unexpected values outside of dtype.categories are treated as

missing values.

In [37]: dtype = CategoricalDtype(['a', 'b', 'd']) # No 'c'

In [38]: pd.read_csv(StringIO(data), dtype={'col1': dtype}).col1

Out[38]:
0 a
1 a
2 NaN
Name: col1, dtype: category
Categories (3, object): [a, b, d]

This matches the behavior of Categorical.set_categories().

Note: With dtype='category', the resulting categories will always be parsed as strings (object dtype). If the
categories are numeric they can be converted using the to_numeric() function, or as appropriate, another converter
such as to_datetime().
When dtype is a CategoricalDtype with homogeneous categories ( all numeric, all datetimes, etc.), the
conversion is done automatically.
In [39]: df = pd.read_csv(StringIO(data), dtype='category')

In [40]: df.dtypes
Out[40]:
col1 category
col2 category
col3 category
dtype: object

In [41]: df['col3']
Out[41]:
0 1
1 2
2 3
Name: col3, dtype: category
Categories (3, object): [1, 2, 3]

In [42]: df['col3'].cat.categories = pd.to_numeric(df['col3'].cat.categories)

In [43]: df['col3']
Out[43]:
0 1
1 2
2 3
Name: col3, dtype: category
Categories (3, int64): [1, 2, 3]

190 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Naming and using columns

Handling column names

A file may or may not have a header row. pandas assumes the first row should be used as the column names:
In [44]: data = ('a,b,c\n'
....: '1,2,3\n'
....: '4,5,6\n'
....: '7,8,9')
....:

In [45]: print(data)
a,b,c
1,2,3
4,5,6
7,8,9

In [46]: pd.read_csv(StringIO(data))
Out[46]:
a b c
0 1 2 3
1 4 5 6
2 7 8 9
By specifying the names argument in conjunction with header you can indicate other names to use and whether or
not to throw away the header row (if any):
In [47]: print(data)
a,b,c
1,2,3
4,5,6
7,8,9

In [48]: pd.read_csv(StringIO(data), names=['foo', 'bar', 'baz'], header=0)

Out[48]:
foo bar baz
0 1 2 3
1 4 5 6
2 7 8 9

In [49]: pd.read_csv(StringIO(data), names=['foo', 'bar', 'baz'], header=None)

Out[49]:
foo bar baz
0 a b c
1 1 2 3
2 4 5 6
3 7 8 9
If the header is in a row other than the first, pass the row number to header. This will skip the preceding rows:

In [50]: data = ('skip this skip it\n'

....: 'a,b,c\n'
....: '1,2,3\n'
....: '4,5,6\n'
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 191

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

....: '7,8,9')
....:

In [51]: pd.read_csv(StringIO(data), header=1)

Out[51]:
a b c
0 1 2 3
1 4 5 6
2 7 8 9

Note: Default behavior is to infer the column names: if no names are passed the behavior is identical to header=0
and column names are inferred from the first non-blank line of the file, if column names are passed explicitly then the
behavior is identical to header=None.

Duplicate names parsing

If the file or header contains duplicate names, pandas will by default distinguish between them so as to prevent
overwriting data:
In [52]: data = ('a,b,a\n'
....: '0,1,2\n'
....: '3,4,5')
....:

In [53]: pd.read_csv(StringIO(data))
Out[53]:
a b a.1
0 0 1 2
1 3 4 5

There is no more duplicate data because mangle_dupe_cols=True by default, which modifies a series of dupli-
cate columns X, , X to become X, X.1, , X.N. If mangle_dupe_cols=False, duplicate data can arise:
In [2]: data = 'a,b,a\n0,1,2\n3,4,5'
In [3]: pd.read_csv(StringIO(data), mangle_dupe_cols=False)
Out[3]:
a b a
0 2 1 2
1 5 4 5

To prevent users from encountering this problem with duplicate data, a ValueError exception is raised if
mangle_dupe_cols != True:
In [2]: data = 'a,b,a\n0,1,2\n3,4,5'
In [3]: pd.read_csv(StringIO(data), mangle_dupe_cols=False)
...
ValueError: Setting mangle_dupe_cols=False is not supported yet

Filtering columns (usecols)

The usecols argument allows you to select any subset of the columns in a file, either using the column names,
position numbers or a callable:

192 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

New in version 0.20.0: support for callable usecols arguments

In [54]: data = 'a,b,c,d\n1,2,3,foo\n4,5,6,bar\n7,8,9,baz'

In [55]: pd.read_csv(StringIO(data))
Out[55]:
a b c d
0 1 2 3 foo
1 4 5 6 bar
2 7 8 9 baz

In [56]: pd.read_csv(StringIO(data), usecols=['b', 'd'])

Out[56]:
b d
0 2 foo
1 5 bar
2 8 baz

In [57]: pd.read_csv(StringIO(data), usecols=[0, 2, 3])

Out[57]:
a c d
0 1 3 foo
1 4 6 bar
2 7 9 baz

In [58]: pd.read_csv(StringIO(data), usecols=lambda x: x.upper() in ['A',

,→'C'])

Out[58]:
a c
0 1 3
1 4 6
2 7 9
The usecols argument can also be used to specify which columns not to use in the final result:

In [59]: pd.read_csv(StringIO(data), usecols=lambda x: x not in ['a', 'c'])

Out[59]:
b d
0 2 foo
1 5 bar
2 8 baz

In this case, the callable is specifying that we exclude the a and c columns from the output.

Comments and empty lines

Ignoring line comments and empty lines

If the comment parameter is specified, then completely commented lines will be ignored. By default, completely
blank lines will be ignored as well.
In [60]: data = ('\n'
....: 'a,b,c\n'
....: ' \n'
....: '# commented line\n'

4.1. IO tools (text, CSV, HDF5, ) 193

pandas: powerful Python data analysis toolkit, Release 0.25.3

....: '1,2,3\n'
....: '\n'
....: '4,5,6')
....:

In [61]: print(data)

a,b,c

# commented line
1,2,3

4,5,6

In [62]: pd.read_csv(StringIO(data), comment='#')

Out[62]:
a b c
0 1 2 3
1 4 5 6
If skip_blank_lines=False, then read_csv will not ignore blank lines:
In [63]: data = ('a,b,c\n'
....: '\n'
....: '1,2,3\n'
....: '\n'
....: '\n'
....: '4,5,6')
....:

In [64]: pd.read_csv(StringIO(data), skip_blank_lines=False)

Out[64]:
a b c
0 NaN NaN NaN
1 1.0 2.0 3.0
2 NaN NaN NaN
3 NaN NaN NaN
4 4.0 5.0 6.0

Warning: The presence of ignored lines might create ambiguities involving line numbers; the parameter header
uses row numbers (ignoring commented/empty lines), while skiprows uses line numbers (including com-
mented/empty lines):
In [65]: data = ('#comment\n'
....: 'a,b,c\n'
....: 'A,B,C\n'
....: '1,2,3')
....:

In [66]: pd.read_csv(StringIO(data), comment='#', header=1)

Out[66]:
A B C
0 1 2 3

In [67]: data = ('A,B,C\n'

....: '#comment\n'
....: 'a,b,c\n'
....: '1,2,3')
....:
194 Chapter 4. User Guide
In [68]: pd.read_csv(StringIO(data), comment='#', skiprows=2)
Out[68]:
a b c
 pandas: powerful Python data analysis toolkit, Release 0.25.3

If both header and skiprows are specified, header will be relative to the end of skiprows. For example:

In [69]: data = ('# empty\n'

....: '# second empty line\n'
....: '# third emptyline\n'
....: 'X,Y,Z\n'
....: '1,2,3\n'
....: 'A,B,C\n'
....: '1,2.,4.\n'
....: '5.,NaN,10.0\n')
....:

In [70]: print(data)
# empty
# second empty line
# third emptyline
X,Y,Z
1,2,3
A,B,C
1,2.,4.
5.,NaN,10.0

In [71]: pd.read_csv(StringIO(data), comment='#', skiprows=4, header=1)

Out[71]:
A B C
0 1.0 2.0 4.0
1 5.0 NaN 10.0

Comments

Sometimes comments or meta data may be included in a file:

In [72]: print(open('tmp.csv').read())
ID,level,category
Patient1,123000,x # really unpleasant
Patient2,23000,y # wouldn't take his medicine
Patient3,1234018,z # awesome

By default, the parser includes the comments in the output:

In [73]: df = pd.read_csv('tmp.csv')

In [74]: df
Out[74]:
ID level category
0 Patient1 123000 x # really unpleasant
1 Patient2 23000 y # wouldn't take his medicine
2 Patient3 1234018 z # awesome

We can suppress the comments using the comment keyword:

4.1. IO tools (text, CSV, HDF5, ) 195

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [75]: df = pd.read_csv('tmp.csv', comment='#')

In [76]: df
Out[76]:
ID level category
0 Patient1 123000 x
1 Patient2 23000 y
2 Patient3 1234018 z

Dealing with Unicode data

The encoding argument should be used for encoded unicode data, which will result in byte strings being decoded
to unicode in the result:
In [77]: data = (b'word,length\n'
....: b'Tr\xc3\xa4umen,7\n'
....: b'Gr\xc3\xbc\xc3\x9fe,5')
....:

In [78]: data = data.decode('utf8').encode('latin-1')

In [79]: df = pd.read_csv(BytesIO(data), encoding='latin-1')

In [80]: df
Out[80]:
word length
0 Träumen 7
1 GrüSSe 5

In [81]: df['word'][1]
Out[81]: 'GrüSSe'
Some formats which encode all characters as multiple bytes, like UTF-16, wont parse correctly at all without specify-
ing the encoding. Full list of Python standard encodings.

Index columns and trailing delimiters

If a file has one more column of data than the number of column names, the first column will be used as the
DataFrames row names:

In [82]: data = ('a,b,c\n'

....: '4,apple,bat,5.7\n'
....: '8,orange,cow,10')
....:

In [83]: pd.read_csv(StringIO(data))
Out[83]:
a b c
4 apple bat 5.7
8 orange cow 10.0

In [84]: data = ('index,a,b,c\n'

....: '4,apple,bat,5.7\n'
(continues on next page)

196 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

....: '8,orange,cow,10')
....:

In [85]: pd.read_csv(StringIO(data), index_col=0)

Out[85]:
a b c
index
4 apple bat 5.7
8 orange cow 10.0

Ordinarily, you can achieve this behavior using the index_col option.
There are some exception cases when a file has been prepared with delimiters at the end of each data line, confusing
the parser. To explicitly disable the index column inference and discard the last column, pass index_col=False:
In [86]: data = ('a,b,c\n'
....: '4,apple,bat,\n'
....: '8,orange,cow,')
....:

In [87]: print(data)
a,b,c
4,apple,bat,
8,orange,cow,

In [88]: pd.read_csv(StringIO(data))
Out[88]:
a b c
4 apple bat NaN
8 orange cow NaN

In [89]: pd.read_csv(StringIO(data), index_col=False)

Out[89]:
a b c
0 4 apple bat
1 8 orange cow
If a subset of data is being parsed using the usecols option, the index_col specification is based on that subset,
not the original data.
In [90]: data = ('a,b,c\n'
....: '4,apple,bat,\n'
....: '8,orange,cow,')
....:

In [91]: print(data)
a,b,c
4,apple,bat,
8,orange,cow,

In [92]: pd.read_csv(StringIO(data), usecols=['b', 'c'])

Out[92]:
b c
4 bat NaN
8 cow NaN

4.1. IO tools (text, CSV, HDF5, ) 197

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [93]: pd.read_csv(StringIO(data), usecols=['b', 'c'], index_col=0)

Out[93]:
b c
4 bat NaN
8 cow NaN

Date Handling

Specifying date columns

To better facilitate working with datetime data, read_csv() uses the keyword arguments parse_dates and
date_parser to allow users to specify a variety of columns and date/time formats to turn the input text data into
datetime objects.
The simplest case is to just pass in parse_dates=True:
# Use a column as an index, and parse it as dates.
In [94]: df = pd.read_csv('foo.csv', index_col=0, parse_dates=True)

In [95]: df
Out[95]:
A B C
date
2009-01-01 a 1 2
2009-01-02 b 3 4
2009-01-03 c 4 5

# These are Python datetime objects

In [96]: df.index
Out[96]: DatetimeIndex(['2009-01-01', '2009-01-02', '2009-01-03'],
,→dtype='datetime64[ns]', name='date', freq=None)

It is often the case that we may want to store date and time data separately, or store various date fields separately. the
parse_dates keyword can be used to specify a combination of columns to parse the dates and/or times from.
You can specify a list of column lists to parse_dates, the resulting date columns will be prepended to the output
(so as to not affect the existing column order) and the new column names will be the concatenation of the component
column names:
In [97]: print(open('tmp.csv').read())
KORD,19990127, 19:00:00, 18:56:00, 0.8100
KORD,19990127, 20:00:00, 19:56:00, 0.0100
KORD,19990127, 21:00:00, 20:56:00, -0.5900
KORD,19990127, 21:00:00, 21:18:00, -0.9900
KORD,19990127, 22:00:00, 21:56:00, -0.5900
KORD,19990127, 23:00:00, 22:56:00, -0.5900

In [98]: df = pd.read_csv('tmp.csv', header=None, parse_dates=[[1, 2], [1, 3]])

In [99]: df
Out[99]:
1_2 1_3 0 4
0 1999-01-27 19:00:00 1999-01-27 18:56:00 KORD 0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00 KORD 0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00 KORD -0.59
(continues on next page)

198 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

3 1999-01-27 21:00:00 1999-01-27 21:18:00 KORD -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00 KORD -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00 KORD -0.59

By default the parser removes the component date columns, but you can choose to retain them via the
keep_date_col keyword:

In [100]: df = pd.read_csv('tmp.csv', header=None, parse_dates=[[1, 2], [1, 3]],

.....: keep_date_col=True)
.....:

In [101]: df
Out[101]:
1_2 1_3 0 1 2 3 4
0 1999-01-27 19:00:00 1999-01-27 18:56:00 KORD 19990127 19:00:00 18:56:00 0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00 KORD 19990127 20:00:00 19:56:00 0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00 KORD 19990127 21:00:00 20:56:00 -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00 KORD 19990127 21:00:00 21:18:00 -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00 KORD 19990127 22:00:00 21:56:00 -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00 KORD 19990127 23:00:00 22:56:00 -0.59

Note that if you wish to combine multiple columns into a single date column, a nested list must be used. In other
words, parse_dates=[1, 2] indicates that the second and third columns should each be parsed as separate date
columns while parse_dates=[[1, 2]] means the two columns should be parsed into a single column.
You can also use a dict to specify custom name columns:

In [102]: date_spec = {'nominal': [1, 2], 'actual': [1, 3]}

In [103]: df = pd.read_csv('tmp.csv', header=None, parse_dates=date_spec)

In [104]: df
Out[104]:
nominal actual 0 4
0 1999-01-27 19:00:00 1999-01-27 18:56:00 KORD 0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00 KORD 0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00 KORD -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00 KORD -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00 KORD -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00 KORD -0.59

It is important to remember that if multiple text columns are to be parsed into a single date column, then a new column
is prepended to the data. The index_col specification is based off of this new set of columns rather than the original
data columns:

In [105]: date_spec = {'nominal': [1, 2], 'actual': [1, 3]}

In [106]: df = pd.read_csv('tmp.csv', header=None, parse_dates=date_spec,

.....: index_col=0) # index is the nominal column
.....:

In [107]: df
Out[107]:
actual 0 4
nominal
1999-01-27 19:00:00 1999-01-27 18:56:00 KORD 0.81
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 199

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

1999-01-27 20:00:00 1999-01-27 19:56:00 KORD 0.01
1999-01-27 21:00:00 1999-01-27 20:56:00 KORD -0.59
1999-01-27 21:00:00 1999-01-27 21:18:00 KORD -0.99
1999-01-27 22:00:00 1999-01-27 21:56:00 KORD -0.59
1999-01-27 23:00:00 1999-01-27 22:56:00 KORD -0.59

Note: If a column or index contains an unparsable date, the entire column or index will be returned unaltered as an
object data type. For non-standard datetime parsing, use to_datetime() after pd.read_csv.

Note: read_csv has a fast_path for parsing datetime strings in iso8601 format, e.g 2000-01-01T00:01:02+00:00 and
similar variations. If you can arrange for your data to store datetimes in this format, load times will be significantly
faster, ~20x has been observed.

Note: When passing a dict as the parse_dates argument, the order of the columns prepended is not guaranteed,
because dict objects do not impose an ordering on their keys. On Python 2.7+ you may use collections.OrderedDict
instead of a regular dict if this matters to you. Because of this, when using a dict for parse_dates in conjunction with
the index_col argument, its best to specify index_col as a column label rather then as an index on the resulting frame.

Date parsing functions

Finally, the parser allows you to specify a custom date_parser function to take full advantage of the flexibility of
the date parsing API:

In [108]: df = pd.read_csv('tmp.csv', header=None, parse_dates=date_spec,

.....: date_parser=pd.io.date_converters.parse_date_time)
.....:

In [109]: df
Out[109]:
nominal actual 0 4
0 1999-01-27 19:00:00 1999-01-27 18:56:00 KORD 0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00 KORD 0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00 KORD -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00 KORD -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00 KORD -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00 KORD -0.59

Pandas will try to call the date_parser function in three different ways. If an exception is raised, the next one is
tried:
1. date_parser is first called with one or more arrays as arguments, as defined using parse_dates (e.g.,
date_parser(['2013', '2013'], ['1', '2'])).
2. If #1 fails, date_parser is called with all the columns concatenated row-wise into a single array (e.g.,
date_parser(['2013 1', '2013 2'])).
3. If #2 fails, date_parser is called once for every row with one or more string arguments from
the columns indicated with parse_dates (e.g., date_parser('2013', '1') for the first row,
date_parser('2013', '2') for the second, etc.).

200 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Note that performance-wise, you should try these methods of parsing dates in order:
1. Try to infer the format using infer_datetime_format=True (see section below).
2. If you know the format, use pd.to_datetime(): date_parser=lambda x: pd.
to_datetime(x, format=...).
3. If you have a really non-standard format, use a custom date_parser function. For optimal performance, this
should be vectorized, i.e., it should accept arrays as arguments.
You can explore the date parsing functionality in date_converters.py and add your own. We would love to turn this
module into a community supported set of date/time parsers. To get you started, date_converters.py contains
functions to parse dual date and time columns, year/month/day columns, and year/month/day/hour/minute/second
columns. It also contains a generic_parser function so you can curry it with a function that deals with a single
date rather than the entire array.

Parsing a CSV with mixed timezones

Pandas cannot natively represent a column or index with mixed timezones. If your CSV file contains columns with a
mixture of timezones, the default result will be an object-dtype column with strings, even with parse_dates.

In [110]: content = """\

.....: a
.....: 2000-01-01T00:00:00+05:00
.....: 2000-01-01T00:00:00+06:00"""
.....:

In [111]: df = pd.read_csv(StringIO(content), parse_dates=['a'])

In [112]: df['a']
Out[112]:
0 2000-01-01 00:00:00+05:00
1 2000-01-01 00:00:00+06:00
Name: a, dtype: object

To parse the mixed-timezone values as a datetime column, pass a partially-applied to_datetime() with
utc=True as the date_parser.

In [113]: df = pd.read_csv(StringIO(content), parse_dates=['a'],

.....: date_parser=lambda col: pd.to_datetime(col, utc=True))
.....:

In [114]: df['a']
Out[114]:
0 1999-12-31 19:00:00+00:00
1 1999-12-31 18:00:00+00:00
Name: a, dtype: datetime64[ns, UTC]

Inferring datetime format

If you have parse_dates enabled for some or all of your columns, and your datetime strings are all formatted the
same way, you may get a large speed up by setting infer_datetime_format=True. If set, pandas will attempt
to guess the format of your datetime strings, and then use a faster means of parsing the strings. 5-10x parsing speeds
have been observed. pandas will fallback to the usual parsing if either the format cannot be guessed or the format that
was guessed cannot properly parse the entire column of strings. So in general, infer_datetime_format should
not have any negative consequences if enabled.

4.1. IO tools (text, CSV, HDF5, ) 201

pandas: powerful Python data analysis toolkit, Release 0.25.3

Here are some examples of datetime strings that can be guessed (All representing December 30th, 2011 at 00:00:00):
• 20111230
• 2011/12/30
• 20111230 00:00:00
• 12/30/2011 00:00:00
• 30/Dec/2011 00:00:00
• 30/December/2011 00:00:00
Note that infer_datetime_format is sensitive to dayfirst. With dayfirst=True, it will guess
01/12/2011 to be December 1st. With dayfirst=False (default) it will guess 01/12/2011 to be January 12th.

# Try to infer the format for the index column

In [115]: df = pd.read_csv('foo.csv', index_col=0, parse_dates=True,
.....: infer_datetime_format=True)
.....:

In [116]: df
Out[116]:
A B C
date
2009-01-01 a 1 2
2009-01-02 b 3 4
2009-01-03 c 4 5

International date formats

While US date formats tend to be MM/DD/YYYY, many international formats use DD/MM/YYYY instead. For
convenience, a dayfirst keyword is provided:
In [117]: print(open('tmp.csv').read())
date,value,cat
1/6/2000,5,a
2/6/2000,10,b
3/6/2000,15,c

In [118]: pd.read_csv('tmp.csv', parse_dates=[0])

Out[118]:
date value cat
0 2000-01-06 5 a
1 2000-02-06 10 b
2 2000-03-06 15 c

In [119]: pd.read_csv('tmp.csv', dayfirst=True, parse_dates=[0])

Out[119]:
date value cat
0 2000-06-01 5 a
1 2000-06-02 10 b
2 2000-06-03 15 c

202 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Specifying method for floating-point conversion

The parameter float_precision can be specified in order to use a specific floating-point converter during parsing
with the C engine. The options are the ordinary converter, the high-precision converter, and the round-trip converter
(which is guaranteed to round-trip values after writing to a file). For example:
In [120]: val = '0.3066101993807095471566981359501369297504425048828125'

In [121]: data = 'a,b,c\n1,2,{0}'.format(val)

In [122]: abs(pd.read_csv(StringIO(data), engine='c',

.....: float_precision=None)['c'][0] - float(val))
.....:
Out[122]: 1.1102230246251565e-16

In [123]: abs(pd.read_csv(StringIO(data), engine='c',

.....: float_precision='high')['c'][0] - float(val))
.....:
Out[123]: 5.551115123125783e-17

In [124]: abs(pd.read_csv(StringIO(data), engine='c',

.....: float_precision='round_trip')['c'][0] - float(val))
.....:
Out[124]: 0.0

Thousand separators

For large numbers that have been written with a thousands separator, you can set the thousands keyword to a string
of length 1 so that integers will be parsed correctly:
By default, numbers with a thousands separator will be parsed as strings:
In [125]: print(open('tmp.csv').read())
ID|level|category
Patient1|123,000|x
Patient2|23,000|y
Patient3|1,234,018|z

In [126]: df = pd.read_csv('tmp.csv', sep='|')

In [127]: df
Out[127]:
ID level category
0 Patient1 123,000 x
1 Patient2 23,000 y
2 Patient3 1,234,018 z

In [128]: df.level.dtype
Out[128]: dtype('O')
The thousands keyword allows integers to be parsed correctly:
In [129]: print(open('tmp.csv').read())
ID|level|category
Patient1|123,000|x
Patient2|23,000|y

4.1. IO tools (text, CSV, HDF5, ) 203

pandas: powerful Python data analysis toolkit, Release 0.25.3

Patient3|1,234,018|z

In [130]: df = pd.read_csv('tmp.csv', sep='|', thousands=',')

In [131]: df
Out[131]:
ID level category
0 Patient1 123000 x
1 Patient2 23000 y
2 Patient3 1234018 z

In [132]: df.level.dtype
Out[132]: dtype('int64')

NA values

To control which values are parsed as missing values (which are signified by NaN), specify a string in na_values.
If you specify a list of strings, then all values in it are considered to be missing values. If you specify a number (a
float, like 5.0 or an integer like 5), the corresponding equivalent values will also imply a missing value (in this
case effectively [5.0, 5] are recognized as NaN).
To completely override the default values that are recognized as missing, specify keep_default_na=False.
The default NaN recognized values are ['-1.#IND', '1.#QNAN', '1.#IND', '-1.#QNAN', '#N/A
N/A', '#N/A', 'N/A', 'n/a', 'NA', '#NA', 'NULL', 'null', 'NaN', '-NaN', 'nan',
'-nan', ''].
Let us consider some examples:

pd.read_csv('path_to_file.csv', na_values=[5])

In the example above 5 and 5.0 will be recognized as NaN, in addition to the defaults. A string will first be interpreted
as a numerical 5, then as a NaN.

pd.read_csv('path_to_file.csv', keep_default_na=False, na_values=[""])

Above, only an empty field will be recognized as NaN.

pd.read_csv('path_to_file.csv', keep_default_na=False, na_values=["NA", "0"])

Above, both NA and 0 as strings are NaN.

pd.read_csv('path_to_file.csv', na_values=["Nope"])

The default values, in addition to the string "Nope" are recognized as NaN.

Infinity

inf like values will be parsed as np.inf (positive infinity), and -inf as -np.inf (negative infinity). These will
ignore the case of the value, meaning Inf, will also be parsed as np.inf.

Returning Series

Using the squeeze keyword, the parser will return output with a single column as a Series:

204 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [133]: print(open('tmp.csv').read())
level
Patient1,123000
Patient2,23000
Patient3,1234018

In [134]: output = pd.read_csv('tmp.csv', squeeze=True)

In [135]: output
Out[135]:
Patient1 123000
Patient2 23000
Patient3 1234018
Name: level, dtype: int64

In [136]: type(output)
Out[136]: pandas.core.series.Series

Boolean values

The common values True, False, TRUE, and FALSE are all recognized as boolean. Occasionally you might want to
recognize other values as being boolean. To do this, use the true_values and false_values options as follows:
In [137]: data = ('a,b,c\n'
.....: '1,Yes,2\n'
.....: '3,No,4')
.....:

In [138]: print(data)
a,b,c
1,Yes,2
3,No,4

In [139]: pd.read_csv(StringIO(data))
Out[139]:
a b c
0 1 Yes 2
1 3 No 4

In [140]: pd.read_csv(StringIO(data), true_values=['Yes'], false_

,→values=['No'])

Out[140]:
a b c
0 1 True 2
1 3 False 4

Handling bad lines

Some files may have malformed lines with too few fields or too many. Lines with too few fields will have NA values
filled in the trailing fields. Lines with too many fields will raise an error by default:
In [141]: data = ('a,b,c\n'
.....: '1,2,3\n'
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 205

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

.....: '4,5,6,7\n'
.....: '8,9,10')
.....:

In [142]: pd.read_csv(StringIO(data))
---------------------------------------------------------------------------
ParserError Traceback (most recent call last)
<ipython-input-142-6388c394e6b8> in <module>
----> 1 pd.read_csv(StringIO(data))

~/sandbox/pandas-release/pandas/pandas/io/parsers.py in parser_f(filepath_or_buffer,
,→sep, delimiter, header, names, index_col, usecols, squeeze, prefix, mangle_dupe_

,→cols, dtype, engine, converters, true_values, false_values, skipinitialspace,

,→skiprows, skipfooter, nrows, na_values, keep_default_na, na_filter, verbose, skip_

,→blank_lines, parse_dates, infer_datetime_format, keep_date_col, date_parser,

,→dayfirst, cache_dates, iterator, chunksize, compression, thousands, decimal,

,→lineterminator, quotechar, quoting, doublequote, escapechar, comment, encoding,

,→dialect, error_bad_lines, warn_bad_lines, delim_whitespace, low_memory, memory_map,

,→float_precision)

683 )
684
--> 685 return _read(filepath_or_buffer, kwds)
686
687 parser_f.__name__ = name

~/sandbox/pandas-release/pandas/pandas/io/parsers.py in _read(filepath_or_buffer,
,→kwds)

461
462 try:
--> 463 data = parser.read(nrows)
464 finally:
465 parser.close()

~/sandbox/pandas-release/pandas/pandas/io/parsers.py in read(self, nrows)

1152 def read(self, nrows=None):
1153 nrows = _validate_integer("nrows", nrows)
-> 1154 ret = self._engine.read(nrows)
1155
1156 # May alter columns / col_dict

~/sandbox/pandas-release/pandas/pandas/io/parsers.py in read(self, nrows)

2057 def read(self, nrows=None):
2058 try:
-> 2059 data = self._reader.read(nrows)
2060 except StopIteration:
2061 if self._first_chunk:

~/sandbox/pandas-release/pandas/pandas/_libs/parsers.pyx in pandas._libs.parsers.
,→TextReader.read()

~/sandbox/pandas-release/pandas/pandas/_libs/parsers.pyx in pandas._libs.parsers.
,→TextReader._read_low_memory()

~/sandbox/pandas-release/pandas/pandas/_libs/parsers.pyx in pandas._libs.parsers.
,→TextReader._read_rows()

~/sandbox/pandas-release/pandas/pandas/_libs/parsers.pyx in pandas._libs.parsers.
,→TextReader._tokenize_rows() (continues on next page)

206 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

~/sandbox/pandas-release/pandas/pandas/_libs/parsers.pyx in pandas._libs.parsers.
,→raise_parser_error()

ParserError: Error tokenizing data. C error: Expected 3 fields in line 3, saw 4

You can elect to skip bad lines:

In [29]: pd.read_csv(StringIO(data), error_bad_lines=False)

Skipping line 3: expected 3 fields, saw 4

Out[29]:
a b c
0 1 2 3
1 8 9 10

You can also use the usecols parameter to eliminate extraneous column data that appear in some lines but not others:

In [30]: pd.read_csv(StringIO(data), usecols=[0, 1, 2])

Out[30]:
a b c
0 1 2 3
1 4 5 6
2 8 9 10

Dialect

The dialect keyword gives greater flexibility in specifying the file format. By default it uses the Excel dialect but
you can specify either the dialect name or a csv.Dialect instance.
Suppose you had data with unenclosed quotes:

In [143]: print(data)
label1,label2,label3
index1,"a,c,e
index2,b,d,f

By default, read_csv uses the Excel dialect and treats the double quote as the quote character, which causes it to
fail when it finds a newline before it finds the closing double quote.
We can get around this using dialect:

In [144]: import csv

In [145]: dia = csv.excel()

In [146]: dia.quoting = csv.QUOTE_NONE

In [147]: pd.read_csv(StringIO(data), dialect=dia)

Out[147]:
label1 label2 label3
index1 "a c e
index2 b d f

All of the dialect options can be specified separately by keyword arguments:

4.1. IO tools (text, CSV, HDF5, ) 207

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [148]: data = 'a,b,c~1,2,3~4,5,6'

In [149]: pd.read_csv(StringIO(data), lineterminator='~')

Out[149]:
a b c
0 1 2 3
1 4 5 6

Another common dialect option is skipinitialspace, to skip any whitespace after a delimiter:
In [150]: data = 'a, b, c\n1, 2, 3\n4, 5, 6'

In [151]: print(data)
a, b, c
1, 2, 3
4, 5, 6

In [152]: pd.read_csv(StringIO(data), skipinitialspace=True)

Out[152]:
a b c
0 1 2 3
1 4 5 6
The parsers make every attempt to do the right thing and not be fragile. Type inference is a pretty big deal. If a column
can be coerced to integer dtype without altering the contents, the parser will do so. Any non-numeric columns will
come through as object dtype as with the rest of pandas objects.

Quoting and Escape Characters

Quotes (and other escape characters) in embedded fields can be handled in any number of ways. One way is to use
backslashes; to properly parse this data, you should pass the escapechar option:
In [153]: data = 'a,b\n"hello, \\"Bob\\", nice to see you",5'

In [154]: print(data)
a,b
"hello, \"Bob\", nice to see you",5

In [155]: pd.read_csv(StringIO(data), escapechar='\\')

Out[155]:
a b
0 hello, "Bob", nice to see you 5

Files with fixed width columns

While read_csv() reads delimited data, the read_fwf() function works with data files that have known and fixed
column widths. The function parameters to read_fwf are largely the same as read_csv with two extra parameters,
and a different usage of the delimiter parameter:
• colspecs: A list of pairs (tuples) giving the extents of the fixed-width fields of each line as half-open intervals
(i.e., [from, to[ ). String value infer can be used to instruct the parser to try detecting the column specifications
from the first 100 rows of the data. Default behavior, if not specified, is to infer.
• widths: A list of field widths which can be used instead of colspecs if the intervals are contiguous.

208 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

• delimiter: Characters to consider as filler characters in the fixed-width file. Can be used to specify the filler
character of the fields if it is not spaces (e.g., ~).
Consider a typical fixed-width data file:

In [156]: print(open('bar.csv').read())
id8141 360.242940 149.910199 11950.7
id1594 444.953632 166.985655 11788.4
id1849 364.136849 183.628767 11806.2
id1230 413.836124 184.375703 11916.8
id1948 502.953953 173.237159 12468.3

In order to parse this file into a DataFrame, we simply need to supply the column specifications to the read_fwf
function along with the file name:

# Column specifications are a list of half-intervals

In [157]: colspecs = [(0, 6), (8, 20), (21, 33), (34, 43)]

In [158]: df = pd.read_fwf('bar.csv', colspecs=colspecs, header=None, index_col=0)

In [159]: df
Out[159]:
1 2 3
0
id8141 360.242940 149.910199 11950.7
id1594 444.953632 166.985655 11788.4
id1849 364.136849 183.628767 11806.2
id1230 413.836124 184.375703 11916.8
id1948 502.953953 173.237159 12468.3

Note how the parser automatically picks column names X.<column number> when header=None argument is spec-
ified. Alternatively, you can supply just the column widths for contiguous columns:

# Widths are a list of integers

In [160]: widths = [6, 14, 13, 10]

In [161]: df = pd.read_fwf('bar.csv', widths=widths, header=None)

In [162]: df
Out[162]:
0 1 2 3
0 id8141 360.242940 149.910199 11950.7
1 id1594 444.953632 166.985655 11788.4
2 id1849 364.136849 183.628767 11806.2
3 id1230 413.836124 184.375703 11916.8
4 id1948 502.953953 173.237159 12468.3

The parser will take care of extra white spaces around the columns so its ok to have extra separation between the
columns in the file.
By default, read_fwf will try to infer the files colspecs by using the first 100 rows of the file. It can do it only
in cases when the columns are aligned and correctly separated by the provided delimiter (default delimiter is
whitespace).

In [163]: df = pd.read_fwf('bar.csv', header=None, index_col=0)

In [164]: df
Out[164]:
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 209

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

1 2 3
0
id8141 360.242940 149.910199 11950.7
id1594 444.953632 166.985655 11788.4
id1849 364.136849 183.628767 11806.2
id1230 413.836124 184.375703 11916.8
id1948 502.953953 173.237159 12468.3

New in version 0.20.0.

read_fwf supports the dtype parameter for specifying the types of parsed columns to be different from the inferred
type.
In [165]: pd.read_fwf('bar.csv', header=None, index_col=0).dtypes
Out[165]:
1 float64
2 float64
3 float64
dtype: object

In [166]: pd.read_fwf('bar.csv', header=None, dtype={2: 'object'}).dtypes

Out[166]:
0 object
1 float64
2 object
3 float64
dtype: object

Indexes

Files with an implicit index column

Consider a file with one less entry in the header than the number of data column:

In [167]: print(open('foo.csv').read())
A,B,C
20090101,a,1,2
20090102,b,3,4
20090103,c,4,5

In this special case, read_csv assumes that the first column is to be used as the index of the DataFrame:

In [168]: pd.read_csv('foo.csv')
Out[168]:
A B C
20090101 a 1 2
20090102 b 3 4
20090103 c 4 5

Note that the dates werent automatically parsed. In that case you would need to do as before:

In [169]: df = pd.read_csv('foo.csv', parse_dates=True)

(continues on next page)

210 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [170]: df.index
Out[170]: DatetimeIndex(['2009-01-01', '2009-01-02', '2009-01-03'], dtype=
,→'datetime64[ns]', freq=None)

Reading an index with a MultiIndex

Suppose you have data indexed by two columns:

In [171]: print(open('data/mindex_ex.csv').read())
year,indiv,zit,xit
1977,"A",1.2,.6
1977,"B",1.5,.5
1977,"C",1.7,.8
1978,"A",.2,.06
1978,"B",.7,.2
1978,"C",.8,.3
1978,"D",.9,.5
1978,"E",1.4,.9
1979,"C",.2,.15
1979,"D",.14,.05
1979,"E",.5,.15
1979,"F",1.2,.5
1979,"G",3.4,1.9
1979,"H",5.4,2.7
1979,"I",6.4,1.2

The index_col argument to read_csv can take a list of column numbers to turn multiple columns into a
MultiIndex for the index of the returned object:
In [172]: df = pd.read_csv("data/mindex_ex.csv", index_col=[0, 1])

In [173]: df
Out[173]:
zit xit
year indiv
1977 A 1.20 0.60
B 1.50 0.50
C 1.70 0.80
1978 A 0.20 0.06
B 0.70 0.20
C 0.80 0.30
D 0.90 0.50
E 1.40 0.90
1979 C 0.20 0.15
D 0.14 0.05
E 0.50 0.15
F 1.20 0.50
G 3.40 1.90
H 5.40 2.70
I 6.40 1.20

In [174]: df.loc[1978]
Out[174]:
zit xit

4.1. IO tools (text, CSV, HDF5, ) 211

pandas: powerful Python data analysis toolkit, Release 0.25.3

indiv
A 0.2 0.06
B 0.7 0.20
C 0.8 0.30
D 0.9 0.50
E 1.4 0.90

Reading columns with a MultiIndex

By specifying list of row locations for the header argument, you can read in a MultiIndex for the columns.
Specifying non-consecutive rows will skip the intervening rows.
In [175]: from pandas.util.testing import makeCustomDataframe as mkdf

In [176]: df = mkdf(5, 3, r_idx_nlevels=2, c_idx_nlevels=4)

In [177]: df.to_csv('mi.csv')

In [178]: print(open('mi.csv').read())
C0,,C_l0_g0,C_l0_g1,C_l0_g2
C1,,C_l1_g0,C_l1_g1,C_l1_g2
C2,,C_l2_g0,C_l2_g1,C_l2_g2
C3,,C_l3_g0,C_l3_g1,C_l3_g2
R0,R1,,,
R_l0_g0,R_l1_g0,R0C0,R0C1,R0C2
R_l0_g1,R_l1_g1,R1C0,R1C1,R1C2
R_l0_g2,R_l1_g2,R2C0,R2C1,R2C2
R_l0_g3,R_l1_g3,R3C0,R3C1,R3C2
R_l0_g4,R_l1_g4,R4C0,R4C1,R4C2

In [179]: pd.read_csv('mi.csv', header=[0, 1, 2, 3], index_col=[0, 1])

Out[179]:
C0 C_l0_g0 C_l0_g1 C_l0_g2
C1 C_l1_g0 C_l1_g1 C_l1_g2
C2 C_l2_g0 C_l2_g1 C_l2_g2
C3 C_l3_g0 C_l3_g1 C_l3_g2
R0 R1
R_l0_g0 R_l1_g0 R0C0 R0C1 R0C2
R_l0_g1 R_l1_g1 R1C0 R1C1 R1C2
R_l0_g2 R_l1_g2 R2C0 R2C1 R2C2
R_l0_g3 R_l1_g3 R3C0 R3C1 R3C2
R_l0_g4 R_l1_g4 R4C0 R4C1 R4C2
read_csv is also able to interpret a more common format of multi-columns indices.
In [180]: print(open('mi2.csv').read())
,a,a,a,b,c,c
,q,r,s,t,u,v
one,1,2,3,4,5,6
two,7,8,9,10,11,12

In [181]: pd.read_csv('mi2.csv', header=[0, 1], index_col=0)

Out[181]:

212 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

a b c
q r s t u v
one 1 2 3 4 5 6
two 7 8 9 10 11 12
Note: If an index_col is not specified (e.g. you dont have an index, or wrote it with df.to_csv(...,
index=False), then any names on the columns index will be lost.

Automatically sniffing the delimiter

read_csv is capable of inferring delimited (not necessarily comma-separated) files, as pandas uses the csv.
Sniffer class of the csv module. For this, you have to specify sep=None.
In [182]: print(open('tmp2.sv').read())
:0:1:2:3
0:1.1214905765122583:-1.1011663421613171:-1.2725711408453018:0.
,→8434589457722285

1:0.8739661419816901:-1.1622548707272122:0.12618578996106738:0.
,→5057848504967111

2:0.6695152369722812:0.4833977900441433:-0.4383565886430891:-0.
,→13952146077085656

3:1.6678766138462109:0.906356209978661:0.8603041052486606:-0.
,→009413710135323125

4:-0.8075485015292924:-0.7848128653629299:-1.3155155066668116:0.
,→6875244729698119

5:-0.1572352664979729:0.30339976035788174:-0.36340691002502046:-0.
,→5526511482544121

6:0.41442095212262187:0.17517103850750262:-0.5295157789486404:-0.
,→06745694327155764

7:1.058814717443789:-0.11789792502832808:-1.8534207864364352:-0.
,→7018494437516053

8:0.26239634172416604:-1.7245959745828128:0.2765803759042711:1.
,→0730241342647273

9:0.6352851164219758:-2.1785482358583024:0.3120437647651685:1.5723784501068536

In [183]: pd.read_csv('tmp2.sv', sep=None, engine='python')

Out[183]:
Unnamed: 0 0 1 2 3
0 0 1.121491 -1.101166 -1.272571 0.843459
1 1 0.873966 -1.162255 0.126186 0.505785
2 2 0.669515 0.483398 -0.438357 -0.139521
3 3 1.667877 0.906356 0.860304 -0.009414
4 4 -0.807549 -0.784813 -1.315516 0.687524
5 5 -0.157235 0.303400 -0.363407 -0.552651
6 6 0.414421 0.175171 -0.529516 -0.067457
7 7 1.058815 -0.117898 -1.853421 -0.701849
8 8 0.262396 -1.724596 0.276580 1.073024
9 9 0.635285 -2.178548 0.312044 1.572378

Reading multiple files to create a single DataFrame

Its best to use concat() to combine multiple files. See the cookbook for an example.

4.1. IO tools (text, CSV, HDF5, ) 213

pandas: powerful Python data analysis toolkit, Release 0.25.3

Iterating through files chunk by chunk

Suppose you wish to iterate through a (potentially very large) file lazily rather than reading the entire file into memory,
such as the following:

In [184]: print(open('tmp.sv').read())
|0|1|2|3
0|1.1214905765122583|-1.1011663421613171|-1.2725711408453018|0.8434589457722285
1|0.8739661419816901|-1.1622548707272122|0.12618578996106738|0.5057848504967111
2|0.6695152369722812|0.4833977900441433|-0.4383565886430891|-0.13952146077085656
3|1.6678766138462109|0.906356209978661|0.8603041052486606|-0.009413710135323125
4|-0.8075485015292924|-0.7848128653629299|-1.3155155066668116|0.6875244729698119
5|-0.1572352664979729|0.30339976035788174|-0.36340691002502046|-0.5526511482544121
6|0.41442095212262187|0.17517103850750262|-0.5295157789486404|-0.06745694327155764
7|1.058814717443789|-0.11789792502832808|-1.8534207864364352|-0.7018494437516053
8|0.26239634172416604|-1.7245959745828128|0.2765803759042711|1.0730241342647273
9|0.6352851164219758|-2.1785482358583024|0.3120437647651685|1.5723784501068536

In [185]: table = pd.read_csv('tmp.sv', sep='|')

In [186]: table
Out[186]:
Unnamed: 0 0 1 2 3
0 0 1.121491 -1.101166 -1.272571 0.843459
1 1 0.873966 -1.162255 0.126186 0.505785
2 2 0.669515 0.483398 -0.438357 -0.139521
3 3 1.667877 0.906356 0.860304 -0.009414
4 4 -0.807549 -0.784813 -1.315516 0.687524
5 5 -0.157235 0.303400 -0.363407 -0.552651
6 6 0.414421 0.175171 -0.529516 -0.067457
7 7 1.058815 -0.117898 -1.853421 -0.701849
8 8 0.262396 -1.724596 0.276580 1.073024
9 9 0.635285 -2.178548 0.312044 1.572378

By specifying a chunksize to read_csv, the return value will be an iterable object of type TextFileReader:
In [187]: reader = pd.read_csv('tmp.sv', sep='|', chunksize=4)

In [188]: reader
Out[188]: <pandas.io.parsers.TextFileReader at 0x1c334ff710>

In [189]: for chunk in reader:

.....: print(chunk)
.....:
Unnamed: 0 0 1 2 3
0 0 1.121491 -1.101166 -1.272571 0.843459
1 1 0.873966 -1.162255 0.126186 0.505785
2 2 0.669515 0.483398 -0.438357 -0.139521
3 3 1.667877 0.906356 0.860304 -0.009414
Unnamed: 0 0 1 2 3
4 4 -0.807549 -0.784813 -1.315516 0.687524
5 5 -0.157235 0.303400 -0.363407 -0.552651
6 6 0.414421 0.175171 -0.529516 -0.067457
7 7 1.058815 -0.117898 -1.853421 -0.701849
Unnamed: 0 0 1 2 3
8 8 0.262396 -1.724596 0.276580 1.073024

214 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

9 9 0.635285 -2.178548 0.312044 1.572378

Specifying iterator=True will also return the TextFileReader object:

In [190]: reader = pd.read_csv('tmp.sv', sep='|', iterator=True)

In [191]: reader.get_chunk(5)
Out[191]:
Unnamed: 0 0 1 2 3
0 0 1.121491 -1.101166 -1.272571 0.843459
1 1 0.873966 -1.162255 0.126186 0.505785
2 2 0.669515 0.483398 -0.438357 -0.139521
3 3 1.667877 0.906356 0.860304 -0.009414
4 4 -0.807549 -0.784813 -1.315516 0.687524

Specifying the parser engine

Under the hood pandas uses a fast and efficient parser implemented in C as well as a Python implementation which is
currently more feature-complete. Where possible pandas uses the C parser (specified as engine='c'), but may fall
back to Python if C-unsupported options are specified. Currently, C-unsupported options include:
• sep other than a single character (e.g. regex separators)
• skipfooter
• sep=None with delim_whitespace=False
Specifying any of the above options will produce a ParserWarning unless the python engine is selected explicitly
using engine='python'.

Reading remote files

You can pass in a URL to a CSV file:

df = pd.read_csv('https://download.bls.gov/pub/time.series/cu/cu.item',
sep='\t')

S3 URLs are handled as well but require installing the S3Fs library:

df = pd.read_csv('s3://pandas-test/tips.csv')

If your S3 bucket requires credentials you will need to set them as environment variables or in the ~/.aws/
credentials config file, refer to the S3Fs documentation on credentials.

Writing out data

Writing to CSV format

The Series and DataFrame objects have an instance method to_csv which allows storing the contents of the
object as a comma-separated-values file. The function takes a number of arguments. Only the first is required.
• path_or_buf: A string path to the file to write or a file object. If a file object it must be opened with newline=
• sep : Field delimiter for the output file (default ,)
• na_rep: A string representation of a missing value (default )

4.1. IO tools (text, CSV, HDF5, ) 215

pandas: powerful Python data analysis toolkit, Release 0.25.3

• float_format: Format string for floating point numbers

• columns: Columns to write (default None)
• header: Whether to write out the column names (default True)
• index: whether to write row (index) names (default True)
• index_label: Column label(s) for index column(s) if desired. If None (default), and header and index are
True, then the index names are used. (A sequence should be given if the DataFrame uses MultiIndex).
• mode : Python write mode, default w
• encoding: a string representing the encoding to use if the contents are non-ASCII, for Python versions prior
to 3
• line_terminator: Character sequence denoting line end (default os.linesep)
• quoting: Set quoting rules as in csv module (default csv.QUOTE_MINIMAL). Note that if you have set
a float_format then floats are converted to strings and csv.QUOTE_NONNUMERIC will treat them as non-
numeric
• quotechar: Character used to quote fields (default )
• doublequote: Control quoting of quotechar in fields (default True)
• escapechar: Character used to escape sep and quotechar when appropriate (default None)
• chunksize: Number of rows to write at a time
• date_format: Format string for datetime objects

Writing a formatted string

The DataFrame object has an instance method to_string which allows control over the string representation of
the object. All arguments are optional:
• buf default None, for example a StringIO object
• columns default None, which columns to write
• col_space default None, minimum width of each column.
• na_rep default NaN, representation of NA value
• formatters default None, a dictionary (by column) of functions each of which takes a single argument and
returns a formatted string
• float_format default None, a function which takes a single (float) argument and returns a formatted string;
to be applied to floats in the DataFrame.
• sparsify default True, set to False for a DataFrame with a hierarchical index to print every MultiIndex key
at each row.
• index_names default True, will print the names of the indices
• index default True, will print the index (ie, row labels)
• header default True, will print the column labels
• justify default left, will print column headers left- or right-justified
The Series object also has a to_string method, but with only the buf, na_rep, float_format arguments.
There is also a length argument which, if set to True, will additionally output the length of the Series.

216 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

4.1.2 JSON

Read and write JSON format files and strings.

Writing JSON

A Series or DataFrame can be converted to a valid JSON string. Use to_json with optional parameters:
• path_or_buf : the pathname or buffer to write the output This can be None in which case a JSON string is
returned
• orient :
Series:
– default is index
– allowed values are {split, records, index}
DataFrame:
– default is columns
– allowed values are {split, records, index, columns, values, table}
The format of the JSON string

split dict like {index -> [index], columns -> [columns], data -> [values]}
records list like [{column -> value}, , {column -> value}]
index dict like {index -> {column -> value}}
columns dict like {column -> {index -> value}}
values just the values array

• date_format : string, type of date conversion, epoch for timestamp, iso for ISO8601.
• double_precision : The number of decimal places to use when encoding floating point values, default 10.
• force_ascii : force encoded string to be ASCII, default True.
• date_unit : The time unit to encode to, governs timestamp and ISO8601 precision. One of s, ms, us or ns
for seconds, milliseconds, microseconds and nanoseconds respectively. Default ms.
• default_handler : The handler to call if an object cannot otherwise be converted to a suitable format for
JSON. Takes a single argument, which is the object to convert, and returns a serializable object.
• lines : If records orient, then will write each record per line as json.
Note NaNs, NaTs and None will be converted to null and datetime objects will be converted based on the
date_format and date_unit parameters.

In [192]: dfj = pd.DataFrame(np.random.randn(5, 2), columns=list('AB'))

In [193]: json = dfj.to_json()

In [194]: json
Out[194]: '{"A":{"0":2.0027932898,"1":-1.1284197337,"2":-0.2671230751,"3":0.
,→0590811856,"4":0.2126018166},"B":{"0":0.009310115,"1":-1.2591311739,"2":1.

,→7549089729,"3":0.9464922966,"4":-0.5276761509}}'

4.1. IO tools (text, CSV, HDF5, ) 217

pandas: powerful Python data analysis toolkit, Release 0.25.3

Orient options

There are a number of different options for the format of the resulting JSON file / string. Consider the following
DataFrame and Series:

In [195]: dfjo = pd.DataFrame(dict(A=range(1, 4), B=range(4, 7), C=range(7, 10)),

.....: columns=list('ABC'), index=list('xyz'))
.....:

In [196]: dfjo
Out[196]:
A B C
x 1 4 7
y 2 5 8
z 3 6 9

In [197]: sjo = pd.Series(dict(x=15, y=16, z=17), name='D')

In [198]: sjo
Out[198]:
x 15
y 16
z 17
Name: D, dtype: int64

Column oriented (the default for DataFrame) serializes the data as nested JSON objects with column labels acting
as the primary index:

In [199]: dfjo.to_json(orient="columns")
Out[199]: '{"A":{"x":1,"y":2,"z":3},"B":{"x":4,"y":5,"z":6},"C":{"x":7,"y":8,"z":9}}'

# Not available for Series

Index oriented (the default for Series) similar to column oriented but the index labels are now primary:
In [200]: dfjo.to_json(orient="index")
Out[200]: '{"x":{"A":1,"B":4,"C":7},"y":{"A":2,"B":5,"C":8},"z":{"A":3,"B":6,
,→"C":9}}'

In [201]: sjo.to_json(orient="index")
Out[201]: '{"x":15,"y":16,"z":17}'
Record oriented serializes the data to a JSON array of column -> value records, index labels are not included. This is
useful for passing DataFrame data to plotting libraries, for example the JavaScript library d3.js:
In [202]: dfjo.to_json(orient="records")
Out[202]: '[{"A":1,"B":4,"C":7},{"A":2,"B":5,"C":8},{"A":3,"B":6,"C":9}]'

In [203]: sjo.to_json(orient="records")
Out[203]: '[15,16,17]'
Value oriented is a bare-bones option which serializes to nested JSON arrays of values only, column and index labels
are not included:

In [204]: dfjo.to_json(orient="values")
Out[204]: '[[1,4,7],[2,5,8],[3,6,9]]'

# Not available for Series

218 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Split oriented serializes to a JSON object containing separate entries for values, index and columns. Name is also
included for Series:
In [205]: dfjo.to_json(orient="split")
Out[205]: '{"columns":["A","B","C"],"index":["x","y","z"],"data":[[1,4,7],[2,
,→5,8],[3,6,9]]}'

In [206]: sjo.to_json(orient="split")
Out[206]: '{"name":"D","index":["x","y","z"],"data":[15,16,17]}'
Table oriented serializes to the JSON Table Schema, allowing for the preservation of metadata including but not
limited to dtypes and index names.

Note: Any orient option that encodes to a JSON object will not preserve the ordering of index and column labels
during round-trip serialization. If you wish to preserve label ordering use the split option as it uses ordered containers.

Date handling

Writing in ISO date format:

In [207]: dfd = pd.DataFrame(np.random.randn(5, 2), columns=list('AB'))

In [208]: dfd['date'] = pd.Timestamp('20130101')

In [209]: dfd = dfd.sort_index(1, ascending=False)

In [210]: json = dfd.to_json(date_format='iso')

In [211]: json
Out[211]: '{"date":{"0":"2013-01-01T00:00:00.000Z","1":"2013-01-01T00:00:00.000Z","2":
,→"2013-01-01T00:00:00.000Z","3":"2013-01-01T00:00:00.000Z","4":"2013-01-01T00:00:00.

,→000Z"},"B":{"0":0.3903383957,"1":-0.5223681486,"2":2.0249145293,"3":2.1144885256,"4

,→":0.5337588359},"A":{"0":-1.0954121534,"1":-0.147141856,"2":0.6305826658,"3":1.

,→5730764249,"4":0.6200376615}}'

Writing in ISO date format, with microseconds:

In [212]: json = dfd.to_json(date_format='iso', date_unit='us')

In [213]: json
Out[213]: '{"date":{"0":"2013-01-01T00:00:00.000000Z","1":"2013-01-01T00:00:00.000000Z
,→","2":"2013-01-01T00:00:00.000000Z","3":"2013-01-01T00:00:00.000000Z","4":"2013-01-

,→01T00:00:00.000000Z"},"B":{"0":0.3903383957,"1":-0.5223681486,"2":2.0249145293,"3

,→":2.1144885256,"4":0.5337588359},"A":{"0":-1.0954121534,"1":-0.147141856,"2":0.

,→6305826658,"3":1.5730764249,"4":0.6200376615}}'

Epoch timestamps, in seconds:

In [214]: json = dfd.to_json(date_format='epoch', date_unit='s')

In [215]: json
Out[215]: '{"date":{"0":1356998400,"1":1356998400,"2":1356998400,"3":1356998400,"4
,→":1356998400},"B":{"0":0.3903383957,"1":-0.5223681486,"2":2.0249145293,"3":2.

,→1144885256,"4":0.5337588359},"A":{"0":-1.0954121534,"1":-0.147141856,"2":0.

,→6305826658,"3":1.5730764249,"4":0.6200376615}}'

4.1. IO tools (text, CSV, HDF5, ) 219

pandas: powerful Python data analysis toolkit, Release 0.25.3

Writing to a file, with a date index and a date column:

In [216]: dfj2 = dfj.copy()

In [217]: dfj2['date'] = pd.Timestamp('20130101')

In [218]: dfj2['ints'] = list(range(5))

In [219]: dfj2['bools'] = True

In [220]: dfj2.index = pd.date_range('20130101', periods=5)

In [221]: dfj2.to_json('test.json')

In [222]: with open('test.json') as fh:

.....: print(fh.read())
.....:
{"A":{"1356998400000":2.0027932898,"1357084800000":-1.1284197337,"1357171200000":-0.
,→2671230751,"1357257600000":0.0590811856,"1357344000000":0.2126018166},"B":{

,→"1356998400000":0.009310115,"1357084800000":-1.2591311739,"1357171200000":1.

,→7549089729,"1357257600000":0.9464922966,"1357344000000":-0.5276761509},"date":{

,→"1356998400000":1356998400000,"1357084800000":1356998400000,"1357171200000

,→":1356998400000,"1357257600000":1356998400000,"1357344000000":1356998400000},"ints":

,→{"1356998400000":0,"1357084800000":1,"1357171200000":2,"1357257600000":3,

,→"1357344000000":4},"bools":{"1356998400000":true,"1357084800000":true,"1357171200000

,→":true,"1357257600000":true,"1357344000000":true}}

Fallback behavior

If the JSON serializer cannot handle the container contents directly it will fall back in the following manner:
• if the dtype is unsupported (e.g. np.complex) then the default_handler, if provided, will be called for
each value, otherwise an exception is raised.
• if an object is unsupported it will attempt the following:
– check if the object has defined a toDict method and call it. A toDict method should return a dict
which will then be JSON serialized.
– invoke the default_handler if one was provided.
– convert the object to a dict by traversing its contents. However this will often fail with an
OverflowError or give unexpected results.
In general the best approach for unsupported objects or dtypes is to provide a default_handler. For example:

>>> DataFrame([1.0, 2.0, complex(1.0, 2.0)]).to_json() # raises

RuntimeError: Unhandled numpy dtype 15

can be dealt with by specifying a simple default_handler:

In [223]: pd.DataFrame([1.0, 2.0, complex(1.0, 2.0)]).to_json(default_handler=str)

Out[223]: '{"0":{"0":"(1+0j)","1":"(2+0j)","2":"(1+2j)"}}'

220 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Reading JSON

Reading a JSON string to pandas object can take a number of parameters. The parser will try to parse a DataFrame
if typ is not supplied or is None. To explicitly force Series parsing, pass typ=series
• filepath_or_buffer : a VALID JSON string or file handle / StringIO. The string could be a URL. Valid
URL schemes include http, ftp, S3, and file. For file URLs, a host is expected. For instance, a local file could be
file ://localhost/path/to/table.json
• typ : type of object to recover (series or frame), default frame
• orient :
Series :
– default is index
– allowed values are {split, records, index}
DataFrame
– default is columns
– allowed values are {split, records, index, columns, values, table}
The format of the JSON string

split dict like {index -> [index], columns -> [columns], data -> [values]}
records list like [{column -> value}, , {column -> value}]
index dict like {index -> {column -> value}}
columns dict like {column -> {index -> value}}
values just the values array
table adhering to the JSON Table Schema

• dtype : if True, infer dtypes, if a dict of column to dtype, then use those, if False, then dont infer dtypes at
all, default is True, apply only to the data.
• convert_axes : boolean, try to convert the axes to the proper dtypes, default is True
• convert_dates : a list of columns to parse for dates; If True, then try to parse date-like columns, default
is True.
• keep_default_dates : boolean, default True. If parsing dates, then parse the default date-like columns.
• numpy : direct decoding to NumPy arrays. default is False; Supports numeric data only, although labels may
be non-numeric. Also note that the JSON ordering MUST be the same for each term if numpy=True.
• precise_float : boolean, default False. Set to enable usage of higher precision (strtod) function when
decoding string to double values. Default (False) is to use fast but less precise builtin functionality.
• date_unit : string, the timestamp unit to detect if converting dates. Default None. By default the timestamp
precision will be detected, if this is not desired then pass one of s, ms, us or ns to force timestamp precision to
seconds, milliseconds, microseconds or nanoseconds respectively.
• lines : reads file as one json object per line.
• encoding : The encoding to use to decode py3 bytes.
• chunksize : when used in combination with lines=True, return a JsonReader which reads in chunksize
lines per iteration.

4.1. IO tools (text, CSV, HDF5, ) 221

pandas: powerful Python data analysis toolkit, Release 0.25.3

The parser will raise one of ValueError/TypeError/AssertionError if the JSON is not parseable.
If a non-default orient was used when encoding to JSON be sure to pass the same option here so that decoding
produces sensible results, see Orient Options for an overview.

Data conversion

The default of convert_axes=True, dtype=True, and convert_dates=True will try to parse the axes, and
all of the data into appropriate types, including dates. If you need to override specific dtypes, pass a dict to dtype.
convert_axes should only be set to False if you need to preserve string-like numbers (e.g. 1, 2) in an axes.

Note: Large integer values may be converted to dates if convert_dates=True and the data and / or column
labels appear date-like. The exact threshold depends on the date_unit specified. date-like means that the column
label meets one of the following criteria:
• it ends with '_at'
• it ends with '_time'
• it begins with 'timestamp'
• it is 'modified'
• it is 'date'

Warning: When reading JSON data, automatic coercing into dtypes has some quirks:
• an index can be reconstructed in a different order from serialization, that is, the returned order is not guaran-
teed to be the same as before serialization
• a column that was float data will be converted to integer if it can be done safely, e.g. a column of 1.
• bool columns will be converted to integer on reconstruction
Thus there are times where you may want to specify specific dtypes via the dtype keyword argument.

Reading from a JSON string:

In [224]: pd.read_json(json)
Out[224]:
date B A
0 2013-01-01 0.390338 -1.095412
1 2013-01-01 -0.522368 -0.147142
2 2013-01-01 2.024915 0.630583
3 2013-01-01 2.114489 1.573076
4 2013-01-01 0.533759 0.620038

Reading from a file:

In [225]: pd.read_json('test.json')
Out[225]:
A B date ints bools
2013-01-01 2.002793 0.009310 2013-01-01 0 True
2013-01-02 -1.128420 -1.259131 2013-01-01 1 True
2013-01-03 -0.267123 1.754909 2013-01-01 2 True
(continues on next page)

222 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2013-01-04 0.059081 0.946492 2013-01-01 3 True
2013-01-05 0.212602 -0.527676 2013-01-01 4 True

Dont convert any data (but still convert axes and dates):

In [226]: pd.read_json('test.json', dtype=object).dtypes

Out[226]:
A object
B object
date object
ints object
bools object
dtype: object

Specify dtypes for conversion:

In [227]: pd.read_json('test.json', dtype={'A': 'float32', 'bools': 'int8'}).dtypes

Out[227]:
A float32
B float64
date datetime64[ns]
ints int64
bools int8
dtype: object

Preserve string indices:

In [228]: si = pd.DataFrame(np.zeros((4, 4)), columns=list(range(4)),
.....: index=[str(i) for i in range(4)])
.....:

In [229]: si
Out[229]:
0 1 2 3
0 0.0 0.0 0.0 0.0
1 0.0 0.0 0.0 0.0
2 0.0 0.0 0.0 0.0
3 0.0 0.0 0.0 0.0

In [230]: si.index
Out[230]: Index(['0', '1', '2', '3'], dtype='object')

In [231]: si.columns
Out[231]: Int64Index([0, 1, 2, 3], dtype='int64')

In [232]: json = si.to_json()

In [233]: sij = pd.read_json(json, convert_axes=False)

In [234]: sij
Out[234]:
0 1 2 3
0 0 0 0 0
1 0 0 0 0
2 0 0 0 0

4.1. IO tools (text, CSV, HDF5, ) 223

pandas: powerful Python data analysis toolkit, Release 0.25.3

3 0 0 0 0

In [235]: sij.index
Out[235]: Index(['0', '1', '2', '3'], dtype='object')

In [236]: sij.columns
Out[236]: Index(['0', '1', '2', '3'], dtype='object')
Dates written in nanoseconds need to be read back in nanoseconds:

In [237]: json = dfj2.to_json(date_unit='ns')

# Try to parse timestamps as milliseconds -> Won't Work

In [238]: dfju = pd.read_json(json, date_unit='ms')

In [239]: dfju
Out[239]:
A B date ints bools
1356998400000000000 2.002793 0.009310 1356998400000000000 0 True
1357084800000000000 -1.128420 -1.259131 1356998400000000000 1 True
1357171200000000000 -0.267123 1.754909 1356998400000000000 2 True
1357257600000000000 0.059081 0.946492 1356998400000000000 3 True
1357344000000000000 0.212602 -0.527676 1356998400000000000 4 True

# Let pandas detect the correct precision

In [240]: dfju = pd.read_json(json)

In [241]: dfju
Out[241]:
A B date ints bools
2013-01-01 2.002793 0.009310 2013-01-01 0 True
2013-01-02 -1.128420 -1.259131 2013-01-01 1 True
2013-01-03 -0.267123 1.754909 2013-01-01 2 True
2013-01-04 0.059081 0.946492 2013-01-01 3 True
2013-01-05 0.212602 -0.527676 2013-01-01 4 True

# Or specify that all timestamps are in nanoseconds

In [242]: dfju = pd.read_json(json, date_unit='ns')

In [243]: dfju
Out[243]:
A B date ints bools
2013-01-01 2.002793 0.009310 2013-01-01 0 True
2013-01-02 -1.128420 -1.259131 2013-01-01 1 True
2013-01-03 -0.267123 1.754909 2013-01-01 2 True
2013-01-04 0.059081 0.946492 2013-01-01 3 True
2013-01-05 0.212602 -0.527676 2013-01-01 4 True

The Numpy parameter

Note: This supports numeric data only. Index and columns labels may be non-numeric, e.g. strings, dates etc.

If numpy=True is passed to read_json an attempt will be made to sniff an appropriate dtype during deserialization
and to subsequently decode directly to NumPy arrays, bypassing the need for intermediate Python objects.

224 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

This can provide speedups if you are deserialising a large amount of numeric data:

In [244]: randfloats = np.random.uniform(-100, 1000, 10000)

In [245]: randfloats.shape = (1000, 10)

In [246]: dffloats = pd.DataFrame(randfloats, columns=list('ABCDEFGHIJ'))

In [247]: jsonfloats = dffloats.to_json()

In [248]: %timeit pd.read_json(jsonfloats)

9.45 ms +- 266 us per loop (mean +- std. dev. of 7 runs, 100 loops each)

In [249]: %timeit pd.read_json(jsonfloats, numpy=True)

7.21 ms +- 356 us per loop (mean +- std. dev. of 7 runs, 100 loops each)

The speedup is less noticeable for smaller datasets:

In [250]: jsonfloats = dffloats.head(100).to_json()

In [251]: %timeit pd.read_json(jsonfloats)

7.05 ms +- 307 us per loop (mean +- std. dev. of 7 runs, 100 loops each)

In [252]: %timeit pd.read_json(jsonfloats, numpy=True)

5.68 ms +- 136 us per loop (mean +- std. dev. of 7 runs, 100 loops each)

Warning: Direct NumPy decoding makes a number of assumptions and may fail or produce unexpected output if
these assumptions are not satisfied:
• data is numeric.
• data is uniform. The dtype is sniffed from the first value decoded. A ValueError may be raised, or
incorrect output may be produced if this condition is not satisfied.
• labels are ordered. Labels are only read from the first container, it is assumed that each subsequent row /
column has been encoded in the same order. This should be satisfied if the data was encoded using to_json
but may not be the case if the JSON is from another source.

Normalization

pandas provides a utility function to take a dict or list of dicts and normalize this semi-structured data into a flat table.

In [253]: from pandas.io.json import json_normalize

In [254]: data = [{'id': 1, 'name': {'first': 'Coleen', 'last': 'Volk'}},

.....: {'name': {'given': 'Mose', 'family': 'Regner'}},
.....: {'id': 2, 'name': 'Faye Raker'}]
.....:

In [255]: json_normalize(data)
Out[255]:
id name.first name.last name.given name.family name
0 1.0 Coleen Volk NaN NaN NaN
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 225

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

1 NaN NaN NaN Mose Regner NaN
2 2.0 NaN NaN NaN NaN Faye Raker

In [256]: data = [{'state': 'Florida',

.....: 'shortname': 'FL',
.....: 'info': {'governor': 'Rick Scott'},
.....: 'counties': [{'name': 'Dade', 'population': 12345},
.....: {'name': 'Broward', 'population': 40000},
.....: {'name': 'Palm Beach', 'population': 60000}]},
.....: {'state': 'Ohio',
.....: 'shortname': 'OH',
.....: 'info': {'governor': 'John Kasich'},
.....: 'counties': [{'name': 'Summit', 'population': 1234},
.....: {'name': 'Cuyahoga', 'population': 1337}]}]
.....:

In [257]: json_normalize(data, 'counties', ['state', 'shortname', ['info', 'governor

,→']])

Out[257]:
name population state shortname info.governor
0 Dade 12345 Florida FL Rick Scott
1 Broward 40000 Florida FL Rick Scott
2 Palm Beach 60000 Florida FL Rick Scott
3 Summit 1234 Ohio OH John Kasich
4 Cuyahoga 1337 Ohio OH John Kasich

The max_level parameter provides more control over which level to end normalization. With max_level=1 the follow-
ing snippet normalizes until 1st nesting level of the provided dict.

In [258]: data = [{'CreatedBy': {'Name': 'User001'},

.....: 'Lookup': {'TextField': 'Some text',
.....: 'UserField': {'Id': 'ID001',
.....: 'Name': 'Name001'}},
.....: 'Image': {'a': 'b'}
.....: }]
.....:

In [259]: json_normalize(data, max_level=1)

Out[259]:
CreatedBy.Name Lookup.TextField Lookup.UserField Image.a
0 User001 Some text {'Id': 'ID001', 'Name': 'Name001'} b

Line delimited json

New in version 0.19.0.

pandas is able to read and write line-delimited json files that are common in data processing pipelines using Hadoop
or Spark.
New in version 0.21.0.
For line-delimited json files, pandas can also return an iterator which reads in chunksize lines at a time. This can
be useful for large files or to read from a stream.
In [260]: jsonl = '''
.....: {"a": 1, "b": 2}

226 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

.....: {"a": 3, "b": 4}

.....: '''
.....:

In [261]: df = pd.read_json(jsonl, lines=True)

In [262]: df
Out[262]:
a b
0 1 2
1 3 4

In [263]: df.to_json(orient='records', lines=True)

Out[263]: '{"a":1,"b":2}\n{"a":3,"b":4}'

# reader is an iterator that returns `chunksize` lines each iteration

In [264]: reader = pd.read_json(StringIO(jsonl), lines=True, chunksize=1)

In [265]: reader
Out[265]: <pandas.io.json._json.JsonReader at 0x1c336c2c10>

In [266]: for chunk in reader:

.....: print(chunk)
.....:
Empty DataFrame
Columns: []
Index: []
a b
0 1 2
a b
1 3 4

Table schema

New in version 0.20.0.

Table Schema is a spec for describing tabular datasets as a JSON object. The JSON includes information on the field
names, types, and other attributes. You can use the orient table to build a JSON string with two fields, schema and
data.
In [267]: df = pd.DataFrame({'A': [1, 2, 3],
.....: 'B': ['a', 'b', 'c'],
.....: 'C': pd.date_range('2016-01-01', freq='d',
,→periods=3)},

.....: index=pd.Index(range(3), name='idx'))

.....:

In [268]: df
Out[268]:
A B C
idx
0 1 a 2016-01-01
1 2 b 2016-01-02
2 3 c 2016-01-03

4.1. IO tools (text, CSV, HDF5, ) 227

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [269]: df.to_json(orient='table', date_format="iso")

Out[269]: '{"schema": {"fields":[{"name":"idx","type":"integer"},
,→{"name":"A","type":"integer"},{"name":"B","type":"string"},{"name":"C",

,→"type":"datetime"}],"primaryKey":["idx"],"pandas_version":"0.20.0"},

,→"data": [{"idx":0,"A":1,"B":"a","C":"2016-01-01T00:00:00.000Z"},{"idx":1,

,→"A":2,"B":"b","C":"2016-01-02T00:00:00.000Z"},{"idx":2,"A":3,"B":"c",

,→"C":"2016-01-03T00:00:00.000Z"}]}'

The schema field contains the fields key, which itself contains a list of column name to type pairs, including the
Index or MultiIndex (see below for a list of types). The schema field also contains a primaryKey field if the
(Multi)index is unique.
The second field, data, contains the serialized data with the records orient. The index is included, and any
datetimes are ISO 8601 formatted, as required by the Table Schema spec.
The full list of types supported are described in the Table Schema spec. This table shows the mapping from pandas
types:

Pandas type Table Schema type

int64 integer
float64 number
bool boolean
datetime64[ns] datetime
timedelta64[ns] duration
categorical any
object str

A few notes on the generated table schema:

• The schema object contains a pandas_version field. This contains the version of pandas dialect of the
schema, and will be incremented with each revision.
• All dates are converted to UTC when serializing. Even timezone naive values, which are treated as UTC with
an offset of 0.
In [270]: from pandas.io.json import build_table_schema

In [271]: s = pd.Series(pd.date_range('2016', periods=4))

In [272]: build_table_schema(s)
Out[272]:
{'fields': [{'name': 'index', 'type': 'integer'},
{'name': 'values', 'type': 'datetime'}],
'primaryKey': ['index'],
'pandas_version': '0.20.0'}

• datetimes with a timezone (before serializing), include an additional field tz with the time zone name (e.g.
'US/Central').
In [273]: s_tz = pd.Series(pd.date_range('2016', periods=12,
.....: tz='US/Central'))
.....:

In [274]: build_table_schema(s_tz)
Out[274]:
{'fields': [{'name': 'index', 'type': 'integer'},
(continues on next page)

228 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

{'name': 'values', 'type': 'datetime', 'tz': 'US/Central'}],
'primaryKey': ['index'],
'pandas_version': '0.20.0'}

• Periods are converted to timestamps before serialization, and so have the same behavior of being converted to
UTC. In addition, periods will contain and additional field freq with the periods frequency, e.g. 'A-DEC'.

In [275]: s_per = pd.Series(1, index=pd.period_range('2016', freq='A-DEC',

.....: periods=4))
.....:

In [276]: build_table_schema(s_per)
Out[276]:
{'fields': [{'name': 'index', 'type': 'datetime', 'freq': 'A-DEC'},
{'name': 'values', 'type': 'integer'}],
'primaryKey': ['index'],
'pandas_version': '0.20.0'}

• Categoricals use the any type and an enum constraint listing the set of possible values. Additionally, an
ordered field is included:

In [277]: s_cat = pd.Series(pd.Categorical(['a', 'b', 'a']))

In [278]: build_table_schema(s_cat)
Out[278]:
{'fields': [{'name': 'index', 'type': 'integer'},
{'name': 'values',
'type': 'any',
'constraints': {'enum': ['a', 'b']},
'ordered': False}],
'primaryKey': ['index'],
'pandas_version': '0.20.0'}

• A primaryKey field, containing an array of labels, is included if the index is unique:

In [279]: s_dupe = pd.Series([1, 2], index=[1, 1])

In [280]: build_table_schema(s_dupe)
Out[280]:
{'fields': [{'name': 'index', 'type': 'integer'},
{'name': 'values', 'type': 'integer'}],
'pandas_version': '0.20.0'}

• The primaryKey behavior is the same with MultiIndexes, but in this case the primaryKey is an array:

In [281]: s_multi = pd.Series(1, index=pd.MultiIndex.from_product([('a', 'b'),

.....: (0, 1)]))
.....:

In [282]: build_table_schema(s_multi)
Out[282]:
{'fields': [{'name': 'level_0', 'type': 'string'},
{'name': 'level_1', 'type': 'integer'},
{'name': 'values', 'type': 'integer'}],
'primaryKey': FrozenList(['level_0', 'level_1']),
'pandas_version': '0.20.0'}

4.1. IO tools (text, CSV, HDF5, ) 229

pandas: powerful Python data analysis toolkit, Release 0.25.3

• The default naming roughly follows these rules:

– For series, the object.name is used. If thats none, then the name is values
– For DataFrames, the stringified version of the column name is used
– For Index (not MultiIndex), index.name is used, with a fallback to index if that is None.
– For MultiIndex, mi.names is used. If any level has no name, then level_<i> is used.
New in version 0.23.0.
read_json also accepts orient='table' as an argument. This allows for the preservation of metadata such as
dtypes and index names in a round-trippable manner.
In [283]: df = pd.DataFrame({'foo': [1, 2, 3, 4],
.....: 'bar': ['a', 'b', 'c', 'd'],
.....: 'baz': pd.date_range('2018-01-01', freq='d', periods=4),
.....: 'qux': pd.Categorical(['a', 'b', 'c', 'c'])
.....: }, index=pd.Index(range(4), name='idx'))
.....:

In [284]: df
Out[284]:
foo bar baz qux
idx
0 1 a 2018-01-01 a
1 2 b 2018-01-02 b
2 3 c 2018-01-03 c
3 4 d 2018-01-04 c

In [285]: df.dtypes
Out[285]:
foo int64
bar object
baz datetime64[ns]
qux category
dtype: object

In [286]: df.to_json('test.json', orient='table')

In [287]: new_df = pd.read_json('test.json', orient='table')

In [288]: new_df
Out[288]:
foo bar baz qux
idx
0 1 a 2018-01-01 a
1 2 b 2018-01-02 b
2 3 c 2018-01-03 c
3 4 d 2018-01-04 c

In [289]: new_df.dtypes
Out[289]:
foo int64
bar object
baz datetime64[ns]

230 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

qux category
dtype: object
Please note that the literal string index as the name of an Index is not round-trippable, nor are any names begin-
ning with 'level_' within a MultiIndex. These are used by default in DataFrame.to_json() to indicate
missing values and the subsequent read cannot distinguish the intent.
In [290]: df.index.name = 'index'

In [291]: df.to_json('test.json', orient='table')

In [292]: new_df = pd.read_json('test.json', orient='table')

In [293]: print(new_df.index.name)
None

4.1.3 HTML

Reading HTML content

Warning: We highly encourage you to read the HTML Table Parsing gotchas below regarding the issues sur-
rounding the BeautifulSoup4/html5lib/lxml parsers.

The top-level read_html() function can accept an HTML string/file/URL and will parse HTML tables into list of
pandas DataFrames. Lets look at a few examples.

Note: read_html returns a list of DataFrame objects, even if there is only a single table contained in the
HTML content.

Read a URL with no options:

In [294]: url = 'https://www.fdic.gov/bank/individual/failed/banklist.html'

In [295]: dfs = pd.read_html(url)

In [296]: dfs
Out[296]:
[ Bank Name City ... Closing Date
,→Updated Date

0 City National Bank of New Jersey Newark ... November 1, 2019

,→November 1, 2019

1 Resolute Bank Maumee ... October 25, 2019

,→October 25, 2019

2 Louisa Community Bank Louisa ... October 25, 2019

,→October 28, 2019

3 The Enloe State Bank Cooper ... May 31, 2019

,→August 22, 2019

4 Washington Federal Bank for Savings Chicago ... December 15, 2017
,→July 24, 2019

.. ... ... ... ...

,→ ...
554 Superior Bank, FSB Hinsdale ... July 27, 2001
,→August 19, 2014
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 231

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

555 Malta National Bank Malta ... May 3, 2001
,→November 18, 2002

556 First Alliance Bank & Trust Co. Manchester ... February 2, 2001
,→February 18, 2003

557 National State Bank of Metropolis Metropolis ... December 14, 2000
,→March 17, 2005

558 Bank of Honolulu Honolulu ... October 13, 2000

,→March 17, 2005

[559 rows x 7 columns]]

Note: The data from the above URL changes every Monday so the resulting data above and the data below may be
slightly different.

Read in the content of the file from the above URL and pass it to read_html as a string:
In [297]: with open(file_path, 'r') as f:
.....: dfs = pd.read_html(f.read())
.....:

In [298]: dfs
Out[298]:
[ Bank Name City ... Closing Date
,→ Updated Date
0 Banks of Wisconsin d/b/a Bank of Kenosha Kenosha ... May 31, 2013
,→ May 31, 2013
1 Central Arizona Bank Scottsdale ... May 14, 2013
,→ May 20, 2013
2 Sunrise Bank Valdosta ... May 10, 2013
,→ May 21, 2013
3 Pisgah Community Bank Asheville ... May 10, 2013
,→ May 14, 2013
4 Douglas County Bank Douglasville ... April 26, 2013
,→ May 16, 2013
.. ... ... ... ...
,→ ...
500 Superior Bank, FSB Hinsdale ... July 27, 2001
,→ June 5, 2012
501 Malta National Bank Malta ... May 3, 2001
,→November 18, 2002

502 First Alliance Bank & Trust Co. Manchester ... February 2, 2001
,→February 18, 2003

503 National State Bank of Metropolis Metropolis ... December 14, 2000
,→ March 17, 2005
504 Bank of Honolulu Honolulu ... October 13, 2000
,→ March 17, 2005

[505 rows x 7 columns]]

You can even pass in an instance of StringIO if you so desire:

In [299]: with open(file_path, 'r') as f:
.....: sio = StringIO(f.read())
.....:

(continues on next page)

232 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [300]: dfs = pd.read_html(sio)

In [301]: dfs
Out[301]:
[ Bank Name City ... Closing Date
,→ Updated Date
0 Banks of Wisconsin d/b/a Bank of Kenosha Kenosha ... May 31, 2013
,→ May 31, 2013
1 Central Arizona Bank Scottsdale ... May 14, 2013
,→ May 20, 2013
2 Sunrise Bank Valdosta ... May 10, 2013
,→ May 21, 2013
3 Pisgah Community Bank Asheville ... May 10, 2013
,→ May 14, 2013
4 Douglas County Bank Douglasville ... April 26, 2013
,→ May 16, 2013
.. ... ... ... ...
,→ ...
500 Superior Bank, FSB Hinsdale ... July 27, 2001
,→ June 5, 2012
501 Malta National Bank Malta ... May 3, 2001
,→November 18, 2002

502 First Alliance Bank & Trust Co. Manchester ... February 2, 2001
,→February 18, 2003

503 National State Bank of Metropolis Metropolis ... December 14, 2000
,→ March 17, 2005
504 Bank of Honolulu Honolulu ... October 13, 2000
,→ March 17, 2005

[505 rows x 7 columns]]

Note: The following examples are not run by the IPython evaluator due to the fact that having so many network-
accessing functions slows down the documentation build. If you spot an error or an example that doesnt run, please do
not hesitate to report it over on pandas GitHub issues page.

Read a URL and match a table that contains specific text:

match = 'Metcalf Bank'

df_list = pd.read_html(url, match=match)

Specify a header row (by default <th> or <td> elements located within a <thead> are used to form the column
index, if multiple rows are contained within <thead> then a MultiIndex is created); if specified, the header row is
taken from the data minus the parsed header elements (<th> elements).

dfs = pd.read_html(url, header=0)

Specify an index column:

dfs = pd.read_html(url, index_col=0)

Specify a number of rows to skip:

dfs = pd.read_html(url, skiprows=0)

Specify a number of rows to skip using a list (xrange (Python 2 only) works as well):

4.1. IO tools (text, CSV, HDF5, ) 233

pandas: powerful Python data analysis toolkit, Release 0.25.3

dfs = pd.read_html(url, skiprows=range(2))

Specify an HTML attribute:

dfs1 = pd.read_html(url, attrs={'id': 'table'})

dfs2 = pd.read_html(url, attrs={'class': 'sortable'})
print(np.array_equal(dfs1[0], dfs2[0])) # Should be True

Specify values that should be converted to NaN:

dfs = pd.read_html(url, na_values=['No Acquirer'])

New in version 0.19.

Specify whether to keep the default set of NaN values:

dfs = pd.read_html(url, keep_default_na=False)

New in version 0.19.

Specify converters for columns. This is useful for numerical text data that has leading zeros. By default columns that
are numerical are cast to numeric types and the leading zeros are lost. To avoid this, we can convert these columns to
strings.

url_mcc = 'https://en.wikipedia.org/wiki/Mobile_country_code'
dfs = pd.read_html(url_mcc, match='Telekom Albania', header=0,
converters={'MNC': str})

New in version 0.19.

Use some combination of the above:

dfs = pd.read_html(url, match='Metcalf Bank', index_col=0)

Read in pandas to_html output (with some loss of floating point precision):

df = pd.DataFrame(np.random.randn(2, 2))
s = df.to_html(float_format='{0:.40g}'.format)
dfin = pd.read_html(s, index_col=0)

The lxml backend will raise an error on a failed parse if that is the only parser you provide. If you only have a single
parser you can provide just a string, but it is considered good practice to pass a list with one string if, for example, the
function expects a sequence of strings. You may use:

dfs = pd.read_html(url, 'Metcalf Bank', index_col=0, flavor=['lxml'])

Or you could pass flavor='lxml' without a list:

dfs = pd.read_html(url, 'Metcalf Bank', index_col=0, flavor='lxml')

However, if you have bs4 and html5lib installed and pass None or ['lxml', 'bs4'] then the parse will most
likely succeed. Note that as soon as a parse succeeds, the function will return.

dfs = pd.read_html(url, 'Metcalf Bank', index_col=0, flavor=['lxml', 'bs4'])

234 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Writing to HTML files

DataFrame objects have an instance method to_html which renders the contents of the DataFrame as an HTML
table. The function arguments are as in the method to_string described above.

Note: Not all of the possible options for DataFrame.to_html are shown here for brevitys sake. See to_html()
for the full set of options.

In [302]: df = pd.DataFrame(np.random.randn(2, 2))

In [303]: df
Out[303]:
0 1
0 -1.050304 1.131622
1 -0.692581 -1.174172

In [304]: print(df.to_html()) # raw html

<table border="1" class="dataframe">
<thead>
<tr style="text-align: right;">
<th></th>
<th>0</th>
<th>1</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
<td>-1.050304</td>
<td>1.131622</td>
</tr>
<tr>
<th>1</th>
<td>-0.692581</td>
<td>-1.174172</td>
</tr>
</tbody>
</table>
HTML:
The columns argument will limit the columns shown:

In [305]: print(df.to_html(columns=[0]))
<table border="1" class="dataframe">
<thead>
<tr style="text-align: right;">
<th></th>
<th>0</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 235

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

<td>-1.050304</td>
</tr>
<tr>
<th>1</th>
<td>-0.692581</td>
</tr>
</tbody>
</table>

HTML:
float_format takes a Python callable to control the precision of floating point values:

In [306]: print(df.to_html(float_format='{0:.10f}'.format))
<table border="1" class="dataframe">
<thead>
<tr style="text-align: right;">
<th></th>
<th>0</th>
<th>1</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
<td>-1.0503044154</td>
<td>1.1316218324</td>
</tr>
<tr>
<th>1</th>
<td>-0.6925807265</td>
<td>-1.1741715747</td>
</tr>
</tbody>
</table>

HTML:
bold_rows will make the row labels bold by default, but you can turn that off:

In [307]: print(df.to_html(bold_rows=False))
<table border="1" class="dataframe">
<thead>
<tr style="text-align: right;">
<th></th>
<th>0</th>
<th>1</th>
</tr>
</thead>
<tbody>
<tr>
<td>0</td>
<td>-1.050304</td>
<td>1.131622</td>
</tr>
<tr>
<td>1</td>
(continues on next page)

236 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

<td>-0.692581</td>
<td>-1.174172</td>
</tr>
</tbody>
</table>

The classes argument provides the ability to give the resulting HTML table CSS classes. Note that these classes
are appended to the existing 'dataframe' class.
In [308]: print(df.to_html(classes=['awesome_table_class', 'even_more_awesome_class
,→']))

<table border="1" class="dataframe awesome_table_class even_more_awesome_class">

<thead>
<tr style="text-align: right;">
<th></th>
<th>0</th>
<th>1</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
<td>-1.050304</td>
<td>1.131622</td>
</tr>
<tr>
<th>1</th>
<td>-0.692581</td>
<td>-1.174172</td>
</tr>
</tbody>
</table>

The render_links argument provides the ability to add hyperlinks to cells that contain URLs.
New in version 0.24.
In [309]: url_df = pd.DataFrame({
.....: 'name': ['Python', 'Pandas'],
.....: 'url': ['https://www.python.org/', 'http://pandas.pydata.org']})
.....:

In [310]: print(url_df.to_html(render_links=True))
<table border="1" class="dataframe">
<thead>
<tr style="text-align: right;">
<th></th>
<th>name</th>
<th>url</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
<td>Python</td>
<td><a href="https://www.python.org/" target="_blank">https://www.python.org/</
,→a></td>

(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 237

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

</tr>
<tr>
<th>1</th>
<td>Pandas</td>
<td><a href="http://pandas.pydata.org" target="_blank">http://pandas.pydata.org
,→</a></td>

</tr>
</tbody>
</table>

HTML:
Finally, the escape argument allows you to control whether the <, > and & characters escaped in the resulting HTML
(by default it is True). So to get the HTML without escaped characters pass escape=False
In [311]: df = pd.DataFrame({'a': list('&<>'), 'b': np.random.randn(3)})

Escaped:
In [312]: print(df.to_html())
<table border="1" class="dataframe">
<thead>
<tr style="text-align: right;">
<th></th>
<th>a</th>
<th>b</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
<td>&amp;</td>
<td>1.254800</td>
</tr>
<tr>
<th>1</th>
<td>&lt;</td>
<td>1.131996</td>
</tr>
<tr>
<th>2</th>
<td>&gt;</td>
<td>-1.311021</td>
</tr>
</tbody>
</table>

Not escaped:
In [313]: print(df.to_html(escape=False))
<table border="1" class="dataframe">
<thead>
<tr style="text-align: right;">
<th></th>
<th>a</th>
<th>b</th>
</tr>
(continues on next page)

238 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

</thead>
<tbody>
<tr>
<th>0</th>
<td>&</td>
<td>1.254800</td>
</tr>
<tr>
<th>1</th>
<td><</td>
<td>1.131996</td>
</tr>
<tr>
<th>2</th>
<td>></td>
<td>-1.311021</td>
</tr>
</tbody>
</table>

Note: Some browsers may not show a difference in the rendering of the previous two HTML tables.

HTML Table Parsing Gotchas

There are some versioning issues surrounding the libraries that are used to parse HTML tables in the top-level pandas
io function read_html.
Issues with lxml
• Benefits
– lxml is very fast.
– lxml requires Cython to install correctly.
• Drawbacks
– lxml does not make any guarantees about the results of its parse unless it is given strictly valid markup.
– In light of the above, we have chosen to allow you, the user, to use the lxml backend, but this backend
will use html5lib if lxml fails to parse
– It is therefore highly recommended that you install both BeautifulSoup4 and html5lib, so that you will
still get a valid result (provided everything else is valid) even if lxml fails.
Issues with BeautifulSoup4 using lxml as a backend
• The above issues hold here as well since BeautifulSoup4 is essentially just a wrapper around a parser backend.
Issues with BeautifulSoup4 using html5lib as a backend
• Benefits
– html5lib is far more lenient than lxml and consequently deals with real-life markup in a much saner way
rather than just, e.g., dropping an element without notifying you.
– html5lib generates valid HTML5 markup from invalid markup automatically. This is extremely important
for parsing HTML tables, since it guarantees a valid document. However, that does NOT mean that it is
correct, since the process of fixing markup does not have a single definition.

4.1. IO tools (text, CSV, HDF5, ) 239

pandas: powerful Python data analysis toolkit, Release 0.25.3

– html5lib is pure Python and requires no additional build steps beyond its own installation.
• Drawbacks
– The biggest drawback to using html5lib is that it is slow as molasses. However consider the fact that many
tables on the web are not big enough for the parsing algorithm runtime to matter. It is more likely that the
bottleneck will be in the process of reading the raw text from the URL over the web, i.e., IO (input-output).
For very large tables, this might not be true.

4.1.4 Excel files

The read_excel() method can read Excel 2003 (.xls) files using the xlrd Python module. Excel 2007+ (.
xlsx) files can be read using either xlrd or openpyxl. The to_excel() instance method is used for saving
a DataFrame to Excel. Generally the semantics are similar to working with csv data. See the cookbook for some
advanced strategies.

Reading Excel files

In the most basic use-case, read_excel takes a path to an Excel file, and the sheet_name indicating which sheet
to parse.

# Returns a DataFrame
pd.read_excel('path_to_file.xls', sheet_name='Sheet1')

ExcelFile class

To facilitate working with multiple sheets from the same file, the ExcelFile class can be used to wrap the file and
can be passed into read_excel There will be a performance benefit for reading multiple sheets as the file is read
into memory only once.

xlsx = pd.ExcelFile('path_to_file.xls')
df = pd.read_excel(xlsx, 'Sheet1')

The ExcelFile class can also be used as a context manager.

with pd.ExcelFile('path_to_file.xls') as xls:

df1 = pd.read_excel(xls, 'Sheet1')
df2 = pd.read_excel(xls, 'Sheet2')

The sheet_names property will generate a list of the sheet names in the file.
The primary use-case for an ExcelFile is parsing multiple sheets with different parameters:

data = {}
# For when Sheet1's format differs from Sheet2
with pd.ExcelFile('path_to_file.xls') as xls:
data['Sheet1'] = pd.read_excel(xls, 'Sheet1', index_col=None,
na_values=['NA'])
data['Sheet2'] = pd.read_excel(xls, 'Sheet2', index_col=1)

Note that if the same parsing parameters are used for all sheets, a list of sheet names can simply be passed to
read_excel with no loss in performance.

240 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

# using the ExcelFile class

data = {}
with pd.ExcelFile('path_to_file.xls') as xls:
data['Sheet1'] = pd.read_excel(xls, 'Sheet1', index_col=None,
na_values=['NA'])
data['Sheet2'] = pd.read_excel(xls, 'Sheet2', index_col=None,
na_values=['NA'])

# equivalent using the read_excel function

data = pd.read_excel('path_to_file.xls', ['Sheet1', 'Sheet2'],
index_col=None, na_values=['NA'])

ExcelFile can also be called with a xlrd.book.Book object as a parameter. This allows the user to control
how the excel file is read. For example, sheets can be loaded on demand by calling xlrd.open_workbook() with
on_demand=True.

import xlrd
xlrd_book = xlrd.open_workbook('path_to_file.xls', on_demand=True)
with pd.ExcelFile(xlrd_book) as xls:
df1 = pd.read_excel(xls, 'Sheet1')
df2 = pd.read_excel(xls, 'Sheet2')

Specifying sheets

Note: The second argument is sheet_name, not to be confused with ExcelFile.sheet_names.

Note: An ExcelFiles attribute sheet_names provides access to a list of sheets.

• The arguments sheet_name allows specifying the sheet or sheets to read.

• The default value for sheet_name is 0, indicating to read the first sheet
• Pass a string to refer to the name of a particular sheet in the workbook.
• Pass an integer to refer to the index of a sheet. Indices follow Python convention, beginning at 0.
• Pass a list of either strings or integers, to return a dictionary of specified sheets.
• Pass a None to return a dictionary of all available sheets.

# Returns a DataFrame
pd.read_excel('path_to_file.xls', 'Sheet1', index_col=None, na_values=['NA'])

Using the sheet index:

# Returns a DataFrame
pd.read_excel('path_to_file.xls', 0, index_col=None, na_values=['NA'])

Using all default values:

# Returns a DataFrame
pd.read_excel('path_to_file.xls')

Using None to get all sheets:

4.1. IO tools (text, CSV, HDF5, ) 241

pandas: powerful Python data analysis toolkit, Release 0.25.3

# Returns a dictionary of DataFrames

pd.read_excel('path_to_file.xls', sheet_name=None)

Using a list to get multiple sheets:

# Returns the 1st and 4th sheet, as a dictionary of DataFrames.

pd.read_excel('path_to_file.xls', sheet_name=['Sheet1', 3])

read_excel can read more than one sheet, by setting sheet_name to either a list of sheet names, a list of sheet
positions, or None to read all sheets. Sheets can be specified by sheet index or sheet name, using an integer or string,
respectively.

Reading a MultiIndex

read_excel can read a MultiIndex index, by passing a list of columns to index_col and a MultiIndex
column by passing a list of rows to header. If either the index or columns have serialized level names those will
be read in as well by specifying the rows/columns that make up the levels.
For example, to read in a MultiIndex index without names:

In [314]: df = pd.DataFrame({'a': [1, 2, 3, 4], 'b': [5, 6, 7, 8]},

.....: index=pd.MultiIndex.from_product([['a', 'b'], ['c', 'd
,→']]))

.....:

In [315]: df.to_excel('path_to_file.xlsx')

In [316]: df = pd.read_excel('path_to_file.xlsx', index_col=[0, 1])

In [317]: df
Out[317]:
a b
a c 1 5
d 2 6
b c 3 7
d 4 8

If the index has level names, they will parsed as well, using the same parameters.

In [318]: df.index = df.index.set_names(['lvl1', 'lvl2'])

In [319]: df.to_excel('path_to_file.xlsx')

In [320]: df = pd.read_excel('path_to_file.xlsx', index_col=[0, 1])

In [321]: df
Out[321]:
a b
lvl1 lvl2
a c 1 5
d 2 6
b c 3 7
d 4 8

If the source file has both MultiIndex index and columns, lists specifying each should be passed to index_col
and header:

242 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [322]: df.columns = pd.MultiIndex.from_product([['a'], ['b', 'd']],

.....: names=['c1', 'c2'])
.....:

In [323]: df.to_excel('path_to_file.xlsx')

In [324]: df = pd.read_excel('path_to_file.xlsx', index_col=[0, 1], header=[0, 1])

In [325]: df
Out[325]:
c1 a
c2 b d
lvl1 lvl2
a c 1 5
d 2 6
b c 3 7
d 4 8

Parsing specific columns

It is often the case that users will insert columns to do temporary computations in Excel and you may not want to read
in those columns. read_excel takes a usecols keyword to allow you to specify a subset of columns to parse.
Deprecated since version 0.24.0.
Passing in an integer for usecols has been deprecated. Please pass in a list of ints from 0 to usecols inclusive
instead.
If usecols is an integer, then it is assumed to indicate the last column to be parsed.

pd.read_excel('path_to_file.xls', 'Sheet1', usecols=2)

You can also specify a comma-delimited set of Excel columns and ranges as a string:

pd.read_excel('path_to_file.xls', 'Sheet1', usecols='A,C:E')

If usecols is a list of integers, then it is assumed to be the file column indices to be parsed.

pd.read_excel('path_to_file.xls', 'Sheet1', usecols=[0, 2, 3])

Element order is ignored, so usecols=[0, 1] is the same as [1, 0].

New in version 0.24.
If usecols is a list of strings, it is assumed that each string corresponds to a column name provided either by the
user in names or inferred from the document header row(s). Those strings define which columns will be parsed:

pd.read_excel('path_to_file.xls', 'Sheet1', usecols=['foo', 'bar'])

Element order is ignored, so usecols=['baz', 'joe'] is the same as ['joe', 'baz'].

New in version 0.24.
If usecols is callable, the callable function will be evaluated against the column names, returning names where the
callable function evaluates to True.

pd.read_excel('path_to_file.xls', 'Sheet1', usecols=lambda x: x.isalpha())

4.1. IO tools (text, CSV, HDF5, ) 243

pandas: powerful Python data analysis toolkit, Release 0.25.3

Parsing dates

Datetime-like values are normally automatically converted to the appropriate dtype when reading the excel file. But
if you have a column of strings that look like dates (but are not actually formatted as dates in excel), you can use the
parse_dates keyword to parse those strings to datetimes:

pd.read_excel('path_to_file.xls', 'Sheet1', parse_dates=['date_strings'])

Cell converters

It is possible to transform the contents of Excel cells via the converters option. For instance, to convert a column
to boolean:

pd.read_excel('path_to_file.xls', 'Sheet1', converters={'MyBools': bool})

This options handles missing values and treats exceptions in the converters as missing data. Transformations are
applied cell by cell rather than to the column as a whole, so the array dtype is not guaranteed. For instance, a column
of integers with missing values cannot be transformed to an array with integer dtype, because NaN is strictly a float.
You can manually mask missing data to recover integer dtype:

def cfun(x):
return int(x) if x else -1

pd.read_excel('path_to_file.xls', 'Sheet1', converters={'MyInts': cfun})

Dtype specifications

New in version 0.20.

As an alternative to converters, the type for an entire column can be specified using the dtype keyword, which takes a
dictionary mapping column names to types. To interpret data with no type inference, use the type str or object.

pd.read_excel('path_to_file.xls', dtype={'MyInts': 'int64', 'MyText': str})

Writing Excel files

Writing Excel files to disk

To write a DataFrame object to a sheet of an Excel file, you can use the to_excel instance method. The arguments
are largely the same as to_csv described above, the first argument being the name of the excel file, and the optional
second argument the name of the sheet to which the DataFrame should be written. For example:

df.to_excel('path_to_file.xlsx', sheet_name='Sheet1')

Files with a .xls extension will be written using xlwt and those with a .xlsx extension will be written using
xlsxwriter (if available) or openpyxl.
The DataFrame will be written in a way that tries to mimic the REPL output. The index_label will be placed
in the second row instead of the first. You can place it in the first row by setting the merge_cells option in
to_excel() to False:

244 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

df.to_excel('path_to_file.xlsx', index_label='label', merge_cells=False)

In order to write separate DataFrames to separate sheets in a single Excel file, one can pass an ExcelWriter.
with pd.ExcelWriter('path_to_file.xlsx') as writer:
df1.to_excel(writer, sheet_name='Sheet1')
df2.to_excel(writer, sheet_name='Sheet2')

Note: Wringing a little more performance out of read_excel Internally, Excel stores all numeric data as floats.
Because this can produce unexpected behavior when reading in data, pandas defaults to trying to convert integers to
floats if it doesnt lose information (1.0 --> 1). You can pass convert_float=False to disable this behavior,
which may give a slight performance improvement.

Writing Excel files to memory

Pandas supports writing Excel files to buffer-like objects such as StringIO or BytesIO using ExcelWriter.
# Safe import for either Python 2.x or 3.x
try:
from io import BytesIO
except ImportError:
from cStringIO import StringIO as BytesIO

bio = BytesIO()

# By setting the 'engine' in the ExcelWriter constructor.

writer = pd.ExcelWriter(bio, engine='xlsxwriter')
df.to_excel(writer, sheet_name='Sheet1')

# Save the workbook

writer.save()

# Seek to the beginning and read to copy the workbook to a variable in memory
bio.seek(0)
workbook = bio.read()

Note: engine is optional but recommended. Setting the engine determines the version of workbook produced.
Setting engine='xlrd' will produce an Excel 2003-format workbook (xls). Using either 'openpyxl' or
'xlsxwriter' will produce an Excel 2007-format workbook (xlsx). If omitted, an Excel 2007-formatted workbook
is produced.

Excel writer engines

Pandas chooses an Excel writer via two methods:

1. the engine keyword argument
2. the filename extension (via the default specified in config options)
By default, pandas uses the XlsxWriter for .xlsx, openpyxl for .xlsm, and xlwt for .xls files. If you have multiple
engines installed, you can set the default engine through setting the config options io.excel.xlsx.writer and
io.excel.xls.writer. pandas will fall back on openpyxl for .xlsx files if Xlsxwriter is not available.

4.1. IO tools (text, CSV, HDF5, ) 245

pandas: powerful Python data analysis toolkit, Release 0.25.3

To specify which writer you want to use, you can pass an engine keyword argument to to_excel and to
ExcelWriter. The built-in engines are:
• openpyxl: version 2.4 or higher is required
• xlsxwriter
• xlwt

# By setting the 'engine' in the DataFrame 'to_excel()' methods.

df.to_excel('path_to_file.xlsx', sheet_name='Sheet1', engine='xlsxwriter')

# By setting the 'engine' in the ExcelWriter constructor.

writer = pd.ExcelWriter('path_to_file.xlsx', engine='xlsxwriter')

# Or via pandas configuration.

from pandas import options # noqa: E402
options.io.excel.xlsx.writer = 'xlsxwriter'

df.to_excel('path_to_file.xlsx', sheet_name='Sheet1')

Style and formatting

The look and feel of Excel worksheets created from pandas can be modified using the following parameters on the
DataFrames to_excel method.
• float_format : Format string for floating point numbers (default None).
• freeze_panes : A tuple of two integers representing the bottommost row and rightmost column to freeze.
Each of these parameters is one-based, so (1, 1) will freeze the first row and first column (default None).
Using the Xlsxwriter engine provides many options for controlling the format of an Excel worksheet created with
the to_excel method. Excellent examples can be found in the Xlsxwriter documentation here: https://xlsxwriter.
readthedocs.io/working_with_pandas.html

4.1.5 OpenDocument Spreadsheets

New in version 0.25.

The read_excel() method can also read OpenDocument spreadsheets using the odfpy module. The semantics
and features for reading OpenDocument spreadsheets match what can be done for Excel files using engine='odf'.

# Returns a DataFrame
pd.read_excel('path_to_file.ods', engine='odf')

Note: Currently pandas only supports reading OpenDocument spreadsheets. Writing is not implemented.

4.1.6 Clipboard

A handy way to grab data is to use the read_clipboard() method, which takes the contents of the clipboard
buffer and passes them to the read_csv method. For instance, you can copy the following text to the clipboard
(CTRL-C on many operating systems):

246 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

A B C
x 1 4 p
y 2 5 q
z 3 6 r

And then import the data directly to a DataFrame by calling:

>>> clipdf = pd.read_clipboard()

>>> clipdf
A B C
x 1 4 p
y 2 5 q
z 3 6 r

The to_clipboard method can be used to write the contents of a DataFrame to the clipboard. Following which
you can paste the clipboard contents into other applications (CTRL-V on many operating systems). Here we illustrate
writing a DataFrame into clipboard and reading it back.

>>> df = pd.DataFrame({'A': [1, 2, 3],

... 'B': [4, 5, 6],
... 'C': ['p', 'q', 'r']},
... index=['x', 'y', 'z'])
>>> df
A B C
x 1 4 p
y 2 5 q
z 3 6 r
>>> df.to_clipboard()
>>> pd.read_clipboard()
A B C
x 1 4 p
y 2 5 q
z 3 6 r

We can see that we got the same content back, which we had earlier written to the clipboard.

Note: You may need to install xclip or xsel (with PyQt5, PyQt4 or qtpy) on Linux to use these methods.

4.1.7 Pickling

All pandas objects are equipped with to_pickle methods which use Pythons cPickle module to save data struc-
tures to disk using the pickle format.

In [326]: df
Out[326]:
c1 a
c2 b d
lvl1 lvl2
a c 1 5
d 2 6
b c 3 7
d 4 8

In [327]: df.to_pickle('foo.pkl')

4.1. IO tools (text, CSV, HDF5, ) 247

pandas: powerful Python data analysis toolkit, Release 0.25.3

The read_pickle function in the pandas namespace can be used to load any pickled pandas object (or any other
pickled object) from file:

In [328]: pd.read_pickle('foo.pkl')
Out[328]:
c1 a
c2 b d
lvl1 lvl2
a c 1 5
d 2 6
b c 3 7
d 4 8

Warning: Loading pickled data received from untrusted sources can be unsafe.
See: https://docs.python.org/3/library/pickle.html

Warning: read_pickle() is only guaranteed backwards compatible back to pandas version 0.20.3

Compressed pickle files

New in version 0.20.0.

read_pickle(), DataFrame.to_pickle() and Series.to_pickle() can read and write compressed
pickle files. The compression types of gzip, bz2, xz are supported for reading and writing. The zip file format
only supports reading and must contain only one data file to be read.
The compression type can be an explicit parameter or be inferred from the file extension. If infer, then use gzip,
bz2, zip, or xz if filename ends in '.gz', '.bz2', '.zip', or '.xz', respectively.

In [329]: df = pd.DataFrame({
.....: 'A': np.random.randn(1000),
.....: 'B': 'foo',
.....: 'C': pd.date_range('20130101', periods=1000, freq='s')})
.....:

In [330]: df
Out[330]:
A B C
0 -0.053113 foo 2013-01-01 00:00:00
1 0.348832 foo 2013-01-01 00:00:01
2 -0.162729 foo 2013-01-01 00:00:02
3 -1.269943 foo 2013-01-01 00:00:03
4 -0.481824 foo 2013-01-01 00:00:04
.. ... ... ...
995 -1.001718 foo 2013-01-01 00:16:35
996 -0.471336 foo 2013-01-01 00:16:36
997 -0.071712 foo 2013-01-01 00:16:37
998 0.578273 foo 2013-01-01 00:16:38
999 0.595708 foo 2013-01-01 00:16:39

[1000 rows x 3 columns]

Using an explicit compression type:

248 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [331]: df.to_pickle("data.pkl.compress", compression="gzip")

In [332]: rt = pd.read_pickle("data.pkl.compress", compression="gzip")

In [333]: rt
Out[333]:
A B C
0 -0.053113 foo 2013-01-01 00:00:00
1 0.348832 foo 2013-01-01 00:00:01
2 -0.162729 foo 2013-01-01 00:00:02
3 -1.269943 foo 2013-01-01 00:00:03
4 -0.481824 foo 2013-01-01 00:00:04
.. ... ... ...
995 -1.001718 foo 2013-01-01 00:16:35
996 -0.471336 foo 2013-01-01 00:16:36
997 -0.071712 foo 2013-01-01 00:16:37
998 0.578273 foo 2013-01-01 00:16:38
999 0.595708 foo 2013-01-01 00:16:39

[1000 rows x 3 columns]

Inferring compression type from the extension:

In [334]: df.to_pickle("data.pkl.xz", compression="infer")

In [335]: rt = pd.read_pickle("data.pkl.xz", compression="infer")

In [336]: rt
Out[336]:
A B C
0 -0.053113 foo 2013-01-01 00:00:00
1 0.348832 foo 2013-01-01 00:00:01
2 -0.162729 foo 2013-01-01 00:00:02
3 -1.269943 foo 2013-01-01 00:00:03
4 -0.481824 foo 2013-01-01 00:00:04
.. ... ... ...
995 -1.001718 foo 2013-01-01 00:16:35
996 -0.471336 foo 2013-01-01 00:16:36
997 -0.071712 foo 2013-01-01 00:16:37
998 0.578273 foo 2013-01-01 00:16:38
999 0.595708 foo 2013-01-01 00:16:39

[1000 rows x 3 columns]

The default is to infer:

In [337]: df.to_pickle("data.pkl.gz")

In [338]: rt = pd.read_pickle("data.pkl.gz")

In [339]: rt
Out[339]:
A B C
0 -0.053113 foo 2013-01-01 00:00:00
1 0.348832 foo 2013-01-01 00:00:01
2 -0.162729 foo 2013-01-01 00:00:02
3 -1.269943 foo 2013-01-01 00:00:03
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 249

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

4 -0.481824 foo 2013-01-01 00:00:04
.. ... ... ...
995 -1.001718 foo 2013-01-01 00:16:35
996 -0.471336 foo 2013-01-01 00:16:36
997 -0.071712 foo 2013-01-01 00:16:37
998 0.578273 foo 2013-01-01 00:16:38
999 0.595708 foo 2013-01-01 00:16:39

[1000 rows x 3 columns]

In [340]: df["A"].to_pickle("s1.pkl.bz2")

In [341]: rt = pd.read_pickle("s1.pkl.bz2")

In [342]: rt
Out[342]:
0 -0.053113
1 0.348832
2 -0.162729
3 -1.269943
4 -0.481824
...
995 -1.001718
996 -0.471336
997 -0.071712
998 0.578273
999 0.595708
Name: A, Length: 1000, dtype: float64

4.1.8 msgpack

pandas supports the msgpack format for object serialization. This is a lightweight portable binary format, similar
to binary JSON, that is highly space efficient, and provides good performance both on the writing (serialization), and
reading (deserialization).

Warning: The msgpack format is deprecated as of 0.25 and will be removed in a future version. It is recommended
to use pyarrow for on-the-wire transmission of pandas objects.

Warning: read_msgpack() is only guaranteed backwards compatible back to pandas version 0.20.3

In [343]: df = pd.DataFrame(np.random.rand(5, 2), columns=list('AB'))

In [344]: df.to_msgpack('foo.msg')

In [345]: pd.read_msgpack('foo.msg')
Out[345]:
A B
0 0.541029 0.554672
1 0.150831 0.503287
2 0.834267 0.881894
(continues on next page)

250 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

3 0.706066 0.726912
4 0.639300 0.067928

In [346]: s = pd.Series(np.random.rand(5), index=pd.date_range('20130101', periods=5))

You can pass a list of objects and you will receive them back on deserialization.
In [347]: pd.to_msgpack('foo.msg', df, 'foo', np.array([1, 2, 3]), s)

In [348]: pd.read_msgpack('foo.msg')
Out[348]:
[ A B
0 0.541029 0.554672
1 0.150831 0.503287
2 0.834267 0.881894
3 0.706066 0.726912
4 0.639300 0.067928, 'foo', array([1, 2, 3]), 2013-01-01 0.753932
2013-01-02 0.676180
2013-01-03 0.924728
2013-01-04 0.338661
2013-01-05 0.592241
Freq: D, dtype: float64]

You can pass iterator=True to iterate over the unpacked results:

In [349]: for o in pd.read_msgpack('foo.msg', iterator=True):
.....: print(o)
.....:
A B
0 0.541029 0.554672
1 0.150831 0.503287
2 0.834267 0.881894
3 0.706066 0.726912
4 0.639300 0.067928
foo
[1 2 3]
2013-01-01 0.753932
2013-01-02 0.676180
2013-01-03 0.924728
2013-01-04 0.338661
2013-01-05 0.592241
Freq: D, dtype: float64

You can pass append=True to the writer to append to an existing pack:

In [350]: df.to_msgpack('foo.msg', append=True)

In [351]: pd.read_msgpack('foo.msg')
Out[351]:
[ A B
0 0.541029 0.554672
1 0.150831 0.503287
2 0.834267 0.881894
3 0.706066 0.726912
4 0.639300 0.067928, 'foo', array([1, 2, 3]), 2013-01-01 0.753932
2013-01-02 0.676180
2013-01-03 0.924728
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 251

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2013-01-04 0.338661
2013-01-05 0.592241
Freq: D, dtype: float64, A B
0 0.541029 0.554672
1 0.150831 0.503287
2 0.834267 0.881894
3 0.706066 0.726912
4 0.639300 0.067928]

Unlike other io methods, to_msgpack is available on both a per-object basis, df.to_msgpack() and using the
top-level pd.to_msgpack(...) where you can pack arbitrary collections of Python lists, dicts, scalars, while
intermixing pandas objects.
In [352]: pd.to_msgpack('foo2.msg', {'dict': [{'df': df}, {'string': 'foo'},
.....: {'scalar': 1.}, {'s': s}]})
.....:

In [353]: pd.read_msgpack('foo2.msg')
Out[353]:
{'dict': ({'df': A B
0 0.541029 0.554672
1 0.150831 0.503287
2 0.834267 0.881894
3 0.706066 0.726912
4 0.639300 0.067928},
{'string': 'foo'},
{'scalar': 1.0},
{'s': 2013-01-01 0.753932
2013-01-02 0.676180
2013-01-03 0.924728
2013-01-04 0.338661
2013-01-05 0.592241
Freq: D, dtype: float64})}

Read/write API

Msgpacks can also be read from and written to strings.

In [354]: df.to_msgpack()
Out[354]: b"\x84\xa3typ\xadblock_
,→manager\xa5klass\xa9DataFrame\xa4axes\x92\x86\xa3typ\xa5index\xa5klass\xa5Index\xa4name\xc0\xa5dtyp

,→index\xa5klass\xaaRangeIndex\xa4name\xc0\xa5start\x00\xa4stop\x05\xa4step\x01\xa6blocks\x91\x86\xa4

,→`)\x0c3kN\xc3?\xac\xa1:JQ\xb2\xea?\x8c|\xa87\x17\x98\xe6?\xf3H\x83*&u\xe4?\xd4S\xff

,→{\xe0\xbf\xe1?\xd3'2\xea\xed\x1a\xe0?6\x00'gy8\xec?S\x98/\xe7\xdcB\xe7?

,→`\xdbr\xed\xbac\xb1?

,→\xa5shape\x92\x02\x05\xa5dtype\xa7float64\xa5klass\xaaFloatBlock\xa8compress\xc0"

Furthermore you can concatenate the strings to produce a list of the original objects.
In [355]: pd.read_msgpack(df.to_msgpack() + s.to_msgpack())
Out[355]:
[ A B
0 0.541029 0.554672
1 0.150831 0.503287
2 0.834267 0.881894
(continues on next page)

252 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

3 0.706066 0.726912
4 0.639300 0.067928, 2013-01-01 0.753932
2013-01-02 0.676180
2013-01-03 0.924728
2013-01-04 0.338661
2013-01-05 0.592241
Freq: D, dtype: float64]

4.1.9 HDF5 (PyTables)

HDFStore is a dict-like object which reads and writes pandas using the high performance HDF5 format using the
excellent PyTables library. See the cookbook for some advanced strategies

Warning: pandas requires PyTables >= 3.0.0. There is a indexing bug in PyTables < 3.2 which may appear
when querying stores using an index. If you see a subset of results being returned, upgrade to PyTables >= 3.2.
Stores created previously will need to be rewritten using the updated version.

In [356]: store = pd.HDFStore('store.h5')

In [357]: print(store)
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

Objects can be written to the file just like adding key-value pairs to a dict:

In [358]: index = pd.date_range('1/1/2000', periods=8)

In [359]: s = pd.Series(np.random.randn(5), index=['a', 'b', 'c', 'd', 'e'])

In [360]: df = pd.DataFrame(np.random.randn(8, 3), index=index,

.....: columns=['A', 'B', 'C'])
.....:

# store.put('s', s) is an equivalent method

In [361]: store['s'] = s

In [362]: store['df'] = df

In [363]: store
Out[363]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

In a current or later Python session, you can retrieve stored objects:

# store.get('df') is an equivalent method
In [364]: store['df']
Out[364]:
A B C
2000-01-01 0.263806 1.913465 -0.274536
2000-01-02 0.283334 1.798001 -0.053258
2000-01-03 0.799684 -0.733715 1.205089
2000-01-04 0.131478 0.100995 -0.764260

4.1. IO tools (text, CSV, HDF5, ) 253

pandas: powerful Python data analysis toolkit, Release 0.25.3

2000-01-05 1.891112 -1.410251 0.752883

2000-01-06 -0.274852 -0.667027 -0.688782
2000-01-07 0.621607 -1.300199 0.050119
2000-01-08 -0.999591 -0.320658 -1.922640

# dotted (attribute) access provides get as well

In [365]: store.df
Out[365]:
A B C
2000-01-01 0.263806 1.913465 -0.274536
2000-01-02 0.283334 1.798001 -0.053258
2000-01-03 0.799684 -0.733715 1.205089
2000-01-04 0.131478 0.100995 -0.764260
2000-01-05 1.891112 -1.410251 0.752883
2000-01-06 -0.274852 -0.667027 -0.688782
2000-01-07 0.621607 -1.300199 0.050119
2000-01-08 -0.999591 -0.320658 -1.922640
Deletion of the object specified by the key:
# store.remove('df') is an equivalent method
In [366]: del store['df']

In [367]: store
Out[367]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

Closing a Store and using a context manager:

In [368]: store.close()

In [369]: store
Out[369]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

In [370]: store.is_open
Out[370]: False

# Working with, and automatically closing the store using a context manager
In [371]: with pd.HDFStore('store.h5') as store:
.....: store.keys()
.....:

Read/write API

HDFStore supports an top-level API using read_hdf for reading and to_hdf for writing, similar to how
read_csv and to_csv work.
In [372]: df_tl = pd.DataFrame({'A': list(range(5)), 'B': list(range(5))})

In [373]: df_tl.to_hdf('store_tl.h5', 'table', append=True)

In [374]: pd.read_hdf('store_tl.h5', 'table', where=['index>2'])

(continues on next page)

254 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

Out[374]:
A B
3 3 3
4 4 4

HDFStore will by default not drop rows that are all missing. This behavior can be changed by setting dropna=True.

In [375]: df_with_missing = pd.DataFrame({'col1': [0, np.nan, 2],

.....: 'col2': [1, np.nan, np.nan]})
.....:

In [376]: df_with_missing
Out[376]:
col1 col2
0 0.0 1.0
1 NaN NaN
2 2.0 NaN

In [377]: df_with_missing.to_hdf('file.h5', 'df_with_missing',

.....: format='table', mode='w')
.....:

In [378]: pd.read_hdf('file.h5', 'df_with_missing')

Out[378]:
col1 col2
0 0.0 1.0
1 NaN NaN
2 2.0 NaN

In [379]: df_with_missing.to_hdf('file.h5', 'df_with_missing',

.....: format='table', mode='w', dropna=True)
.....:

In [380]: pd.read_hdf('file.h5', 'df_with_missing')

Out[380]:
col1 col2
0 0.0 1.0
2 2.0 NaN

Fixed format

The examples above show storing using put, which write the HDF5 to PyTables in a fixed array format, called
the fixed format. These types of stores are not appendable once written (though you can simply remove them and
rewrite). Nor are they queryable; they must be retrieved in their entirety. They also do not support dataframes with
non-unique column names. The fixed format stores offer very fast writing and slightly faster reading than table
stores. This format is specified by default when using put or to_hdf or by format='fixed' or format='f'.

Warning: A fixed format will raise a TypeError if you try to retrieve using a where:
>>> pd.DataFrame(np.random.randn(10, 2)).to_hdf('test_fixed.h5', 'df')
>>> pd.read_hdf('test_fixed.h5', 'df', where='index>5')
TypeError: cannot pass a where specification when reading a fixed format.
this store must be selected in its entirety

4.1. IO tools (text, CSV, HDF5, ) 255

pandas: powerful Python data analysis toolkit, Release 0.25.3

Table format

HDFStore supports another PyTables format on disk, the table format. Conceptually a table is shaped very
much like a DataFrame, with rows and columns. A table may be appended to in the same or other sessions.
In addition, delete and query type operations are supported. This format is specified by format='table' or
format='t' to append or put or to_hdf.
This format can be set as an option as well pd.set_option('io.hdf.default_format','table') to
enable put/append/to_hdf to by default store in the table format.
In [381]: store = pd.HDFStore('store.h5')

In [382]: df1 = df[0:4]

In [383]: df2 = df[4:]

# append data (creates a table automatically)

In [384]: store.append('df', df1)

In [385]: store.append('df', df2)

In [386]: store
Out[386]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

# select the entire object

In [387]: store.select('df')
Out[387]:
A B C
2000-01-01 0.263806 1.913465 -0.274536
2000-01-02 0.283334 1.798001 -0.053258
2000-01-03 0.799684 -0.733715 1.205089
2000-01-04 0.131478 0.100995 -0.764260
2000-01-05 1.891112 -1.410251 0.752883
2000-01-06 -0.274852 -0.667027 -0.688782
2000-01-07 0.621607 -1.300199 0.050119
2000-01-08 -0.999591 -0.320658 -1.922640

# the type of stored data

In [388]: store.root.df._v_attrs.pandas_type
Out[388]: 'frame_table'

Note: You can also create a table by passing format='table' or format='t' to a put operation.

Hierarchical keys

Keys to a store can be specified as a string. These can be in a hierarchical path-name like format (e.g. foo/bar/
bah), which will generate a hierarchy of sub-stores (or Groups in PyTables parlance). Keys can be specified with
out the leading / and are always absolute (e.g. foo refers to /foo). Removal operations can remove everything in the
sub-store and below, so be careful.
In [389]: store.put('foo/bar/bah', df)

256 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [390]: store.append('food/orange', df)

In [391]: store.append('food/apple', df)

In [392]: store
Out[392]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

# a list of keys are returned

In [393]: store.keys()
Out[393]: ['/df', '/food/apple', '/food/orange', '/foo/bar/bah']

# remove all nodes under this level

In [394]: store.remove('food')

In [395]: store
Out[395]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
You can walk through the group hierarchy using the walk method which will yield a tuple for each group key along
with the relative keys of its contents.
New in version 0.24.0.

In [396]: for (path, subgroups, subkeys) in store.walk():

.....: for subgroup in subgroups:
.....: print('GROUP: {}/{}'.format(path, subgroup))
.....: for subkey in subkeys:
.....: key = '/'.join([path, subkey])
.....: print('KEY: {}'.format(key))
.....: print(store.get(key))
.....:
GROUP: /foo
KEY: /df
A B C
2000-01-01 0.263806 1.913465 -0.274536
2000-01-02 0.283334 1.798001 -0.053258
2000-01-03 0.799684 -0.733715 1.205089
2000-01-04 0.131478 0.100995 -0.764260
2000-01-05 1.891112 -1.410251 0.752883
2000-01-06 -0.274852 -0.667027 -0.688782
2000-01-07 0.621607 -1.300199 0.050119
2000-01-08 -0.999591 -0.320658 -1.922640
GROUP: /foo/bar
KEY: /foo/bar/bah
A B C
2000-01-01 0.263806 1.913465 -0.274536
2000-01-02 0.283334 1.798001 -0.053258
2000-01-03 0.799684 -0.733715 1.205089
2000-01-04 0.131478 0.100995 -0.764260
2000-01-05 1.891112 -1.410251 0.752883
2000-01-06 -0.274852 -0.667027 -0.688782
2000-01-07 0.621607 -1.300199 0.050119
2000-01-08 -0.999591 -0.320658 -1.922640

4.1. IO tools (text, CSV, HDF5, ) 257

pandas: powerful Python data analysis toolkit, Release 0.25.3

Warning: Hierarchical keys cannot be retrieved as dotted (attribute) access as described above for items stored
under the root node.
In [8]: store.foo.bar.bah
AttributeError: 'HDFStore' object has no attribute 'foo'

# you can directly access the actual PyTables node but using the root node
In [9]: store.root.foo.bar.bah
Out[9]:
/foo/bar/bah (Group) ''
children := ['block0_items' (Array), 'block0_values' (Array), 'axis0' (Array),
,→'axis1' (Array)]

Instead, use explicit string based keys:

In [397]: store['foo/bar/bah']
Out[397]:
A B C
2000-01-01 0.263806 1.913465 -0.274536
2000-01-02 0.283334 1.798001 -0.053258
2000-01-03 0.799684 -0.733715 1.205089
2000-01-04 0.131478 0.100995 -0.764260
2000-01-05 1.891112 -1.410251 0.752883
2000-01-06 -0.274852 -0.667027 -0.688782
2000-01-07 0.621607 -1.300199 0.050119
2000-01-08 -0.999591 -0.320658 -1.922640

Storing types

Storing mixed types in a table

Storing mixed-dtype data is supported. Strings are stored as a fixed-width using the maximum size of the appended
column. Subsequent attempts at appending longer strings will raise a ValueError.
Passing min_itemsize={`values`: size} as a parameter to append will set a larger minimum for the string
columns. Storing floats, strings, ints, bools, datetime64 are currently supported. For string
columns, passing nan_rep = 'nan' to append will change the default nan representation on disk (which con-
verts to/from np.nan), this defaults to nan.
In [398]: df_mixed = pd.DataFrame({'A': np.random.randn(8),
.....: 'B': np.random.randn(8),
.....: 'C': np.array(np.random.randn(8),
,→dtype='float32'),

.....: 'string': 'string',

.....: 'int': 1,
.....: 'bool': True,
.....: 'datetime64': pd.Timestamp('20010102')},
.....: index=list(range(8)))
.....:

In [399]: df_mixed.loc[df_mixed.index[3:5],
.....: ['A', 'B', 'string', 'datetime64']] = np.nan
.....:

In [400]: store.append('df_mixed', df_mixed, min_itemsize={'values': 50})

258 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [401]: df_mixed1 = store.select('df_mixed')

In [402]: df_mixed1
Out[402]:
A B C string int bool datetime64
0 0.894171 -1.452159 -0.105646 string 1 True 2001-01-02
1 -1.539066 1.018959 0.028593 string 1 True 2001-01-02
2 -0.114019 -0.087476 0.693070 string 1 True 2001-01-02
3 NaN NaN -0.646571 NaN 1 True NaT
4 NaN NaN -0.174558 NaN 1 True NaT
5 -2.110838 -1.234633 -1.257271 string 1 True 2001-01-02
6 -0.704558 0.463419 -0.917264 string 1 True 2001-01-02
7 -0.929182 0.841053 0.414183 string 1 True 2001-01-02

In [403]: df_mixed1.dtypes.value_counts()
Out[403]:
float64 2
int64 1
datetime64[ns] 1
object 1
bool 1
float32 1
dtype: int64

# we have provided a minimum string column size

In [404]: store.root.df_mixed.table
Out[404]:
/df_mixed/table (Table(8,)) ''
description := {
"index": Int64Col(shape=(), dflt=0, pos=0),
"values_block_0": Float64Col(shape=(2,), dflt=0.0, pos=1),
"values_block_1": Float32Col(shape=(1,), dflt=0.0, pos=2),
"values_block_2": Int64Col(shape=(1,), dflt=0, pos=3),
"values_block_3": Int64Col(shape=(1,), dflt=0, pos=4),
"values_block_4": BoolCol(shape=(1,), dflt=False, pos=5),
"values_block_5": StringCol(itemsize=50, shape=(1,), dflt=b'', pos=6)}
byteorder := 'little'
chunkshape := (689,)
autoindex := True
colindexes := {
"index": Index(6, medium, shuffle, zlib(1)).is_csi=False}

Storing MultiIndex DataFrames

Storing MultiIndex DataFrames as tables is very similar to storing/selecting from homogeneous index
DataFrames.
In [405]: index = pd.MultiIndex(levels=[['foo', 'bar', 'baz', 'qux'],
.....: ['one', 'two', 'three']],
.....: codes=[[0, 0, 0, 1, 1, 2, 2, 3, 3, 3],
.....: [0, 1, 2, 0, 1, 1, 2, 0, 1, 2]],
.....: names=['foo', 'bar'])

4.1. IO tools (text, CSV, HDF5, ) 259

pandas: powerful Python data analysis toolkit, Release 0.25.3

.....:

In [406]: df_mi = pd.DataFrame(np.random.randn(10, 3), index=index,

.....: columns=['A', 'B', 'C'])
.....:

In [407]: df_mi
Out[407]:
A B C
foo bar
foo one 0.072648 -0.851494 0.140402
two -0.568937 0.439050 2.531582
three 0.539277 -1.398668 0.740635
bar one -1.892064 -0.830925 1.775692
two 2.183350 -1.565258 1.016985
baz two -0.476773 -0.566776 -0.665680
three 0.935387 0.551846 0.786999
qux one 0.481318 0.001118 -0.005084
two 0.238900 -1.888197 -0.943224
three -0.761786 -0.706338 -0.594234

In [408]: store.append('df_mi', df_mi)

In [409]: store.select('df_mi')
Out[409]:
A B C
foo bar
foo one 0.072648 -0.851494 0.140402
two -0.568937 0.439050 2.531582
three 0.539277 -1.398668 0.740635
bar one -1.892064 -0.830925 1.775692
two 2.183350 -1.565258 1.016985
baz two -0.476773 -0.566776 -0.665680
three 0.935387 0.551846 0.786999
qux one 0.481318 0.001118 -0.005084
two 0.238900 -1.888197 -0.943224
three -0.761786 -0.706338 -0.594234

# the levels are automatically included as data columns

In [410]: store.select('df_mi', 'foo=bar')
Out[410]:
A B C
foo bar
bar one -1.892064 -0.830925 1.775692
two 2.183350 -1.565258 1.016985

Querying

Querying a table

select and delete operations have an optional criterion that can be specified to select/delete only a subset of the
data. This allows one to have a very large on-disk table and retrieve only a portion of the data.

260 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

A query is specified using the Term class under the hood, as a boolean expression.
• index and columns are supported indexers of a DataFrames.
• if data_columns are specified, these can be used as additional indexers.
Valid comparison operators are:
=, ==, !=, >, >=, <, <=
Valid boolean expressions are combined with:
• | : or
• & : and
• ( and ) : for grouping
These rules are similar to how boolean expressions are used in pandas for indexing.

Note:
• = will be automatically expanded to the comparison operator ==
• ~ is the not operator, but can only be used in very limited circumstances
• If a list/tuple of expressions is passed they will be combined via &

The following are valid expressions:

• 'index >= date'
• "columns = ['A', 'D']"
• "columns in ['A', 'D']"
• 'columns = A'
• 'columns == A'
• "~(columns = ['A', 'B'])"
• 'index > df.index[3] & string = "bar"'
• '(index > df.index[3] & index <= df.index[6]) | string = "bar"'
• "ts >= Timestamp('2012-02-01')"
• "major_axis>=20130101"
The indexers are on the left-hand side of the sub-expression:
columns, major_axis, ts
The right-hand side of the sub-expression (after a comparison operator) can be:
• functions that will be evaluated, e.g. Timestamp('2012-02-01')
• strings, e.g. "bar"
• date-like, e.g. 20130101, or "20130101"
• lists, e.g. "['A', 'B']"
• variables that are defined in the local names space, e.g. date

4.1. IO tools (text, CSV, HDF5, ) 261

pandas: powerful Python data analysis toolkit, Release 0.25.3

Note: Passing a string to a query by interpolating it into the query expression is not recommended. Simply assign the
string of interest to a variable and use that variable in an expression. For example, do this

string = "HolyMoly'"
store.select('df', 'index == string')

instead of this

string = "HolyMoly'"
store.select('df', 'index == %s' % string)

The latter will not work and will raise a SyntaxError.Note that theres a single quote followed by a double quote
in the string variable.
If you must interpolate, use the '%r' format specifier

store.select('df', 'index == %r' % string)

which will quote string.

Here are some examples:

In [411]: dfq = pd.DataFrame(np.random.randn(10, 4), columns=list('ABCD'),

.....: index=pd.date_range('20130101', periods=10))
.....:

In [412]: store.append('dfq', dfq, format='table', data_columns=True)

Use boolean expressions, with in-line function evaluation.

In [413]: store.select('dfq', "index>pd.Timestamp('20130104') & columns=['A', 'B']")

Out[413]:
A B
2013-01-05 0.141437 -0.757467
2013-01-06 0.428370 0.281642
2013-01-07 -0.122746 0.909453
2013-01-08 0.319459 0.828792
2013-01-09 -0.446442 0.030712
2013-01-10 -0.627425 0.599256

Use and inline column reference

In [414]: store.select('dfq', where="A>0 or C>0")

Out[414]:
A B C D
2013-01-03 0.099259 -0.456319 0.079027 1.367963
2013-01-04 1.339489 1.086947 -0.918047 0.552759
2013-01-05 0.141437 -0.757467 -1.536944 -0.641387
2013-01-06 0.428370 0.281642 2.297828 -1.406053
2013-01-07 -0.122746 0.909453 1.434186 2.485565
2013-01-08 0.319459 0.828792 -0.221890 0.693093
2013-01-09 -0.446442 0.030712 1.330275 1.509187

The columns keyword can be supplied to select a list of columns to be returned, this is equivalent to passing a
'columns=list_of_columns_to_filter':

262 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [415]: store.select('df', "columns=['A', 'B']")

Out[415]:
A B
2000-01-01 0.263806 1.913465
2000-01-02 0.283334 1.798001
2000-01-03 0.799684 -0.733715
2000-01-04 0.131478 0.100995
2000-01-05 1.891112 -1.410251
2000-01-06 -0.274852 -0.667027
2000-01-07 0.621607 -1.300199
2000-01-08 -0.999591 -0.320658

start and stop parameters can be specified to limit the total search space. These are in terms of the total number
of rows in a table.

Note: select will raise a ValueError if the query expression has an unknown variable reference. Usually this
means that you are trying to select on a column that is not a data_column.
select will raise a SyntaxError if the query expression is not valid.

Using timedelta64[ns]

You can store and query using the timedelta64[ns] type. Terms can be specified in the format:
<float>(<unit>), where float may be signed (and fractional), and unit can be D,s,ms,us,ns for the timedelta.
Heres an example:

In [416]: from datetime import timedelta

In [417]: dftd = pd.DataFrame({'A': pd.Timestamp('20130101'),

.....: 'B': [pd.Timestamp('20130101') + timedelta(days=i,
.....: seconds=10)
.....: for i in range(10)]})
.....:

In [418]: dftd['C'] = dftd['A'] - dftd['B']

In [419]: dftd
Out[419]:
A B C
0 2013-01-01 2013-01-01 00:00:10 -1 days +23:59:50
1 2013-01-01 2013-01-02 00:00:10 -2 days +23:59:50
2 2013-01-01 2013-01-03 00:00:10 -3 days +23:59:50
3 2013-01-01 2013-01-04 00:00:10 -4 days +23:59:50
4 2013-01-01 2013-01-05 00:00:10 -5 days +23:59:50
5 2013-01-01 2013-01-06 00:00:10 -6 days +23:59:50
6 2013-01-01 2013-01-07 00:00:10 -7 days +23:59:50
7 2013-01-01 2013-01-08 00:00:10 -8 days +23:59:50
8 2013-01-01 2013-01-09 00:00:10 -9 days +23:59:50
9 2013-01-01 2013-01-10 00:00:10 -10 days +23:59:50

In [420]: store.append('dftd', dftd, data_columns=True)

In [421]: store.select('dftd', "C<'-3.5D'")

Out[421]:
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 263

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

A B C
4 2013-01-01 2013-01-05 00:00:10 -5 days +23:59:50
5 2013-01-01 2013-01-06 00:00:10 -6 days +23:59:50
6 2013-01-01 2013-01-07 00:00:10 -7 days +23:59:50
7 2013-01-01 2013-01-08 00:00:10 -8 days +23:59:50
8 2013-01-01 2013-01-09 00:00:10 -9 days +23:59:50
9 2013-01-01 2013-01-10 00:00:10 -10 days +23:59:50

Indexing

You can create/modify an index for a table with create_table_index after data is already in the table (after and
append/put operation). Creating a table index is highly encouraged. This will speed your queries a great deal
when you use a select with the indexed dimension as the where.

Note: Indexes are automagically created on the indexables and any data columns you specify. This behavior can be
turned off by passing index=False to append.

# we have automagically already created an index (in the first section)

In [422]: i = store.root.df.table.cols.index.index

In [423]: i.optlevel, i.kind

Out[423]: (6, 'medium')

# change an index by passing new parameters

In [424]: store.create_table_index('df', optlevel=9, kind='full')

In [425]: i = store.root.df.table.cols.index.index

In [426]: i.optlevel, i.kind

Out[426]: (9, 'full')

Oftentimes when appending large amounts of data to a store, it is useful to turn off index creation for each append,
then recreate at the end.

In [427]: df_1 = pd.DataFrame(np.random.randn(10, 2), columns=list('AB'))

In [428]: df_2 = pd.DataFrame(np.random.randn(10, 2), columns=list('AB'))

In [429]: st = pd.HDFStore('appends.h5', mode='w')

In [430]: st.append('df', df_1, data_columns=['B'], index=False)

In [431]: st.append('df', df_2, data_columns=['B'], index=False)

In [432]: st.get_storer('df').table
Out[432]:
/df/table (Table(20,)) ''
description := {
"index": Int64Col(shape=(), dflt=0, pos=0),
"values_block_0": Float64Col(shape=(1,), dflt=0.0, pos=1),
"B": Float64Col(shape=(), dflt=0.0, pos=2)}
byteorder := 'little'
chunkshape := (2730,)

264 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Then create the index when finished appending.

In [433]: st.create_table_index('df', columns=['B'], optlevel=9, kind='full')

In [434]: st.get_storer('df').table
Out[434]:
/df/table (Table(20,)) ''
description := {
"index": Int64Col(shape=(), dflt=0, pos=0),
"values_block_0": Float64Col(shape=(1,), dflt=0.0, pos=1),
"B": Float64Col(shape=(), dflt=0.0, pos=2)}
byteorder := 'little'
chunkshape := (2730,)
autoindex := True
colindexes := {
"B": Index(9, full, shuffle, zlib(1)).is_csi=True}

In [435]: st.close()

See here for how to create a completely-sorted-index (CSI) on an existing store.

Query via data columns

You can designate (and index) certain columns that you want to be able to perform queries (other than the indexable
columns, which you can always query). For instance say you want to perform this common operation, on-disk, and
return just the frame that matches this query. You can specify data_columns = True to force all columns to be
data_columns.
In [436]: df_dc = df.copy()

In [437]: df_dc['string'] = 'foo'

In [438]: df_dc.loc[df_dc.index[4:6], 'string'] = np.nan

In [439]: df_dc.loc[df_dc.index[7:9], 'string'] = 'bar'

In [440]: df_dc['string2'] = 'cool'

In [441]: df_dc.loc[df_dc.index[1:3], ['B', 'C']] = 1.0

In [442]: df_dc
Out[442]:
A B C string string2
2000-01-01 0.263806 1.913465 -0.274536 foo cool
2000-01-02 0.283334 1.000000 1.000000 foo cool
2000-01-03 0.799684 1.000000 1.000000 foo cool
2000-01-04 0.131478 0.100995 -0.764260 foo cool
2000-01-05 1.891112 -1.410251 0.752883 NaN cool
2000-01-06 -0.274852 -0.667027 -0.688782 NaN cool
2000-01-07 0.621607 -1.300199 0.050119 foo cool
2000-01-08 -0.999591 -0.320658 -1.922640 bar cool

# on-disk operations
In [443]: store.append('df_dc', df_dc, data_columns=['B', 'C', 'string',
,→'string2'])

4.1. IO tools (text, CSV, HDF5, ) 265

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [444]: store.select('df_dc', where='B > 0')

Out[444]:
A B C string string2
2000-01-01 0.263806 1.913465 -0.274536 foo cool
2000-01-02 0.283334 1.000000 1.000000 foo cool
2000-01-03 0.799684 1.000000 1.000000 foo cool
2000-01-04 0.131478 0.100995 -0.764260 foo cool

# getting creative
In [445]: store.select('df_dc', 'B > 0 & C > 0 & string == foo')
Out[445]:
A B C string string2
2000-01-02 0.283334 1.0 1.0 foo cool
2000-01-03 0.799684 1.0 1.0 foo cool

# this is in-memory version of this type of selection

In [446]: df_dc[(df_dc.B > 0) & (df_dc.C > 0) & (df_dc.string == 'foo')]
Out[446]:
A B C string string2
2000-01-02 0.283334 1.0 1.0 foo cool
2000-01-03 0.799684 1.0 1.0 foo cool

# we have automagically created this index and the B/C/string/string2

# columns are stored separately as ``PyTables`` columns
In [447]: store.root.df_dc.table
Out[447]:
/df_dc/table (Table(8,)) ''
description := {
"index": Int64Col(shape=(), dflt=0, pos=0),
"values_block_0": Float64Col(shape=(1,), dflt=0.0, pos=1),
"B": Float64Col(shape=(), dflt=0.0, pos=2),
"C": Float64Col(shape=(), dflt=0.0, pos=3),
"string": StringCol(itemsize=3, shape=(), dflt=b'', pos=4),
"string2": StringCol(itemsize=4, shape=(), dflt=b'', pos=5)}
byteorder := 'little'
chunkshape := (1680,)
autoindex := True
colindexes := {
"index": Index(6, medium, shuffle, zlib(1)).is_csi=False,
"B": Index(6, medium, shuffle, zlib(1)).is_csi=False,
"C": Index(6, medium, shuffle, zlib(1)).is_csi=False,
"string": Index(6, medium, shuffle, zlib(1)).is_csi=False,
"string2": Index(6, medium, shuffle, zlib(1)).is_csi=False}
There is some performance degradation by making lots of columns into data columns, so it is up to the user to designate
these. In addition, you cannot change data columns (nor indexables) after the first append/put operation (Of course
you can simply read in the data and create a new table!).

Iterator

You can pass iterator=True or chunksize=number_in_a_chunk to select and

select_as_multiple to return an iterator on the results. The default is 50,000 rows returned in a chunk.

266 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [448]: for df in store.select('df', chunksize=3):

.....: print(df)
.....:
A B C
2000-01-01 0.263806 1.913465 -0.274536
2000-01-02 0.283334 1.798001 -0.053258
2000-01-03 0.799684 -0.733715 1.205089
A B C
2000-01-04 0.131478 0.100995 -0.764260
2000-01-05 1.891112 -1.410251 0.752883
2000-01-06 -0.274852 -0.667027 -0.688782
A B C
2000-01-07 0.621607 -1.300199 0.050119
2000-01-08 -0.999591 -0.320658 -1.922640

Note: You can also use the iterator with read_hdf which will open, then automatically close the store when finished
iterating.

for df in pd.read_hdf('store.h5', 'df', chunksize=3):

print(df)

Note, that the chunksize keyword applies to the source rows. So if you are doing a query, then the chunksize will
subdivide the total rows in the table and the query applied, returning an iterator on potentially unequal sized chunks.
Here is a recipe for generating a query and using it to create equal sized return chunks.

In [449]: dfeq = pd.DataFrame({'number': np.arange(1, 11)})

In [450]: dfeq
Out[450]:
number
0 1
1 2
2 3
3 4
4 5
5 6
6 7
7 8
8 9
9 10

In [451]: store.append('dfeq', dfeq, data_columns=['number'])

In [452]: def chunks(l, n):

.....: return [l[i:i + n] for i in range(0, len(l), n)]
.....:

In [453]: evens = [2, 4, 6, 8, 10]

In [454]: coordinates = store.select_as_coordinates('dfeq', 'number=evens')

In [455]: for c in chunks(coordinates, 2):

.....: print(store.select('dfeq', where=c))
.....:
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 267

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

number
1 2
3 4
number
5 6
7 8
number
9 10

Advanced queries

Select a single column

To retrieve a single indexable or data column, use the method select_column. This will, for example, enable you
to get the index very quickly. These return a Series of the result, indexed by the row number. These do not currently
accept the where selector.
In [456]: store.select_column('df_dc', 'index')
Out[456]:
0 2000-01-01
1 2000-01-02
2 2000-01-03
3 2000-01-04
4 2000-01-05
5 2000-01-06
6 2000-01-07
7 2000-01-08
Name: index, dtype: datetime64[ns]

In [457]: store.select_column('df_dc', 'string')

Out[457]:
0 foo
1 foo
2 foo
3 foo
4 NaN
5 NaN
6 foo
7 bar
Name: string, dtype: object

Selecting coordinates

Sometimes you want to get the coordinates (a.k.a the index locations) of your query. This returns an Int64Index
of the resulting locations. These coordinates can also be passed to subsequent where operations.
In [458]: df_coord = pd.DataFrame(np.random.randn(1000, 2),
.....: index=pd.date_range('20000101',
,→periods=1000))

.....:

In [459]: store.append('df_coord', df_coord)

268 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [460]: c = store.select_as_coordinates('df_coord', 'index > 20020101')

In [461]: c
Out[461]:
Int64Index([732, 733, 734, 735, 736, 737, 738, 739, 740, 741,
...
990, 991, 992, 993, 994, 995, 996, 997, 998, 999],
dtype='int64', length=268)

In [462]: store.select('df_coord', where=c)

Out[462]:
0 1
2002-01-02 -0.479047 -1.957543
2002-01-03 1.254799 1.315447
2002-01-04 -0.233731 -1.630779
2002-01-05 -1.098851 -1.299909
2002-01-06 0.667606 0.143470
... ... ...
2002-09-22 -1.066208 -0.894186
2002-09-23 2.079421 -0.266075
2002-09-24 1.496228 -0.668430
2002-09-25 0.402749 -0.187683
2002-09-26 0.350581 1.152795

[268 rows x 2 columns]

Selecting using a where mask

Sometime your query can involve creating a list of rows to select. Usually this mask would be a resulting index
from an indexing operation. This example selects the months of a datetimeindex which are 5.

In [463]: df_mask = pd.DataFrame(np.random.randn(1000, 2),

.....: index=pd.date_range('20000101', periods=1000))
.....:

In [464]: store.append('df_mask', df_mask)

In [465]: c = store.select_column('df_mask', 'index')

In [466]: where = c[pd.DatetimeIndex(c).month == 5].index

In [467]: store.select('df_mask', where=where)

Out[467]:
0 1
2000-05-01 0.673172 0.388669
2000-05-02 0.761613 -0.368644
2000-05-03 0.435121 -0.062180
2000-05-04 1.677448 -0.019684
2000-05-05 0.562315 1.464591
... ... ...
2002-05-27 0.437853 1.315529
2002-05-28 0.931011 0.532589
2002-05-29 0.349402 -0.138098
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 269

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2002-05-30 0.752838 0.254404
2002-05-31 0.013580 1.078755

[93 rows x 2 columns]

Storer object

If you want to inspect the stored object, retrieve via get_storer. You could use this programmatically to say get
the number of rows in an object.

In [468]: store.get_storer('df_dc').nrows
Out[468]: 8

Multiple table queries

The methods append_to_multiple and select_as_multiple can perform appending/selecting from mul-
tiple tables at once. The idea is to have one table (call it the selector table) that you index most/all of the columns, and
perform your queries. The other table(s) are data tables with an index matching the selector tables index. You can then
perform a very fast query on the selector table, yet get lots of data back. This method is similar to having a very wide
table, but enables more efficient queries.
The append_to_multiple method splits a given single DataFrame into multiple tables according to d, a dictio-
nary that maps the table names to a list of columns you want in that table. If None is used in place of a list, that
table will have the remaining unspecified columns of the given DataFrame. The argument selector defines which
table is the selector table (which you can make queries from). The argument dropna will drop rows from the input
DataFrame to ensure tables are synchronized. This means that if a row for one of the tables being written to is
entirely np.NaN, that row will be dropped from all tables.
If dropna is False, THE USER IS RESPONSIBLE FOR SYNCHRONIZING THE TABLES. Remember that
entirely np.Nan rows are not written to the HDFStore, so if you choose to call dropna=False, some tables may
have more rows than others, and therefore select_as_multiple may not work or it may return unexpected
results.
In [469]: df_mt = pd.DataFrame(np.random.randn(8, 6),
.....: index=pd.date_range('1/1/2000', periods=8),
.....: columns=['A', 'B', 'C', 'D', 'E', 'F'])
.....:

In [470]: df_mt['foo'] = 'bar'

In [471]: df_mt.loc[df_mt.index[1], ('A', 'B')] = np.nan

# you can also create the tables individually

In [472]: store.append_to_multiple({'df1_mt': ['A', 'B'], 'df2_mt': None},
.....: df_mt, selector='df1_mt')
.....:

In [473]: store
Out[473]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

270 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

# individual tables were created

In [474]: store.select('df1_mt')
Out[474]:
A B
2000-01-01 1.222083 -1.048640
2000-01-02 NaN NaN
2000-01-03 -0.403428 -0.359149
2000-01-04 -0.445969 0.331799
2000-01-05 -0.358375 -0.713913
2000-01-06 0.767579 -1.518142
2000-01-07 -0.125142 0.285073
2000-01-08 1.014284 -0.305314

In [475]: store.select('df2_mt')
Out[475]:
C D E F foo
2000-01-01 0.519195 -1.005739 -1.126444 -0.331994 bar
2000-01-02 -0.220187 -2.576683 -1.531751 1.545878 bar
2000-01-03 0.164228 -0.920410 0.184463 -0.179357 bar
2000-01-04 -1.493748 0.016206 -1.594605 -0.077495 bar
2000-01-05 -0.818602 -0.271994 1.188345 0.345087 bar
2000-01-06 -1.514329 -1.344535 -1.243543 0.231915 bar
2000-01-07 -0.702845 2.420812 0.309805 -1.101996 bar
2000-01-08 0.441337 -0.846243 -0.984939 -1.159608 bar

# as a multiple
In [476]: store.select_as_multiple(['df1_mt', 'df2_mt'], where=['A>0', 'B>0'],
.....: selector='df1_mt')
.....:
Out[476]:
Empty DataFrame
Columns: [A, B, C, D, E, F, foo]
Index: []

Delete from a table

You can delete from a table selectively by specifying a where. In deleting rows, it is important to understand the
PyTables deletes rows by erasing the rows, then moving the following data. Thus deleting can potentially be a very
expensive operation depending on the orientation of your data. To get optimal performance, its worthwhile to have the
dimension you are deleting be the first of the indexables.
Data is ordered (on the disk) in terms of the indexables. Heres a simple use case. You store panel-type data, with
dates in the major_axis and ids in the minor_axis. The data is then interleaved like this:
• date_1
– id_1
– id_2
– .
– id_n
• date_2
– id_1

4.1. IO tools (text, CSV, HDF5, ) 271

pandas: powerful Python data analysis toolkit, Release 0.25.3

– .
– id_n
It should be clear that a delete operation on the major_axis will be fairly quick, as one chunk is removed, then the
following data moved. On the other hand a delete operation on the minor_axis will be very expensive. In this case
it would almost certainly be faster to rewrite the table using a where that selects all but the missing data.

Warning: Please note that HDF5 DOES NOT RECLAIM SPACE in the h5 files automatically. Thus, repeatedly
deleting (or removing nodes) and adding again, WILL TEND TO INCREASE THE FILE SIZE.
To repack and clean the file, use ptrepack.

Notes & caveats

Compression

PyTables allows the stored data to be compressed. This applies to all kinds of stores, not just tables. Two parameters
are used to control compression: complevel and complib.
complevel specifies if and how hard data is to be compressed. complevel=0 and complevel=None dis-
ables compression and 0<complevel<10 enables compression.
complib specifies which compression library to use. If nothing is specified the default library zlib is used. A
compression library usually optimizes for either good compression rates or speed and the results will depend
on the type of data. Which type of compression to choose depends on your specific needs and data. The list of
supported compression libraries:
• zlib: The default compression library. A classic in terms of compression, achieves good com-
pression rates but is somewhat slow.
• lzo: Fast compression and decompression.
• bzip2: Good compression rates.
• blosc: Fast compression and decompression.
New in version 0.20.2: Support for alternative blosc compressors:
• blosc:blosclz This is the default compressor for blosc
• blosc:lz4: A compact, very popular and fast compressor.
• blosc:lz4hc: A tweaked version of LZ4, produces better compression ratios at the expense of
speed.
• blosc:snappy: A popular compressor used in many places.
• blosc:zlib: A classic; somewhat slower than the previous ones, but achieving better compression
ratios.
• blosc:zstd: An extremely well balanced codec; it provides the best compression ratios among
the others above, and at reasonably fast speed.
If complib is defined as something other than the listed libraries a ValueError exception is
issued.

272 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Note: If the library specified with the complib option is missing on your platform, compression defaults to zlib
without further ado.

Enable compression for all objects within the file:

store_compressed = pd.HDFStore('store_compressed.h5', complevel=9,

complib='blosc:blosclz')

Or on-the-fly compression (this only applies to tables) in stores where compression is not enabled:

store.append('df', df, complib='zlib', complevel=5)

ptrepack

PyTables offers better write performance when tables are compressed after they are written, as opposed to turning on
compression at the very beginning. You can use the supplied PyTables utility ptrepack. In addition, ptrepack
can change compression levels after the fact.

ptrepack --chunkshape=auto --propindexes --complevel=9 --complib=blosc in.h5 out.h5

Furthermore ptrepack in.h5 out.h5 will repack the file to allow you to reuse previously deleted space. Alter-
natively, one can simply remove the file and write again, or use the copy method.

Caveats

Warning: HDFStore is not-threadsafe for writing. The underlying PyTables only supports concurrent
reads (via threading or processes). If you need reading and writing at the same time, you need to serialize these
operations in a single thread in a single process. You will corrupt your data otherwise. See the (GH2397) for more
information.

• If you use locks to manage write access between multiple processes, you may want to use fsync() before
releasing write locks. For convenience you can use store.flush(fsync=True) to do this for you.
• Once a table is created columns (DataFrame) are fixed; only exactly the same columns can be appended
• Be aware that timezones (e.g., pytz.timezone('US/Eastern')) are not necessarily equal across time-
zone versions. So if data is localized to a specific timezone in the HDFStore using one version of a timezone
library and that data is updated with another version, the data will be converted to UTC since these timezones
are not considered equal. Either use the same version of timezone library or use tz_convert with the updated
timezone definition.

Warning: PyTables will show a NaturalNameWarning if a column name cannot be used as an attribute
selector. Natural identifiers contain only letters, numbers, and underscores, and may not begin with a number.
Other identifiers cannot be used in a where clause and are generally a bad idea.

4.1. IO tools (text, CSV, HDF5, ) 273

pandas: powerful Python data analysis toolkit, Release 0.25.3

DataTypes

HDFStore will map an object dtype to the PyTables underlying dtype. This means the following types are known
to work:

Type Represents missing values

floating : float64, float32, float16 np.nan
integer : int64, int32, int8, uint64,uint32, uint8
boolean
datetime64[ns] NaT
timedelta64[ns] NaT
categorical : see the section below
object : strings np.nan

unicode columns are not supported, and WILL FAIL.

Categorical data

You can write data that contains category dtypes to a HDFStore. Queries work the same as if it was an object
array. However, the category dtyped data is stored in a more efficient manner.
In [477]: dfcat = pd.DataFrame({'A': pd.Series(list('aabbcdba')).
,→astype('category'),

.....: 'B': np.random.randn(8)})

.....:

In [478]: dfcat
Out[478]:
A B
0 a -0.182915
1 a -1.740077
2 b -0.227312
3 b -0.351706
4 c -0.484542
5 d 1.150367
6 b -1.223195
7 a 1.022514

In [479]: dfcat.dtypes
Out[479]:
A category
B float64
dtype: object

In [480]: cstore = pd.HDFStore('cats.h5', mode='w')

In [481]: cstore.append('dfcat', dfcat, format='table', data_columns=['A'])

In [482]: result = cstore.select('dfcat', where="A in ['b', 'c']")

In [483]: result
Out[483]:
A B

274 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

2 b -0.227312
3 b -0.351706
4 c -0.484542
6 b -1.223195

In [484]: result.dtypes
Out[484]:
A category
B float64
dtype: object

String columns

min_itemsize
The underlying implementation of HDFStore uses a fixed column width (itemsize) for string columns. A string
column itemsize is calculated as the maximum of the length of data (for that column) that is passed to the HDFStore,
in the first append. Subsequent appends, may introduce a string for a column larger than the column can hold, an
Exception will be raised (otherwise you could have a silent truncation of these columns, leading to loss of information).
In the future we may relax this and allow a user-specified truncation to occur.
Pass min_itemsize on the first table creation to a-priori specify the minimum length of a particular string column.
min_itemsize can be an integer, or a dict mapping a column name to an integer. You can pass values as a key
to allow all indexables or data_columns to have this min_itemsize.
Passing a min_itemsize dict will cause all passed columns to be created as data_columns automatically.

Note: If you are not passing any data_columns, then the min_itemsize will be the maximum of the length of
any string passed

In [485]: dfs = pd.DataFrame({'A': 'foo', 'B': 'bar'}, index=list(range(5)))

In [486]: dfs
Out[486]:
A B
0 foo bar
1 foo bar
2 foo bar
3 foo bar
4 foo bar

# A and B have a size of 30

In [487]: store.append('dfs', dfs, min_itemsize=30)

In [488]: store.get_storer('dfs').table
Out[488]:
/dfs/table (Table(5,)) ''
description := {
"index": Int64Col(shape=(), dflt=0, pos=0),
"values_block_0": StringCol(itemsize=30, shape=(2,), dflt=b'', pos=1)}
byteorder := 'little'
chunkshape := (963,)
autoindex := True
colindexes := {
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 275

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

"index": Index(6, medium, shuffle, zlib(1)).is_csi=False}

# A is created as a data_column with a size of 30

# B is size is calculated
In [489]: store.append('dfs2', dfs, min_itemsize={'A': 30})

In [490]: store.get_storer('dfs2').table
Out[490]:
/dfs2/table (Table(5,)) ''
description := {
"index": Int64Col(shape=(), dflt=0, pos=0),
"values_block_0": StringCol(itemsize=3, shape=(1,), dflt=b'', pos=1),
"A": StringCol(itemsize=30, shape=(), dflt=b'', pos=2)}
byteorder := 'little'
chunkshape := (1598,)
autoindex := True
colindexes := {
"index": Index(6, medium, shuffle, zlib(1)).is_csi=False,
"A": Index(6, medium, shuffle, zlib(1)).is_csi=False}

nan_rep
String columns will serialize a np.nan (a missing value) with the nan_rep string representation. This defaults to
the string value nan. You could inadvertently turn an actual nan value into a missing value.

In [491]: dfss = pd.DataFrame({'A': ['foo', 'bar', 'nan']})

In [492]: dfss
Out[492]:
A
0 foo
1 bar
2 nan

In [493]: store.append('dfss', dfss)

In [494]: store.select('dfss')
Out[494]:
A
0 foo
1 bar
2 NaN

# here you need to specify a different nan rep

In [495]: store.append('dfss2', dfss, nan_rep='_nan_')

In [496]: store.select('dfss2')
Out[496]:
A
0 foo
1 bar
2 nan

276 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

External compatibility

HDFStore writes table format objects in specific formats suitable for producing loss-less round trips to pandas
objects. For external compatibility, HDFStore can read native PyTables format tables.
It is possible to write an HDFStore object that can easily be imported into R using the rhdf5 library (Package
website). Create a table format store like this:
In [497]: df_for_r = pd.DataFrame({"first": np.random.rand(100),
.....: "second": np.random.rand(100),
.....: "class": np.random.randint(0, 2, (100, ))},
.....: index=range(100))
.....:

In [498]: df_for_r.head()
Out[498]:
first second class
0 0.253517 0.473526 0
1 0.232906 0.331008 0
2 0.069221 0.532945 1
3 0.290835 0.069538 1
4 0.912722 0.346792 0

In [499]: store_export = pd.HDFStore('export.h5')

In [500]: store_export.append('df_for_r', df_for_r, data_columns=df_dc.columns)

In [501]: store_export
Out[501]:
<class 'pandas.io.pytables.HDFStore'>
File path: export.h5

In R this file can be read into a data.frame object using the rhdf5 library. The following example function reads
the corresponding column names and data values from the values and assembles them into a data.frame:
# Load values and column names for all datasets from corresponding nodes and
# insert them into one data.frame object.

library(rhdf5)

loadhdf5data <- function(h5File) {

listing <- h5ls(h5File)

# Find all data nodes, values are stored in *_values and corresponding column
# titles in *_items
data_nodes <- grep("_values", listing$name)
name_nodes <- grep("_items", listing$name)
data_paths = paste(listing$group[data_nodes], listing$name[data_nodes], sep = "/")
name_paths = paste(listing$group[name_nodes], listing$name[name_nodes], sep = "/")
columns = list()
for (idx in seq(data_paths)) {
# NOTE: matrices returned by h5read have to be transposed to obtain
# required Fortran order!
data <- data.frame(t(h5read(h5File, data_paths[idx])))
names <- t(h5read(h5File, name_paths[idx]))
entry <- data.frame(data)
colnames(entry) <- names
columns <- append(columns, entry)
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 277

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

}

data <- data.frame(columns)

return(data)
}

Now you can import the DataFrame into R:

> data = loadhdf5data("transfer.hdf5")

> head(data)
first second class
1 0.4170220047 0.3266449 0
2 0.7203244934 0.5270581 0
3 0.0001143748 0.8859421 1
4 0.3023325726 0.3572698 1
5 0.1467558908 0.9085352 1
6 0.0923385948 0.6233601 1

Note: The R function lists the entire HDF5 files contents and assembles the data.frame object from all matching
nodes, so use this only as a starting point if you have stored multiple DataFrame objects to a single HDF5 file.

Performance

• tables format come with a writing performance penalty as compared to fixed stores. The benefit is the
ability to append/delete and query (potentially very large amounts of data). Write times are generally longer as
compared with regular stores. Query times can be quite fast, especially on an indexed axis.
• You can pass chunksize=<int> to append, specifying the write chunksize (default is 50000). This will
significantly lower your memory usage on writing.
• You can pass expectedrows=<int> to the first append, to set the TOTAL number of expected rows that
PyTables will expected. This will optimize read/write performance.
• Duplicate rows can be written to tables, but are filtered out in selection (with the last items being selected; thus
a table is unique on major, minor pairs)
• A PerformanceWarning will be raised if you are attempting to store types that will be pickled by PyTables
(rather than stored as endemic types). See Here for more information and some solutions.

4.1.10 Feather

New in version 0.20.0.

Feather provides binary columnar serialization for data frames. It is designed to make reading and writing data frames
efficient, and to make sharing data across data analysis languages easy.
Feather is designed to faithfully serialize and de-serialize DataFrames, supporting all of the pandas dtypes, including
extension dtypes such as categorical and datetime with tz.
Several caveats.
• This is a newer library, and the format, though stable, is not guaranteed to be backward compatible to the earlier
versions.

278 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

• The format will NOT write an Index, or MultiIndex for the DataFrame and will raise an error if a non-
default one is provided. You can .reset_index() to store the index or .reset_index(drop=True)
to ignore it.
• Duplicate column names and non-string columns names are not supported
• Non supported types include Period and actual Python object types. These will raise a helpful error message
on an attempt at serialization.
See the Full Documentation.
In [502]: df = pd.DataFrame({'a': list('abc'),
.....: 'b': list(range(1, 4)),
.....: 'c': np.arange(3, 6).astype('u1'),
.....: 'd': np.arange(4.0, 7.0, dtype='float64'),
.....: 'e': [True, False, True],
.....: 'f': pd.Categorical(list('abc')),
.....: 'g': pd.date_range('20130101', periods=3),
.....: 'h': pd.date_range('20130101', periods=3, tz='US/
,→Eastern'),

.....: 'i': pd.date_range('20130101', periods=3,

,→freq='ns')})

.....:

In [503]: df
Out[503]:
a b c d e f g h
,→ i
0 a 1 3 4.0 True a 2013-01-01 2013-01-01 00:00:00-05:00 2013-01-01
,→00:00:00.000000000
1 b 2 4 5.0 False b 2013-01-02 2013-01-02 00:00:00-05:00 2013-01-01
,→00:00:00.000000001

2 c 3 5 6.0 True c 2013-01-03 2013-01-03 00:00:00-05:00 2013-01-01

,→00:00:00.000000002

In [504]: df.dtypes
Out[504]:
a object
b int64
c uint8
d float64
e bool
f category
g datetime64[ns]
h datetime64[ns, US/Eastern]
i datetime64[ns]
dtype: object
Write to a feather file.

In [505]: df.to_feather('example.feather')
---------------------------------------------------------------------------
ImportError Traceback (most recent call last)
<ipython-input-505-b832ec9cf6be> in <module>
----> 1 df.to_feather('example.feather')

(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 279

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

~/sandbox/pandas-release/pandas/pandas/core/frame.py in to_feather(self, fname)
2150 from pandas.io.feather_format import to_feather
2151
-> 2152 to_feather(self, fname)
2153
2154 def to_parquet(

~/sandbox/pandas-release/pandas/pandas/io/feather_format.py in to_feather(df, path)

21
22 """
---> 23 import_optional_dependency("pyarrow")
24 from pyarrow import feather
25

~/sandbox/pandas-release/pandas/pandas/compat/_optional.py in import_optional_
,→dependency(name, extra, raise_on_missing, on_version)

91 except ImportError:
92 if raise_on_missing:
---> 93 raise ImportError(message.format(name=name, extra=extra)) from
,→None

94 else:
95 return None

ImportError: Missing optional dependency 'pyarrow'. Use pip or conda to install

,→pyarrow.

Read from a feather file.

In [506]: result = pd.read_feather('example.feather')
---------------------------------------------------------------------------
ImportError Traceback (most recent call last)
<ipython-input-506-9efa72931ac9> in <module>
----> 1 result = pd.read_feather('example.feather')

~/sandbox/pandas-release/pandas/pandas/util/_decorators.py in wrapper(*args,
,→**kwargs)

206 else:
207 kwargs[new_arg_name] = new_arg_value
--> 208 return func(*args, **kwargs)
209
210 return wrapper

~/sandbox/pandas-release/pandas/pandas/io/feather_format.py in read_
,→feather(path, columns, use_threads)

106 type of object stored in file

107 """
--> 108 pyarrow = import_optional_dependency("pyarrow")
109 from pyarrow import feather
110

~/sandbox/pandas-release/pandas/pandas/compat/_optional.py in import_optional_
,→dependency(name, extra, raise_on_missing, on_version)

91 except ImportError:
92 if raise_on_missing:
---> 93 raise ImportError(message.format(name=name, extra=extra))

280 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

,→from None
94 else:
95 return None

ImportError: Missing optional dependency 'pyarrow'. Use pip or conda to

,→install pyarrow.

In [507]: result
Out[507]:
A B
2 b -0.227312
3 b -0.351706
4 c -0.484542
6 b -1.223195

# we preserve dtypes
In [508]: result.dtypes
Out[508]:
A category
B float64
dtype: object

4.1.11 Parquet

New in version 0.21.0.

Apache Parquet provides a partitioned binary columnar serialization for data frames. It is designed to make reading and
writing data frames efficient, and to make sharing data across data analysis languages easy. Parquet can use a variety
of compression techniques to shrink the file size as much as possible while still maintaining good read performance.
Parquet is designed to faithfully serialize and de-serialize DataFrame s, supporting all of the pandas dtypes, includ-
ing extension dtypes such as datetime with tz.
Several caveats.
• Duplicate column names and non-string columns names are not supported.
• The pyarrow engine always writes the index to the output, but fastparquet only writes non-default in-
dexes. This extra column can cause problems for non-Pandas consumers that are not expecting it. You can force
including or omitting indexes with the index argument, regardless of the underlying engine.
• Index level names, if specified, must be strings.
• Categorical dtypes can be serialized to parquet, but will de-serialize as object dtype.
• Non supported types include Period and actual Python object types. These will raise a helpful error message
on an attempt at serialization.
You can specify an engine to direct the serialization. This can be one of pyarrow, or fastparquet, or auto.
If the engine is NOT specified, then the pd.options.io.parquet.engine option is checked; if this is also
auto, then pyarrow is tried, and falling back to fastparquet.
See the documentation for pyarrow and fastparquet.

Note: These engines are very similar and should read/write nearly identical parquet format files. Currently pyarrow
does not support timedelta data, fastparquet>=0.1.4 supports timezone aware datetimes. These libraries differ
by having different underlying dependencies (fastparquet by using numba, while pyarrow uses a c-library).

4.1. IO tools (text, CSV, HDF5, ) 281

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [509]: df = pd.DataFrame({'a': list('abc'),

.....: 'b': list(range(1, 4)),
.....: 'c': np.arange(3, 6).astype('u1'),
.....: 'd': np.arange(4.0, 7.0, dtype='float64'),
.....: 'e': [True, False, True],
.....: 'f': pd.date_range('20130101', periods=3),
.....: 'g': pd.date_range('20130101', periods=3, tz='US/
,→Eastern')})

.....:

In [510]: df
Out[510]:
a b c d e f g
0 a 1 3 4.0 True 2013-01-01 2013-01-01 00:00:00-05:00
1 b 2 4 5.0 False 2013-01-02 2013-01-02 00:00:00-05:00
2 c 3 5 6.0 True 2013-01-03 2013-01-03 00:00:00-05:00

In [511]: df.dtypes
Out[511]:
a object
b int64
c uint8
d float64
e bool
f datetime64[ns]
g datetime64[ns, US/Eastern]
dtype: object
Write to a parquet file.
In [512]: df.to_parquet('example_pa.parquet', engine='pyarrow')
---------------------------------------------------------------------------
ImportError Traceback (most recent call last)
<ipython-input-512-96e1a42ae5ed> in <module>
----> 1 df.to_parquet('example_pa.parquet', engine='pyarrow')

~/sandbox/pandas-release/pandas/pandas/core/frame.py in to_parquet(self,
,→fname, engine, compression, index, partition_cols, **kwargs)

2235 index=index,
2236 partition_cols=partition_cols,
-> 2237 **kwargs
2238 )
2239

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in to_parquet(df, path,

,→engine, compression, index, partition_cols, **kwargs)

245 Additional keyword arguments passed to the engine

246 """
--> 247 impl = get_engine(engine)
248 return impl.write(
249 df,

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in get_engine(engine)
40
41 if engine == "pyarrow":

282 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

---> 42 return PyArrowImpl()

43 elif engine == "fastparquet":
44 return FastParquetImpl()

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in __init__(self)
76 def __init__(self):
77 pyarrow = import_optional_dependency(
---> 78 "pyarrow", extra="pyarrow is required for parquet support.
,→"

79 )
80 import pyarrow.parquet

~/sandbox/pandas-release/pandas/pandas/compat/_optional.py in import_optional_
,→dependency(name, extra, raise_on_missing, on_version)

91 except ImportError:
92 if raise_on_missing:
---> 93 raise ImportError(message.format(name=name, extra=extra))
,→from None

94 else:
95 return None

ImportError: Missing optional dependency 'pyarrow'. pyarrow is required for

,→parquet support. Use pip or conda to install pyarrow.

In [513]: df.to_parquet('example_fp.parquet', engine='fastparquet')

---------------------------------------------------------------------------
ImportError Traceback (most recent call last)
<ipython-input-513-8ba55b61d48c> in <module>
----> 1 df.to_parquet('example_fp.parquet', engine='fastparquet')

~/sandbox/pandas-release/pandas/pandas/core/frame.py in to_parquet(self,
,→fname, engine, compression, index, partition_cols, **kwargs)

2235 index=index,
2236 partition_cols=partition_cols,
-> 2237 **kwargs
2238 )
2239

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in to_parquet(df, path,

,→engine, compression, index, partition_cols, **kwargs)

245 Additional keyword arguments passed to the engine

246 """
--> 247 impl = get_engine(engine)
248 return impl.write(
249 df,

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in get_engine(engine)
42 return PyArrowImpl()
43 elif engine == "fastparquet":
---> 44 return FastParquetImpl()
45
46

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in __init__(self)

4.1. IO tools (text, CSV, HDF5, ) 283

pandas: powerful Python data analysis toolkit, Release 0.25.3

139 # we need to import on first use

140 fastparquet = import_optional_dependency(
--> 141 "fastparquet", extra="fastparquet is required for parquet
,→support."

142 )
143 self.api = fastparquet

~/sandbox/pandas-release/pandas/pandas/compat/_optional.py in import_optional_
,→dependency(name, extra, raise_on_missing, on_version)

91 except ImportError:
92 if raise_on_missing:
---> 93 raise ImportError(message.format(name=name, extra=extra))
,→from None

94 else:
95 return None

ImportError: Missing optional dependency 'fastparquet'. fastparquet is

,→required for parquet support. Use pip or conda to install fastparquet.

Read from a parquet file.

In [514]: result = pd.read_parquet('example_fp.parquet', engine='fastparquet')
---------------------------------------------------------------------------
ImportError Traceback (most recent call last)
<ipython-input-514-d4bce7d5df53> in <module>
----> 1 result = pd.read_parquet('example_fp.parquet', engine='fastparquet')

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in read_parquet(path,
,→engine, columns, **kwargs)

293 """
294
--> 295 impl = get_engine(engine)
296 return impl.read(path, columns=columns, **kwargs)

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in get_engine(engine)
42 return PyArrowImpl()
43 elif engine == "fastparquet":
---> 44 return FastParquetImpl()
45
46

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in __init__(self)
139 # we need to import on first use
140 fastparquet = import_optional_dependency(
--> 141 "fastparquet", extra="fastparquet is required for parquet
,→support."

142 )
143 self.api = fastparquet

~/sandbox/pandas-release/pandas/pandas/compat/_optional.py in import_optional_
,→dependency(name, extra, raise_on_missing, on_version)

91 except ImportError:
92 if raise_on_missing:
---> 93 raise ImportError(message.format(name=name, extra=extra))
,→from None

284 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

94 else:
95 return None

ImportError: Missing optional dependency 'fastparquet'. fastparquet is

,→required for parquet support. Use pip or conda to install fastparquet.

In [515]: result = pd.read_parquet('example_pa.parquet', engine='pyarrow')

---------------------------------------------------------------------------
ImportError Traceback (most recent call last)
<ipython-input-515-f9990387c867> in <module>
----> 1 result = pd.read_parquet('example_pa.parquet', engine='pyarrow')

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in read_parquet(path,
,→engine, columns, **kwargs)

293 """
294
--> 295 impl = get_engine(engine)
296 return impl.read(path, columns=columns, **kwargs)

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in get_engine(engine)
40
41 if engine == "pyarrow":
---> 42 return PyArrowImpl()
43 elif engine == "fastparquet":
44 return FastParquetImpl()

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in __init__(self)
76 def __init__(self):
77 pyarrow = import_optional_dependency(
---> 78 "pyarrow", extra="pyarrow is required for parquet support.
,→"

79 )
80 import pyarrow.parquet

~/sandbox/pandas-release/pandas/pandas/compat/_optional.py in import_optional_
,→dependency(name, extra, raise_on_missing, on_version)

91 except ImportError:
92 if raise_on_missing:
---> 93 raise ImportError(message.format(name=name, extra=extra))
,→from None

94 else:
95 return None

ImportError: Missing optional dependency 'pyarrow'. pyarrow is required for

,→parquet support. Use pip or conda to install pyarrow.

In [516]: result.dtypes
Out[516]:
A category
B float64
dtype: object
Read only certain columns of a parquet file.
In [517]: result = pd.read_parquet('example_fp.parquet',

4.1. IO tools (text, CSV, HDF5, ) 285

pandas: powerful Python data analysis toolkit, Release 0.25.3

.....: engine='fastparquet', columns=['a', 'b'])

.....:
---------------------------------------------------------------------------
ImportError Traceback (most recent call last)
<ipython-input-517-e4f7add80b89> in <module>
1 result = pd.read_parquet('example_fp.parquet',
----> 2 engine='fastparquet', columns=['a', 'b'])

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in read_parquet(path,
,→engine, columns, **kwargs)

293 """
294
--> 295 impl = get_engine(engine)
296 return impl.read(path, columns=columns, **kwargs)

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in get_engine(engine)
42 return PyArrowImpl()
43 elif engine == "fastparquet":
---> 44 return FastParquetImpl()
45
46

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in __init__(self)
139 # we need to import on first use
140 fastparquet = import_optional_dependency(
--> 141 "fastparquet", extra="fastparquet is required for parquet
,→support."

142 )
143 self.api = fastparquet

~/sandbox/pandas-release/pandas/pandas/compat/_optional.py in import_optional_
,→dependency(name, extra, raise_on_missing, on_version)

91 except ImportError:
92 if raise_on_missing:
---> 93 raise ImportError(message.format(name=name, extra=extra))
,→from None

94 else:
95 return None

ImportError: Missing optional dependency 'fastparquet'. fastparquet is

,→required for parquet support. Use pip or conda to install fastparquet.

In [518]: result = pd.read_parquet('example_pa.parquet',

.....: engine='pyarrow', columns=['a', 'b'])
.....:
---------------------------------------------------------------------------
ImportError Traceback (most recent call last)
<ipython-input-518-be22c54a7770> in <module>
1 result = pd.read_parquet('example_pa.parquet',
----> 2 engine='pyarrow', columns=['a', 'b'])

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in read_parquet(path,
,→engine, columns, **kwargs)

293 """

286 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

294
--> 295 impl = get_engine(engine)
296 return impl.read(path, columns=columns, **kwargs)

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in get_engine(engine)
40
41 if engine == "pyarrow":
---> 42 return PyArrowImpl()
43 elif engine == "fastparquet":
44 return FastParquetImpl()

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in __init__(self)
76 def __init__(self):
77 pyarrow = import_optional_dependency(
---> 78 "pyarrow", extra="pyarrow is required for parquet support.
,→"

79 )
80 import pyarrow.parquet

~/sandbox/pandas-release/pandas/pandas/compat/_optional.py in import_optional_
,→dependency(name, extra, raise_on_missing, on_version)

91 except ImportError:
92 if raise_on_missing:
---> 93 raise ImportError(message.format(name=name, extra=extra))
,→from None

94 else:
95 return None

ImportError: Missing optional dependency 'pyarrow'. pyarrow is required for

,→parquet support. Use pip or conda to install pyarrow.

In [519]: result.dtypes
Out[519]:
A category
B float64
dtype: object

Handling indexes

Serializing a DataFrame to parquet may include the implicit index as one or more columns in the output file. Thus,
this code:
In [520]: df = pd.DataFrame({'a': [1, 2], 'b': [3, 4]})

In [521]: df.to_parquet('test.parquet', engine='pyarrow')

---------------------------------------------------------------------------
ImportError Traceback (most recent call last)
<ipython-input-521-bebaa85a69fa> in <module>
----> 1 df.to_parquet('test.parquet', engine='pyarrow')

~/sandbox/pandas-release/pandas/pandas/core/frame.py in to_parquet(self, fname,

,→engine, compression, index, partition_cols, **kwargs)

2235 index=index,
2236 partition_cols=partition_cols,
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 287

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

-> 2237 **kwargs
2238 )
2239

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in to_parquet(df, path, engine,

,→compression, index, partition_cols, **kwargs)

245 Additional keyword arguments passed to the engine

246 """
--> 247 impl = get_engine(engine)
248 return impl.write(
249 df,

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in get_engine(engine)
40
41 if engine == "pyarrow":
---> 42 return PyArrowImpl()
43 elif engine == "fastparquet":
44 return FastParquetImpl()

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in __init__(self)
76 def __init__(self):
77 pyarrow = import_optional_dependency(
---> 78 "pyarrow", extra="pyarrow is required for parquet support."
79 )
80 import pyarrow.parquet

~/sandbox/pandas-release/pandas/pandas/compat/_optional.py in import_optional_
,→dependency(name, extra, raise_on_missing, on_version)

91 except ImportError:
92 if raise_on_missing:
---> 93 raise ImportError(message.format(name=name, extra=extra)) from
,→None

94 else:
95 return None

ImportError: Missing optional dependency 'pyarrow'. pyarrow is required for parquet

,→support. Use pip or conda to install pyarrow.

creates a parquet file with three columns if you use pyarrow for serialization: a, b, and __index_level_0__.
If youre using fastparquet, the index may or may not be written to the file.
This unexpected extra column causes some databases like Amazon Redshift to reject the file, because that column
doesnt exist in the target table.
If you want to omit a dataframes indexes when writing, pass index=False to to_parquet():

In [522]: df.to_parquet('test.parquet', index=False)

---------------------------------------------------------------------------
ImportError Traceback (most recent call last)
<ipython-input-522-92139e5cd2c3> in <module>
----> 1 df.to_parquet('test.parquet', index=False)

~/sandbox/pandas-release/pandas/pandas/core/frame.py in to_parquet(self, fname,

,→engine, compression, index, partition_cols, **kwargs)

2235 index=index,
2236 partition_cols=partition_cols,
-> 2237 **kwargs
(continues on next page)

288 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2238 )
2239

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in to_parquet(df, path, engine,

,→compression, index, partition_cols, **kwargs)

245 Additional keyword arguments passed to the engine

246 """
--> 247 impl = get_engine(engine)
248 return impl.write(
249 df,

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in get_engine(engine)
30
31 raise ImportError(
---> 32 "Unable to find a usable engine; "
33 "tried using: 'pyarrow', 'fastparquet'.\n"
34 "pyarrow or fastparquet is required for parquet "

ImportError: Unable to find a usable engine; tried using: 'pyarrow', 'fastparquet'.

pyarrow or fastparquet is required for parquet support

This creates a parquet file with just the two expected columns, a and b. If your DataFrame has a custom index, you
wont get it back when you load this file into a DataFrame.
Passing index=True will always write the index, even if thats not the underlying engines default behavior.

Partitioning Parquet files

New in version 0.24.0.

Parquet supports partitioning of data based on the values of one or more columns.

In [523]: df = pd.DataFrame({'a': [0, 0, 1, 1], 'b': [0, 1, 0, 1]})

In [524]: df.to_parquet(fname='test', engine='pyarrow',

.....: partition_cols=['a'], compression=None)
.....:
---------------------------------------------------------------------------
ImportError Traceback (most recent call last)
<ipython-input-524-a65c39059f84> in <module>
1 df.to_parquet(fname='test', engine='pyarrow',
----> 2 partition_cols=['a'], compression=None)

~/sandbox/pandas-release/pandas/pandas/core/frame.py in to_parquet(self, fname,

,→engine, compression, index, partition_cols, **kwargs)

2235 index=index,
2236 partition_cols=partition_cols,
-> 2237 **kwargs
2238 )
2239

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in to_parquet(df, path, engine,

,→compression, index, partition_cols, **kwargs)

245 Additional keyword arguments passed to the engine

246 """
--> 247 impl = get_engine(engine)
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 289

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

248 return impl.write(
249 df,

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in get_engine(engine)
40
41 if engine == "pyarrow":
---> 42 return PyArrowImpl()
43 elif engine == "fastparquet":
44 return FastParquetImpl()

~/sandbox/pandas-release/pandas/pandas/io/parquet.py in __init__(self)
76 def __init__(self):
77 pyarrow = import_optional_dependency(
---> 78 "pyarrow", extra="pyarrow is required for parquet support."
79 )
80 import pyarrow.parquet

~/sandbox/pandas-release/pandas/pandas/compat/_optional.py in import_optional_
,→dependency(name, extra, raise_on_missing, on_version)

91 except ImportError:
92 if raise_on_missing:
---> 93 raise ImportError(message.format(name=name, extra=extra)) from
,→None

94 else:
95 return None

ImportError: Missing optional dependency 'pyarrow'. pyarrow is required for parquet

,→support. Use pip or conda to install pyarrow.

The fname specifies the parent directory to which data will be saved. The partition_cols are the column names by
which the dataset will be partitioned. Columns are partitioned in the order they are given. The partition splits are
determined by the unique values in the partition columns. The above example creates a partitioned dataset that may
look like:

test
a=0
0bac803e32dc42ae83fddfd029cbdebc.parquet
...
a=1
e6ab24a4f45147b49b54a662f0c412a3.parquet
...

4.1.12 SQL queries

The pandas.io.sql module provides a collection of query wrappers to both facilitate data retrieval and to reduce
dependency on DB-specific API. Database abstraction is provided by SQLAlchemy if installed. In addition you will
need a driver library for your database. Examples of such drivers are psycopg2 for PostgreSQL or pymysql for
MySQL. For SQLite this is included in Pythons standard library by default. You can find an overview of supported
drivers for each SQL dialect in the SQLAlchemy docs.
If SQLAlchemy is not installed, a fallback is only provided for sqlite (and for mysql for backwards compatibility,
but this is deprecated and will be removed in a future version). This mode requires a Python database adapter which
respect the Python DB-API.
See also some cookbook examples for some advanced strategies.

290 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

The key functions are:

read_sql_table(table_name, con[, schema, ]) Read SQL database table into a DataFrame.

read_sql_query(sql, con[, index_col, ]) Read SQL query into a DataFrame.
read_sql(sql, con[, index_col, ]) Read SQL query or database table into a DataFrame.
DataFrame.to_sql(self, name, con[, schema, ]) Write records stored in a DataFrame to a SQL database.

pandas.read_sql_table

pandas.read_sql_table(table_name, con, schema=None, index_col=None, coerce_float=True,

parse_dates=None, columns=None, chunksize=None)
Read SQL database table into a DataFrame.
Given a table name and a SQLAlchemy connectable, returns a DataFrame. This function does not support
DBAPI connections.
Parameters
table_name [str] Name of SQL table in database.
con [SQLAlchemy connectable or str] A database URI could be provided as as str. SQLite
DBAPI connection mode not supported.
schema [str, default None] Name of SQL schema in database to query (if database flavor sup-
ports this). Uses default schema if None (default).
index_col [str or list of str, optional, default: None] Column(s) to set as index(MultiIndex).
coerce_float [bool, default True] Attempts to convert values of non-string, non-numeric objects
(like decimal.Decimal) to floating point. Can result in loss of Precision.
parse_dates [list or dict, default None]
• List of column names to parse as dates.
• Dict of {column_name: format string} where format string is strftime com-
patible in case of parsing string times or is one of (D, s, ns, ms, us) in case of parsing
integer timestamps.
• Dict of {column_name: arg dict}, where the arg dict corresponds to the key-
word arguments of pandas.to_datetime() Especially useful with databases with-
out native Datetime support, such as SQLite.
columns [list, default None] List of column names to select from SQL table.
chunksize [int, default None] If specified, returns an iterator where chunksize is the number of
rows to include in each chunk.
Returns
DataFrame A SQL table is returned as two-dimensional data structure with labeled axes.
See also:

read_sql_query Read SQL query into a DataFrame.

read_sql Read SQL query or database table into a DataFrame.

4.1. IO tools (text, CSV, HDF5, ) 291

pandas: powerful Python data analysis toolkit, Release 0.25.3

Notes

Any datetime values with time zone information will be converted to UTC.

Examples

>>> pd.read_sql_table('table_name', 'postgres:///db_name') # doctest:+SKIP

pandas.read_sql_query

pandas.read_sql_query(sql, con, index_col=None, coerce_float=True, params=None,

parse_dates=None, chunksize=None)
Read SQL query into a DataFrame.
Returns a DataFrame corresponding to the result set of the query string. Optionally provide an index_col pa-
rameter to use one of the columns as the index, otherwise default integer index will be used.
Parameters
sql [string SQL query or SQLAlchemy Selectable (select or text object)] SQL query to be exe-
cuted.
con [SQLAlchemy connectable(engine/connection), database string URI,] or sqlite3 DBAPI2
connection Using SQLAlchemy makes it possible to use any DB supported by that library.
If a DBAPI2 object, only sqlite3 is supported.
index_col [string or list of strings, optional, default: None] Column(s) to set as in-
dex(MultiIndex).
coerce_float [boolean, default True] Attempts to convert values of non-string, non-numeric ob-
jects (like decimal.Decimal) to floating point. Useful for SQL result sets.
params [list, tuple or dict, optional, default: None] List of parameters to pass to execute
method. The syntax used to pass parameters is database driver dependent. Check your
database driver documentation for which of the five syntax styles, described in PEP 249s
paramstyle, is supported. Eg. for psycopg2, uses %(name)s so use params={name : value}
parse_dates [list or dict, default: None]
• List of column names to parse as dates.
• Dict of {column_name: format string} where format string is strftime com-
patible in case of parsing string times, or is one of (D, s, ns, ms, us) in case of parsing
integer timestamps.
• Dict of {column_name: arg dict}, where the arg dict corresponds to the key-
word arguments of pandas.to_datetime() Especially useful with databases with-
out native Datetime support, such as SQLite.
chunksize [int, default None] If specified, return an iterator where chunksize is the number of
rows to include in each chunk.
Returns
DataFrame
See also:

read_sql_table Read SQL database table into a DataFrame.

292 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

read_sql

Notes

Any datetime values with time zone information parsed via the parse_dates parameter will be converted to UTC.

pandas.read_sql

pandas.read_sql(sql, con, index_col=None, coerce_float=True, params=None, parse_dates=None,

columns=None, chunksize=None)
Read SQL query or database table into a DataFrame.
This function is a convenience wrapper around read_sql_table and read_sql_query (for backward
compatibility). It will delegate to the specific function depending on the provided input. A SQL query will be
routed to read_sql_query, while a database table name will be routed to read_sql_table. Note that
the delegated function might have more specific notes about their functionality not listed here.
Parameters
sql [string or SQLAlchemy Selectable (select or text object)] SQL query to be executed or a
table name.
con [SQLAlchemy connectable (engine/connection) or database string URI] or DBAPI2 con-
nection (fallback mode)
Using SQLAlchemy makes it possible to use any DB supported by that library. If a DBAPI2
object, only sqlite3 is supported.
index_col [string or list of strings, optional, default: None] Column(s) to set as in-
dex(MultiIndex).
coerce_float [boolean, default True] Attempts to convert values of non-string, non-numeric ob-
jects (like decimal.Decimal) to floating point, useful for SQL result sets.
params [list, tuple or dict, optional, default: None] List of parameters to pass to execute
method. The syntax used to pass parameters is database driver dependent. Check your
database driver documentation for which of the five syntax styles, described in PEP 249s
paramstyle, is supported. Eg. for psycopg2, uses %(name)s so use params={name : value}
parse_dates [list or dict, default: None]
• List of column names to parse as dates.
• Dict of {column_name: format string} where format string is strftime com-
patible in case of parsing string times, or is one of (D, s, ns, ms, us) in case of parsing
integer timestamps.
• Dict of {column_name: arg dict}, where the arg dict corresponds to the key-
word arguments of pandas.to_datetime() Especially useful with databases with-
out native Datetime support, such as SQLite.
columns [list, default: None] List of column names to select from SQL table (only used when
reading a table).
chunksize [int, default None] If specified, return an iterator where chunksize is the number of
rows to include in each chunk.
Returns
DataFrame

4.1. IO tools (text, CSV, HDF5, ) 293

pandas: powerful Python data analysis toolkit, Release 0.25.3

See also:

read_sql_table Read SQL database table into a DataFrame.

read_sql_query Read SQL query into a DataFrame.

pandas.DataFrame.to_sql

DataFrame.to_sql(self, name, con, schema=None, if_exists=’fail’, index=True, index_label=None,

chunksize=None, dtype=None, method=None)
Write records stored in a DataFrame to a SQL database.
Databases supported by SQLAlchemy [?] are supported. Tables can be newly created, appended to, or overwrit-
ten.
Parameters
name [string] Name of SQL table.
con [sqlalchemy.engine.Engine or sqlite3.Connection] Using SQLAlchemy makes it possible
to use any DB supported by that library. Legacy support is provided for sqlite3.Connection
objects.
schema [string, optional] Specify the schema (if database flavor supports this). If None, use
default schema.
if_exists [{fail, replace, append}, default fail] How to behave if the table already exists.
• fail: Raise a ValueError.
• replace: Drop the table before inserting new values.
• append: Insert new values to the existing table.
index [bool, default True] Write DataFrame index as a column. Uses index_label as the column
name in the table.
index_label [string or sequence, default None] Column label for index column(s). If None is
given (default) and index is True, then the index names are used. A sequence should be
given if the DataFrame uses MultiIndex.
chunksize [int, optional] Rows will be written in batches of this size at a time. By default, all
rows will be written at once.
dtype [dict, optional] Specifying the datatype for columns. The keys should be the column
names and the values should be the SQLAlchemy types or strings for the sqlite3 legacy
mode.
method [{None, multi, callable}, default None] Controls the SQL insertion clause used:
• None : Uses standard SQL INSERT clause (one per row).
• multi: Pass multiple values in a single INSERT clause.
• callable with signature (pd_table, conn, keys, data_iter).
Details and a sample callable implementation can be found in the section insert method.
New in version 0.24.0.
Raises
ValueError When the table already exists and if_exists is fail (the default).
See also:

294 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

read_sql Read a DataFrame from a table.

Notes

Timezone aware datetime columns will be written as Timestamp with timezone type with SQLAlchemy
if supported by the database. Otherwise, the datetimes will be stored as timezone unaware timestamps local to
the original timezone.
New in version 0.24.0.

References

[?], [?]

Examples

Create an in-memory SQLite database.

>>> from sqlalchemy import create_engine
>>> engine = create_engine('sqlite://', echo=False)

Create a table from scratch with 3 rows.

>>> df = pd.DataFrame({'name' : ['User 1', 'User 2', 'User 3']})
>>> df
name
0 User 1
1 User 2
2 User 3

>>> df.to_sql('users', con=engine)

>>> engine.execute("SELECT * FROM users").fetchall()
[(0, 'User 1'), (1, 'User 2'), (2, 'User 3')]

>>> df1 = pd.DataFrame({'name' : ['User 4', 'User 5']})

>>> df1.to_sql('users', con=engine, if_exists='append')
>>> engine.execute("SELECT * FROM users").fetchall()
[(0, 'User 1'), (1, 'User 2'), (2, 'User 3'),
(0, 'User 4'), (1, 'User 5')]

Overwrite the table with just df1.

>>> df1.to_sql('users', con=engine, if_exists='replace',
... index_label='id')
>>> engine.execute("SELECT * FROM users").fetchall()
[(0, 'User 4'), (1, 'User 5')]

Specify the dtype (especially useful for integers with missing values). Notice that while pandas is forced to store
the data as floating point, the database supports nullable integers. When fetching the data with Python, we get
back integer scalars.
>>> df = pd.DataFrame({"A": [1, None, 2]})
>>> df
A
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 295

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

0 1.0
1 NaN
2 2.0

>>> from sqlalchemy.types import Integer

>>> df.to_sql('integers', con=engine, index=False,
... dtype={"A": Integer()})

>>> engine.execute("SELECT * FROM integers").fetchall()

[(1,), (None,), (2,)]

Note: The function read_sql() is a convenience wrapper around read_sql_table() and

read_sql_query() (and for backward compatibility) and will delegate to specific function depending on the
provided input (database table name or sql query). Table names do not need to be quoted if they have special charac-
ters.

In the following example, we use the SQlite SQL database engine. You can use a temporary SQLite database where
data are stored in memory.
To connect with SQLAlchemy you use the create_engine() function to create an engine object from database
URI. You only need to create the engine once per database you are connecting to. For more information on
create_engine() and the URI formatting, see the examples below and the SQLAlchemy documentation

In [525]: from sqlalchemy import create_engine

# Create your engine.

In [526]: engine = create_engine('sqlite:///:memory:')

If you want to manage your own connections you can pass one of those instead:

with engine.connect() as conn, conn.begin():

data = pd.read_sql_table('data', conn)

Writing DataFrames

Assuming the following data is in a DataFrame data, we can insert it into the database using to_sql().

id Date Col_1 Col_2 Col_3

26 2012-10-18 X 25.7 True
42 2012-10-19 Y -12.4 False
63 2012-10-20 Z 5.73 True

In [527]: data
Out[527]:
id Date Col_1 Col_2 Col_3
0 26 2010-10-18 X 27.50 True
1 42 2010-10-19 Y -12.50 False
2 63 2010-10-20 Z 5.73 True

In [528]: data.to_sql('data', engine)

296 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

With some databases, writing large DataFrames can result in errors due to packet size limitations being exceeded. This
can be avoided by setting the chunksize parameter when calling to_sql. For example, the following writes data
to the database in batches of 1000 rows at a time:

In [529]: data.to_sql('data_chunked', engine, chunksize=1000)

SQL data types

to_sql() will try to map your data to an appropriate SQL data type based on the dtype of the data. When you have
columns of dtype object, pandas will try to infer the data type.
You can always override the default type by specifying the desired SQL type of any of the columns by using the
dtype argument. This argument needs a dictionary mapping column names to SQLAlchemy types (or strings for the
sqlite3 fallback mode). For example, specifying to use the sqlalchemy String type instead of the default Text type
for string columns:

In [530]: from sqlalchemy.types import String

In [531]: data.to_sql('data_dtype', engine, dtype={'Col_1': String})

Note: Due to the limited support for timedeltas in the different database flavors, columns with type timedelta64
will be written as integer values as nanoseconds to the database and a warning will be raised.

Note: Columns of category dtype will be converted to the dense representation as you would get with np.
asarray(categorical) (e.g. for string categories this gives an array of strings). Because of this, reading the
database table back in does not generate a categorical.

Datetime data types

Using SQLAlchemy, to_sql() is capable of writing datetime data that is timezone naive or timezone aware. How-
ever, the resulting data stored in the database ultimately depends on the supported data type for datetime data of the
database system being used.
The following table lists supported data types for datetime data for some common databases. Other database dialects
may have different data types for datetime data.

Database SQL Datetime Types Timezone Support

SQLite TEXT No
MySQL TIMESTAMP or DATETIME No
PostgreSQL TIMESTAMP or TIMESTAMP WITH TIME ZONE Yes

When writing timezone aware data to databases that do not support timezones, the data will be written as timezone
naive timestamps that are in local time with respect to the timezone.
read_sql_table() is also capable of reading datetime data that is timezone aware or naive. When reading
TIMESTAMP WITH TIME ZONE types, pandas will convert the data to UTC.

4.1. IO tools (text, CSV, HDF5, ) 297

pandas: powerful Python data analysis toolkit, Release 0.25.3

Insertion method

New in version 0.24.0.

The parameter method controls the SQL insertion clause used. Possible values are:
• None: Uses standard SQL INSERT clause (one per row).
• 'multi': Pass multiple values in a single INSERT clause. It uses a special SQL syntax not supported by
all backends. This usually provides better performance for analytic databases like Presto and Redshift, but has
worse performance for traditional SQL backend if the table contains many columns. For more information
check the SQLAlchemy documention.
• callable with signature (pd_table, conn, keys, data_iter): This can be used to implement a more
performant insertion method based on specific backend dialect features.
Example of a callable using PostgreSQL COPY clause:

# Alternative to_sql() *method* for DBs that support COPY FROM

import csv
from io import StringIO

def psql_insert_copy(table, conn, keys, data_iter):

# gets a DBAPI connection that can provide a cursor
dbapi_conn = conn.connection
with dbapi_conn.cursor() as cur:
s_buf = StringIO()
writer = csv.writer(s_buf)
writer.writerows(data_iter)
s_buf.seek(0)

columns = ', '.join('"{}"'.format(k) for k in keys)

if table.schema:
table_name = '{}.{}'.format(table.schema, table.name)
else:
table_name = table.name

sql = 'COPY {} ({}) FROM STDIN WITH CSV'.format(

table_name, columns)
cur.copy_expert(sql=sql, file=s_buf)

Reading tables

read_sql_table() will read a database table given the table name and optionally a subset of columns to read.

Note: In order to use read_sql_table(), you must have the SQLAlchemy optional dependency installed.

In [532]: pd.read_sql_table('data', engine)

Out[532]:
index id Date Col_1 Col_2 Col_3
0 0 26 2010-10-18 X 27.50 True
1 1 42 2010-10-19 Y -12.50 False
2 2 63 2010-10-20 Z 5.73 True

You can also specify the name of the column as the DataFrame index, and specify a subset of columns to be read.

298 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [533]: pd.read_sql_table('data', engine, index_col='id')

Out[533]:
index Date Col_1 Col_2 Col_3
id
26 0 2010-10-18 X 27.50 True
42 1 2010-10-19 Y -12.50 False
63 2 2010-10-20 Z 5.73 True

In [534]: pd.read_sql_table('data', engine, columns=['Col_1', 'Col_2'])

Out[534]:
Col_1 Col_2
0 X 27.50
1 Y -12.50
2 Z 5.73
And you can explicitly force columns to be parsed as dates:

In [535]: pd.read_sql_table('data', engine, parse_dates=['Date'])

Out[535]:
index id Date Col_1 Col_2 Col_3
0 0 26 2010-10-18 X 27.50 True
1 1 42 2010-10-19 Y -12.50 False
2 2 63 2010-10-20 Z 5.73 True

If needed you can explicitly specify a format string, or a dict of arguments to pass to pandas.to_datetime():

pd.read_sql_table('data', engine, parse_dates={'Date': '%Y-%m-%d'})

pd.read_sql_table('data', engine,
parse_dates={'Date': {'format': '%Y-%m-%d %H:%M:%S'}})

You can check if a table exists using has_table()

Schema support

Reading from and writing to different schemas is supported through the schema keyword in the
read_sql_table() and to_sql() functions. Note however that this depends on the database flavor (sqlite
does not have schemas). For example:

df.to_sql('table', engine, schema='other_schema')

pd.read_sql_table('table', engine, schema='other_schema')

Querying

You can query using raw SQL in the read_sql_query() function. In this case you must use the SQL variant
appropriate for your database. When using SQLAlchemy, you can also pass SQLAlchemy Expression language
constructs, which are database-agnostic.

In [536]: pd.read_sql_query('SELECT * FROM data', engine)

Out[536]:
index id Date Col_1 Col_2 Col_3
0 0 26 2010-10-18 00:00:00.000000 X 27.50 1
1 1 42 2010-10-19 00:00:00.000000 Y -12.50 0
2 2 63 2010-10-20 00:00:00.000000 Z 5.73 1

Of course, you can specify a more complex query.

4.1. IO tools (text, CSV, HDF5, ) 299

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [537]: pd.read_sql_query("SELECT id, Col_1, Col_2 FROM data WHERE id = 42;",

,→engine)

Out[537]:
id Col_1 Col_2
0 42 Y -12.5

The read_sql_query() function supports a chunksize argument. Specifying this will return an iterator through
chunks of the query result:

In [538]: df = pd.DataFrame(np.random.randn(20, 3), columns=list('abc'))

In [539]: df.to_sql('data_chunks', engine, index=False)

In [540]: for chunk in pd.read_sql_query("SELECT * FROM data_chunks",

.....: engine, chunksize=5):
.....: print(chunk)
.....:
a b c
0 1.305880 1.962862 -0.365194
1 0.193323 2.602824 -0.896993
2 -2.389181 -1.261600 0.868444
3 0.393496 0.609451 -0.845067
4 -0.322214 -0.447158 0.284659
a b c
0 0.201085 -0.048657 0.720448
1 -0.137165 -0.684949 -0.519145
2 1.271559 -0.861960 0.105051
3 -0.541846 -0.103767 -1.095443
4 -1.361663 1.684795 -0.891315
a b c
0 -0.539063 -0.544872 1.480449
1 0.613611 0.709609 -0.599927
2 0.100530 0.380645 -1.272573
3 -0.868754 0.111400 -1.851400
4 -0.034434 0.323798 -0.247849
a b c
0 1.900594 -0.746429 -0.317105
1 -0.095588 -0.781659 0.205437
2 -0.407292 -0.007602 -0.257895
3 2.356896 -0.650727 0.813023
4 1.094871 0.784225 -0.217173

You can also run a plain query without creating a DataFrame with execute(). This is useful for queries that dont
return values, such as INSERT. This is functionally equivalent to calling execute on the SQLAlchemy engine or db
connection object. Again, you must use the SQL syntax variant appropriate for your database.

from pandas.io import sql

sql.execute('SELECT * FROM table_name', engine)
sql.execute('INSERT INTO table_name VALUES(?, ?, ?)', engine,
params=[('id', 1, 12.2, True)])

Engine connection examples

To connect with SQLAlchemy you use the create_engine() function to create an engine object from database
URI. You only need to create the engine once per database you are connecting to.

300 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

from sqlalchemy import create_engine

engine = create_engine('postgresql://scott:tiger@localhost:5432/mydatabase')

engine = create_engine('mysql+mysqldb://scott:tiger@localhost/foo')

engine = create_engine('oracle://scott:[email protected]:1521/sidname')

engine = create_engine('mssql+pyodbc://mydsn')

# sqlite://<nohostname>/<path>
# where <path> is relative:
engine = create_engine('sqlite:///foo.db')

# or absolute, starting with a slash:

engine = create_engine('sqlite:////absolute/path/to/foo.db')

For more information see the examples the SQLAlchemy documentation

Advanced SQLAlchemy queries

You can use SQLAlchemy constructs to describe your query.

Use sqlalchemy.text() to specify query parameters in a backend-neutral way

In [541]: import sqlalchemy as sa

In [542]: pd.read_sql(sa.text('SELECT * FROM data where Col_1=:col1'),

.....: engine, params={'col1': 'X'})
.....:
Out[542]:
index id Date Col_1 Col_2 Col_3
0 0 26 2010-10-18 00:00:00.000000 X 27.5 1

If you have an SQLAlchemy description of your database you can express where conditions using SQLAlchemy
expressions

In [543]: metadata = sa.MetaData()

In [544]: data_table = sa.Table('data', metadata,

.....: sa.Column('index', sa.Integer),
.....: sa.Column('Date', sa.DateTime),
.....: sa.Column('Col_1', sa.String),
.....: sa.Column('Col_2', sa.Float),
.....: sa.Column('Col_3', sa.Boolean),
.....: )
.....:

In [545]: pd.read_sql(sa.select([data_table]).where(data_table.c.Col_3 is True),

,→engine)

Out[545]:
Empty DataFrame
Columns: [index, Date, Col_1, Col_2, Col_3]
Index: []

You can combine SQLAlchemy expressions with parameters passed to read_sql() using sqlalchemy.
bindparam()

4.1. IO tools (text, CSV, HDF5, ) 301

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [546]: import datetime as dt

In [547]: expr = sa.select([data_table]).where(data_table.c.Date > sa.bindparam('date

,→'))

In [548]: pd.read_sql(expr, engine, params={'date': dt.datetime(2010, 10, 18)})

Out[548]:
index Date Col_1 Col_2 Col_3
0 1 2010-10-19 Y -12.50 False
1 2 2010-10-20 Z 5.73 True

Sqlite fallback

The use of sqlite is supported without using SQLAlchemy. This mode requires a Python database adapter which
respect the Python DB-API.
You can create connections like so:

import sqlite3
con = sqlite3.connect(':memory:')

And then issue the following queries:

data.to_sql('data', con)
pd.read_sql_query("SELECT * FROM data", con)

4.1.13 Google BigQuery

Warning: Starting in 0.20.0, pandas has split off Google BigQuery support into the separate package
pandas-gbq. You can pip install pandas-gbq to get it.

The pandas-gbq package provides functionality to read/write from Google BigQuery.

pandas integrates with this external package. if pandas-gbq is installed, you can use the pandas methods pd.
read_gbq and DataFrame.to_gbq, which will call the respective functions from pandas-gbq.
Full documentation can be found here.

4.1.14 Stata format

Writing to stata format

The method to_stata() will write a DataFrame into a .dta file. The format version of this file is always 115 (Stata
12).

In [549]: df = pd.DataFrame(np.random.randn(10, 2), columns=list('AB'))

In [550]: df.to_stata('stata.dta')

Stata data files have limited data type support; only strings with 244 or fewer characters, int8, int16, int32,
float32 and float64 can be stored in .dta files. Additionally, Stata reserves certain values to represent missing

302 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

data. Exporting a non-missing value that is outside of the permitted range in Stata for a particular data type will retype
the variable to the next larger size. For example, int8 values are restricted to lie between -127 and 100 in Stata, and
so variables with values above 100 will trigger a conversion to int16. nan values in floating points data types are
stored as the basic missing data type (. in Stata).

Note: It is not possible to export missing data values for integer data types.

The Stata writer gracefully handles other data types including int64, bool, uint8, uint16, uint32 by casting
to the smallest supported type that can represent the data. For example, data with a type of uint8 will be cast to
int8 if all values are less than 100 (the upper bound for non-missing int8 data in Stata), or, if values are outside of
this range, the variable is cast to int16.

Warning: Conversion from int64 to float64 may result in a loss of precision if int64 values are larger than
2**53.

Warning: StataWriter and to_stata() only support fixed width strings containing up to 244 characters,
a limitation imposed by the version 115 dta file format. Attempting to write Stata dta files with strings longer than
244 characters raises a ValueError.

Reading from Stata format

The top-level function read_stata will read a dta file and return either a DataFrame or a StataReader that
can be used to read the file incrementally.
In [551]: pd.read_stata('stata.dta')
Out[551]:
index A B
0 0 -2.802620 -1.031153
1 1 -0.471722 -1.004288
2 2 -0.809833 1.537958
3 3 -0.833349 2.502008
4 4 -1.016559 -0.583782
5 5 -0.369422 0.146956
6 6 -0.815559 -1.032447
7 7 0.676818 -0.410341
8 8 1.171674 2.227913
9 9 0.764637 1.540163

Specifying a chunksize yields a StataReader instance that can be used to read chunksize lines from the file
at a time. The StataReader object can be used as an iterator.
In [552]: reader = pd.read_stata('stata.dta', chunksize=3)

In [553]: for df in reader:

.....: print(df.shape)
.....:
(3, 3)
(3, 3)
(3, 3)
(1, 3)

For more fine-grained control, use iterator=True and specify chunksize with each call to read().

4.1. IO tools (text, CSV, HDF5, ) 303

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [554]: reader = pd.read_stata('stata.dta', iterator=True)

In [555]: chunk1 = reader.read(5)

In [556]: chunk2 = reader.read(5)

Currently the index is retrieved as a column.

The parameter convert_categoricals indicates whether value labels should be read and used to create a
Categorical variable from them. Value labels can also be retrieved by the function value_labels, which
requires read() to be called before use.
The parameter convert_missing indicates whether missing value representations in Stata should be preserved.
If False (the default), missing values are represented as np.nan. If True, missing values are represented using
StataMissingValue objects, and columns containing missing values will have object data type.

Note: read_stata() and StataReader support .dta formats 113-115 (Stata 10-12), 117 (Stata 13), and 118
(Stata 14).

Note: Setting preserve_dtypes=False will upcast to the standard pandas data types: int64 for all integer
types and float64 for floating point data. By default, the Stata data types are preserved when importing.

Categorical data

Categorical data can be exported to Stata data files as value labeled data. The exported data consists of the
underlying category codes as integer data values and the categories as value labels. Stata does not have an explicit
equivalent to a Categorical and information about whether the variable is ordered is lost when exporting.

Warning: Stata only supports string value labels, and so str is called on the categories when exporting data.
Exporting Categorical variables with non-string categories produces a warning, and can result a loss of infor-
mation if the str representations of the categories are not unique.

Labeled data can similarly be imported from Stata data files as Categorical variables using the keyword argu-
ment convert_categoricals (True by default). The keyword argument order_categoricals (True by
default) determines whether imported Categorical variables are ordered.

Note: When importing categorical data, the values of the variables in the Stata data file are not preserved
since Categorical variables always use integer data types between -1 and n-1 where n is the number
of categories. If the original values in the Stata data file are required, these can be imported by setting
convert_categoricals=False, which will import original data (but not the variable labels). The original
values can be matched to the imported categorical data since there is a simple mapping between the original Stata
data values and the category codes of imported Categorical variables: missing values are assigned code -1, and the
smallest original value is assigned 0, the second smallest is assigned 1 and so on until the largest original value is
assigned the code n-1.

Note: Stata supports partially labeled series. These series have value labels for some but not all data values. Importing
a partially labeled series will produce a Categorical with string categories for the values that are labeled and

304 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

numeric categories for values with no label.

4.1.15 SAS formats

The top-level function read_sas() can read (but not write) SAS xport (.XPT) and (since v0.18.0) SAS7BDAT
(.sas7bdat) format files.
SAS files only contain two value types: ASCII text and floating point values (usually 8 bytes but sometimes truncated).
For xport files, there is no automatic type conversion to integers, dates, or categoricals. For SAS7BDAT files, the
format codes may allow date variables to be automatically converted to dates. By default the whole file is read and
returned as a DataFrame.
Specify a chunksize or use iterator=True to obtain reader objects (XportReader or SAS7BDATReader)
for incrementally reading the file. The reader objects also have attributes that contain additional information about the
file and its variables.
Read a SAS7BDAT file:

df = pd.read_sas('sas_data.sas7bdat')

Obtain an iterator and read an XPORT file 100,000 lines at a time:

def do_something(chunk):
pass

rdr = pd.read_sas('sas_xport.xpt', chunk=100000)

for chunk in rdr:
do_something(chunk)

The specification for the xport file format is available from the SAS web site.
No official documentation is available for the SAS7BDAT format.

4.1.16 Other file formats

pandas itself only supports IO with a limited set of file formats that map cleanly to its tabular data model. For reading
and writing other file formats into and from pandas, we recommend these packages from the broader community.

netCDF

xarray provides data structures inspired by the pandas DataFrame for working with multi-dimensional datasets, with
a focus on the netCDF file format and easy conversion to and from pandas.

4.1.17 Performance considerations

This is an informal comparison of various IO methods, using pandas 0.20.3. Timings are machine dependent and small
differences should be ignored.

In [1]: sz = 1000000
In [2]: df = pd.DataFrame({'A': np.random.randn(sz), 'B': [1] * sz})

In [3]: df.info()
(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 305

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 1000000 entries, 0 to 999999
Data columns (total 2 columns):
A 1000000 non-null float64
B 1000000 non-null int64
dtypes: float64(1), int64(1)
memory usage: 15.3 MB

Given the next test set:

from numpy.random import randn

sz = 1000000
df = pd.DataFrame({'A': randn(sz), 'B': [1] * sz})

def test_sql_write(df):
if os.path.exists('test.sql'):
os.remove('test.sql')
sql_db = sqlite3.connect('test.sql')
df.to_sql(name='test_table', con=sql_db)
sql_db.close()

def test_sql_read():
sql_db = sqlite3.connect('test.sql')
pd.read_sql_query("select * from test_table", sql_db)
sql_db.close()

def test_hdf_fixed_write(df):
df.to_hdf('test_fixed.hdf', 'test', mode='w')

def test_hdf_fixed_read():
pd.read_hdf('test_fixed.hdf', 'test')

def test_hdf_fixed_write_compress(df):
df.to_hdf('test_fixed_compress.hdf', 'test', mode='w', complib='blosc')

def test_hdf_fixed_read_compress():
pd.read_hdf('test_fixed_compress.hdf', 'test')

def test_hdf_table_write(df):
df.to_hdf('test_table.hdf', 'test', mode='w', format='table')

def test_hdf_table_read():
pd.read_hdf('test_table.hdf', 'test')

def test_hdf_table_write_compress(df):
df.to_hdf('test_table_compress.hdf', 'test', mode='w',
complib='blosc', format='table')
(continues on next page)

306 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

def test_hdf_table_read_compress():
pd.read_hdf('test_table_compress.hdf', 'test')

def test_csv_write(df):
df.to_csv('test.csv', mode='w')

def test_csv_read():
pd.read_csv('test.csv', index_col=0)

def test_feather_write(df):
df.to_feather('test.feather')

def test_feather_read():
pd.read_feather('test.feather')

def test_pickle_write(df):
df.to_pickle('test.pkl')

def test_pickle_read():
pd.read_pickle('test.pkl')

def test_pickle_write_compress(df):
df.to_pickle('test.pkl.compress', compression='xz')

def test_pickle_read_compress():
pd.read_pickle('test.pkl.compress', compression='xz')

When writing, the top-three functions in terms of speed are are test_pickle_write, test_feather_write
and test_hdf_fixed_write_compress.

In [14]: %timeit test_sql_write(df)

2.37 s ± 36.6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [15]: %timeit test_hdf_fixed_write(df)

194 ms ± 65.9 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [26]: %timeit test_hdf_fixed_write_compress(df)

119 ms ± 2.15 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [16]: %timeit test_hdf_table_write(df)

623 ms ± 125 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [27]: %timeit test_hdf_table_write_compress(df)

563 ms ± 23.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [17]: %timeit test_csv_write(df)

(continues on next page)

4.1. IO tools (text, CSV, HDF5, ) 307

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

3.13 s ± 49.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [30]: %timeit test_feather_write(df)

103 ms ± 5.88 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [31]: %timeit test_pickle_write(df)

109 ms ± 3.72 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [32]: %timeit test_pickle_write_compress(df)

3.33 s ± 55.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

When reading, the top three are test_feather_read, test_pickle_read and test_hdf_fixed_read.

In [18]: %timeit test_sql_read()

1.35 s ± 14.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [19]: %timeit test_hdf_fixed_read()

14.3 ms ± 438 ţs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [28]: %timeit test_hdf_fixed_read_compress()

23.5 ms ± 672 ţs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [20]: %timeit test_hdf_table_read()

35.4 ms ± 314 ţs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [29]: %timeit test_hdf_table_read_compress()

42.6 ms ± 2.1 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [22]: %timeit test_csv_read()

516 ms ± 27.1 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [33]: %timeit test_feather_read()

4.06 ms ± 115 ţs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [34]: %timeit test_pickle_read()

6.5 ms ± 172 ţs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [35]: %timeit test_pickle_read_compress()

588 ms ± 3.57 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Space on disk (in bytes)

34816000 Aug 21 18:00 test.sql

24009240 Aug 21 18:00 test_fixed.hdf
7919610 Aug 21 18:00 test_fixed_compress.hdf
24458892 Aug 21 18:00 test_table.hdf
8657116 Aug 21 18:00 test_table_compress.hdf
28520770 Aug 21 18:00 test.csv
16000248 Aug 21 18:00 test.feather
16000848 Aug 21 18:00 test.pkl
7554108 Aug 21 18:00 test.pkl.compress

{{ header }}

308 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

4.2 Indexing and selecting data

The axis labeling information in pandas objects serves many purposes:

• Identifies data (i.e. provides metadata) using known indicators, important for analysis, visualization, and inter-
active console display.
• Enables automatic and explicit data alignment.
• Allows intuitive getting and setting of subsets of the data set.
In this section, we will focus on the final point: namely, how to slice, dice, and generally get and set subsets of pandas
objects. The primary focus will be on Series and DataFrame as they have received more development attention in this
area.

Note: The Python and NumPy indexing operators [] and attribute operator . provide quick and easy access to pandas
data structures across a wide range of use cases. This makes interactive work intuitive, as theres little new to learn if
you already know how to deal with Python dictionaries and NumPy arrays. However, since the type of the data to be
accessed isnt known in advance, directly using standard operators has some optimization limits. For production code,
we recommended that you take advantage of the optimized pandas data access methods exposed in this chapter.

Warning: Whether a copy or a reference is returned for a setting operation, may depend on the context. This is
sometimes called chained assignment and should be avoided. See Returning a View versus Copy.

Warning: Indexing on an integer-based Index with floats has been clarified in 0.18.0, for a summary of the
changes, see here.

See the MultiIndex / Advanced Indexing for MultiIndex and more advanced indexing documentation.
See the cookbook for some advanced strategies.

4.2.1 Different choices for indexing

Object selection has had a number of user-requested additions in order to support more explicit location based index-
ing. Pandas now supports three types of multi-axis indexing.
• .loc is primarily label based, but may also be used with a boolean array. .loc will raise KeyError when
the items are not found. Allowed inputs are:
– A single label, e.g. 5 or 'a' (Note that 5 is interpreted as a label of the index. This use is not an integer
position along the index.).
– A list or array of labels ['a', 'b', 'c'].
– A slice object with labels 'a':'f' (Note that contrary to usual python slices, both the start and the stop
are included, when present in the index! See Slicing with labels and Endpoints are inclusive.)
– A boolean array
– A callable function with one argument (the calling Series or DataFrame) and that returns valid output
for indexing (one of the above).
New in version 0.18.1.

4.2. Indexing and selecting data 309

pandas: powerful Python data analysis toolkit, Release 0.25.3

See more at Selection by Label.

• .iloc is primarily integer position based (from 0 to length-1 of the axis), but may also be used with a
boolean array. .iloc will raise IndexError if a requested indexer is out-of-bounds, except slice indexers
which allow out-of-bounds indexing. (this conforms with Python/NumPy slice semantics). Allowed inputs are:
– An integer e.g. 5.
– A list or array of integers [4, 3, 0].
– A slice object with ints 1:7.
– A boolean array.
– A callable function with one argument (the calling Series or DataFrame) and that returns valid output
for indexing (one of the above).
New in version 0.18.1.
See more at Selection by Position, Advanced Indexing and Advanced Hierarchical.
• .loc, .iloc, and also [] indexing can accept a callable as indexer. See more at Selection By Callable.
Getting values from an object with multi-axes selection uses the following notation (using .loc as an example, but
the following applies to .iloc as well). Any of the axes accessors may be the null slice :. Axes left out of the
specification are assumed to be :, e.g. p.loc['a'] is equivalent to p.loc['a', :, :].

Object Type Indexers

Series s.loc[indexer]
DataFrame df.loc[row_indexer,column_indexer]

4.2.2 Basics

As mentioned when introducing the data structures in the last section, the primary function of indexing with [] (a.k.a.
__getitem__ for those familiar with implementing class behavior in Python) is selecting out lower-dimensional
slices. The following table shows return type values when indexing pandas objects with []:

Object Type Selection Return Value Type

Series series[label] scalar value
DataFrame frame[colname] Series corresponding to colname

Here we construct a simple time series data set to use for illustrating the indexing functionality:

In [1]: dates = pd.date_range('1/1/2000', periods=8)

In [2]: df = pd.DataFrame(np.random.randn(8, 4),

...: index=dates, columns=['A', 'B', 'C', 'D'])
...:

In [3]: df
Out[3]:
A B C D
2000-01-01 -1.157426 -0.096491 0.999344 -1.482012
2000-01-02 0.189291 0.926828 -0.029095 1.776600
2000-01-03 -1.334294 2.085399 -0.633036 0.208208
2000-01-04 -1.723333 -0.355486 -0.143959 0.177635
2000-01-05 1.071746 -0.516876 -0.382709 0.888600
(continues on next page)

310 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2000-01-06 -0.156260 -0.720254 -0.837161 -0.426902
2000-01-07 -0.354174 0.510804 0.156535 0.294767
2000-01-08 -1.448608 -1.191084 -0.128338 -0.687717

Note: None of the indexing functionality is time series specific unless specifically stated.

Thus, as per above, we have the most basic indexing using []:

In [4]: s = df['A']

In [5]: s[dates[5]]
Out[5]: -0.15625969875302725

You can pass a list of columns to [] to select columns in that order. If a column is not contained in the DataFrame, an
exception will be raised. Multiple columns can also be set in this manner:

In [6]: df
Out[6]:
A B C D
2000-01-01 -1.157426 -0.096491 0.999344 -1.482012
2000-01-02 0.189291 0.926828 -0.029095 1.776600
2000-01-03 -1.334294 2.085399 -0.633036 0.208208
2000-01-04 -1.723333 -0.355486 -0.143959 0.177635
2000-01-05 1.071746 -0.516876 -0.382709 0.888600
2000-01-06 -0.156260 -0.720254 -0.837161 -0.426902
2000-01-07 -0.354174 0.510804 0.156535 0.294767
2000-01-08 -1.448608 -1.191084 -0.128338 -0.687717

In [7]: df[['B', 'A']] = df[['A', 'B']]

In [8]: df
Out[8]:
A B C D
2000-01-01 -0.096491 -1.157426 0.999344 -1.482012
2000-01-02 0.926828 0.189291 -0.029095 1.776600
2000-01-03 2.085399 -1.334294 -0.633036 0.208208
2000-01-04 -0.355486 -1.723333 -0.143959 0.177635
2000-01-05 -0.516876 1.071746 -0.382709 0.888600
2000-01-06 -0.720254 -0.156260 -0.837161 -0.426902
2000-01-07 0.510804 -0.354174 0.156535 0.294767
2000-01-08 -1.191084 -1.448608 -0.128338 -0.687717

You may find this useful for applying a transform (in-place) to a subset of the columns.

Warning: pandas aligns all AXES when setting Series and DataFrame from .loc, and .iloc.
This will not modify df because the column alignment is before value assignment.

4.2. Indexing and selecting data 311

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [9]: df[['A', 'B']]

Out[9]:
A B
2000-01-01 -0.096491 -1.157426
2000-01-02 0.926828 0.189291
2000-01-03 2.085399 -1.334294
2000-01-04 -0.355486 -1.723333
2000-01-05 -0.516876 1.071746
2000-01-06 -0.720254 -0.156260
2000-01-07 0.510804 -0.354174
2000-01-08 -1.191084 -1.448608

In [10]: df.loc[:, ['B', 'A']] = df[['A', 'B']]

In [11]: df[['A', 'B']]

Out[11]:
A B
2000-01-01 -0.096491 -1.157426
2000-01-02 0.926828 0.189291
2000-01-03 2.085399 -1.334294
2000-01-04 -0.355486 -1.723333
2000-01-05 -0.516876 1.071746
2000-01-06 -0.720254 -0.156260
2000-01-07 0.510804 -0.354174
2000-01-08 -1.191084 -1.448608

The correct way to swap column values is by using raw values:

In [12]: df.loc[:, ['B', 'A']] = df[['A', 'B']].to_numpy()

In [13]: df[['A', 'B']]

Out[13]:
A B
2000-01-01 -1.157426 -0.096491
2000-01-02 0.189291 0.926828
2000-01-03 -1.334294 2.085399
2000-01-04 -1.723333 -0.355486
2000-01-05 1.071746 -0.516876
2000-01-06 -0.156260 -0.720254
2000-01-07 -0.354174 0.510804
2000-01-08 -1.448608 -1.191084

4.2.3 Attribute access

You may access an index on a Series or column on a DataFrame directly as an attribute:

In [14]: sa = pd.Series([1, 2, 3], index=list('abc'))

In [15]: dfa = df.copy()

In [16]: sa.b
Out[16]: 2

In [17]: dfa.A
Out[17]:

312 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

2000-01-01 -1.157426
2000-01-02 0.189291
2000-01-03 -1.334294
2000-01-04 -1.723333
2000-01-05 1.071746
2000-01-06 -0.156260
2000-01-07 -0.354174
2000-01-08 -1.448608
Freq: D, Name: A, dtype: float64

In [18]: sa.a = 5

In [19]: sa
Out[19]:
a 5
b 2
c 3
dtype: int64

In [20]: dfa.A = list(range(len(dfa.index))) # ok if A already exists

In [21]: dfa
Out[21]:
A B C D
2000-01-01 0 -0.096491 0.999344 -1.482012
2000-01-02 1 0.926828 -0.029095 1.776600
2000-01-03 2 2.085399 -0.633036 0.208208
2000-01-04 3 -0.355486 -0.143959 0.177635
2000-01-05 4 -0.516876 -0.382709 0.888600
2000-01-06 5 -0.720254 -0.837161 -0.426902
2000-01-07 6 0.510804 0.156535 0.294767
2000-01-08 7 -1.191084 -0.128338 -0.687717

In [22]: dfa['A'] = list(range(len(dfa.index))) # use this form to create a new

,→column

In [23]: dfa
Out[23]:
A B C D
2000-01-01 0 -0.096491 0.999344 -1.482012
2000-01-02 1 0.926828 -0.029095 1.776600
2000-01-03 2 2.085399 -0.633036 0.208208
2000-01-04 3 -0.355486 -0.143959 0.177635
2000-01-05 4 -0.516876 -0.382709 0.888600
2000-01-06 5 -0.720254 -0.837161 -0.426902
2000-01-07 6 0.510804 0.156535 0.294767
2000-01-08 7 -1.191084 -0.128338 -0.687717

Warning:
• You can use this access only if the index element is a valid Python identifier, e.g. s.1 is not allowed. See
here for an explanation of valid identifiers.
• The attribute will not be available if it conflicts with an existing method name, e.g. s.min is not allowed.
• Similarly, the attribute will not be available if it conflicts with any of the following list: index,
major_axis, minor_axis, items.

4.2. Indexing and selecting data 313

pandas: powerful Python data analysis toolkit, Release 0.25.3

• In any of these cases, standard indexing will still work, e.g. s['1'], s['min'], and s['index'] will
access the corresponding element or column.

If you are using the IPython environment, you may also use tab-completion to see these accessible attributes.
You can also assign a dict to a row of a DataFrame:

In [24]: x = pd.DataFrame({'x': [1, 2, 3], 'y': [3, 4, 5]})

In [25]: x.iloc[1] = {'x': 9, 'y': 99}

In [26]: x
Out[26]:
x y
0 1 3
1 9 99
2 3 5

You can use attribute access to modify an existing element of a Series or column of a DataFrame, but be careful; if
you try to use attribute access to create a new column, it creates a new attribute rather than a new column. In 0.21.0
and later, this will raise a UserWarning:

In [1]: df = pd.DataFrame({'one': [1., 2., 3.]})

In [2]: df.two = [4, 5, 6]
UserWarning: Pandas doesn't allow Series to be assigned into nonexistent columns -
,→see https://pandas.pydata.org/pandas-docs/stable/indexing.html#attribute_access

In [3]: df
Out[3]:
one
0 1.0
1 2.0
2 3.0

4.2.4 Slicing ranges

The most robust and consistent way of slicing ranges along arbitrary axes is described in the Selection by Position
section detailing the .iloc method. For now, we explain the semantics of slicing using the [] operator.
With Series, the syntax works exactly as with an ndarray, returning a slice of the values and the corresponding labels:
In [27]: s[:5]
Out[27]:
2000-01-01 -1.157426
2000-01-02 0.189291
2000-01-03 -1.334294
2000-01-04 -1.723333
2000-01-05 1.071746
Freq: D, Name: A, dtype: float64

In [28]: s[::2]
Out[28]:
2000-01-01 -1.157426
2000-01-03 -1.334294
2000-01-05 1.071746
2000-01-07 -0.354174

314 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Freq: 2D, Name: A, dtype: float64

In [29]: s[::-1]
Out[29]:
2000-01-08 -1.448608
2000-01-07 -0.354174
2000-01-06 -0.156260
2000-01-05 1.071746
2000-01-04 -1.723333
2000-01-03 -1.334294
2000-01-02 0.189291
2000-01-01 -1.157426
Freq: -1D, Name: A, dtype: float64
Note that setting works as well:

In [30]: s2 = s.copy()

In [31]: s2[:5] = 0

In [32]: s2
Out[32]:
2000-01-01 0.000000
2000-01-02 0.000000
2000-01-03 0.000000
2000-01-04 0.000000
2000-01-05 0.000000
2000-01-06 -0.156260
2000-01-07 -0.354174
2000-01-08 -1.448608
Freq: D, Name: A, dtype: float64

With DataFrame, slicing inside of [] slices the rows. This is provided largely as a convenience since it is such a
common operation.
In [33]: df[:3]
Out[33]:
A B C D
2000-01-01 -1.157426 -0.096491 0.999344 -1.482012
2000-01-02 0.189291 0.926828 -0.029095 1.776600
2000-01-03 -1.334294 2.085399 -0.633036 0.208208

In [34]: df[::-1]
Out[34]:
A B C D
2000-01-08 -1.448608 -1.191084 -0.128338 -0.687717
2000-01-07 -0.354174 0.510804 0.156535 0.294767
2000-01-06 -0.156260 -0.720254 -0.837161 -0.426902
2000-01-05 1.071746 -0.516876 -0.382709 0.888600
2000-01-04 -1.723333 -0.355486 -0.143959 0.177635
2000-01-03 -1.334294 2.085399 -0.633036 0.208208
2000-01-02 0.189291 0.926828 -0.029095 1.776600
2000-01-01 -1.157426 -0.096491 0.999344 -1.482012

4.2. Indexing and selecting data 315

pandas: powerful Python data analysis toolkit, Release 0.25.3

4.2.5 Selection by label

Warning: Whether a copy or a reference is returned for a setting operation, may depend on the context. This is
sometimes called chained assignment and should be avoided. See Returning a View versus Copy.

Warning:
.loc is strict when you present slicers that are not compatible (or convertible) with the index type.
For example using integers in a DatetimeIndex. These will raise a TypeError.
In [35]: dfl = pd.DataFrame(np.random.randn(5, 4),
....: columns=list('ABCD'),
....: index=pd.date_range('20130101', periods=5))
....:

In [36]: dfl
Out[36]:
A B C D
2013-01-01 -2.100928 1.025928 -0.007973 -1.336035
2013-01-02 0.915382 -0.130655 0.022627 -1.425459
2013-01-03 0.271939 0.169543 0.692513 -1.231139
2013-01-04 1.692870 0.783855 -0.721626 1.698994
2013-01-05 -0.349882 1.586451 1.454199 0.149458

In [4]: dfl.loc[2:3]
TypeError: cannot do slice indexing on <class 'pandas.tseries.index.DatetimeIndex'>
,→with these indexers [2] of <type 'int'>

String likes in slicing can be convertible to the type of the index and lead to natural slicing.
In [37]: dfl.loc['20130102':'20130104']
Out[37]:
A B C D
2013-01-02 0.915382 -0.130655 0.022627 -1.425459
2013-01-03 0.271939 0.169543 0.692513 -1.231139
2013-01-04 1.692870 0.783855 -0.721626 1.698994

Warning: Starting in 0.21.0, pandas will show a FutureWarning if indexing with a list with missing labels.
In the future this will raise a KeyError. See list-like Using loc with missing keys in a list is Deprecated.

pandas provides a suite of methods in order to have purely label based indexing. This is a strict inclusion based
protocol. Every label asked for must be in the index, or a KeyError will be raised. When slicing, both the start
bound AND the stop bound are included, if present in the index. Integers are valid labels, but they refer to the label
and not the position.
The .loc attribute is the primary access method. The following are valid inputs:
• A single label, e.g. 5 or 'a' (Note that 5 is interpreted as a label of the index. This use is not an integer
position along the index.).
• A list or array of labels ['a', 'b', 'c'].
• A slice object with labels 'a':'f' (Note that contrary to usual python slices, both the start and the stop are
included, when present in the index! See Slicing with labels.

316 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

• A boolean array.
• A callable, see Selection By Callable.
In [38]: s1 = pd.Series(np.random.randn(6), index=list('abcdef'))

In [39]: s1
Out[39]:
a 1.018393
b -0.565651
c -1.590951
d -0.651777
e 0.456407
f 1.134054
dtype: float64

In [40]: s1.loc['c':]
Out[40]:
c -1.590951
d -0.651777
e 0.456407
f 1.134054
dtype: float64

In [41]: s1.loc['b']
Out[41]: -0.5656513106035832
Note that setting works as well:

In [42]: s1.loc['c':] = 0

In [43]: s1
Out[43]:
a 1.018393
b -0.565651
c 0.000000
d 0.000000
e 0.000000
f 0.000000
dtype: float64

With a DataFrame:
In [44]: df1 = pd.DataFrame(np.random.randn(6, 4),
....: index=list('abcdef'),
....: columns=list('ABCD'))
....:

In [45]: df1
Out[45]:
A B C D
a 1.695493 1.632303 -1.726092 0.486227
b -0.625187 0.386616 -0.048112 -2.598355
c -0.871135 -0.209156 0.004590 0.449006
d 0.573428 0.697186 -2.442512 -1.423556
e -0.304997 0.672794 0.954090 1.323584
f 0.015720 -0.815293 0.164562 0.576599

4.2. Indexing and selecting data 317

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [46]: df1.loc[['a', 'b', 'd'], :]

Out[46]:
A B C D
a 1.695493 1.632303 -1.726092 0.486227
b -0.625187 0.386616 -0.048112 -2.598355
d 0.573428 0.697186 -2.442512 -1.423556
Accessing via label slices:

In [47]: df1.loc['d':, 'A':'C']

Out[47]:
A B C
d 0.573428 0.697186 -2.442512
e -0.304997 0.672794 0.954090
f 0.015720 -0.815293 0.164562

For getting a cross section using a label (equivalent to df.xs('a')):

In [48]: df1.loc['a']
Out[48]:
A 1.695493
B 1.632303
C -1.726092
D 0.486227
Name: a, dtype: float64

For getting values with a boolean array:

In [49]: df1.loc['a'] > 0
Out[49]:
A True
B True
C False
D True
Name: a, dtype: bool

In [50]: df1.loc[:, df1.loc['a'] > 0]

Out[50]:
A B D
a 1.695493 1.632303 0.486227
b -0.625187 0.386616 -2.598355
c -0.871135 -0.209156 0.449006
d 0.573428 0.697186 -1.423556
e -0.304997 0.672794 1.323584
f 0.015720 -0.815293 0.576599
For getting a value explicitly (equivalent to deprecated df.get_value('a','A')):

# this is also equivalent to ``df1.at['a','A']``

In [51]: df1.loc['a', 'A']
Out[51]: 1.6954931738440173

318 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Slicing with labels

When using .loc with slices, if both the start and the stop labels are present in the index, then elements located
between the two (including them) are returned:
In [52]: s = pd.Series(list('abcde'), index=[0, 3, 2, 5, 4])

In [53]: s.loc[3:5]
Out[53]:
3 b
2 c
5 d
dtype: object

If at least one of the two is absent, but the index is sorted, and can be compared against start and stop labels, then
slicing will still work as expected, by selecting labels which rank between the two:
In [54]: s.sort_index()
Out[54]:
0 a
2 c
3 b
4 e
5 d
dtype: object

In [55]: s.sort_index().loc[1:6]
Out[55]:
2 c
3 b
4 e
5 d
dtype: object
However, if at least one of the two is absent and the index is not sorted, an error will be raised (since doing otherwise
would be computationally expensive, as well as potentially ambiguous for mixed type indexes). For instance, in the
above example, s.loc[1:6] would raise KeyError.
For the rationale behind this behavior, see Endpoints are inclusive.

4.2.6 Selection by position

Warning: Whether a copy or a reference is returned for a setting operation, may depend on the context. This is
sometimes called chained assignment and should be avoided. See Returning a View versus Copy.

Pandas provides a suite of methods in order to get purely integer based indexing. The semantics follow closely
Python and NumPy slicing. These are 0-based indexing. When slicing, the start bound is included, while the upper
bound is excluded. Trying to use a non-integer, even a valid label will raise an IndexError.
The .iloc attribute is the primary access method. The following are valid inputs:
• An integer e.g. 5.
• A list or array of integers [4, 3, 0].
• A slice object with ints 1:7.

4.2. Indexing and selecting data 319

pandas: powerful Python data analysis toolkit, Release 0.25.3

• A boolean array.
• A callable, see Selection By Callable.
In [56]: s1 = pd.Series(np.random.randn(5), index=list(range(0, 10, 2)))

In [57]: s1
Out[57]:
0 -0.813837
2 1.020094
4 -0.538755
6 -0.273898
8 1.374350
dtype: float64

In [58]: s1.iloc[:3]
Out[58]:
0 -0.813837
2 1.020094
4 -0.538755
dtype: float64

In [59]: s1.iloc[3]
Out[59]: -0.2738980637291465
Note that setting works as well:

In [60]: s1.iloc[:3] = 0

In [61]: s1
Out[61]:
0 0.000000
2 0.000000
4 0.000000
6 -0.273898
8 1.374350
dtype: float64

With a DataFrame:

In [62]: df1 = pd.DataFrame(np.random.randn(6, 4),

....: index=list(range(0, 12, 2)),
....: columns=list(range(0, 8, 2)))
....:

In [63]: df1
Out[63]:
0 2 4 6
0 -0.769208 -0.094955 -0.339642 1.131238
2 -1.165074 0.191823 -0.424832 0.641310
4 -1.389117 0.367491 2.164790 1.126079
6 1.550817 0.826973 -0.677486 2.087563
8 0.117134 -0.855723 0.082120 1.276149
10 0.270969 1.210188 -0.988631 -1.253327

Select via integer slicing:

In [64]: df1.iloc[:3]

320 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[64]:
0 2 4 6
0 -0.769208 -0.094955 -0.339642 1.131238
2 -1.165074 0.191823 -0.424832 0.641310
4 -1.389117 0.367491 2.164790 1.126079

In [65]: df1.iloc[1:5, 2:4]

Out[65]:
4 6
2 -0.424832 0.641310
4 2.164790 1.126079
6 -0.677486 2.087563
8 0.082120 1.276149
Select via integer list:

In [66]: df1.iloc[[1, 3, 5], [1, 3]]

Out[66]:
2 6
2 0.191823 0.641310
6 0.826973 2.087563
10 1.210188 -1.253327

In [67]: df1.iloc[1:3, :]
Out[67]:
0 2 4 6
2 -1.165074 0.191823 -0.424832 0.641310
4 -1.389117 0.367491 2.164790 1.126079

In [68]: df1.iloc[:, 1:3]

Out[68]:
2 4
0 -0.094955 -0.339642
2 0.191823 -0.424832
4 0.367491 2.164790
6 0.826973 -0.677486
8 -0.855723 0.082120
10 1.210188 -0.988631

# this is also equivalent to ``df1.iat[1,1]``

In [69]: df1.iloc[1, 1]
Out[69]: 0.1918230563556888

For getting a cross section using an integer position (equiv to df.xs(1)):

In [70]: df1.iloc[1]
Out[70]:
0 -1.165074
2 0.191823
4 -0.424832
6 0.641310
Name: 2, dtype: float64

Out of range slice indexes are handled gracefully just as in Python/Numpy.

# these are allowed in python/numpy.
In [71]: x = list('abcdef')

4.2. Indexing and selecting data 321

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [72]: x
Out[72]: ['a', 'b', 'c', 'd', 'e', 'f']

In [73]: x[4:10]
Out[73]: ['e', 'f']

In [74]: x[8:10]
Out[74]: []

In [75]: s = pd.Series(x)

In [76]: s
Out[76]:
0 a
1 b
2 c
3 d
4 e
5 f
dtype: object

In [77]: s.iloc[4:10]
Out[77]:
4 e
5 f
dtype: object

In [78]: s.iloc[8:10]
Out[78]: Series([], dtype: object)
Note that using slices that go out of bounds can result in an empty axis (e.g. an empty DataFrame being returned).
In [79]: dfl = pd.DataFrame(np.random.randn(5, 2), columns=list('AB'))

In [80]: dfl
Out[80]:
A B
0 -1.141824 0.487982
1 1.110710 1.775785
2 -0.158929 1.256688
3 0.480722 0.545781
4 -1.214729 0.259405

In [81]: dfl.iloc[:, 2:3]

Out[81]:
Empty DataFrame
Columns: []
Index: [0, 1, 2, 3, 4]

In [82]: dfl.iloc[:, 1:3]

Out[82]:
B
0 0.487982
1 1.775785

322 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

2 1.256688
3 0.545781
4 0.259405

In [83]: dfl.iloc[4:6]
Out[83]:
A B
4 -1.214729 0.259405
A single indexer that is out of bounds will raise an IndexError. A list of indexers where any element is out of
bounds will raise an IndexError.

>>> dfl.iloc[[4, 5, 6]]

IndexError: positional indexers are out-of-bounds

>>> dfl.iloc[:, 4]
IndexError: single positional indexer is out-of-bounds

4.2.7 Selection by callable

New in version 0.18.1.

.loc, .iloc, and also [] indexing can accept a callable as indexer. The callable must be a function with
one argument (the calling Series or DataFrame) that returns valid output for indexing.
In [84]: df1 = pd.DataFrame(np.random.randn(6, 4),
....: index=list('abcdef'),
....: columns=list('ABCD'))
....:

In [85]: df1
Out[85]:
A B C D
a -1.898204 0.933280 0.410757 -1.209116
b 1.072207 -2.076376 -0.032087 -1.179905
c 0.819041 0.169362 0.395066 1.793339
d 0.620358 0.687095 0.924752 -0.953211
e 0.272744 -0.264613 -0.299304 0.828769
f 1.384847 1.408420 -0.599304 1.455457

In [86]: df1.loc[lambda df: df.A > 0, :]

Out[86]:
A B C D
b 1.072207 -2.076376 -0.032087 -1.179905
c 0.819041 0.169362 0.395066 1.793339
d 0.620358 0.687095 0.924752 -0.953211
e 0.272744 -0.264613 -0.299304 0.828769
f 1.384847 1.408420 -0.599304 1.455457

In [87]: df1.loc[:, lambda df: ['A', 'B']]

Out[87]:
A B
a -1.898204 0.933280
b 1.072207 -2.076376
c 0.819041 0.169362

4.2. Indexing and selecting data 323

pandas: powerful Python data analysis toolkit, Release 0.25.3

d 0.620358 0.687095
e 0.272744 -0.264613
f 1.384847 1.408420

In [88]: df1.iloc[:, lambda df: [0, 1]]

Out[88]:
A B
a -1.898204 0.933280
b 1.072207 -2.076376
c 0.819041 0.169362
d 0.620358 0.687095
e 0.272744 -0.264613
f 1.384847 1.408420

In [89]: df1[lambda df: df.columns[0]]

Out[89]:
a -1.898204
b 1.072207
c 0.819041
d 0.620358
e 0.272744
f 1.384847
Name: A, dtype: float64
You can use callable indexing in Series.
In [90]: df1.A.loc[lambda s: s > 0]
Out[90]:
b 1.072207
c 0.819041
d 0.620358
e 0.272744
f 1.384847
Name: A, dtype: float64

Using these methods / indexers, you can chain data selection operations without using a temporary variable.
In [91]: bb = pd.read_csv('data/baseball.csv', index_col='id')

In [92]: (bb.groupby(['year', 'team']).sum()

....: .loc[lambda df: df.r > 100])
....:
Out[92]:
stint g ab r h X2b X3b hr rbi sb cs bb so
,→ibb hbp sh sf gidp
year team
,→

2007 CIN 6 379 745 101 203 35 2 36 125.0 10.0 1.0 105 127.0 14.
,→0 1.0 1.0 15.0 18.0
DET 5 301 1062 162 283 54 4 37 144.0 24.0 7.0 97 176.0 3.
,→0 10.0 4.0 8.0 28.0
HOU 4 311 926 109 218 47 6 14 77.0 10.0 4.0 60 212.0 3.
,→0 9.0 16.0 6.0 17.0
LAN 11 413 1021 153 293 61 3 36 154.0 7.0 5.0 114 141.0 8.
,→0 9.0 3.0 8.0 29.0
NYN 13 622 1854 240 509 101 3 61 243.0 22.0 4.0 174 310.0 24.
,→0 23.0 18.0 15.0 48.0
(continues on next page)

324 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

SFN 5 482 1305 198 337 67 6 40 171.0 26.0 7.0 235 188.0 51.
,→ 0 8.0 16.0 6.0 41.0
TEX 2 198 729 115 200 40 4 28 115.0 21.0 4.0 73 140.0 4.
,→ 0 5.0 2.0 8.0 16.0
TOR 4 459 1408 187 378 96 2 58 223.0 4.0 2.0 190 265.0 16.
,→ 0 12.0 4.0 16.0 38.0

4.2.8 IX indexer is deprecated

Warning: Starting in 0.20.0, the .ix indexer is deprecated, in favor of the more strict .iloc and .loc indexers.

.ix offers a lot of magic on the inference of what the user wants to do. To wit, .ix can decide to index positionally
OR via labels depending on the data type of the index. This has caused quite a bit of user confusion over the years.
The recommended methods of indexing are:
• .loc if you want to label index.
• .iloc if you want to positionally index.

In [93]: dfd = pd.DataFrame({'A': [1, 2, 3],

....: 'B': [4, 5, 6]},
....: index=list('abc'))
....:

In [94]: dfd
Out[94]:
A B
a 1 4
b 2 5
c 3 6

Previous behavior, where you wish to get the 0th and the 2nd elements from the index in the A column.

In [3]: dfd.ix[[0, 2], 'A']

Out[3]:
a 1
c 3
Name: A, dtype: int64

Using .loc. Here we will select the appropriate indexes from the index, then use label indexing.

In [95]: dfd.loc[dfd.index[[0, 2]], 'A']

Out[95]:
a 1
c 3
Name: A, dtype: int64

This can also be expressed using .iloc, by explicitly getting locations on the indexers, and using positional indexing
to select things.

In [96]: dfd.iloc[[0, 2], dfd.columns.get_loc('A')]

Out[96]:
(continues on next page)

4.2. Indexing and selecting data 325

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

a 1
c 3
Name: A, dtype: int64

For getting multiple indexers, using .get_indexer:

In [97]: dfd.iloc[[0, 2], dfd.columns.get_indexer(['A', 'B'])]
Out[97]:
A B
a 1 4
c 3 6

4.2.9 Indexing with list with missing labels is deprecated

Warning: Starting in 0.21.0, using .loc or [] with a list with one or more missing labels, is deprecated, in favor
of .reindex.

In prior versions, using .loc[list-of-labels] would work as long as at least 1 of the keys was found (oth-
erwise it would raise a KeyError). This behavior is deprecated and will show a warning message pointing to this
section. The recommended alternative is to use .reindex().
For example.
In [98]: s = pd.Series([1, 2, 3])

In [99]: s
Out[99]:
0 1
1 2
2 3
dtype: int64

Selection with all keys found is unchanged.

In [100]: s.loc[[1, 2]]
Out[100]:
1 2
2 3
dtype: int64

Previous behavior
In [4]: s.loc[[1, 2, 3]]
Out[4]:
1 2.0
2 3.0
3 NaN
dtype: float64

Current behavior
In [4]: s.loc[[1, 2, 3]]
Passing list-likes to .loc with any non-matching elements will raise
(continues on next page)

326 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

KeyError in the future, you can use .reindex() as an alternative.

See the documentation here:

http://pandas.pydata.org/pandas-docs/stable/indexing.html#deprecate-loc-reindex-
,→listlike

Out[4]:
1 2.0
2 3.0
3 NaN
dtype: float64

Reindexing

The idiomatic way to achieve selecting potentially not-found elements is via .reindex(). See also the section on
reindexing.

In [101]: s.reindex([1, 2, 3])

Out[101]:
1 2.0
2 3.0
3 NaN
dtype: float64

Alternatively, if you want to select only valid keys, the following is idiomatic and efficient; it is guaranteed to preserve
the dtype of the selection.

In [102]: labels = [1, 2, 3]

In [103]: s.loc[s.index.intersection(labels)]
Out[103]:
1 2
2 3
dtype: int64

Having a duplicated index will raise for a .reindex():

In [104]: s = pd.Series(np.arange(4), index=['a', 'a', 'b', 'c'])

In [105]: labels = ['c', 'd']

In [17]: s.reindex(labels)
ValueError: cannot reindex from a duplicate axis

Generally, you can intersect the desired labels with the current axis, and then reindex.

In [106]: s.loc[s.index.intersection(labels)].reindex(labels)
Out[106]:
c 3.0
d NaN
dtype: float64

However, this would still raise if your resulting index is duplicated.

4.2. Indexing and selecting data 327

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [41]: labels = ['a', 'd']

In [42]: s.loc[s.index.intersection(labels)].reindex(labels)
ValueError: cannot reindex from a duplicate axis

4.2.10 Selecting random samples

A random selection of rows or columns from a Series or DataFrame with the sample() method. The method will
sample rows by default, and accepts a specific number of rows/columns to return, or a fraction of rows.
In [107]: s = pd.Series([0, 1, 2, 3, 4, 5])

# When no arguments are passed, returns 1 row.

In [108]: s.sample()
Out[108]:
0 0
dtype: int64

# One may specify either a number of rows:

In [109]: s.sample(n=3)
Out[109]:
2 2
1 1
5 5
dtype: int64

# Or a fraction of the rows:

In [110]: s.sample(frac=0.5)
Out[110]:
0 0
2 2
3 3
dtype: int64
By default, sample will return each row at most once, but one can also sample with replacement using the replace
option:
In [111]: s = pd.Series([0, 1, 2, 3, 4, 5])

# Without replacement (default):

In [112]: s.sample(n=6, replace=False)
Out[112]:
3 3
2 2
4 4
0 0
5 5
1 1
dtype: int64

# With replacement:
In [113]: s.sample(n=6, replace=True)
Out[113]:
0 0

328 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

2 2
0 0
3 3
1 1
2 2
dtype: int64
By default, each row has an equal probability of being selected, but if you want rows to have different probabilities,
you can pass the sample function sampling weights as weights. These weights can be a list, a NumPy array, or a
Series, but they must be of the same length as the object you are sampling. Missing values will be treated as a weight
of zero, and inf values are not allowed. If weights do not sum to 1, they will be re-normalized by dividing all weights
by the sum of the weights. For example:

In [114]: s = pd.Series([0, 1, 2, 3, 4, 5])

In [115]: example_weights = [0, 0, 0.2, 0.2, 0.2, 0.4]

In [116]: s.sample(n=3, weights=example_weights)

Out[116]:
5 5
4 4
3 3
dtype: int64

# Weights will be re-normalized automatically

In [117]: example_weights2 = [0.5, 0, 0, 0, 0, 0]

In [118]: s.sample(n=1, weights=example_weights2)

Out[118]:
0 0
dtype: int64

When applied to a DataFrame, you can use a column of the DataFrame as sampling weights (provided you are sampling
rows and not columns) by simply passing the name of the column as a string.

In [119]: df2 = pd.DataFrame({'col1': [9, 8, 7, 6],

.....: 'weight_column': [0.5, 0.4, 0.1, 0]})
.....:

In [120]: df2.sample(n=3, weights='weight_column')

Out[120]:
col1 weight_column
0 9 0.5
2 7 0.1
1 8 0.4

sample also allows users to sample columns instead of rows using the axis argument.

In [121]: df3 = pd.DataFrame({'col1': [1, 2, 3], 'col2': [2, 3, 4]})

In [122]: df3.sample(n=1, axis=1)

Out[122]:
col1
0 1
1 2
2 3

Finally, one can also set a seed for samples random number generator using the random_state argument, which

4.2. Indexing and selecting data 329

pandas: powerful Python data analysis toolkit, Release 0.25.3

will accept either an integer (as a seed) or a NumPy RandomState object.

In [123]: df4 = pd.DataFrame({'col1': [1, 2, 3], 'col2': [2, 3, 4]})

# With a given seed, the sample will always draw the same rows.
In [124]: df4.sample(n=2, random_state=2)
Out[124]:
col1 col2
2 3 4
1 2 3

In [125]: df4.sample(n=2, random_state=2)

Out[125]:
col1 col2
2 3 4
1 2 3

4.2.11 Setting with enlargement

The .loc/[] operations can perform enlargement when setting a non-existent key for that axis.
In the Series case this is effectively an appending operation.

In [126]: se = pd.Series([1, 2, 3])

In [127]: se
Out[127]:
0 1
1 2
2 3
dtype: int64

In [128]: se[5] = 5.

In [129]: se
Out[129]:
0 1.0
1 2.0
2 3.0
5 5.0
dtype: float64

A DataFrame can be enlarged on either axis via .loc.

In [130]: dfi = pd.DataFrame(np.arange(6).reshape(3, 2),

.....: columns=['A', 'B'])
.....:

In [131]: dfi
Out[131]:
A B
0 0 1
1 2 3
2 4 5

In [132]: dfi.loc[:, 'C'] = dfi.loc[:, 'A']

(continues on next page)

330 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [133]: dfi
Out[133]:
A B C
0 0 1 0
1 2 3 2
2 4 5 4

This is like an append operation on the DataFrame.

In [134]: dfi.loc[3] = 5

In [135]: dfi
Out[135]:
A B C
0 0 1 0
1 2 3 2
2 4 5 4
3 5 5 5

4.2.12 Fast scalar value getting and setting

Since indexing with [] must handle a lot of cases (single-label access, slicing, boolean indexing, etc.), it has a bit of
overhead in order to figure out what youre asking for. If you only want to access a scalar value, the fastest way is to
use the at and iat methods, which are implemented on all of the data structures.
Similarly to loc, at provides label based scalar lookups, while, iat provides integer based lookups analogously to
iloc
In [136]: s.iat[5]
Out[136]: 5

In [137]: df.at[dates[5], 'A']

Out[137]: -0.15625969875302725

In [138]: df.iat[3, 0]
Out[138]: -1.7233332966009836
You can also set using these same indexers.

In [139]: df.at[dates[5], 'E'] = 7

In [140]: df.iat[3, 0] = 7

at may enlarge the object in-place as above if the indexer is missing.

In [141]: df.at[dates[-1] + pd.Timedelta('1 day'), 0] = 7

In [142]: df
Out[142]:
A B C D E 0
2000-01-01 -1.157426 -0.096491 0.999344 -1.482012 NaN NaN
2000-01-02 0.189291 0.926828 -0.029095 1.776600 NaN NaN
2000-01-03 -1.334294 2.085399 -0.633036 0.208208 NaN NaN
2000-01-04 7.000000 -0.355486 -0.143959 0.177635 NaN NaN
(continues on next page)

4.2. Indexing and selecting data 331

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2000-01-05 1.071746 -0.516876 -0.382709 0.888600 NaN NaN
2000-01-06 -0.156260 -0.720254 -0.837161 -0.426902 7.0 NaN
2000-01-07 -0.354174 0.510804 0.156535 0.294767 NaN NaN
2000-01-08 -1.448608 -1.191084 -0.128338 -0.687717 NaN NaN
2000-01-09 NaN NaN NaN NaN NaN 7.0

4.2.13 Boolean indexing

Another common operation is the use of boolean vectors to filter the data. The operators are: | for or, & for and, and
~ for not. These must be grouped by using parentheses, since by default Python will evaluate an expression such as
df.A > 2 & df.B < 3 as df.A > (2 & df.B) < 3, while the desired evaluation order is (df.A > 2)
& (df.B < 3).
Using a boolean vector to index a Series works exactly as in a NumPy ndarray:
In [143]: s = pd.Series(range(-3, 4))

In [144]: s
Out[144]:
0 -3
1 -2
2 -1
3 0
4 1
5 2
6 3
dtype: int64

In [145]: s[s > 0]

Out[145]:
4 1
5 2
6 3
dtype: int64

In [146]: s[(s < -1) | (s > 0.5)]

Out[146]:
0 -3
1 -2
4 1
5 2
6 3
dtype: int64

In [147]: s[~(s < 0)]

Out[147]:
3 0
4 1
5 2
6 3
dtype: int64
You may select rows from a DataFrame using a boolean vector the same length as the DataFrames index (for example,

332 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

something derived from one of the columns of the DataFrame):

In [148]: df[df['A'] > 0]

Out[148]:
A B C D E 0
2000-01-02 0.189291 0.926828 -0.029095 1.776600 NaN NaN
2000-01-04 7.000000 -0.355486 -0.143959 0.177635 NaN NaN
2000-01-05 1.071746 -0.516876 -0.382709 0.888600 NaN NaN

List comprehensions and the map method of Series can also be used to produce more complex criteria:
In [149]: df2 = pd.DataFrame({'a': ['one', 'one', 'two', 'three', 'two',
,→'one', 'six'],

.....: 'b': ['x', 'y', 'y', 'x', 'y', 'x', 'x'],

.....: 'c': np.random.randn(7)})
.....:

# only want 'two' or 'three'

In [150]: criterion = df2['a'].map(lambda x: x.startswith('t'))

In [151]: df2[criterion]
Out[151]:
a b c
2 two y 0.344065
3 three x 1.275247
4 two y 1.303763

# equivalent but slower

In [152]: df2[[x.startswith('t') for x in df2['a']]]
Out[152]:
a b c
2 two y 0.344065
3 three x 1.275247
4 two y 1.303763

# Multiple criteria
In [153]: df2[criterion & (df2['b'] == 'x')]
Out[153]:
a b c
3 three x 1.275247
With the choice methods Selection by Label, Selection by Position, and Advanced Indexing you may select along more
than one axis using boolean vectors combined with other indexing expressions.

In [154]: df2.loc[criterion & (df2['b'] == 'x'), 'b':'c']

Out[154]:
b c
3 x 1.275247

4.2.14 Indexing with isin

Consider the isin() method of Series, which returns a boolean vector that is true wherever the Series elements
exist in the passed list. This allows you to select rows where one or more columns have values you want:
In [155]: s = pd.Series(np.arange(5), index=np.arange(5)[::-1], dtype='int64')

4.2. Indexing and selecting data 333

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [156]: s
Out[156]:
4 0
3 1
2 2
1 3
0 4
dtype: int64

In [157]: s.isin([2, 4, 6])

Out[157]:
4 False
3 False
2 True
1 False
0 True
dtype: bool

In [158]: s[s.isin([2, 4, 6])]

Out[158]:
2 2
0 4
dtype: int64
The same method is available for Index objects and is useful for the cases when you dont know which of the sought
labels are in fact present:
In [159]: s[s.index.isin([2, 4, 6])]
Out[159]:
4 0
2 2
dtype: int64

# compare it to the following

In [160]: s.reindex([2, 4, 6])
Out[160]:
2 2.0
4 0.0
6 NaN
dtype: float64
In addition to that, MultiIndex allows selecting a separate level to use in the membership check:
In [161]: s_mi = pd.Series(np.arange(6),
.....: index=pd.MultiIndex.from_product([[0, 1], ['a',
,→'b', 'c']]))

.....:

In [162]: s_mi
Out[162]:
0 a 0
b 1
c 2
1 a 3
b 4

334 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

c 5
dtype: int64

In [163]: s_mi.iloc[s_mi.index.isin([(1, 'a'), (2, 'b'), (0, 'c')])]

Out[163]:
0 c 2
1 a 3
dtype: int64

In [164]: s_mi.iloc[s_mi.index.isin(['a', 'c', 'e'], level=1)]

Out[164]:
0 a 0
c 2
1 a 3
c 5
dtype: int64
DataFrame also has an isin() method. When calling isin, pass a set of values as either an array or dict. If values is
an array, isin returns a DataFrame of booleans that is the same shape as the original DataFrame, with True wherever
the element is in the sequence of values.
In [165]: df = pd.DataFrame({'vals': [1, 2, 3, 4], 'ids': ['a', 'b', 'f', 'n'],
.....: 'ids2': ['a', 'n', 'c', 'n']})
.....:

In [166]: values = ['a', 'b', 1, 3]

In [167]: df.isin(values)
Out[167]:
vals ids ids2
0 True True True
1 False True False
2 True False False
3 False False False

Oftentimes youll want to match certain values with certain columns. Just make values a dict where the key is the
column, and the value is a list of items you want to check for.
In [168]: values = {'ids': ['a', 'b'], 'vals': [1, 3]}

In [169]: df.isin(values)
Out[169]:
vals ids ids2
0 True True False
1 False True False
2 True False False
3 False False False

Combine DataFrames isin with the any() and all() methods to quickly select subsets of your data that meet a
given criteria. To select a row where each column meets its own criterion:
In [170]: values = {'ids': ['a', 'b'], 'ids2': ['a', 'c'], 'vals': [1, 3]}

In [171]: row_mask = df.isin(values).all(1)

In [172]: df[row_mask]
Out[172]:
(continues on next page)

4.2. Indexing and selecting data 335

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

vals ids ids2
0 1 a a

4.2.15 The where() Method and Masking

Selecting values from a Series with a boolean vector generally returns a subset of the data. To guarantee that selection
output has the same shape as the original data, you can use the where method in Series and DataFrame.
To return only the selected rows:
In [173]: s[s > 0]
Out[173]:
3 1
2 2
1 3
0 4
dtype: int64

To return a Series of the same shape as the original:

In [174]: s.where(s > 0)
Out[174]:
4 NaN
3 1.0
2 2.0
1 3.0
0 4.0
dtype: float64

Selecting values from a DataFrame with a boolean criterion now also preserves input data shape. where is used under
the hood as the implementation. The code below is equivalent to df.where(df < 0).
In [175]: df[df < 0]
Out[175]:
A B C D
2000-01-01 -0.986205 -1.719758 -2.230079 -0.439106
2000-01-02 -2.397242 NaN NaN -0.237058
2000-01-03 -1.482014 NaN -0.782186 NaN
2000-01-04 -0.480306 -1.051903 -0.987736 -0.182060
2000-01-05 -0.379467 NaN -0.138556 NaN
2000-01-06 -0.514897 -0.117796 -0.108906 -1.142649
2000-01-07 -1.349120 NaN -0.128845 -1.352644
2000-01-08 NaN NaN NaN -1.495080

In addition, where takes an optional other argument for replacement of values where the condition is False, in the
returned copy.
In [176]: df.where(df < 0, -df)
Out[176]:
A B C D
2000-01-01 -0.986205 -1.719758 -2.230079 -0.439106
2000-01-02 -2.397242 -0.124508 -1.493995 -0.237058
2000-01-03 -1.482014 -0.429889 -0.782186 -0.389666
2000-01-04 -0.480306 -1.051903 -0.987736 -0.182060
2000-01-05 -0.379467 -0.273248 -0.138556 -0.881904
(continues on next page)

336 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2000-01-06 -0.514897 -0.117796 -0.108906 -1.142649
2000-01-07 -1.349120 -0.316880 -0.128845 -1.352644
2000-01-08 -0.161458 -0.739064 -0.165377 -1.495080

You may wish to set values based on some boolean criteria. This can be done intuitively like so:

In [177]: s2 = s.copy()

In [178]: s2[s2 < 0] = 0

In [179]: s2
Out[179]:
4 0
3 1
2 2
1 3
0 4
dtype: int64

In [180]: df2 = df.copy()

In [181]: df2[df2 < 0] = 0

In [182]: df2
Out[182]:
A B C D
2000-01-01 0.000000 0.000000 0.000000 0.000000
2000-01-02 0.000000 0.124508 1.493995 0.000000
2000-01-03 0.000000 0.429889 0.000000 0.389666
2000-01-04 0.000000 0.000000 0.000000 0.000000
2000-01-05 0.000000 0.273248 0.000000 0.881904
2000-01-06 0.000000 0.000000 0.000000 0.000000
2000-01-07 0.000000 0.316880 0.000000 0.000000
2000-01-08 0.161458 0.739064 0.165377 0.000000

By default, where returns a modified copy of the data. There is an optional parameter inplace so that the original
data can be modified without creating a copy:

In [183]: df_orig = df.copy()

In [184]: df_orig.where(df > 0, -df, inplace=True)

In [185]: df_orig
Out[185]:
A B C D
2000-01-01 0.986205 1.719758 2.230079 0.439106
2000-01-02 2.397242 0.124508 1.493995 0.237058
2000-01-03 1.482014 0.429889 0.782186 0.389666
2000-01-04 0.480306 1.051903 0.987736 0.182060
2000-01-05 0.379467 0.273248 0.138556 0.881904
2000-01-06 0.514897 0.117796 0.108906 1.142649
2000-01-07 1.349120 0.316880 0.128845 1.352644
2000-01-08 0.161458 0.739064 0.165377 1.495080

Note: The signature for DataFrame.where() differs from numpy.where(). Roughly df1.where(m,
df2) is equivalent to np.where(m, df1, df2).

4.2. Indexing and selecting data 337

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [186]: df.where(df < 0, -df) == np.where(df < 0, df, -df)

Out[186]:
A B C D
2000-01-01 True True True True
2000-01-02 True True True True
2000-01-03 True True True True
2000-01-04 True True True True
2000-01-05 True True True True
2000-01-06 True True True True
2000-01-07 True True True True
2000-01-08 True True True True

Alignment
Furthermore, where aligns the input boolean condition (ndarray or DataFrame), such that partial selection with setting
is possible. This is analogous to partial setting via .loc (but on the contents rather than the axis labels).

In [187]: df2 = df.copy()

In [188]: df2[df2[1:4] > 0] = 3

In [189]: df2
Out[189]:
A B C D
2000-01-01 -0.986205 -1.719758 -2.230079 -0.439106
2000-01-02 -2.397242 3.000000 3.000000 -0.237058
2000-01-03 -1.482014 3.000000 -0.782186 3.000000
2000-01-04 -0.480306 -1.051903 -0.987736 -0.182060
2000-01-05 -0.379467 0.273248 -0.138556 0.881904
2000-01-06 -0.514897 -0.117796 -0.108906 -1.142649
2000-01-07 -1.349120 0.316880 -0.128845 -1.352644
2000-01-08 0.161458 0.739064 0.165377 -1.495080

Where can also accept axis and level parameters to align the input when performing the where.

In [190]: df2 = df.copy()

In [191]: df2.where(df2 > 0, df2['A'], axis='index')

Out[191]:
A B C D
2000-01-01 -0.986205 -0.986205 -0.986205 -0.986205
2000-01-02 -2.397242 0.124508 1.493995 -2.397242
2000-01-03 -1.482014 0.429889 -1.482014 0.389666
2000-01-04 -0.480306 -0.480306 -0.480306 -0.480306
2000-01-05 -0.379467 0.273248 -0.379467 0.881904
2000-01-06 -0.514897 -0.514897 -0.514897 -0.514897
2000-01-07 -1.349120 0.316880 -1.349120 -1.349120
2000-01-08 0.161458 0.739064 0.165377 0.161458

This is equivalent to (but faster than) the following.

In [192]: df2 = df.copy()

In [193]: df.apply(lambda x, y: x.where(x > 0, y), y=df['A'])

Out[193]:
A B C D
(continues on next page)

338 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2000-01-01 -0.986205 -0.986205 -0.986205 -0.986205
2000-01-02 -2.397242 0.124508 1.493995 -2.397242
2000-01-03 -1.482014 0.429889 -1.482014 0.389666
2000-01-04 -0.480306 -0.480306 -0.480306 -0.480306
2000-01-05 -0.379467 0.273248 -0.379467 0.881904
2000-01-06 -0.514897 -0.514897 -0.514897 -0.514897
2000-01-07 -1.349120 0.316880 -1.349120 -1.349120
2000-01-08 0.161458 0.739064 0.165377 0.161458

New in version 0.18.1.

Where can accept a callable as condition and other arguments. The function must be with one argument (the calling
Series or DataFrame) and that returns valid output as condition and other argument.
In [194]: df3 = pd.DataFrame({'A': [1, 2, 3],
.....: 'B': [4, 5, 6],
.....: 'C': [7, 8, 9]})
.....:

In [195]: df3.where(lambda x: x > 4, lambda x: x + 10)

Out[195]:
A B C
0 11 14 7
1 12 5 8
2 13 6 9

Mask

mask() is the inverse boolean operation of where.

In [196]: s.mask(s >= 0)
Out[196]:
4 NaN
3 NaN
2 NaN
1 NaN
0 NaN
dtype: float64

In [197]: df.mask(df >= 0)

Out[197]:
A B C D
2000-01-01 -0.986205 -1.719758 -2.230079 -0.439106
2000-01-02 -2.397242 NaN NaN -0.237058
2000-01-03 -1.482014 NaN -0.782186 NaN
2000-01-04 -0.480306 -1.051903 -0.987736 -0.182060
2000-01-05 -0.379467 NaN -0.138556 NaN
2000-01-06 -0.514897 -0.117796 -0.108906 -1.142649
2000-01-07 -1.349120 NaN -0.128845 -1.352644
2000-01-08 NaN NaN NaN -1.495080

4.2.16 The query() Method

DataFrame objects have a query() method that allows selection using an expression.

4.2. Indexing and selecting data 339

pandas: powerful Python data analysis toolkit, Release 0.25.3

You can get the value of the frame where column b has values between the values of columns a and c. For example:
In [198]: n = 10

In [199]: df = pd.DataFrame(np.random.rand(n, 3), columns=list('abc'))

In [200]: df
Out[200]:
a b c
0 0.564354 0.446221 0.031135
1 0.902823 0.612305 0.538181
2 0.663245 0.164731 0.782808
3 0.194429 0.253263 0.923684
4 0.698026 0.425477 0.840259
5 0.739830 0.738261 0.105783
6 0.214871 0.824421 0.246428
7 0.821225 0.243330 0.650292
8 0.373653 0.171969 0.479279
9 0.450277 0.365091 0.217416

# pure python
In [201]: df[(df.a < df.b) & (df.b < df.c)]
Out[201]:
a b c
3 0.194429 0.253263 0.923684

# query
In [202]: df.query('(a < b) & (b < c)')
Out[202]:
a b c
3 0.194429 0.253263 0.923684
Do the same thing but fall back on a named index if there is no column with the name a.
In [203]: df = pd.DataFrame(np.random.randint(n / 2, size=(n, 2)),
,→columns=list('bc'))

In [204]: df.index.name = 'a'

In [205]: df
Out[205]:
b c
a
0 4 0
1 3 1
2 4 3
3 0 2
4 1 4
5 4 4
6 4 2
7 2 0
8 0 0
9 1 3

In [206]: df.query('a < b and b < c')

340 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[206]:
Empty DataFrame
Columns: [b, c]
Index: []
If instead you dont want to or cannot name your index, you can use the name index in your query expression:
In [207]: df = pd.DataFrame(np.random.randint(n, size=(n, 2)),
,→columns=list('bc'))

In [208]: df
Out[208]:
b c
0 4 5
1 4 1
2 0 8
3 4 7
4 0 0
5 8 5
6 1 2
7 4 3
8 1 0
9 6 3

In [209]: df.query('index < b < c')

Out[209]:
b c
0 4 5
3 4 7

Note: If the name of your index overlaps with a column name, the column name is given precedence. For example,

In [210]: df = pd.DataFrame({'a': np.random.randint(5, size=5)})

In [211]: df.index.name = 'a'

In [212]: df.query('a > 2') # uses the column 'a', not the index
Out[212]:
a
a
0 4
1 4
2 3

You can still use the index in a query expression by using the special identifier index:

In [213]: df.query('index > 2')

Out[213]:
a
a
3 1
4 1

If for some reason you have a column named index, then you can refer to the index as ilevel_0 as well, but at
this point you should consider renaming your columns to something less ambiguous.

4.2. Indexing and selecting data 341

pandas: powerful Python data analysis toolkit, Release 0.25.3

MultiIndex query() Syntax

You can also use the levels of a DataFrame with a MultiIndex as if they were columns in the frame:
In [214]: n = 10

In [215]: colors = np.random.choice(['red', 'green'], size=n)

In [216]: foods = np.random.choice(['eggs', 'ham'], size=n)

In [217]: colors
Out[217]:
array(['red', 'red', 'green', 'green', 'red', 'red', 'green', 'green',
'green', 'red'], dtype='<U5')

In [218]: foods
Out[218]:
array(['eggs', 'ham', 'ham', 'ham', 'ham', 'ham', 'eggs', 'eggs', 'ham',
'eggs'], dtype='<U4')

In [219]: index = pd.MultiIndex.from_arrays([colors, foods], names=['color',

,→'food'])

In [220]: df = pd.DataFrame(np.random.randn(n, 2), index=index)

In [221]: df
Out[221]:
0 1
color food
red eggs 0.397240 1.722883
ham 0.634589 1.761948
green ham 1.191222 -0.748678
ham -0.013401 -0.982325
red ham 0.272726 1.042615
ham 0.267082 0.191461
green eggs -0.435659 -0.035917
eggs 0.194931 0.970348
ham 2.187055 0.383666
red eggs -0.812383 -0.497327

In [222]: df.query('color == "red"')

Out[222]:
0 1
color food
red eggs 0.397240 1.722883
ham 0.634589 1.761948
ham 0.272726 1.042615
ham 0.267082 0.191461
eggs -0.812383 -0.497327
If the levels of the MultiIndex are unnamed, you can refer to them using special names:
In [223]: df.index.names = [None, None]

In [224]: df

342 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[224]:
0 1
red eggs 0.397240 1.722883
ham 0.634589 1.761948
green ham 1.191222 -0.748678
ham -0.013401 -0.982325
red ham 0.272726 1.042615
ham 0.267082 0.191461
green eggs -0.435659 -0.035917
eggs 0.194931 0.970348
ham 2.187055 0.383666
red eggs -0.812383 -0.497327

In [225]: df.query('ilevel_0 == "red"')

Out[225]:
0 1
red eggs 0.397240 1.722883
ham 0.634589 1.761948
ham 0.272726 1.042615
ham 0.267082 0.191461
eggs -0.812383 -0.497327
The convention is ilevel_0, which means index level 0 for the 0th level of the index.

query() Use Cases

A use case for query() is when you have a collection of DataFrame objects that have a subset of column names
(or index levels/names) in common. You can pass the same query to both frames without having to specify which
frame youre interested in querying

In [226]: df = pd.DataFrame(np.random.rand(n, 3), columns=list('abc'))

In [227]: df
Out[227]:
a b c
0 0.483974 0.645639 0.413412
1 0.611039 0.585546 0.848970
2 0.523271 0.811649 0.517849
3 0.947506 0.143525 0.055154
4 0.934891 0.214973 0.271028
5 0.832143 0.777114 0.572133
6 0.304056 0.712288 0.960006
7 0.965451 0.803696 0.866318
8 0.965355 0.383391 0.647743
9 0.639263 0.218103 0.886788

In [228]: df2 = pd.DataFrame(np.random.rand(n + 2, 3), columns=df.columns)

In [229]: df2
Out[229]:
a b c
0 0.563472 0.298326 0.361543
1 0.021200 0.761846 0.279478
2 0.274321 0.127032 0.025433
3 0.789059 0.154680 0.999703
(continues on next page)

4.2. Indexing and selecting data 343

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

4 0.195936 0.042450 0.475367
5 0.970329 0.053024 0.293762
6 0.877607 0.352530 0.300746
7 0.259895 0.666779 0.920354
8 0.861035 0.176572 0.638339
9 0.083984 0.834057 0.673247
10 0.190267 0.647251 0.586836
11 0.395322 0.575815 0.184662

In [230]: expr = '0.0 <= a <= c <= 0.5'

In [231]: map(lambda frame: frame.query(expr), [df, df2])

Out[231]: <map at 0x125179610>

query() Python versus pandas Syntax Comparison

Full numpy-like syntax:

In [232]: df = pd.DataFrame(np.random.randint(n, size=(n, 3)),
,→columns=list('abc'))

In [233]: df
Out[233]:
a b c
0 4 1 0
1 9 3 7
2 6 7 9
3 1 9 4
4 6 9 9
5 1 1 2
6 6 6 6
7 9 6 3
8 0 7 9
9 3 8 3

In [234]: df.query('(a < b) & (b < c)')

Out[234]:
a b c
2 6 7 9
8 0 7 9

In [235]: df[(df.a < df.b) & (df.b < df.c)]

Out[235]:
a b c
2 6 7 9
8 0 7 9
Slightly nicer by removing the parentheses (by binding making comparison operators bind tighter than & and |).

In [236]: df.query('a < b & b < c')

Out[236]:
a b c
2 6 7 9
8 0 7 9

344 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Use English instead of symbols:

In [237]: df.query('a < b and b < c')
Out[237]:
a b c
2 6 7 9
8 0 7 9

Pretty close to how you might write it on paper:

In [238]: df.query('a < b < c')
Out[238]:
a b c
2 6 7 9
8 0 7 9

The in and not in operators

query() also supports special use of Pythons in and not in comparison operators, providing a succinct syntax
for calling the isin method of a Series or DataFrame.
# get all rows where columns "a" and "b" have overlapping values
In [239]: df = pd.DataFrame({'a': list('aabbccddeeff'), 'b':
,→list('aaaabbbbcccc'),

.....: 'c': np.random.randint(5, size=12),

.....: 'd': np.random.randint(9, size=12)})
.....:

In [240]: df
Out[240]:
a b c d
0 a a 1 5
1 a a 1 1
2 b a 4 1
3 b a 4 3
4 c b 4 5
5 c b 0 0
6 d b 0 4
7 d b 3 3
8 e c 4 5
9 e c 2 0
10 f c 2 2
11 f c 4 7

In [241]: df.query('a in b')

Out[241]:
a b c d
0 a a 1 5
1 a a 1 1
2 b a 4 1
3 b a 4 3
4 c b 4 5
5 c b 0 0

# How you'd do it in pure Python

4.2. Indexing and selecting data 345

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [242]: df[df.a.isin(df.b)]
Out[242]:
a b c d
0 a a 1 5
1 a a 1 1
2 b a 4 1
3 b a 4 3
4 c b 4 5
5 c b 0 0

In [243]: df.query('a not in b')

Out[243]:
a b c d
6 d b 0 4
7 d b 3 3
8 e c 4 5
9 e c 2 0
10 f c 2 2
11 f c 4 7

# pure Python
In [244]: df[~df.a.isin(df.b)]
Out[244]:
a b c d
6 d b 0 4
7 d b 3 3
8 e c 4 5
9 e c 2 0
10 f c 2 2
11 f c 4 7
You can combine this with other expressions for very succinct queries:
# rows where cols a and b have overlapping values
# and col c's values are less than col d's
In [245]: df.query('a in b and c < d')
Out[245]:
a b c d
0 a a 1 5
4 c b 4 5

# pure Python
In [246]: df[df.b.isin(df.a) & (df.c < df.d)]
Out[246]:
a b c d
0 a a 1 5
4 c b 4 5
6 d b 0 4
8 e c 4 5
11 f c 4 7

Note: Note that in and not in are evaluated in Python, since numexpr has no equivalent of this operation.
However, only the in/not in expression itself is evaluated in vanilla Python. For example, in the expression

346 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

df.query('a in b + c + d')

(b + c + d) is evaluated by numexpr and then the in operation is evaluated in plain Python. In general, any
operations that can be evaluated using numexpr will be.

Special use of the == operator with list objects

Comparing a list of values to a column using ==/!= works similarly to in/not in.
In [247]: df.query('b == ["a", "b", "c"]')
Out[247]:
a b c d
0 a a 1 5
1 a a 1 1
2 b a 4 1
3 b a 4 3
4 c b 4 5
5 c b 0 0
6 d b 0 4
7 d b 3 3
8 e c 4 5
9 e c 2 0
10 f c 2 2
11 f c 4 7

# pure Python
In [248]: df[df.b.isin(["a", "b", "c"])]
Out[248]:
a b c d
0 a a 1 5
1 a a 1 1
2 b a 4 1
3 b a 4 3
4 c b 4 5
5 c b 0 0
6 d b 0 4
7 d b 3 3
8 e c 4 5
9 e c 2 0
10 f c 2 2
11 f c 4 7

In [249]: df.query('c == [1, 2]')

Out[249]:
a b c d
0 a a 1 5
1 a a 1 1
9 e c 2 0
10 f c 2 2

In [250]: df.query('c != [1, 2]')

Out[250]:

4.2. Indexing and selecting data 347

pandas: powerful Python data analysis toolkit, Release 0.25.3

a b c d
2 b a 4 1
3 b a 4 3
4 c b 4 5
5 c b 0 0
6 d b 0 4
7 d b 3 3
8 e c 4 5
11 f c 4 7

# using in/not in
In [251]: df.query('[1, 2] in c')
Out[251]:
a b c d
0 a a 1 5
1 a a 1 1
9 e c 2 0
10 f c 2 2

In [252]: df.query('[1, 2] not in c')

Out[252]:
a b c d
2 b a 4 1
3 b a 4 3
4 c b 4 5
5 c b 0 0
6 d b 0 4
7 d b 3 3
8 e c 4 5
11 f c 4 7

# pure Python
In [253]: df[df.c.isin([1, 2])]
Out[253]:
a b c d
0 a a 1 5
1 a a 1 1
9 e c 2 0
10 f c 2 2

Boolean operators

You can negate boolean expressions with the word not or the ~ operator.
In [254]: df = pd.DataFrame(np.random.rand(n, 3), columns=list('abc'))

In [255]: df['bools'] = np.random.rand(len(df)) > 0.5

In [256]: df.query('~bools')
Out[256]:
a b c bools
0 0.499604 0.981560 0.137759 False
4 0.104751 0.782568 0.977198 False
7 0.973197 0.245000 0.977406 False

348 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

9 0.805940 0.451425 0.070470 False

In [257]: df.query('not bools')

Out[257]:
a b c bools
0 0.499604 0.981560 0.137759 False
4 0.104751 0.782568 0.977198 False
7 0.973197 0.245000 0.977406 False
9 0.805940 0.451425 0.070470 False

In [258]: df.query('not bools') == df[~df.bools]

Out[258]:
a b c bools
0 True True True True
4 True True True True
7 True True True True
9 True True True True
Of course, expressions can be arbitrarily complex too:
# short query syntax
In [259]: shorter = df.query('a < b < c and (not bools) or bools > 2')

# equivalent in pure Python

In [260]: longer = df[(df.a < df.b) & (df.b < df.c) & (~df.bools) | (df.bools
,→> 2)]

In [261]: shorter
Out[261]:
a b c bools
4 0.104751 0.782568 0.977198 False

In [262]: longer
Out[262]:
a b c bools
4 0.104751 0.782568 0.977198 False

In [263]: shorter == longer

Out[263]:
a b c bools
4 True True True True

Performance of query()

DataFrame.query() using numexpr is slightly faster than Python for large frames.

4.2. Indexing and selecting data 349

pandas: powerful Python data analysis toolkit, Release 0.25.3

Note: You will only see the performance benefits of using the numexpr engine with DataFrame.query() if
your frame has more than approximately 200,000 rows.

This plot was created using a DataFrame with 3 columns each containing floating point values generated using
numpy.random.randn().

4.2.17 Duplicate data

If you want to identify and remove duplicate rows in a DataFrame, there are two methods that will help: duplicated
and drop_duplicates. Each takes as an argument the columns to use to identify duplicated rows.
• duplicated returns a boolean vector whose length is the number of rows, and which indicates whether a row
is duplicated.

350 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

• drop_duplicates removes duplicate rows.

By default, the first observed row of a duplicate set is considered unique, but each method has a keep parameter to
specify targets to be kept.
• keep='first' (default): mark / drop duplicates except for the first occurrence.
• keep='last': mark / drop duplicates except for the last occurrence.
• keep=False: mark / drop all duplicates.
In [264]: df2 = pd.DataFrame({'a': ['one', 'one', 'two', 'two', 'two',
,→'three', 'four'],

.....: 'b': ['x', 'y', 'x', 'y', 'x', 'x', 'x'],

.....: 'c': np.random.randn(7)})
.....:

In [265]: df2
Out[265]:
a b c
0 one x -0.086643
1 one y -0.862428
2 two x 1.155986
3 two y -0.583644
4 two x -1.416461
5 three x 0.799196
6 four x -2.063856

In [266]: df2.duplicated('a')
Out[266]:
0 False
1 True
2 False
3 True
4 True
5 False
6 False
dtype: bool

In [267]: df2.duplicated('a', keep='last')

Out[267]:
0 True
1 False
2 True
3 True
4 False
5 False
6 False
dtype: bool

In [268]: df2.duplicated('a', keep=False)

Out[268]:
0 True
1 True
2 True
3 True

4.2. Indexing and selecting data 351

pandas: powerful Python data analysis toolkit, Release 0.25.3

4 True
5 False
6 False
dtype: bool

In [269]: df2.drop_duplicates('a')
Out[269]:
a b c
0 one x -0.086643
2 two x 1.155986
5 three x 0.799196
6 four x -2.063856

In [270]: df2.drop_duplicates('a', keep='last')

Out[270]:
a b c
1 one y -0.862428
4 two x -1.416461
5 three x 0.799196
6 four x -2.063856

In [271]: df2.drop_duplicates('a', keep=False)

Out[271]:
a b c
5 three x 0.799196
6 four x -2.063856
Also, you can pass a list of columns to identify duplications.
In [272]: df2.duplicated(['a', 'b'])
Out[272]:
0 False
1 False
2 False
3 False
4 True
5 False
6 False
dtype: bool

In [273]: df2.drop_duplicates(['a', 'b'])

Out[273]:
a b c
0 one x -0.086643
1 one y -0.862428
2 two x 1.155986
3 two y -0.583644
5 three x 0.799196
6 four x -2.063856
To drop duplicates by index value, use Index.duplicated then perform slicing. The same set of options are
available for the keep parameter.
In [274]: df3 = pd.DataFrame({'a': np.arange(6),
.....: 'b': np.random.randn(6)},
.....: index=['a', 'a', 'b', 'c', 'b', 'a'])

352 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

.....:

In [275]: df3
Out[275]:
a b
a 0 -0.457673
a 1 0.315795
b 2 -0.013959
c 3 -0.376069
b 4 -0.715356
a 5 1.802760

In [276]: df3.index.duplicated()
Out[276]: array([False, True, False, False, True, True])

In [277]: df3[~df3.index.duplicated()]
Out[277]:
a b
a 0 -0.457673
b 2 -0.013959
c 3 -0.376069

In [278]: df3[~df3.index.duplicated(keep='last')]
Out[278]:
a b
c 3 -0.376069
b 4 -0.715356
a 5 1.802760

In [279]: df3[~df3.index.duplicated(keep=False)]
Out[279]:
a b
c 3 -0.376069

4.2.18 Dictionary-like get() method

Each of Series or DataFrame have a get method which can return a default value.
In [280]: s = pd.Series([1, 2, 3], index=['a', 'b', 'c'])

In [281]: s.get('a') # equivalent to s['a']

Out[281]: 1

In [282]: s.get('x', default=-1)

Out[282]: -1

4.2.19 The lookup() method

Sometimes you want to extract a set of values given a sequence of row labels and column labels, and the lookup
method allows for this and returns a NumPy array. For instance:

4.2. Indexing and selecting data 353

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [283]: dflookup = pd.DataFrame(np.random.rand(20, 4), columns = ['A', 'B', 'C', 'D

,→'])

In [284]: dflookup.lookup(list(range(0, 10, 2)), ['B', 'C', 'A', 'B', 'D'])

Out[284]: array([0.85612986, 0.02825952, 0.99403226, 0.02263787, 0.37166623])

4.2.20 Index objects

The pandas Index class and its subclasses can be viewed as implementing an ordered multiset. Duplicates are
allowed. However, if you try to convert an Index object with duplicate entries into a set, an exception will be
raised.
Index also provides the infrastructure necessary for lookups, data alignment, and reindexing. The easiest way to
create an Index directly is to pass a list or other sequence to Index:
In [285]: index = pd.Index(['e', 'd', 'a', 'b'])

In [286]: index
Out[286]: Index(['e', 'd', 'a', 'b'], dtype='object')

In [287]: 'd' in index

Out[287]: True
You can also pass a name to be stored in the index:

In [288]: index = pd.Index(['e', 'd', 'a', 'b'], name='something')

In [289]: index.name
Out[289]: 'something'

The name, if set, will be shown in the console display:

In [290]: index = pd.Index(list(range(5)), name='rows')

In [291]: columns = pd.Index(['A', 'B', 'C'], name='cols')

In [292]: df = pd.DataFrame(np.random.randn(5, 3), index=index,

,→columns=columns)

In [293]: df
Out[293]:
cols A B C
rows
0 -1.539526 0.083551 1.819217
1 -0.556258 -1.013751 0.804958
2 1.138849 -0.913212 2.047493
3 -0.894783 1.059103 -0.605857
4 -1.096832 0.217643 2.122047

In [294]: df['A']
Out[294]:
rows
0 -1.539526
1 -0.556258
2 1.138849

354 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

3 -0.894783
4 -1.096832
Name: A, dtype: float64

Setting metadata

Indexes are mostly immutable, but it is possible to set and change their metadata, like the index name (or, for
MultiIndex, levels and codes).
You can use the rename, set_names, set_levels, and set_codes to set these attributes directly. They default
to returning a copy; however, you can specify inplace=True to have the data change in place.
See Advanced Indexing for usage of MultiIndexes.
In [295]: ind = pd.Index([1, 2, 3])

In [296]: ind.rename("apple")
Out[296]: Int64Index([1, 2, 3], dtype='int64', name='apple')

In [297]: ind
Out[297]: Int64Index([1, 2, 3], dtype='int64')

In [298]: ind.set_names(["apple"], inplace=True)

In [299]: ind.name = "bob"

In [300]: ind
Out[300]: Int64Index([1, 2, 3], dtype='int64', name='bob')
set_names, set_levels, and set_codes also take an optional level argument
In [301]: index = pd.MultiIndex.from_product([range(3), ['one', 'two']],
,→names=['first', 'second'])

In [302]: index
Out[302]:
MultiIndex([(0, 'one'),
(0, 'two'),
(1, 'one'),
(1, 'two'),
(2, 'one'),
(2, 'two')],
names=['first', 'second'])

In [303]: index.levels[1]
Out[303]: Index(['one', 'two'], dtype='object', name='second')

In [304]: index.set_levels(["a", "b"], level=1)

Out[304]:
MultiIndex([(0, 'a'),
(0, 'b'),
(1, 'a'),
(1, 'b'),
(2, 'a'),
(2, 'b')],
names=['first', 'second'])

4.2. Indexing and selecting data 355

pandas: powerful Python data analysis toolkit, Release 0.25.3

Set operations on Index objects

The two main operations are union (|) and intersection (&). These can be directly called as instance
methods or used via overloaded operators. Difference is provided via the .difference() method.
In [305]: a = pd.Index(['c', 'b', 'a'])

In [306]: b = pd.Index(['c', 'e', 'd'])

In [307]: a | b
Out[307]: Index(['a', 'b', 'c', 'd', 'e'], dtype='object')

In [308]: a & b
Out[308]: Index(['c'], dtype='object')

In [309]: a.difference(b)
Out[309]: Index(['a', 'b'], dtype='object')
Also available is the symmetric_difference (^) operation, which returns elements that appear in either idx1
or idx2, but not in both. This is equivalent to the Index created by idx1.difference(idx2).union(idx2.
difference(idx1)), with duplicates dropped.
In [310]: idx1 = pd.Index([1, 2, 3, 4])

In [311]: idx2 = pd.Index([2, 3, 4, 5])

In [312]: idx1.symmetric_difference(idx2)
Out[312]: Int64Index([1, 5], dtype='int64')

In [313]: idx1 ^ idx2

Out[313]: Int64Index([1, 5], dtype='int64')

Note: The resulting index from a set operation will be sorted in ascending order.

When performing Index.union() between indexes with different dtypes, the indexes must be cast to a common
dtype. Typically, though not always, this is object dtype. The exception is when performing a union between integer
and float data. In this case, the integer values are converted to float

In [314]: idx1 = pd.Index([0, 1, 2])

In [315]: idx2 = pd.Index([0.5, 1.5])

In [316]: idx1 | idx2

Out[316]: Float64Index([0.0, 0.5, 1.0, 1.5, 2.0], dtype='float64')

Missing values

Important: Even though Index can hold missing values (NaN), it should be avoided if you do not want any
unexpected results. For example, some operations exclude missing values implicitly.

Index.fillna fills missing values with specified scalar value.

In [317]: idx1 = pd.Index([1, np.nan, 3, 4])

356 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [318]: idx1
Out[318]: Float64Index([1.0, nan, 3.0, 4.0], dtype='float64')

In [319]: idx1.fillna(2)
Out[319]: Float64Index([1.0, 2.0, 3.0, 4.0], dtype='float64')

In [320]: idx2 = pd.DatetimeIndex([pd.Timestamp('2011-01-01'),

.....: pd.NaT,
.....: pd.Timestamp('2011-01-03')])
.....:

In [321]: idx2
Out[321]: DatetimeIndex(['2011-01-01', 'NaT', '2011-01-03'],
,→dtype='datetime64[ns]', freq=None)

In [322]: idx2.fillna(pd.Timestamp('2011-01-02'))
Out[322]: DatetimeIndex(['2011-01-01', '2011-01-02', '2011-01-03'],
,→dtype='datetime64[ns]', freq=None)

4.2.21 Set / reset index

Occasionally you will load or create a data set into a DataFrame and want to add an index after youve already done
so. There are a couple of different ways.

Set an index

DataFrame has a set_index() method which takes a column name (for a regular Index) or a list of column names
(for a MultiIndex). To create a new, re-indexed DataFrame:
In [323]: data
Out[323]:
a b c d
0 bar one z 1.0
1 bar two y 2.0
2 foo one x 3.0
3 foo two w 4.0

In [324]: indexed1 = data.set_index('c')

In [325]: indexed1
Out[325]:
a b d
c
z bar one 1.0
y bar two 2.0
x foo one 3.0
w foo two 4.0

In [326]: indexed2 = data.set_index(['a', 'b'])

In [327]: indexed2
Out[327]:
c d
(continues on next page)

4.2. Indexing and selecting data 357

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

a b
bar one z 1.0
two y 2.0
foo one x 3.0
two w 4.0

The append keyword option allow you to keep the existing index and append the given columns to a MultiIndex:

In [328]: frame = data.set_index('c', drop=False)

In [329]: frame = frame.set_index(['a', 'b'], append=True)

In [330]: frame
Out[330]:
c d
c a b
z bar one z 1.0
y bar two y 2.0
x foo one x 3.0
w foo two w 4.0

Other options in set_index allow you not drop the index columns or to add the index in-place (without creating a
new object):

In [331]: data.set_index('c', drop=False)

Out[331]:
a b c d
c
z bar one z 1.0
y bar two y 2.0
x foo one x 3.0
w foo two w 4.0

In [332]: data.set_index(['a', 'b'], inplace=True)

In [333]: data
Out[333]:
c d
a b
bar one z 1.0
two y 2.0
foo one x 3.0
two w 4.0

Reset the index

As a convenience, there is a new function on DataFrame called reset_index() which transfers the index values
into the DataFrames columns and sets a simple integer index. This is the inverse operation of set_index().
In [334]: data
Out[334]:
c d
a b
bar one z 1.0
two y 2.0

358 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

foo one x 3.0

two w 4.0

In [335]: data.reset_index()
Out[335]:
a b c d
0 bar one z 1.0
1 bar two y 2.0
2 foo one x 3.0
3 foo two w 4.0
The output is more similar to a SQL table or a record array. The names for the columns derived from the index are the
ones stored in the names attribute.
You can use the level keyword to remove only a portion of the index:
In [336]: frame
Out[336]:
c d
c a b
z bar one z 1.0
y bar two y 2.0
x foo one x 3.0
w foo two w 4.0

In [337]: frame.reset_index(level=1)
Out[337]:
a c d
c b
z one bar z 1.0
y two bar y 2.0
x one foo x 3.0
w two foo w 4.0
reset_index takes an optional parameter drop which if true simply discards the index, instead of putting index
values in the DataFrames columns.

Adding an ad hoc index

If you create an index yourself, you can just assign it to the index field:
data.index = index

4.2.22 Returning a view versus a copy

When setting values in a pandas object, care must be taken to avoid what is called chained indexing. Here is an
example.
In [338]: dfmi = pd.DataFrame([list('abcd'),
.....: list('efgh'),
.....: list('ijkl'),
.....: list('mnop')],
.....: columns=pd.MultiIndex.from_product([['one', 'two'],
.....: ['first', 'second
,→']]))
(continues on next page)

4.2. Indexing and selecting data 359

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

.....:

In [339]: dfmi
Out[339]:
one two
first second first second
0 a b c d
1 e f g h
2 i j k l
3 m n o p

Compare these two access methods:

In [340]: dfmi['one']['second']
Out[340]:
0 b
1 f
2 j
3 n
Name: second, dtype: object

In [341]: dfmi.loc[:, ('one', 'second')]

Out[341]:
0 b
1 f
2 j
3 n
Name: (one, second), dtype: object

These both yield the same results, so which should you use? It is instructive to understand the order of operations on
these and why method 2 (.loc) is much preferred over method 1 (chained []).
dfmi['one'] selects the first level of the columns and returns a DataFrame that is singly-indexed. Then an-
other Python operation dfmi_with_one['second'] selects the series indexed by 'second'. This is indicated
by the variable dfmi_with_one because pandas sees these operations as separate events. e.g. separate calls to
__getitem__, so it has to treat them as linear operations, they happen one after another.
Contrast this to df.loc[:,('one','second')] which passes a nested tuple of (slice(None),('one',
'second')) to a single call to __getitem__. This allows pandas to deal with this as a single entity. Furthermore
this order of operations can be significantly faster, and allows one to index both axes if so desired.

Why does assignment fail when using chained indexing?

The problem in the previous section is just a performance issue. Whats up with the SettingWithCopy warning?
We dont usually throw warnings around when you do something that might cost a few extra milliseconds!
But it turns out that assigning to the product of chained indexing has inherently unpredictable results. To see this,
think about how the Python interpreter executes this code:

dfmi.loc[:, ('one', 'second')] = value

# becomes
dfmi.loc.__setitem__((slice(None), ('one', 'second')), value)

But this code is handled differently:

360 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

dfmi['one']['second'] = value
# becomes
dfmi.__getitem__('one').__setitem__('second', value)

See that __getitem__ in there? Outside of simple cases, its very hard to predict whether it will return a view or a
copy (it depends on the memory layout of the array, about which pandas makes no guarantees), and therefore whether
the __setitem__ will modify dfmi or a temporary object that gets thrown out immediately afterward. Thats what
SettingWithCopy is warning you about!

Note: You may be wondering whether we should be concerned about the loc property in the first example. But
dfmi.loc is guaranteed to be dfmi itself with modified indexing behavior, so dfmi.loc.__getitem__ /
dfmi.loc.__setitem__ operate on dfmi directly. Of course, dfmi.loc.__getitem__(idx) may be
a view or a copy of dfmi.

Sometimes a SettingWithCopy warning will arise at times when theres no obvious chained indexing going on.
These are the bugs that SettingWithCopy is designed to catch! Pandas is probably trying to warn you that youve
done this:

def do_something(df):
foo = df[['bar', 'baz']] # Is foo a view? A copy? Nobody knows!
# ... many lines here ...
# We don't know whether this will modify df or not!
foo['quux'] = value
return foo

Yikes!

Evaluation order matters

When you use chained indexing, the order and type of the indexing operation partially determine whether the result is
a slice into the original object, or a copy of the slice.
Pandas has the SettingWithCopyWarning because assigning to a copy of a slice is frequently not intentional,
but a mistake caused by chained indexing returning a copy where a slice was expected.
If you would like pandas to be more or less trusting about assignment to a chained indexing expression, you can set
the option mode.chained_assignment to one of these values:
• 'warn', the default, means a SettingWithCopyWarning is printed.
• 'raise' means pandas will raise a SettingWithCopyException you have to deal with.
• None will suppress the warnings entirely.

In [342]: dfb = pd.DataFrame({'a': ['one', 'one', 'two',

.....: 'three', 'two', 'one', 'six'],
.....: 'c': np.arange(7)})
.....:

# This will show the SettingWithCopyWarning

# but the frame values will be set
In [343]: dfb['c'][dfb.a.str.startswith('o')] = 42

This however is operating on a copy and will not work.

4.2. Indexing and selecting data 361

pandas: powerful Python data analysis toolkit, Release 0.25.3

>>> pd.set_option('mode.chained_assignment','warn')
>>> dfb[dfb.a.str.startswith('o')]['c'] = 42
Traceback (most recent call last)
...
SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_index,col_indexer] = value instead

A chained assignment can also crop up in setting in a mixed dtype frame.

Note: These setting rules apply to all of .loc/.iloc.

This is the correct access method:

In [344]: dfc = pd.DataFrame({'A': ['aaa', 'bbb', 'ccc'], 'B': [1, 2, 3]})

In [345]: dfc.loc[0, 'A'] = 11

In [346]: dfc
Out[346]:
A B
0 11 1
1 bbb 2
2 ccc 3

This can work at times, but it is not guaranteed to, and therefore should be avoided:

In [347]: dfc = dfc.copy()

In [348]: dfc['A'][0] = 111

In [349]: dfc
Out[349]:
A B
0 111 1
1 bbb 2
2 ccc 3

This will not work at all, and so should be avoided:

>>> pd.set_option('mode.chained_assignment','raise')
>>> dfc.loc[0]['A'] = 1111
Traceback (most recent call last)
...
SettingWithCopyException:
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_index,col_indexer] = value instead

Warning: The chained assignment warnings / exceptions are aiming to inform the user of a possibly invalid
assignment. There may be false positives; situations where a chained assignment is inadvertently reported.

{{ header }}

362 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

4.3 MultiIndex / advanced indexing

This section covers indexing with a MultiIndex and other advanced indexing features.
See the Indexing and Selecting Data for general indexing documentation.

Warning: Whether a copy or a reference is returned for a setting operation may depend on the context. This is
sometimes called chained assignment and should be avoided. See Returning a View versus Copy.

See the cookbook for some advanced strategies.

4.3.1 Hierarchical indexing (MultiIndex)

Hierarchical / Multi-level indexing is very exciting as it opens the door to some quite sophisticated data analysis and
manipulation, especially for working with higher dimensional data. In essence, it enables you to store and manipulate
data with an arbitrary number of dimensions in lower dimensional data structures like Series (1d) and DataFrame
(2d).
In this section, we will show what exactly we mean by hierarchical indexing and how it integrates with all of the
pandas indexing functionality described above and in prior sections. Later, when discussing group by and pivoting and
reshaping data, well show non-trivial applications to illustrate how it aids in structuring data for analysis.
See the cookbook for some advanced strategies.
Changed in version 0.24.0: MultiIndex.labels has been renamed to MultiIndex.codes and
MultiIndex.set_labels to MultiIndex.set_codes.

Creating a MultiIndex (hierarchical index) object

The MultiIndex object is the hierarchical analogue of the standard Index object which typically stores the axis
labels in pandas objects. You can think of MultiIndex as an array of tuples where each tuple is unique. A
MultiIndex can be created from a list of arrays (using MultiIndex.from_arrays()), an array of tuples
(using MultiIndex.from_tuples()), a crossed set of iterables (using MultiIndex.from_product()),
or a DataFrame (using MultiIndex.from_frame()). The Index constructor will attempt to return a
MultiIndex when it is passed a list of tuples. The following examples demonstrate different ways to initialize
MultiIndexes.

In [1]: arrays = [['bar', 'bar', 'baz', 'baz', 'foo', 'foo', 'qux', 'qux'],
...: ['one', 'two', 'one', 'two', 'one', 'two', 'one', 'two']]
...:

In [2]: tuples = list(zip(*arrays))

In [3]: tuples
Out[3]:
[('bar', 'one'),
('bar', 'two'),
('baz', 'one'),
('baz', 'two'),
('foo', 'one'),
('foo', 'two'),
('qux', 'one'),
('qux', 'two')]
(continues on next page)

4.3. MultiIndex / advanced indexing 363

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [4]: index = pd.MultiIndex.from_tuples(tuples, names=['first', 'second'])

In [5]: index
Out[5]:
MultiIndex([('bar', 'one'),
('bar', 'two'),
('baz', 'one'),
('baz', 'two'),
('foo', 'one'),
('foo', 'two'),
('qux', 'one'),
('qux', 'two')],
names=['first', 'second'])

In [6]: s = pd.Series(np.random.randn(8), index=index)

In [7]: s
Out[7]:
first second
bar one -1.386024
two -0.148663
baz one -2.607819
two -1.005003
foo one 1.204157
two 0.866884
qux one -0.284878
two 1.160953
dtype: float64

When you want every pairing of the elements in two iterables, it can be easier to use the MultiIndex.
from_product() method:
In [8]: iterables = [['bar', 'baz', 'foo', 'qux'], ['one', 'two']]

In [9]: pd.MultiIndex.from_product(iterables, names=['first', 'second'])

Out[9]:
MultiIndex([('bar', 'one'),
('bar', 'two'),
('baz', 'one'),
('baz', 'two'),
('foo', 'one'),
('foo', 'two'),
('qux', 'one'),
('qux', 'two')],
names=['first', 'second'])

You can also construct a MultiIndex from a DataFrame directly, using the method MultiIndex.
from_frame(). This is a complementary method to MultiIndex.to_frame().
New in version 0.24.0.
In [10]: df = pd.DataFrame([['bar', 'one'], ['bar', 'two'],
....: ['foo', 'one'], ['foo', 'two']],
....: columns=['first', 'second'])
....:

(continues on next page)

364 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [11]: pd.MultiIndex.from_frame(df)
Out[11]:
MultiIndex([('bar', 'one'),
('bar', 'two'),
('foo', 'one'),
('foo', 'two')],
names=['first', 'second'])

As a convenience, you can pass a list of arrays directly into Series or DataFrame to construct a MultiIndex
automatically:

In [12]: arrays = [np.array(['bar', 'bar', 'baz', 'baz', 'foo', 'foo', 'qux', 'qux']),
....: np.array(['one', 'two', 'one', 'two', 'one', 'two', 'one', 'two'])]
....:

In [13]: s = pd.Series(np.random.randn(8), index=arrays)

In [14]: s
Out[14]:
bar one 1.678004
two 0.957360
baz one 0.866120
two -0.348147
foo one 0.234443
two 1.279886
qux one -0.079828
two -0.398588
dtype: float64

In [15]: df = pd.DataFrame(np.random.randn(8, 4), index=arrays)

In [16]: df
Out[16]:
0 1 2 3
bar one -0.479845 -1.066579 0.304694 -0.092407
two -1.470428 -1.023536 0.513919 0.429166
baz one -0.869518 0.054924 -0.379843 -0.815517
two -1.025934 -0.748570 0.938954 1.472919
foo one 1.725907 1.384912 0.795282 -0.576135
two -1.383644 -0.854899 -0.888839 -1.298461
qux one -0.227461 2.302300 1.989458 0.867323
two -0.785156 -0.647147 -0.146113 -0.754742

All of the MultiIndex constructors accept a names argument which stores string names for the levels themselves.
If no names are provided, None will be assigned:

In [17]: df.index.names
Out[17]: FrozenList([None, None])

This index can back any axis of a pandas object, and the number of levels of the index is up to you:
In [18]: df = pd.DataFrame(np.random.randn(3, 8), index=['A', 'B', 'C'],
,→columns=index)

In [19]: df
Out[19]:
first bar baz foo qux

4.3. MultiIndex / advanced indexing 365

pandas: powerful Python data analysis toolkit, Release 0.25.3

second one two one two one two one

,→ two
A -0.134175 0.225218 -0.338687 -1.142207 -0.981045 -0.474525 0.857580 -
,→0.161414

B 1.188426 -0.286643 -1.643376 -0.739265 0.841529 -0.443695 -2.770279 -

,→0.947253

C 0.866885 0.954773 2.018904 -0.650307 -0.347872 1.068422 -0.721383

,→1.184454

In [20]: pd.DataFrame(np.random.randn(6, 6), index=index[:6],

,→columns=index[:6])

Out[20]:
first bar baz foo
second one two one two one two
first second
bar one 2.254090 0.665628 0.209523 -1.872699 0.277580 0.842965
two -0.690984 -0.336425 0.688883 0.848639 0.005431 0.997845
baz one 0.386518 0.763158 -0.864808 -1.670655 -1.049607 0.028482
two -1.761586 0.487875 0.131290 -0.262495 -0.324123 -0.108842
foo one 2.343179 0.747274 -0.456691 -0.014325 -0.397298 -0.242768
two 1.318750 -1.119168 1.061544 -2.664141 -0.600815 0.500161
Weve sparsified the higher levels of the indexes to make the console output a bit easier on the eyes. Note that how the
index is displayed can be controlled using the multi_sparse option in pandas.set_options():

In [21]: with pd.option_context('display.multi_sparse', False):

....: df
....:

Its worth keeping in mind that theres nothing preventing you from using tuples as atomic labels on an axis:

In [22]: pd.Series(np.random.randn(8), index=tuples)

Out[22]:
(bar, one) -1.253707
(bar, two) -1.439613
(baz, one) 0.024850
(baz, two) 0.166637
(foo, one) -0.962691
(foo, two) -1.527951
(qux, one) 0.456415
(qux, two) 0.323368
dtype: float64

The reason that the MultiIndex matters is that it can allow you to do grouping, selection, and reshaping operations
as we will describe below and in subsequent areas of the documentation. As you will see in later sections, you can find
yourself working with hierarchically-indexed data without creating a MultiIndex explicitly yourself. However,
when loading data from a file, you may wish to generate your own MultiIndex when preparing the data set.

Reconstructing the level labels

The method get_level_values() will return a vector of the labels for each location at a particular level:
In [23]: index.get_level_values(0)
Out[23]: Index(['bar', 'bar', 'baz', 'baz', 'foo', 'foo', 'qux', 'qux'],
,→dtype='object', name='first')

366 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [24]: index.get_level_values('second')
Out[24]: Index(['one', 'two', 'one', 'two', 'one', 'two', 'one', 'two'],
,→dtype='object', name='second')

Basic indexing on axis with MultiIndex

One of the important features of hierarchical indexing is that you can select data by a partial label identifying a
subgroup in the data. Partial selection drops levels of the hierarchical index in the result in a completely analogous
way to selecting a column in a regular DataFrame:
In [25]: df['bar']
Out[25]:
second one two
A -0.134175 0.225218
B 1.188426 -0.286643
C 0.866885 0.954773

In [26]: df['bar', 'one']

Out[26]:
A -0.134175
B 1.188426
C 0.866885
Name: (bar, one), dtype: float64

In [27]: df['bar']['one']
Out[27]:
A -0.134175
B 1.188426
C 0.866885
Name: one, dtype: float64

In [28]: s['qux']
Out[28]:
one -0.079828
two -0.398588
dtype: float64
See Cross-section with hierarchical index for how to select on a deeper level.

Defined levels

The MultiIndex keeps all the defined levels of an index, even if they are not actually used. When slicing an index,
you may notice this. For example:
In [29]: df.columns.levels # original MultiIndex
Out[29]: FrozenList([['bar', 'baz', 'foo', 'qux'], ['one', 'two']])

In [30]: df[['foo','qux']].columns.levels # sliced

Out[30]: FrozenList([['bar', 'baz', 'foo', 'qux'], ['one', 'two']])
This is done to avoid a recomputation of the levels in order to make slicing highly performant. If you want to see only
the used levels, you can use the get_level_values() method.
In [31]: df[['foo', 'qux']].columns.to_numpy()
Out[31]:

4.3. MultiIndex / advanced indexing 367

pandas: powerful Python data analysis toolkit, Release 0.25.3

array([('foo', 'one'), ('foo', 'two'), ('qux', 'one'), ('qux', 'two')],

dtype=object)

# for a specific level

In [32]: df[['foo', 'qux']].columns.get_level_values(0)
Out[32]: Index(['foo', 'foo', 'qux', 'qux'], dtype='object', name='first')
To reconstruct the MultiIndex with only the used levels, the remove_unused_levels() method may be used.
New in version 0.20.0.

In [33]: new_mi = df[['foo', 'qux']].columns.remove_unused_levels()

In [34]: new_mi.levels
Out[34]: FrozenList([['foo', 'qux'], ['one', 'two']])

Data alignment and using reindex

Operations between differently-indexed objects having MultiIndex on the axes will work as you expect; data
alignment will work the same as an Index of tuples:
In [35]: s + s[:-2]
Out[35]:
bar one 3.356008
two 1.914720
baz one 1.732240
two -0.696293
foo one 0.468886
two 2.559772
qux one NaN
two NaN
dtype: float64

In [36]: s + s[::2]
Out[36]:
bar one 3.356008
two NaN
baz one 1.732240
two NaN
foo one 0.468886
two NaN
qux one -0.159656
two NaN
dtype: float64
The reindex() method of Series/DataFrames can be called with another MultiIndex, or even a list or array
of tuples:
In [37]: s.reindex(index[:3])
Out[37]:
first second
bar one 1.678004
two 0.957360
baz one 0.866120
dtype: float64

368 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [38]: s.reindex([('foo', 'two'), ('bar', 'one'), ('qux', 'one'), ('baz',

,→'one')])

Out[38]:
foo two 1.279886
bar one 1.678004
qux one -0.079828
baz one 0.866120
dtype: float64

4.3.2 Advanced indexing with hierarchical index

Syntactically integrating MultiIndex in advanced indexing with .loc is a bit challenging, but weve made every
effort to do so. In general, MultiIndex keys take the form of tuples. For example, the following works as you would
expect:
In [39]: df = df.T

In [40]: df
Out[40]:
A B C
first second
bar one -0.134175 1.188426 0.866885
two 0.225218 -0.286643 0.954773
baz one -0.338687 -1.643376 2.018904
two -1.142207 -0.739265 -0.650307
foo one -0.981045 0.841529 -0.347872
two -0.474525 -0.443695 1.068422
qux one 0.857580 -2.770279 -0.721383
two -0.161414 -0.947253 1.184454

In [41]: df.loc[('bar', 'two')]

Out[41]:
A 0.225218
B -0.286643
C 0.954773
Name: (bar, two), dtype: float64
Note that df.loc['bar', 'two'] would also work in this example, but this shorthand notation can lead to
ambiguity in general.
If you also want to index a specific column with .loc, you must use a tuple like this:

In [42]: df.loc[('bar', 'two'), 'A']

Out[42]: 0.22521771909947486

You dont have to specify all levels of the MultiIndex by passing only the first elements of the tuple. For example,
you can use partial indexing to get all elements with bar in the first level as follows:
df.loc[bar]
This is a shortcut for the slightly more verbose notation df.loc[('bar',),] (equivalent to df.loc['bar',]
in this example).
Partial slicing also works quite nicely.

4.3. MultiIndex / advanced indexing 369

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [43]: df.loc['baz':'foo']
Out[43]:
A B C
first second
baz one -0.338687 -1.643376 2.018904
two -1.142207 -0.739265 -0.650307
foo one -0.981045 0.841529 -0.347872
two -0.474525 -0.443695 1.068422

You can slice with a range of values, by providing a slice of tuples.

In [44]: df.loc[('baz', 'two'):('qux', 'one')]
Out[44]:
A B C
first second
baz two -1.142207 -0.739265 -0.650307
foo one -0.981045 0.841529 -0.347872
two -0.474525 -0.443695 1.068422
qux one 0.857580 -2.770279 -0.721383

In [45]: df.loc[('baz', 'two'):'foo']

Out[45]:
A B C
first second
baz two -1.142207 -0.739265 -0.650307
foo one -0.981045 0.841529 -0.347872
two -0.474525 -0.443695 1.068422
Passing a list of labels or tuples works similar to reindexing:

In [46]: df.loc[[('bar', 'two'), ('qux', 'one')]]

Out[46]:
A B C
first second
bar two 0.225218 -0.286643 0.954773
qux one 0.857580 -2.770279 -0.721383

Note: It is important to note that tuples and lists are not treated identically in pandas when it comes to indexing.
Whereas a tuple is interpreted as one multi-level key, a list is used to specify several keys. Or in other words, tuples
go horizontally (traversing levels), lists go vertically (scanning levels).

Importantly, a list of tuples indexes several complete MultiIndex keys, whereas a tuple of lists refer to several
values within a level:
In [47]: s = pd.Series([1, 2, 3, 4, 5, 6],
....: index=pd.MultiIndex.from_product([["A", "B"], ["c",
,→"d", "e"]]))

....:

In [48]: s.loc[[("A", "c"), ("B", "d")]] # list of tuples

Out[48]:
A c 1
B d 5
dtype: int64

370 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [49]: s.loc[(["A", "B"], ["c", "d"])] # tuple of lists

Out[49]:
A c 1
d 2
B c 4
d 5
dtype: int64

Using slicers

You can slice a MultiIndex by providing multiple indexers.

You can provide any of the selectors as if you are indexing by label, see Selection by Label, including slices, lists of
labels, labels, and boolean indexers.
You can use slice(None) to select all the contents of that level. You do not need to specify all the deeper levels,
they will be implied as slice(None).
As usual, both sides of the slicers are included as this is label indexing.

Warning: You should specify all axes in the .loc specifier, meaning the indexer for the index and for the
columns. There are some ambiguous cases where the passed indexer could be mis-interpreted as indexing both
axes, rather than into say the MultiIndex for the rows.
You should do this:
df.loc[(slice('A1', 'A3'), ...), :] # noqa: E999

You should not do this:

df.loc[(slice('A1', 'A3'), ...)] # noqa: E999

In [50]: def mklbl(prefix, n):

....: return ["%s%s" % (prefix, i) for i in range(n)]
....:

In [51]: miindex = pd.MultiIndex.from_product([mklbl('A', 4),

....: mklbl('B', 2),
....: mklbl('C', 4),
....: mklbl('D', 2)])
....:

In [52]: micolumns = pd.MultiIndex.from_tuples([('a', 'foo'), ('a', 'bar'),

....: ('b', 'foo'), ('b', 'bah')],
....: names=['lvl0', 'lvl1'])
....:

In [53]: dfmi = pd.DataFrame(np.arange(len(miindex) * len(micolumns))

....: .reshape((len(miindex), len(micolumns))),
....: index=miindex,
....: columns=micolumns).sort_index().sort_index(axis=1)
....:

In [54]: dfmi
Out[54]:
lvl0 a b
(continues on next page)

4.3. MultiIndex / advanced indexing 371

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

lvl1 bar foo bah foo
A0 B0 C0 D0 1 0 3 2
D1 5 4 7 6
C1 D0 9 8 11 10
D1 13 12 15 14
C2 D0 17 16 19 18
... ... ... ... ...
A3 B1 C1 D1 237 236 239 238
C2 D0 241 240 243 242
D1 245 244 247 246
C3 D0 249 248 251 250
D1 253 252 255 254

[64 rows x 4 columns]

Basic MultiIndex slicing using slices, lists, and labels.

In [55]: dfmi.loc[(slice('A1', 'A3'), slice(None), ['C1', 'C3']), :]
Out[55]:
lvl0 a b
lvl1 bar foo bah foo
A1 B0 C1 D0 73 72 75 74
D1 77 76 79 78
C3 D0 89 88 91 90
D1 93 92 95 94
B1 C1 D0 105 104 107 106
D1 109 108 111 110
C3 D0 121 120 123 122
D1 125 124 127 126
A2 B0 C1 D0 137 136 139 138
D1 141 140 143 142
C3 D0 153 152 155 154
D1 157 156 159 158
B1 C1 D0 169 168 171 170
D1 173 172 175 174
C3 D0 185 184 187 186
D1 189 188 191 190
A3 B0 C1 D0 201 200 203 202
D1 205 204 207 206
C3 D0 217 216 219 218
D1 221 220 223 222
B1 C1 D0 233 232 235 234
D1 237 236 239 238
C3 D0 249 248 251 250
D1 253 252 255 254

You can use pandas.IndexSlice to facilitate a more natural syntax using :, rather than using slice(None).
In [56]: idx = pd.IndexSlice

In [57]: dfmi.loc[idx[:, :, ['C1', 'C3']], idx[:, 'foo']]

Out[57]:
lvl0 a b
lvl1 foo foo
A0 B0 C1 D0 8 10
D1 12 14
C3 D0 24 26
(continues on next page)

372 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

D1 28 30
B1 C1 D0 40 42
D1 44 46
C3 D0 56 58
D1 60 62
A1 B0 C1 D0 72 74
D1 76 78
C3 D0 88 90
D1 92 94
B1 C1 D0 104 106
D1 108 110
C3 D0 120 122
D1 124 126
A2 B0 C1 D0 136 138
D1 140 142
C3 D0 152 154
D1 156 158
B1 C1 D0 168 170
D1 172 174
C3 D0 184 186
D1 188 190
A3 B0 C1 D0 200 202
D1 204 206
C3 D0 216 218
D1 220 222
B1 C1 D0 232 234
D1 236 238
C3 D0 248 250
D1 252 254

It is possible to perform quite complicated selections using this method on multiple axes at the same time.
In [58]: dfmi.loc['A1', (slice(None), 'foo')]
Out[58]:
lvl0 a b
lvl1 foo foo
B0 C0 D0 64 66
D1 68 70
C1 D0 72 74
D1 76 78
C2 D0 80 82
D1 84 86
C3 D0 88 90
D1 92 94
B1 C0 D0 96 98
D1 100 102
C1 D0 104 106
D1 108 110
C2 D0 112 114
D1 116 118
C3 D0 120 122
D1 124 126

In [59]: dfmi.loc[idx[:, :, ['C1', 'C3']], idx[:, 'foo']]

Out[59]:
lvl0 a b

4.3. MultiIndex / advanced indexing 373

pandas: powerful Python data analysis toolkit, Release 0.25.3

lvl1 foo foo

A0 B0 C1 D0 8 10
D1 12 14
C3 D0 24 26
D1 28 30
B1 C1 D0 40 42
D1 44 46
C3 D0 56 58
D1 60 62
A1 B0 C1 D0 72 74
D1 76 78
C3 D0 88 90
D1 92 94
B1 C1 D0 104 106
D1 108 110
C3 D0 120 122
D1 124 126
A2 B0 C1 D0 136 138
D1 140 142
C3 D0 152 154
D1 156 158
B1 C1 D0 168 170
D1 172 174
C3 D0 184 186
D1 188 190
A3 B0 C1 D0 200 202
D1 204 206
C3 D0 216 218
D1 220 222
B1 C1 D0 232 234
D1 236 238
C3 D0 248 250
D1 252 254
Using a boolean indexer you can provide selection related to the values.

In [60]: mask = dfmi[('a', 'foo')] > 200

In [61]: dfmi.loc[idx[mask, :, ['C1', 'C3']], idx[:, 'foo']]

Out[61]:
lvl0 a b
lvl1 foo foo
A3 B0 C1 D1 204 206
C3 D0 216 218
D1 220 222
B1 C1 D0 232 234
D1 236 238
C3 D0 248 250
D1 252 254

You can also specify the axis argument to .loc to interpret the passed slicers on a single axis.

In [62]: dfmi.loc(axis=0)[:, :, ['C1', 'C3']]

Out[62]:
lvl0 a b
(continues on next page)

374 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

lvl1 bar foo bah foo
A0 B0 C1 D0 9 8 11 10
D1 13 12 15 14
C3 D0 25 24 27 26
D1 29 28 31 30
B1 C1 D0 41 40 43 42
D1 45 44 47 46
C3 D0 57 56 59 58
D1 61 60 63 62
A1 B0 C1 D0 73 72 75 74
D1 77 76 79 78
C3 D0 89 88 91 90
D1 93 92 95 94
B1 C1 D0 105 104 107 106
D1 109 108 111 110
C3 D0 121 120 123 122
D1 125 124 127 126
A2 B0 C1 D0 137 136 139 138
D1 141 140 143 142
C3 D0 153 152 155 154
D1 157 156 159 158
B1 C1 D0 169 168 171 170
D1 173 172 175 174
C3 D0 185 184 187 186
D1 189 188 191 190
A3 B0 C1 D0 201 200 203 202
D1 205 204 207 206
C3 D0 217 216 219 218
D1 221 220 223 222
B1 C1 D0 233 232 235 234
D1 237 236 239 238
C3 D0 249 248 251 250
D1 253 252 255 254

Furthermore, you can set the values using the following methods.

In [63]: df2 = dfmi.copy()

In [64]: df2.loc(axis=0)[:, :, ['C1', 'C3']] = -10

In [65]: df2
Out[65]:
lvl0 a b
lvl1 bar foo bah foo
A0 B0 C0 D0 1 0 3 2
D1 5 4 7 6
C1 D0 -10 -10 -10 -10
D1 -10 -10 -10 -10
C2 D0 17 16 19 18
... ... ... ... ...
A3 B1 C1 D1 -10 -10 -10 -10
C2 D0 241 240 243 242
D1 245 244 247 246
C3 D0 -10 -10 -10 -10
D1 -10 -10 -10 -10

[64 rows x 4 columns]

4.3. MultiIndex / advanced indexing 375

pandas: powerful Python data analysis toolkit, Release 0.25.3

You can use a right-hand-side of an alignable object as well.

In [66]: df2 = dfmi.copy()

In [67]: df2.loc[idx[:, :, ['C1', 'C3']], :] = df2 * 1000

In [68]: df2
Out[68]:
lvl0 a b
lvl1 bar foo bah foo
A0 B0 C0 D0 1 0 3 2
D1 5 4 7 6
C1 D0 9000 8000 11000 10000
D1 13000 12000 15000 14000
C2 D0 17 16 19 18
... ... ... ... ...
A3 B1 C1 D1 237000 236000 239000 238000
C2 D0 241 240 243 242
D1 245 244 247 246
C3 D0 249000 248000 251000 250000
D1 253000 252000 255000 254000

[64 rows x 4 columns]

Cross-section

The xs() method of DataFrame additionally takes a level argument to make selecting data at a particular level of a
MultiIndex easier.
In [69]: df
Out[69]:
A B C
first second
bar one -0.134175 1.188426 0.866885
two 0.225218 -0.286643 0.954773
baz one -0.338687 -1.643376 2.018904
two -1.142207 -0.739265 -0.650307
foo one -0.981045 0.841529 -0.347872
two -0.474525 -0.443695 1.068422
qux one 0.857580 -2.770279 -0.721383
two -0.161414 -0.947253 1.184454

In [70]: df.xs('one', level='second')

Out[70]:
A B C
first
bar -0.134175 1.188426 0.866885
baz -0.338687 -1.643376 2.018904
foo -0.981045 0.841529 -0.347872
qux 0.857580 -2.770279 -0.721383

# using the slicers

In [71]: df.loc[(slice(None), 'one'), :]
Out[71]:
A B C
(continues on next page)

376 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

first second
bar one -0.134175 1.188426 0.866885
baz one -0.338687 -1.643376 2.018904
foo one -0.981045 0.841529 -0.347872
qux one 0.857580 -2.770279 -0.721383

You can also select on the columns with xs, by providing the axis argument.

In [72]: df = df.T

In [73]: df.xs('one', level='second', axis=1)

Out[73]:
first bar baz foo qux
A -0.134175 -0.338687 -0.981045 0.857580
B 1.188426 -1.643376 0.841529 -2.770279
C 0.866885 2.018904 -0.347872 -0.721383

# using the slicers

In [74]: df.loc[:, (slice(None), 'one')]
Out[74]:
first bar baz foo qux
second one one one one
A -0.134175 -0.338687 -0.981045 0.857580
B 1.188426 -1.643376 0.841529 -2.770279
C 0.866885 2.018904 -0.347872 -0.721383

xs also allows selection with multiple keys.

In [75]: df.xs(('one', 'bar'), level=('second', 'first'), axis=1)

Out[75]:
first bar
second one
A -0.134175
B 1.188426
C 0.866885

# using the slicers

In [76]: df.loc[:, ('bar', 'one')]
Out[76]:
A -0.134175
B 1.188426
C 0.866885
Name: (bar, one), dtype: float64

You can pass drop_level=False to xs to retain the level that was selected.

In [77]: df.xs('one', level='second', axis=1, drop_level=False)

Out[77]:
first bar baz foo qux
second one one one one
A -0.134175 -0.338687 -0.981045 0.857580
B 1.188426 -1.643376 0.841529 -2.770279
C 0.866885 2.018904 -0.347872 -0.721383

Compare the above with the result using drop_level=True (the default value).

4.3. MultiIndex / advanced indexing 377

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [78]: df.xs('one', level='second', axis=1, drop_level=True)

Out[78]:
first bar baz foo qux
A -0.134175 -0.338687 -0.981045 0.857580
B 1.188426 -1.643376 0.841529 -2.770279
C 0.866885 2.018904 -0.347872 -0.721383

Advanced reindexing and alignment

Using the parameter level in the reindex() and align() methods of pandas objects is useful to broadcast
values across a level. For instance:
In [79]: midx = pd.MultiIndex(levels=[['zero', 'one'], ['x', 'y']],
....: codes=[[1, 1, 0, 0], [1, 0, 1, 0]])
....:

In [80]: df = pd.DataFrame(np.random.randn(4, 2), index=midx)

In [81]: df
Out[81]:
0 1
one y 0.229891 -0.556291
x 0.790519 -0.207853
zero y -0.182679 0.496230
x 0.177932 1.885439

In [82]: df2 = df.mean(level=0)

In [83]: df2
Out[83]:
0 1
one 0.510205 -0.382072
zero -0.002373 1.190835

In [84]: df2.reindex(df.index, level=0)

Out[84]:
0 1
one y 0.510205 -0.382072
x 0.510205 -0.382072
zero y -0.002373 1.190835
x -0.002373 1.190835

# aligning
In [85]: df_aligned, df2_aligned = df.align(df2, level=0)

In [86]: df_aligned
Out[86]:
0 1
one y 0.229891 -0.556291
x 0.790519 -0.207853
zero y -0.182679 0.496230
x 0.177932 1.885439

378 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [87]: df2_aligned
Out[87]:
0 1
one y 0.510205 -0.382072
x 0.510205 -0.382072
zero y -0.002373 1.190835
x -0.002373 1.190835

Swapping levels with swaplevel

The swaplevel() method can switch the order of two levels:

In [88]: df[:5]
Out[88]:
0 1
one y 0.229891 -0.556291
x 0.790519 -0.207853
zero y -0.182679 0.496230
x 0.177932 1.885439

In [89]: df[:5].swaplevel(0, 1, axis=0)

Out[89]:
0 1
y one 0.229891 -0.556291
x one 0.790519 -0.207853
y zero -0.182679 0.496230
x zero 0.177932 1.885439

Reordering levels with reorder_levels

The reorder_levels() method generalizes the swaplevel method, allowing you to permute the hierarchical
index levels in one step:

In [90]: df[:5].reorder_levels([1, 0], axis=0)

Out[90]:
0 1
y one 0.229891 -0.556291
x one 0.790519 -0.207853
y zero -0.182679 0.496230
x zero 0.177932 1.885439

Renaming names of an Index or MultiIndex

The rename() method is used to rename the labels of a MultiIndex, and is typically used to rename the columns
of a DataFrame. The columns argument of rename allows a dictionary to be specified that includes only the
columns you wish to rename.

In [91]: df.rename(columns={0: "col0", 1: "col1"})

Out[91]:
col0 col1
one y 0.229891 -0.556291
x 0.790519 -0.207853
(continues on next page)

4.3. MultiIndex / advanced indexing 379

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

zero y -0.182679 0.496230
x 0.177932 1.885439

This method can also be used to rename specific labels of the main index of the DataFrame.
In [92]: df.rename(index={"one": "two", "y": "z"})
Out[92]:
0 1
two z 0.229891 -0.556291
x 0.790519 -0.207853
zero z -0.182679 0.496230
x 0.177932 1.885439

The rename_axis() method is used to rename the name of a Index or MultiIndex. In particular, the names of
the levels of a MultiIndex can be specified, which is useful if reset_index() is later used to move the values
from the MultiIndex to a column.
In [93]: df.rename_axis(index=['abc', 'def'])
Out[93]:
0 1
abc def
one y 0.229891 -0.556291
x 0.790519 -0.207853
zero y -0.182679 0.496230
x 0.177932 1.885439

Note that the columns of a DataFrame are an index, so that using rename_axis with the columns argument
will change the name of that index.
In [94]: df.rename_axis(columns="Cols").columns
Out[94]: RangeIndex(start=0, stop=2, step=1, name='Cols')

Both rename and rename_axis support specifying a dictionary, Series or a mapping function to map la-
bels/names to new values.

4.3.3 Sorting a MultiIndex

For MultiIndex-ed objects to be indexed and sliced effectively, they need to be sorted. As with any index, you can
use sort_index().
In [95]: import random

In [96]: random.shuffle(tuples)

In [97]: s = pd.Series(np.random.randn(8), index=pd.MultiIndex.from_

,→tuples(tuples))

In [98]: s
Out[98]:
baz one 1.432295
qux two -0.999104
foo one 0.099556
bar one 0.189457
foo two 1.070405
baz two -0.015460

380 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

qux one 1.018696

bar two -0.966176
dtype: float64

In [99]: s.sort_index()
Out[99]:
bar one 0.189457
two -0.966176
baz one 1.432295
two -0.015460
foo one 0.099556
two 1.070405
qux one 1.018696
two -0.999104
dtype: float64

In [100]: s.sort_index(level=0)
Out[100]:
bar one 0.189457
two -0.966176
baz one 1.432295
two -0.015460
foo one 0.099556
two 1.070405
qux one 1.018696
two -0.999104
dtype: float64

In [101]: s.sort_index(level=1)
Out[101]:
bar one 0.189457
baz one 1.432295
foo one 0.099556
qux one 1.018696
bar two -0.966176
baz two -0.015460
foo two 1.070405
qux two -0.999104
dtype: float64
You may also pass a level name to sort_index if the MultiIndex levels are named.
In [102]: s.index.set_names(['L1', 'L2'], inplace=True)

In [103]: s.sort_index(level='L1')
Out[103]:
L1 L2
bar one 0.189457
two -0.966176
baz one 1.432295
two -0.015460
foo one 0.099556
two 1.070405
qux one 1.018696
two -0.999104

4.3. MultiIndex / advanced indexing 381

pandas: powerful Python data analysis toolkit, Release 0.25.3

dtype: float64

In [104]: s.sort_index(level='L2')
Out[104]:
L1 L2
bar one 0.189457
baz one 1.432295
foo one 0.099556
qux one 1.018696
bar two -0.966176
baz two -0.015460
foo two 1.070405
qux two -0.999104
dtype: float64
On higher dimensional objects, you can sort any of the other axes by level if they have a MultiIndex:

In [105]: df.T.sort_index(level=1, axis=1)

Out[105]:
one zero one zero
x x y y
0 0.790519 0.177932 0.229891 -0.182679
1 -0.207853 1.885439 -0.556291 0.496230

Indexing will work even if the data are not sorted, but will be rather inefficient (and show a PerformanceWarning).
It will also return a copy of the data rather than a view:

In [106]: dfm = pd.DataFrame({'jim': [0, 0, 1, 1],

.....: 'joe': ['x', 'x', 'z', 'y'],
.....: 'jolie': np.random.rand(4)})
.....:

In [107]: dfm = dfm.set_index(['jim', 'joe'])

In [108]: dfm
Out[108]:
jolie
jim joe
0 x 0.238982
x 0.824351
1 z 0.159509
y 0.076877

In [4]: dfm.loc[(1, 'z')]

PerformanceWarning: indexing past lexsort depth may impact performance.

Out[4]:
jolie
jim joe
1 z 0.64094

Furthermore, if you try to index something that is not fully lexsorted, this can raise:

In [5]: dfm.loc[(0, 'y'):(1, 'z')]

UnsortedIndexError: 'Key length (2) was greater than MultiIndex lexsort depth (1)'

The is_lexsorted() method on a MultiIndex shows if the index is sorted, and the lexsort_depth prop-

382 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

erty returns the sort depth:

In [109]: dfm.index.is_lexsorted()
Out[109]: False

In [110]: dfm.index.lexsort_depth
Out[110]: 1
In [111]: dfm = dfm.sort_index()

In [112]: dfm
Out[112]:
jolie
jim joe
0 x 0.238982
x 0.824351
1 y 0.076877
z 0.159509

In [113]: dfm.index.is_lexsorted()
Out[113]: True

In [114]: dfm.index.lexsort_depth
Out[114]: 2
And now selection works as expected.

In [115]: dfm.loc[(0, 'y'):(1, 'z')]

Out[115]:
jolie
jim joe
1 y 0.076877
z 0.159509

4.3.4 Take methods

Similar to NumPy ndarrays, pandas Index, Series, and DataFrame also provides the take() method that
retrieves elements along a given axis at the given indices. The given indices must be either a list or an ndarray of
integer index positions. take will also accept negative integers as relative positions to the end of the object.
In [116]: index = pd.Index(np.random.randint(0, 1000, 10))

In [117]: index
Out[117]: Int64Index([951, 384, 317, 822, 525, 60, 525, 328, 210, 945],
,→dtype='int64')

In [118]: positions = [0, 9, 3]

In [119]: index[positions]
Out[119]: Int64Index([951, 945, 822], dtype='int64')

In [120]: index.take(positions)
Out[120]: Int64Index([951, 945, 822], dtype='int64')

In [121]: ser = pd.Series(np.random.randn(10))

4.3. MultiIndex / advanced indexing 383

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [122]: ser.iloc[positions]
Out[122]:
0 0.776615
9 -1.120902
3 -0.232684
dtype: float64

In [123]: ser.take(positions)
Out[123]:
0 0.776615
9 -1.120902
3 -0.232684
dtype: float64
For DataFrames, the given indices should be a 1d list or ndarray that specifies row or column positions.
In [124]: frm = pd.DataFrame(np.random.randn(5, 3))

In [125]: frm.take([1, 4, 3])

Out[125]:
0 1 2
1 0.122843 -0.001117 -0.571039
4 -1.630543 0.053506 -1.312211
3 -1.131866 1.675784 -0.420701

In [126]: frm.take([0, 2], axis=1)

Out[126]:
0 2
0 -0.420297 0.331846
1 0.122843 -0.571039
2 1.182088 1.695155
3 -1.131866 -0.420701
4 -1.630543 -1.312211
It is important to note that the take method on pandas objects are not intended to work on boolean indices and may
return unexpected results.
In [127]: arr = np.random.randn(10)

In [128]: arr.take([False, False, True, True])

Out[128]: array([ 1.62593494, 1.62593494, -0.90445976, -0.90445976])

In [129]: arr[[0, 1]]

Out[129]: array([ 1.62593494, -0.90445976])

In [130]: ser = pd.Series(np.random.randn(10))

In [131]: ser.take([False, False, True, True])

Out[131]:
0 1.186759
0 1.186759
1 1.194373
1 1.194373
dtype: float64

384 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [132]: ser.iloc[[0, 1]]

Out[132]:
0 1.186759
1 1.194373
dtype: float64
Finally, as a small note on performance, because the take method handles a narrower range of inputs, it can offer
performance that is a good deal faster than fancy indexing.

In [133]: arr = np.random.randn(10000, 5)

In [134]: indexer = np.arange(10000)

In [135]: random.shuffle(indexer)

In [136]: %timeit arr[indexer]

.....: %timeit arr.take(indexer, axis=0)
.....:
200 us +- 3.32 us per loop (mean +- std. dev. of 7 runs, 10000 loops each)
76.6 us +- 1.26 us per loop (mean +- std. dev. of 7 runs, 10000 loops each)

In [137]: ser = pd.Series(arr[:, 0])

In [138]: %timeit ser.iloc[indexer]

.....: %timeit ser.take(indexer)
.....:
117 us +- 2.31 us per loop (mean +- std. dev. of 7 runs, 10000 loops each)
106 us +- 1.87 us per loop (mean +- std. dev. of 7 runs, 10000 loops each)

4.3.5 Index types

We have discussed MultiIndex in the previous sections pretty extensively. Documentation about
DatetimeIndex and PeriodIndex are shown here, and documentation about TimedeltaIndex is found
here.
In the following sub-sections we will highlight some other index types.

CategoricalIndex

CategoricalIndex is a type of index that is useful for supporting indexing with duplicates. This is a container
around a Categorical and allows efficient indexing and storage of an index with a large number of duplicated
elements.
In [139]: from pandas.api.types import CategoricalDtype

In [140]: df = pd.DataFrame({'A': np.arange(6),

.....: 'B': list('aabbca')})
.....:

In [141]: df['B'] = df['B'].astype(CategoricalDtype(list('cab')))

In [142]: df
Out[142]:
A B
0 0 a

4.3. MultiIndex / advanced indexing 385

pandas: powerful Python data analysis toolkit, Release 0.25.3

1 1 a
2 2 b
3 3 b
4 4 c
5 5 a

In [143]: df.dtypes
Out[143]:
A int64
B category
dtype: object

In [144]: df.B.cat.categories
Out[144]: Index(['c', 'a', 'b'], dtype='object')
Setting the index will create a CategoricalIndex.

In [145]: df2 = df.set_index('B')

In [146]: df2.index
Out[146]: CategoricalIndex(['a', 'a', 'b', 'b', 'c', 'a'], categories=['c', 'a', 'b'],
,→ ordered=False, name='B', dtype='category')

Indexing with __getitem__/.iloc/.loc works similarly to an Index with duplicates. The indexers must be
in the category or the operation will raise a KeyError.

In [147]: df2.loc['a']
Out[147]:
A
B
a 0
a 1
a 5

The CategoricalIndex is preserved after indexing:

In [148]: df2.loc['a'].index
Out[148]: CategoricalIndex(['a', 'a', 'a'], categories=['c', 'a', 'b'], ordered=False,
,→ name='B', dtype='category')

Sorting the index will sort by the order of the categories (recall that we created the index with
CategoricalDtype(list('cab')), so the sorted order is cab).

In [149]: df2.sort_index()
Out[149]:
A
B
c 4
a 0
a 1
a 5
b 2
b 3

Groupby operations on the index will preserve the index nature as well.
In [150]: df2.groupby(level=0).sum()
Out[150]:

386 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

A
B
c 4
a 6
b 5

In [151]: df2.groupby(level=0).sum().index
Out[151]: CategoricalIndex(['c', 'a', 'b'], categories=['c', 'a', 'b'],
,→ordered=False, name='B', dtype='category')

Reindexing operations will return a resulting index based on the type of the passed indexer. Passing a list will return
a plain-old Index; indexing with a Categorical will return a CategoricalIndex, indexed according to the
categories of the passed Categorical dtype. This allows one to arbitrarily index these even with values not in the
categories, similarly to how you can reindex any pandas index.
In [152]: df2.reindex(['a', 'e'])
Out[152]:
A
B
a 0.0
a 1.0
a 5.0
e NaN

In [153]: df2.reindex(['a', 'e']).index

Out[153]: Index(['a', 'a', 'a', 'e'], dtype='object', name='B')

In [154]: df2.reindex(pd.Categorical(['a', 'e'], categories=list('abcde')))

Out[154]:
A
B
a 0.0
a 1.0
a 5.0
e NaN

In [155]: df2.reindex(pd.Categorical(['a', 'e'], categories=list('abcde'))).

,→index

Out[155]: CategoricalIndex(['a', 'a', 'a', 'e'], categories=['a', 'b', 'c',

,→'d', 'e'], ordered=False, name='B', dtype='category')

Warning: Reshaping and Comparison operations on a CategoricalIndex must have the same categories or
a TypeError will be raised.
In [9]: df3 = pd.DataFrame({'A': np.arange(6), 'B': pd.Series(list('aabbca')).
,→astype('category')})

In [11]: df3 = df3.set_index('B')

In [11]: df3.index
Out[11]: CategoricalIndex(['a', 'a', 'b', 'b', 'c', 'a'], categories=['a', 'b', 'c
,→'], ordered=False, name='B', dtype='category')

In [12]: pd.concat([df2, df3])

TypeError: categories must match existing categories when appending

4.3. MultiIndex / advanced indexing 387

pandas: powerful Python data analysis toolkit, Release 0.25.3

Int64Index and RangeIndex

Warning: Indexing on an integer-based Index with floats has been clarified in 0.18.0, for a summary of the
changes, see here.

Int64Index is a fundamental basic index in pandas. This is an immutable array implementing an ordered, sliceable
set. Prior to 0.18.0, the Int64Index would provide the default index for all NDFrame objects.
RangeIndex is a sub-class of Int64Index added in version 0.18.0, now providing the default index for all
NDFrame objects. RangeIndex is an optimized version of Int64Index that can represent a monotonic ordered
set. These are analogous to Python range types.

Float64Index

By default a Float64Index will be automatically created when passing floating, or mixed-integer-floating values
in index creation. This enables a pure label-based slicing paradigm that makes [],ix,loc for scalar indexing and
slicing work exactly the same.

In [156]: indexf = pd.Index([1.5, 2, 3, 4.5, 5])

In [157]: indexf
Out[157]: Float64Index([1.5, 2.0, 3.0, 4.5, 5.0], dtype='float64')

In [158]: sf = pd.Series(range(5), index=indexf)

In [159]: sf
Out[159]:
1.5 0
2.0 1
3.0 2
4.5 3
5.0 4
dtype: int64

Scalar selection for [],.loc will always be label based. An integer will match an equal float index (e.g. 3 is
equivalent to 3.0).
In [160]: sf[3]
Out[160]: 2

In [161]: sf[3.0]
Out[161]: 2

In [162]: sf.loc[3]
Out[162]: 2

In [163]: sf.loc[3.0]
Out[163]: 2
The only positional indexing is via iloc.

In [164]: sf.iloc[3]
Out[164]: 3

388 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

A scalar index that is not found will raise a KeyError. Slicing is primarily on the values of the index when using
[],ix,loc, and always positional when using iloc. The exception is when the slice is boolean, in which case it
will always be positional.
In [165]: sf[2:4]
Out[165]:
2.0 1
3.0 2
dtype: int64

In [166]: sf.loc[2:4]
Out[166]:
2.0 1
3.0 2
dtype: int64

In [167]: sf.iloc[2:4]
Out[167]:
3.0 2
4.5 3
dtype: int64
In float indexes, slicing using floats is allowed.
In [168]: sf[2.1:4.6]
Out[168]:
3.0 2
4.5 3
dtype: int64

In [169]: sf.loc[2.1:4.6]
Out[169]:
3.0 2
4.5 3
dtype: int64
In non-float indexes, slicing using floats will raise a TypeError.

In [1]: pd.Series(range(5))[3.5]
TypeError: the label [3.5] is not a proper indexer for this index type (Int64Index)

In [1]: pd.Series(range(5))[3.5:4.5]
TypeError: the slice start [3.5] is not a proper indexer for this index type
,→(Int64Index)

Warning: Using a scalar float indexer for .iloc has been removed in 0.18.0, so the following will raise a
TypeError:
In [3]: pd.Series(range(5)).iloc[3.0]
TypeError: cannot do positional indexing on <class 'pandas.indexes.range.RangeIndex
,→'> with these indexers [3.0] of <type 'float'>

Here is a typical use-case for using this type of indexing. Imagine that you have a somewhat irregular timedelta-like
indexing scheme, but the data is recorded as floats. This could, for example, be millisecond offsets.

4.3. MultiIndex / advanced indexing 389

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [170]: dfir = pd.concat([pd.DataFrame(np.random.randn(5, 2),

.....: index=np.arange(5) * 250.0,
.....: columns=list('AB')),
.....: pd.DataFrame(np.random.randn(6, 2),
.....: index=np.arange(4, 10) * 250.1,
.....: columns=list('AB'))])
.....:

In [171]: dfir
Out[171]:
A B
0.0 -1.902297 -0.933899
250.0 0.035744 -0.338195
500.0 -1.851855 0.635651
750.0 -0.207053 0.419207
1000.0 0.337604 1.324050
1000.4 -0.623756 0.066847
1250.5 0.597810 0.597494
1500.6 -0.028424 -1.160777
1750.7 -1.591904 -1.183292
2000.8 0.868619 0.417332
2250.9 0.265403 -0.379946

Selection operations then will always work on a value basis, for all selection operators.
In [172]: dfir[0:1000.4]
Out[172]:
A B
0.0 -1.902297 -0.933899
250.0 0.035744 -0.338195
500.0 -1.851855 0.635651
750.0 -0.207053 0.419207
1000.0 0.337604 1.324050
1000.4 -0.623756 0.066847

In [173]: dfir.loc[0:1001, 'A']

Out[173]:
0.0 -1.902297
250.0 0.035744
500.0 -1.851855
750.0 -0.207053
1000.0 0.337604
1000.4 -0.623756
Name: A, dtype: float64

In [174]: dfir.loc[1000.4]
Out[174]:
A -0.623756
B 0.066847
Name: 1000.4, dtype: float64
You could retrieve the first 1 second (1000 ms) of data as such:
In [175]: dfir[0:1000]
Out[175]:
A B
(continues on next page)

390 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

0.0 -1.902297 -0.933899
250.0 0.035744 -0.338195
500.0 -1.851855 0.635651
750.0 -0.207053 0.419207
1000.0 0.337604 1.324050

If you need integer based selection, you should use iloc:

In [176]: dfir.iloc[0:5]
Out[176]:
A B
0.0 -1.902297 -0.933899
250.0 0.035744 -0.338195
500.0 -1.851855 0.635651
750.0 -0.207053 0.419207
1000.0 0.337604 1.324050

IntervalIndex

New in version 0.20.0.

IntervalIndex together with its own dtype, IntervalDtype as well as the Interval scalar type, allow
first-class support in pandas for interval notation.
The IntervalIndex allows some unique indexing and is also used as a return type for the categories in cut()
and qcut().

Indexing with an IntervalIndex

An IntervalIndex can be used in Series and in DataFrame as the index.

In [177]: df = pd.DataFrame({'A': [1, 2, 3, 4]},

.....: index=pd.IntervalIndex.from_breaks([0, 1, 2, 3, 4]))
.....:

In [178]: df
Out[178]:
A
(0, 1] 1
(1, 2] 2
(2, 3] 3
(3, 4] 4

Label based indexing via .loc along the edges of an interval works as you would expect, selecting that particular
interval.
In [179]: df.loc[2]
Out[179]:
A 2
Name: (1, 2], dtype: int64

In [180]: df.loc[[2, 3]]

Out[180]:
A

4.3. MultiIndex / advanced indexing 391

pandas: powerful Python data analysis toolkit, Release 0.25.3

(1, 2] 2
(2, 3] 3
If you select a label contained within an interval, this will also select the interval.
In [181]: df.loc[2.5]
Out[181]:
A 3
Name: (2, 3], dtype: int64

In [182]: df.loc[[2.5, 3.5]]

Out[182]:
A
(2, 3] 3
(3, 4] 4
Selecting using an Interval will only return exact matches (starting from pandas 0.25.0).

In [183]: df.loc[pd.Interval(1, 2)]

Out[183]:
A 2
Name: (1, 2], dtype: int64

Trying to select an Interval that is not exactly contained in the IntervalIndex will raise a KeyError.

In [7]: df.loc[pd.Interval(0.5, 2.5)]

---------------------------------------------------------------------------
KeyError: Interval(0.5, 2.5, closed='right')

Selecting all Intervals that overlap a given Interval can be performed using the overlaps() method to
create a boolean indexer.
In [184]: idxr = df.index.overlaps(pd.Interval(0.5, 2.5))

In [185]: idxr
Out[185]: array([ True, True, True, False])

In [186]: df[idxr]
Out[186]:
A
(0, 1] 1
(1, 2] 2
(2, 3] 3

Binning data with cut and qcut

cut() and qcut() both return a Categorical object, and the bins they create are stored as an IntervalIndex
in its .categories attribute.
In [187]: c = pd.cut(range(4), bins=2)

In [188]: c
Out[188]:
[(-0.003, 1.5], (-0.003, 1.5], (1.5, 3.0], (1.5, 3.0]]
Categories (2, interval[float64]): [(-0.003, 1.5] < (1.5, 3.0]]

392 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [189]: c.categories
Out[189]:
IntervalIndex([(-0.003, 1.5], (1.5, 3.0]],
closed='right',
dtype='interval[float64]')
cut() also accepts an IntervalIndex for its bins argument, which enables a useful pandas idiom. First, We
call cut() with some data and bins set to a fixed number, to generate the bins. Then, we pass the values of .
categories as the bins argument in subsequent calls to cut(), supplying new data which will be binned into
the same bins.

In [190]: pd.cut([0, 3, 5, 1], bins=c.categories)

Out[190]:
[(-0.003, 1.5], (1.5, 3.0], NaN, (-0.003, 1.5]]
Categories (2, interval[float64]): [(-0.003, 1.5] < (1.5, 3.0]]

Any value which falls outside all bins will be assigned a NaN value.

Generating ranges of intervals

If we need intervals on a regular frequency, we can use the interval_range() function to create an
IntervalIndex using various combinations of start, end, and periods. The default frequency for
interval_range is a 1 for numeric intervals, and calendar day for datetime-like intervals:
In [191]: pd.interval_range(start=0, end=5)
Out[191]:
IntervalIndex([(0, 1], (1, 2], (2, 3], (3, 4], (4, 5]],
closed='right',
dtype='interval[int64]')

In [192]: pd.interval_range(start=pd.Timestamp('2017-01-01'), periods=4)

Out[192]:
IntervalIndex([(2017-01-01, 2017-01-02], (2017-01-02, 2017-01-03], (2017-01-
,→03, 2017-01-04], (2017-01-04, 2017-01-05]],

closed='right',
dtype='interval[datetime64[ns]]')

In [193]: pd.interval_range(end=pd.Timedelta('3 days'), periods=3)

Out[193]:
IntervalIndex([(0 days 00:00:00, 1 days 00:00:00], (1 days 00:00:00, 2 days
,→00:00:00], (2 days 00:00:00, 3 days 00:00:00]],

closed='right',
dtype='interval[timedelta64[ns]]')
The freq parameter can used to specify non-default frequencies, and can utilize a variety of frequency aliases with
datetime-like intervals:
In [194]: pd.interval_range(start=0, periods=5, freq=1.5)
Out[194]:
IntervalIndex([(0.0, 1.5], (1.5, 3.0], (3.0, 4.5], (4.5, 6.0], (6.0, 7.5]],
closed='right',
dtype='interval[float64]')

In [195]: pd.interval_range(start=pd.Timestamp('2017-01-01'), periods=4,

,→freq='W')

4.3. MultiIndex / advanced indexing 393

pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[195]:
IntervalIndex([(2017-01-01, 2017-01-08], (2017-01-08, 2017-01-15], (2017-01-
,→15, 2017-01-22], (2017-01-22, 2017-01-29]],

closed='right',
dtype='interval[datetime64[ns]]')

In [196]: pd.interval_range(start=pd.Timedelta('0 days'), periods=3,

,→freq='9H')

Out[196]:
IntervalIndex([(0 days 00:00:00, 0 days 09:00:00], (0 days 09:00:00, 0 days
,→18:00:00], (0 days 18:00:00, 1 days 03:00:00]],

closed='right',
dtype='interval[timedelta64[ns]]')
Additionally, the closed parameter can be used to specify which side(s) the intervals are closed on. Intervals are
closed on the right side by default.
In [197]: pd.interval_range(start=0, end=4, closed='both')
Out[197]:
IntervalIndex([[0, 1], [1, 2], [2, 3], [3, 4]],
closed='both',
dtype='interval[int64]')

In [198]: pd.interval_range(start=0, end=4, closed='neither')

Out[198]:
IntervalIndex([(0, 1), (1, 2), (2, 3), (3, 4)],
closed='neither',
dtype='interval[int64]')
New in version 0.23.0.
Specifying start, end, and periods will generate a range of evenly spaced intervals from start to end inclu-
sively, with periods number of elements in the resulting IntervalIndex:
In [199]: pd.interval_range(start=0, end=6, periods=4)
Out[199]:
IntervalIndex([(0.0, 1.5], (1.5, 3.0], (3.0, 4.5], (4.5, 6.0]],
closed='right',
dtype='interval[float64]')

In [200]: pd.interval_range(pd.Timestamp('2018-01-01'),
.....: pd.Timestamp('2018-02-28'), periods=3)
.....:
Out[200]:
IntervalIndex([(2018-01-01, 2018-01-20 08:00:00], (2018-01-20 08:00:00, 2018-
,→02-08 16:00:00], (2018-02-08 16:00:00, 2018-02-28]],

closed='right',
dtype='interval[datetime64[ns]]')

4.3.6 Miscellaneous indexing FAQ

Integer indexing

Label-based indexing with integer axis labels is a thorny topic. It has been discussed heavily on mailing lists and
among various members of the scientific Python community. In pandas, our general viewpoint is that labels matter

394 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

more than integer locations. Therefore, with an integer axis index only label-based indexing is possible with the
standard tools like .loc. The following code will generate exceptions:
In [201]: s = pd.Series(range(5))

In [202]: s[-1]
---------------------------------------------------------------------------
KeyError Traceback (most recent call last)
<ipython-input-202-76c3dce40054> in <module>
----> 1 s[-1]

~/sandbox/pandas-release/pandas/pandas/core/series.py in __getitem__(self,
,→key)

1069 key = com.apply_if_callable(key, self)

1070 try:
-> 1071 result = self.index.get_value(self, key)
1072
1073 if not is_scalar(result):

~/sandbox/pandas-release/pandas/pandas/core/indexes/base.py in get_value(self,
,→ series, key)

4728 k = self._convert_scalar_indexer(k, kind="getitem")

4729 try:
-> 4730 return self._engine.get_value(s, k, tz=getattr(series.
,→dtype, "tz", None))

4731 except KeyError as e1:

4732 if len(self) > 0 and (self.holds_integer() or self.is_
,→boolean()):

~/sandbox/pandas-release/pandas/pandas/_libs/index.pyx in pandas._libs.index.
,→IndexEngine.get_value()

~/sandbox/pandas-release/pandas/pandas/_libs/index.pyx in pandas._libs.index.
,→IndexEngine.get_value()

~/sandbox/pandas-release/pandas/pandas/_libs/index.pyx in pandas._libs.index.
,→IndexEngine.get_loc()

~/sandbox/pandas-release/pandas/pandas/_libs/hashtable_class_helper.pxi in
,→pandas._libs.hashtable.Int64HashTable.get_item()

~/sandbox/pandas-release/pandas/pandas/_libs/hashtable_class_helper.pxi in
,→pandas._libs.hashtable.Int64HashTable.get_item()

KeyError: -1

In [203]: df = pd.DataFrame(np.random.randn(5, 4))

In [204]: df
Out[204]:
0 1 2 3
0 0.066370 0.481162 -0.247207 0.021250
1 0.768621 -0.053824 1.219183 2.211600
2 -0.381577 -0.068826 -1.112808 -0.976648

4.3. MultiIndex / advanced indexing 395

pandas: powerful Python data analysis toolkit, Release 0.25.3

3 -0.177430 -1.029073 0.118622 0.872648

4 -1.240728 1.519051 -0.122180 -0.665184

In [205]: df.loc[-2:]
Out[205]:
0 1 2 3
0 0.066370 0.481162 -0.247207 0.021250
1 0.768621 -0.053824 1.219183 2.211600
2 -0.381577 -0.068826 -1.112808 -0.976648
3 -0.177430 -1.029073 0.118622 0.872648
4 -1.240728 1.519051 -0.122180 -0.665184
This deliberate decision was made to prevent ambiguities and subtle bugs (many users reported finding bugs when the
API change was made to stop falling back on position-based indexing).

Non-monotonic indexes require exact matches

If the index of a Series or DataFrame is monotonically increasing or decreasing, then the bounds of a label-based
slice can be outside the range of the index, much like slice indexing a normal Python list. Monotonicity of an index
can be tested with the is_monotonic_increasing() and is_monotonic_decreasing() attributes.
In [206]: df = pd.DataFrame(index=[2, 3, 3, 4, 5], columns=['data'],
,→data=list(range(5)))

In [207]: df.index.is_monotonic_increasing
Out[207]: True

# no rows 0 or 1, but still returns rows 2, 3 (both of them), and 4:

In [208]: df.loc[0:4, :]
Out[208]:
data
2 0
3 1
3 2
4 3

# slice is are outside the index, so empty DataFrame is returned

In [209]: df.loc[13:15, :]
Out[209]:
Empty DataFrame
Columns: [data]
Index: []
On the other hand, if the index is not monotonic, then both slice bounds must be unique members of the index.
In [210]: df = pd.DataFrame(index=[2, 3, 1, 4, 3, 5],
.....: columns=['data'], data=list(range(6)))
.....:

In [211]: df.index.is_monotonic_increasing
Out[211]: False

# OK because 2 and 4 are in the index

In [212]: df.loc[2:4, :]
Out[212]:

396 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

data
2 0
3 1
1 2
4 3

# 0 is not in the index

In [9]: df.loc[0:4, :]
KeyError: 0

# 3 is not a unique label

In [11]: df.loc[2:3, :]
KeyError: 'Cannot get right slice bound for non-unique label: 3'

Index.is_monotonic_increasing and Index.is_monotonic_decreasing only check that an index

is weakly monotonic. To check for strict monotonicity, you can combine one of those with the is_unique()
attribute.
In [213]: weakly_monotonic = pd.Index(['a', 'b', 'c', 'c'])

In [214]: weakly_monotonic
Out[214]: Index(['a', 'b', 'c', 'c'], dtype='object')

In [215]: weakly_monotonic.is_monotonic_increasing
Out[215]: True

In [216]: weakly_monotonic.is_monotonic_increasing & weakly_monotonic.is_

,→unique

Out[216]: False

Endpoints are inclusive

Compared with standard Python sequence slicing in which the slice endpoint is not inclusive, label-based slicing in
pandas is inclusive. The primary reason for this is that it is often not possible to easily determine the successor or next
element after a particular label in an index. For example, consider the following Series:

In [217]: s = pd.Series(np.random.randn(6), index=list('abcdef'))

In [218]: s
Out[218]:
a 0.895739
b 1.349538
c 0.040227
d 1.154586
e 0.162300
f -1.213270
dtype: float64

Suppose we wished to slice from c to e, using integers this would be accomplished as such:

In [219]: s[2:5]
Out[219]:
c 0.040227
d 1.154586
e 0.162300
dtype: float64

4.3. MultiIndex / advanced indexing 397

pandas: powerful Python data analysis toolkit, Release 0.25.3

However, if you only had c and e, determining the next element in the index can be somewhat complicated. For
example, the following does not work:

s.loc['c':'e' + 1]

A very common use case is to limit a time series to start and end at two specific dates. To enable this, we made the
design choice to make label-based slicing include both endpoints:

In [220]: s.loc['c':'e']
Out[220]:
c 0.040227
d 1.154586
e 0.162300
dtype: float64

This is most definitely a practicality beats purity sort of thing, but it is something to watch out for if you expect
label-based slicing to behave exactly in the way that standard Python integer slicing works.

Indexing potentially changes underlying Series dtype

The different indexing operation can potentially change the dtype of a Series.
In [221]: series1 = pd.Series([1, 2, 3])

In [222]: series1.dtype
Out[222]: dtype('int64')

In [223]: res = series1.reindex([0, 4])

In [224]: res.dtype
Out[224]: dtype('float64')

In [225]: res
Out[225]:
0 1.0
4 NaN
dtype: float64
In [226]: series2 = pd.Series([True])

In [227]: series2.dtype
Out[227]: dtype('bool')

In [228]: res = series2.reindex_like(series1)

In [229]: res.dtype
Out[229]: dtype('O')

In [230]: res
Out[230]:
0 True
1 NaN
2 NaN
dtype: object

398 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

This is because the (re)indexing operations above silently inserts NaNs and the dtype changes accordingly. This can
cause some issues when using numpy ufuncs such as numpy.logical_and.
See the this old issue for a more detailed discussion. {{ header }}

4.4 Merge, join, and concatenate

pandas provides various facilities for easily combining together Series or DataFrame with various kinds of set logic
for the indexes and relational algebra functionality in the case of join / merge-type operations.

4.4.1 Concatenating objects

The concat() function (in the main pandas namespace) does all of the heavy lifting of performing concatenation
operations along an axis while performing optional set logic (union or intersection) of the indexes (if any) on the other
axes. Note that I say if any because there is only a single possible axis of concatenation for Series.
Before diving into all of the details of concat and what it can do, here is a simple example:

In [1]: df1 = pd.DataFrame({'A': ['A0', 'A1', 'A2', 'A3'],

...: 'B': ['B0', 'B1', 'B2', 'B3'],
...: 'C': ['C0', 'C1', 'C2', 'C3'],
...: 'D': ['D0', 'D1', 'D2', 'D3']},
...: index=[0, 1, 2, 3])
...:

In [2]: df2 = pd.DataFrame({'A': ['A4', 'A5', 'A6', 'A7'],

...: 'B': ['B4', 'B5', 'B6', 'B7'],
...: 'C': ['C4', 'C5', 'C6', 'C7'],
...: 'D': ['D4', 'D5', 'D6', 'D7']},
...: index=[4, 5, 6, 7])
...:

In [3]: df3 = pd.DataFrame({'A': ['A8', 'A9', 'A10', 'A11'],

...: 'B': ['B8', 'B9', 'B10', 'B11'],
...: 'C': ['C8', 'C9', 'C10', 'C11'],
...: 'D': ['D8', 'D9', 'D10', 'D11']},
...: index=[8, 9, 10, 11])
...:

In [4]: frames = [df1, df2, df3]

In [5]: result = pd.concat(frames)

4.4. Merge, join, and concatenate 399

pandas: powerful Python data analysis toolkit, Release 0.25.3

Like its sibling function on ndarrays, numpy.concatenate, pandas.concat takes a list or dict of
homogeneously-typed objects and concatenates them with some configurable handling of what to do with the other
axes:

pd.concat(objs, axis=0, join='outer', ignore_index=False, keys=None,

levels=None, names=None, verify_integrity=False, copy=True)

• objs : a sequence or mapping of Series or DataFrame objects. If a dict is passed, the sorted keys will be used as
the keys argument, unless it is passed, in which case the values will be selected (see below). Any None objects
will be dropped silently unless they are all None in which case a ValueError will be raised.
• axis : {0, 1, }, default 0. The axis to concatenate along.
• join : {inner, outer}, default outer. How to handle indexes on other axis(es). Outer for union and inner for
intersection.
• ignore_index : boolean, default False. If True, do not use the index values on the concatenation axis. The
resulting axis will be labeled 0, , n - 1. This is useful if you are concatenating objects where the concatenation
axis does not have meaningful indexing information. Note the index values on the other axes are still respected
in the join.
• keys : sequence, default None. Construct hierarchical index using the passed keys as the outermost level. If
multiple levels passed, should contain tuples.
• levels : list of sequences, default None. Specific levels (unique values) to use for constructing a MultiIndex.
Otherwise they will be inferred from the keys.
• names : list, default None. Names for the levels in the resulting hierarchical index.
• verify_integrity : boolean, default False. Check whether the new concatenated axis contains duplicates.
This can be very expensive relative to the actual data concatenation.
• copy : boolean, default True. If False, do not copy data unnecessarily.

400 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Without a little bit of context many of these arguments dont make much sense. Lets revisit the above example. Suppose
we wanted to associate specific keys with each of the pieces of the chopped up DataFrame. We can do this using the
keys argument:

In [6]: result = pd.concat(frames, keys=['x', 'y', 'z'])

As you can see (if youve read the rest of the documentation), the resulting objects index has a hierarchical index. This
means that we can now select out each chunk by key:

In [7]: result.loc['y']
Out[7]:
A B C D
4 A4 B4 C4 D4
5 A5 B5 C5 D5
6 A6 B6 C6 D6
7 A7 B7 C7 D7

Its not a stretch to see how this can be very useful. More detail on this functionality below.

Note: It is worth noting that concat() (and therefore append()) makes a full copy of the data, and that constantly
reusing this function can create a significant performance hit. If you need to use the operation over several datasets,
use a list comprehension.

frames = [ process_your_file(f) for f in files ]

result = pd.concat(frames)

4.4. Merge, join, and concatenate 401

pandas: powerful Python data analysis toolkit, Release 0.25.3

Set logic on the other axes

When gluing together multiple DataFrames, you have a choice of how to handle the other axes (other than the one
being concatenated). This can be done in the following two ways:
• Take the union of them all, join='outer'. This is the default option as it results in zero information loss.
• Take the intersection, join='inner'.
Here is an example of each of these methods. First, the default join='outer' behavior:

In [8]: df4 = pd.DataFrame({'B': ['B2', 'B3', 'B6', 'B7'],

...: 'D': ['D2', 'D3', 'D6', 'D7'],
...: 'F': ['F2', 'F3', 'F6', 'F7']},
...: index=[2, 3, 6, 7])
...:

In [9]: result = pd.concat([df1, df4], axis=1, sort=False)

Warning: Changed in version 0.23.0.

The default behavior with join='outer' is to sort the other axis (columns in this case). In a future version of
pandas, the default will be to not sort. We specified sort=False to opt in to the new behavior now.

Here is the same thing with join='inner':

In [10]: result = pd.concat([df1, df4], axis=1, join='inner')

Lastly, suppose we just wanted to reuse the exact index from the original DataFrame:

402 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [11]: result = pd.concat([df1, df4], axis=1).reindex(df1.index)

Similarly, we could index before the concatenation:

In [12]: pd.concat([df1, df4.reindex(df1.index)], axis=1)

Out[12]:
A B C D B D F
0 A0 B0 C0 D0 NaN NaN NaN
1 A1 B1 C1 D1 NaN NaN NaN
2 A2 B2 C2 D2 B2 D2 F2
3 A3 B3 C3 D3 B3 D3 F3

Concatenating using append

A useful shortcut to concat() are the append() instance methods on Series and DataFrame. These methods
actually predated concat. They concatenate along axis=0, namely the index:

In [13]: result = df1.append(df2)

In the case of DataFrame, the indexes must be disjoint but the columns do not need to be:

In [14]: result = df1.append(df4, sort=False)

4.4. Merge, join, and concatenate 403

pandas: powerful Python data analysis toolkit, Release 0.25.3

append may take multiple objects to concatenate:

In [15]: result = df1.append([df2, df3])

Note: Unlike the append() method, which appends to the original list and returns None, append() here does
not modify df1 and returns its copy with df2 appended.

404 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Ignoring indexes on the concatenation axis

For DataFrame objects which dont have a meaningful index, you may wish to append them and ignore the fact that
they may have overlapping indexes. To do this, use the ignore_index argument:

In [16]: result = pd.concat([df1, df4], ignore_index=True, sort=False)

This is also a valid argument to DataFrame.append():

In [17]: result = df1.append(df4, ignore_index=True, sort=False)

Concatenating with mixed ndims

You can concatenate a mix of Series and DataFrame objects. The Series will be transformed to DataFrame
with the column name as the name of the Series.

4.4. Merge, join, and concatenate 405

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [18]: s1 = pd.Series(['X0', 'X1', 'X2', 'X3'], name='X')

In [19]: result = pd.concat([df1, s1], axis=1)

Note: Since were concatenating a Series to a DataFrame, we could have achieved the same result with
DataFrame.assign(). To concatenate an arbitrary number of pandas objects (DataFrame or Series), use
concat.

If unnamed Series are passed they will be numbered consecutively.

In [20]: s2 = pd.Series(['_0', '_1', '_2', '_3'])

In [21]: result = pd.concat([df1, s2, s2, s2], axis=1)

Passing ignore_index=True will drop all name references.

In [22]: result = pd.concat([df1, s1], axis=1, ignore_index=True)

406 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

More concatenating with group keys

A fairly common use of the keys argument is to override the column names when creating a new DataFrame based
on existing Series. Notice how the default behaviour consists on letting the resulting DataFrame inherit the parent
Series name, when these existed.

In [23]: s3 = pd.Series([0, 1, 2, 3], name='foo')

In [24]: s4 = pd.Series([0, 1, 2, 3])

In [25]: s5 = pd.Series([0, 1, 4, 5])

In [26]: pd.concat([s3, s4, s5], axis=1)

Out[26]:
foo 0 1
0 0 0 0
1 1 1 1
2 2 2 4
3 3 3 5

Through the keys argument we can override the existing column names.

In [27]: pd.concat([s3, s4, s5], axis=1, keys=['red', 'blue', 'yellow'])

Out[27]:
red blue yellow
0 0 0 0
1 1 1 1
2 2 2 4
3 3 3 5

Lets consider a variation of the very first example presented:

In [28]: result = pd.concat(frames, keys=['x', 'y', 'z'])

4.4. Merge, join, and concatenate 407

pandas: powerful Python data analysis toolkit, Release 0.25.3

You can also pass a dict to concat in which case the dict keys will be used for the keys argument (unless other keys
are specified):

In [29]: pieces = {'x': df1, 'y': df2, 'z': df3}

In [30]: result = pd.concat(pieces)

408 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [31]: result = pd.concat(pieces, keys=['z', 'y'])

The MultiIndex created has levels that are constructed from the passed keys and the index of the DataFrame pieces:

4.4. Merge, join, and concatenate 409

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [32]: result.index.levels
Out[32]: FrozenList([['z', 'y'], [4, 5, 6, 7, 8, 9, 10, 11]])

If you wish to specify other levels (as will occasionally be the case), you can do so using the levels argument:

In [33]: result = pd.concat(pieces, keys=['x', 'y', 'z'],

....: levels=[['z', 'y', 'x', 'w']],
....: names=['group_key'])
....:

In [34]: result.index.levels
Out[34]: FrozenList([['z', 'y', 'x', 'w'], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11]])

This is fairly esoteric, but it is actually necessary for implementing things like GroupBy where the order of a categorical
variable is meaningful.

Appending rows to a DataFrame

While not especially efficient (since a new object must be created), you can append a single row to a DataFrame by
passing a Series or dict to append, which returns a new DataFrame as above.

In [35]: s2 = pd.Series(['X0', 'X1', 'X2', 'X3'], index=['A', 'B', 'C', 'D'])

In [36]: result = df1.append(s2, ignore_index=True)

410 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

You should use ignore_index with this method to instruct DataFrame to discard its index. If you wish to preserve
the index, you should construct an appropriately-indexed DataFrame and append or concatenate those objects.
You can also pass a list of dicts or Series:
In [37]: dicts = [{'A': 1, 'B': 2, 'C': 3, 'X': 4},
....: {'A': 5, 'B': 6, 'C': 7, 'Y': 8}]
....:

In [38]: result = df1.append(dicts, ignore_index=True, sort=False)

4.4.2 Database-style DataFrame or named Series joining/merging

pandas has full-featured, high performance in-memory join operations idiomatically very similar to relational
databases like SQL. These methods perform significantly better (in some cases well over an order of magnitude better)
than other open source implementations (like base::merge.data.frame in R). The reason for this is careful
algorithmic design and the internal layout of the data in DataFrame.
See the cookbook for some advanced strategies.
Users who are familiar with SQL but new to pandas might be interested in a comparison with SQL.
pandas provides a single function, merge(), as the entry point for all standard database join operations between
DataFrame or named Series objects:

4.4. Merge, join, and concatenate 411

pandas: powerful Python data analysis toolkit, Release 0.25.3

pd.merge(left, right, how='inner', on=None, left_on=None, right_on=None,

left_index=False, right_index=False, sort=True,
suffixes=('_x', '_y'), copy=True, indicator=False,
validate=None)

• left: A DataFrame or named Series object.

• right: Another DataFrame or named Series object.
• on: Column or index level names to join on. Must be found in both the left and right DataFrame and/or Series
objects. If not passed and left_index and right_index are False, the intersection of the columns in
the DataFrames and/or Series will be inferred to be the join keys.
• left_on: Columns or index levels from the left DataFrame or Series to use as keys. Can either be column
names, index level names, or arrays with length equal to the length of the DataFrame or Series.
• right_on: Columns or index levels from the right DataFrame or Series to use as keys. Can either be column
names, index level names, or arrays with length equal to the length of the DataFrame or Series.
• left_index: If True, use the index (row labels) from the left DataFrame or Series as its join key(s). In the
case of a DataFrame or Series with a MultiIndex (hierarchical), the number of levels must match the number of
join keys from the right DataFrame or Series.
• right_index: Same usage as left_index for the right DataFrame or Series
• how: One of 'left', 'right', 'outer', 'inner'. Defaults to inner. See below for more detailed
description of each method.
• sort: Sort the result DataFrame by the join keys in lexicographical order. Defaults to True, setting to False
will improve performance substantially in many cases.
• suffixes: A tuple of string suffixes to apply to overlapping columns. Defaults to ('_x', '_y').
• copy: Always copy data (default True) from the passed DataFrame or named Series objects, even when
reindexing is not necessary. Cannot be avoided in many cases but may improve performance / memory usage.
The cases where copying can be avoided are somewhat pathological but this option is provided nonetheless.
• indicator: Add a column to the output DataFrame called _merge with information on the source of each
row. _merge is Categorical-type and takes on a value of left_only for observations whose merge key only
appears in 'left' DataFrame or Series, right_only for observations whose merge key only appears in
'right' DataFrame or Series, and both if the observations merge key is found in both.
• validate : string, default None. If specified, checks if merge is of specified type.
– one_to_one or 1:1: checks if merge keys are unique in both left and right datasets.
– one_to_many or 1:m: checks if merge keys are unique in left dataset.
– many_to_one or m:1: checks if merge keys are unique in right dataset.
– many_to_many or m:m: allowed, but does not result in checks.
New in version 0.21.0.

Note: Support for specifying index levels as the on, left_on, and right_on parameters was added in version
0.23.0. Support for merging named Series objects was added in version 0.24.0.

The return type will be the same as left. If left is a DataFrame or named Series and right is a subclass of
DataFrame, the return type will still be DataFrame.
merge is a function in the pandas namespace, and it is also available as a DataFrame instance method merge(),
with the calling DataFrame being implicitly considered the left object in the join.

412 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

The related join() method, uses merge internally for the index-on-index (by default) and column(s)-on-index join.
If you are joining on index only, you may wish to use DataFrame.join to save yourself some typing.

Brief primer on merge methods (relational algebra)

Experienced users of relational databases like SQL will be familiar with the terminology used to describe join oper-
ations between two SQL-table like structures (DataFrame objects). There are several cases to consider which are
very important to understand:
• one-to-one joins: for example when joining two DataFrame objects on their indexes (which must contain
unique values).
• many-to-one joins: for example when joining an index (unique) to one or more columns in a different
DataFrame.
• many-to-many joins: joining columns on columns.

Note: When joining columns on columns (potentially a many-to-many join), any indexes on the passed DataFrame
objects will be discarded.

It is worth spending some time understanding the result of the many-to-many join case. In SQL / standard relational
algebra, if a key combination appears more than once in both tables, the resulting table will have the Cartesian
product of the associated data. Here is a very basic example with one unique key combination:

In [39]: left = pd.DataFrame({'key': ['K0', 'K1', 'K2', 'K3'],

....: 'A': ['A0', 'A1', 'A2', 'A3'],
....: 'B': ['B0', 'B1', 'B2', 'B3']})
....:

In [40]: right = pd.DataFrame({'key': ['K0', 'K1', 'K2', 'K3'],

....: 'C': ['C0', 'C1', 'C2', 'C3'],
....: 'D': ['D0', 'D1', 'D2', 'D3']})
....:

In [41]: result = pd.merge(left, right, on='key')

Here is a more complicated example with multiple join keys. Only the keys appearing in left and right are present
(the intersection), since how='inner' by default.

In [42]: left = pd.DataFrame({'key1': ['K0', 'K0', 'K1', 'K2'],

....: 'key2': ['K0', 'K1', 'K0', 'K1'],
....: 'A': ['A0', 'A1', 'A2', 'A3'],
....: 'B': ['B0', 'B1', 'B2', 'B3']})
....:
(continues on next page)

4.4. Merge, join, and concatenate 413

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [43]: right = pd.DataFrame({'key1': ['K0', 'K1', 'K1', 'K2'],

....: 'key2': ['K0', 'K0', 'K0', 'K0'],
....: 'C': ['C0', 'C1', 'C2', 'C3'],
....: 'D': ['D0', 'D1', 'D2', 'D3']})
....:

In [44]: result = pd.merge(left, right, on=['key1', 'key2'])

The how argument to merge specifies how to determine which keys are to be included in the resulting table. If a
key combination does not appear in either the left or right tables, the values in the joined table will be NA. Here is a
summary of the how options and their SQL equivalent names:

Merge method SQL Join Name Description

left LEFT OUTER Use keys from left frame only
JOIN
right RIGHT OUTER Use keys from right frame only
JOIN
outer FULL OUTER Use union of keys from both frames
JOIN
inner INNER JOIN Use intersection of keys from both frames

In [45]: result = pd.merge(left, right, how='left', on=['key1', 'key2'])

In [46]: result = pd.merge(left, right, how='right', on=['key1', 'key2'])

414 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [47]: result = pd.merge(left, right, how='outer', on=['key1', 'key2'])

In [48]: result = pd.merge(left, right, how='inner', on=['key1', 'key2'])

Here is another example with duplicate join keys in DataFrames:

In [49]: left = pd.DataFrame({'A': [1, 2], 'B': [2, 2]})

In [50]: right = pd.DataFrame({'A': [4, 5, 6], 'B': [2, 2, 2]})

In [51]: result = pd.merge(left, right, on='B', how='outer')

4.4. Merge, join, and concatenate 415

pandas: powerful Python data analysis toolkit, Release 0.25.3

Warning: Joining / merging on duplicate keys can cause a returned frame that is the multiplication of the row
dimensions, which may result in memory overflow. It is the user s responsibility to manage duplicate values in
keys before joining large DataFrames.

Checking for duplicate keys

New in version 0.21.0.

Users can use the validate argument to automatically check whether there are unexpected duplicates in their merge
keys. Key uniqueness is checked before merge operations and so should protect against memory overflows. Checking
key uniqueness is also a good way to ensure user data structures are as expected.
In the following example, there are duplicate values of B in the right DataFrame. As this is not a one-to-one merge
– as specified in the validate argument – an exception will be raised.

In [52]: left = pd.DataFrame({'A' : [1,2], 'B' : [1, 2]})

In [53]: right = pd.DataFrame({'A' : [4,5,6], 'B': [2, 2, 2]})

In [53]: result = pd.merge(left, right, on='B', how='outer', validate="one_to_one")

...
MergeError: Merge keys are not unique in right dataset; not a one-to-one merge

If the user is aware of the duplicates in the right DataFrame but wants to ensure there are no duplicates in the left
DataFrame, one can use the validate='one_to_many' argument instead, which will not raise an exception.

In [54]: pd.merge(left, right, on='B', how='outer', validate="one_to_many")

Out[54]:
A_x B A_y
0 1 1 NaN
1 2 2 4.0
2 2 2 5.0
3 2 2 6.0

The merge indicator

merge() accepts the argument indicator. If True, a Categorical-type column called _merge will be added to
the output object that takes on values:

416 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Observation Origin _merge value

Merge key only in 'left' frame left_only
Merge key only in 'right' frame right_only
Merge key in both frames both

In [55]: df1 = pd.DataFrame({'col1': [0, 1], 'col_left': ['a', 'b']})

In [56]: df2 = pd.DataFrame({'col1': [1, 2, 2], 'col_right': [2, 2, 2]})

In [57]: pd.merge(df1, df2, on='col1', how='outer', indicator=True)

Out[57]:
col1 col_left col_right _merge
0 0 a NaN left_only
1 1 b 2.0 both
2 2 NaN 2.0 right_only
3 2 NaN 2.0 right_only

The indicator argument will also accept string arguments, in which case the indicator function will use the value
of the passed string as the name for the indicator column.

In [58]: pd.merge(df1, df2, on='col1', how='outer', indicator='indicator_column')

Out[58]:
col1 col_left col_right indicator_column
0 0 a NaN left_only
1 1 b 2.0 both
2 2 NaN 2.0 right_only
3 2 NaN 2.0 right_only

Merge dtypes

New in version 0.19.0.

Merging will preserve the dtype of the join keys.

In [59]: left = pd.DataFrame({'key': [1], 'v1': [10]})

In [60]: left
Out[60]:
key v1
0 1 10

In [61]: right = pd.DataFrame({'key': [1, 2], 'v1': [20, 30]})

In [62]: right
Out[62]:
key v1
0 1 20
1 2 30

We are able to preserve the join keys:

In [63]: pd.merge(left, right, how='outer')
Out[63]:
key v1
0 1 10

4.4. Merge, join, and concatenate 417

pandas: powerful Python data analysis toolkit, Release 0.25.3

1 1 20
2 2 30

In [64]: pd.merge(left, right, how='outer').dtypes

Out[64]:
key int64
v1 int64
dtype: object
Of course if you have missing values that are introduced, then the resulting dtype will be upcast.
In [65]: pd.merge(left, right, how='outer', on='key')
Out[65]:
key v1_x v1_y
0 1 10.0 20
1 2 NaN 30

In [66]: pd.merge(left, right, how='outer', on='key').dtypes

Out[66]:
key int64
v1_x float64
v1_y int64
dtype: object
New in version 0.20.0.
Merging will preserve category dtypes of the mergands. See also the section on categoricals.
The left frame.
In [67]: from pandas.api.types import CategoricalDtype

In [68]: X = pd.Series(np.random.choice(['foo', 'bar'], size=(10,)))

In [69]: X = X.astype(CategoricalDtype(categories=['foo', 'bar']))

In [70]: left = pd.DataFrame({'X': X,

....: 'Y': np.random.choice(['one', 'two', 'three'],
....: size=(10,))})
....:

In [71]: left
Out[71]:
X Y
0 foo three
1 bar three
2 foo two
3 foo three
4 foo one
5 foo one
6 bar one
7 foo two
8 foo one
9 bar two

In [72]: left.dtypes
Out[72]:

418 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

X category
Y object
dtype: object
The right frame.
In [73]: right = pd.DataFrame({'X': pd.Series(['foo', 'bar'],
....: dtype=CategoricalDtype(['foo',
,→'bar'])),

....: 'Z': [1, 2]})

....:

In [74]: right
Out[74]:
X Z
0 foo 1
1 bar 2

In [75]: right.dtypes
Out[75]:
X category
Z int64
dtype: object
The merged result:
In [76]: result = pd.merge(left, right, how='outer')

In [77]: result
Out[77]:
X Y Z
0 foo three 1
1 foo two 1
2 foo three 1
3 foo one 1
4 foo one 1
5 foo two 1
6 foo one 1
7 bar three 2
8 bar one 2
9 bar two 2

In [78]: result.dtypes
Out[78]:
X category
Y object
Z int64
dtype: object

Note: The category dtypes must be exactly the same, meaning the same categories and the ordered attribute. Other-
wise the result will coerce to object dtype.

Note: Merging on category dtypes that are the same can be quite performant compared to object dtype merging.

4.4. Merge, join, and concatenate 419

pandas: powerful Python data analysis toolkit, Release 0.25.3

Joining on index

DataFrame.join() is a convenient method for combining the columns of two potentially differently-indexed
DataFrames into a single result DataFrame. Here is a very basic example:

In [79]: left = pd.DataFrame({'A': ['A0', 'A1', 'A2'],

....: 'B': ['B0', 'B1', 'B2']},
....: index=['K0', 'K1', 'K2'])
....:

In [80]: right = pd.DataFrame({'C': ['C0', 'C2', 'C3'],

....: 'D': ['D0', 'D2', 'D3']},
....: index=['K0', 'K2', 'K3'])
....:

In [81]: result = left.join(right)

In [82]: result = left.join(right, how='outer')

The same as above, but with how='inner'.

In [83]: result = left.join(right, how='inner')

The data alignment here is on the indexes (row labels). This same behavior can be achieved using merge plus
additional arguments instructing it to use the indexes:

420 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [84]: result = pd.merge(left, right, left_index=True, right_index=True, how='outer

,→')

In [85]: result = pd.merge(left, right, left_index=True, right_index=True, how='inner

,→');

Joining key columns on an index

join() takes an optional on argument which may be a column or multiple column names, which specifies that the
passed DataFrame is to be aligned on that column in the DataFrame. These two function calls are completely
equivalent:

left.join(right, on=key_or_keys)
pd.merge(left, right, left_on=key_or_keys, right_index=True,
how='left', sort=False)

Obviously you can choose whichever form you find more convenient. For many-to-one joins (where one of the
DataFrames is already indexed by the join key), using join may be more convenient. Here is a simple example:

In [86]: left = pd.DataFrame({'A': ['A0', 'A1', 'A2', 'A3'],

....: 'B': ['B0', 'B1', 'B2', 'B3'],
....: 'key': ['K0', 'K1', 'K0', 'K1']})
....:

In [87]: right = pd.DataFrame({'C': ['C0', 'C1'],

....: 'D': ['D0', 'D1']},
....: index=['K0', 'K1'])
....:

In [88]: result = left.join(right, on='key')

4.4. Merge, join, and concatenate 421

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [89]: result = pd.merge(left, right, left_on='key', right_index=True,

....: how='left', sort=False);
....:

To join on multiple keys,

the passed DataFrame must have a MultiIndex:

In [90]: left = pd.DataFrame({'A': ['A0', 'A1', 'A2', 'A3'],

....: 'B': ['B0', 'B1', 'B2', 'B3'],
....: 'key1': ['K0', 'K0', 'K1', 'K2'],
....: 'key2': ['K0', 'K1', 'K0', 'K1']})
....:

In [91]: index = pd.MultiIndex.from_tuples([('K0', 'K0'), ('K1', 'K0'),

....: ('K2', 'K0'), ('K2', 'K1')])
....:

In [92]: right = pd.DataFrame({'C': ['C0', 'C1', 'C2', 'C3'],

....: 'D': ['D0', 'D1', 'D2', 'D3']},
....: index=index)
....:

Now this can be joined by passing the two key column names:

In [93]: result = left.join(right, on=['key1', 'key2'])

The
default for DataFrame.join is to perform a left join (essentially a VLOOKUP operation, for Excel users), which
uses only the keys found in the calling DataFrame. Other join types, for example inner join, can be just as easily

422 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

performed:

In [94]: result = left.join(right, on=['key1', 'key2'], how='inner')

As you can see, this drops any rows where there was no match.

Joining a single Index to a MultiIndex

You can join a singly-indexed DataFrame with a level of a MultiIndexed DataFrame. The level will match on the
name of the index of the singly-indexed frame against a level name of the MultiIndexed frame.

In [95]: left = pd.DataFrame({'A': ['A0', 'A1', 'A2'],

....: 'B': ['B0', 'B1', 'B2']},
....: index=pd.Index(['K0', 'K1', 'K2'], name='key'))
....:

In [96]: index = pd.MultiIndex.from_tuples([('K0', 'Y0'), ('K1', 'Y1'),

....: ('K2', 'Y2'), ('K2', 'Y3')],
....: names=['key', 'Y'])
....:

In [97]: right = pd.DataFrame({'C': ['C0', 'C1', 'C2', 'C3'],

....: 'D': ['D0', 'D1', 'D2', 'D3']},
....: index=index)
....:

In [98]: result = left.join(right, how='inner')

This is equivalent but less verbose and more memory efficient / faster than this.

In [99]: result = pd.merge(left.reset_index(), right.reset_index(),

....: on=['key'], how='inner').set_index(['key','Y'])
....:

4.4. Merge, join, and concatenate 423

pandas: powerful Python data analysis toolkit, Release 0.25.3

Joining with two MultiIndexes

This is supported in a limited way, provided that the index for the right argument is completely used in the join, and is
a subset of the indices in the left argument, as in this example:
In [100]: leftindex = pd.MultiIndex.from_product([list('abc'), list('xy'), [1,
,→ 2]],

.....: names=['abc', 'xy', 'num'])

.....:

In [101]: left = pd.DataFrame({'v1': range(12)}, index=leftindex)

In [102]: left
Out[102]:
v1
abc xy num
a x 1 0
2 1
y 1 2
2 3
b x 1 4
2 5
y 1 6
2 7
c x 1 8
2 9
y 1 10
2 11

In [103]: rightindex = pd.MultiIndex.from_product([list('abc'), list('xy')],

.....: names=['abc', 'xy'])
.....:

In [104]: right = pd.DataFrame({'v2': [100 * i for i in range(1, 7)]},

,→index=rightindex)

In [105]: right
Out[105]:
v2
abc xy
a x 100
y 200

424 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

b x 300
y 400
c x 500
y 600

In [106]: left.join(right, on=['abc', 'xy'], how='inner')

Out[106]:
v1 v2
abc xy num
a x 1 0 100
2 1 100
y 1 2 200
2 3 200
b x 1 4 300
2 5 300
y 1 6 400
2 7 400
c x 1 8 500
2 9 500
y 1 10 600
2 11 600
If that condition is not satisfied, a join with two multi-indexes can be done using the following code.

In [107]: leftindex = pd.MultiIndex.from_tuples([('K0', 'X0'), ('K0', 'X1'),

.....: ('K1', 'X2')],
.....: names=['key', 'X'])
.....:

In [108]: left = pd.DataFrame({'A': ['A0', 'A1', 'A2'],

.....: 'B': ['B0', 'B1', 'B2']},
.....: index=leftindex)
.....:

In [109]: rightindex = pd.MultiIndex.from_tuples([('K0', 'Y0'), ('K1', 'Y1'),

.....: ('K2', 'Y2'), ('K2', 'Y3')],
.....: names=['key', 'Y'])
.....:

In [110]: right = pd.DataFrame({'C': ['C0', 'C1', 'C2', 'C3'],

.....: 'D': ['D0', 'D1', 'D2', 'D3']},
.....: index=rightindex)
.....:

In [111]: result = pd.merge(left.reset_index(), right.reset_index(),

.....: on=['key'], how='inner').set_index(['key', 'X', 'Y'])
.....:

4.4. Merge, join, and concatenate 425

pandas: powerful Python data analysis toolkit, Release 0.25.3

Merging on a combination of columns and index levels

New in version 0.23.

Strings passed as the on, left_on, and right_on parameters may refer to either column names or index level
names. This enables merging DataFrame instances on a combination of index levels and columns without resetting
indexes.

In [112]: left_index = pd.Index(['K0', 'K0', 'K1', 'K2'], name='key1')

In [113]: left = pd.DataFrame({'A': ['A0', 'A1', 'A2', 'A3'],

.....: 'B': ['B0', 'B1', 'B2', 'B3'],
.....: 'key2': ['K0', 'K1', 'K0', 'K1']},
.....: index=left_index)
.....:

In [114]: right_index = pd.Index(['K0', 'K1', 'K2', 'K2'], name='key1')

In [115]: right = pd.DataFrame({'C': ['C0', 'C1', 'C2', 'C3'],

.....: 'D': ['D0', 'D1', 'D2', 'D3'],
.....: 'key2': ['K0', 'K0', 'K0', 'K1']},
.....: index=right_index)
.....:

In [116]: result = left.merge(right, on=['key1', 'key2'])

Note: When DataFrames are merged on a string that matches an index level in both frames, the index level is
preserved as an index level in the resulting DataFrame.

Note: When DataFrames are merged using only some of the levels of a MultiIndex, the extra levels will be dropped
from the resulting merge. In order to preserve those levels, use reset_index on those level names to move those
levels to columns prior to doing the merge.

426 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Note: If a string matches both a column name and an index level name, then a warning is issued and the column takes
precedence. This will result in an ambiguity error in a future version.

Overlapping value columns

The merge suffixes argument takes a tuple of list of strings to append to overlapping column names in the input
DataFrames to disambiguate the result columns:

In [117]: left = pd.DataFrame({'k': ['K0', 'K1', 'K2'], 'v': [1, 2, 3]})

In [118]: right = pd.DataFrame({'k': ['K0', 'K0', 'K3'], 'v': [4, 5, 6]})

In [119]: result = pd.merge(left, right, on='k')

In [120]: result = pd.merge(left, right, on='k', suffixes=['_l', '_r'])

DataFrame.join() has lsuffix and rsuffix arguments which behave similarly.

In [121]: left = left.set_index('k')

In [122]: right = right.set_index('k')

In [123]: result = left.join(right, lsuffix='_l', rsuffix='_r')

4.4. Merge, join, and concatenate 427

pandas: powerful Python data analysis toolkit, Release 0.25.3

Joining multiple DataFrames

A list or tuple of DataFrames can also be passed to join() to join them together on their indexes.

In [124]: right2 = pd.DataFrame({'v': [7, 8, 9]}, index=['K1', 'K1', 'K2'])

In [125]: result = left.join([right, right2])

Merging together values within Series or DataFrame columns

Another fairly common situation is to have two like-indexed (or similarly indexed) Series or DataFrame objects
and wanting to patch values in one object from values for matching indices in the other. Here is an example:

In [126]: df1 = pd.DataFrame([[np.nan, 3., 5.], [-4.6, np.nan, np.nan],

.....: [np.nan, 7., np.nan]])
.....:

In [127]: df2 = pd.DataFrame([[-42.6, np.nan, -8.2], [-5., 1.6, 4]],

.....: index=[1, 2])
.....:

For this, use the combine_first() method:

In [128]: result = df1.combine_first(df2)

Note that this method only takes values from the right DataFrame if they are missing in the left DataFrame. A
related method, update(), alters non-NA values in place:

In [129]: df1.update(df2)

428 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

4.4.3 Timeseries friendly merging

Merging ordered data

A merge_ordered() function allows combining time series and other ordered data. In particular it has an optional
fill_method keyword to fill/interpolate missing data:

In [130]: left = pd.DataFrame({'k': ['K0', 'K1', 'K1', 'K2'],

.....: 'lv': [1, 2, 3, 4],
.....: 's': ['a', 'b', 'c', 'd']})
.....:

In [131]: right = pd.DataFrame({'k': ['K1', 'K2', 'K4'],

.....: 'rv': [1, 2, 3]})
.....:

In [132]: pd.merge_ordered(left, right, fill_method='ffill', left_by='s')

Out[132]:
k lv s rv
0 K0 1.0 a NaN
1 K1 1.0 a 1.0
2 K2 1.0 a 2.0
3 K4 1.0 a 3.0
4 K1 2.0 b 1.0
5 K2 2.0 b 2.0
6 K4 2.0 b 3.0
7 K1 3.0 c 1.0
8 K2 3.0 c 2.0
9 K4 3.0 c 3.0
10 K1 NaN d 1.0
11 K2 4.0 d 2.0
12 K4 4.0 d 3.0

Merging asof

New in version 0.19.0.

A merge_asof() is similar to an ordered left-join except that we match on nearest key rather than equal keys. For
each row in the left DataFrame, we select the last row in the right DataFrame whose on key is less than the
lefts key. Both DataFrames must be sorted by the key.
Optionally an asof merge can perform a group-wise merge. This matches the by key equally, in addition to the nearest
match on the on key.
For example; we might have trades and quotes and we want to asof merge them.

4.4. Merge, join, and concatenate 429

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [133]: trades = pd.DataFrame({

.....: 'time': pd.to_datetime(['20160525 13:30:00.023',
.....: '20160525 13:30:00.038',
.....: '20160525 13:30:00.048',
.....: '20160525 13:30:00.048',
.....: '20160525 13:30:00.048']),
.....: 'ticker': ['MSFT', 'MSFT',
.....: 'GOOG', 'GOOG', 'AAPL'],
.....: 'price': [51.95, 51.95,
.....: 720.77, 720.92, 98.00],
.....: 'quantity': [75, 155,
.....: 100, 100, 100]},
.....: columns=['time', 'ticker', 'price', 'quantity'])
.....:

In [134]: quotes = pd.DataFrame({

.....: 'time': pd.to_datetime(['20160525 13:30:00.023',
.....: '20160525 13:30:00.023',
.....: '20160525 13:30:00.030',
.....: '20160525 13:30:00.041',
.....: '20160525 13:30:00.048',
.....: '20160525 13:30:00.049',
.....: '20160525 13:30:00.072',
.....: '20160525 13:30:00.075']),
.....: 'ticker': ['GOOG', 'MSFT', 'MSFT',
.....: 'MSFT', 'GOOG', 'AAPL', 'GOOG',
.....: 'MSFT'],
.....: 'bid': [720.50, 51.95, 51.97, 51.99,
.....: 720.50, 97.99, 720.50, 52.01],
.....: 'ask': [720.93, 51.96, 51.98, 52.00,
.....: 720.93, 98.01, 720.88, 52.03]},
.....: columns=['time', 'ticker', 'bid', 'ask'])
.....:

In [135]: trades
Out[135]:
time ticker price quantity
0 2016-05-25 13:30:00.023 MSFT 51.95 75
1 2016-05-25 13:30:00.038 MSFT 51.95 155
2 2016-05-25 13:30:00.048 GOOG 720.77 100
3 2016-05-25 13:30:00.048 GOOG 720.92 100
4 2016-05-25 13:30:00.048 AAPL 98.00 100

In [136]: quotes
Out[136]:
time ticker bid ask
0 2016-05-25 13:30:00.023 GOOG 720.50 720.93
1 2016-05-25 13:30:00.023 MSFT 51.95 51.96
2 2016-05-25 13:30:00.030 MSFT 51.97 51.98
3 2016-05-25 13:30:00.041 MSFT 51.99 52.00
4 2016-05-25 13:30:00.048 GOOG 720.50 720.93
5 2016-05-25 13:30:00.049 AAPL 97.99 98.01
6 2016-05-25 13:30:00.072 GOOG 720.50 720.88
7 2016-05-25 13:30:00.075 MSFT 52.01 52.03
By default we are taking the asof of the quotes.

430 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [137]: pd.merge_asof(trades, quotes,

.....: on='time',
.....: by='ticker')
.....:
Out[137]:
time ticker price quantity bid ask
0 2016-05-25 13:30:00.023 MSFT 51.95 75 51.95 51.96
1 2016-05-25 13:30:00.038 MSFT 51.95 155 51.97 51.98
2 2016-05-25 13:30:00.048 GOOG 720.77 100 720.50 720.93
3 2016-05-25 13:30:00.048 GOOG 720.92 100 720.50 720.93
4 2016-05-25 13:30:00.048 AAPL 98.00 100 NaN NaN

We only asof within 2ms between the quote time and the trade time.

In [138]: pd.merge_asof(trades, quotes,

.....: on='time',
.....: by='ticker',
.....: tolerance=pd.Timedelta('2ms'))
.....:
Out[138]:
time ticker price quantity bid ask
0 2016-05-25 13:30:00.023 MSFT 51.95 75 51.95 51.96
1 2016-05-25 13:30:00.038 MSFT 51.95 155 NaN NaN
2 2016-05-25 13:30:00.048 GOOG 720.77 100 720.50 720.93
3 2016-05-25 13:30:00.048 GOOG 720.92 100 720.50 720.93
4 2016-05-25 13:30:00.048 AAPL 98.00 100 NaN NaN

We only asof within 10ms between the quote time and the trade time and we exclude exact matches on time. Note
that though we exclude the exact matches (of the quotes), prior quotes do propagate to that point in time.

In [139]: pd.merge_asof(trades, quotes,

.....: on='time',
.....: by='ticker',
.....: tolerance=pd.Timedelta('10ms'),
.....: allow_exact_matches=False)
.....:
Out[139]:
time ticker price quantity bid ask
0 2016-05-25 13:30:00.023 MSFT 51.95 75 NaN NaN
1 2016-05-25 13:30:00.038 MSFT 51.95 155 51.97 51.98
2 2016-05-25 13:30:00.048 GOOG 720.77 100 NaN NaN
3 2016-05-25 13:30:00.048 GOOG 720.92 100 NaN NaN
4 2016-05-25 13:30:00.048 AAPL 98.00 100 NaN NaN

{{ header }}

4.4. Merge, join, and concatenate 431

pandas: powerful Python data analysis toolkit, Release 0.25.3

4.5 Reshaping and pivot tables

4.5.1 Reshaping by pivoting DataFrame objects

Data is often stored in so-called stacked or record format:

In [1]: df
Out[1]:
date variable value
0 2000-01-03 A 0.259126
1 2000-01-04 A 0.774683
2 2000-01-05 A 0.713390
3 2000-01-03 B -0.837638
4 2000-01-04 B 0.064089
5 2000-01-05 B -0.034934
6 2000-01-03 C -0.452395
7 2000-01-04 C 2.488398
8 2000-01-05 C 1.429389
9 2000-01-03 D 0.547623
10 2000-01-04 D 0.008959
11 2000-01-05 D 1.097237

For the curious here is how the above DataFrame was created:

432 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

import pandas.util.testing as tm

tm.N = 3

def unpivot(frame):
N, K = frame.shape
data = {'value': frame.to_numpy().ravel('F'),
'variable': np.asarray(frame.columns).repeat(N),
'date': np.tile(np.asarray(frame.index), K)}
return pd.DataFrame(data, columns=['date', 'variable', 'value'])

df = unpivot(tm.makeTimeDataFrame())

To select out everything for variable A we could do:

In [2]: df[df['variable'] == 'A']
Out[2]:
date variable value
0 2000-01-03 A 0.259126
1 2000-01-04 A 0.774683
2 2000-01-05 A 0.713390

But suppose we wish to do time series operations with the variables. A better representation would be where the
columns are the unique variables and an index of dates identifies individual observations. To reshape the data into
this form, we use the DataFrame.pivot() method (also implemented as a top level function pivot()):
In [3]: df.pivot(index='date', columns='variable', values='value')
Out[3]:
variable A B C D
date
2000-01-03 0.259126 -0.837638 -0.452395 0.547623
2000-01-04 0.774683 0.064089 2.488398 0.008959
2000-01-05 0.713390 -0.034934 1.429389 1.097237

If the values argument is omitted, and the input DataFrame has more than one column of values which are not
used as column or index inputs to pivot, then the resulting pivoted DataFrame will have hierarchical columns
whose topmost level indicates the respective value column:
In [4]: df['value2'] = df['value'] * 2

In [5]: pivoted = df.pivot(index='date', columns='variable')

In [6]: pivoted
Out[6]:
value value2
,→

variable A B C D A B C
,→ D
date
,→

2000-01-03 0.259126 -0.837638 -0.452395 0.547623 0.518252 -1.675276 -0.904790 1.

,→095246

2000-01-04 0.774683 0.064089 2.488398 0.008959 1.549366 0.128178 4.976796 0.

,→017919

2000-01-05 0.713390 -0.034934 1.429389 1.097237 1.426780 -0.069868 2.858779 2.

,→194473
(continues on next page)

4.5. Reshaping and pivot tables 433

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

You can then select subsets from the pivoted DataFrame:

In [7]: pivoted['value2']
Out[7]:
variable A B C D
date
2000-01-03 0.518252 -1.675276 -0.904790 1.095246
2000-01-04 1.549366 0.128178 4.976796 0.017919
2000-01-05 1.426780 -0.069868 2.858779 2.194473

Note that this returns a view on the underlying data in the case where the data are homogeneously-typed.

Note: pivot() will error with a ValueError: Index contains duplicate entries, cannot
reshape if the index/column pair is not unique. In this case, consider using pivot_table() which is a general-
ization of pivot that can handle duplicate values for one index/column pair.

434 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

4.5.2 Reshaping by stacking and unstacking

Closely related to the pivot() method are the related stack() and unstack() methods available on Series
and DataFrame. These methods are designed to work together with MultiIndex objects (see the section on
hierarchical indexing). Here are essentially what these methods do:
• stack: pivot a level of the (possibly hierarchical) column labels, returning a DataFrame with an index with
a new inner-most level of row labels.
• unstack: (inverse operation of stack) pivot a level of the (possibly hierarchical) row index to the column
axis, producing a reshaped DataFrame with a new inner-most level of column labels.

4.5. Reshaping and pivot tables 435

pandas: powerful Python data analysis toolkit, Release 0.25.3

The clearest way to explain is by example. Lets take a prior example data set from the hierarchical indexing section:

In [8]: tuples = list(zip(*[['bar', 'bar', 'baz', 'baz',

...: 'foo', 'foo', 'qux', 'qux'],
...: ['one', 'two', 'one', 'two',
...: 'one', 'two', 'one', 'two']]))
...:

In [9]: index = pd.MultiIndex.from_tuples(tuples, names=['first', 'second'])

In [10]: df = pd.DataFrame(np.random.randn(8, 2), index=index, columns=['A', 'B'])

In [11]: df2 = df[:4]

In [12]: df2
Out[12]:
A B
first second
bar one -1.427767 1.011174
two -0.227837 0.260297
(continues on next page)

436 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

baz one -0.664499 -1.085553
two -1.392521 -0.426500

The stack function compresses a level in the DataFrames columns to produce either:
• A Series, in the case of a simple column Index.
• A DataFrame, in the case of a MultiIndex in the columns.
If the columns have a MultiIndex, you can choose which level to stack. The stacked level becomes the new lowest
level in a MultiIndex on the columns:

In [13]: stacked = df2.stack()

In [14]: stacked
Out[14]:
first second
bar one A -1.427767
B 1.011174
two A -0.227837
B 0.260297
baz one A -0.664499
B -1.085553
two A -1.392521
B -0.426500
dtype: float64

With a stacked DataFrame or Series (having a MultiIndex as the index), the inverse operation of stack is
unstack, which by default unstacks the last level:
In [15]: stacked.unstack()
Out[15]:
A B
first second
bar one -1.427767 1.011174
two -0.227837 0.260297
baz one -0.664499 -1.085553
two -1.392521 -0.426500

In [16]: stacked.unstack(1)
Out[16]:
second one two
first
bar A -1.427767 -0.227837
B 1.011174 0.260297
baz A -0.664499 -1.392521
B -1.085553 -0.426500

In [17]: stacked.unstack(0)
Out[17]:
first bar baz
second
one A -1.427767 -0.664499
B 1.011174 -1.085553
two A -0.227837 -1.392521
B 0.260297 -0.426500

4.5. Reshaping and pivot tables 437

pandas: powerful Python data analysis toolkit, Release 0.25.3

If the indexes have names, you can use the level names instead of specifying the level numbers:

In [18]: stacked.unstack('second')
Out[18]:
second one two
first
bar A -1.427767 -0.227837
B 1.011174 0.260297
baz A -0.664499 -1.392521
B -1.085553 -0.426500

438 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Notice that the stack and unstack methods implicitly sort the index levels involved. Hence a call to stack and
then unstack, or vice versa, will result in a sorted copy of the original DataFrame or Series:
In [19]: index = pd.MultiIndex.from_product([[2, 1], ['a', 'b']])

In [20]: df = pd.DataFrame(np.random.randn(4), index=index, columns=['A'])

In [21]: df
Out[21]:
A
2 a -1.727109
b -1.966122
1 a -0.004308
b 0.627249

In [22]: all(df.unstack().stack() == df.sort_index())

Out[22]: True
The above code will raise a TypeError if the call to sort_index is removed.

4.5. Reshaping and pivot tables 439

pandas: powerful Python data analysis toolkit, Release 0.25.3

Multiple levels

You may also stack or unstack more than one level at a time by passing a list of levels, in which case the end result is
as if each level in the list were processed individually.
In [23]: columns = pd.MultiIndex.from_tuples([
....: ('A', 'cat', 'long'), ('B', 'cat', 'long'),
....: ('A', 'dog', 'short'), ('B', 'dog', 'short')],
....: names=['exp', 'animal', 'hair_length']
....: )
....:

In [24]: df = pd.DataFrame(np.random.randn(4, 4), columns=columns)

In [25]: df
Out[25]:
exp A B A B
animal cat cat dog dog
hair_length long long short short
0 -2.260817 0.023953 -1.328037 -0.091360
1 -0.063483 0.691641 1.062240 -1.912934
2 -0.967661 1.438160 1.796396 0.364482
3 0.384514 0.774313 -1.215737 1.533214

In [26]: df.stack(level=['animal', 'hair_length'])

Out[26]:
exp A B
animal hair_length
0 cat long -2.260817 0.023953
dog short -1.328037 -0.091360
1 cat long -0.063483 0.691641
dog short 1.062240 -1.912934
2 cat long -0.967661 1.438160
dog short 1.796396 0.364482
3 cat long 0.384514 0.774313
dog short -1.215737 1.533214
The list of levels can contain either level names or level numbers (but not a mixture of the two).

# df.stack(level=['animal', 'hair_length'])
# from above is equivalent to:
In [27]: df.stack(level=[1, 2])
Out[27]:
exp A B
animal hair_length
0 cat long -2.260817 0.023953
dog short -1.328037 -0.091360
1 cat long -0.063483 0.691641
dog short 1.062240 -1.912934
2 cat long -0.967661 1.438160
dog short 1.796396 0.364482
3 cat long 0.384514 0.774313
dog short -1.215737 1.533214

440 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Missing data

These functions are intelligent about handling missing data and do not expect each subgroup within the hierarchical
index to have the same set of labels. They also can handle the index being unsorted (but you can make it sorted by
calling sort_index, of course). Here is a more complex example:

In [28]: columns = pd.MultiIndex.from_tuples([('A', 'cat'), ('B', 'dog'),

....: ('B', 'cat'), ('A', 'dog')],
....: names=['exp', 'animal'])
....:

In [29]: index = pd.MultiIndex.from_product([('bar', 'baz', 'foo', 'qux'),

....: ('one', 'two')],
....: names=['first', 'second'])
....:

In [30]: df = pd.DataFrame(np.random.randn(8, 4), index=index, columns=columns)

In [31]: df2 = df.iloc[[0, 1, 2, 4, 5, 7]]

In [32]: df2
Out[32]:
exp A B A
animal cat dog cat dog
first second
bar one -1.436953 0.561814 -0.346880 -0.546060
two -1.993054 -0.689365 -0.877031 1.507935
baz one -1.866380 0.043384 0.252683 -0.479004
foo one -1.613149 -0.622599 0.291003 0.792238
two 0.807151 -0.758613 -2.393856 1.098272
qux two -0.411186 1.584705 -1.042868 -0.295906

As mentioned above, stack can be called with a level argument to select which level in the columns to stack:
In [33]: df2.stack('exp')
Out[33]:
animal cat dog
first second exp
bar one A -1.436953 -0.546060
B -0.346880 0.561814
two A -1.993054 1.507935
B -0.877031 -0.689365
baz one A -1.866380 -0.479004
B 0.252683 0.043384
foo one A -1.613149 0.792238
B 0.291003 -0.622599
two A 0.807151 1.098272
B -2.393856 -0.758613
qux two A -0.411186 -0.295906
B -1.042868 1.584705

In [34]: df2.stack('animal')
Out[34]:
exp A B
first second animal
bar one cat -1.436953 -0.346880
dog -0.546060 0.561814

4.5. Reshaping and pivot tables 441

pandas: powerful Python data analysis toolkit, Release 0.25.3

two cat -1.993054 -0.877031

dog 1.507935 -0.689365
baz one cat -1.866380 0.252683
dog -0.479004 0.043384
foo one cat -1.613149 0.291003
dog 0.792238 -0.622599
two cat 0.807151 -2.393856
dog 1.098272 -0.758613
qux two cat -0.411186 -1.042868
dog -0.295906 1.584705
Unstacking can result in missing values if subgroups do not have the same set of labels. By default, missing values
will be replaced with the default fill value for that data type, NaN for float, NaT for datetimelike, etc. For integer types,
by default data will converted to float and missing values will be set to NaN.
In [35]: df3 = df.iloc[[0, 1, 4, 7], [1, 2]]

In [36]: df3
Out[36]:
exp B
animal dog cat
first second
bar one 0.561814 -0.346880
two -0.689365 -0.877031
foo one -0.622599 0.291003
qux two 1.584705 -1.042868

In [37]: df3.unstack()
Out[37]:
exp B
animal dog cat
second one two one two
first
bar 0.561814 -0.689365 -0.346880 -0.877031
foo -0.622599 NaN 0.291003 NaN
qux NaN 1.584705 NaN -1.042868
New in version 0.18.0.
Alternatively, unstack takes an optional fill_value argument, for specifying the value of missing data.

In [38]: df3.unstack(fill_value=-1e9)
Out[38]:
exp B
animal dog cat
second one two one two
first
bar 5.618142e-01 -6.893650e-01 -3.468802e-01 -8.770315e-01
foo -6.225992e-01 -1.000000e+09 2.910029e-01 -1.000000e+09
qux -1.000000e+09 1.584705e+00 -1.000000e+09 -1.042868e+00

With a MultiIndex

Unstacking when the columns are a MultiIndex is also careful about doing the right thing:
In [39]: df[:3].unstack(0)

442 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[39]:
exp A B A
animal cat dog cat dog
first bar baz bar baz bar baz bar
,→ baz
second
one -1.436953 -1.86638 0.561814 0.043384 -0.346880 0.252683 -0.546060 -
,→0.479004

two -1.993054 NaN -0.689365 NaN -0.877031 NaN 1.507935

,→ NaN

In [40]: df2.unstack(1)
Out[40]:
exp A B A
animal cat dog cat dog
second one two one two one two one
,→ two
first
bar -1.436953 -1.993054 0.561814 -0.689365 -0.346880 -0.877031 -0.546060
,→1.507935

baz -1.866380 NaN 0.043384 NaN 0.252683 NaN -0.479004

,→ NaN
foo -1.613149 0.807151 -0.622599 -0.758613 0.291003 -2.393856 0.792238
,→1.098272

qux NaN -0.411186 NaN 1.584705 NaN -1.042868 NaN -

,→0.295906

4.5. Reshaping and pivot tables 443

pandas: powerful Python data analysis toolkit, Release 0.25.3

4.5.3 Reshaping by Melt

The top-level melt() function and the corresponding DataFrame.melt() are useful to massage a DataFrame
into a format where one or more columns are identifier variables, while all other columns, considered measured
variables, are unpivoted to the row axis, leaving just two non-identifier columns, variable and value. The names of
those columns can be customized by supplying the var_name and value_name parameters.
For instance,
In [41]: cheese = pd.DataFrame({'first': ['John', 'Mary'],
....: 'last': ['Doe', 'Bo'],
....: 'height': [5.5, 6.0],
....: 'weight': [130, 150]})
....:

In [42]: cheese
Out[42]:
first last height weight
0 John Doe 5.5 130
1 Mary Bo 6.0 150

In [43]: cheese.melt(id_vars=['first', 'last'])

Out[43]:
first last variable value

444 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

0 John Doe height 5.5

1 Mary Bo height 6.0
2 John Doe weight 130.0
3 Mary Bo weight 150.0

In [44]: cheese.melt(id_vars=['first', 'last'], var_name='quantity')

Out[44]:
first last quantity value
0 John Doe height 5.5
1 Mary Bo height 6.0
2 John Doe weight 130.0
3 Mary Bo weight 150.0
Another way to transform is to use the wide_to_long() panel data convenience function. It is less flexible than
melt(), but more user-friendly.
In [45]: dft = pd.DataFrame({"A1970": {0: "a", 1: "b", 2: "c"},
....: "A1980": {0: "d", 1: "e", 2: "f"},
....: "B1970": {0: 2.5, 1: 1.2, 2: .7},
....: "B1980": {0: 3.2, 1: 1.3, 2: .1},
....: "X": dict(zip(range(3), np.random.randn(3)))
....: })
....:

In [46]: dft["id"] = dft.index

In [47]: dft
Out[47]:
A1970 A1980 B1970 B1980 X id
0 a d 2.5 3.2 -2.013095 0
1 b e 1.2 1.3 -1.711797 1
2 c f 0.7 0.1 0.975018 2

In [48]: pd.wide_to_long(dft, ["A", "B"], i="id", j="year")

Out[48]:
X A B
id year
0 1970 -2.013095 a 2.5
1 1970 -1.711797 b 1.2
2 1970 0.975018 c 0.7
0 1980 -2.013095 d 3.2
1 1980 -1.711797 e 1.3
2 1980 0.975018 f 0.1

4.5.4 Combining with stats and GroupBy

It should be no shock that combining pivot / stack / unstack with GroupBy and the basic Series and DataFrame
statistical functions can produce some very expressive and fast data manipulations.
In [49]: df
Out[49]:
exp A B A
animal cat dog cat dog
first second
bar one -1.436953 0.561814 -0.346880 -0.546060

4.5. Reshaping and pivot tables 445

pandas: powerful Python data analysis toolkit, Release 0.25.3

two -1.993054 -0.689365 -0.877031 1.507935

baz one -1.866380 0.043384 0.252683 -0.479004
two -0.190425 -1.520207 1.697480 0.923916
foo one -1.613149 -0.622599 0.291003 0.792238
two 0.807151 -0.758613 -2.393856 1.098272
qux one -0.168090 -0.243576 0.316513 -0.633213
two -0.411186 1.584705 -1.042868 -0.295906

In [50]: df.stack().mean(1).unstack()
Out[50]:
animal cat dog
first second
bar one -0.891917 0.007877
two -1.435043 0.409285
baz one -0.806849 -0.217810
two 0.753528 -0.298146
foo one -0.661073 0.084819
two -0.793352 0.169830
qux one 0.074211 -0.438394
two -0.727027 0.644399

# same result, another way

In [51]: df.groupby(level=1, axis=1).mean()
Out[51]:
animal cat dog
first second
bar one -0.891917 0.007877
two -1.435043 0.409285
baz one -0.806849 -0.217810
two 0.753528 -0.298146
foo one -0.661073 0.084819
two -0.793352 0.169830
qux one 0.074211 -0.438394
two -0.727027 0.644399

In [52]: df.stack().groupby(level=1).mean()
Out[52]:
exp A B
second
one -0.743827 0.031543
two 0.180838 -0.499969

In [53]: df.mean().unstack(0)
Out[53]:
exp A B
animal
cat -0.859011 -0.262870
dog 0.296022 -0.205557

4.5.5 Pivot tables

While pivot() provides general purpose pivoting with various data types (strings, numerics, etc.), pandas also
provides pivot_table() for pivoting with aggregation of numeric data.

446 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

The function pivot_table() can be used to create spreadsheet-style pivot tables. See the cookbook for some
advanced strategies.
It takes a number of arguments:
• data: a DataFrame object.
• values: a column or a list of columns to aggregate.
• index: a column, Grouper, array which has the same length as data, or list of them. Keys to group by on the
pivot table index. If an array is passed, it is being used as the same manner as column values.
• columns: a column, Grouper, array which has the same length as data, or list of them. Keys to group by on
the pivot table column. If an array is passed, it is being used as the same manner as column values.
• aggfunc: function to use for aggregation, defaulting to numpy.mean.
Consider a data set like this:

In [54]: import datetime

In [55]: df = pd.DataFrame({'A': ['one', 'one', 'two', 'three'] * 6,

....: 'B': ['A', 'B', 'C'] * 8,
....: 'C': ['foo', 'foo', 'foo', 'bar', 'bar', 'bar'] * 4,
....: 'D': np.random.randn(24),
....: 'E': np.random.randn(24),
....: 'F': [datetime.datetime(2013, i, 1) for i in range(1, 13)]
....: + [datetime.datetime(2013, i, 15) for i in range(1, 13)]})
....:

In [56]: df
Out[56]:
A B C D E F
0 one A foo 1.772891 -1.594231 2013-01-01
1 one B foo 0.498050 -1.165583 2013-02-01
2 two C foo 0.476510 0.451407 2013-03-01
3 three A bar 0.877022 -1.724523 2013-04-01
4 one B bar -0.301831 2.097509 2013-05-01
5 one C bar -0.733711 1.078343 2013-06-01
6 two A foo 0.826360 -0.394125 2013-07-01
7 three B foo 0.466766 -0.726083 2013-08-01
8 one C foo 1.634589 0.490827 2013-09-01
9 one A bar 0.958132 -0.559443 2013-10-01
10 two B bar -0.970543 -0.395876 2013-11-01
11 three C bar -0.932172 -0.927091 2013-12-01
12 one A foo -1.273567 -0.417776 2013-01-15
13 one B foo 0.270030 -0.708679 2013-02-15
14 two C foo -1.388749 -1.557855 2013-03-15
15 three A bar 2.090035 -1.742504 2013-04-15
16 one B bar -0.815984 -0.201498 2013-05-15
17 one C bar -0.020855 0.070382 2013-06-15
18 two A foo -0.108292 -0.749683 2013-07-15
19 three B foo 0.093700 1.434318 2013-08-15
20 one C foo -0.851281 1.226799 2013-09-15
21 one A bar -2.154727 -0.706492 2013-10-15
22 two B bar -0.120914 0.299114 2013-11-15
23 three C bar -0.976377 -0.081937 2013-12-15

We can produce pivot tables from this data very easily:

In [57]: pd.pivot_table(df, values='D', index=['A', 'B'], columns=['C'])

4.5. Reshaping and pivot tables 447

pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[57]:
C bar foo
A B
one A -0.598297 0.249662
B -0.558907 0.384040
C -0.377283 0.391654
three A 1.483528 NaN
B NaN 0.280233
C -0.954274 NaN
two A NaN 0.359034
B -0.545728 NaN
C NaN -0.456119

In [58]: pd.pivot_table(df, values='D', index=['B'], columns=['A', 'C'],

,→aggfunc=np.sum)

Out[58]:
A one three two
C bar foo bar foo bar foo
B
A -1.196595 0.499324 2.967057 NaN NaN 0.718067
B -1.117815 0.768081 NaN 0.560466 -1.091457 NaN
C -0.754566 0.783308 -1.908549 NaN NaN -0.912238

In [59]: pd.pivot_table(df, values=['D', 'E'], index=['B'], columns=['A',

,→'C'],

....: aggfunc=np.sum)
....:
Out[59]:
D ... E
A one three two ... one three
,→ two
C bar foo bar foo bar ... foo bar
,→ foo bar foo
B ...
A -1.196595 0.499324 2.967057 NaN NaN ... -2.012007 -3.467026
,→ NaN NaN -1.143807
B -1.117815 0.768081 NaN 0.560466 -1.091457 ... -1.874262 NaN
,→0.708235 -0.096762 NaN
C -0.754566 0.783308 -1.908549 NaN NaN ... 1.717626 -1.009028
,→ NaN NaN -1.106447

[3 rows x 12 columns]
The result object is a DataFrame having potentially hierarchical indexes on the rows and columns. If the values
column name is not given, the pivot table will include all of the data that can be aggregated in an additional level of
hierarchy in the columns:

In [60]: pd.pivot_table(df, index=['A', 'B'], columns=['C'])

Out[60]:
D E
C bar foo bar foo
A B
one A -0.598297 0.249662 -0.632968 -1.006004
B -0.558907 0.384040 0.948005 -0.937131
(continues on next page)

448 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

C -0.377283 0.391654 0.574362 0.858813
three A 1.483528 NaN -1.733513 NaN
B NaN 0.280233 NaN 0.354117
C -0.954274 NaN -0.504514 NaN
two A NaN 0.359034 NaN -0.571904
B -0.545728 NaN -0.048381 NaN
C NaN -0.456119 NaN -0.553224

Also, you can use Grouper for index and columns keywords. For detail of Grouper, see Grouping with a
Grouper specification.

In [61]: pd.pivot_table(df, values='D', index=pd.Grouper(freq='M', key='F'),

....: columns='C')
....:
Out[61]:
C bar foo
F
2013-01-31 NaN 0.249662
2013-02-28 NaN 0.384040
2013-03-31 NaN -0.456119
2013-04-30 1.483528 NaN
2013-05-31 -0.558907 NaN
2013-06-30 -0.377283 NaN
2013-07-31 NaN 0.359034
2013-08-31 NaN 0.280233
2013-09-30 NaN 0.391654
2013-10-31 -0.598297 NaN
2013-11-30 -0.545728 NaN
2013-12-31 -0.954274 NaN

You can render a nice output of the table omitting the missing values by calling to_string if you wish:

In [62]: table = pd.pivot_table(df, index=['A', 'B'], columns=['C'])

In [63]: print(table.to_string(na_rep=''))
D E
C bar foo bar foo
A B
one A -0.598297 0.249662 -0.632968 -1.006004
B -0.558907 0.384040 0.948005 -0.937131
C -0.377283 0.391654 0.574362 0.858813
three A 1.483528 -1.733513
B 0.280233 0.354117
C -0.954274 -0.504514
two A 0.359034 -0.571904
B -0.545728 -0.048381
C -0.456119 -0.553224

Note that pivot_table is also available as an instance method on DataFrame, i.e. DataFrame.
pivot_table().

Adding margins

If you pass margins=True to pivot_table, special All columns and rows will be added with partial group
aggregates across the categories on the rows and columns:

4.5. Reshaping and pivot tables 449

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [64]: df.pivot_table(index=['A', 'B'], columns='C', margins=True, aggfunc=np.std)

Out[64]:
D E
C bar foo All bar foo All
A B
one A 2.201124 2.154171 1.844310 0.103979 0.831880 0.529777
B 0.363561 0.161235 0.590854 1.625643 0.323080 1.449234
C 0.504065 1.757775 1.145298 0.712737 0.520411 0.535329
three A 0.857730 NaN 0.857730 0.012714 NaN 0.012714
B NaN 0.263797 0.263797 NaN 1.527634 1.527634
C 0.031258 NaN 0.031258 0.597614 NaN 0.597614
two A NaN 0.660899 0.660899 NaN 0.251417 0.251417
B 0.600778 NaN 0.600778 0.491432 NaN 0.491432
C NaN 1.318937 1.318937 NaN 1.420763 1.420763
All 1.126982 1.003301 1.047076 1.082597 1.004895 1.000732

4.5.6 Cross tabulations

Use crosstab() to compute a cross-tabulation of two (or more) factors. By default crosstab computes a fre-
quency table of the factors unless an array of values and an aggregation function are passed.
It takes a number of arguments
• index: array-like, values to group by in the rows.
• columns: array-like, values to group by in the columns.
• values: array-like, optional, array of values to aggregate according to the factors.
• aggfunc: function, optional, If no values array is passed, computes a frequency table.
• rownames: sequence, default None, must match number of row arrays passed.
• colnames: sequence, default None, if passed, must match number of column arrays passed.
• margins: boolean, default False, Add row/column margins (subtotals)
• normalize: boolean, {all, index, columns}, or {0,1}, default False. Normalize by dividing all values by
the sum of values.
Any Series passed will have their name attributes used unless row or column names for the cross-tabulation are
specified
For example:

In [65]: foo, bar, dull, shiny, one, two = 'foo', 'bar', 'dull', 'shiny', 'one', 'two'

In [66]: a = np.array([foo, foo, bar, bar, foo, foo], dtype=object)

In [67]: b = np.array([one, one, two, one, two, one], dtype=object)

In [68]: c = np.array([dull, dull, shiny, dull, dull, shiny], dtype=object)

In [69]: pd.crosstab(a, [b, c], rownames=['a'], colnames=['b', 'c'])

Out[69]:
b one two
c dull shiny dull shiny
a
bar 1 0 0 1
foo 2 1 1 0

450 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

If crosstab receives only two Series, it will provide a frequency table.

In [70]: df = pd.DataFrame({'A': [1, 2, 2, 2, 2], 'B': [3, 3, 4, 4, 4],
....: 'C': [1, 1, np.nan, 1, 1]})
....:

In [71]: df
Out[71]:
A B C
0 1 3 1.0
1 2 3 1.0
2 2 4 NaN
3 2 4 1.0
4 2 4 1.0

In [72]: pd.crosstab(df.A, df.B)

Out[72]:
B 3 4
A
1 1 0
2 1 3
Any input passed containing Categorical data will have all of its categories included in the cross-tabulation, even
if the actual data does not contain any instances of a particular category.

In [73]: foo = pd.Categorical(['a', 'b'], categories=['a', 'b', 'c'])

In [74]: bar = pd.Categorical(['d', 'e'], categories=['d', 'e', 'f'])

In [75]: pd.crosstab(foo, bar)

Out[75]:
col_0 d e
row_0
a 1 0
b 0 1

Normalization

New in version 0.18.1.

Frequency tables can also be normalized to show percentages rather than counts using the normalize argument:

In [76]: pd.crosstab(df.A, df.B, normalize=True)

Out[76]:
B 3 4
A
1 0.2 0.0
2 0.2 0.6

normalize can also normalize values within each row or within each column:

In [77]: pd.crosstab(df.A, df.B, normalize='columns')

Out[77]:
B 3 4
A
(continues on next page)

4.5. Reshaping and pivot tables 451

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

1 0.5 0.0
2 0.5 1.0

crosstab can also be passed a third Series and an aggregation function (aggfunc) that will be applied to the
values of the third Series within each group defined by the first two Series:

In [78]: pd.crosstab(df.A, df.B, values=df.C, aggfunc=np.sum)

Out[78]:
B 3 4
A
1 1.0 NaN
2 1.0 2.0

Adding margins

Finally, one can also add margins or normalize this output.

In [79]: pd.crosstab(df.A, df.B, values=df.C, aggfunc=np.sum, normalize=True,

....: margins=True)
....:
Out[79]:
B 3 4 All
A
1 0.25 0.0 0.25
2 0.25 0.5 0.75
All 0.50 0.5 1.00

4.5.7 Tiling

The cut() function computes groupings for the values of the input array and is often used to transform continuous
variables to discrete or categorical variables:

In [80]: ages = np.array([10, 15, 13, 12, 23, 25, 28, 59, 60])

In [81]: pd.cut(ages, bins=3)

Out[81]:
[(9.95, 26.667], (9.95, 26.667], (9.95, 26.667], (9.95, 26.667], (9.95, 26.667], (9.
,→95, 26.667], (26.667, 43.333], (43.333, 60.0], (43.333, 60.0]]

Categories (3, interval[float64]): [(9.95, 26.667] < (26.667, 43.333] < (43.333, 60.
,→0]]

If the bins keyword is an integer, then equal-width bins are formed. Alternatively we can specify custom bin-edges:

In [82]: c = pd.cut(ages, bins=[0, 18, 35, 70])

In [83]: c
Out[83]:
[(0, 18], (0, 18], (0, 18], (0, 18], (18, 35], (18, 35], (18, 35], (35, 70], (35, 70]]
Categories (3, interval[int64]): [(0, 18] < (18, 35] < (35, 70]]

New in version 0.20.0.

If the bins keyword is an IntervalIndex, then these will be used to bin the passed data.:

452 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

pd.cut([25, 20, 50], bins=c.categories)

4.5.8 Computing indicator / dummy variables

To convert a categorical variable into a dummy or indicator DataFrame, for example a column in a DataFrame
(a Series) which has k distinct values, can derive a DataFrame containing k columns of 1s and 0s using
get_dummies():

In [84]: df = pd.DataFrame({'key': list('bbacab'), 'data1': range(6)})

In [85]: pd.get_dummies(df['key'])
Out[85]:
a b c
0 0 1 0
1 0 1 0
2 1 0 0
3 0 0 1
4 1 0 0
5 0 1 0

Sometimes its useful to prefix the column names, for example when merging the result with the original DataFrame:
In [86]: dummies = pd.get_dummies(df['key'], prefix='key')

In [87]: dummies
Out[87]:
key_a key_b key_c
0 0 1 0
1 0 1 0
2 1 0 0
3 0 0 1
4 1 0 0
5 0 1 0

In [88]: df[['data1']].join(dummies)
Out[88]:
data1 key_a key_b key_c
0 0 0 1 0
1 1 0 1 0
2 2 1 0 0
3 3 0 0 1
4 4 1 0 0
5 5 0 1 0
This function is often used along with discretization functions like cut:

In [89]: values = np.random.randn(10)

In [90]: values
Out[90]:
array([-0.7350435 , 1.42920375, -1.11519984, -0.97015174, -1.19270064,
0.02125661, -0.20556342, -0.66677255, 2.12277401, -0.10814128])

In [91]: bins = [0, 0.2, 0.4, 0.6, 0.8, 1]

(continues on next page)

4.5. Reshaping and pivot tables 453

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [92]: pd.get_dummies(pd.cut(values, bins))

Out[92]:
(0.0, 0.2] (0.2, 0.4] (0.4, 0.6] (0.6, 0.8] (0.8, 1.0]
0 0 0 0 0 0
1 0 0 0 0 0
2 0 0 0 0 0
3 0 0 0 0 0
4 0 0 0 0 0
5 1 0 0 0 0
6 0 0 0 0 0
7 0 0 0 0 0
8 0 0 0 0 0
9 0 0 0 0 0

See also Series.str.get_dummies.

get_dummies() also accepts a DataFrame. By default all categorical variables (categorical in the statistical
sense, those with object or categorical dtype) are encoded as dummy variables.

In [93]: df = pd.DataFrame({'A': ['a', 'b', 'a'], 'B': ['c', 'c', 'b'],

....: 'C': [1, 2, 3]})
....:

In [94]: pd.get_dummies(df)
Out[94]:
C A_a A_b B_b B_c
0 1 1 0 0 1
1 2 0 1 0 1
2 3 1 0 1 0

All non-object columns are included untouched in the output. You can control the columns that are encoded with the
columns keyword.

In [95]: pd.get_dummies(df, columns=['A'])

Out[95]:
B C A_a A_b
0 c 1 1 0
1 c 2 0 1
2 b 3 1 0

Notice that the B column is still included in the output, it just hasnt been encoded. You can drop B before calling
get_dummies if you dont want to include it in the output.
As with the Series version, you can pass values for the prefix and prefix_sep. By default the column name
is used as the prefix, and _ as the prefix separator. You can specify prefix and prefix_sep in 3 ways:
• string: Use the same value for prefix or prefix_sep for each column to be encoded.
• list: Must be the same length as the number of columns being encoded.
• dict: Mapping column name to prefix.

In [96]: simple = pd.get_dummies(df, prefix='new_prefix')

In [97]: simple
Out[97]:
C new_prefix_a new_prefix_b new_prefix_b new_prefix_c
(continues on next page)

454 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

0 1 1 0 0 1
1 2 0 1 0 1
2 3 1 0 1 0

In [98]: from_list = pd.get_dummies(df, prefix=['from_A', 'from_B'])

In [99]: from_list
Out[99]:
C from_A_a from_A_b from_B_b from_B_c
0 1 1 0 0 1
1 2 0 1 0 1
2 3 1 0 1 0

In [100]: from_dict = pd.get_dummies(df, prefix={'B': 'from_B', 'A': 'from_A'})

In [101]: from_dict
Out[101]:
C from_A_a from_A_b from_B_b from_B_c
0 1 1 0 0 1
1 2 0 1 0 1
2 3 1 0 1 0

New in version 0.18.0.

Sometimes it will be useful to only keep k-1 levels of a categorical variable to avoid collinearity when feeding the
result to statistical models. You can switch to this mode by turn on drop_first.
In [102]: s = pd.Series(list('abcaa'))

In [103]: pd.get_dummies(s)
Out[103]:
a b c
0 1 0 0
1 0 1 0
2 0 0 1
3 1 0 0
4 1 0 0

In [104]: pd.get_dummies(s, drop_first=True)

Out[104]:
b c
0 0 0
1 1 0
2 0 1
3 0 0
4 0 0
When a column contains only one level, it will be omitted in the result.
In [105]: df = pd.DataFrame({'A': list('aaaaa'), 'B': list('ababc')})

In [106]: pd.get_dummies(df)
Out[106]:
A_a B_a B_b B_c
0 1 1 0 0
1 1 0 1 0

4.5. Reshaping and pivot tables 455

pandas: powerful Python data analysis toolkit, Release 0.25.3

2 1 1 0 0
3 1 0 1 0
4 1 0 0 1

In [107]: pd.get_dummies(df, drop_first=True)

Out[107]:
B_b B_c
0 0 0
1 1 0
2 0 0
3 1 0
4 0 1
By default new columns will have np.uint8 dtype. To choose another dtype, use the dtype argument:

In [108]: df = pd.DataFrame({'A': list('abc'), 'B': [1.1, 2.2, 3.3]})

In [109]: pd.get_dummies(df, dtype=bool).dtypes

Out[109]:
B float64
A_a bool
A_b bool
A_c bool
dtype: object

New in version 0.23.0.

4.5.9 Factorizing values

To encode 1-d values as an enumerated type use factorize():

In [110]: x = pd.Series(['A', 'A', np.nan, 'B', 3.14, np.inf])

In [111]: x
Out[111]:
0 A
1 A
2 NaN
3 B
4 3.14
5 inf
dtype: object

In [112]: labels, uniques = pd.factorize(x)

In [113]: labels
Out[113]: array([ 0, 0, -1, 1, 2, 3])

In [114]: uniques
Out[114]: Index(['A', 'B', 3.14, inf], dtype='object')
Note that factorize is similar to numpy.unique, but differs in its handling of NaN:

Note: The following numpy.unique will fail under Python 3 with a TypeError because of an ordering bug. See

456 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

also here.

In [1]: x = pd.Series(['A', 'A', np.nan, 'B', 3.14, np.inf])

In [2]: pd.factorize(x, sort=True)
Out[2]:
(array([ 2, 2, -1, 3, 0, 1]),
Index([3.14, inf, 'A', 'B'], dtype='object'))

In [3]: np.unique(x, return_inverse=True)[::-1]

Out[3]: (array([3, 3, 0, 4, 1, 2]), array([nan, 3.14, inf, 'A', 'B'], dtype=object))

Note: If you just want to handle one column as a categorical variable (like Rs factor), you can use df["cat_col"]
= pd.Categorical(df["col"]) or df["cat_col"] = df["col"].astype("category"). For
full docs on Categorical, see the Categorical introduction and the API documentation.

4.5.10 Examples

In this section, we will review frequently asked questions and examples. The column names and relevant column
values are named to correspond with how this DataFrame will be pivoted in the answers below.

In [115]: np.random.seed([3, 1415])

In [116]: n = 20

In [117]: cols = np.array(['key', 'row', 'item', 'col'])

In [118]: df = cols + pd.DataFrame((np.random.randint(5, size=(n, 4))

.....: // [2, 1, 2, 1]).astype(str))
.....:

In [119]: df.columns = cols

In [120]: df = df.join(pd.DataFrame(np.random.rand(n, 2).round(2)).add_prefix('val'))

In [121]: df
Out[121]:
key row item col val0 val1
0 key0 row3 item1 col3 0.81 0.04
1 key1 row2 item1 col2 0.44 0.07
2 key1 row0 item1 col0 0.77 0.01
3 key0 row4 item0 col2 0.15 0.59
4 key1 row0 item2 col1 0.81 0.64
5 key1 row2 item2 col4 0.13 0.88
6 key2 row4 item1 col3 0.88 0.39
7 key1 row4 item1 col1 0.10 0.07
8 key1 row0 item2 col4 0.65 0.02
9 key1 row2 item0 col2 0.35 0.61
10 key2 row0 item2 col1 0.40 0.85
11 key2 row4 item1 col2 0.64 0.25
12 key0 row2 item2 col3 0.50 0.44
13 key0 row4 item1 col4 0.24 0.46
14 key1 row3 item2 col3 0.28 0.11
15 key0 row3 item1 col1 0.31 0.23
(continues on next page)

4.5. Reshaping and pivot tables 457

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

16 key0 row0 item2 col3 0.86 0.01
17 key0 row4 item0 col3 0.64 0.21
18 key2 row2 item2 col0 0.13 0.45
19 key0 row2 item0 col4 0.37 0.70

Pivoting with single aggregations

Suppose we wanted to pivot df such that the col values are columns, row values are the index, and the mean of
val0 are the values? In particular, the resulting DataFrame should look like:

Note: col col0 col1 col2 col3 col4 row row0 0.77 0.605 NaN 0.860 0.65 row2 0.13 NaN 0.395 0.500 0.25 row3 NaN
0.310 NaN 0.545 NaN row4 NaN 0.100 0.395 0.760 0.24

This solution uses pivot_table(). Also note that aggfunc='mean' is the default. It is included here to be
explicit.

In [122]: df.pivot_table(
.....: values='val0', index='row', columns='col', aggfunc='mean')
.....:
Out[122]:
col col0 col1 col2 col3 col4
row
row0 0.77 0.605 NaN 0.860 0.65
row2 0.13 NaN 0.395 0.500 0.25
row3 NaN 0.310 NaN 0.545 NaN
row4 NaN 0.100 0.395 0.760 0.24

Note that we can also replace the missing values by using the fill_value parameter.

In [123]: df.pivot_table(
.....: values='val0', index='row', columns='col', aggfunc='mean', fill_value=0)
.....:
Out[123]:
col col0 col1 col2 col3 col4
row
row0 0.77 0.605 0.000 0.860 0.65
row2 0.13 0.000 0.395 0.500 0.25
row3 0.00 0.310 0.000 0.545 0.00
row4 0.00 0.100 0.395 0.760 0.24

Also note that we can pass in other aggregation functions as well. For example, we can also pass in sum.

In [124]: df.pivot_table(
.....: values='val0', index='row', columns='col', aggfunc='sum', fill_value=0)
.....:
Out[124]:
col col0 col1 col2 col3 col4
row
row0 0.77 1.21 0.00 0.86 0.65
row2 0.13 0.00 0.79 0.50 0.50
row3 0.00 0.31 0.00 1.09 0.00
row4 0.00 0.10 0.79 1.52 0.24

458 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Another aggregation we can do is calculate the frequency in which the columns and rows occur together a.k.a. cross
tabulation. To do this, we can pass size to the aggfunc parameter.

In [125]: df.pivot_table(index='row', columns='col', fill_value=0, aggfunc='size')

Out[125]:
col col0 col1 col2 col3 col4
row
row0 1 2 0 1 1
row2 1 0 2 1 2
row3 0 1 0 2 0
row4 0 1 2 2 1

Pivoting with multiple aggregations

We can also perform multiple aggregations. For example, to perform both a sum and mean, we can pass in a list to
the aggfunc argument.

In [126]: df.pivot_table(
.....: values='val0', index='row', columns='col', aggfunc=['mean', 'sum'])
.....:
Out[126]:
mean sum
col col0 col1 col2 col3 col4 col0 col1 col2 col3 col4
row
row0 0.77 0.605 NaN 0.860 0.65 0.77 1.21 NaN 0.86 0.65
row2 0.13 NaN 0.395 0.500 0.25 0.13 NaN 0.79 0.50 0.50
row3 NaN 0.310 NaN 0.545 NaN NaN 0.31 NaN 1.09 NaN
row4 NaN 0.100 0.395 0.760 0.24 NaN 0.10 0.79 1.52 0.24

Note to aggregate over multiple value columns, we can pass in a list to the values parameter.

In [127]: df.pivot_table(
.....: values=['val0', 'val1'], index='row', columns='col', aggfunc=['mean'])
.....:
Out[127]:
mean
val0 val1
col col0 col1 col2 col3 col4 col0 col1 col2 col3 col4
row
row0 0.77 0.605 NaN 0.860 0.65 0.01 0.745 NaN 0.010 0.02
row2 0.13 NaN 0.395 0.500 0.25 0.45 NaN 0.34 0.440 0.79
row3 NaN 0.310 NaN 0.545 NaN NaN 0.230 NaN 0.075 NaN
row4 NaN 0.100 0.395 0.760 0.24 NaN 0.070 0.42 0.300 0.46

Note to subdivide over multiple columns we can pass in a list to the columns parameter.

In [128]: df.pivot_table(
.....: values=['val0'], index='row', columns=['item', 'col'], aggfunc=['mean'])
.....:
Out[128]:
mean
val0
item item0 item1 item2
col col2 col3 col4 col0 col1 col2 col3 col4 col0 col1 col3 col4
row
row0 NaN NaN NaN 0.77 NaN NaN NaN NaN NaN 0.605 0.86 0.65
(continues on next page)

4.5. Reshaping and pivot tables 459

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

row2 0.35 NaN 0.37 NaN NaN 0.44 NaN NaN 0.13 NaN 0.50 0.13
row3 NaN NaN NaN NaN 0.31 NaN 0.81 NaN NaN NaN 0.28 NaN
row4 0.15 0.64 NaN NaN 0.10 0.64 0.88 0.24 NaN NaN NaN NaN

4.5.11 Exploding a list-like column

New in version 0.25.0.

Sometimes the values in a column are list-like.

In [129]: keys = ['panda1', 'panda2', 'panda3']

In [130]: values = [['eats', 'shoots'], ['shoots', 'leaves'], ['eats', 'leaves']]

In [131]: df = pd.DataFrame({'keys': keys, 'values': values})

In [132]: df
Out[132]:
keys values
0 panda1 [eats, shoots]
1 panda2 [shoots, leaves]
2 panda3 [eats, leaves]

We can explode the values column, transforming each list-like to a separate row, by using explode(). This will
replicate the index values from the original row:

In [133]: df['values'].explode()
Out[133]:
0 eats
0 shoots
1 shoots
1 leaves
2 eats
2 leaves
Name: values, dtype: object

You can also explode the column in the DataFrame.

In [134]: df.explode('values')
Out[134]:
keys values
0 panda1 eats
0 panda1 shoots
1 panda2 shoots
1 panda2 leaves
2 panda3 eats
2 panda3 leaves

Series.explode() will replace empty lists with np.nan and preserve scalar entries. The dtype of the resulting
Series is always object.
In [135]: s = pd.Series([[1, 2, 3], 'foo', [], ['a', 'b']])

In [136]: s
Out[136]:
0 [1, 2, 3]

460 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

1 foo
2 []
3 [a, b]
dtype: object

In [137]: s.explode()
Out[137]:
0 1
0 2
0 3
1 foo
2 NaN
3 a
3 b
dtype: object
Here is a typical usecase. You have comma separated strings in a column and want to expand this.

In [138]: df = pd.DataFrame([{'var1': 'a,b,c', 'var2': 1},

.....: {'var1': 'd,e,f', 'var2': 2}])
.....:

In [139]: df
Out[139]:
var1 var2
0 a,b,c 1
1 d,e,f 2

Creating a long form DataFrame is now straightforward using explode and chained operations

In [140]: df.assign(var1=df.var1.str.split(',')).explode('var1')
Out[140]:
var1 var2
0 a 1
0 b 1
0 c 1
1 d 2
1 e 2
1 f 2

{{ header }}

4.6 Working with text data

Series and Index are equipped with a set of string processing methods that make it easy to operate on each element of
the array. Perhaps most importantly, these methods exclude missing/NA values automatically. These are accessed via
the str attribute and generally have names matching the equivalent (scalar) built-in string methods:
In [1]: s = pd.Series(['A', 'B', 'C', 'Aaba', 'Baca', np.nan, 'CABA', 'dog',
,→'cat'])

In [2]: s.str.lower()
Out[2]:
0 a
1 b

4.6. Working with text data 461

pandas: powerful Python data analysis toolkit, Release 0.25.3

2 c
3 aaba
4 baca
5 NaN
6 caba
7 dog
8 cat
dtype: object

In [3]: s.str.upper()
Out[3]:
0 A
1 B
2 C
3 AABA
4 BACA
5 NaN
6 CABA
7 DOG
8 CAT
dtype: object

In [4]: s.str.len()
Out[4]:
0 1.0
1 1.0
2 1.0
3 4.0
4 4.0
5 NaN
6 4.0
7 3.0
8 3.0
dtype: float64
In [5]: idx = pd.Index([' jack', 'jill ', ' jesse ', 'frank'])

In [6]: idx.str.strip()
Out[6]: Index(['jack', 'jill', 'jesse', 'frank'], dtype='object')

In [7]: idx.str.lstrip()
Out[7]: Index(['jack', 'jill ', 'jesse ', 'frank'], dtype='object')

In [8]: idx.str.rstrip()
Out[8]: Index([' jack', 'jill', ' jesse', 'frank'], dtype='object')
The string methods on Index are especially useful for cleaning up or transforming DataFrame columns. For instance,
you may have columns with leading or trailing whitespace:
In [9]: df = pd.DataFrame(np.random.randn(3, 2),
...: columns=[' Column A ', ' Column B '], index=range(3))
...:

In [10]: df
Out[10]:
(continues on next page)

462 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

Column A Column B
0 -0.061542 0.528296
1 -0.952849 2.392803
2 -1.524186 2.642204

Since df.columns is an Index object, we can use the .str accessor

In [11]: df.columns.str.strip()
Out[11]: Index(['Column A', 'Column B'], dtype='object')

In [12]: df.columns.str.lower()
Out[12]: Index([' column a ', ' column b '], dtype='object')
These string methods can then be used to clean up the columns as needed. Here we are removing leading and trailing
whitespaces, lower casing all names, and replacing any remaining whitespaces with underscores:
In [13]: df.columns = df.columns.str.strip().str.lower().str.replace(' ', '_')

In [14]: df
Out[14]:
column_a column_b
0 -0.061542 0.528296
1 -0.952849 2.392803
2 -1.524186 2.642204

Note: If you have a Series where lots of elements are repeated (i.e. the number of unique elements in the
Series is a lot smaller than the length of the Series), it can be faster to convert the original Series to one of
type category and then use .str.<method> or .dt.<property> on that. The performance difference comes
from the fact that, for Series of type category, the string operations are done on the .categories and not on
each element of the Series.
Please note that a Series of type category with string .categories has some limitations in comparison to
Series of type string (e.g. you cant add strings to each other: s + " " + s wont work if s is a Series of type
category). Also, .str methods which operate on elements of type list are not available on such a Series.

Warning: Before v.0.25.0, the .str-accessor did only the most rudimentary type checks. Starting with v.0.25.0,
the type of the Series is inferred and the allowed types (i.e. strings) are enforced more rigorously.
Generally speaking, the .str accessor is intended to work only on strings. With very few exceptions, other uses
are not supported, and may be disabled at a later point.

4.6.1 Splitting and replacing strings

Methods like split return a Series of lists:

In [15]: s2 = pd.Series(['a_b_c', 'c_d_e', np.nan, 'f_g_h'])

In [16]: s2.str.split('_')
Out[16]:
0 [a, b, c]
1 [c, d, e]
2 NaN
(continues on next page)

4.6. Working with text data 463

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

3 [f, g, h]
dtype: object

Elements in the split lists can be accessed using get or [] notation:

In [17]: s2.str.split('_').str.get(1)
Out[17]:
0 b
1 d
2 NaN
3 g
dtype: object

In [18]: s2.str.split('_').str[1]
Out[18]:
0 b
1 d
2 NaN
3 g
dtype: object
It is easy to expand this to return a DataFrame using expand.

In [19]: s2.str.split('_', expand=True)

Out[19]:
0 1 2
0 a b c
1 c d e
2 NaN NaN NaN
3 f g h

It is also possible to limit the number of splits:

In [20]: s2.str.split('_', expand=True, n=1)

Out[20]:
0 1
0 a b_c
1 c d_e
2 NaN NaN
3 f g_h

rsplit is similar to split except it works in the reverse direction, i.e., from the end of the string to the beginning
of the string:

In [21]: s2.str.rsplit('_', expand=True, n=1)

Out[21]:
0 1
0 a_b c
1 c_d e
2 NaN NaN
3 f_g h

replace by default replaces regular expressions:

In [22]: s3 = pd.Series(['A', 'B', 'C', 'Aaba', 'Baca',
....: '', np.nan, 'CABA', 'dog', 'cat'])
....:

464 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [23]: s3
Out[23]:
0 A
1 B
2 C
3 Aaba
4 Baca
5
6 NaN
7 CABA
8 dog
9 cat
dtype: object

In [24]: s3.str.replace('^.a|dog', 'XX-XX ', case=False)

Out[24]:
0 A
1 B
2 C
3 XX-XX ba
4 XX-XX ca
5
6 NaN
7 XX-XX BA
8 XX-XX
9 XX-XX t
dtype: object
Some caution must be taken to keep regular expressions in mind! For example, the following code will cause trouble
because of the regular expression meaning of $:
# Consider the following badly formatted financial data
In [25]: dollars = pd.Series(['12', '-$10', '$10,000'])

# This does what you'd naively expect:

In [26]: dollars.str.replace('$', '')
Out[26]:
0 12
1 -10
2 10,000
dtype: object

# But this doesn't:

In [27]: dollars.str.replace('-$', '-')
Out[27]:
0 12
1 -$10
2 $10,000
dtype: object

# We need to escape the special character (for >1 len patterns)

In [28]: dollars.str.replace(r'-\$', '-')
Out[28]:
0 12

4.6. Working with text data 465

pandas: powerful Python data analysis toolkit, Release 0.25.3

1 -10
2 $10,000
dtype: object
New in version 0.23.0.
If you do want literal replacement of a string (equivalent to str.replace()), you can set the optional regex
parameter to False, rather than escaping each character. In this case both pat and repl must be strings:
# These lines are equivalent
In [29]: dollars.str.replace(r'-\$', '-')
Out[29]:
0 12
1 -10
2 $10,000
dtype: object

In [30]: dollars.str.replace('-$', '-', regex=False)

Out[30]:
0 12
1 -10
2 $10,000
dtype: object
New in version 0.20.0.
The replace method can also take a callable as replacement. It is called on every pat using re.sub(). The
callable should expect one positional argument (a regex object) and return a string.

# Reverse every lowercase alphabetic word

In [31]: pat = r'[a-z]+'

In [32]: def repl(m):

....: return m.group(0)[::-1]
....:

In [33]: pd.Series(['foo 123', 'bar baz', np.nan]).str.replace(pat, repl)

Out[33]:
0 oof 123
1 rab zab
2 NaN
dtype: object

# Using regex groups

In [34]: pat = r"(?P<one>\w+) (?P<two>\w+) (?P<three>\w+)"

In [35]: def repl(m):

....: return m.group('two').swapcase()
....:

In [36]: pd.Series(['Foo Bar Baz', np.nan]).str.replace(pat, repl)

Out[36]:
0 bAR
1 NaN
dtype: object

New in version 0.20.0.

The replace method also accepts a compiled regular expression object from re.compile() as a pattern. All

466 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

flags should be included in the compiled regular expression object.

In [37]: import re

In [38]: regex_pat = re.compile(r'^.a|dog', flags=re.IGNORECASE)

In [39]: s3.str.replace(regex_pat, 'XX-XX ')

Out[39]:
0 A
1 B
2 C
3 XX-XX ba
4 XX-XX ca
5
6 NaN
7 XX-XX BA
8 XX-XX
9 XX-XX t
dtype: object

Including a flags argument when calling replace with a compiled regular expression object will raise a
ValueError.

In [40]: s3.str.replace(regex_pat, 'XX-XX ', flags=re.IGNORECASE)

---------------------------------------------------------------------------
ValueError: case and flags cannot be set when pat is a compiled regex

4.6.2 Concatenation

There are several ways to concatenate a Series or Index, either with itself or others, all based on cat(), resp.
Index.str.cat.

Concatenating a single Series into a string

The content of a Series (or Index) can be concatenated:

In [41]: s = pd.Series(['a', 'b', 'c', 'd'])

In [42]: s.str.cat(sep=',')
Out[42]: 'a,b,c,d'

If not specified, the keyword sep for the separator defaults to the empty string, sep='':

In [43]: s.str.cat()
Out[43]: 'abcd'

By default, missing values are ignored. Using na_rep, they can be given a representation:
In [44]: t = pd.Series(['a', 'b', np.nan, 'd'])

In [45]: t.str.cat(sep=',')
Out[45]: 'a,b,d'

In [46]: t.str.cat(sep=',', na_rep='-')

Out[46]: 'a,b,-,d'

4.6. Working with text data 467

pandas: powerful Python data analysis toolkit, Release 0.25.3

Concatenating a Series and something list-like into a Series

The first argument to cat() can be a list-like object, provided that it matches the length of the calling Series (or
Index).

In [47]: s.str.cat(['A', 'B', 'C', 'D'])

Out[47]:
0 aA
1 bB
2 cC
3 dD
dtype: object

Missing values on either side will result in missing values in the result as well, unless na_rep is specified:
In [48]: s.str.cat(t)
Out[48]:
0 aa
1 bb
2 NaN
3 dd
dtype: object

In [49]: s.str.cat(t, na_rep='-')

Out[49]:
0 aa
1 bb
2 c-
3 dd
dtype: object

Concatenating a Series and something array-like into a Series

New in version 0.23.0.

The parameter others can also be two-dimensional. In this case, the number or rows must match the lengths of the
calling Series (or Index).
In [50]: d = pd.concat([t, s], axis=1)

In [51]: s
Out[51]:
0 a
1 b
2 c
3 d
dtype: object

In [52]: d
Out[52]:
0 1
0 a a
1 b b
2 NaN c
3 d d

468 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [53]: s.str.cat(d, na_rep='-')

Out[53]:
0 aaa
1 bbb
2 c-c
3 ddd
dtype: object

Concatenating a Series and an indexed object into a Series, with alignment

New in version 0.23.0.

For concatenation with a Series or DataFrame, it is possible to align the indexes before concatenation by setting
the join-keyword.
In [54]: u = pd.Series(['b', 'd', 'a', 'c'], index=[1, 3, 0, 2])

In [55]: s
Out[55]:
0 a
1 b
2 c
3 d
dtype: object

In [56]: u
Out[56]:
1 b
3 d
0 a
2 c
dtype: object

In [57]: s.str.cat(u)
Out[57]:
0 ab
1 bd
2 ca
3 dc
dtype: object

In [58]: s.str.cat(u, join='left')

Out[58]:
0 aa
1 bb
2 cc
3 dd
dtype: object

Warning: If the join keyword is not passed, the method cat() will currently fall back to the behavior before
version 0.23.0 (i.e. no alignment), but a FutureWarning will be raised if any of the involved indexes differ,
since this default will change to join='left' in a future version.

4.6. Working with text data 469

pandas: powerful Python data analysis toolkit, Release 0.25.3

The usual options are available for join (one of 'left', 'outer', 'inner', 'right'). In particular,
alignment also means that the different lengths do not need to coincide anymore.
In [59]: v = pd.Series(['z', 'a', 'b', 'd', 'e'], index=[-1, 0, 1, 3, 4])

In [60]: s
Out[60]:
0 a
1 b
2 c
3 d
dtype: object

In [61]: v
Out[61]:
-1 z
0 a
1 b
3 d
4 e
dtype: object

In [62]: s.str.cat(v, join='left', na_rep='-')

Out[62]:
0 aa
1 bb
2 c-
3 dd
dtype: object

In [63]: s.str.cat(v, join='outer', na_rep='-')

Out[63]:
-1 -z
0 aa
1 bb
2 c-
3 dd
4 -e
dtype: object
The same alignment can be used when others is a DataFrame:
In [64]: f = d.loc[[3, 2, 1, 0], :]

In [65]: s
Out[65]:
0 a
1 b
2 c
3 d
dtype: object

In [66]: f
Out[66]:
0 1

470 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

3 d d
2 NaN c
1 b b
0 a a

In [67]: s.str.cat(f, join='left', na_rep='-')

Out[67]:
0 aaa
1 bbb
2 c-c
3 ddd
dtype: object

Concatenating a Series and many objects into a Series

Several array-like items (specifically: Series, Index, and 1-dimensional variants of np.ndarray) can be com-
bined in a list-like container (including iterators, dict-views, etc.).
In [68]: s
Out[68]:
0 a
1 b
2 c
3 d
dtype: object

In [69]: u
Out[69]:
1 b
3 d
0 a
2 c
dtype: object

In [70]: s.str.cat([u, u.to_numpy()], join='left')

Out[70]:
0 aab
1 bbd
2 cca
3 ddc
dtype: object
All elements without an index (e.g. np.ndarray) within the passed list-like must match in length to the calling
Series (or Index), but Series and Index may have arbitrary length (as long as alignment is not disabled with
join=None):
In [71]: v
Out[71]:
-1 z
0 a
1 b
3 d
4 e
dtype: object

4.6. Working with text data 471

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [72]: s.str.cat([v, u, u.to_numpy()], join='outer', na_rep='-')

Out[72]:
-1 -z--
0 aaab
1 bbbd
2 c-ca
3 dddc
4 -e--
dtype: object
If using join='right' on a list-like of others that contains different indexes, the union of these indexes will be
used as the basis for the final concatenation:
In [73]: u.loc[[3]]
Out[73]:
3 d
dtype: object

In [74]: v.loc[[-1, 0]]

Out[74]:
-1 z
0 a
dtype: object

In [75]: s.str.cat([u.loc[[3]], v.loc[[-1, 0]]], join='right', na_rep='-')

Out[75]:
-1 --z
0 a-a
3 dd-
dtype: object

4.6.3 Indexing with .str

You can use [] notation to directly index by position locations. If you index past the end of the string, the result will
be a NaN.
In [76]: s = pd.Series(['A', 'B', 'C', 'Aaba', 'Baca', np.nan,
....: 'CABA', 'dog', 'cat'])
....:

In [77]: s.str[0]
Out[77]:
0 A
1 B
2 C
3 A
4 B
5 NaN
6 C
7 d
8 c
dtype: object

In [78]: s.str[1]
Out[78]:

472 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

0 NaN
1 NaN
2 NaN
3 a
4 a
5 NaN
6 A
7 o
8 a
dtype: object

4.6.4 Extracting substrings

Extract first match in each subject (extract)

Warning: In version 0.18.0, extract gained the expand argument. When expand=False it returns a
Series, Index, or DataFrame, depending on the subject and regular expression pattern (same behavior as
pre-0.18.0). When expand=True it always returns a DataFrame, which is more consistent and less confusing
from the perspective of a user. expand=True is the default since version 0.23.0.

The extract method accepts a regular expression with at least one capture group.
Extracting a regular expression with more than one group returns a DataFrame with one column per group.

In [79]: pd.Series(['a1', 'b2', 'c3']).str.extract(r'([ab])(\d)', expand=False)

Out[79]:
0 1
0 a 1
1 b 2
2 NaN NaN

Elements that do not match return a row filled with NaN. Thus, a Series of messy strings can be converted into a
like-indexed Series or DataFrame of cleaned-up or more useful strings, without necessitating get() to access tuples
or re.match objects. The dtype of the result is always object, even if no match is found and the result only contains
NaN.
Named groups like

In [80]: pd.Series(['a1', 'b2', 'c3']).str.extract(r'(?P<letter>[ab])(?P<digit>\d)',

....: expand=False)
....:
Out[80]:
letter digit
0 a 1
1 b 2
2 NaN NaN

and optional groups like

In [81]: pd.Series(['a1', 'b2', '3']).str.extract(r'([ab])?(\d)', expand=False)

Out[81]:
0 1
0 a 1
(continues on next page)

4.6. Working with text data 473

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

1 b 2
2 NaN 3

can also be used. Note that any capture group names in the regular expression will be used for column names;
otherwise capture group numbers will be used.
Extracting a regular expression with one group returns a DataFrame with one column if expand=True.

In [82]: pd.Series(['a1', 'b2', 'c3']).str.extract(r'[ab](\d)', expand=True)

Out[82]:
0
0 1
1 2
2 NaN

It returns a Series if expand=False.

In [83]: pd.Series(['a1', 'b2', 'c3']).str.extract(r'[ab](\d)', expand=False)

Out[83]:
0 1
1 2
2 NaN
dtype: object

Calling on an Index with a regex with exactly one capture group returns a DataFrame with one column if
expand=True.
In [84]: s = pd.Series(["a1", "b2", "c3"], ["A11", "B22", "C33"])

In [85]: s
Out[85]:
A11 a1
B22 b2
C33 c3
dtype: object

In [86]: s.index.str.extract("(?P<letter>[a-zA-Z])", expand=True)

Out[86]:
letter
0 A
1 B
2 C
It returns an Index if expand=False.

In [87]: s.index.str.extract("(?P<letter>[a-zA-Z])", expand=False)

Out[87]: Index(['A', 'B', 'C'], dtype='object', name='letter')

Calling on an Index with a regex with more than one capture group returns a DataFrame if expand=True.

In [88]: s.index.str.extract("(?P<letter>[a-zA-Z])([0-9]+)", expand=True)

Out[88]:
letter 1
0 A 11
1 B 22
2 C 33

474 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

It raises ValueError if expand=False.

>>> s.index.str.extract("(?P<letter>[a-zA-Z])([0-9]+)", expand=False)
ValueError: only one regex group is supported with Index

The table below summarizes the behavior of extract(expand=False) (input subject in first column, number of
groups in regex in first row)

1 group >1 group

Index Index ValueError
Series Series DataFrame

Extract all matches in each subject (extractall)

New in version 0.18.0.

Unlike extract (which returns only the first match),
In [89]: s = pd.Series(["a1a2", "b1", "c1"], index=["A", "B", "C"])

In [90]: s
Out[90]:
A a1a2
B b1
C c1
dtype: object

In [91]: two_groups = '(?P<letter>[a-z])(?P<digit>[0-9])'

In [92]: s.str.extract(two_groups, expand=True)

Out[92]:
letter digit
A a 1
B b 1
C c 1

the extractall method returns every match. The result of extractall is always a DataFrame with a
MultiIndex on its rows. The last level of the MultiIndex is named match and indicates the order in the
subject.
In [93]: s.str.extractall(two_groups)
Out[93]:
letter digit
match
A 0 a 1
1 a 2
B 0 b 1
C 0 c 1

When each subject string in the Series has exactly one match,
In [94]: s = pd.Series(['a3', 'b3', 'c2'])

In [95]: s
Out[95]:
0 a3
(continues on next page)

4.6. Working with text data 475

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

1 b3
2 c2
dtype: object

then extractall(pat).xs(0, level='match') gives the same result as extract(pat).

In [96]: extract_result = s.str.extract(two_groups, expand=True)

In [97]: extract_result
Out[97]:
letter digit
0 a 3
1 b 3
2 c 2

In [98]: extractall_result = s.str.extractall(two_groups)

In [99]: extractall_result
Out[99]:
letter digit
match
0 0 a 3
1 0 b 3
2 0 c 2

In [100]: extractall_result.xs(0, level="match")

Out[100]:
letter digit
0 a 3
1 b 3
2 c 2
Index also supports .str.extractall. It returns a DataFrame which has the same result as a Series.str.
extractall with a default index (starts from 0).
New in version 0.19.0.
In [101]: pd.Index(["a1a2", "b1", "c1"]).str.extractall(two_groups)
Out[101]:
letter digit
match
0 0 a 1
1 a 2
1 0 b 1
2 0 c 1

In [102]: pd.Series(["a1a2", "b1", "c1"]).str.extractall(two_groups)

Out[102]:
letter digit
match
0 0 a 1
1 a 2
1 0 b 1
2 0 c 1

476 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

4.6.5 Testing for Strings that match or contain a pattern

You can check whether elements contain a pattern:

In [103]: pattern = r'[0-9][a-z]'

In [104]: pd.Series(['1', '2', '3a', '3b', '03c']).str.contains(pattern)

Out[104]:
0 False
1 False
2 True
3 True
4 True
dtype: bool

Or whether elements match a pattern:

In [105]: pd.Series(['1', '2', '3a', '3b', '03c']).str.match(pattern)

Out[105]:
0 False
1 False
2 True
3 True
4 False
dtype: bool

The distinction between match and contains is strictness: match relies on strict re.match, while contains
relies on re.search.
Methods like match, contains, startswith, and endswith take an extra na argument so missing values can
be considered True or False:

In [106]: s4 = pd.Series(['A', 'B', 'C', 'Aaba', 'Baca', np.nan, 'CABA', 'dog', 'cat
,→'])

In [107]: s4.str.contains('A', na=False)

Out[107]:
0 True
1 False
2 False
3 True
4 False
5 False
6 True
7 False
8 False
dtype: bool

4.6.6 Creating indicator variables

You can extract dummy variables from string columns. For example if they are separated by a '|':

In [108]: s = pd.Series(['a', 'a|b', np.nan, 'a|c'])

In [109]: s.str.get_dummies(sep='|')
Out[109]:
(continues on next page)

4.6. Working with text data 477

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

a b c
0 1 0 0
1 1 1 0
2 0 0 0
3 1 0 1

String Index also supports get_dummies which returns a MultiIndex.

New in version 0.18.1.

In [110]: idx = pd.Index(['a', 'a|b', np.nan, 'a|c'])

In [111]: idx.str.get_dummies(sep='|')
Out[111]:
MultiIndex([(1, 0, 0),
(1, 1, 0),
(0, 0, 0),
(1, 0, 1)],
names=['a', 'b', 'c'])

See also get_dummies().

4.6.7 Method summary

Method Description
cat() Concatenate strings
split() Split strings on delimiter
rsplit() Split strings on delimiter working from the end of the string
get() Index into each element (retrieve i-th element)
join() Join strings in each element of the Series with passed separator
get_dummies() Split strings on the delimiter returning DataFrame of dummy variables
contains() Return boolean array if each string contains pattern/regex
replace() Replace occurrences of pattern/regex/string with some other string or the return value of a
callable given the occurrence
repeat() Duplicate values (s.str.repeat(3) equivalent to x * 3)
pad() Add whitespace to left, right, or both sides of strings
center() Equivalent to str.center
ljust() Equivalent to str.ljust
rjust() Equivalent to str.rjust
zfill() Equivalent to str.zfill
wrap() Split long strings into lines with length less than a given width
slice() Slice each string in the Series
slice_replace() Replace slice in each string with passed value
count() Count occurrences of pattern
startswith() Equivalent to str.startswith(pat) for each element
endswith() Equivalent to str.endswith(pat) for each element
findall() Compute list of all occurrences of pattern/regex for each string
match() Call re.match on each element, returning matched groups as list
extract() Call re.search on each element, returning DataFrame with one row for each element
and one column for each regex capture group
Continued on next page

478 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Table 2 – continued from previous page

Method Description
extractall() Call re.findall on each element, returning DataFrame with one row for each match
and one column for each regex capture group
len() Compute string lengths
strip() Equivalent to str.strip
rstrip() Equivalent to str.rstrip
lstrip() Equivalent to str.lstrip
partition() Equivalent to str.partition
rpartition() Equivalent to str.rpartition
lower() Equivalent to str.lower
casefold() Equivalent to str.casefold
upper() Equivalent to str.upper
find() Equivalent to str.find
rfind() Equivalent to str.rfind
index() Equivalent to str.index
rindex() Equivalent to str.rindex
capitalize() Equivalent to str.capitalize
swapcase() Equivalent to str.swapcase
normalize() Return Unicode normal form. Equivalent to unicodedata.normalize
translate() Equivalent to str.translate
isalnum() Equivalent to str.isalnum
isalpha() Equivalent to str.isalpha
isdigit() Equivalent to str.isdigit
isspace() Equivalent to str.isspace
islower() Equivalent to str.islower
isupper() Equivalent to str.isupper
istitle() Equivalent to str.istitle
isnumeric() Equivalent to str.isnumeric
isdecimal() Equivalent to str.isdecimal

{{ header }}

4.7 Working with missing data

In this section, we will discuss missing (also referred to as NA) values in pandas.

Note: The choice of using NaN internally to denote missing data was largely for simplicity and performance reasons.
It differs from the MaskedArray approach of, for example, scikits.timeseries. We are hopeful that NumPy
will soon be able to provide a native NA type solution (similar to R) performant enough to be used in pandas.

See the cookbook for some advanced strategies.

4.7.1 Values considered missing

As data comes in many shapes and forms, pandas aims to be flexible with regard to handling missing data. While
NaN is the default missing value marker for reasons of computational speed and convenience, we need to be able to
easily detect this value with data of different types: floating point, integer, boolean, and general object. In many cases,
however, the Python None will arise and we wish to also consider that missing or not available or NA.

4.7. Working with missing data 479

pandas: powerful Python data analysis toolkit, Release 0.25.3

Note: If you want to consider inf and -inf to be NA in computations, you can set pandas.options.mode.
use_inf_as_na = True.

In [1]: df = pd.DataFrame(np.random.randn(5, 3), index=['a', 'c', 'e', 'f', 'h'],

...: columns=['one', 'two', 'three'])
...:

In [2]: df['four'] = 'bar'

In [3]: df['five'] = df['one'] > 0

In [4]: df
Out[4]:
one two three four five
a 0.070821 -0.093641 0.014099 bar True
c 0.702961 0.870484 -1.521966 bar True
e 0.219802 1.546457 -0.174262 bar True
f 0.831417 1.332054 0.438532 bar True
h 0.786117 -0.304618 0.956171 bar True

In [5]: df2 = df.reindex(['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h'])

In [6]: df2
Out[6]:
one two three four five
a 0.070821 -0.093641 0.014099 bar True
b NaN NaN NaN NaN NaN
c 0.702961 0.870484 -1.521966 bar True
d NaN NaN NaN NaN NaN
e 0.219802 1.546457 -0.174262 bar True
f 0.831417 1.332054 0.438532 bar True
g NaN NaN NaN NaN NaN
h 0.786117 -0.304618 0.956171 bar True

To make detecting missing values easier (and across different array dtypes), pandas provides the isna() and
notna() functions, which are also methods on Series and DataFrame objects:
In [7]: df2['one']
Out[7]:
a 0.070821
b NaN
c 0.702961
d NaN
e 0.219802
f 0.831417
g NaN
h 0.786117
Name: one, dtype: float64

In [8]: pd.isna(df2['one'])
Out[8]:
a False
b True
c False
d True

480 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

e False
f False
g True
h False
Name: one, dtype: bool

In [9]: df2['four'].notna()
Out[9]:
a True
b False
c True
d False
e True
f True
g False
h True
Name: four, dtype: bool

In [10]: df2.isna()
Out[10]:
one two three four five
a False False False False False
b True True True True True
c False False False False False
d True True True True True
e False False False False False
f False False False False False
g True True True True True
h False False False False False

Warning: One has to be mindful that in Python (and NumPy), the nan's dont compare equal, but None's do.
Note that pandas/NumPy uses the fact that np.nan != np.nan, and treats None like np.nan.
In [11]: None == None #
,→noqa: E711

Out[11]: True

In [12]: np.nan == np.nan

Out[12]: False
So as compared to above, a scalar equality comparison versus a None/np.nan doesnt provide useful information.
In [13]: df2['one'] == np.nan
Out[13]:
a False
b False
c False
d False
e False
f False
g False
h False
Name: one, dtype: bool

4.7. Working with missing data 481

pandas: powerful Python data analysis toolkit, Release 0.25.3

Integer dtypes and missing data

Because NaN is a float, a column of integers with even one missing values is cast to floating-point dtype (see Support
for integer NA for more). Pandas provides a nullable integer array, which can be used by explicitly requesting the
dtype:

In [14]: pd.Series([1, 2, np.nan, 4], dtype=pd.Int64Dtype())

Out[14]:
0 1
1 2
2 NaN
3 4
dtype: Int64

Alternatively, the string alias dtype='Int64' (note the capital "I") can be used.
See Nullable integer data type for more.

Datetimes

For datetime64[ns] types, NaT represents missing values. This is a pseudo-native sentinel value that can be represented
by NumPy in a singular dtype (datetime64[ns]). pandas objects provide compatibility between NaT and NaN.
In [15]: df2 = df.copy()

In [16]: df2['timestamp'] = pd.Timestamp('20120101')

In [17]: df2
Out[17]:
one two three four five timestamp
a 0.070821 -0.093641 0.014099 bar True 2012-01-01
c 0.702961 0.870484 -1.521966 bar True 2012-01-01
e 0.219802 1.546457 -0.174262 bar True 2012-01-01
f 0.831417 1.332054 0.438532 bar True 2012-01-01
h 0.786117 -0.304618 0.956171 bar True 2012-01-01

In [18]: df2.loc[['a', 'c', 'h'], ['one', 'timestamp']] = np.nan

In [19]: df2
Out[19]:
one two three four five timestamp
a NaN -0.093641 0.014099 bar True NaT
c NaN 0.870484 -1.521966 bar True NaT
e 0.219802 1.546457 -0.174262 bar True 2012-01-01
f 0.831417 1.332054 0.438532 bar True 2012-01-01
h NaN -0.304618 0.956171 bar True NaT

In [20]: df2.dtypes.value_counts()
Out[20]:
float64 3
datetime64[ns] 1
object 1
bool 1
dtype: int64

482 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Inserting missing data

You can insert missing values by simply assigning to containers. The actual missing value used will be chosen based
on the dtype.
For example, numeric containers will always use NaN regardless of the missing value type chosen:

In [21]: s = pd.Series([1, 2, 3])

In [22]: s.loc[0] = None

In [23]: s
Out[23]:
0 NaN
1 2.0
2 3.0
dtype: float64

Likewise, datetime containers will always use NaT.

For object containers, pandas will use the value given:

In [24]: s = pd.Series(["a", "b", "c"])

In [25]: s.loc[0] = None

In [26]: s.loc[1] = np.nan

In [27]: s
Out[27]:
0 None
1 NaN
2 c
dtype: object

Calculations with missing data

Missing values propagate naturally through arithmetic operations between pandas objects.
In [28]: a
Out[28]:
one two
a NaN -0.093641
c NaN 0.870484
e 0.219802 1.546457
f 0.831417 1.332054
h 0.831417 -0.304618

In [29]: b
Out[29]:
one two three
a NaN -0.093641 0.014099
c NaN 0.870484 -1.521966
e 0.219802 1.546457 -0.174262
f 0.831417 1.332054 0.438532
h NaN -0.304618 0.956171

4.7. Working with missing data 483

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [30]: a + b
Out[30]:
one three two
a NaN NaN -0.187283
c NaN NaN 1.740968
e 0.439605 NaN 3.092914
f 1.662833 NaN 2.664108
h NaN NaN -0.609235
The descriptive statistics and computational methods discussed in the data structure overview (and listed here and
here) are all written to account for missing data. For example:
• When summing data, NA (missing) values will be treated as zero.
• If the data are all NA, the result will be 0.
• Cumulative methods like cumsum() and cumprod() ignore NA values by default, but preserve them in the
resulting arrays. To override this behaviour and include NA values, use skipna=False.
In [31]: df
Out[31]:
one two three
a NaN -0.093641 0.014099
c NaN 0.870484 -1.521966
e 0.219802 1.546457 -0.174262
f 0.831417 1.332054 0.438532
h NaN -0.304618 0.956171

In [32]: df['one'].sum()
Out[32]: 1.0512188858617544

In [33]: df.mean(1)
Out[33]:
a -0.039771
c -0.325741
e 0.530666
f 0.867334
h 0.325777
dtype: float64

In [34]: df.cumsum()
Out[34]:
one two three
a NaN -0.093641 0.014099
c NaN 0.776843 -1.507866
e 0.219802 2.323300 -1.682128
f 1.051219 3.655354 -1.243596
h NaN 3.350736 -0.287425

In [35]: df.cumsum(skipna=False)
Out[35]:
one two three
a NaN -0.093641 0.014099
c NaN 0.776843 -1.507866
e NaN 2.323300 -1.682128

484 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

f NaN 3.655354 -1.243596

h NaN 3.350736 -0.287425

4.7.2 Sum/prod of empties/nans

Warning: This behavior is now standard as of v0.22.0 and is consistent with the default in numpy; previously
sum/prod of all-NA or empty Series/DataFrames would return NaN. See v0.22.0 whatsnew for more.

The sum of an empty or all-NA Series or column of a DataFrame is 0.

In [36]: pd.Series([np.nan]).sum()
Out[36]: 0.0

In [37]: pd.Series([]).sum()
Out[37]: 0.0
The product of an empty or all-NA Series or column of a DataFrame is 1.
In [38]: pd.Series([np.nan]).prod()
Out[38]: 1.0

In [39]: pd.Series([]).prod()
Out[39]: 1.0

4.7.3 NA values in GroupBy

NA groups in GroupBy are automatically excluded. This behavior is consistent with R, for example:
In [40]: df
Out[40]:
one two three
a NaN -0.093641 0.014099
c NaN 0.870484 -1.521966
e 0.219802 1.546457 -0.174262
f 0.831417 1.332054 0.438532
h NaN -0.304618 0.956171

In [41]: df.groupby('one').mean()
Out[41]:
two three
one
0.219802 1.546457 -0.174262
0.831417 1.332054 0.438532
See the groupby section here for more information.

Cleaning / filling missing data

pandas objects are equipped with various data manipulation methods for dealing with missing data.

4.7. Working with missing data 485

pandas: powerful Python data analysis toolkit, Release 0.25.3

4.7.4 Filling missing values: fillna

fillna() can fill in NA values with non-NA data in a couple of ways, which we illustrate:
Replace NA with a scalar value
In [42]: df2
Out[42]:
one two three four five timestamp
a NaN -0.093641 0.014099 bar True NaT
c NaN 0.870484 -1.521966 bar True NaT
e 0.219802 1.546457 -0.174262 bar True 2012-01-01
f 0.831417 1.332054 0.438532 bar True 2012-01-01
h NaN -0.304618 0.956171 bar True NaT

In [43]: df2.fillna(0)
Out[43]:
one two three four five timestamp
a 0.000000 -0.093641 0.014099 bar True 0
c 0.000000 0.870484 -1.521966 bar True 0
e 0.219802 1.546457 -0.174262 bar True 2012-01-01 00:00:00
f 0.831417 1.332054 0.438532 bar True 2012-01-01 00:00:00
h 0.000000 -0.304618 0.956171 bar True 0

In [44]: df2['one'].fillna('missing')
Out[44]:
a missing
c missing
e 0.219802
f 0.831417
h missing
Name: one, dtype: object
Fill gaps forward or backward
Using the same filling arguments as reindexing, we can propagate non-NA values forward or backward:
In [45]: df
Out[45]:
one two three
a NaN -0.093641 0.014099
c NaN 0.870484 -1.521966
e 0.219802 1.546457 -0.174262
f 0.831417 1.332054 0.438532
h NaN -0.304618 0.956171

In [46]: df.fillna(method='pad')
Out[46]:
one two three
a NaN -0.093641 0.014099
c NaN 0.870484 -1.521966
e 0.219802 1.546457 -0.174262
f 0.831417 1.332054 0.438532
h 0.831417 -0.304618 0.956171
Limit the amount of filling

486 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

If we only want consecutive gaps filled up to a certain number of data points, we can use the limit keyword:
In [47]: df
Out[47]:
one two three
a NaN -0.093641 0.014099
c NaN 0.870484 -1.521966
e NaN NaN NaN
f NaN NaN NaN
h NaN -0.304618 0.956171

In [48]: df.fillna(method='pad', limit=1)

Out[48]:
one two three
a NaN -0.093641 0.014099
c NaN 0.870484 -1.521966
e NaN 0.870484 -1.521966
f NaN NaN NaN
h NaN -0.304618 0.956171
To remind you, these are the available filling methods:

Method Action
pad / ffill Fill values forward
bfill / backfill Fill values backward

With time series data, using pad/ffill is extremely common so that the last known value is available at every time point.
ffill() is equivalent to fillna(method='ffill') and bfill() is equivalent to
fillna(method='bfill')

4.7.5 Filling with a PandasObject

You can also fillna using a dict or Series that is alignable. The labels of the dict or index of the Series must match the
columns of the frame you wish to fill. The use case of this is to fill a DataFrame with the mean of that column.
In [49]: dff = pd.DataFrame(np.random.randn(10, 3), columns=list('ABC'))

In [50]: dff.iloc[3:5, 0] = np.nan

In [51]: dff.iloc[4:6, 1] = np.nan

In [52]: dff.iloc[5:8, 2] = np.nan

In [53]: dff
Out[53]:
A B C
0 0.632107 -0.123200 0.579811
1 -0.617833 0.730289 -1.930809
2 -1.795903 -2.012672 -0.884042
3 NaN 0.809658 0.727889
4 NaN NaN 1.683552
5 -1.134942 NaN NaN
6 -1.654372 -0.175245 NaN
7 0.332654 -1.208013 NaN

4.7. Working with missing data 487

pandas: powerful Python data analysis toolkit, Release 0.25.3

8 0.028692 0.139178 0.877023

9 0.780292 0.682156 0.993475

In [54]: dff.fillna(dff.mean())
Out[54]:
A B C
0 0.632107 -0.123200 0.579811
1 -0.617833 0.730289 -1.930809
2 -1.795903 -2.012672 -0.884042
3 -0.428663 0.809658 0.727889
4 -0.428663 -0.144731 1.683552
5 -1.134942 -0.144731 0.292414
6 -1.654372 -0.175245 0.292414
7 0.332654 -1.208013 0.292414
8 0.028692 0.139178 0.877023
9 0.780292 0.682156 0.993475

In [55]: dff.fillna(dff.mean()['B':'C'])
Out[55]:
A B C
0 0.632107 -0.123200 0.579811
1 -0.617833 0.730289 -1.930809
2 -1.795903 -2.012672 -0.884042
3 NaN 0.809658 0.727889
4 NaN -0.144731 1.683552
5 -1.134942 -0.144731 0.292414
6 -1.654372 -0.175245 0.292414
7 0.332654 -1.208013 0.292414
8 0.028692 0.139178 0.877023
9 0.780292 0.682156 0.993475
Same result as above, but is aligning the fill value which is a Series in this case.

In [56]: dff.where(pd.notna(dff), dff.mean(), axis='columns')

Out[56]:
A B C
0 0.632107 -0.123200 0.579811
1 -0.617833 0.730289 -1.930809
2 -1.795903 -2.012672 -0.884042
3 -0.428663 0.809658 0.727889
4 -0.428663 -0.144731 1.683552
5 -1.134942 -0.144731 0.292414
6 -1.654372 -0.175245 0.292414
7 0.332654 -1.208013 0.292414
8 0.028692 0.139178 0.877023
9 0.780292 0.682156 0.993475

4.7.6 Dropping axis labels with missing data: dropna

You may wish to simply exclude labels from a data set which refer to missing data. To do this, use dropna():
In [57]: df
Out[57]:
one two three
a NaN -0.093641 0.014099

488 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

c NaN 0.870484 -1.521966

e NaN 0.000000 0.000000
f NaN 0.000000 0.000000
h NaN -0.304618 0.956171

In [58]: df.dropna(axis=0)
Out[58]:
Empty DataFrame
Columns: [one, two, three]
Index: []

In [59]: df.dropna(axis=1)
Out[59]:
two three
a -0.093641 0.014099
c 0.870484 -1.521966
e 0.000000 0.000000
f 0.000000 0.000000
h -0.304618 0.956171

In [60]: df['one'].dropna()
Out[60]: Series([], Name: one, dtype: float64)
An equivalent dropna() is available for Series. DataFrame.dropna has considerably more options than Se-
ries.dropna, which can be examined in the API.

4.7.7 Interpolation

New in version 0.23.0: The limit_area keyword argument was added.

Both Series and DataFrame objects have interpolate() that, by default, performs linear interpolation at missing
data points.
In [61]: ts
Out[61]:
2000-01-31 0.469112
2000-02-29 NaN
2000-03-31 NaN
2000-04-28 NaN
2000-05-31 NaN
...
2007-12-31 -6.950267
2008-01-31 -7.904475
2008-02-29 -6.441779
2008-03-31 -8.184940
2008-04-30 -9.011531
Freq: BM, Length: 100, dtype: float64

In [62]: ts.count()
Out[62]: 66

In [63]: ts.plot()
Out[63]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3f238410>

4.7. Working with missing data 489

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [64]: ts.interpolate()
Out[64]:
2000-01-31 0.469112
2000-02-29 0.434469
2000-03-31 0.399826
2000-04-28 0.365184
2000-05-31 0.330541
...
2007-12-31 -6.950267
2008-01-31 -7.904475
2008-02-29 -6.441779
2008-03-31 -8.184940
2008-04-30 -9.011531
Freq: BM, Length: 100, dtype: float64

In [65]: ts.interpolate().count()
Out[65]: 100

In [66]: ts.interpolate().plot()
Out[66]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3b392d90>

490 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Index aware interpolation is available via the method keyword:

In [67]: ts2
Out[67]:
2000-01-31 0.469112
2000-02-29 NaN
2002-07-31 -5.785037
2005-01-31 NaN
2008-04-30 -9.011531
dtype: float64

In [68]: ts2.interpolate()
Out[68]:
2000-01-31 0.469112
2000-02-29 -2.657962
2002-07-31 -5.785037
2005-01-31 -7.398284
2008-04-30 -9.011531
dtype: float64

In [69]: ts2.interpolate(method='time')
Out[69]:
2000-01-31 0.469112
2000-02-29 0.270241

4.7. Working with missing data 491

pandas: powerful Python data analysis toolkit, Release 0.25.3

2002-07-31 -5.785037
2005-01-31 -7.190866
2008-04-30 -9.011531
dtype: float64
For a floating-point index, use method='values':
In [70]: ser
Out[70]:
0.0 0.0
1.0 NaN
10.0 10.0
dtype: float64

In [71]: ser.interpolate()
Out[71]:
0.0 0.0
1.0 5.0
10.0 10.0
dtype: float64

In [72]: ser.interpolate(method='values')
Out[72]:
0.0 0.0
1.0 1.0
10.0 10.0
dtype: float64
You can also interpolate with a DataFrame:
In [73]: df = pd.DataFrame({'A': [1, 2.1, np.nan, 4.7, 5.6, 6.8],
....: 'B': [.25, np.nan, np.nan, 4, 12.2, 14.4]})
....:

In [74]: df
Out[74]:
A B
0 1.0 0.25
1 2.1 NaN
2 NaN NaN
3 4.7 4.00
4 5.6 12.20
5 6.8 14.40

In [75]: df.interpolate()
Out[75]:
A B
0 1.0 0.25
1 2.1 1.50
2 3.4 2.75
3 4.7 4.00
4 5.6 12.20
5 6.8 14.40
The method argument gives access to fancier interpolation methods. If you have scipy installed, you can pass the
name of a 1-d interpolation routine to method. Youll want to consult the full scipy interpolation documentation and

492 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

reference guide for details. The appropriate interpolation method will depend on the type of data you are working
with.
• If you are dealing with a time series that is growing at an increasing rate, method='quadratic' may be
appropriate.
• If you have values approximating a cumulative distribution function, then method='pchip' should work
well.
• To fill missing values with goal of smooth plotting, consider method='akima'.

Warning: These methods require scipy.

In [76]: df.interpolate(method='barycentric')
Out[76]:
A B
0 1.00 0.250
1 2.10 -7.660
2 3.53 -4.515
3 4.70 4.000
4 5.60 12.200
5 6.80 14.400

In [77]: df.interpolate(method='pchip')
Out[77]:
A B
0 1.00000 0.250000
1 2.10000 0.672808
2 3.43454 1.928950
3 4.70000 4.000000
4 5.60000 12.200000
5 6.80000 14.400000

In [78]: df.interpolate(method='akima')
Out[78]:
A B
0 1.000000 0.250000
1 2.100000 -0.873316
2 3.406667 0.320034
3 4.700000 4.000000
4 5.600000 12.200000
5 6.800000 14.400000
When interpolating via a polynomial or spline approximation, you must also specify the degree or order of the approx-
imation:
In [79]: df.interpolate(method='spline', order=2)
Out[79]:
A B
0 1.000000 0.250000
1 2.100000 -0.428598
2 3.404545 1.206900
3 4.700000 4.000000
4 5.600000 12.200000
5 6.800000 14.400000

4.7. Working with missing data 493

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [80]: df.interpolate(method='polynomial', order=2)

Out[80]:
A B
0 1.000000 0.250000
1 2.100000 -2.703846
2 3.451351 -1.453846
3 4.700000 4.000000
4 5.600000 12.200000
5 6.800000 14.400000
Compare several methods:

In [81]: np.random.seed(2)

In [82]: ser = pd.Series(np.arange(1, 10.1, .25) ** 2 + np.random.randn(37))

In [83]: missing = np.array([4, 13, 14, 15, 16, 17, 18, 20, 29])

In [84]: ser[missing] = np.nan

In [85]: methods = ['linear', 'quadratic', 'cubic']

In [86]: df = pd.DataFrame({m: ser.interpolate(method=m) for m in methods})

In [87]: df.plot()
Out[87]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3f319490>

494 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Another use case is interpolation at new values. Suppose you have 100 observations from some distribution. And lets
suppose that youre particularly interested in whats happening around the middle. You can mix pandas reindex and
interpolate methods to interpolate at the new values.

In [88]: ser = pd.Series(np.sort(np.random.uniform(size=100)))

# interpolate at new_index
In [89]: new_index = ser.index | pd.Index([49.25, 49.5, 49.75, 50.25, 50.5, 50.75])

In [90]: interp_s = ser.reindex(new_index).interpolate(method='pchip')

In [91]: interp_s[49:51]
Out[91]:
49.00 0.471410
49.25 0.476841
49.50 0.481780
49.75 0.485998
50.00 0.489266
50.25 0.491814
50.50 0.493995
50.75 0.495763
51.00 0.497074
dtype: float64

4.7. Working with missing data 495

pandas: powerful Python data analysis toolkit, Release 0.25.3

Interpolation limits

Like other pandas fill methods, interpolate() accepts a limit keyword argument. Use this argument to limit
the number of consecutive NaN values filled since the last valid observation:
In [92]: ser = pd.Series([np.nan, np.nan, 5, np.nan, np.nan,
....: np.nan, 13, np.nan, np.nan])
....:

In [93]: ser
Out[93]:
0 NaN
1 NaN
2 5.0
3 NaN
4 NaN
5 NaN
6 13.0
7 NaN
8 NaN
dtype: float64

# fill all consecutive values in a forward direction

In [94]: ser.interpolate()
Out[94]:
0 NaN
1 NaN
2 5.0
3 7.0
4 9.0
5 11.0
6 13.0
7 13.0
8 13.0
dtype: float64

# fill one consecutive value in a forward direction

In [95]: ser.interpolate(limit=1)
Out[95]:
0 NaN
1 NaN
2 5.0
3 7.0
4 NaN
5 NaN
6 13.0
7 13.0
8 NaN
dtype: float64
By default, NaN values are filled in a forward direction. Use limit_direction parameter to fill backward or
from both directions.
# fill one consecutive value backwards
In [96]: ser.interpolate(limit=1, limit_direction='backward')

496 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[96]:
0 NaN
1 5.0
2 5.0
3 NaN
4 NaN
5 11.0
6 13.0
7 NaN
8 NaN
dtype: float64

# fill one consecutive value in both directions

In [97]: ser.interpolate(limit=1, limit_direction='both')
Out[97]:
0 NaN
1 5.0
2 5.0
3 7.0
4 NaN
5 11.0
6 13.0
7 13.0
8 NaN
dtype: float64

# fill all consecutive values in both directions

In [98]: ser.interpolate(limit_direction='both')
Out[98]:
0 5.0
1 5.0
2 5.0
3 7.0
4 9.0
5 11.0
6 13.0
7 13.0
8 13.0
dtype: float64
By default, NaN values are filled whether they are inside (surrounded by) existing valid values, or outside existing
valid values. Introduced in v0.23 the limit_area parameter restricts filling to either inside or outside values.
# fill one consecutive inside value in both directions
In [99]: ser.interpolate(limit_direction='both', limit_area='inside', limit=1)
Out[99]:
0 NaN
1 NaN
2 5.0
3 7.0
4 NaN
5 11.0
6 13.0
7 NaN
8 NaN

4.7. Working with missing data 497

pandas: powerful Python data analysis toolkit, Release 0.25.3

dtype: float64

# fill all consecutive outside values backward

In [100]: ser.interpolate(limit_direction='backward', limit_area='outside')
Out[100]:
0 5.0
1 5.0
2 5.0
3 NaN
4 NaN
5 NaN
6 13.0
7 NaN
8 NaN
dtype: float64

# fill all consecutive outside values in both directions

In [101]: ser.interpolate(limit_direction='both', limit_area='outside')
Out[101]:
0 5.0
1 5.0
2 5.0
3 NaN
4 NaN
5 NaN
6 13.0
7 13.0
8 13.0
dtype: float64

4.7.8 Replacing generic values

Often times we want to replace arbitrary values with other values.

replace() in Series and replace() in DataFrame provides an efficient yet flexible way to perform such replace-
ments.
For a Series, you can replace a single value or a list of values by another value:
In [102]: ser = pd.Series([0., 1., 2., 3., 4.])

In [103]: ser.replace(0, 5)
Out[103]:
0 5.0
1 1.0
2 2.0
3 3.0
4 4.0
dtype: float64

You can replace a list of values by a list of other values:

In [104]: ser.replace([0, 1, 2, 3, 4], [4, 3, 2, 1, 0])
Out[104]:
0 4.0
(continues on next page)

498 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

1 3.0
2 2.0
3 1.0
4 0.0
dtype: float64

You can also specify a mapping dict:

In [105]: ser.replace({0: 10, 1: 100})
Out[105]:
0 10.0
1 100.0
2 2.0
3 3.0
4 4.0
dtype: float64

For a DataFrame, you can specify individual values by column:

In [106]: df = pd.DataFrame({'a': [0, 1, 2, 3, 4], 'b': [5, 6, 7, 8, 9]})

In [107]: df.replace({'a': 0, 'b': 5}, 100)

Out[107]:
a b
0 100 100
1 1 6
2 2 7
3 3 8
4 4 9

Instead of replacing with specified values, you can treat all given values as missing and interpolate over them:
In [108]: ser.replace([1, 2, 3], method='pad')
Out[108]:
0 0.0
1 0.0
2 0.0
3 0.0
4 4.0
dtype: float64

4.7.9 String/regular expression replacement

Note: Python strings prefixed with the r character such as r'hello world' are so-called raw strings. They
have different semantics regarding backslashes than strings without this prefix. Backslashes in raw strings will be
interpreted as an escaped backslash, e.g., r'\' == '\\'. You should read about them if this is unclear.

Replace the . with NaN (str -> str):

In [109]: d = {'a': list(range(4)), 'b': list('ab..'), 'c': ['a', 'b', np.nan, 'd']}

In [110]: df = pd.DataFrame(d)

(continues on next page)

4.7. Working with missing data 499

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [111]: df.replace('.', np.nan)
Out[111]:
a b c
0 0 a a
1 1 b b
2 2 NaN NaN
3 3 NaN d

Now do it with a regular expression that removes surrounding whitespace (regex -> regex):

In [112]: df.replace(r'\s*\.\s*', np.nan, regex=True)

Out[112]:
a b c
0 0 a a
1 1 b b
2 2 NaN NaN
3 3 NaN d

Replace a few different values (list -> list):

In [113]: df.replace(['a', '.'], ['b', np.nan])

Out[113]:
a b c
0 0 b b
1 1 b b
2 2 NaN NaN
3 3 NaN d

list of regex -> list of regex:

In [114]: df.replace([r'\.', r'(a)'], ['dot', r'\1stuff'], regex=True)

Out[114]:
a b c
0 0 astuff astuff
1 1 b b
2 2 dot NaN
3 3 dot d

Only search in column 'b' (dict -> dict):

In [115]: df.replace({'b': '.'}, {'b': np.nan})

Out[115]:
a b c
0 0 a a
1 1 b b
2 2 NaN NaN
3 3 NaN d

Same as the previous example, but use a regular expression for searching instead (dict of regex -> dict):

In [116]: df.replace({'b': r'\s*\.\s*'}, {'b': np.nan}, regex=True)

Out[116]:
a b c
0 0 a a
1 1 b b
2 2 NaN NaN
3 3 NaN d

500 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

You can pass nested dictionaries of regular expressions that use regex=True:

In [117]: df.replace({'b': {'b': r''}}, regex=True)

Out[117]:
a b c
0 0 a a
1 1 b
2 2 . NaN
3 3 . d

Alternatively, you can pass the nested dictionary like so:

In [118]: df.replace(regex={'b': {r'\s*\.\s*': np.nan}})

Out[118]:
a b c
0 0 a a
1 1 b b
2 2 NaN NaN
3 3 NaN d

You can also use the group of a regular expression match when replacing (dict of regex -> dict of regex), this works
for lists as well.

In [119]: df.replace({'b': r'\s*(\.)\s*'}, {'b': r'\1ty'}, regex=True)

Out[119]:
a b c
0 0 a a
1 1 b b
2 2 .ty NaN
3 3 .ty d

You can pass a list of regular expressions, of which those that match will be replaced with a scalar (list of regex ->
regex).

In [120]: df.replace([r'\s*\.\s*', r'a|b'], np.nan, regex=True)

Out[120]:
a b c
0 0 NaN NaN
1 1 NaN NaN
2 2 NaN NaN
3 3 NaN d

All of the regular expression examples can also be passed with the to_replace argument as the regex argument.
In this case the value argument must be passed explicitly by name or regex must be a nested dictionary. The
previous example, in this case, would then be:

In [121]: df.replace(regex=[r'\s*\.\s*', r'a|b'], value=np.nan)

Out[121]:
a b c
0 0 NaN NaN
1 1 NaN NaN
2 2 NaN NaN
3 3 NaN d

This can be convenient if you do not want to pass regex=True every time you want to use a regular expression.

Note: Anywhere in the above replace examples that you see a regular expression a compiled regular expression is

4.7. Working with missing data 501

pandas: powerful Python data analysis toolkit, Release 0.25.3

valid as well.

4.7.10 Numeric replacement

replace() is similar to fillna().

In [122]: df = pd.DataFrame(np.random.randn(10, 2))

In [123]: df[np.random.rand(df.shape[0]) > 0.5] = 1.5

In [124]: df.replace(1.5, np.nan)

Out[124]:
0 1
0 -0.844214 -1.021415
1 0.432396 -0.323580
2 0.423825 0.799180
3 1.262614 0.751965
4 NaN NaN
5 NaN NaN
6 -0.498174 -1.060799
7 0.591667 -0.183257
8 1.019855 -1.482465
9 NaN NaN

Replacing more than one value is possible by passing a list.

In [125]: df00 = df.iloc[0, 0]

In [126]: df.replace([1.5, df00], [np.nan, 'a'])

Out[126]:
0 1
0 a -1.02141
1 0.432396 -0.32358
2 0.423825 0.79918
3 1.26261 0.751965
4 NaN NaN
5 NaN NaN
6 -0.498174 -1.0608
7 0.591667 -0.183257
8 1.01985 -1.48247
9 NaN NaN

In [127]: df[1].dtype
Out[127]: dtype('float64')
You can also operate on the DataFrame in place:
In [128]: df.replace(1.5, np.nan, inplace=True)

Warning: When replacing multiple bool or datetime64 objects, the first argument to replace
(to_replace) must match the type of the value being replaced. For example,
>>> s = pd.Series([True, False, True])
>>> s.replace({'a string': 'new value', True: False}) # raises
TypeError: Cannot compare types 'ndarray(dtype=bool)' and 'str'

502 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

will raise a TypeError because one of the dict keys is not of the correct type for replacement.
However, when replacing a single object such as,
In [129]: s = pd.Series([True, False, True])

In [130]: s.replace('a string', 'another string')

Out[130]:
0 True
1 False
2 True
dtype: bool

the original NDFrame object will be returned untouched. Were working on unifying this API, but for backwards
compatibility reasons we cannot break the latter behavior. See GH6354 for more details.

Missing data casting rules and indexing

While pandas supports storing arrays of integer and boolean type, these types are not capable of storing missing data.
Until we can switch to using a native NA type in NumPy, weve established some casting rules. When a reindexing
operation introduces missing data, the Series will be cast according to the rules introduced in the table below.

data type Cast to

integer float
boolean object
float no cast
object no cast

For example:
In [131]: s = pd.Series(np.random.randn(5), index=[0, 2, 4, 6, 7])

In [132]: s > 0
Out[132]:
0 True
2 True
4 True
6 True
7 True
dtype: bool

In [133]: (s > 0).dtype

Out[133]: dtype('bool')

In [134]: crit = (s > 0).reindex(list(range(8)))

In [135]: crit
Out[135]:
0 True
1 NaN
2 True
3 NaN
4 True
5 NaN

4.7. Working with missing data 503

pandas: powerful Python data analysis toolkit, Release 0.25.3

6 True
7 True
dtype: object

In [136]: crit.dtype
Out[136]: dtype('O')
Ordinarily NumPy will complain if you try to use an object array (even if it contains boolean values) instead of a
boolean array to get or set values from an ndarray (e.g. selecting values based on some criteria). If a boolean vector
contains NAs, an exception will be generated:

In [137]: reindexed = s.reindex(list(range(8))).fillna(0)

In [138]: reindexed[crit]
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-138-0dac417a4890> in <module>
----> 1 reindexed[crit]

~/sandbox/pandas-release/pandas/pandas/core/series.py in __getitem__(self, key)

1108 key = list(key)
1109
-> 1110 if com.is_bool_indexer(key):
1111 key = check_bool_indexer(self.index, key)
1112

~/sandbox/pandas-release/pandas/pandas/core/common.py in is_bool_indexer(key)
128 if not lib.is_bool_array(key):
129 if isna(key).any():
--> 130 raise ValueError(na_msg)
131 return False
132 return True

ValueError: cannot index with vector containing NA / NaN values

However, these can be filled in using fillna() and it will work fine:
In [139]: reindexed[crit.fillna(False)]
Out[139]:
0 0.126504
2 0.696198
4 0.697416
6 0.601516
7 0.003659
dtype: float64

In [140]: reindexed[crit.fillna(True)]
Out[140]:
0 0.126504
1 0.000000
2 0.696198
3 0.000000
4 0.697416
5 0.000000
6 0.601516
7 0.003659
dtype: float64

504 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Pandas provides a nullable integer dtype, but you must explicitly request it when creating the series or column. Notice
that we use a capital I in the dtype="Int64".

In [141]: s = pd.Series([0, 1, np.nan, 3, 4], dtype="Int64")

In [142]: s
Out[142]:
0 0
1 1
2 NaN
3 3
4 4
dtype: Int64

See Nullable integer data type for more. {{ header }}

4.8 Categorical data

This is an introduction to pandas categorical data type, including a short comparison with Rs factor.
Categoricals are a pandas data type corresponding to categorical variables in statistics. A categorical variable takes
on a limited, and usually fixed, number of possible values (categories; levels in R). Examples are gender, social class,
blood type, country affiliation, observation time or rating via Likert scales.
In contrast to statistical categorical variables, categorical data might have an order (e.g. strongly agree vs agree or first
observation vs. second observation), but numerical operations (additions, divisions, ) are not possible.
All values of categorical data are either in categories or np.nan. Order is defined by the order of categories, not lexical
order of the values. Internally, the data structure consists of a categories array and an integer array of codes which
point to the real value in the categories array.
The categorical data type is useful in the following cases:
• A string variable consisting of only a few different values. Converting such a string variable to a categorical
variable will save some memory, see here.
• The lexical order of a variable is not the same as the logical order (one, two, three). By converting to a categorical
and specifying an order on the categories, sorting and min/max will use the logical order instead of the lexical
order, see here.
• As a signal to other Python libraries that this column should be treated as a categorical variable (e.g. to use
suitable statistical methods or plot types).
See also the API docs on categoricals.

4.8.1 Object creation

Series creation

Categorical Series or columns in a DataFrame can be created in several ways:

By specifying dtype="category" when constructing a Series:

In [1]: s = pd.Series(["a", "b", "c", "a"], dtype="category")

In [2]: s
(continues on next page)

4.8. Categorical data 505

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

Out[2]:
0 a
1 b
2 c
3 a
dtype: category
Categories (3, object): [a, b, c]

By converting an existing Series or column to a category dtype:

In [3]: df = pd.DataFrame({"A": ["a", "b", "c", "a"]})

In [4]: df["B"] = df["A"].astype('category')

In [5]: df
Out[5]:
A B
0 a a
1 b b
2 c c
3 a a

By using special functions, such as cut(), which groups data into discrete bins. See the example on tiling in the
docs.

In [6]: df = pd.DataFrame({'value': np.random.randint(0, 100, 20)})

In [7]: labels = ["{0} - {1}".format(i, i + 9) for i in range(0, 100, 10)]

In [8]: df['group'] = pd.cut(df.value, range(0, 105, 10), right=False, labels=labels)

In [9]: df.head(10)
Out[9]:
value group
0 82 80 - 89
1 8 0 - 9
2 67 60 - 69
3 2 0 - 9
4 41 40 - 49
5 44 40 - 49
6 79 70 - 79
7 66 60 - 69
8 4 0 - 9
9 62 60 - 69

By passing a pandas.Categorical object to a Series or assigning it to a DataFrame.

In [10]: raw_cat = pd.Categorical(["a", "b", "c", "a"], categories=["b", "c", "d"],

....: ordered=False)
....:

In [11]: s = pd.Series(raw_cat)

In [12]: s
Out[12]:
0 NaN
(continues on next page)

506 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

1 b
2 c
3 NaN
dtype: category
Categories (3, object): [b, c, d]

In [13]: df = pd.DataFrame({"A": ["a", "b", "c", "a"]})

In [14]: df["B"] = raw_cat

In [15]: df
Out[15]:
A B
0 a NaN
1 b b
2 c c
3 a NaN

Categorical data has a specific category dtype:

In [16]: df.dtypes
Out[16]:
A object
B category
dtype: object

DataFrame creation

Similar to the previous section where a single column was converted to categorical, all columns in a DataFrame can
be batch converted to categorical either during or after construction.
This can be done during construction by specifying dtype="category" in the DataFrame constructor:

In [17]: df = pd.DataFrame({'A': list('abca'), 'B': list('bccd')}, dtype="category")

In [18]: df.dtypes
Out[18]:
A category
B category
dtype: object

Note that the categories present in each column differ; the conversion is done column by column, so only labels present
in a given column are categories:
In [19]: df['A']
Out[19]:
0 a
1 b
2 c
3 a
Name: A, dtype: category
Categories (3, object): [a, b, c]

In [20]: df['B']
Out[20]:

4.8. Categorical data 507

pandas: powerful Python data analysis toolkit, Release 0.25.3

0 b
1 c
2 c
3 d
Name: B, dtype: category
Categories (3, object): [b, c, d]
New in version 0.23.0.
Analogously, all columns in an existing DataFrame can be batch converted using DataFrame.astype():

In [21]: df = pd.DataFrame({'A': list('abca'), 'B': list('bccd')})

In [22]: df_cat = df.astype('category')

In [23]: df_cat.dtypes
Out[23]:
A category
B category
dtype: object

This conversion is likewise done column by column:

In [24]: df_cat['A']
Out[24]:
0 a
1 b
2 c
3 a
Name: A, dtype: category
Categories (3, object): [a, b, c]

In [25]: df_cat['B']
Out[25]:
0 b
1 c
2 c
3 d
Name: B, dtype: category
Categories (3, object): [b, c, d]

Controlling behavior

In the examples above where we passed dtype='category', we used the default behavior:
1. Categories are inferred from the data.
2. Categories are unordered.
To control those behaviors, instead of passing 'category', use an instance of CategoricalDtype.

In [26]: from pandas.api.types import CategoricalDtype

In [27]: s = pd.Series(["a", "b", "c", "a"])

In [28]: cat_type = CategoricalDtype(categories=["b", "c", "d"],

....: ordered=True)
(continues on next page)

508 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

....:

In [29]: s_cat = s.astype(cat_type)

In [30]: s_cat
Out[30]:
0 NaN
1 b
2 c
3 NaN
dtype: category
Categories (3, object): [b < c < d]

Similarly, a CategoricalDtype can be used with a DataFrame to ensure that categories are consistent among
all columns.
In [31]: from pandas.api.types import CategoricalDtype

In [32]: df = pd.DataFrame({'A': list('abca'), 'B': list('bccd')})

In [33]: cat_type = CategoricalDtype(categories=list('abcd'),

....: ordered=True)
....:

In [34]: df_cat = df.astype(cat_type)

In [35]: df_cat['A']
Out[35]:
0 a
1 b
2 c
3 a
Name: A, dtype: category
Categories (4, object): [a < b < c < d]

In [36]: df_cat['B']
Out[36]:
0 b
1 c
2 c
3 d
Name: B, dtype: category
Categories (4, object): [a < b < c < d]

Note: To perform table-wise conversion, where all labels in the entire DataFrame are used as categories for each
column, the categories parameter can be determined programmatically by categories = pd.unique(df.
to_numpy().ravel()).

If you already have codes and categories, you can use the from_codes() constructor to save the factorize
step during normal constructor mode:
In [37]: splitter = np.random.choice([0, 1], 5, p=[0.5, 0.5])

(continues on next page)

4.8. Categorical data 509

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [38]: s = pd.Series(pd.Categorical.from_codes(splitter,
....: categories=["train", "test"]))
....:

Regaining original data

To get back to the original Series or NumPy array, use Series.astype(original_dtype) or np.
asarray(categorical):
In [39]: s = pd.Series(["a", "b", "c", "a"])

In [40]: s
Out[40]:
0 a
1 b
2 c
3 a
dtype: object

In [41]: s2 = s.astype('category')

In [42]: s2
Out[42]:
0 a
1 b
2 c
3 a
dtype: category
Categories (3, object): [a, b, c]

In [43]: s2.astype(str)
Out[43]:
0 a
1 b
2 c
3 a
dtype: object

In [44]: np.asarray(s2)
Out[44]: array(['a', 'b', 'c', 'a'], dtype=object)

Note: In contrast to Rs factor function, categorical data is not converting input values to strings; categories will end
up the same data type as the original values.

Note: In contrast to Rs factor function, there is currently no way to assign/change labels at creation time. Use
categories to change the categories after creation time.

510 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

4.8.2 CategoricalDtype

Changed in version 0.21.0.

A categoricals type is fully described by
1. categories: a sequence of unique values and no missing values
2. ordered: a boolean
This information can be stored in a CategoricalDtype. The categories argument is optional, which implies
that the actual categories should be inferred from whatever is present in the data when the pandas.Categorical
is created. The categories are assumed to be unordered by default.
In [45]: from pandas.api.types import CategoricalDtype

In [46]: CategoricalDtype(['a', 'b', 'c'])

Out[46]: CategoricalDtype(categories=['a', 'b', 'c'], ordered=None)

In [47]: CategoricalDtype(['a', 'b', 'c'], ordered=True)

Out[47]: CategoricalDtype(categories=['a', 'b', 'c'], ordered=True)

In [48]: CategoricalDtype()
Out[48]: CategoricalDtype(categories=None, ordered=None)
A CategoricalDtype can be used in any place pandas expects a dtype. For example pandas.read_csv(),
pandas.DataFrame.astype(), or in the Series constructor.

Note: As a convenience, you can use the string 'category' in place of a CategoricalDtype when you want
the default behavior of the categories being unordered, and equal to the set values present in the array. In other words,
dtype='category' is equivalent to dtype=CategoricalDtype().

Equality semantics

Two instances of CategoricalDtype compare equal whenever they have the same categories and order. When
comparing two unordered categoricals, the order of the categories is not considered.
In [49]: c1 = CategoricalDtype(['a', 'b', 'c'], ordered=False)

# Equal, since order is not considered when ordered=False

In [50]: c1 == CategoricalDtype(['b', 'c', 'a'], ordered=False)
Out[50]: True

# Unequal, since the second CategoricalDtype is ordered

In [51]: c1 == CategoricalDtype(['a', 'b', 'c'], ordered=True)
Out[51]: False
All instances of CategoricalDtype compare equal to the string 'category'.

In [52]: c1 == 'category'
Out[52]: True

4.8. Categorical data 511

pandas: powerful Python data analysis toolkit, Release 0.25.3

Warning: Since dtype='category' is essentially CategoricalDtype(None, False), and since

all instances CategoricalDtype compare equal to 'category', all instances of CategoricalDtype
compare equal to a CategoricalDtype(None, False), regardless of categories or ordered.

4.8.3 Description

Using describe() on categorical data will produce similar output to a Series or DataFrame of type string.
In [53]: cat = pd.Categorical(["a", "c", "c", np.nan], categories=["b", "a",
,→"c"])

In [54]: df = pd.DataFrame({"cat": cat, "s": ["a", "c", "c", np.nan]})

In [55]: df.describe()
Out[55]:
cat s
count 3 3
unique 2 2
top c c
freq 2 2

In [56]: df["cat"].describe()
Out[56]:
count 3
unique 2
top c
freq 2
Name: cat, dtype: object

4.8.4 Working with categories

Categorical data has a categories and a ordered property, which list their possible values and whether the ordering
matters or not. These properties are exposed as s.cat.categories and s.cat.ordered. If you dont manually
specify categories and ordering, they are inferred from the passed arguments.
In [57]: s = pd.Series(["a", "b", "c", "a"], dtype="category")

In [58]: s.cat.categories
Out[58]: Index(['a', 'b', 'c'], dtype='object')

In [59]: s.cat.ordered
Out[59]: False
Its also possible to pass in the categories in a specific order:
In [60]: s = pd.Series(pd.Categorical(["a", "b", "c", "a"],
....: categories=["c", "b", "a"]))
....:

In [61]: s.cat.categories
Out[61]: Index(['c', 'b', 'a'], dtype='object')

In [62]: s.cat.ordered

512 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[62]: False

Note: New categorical data are not automatically ordered. You must explicitly pass ordered=True to indicate an
ordered Categorical.

Note: The result of unique() is not always the same as Series.cat.categories, because Series.
unique() has a couple of guarantees, namely that it returns categories in the order of appearance, and it only
includes values that are actually present.
In [63]: s = pd.Series(list('babc')).astype(CategoricalDtype(list('abcd')))

In [64]: s
Out[64]:
0 b
1 a
2 b
3 c
dtype: category
Categories (4, object): [a, b, c, d]

# categories
In [65]: s.cat.categories
Out[65]: Index(['a', 'b', 'c', 'd'], dtype='object')

# uniques
In [66]: s.unique()
Out[66]:
[b, a, c]
Categories (3, object): [b, a, c]

Renaming categories

Renaming categories is done by assigning new values to the Series.cat.categories property or by using the
rename_categories() method:
In [67]: s = pd.Series(["a", "b", "c", "a"], dtype="category")

In [68]: s
Out[68]:
0 a
1 b
2 c
3 a
dtype: category
Categories (3, object): [a, b, c]

In [69]: s.cat.categories = ["Group %s" % g for g in s.cat.categories]

In [70]: s
Out[70]:
0 Group a
1 Group b
(continues on next page)

4.8. Categorical data 513

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2 Group c
3 Group a
dtype: category
Categories (3, object): [Group a, Group b, Group c]

In [71]: s = s.cat.rename_categories([1, 2, 3])

In [72]: s
Out[72]:
0 1
1 2
2 3
3 1
dtype: category
Categories (3, int64): [1, 2, 3]

# You can also pass a dict-like object to map the renaming

In [73]: s = s.cat.rename_categories({1: 'x', 2: 'y', 3: 'z'})

In [74]: s
Out[74]:
0 x
1 y
2 z
3 x
dtype: category
Categories (3, object): [x, y, z]

Note: In contrast to Rs factor, categorical data can have categories of other types than string.

Note: Be aware that assigning new categories is an inplace operation, while most other operations under Series.
cat per default return a new Series of dtype category.

Categories must be unique or a ValueError is raised:

In [75]: try:
....: s.cat.categories = [1, 1, 1]
....: except ValueError as e:
....: print("ValueError:", str(e))
....:
ValueError: Categorical categories must be unique

Categories must also not be NaN or a ValueError is raised:

In [76]: try:
....: s.cat.categories = [1, 2, np.nan]
....: except ValueError as e:
....: print("ValueError:", str(e))
....:
ValueError: Categorial categories cannot be null

514 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Appending new categories

Appending categories can be done by using the add_categories() method:

In [77]: s = s.cat.add_categories([4])

In [78]: s.cat.categories
Out[78]: Index(['x', 'y', 'z', 4], dtype='object')

In [79]: s
Out[79]:
0 x
1 y
2 z
3 x
dtype: category
Categories (4, object): [x, y, z, 4]

Removing categories

Removing categories can be done by using the remove_categories() method. Values which are removed are
replaced by np.nan.:

In [80]: s = s.cat.remove_categories([4])

In [81]: s
Out[81]:
0 x
1 y
2 z
3 x
dtype: category
Categories (3, object): [x, y, z]

Removing unused categories

Removing unused categories can also be done:

In [82]: s = pd.Series(pd.Categorical(["a", "b", "a"],
....: categories=["a", "b", "c", "d"]))
....:

In [83]: s
Out[83]:
0 a
1 b
2 a
dtype: category
Categories (4, object): [a, b, c, d]

In [84]: s.cat.remove_unused_categories()
Out[84]:
0 a
1 b

4.8. Categorical data 515

pandas: powerful Python data analysis toolkit, Release 0.25.3

2 a
dtype: category
Categories (2, object): [a, b]

Setting categories

If you want to do remove and add new categories in one step (which has some speed advantage), or simply set the
categories to a predefined scale, use set_categories().

In [85]: s = pd.Series(["one", "two", "four", "-"], dtype="category")

In [86]: s
Out[86]:
0 one
1 two
2 four
3 -
dtype: category
Categories (4, object): [-, four, one, two]

In [87]: s = s.cat.set_categories(["one", "two", "three", "four"])

In [88]: s
Out[88]:
0 one
1 two
2 four
3 NaN
dtype: category
Categories (4, object): [one, two, three, four]

Note: Be aware that Categorical.set_categories() cannot know whether some category is omitted in-
tentionally or because it is misspelled or (under Python3) due to a type difference (e.g., NumPy S1 dtype and Python
strings). This can result in surprising behaviour!

4.8.5 Sorting and order

If categorical data is ordered (s.cat.ordered == True), then the order of the categories has a meaning and
certain operations are possible. If the categorical is unordered, .min()/.max() will raise a TypeError.
In [89]: s = pd.Series(pd.Categorical(["a", "b", "c", "a"], ordered=False))

In [90]: s.sort_values(inplace=True)

In [91]: s = pd.Series(["a", "b", "c", "a"]).astype(

....: CategoricalDtype(ordered=True)
....: )
....:

In [92]: s.sort_values(inplace=True)

In [93]: s
Out[93]:

516 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

0 a
3 a
1 b
2 c
dtype: category
Categories (3, object): [a < b < c]

In [94]: s.min(), s.max()

Out[94]: ('a', 'c')
You can set categorical data to be ordered by using as_ordered() or unordered by using as_unordered().
These will by default return a new object.
In [95]: s.cat.as_ordered()
Out[95]:
0 a
3 a
1 b
2 c
dtype: category
Categories (3, object): [a < b < c]

In [96]: s.cat.as_unordered()
Out[96]:
0 a
3 a
1 b
2 c
dtype: category
Categories (3, object): [a, b, c]
Sorting will use the order defined by categories, not any lexical order present on the data type. This is even true for
strings and numeric data:
In [97]: s = pd.Series([1, 2, 3, 1], dtype="category")

In [98]: s = s.cat.set_categories([2, 3, 1], ordered=True)

In [99]: s
Out[99]:
0 1
1 2
2 3
3 1
dtype: category
Categories (3, int64): [2 < 3 < 1]

In [100]: s.sort_values(inplace=True)

In [101]: s
Out[101]:
1 2
2 3
0 1
3 1
dtype: category

4.8. Categorical data 517

pandas: powerful Python data analysis toolkit, Release 0.25.3

Categories (3, int64): [2 < 3 < 1]

In [102]: s.min(), s.max()

Out[102]: (2, 1)

Reordering

Reordering the categories is possible via the Categorical.reorder_categories() and the

Categorical.set_categories() methods. For Categorical.reorder_categories(), all old
categories must be included in the new categories and no new categories are allowed. This will necessarily make the
sort order the same as the categories order.
In [103]: s = pd.Series([1, 2, 3, 1], dtype="category")

In [104]: s = s.cat.reorder_categories([2, 3, 1], ordered=True)

In [105]: s
Out[105]:
0 1
1 2
2 3
3 1
dtype: category
Categories (3, int64): [2 < 3 < 1]

In [106]: s.sort_values(inplace=True)

In [107]: s
Out[107]:
1 2
2 3
0 1
3 1
dtype: category
Categories (3, int64): [2 < 3 < 1]

In [108]: s.min(), s.max()

Out[108]: (2, 1)

Note: Note the difference between assigning new categories and reordering the categories: the first renames categories
and therefore the individual values in the Series, but if the first position was sorted last, the renamed value will still
be sorted last. Reordering means that the way values are sorted is different afterwards, but not that individual values
in the Series are changed.

Note: If the Categorical is not ordered, Series.min() and Series.max() will raise TypeError. Nu-
meric operations like +, -, *, / and operations based on them (e.g. Series.median(), which would need to
compute the mean between two values if the length of an array is even) do not work and raise a TypeError.

518 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Multi column sorting

A categorical dtyped column will participate in a multi-column sort in a similar manner to other columns. The ordering
of the categorical is determined by the categories of that column.

In [109]: dfs = pd.DataFrame({'A': pd.Categorical(list('bbeebbaa'),

.....: categories=['e', 'a', 'b'],
.....: ordered=True),
.....: 'B': [1, 2, 1, 2, 2, 1, 2, 1]})
.....:

In [110]: dfs.sort_values(by=['A', 'B'])

Out[110]:
A B
2 e 1
3 e 2
7 a 1
6 a 2
0 b 1
5 b 1
1 b 2
4 b 2

Reordering the categories changes a future sort.

In [111]: dfs['A'] = dfs['A'].cat.reorder_categories(['a', 'b', 'e'])

In [112]: dfs.sort_values(by=['A', 'B'])

Out[112]:
A B
7 a 1
6 a 2
0 b 1
5 b 1
1 b 2
4 b 2
2 e 1
3 e 2

4.8.6 Comparisons

Comparing categorical data with other objects is possible in three cases:

• Comparing equality (== and !=) to a list-like object (list, Series, array, ) of the same length as the categorical
data.
• All comparisons (==, !=, >, >=, <, and <=) of categorical data to another categorical Series, when
ordered==True and the categories are the same.
• All comparisons of a categorical data to a scalar.
All other comparisons, especially non-equality comparisons of two categoricals with different categories or a categor-
ical with any list-like object, will raise a TypeError.

Note: Any non-equality comparisons of categorical data with a Series, np.array, list or categorical data with
different categories or ordering will raise a TypeError because custom categories ordering could be interpreted in

4.8. Categorical data 519

pandas: powerful Python data analysis toolkit, Release 0.25.3

two ways: one with taking into account the ordering and one without.

In [113]: cat = pd.Series([1, 2, 3]).astype(

.....: CategoricalDtype([3, 2, 1], ordered=True)
.....: )
.....:

In [114]: cat_base = pd.Series([2, 2, 2]).astype(

.....: CategoricalDtype([3, 2, 1], ordered=True)
.....: )
.....:

In [115]: cat_base2 = pd.Series([2, 2, 2]).astype(

.....: CategoricalDtype(ordered=True)
.....: )
.....:

In [116]: cat
Out[116]:
0 1
1 2
2 3
dtype: category
Categories (3, int64): [3 < 2 < 1]

In [117]: cat_base
Out[117]:
0 2
1 2
2 2
dtype: category
Categories (3, int64): [3 < 2 < 1]

In [118]: cat_base2
Out[118]:
0 2
1 2
2 2
dtype: category
Categories (1, int64): [2]
Comparing to a categorical with the same categories and ordering or to a scalar works:
In [119]: cat > cat_base
Out[119]:
0 True
1 False
2 False
dtype: bool

In [120]: cat > 2

Out[120]:
0 True
1 False

520 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

2 False
dtype: bool
Equality comparisons work with any list-like object of same length and scalars:
In [121]: cat == cat_base
Out[121]:
0 False
1 True
2 False
dtype: bool

In [122]: cat == np.array([1, 2, 3])

Out[122]:
0 True
1 True
2 True
dtype: bool

In [123]: cat == 2
Out[123]:
0 False
1 True
2 False
dtype: bool
This doesnt work because the categories are not the same:
In [124]: try:
.....: cat > cat_base2
.....: except TypeError as e:
.....: print("TypeError:", str(e))
.....:
TypeError: Categoricals can only be compared if 'categories' are the same. Categories
,→are different lengths

If you want to do a non-equality comparison of a categorical series with a list-like object which is not categorical data,
you need to be explicit and convert the categorical data back to the original values:
In [125]: base = np.array([1, 2, 3])

In [126]: try:
.....: cat > base
.....: except TypeError as e:
.....: print("TypeError:", str(e))
.....:
TypeError: Cannot compare a Categorical for op __gt__ with type <class 'numpy.
,→ndarray'>.

If you want to compare values, use 'np.asarray(cat) <op> other'.

In [127]: np.asarray(cat) > base

Out[127]: array([False, False, False])
When you compare two unordered categoricals with the same categories, the order is not considered:
In [128]: c1 = pd.Categorical(['a', 'b'], categories=['a', 'b'], ordered=False)

(continues on next page)

4.8. Categorical data 521

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [129]: c2 = pd.Categorical(['a', 'b'], categories=['b', 'a'], ordered=False)

In [130]: c1 == c2
Out[130]: array([ True, True])

4.8.7 Operations

Apart from Series.min(), Series.max() and Series.mode(), the following operations are possible with
categorical data:
Series methods like Series.value_counts() will use all categories, even if some categories are not present
in the data:

In [131]: s = pd.Series(pd.Categorical(["a", "b", "c", "c"],

.....: categories=["c", "a", "b", "d"]))
.....:

In [132]: s.value_counts()
Out[132]:
c 2
b 1
a 1
d 0
dtype: int64

Groupby will also show unused categories:

In [133]: cats = pd.Categorical(["a", "b", "b", "b", "c", "c", "c"],

.....: categories=["a", "b", "c", "d"])
.....:

In [134]: df = pd.DataFrame({"cats": cats, "values": [1, 2, 2, 2, 3, 4, 5]})

In [135]: df.groupby("cats").mean()
Out[135]:
values
cats
a 1.0
b 2.0
c 4.0
d NaN

In [136]: cats2 = pd.Categorical(["a", "a", "b", "b"], categories=["a", "b", "c"])

In [137]: df2 = pd.DataFrame({"cats": cats2,

.....: "B": ["c", "d", "c", "d"],
.....: "values": [1, 2, 3, 4]})
.....:

In [138]: df2.groupby(["cats", "B"]).mean()

Out[138]:
values
cats B
a c 1.0
d 2.0
(continues on next page)

522 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

b c 3.0
d 4.0
c c NaN
d NaN

Pivot tables:

In [139]: raw_cat = pd.Categorical(["a", "a", "b", "b"], categories=["a", "b", "c"])

In [140]: df = pd.DataFrame({"A": raw_cat,

.....: "B": ["c", "d", "c", "d"],
.....: "values": [1, 2, 3, 4]})
.....:

In [141]: pd.pivot_table(df, values='values', index=['A', 'B'])

Out[141]:
values
A B
a c 1
d 2
b c 3
d 4

4.8.8 Data munging

The optimized pandas data access methods .loc, .iloc, .at, and .iat, work as normal. The only difference is
the return type (for getting) and that only values already in categories can be assigned.

Getting

If the slicing operation returns either a DataFrame or a column of type Series, the category dtype is preserved.
In [142]: idx = pd.Index(["h", "i", "j", "k", "l", "m", "n"])

In [143]: cats = pd.Series(["a", "b", "b", "b", "c", "c", "c"],

.....: dtype="category", index=idx)
.....:

In [144]: values = [1, 2, 2, 2, 3, 4, 5]

In [145]: df = pd.DataFrame({"cats": cats, "values": values}, index=idx)

In [146]: df.iloc[2:4, :]
Out[146]:
cats values
j b 2
k b 2

In [147]: df.iloc[2:4, :].dtypes

Out[147]:
cats category
values int64
dtype: object

4.8. Categorical data 523

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [148]: df.loc["h":"j", "cats"]

Out[148]:
h a
i b
j b
Name: cats, dtype: category
Categories (3, object): [a, b, c]

In [149]: df[df["cats"] == "b"]

Out[149]:
cats values
i b 2
j b 2
k b 2
An example where the category type is not preserved is if you take one single row: the resulting Series is of dtype
object:

# get the complete "h" row as a Series

In [150]: df.loc["h", :]
Out[150]:
cats a
values 1
Name: h, dtype: object

Returning a single item from categorical data will also return the value, not a categorical of length 1.

In [151]: df.iat[0, 0]
Out[151]: 'a'

In [152]: df["cats"].cat.categories = ["x", "y", "z"]

In [153]: df.at["h", "cats"] # returns a string

Out[153]: 'x'

Note: The is in contrast to Rs factor function, where factor(c(1,2,3))[1] returns a single value factor.

To get a single value Series of type category, you pass in a list with a single value:

In [154]: df.loc[["h"], "cats"]

Out[154]:
h x
Name: cats, dtype: category
Categories (3, object): [x, y, z]

String and datetime accessors

The accessors .dt and .str will work if the s.cat.categories are of an appropriate type:
In [155]: str_s = pd.Series(list('aabb'))

In [156]: str_cat = str_s.astype('category')

524 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [157]: str_cat
Out[157]:
0 a
1 a
2 b
3 b
dtype: category
Categories (2, object): [a, b]

In [158]: str_cat.str.contains("a")
Out[158]:
0 True
1 True
2 False
3 False
dtype: bool

In [159]: date_s = pd.Series(pd.date_range('1/1/2015', periods=5))

In [160]: date_cat = date_s.astype('category')

In [161]: date_cat
Out[161]:
0 2015-01-01
1 2015-01-02
2 2015-01-03
3 2015-01-04
4 2015-01-05
dtype: category
Categories (5, datetime64[ns]): [2015-01-01, 2015-01-02, 2015-01-03, 2015-01-
,→04, 2015-01-05]

In [162]: date_cat.dt.day
Out[162]:
0 1
1 2
2 3
3 4
4 5
dtype: int64

Note: The returned Series (or DataFrame) is of the same type as if you used the .str.<method> / .dt.
<method> on a Series of that type (and not of type category!).

That means, that the returned values from methods and properties on the accessors of a Series and the returned
values from methods and properties on the accessors of this Series transformed to one of type category will be
equal:
In [163]: ret_s = str_s.str.contains("a")

In [164]: ret_cat = str_cat.str.contains("a")

In [165]: ret_s.dtype == ret_cat.dtype

4.8. Categorical data 525

pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[165]: True

In [166]: ret_s == ret_cat

Out[166]:
0 True
1 True
2 True
3 True
dtype: bool

Note: The work is done on the categories and then a new Series is constructed. This has some performance
implication if you have a Series of type string, where lots of elements are repeated (i.e. the number of unique
elements in the Series is a lot smaller than the length of the Series). In this case it can be faster to convert the
original Series to one of type category and use .str.<method> or .dt.<property> on that.

Setting

Setting values in a categorical column (or Series) works as long as the value is included in the categories:
In [167]: idx = pd.Index(["h", "i", "j", "k", "l", "m", "n"])

In [168]: cats = pd.Categorical(["a", "a", "a", "a", "a", "a", "a"],

.....: categories=["a", "b"])
.....:

In [169]: values = [1, 1, 1, 1, 1, 1, 1]

In [170]: df = pd.DataFrame({"cats": cats, "values": values}, index=idx)

In [171]: df.iloc[2:4, :] = [["b", 2], ["b", 2]]

In [172]: df
Out[172]:
cats values
h a 1
i a 1
j b 2
k b 2
l a 1
m a 1
n a 1

In [173]: try:
.....: df.iloc[2:4, :] = [["c", 3], ["c", 3]]
.....: except ValueError as e:
.....: print("ValueError:", str(e))
.....:
ValueError: Cannot setitem on a Categorical with a new category, set the
,→categories first

Setting values by assigning categorical data will also check that the categories match:
In [174]: df.loc["j":"k", "cats"] = pd.Categorical(["a", "a"],

526 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

,→categories=["a", "b"])

In [175]: df
Out[175]:
cats values
h a 1
i a 1
j a 2
k a 2
l a 1
m a 1
n a 1

In [176]: try:
.....: df.loc["j":"k", "cats"] = pd.Categorical(["b", "b"],
.....: categories=["a", "b",
,→"c"])

.....: except ValueError as e:

.....: print("ValueError:", str(e))
.....:
ValueError: Cannot set a Categorical with another, without identical
,→categories

Assigning a Categorical to parts of a column of other types will use the values:
In [177]: df = pd.DataFrame({"a": [1, 1, 1, 1, 1], "b": ["a", "a", "a", "a",
,→"a"]})

In [178]: df.loc[1:2, "a"] = pd.Categorical(["b", "b"], categories=["a", "b"])

In [179]: df.loc[2:3, "b"] = pd.Categorical(["b", "b"], categories=["a", "b"])

In [180]: df
Out[180]:
a b
0 1 a
1 b a
2 b b
3 1 b
4 1 a

In [181]: df.dtypes
Out[181]:
a object
b object
dtype: object

Merging

You can concat two DataFrames containing categorical data together, but the categories of these categoricals need
to be the same:
In [182]: cat = pd.Series(["a", "b"], dtype="category")

In [183]: vals = [1, 2]

4.8. Categorical data 527

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [184]: df = pd.DataFrame({"cats": cat, "vals": vals})

In [185]: res = pd.concat([df, df])

In [186]: res
Out[186]:
cats vals
0 a 1
1 b 2
0 a 1
1 b 2

In [187]: res.dtypes
Out[187]:
cats category
vals int64
dtype: object
In this case the categories are not the same, and therefore an error is raised:

In [188]: df_different = df.copy()

In [189]: df_different["cats"].cat.categories = ["c", "d"]

In [190]: try:
.....: pd.concat([df, df_different])
.....: except ValueError as e:
.....: print("ValueError:", str(e))
.....:

The same applies to df.append(df_different).

See also the section on merge dtypes for notes about preserving merge dtypes and performance.

Unioning

New in version 0.19.0.

If you want to combine categoricals that do not necessarily have the same categories, the union_categoricals()
function will combine a list-like of categoricals. The new categories will be the union of the categories being combined.

In [191]: from pandas.api.types import union_categoricals

In [192]: a = pd.Categorical(["b", "c"])

In [193]: b = pd.Categorical(["a", "b"])

In [194]: union_categoricals([a, b])

Out[194]:
[b, c, a, b]
Categories (3, object): [b, c, a]

By default, the resulting categories will be ordered as they appear in the data. If you want the categories to be lexsorted,
use sort_categories=True argument.

528 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [195]: union_categoricals([a, b], sort_categories=True)

Out[195]:
[b, c, a, b]
Categories (3, object): [a, b, c]

union_categoricals also works with the easy case of combining two categoricals of the same categories and
order information (e.g. what you could also append for).

In [196]: a = pd.Categorical(["a", "b"], ordered=True)

In [197]: b = pd.Categorical(["a", "b", "a"], ordered=True)

In [198]: union_categoricals([a, b])

Out[198]:
[a, b, a, b, a]
Categories (2, object): [a < b]

The below raises TypeError because the categories are ordered and not identical.

In [1]: a = pd.Categorical(["a", "b"], ordered=True)

In [2]: b = pd.Categorical(["a", "b", "c"], ordered=True)
In [3]: union_categoricals([a, b])
Out[3]:
TypeError: to union ordered Categoricals, all categories must be the same

New in version 0.20.0.

Ordered categoricals with different categories or orderings can be combined by using the ignore_ordered=True
argument.

In [199]: a = pd.Categorical(["a", "b", "c"], ordered=True)

In [200]: b = pd.Categorical(["c", "b", "a"], ordered=True)

In [201]: union_categoricals([a, b], ignore_order=True)

Out[201]:
[a, b, c, c, b, a]
Categories (3, object): [a, b, c]

union_categoricals() also works with a CategoricalIndex, or Series containing categorical data, but
note that the resulting array will always be a plain Categorical:

In [202]: a = pd.Series(["b", "c"], dtype='category')

In [203]: b = pd.Series(["a", "b"], dtype='category')

In [204]: union_categoricals([a, b])

Out[204]:
[b, c, a, b]
Categories (3, object): [b, c, a]

Note: union_categoricals may recode the integer codes for categories when combining categoricals. This is
likely what you want, but if you are relying on the exact numbering of the categories, be aware.
In [205]: c1 = pd.Categorical(["b", "c"])

In [206]: c2 = pd.Categorical(["a", "b"])

4.8. Categorical data 529

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [207]: c1
Out[207]:
[b, c]
Categories (2, object): [b, c]

# "b" is coded to 0
In [208]: c1.codes
Out[208]: array([0, 1], dtype=int8)

In [209]: c2
Out[209]:
[a, b]
Categories (2, object): [a, b]

# "b" is coded to 1
In [210]: c2.codes
Out[210]: array([0, 1], dtype=int8)

In [211]: c = union_categoricals([c1, c2])

In [212]: c
Out[212]:
[b, c, a, b]
Categories (3, object): [b, c, a]

# "b" is coded to 0 throughout, same as c1, different from c2

In [213]: c.codes
Out[213]: array([0, 1, 2, 0], dtype=int8)

Concatenation

This section describes concatenations specific to category dtype. See Concatenating objects for general description.
By default, Series or DataFrame concatenation which contains the same categories results in category dtype,
otherwise results in object dtype. Use .astype or union_categoricals to get category result.
# same categories
In [214]: s1 = pd.Series(['a', 'b'], dtype='category')

In [215]: s2 = pd.Series(['a', 'b', 'a'], dtype='category')

In [216]: pd.concat([s1, s2])

Out[216]:
0 a
1 b
0 a
1 b
2 a
dtype: category
Categories (2, object): [a, b]

# different categories

530 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [217]: s3 = pd.Series(['b', 'c'], dtype='category')

In [218]: pd.concat([s1, s3])

Out[218]:
0 a
1 b
0 b
1 c
dtype: object

In [219]: pd.concat([s1, s3]).astype('category')

Out[219]:
0 a
1 b
0 b
1 c
dtype: category
Categories (3, object): [a, b, c]

In [220]: union_categoricals([s1.array, s3.array])

Out[220]:
[a, b, b, c]
Categories (3, object): [a, b, c]
Following table summarizes the results of Categoricals related concatenations.

arg1 arg2 result

category category (identical categories) category
category category (different categories, both not ordered) object (dtype is inferred)
category category (different categories, either one is ordered) object (dtype is inferred)
category not category object (dtype is inferred)

4.8.9 Getting data in/out

You can write data that contains category dtypes to a HDFStore. See here for an example and caveats.
It is also possible to write data to and reading data from Stata format files. See here for an example and caveats.
Writing to a CSV file will convert the data, effectively removing any information about the categorical (categories and
ordering). So if you read back the CSV file you have to convert the relevant columns back to category and assign the
right categories and categories ordering.
In [221]: import io

In [222]: s = pd.Series(pd.Categorical(['a', 'b', 'b', 'a', 'a', 'd']))

# rename the categories

In [223]: s.cat.categories = ["very good", "good", "bad"]

# reorder the categories and add missing categories

In [224]: s = s.cat.set_categories(["very bad", "bad", "medium", "good",
,→"very good"])

In [225]: df = pd.DataFrame({"cats": s, "vals": [1, 2, 3, 4, 5, 6]})

4.8. Categorical data 531

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [226]: csv = io.StringIO()

In [227]: df.to_csv(csv)

In [228]: df2 = pd.read_csv(io.StringIO(csv.getvalue()))

In [229]: df2.dtypes
Out[229]:
Unnamed: 0 int64
cats object
vals int64
dtype: object

In [230]: df2["cats"]
Out[230]:
0 very good
1 good
2 good
3 very good
4 very good
5 bad
Name: cats, dtype: object

# Redo the category

In [231]: df2["cats"] = df2["cats"].astype("category")

In [232]: df2["cats"].cat.set_categories(["very bad", "bad", "medium",

.....: "good", "very good"],
.....: inplace=True)
.....:

In [233]: df2.dtypes
Out[233]:
Unnamed: 0 int64
cats category
vals int64
dtype: object

In [234]: df2["cats"]
Out[234]:
0 very good
1 good
2 good
3 very good
4 very good
5 bad
Name: cats, dtype: category
Categories (5, object): [very bad, bad, medium, good, very good]
The same holds for writing to a SQL database with to_sql.

532 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

4.8.10 Missing data

pandas primarily uses the value np.nan to represent missing data. It is by default not included in computations. See
the Missing Data section.
Missing values should not be included in the Categoricals categories, only in the values. Instead, it is under-
stood that NaN is different, and is always a possibility. When working with the Categoricals codes, missing values
will always have a code of -1.
In [235]: s = pd.Series(["a", "b", np.nan, "a"], dtype="category")

# only two categories

In [236]: s
Out[236]:
0 a
1 b
2 NaN
3 a
dtype: category
Categories (2, object): [a, b]

In [237]: s.cat.codes
Out[237]:
0 0
1 1
2 -1
3 0
dtype: int8
Methods for working with missing data, e.g. isna(), fillna(), dropna(), all work normally:
In [238]: s = pd.Series(["a", "b", np.nan], dtype="category")

In [239]: s
Out[239]:
0 a
1 b
2 NaN
dtype: category
Categories (2, object): [a, b]

In [240]: pd.isna(s)
Out[240]:
0 False
1 False
2 True
dtype: bool

In [241]: s.fillna("a")
Out[241]:
0 a
1 b
2 a
dtype: category
Categories (2, object): [a, b]

4.8. Categorical data 533

pandas: powerful Python data analysis toolkit, Release 0.25.3

4.8.11 Differences to Rs factor

The following differences to Rs factor functions can be observed:

• Rs levels are named categories.
• Rs levels are always of type string, while categories in pandas can be of any dtype.
• Its not possible to specify labels at creation time. Use s.cat.rename_categories(new_labels) af-
terwards.
• In contrast to Rs factor function, using categorical data as the sole input to create a new categorical series will
not remove unused categories but create a new categorical series which is equal to the passed in one!
• R allows for missing values to be included in its levels (pandas categories). Pandas does not allow NaN cate-
gories, but missing values can still be in the values.

4.8.12 Gotchas

Memory usage

The memory usage of a Categorical is proportional to the number of categories plus the length of the data. In
contrast, an object dtype is a constant times the length of the data.
In [242]: s = pd.Series(['foo', 'bar'] * 1000)

# object dtype
In [243]: s.nbytes
Out[243]: 16000

# category dtype
In [244]: s.astype('category').nbytes
Out[244]: 2016

Note: If the number of categories approaches the length of the data, the Categorical will use nearly the same or
more memory than an equivalent object dtype representation.
In [245]: s = pd.Series(['foo%04d' % i for i in range(2000)])

# object dtype
In [246]: s.nbytes
Out[246]: 16000

# category dtype
In [247]: s.astype('category').nbytes
Out[247]: 20000

Categorical is not a numpy array

Currently, categorical data and the underlying Categorical is implemented as a Python object and not as a low-
level NumPy array dtype. This leads to some problems.
NumPy itself doesnt know about the new dtype:

534 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [248]: try:
.....: np.dtype("category")
.....: except TypeError as e:
.....: print("TypeError:", str(e))
.....:
TypeError: data type "category" not understood

In [249]: dtype = pd.Categorical(["a"]).dtype

In [250]: try:
.....: np.dtype(dtype)
.....: except TypeError as e:
.....: print("TypeError:", str(e))
.....:
TypeError: data type not understood

Dtype comparisons work:

In [251]: dtype == np.str_
Out[251]: False

In [252]: np.str_ == dtype

Out[252]: False
To check if a Series contains Categorical data, use hasattr(s, 'cat'):
In [253]: hasattr(pd.Series(['a'], dtype='category'), 'cat')
Out[253]: True

In [254]: hasattr(pd.Series(['a']), 'cat')

Out[254]: False
Using NumPy functions on a Series of type category should not work as Categoricals are not numeric data (even
in the case that .categories is numeric).

In [255]: s = pd.Series(pd.Categorical([1, 2, 3, 4]))

In [256]: try:
.....: np.sum(s)
.....: except TypeError as e:
.....: print("TypeError:", str(e))
.....:
TypeError: Categorical cannot perform the operation sum

Note: If such a function works, please file a bug at https://github.com/pandas-dev/pandas!

dtype in apply

Pandas currently does not preserve the dtype in apply functions: If you apply along rows you get a Series of object
dtype (same as getting a row -> getting one element will return a basic type) and applying along columns will also
convert to object. NaN values are unaffected. You can use fillna to handle missing values before applying a
function.
In [257]: df = pd.DataFrame({"a": [1, 2, 3, 4],
.....: "b": ["a", "b", "c", "d"],

4.8. Categorical data 535

pandas: powerful Python data analysis toolkit, Release 0.25.3

.....: "cats": pd.Categorical([1, 2, 3, 2])})

.....:

In [258]: df.apply(lambda row: type(row["cats"]), axis=1)

Out[258]:
0 <class 'int'>
1 <class 'int'>
2 <class 'int'>
3 <class 'int'>
dtype: object

In [259]: df.apply(lambda col: col.dtype, axis=0)

Out[259]:
a int64
b object
cats category
dtype: object

Categorical index

CategoricalIndex is a type of index that is useful for supporting indexing with duplicates. This is a container
around a Categorical and allows efficient indexing and storage of an index with a large number of duplicated
elements. See the advanced indexing docs for a more detailed explanation.
Setting the index will create a CategoricalIndex:
In [260]: cats = pd.Categorical([1, 2, 3, 4], categories=[4, 2, 3, 1])

In [261]: strings = ["a", "b", "c", "d"]

In [262]: values = [4, 2, 3, 1]

In [263]: df = pd.DataFrame({"strings": strings, "values": values},

,→index=cats)

In [264]: df.index
Out[264]: CategoricalIndex([1, 2, 3, 4], categories=[4, 2, 3, 1],
,→ordered=False, dtype='category')

# This now sorts by the categories order

In [265]: df.sort_index()
Out[265]:
strings values
4 d 1
2 b 2
3 c 3
1 a 4

Side effects

Constructing a Series from a Categorical will not copy the input Categorical. This means that changes to
the Series will in most cases change the original Categorical:

536 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [266]: cat = pd.Categorical([1, 2, 3, 10], categories=[1, 2, 3, 4, 10])

In [267]: s = pd.Series(cat, name="cat")

In [268]: cat
Out[268]:
[1, 2, 3, 10]
Categories (5, int64): [1, 2, 3, 4, 10]

In [269]: s.iloc[0:2] = 10

In [270]: cat
Out[270]:
[10, 10, 3, 10]
Categories (5, int64): [1, 2, 3, 4, 10]

In [271]: df = pd.DataFrame(s)

In [272]: df["cat"].cat.categories = [1, 2, 3, 4, 5]

In [273]: cat
Out[273]:
[5, 5, 3, 5]
Categories (5, int64): [1, 2, 3, 4, 5]

Use copy=True to prevent such a behaviour or simply dont reuse Categoricals:

In [274]: cat = pd.Categorical([1, 2, 3, 10], categories=[1, 2, 3, 4, 10])

In [275]: s = pd.Series(cat, name="cat", copy=True)

In [276]: cat
Out[276]:
[1, 2, 3, 10]
Categories (5, int64): [1, 2, 3, 4, 10]

In [277]: s.iloc[0:2] = 10

In [278]: cat
Out[278]:
[1, 2, 3, 10]
Categories (5, int64): [1, 2, 3, 4, 10]

Note: This also happens in some cases when you supply a NumPy array instead of a Categorical: using an
int array (e.g. np.array([1,2,3,4])) will exhibit the same behavior, while using a string array (e.g. np.
array(["a","b","c","a"])) will not.

{{ header }}

4.9 Nullable integer data type

New in version 0.24.0.

4.9. Nullable integer data type 537

pandas: powerful Python data analysis toolkit, Release 0.25.3

Note: IntegerArray is currently experimental. Its API or implementation may change without warning.

In Missing Data, we saw that pandas primarily uses NaN to represent missing data. Because NaN is a float, this forces
an array of integers with any missing values to become floating point. In some cases, this may not matter much. But if
your integer column is, say, an identifier, casting to float can be problematic. Some integers cannot even be represented
as floating point numbers.
Pandas can represent integer data with possibly missing values using arrays.IntegerArray. This is an extension
types implemented within pandas. It is not the default dtype for integers, and will not be inferred; you must explicitly
pass the dtype into array() or Series:

In [1]: arr = pd.array([1, 2, np.nan], dtype=pd.Int64Dtype())

In [2]: arr
Out[2]:
<IntegerArray>
[1, 2, NaN]
Length: 3, dtype: Int64

Or the string alias "Int64" (note the capital "I", to differentiate from NumPys 'int64' dtype:

In [3]: pd.array([1, 2, np.nan], dtype="Int64")

Out[3]:
<IntegerArray>
[1, 2, NaN]
Length: 3, dtype: Int64

This array can be stored in a DataFrame or Series like any NumPy array.

In [4]: pd.Series(arr)
Out[4]:
0 1
1 2
2 NaN
dtype: Int64

You can also pass the list-like object to the Series constructor with the dtype.

In [5]: s = pd.Series([1, 2, np.nan], dtype="Int64")

In [6]: s
Out[6]:
0 1
1 2
2 NaN
dtype: Int64

By default (if you dont specify dtype), NumPy is used, and youll end up with a float64 dtype Series:

In [7]: pd.Series([1, 2, np.nan])

Out[7]:
0 1.0
1 2.0
2 NaN
dtype: float64

538 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Operations involving an integer array will behave similar to NumPy arrays. Missing values will be propagated, and
and the data will be coerced to another dtype if needed.
# arithmetic
In [8]: s + 1
Out[8]:
0 2
1 3
2 NaN
dtype: Int64

# comparison
In [9]: s == 1
Out[9]:
0 True
1 False
2 False
dtype: bool

# indexing
In [10]: s.iloc[1:3]
Out[10]:
1 2
2 NaN
dtype: Int64

# operate with other dtypes

In [11]: s + s.iloc[1:3].astype('Int8')
Out[11]:
0 NaN
1 4
2 NaN
dtype: Int64

# coerce when needed

In [12]: s + 0.01
Out[12]:
0 1.01
1 2.01
2 NaN
dtype: float64
These dtypes can operate as part of of DataFrame.
In [13]: df = pd.DataFrame({'A': s, 'B': [1, 1, 3], 'C': list('aab')})

In [14]: df
Out[14]:
A B C
0 1 1 a
1 2 1 a
2 NaN 3 b

In [15]: df.dtypes
Out[15]:

4.9. Nullable integer data type 539

pandas: powerful Python data analysis toolkit, Release 0.25.3

A Int64
B int64
C object
dtype: object
These dtypes can be merged & reshaped & casted.
In [16]: pd.concat([df[['A']], df[['B', 'C']]], axis=1).dtypes
Out[16]:
A Int64
B int64
C object
dtype: object

In [17]: df['A'].astype(float)
Out[17]:
0 1.0
1 2.0
2 NaN
Name: A, dtype: float64
Reduction and groupby operations such as sum work as well.
In [18]: df.sum()
Out[18]:
A 3
B 5
C aab
dtype: object

In [19]: df.groupby('B').A.sum()
Out[19]:
B
1 3
3 0
Name: A, dtype: Int64
{{ header }}

4.10 Visualization

We use the standard convention for referencing the matplotlib API:

In [1]: import matplotlib.pyplot as plt

In [2]: plt.close('all')

We provide the basics in pandas to easily create decent looking plots. See the ecosystem section for visualization
libraries that go beyond the basics documented here.

Note: All calls to np.random are seeded with 123456.

540 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

4.10.1 Basic plotting: plot

We will demonstrate the basics, see the cookbook for some advanced strategies.
The plot method on Series and DataFrame is just a simple wrapper around plt.plot():

In [3]: ts = pd.Series(np.random.randn(1000),
...: index=pd.date_range('1/1/2000', periods=1000))
...:

In [4]: ts = ts.cumsum()

In [5]: ts.plot()
Out[5]: <matplotlib.axes._subplots.AxesSubplot at 0x1c36223c90>

If the index consists of dates, it calls gcf().autofmt_xdate() to try to format the x-axis nicely as per above.
On DataFrame, plot() is a convenience to plot all of the columns with labels:

In [6]: df = pd.DataFrame(np.random.randn(1000, 4),

...: index=ts.index, columns=list('ABCD'))
...:

In [7]: df = df.cumsum()

(continues on next page)

4.10. Visualization 541

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [8]: plt.figure();

In [9]: df.plot();

You can plot one column versus another using the x and y keywords in plot():

In [10]: df3 = pd.DataFrame(np.random.randn(1000, 2), columns=['B', 'C']).cumsum()

In [11]: df3['A'] = pd.Series(list(range(len(df))))

In [12]: df3.plot(x='A', y='B')

Out[12]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3a20a490>

542 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Note: For more formatting and styling options, see formatting below.

4.10.2 Other plots

Plotting methods allow for a handful of plot styles other than the default line plot. These methods can be provided as
the kind keyword argument to plot(), and include:
• bar or barh for bar plots
• hist for histogram
• box for boxplot
• kde or density for density plots
• area for area plots
• scatter for scatter plots
• hexbin for hexagonal bin plots
• pie for pie plots
For example, a bar plot can be created the following way:

4.10. Visualization 543

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [13]: plt.figure();

In [14]: df.iloc[5].plot(kind='bar');

You can also create these other plots using the methods DataFrame.plot.<kind> instead of providing the kind
keyword argument. This makes it easier to discover plot methods and the specific arguments they use:

In [15]: df = pd.DataFrame()

In [16]: df.plot.<TAB> # noqa: E225, E999

df.plot.area df.plot.barh df.plot.density df.plot.hist df.plot.line
,→df.plot.scatter

df.plot.bar df.plot.box df.plot.hexbin df.plot.kde df.plot.pie

In addition to these kind s, there are the DataFrame.hist(), and DataFrame.boxplot() methods, which use a separate
interface.
Finally, there are several plotting functions in pandas.plotting that take a Series or DataFrame as an argu-
ment. These include:
• Scatter Matrix
• Andrews Curves
• Parallel Coordinates

544 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

• Lag Plot
• Autocorrelation Plot
• Bootstrap Plot
• RadViz
Plots may also be adorned with errorbars or tables.

Bar plots

For labeled, non-time series data, you may wish to produce a bar plot:

In [17]: plt.figure();

In [18]: df.iloc[5].plot.bar()
Out[18]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3aa7aa50>

In [19]: plt.axhline(0, color='k');

Calling a DataFrames plot.bar() method produces a multiple bar plot:

4.10. Visualization 545

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [20]: df2 = pd.DataFrame(np.random.rand(10, 4), columns=['a', 'b', 'c', 'd'])

In [21]: df2.plot.bar();

To produce a stacked bar plot, pass stacked=True:

In [22]: df2.plot.bar(stacked=True);

546 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

To get horizontal bar plots, use the barh method:

In [23]: df2.plot.barh(stacked=True);

4.10. Visualization 547

pandas: powerful Python data analysis toolkit, Release 0.25.3

Histograms

Histograms can be drawn by using the DataFrame.plot.hist() and Series.plot.hist() methods.

In [24]: df4 = pd.DataFrame({'a': np.random.randn(1000) + 1, 'b': np.random.

,→randn(1000),

....: 'c': np.random.randn(1000) - 1}, columns=['a', 'b', 'c'])

....:

In [25]: plt.figure();

In [26]: df4.plot.hist(alpha=0.5)
Out[26]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3af8db50>

548 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

A histogram can be stacked using stacked=True. Bin size can be changed using the bins keyword.

In [27]: plt.figure();

In [28]: df4.plot.hist(stacked=True, bins=20)

Out[28]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3ae12d90>

4.10. Visualization 549

pandas: powerful Python data analysis toolkit, Release 0.25.3

You can pass other keywords supported by matplotlib hist. For example, horizontal and cumulative histograms can
be drawn by orientation='horizontal' and cumulative=True.

In [29]: plt.figure();

In [30]: df4['a'].plot.hist(orientation='horizontal', cumulative=True)

Out[30]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3a171890>

550 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

See the hist method and the matplotlib hist documentation for more.
The existing interface DataFrame.hist to plot histogram still can be used.

In [31]: plt.figure();

In [32]: df['A'].diff().hist()
Out[32]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3a1822d0>

4.10. Visualization 551

pandas: powerful Python data analysis toolkit, Release 0.25.3

DataFrame.hist() plots the histograms of the columns on multiple subplots:

In [33]: plt.figure()
Out[33]: <Figure size 640x480 with 0 Axes>

In [34]: df.diff().hist(color='k', alpha=0.5, bins=50)

Out[34]:
array([[<matplotlib.axes._subplots.AxesSubplot object at 0x1c39ff60d0>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c3a3809d0>],
[<matplotlib.axes._subplots.AxesSubplot object at 0x1c36c54c90>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c36b5df90>]],
dtype=object)

552 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

The by keyword can be specified to plot grouped histograms:

In [35]: data = pd.Series(np.random.randn(1000))

In [36]: data.hist(by=np.random.randint(0, 4, 1000), figsize=(6, 4))

Out[36]:
array([[<matplotlib.axes._subplots.AxesSubplot object at 0x1c39ff6110>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c4042b810>],
[<matplotlib.axes._subplots.AxesSubplot object at 0x1c404b3e50>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c404ed6d0>]],
dtype=object)

4.10. Visualization 553

pandas: powerful Python data analysis toolkit, Release 0.25.3

Box plots

Boxplot can be drawn calling Series.plot.box() and DataFrame.plot.box(), or DataFrame.

boxplot() to visualize the distribution of values within each column.
For instance, here is a boxplot representing five trials of 10 observations of a uniform random variable on [0,1).

In [37]: df = pd.DataFrame(np.random.rand(10, 5), columns=['A', 'B', 'C', 'D', 'E'])

In [38]: df.plot.box()
Out[38]: <matplotlib.axes._subplots.AxesSubplot at 0x1c405fc410>

554 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Boxplot can be colorized by passing color keyword. You can pass a dict whose keys are boxes, whiskers,
medians and caps. If some keys are missing in the dict, default colors are used for the corresponding artists.
Also, boxplot has sym keyword to specify fliers style.
When you pass other type of arguments via color keyword, it will be directly passed to matplotlib for all the boxes,
whiskers, medians and caps colorization.
The colors are applied to every boxes to be drawn. If you want more complicated colorization, you can get each drawn
artists by passing return_type.

In [39]: color = {'boxes': 'DarkGreen', 'whiskers': 'DarkOrange',

....: 'medians': 'DarkBlue', 'caps': 'Gray'}
....:

In [40]: df.plot.box(color=color, sym='r+')

Out[40]: <matplotlib.axes._subplots.AxesSubplot at 0x1c4079d710>

4.10. Visualization 555

pandas: powerful Python data analysis toolkit, Release 0.25.3

Also, you can pass other keywords supported by matplotlib boxplot. For example, horizontal and custom-positioned
boxplot can be drawn by vert=False and positions keywords.

In [41]: df.plot.box(vert=False, positions=[1, 4, 5, 6, 8])

Out[41]: <matplotlib.axes._subplots.AxesSubplot at 0x1c4074dbd0>

556 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

See the boxplot method and the matplotlib boxplot documentation for more.
The existing interface DataFrame.boxplot to plot boxplot still can be used.

In [42]: df = pd.DataFrame(np.random.rand(10, 5))

In [43]: plt.figure();

In [44]: bp = df.boxplot()

4.10. Visualization 557

pandas: powerful Python data analysis toolkit, Release 0.25.3

You can create a stratified boxplot using the by keyword argument to create groupings. For instance,

In [45]: df = pd.DataFrame(np.random.rand(10, 2), columns=['Col1', 'Col2'])

In [46]: df['X'] = pd.Series(['A', 'A', 'A', 'A', 'A', 'B', 'B', 'B', 'B', 'B'])

In [47]: plt.figure();

In [48]: bp = df.boxplot(by='X')

558 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

You can also pass a subset of columns to plot, as well as group by multiple columns:

In [49]: df = pd.DataFrame(np.random.rand(10, 3), columns=['Col1', 'Col2', 'Col3'])

In [50]: df['X'] = pd.Series(['A', 'A', 'A', 'A', 'A', 'B', 'B', 'B', 'B', 'B'])

In [51]: df['Y'] = pd.Series(['A', 'B', 'A', 'B', 'A', 'B', 'A', 'B', 'A', 'B'])

In [52]: plt.figure();

In [53]: bp = df.boxplot(column=['Col1', 'Col2'], by=['X', 'Y'])

4.10. Visualization 559

pandas: powerful Python data analysis toolkit, Release 0.25.3

Warning: The default changed from 'dict' to 'axes' in version 0.19.0.

In boxplot, the return type can be controlled by the return_type, keyword. The valid choices are {"axes",
"dict", "both", None}. Faceting, created by DataFrame.boxplot with the by keyword, will affect the
output type as well:

return_type= Faceted Output type

None No axes
None Yes 2-D ndarray of axes
'axes' No axes
'axes' Yes Series of axes
'dict' No dict of artists
'dict' Yes Series of dicts of artists
'both' No namedtuple
'both' Yes Series of namedtuples

Groupby.boxplot always returns a Series of return_type.

560 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [54]: np.random.seed(1234)

In [55]: df_box = pd.DataFrame(np.random.randn(50, 2))

In [56]: df_box['g'] = np.random.choice(['A', 'B'], size=50)

In [57]: df_box.loc[df_box['g'] == 'B', 1] += 3

In [58]: bp = df_box.boxplot(by='g')

The subplots above are split by the numeric columns first, then the value of the g column. Below the subplots are first
split by the value of g, then by the numeric columns.

In [59]: bp = df_box.groupby('g').boxplot()

4.10. Visualization 561

pandas: powerful Python data analysis toolkit, Release 0.25.3

Area plot

You can create area plots with Series.plot.area() and DataFrame.plot.area(). Area plots are stacked
by default. To produce stacked area plot, each column must be either all positive or all negative values.
When input data contains NaN, it will be automatically filled by 0. If you want to drop or fill by different values, use
dataframe.dropna() or dataframe.fillna() before calling plot.

In [60]: df = pd.DataFrame(np.random.rand(10, 4), columns=['a', 'b', 'c', 'd'])

In [61]: df.plot.area();

562 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

To produce an unstacked plot, pass stacked=False. Alpha value is set to 0.5 unless otherwise specified:

In [62]: df.plot.area(stacked=False);

4.10. Visualization 563

pandas: powerful Python data analysis toolkit, Release 0.25.3

Scatter plot

Scatter plot can be drawn by using the DataFrame.plot.scatter() method. Scatter plot requires numeric
columns for the x and y axes. These can be specified by the x and y keywords.

In [63]: df = pd.DataFrame(np.random.rand(50, 4), columns=['a', 'b', 'c', 'd'])

In [64]: df.plot.scatter(x='a', y='b');

564 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

To plot multiple column groups in a single axes, repeat plot method specifying target ax. It is recommended to
specify color and label keywords to distinguish each groups.

In [65]: ax = df.plot.scatter(x='a', y='b', color='DarkBlue', label='Group 1');

In [66]: df.plot.scatter(x='c', y='d', color='DarkGreen', label='Group 2', ax=ax);

4.10. Visualization 565

pandas: powerful Python data analysis toolkit, Release 0.25.3

The keyword c may be given as the name of a column to provide colors for each point:

In [67]: df.plot.scatter(x='a', y='b', c='c', s=50);

566 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

You can pass other keywords supported by matplotlib scatter. The example below shows a bubble chart using a
column of the DataFrame as the bubble size.

In [68]: df.plot.scatter(x='a', y='b', s=df['c'] * 200);

4.10. Visualization 567

pandas: powerful Python data analysis toolkit, Release 0.25.3

See the scatter method and the matplotlib scatter documentation for more.

Hexagonal bin plot

You can create hexagonal bin plots with DataFrame.plot.hexbin(). Hexbin plots can be a useful alternative
to scatter plots if your data are too dense to plot each point individually.

In [69]: df = pd.DataFrame(np.random.randn(1000, 2), columns=['a', 'b'])

In [70]: df['b'] = df['b'] + np.arange(1000)

In [71]: df.plot.hexbin(x='a', y='b', gridsize=25)

Out[71]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3bc43c90>

568 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

A useful keyword argument is gridsize; it controls the number of hexagons in the x-direction, and defaults to 100.
A larger gridsize means more, smaller bins.
By default, a histogram of the counts around each (x, y) point is computed. You can specify alternative aggregations
by passing values to the C and reduce_C_function arguments. C specifies the value at each (x, y) point and
reduce_C_function is a function of one argument that reduces all the values in a bin to a single number (e.g.
mean, max, sum, std). In this example the positions are given by columns a and b, while the value is given by
column z. The bins are aggregated with NumPys max function.

In [72]: df = pd.DataFrame(np.random.randn(1000, 2), columns=['a', 'b'])

In [73]: df['b'] = df['b'] = df['b'] + np.arange(1000)

In [74]: df['z'] = np.random.uniform(0, 3, 1000)

In [75]: df.plot.hexbin(x='a', y='b', C='z', reduce_C_function=np.max, gridsize=25)

Out[75]: <matplotlib.axes._subplots.AxesSubplot at 0x1c39a8cf90>

4.10. Visualization 569

pandas: powerful Python data analysis toolkit, Release 0.25.3

See the hexbin method and the matplotlib hexbin documentation for more.

Pie plot

You can create a pie plot with DataFrame.plot.pie() or Series.plot.pie(). If your data includes any
NaN, they will be automatically filled with 0. A ValueError will be raised if there are any negative values in your
data.

In [76]: series = pd.Series(3 * np.random.rand(4),

....: index=['a', 'b', 'c', 'd'], name='series')
....:

In [77]: series.plot.pie(figsize=(6, 6))

Out[77]: <matplotlib.axes._subplots.AxesSubplot at 0x1257b7a10>

570 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

For pie plots its best to use square figures, i.e. a figure aspect ratio 1. You can create the figure with equal width and
height, or force the aspect ratio to be equal after plotting by calling ax.set_aspect('equal') on the returned
axes object.
Note that pie plot with DataFrame requires that you either specify a target column by the y argument or
subplots=True. When y is specified, pie plot of selected column will be drawn. If subplots=True is spec-
ified, pie plots for each column are drawn as subplots. A legend will be drawn in each pie plots by default; specify
legend=False to hide it.

In [78]: df = pd.DataFrame(3 * np.random.rand(4, 2),

....: index=['a', 'b', 'c', 'd'], columns=['x', 'y'])
....:

In [79]: df.plot.pie(subplots=True, figsize=(8, 4))

Out[79]:
array([<matplotlib.axes._subplots.AxesSubplot object at 0x1c39e89a10>,
(continues on next page)

4.10. Visualization 571

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

<matplotlib.axes._subplots.AxesSubplot object at 0x1c399678d0>],
dtype=object)

You can use the labels and colors keywords to specify the labels and colors of each wedge.

Warning: Most pandas plots use the label and color arguments (note the lack of s on those). To be consistent
with matplotlib.pyplot.pie() you must use labels and colors.

If you want to hide wedge labels, specify labels=None. If fontsize is specified, the value will be applied to
wedge labels. Also, other keywords supported by matplotlib.pyplot.pie() can be used.

In [80]: series.plot.pie(labels=['AA', 'BB', 'CC', 'DD'], colors=['r', 'g', 'b', 'c'],

....: autopct='%.2f', fontsize=20, figsize=(6, 6))
....:
Out[80]: <matplotlib.axes._subplots.AxesSubplot at 0x1a2e007310>

572 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

If you pass values whose sum total is less than 1.0, matplotlib draws a semicircle.

In [81]: series = pd.Series([0.1] * 4, index=['a', 'b', 'c', 'd'], name='series2')

In [82]: series.plot.pie(figsize=(6, 6))

Out[82]: <matplotlib.axes._subplots.AxesSubplot at 0x1c38678390>

4.10. Visualization 573

pandas: powerful Python data analysis toolkit, Release 0.25.3

See the matplotlib pie documentation for more.

4.10.3 Plotting with missing data

Pandas tries to be pragmatic about plotting DataFrames or Series that contain missing data. Missing values are
dropped, left out, or filled depending on the plot type.

574 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Plot Type NaN Handling

Line Leave gaps at NaNs
Line (stacked) Fill 0s
Bar Fill 0s
Scatter Drop NaNs
Histogram Drop NaNs (column-wise)
Box Drop NaNs (column-wise)
Area Fill 0s
KDE Drop NaNs (column-wise)
Hexbin Drop NaNs
Pie Fill 0s

If any of these defaults are not what you want, or if you want to be explicit about how missing values are handled,
consider using fillna() or dropna() before plotting.

4.10.4 Plotting Tools

These functions can be imported from pandas.plotting and take a Series or DataFrame as an argument.

Scatter matrix plot

You can create a scatter plot matrix using the scatter_matrix method in pandas.plotting:

In [83]: from pandas.plotting import scatter_matrix

In [84]: df = pd.DataFrame(np.random.randn(1000, 4), columns=['a', 'b', 'c', 'd'])

In [85]: scatter_matrix(df, alpha=0.2, figsize=(6, 6), diagonal='kde')

Out[85]:
array([[<matplotlib.axes._subplots.AxesSubplot object at 0x1258eb5d0>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1258fbc50>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c38bf6fd0>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c3e94ecd0>],
[<matplotlib.axes._subplots.AxesSubplot object at 0x1c3e97f510>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c38ea6d10>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c3e7d1550>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c3ee62d50>],
[<matplotlib.axes._subplots.AxesSubplot object at 0x1c3ee898d0>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c3dea1290>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c39ecf5d0>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c38e3bdd0>],
[<matplotlib.axes._subplots.AxesSubplot object at 0x1c3a497610>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c38683e10>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c38d1b650>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1c3e14ee50>]],
dtype=object)

4.10. Visualization 575

pandas: powerful Python data analysis toolkit, Release 0.25.3

Density plot

You can create density plots using the Series.plot.kde() and DataFrame.plot.kde() methods.

In [86]: ser = pd.Series(np.random.randn(1000))

In [87]: ser.plot.kde()
Out[87]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3beb8450>

576 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Andrews curves

Andrews curves allow one to plot multivariate data as a large number of curves that are created using the attributes
of samples as coefficients for Fourier series, see the Wikipedia entry for more information. By coloring these curves
differently for each class it is possible to visualize data clustering. Curves belonging to samples of the same class will
usually be closer together and form larger structures.
Note: The Iris dataset is available here.
In [88]: from pandas.plotting import andrews_curves

In [89]: data = pd.read_csv('data/iris.data')

In [90]: plt.figure()
Out[90]: <Figure size 640x480 with 0 Axes>

In [91]: andrews_curves(data, 'Name')

Out[91]: <matplotlib.axes._subplots.AxesSubplot at 0x1c39b21c10>

4.10. Visualization 577

pandas: powerful Python data analysis toolkit, Release 0.25.3

Parallel coordinates

Parallel coordinates is a plotting technique for plotting multivariate data, see the Wikipedia entry for an introduction.
Parallel coordinates allows one to see clusters in data and to estimate other statistics visually. Using parallel coordinates
points are represented as connected line segments. Each vertical line represents one attribute. One set of connected
line segments represents one data point. Points that tend to cluster will appear closer together.
In [92]: from pandas.plotting import parallel_coordinates

In [93]: data = pd.read_csv('data/iris.data')

In [94]: plt.figure()
Out[94]: <Figure size 640x480 with 0 Axes>

In [95]: parallel_coordinates(data, 'Name')

Out[95]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3bd54bd0>

578 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Lag plot

Lag plots are used to check if a data set or time series is random. Random data should not exhibit any structure in the
lag plot. Non-random structure implies that the underlying data are not random. The lag argument may be passed,
and when lag=1 the plot is essentially data[:-1] vs. data[1:].

In [96]: from pandas.plotting import lag_plot

In [97]: plt.figure()
Out[97]: <Figure size 640x480 with 0 Axes>

In [98]: spacing = np.linspace(-99 * np.pi, 99 * np.pi, num=1000)

In [99]: data = pd.Series(0.1 * np.random.rand(1000) + 0.9 * np.sin(spacing))

In [100]: lag_plot(data)
Out[100]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3efd8350>

4.10. Visualization 579

pandas: powerful Python data analysis toolkit, Release 0.25.3

Autocorrelation plot

Autocorrelation plots are often used for checking randomness in time series. This is done by computing autocorrela-
tions for data values at varying time lags. If time series is random, such autocorrelations should be near zero for any
and all time-lag separations. If time series is non-random then one or more of the autocorrelations will be significantly
non-zero. The horizontal lines displayed in the plot correspond to 95% and 99% confidence bands. The dashed line is
99% confidence band. See the Wikipedia entry for more about autocorrelation plots.

In [101]: from pandas.plotting import autocorrelation_plot

In [102]: plt.figure()
Out[102]: <Figure size 640x480 with 0 Axes>

In [103]: spacing = np.linspace(-9 * np.pi, 9 * np.pi, num=1000)

In [104]: data = pd.Series(0.7 * np.random.rand(1000) + 0.3 * np.sin(spacing))

In [105]: autocorrelation_plot(data)
Out[105]: <matplotlib.axes._subplots.AxesSubplot at 0x1c39af5c10>

580 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Bootstrap plot

Bootstrap plots are used to visually assess the uncertainty of a statistic, such as mean, median, midrange, etc. A
random subset of a specified size is selected from a data set, the statistic in question is computed for this subset and
the process is repeated a specified number of times. Resulting plots and histograms are what constitutes the bootstrap
plot.

In [106]: from pandas.plotting import bootstrap_plot

In [107]: data = pd.Series(np.random.rand(1000))

In [108]: bootstrap_plot(data, size=50, samples=500, color='grey')

Out[108]: <Figure size 640x480 with 6 Axes>

4.10. Visualization 581

pandas: powerful Python data analysis toolkit, Release 0.25.3

RadViz

RadViz is a way of visualizing multi-variate data. It is based on a simple spring tension minimization algorithm.
Basically you set up a bunch of points in a plane. In our case they are equally spaced on a unit circle. Each point
represents a single attribute. You then pretend that each sample in the data set is attached to each of these points
by a spring, the stiffness of which is proportional to the numerical value of that attribute (they are normalized to
unit interval). The point in the plane, where our sample settles to (where the forces acting on our sample are at an
equilibrium) is where a dot representing our sample will be drawn. Depending on which class that sample belongs it
will be colored differently. See the R package Radviz for more information.
Note: The Iris dataset is available here.
In [109]: from pandas.plotting import radviz

In [110]: data = pd.read_csv('data/iris.data')

In [111]: plt.figure()
Out[111]: <Figure size 640x480 with 0 Axes>

In [112]: radviz(data, 'Name')

Out[112]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3e9f0110>

582 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

4.10.5 Plot Formatting

Setting the plot style

From version 1.5 and up, matplotlib offers a range of pre-configured plotting styles. Setting the style can be
used to easily give plots the general look that you want. Setting the style is as easy as calling matplotlib.
style.use(my_plot_style) before creating your plot. For example you could write matplotlib.style.
use('ggplot') for ggplot-style plots.
You can see the various available style names at matplotlib.style.available and its very easy to try them
out.

General plot style arguments

Most plotting methods have a set of keyword arguments that control the layout and formatting of the returned plot:

In [113]: plt.figure();

In [114]: ts.plot(style='k--', label='Series');

4.10. Visualization 583

pandas: powerful Python data analysis toolkit, Release 0.25.3

For each kind of plot (e.g. line, bar, scatter) any additional arguments keywords are passed along to the corresponding
matplotlib function (ax.plot(), ax.bar(), ax.scatter()). These can be used to control additional styling,
beyond what pandas provides.

Controlling the legend

You may set the legend argument to False to hide the legend, which is shown by default.

In [115]: df = pd.DataFrame(np.random.randn(1000, 4),

.....: index=ts.index, columns=list('ABCD'))
.....:

In [116]: df = df.cumsum()

In [117]: df.plot(legend=False)
Out[117]: <matplotlib.axes._subplots.AxesSubplot at 0x1c386d8ed0>

584 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Scales

You may pass logy to get a log-scale Y axis.

In [118]: ts = pd.Series(np.random.randn(1000),
.....: index=pd.date_range('1/1/2000', periods=1000))
.....:

In [119]: ts = np.exp(ts.cumsum())

In [120]: ts.plot(logy=True)
Out[120]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3e544ad0>

4.10. Visualization 585

pandas: powerful Python data analysis toolkit, Release 0.25.3

See also the logx and loglog keyword arguments.

Plotting on a secondary y-axis

To plot data on a secondary y-axis, use the secondary_y keyword:

In [121]: df.A.plot()
Out[121]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3ec4f890>

In [122]: df.B.plot(secondary_y=True, style='g')

Out[122]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3ec66150>

586 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

To plot some columns in a DataFrame, give the column names to the secondary_y keyword:
In [123]: plt.figure()
Out[123]: <Figure size 640x480 with 0 Axes>

In [124]: ax = df.plot(secondary_y=['A', 'B'])

In [125]: ax.set_ylabel('CD scale')

Out[125]: Text(0, 0.5, 'CD scale')

In [126]: ax.right_ax.set_ylabel('AB scale')

Out[126]: Text(0, 0.5, 'AB scale')

4.10. Visualization 587

pandas: powerful Python data analysis toolkit, Release 0.25.3

Note that the columns plotted on the secondary y-axis is automatically marked with (right) in the legend. To turn off
the automatic marking, use the mark_right=False keyword:
In [127]: plt.figure()
Out[127]: <Figure size 640x480 with 0 Axes>

In [128]: df.plot(secondary_y=['A', 'B'], mark_right=False)

Out[128]: <matplotlib.axes._subplots.AxesSubplot at 0x1c3e18f850>

588 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Suppressing tick resolution adjustment

pandas includes automatic tick resolution adjustment for regular frequency time-series data. For limited cases where
pandas cannot infer the frequency information (e.g., in an externally created twinx), you can choose to suppress this
behavior for alignment purposes.
Here is the default behavior, notice how the x-axis tick labeling is performed:
In [129]: plt.figure()
Out[129]: <Figure size 640x480 with 0 Axes>

In [130]: df.A.plot()
Out[130]: <matplotlib.axes._subplots.AxesSubplot at 0x1c39659450>

4.10. Visualization 589

pandas: powerful Python data analysis toolkit, Release 0.25.3

Using the x_compat parameter, you can suppress this behavior:

In [131]: plt.figure()
Out[131]: <Figure size 640x480 with 0 Axes>

In [132]: df.A.plot(x_compat=True)
Out[132]: <matplotlib.axes._subplots.AxesSubplot at 0x1a30f11d90>

590 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

If you have more than one plot that needs to be suppressed, the use method in pandas.plotting.
plot_params can be used in a with statement:

In [133]: plt.figure()
Out[133]: <Figure size 640x480 with 0 Axes>

In [134]: with pd.plotting.plot_params.use('x_compat', True):

.....: df.A.plot(color='r')
.....: df.B.plot(color='g')
.....: df.C.plot(color='b')
.....:

4.10. Visualization 591

pandas: powerful Python data analysis toolkit, Release 0.25.3

Automatic date tick adjustment

New in version 0.20.0.

TimedeltaIndex now uses the native matplotlib tick locator methods, it is useful to call the automatic date tick
adjustment from matplotlib for figures whose ticklabels overlap.
See the autofmt_xdate method and the matplotlib documentation for more.

Subplots

Each Series in a DataFrame can be plotted on a different axis with the subplots keyword:

In [135]: df.plot(subplots=True, figsize=(6, 6));

592 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Using layout and targeting multiple axes

The layout of subplots can be specified by the layout keyword. It can accept (rows, columns). The layout
keyword can be used in hist and boxplot also. If the input is invalid, a ValueError will be raised.
The number of axes which can be contained by rows x columns specified by layout must be larger than the number
of required subplots. If layout can contain more axes than required, blank axes are not drawn. Similar to a NumPy
arrays reshape method, you can use -1 for one dimension to automatically calculate the number of rows or columns
needed, given the other.

In [136]: df.plot(subplots=True, layout=(2, 3), figsize=(6, 6), sharex=False);

4.10. Visualization 593

pandas: powerful Python data analysis toolkit, Release 0.25.3

The above example is identical to using:

In [137]: df.plot(subplots=True, layout=(2, -1), figsize=(6, 6), sharex=False);

The required number of columns (3) is inferred from the number of series to plot and the given number of rows (2).
You can pass multiple axes created beforehand as list-like via ax keyword. This allows more complicated layouts.
The passed axes must be the same number as the subplots being drawn.
When multiple axes are passed via the ax keyword, layout, sharex and sharey keywords dont affect to the
output. You should explicitly pass sharex=False and sharey=False, otherwise you will see a warning.
In [138]: fig, axes = plt.subplots(4, 4, figsize=(6, 6))

In [139]: plt.subplots_adjust(wspace=0.5, hspace=0.5)

In [140]: target1 = [axes[0][0], axes[1][1], axes[2][2], axes[3][3]]

(continues on next page)

594 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [141]: target2 = [axes[3][0], axes[2][1], axes[1][2], axes[0][3]]

In [142]: df.plot(subplots=True, ax=target1, legend=False, sharex=False,

,→sharey=False);

In [143]: (-df).plot(subplots=True, ax=target2, legend=False,

.....: sharex=False, sharey=False);
.....:

Another option is passing an ax argument to Series.plot() to plot on a particular axis:

In [144]: fig, axes = plt.subplots(nrows=2, ncols=2)

In [145]: df['A'].plot(ax=axes[0, 0]);

(continues on next page)

4.10. Visualization 595

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [146]: axes[0, 0].set_title('A');

In [147]: df['B'].plot(ax=axes[0, 1]);

In [148]: axes[0, 1].set_title('B');

In [149]: df['C'].plot(ax=axes[1, 0]);

In [150]: axes[1, 0].set_title('C');

In [151]: df['D'].plot(ax=axes[1, 1]);

In [152]: axes[1, 1].set_title('D');

Plotting with error bars

Plotting with error bars is supported in DataFrame.plot() and Series.plot().

Horizontal and vertical error bars can be supplied to the xerr and yerr keyword arguments to plot(). The error
values can be specified using a variety of formats:
• As a DataFrame or dict of errors with column names matching the columns attribute of the plotting

596 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

DataFrame or matching the name attribute of the Series.

• As a str indicating which of the columns of plotting DataFrame contain the error values.
• As raw values (list, tuple, or np.ndarray). Must be the same length as the plotting
DataFrame/Series.
Asymmetrical error bars are also supported, however raw error values must be provided in this case. For a M
length Series, a Mx2 array should be provided indicating lower and upper (or left and right) errors. For a MxN
DataFrame, asymmetrical errors should be in a Mx2xN array.
Here is an example of one way to easily plot group means with standard deviations from the raw data.
# Generate the data
In [153]: ix3 = pd.MultiIndex.from_arrays([
.....: ['a', 'a', 'a', 'a', 'b', 'b', 'b', 'b'],
.....: ['foo', 'foo', 'bar', 'bar', 'foo', 'foo', 'bar', 'bar']],
.....: names=['letter', 'word'])
.....:

In [154]: df3 = pd.DataFrame({'data1': [3, 2, 4, 3, 2, 4, 3, 2],

.....: 'data2': [6, 5, 7, 5, 4, 5, 6, 5]}, index=ix3)
.....:

# Group by index labels and take the means and standard deviations
# for each group
In [155]: gp3 = df3.groupby(level=('letter', 'word'))

In [156]: means = gp3.mean()

In [157]: errors = gp3.std()

In [158]: means
Out[158]:
data1 data2
letter word
a bar 3.5 6.0
foo 2.5 5.5
b bar 2.5 5.5
foo 3.0 4.5

In [159]: errors
Out[159]:
data1 data2
letter word
a bar 0.707107 1.414214
foo 0.707107 0.707107
b bar 0.707107 0.707107
foo 1.414214 0.707107

# Plot
In [160]: fig, ax = plt.subplots()

In [161]: means.plot.bar(yerr=errors, ax=ax, capsize=4)

Out[161]: <matplotlib.axes._subplots.AxesSubplot at 0x1c4327bc50>

4.10. Visualization 597

pandas: powerful Python data analysis toolkit, Release 0.25.3

Plotting tables

Plotting with matplotlib table is now supported in DataFrame.plot() and Series.plot() with a table
keyword. The table keyword can accept bool, DataFrame or Series. The simple way to draw a table is to
specify table=True. Data will be transposed to meet matplotlibs default layout.

In [162]: fig, ax = plt.subplots(1, 1)

In [163]: df = pd.DataFrame(np.random.rand(5, 3), columns=['a', 'b', 'c'])

In [164]: ax.get_xaxis().set_visible(False) # Hide Ticks

In [165]: df.plot(table=True, ax=ax)

Out[165]: <matplotlib.axes._subplots.AxesSubplot at 0x1c43469350>

598 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Also, you can pass a different DataFrame or Series to the table keyword. The data will be drawn as displayed
in print method (not transposed automatically). If required, it should be transposed manually as seen in the example
below.

In [166]: fig, ax = plt.subplots(1, 1)

In [167]: ax.get_xaxis().set_visible(False) # Hide Ticks

In [168]: df.plot(table=np.round(df.T, 2), ax=ax)

Out[168]: <matplotlib.axes._subplots.AxesSubplot at 0x1c4363e6d0>

4.10. Visualization 599

pandas: powerful Python data analysis toolkit, Release 0.25.3

There also exists a helper function pandas.plotting.table, which creates a table from DataFrame or
Series, and adds it to an matplotlib.Axes instance. This function can accept keywords which the matplotlib
table has.
In [169]: from pandas.plotting import table

In [170]: fig, ax = plt.subplots(1, 1)

In [171]: table(ax, np.round(df.describe(), 2),

.....: loc='upper right', colWidths=[0.2, 0.2, 0.2])
.....:
Out[171]: <matplotlib.table.Table at 0x1c43858fd0>

In [172]: df.plot(ax=ax, ylim=(0, 2), legend=None)

Out[172]: <matplotlib.axes._subplots.AxesSubplot at 0x1c4381c090>

600 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Note: You can get table instances on the axes using axes.tables property for further decorations. See the mat-
plotlib table documentation for more.

Colormaps

A potential issue when plotting a large number of columns is that it can be difficult to distinguish some series due to
repetition in the default colors. To remedy this, DataFrame plotting supports the use of the colormap argument,
which accepts either a Matplotlib colormap or a string that is a name of a colormap registered with Matplotlib. A
visualization of the default matplotlib colormaps is available here.
As matplotlib does not directly support colormaps for line-based plots, the colors are selected based on an even spacing
determined by the number of columns in the DataFrame. There is no consideration made for background color, so
some colormaps will produce lines that are not easily visible.
To use the cubehelix colormap, we can pass colormap='cubehelix'.
In [173]: df = pd.DataFrame(np.random.randn(1000, 10), index=ts.index)

In [174]: df = df.cumsum()

In [175]: plt.figure()
Out[175]: <Figure size 640x480 with 0 Axes>

In [176]: df.plot(colormap='cubehelix')

4.10. Visualization 601

pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[176]: <matplotlib.axes._subplots.AxesSubplot at 0x1c43a29e50>

Alternatively, we can pass the colormap itself:

In [177]: from matplotlib import cm

In [178]: plt.figure()
Out[178]: <Figure size 640x480 with 0 Axes>

In [179]: df.plot(colormap=cm.cubehelix)
Out[179]: <matplotlib.axes._subplots.AxesSubplot at 0x1c43d3d210>

602 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Colormaps can also be used other plot types, like bar charts:
In [180]: dd = pd.DataFrame(np.random.randn(10, 10)).applymap(abs)

In [181]: dd = dd.cumsum()

In [182]: plt.figure()
Out[182]: <Figure size 640x480 with 0 Axes>

In [183]: dd.plot.bar(colormap='Greens')
Out[183]: <matplotlib.axes._subplots.AxesSubplot at 0x1c44056f10>

4.10. Visualization 603

pandas: powerful Python data analysis toolkit, Release 0.25.3

Parallel coordinates charts:

In [184]: plt.figure()
Out[184]: <Figure size 640x480 with 0 Axes>

In [185]: parallel_coordinates(data, 'Name', colormap='gist_rainbow')

Out[185]: <matplotlib.axes._subplots.AxesSubplot at 0x1c443518d0>

604 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Andrews curves charts:

In [186]: plt.figure()
Out[186]: <Figure size 640x480 with 0 Axes>

In [187]: andrews_curves(data, 'Name', colormap='winter')

Out[187]: <matplotlib.axes._subplots.AxesSubplot at 0x1c441ff650>

4.10. Visualization 605

pandas: powerful Python data analysis toolkit, Release 0.25.3

4.10.6 Plotting directly with matplotlib

In some situations it may still be preferable or necessary to prepare plots directly with matplotlib, for instance when a
certain type of plot or customization is not (yet) supported by pandas. Series and DataFrame objects behave like
arrays and can therefore be passed directly to matplotlib functions without explicit casts.
pandas also automatically registers formatters and locators that recognize date indices, thereby extending date and
time support to practically all plot types available in matplotlib. Although this formatting does not provide the same
level of refinement you would get when plotting via pandas, it can be faster when plotting a large number of points.
In [188]: price = pd.Series(np.random.randn(150).cumsum(),
.....: index=pd.date_range('2000-1-1', periods=150,
,→freq='B'))

.....:

In [189]: ma = price.rolling(20).mean()

In [190]: mstd = price.rolling(20).std()

In [191]: plt.figure()
Out[191]: <Figure size 640x480 with 0 Axes>

In [192]: plt.plot(price.index, price, 'k')

606 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[192]: [<matplotlib.lines.Line2D at 0x1c44aaa350>]

In [193]: plt.plot(ma.index, ma, 'b')

Out[193]: [<matplotlib.lines.Line2D at 0x1c44a74ed0>]

In [194]: plt.fill_between(mstd.index, ma - 2 * mstd, ma + 2 * mstd,

.....: color='b', alpha=0.2)
.....:
Out[194]: <matplotlib.collections.PolyCollection at 0x1c44adea10>

4.10.7 Trellis plotting interface

Warning: The rplot trellis plotting interface has been removed. Please use external packages like seaborn for
similar but more refined functionality and refer to our 0.18.1 documentation here for how to convert to using it.

{{ header }}

4.10. Visualization 607

pandas: powerful Python data analysis toolkit, Release 0.25.3

4.11 Computational tools

4.11.1 Statistical functions

Percent change

Series and DataFrame have a method pct_change() to compute the percent change over a given number of
periods (using fill_method to fill NA/null values before computing the percent change).

In [1]: ser = pd.Series(np.random.randn(8))

In [2]: ser.pct_change()
Out[2]:
0 NaN
1 13.547371
2 -1.548836
3 -2.936641
4 -2.449972
5 -1.103130
6 16.241433
7 -1.573605
dtype: float64

In [3]: df = pd.DataFrame(np.random.randn(10, 4))

In [4]: df.pct_change(periods=3)
Out[4]:
0 1 2 3
0 NaN NaN NaN NaN
1 NaN NaN NaN NaN
2 NaN NaN NaN NaN
3 -6.381415 -7.189288 -0.796567 5.324072
4 -1.594458 -0.193537 -1.514930 -0.943232
5 -2.689858 -2.369493 0.155967 -1.569737
6 -1.372184 0.635387 -2.303971 -0.509280
7 -6.194810 -2.482964 -0.764169 -25.332016
8 -0.923376 -2.560822 -2.479379 -2.656377
9 -2.978688 -0.504129 -5.959571 -5.033217

Covariance

Series.cov() can be used to compute covariance between series (excluding missing values).

In [5]: s1 = pd.Series(np.random.randn(1000))

In [6]: s2 = pd.Series(np.random.randn(1000))

In [7]: s1.cov(s2)
Out[7]: -0.035818144141082434

Analogously, DataFrame.cov() to compute pairwise covariances among the series in the DataFrame, also exclud-
ing NA/null values.

Note: Assuming the missing data are missing at random this results in an estimate for the covariance matrix which

608 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

is unbiased. However, for many applications this estimate may not be acceptable because the estimated covariance
matrix is not guaranteed to be positive semi-definite. This could lead to estimated correlations having absolute values
which are greater than one, and/or a non-invertible covariance matrix. See Estimation of covariance matrices for more
details.

In [8]: frame = pd.DataFrame(np.random.randn(1000, 5),

...: columns=['a', 'b', 'c', 'd', 'e'])
...:

In [9]: frame.cov()
Out[9]:
a b c d e
a 0.969905 -0.031318 -0.021847 -0.002147 -0.045109
b -0.031318 0.948570 0.003914 -0.022931 -0.013969
c -0.021847 0.003914 1.039872 0.008068 -0.017560
d -0.002147 -0.022931 0.008068 1.059395 -0.045589
e -0.045109 -0.013969 -0.017560 -0.045589 1.105350

DataFrame.cov also supports an optional min_periods keyword that specifies the required minimum number
of observations for each column pair in order to have a valid result.
In [10]: frame = pd.DataFrame(np.random.randn(20, 3), columns=['a', 'b', 'c'])

In [11]: frame.loc[frame.index[:5], 'a'] = np.nan

In [12]: frame.loc[frame.index[5:10], 'b'] = np.nan

In [13]: frame.cov()
Out[13]:
a b c
a 1.000881 0.151591 -0.143082
b 0.151591 0.869675 -0.035753
c -0.143082 -0.035753 0.737483

In [14]: frame.cov(min_periods=12)
Out[14]:
a b c
a 1.000881 NaN -0.143082
b NaN 0.869675 -0.035753
c -0.143082 -0.035753 0.737483

Correlation

Correlation may be computed using the corr() method. Using the method parameter, several methods for com-
puting correlations are provided:

Method name Description

pearson Standard correlation coefficient
(default)
kendall Kendall Tau correlation coefficient
spearman Spearman rank correlation coefficient

All of these are currently computed using pairwise complete observations. Wikipedia has articles covering the above
correlation coefficients:

4.11. Computational tools 609

pandas: powerful Python data analysis toolkit, Release 0.25.3

• Pearson correlation coefficient

• Kendall rank correlation coefficient
• Spearmans rank correlation coefficient

Note: Please see the caveats associated with this method of calculating correlation matrices in the covariance section.

In [15]: frame = pd.DataFrame(np.random.randn(1000, 5),

....: columns=['a', 'b', 'c', 'd', 'e'])
....:

In [16]: frame.iloc[::2] = np.nan

# Series with Series

In [17]: frame['a'].corr(frame['b'])
Out[17]: 0.016886089025056897

In [18]: frame['a'].corr(frame['b'], method='spearman')

Out[18]: -0.02348697394789579

# Pairwise correlation of DataFrame columns

In [19]: frame.corr()
Out[19]:
a b c d e
a 1.000000 0.016886 -0.028542 0.034697 -0.008209
b 0.016886 1.000000 0.000628 -0.004688 -0.028515
c -0.028542 0.000628 1.000000 -0.025158 0.104966
d 0.034697 -0.004688 -0.025158 1.000000 -0.014990
e -0.008209 -0.028515 0.104966 -0.014990 1.000000
Note that non-numeric columns will be automatically excluded from the correlation calculation.
Like cov, corr also supports the optional min_periods keyword:
In [20]: frame = pd.DataFrame(np.random.randn(20, 3), columns=['a', 'b', 'c'])

In [21]: frame.loc[frame.index[:5], 'a'] = np.nan

In [22]: frame.loc[frame.index[5:10], 'b'] = np.nan

In [23]: frame.corr()
Out[23]:
a b c
a 1.000000 0.227635 -0.080537
b 0.227635 1.000000 0.034068
c -0.080537 0.034068 1.000000

In [24]: frame.corr(min_periods=12)
Out[24]:
a b c
a 1.000000 NaN -0.080537
b NaN 1.000000 0.034068
c -0.080537 0.034068 1.000000
New in version 0.24.0.

610 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

The method argument can also be a callable for a generic correlation calculation. In this case, it should be a single
function that produces a single value from two ndarray inputs. Suppose we wanted to compute the correlation based
on histogram intersection:
# histogram intersection
In [25]: def histogram_intersection(a, b):
....: return np.minimum(np.true_divide(a, a.sum()),
....: np.true_divide(b, b.sum())).sum()
....:

In [26]: frame.corr(method=histogram_intersection)
Out[26]:
a b c
a 1.000000 -7.283861 -31.104637
b -7.283861 1.000000 -7.810865
c -31.104637 -7.810865 1.000000

A related method corrwith() is implemented on DataFrame to compute the correlation between like-labeled Series
contained in different DataFrame objects.
In [27]: index = ['a', 'b', 'c', 'd', 'e']

In [28]: columns = ['one', 'two', 'three', 'four']

In [29]: df1 = pd.DataFrame(np.random.randn(5, 4), index=index,

,→columns=columns)

In [30]: df2 = pd.DataFrame(np.random.randn(4, 4), index=index[:4],

,→columns=columns)

In [31]: df1.corrwith(df2)
Out[31]:
one 0.526632
two 0.611188
three 0.434034
four 0.000843
dtype: float64

In [32]: df2.corrwith(df1, axis=1)

Out[32]:
a 0.791016
b 0.285772
c 0.798940
d -0.946595
e NaN
dtype: float64

Data ranking

The rank() method produces a data ranking with ties being assigned the mean of the ranks (by default) for the group:
In [33]: s = pd.Series(np.random.randn(5), index=list('abcde'))

In [34]: s['d'] = s['b'] # so there's a tie

(continues on next page)

4.11. Computational tools 611

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [35]: s.rank()
Out[35]:
a 4.0
b 1.5
c 5.0
d 1.5
e 3.0
dtype: float64

rank() is also a DataFrame method and can rank either the rows (axis=0) or the columns (axis=1). NaN values
are excluded from the ranking.
In [36]: df = pd.DataFrame(np.random.randn(10, 6))

In [37]: df[4] = df[2][:5] # some ties

In [38]: df
Out[38]:
0 1 2 3 4 5
0 0.944050 -0.657738 -0.125993 0.286736 -0.125993 -0.864790
1 0.055086 -1.056551 0.458775 1.329940 0.458775 -0.697985
2 -0.567600 0.693650 0.962299 -1.216697 0.962299 1.092375
3 0.343301 0.407707 -0.698956 -1.110074 -0.698956 0.126931
4 1.084096 0.971689 -1.368406 2.444578 -1.368406 0.143508
5 0.349580 -1.681880 -0.403970 -1.023733 NaN 0.166872
6 -0.120038 0.797128 0.389633 -1.661451 NaN -0.673658
7 -1.121792 -1.485899 -0.804363 1.317388 NaN -1.382258
8 0.470715 -0.179893 -0.299046 0.153052 NaN -0.265199
9 -0.981253 -0.055122 1.602212 1.304666 NaN 0.443094

In [39]: df.rank(1)
Out[39]:
0 1 2 3 4 5
0 6.0 2.0 3.5 5.0 3.5 1.0
1 3.0 1.0 4.5 6.0 4.5 2.0
2 2.0 3.0 4.5 1.0 4.5 6.0
3 5.0 6.0 2.5 1.0 2.5 4.0
4 5.0 4.0 1.5 6.0 1.5 3.0
5 5.0 1.0 3.0 2.0 NaN 4.0
6 3.0 5.0 4.0 1.0 NaN 2.0
7 3.0 1.0 4.0 5.0 NaN 2.0
8 5.0 3.0 1.0 4.0 NaN 2.0
9 1.0 2.0 5.0 4.0 NaN 3.0
rank optionally takes a parameter ascending which by default is true; when false, data is reverse-ranked, with
larger values assigned a smaller rank.
rank supports different tie-breaking methods, specified with the method parameter:
• average : average rank of tied group
• min : lowest rank in the group
• max : highest rank in the group
• first : ranks assigned in the order they appear in the array

612 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

4.11.2 Window Functions

For working with data, a number of window functions are provided for computing common window or rolling statistics.
Among these are count, sum, mean, median, correlation, variance, covariance, standard deviation, skewness, and
kurtosis.
The rolling() and expanding() functions can be used directly from DataFrameGroupBy objects, see the
groupby docs.

Note: The API for window statistics is quite similar to the way one works with GroupBy objects, see the documen-
tation here.

We work with rolling, expanding and exponentially weighted data through the corresponding objects,
Rolling, Expanding and EWM.

In [40]: s = pd.Series(np.random.randn(1000),
....: index=pd.date_range('1/1/2000', periods=1000))
....:

In [41]: s = s.cumsum()

In [42]: s
Out[42]:
2000-01-01 -0.176917
2000-01-02 0.316541
2000-01-03 -0.787075
2000-01-04 -2.107309
2000-01-05 -3.235250
...
2002-09-22 -22.662237
2002-09-23 -24.731927
2002-09-24 -25.898290
2002-09-25 -25.522878
2002-09-26 -25.091748
Freq: D, Length: 1000, dtype: float64

These are created from methods on Series and DataFrame.

In [43]: r = s.rolling(window=60)

In [44]: r
Out[44]: Rolling [window=60,center=False,axis=0]

These object provide tab-completion of the available methods and properties.

In [14]: r.<TAB> # noqa: E225, E999

r.agg r.apply r.count r.exclusions r.max r.median r.
,→name r.skew r.sum
r.aggregate r.corr r.cov r.kurt r.mean r.min r.
,→quantile r.std r.var

Generally these methods all have the same interface. They all accept the following arguments:
• window: size of moving window
• min_periods: threshold of non-null data points to require (otherwise result is NA)
• center: boolean, whether to set the labels at the center (default is False)

4.11. Computational tools 613

pandas: powerful Python data analysis toolkit, Release 0.25.3

We can then call methods on these rolling objects. These return like-indexed objects:

In [45]: r.mean()
Out[45]:
2000-01-01 NaN
2000-01-02 NaN
2000-01-03 NaN
2000-01-04 NaN
2000-01-05 NaN
...
2002-09-22 -18.067286
2002-09-23 -18.203986
2002-09-24 -18.378035
2002-09-25 -18.539518
2002-09-26 -18.676091
Freq: D, Length: 1000, dtype: float64

In [46]: s.plot(style='k--')
Out[46]: <matplotlib.axes._subplots.AxesSubplot at 0x129c78b50>

In [47]: r.mean().plot(style='k')
Out[47]: <matplotlib.axes._subplots.AxesSubplot at 0x129c78b50>

They can also be applied to DataFrame objects. This is really just syntactic sugar for applying the moving window
operator to all of the DataFrames columns:

614 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [48]: df = pd.DataFrame(np.random.randn(1000, 4),

....: index=pd.date_range('1/1/2000', periods=1000),
....: columns=['A', 'B', 'C', 'D'])
....:

In [49]: df = df.cumsum()

In [50]: df.rolling(window=60).sum().plot(subplots=True)
Out[50]:
array([<matplotlib.axes._subplots.AxesSubplot object at 0x125586b10>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1a30878f90>,
<matplotlib.axes._subplots.AxesSubplot object at 0x125552f10>,
<matplotlib.axes._subplots.AxesSubplot object at 0x1a30415a90>],
dtype=object)

Method summary

We provide a number of common statistical functions:

4.11. Computational tools 615

pandas: powerful Python data analysis toolkit, Release 0.25.3

Method Description
count() Number of non-null observations
sum() Sum of values
mean() Mean of values
median() Arithmetic median of values
min() Minimum
max() Maximum
std() Bessel-corrected sample standard deviation
var() Unbiased variance
skew() Sample skewness (3rd moment)
kurt() Sample kurtosis (4th moment)
quantile() Sample quantile (value at %)
apply() Generic apply
cov() Unbiased covariance (binary)
corr() Correlation (binary)

The apply() function takes an extra func argument and performs generic rolling computations. The func argu-
ment should be a single function that produces a single value from an ndarray input. Suppose we wanted to compute
the mean absolute deviation on a rolling basis:

In [51]: def mad(x):

....: return np.fabs(x - x.mean()).mean()
....:

In [52]: s.rolling(window=60).apply(mad, raw=True).plot(style='k')

Out[52]: <matplotlib.axes._subplots.AxesSubplot at 0x1253717d0>

616 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Rolling windows

Passing win_type to .rolling generates a generic rolling window computation, that is weighted according the
win_type. The following methods are available:

Method Description
sum() Sum of values
mean() Mean of values

The weights used in the window are specified by the win_type keyword. The list of recognized types are the
scipy.signal window functions:
• boxcar
• triang
• blackman
• hamming
• bartlett
• parzen
• bohman

4.11. Computational tools 617

pandas: powerful Python data analysis toolkit, Release 0.25.3

• blackmanharris
• nuttall
• barthann
• kaiser (needs beta)
• gaussian (needs std)
• general_gaussian (needs power, width)
• slepian (needs width)
• exponential (needs tau).
In [53]: ser = pd.Series(np.random.randn(10),
....: index=pd.date_range('1/1/2000', periods=10))
....:

In [54]: ser.rolling(window=5, win_type='triang').mean()

Out[54]:
2000-01-01 NaN
2000-01-02 NaN
2000-01-03 NaN
2000-01-04 NaN
2000-01-05 0.380626
2000-01-06 0.573337
2000-01-07 0.378764
2000-01-08 0.004447
2000-01-09 -0.331045
2000-01-10 -0.178263
Freq: D, dtype: float64

Note that the boxcar window is equivalent to mean().

In [55]: ser.rolling(window=5, win_type='boxcar').mean()
Out[55]:
2000-01-01 NaN
2000-01-02 NaN
2000-01-03 NaN
2000-01-04 NaN
2000-01-05 0.540420
2000-01-06 0.306313
2000-01-07 0.081345
2000-01-08 0.254765
2000-01-09 -0.095579
2000-01-10 -0.237109
Freq: D, dtype: float64

In [56]: ser.rolling(window=5).mean()
Out[56]:
2000-01-01 NaN
2000-01-02 NaN
2000-01-03 NaN
2000-01-04 NaN
2000-01-05 0.540420
2000-01-06 0.306313
2000-01-07 0.081345
2000-01-08 0.254765

618 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

2000-01-09 -0.095579
2000-01-10 -0.237109
Freq: D, dtype: float64
For some windowing functions, additional parameters must be specified:

In [57]: ser.rolling(window=5, win_type='gaussian').mean(std=0.1)

Out[57]:
2000-01-01 NaN
2000-01-02 NaN
2000-01-03 NaN
2000-01-04 NaN
2000-01-05 -0.254437
2000-01-06 1.434351
2000-01-07 1.014205
2000-01-08 -0.460612
2000-01-09 -1.326783
2000-01-10 0.612663
Freq: D, dtype: float64

Note: For .sum() with a win_type, there is no normalization done to the weights for the window. Passing custom
weights of [1, 1, 1] will yield a different result than passing weights of [2, 2, 2], for example. When passing
a win_type instead of explicitly specifying the weights, the weights are already normalized so that the largest weight
is 1.
In contrast, the nature of the .mean() calculation is such that the weights are normalized with respect to each other.
Weights of [1, 1, 1] and [2, 2, 2] yield the same result.

Time-aware rolling

New in version 0.19.0.

New in version 0.19.0 are the ability to pass an offset (or convertible) to a .rolling() method and have it produce
variable sized windows based on the passed time window. For each time point, this includes all preceding values
occurring within the indicated time delta.
This can be particularly useful for a non-regular time frequency index.

In [58]: dft = pd.DataFrame({'B': [0, 1, 2, np.nan, 4]},

....: index=pd.date_range('20130101 09:00:00',
....: periods=5,
....: freq='s'))
....:

In [59]: dft
Out[59]:
B
2013-01-01 09:00:00 0.0
2013-01-01 09:00:01 1.0
2013-01-01 09:00:02 2.0
2013-01-01 09:00:03 NaN
2013-01-01 09:00:04 4.0

This is a regular frequency index. Using an integer window parameter works to roll along the window frequency.
In [60]: dft.rolling(2).sum()

4.11. Computational tools 619

pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[60]:
B
2013-01-01 09:00:00 NaN
2013-01-01 09:00:01 1.0
2013-01-01 09:00:02 3.0
2013-01-01 09:00:03 NaN
2013-01-01 09:00:04 NaN

In [61]: dft.rolling(2, min_periods=1).sum()

Out[61]:
B
2013-01-01 09:00:00 0.0
2013-01-01 09:00:01 1.0
2013-01-01 09:00:02 3.0
2013-01-01 09:00:03 2.0
2013-01-01 09:00:04 4.0
Specifying an offset allows a more intuitive specification of the rolling frequency.

In [62]: dft.rolling('2s').sum()
Out[62]:
B
2013-01-01 09:00:00 0.0
2013-01-01 09:00:01 1.0
2013-01-01 09:00:02 3.0
2013-01-01 09:00:03 2.0
2013-01-01 09:00:04 4.0

Using a non-regular, but still monotonic index, rolling with an integer window does not impart any special calculation.
In [63]: dft = pd.DataFrame({'B': [0, 1, 2, np.nan, 4]},
....: index=pd.Index([pd.Timestamp('20130101 09:00:00'),
....: pd.Timestamp('20130101 09:00:02'),
....: pd.Timestamp('20130101 09:00:03'),
....: pd.Timestamp('20130101 09:00:05'),
....: pd.Timestamp('20130101
,→09:00:06')],

....: name='foo'))
....:

In [64]: dft
Out[64]:
B
foo
2013-01-01 09:00:00 0.0
2013-01-01 09:00:02 1.0
2013-01-01 09:00:03 2.0
2013-01-01 09:00:05 NaN
2013-01-01 09:00:06 4.0

In [65]: dft.rolling(2).sum()
Out[65]:
B
foo
2013-01-01 09:00:00 NaN
2013-01-01 09:00:02 1.0

620 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

2013-01-01 09:00:03 3.0

2013-01-01 09:00:05 NaN
2013-01-01 09:00:06 NaN
Using the time-specification generates variable windows for this sparse data.

In [66]: dft.rolling('2s').sum()
Out[66]:
B
foo
2013-01-01 09:00:00 0.0
2013-01-01 09:00:02 1.0
2013-01-01 09:00:03 3.0
2013-01-01 09:00:05 NaN
2013-01-01 09:00:06 4.0

Furthermore, we now allow an optional on parameter to specify a column (rather than the default of the index) in a
DataFrame.
In [67]: dft = dft.reset_index()

In [68]: dft
Out[68]:
foo B
0 2013-01-01 09:00:00 0.0
1 2013-01-01 09:00:02 1.0
2 2013-01-01 09:00:03 2.0
3 2013-01-01 09:00:05 NaN
4 2013-01-01 09:00:06 4.0

In [69]: dft.rolling('2s', on='foo').sum()

Out[69]:
foo B
0 2013-01-01 09:00:00 0.0
1 2013-01-01 09:00:02 1.0
2 2013-01-01 09:00:03 3.0
3 2013-01-01 09:00:05 NaN
4 2013-01-01 09:00:06 4.0

Rolling window endpoints

New in version 0.20.0.

The inclusion of the interval endpoints in rolling window calculations can be specified with the closed parameter:

closed Description Default for

right close right endpoint time-based windows
left close left endpoint
both close both endpoints fixed windows
neither open endpoints

For example, having the right endpoint open is useful in many problems that require that there is no contamination
from present information back to past information. This allows the rolling window to compute statistics up to that
point in time, but not including that point in time.

4.11. Computational tools 621

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [70]: df = pd.DataFrame({'x': 1},

....: index=[pd.Timestamp('20130101 09:00:01'),
....: pd.Timestamp('20130101 09:00:02'),
....: pd.Timestamp('20130101 09:00:03'),
....: pd.Timestamp('20130101 09:00:04'),
....: pd.Timestamp('20130101 09:00:06')])
....:

In [71]: df["right"] = df.rolling('2s', closed='right').x.sum() # default

In [72]: df["both"] = df.rolling('2s', closed='both').x.sum()

In [73]: df["left"] = df.rolling('2s', closed='left').x.sum()

In [74]: df["neither"] = df.rolling('2s', closed='neither').x.sum()

In [75]: df
Out[75]:
x right both left neither
2013-01-01 09:00:01 1 1.0 1.0 NaN NaN
2013-01-01 09:00:02 1 2.0 2.0 1.0 1.0
2013-01-01 09:00:03 1 2.0 3.0 2.0 1.0
2013-01-01 09:00:04 1 2.0 3.0 2.0 1.0
2013-01-01 09:00:06 1 1.0 2.0 1.0 NaN

Currently, this feature is only implemented for time-based windows. For fixed windows, the closed parameter cannot
be set and the rolling window will always have both endpoints closed.

Time-aware rolling vs. resampling

Using .rolling() with a time-based index is quite similar to resampling. They both operate and perform reductive
operations on time-indexed pandas objects.
When using .rolling() with an offset. The offset is a time-delta. Take a backwards-in-time looking window, and
aggregate all of the values in that window (including the end-point, but not the start-point). This is the new value at
that point in the result. These are variable sized windows in time-space for each point of the input. You will get a same
sized result as the input.
When using .resample() with an offset. Construct a new index that is the frequency of the offset. For each
frequency bin, aggregate points from the input within a backwards-in-time looking window that fall in that bin. The
result of this aggregation is the output for that frequency point. The windows are fixed size in the frequency space.
Your result will have the shape of a regular frequency between the min and the max of the original input object.
To summarize, .rolling() is a time-based window operation, while .resample() is a frequency-based window
operation.

Centering windows

By default the labels are set to the right edge of the window, but a center keyword is available so the labels can be
set at the center.
In [76]: ser.rolling(window=5).mean()
Out[76]:
2000-01-01 NaN
2000-01-02 NaN
2000-01-03 NaN

622 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

2000-01-04 NaN
2000-01-05 0.540420
2000-01-06 0.306313
2000-01-07 0.081345
2000-01-08 0.254765
2000-01-09 -0.095579
2000-01-10 -0.237109
Freq: D, dtype: float64

In [77]: ser.rolling(window=5, center=True).mean()

Out[77]:
2000-01-01 NaN
2000-01-02 NaN
2000-01-03 0.540420
2000-01-04 0.306313
2000-01-05 0.081345
2000-01-06 0.254765
2000-01-07 -0.095579
2000-01-08 -0.237109
2000-01-09 NaN
2000-01-10 NaN
Freq: D, dtype: float64

Binary window functions

cov() and corr() can compute moving window statistics about two Series or any combination of DataFrame/
Series or DataFrame/DataFrame. Here is the behavior in each case:
• two Series: compute the statistic for the pairing.
• DataFrame/Series: compute the statistics for each column of the DataFrame with the passed Series, thus
returning a DataFrame.
• DataFrame/DataFrame: by default compute the statistic for matching column names, returning a
DataFrame. If the keyword argument pairwise=True is passed then computes the statistic for each pair
of columns, returning a MultiIndexed DataFrame whose index are the dates in question (see the next
section).
For example:

In [78]: df = pd.DataFrame(np.random.randn(1000, 4),

....: index=pd.date_range('1/1/2000', periods=1000),
....: columns=['A', 'B', 'C', 'D'])
....:

In [79]: df = df.cumsum()

In [80]: df2 = df[:20]

In [81]: df2.rolling(window=5).corr(df2['B'])
Out[81]:
A B C D
2000-01-01 NaN NaN NaN NaN
2000-01-02 NaN NaN NaN NaN
2000-01-03 NaN NaN NaN NaN
2000-01-04 NaN NaN NaN NaN
(continues on next page)

4.11. Computational tools 623

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2000-01-05 -0.410454 1.0 0.523221 -0.860934
2000-01-06 0.691714 1.0 -0.006852 -0.883151
2000-01-07 0.127010 1.0 -0.819233 -0.920657
2000-01-08 0.083226 1.0 -0.707348 -0.612911
2000-01-09 0.366106 1.0 -0.868906 0.814301
2000-01-10 -0.027745 1.0 -0.948739 -0.236256
2000-01-11 -0.503110 1.0 -0.512448 -0.387532
2000-01-12 0.245194 1.0 -0.269958 -0.284629
2000-01-13 0.045798 1.0 -0.512087 -0.209225
2000-01-14 -0.226867 1.0 -0.580830 0.203087
2000-01-15 -0.152839 1.0 -0.566645 0.034572
2000-01-16 0.385888 1.0 -0.283130 0.108951
2000-01-17 0.057373 1.0 0.493143 -0.044273
2000-01-18 -0.162383 1.0 0.449045 -0.189665
2000-01-19 -0.174013 1.0 0.266842 -0.661392
2000-01-20 0.486804 1.0 0.120970 0.139172

Computing rolling pairwise covariances and correlations

In financial data analysis and other fields its common to compute covariance and correlation matrices for a collection
of time series. Often one is also interested in moving-window covariance and correlation matrices. This can be done
by passing the pairwise keyword argument, which in the case of DataFrame inputs will yield a MultiIndexed
DataFrame whose index are the dates in question. In the case of a single DataFrame argument the pairwise
argument can even be omitted:

Note: Missing values are ignored and each entry is computed using the pairwise complete observations. Please see
the covariance section for caveats associated with this method of calculating covariance and correlation matrices.

In [82]: covs = (df[['B', 'C', 'D']].rolling(window=50)

....: .cov(df[['A', 'B', 'C']], pairwise=True))
....:

In [83]: covs.loc['2002-09-22':]
Out[83]:
B C D
2002-09-22 A -11.788482 3.788620 -9.532202
B 27.214936 0.218853 13.566773
C 0.218853 5.933531 -4.591019
2002-09-23 A -12.096941 4.076409 -10.208554
B 27.436671 -0.830191 15.370018
C -0.830191 5.582126 -4.811035
2002-09-24 A -12.239394 4.330590 -10.742513
B 27.226658 -1.658535 16.437777
C -1.658535 5.143402 -4.960592
2002-09-25 A -12.752127 4.548132 -11.577657
B 27.467830 -2.486227 17.845438
C -2.486227 4.540150 -4.973443
2002-09-26 A -13.236082 4.531937 -12.379683
B 27.459811 -2.780055 19.188470
C -2.780055 4.387162 -4.798969

624 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [84]: correls = df.rolling(window=50).corr()

In [85]: correls.loc['2002-09-22':]
Out[85]:
A B C D
2002-09-22 A 1.000000 -0.632380 0.435259 -0.758112
B -0.632380 1.000000 0.017222 0.739077
C 0.435259 0.017222 1.000000 -0.535635
D -0.758112 0.739077 -0.535635 1.000000
2002-09-23 A 1.000000 -0.642018 0.479640 -0.757166
B -0.642018 1.000000 -0.067083 0.782885
C 0.479640 -0.067083 1.000000 -0.543285
D -0.757166 0.782885 -0.543285 1.000000
2002-09-24 A 1.000000 -0.648609 0.528010 -0.758338
B -0.648609 1.000000 -0.140153 0.804234
C 0.528010 -0.140153 1.000000 -0.558399
D -0.758338 0.804234 -0.558399 1.000000
2002-09-25 A 1.000000 -0.662761 0.581413 -0.768971
B -0.662761 1.000000 -0.222635 0.830267
C 0.581413 -0.222635 1.000000 -0.569147
D -0.768971 0.830267 -0.569147 1.000000
2002-09-26 A 1.000000 -0.678836 0.581495 -0.774662
B -0.678836 1.000000 -0.253287 0.852590
C 0.581495 -0.253287 1.000000 -0.533463
D -0.774662 0.852590 -0.533463 1.000000

You can efficiently retrieve the time series of correlations between two columns by reshaping and indexing:

In [86]: correls.unstack(1)[('A', 'C')].plot()

Out[86]: <matplotlib.axes._subplots.AxesSubplot at 0x1a30dc52d0>

4.11. Computational tools 625

pandas: powerful Python data analysis toolkit, Release 0.25.3

4.11.3 Aggregation

Once the Rolling, Expanding or EWM objects have been created, several methods are available to perform multiple
computations on the data. These operations are similar to the aggregating API, groupby API, and resample API.

In [87]: dfa = pd.DataFrame(np.random.randn(1000, 3),

....: index=pd.date_range('1/1/2000', periods=1000),
....: columns=['A', 'B', 'C'])
....:

In [88]: r = dfa.rolling(window=60, min_periods=1)

In [89]: r
Out[89]: Rolling [window=60,min_periods=1,center=False,axis=0]

We can aggregate by passing a function to the entire DataFrame, or select a Series (or multiple Series) via standard
__getitem__.
In [90]: r.aggregate(np.sum)
Out[90]:
A B C
2000-01-01 1.467445 0.520922 0.333619
2000-01-02 0.913815 0.659062 0.678084

626 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

2000-01-03 1.697014 2.104923 1.540576

2000-01-04 1.205306 2.617423 0.435920
2000-01-05 0.474929 1.072642 0.866405
... ... ... ...
2002-09-22 -2.120855 1.278532 0.615368
2002-09-23 -1.517675 2.423909 4.625449
2002-09-24 -0.633229 1.370941 5.138608
2002-09-25 -1.552531 0.546496 6.269487
2002-09-26 0.711361 2.818447 4.596824

[1000 rows x 3 columns]

In [91]: r['A'].aggregate(np.sum)
Out[91]:
2000-01-01 1.467445
2000-01-02 0.913815
2000-01-03 1.697014
2000-01-04 1.205306
2000-01-05 0.474929
...
2002-09-22 -2.120855
2002-09-23 -1.517675
2002-09-24 -0.633229
2002-09-25 -1.552531
2002-09-26 0.711361
Freq: D, Name: A, Length: 1000, dtype: float64

In [92]: r[['A', 'B']].aggregate(np.sum)

Out[92]:
A B
2000-01-01 1.467445 0.520922
2000-01-02 0.913815 0.659062
2000-01-03 1.697014 2.104923
2000-01-04 1.205306 2.617423
2000-01-05 0.474929 1.072642
... ... ...
2002-09-22 -2.120855 1.278532
2002-09-23 -1.517675 2.423909
2002-09-24 -0.633229 1.370941
2002-09-25 -1.552531 0.546496
2002-09-26 0.711361 2.818447

[1000 rows x 2 columns]

As you can see, the result of the aggregation will have the selected columns, or all columns if none are selected.

Applying multiple functions

With windowed Series you can also pass a list of functions to do aggregation with, outputting a DataFrame:

In [93]: r['A'].agg([np.sum, np.mean, np.std])

Out[93]:
sum mean std
(continues on next page)

4.11. Computational tools 627

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2000-01-01 1.467445 1.467445 NaN
2000-01-02 0.913815 0.456908 1.429116
2000-01-03 1.697014 0.565671 1.027947
2000-01-04 1.205306 0.301326 0.991949
2000-01-05 0.474929 0.094986 0.975118
... ... ... ...
2002-09-22 -2.120855 -0.035348 1.000696
2002-09-23 -1.517675 -0.025295 0.990434
2002-09-24 -0.633229 -0.010554 0.995740
2002-09-25 -1.552531 -0.025876 1.003939
2002-09-26 0.711361 0.011856 1.018469

[1000 rows x 3 columns]

On a windowed DataFrame, you can pass a list of functions to apply to each column, which produces an aggregated
result with a hierarchical index:

In [94]: r.agg([np.sum, np.mean])

Out[94]:
A B C
sum mean sum mean sum mean
2000-01-01 1.467445 1.467445 0.520922 0.520922 0.333619 0.333619
2000-01-02 0.913815 0.456908 0.659062 0.329531 0.678084 0.339042
2000-01-03 1.697014 0.565671 2.104923 0.701641 1.540576 0.513525
2000-01-04 1.205306 0.301326 2.617423 0.654356 0.435920 0.108980
2000-01-05 0.474929 0.094986 1.072642 0.214528 0.866405 0.173281
... ... ... ... ... ... ...
2002-09-22 -2.120855 -0.035348 1.278532 0.021309 0.615368 0.010256
2002-09-23 -1.517675 -0.025295 2.423909 0.040398 4.625449 0.077091
2002-09-24 -0.633229 -0.010554 1.370941 0.022849 5.138608 0.085643
2002-09-25 -1.552531 -0.025876 0.546496 0.009108 6.269487 0.104491
2002-09-26 0.711361 0.011856 2.818447 0.046974 4.596824 0.076614

[1000 rows x 6 columns]

Passing a dict of functions has different behavior by default, see the next section.

Applying different functions to DataFrame columns

By passing a dict to aggregate you can apply a different aggregation to the columns of a DataFrame:

In [95]: r.agg({'A': np.sum, 'B': lambda x: np.std(x, ddof=1)})

Out[95]:
A B
2000-01-01 1.467445 NaN
2000-01-02 0.913815 0.270668
2000-01-03 1.697014 0.672331
2000-01-04 1.205306 0.557042
2000-01-05 0.474929 1.095428
... ... ...
2002-09-22 -2.120855 1.100184
2002-09-23 -1.517675 1.106164
2002-09-24 -0.633229 1.099861
2002-09-25 -1.552531 1.099031
2002-09-26 0.711361 1.159605
(continues on next page)

628 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

[1000 rows x 2 columns]

The function names can also be strings. In order for a string to be valid it must be implemented on the windowed
object

In [96]: r.agg({'A': 'sum', 'B': 'std'})

Out[96]:
A B
2000-01-01 1.467445 NaN
2000-01-02 0.913815 0.270668
2000-01-03 1.697014 0.672331
2000-01-04 1.205306 0.557042
2000-01-05 0.474929 1.095428
... ... ...
2002-09-22 -2.120855 1.100184
2002-09-23 -1.517675 1.106164
2002-09-24 -0.633229 1.099861
2002-09-25 -1.552531 1.099031
2002-09-26 0.711361 1.159605

[1000 rows x 2 columns]

Furthermore you can pass a nested dict to indicate different aggregations on different columns.

In [97]: r.agg({'A': ['sum', 'std'], 'B': ['mean', 'std']})

Out[97]:
A B
sum std mean std
2000-01-01 1.467445 NaN 0.520922 NaN
2000-01-02 0.913815 1.429116 0.329531 0.270668
2000-01-03 1.697014 1.027947 0.701641 0.672331
2000-01-04 1.205306 0.991949 0.654356 0.557042
2000-01-05 0.474929 0.975118 0.214528 1.095428
... ... ... ... ...
2002-09-22 -2.120855 1.000696 0.021309 1.100184
2002-09-23 -1.517675 0.990434 0.040398 1.106164
2002-09-24 -0.633229 0.995740 0.022849 1.099861
2002-09-25 -1.552531 1.003939 0.009108 1.099031
2002-09-26 0.711361 1.018469 0.046974 1.159605

[1000 rows x 4 columns]

4.11.4 Expanding windows

A common alternative to rolling statistics is to use an expanding window, which yields the value of the statistic with
all the data available up to that point in time.
These follow a similar interface to .rolling, with the .expanding method returning an Expanding object.
As these calculations are a special case of rolling statistics, they are implemented in pandas such that the following
two calls are equivalent:
In [98]: df.rolling(window=len(df), min_periods=1).mean()[:5]
Out[98]:
A B C D

4.11. Computational tools 629

pandas: powerful Python data analysis toolkit, Release 0.25.3

2000-01-01 0.382444 0.209948 -0.510030 -0.394100

2000-01-02 0.030851 0.445864 -0.374918 -1.316470
2000-01-03 -0.100175 1.104252 -0.160086 -1.775538
2000-01-04 -0.192204 1.391755 0.077095 -1.869501
2000-01-05 -0.181876 1.961661 0.071067 -2.189869

In [99]: df.expanding(min_periods=1).mean()[:5]
Out[99]:
A B C D
2000-01-01 0.382444 0.209948 -0.510030 -0.394100
2000-01-02 0.030851 0.445864 -0.374918 -1.316470
2000-01-03 -0.100175 1.104252 -0.160086 -1.775538
2000-01-04 -0.192204 1.391755 0.077095 -1.869501
2000-01-05 -0.181876 1.961661 0.071067 -2.189869
These have a similar set of methods to .rolling methods.

Method summary

Function Description
count() Number of non-null observations
sum() Sum of values
mean() Mean of values
median() Arithmetic median of values
min() Minimum
max() Maximum
std() Unbiased standard deviation
var() Unbiased variance
skew() Unbiased skewness (3rd moment)
kurt() Unbiased kurtosis (4th moment)
quantile() Sample quantile (value at %)
apply() Generic apply
cov() Unbiased covariance (binary)
corr() Correlation (binary)

Aside from not having a window parameter, these functions have the same interfaces as their .rolling counter-
parts. Like above, the parameters they all accept are:
• min_periods: threshold of non-null data points to require. Defaults to minimum needed to compute statistic.
No NaNs will be output once min_periods non-null data points have been seen.
• center: boolean, whether to set the labels at the center (default is False).

Note: The output of the .rolling and .expanding methods do not return a NaN if there are at least
min_periods non-null values in the current window. For example:
In [100]: sn = pd.Series([1, 2, np.nan, 3, np.nan, 4])

In [101]: sn
Out[101]:
0 1.0
1 2.0
2 NaN

630 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

3 3.0
4 NaN
5 4.0
dtype: float64

In [102]: sn.rolling(2).max()
Out[102]:
0 NaN
1 2.0
2 NaN
3 NaN
4 NaN
5 NaN
dtype: float64

In [103]: sn.rolling(2, min_periods=1).max()

Out[103]:
0 1.0
1 2.0
2 2.0
3 3.0
4 3.0
5 4.0
dtype: float64
In case of expanding functions, this differs from cumsum(), cumprod(), cummax(), and cummin(), which
return NaN in the output wherever a NaN is encountered in the input. In order to match the output of cumsum with
expanding, use fillna():
In [104]: sn.expanding().sum()
Out[104]:
0 1.0
1 3.0
2 3.0
3 6.0
4 6.0
5 10.0
dtype: float64

In [105]: sn.cumsum()
Out[105]:
0 1.0
1 3.0
2 NaN
3 6.0
4 NaN
5 10.0
dtype: float64

In [106]: sn.cumsum().fillna(method='ffill')
Out[106]:
0 1.0
1 3.0
2 3.0
3 6.0

4.11. Computational tools 631

pandas: powerful Python data analysis toolkit, Release 0.25.3

4 6.0
5 10.0
dtype: float64

An expanding window statistic will be more stable (and less responsive) than its rolling window counterpart as the
increasing window size decreases the relative impact of an individual data point. As an example, here is the mean()
output for the previous time series dataset:
In [107]: s.plot(style='k--')
Out[107]: <matplotlib.axes._subplots.AxesSubplot at 0x1a313078d0>

In [108]: s.expanding().mean().plot(style='k')
Out[108]: <matplotlib.axes._subplots.AxesSubplot at 0x1a313078d0>

4.11.5 Exponentially weighted windows

A related set of functions are exponentially weighted versions of several of the above statistics. A similar interface
to .rolling and .expanding is accessed through the .ewm method to receive an EWM object. A number of
expanding EW (exponentially weighted) methods are provided:

632 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Function Description
mean() EW moving average
var() EW moving variance
std() EW moving standard deviation
corr() EW moving correlation
cov() EW moving covariance

In general, a weighted moving average is calculated as

∑t
i=0 wi xt−i
yt = ∑ t ,
i=0 wi

where xt is the input, yt is the result and the wi are the weights.
The EW functions support two variants of exponential weights. The default, adjust=True, uses the weights wi =
(1 − α)i which gives

xt + (1 − α)xt−1 + (1 − α)2 xt−2 + ... + (1 − α)t x0

yt =
1 + (1 − α) + (1 − α)2 + ... + (1 − α)t

When adjust=False is specified, moving averages are calculated as

y0 = x0
yt = (1 − α)yt−1 + αxt ,

which is equivalent to using weights

{
α(1 − α)i if i < t
wi =
(1 − α)i if i = t.

Note: These equations are sometimes written in terms of α′ = 1 − α, e.g.

yt = α′ yt−1 + (1 − α′ )xt .

The difference between the above two variants arises because we are dealing with series which have finite history.
Consider a series of infinite history, with adjust=True:

xt + (1 − α)xt−1 + (1 − α)2 xt−2 + ...

yt =
1 + (1 − α) + (1 − α)2 + ...

Noting that the denominator is a geometric series with initial term equal to 1 and a ratio of 1 − α we have

xt + (1 − α)xt−1 + (1 − α)2 xt−2 + ...

yt = 1
1−(1−α)

= [xt + (1 − α)xt−1 + (1 − α)2 xt−2 + ...]α

= αxt + [(1 − α)xt−1 + (1 − α)2 xt−2 + ...]α
= αxt + (1 − α)[xt−1 + (1 − α)xt−2 + ...]α
= αxt + (1 − α)yt−1

which is the same expression as adjust=False above and therefore shows the equivalence of the two variants for
infinite series. When adjust=False, we have y0 = x0 and yt = αxt + (1 − α)yt−1 . Therefore, there is an

4.11. Computational tools 633

pandas: powerful Python data analysis toolkit, Release 0.25.3

assumption that x0 is not an ordinary value but rather an exponentially weighted moment of the infinite series up to
that point.
One must have 0 < α ≤ 1, and while since version 0.18.0 it has been possible to pass α directly, its often easier to
think about either the span, center of mass (com) or half-life of an EW moment:

 2
 s+1 , for span s ≥ 1
1
α = 1+c , for center of mass c ≥ 0

 log 0.5
1 − exp h , for half-life h > 0

One must specify precisely one of span, center of mass, half-life and alpha to the EW functions:
• Span corresponds to what is commonly called an N-day EW moving average.
• Center of mass has a more physical interpretation and can be thought of in terms of span: c = (s − 1)/2.
• Half-life is the period of time for the exponential weight to reduce to one half.
• Alpha specifies the smoothing factor directly.
Here is an example for a univariate time series:
In [109]: s.plot(style='k--')
Out[109]: <matplotlib.axes._subplots.AxesSubplot at 0x1a312dc6d0>

In [110]: s.ewm(span=20).mean().plot(style='k')
Out[110]: <matplotlib.axes._subplots.AxesSubplot at 0x1a312dc6d0>

634 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

EWM has a min_periods argument, which has the same meaning it does for all the .expanding and .rolling
methods: no output values will be set until at least min_periods non-null values are encountered in the (expanding)
window.
EWM also has an ignore_na argument, which determines how intermediate null values affect the calculation of
the weights. When ignore_na=False (the default), weights are calculated based on absolute positions, so that
intermediate null values affect the result. When ignore_na=True, weights are calculated by ignoring intermediate
null values. For example, assuming adjust=True, if ignore_na=False, the weighted average of 3, NaN, 5
would be calculated as
(1 − α)2 · 3 + 1 · 5
.
(1 − α)2 + 1
Whereas if ignore_na=True, the weighted average would be calculated as
(1 − α) · 3 + 1 · 5
.
(1 − α) + 1
The var(), std(), and cov() functions have a bias argument, specifying whether the result should con-
tain biased or unbiased statistics. For example, if bias=True, ewmvar(x) is calculated as ewmvar(x) =
ewma(x**2) - ewma(x)**2; whereas if bias=False (the default), the biased variance statistics are scaled
by debiasing factors
(∑ )2
t
i=0 wi
(∑ )2 ∑ .
t t
i=0 iw − w
i=0 i
2

(For wi = 1, this reduces to the usual N/(N − 1) factor, with N = t + 1.) See Weighted Sample Variance on
Wikipedia for further details. {{ header }}

4.12 Group By: split-apply-combine

By group by we are referring to a process involving one or more of the following steps:
• Splitting the data into groups based on some criteria.
• Applying a function to each group independently.
• Combining the results into a data structure.
Out of these, the split step is the most straightforward. In fact, in many situations we may wish to split the data set into
groups and do something with those groups. In the apply step, we might wish to do one of the following:
• Aggregation: compute a summary statistic (or statistics) for each group. Some examples:
– Compute group sums or means.
– Compute group sizes / counts.
• Transformation: perform some group-specific computations and return a like-indexed object. Some examples:
– Standardize data (zscore) within a group.
– Filling NAs within groups with a value derived from each group.
• Filtration: discard some groups, according to a group-wise computation that evaluates True or False. Some
examples:
– Discard data that belongs to groups with only a few members.
– Filter out data based on the group sum or mean.

4.12. Group By: split-apply-combine 635

pandas: powerful Python data analysis toolkit, Release 0.25.3

• Some combination of the above: GroupBy will examine the results of the apply step and try to return a sensibly
combined result if it doesnt fit into either of the above two categories.
Since the set of object instance methods on pandas data structures are generally rich and expressive, we often simply
want to invoke, say, a DataFrame function on each group. The name GroupBy should be quite familiar to those who
have used a SQL-based tool (or itertools), in which you can write code like:

SELECT Column1, Column2, mean(Column3), sum(Column4)

FROM SomeTable
GROUP BY Column1, Column2

We aim to make operations like this natural and easy to express using pandas. Well address each area of GroupBy
functionality then provide some non-trivial examples / use cases.
See the cookbook for some advanced strategies.

4.12.1 Splitting an object into groups

pandas objects can be split on any of their axes. The abstract definition of grouping is to provide a mapping of labels
to group names. To create a GroupBy object (more on what the GroupBy object is later), you may do the following:

In [1]: df = pd.DataFrame([('bird', 'Falconiformes', 389.0),

...: ('bird', 'Psittaciformes', 24.0),
...: ('mammal', 'Carnivora', 80.2),
...: ('mammal', 'Primates', np.nan),
...: ('mammal', 'Carnivora', 58)],
...: index=['falcon', 'parrot', 'lion', 'monkey', 'leopard'],
...: columns=('class', 'order', 'max_speed'))
...:

In [2]: df
Out[2]:
class order max_speed
falcon bird Falconiformes 389.0
parrot bird Psittaciformes 24.0
lion mammal Carnivora 80.2
monkey mammal Primates NaN
leopard mammal Carnivora 58.0

# default is axis=0
In [3]: grouped = df.groupby('class')

In [4]: grouped = df.groupby('order', axis='columns')

In [5]: grouped = df.groupby(['class', 'order'])

The mapping can be specified many different ways:

• A Python function, to be called on each of the axis labels.
• A list or NumPy array of the same length as the selected axis.
• A dict or Series, providing a label -> group name mapping.
• For DataFrame objects, a string indicating a column to be used to group. Of course df.groupby('A') is
just syntactic sugar for df.groupby(df['A']), but it makes life simpler.
• For DataFrame objects, a string indicating an index level to be used to group.
• A list of any of the above things.

636 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Collectively we refer to the grouping objects as the keys. For example, consider the following DataFrame:

Note: A string passed to groupby may refer to either a column or an index level. If a string matches both a column
name and an index level name, a ValueError will be raised.

In [6]: df = pd.DataFrame({'A': ['foo', 'bar', 'foo', 'bar',

...: 'foo', 'bar', 'foo', 'foo'],
...: 'B': ['one', 'one', 'two', 'three',
...: 'two', 'two', 'one', 'three'],
...: 'C': np.random.randn(8),
...: 'D': np.random.randn(8)})
...:

In [7]: df
Out[7]:
A B C D
0 foo one -0.024896 0.238006
1 bar one 0.286068 -0.761142
2 foo two 0.177576 2.097385
3 bar three -0.143033 0.354051
4 foo two 0.960747 -0.687808
5 bar two -1.675479 1.725302
6 foo one 0.227102 -0.755041
7 foo three -0.557170 0.230925

On a DataFrame, we obtain a GroupBy object by calling groupby(). We could naturally group by either the A or B
columns, or both:
In [8]: grouped = df.groupby('A')

In [9]: grouped = df.groupby(['A', 'B'])

New in version 0.24.

If we also have a MultiIndex on columns A and B, we can group by all but the specified columns
In [10]: df2 = df.set_index(['A', 'B'])

In [11]: grouped = df2.groupby(level=df2.index.names.difference(['B']))

In [12]: grouped.sum()
Out[12]:
C D
A
bar -1.532443 1.318211
foo 0.783358 1.123467

These will split the DataFrame on its index (rows). We could also split by the columns:
In [13]: def get_letter_type(letter):
....: if letter.lower() in 'aeiou':
....: return 'vowel'
....: else:
....: return 'consonant'
....:

In [14]: grouped = df.groupby(get_letter_type, axis=1)

4.12. Group By: split-apply-combine 637

pandas: powerful Python data analysis toolkit, Release 0.25.3

pandas Index objects support duplicate values. If a non-unique index is used as the group key in a groupby operation,
all values for the same index value will be considered to be in one group and thus the output of aggregation functions
will only contain unique index values:
In [15]: lst = [1, 2, 3, 1, 2, 3]

In [16]: s = pd.Series([1, 2, 3, 10, 20, 30], lst)

In [17]: grouped = s.groupby(level=0)

In [18]: grouped.first()
Out[18]:
1 1
2 2
3 3
dtype: int64

In [19]: grouped.last()
Out[19]:
1 10
2 20
3 30
dtype: int64

In [20]: grouped.sum()
Out[20]:
1 11
2 22
3 33
dtype: int64
Note that no splitting occurs until its needed. Creating the GroupBy object only verifies that youve passed a valid
mapping.

Note: Many kinds of complicated data manipulations can be expressed in terms of GroupBy operations (though cant
be guaranteed to be the most efficient). You can get quite creative with the label mapping functions.

GroupBy sorting

By default the group keys are sorted during the groupby operation. You may however pass sort=False for
potential speedups:
In [21]: df2 = pd.DataFrame({'X': ['B', 'B', 'A', 'A'], 'Y': [1, 2, 3, 4]})

In [22]: df2.groupby(['X']).sum()
Out[22]:
Y
X
A 7
B 3

In [23]: df2.groupby(['X'], sort=False).sum()

Out[23]:

638 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Y
X
B 3
A 7
Note that groupby will preserve the order in which observations are sorted within each group. For example, the
groups created by groupby() below are in the order they appeared in the original DataFrame:
In [24]: df3 = pd.DataFrame({'X': ['A', 'B', 'A', 'B'], 'Y': [1, 4, 3, 2]})

In [25]: df3.groupby(['X']).get_group('A')
Out[25]:
X Y
0 A 1
2 A 3

In [26]: df3.groupby(['X']).get_group('B')
Out[26]:
X Y
1 B 4
3 B 2

GroupBy object attributes

The groups attribute is a dict whose keys are the computed unique groups and corresponding values being the axis
labels belonging to each group. In the above example we have:
In [27]: df.groupby('A').groups
Out[27]:
{'bar': Int64Index([1, 3, 5], dtype='int64'),
'foo': Int64Index([0, 2, 4, 6, 7], dtype='int64')}

In [28]: df.groupby(get_letter_type, axis=1).groups

Out[28]:
{'consonant': Index(['B', 'C', 'D'], dtype='object'),
'vowel': Index(['A'], dtype='object')}
Calling the standard Python len function on the GroupBy object just returns the length of the groups dict, so it is
largely just a convenience:
In [29]: grouped = df.groupby(['A', 'B'])

In [30]: grouped.groups
Out[30]:
{('bar', 'one'): Int64Index([1], dtype='int64'),
('bar', 'three'): Int64Index([3], dtype='int64'),
('bar', 'two'): Int64Index([5], dtype='int64'),
('foo', 'one'): Int64Index([0, 6], dtype='int64'),
('foo', 'three'): Int64Index([7], dtype='int64'),
('foo', 'two'): Int64Index([2, 4], dtype='int64')}

In [31]: len(grouped)
Out[31]: 6
GroupBy will tab complete column names (and other attributes):

4.12. Group By: split-apply-combine 639

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [32]: df
Out[32]:
height weight gender
2000-01-01 66.625231 167.442586 female
2000-01-02 67.783272 170.355866 male
2000-01-03 59.226150 132.525248 female
2000-01-04 45.989647 132.927883 female
2000-01-05 55.485350 142.663201 male
2000-01-06 49.650341 169.739406 male
2000-01-07 70.884977 152.414050 female
2000-01-08 57.512599 178.333864 male
2000-01-09 57.653796 177.259389 female
2000-01-10 59.657849 146.833646 female

In [33]: gb = df.groupby('gender')

In [34]: gb.<TAB> # noqa: E225, E999

gb.agg gb.boxplot gb.cummin gb.describe gb.filter gb.get_group
,→gb.height gb.last gb.median gb.ngroups gb.plot gb.rank
,→gb.std gb.transform
gb.aggregate gb.count gb.cumprod gb.dtype gb.first gb.groups
,→gb.hist gb.max gb.min gb.nth gb.prod gb.resample
,→gb.sum gb.var
gb.apply gb.cummax gb.cumsum gb.fillna gb.gender gb.head
,→gb.indices gb.mean gb.name gb.ohlc gb.quantile gb.size
,→gb.tail gb.weight

GroupBy with MultiIndex

With hierarchically-indexed data, its quite natural to group by one of the levels of the hierarchy.
Lets create a Series with a two-level MultiIndex.

In [35]: arrays = [['bar', 'bar', 'baz', 'baz', 'foo', 'foo', 'qux', 'qux'],
....: ['one', 'two', 'one', 'two', 'one', 'two', 'one', 'two']]
....:

In [36]: index = pd.MultiIndex.from_arrays(arrays, names=['first', 'second'])

In [37]: s = pd.Series(np.random.randn(8), index=index)

In [38]: s
Out[38]:
first second
bar one 0.589964
two -1.842844
baz one 1.601648
two 0.322967
foo one 1.668732
two 1.291467
qux one -0.492679
two -0.557110
dtype: float64

We can then group by one of the levels in s.

640 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [39]: grouped = s.groupby(level=0)

In [40]: grouped.sum()
Out[40]:
first
bar -1.252880
baz 1.924614
foo 2.960199
qux -1.049789
dtype: float64

If the MultiIndex has names specified, these can be passed instead of the level number:

In [41]: s.groupby(level='second').sum()
Out[41]:
second
one 3.367665
two -0.785521
dtype: float64

The aggregation functions such as sum will take the level parameter directly. Additionally, the resulting index will be
named according to the chosen level:

In [42]: s.sum(level='second')
Out[42]:
second
one 3.367665
two -0.785521
dtype: float64

Grouping with multiple levels is supported.

In [43]: s
Out[43]:
first second third
bar doo one -2.165187
two -0.293011
baz bee one 0.416483
two 1.027560
foo bop one -0.422062
two -0.840953
qux bop one -0.390604
two -0.242654
dtype: float64

In [44]: s.groupby(level=['first', 'second']).sum()

Out[44]:
first second
bar doo -2.458198
baz bee 1.444043
foo bop -1.263015
qux bop -0.633258
dtype: float64
New in version 0.20.
Index level names may be supplied as keys.

4.12. Group By: split-apply-combine 641

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [45]: s.groupby(['first', 'second']).sum()

Out[45]:
first second
bar doo -2.458198
baz bee 1.444043
foo bop -1.263015
qux bop -0.633258
dtype: float64

More on the sum function and aggregation later.

Grouping DataFrame with Index levels and columns

A DataFrame may be grouped by a combination of columns and index levels by specifying the column names as
strings and the index levels as pd.Grouper objects.

In [46]: arrays = [['bar', 'bar', 'baz', 'baz', 'foo', 'foo', 'qux', 'qux'],
....: ['one', 'two', 'one', 'two', 'one', 'two', 'one', 'two']]
....:

In [47]: index = pd.MultiIndex.from_arrays(arrays, names=['first', 'second'])

In [48]: df = pd.DataFrame({'A': [1, 1, 1, 1, 2, 2, 3, 3],

....: 'B': np.arange(8)},
....: index=index)
....:

In [49]: df
Out[49]:
A B
first second
bar one 1 0
two 1 1
baz one 1 2
two 1 3
foo one 2 4
two 2 5
qux one 3 6
two 3 7

The following example groups df by the second index level and the A column.

In [50]: df.groupby([pd.Grouper(level=1), 'A']).sum()

Out[50]:
B
second A
one 1 2
2 4
3 6
two 1 4
2 5
3 7

Index levels may also be specified by name.

642 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [51]: df.groupby([pd.Grouper(level='second'), 'A']).sum()

Out[51]:
B
second A
one 1 2
2 4
3 6
two 1 4
2 5
3 7

New in version 0.20.

Index level names may be specified as keys directly to groupby.
In [52]: df.groupby(['second', 'A']).sum()
Out[52]:
B
second A
one 1 2
2 4
3 6
two 1 4
2 5
3 7

DataFrame column selection in GroupBy

Once you have created the GroupBy object from a DataFrame, you might want to do something different for each of
the columns. Thus, using [] similar to getting a column from a DataFrame, you can do:
In [53]: grouped = df.groupby(['A'])

In [54]: grouped_C = grouped['C']

In [55]: grouped_D = grouped['D']

This is mainly syntactic sugar for the alternative and much more verbose:
In [56]: df['C'].groupby(df['A'])
Out[56]: <pandas.core.groupby.generic.SeriesGroupBy object at 0x1c398c76d0>

Additionally this method avoids recomputing the internal grouping information derived from the passed key.

4.12.2 Iterating through groups

With the GroupBy object in hand, iterating through the grouped data is very natural and functions similarly to
itertools.groupby():
In [57]: grouped = df.groupby('A')

In [58]: for name, group in grouped:

....: print(name)
....: print(group)
....:
(continues on next page)

4.12. Group By: split-apply-combine 643

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

bar
A B C D
1 bar one -0.732034 0.531245
3 bar three 0.503626 -1.293798
5 bar two 2.485152 -0.028109
foo
A B C D
0 foo one 1.018905 1.623114
2 foo two -0.242912 -1.390677
4 foo two 0.195596 1.312507
6 foo one 0.299609 1.312379
7 foo three 1.073480 1.869813

In the case of grouping by multiple keys, the group name will be a tuple:

In [59]: for name, group in df.groupby(['A', 'B']):

....: print(name)
....: print(group)
....:
('bar', 'one')
A B C D
1 bar one -0.732034 0.531245
('bar', 'three')
A B C D
3 bar three 0.503626 -1.293798
('bar', 'two')
A B C D
5 bar two 2.485152 -0.028109
('foo', 'one')
A B C D
0 foo one 1.018905 1.623114
6 foo one 0.299609 1.312379
('foo', 'three')
A B C D
7 foo three 1.07348 1.869813
('foo', 'two')
A B C D
2 foo two -0.242912 -1.390677
4 foo two 0.195596 1.312507

See Iterating through groups.

4.12.3 Selecting a group

A single group can be selected using get_group():

In [60]: grouped.get_group('bar')
Out[60]:
A B C D
1 bar one -0.732034 0.531245
3 bar three 0.503626 -1.293798
5 bar two 2.485152 -0.028109

Or for an object grouped on multiple columns:

644 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [61]: df.groupby(['A', 'B']).get_group(('bar', 'one'))

Out[61]:
A B C D
1 bar one -0.732034 0.531245

4.12.4 Aggregation

Once the GroupBy object has been created, several methods are available to perform a computation on the grouped
data. These operations are similar to the aggregating API, window functions API, and resample API.
An obvious one is aggregation via the aggregate() or equivalently agg() method:

In [62]: grouped = df.groupby('A')

In [63]: grouped.aggregate(np.sum)
Out[63]:
C D
A
bar 2.256744 -0.790662
foo 2.344678 4.727137

In [64]: grouped = df.groupby(['A', 'B'])

In [65]: grouped.aggregate(np.sum)
Out[65]:
C D
A B
bar one -0.732034 0.531245
three 0.503626 -1.293798
two 2.485152 -0.028109
foo one 1.318514 2.935493
three 1.073480 1.869813
two -0.047316 -0.078170

As you can see, the result of the aggregation will have the group names as the new index along the grouped axis. In
the case of multiple keys, the result is a MultiIndex by default, though this can be changed by using the as_index
option:
In [66]: grouped = df.groupby(['A', 'B'], as_index=False)

In [67]: grouped.aggregate(np.sum)
Out[67]:
A B C D
0 bar one -0.732034 0.531245
1 bar three 0.503626 -1.293798
2 bar two 2.485152 -0.028109
3 foo one 1.318514 2.935493
4 foo three 1.073480 1.869813
5 foo two -0.047316 -0.078170

In [68]: df.groupby('A', as_index=False).sum()

Out[68]:
A C D
0 bar 2.256744 -0.790662
1 foo 2.344678 4.727137

4.12. Group By: split-apply-combine 645

pandas: powerful Python data analysis toolkit, Release 0.25.3

Note that you could use the reset_index DataFrame function to achieve the same result as the column names are
stored in the resulting MultiIndex:

In [69]: df.groupby(['A', 'B']).sum().reset_index()

Out[69]:
A B C D
0 bar one -0.732034 0.531245
1 bar three 0.503626 -1.293798
2 bar two 2.485152 -0.028109
3 foo one 1.318514 2.935493
4 foo three 1.073480 1.869813
5 foo two -0.047316 -0.078170

Another simple aggregation example is to compute the size of each group. This is included in GroupBy as the size
method. It returns a Series whose index are the group names and whose values are the sizes of each group.

In [70]: grouped.size()
Out[70]:
A B
bar one 1
three 1
two 1
foo one 2
three 1
two 2
dtype: int64

In [71]: grouped.describe()
Out[71]:
C ... D
,→

count mean std min 25% ... min 25% 50%

,→ 75% max
0 1.0 -0.732034 NaN -0.732034 -0.732034 ... 0.531245 0.531245 0.531245 0.
,→531245 0.531245
1 1.0 0.503626 NaN 0.503626 0.503626 ... -1.293798 -1.293798 -1.293798 -1.
,→293798 -1.293798

2 1.0 2.485152 NaN 2.485152 2.485152 ... -0.028109 -0.028109 -0.028109 -0.
,→028109 -0.028109

3 2.0 0.659257 0.508619 0.299609 0.479433 ... 1.312379 1.390063 1.467747 1.

,→545430 1.623114
4 1.0 1.073480 NaN 1.073480 1.073480 ... 1.869813 1.869813 1.869813 1.
,→869813 1.869813
5 2.0 -0.023658 0.310072 -0.242912 -0.133285 ... -1.390677 -0.714881 -0.039085 0.
,→636711 1.312507

[6 rows x 16 columns]

Note: Aggregation functions will not return the groups that you are aggregating over if they are named columns,
when as_index=True, the default. The grouped columns will be the indices of the returned object.
Passing as_index=False will return the groups that you are aggregating over, if they are named columns.

Aggregating functions are the ones that reduce the dimension of the returned objects. Some common aggregating
functions are tabulated below:

646 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Function Description
mean() Compute mean of groups
sum() Compute sum of group values
size() Compute group sizes
count() Compute count of group
std() Standard deviation of groups
var() Compute variance of groups
sem() Standard error of the mean of groups
describe() Generates descriptive statistics
first() Compute first of group values
last() Compute last of group values
nth() Take nth value, or a subset if n is a list
min() Compute min of group values
max() Compute max of group values

The aggregating functions above will exclude NA values. Any function which reduces a Series to a scalar value is
an aggregation function and will work, a trivial example is df.groupby('A').agg(lambda ser: 1). Note
that nth() can act as a reducer or a filter, see here.

Applying multiple functions at once

With grouped Series you can also pass a list or dict of functions to do aggregation with, outputting a DataFrame:
In [72]: grouped = df.groupby('A')

In [73]: grouped['C'].agg([np.sum, np.mean, np.std])

Out[73]:
sum mean std
A
bar 2.256744 0.752248 1.622939
foo 2.344678 0.468936 0.565255

On a grouped DataFrame, you can pass a list of functions to apply to each column, which produces an aggregated
result with a hierarchical index:
In [74]: grouped.agg([np.sum, np.mean, np.std])
Out[74]:
C D
sum mean std sum mean std
A
bar 2.256744 0.752248 1.622939 -0.790662 -0.263554 0.935025
foo 2.344678 0.468936 0.565255 4.727137 0.945427 1.326700

The resulting aggregations are named for the functions themselves. If you need to rename, then you can add in a
chained operation for a Series like this:
In [75]: (grouped['C'].agg([np.sum, np.mean, np.std])
....: .rename(columns={'sum': 'foo',
....: 'mean': 'bar',
....: 'std': 'baz'}))
....:
Out[75]:
foo bar baz
A
(continues on next page)

4.12. Group By: split-apply-combine 647

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

bar 2.256744 0.752248 1.622939
foo 2.344678 0.468936 0.565255

For a grouped DataFrame, you can rename in a similar manner:

In [76]: (grouped.agg([np.sum, np.mean, np.std])
....: .rename(columns={'sum': 'foo',
....: 'mean': 'bar',
....: 'std': 'baz'}))
....:
Out[76]:
C D
foo bar baz foo bar baz
A
bar 2.256744 0.752248 1.622939 -0.790662 -0.263554 0.935025
foo 2.344678 0.468936 0.565255 4.727137 0.945427 1.326700

Note: In general, the output column names should be unique. You cant apply the same function (or two functions
with the same name) to the same column.
In [77]: grouped['C'].agg(['sum', 'sum'])
---------------------------------------------------------------------------
SpecificationError Traceback (most recent call last)
<ipython-input-77-7be02859f395> in <module>
----> 1 grouped['C'].agg(['sum', 'sum'])

~/sandbox/pandas-release/pandas/pandas/core/groupby/generic.py in aggregate(self,
,→func_or_funcs, *args, **kwargs)

849 # but not the class list / tuple itself.

850 func_or_funcs = _maybe_mangle_lambdas(func_or_funcs)
--> 851 ret = self._aggregate_multiple_funcs(func_or_funcs, (_level or 0)
,→+ 1)

852 if relabeling:
853 ret.columns = columns

~/sandbox/pandas-release/pandas/pandas/core/groupby/generic.py in _aggregate_multiple_
,→funcs(self, arg, _level)

919 raise SpecificationError(

920 "Function names must be unique, found multiple named "
--> 921 "{}".format(name)
922 )
923

SpecificationError: Function names must be unique, found multiple named sum

Pandas does allow you to provide multiple lambdas. In this case, pandas will mangle the name of the (nameless)
lambda functions, appending _<i> to each subsequent lambda.
In [78]: grouped['C'].agg([lambda x: x.max() - x.min(),
....: lambda x: x.median() - x.mean()])
....:
Out[78]:
<lambda_0> <lambda_1>
A
bar 3.217186 -0.248622
foo 1.316392 -0.169326

648 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Named aggregation

New in version 0.25.0.

To support column-specific aggregation with control over the output column names, pandas accepts the special syntax
in GroupBy.agg(), known as named aggregation, where
• The keywords are the output column names
• The values are tuples whose first element is the column to select and the second element is the aggregation to
apply to that column. Pandas provides the pandas.NamedAgg namedtuple with the fields ['column',
'aggfunc'] to make it clearer what the arguments are. As usual, the aggregation can be a callable or a string
alias.
In [79]: animals = pd.DataFrame({'kind': ['cat', 'dog', 'cat', 'dog'],
....: 'height': [9.1, 6.0, 9.5, 34.0],
....: 'weight': [7.9, 7.5, 9.9, 198.0]})
....:

In [80]: animals
Out[80]:
kind height weight
0 cat 9.1 7.9
1 dog 6.0 7.5
2 cat 9.5 9.9
3 dog 34.0 198.0

In [81]: animals.groupby("kind").agg(
....: min_height=pd.NamedAgg(column='height', aggfunc='min'),
....: max_height=pd.NamedAgg(column='height', aggfunc='max'),
....: average_weight=pd.NamedAgg(column='weight', aggfunc=np.mean),
....: )
....:
Out[81]:
min_height max_height average_weight
kind
cat 9.1 9.5 8.90
dog 6.0 34.0 102.75
pandas.NamedAgg is just a namedtuple. Plain tuples are allowed as well.

In [82]: animals.groupby("kind").agg(
....: min_height=('height', 'min'),
....: max_height=('height', 'max'),
....: average_weight=('weight', np.mean),
....: )
....:
Out[82]:
min_height max_height average_weight
kind
cat 9.1 9.5 8.90
dog 6.0 34.0 102.75

If your desired output column names are not valid python keywords, construct a dictionary and unpack the keyword
arguments

4.12. Group By: split-apply-combine 649

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [83]: animals.groupby("kind").agg(**{
....: 'total weight': pd.NamedAgg(column='weight', aggfunc=sum),
....: })
....:
Out[83]:
total weight
kind
cat 17.8
dog 205.5

Additional keyword arguments are not passed through to the aggregation functions. Only pairs of (column,
aggfunc) should be passed as **kwargs. If your aggregation functions requires additional arguments, partially
apply them with functools.partial().

Note: For Python 3.5 and earlier, the order of **kwargs in a functions was not preserved. This means that the
output column ordering would not be consistent. To ensure consistent ordering, the keys (and so output columns) will
always be sorted for Python 3.5.

Named aggregation is also valid for Series groupby aggregations. In this case theres no column selection, so the values
are just the functions.

In [84]: animals.groupby("kind").height.agg(
....: min_height='min',
....: max_height='max',
....: )
....:
Out[84]:
min_height max_height
kind
cat 9.1 9.5
dog 6.0 34.0

Applying different functions to DataFrame columns

By passing a dict to aggregate you can apply a different aggregation to the columns of a DataFrame:

In [85]: grouped.agg({'C': np.sum,

....: 'D': lambda x: np.std(x, ddof=1)})
....:
Out[85]:
C D
A
bar 2.256744 0.935025
foo 2.344678 1.326700

The function names can also be strings. In order for a string to be valid it must be either implemented on GroupBy or
available via dispatching:

In [86]: grouped.agg({'C': 'sum', 'D': 'std'})

Out[86]:
C D
A
bar 2.256744 0.935025
foo 2.344678 1.326700

650 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Cython-optimized aggregation functions

Some common aggregations, currently only sum, mean, std, and sem, have optimized Cython implementations:
In [87]: df.groupby('A').sum()
Out[87]:
C D
A
bar 2.256744 -0.790662
foo 2.344678 4.727137

In [88]: df.groupby(['A', 'B']).mean()

Out[88]:
C D
A B
bar one -0.732034 0.531245
three 0.503626 -1.293798
two 2.485152 -0.028109
foo one 0.659257 1.467747
three 1.073480 1.869813
two -0.023658 -0.039085
Of course sum and mean are implemented on pandas objects, so the above code would work even without the special
versions via dispatching (see below).

4.12.5 Transformation

The transform method returns an object that is indexed the same (same size) as the one being grouped. The
transform function must:
• Return a result that is either the same size as the group chunk or broadcastable to the size of the group chunk
(e.g., a scalar, grouped.transform(lambda x: x.iloc[-1])).
• Operate column-by-column on the group chunk. The transform is applied to the first group chunk using
chunk.apply.
• Not perform in-place operations on the group chunk. Group chunks should be treated as immutable, and changes
to a group chunk may produce unexpected results. For example, when using fillna, inplace must be
False (grouped.transform(lambda x: x.fillna(inplace=False))).
• (Optionally) operates on the entire group chunk. If this is supported, a fast path is used starting from the second
chunk.
For example, suppose we wished to standardize the data within each group:
In [89]: index = pd.date_range('10/1/1999', periods=1100)

In [90]: ts = pd.Series(np.random.normal(0.5, 2, 1100), index)

In [91]: ts = ts.rolling(window=100, min_periods=100).mean().dropna()

In [92]: ts.head()
Out[92]:
2000-01-08 0.482962
2000-01-09 0.486928
2000-01-10 0.479684
2000-01-11 0.443526

4.12. Group By: split-apply-combine 651

pandas: powerful Python data analysis toolkit, Release 0.25.3

2000-01-12 0.462199
Freq: D, dtype: float64

In [93]: ts.tail()
Out[93]:
2002-09-30 0.712722
2002-10-01 0.717748
2002-10-02 0.710348
2002-10-03 0.712006
2002-10-04 0.688134
Freq: D, dtype: float64

In [94]: transformed = (ts.groupby(lambda x: x.year)

....: .transform(lambda x: (x - x.mean()) / x.std()))
....:
We would expect the result to now have mean 0 and standard deviation 1 within each group, which we can easily
check:
# Original Data
In [95]: grouped = ts.groupby(lambda x: x.year)

In [96]: grouped.mean()
Out[96]:
2000 0.597322
2001 0.677152
2002 0.705474
dtype: float64

In [97]: grouped.std()
Out[97]:
2000 0.171206
2001 0.137719
2002 0.125031
dtype: float64

# Transformed Data
In [98]: grouped_trans = transformed.groupby(lambda x: x.year)

In [99]: grouped_trans.mean()
Out[99]:
2000 4.221322e-15
2001 5.161168e-15
2002 1.169542e-15
dtype: float64

In [100]: grouped_trans.std()
Out[100]:
2000 1.0
2001 1.0
2002 1.0
dtype: float64
We can also visually compare the original and transformed data sets.

652 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [101]: compare = pd.DataFrame({'Original': ts, 'Transformed': transformed})

In [102]: compare.plot()
Out[102]: <matplotlib.axes._subplots.AxesSubplot at 0x1a312e2650>

Transformation functions that have lower dimension outputs are broadcast to match the shape of the input array.

In [103]: ts.groupby(lambda x: x.year).transform(lambda x: x.max() - x.min())

Out[103]:
2000-01-08 0.706553
2000-01-09 0.706553
2000-01-10 0.706553
2000-01-11 0.706553
2000-01-12 0.706553
...
2002-09-30 0.528844
2002-10-01 0.528844
2002-10-02 0.528844
2002-10-03 0.528844
2002-10-04 0.528844
Freq: D, Length: 1001, dtype: float64

Alternatively, the built-in methods could be used to produce the same outputs.

4.12. Group By: split-apply-combine 653

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [104]: max = ts.groupby(lambda x: x.year).transform('max')

In [105]: min = ts.groupby(lambda x: x.year).transform('min')

In [106]: max - min

Out[106]:
2000-01-08 0.706553
2000-01-09 0.706553
2000-01-10 0.706553
2000-01-11 0.706553
2000-01-12 0.706553
...
2002-09-30 0.528844
2002-10-01 0.528844
2002-10-02 0.528844
2002-10-03 0.528844
2002-10-04 0.528844
Freq: D, Length: 1001, dtype: float64

Another common data transform is to replace missing data with the group mean.

In [107]: data_df
Out[107]:
A B C
0 0.869121 1.473802 0.787744
1 0.464706 1.323886 -1.113432
2 -0.777595 0.143230 -0.718546
3 -0.501424 -2.031157 NaN
4 -0.489464 0.187423 -1.036766
.. ... ... ...
995 1.742327 -0.883907 -0.848783
996 0.168746 -1.180773 NaN
997 NaN -0.678201 0.080136
998 -1.389014 -0.372892 0.572905
999 -0.858635 -1.784332 NaN

[1000 rows x 3 columns]

In [108]: countries = np.array(['US', 'UK', 'GR', 'JP'])

In [109]: key = countries[np.random.randint(0, 4, 1000)]

In [110]: grouped = data_df.groupby(key)

# Non-NA count in each group

In [111]: grouped.count()
Out[111]:
A B C
GR 244 256 212
JP 213 223 201
UK 224 242 211
US 221 231 199

In [112]: transformed = grouped.transform(lambda x: x.fillna(x.mean()))

We can verify that the group means have not changed in the transformed data and that the transformed data contains
no NAs.

654 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [113]: grouped_trans = transformed.groupby(key)

In [114]: grouped.mean() # original group means

Out[114]:
A B C
GR -0.058842 0.078292 -0.081559
JP 0.108123 -0.038631 0.087081
UK 0.130078 0.067902 -0.076353
US 0.002237 0.048024 0.092084

In [115]: grouped_trans.mean() # transformation did not change group means

Out[115]:
A B C
GR -0.058842 0.078292 -0.081559
JP 0.108123 -0.038631 0.087081
UK 0.130078 0.067902 -0.076353
US 0.002237 0.048024 0.092084

In [116]: grouped.count() # original has some missing data points

Out[116]:
A B C
GR 244 256 212
JP 213 223 201
UK 224 242 211
US 221 231 199

In [117]: grouped_trans.count() # counts after transformation

Out[117]:
A B C
GR 266 266 266
JP 235 235 235
UK 252 252 252
US 247 247 247

In [118]: grouped_trans.size() # Verify non-NA count equals group size

Out[118]:
GR 266
JP 235
UK 252
US 247
dtype: int64

Note: Some functions will automatically transform the input when applied to a GroupBy object, but returning an
object of the same shape as the original. Passing as_index=False will not affect these transformation methods.
For example: fillna, ffill, bfill, shift..

In [119]: grouped.ffill()
Out[119]:
A B C
0 0.869121 1.473802 0.787744
1 0.464706 1.323886 -1.113432
2 -0.777595 0.143230 -0.718546
3 -0.501424 -2.031157 -0.718546
(continues on next page)

4.12. Group By: split-apply-combine 655

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

4 -0.489464 0.187423 -1.036766
.. ... ... ...
995 1.742327 -0.883907 -0.848783
996 0.168746 -1.180773 -0.838714
997 -1.159006 -0.678201 0.080136
998 -1.389014 -0.372892 0.572905
999 -0.858635 -1.784332 -0.848783

[1000 rows x 3 columns]

New syntax to window and resample operations

New in version 0.18.1.

Working with the resample, expanding or rolling operations on the groupby level used to require the application of
helper functions. However, now it is possible to use resample(), expanding() and rolling() as methods
on groupbys.
The example below will apply the rolling() method on the samples of the column B based on the groups of
column A.
In [120]: df_re = pd.DataFrame({'A': [1] * 10 + [5] * 10,
.....: 'B': np.arange(20)})
.....:

In [121]: df_re
Out[121]:
A B
0 1 0
1 1 1
2 1 2
3 1 3
4 1 4
5 1 5
6 1 6
7 1 7
8 1 8
9 1 9
10 5 10
11 5 11
12 5 12
13 5 13
14 5 14
15 5 15
16 5 16
17 5 17
18 5 18
19 5 19

In [122]: df_re.groupby('A').rolling(4).B.mean()
Out[122]:
A
1 0 NaN

656 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

1 NaN
2 NaN
3 1.5
4 2.5
5 3.5
6 4.5
7 5.5
8 6.5
9 7.5
5 10 NaN
11 NaN
12 NaN
13 11.5
14 12.5
15 13.5
16 14.5
17 15.5
18 16.5
19 17.5
Name: B, dtype: float64
The expanding() method will accumulate a given operation (sum() in the example) for all the members of each
particular group.

In [123]: df_re.groupby('A').expanding().sum()
Out[123]:
A B
A
1 0 1.0 0.0
1 2.0 1.0
2 3.0 3.0
3 4.0 6.0
4 5.0 10.0
5 6.0 15.0
6 7.0 21.0
7 8.0 28.0
8 9.0 36.0
9 10.0 45.0
5 10 5.0 10.0
11 10.0 21.0
12 15.0 33.0
13 20.0 46.0
14 25.0 60.0
15 30.0 75.0
16 35.0 91.0
17 40.0 108.0
18 45.0 126.0
19 50.0 145.0

Suppose you want to use the resample() method to get a daily frequency in each group of your dataframe and wish
to complete the missing values with the ffill() method.
In [124]: df_re = pd.DataFrame({'date': pd.date_range(start='2016-01-01',
,→periods=4,

.....: freq='W'),
.....: 'group': [1, 1, 2, 2],
.....: 'val': [5, 6, 7, 8]}).set_index('date')

4.12. Group By: split-apply-combine 657

pandas: powerful Python data analysis toolkit, Release 0.25.3

.....:

In [125]: df_re
Out[125]:
group val
date
2016-01-03 1 5
2016-01-10 1 6
2016-01-17 2 7
2016-01-24 2 8

In [126]: df_re.groupby('group').resample('1D').ffill()
Out[126]:
group val
group date
1 2016-01-03 1 5
2016-01-04 1 5
2016-01-05 1 5
2016-01-06 1 5
2016-01-07 1 5
2016-01-08 1 5
2016-01-09 1 5
2016-01-10 1 6
2 2016-01-17 2 7
2016-01-18 2 7
2016-01-19 2 7
2016-01-20 2 7
2016-01-21 2 7
2016-01-22 2 7
2016-01-23 2 7
2016-01-24 2 8

4.12.6 Filtration

The filter method returns a subset of the original object. Suppose we want to take only elements that belong to
groups with a group sum greater than 2.

In [127]: sf = pd.Series([1, 1, 2, 3, 3, 3])

In [128]: sf.groupby(sf).filter(lambda x: x.sum() > 2)

Out[128]:
3 3
4 3
5 3
dtype: int64

The argument of filter must be a function that, applied to the group as a whole, returns True or False.
Another useful operation is filtering out elements that belong to groups with only a couple members.

In [129]: dff = pd.DataFrame({'A': np.arange(8), 'B': list('aabbbbcc')})

In [130]: dff.groupby('B').filter(lambda x: len(x) > 2)

Out[130]:
(continues on next page)

658 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

A B
2 2 b
3 3 b
4 4 b
5 5 b

Alternatively, instead of dropping the offending groups, we can return a like-indexed objects where the groups that do
not pass the filter are filled with NaNs.

In [131]: dff.groupby('B').filter(lambda x: len(x) > 2, dropna=False)

Out[131]:
A B
0 NaN NaN
1 NaN NaN
2 2.0 b
3 3.0 b
4 4.0 b
5 5.0 b
6 NaN NaN
7 NaN NaN

For DataFrames with multiple columns, filters should explicitly specify a column as the filter criterion.

In [132]: dff['C'] = np.arange(8)

In [133]: dff.groupby('B').filter(lambda x: len(x['C']) > 2)

Out[133]:
A B C
2 2 b 2
3 3 b 3
4 4 b 4
5 5 b 5

Note: Some functions when applied to a groupby object will act as a filter on the input, returning a reduced shape of
the original (and potentially eliminating groups), but with the index unchanged. Passing as_index=False will not
affect these transformation methods.
For example: head, tail.

In [134]: dff.groupby('B').head(2)
Out[134]:
A B C
0 0 a 0
1 1 a 1
2 2 b 2
3 3 b 3
6 6 c 6
7 7 c 7

4.12.7 Dispatching to instance methods

When doing an aggregation or transformation, you might just want to call an instance method on each data group.
This is pretty easy to do by passing lambda functions:

4.12. Group By: split-apply-combine 659

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [135]: grouped = df.groupby('A')

In [136]: grouped.agg(lambda x: x.std())

Out[136]:
C D
A
bar 1.622939 0.935025
foo 0.565255 1.326700

But, its rather verbose and can be untidy if you need to pass additional arguments. Using a bit of metaprogramming
cleverness, GroupBy now has the ability to dispatch method calls to the groups:

In [137]: grouped.std()
Out[137]:
C D
A
bar 1.622939 0.935025
foo 0.565255 1.326700

What is actually happening here is that a function wrapper is being generated. When invoked, it takes any passed
arguments and invokes the function with any arguments on each group (in the above example, the std function). The
results are then combined together much in the style of agg and transform (it actually uses apply to infer the
gluing, documented next). This enables some operations to be carried out rather succinctly:

In [138]: tsdf = pd.DataFrame(np.random.randn(1000, 3),

.....: index=pd.date_range('1/1/2000', periods=1000),
.....: columns=['A', 'B', 'C'])
.....:

In [139]: tsdf.iloc[::2] = np.nan

In [140]: grouped = tsdf.groupby(lambda x: x.year)

In [141]: grouped.fillna(method='pad')
Out[141]:
A B C
2000-01-01 NaN NaN NaN
2000-01-02 -0.183449 -0.611494 -0.178486
2000-01-03 -0.183449 -0.611494 -0.178486
2000-01-04 -2.105115 -1.569569 0.553558
2000-01-05 -2.105115 -1.569569 0.553558
... ... ... ...
2002-09-22 -1.138283 1.475569 -0.063934
2002-09-23 -1.138283 1.475569 -0.063934
2002-09-24 0.069982 1.189623 -0.212112
2002-09-25 0.069982 1.189623 -0.212112
2002-09-26 -1.601811 0.215340 1.432049

[1000 rows x 3 columns]

In this example, we chopped the collection of time series into yearly chunks then independently called fillna on the
groups.
The nlargest and nsmallest methods work on Series style groupbys:
In [142]: s = pd.Series([9, 8, 7, 5, 19, 1, 4.2, 3.3])

In [143]: g = pd.Series(list('abababab'))

660 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [144]: gb = s.groupby(g)

In [145]: gb.nlargest(3)
Out[145]:
a 4 19.0
0 9.0
2 7.0
b 1 8.0
3 5.0
7 3.3
dtype: float64

In [146]: gb.nsmallest(3)
Out[146]:
a 6 4.2
2 7.0
0 9.0
b 5 1.0
7 3.3
3 5.0
dtype: float64

4.12.8 Flexible apply

Some operations on the grouped data might not fit into either the aggregate or transform categories. Or, you may simply
want GroupBy to infer how to combine the results. For these, use the apply function, which can be substituted for
both aggregate and transform in many standard use cases. However, apply can handle some exceptional use
cases, for example:
In [147]: df
Out[147]:
A B C D
0 foo one 1.018905 1.623114
1 bar one -0.732034 0.531245
2 foo two -0.242912 -1.390677
3 bar three 0.503626 -1.293798
4 foo two 0.195596 1.312507
5 bar two 2.485152 -0.028109
6 foo one 0.299609 1.312379
7 foo three 1.073480 1.869813

In [148]: grouped = df.groupby('A')

# could also just call .describe()

In [149]: grouped['C'].apply(lambda x: x.describe())
Out[149]:
A
bar count 3.000000
mean 0.752248
std 1.622939
min -0.732034
25% -0.114204
50% 0.503626
75% 1.494389
(continues on next page)

4.12. Group By: split-apply-combine 661

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

max 2.485152
foo count 5.000000
mean 0.468936
std 0.565255
min -0.242912
25% 0.195596
50% 0.299609
75% 1.018905
max 1.073480
Name: C, dtype: float64

The dimension of the returned result can also change:

In [150]: grouped = df.groupby('A')['C']

In [151]: def f(group):

.....: return pd.DataFrame({'original': group,
.....: 'demeaned': group - group.mean()})
.....:

In [152]: grouped.apply(f)
Out[152]:
original demeaned
0 1.018905 0.549970
1 -0.732034 -1.484282
2 -0.242912 -0.711848
3 0.503626 -0.248622
4 0.195596 -0.273340
5 2.485152 1.732904
6 0.299609 -0.169326
7 1.073480 0.604544

apply on a Series can operate on a returned value from the applied function, that is itself a series, and possibly upcast
the result to a DataFrame:
In [153]: def f(x):
.....: return pd.Series([x, x ** 2], index=['x', 'x^2'])
.....:

In [154]: s = pd.Series(np.random.rand(5))

In [155]: s
Out[155]:
0 0.656427
1 0.691656
2 0.352607
3 0.059733
4 0.451923
dtype: float64

In [156]: s.apply(f)
Out[156]:
x x^2
0 0.656427 0.430896
1 0.691656 0.478388
2 0.352607 0.124332

662 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

3 0.059733 0.003568
4 0.451923 0.204234

Note: apply can act as a reducer, transformer, or filter function, depending on exactly what is passed to it. So
depending on the path taken, and exactly what you are grouping. Thus the grouped columns(s) may be included in the
output as well as set the indices.

4.12.9 Other useful features

Automatic exclusion of nuisance columns

Again consider the example DataFrame weve been looking at:

In [157]: df
Out[157]:
A B C D
0 foo one 1.018905 1.623114
1 bar one -0.732034 0.531245
2 foo two -0.242912 -1.390677
3 bar three 0.503626 -1.293798
4 foo two 0.195596 1.312507
5 bar two 2.485152 -0.028109
6 foo one 0.299609 1.312379
7 foo three 1.073480 1.869813

Suppose we wish to compute the standard deviation grouped by the A column. There is a slight problem, namely that
we dont care about the data in column B. We refer to this as a nuisance column. If the passed aggregation function
cant be applied to some columns, the troublesome columns will be (silently) dropped. Thus, this does not pose any
problems:

In [158]: df.groupby('A').std()
Out[158]:
C D
A
bar 1.622939 0.935025
foo 0.565255 1.326700

Note that df.groupby('A').colname.std(). is more efficient than df.groupby('A').std().

colname, so if the result of an aggregation function is only interesting over one column (here colname), it may be
filtered before applying the aggregation function.

Note: Any object column, also if it contains numerical values such as Decimal objects, is considered as a nuisance
columns. They are excluded from aggregate functions automatically in groupby.
If you do wish to include decimal or object columns in an aggregation with other non-nuisance data types, you must
do so explicitly.

In [159]: from decimal import Decimal

In [160]: df_dec = pd.DataFrame(

.....: {'id': [1, 2, 1, 2],
.....: 'int_column': [1, 2, 3, 4],
.....: 'dec_column': [Decimal('0.50'), Decimal('0.15'),

4.12. Group By: split-apply-combine 663

pandas: powerful Python data analysis toolkit, Release 0.25.3

.....: Decimal('0.25'), Decimal('0.40')]

.....: }
.....: )
.....:

# Decimal columns can be sum'd explicitly by themselves...

In [161]: df_dec.groupby(['id'])[['dec_column']].sum()
Out[161]:
dec_column
id
1 0.75
2 0.55

# ...but cannot be combined with standard data types or they will be excluded
In [162]: df_dec.groupby(['id'])[['int_column', 'dec_column']].sum()
Out[162]:
int_column
id
1 4
2 6

# Use .agg function to aggregate over standard and "nuisance" data types
# at the same time
In [163]: df_dec.groupby(['id']).agg({'int_column': 'sum', 'dec_column':
,→'sum'})

Out[163]:
int_column dec_column
id
1 4 0.75
2 6 0.55

Handling of (un)observed Categorical values

When using a Categorical grouper (as a single grouper, or as part of multiple groupers), the observed keyword
controls whether to return a cartesian product of all possible groupers values (observed=False) or only those that
are observed groupers (observed=True).
Show all values:

In [164]: pd.Series([1, 1, 1]).groupby(pd.Categorical(['a', 'a', 'a'],

.....: categories=['a', 'b']),
.....: observed=False).count()
.....:
Out[164]:
a 3
b 0
dtype: int64

Show only the observed values:

In [165]: pd.Series([1, 1, 1]).groupby(pd.Categorical(['a', 'a', 'a'],

.....: categories=['a', 'b']),
.....: observed=True).count()
.....:
(continues on next page)

664 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

Out[165]:
a 3
dtype: int64

The returned dtype of the grouped will always include all of the categories that were grouped.

In [166]: s = pd.Series([1, 1, 1]).groupby(pd.Categorical(['a', 'a', 'a'],

.....: categories=['a', 'b']),
.....: observed=False).count()
.....:

In [167]: s.index.dtype
Out[167]: CategoricalDtype(categories=['a', 'b'], ordered=False)

NA and NaT group handling

If there are any NaN or NaT values in the grouping key, these will be automatically excluded. In other words, there
will never be an NA group or NaT group. This was not the case in older versions of pandas, but users were generally
discarding the NA group anyway (and supporting it was an implementation headache).

Grouping with ordered factors

Categorical variables represented as instance of pandass Categorical class can be used as group keys. If so, the
order of the levels will be preserved:

In [168]: data = pd.Series(np.random.randn(100))

In [169]: factor = pd.qcut(data, [0, .25, .5, .75, 1.])

In [170]: data.groupby(factor).mean()
Out[170]:
(-3.221, -0.48] -1.238032
(-0.48, 0.119] -0.216324
(0.119, 0.702] 0.381676
(0.702, 2.538] 1.216886
dtype: float64

Grouping with a grouper specification

You may need to specify a bit more data to properly group. You can use the pd.Grouper to provide this local
control.

In [171]: import datetime

In [172]: df = pd.DataFrame({'Branch': 'A A A A A A A B'.split(),

.....: 'Buyer': 'Carl Mark Carl Carl Joe Joe Joe Carl'.split(),
.....: 'Quantity': [1, 3, 5, 1, 8, 1, 9, 3],
.....: 'Date': [
.....: datetime.datetime(2013, 1, 1, 13, 0),
.....: datetime.datetime(2013, 1, 1, 13, 5),
.....: datetime.datetime(2013, 10, 1, 20, 0),
.....: datetime.datetime(2013, 10, 2, 10, 0),
(continues on next page)

4.12. Group By: split-apply-combine 665

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

.....: datetime.datetime(2013, 10, 1, 20, 0),
.....: datetime.datetime(2013, 10, 2, 10, 0),
.....: datetime.datetime(2013, 12, 2, 12, 0),
.....: datetime.datetime(2013, 12, 2, 14, 0)]
.....: })
.....:

In [173]: df
Out[173]:
Branch Buyer Quantity Date
0 A Carl 1 2013-01-01 13:00:00
1 A Mark 3 2013-01-01 13:05:00
2 A Carl 5 2013-10-01 20:00:00
3 A Carl 1 2013-10-02 10:00:00
4 A Joe 8 2013-10-01 20:00:00
5 A Joe 1 2013-10-02 10:00:00
6 A Joe 9 2013-12-02 12:00:00
7 B Carl 3 2013-12-02 14:00:00

Groupby a specific column with the desired frequency. This is like resampling.
In [174]: df.groupby([pd.Grouper(freq='1M', key='Date'), 'Buyer']).sum()
Out[174]:
Quantity
Date Buyer
2013-01-31 Carl 1
Mark 3
2013-10-31 Carl 6
Joe 9
2013-12-31 Carl 3
Joe 9

You have an ambiguous specification in that you have a named index and a column that could be potential groupers.
In [175]: df = df.set_index('Date')

In [176]: df['Date'] = df.index + pd.offsets.MonthEnd(2)

In [177]: df.groupby([pd.Grouper(freq='6M', key='Date'), 'Buyer']).sum()

Out[177]:
Quantity
Date Buyer
2013-02-28 Carl 1
Mark 3
2014-02-28 Carl 9
Joe 18

In [178]: df.groupby([pd.Grouper(freq='6M', level='Date'), 'Buyer']).sum()

Out[178]:
Quantity
Date Buyer
2013-01-31 Carl 1
Mark 3
2014-01-31 Carl 9
Joe 18

666 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Taking the first rows of each group

Just like for a DataFrame or Series you can call head and tail on a groupby:
In [179]: df = pd.DataFrame([[1, 2], [1, 4], [5, 6]], columns=['A', 'B'])

In [180]: df
Out[180]:
A B
0 1 2
1 1 4
2 5 6

In [181]: g = df.groupby('A')

In [182]: g.head(1)
Out[182]:
A B
0 1 2
2 5 6

In [183]: g.tail(1)
Out[183]:
A B
1 1 4
2 5 6
This shows the first or last n rows from each group.

Taking the nth row of each group

To select from a DataFrame or Series the nth item, use nth(). This is a reduction method, and will return a single
row (or no row) per group if you pass an int for n:
In [184]: df = pd.DataFrame([[1, np.nan], [1, 4], [5, 6]], columns=['A', 'B'])

In [185]: g = df.groupby('A')

In [186]: g.nth(0)
Out[186]:
B
A
1 NaN
5 6.0

In [187]: g.nth(-1)
Out[187]:
B
A
1 4.0
5 6.0

In [188]: g.nth(1)
Out[188]:
B

4.12. Group By: split-apply-combine 667

pandas: powerful Python data analysis toolkit, Release 0.25.3

A
1 4.0
If you want to select the nth not-null item, use the dropna kwarg. For a DataFrame this should be either 'any' or
'all' just like you would pass to dropna:
# nth(0) is the same as g.first()
In [189]: g.nth(0, dropna='any')
Out[189]:
B
A
1 4.0
5 6.0

In [190]: g.first()
Out[190]:
B
A
1 4.0
5 6.0

# nth(-1) is the same as g.last()

In [191]: g.nth(-1, dropna='any') # NaNs denote group exhausted when using
,→dropna

Out[191]:
B
A
1 4.0
5 6.0

In [192]: g.last()
Out[192]:
B
A
1 4.0
5 6.0

In [193]: g.B.nth(0, dropna='all')

Out[193]:
A
1 4.0
5 6.0
Name: B, dtype: float64
As with other methods, passing as_index=False, will achieve a filtration, which returns the grouped row.
In [194]: df = pd.DataFrame([[1, np.nan], [1, 4], [5, 6]], columns=['A', 'B'])

In [195]: g = df.groupby('A', as_index=False)

In [196]: g.nth(0)
Out[196]:
A B
0 1 NaN
2 5 6.0

668 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [197]: g.nth(-1)
Out[197]:
A B
1 1 4.0
2 5 6.0
You can also select multiple rows from each group by specifying multiple nth values as a list of ints.

In [198]: business_dates = pd.date_range(start='4/1/2014', end='6/30/2014', freq='B')

In [199]: df = pd.DataFrame(1, index=business_dates, columns=['a', 'b'])

# get the first, 4th, and last date index for each month
In [200]: df.groupby([df.index.year, df.index.month]).nth([0, 3, -1])
Out[200]:
a b
2014 4 1 1
4 1 1
4 1 1
5 1 1
5 1 1
5 1 1
6 1 1
6 1 1
6 1 1

Enumerate group items

To see the order in which each row appears within its group, use the cumcount method:
In [201]: dfg = pd.DataFrame(list('aaabba'), columns=['A'])

In [202]: dfg
Out[202]:
A
0 a
1 a
2 a
3 b
4 b
5 a

In [203]: dfg.groupby('A').cumcount()
Out[203]:
0 0
1 1
2 2
3 0
4 1
5 3
dtype: int64

In [204]: dfg.groupby('A').cumcount(ascending=False)
Out[204]:
0 3

4.12. Group By: split-apply-combine 669

pandas: powerful Python data analysis toolkit, Release 0.25.3

1 2
2 1
3 1
4 0
5 0
dtype: int64

Enumerate groups

New in version 0.20.2.

To see the ordering of the groups (as opposed to the order of rows within a group given by cumcount) you can use
ngroup().
Note that the numbers given to the groups match the order in which the groups would be seen when iterating over the
groupby object, not the order they are first observed.
In [205]: dfg = pd.DataFrame(list('aaabba'), columns=['A'])

In [206]: dfg
Out[206]:
A
0 a
1 a
2 a
3 b
4 b
5 a

In [207]: dfg.groupby('A').ngroup()
Out[207]:
0 0
1 0
2 0
3 1
4 1
5 0
dtype: int64

In [208]: dfg.groupby('A').ngroup(ascending=False)
Out[208]:
0 1
1 1
2 1
3 0
4 0
5 1
dtype: int64

Plotting

Groupby also works with some plotting methods. For example, suppose we suspect that some features in a DataFrame
may differ by group, in this case, the values in column 1 where the group is B are 3 higher on average.

670 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [209]: np.random.seed(1234)

In [210]: df = pd.DataFrame(np.random.randn(50, 2))

In [211]: df['g'] = np.random.choice(['A', 'B'], size=50)

In [212]: df.loc[df['g'] == 'B', 1] += 3

We can easily visualize this with a boxplot:

In [213]: df.groupby('g').boxplot()
Out[213]:
A AxesSubplot(0.1,0.15;0.363636x0.75)
B AxesSubplot(0.536364,0.15;0.363636x0.75)
dtype: object

The result of calling boxplot is a dictionary whose keys are the values of our grouping column g (A and B).
The values of the resulting dictionary can be controlled by the return_type keyword of boxplot. See the
visualization documentation for more.

Warning: For historical reasons, df.groupby("g").boxplot() is not equivalent to df.

boxplot(by="g"). See here for an explanation.

4.12. Group By: split-apply-combine 671

pandas: powerful Python data analysis toolkit, Release 0.25.3

Piping function calls

New in version 0.21.0.

Similar to the functionality provided by DataFrame and Series, functions that take GroupBy objects can be
chained together using a pipe method to allow for a cleaner, more readable syntax. To read about .pipe in general
terms, see here.
Combining .groupby and .pipe is often useful when you need to reuse GroupBy objects.
As an example, imagine having a DataFrame with columns for stores, products, revenue and quantity sold. Wed like to
do a groupwise calculation of prices (i.e. revenue/quantity) per store and per product. We could do this in a multi-step
operation, but expressing it in terms of piping can make the code more readable. First we set the data:

In [214]: n = 1000

In [215]: df = pd.DataFrame({'Store': np.random.choice(['Store_1', 'Store_2'], n),

.....: 'Product': np.random.choice(['Product_1',
.....: 'Product_2'], n),
.....: 'Revenue': (np.random.random(n) * 50 + 10).round(2),
.....: 'Quantity': np.random.randint(1, 10, size=n)})
.....:

In [216]: df.head(2)
Out[216]:
Store Product Revenue Quantity
0 Store_2 Product_1 26.12 1
1 Store_2 Product_1 28.86 1

Now, to find prices per store/product, we can simply do:

In [217]: (df.groupby(['Store', 'Product'])

.....: .pipe(lambda grp: grp.Revenue.sum() / grp.Quantity.sum())
.....: .unstack().round(2))
.....:
Out[217]:
Product Product_1 Product_2
Store
Store_1 6.82 7.05
Store_2 6.30 6.64

Piping can also be expressive when you want to deliver a grouped object to some arbitrary function, for example:

In [218]: def mean(groupby):

.....: return groupby.mean()
.....:

In [219]: df.groupby(['Store', 'Product']).pipe(mean)

Out[219]:
Revenue Quantity
Store Product
Store_1 Product_1 34.622727 5.075758
Product_2 35.482815 5.029630
Store_2 Product_1 32.972837 5.237589
Product_2 34.684360 5.224000

where mean takes a GroupBy object and finds the mean of the Revenue and Quantity columns respectively for each
Store-Product combination. The mean function can be any function that takes in a GroupBy object; the .pipe will
pass the GroupBy object as a parameter into the function you specify.

672 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

4.12.10 Examples

Regrouping by factor

Regroup columns of a DataFrame according to their sum, and sum the aggregated ones.
In [220]: df = pd.DataFrame({'a': [1, 0, 0], 'b': [0, 1, 0],
.....: 'c': [1, 0, 0], 'd': [2, 3, 4]})
.....:

In [221]: df
Out[221]:
a b c d
0 1 0 1 2
1 0 1 0 3
2 0 0 0 4

In [222]: df.groupby(df.sum(), axis=1).sum()

Out[222]:
1 9
0 2 2
1 1 3
2 0 4

Multi-column factorization

By using ngroup(), we can extract information about the groups in a way similar to factorize() (as described
further in the reshaping API) but which applies naturally to multiple columns of mixed type and different sources.
This can be useful as an intermediate categorical-like step in processing, when the relationships between the group
rows are more important than their content, or as input to an algorithm which only accepts the integer encoding.
(For more information about support in pandas for full categorical data, see the Categorical introduction and the API
documentation.)
In [223]: dfg = pd.DataFrame({"A": [1, 1, 2, 3, 2], "B": list("aaaba")})

In [224]: dfg
Out[224]:
A B
0 1 a
1 1 a
2 2 a
3 3 b
4 2 a

In [225]: dfg.groupby(["A", "B"]).ngroup()

Out[225]:
0 0
1 0
2 1
3 2
4 1
dtype: int64

In [226]: dfg.groupby(["A", [0, 0, 0, 1, 1]]).ngroup()

4.12. Group By: split-apply-combine 673

pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[226]:
0 0
1 0
2 1
3 3
4 2
dtype: int64

Groupby by indexer to resample data

Resampling produces new hypothetical samples (resamples) from already existing observed data or from a model that
generates data. These new samples are similar to the pre-existing samples.
In order to resample to work on indices that are non-datetimelike, the following procedure can be utilized.
In the following examples, df.index // 5 returns a binary array which is used to determine what gets selected for the
groupby operation.

Note: The below example shows how we can downsample by consolidation of samples into fewer samples. Here by
using df.index // 5, we are aggregating the samples in bins. By applying std() function, we aggregate the information
contained in many samples into a small subset of values which is their standard deviation thereby reducing the number
of samples.

In [227]: df = pd.DataFrame(np.random.randn(10, 2))

In [228]: df
Out[228]:
0 1
0 -0.793893 0.321153
1 0.342250 1.618906
2 -0.975807 1.918201
3 -0.810847 -1.405919
4 -1.977759 0.461659
5 0.730057 -1.316938
6 -0.751328 0.528290
7 -0.257759 -1.081009
8 0.505895 -1.701948
9 -1.006349 0.020208

In [229]: df.index // 5
Out[229]: Int64Index([0, 0, 0, 0, 0, 1, 1, 1, 1, 1], dtype='int64')

In [230]: df.groupby(df.index // 5).std()

Out[230]:
0 1
0 0.823647 1.312912
1 0.760109 0.942941

Returning a Series to propagate names

Group DataFrame columns, compute a set of metrics and return a named Series. The Series name is used as the name
for the column index. This is especially useful in conjunction with reshaping operations such as stacking in which the
column index name will be used as the name of the inserted column:

674 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

In [231]: df = pd.DataFrame({'a': [0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2],

.....: 'b': [0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1],
.....: 'c': [1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0],
.....: 'd': [0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1]})
.....:

In [232]: def compute_metrics(x):

.....: result = {'b_sum': x['b'].sum(), 'c_mean': x['c'].mean()}
.....: return pd.Series(result, name='metrics')
.....:

In [233]: result = df.groupby('a').apply(compute_metrics)

In [234]: result
Out[234]:
metrics b_sum c_mean
a
0 2.0 0.5
1 2.0 0.5
2 2.0 0.5

In [235]: result.stack()
Out[235]:
a metrics
0 b_sum 2.0
c_mean 0.5
1 b_sum 2.0
c_mean 0.5
2 b_sum 2.0
c_mean 0.5
dtype: float64
{{ header }}

4.13 Time series / date functionality

pandas contains extensive capabilities and features for working with time series data for all domains. Using the
NumPy datetime64 and timedelta64 dtypes, pandas has consolidated a large number of features from other
Python libraries like scikits.timeseries as well as created a tremendous amount of new functionality for
manipulating time series data.
For example, pandas supports:
Parsing time series information from various sources and formats

In [1]: import datetime

In [2]: dti = pd.to_datetime(['1/1/2018', np.datetime64('2018-01-01'),

...: datetime.datetime(2018, 1, 1)])
...:

In [3]: dti
Out[3]: DatetimeIndex(['2018-01-01', '2018-01-01', '2018-01-01'], dtype=
,→'datetime64[ns]', freq=None)

4.13. Time series / date functionality 675

pandas: powerful Python data analysis toolkit, Release 0.25.3

Generate sequences of fixed-frequency dates and time spans

In [4]: dti = pd.date_range('2018-01-01', periods=3, freq='H')

In [5]: dti
Out[5]:
DatetimeIndex(['2018-01-01 00:00:00', '2018-01-01 01:00:00',
'2018-01-01 02:00:00'],
dtype='datetime64[ns]', freq='H')

Manipulating and converting date times with timezone information

In [6]: dti = dti.tz_localize('UTC')

In [7]: dti
Out[7]:
DatetimeIndex(['2018-01-01 00:00:00+00:00', '2018-01-01 01:00:00+00:00',
'2018-01-01 02:00:00+00:00'],
dtype='datetime64[ns, UTC]', freq='H')

In [8]: dti.tz_convert('US/Pacific')
Out[8]:
DatetimeIndex(['2017-12-31 16:00:00-08:00', '2017-12-31 17:00:00-08:00',
'2017-12-31 18:00:00-08:00'],
dtype='datetime64[ns, US/Pacific]', freq='H')
Resampling or converting a time series to a particular frequency
In [9]: idx = pd.date_range('2018-01-01', periods=5, freq='H')

In [10]: ts = pd.Series(range(len(idx)), index=idx)

In [11]: ts
Out[11]:
2018-01-01 00:00:00 0
2018-01-01 01:00:00 1
2018-01-01 02:00:00 2
2018-01-01 03:00:00 3
2018-01-01 04:00:00 4
Freq: H, dtype: int64

In [12]: ts.resample('2H').mean()
Out[12]:
2018-01-01 00:00:00 0.5
2018-01-01 02:00:00 2.5
2018-01-01 04:00:00 4.0
Freq: 2H, dtype: float64
Performing date and time arithmetic with absolute or relative time increments
In [13]: friday = pd.Timestamp('2018-01-05')

In [14]: friday.day_name()
Out[14]: 'Friday'

# Add 1 day
In [15]: saturday = friday + pd.Timedelta('1 day')
(continues on next page)

676 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [16]: saturday.day_name()
Out[16]: 'Saturday'

# Add 1 business day (Friday --> Monday)

In [17]: monday = friday + pd.offsets.BDay()

In [18]: monday.day_name()
Out[18]: 'Monday'

pandas provides a relatively compact and self-contained set of tools for performing the above tasks and more.

4.13.1 Overview

pandas captures 4 general time related concepts:

1. Date times: A specific date and time with timezone support. Similar to datetime.datetime from the
standard library.
2. Time deltas: An absolute time duration. Similar to datetime.timedelta from the standard library.
3. Time spans: A span of time defined by a point in time and its associated frequency.
4. Date offsets: A relative time duration that respects calendar arithmetic. Similar to dateutil.
relativedelta.relativedelta from the dateutil package.

Con- Scalar Array Class pandas Data Type Primary Creation Method
cept Class
Date Timestamp DatetimeIndex
datetime64[ns] or to_datetime or
times datetime64[ns, tz] date_range
Time Timedelta TimedeltaIndex
timedelta64[ns] to_timedelta or
deltas timedelta_range
Time Period PeriodIndex period[freq] Period or period_range
spans
Date DateOffsetNone None DateOffset
offsets

For time series data, its conventional to represent the time component in the index of a Series or DataFrame so
manipulations can be performed with respect to the time element.

In [19]: pd.Series(range(3), index=pd.date_range('2000', freq='D', periods=3))

Out[19]:
2000-01-01 0
2000-01-02 1
2000-01-03 2
Freq: D, dtype: int64

However, Series and DataFrame can directly also support the time component as data itself.

In [20]: pd.Series(pd.date_range('2000', freq='D', periods=3))

Out[20]:
0 2000-01-01
1 2000-01-02
2 2000-01-03
dtype: datetime64[ns]

4.13. Time series / date functionality 677

pandas: powerful Python data analysis toolkit, Release 0.25.3

Series and DataFrame have extended data type support and functionality for datetime, timedelta and
Period data when passed into those constructors. DateOffset data however will be stored as object data.
In [21]: pd.Series(pd.period_range('1/1/2011', freq='M', periods=3))
Out[21]:
0 2011-01
1 2011-02
2 2011-03
dtype: period[M]

In [22]: pd.Series([pd.DateOffset(1), pd.DateOffset(2)])

Out[22]:
0 <DateOffset>
1 <2 * DateOffsets>
dtype: object

In [23]: pd.Series(pd.date_range('1/1/2011', freq='M', periods=3))

Out[23]:
0 2011-01-31
1 2011-02-28
2 2011-03-31
dtype: datetime64[ns]
Lastly, pandas represents null date times, time deltas, and time spans as NaT which is useful for representing missing
or null date like values and behaves similar as np.nan does for float data.
In [24]: pd.Timestamp(pd.NaT)
Out[24]: NaT

In [25]: pd.Timedelta(pd.NaT)
Out[25]: NaT

In [26]: pd.Period(pd.NaT)
Out[26]: NaT

# Equality acts as np.nan would

In [27]: pd.NaT == pd.NaT
Out[27]: False

4.13.2 Timestamps vs. Time Spans

Timestamped data is the most basic type of time series data that associates values with points in time. For pandas
objects it means using the points in time.
In [28]: pd.Timestamp(datetime.datetime(2012, 5, 1))
Out[28]: Timestamp('2012-05-01 00:00:00')

In [29]: pd.Timestamp('2012-05-01')
Out[29]: Timestamp('2012-05-01 00:00:00')

In [30]: pd.Timestamp(2012, 5, 1)
Out[30]: Timestamp('2012-05-01 00:00:00')
However, in many cases it is more natural to associate things like change variables with a time span instead. The span
represented by Period can be specified explicitly, or inferred from datetime string format.

678 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

For example:
In [31]: pd.Period('2011-01')
Out[31]: Period('2011-01', 'M')

In [32]: pd.Period('2012-05', freq='D')

Out[32]: Period('2012-05-01', 'D')
Timestamp and Period can serve as an index. Lists of Timestamp and Period are automatically coerced to
DatetimeIndex and PeriodIndex respectively.
In [33]: dates = [pd.Timestamp('2012-05-01'),
....: pd.Timestamp('2012-05-02'),
....: pd.Timestamp('2012-05-03')]
....:

In [34]: ts = pd.Series(np.random.randn(3), dates)

In [35]: type(ts.index)
Out[35]: pandas.core.indexes.datetimes.DatetimeIndex

In [36]: ts.index
Out[36]: DatetimeIndex(['2012-05-01', '2012-05-02', '2012-05-03'],
,→dtype='datetime64[ns]', freq=None)

In [37]: ts
Out[37]:
2012-05-01 1.212707
2012-05-02 -2.219105
2012-05-03 0.930394
dtype: float64

In [38]: periods = [pd.Period('2012-01'), pd.Period('2012-02'), pd.

,→Period('2012-03')]

In [39]: ts = pd.Series(np.random.randn(3), periods)

In [40]: type(ts.index)
Out[40]: pandas.core.indexes.period.PeriodIndex

In [41]: ts.index
Out[41]: PeriodIndex(['2012-01', '2012-02', '2012-03'], dtype='period[M]',
,→freq='M')

In [42]: ts
Out[42]:
2012-01 0.703754
2012-02 0.580511
2012-03 -0.135776
Freq: M, dtype: float64
pandas allows you to capture both representations and convert between them. Under the hood, pandas represents
timestamps using instances of Timestamp and sequences of timestamps using instances of DatetimeIndex. For
regular time spans, pandas uses Period objects for scalar values and PeriodIndex for sequences of spans. Better
support for irregular intervals with arbitrary start and end points are forth-coming in future releases.

4.13. Time series / date functionality 679

pandas: powerful Python data analysis toolkit, Release 0.25.3

4.13.3 Converting to timestamps

To convert a Series or list-like object of date-like objects e.g. strings, epochs, or a mixture, you can use the
to_datetime function. When passed a Series, this returns a Series (with the same index), while a list-like is
converted to a DatetimeIndex:
In [43]: pd.to_datetime(pd.Series(['Jul 31, 2009', '2010-01-10', None]))
Out[43]:
0 2009-07-31
1 2010-01-10
2 NaT
dtype: datetime64[ns]

In [44]: pd.to_datetime(['2005/11/23', '2010.12.31'])

Out[44]: DatetimeIndex(['2005-11-23', '2010-12-31'], dtype='datetime64[ns]',
,→freq=None)

If you use dates which start with the day first (i.e. European style), you can pass the dayfirst flag:
In [45]: pd.to_datetime(['04-01-2012 10:00'], dayfirst=True)
Out[45]: DatetimeIndex(['2012-01-04 10:00:00'], dtype='datetime64[ns]',
,→freq=None)

In [46]: pd.to_datetime(['14-01-2012', '01-14-2012'], dayfirst=True)

Out[46]: DatetimeIndex(['2012-01-14', '2012-01-14'], dtype='datetime64[ns]',
,→freq=None)

Warning: You see in the above example that dayfirst isnt strict, so if a date cant be parsed with the day being
first it will be parsed as if dayfirst were False.

If you pass a single string to to_datetime, it returns a single Timestamp. Timestamp can also accept string
input, but it doesnt accept string parsing options like dayfirst or format, so use to_datetime if these are
required.
In [47]: pd.to_datetime('2010/11/12')
Out[47]: Timestamp('2010-11-12 00:00:00')

In [48]: pd.Timestamp('2010/11/12')
Out[48]: Timestamp('2010-11-12 00:00:00')
You can also use the DatetimeIndex constructor directly:

In [49]: pd.DatetimeIndex(['2018-01-01', '2018-01-03', '2018-01-05'])

Out[49]: DatetimeIndex(['2018-01-01', '2018-01-03', '2018-01-05'], dtype=
,→'datetime64[ns]', freq=None)

The string infer can be passed in order to set the frequency of the index as the inferred frequency upon creation:

In [50]: pd.DatetimeIndex(['2018-01-01', '2018-01-03', '2018-01-05'], freq='infer')

Out[50]: DatetimeIndex(['2018-01-01', '2018-01-03', '2018-01-05'], dtype=
,→'datetime64[ns]', freq='2D')

680 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Providing a format argument

In addition to the required datetime string, a format argument can be passed to ensure specific parsing. This could
also potentially speed up the conversion considerably.
In [51]: pd.to_datetime('2010/11/12', format='%Y/%m/%d')
Out[51]: Timestamp('2010-11-12 00:00:00')

In [52]: pd.to_datetime('12-11-2010 00:00', format='%d-%m-%Y %H:%M')

Out[52]: Timestamp('2010-11-12 00:00:00')
For more information on the choices available when specifying the format option, see the Python datetime docu-
mentation.

Assembling datetime from multiple DataFrame columns

New in version 0.18.1.

You can also pass a DataFrame of integer or string columns to assemble into a Series of Timestamps.

In [53]: df = pd.DataFrame({'year': [2015, 2016],

....: 'month': [2, 3],
....: 'day': [4, 5],
....: 'hour': [2, 3]})
....:

In [54]: pd.to_datetime(df)
Out[54]:
0 2015-02-04 02:00:00
1 2016-03-05 03:00:00
dtype: datetime64[ns]

You can pass only the columns that you need to assemble.

In [55]: pd.to_datetime(df[['year', 'month', 'day']])

Out[55]:
0 2015-02-04
1 2016-03-05
dtype: datetime64[ns]

pd.to_datetime looks for standard designations of the datetime component in the column names, including:
• required: year, month, day
• optional: hour, minute, second, millisecond, microsecond, nanosecond

Invalid data

The default behavior, errors='raise', is to raise when unparseable:

In [2]: pd.to_datetime(['2009/07/31', 'asd'], errors='raise')

ValueError: Unknown string format

Pass errors='ignore' to return the original input when unparseable:

In [56]: pd.to_datetime(['2009/07/31', 'asd'], errors='ignore')

Out[56]: Index(['2009/07/31', 'asd'], dtype='object')

4.13. Time series / date functionality 681

pandas: powerful Python data analysis toolkit, Release 0.25.3

Pass errors='coerce' to convert unparseable data to NaT (not a time):

In [57]: pd.to_datetime(['2009/07/31', 'asd'], errors='coerce')

Out[57]: DatetimeIndex(['2009-07-31', 'NaT'], dtype='datetime64[ns]', freq=None)

Epoch timestamps

pandas supports converting integer or float epoch times to Timestamp and DatetimeIndex. The default unit is
nanoseconds, since that is how Timestamp objects are stored internally. However, epochs are often stored in another
unit which can be specified. These are computed from the starting point specified by the origin parameter.
In [58]: pd.to_datetime([1349720105, 1349806505, 1349892905,
....: 1349979305, 1350065705], unit='s')
....:
Out[58]:
DatetimeIndex(['2012-10-08 18:15:05', '2012-10-09 18:15:05',
'2012-10-10 18:15:05', '2012-10-11 18:15:05',
'2012-10-12 18:15:05'],
dtype='datetime64[ns]', freq=None)

In [59]: pd.to_datetime([1349720105100, 1349720105200, 1349720105300,

....: 1349720105400, 1349720105500], unit='ms')
....:
Out[59]:
DatetimeIndex(['2012-10-08 18:15:05.100000', '2012-10-08 18:15:05.200000',
'2012-10-08 18:15:05.300000', '2012-10-08 18:15:05.400000',
'2012-10-08 18:15:05.500000'],
dtype='datetime64[ns]', freq=None)
Constructing a Timestamp or DatetimeIndex with an epoch timestamp with the tz argument specified will
currently localize the epoch timestamps to UTC first then convert the result to the specified time zone. However, this
behavior is deprecated, and if you have epochs in wall time in another timezone, it is recommended to read the epochs
as timezone-naive timestamps and then localize to the appropriate timezone:
In [60]: pd.Timestamp(1262347200000000000).tz_localize('US/Pacific')
Out[60]: Timestamp('2010-01-01 12:00:00-0800', tz='US/Pacific')

In [61]: pd.DatetimeIndex([1262347200000000000]).tz_localize('US/Pacific')
Out[61]: DatetimeIndex(['2010-01-01 12:00:00-08:00'], dtype='datetime64[ns,
,→US/Pacific]', freq=None)

Note: Epoch times will be rounded to the nearest nanosecond.

Warning: Conversion of float epoch times can lead to inaccurate and unexpected results. Python floats have
about 15 digits precision in decimal. Rounding during conversion from float to high precision Timestamp is
unavoidable. The only way to achieve exact precision is to use a fixed-width types (e.g. an int64).
In [62]: pd.to_datetime([1490195805.433, 1490195805.433502912], unit='s')
Out[62]: DatetimeIndex(['2017-03-22 15:16:45.433000088', '2017-03-22
,→15:16:45.433502913'], dtype='datetime64[ns]', freq=None)

In [63]: pd.to_datetime(1490195805433502912, unit='ns')

682 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

Out[63]: Timestamp('2017-03-22 15:16:45.433502912')

See also:
Using the origin Parameter

From timestamps to epoch

To invert the operation from above, namely, to convert from a Timestamp to a unix epoch:

In [64]: stamps = pd.date_range('2012-10-08 18:15:05', periods=4, freq='D')

In [65]: stamps
Out[65]:
DatetimeIndex(['2012-10-08 18:15:05', '2012-10-09 18:15:05',
'2012-10-10 18:15:05', '2012-10-11 18:15:05'],
dtype='datetime64[ns]', freq='D')

We subtract the epoch (midnight at January 1, 1970 UTC) and then floor divide by the unit (1 second).

In [66]: (stamps - pd.Timestamp("1970-01-01")) // pd.Timedelta('1s')

Out[66]: Int64Index([1349720105, 1349806505, 1349892905, 1349979305], dtype='int64')

Using the origin Parameter

New in version 0.20.0.

Using the origin parameter, one can specify an alternative starting point for creation of a DatetimeIndex. For
example, to use 1960-01-01 as the starting date:

In [67]: pd.to_datetime([1, 2, 3], unit='D', origin=pd.Timestamp('1960-01-01'))

Out[67]: DatetimeIndex(['1960-01-02', '1960-01-03', '1960-01-04'], dtype=
,→'datetime64[ns]', freq=None)

The default is set at origin='unix', which defaults to 1970-01-01 00:00:00. Commonly called unix epoch
or POSIX time.

In [68]: pd.to_datetime([1, 2, 3], unit='D')

Out[68]: DatetimeIndex(['1970-01-02', '1970-01-03', '1970-01-04'], dtype=
,→'datetime64[ns]', freq=None)

4.13.4 Generating ranges of timestamps

To generate an index with timestamps, you can use either the DatetimeIndex or Index constructor and pass in a
list of datetime objects:

In [69]: dates = [datetime.datetime(2012, 5, 1),

....: datetime.datetime(2012, 5, 2),
....: datetime.datetime(2012, 5, 3)]
....:

# Note the frequency information

(continues on next page)

4.13. Time series / date functionality 683

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

In [70]: index = pd.DatetimeIndex(dates)

In [71]: index
Out[71]: DatetimeIndex(['2012-05-01', '2012-05-02', '2012-05-03'], dtype=
,→'datetime64[ns]', freq=None)

# Automatically converted to DatetimeIndex

In [72]: index = pd.Index(dates)

In [73]: index
Out[73]: DatetimeIndex(['2012-05-01', '2012-05-02', '2012-05-03'], dtype=
,→'datetime64[ns]', freq=None)

In practice this becomes very cumbersome because we often need a very long index with a large number of timestamps.
If we need timestamps on a regular frequency, we can use the date_range() and bdate_range() functions
to create a DatetimeIndex. The default frequency for date_range is a calendar day while the default for
bdate_range is a business day:

In [74]: start = datetime.datetime(2011, 1, 1)

In [75]: end = datetime.datetime(2012, 1, 1)

In [76]: index = pd.date_range(start, end)

In [77]: index
Out[77]:
DatetimeIndex(['2011-01-01', '2011-01-02', '2011-01-03', '2011-01-04',
'2011-01-05', '2011-01-06', '2011-01-07', '2011-01-08',
'2011-01-09', '2011-01-10',
...
'2011-12-23', '2011-12-24', '2011-12-25', '2011-12-26',
'2011-12-27', '2011-12-28', '2011-12-29', '2011-12-30',
'2011-12-31', '2012-01-01'],
dtype='datetime64[ns]', length=366, freq='D')

In [78]: index = pd.bdate_range(start, end)

In [79]: index
Out[79]:
DatetimeIndex(['2011-01-03', '2011-01-04', '2011-01-05', '2011-01-06',
'2011-01-07', '2011-01-10', '2011-01-11', '2011-01-12',
'2011-01-13', '2011-01-14',
...
'2011-12-19', '2011-12-20', '2011-12-21', '2011-12-22',
'2011-12-23', '2011-12-26', '2011-12-27', '2011-12-28',
'2011-12-29', '2011-12-30'],
dtype='datetime64[ns]', length=260, freq='B')

Convenience functions like date_range and bdate_range can utilize a variety of frequency aliases:
In [80]: pd.date_range(start, periods=1000, freq='M')
Out[80]:
DatetimeIndex(['2011-01-31', '2011-02-28', '2011-03-31', '2011-04-30',
'2011-05-31', '2011-06-30', '2011-07-31', '2011-08-31',
'2011-09-30', '2011-10-31',
...
'2093-07-31', '2093-08-31', '2093-09-30', '2093-10-31',

684 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

'2093-11-30', '2093-12-31', '2094-01-31', '2094-02-28',

'2094-03-31', '2094-04-30'],
dtype='datetime64[ns]', length=1000, freq='M')

In [81]: pd.bdate_range(start, periods=250, freq='BQS')

Out[81]:
DatetimeIndex(['2011-01-03', '2011-04-01', '2011-07-01', '2011-10-03',
'2012-01-02', '2012-04-02', '2012-07-02', '2012-10-01',
'2013-01-01', '2013-04-01',
...
'2071-01-01', '2071-04-01', '2071-07-01', '2071-10-01',
'2072-01-01', '2072-04-01', '2072-07-01', '2072-10-03',
'2073-01-02', '2073-04-03'],
dtype='datetime64[ns]', length=250, freq='BQS-JAN')
date_range and bdate_range make it easy to generate a range of dates using various combinations of parame-
ters like start, end, periods, and freq. The start and end dates are strictly inclusive, so dates outside of those
specified will not be generated:
In [82]: pd.date_range(start, end, freq='BM')
Out[82]:
DatetimeIndex(['2011-01-31', '2011-02-28', '2011-03-31', '2011-04-29',
'2011-05-31', '2011-06-30', '2011-07-29', '2011-08-31',
'2011-09-30', '2011-10-31', '2011-11-30', '2011-12-30'],
dtype='datetime64[ns]', freq='BM')

In [83]: pd.date_range(start, end, freq='W')

Out[83]:
DatetimeIndex(['2011-01-02', '2011-01-09', '2011-01-16', '2011-01-23',
'2011-01-30', '2011-02-06', '2011-02-13', '2011-02-20',
'2011-02-27', '2011-03-06', '2011-03-13', '2011-03-20',
'2011-03-27', '2011-04-03', '2011-04-10', '2011-04-17',
'2011-04-24', '2011-05-01', '2011-05-08', '2011-05-15',
'2011-05-22', '2011-05-29', '2011-06-05', '2011-06-12',
'2011-06-19', '2011-06-26', '2011-07-03', '2011-07-10',
'2011-07-17', '2011-07-24', '2011-07-31', '2011-08-07',
'2011-08-14', '2011-08-21', '2011-08-28', '2011-09-04',
'2011-09-11', '2011-09-18', '2011-09-25', '2011-10-02',
'2011-10-09', '2011-10-16', '2011-10-23', '2011-10-30',
'2011-11-06', '2011-11-13', '2011-11-20', '2011-11-27',
'2011-12-04', '2011-12-11', '2011-12-18', '2011-12-25',
'2012-01-01'],
dtype='datetime64[ns]', freq='W-SUN')

In [84]: pd.bdate_range(end=end, periods=20)

Out[84]:
DatetimeIndex(['2011-12-05', '2011-12-06', '2011-12-07', '2011-12-08',
'2011-12-09', '2011-12-12', '2011-12-13', '2011-12-14',
'2011-12-15', '2011-12-16', '2011-12-19', '2011-12-20',
'2011-12-21', '2011-12-22', '2011-12-23', '2011-12-26',
'2011-12-27', '2011-12-28', '2011-12-29', '2011-12-30'],
dtype='datetime64[ns]', freq='B')

In [85]: pd.bdate_range(start=start, periods=20)

Out[85]:

4.13. Time series / date functionality 685

pandas: powerful Python data analysis toolkit, Release 0.25.3

DatetimeIndex(['2011-01-03', '2011-01-04', '2011-01-05', '2011-01-06',

'2011-01-07', '2011-01-10', '2011-01-11', '2011-01-12',
'2011-01-13', '2011-01-14', '2011-01-17', '2011-01-18',
'2011-01-19', '2011-01-20', '2011-01-21', '2011-01-24',
'2011-01-25', '2011-01-26', '2011-01-27', '2011-01-28'],
dtype='datetime64[ns]', freq='B')
New in version 0.23.0.
Specifying start, end, and periods will generate a range of evenly spaced dates from start to end inclusively,
with periods number of elements in the resulting DatetimeIndex:
In [86]: pd.date_range('2018-01-01', '2018-01-05', periods=5)
Out[86]:
DatetimeIndex(['2018-01-01', '2018-01-02', '2018-01-03', '2018-01-04',
'2018-01-05'],
dtype='datetime64[ns]', freq=None)

In [87]: pd.date_range('2018-01-01', '2018-01-05', periods=10)

Out[87]:
DatetimeIndex(['2018-01-01 00:00:00', '2018-01-01 10:40:00',
'2018-01-01 21:20:00', '2018-01-02 08:00:00',
'2018-01-02 18:40:00', '2018-01-03 05:20:00',
'2018-01-03 16:00:00', '2018-01-04 02:40:00',
'2018-01-04 13:20:00', '2018-01-05 00:00:00'],
dtype='datetime64[ns]', freq=None)

Custom frequency ranges

bdate_range can also generate a range of custom frequency dates by using the weekmask and holidays pa-
rameters. These parameters will only be used if a custom frequency string is passed.
In [88]: weekmask = 'Mon Wed Fri'

In [89]: holidays = [datetime.datetime(2011, 1, 5), datetime.datetime(2011, 3,

,→ 14)]

In [90]: pd.bdate_range(start, end, freq='C', weekmask=weekmask,

,→holidays=holidays)

Out[90]:
DatetimeIndex(['2011-01-03', '2011-01-07', '2011-01-10', '2011-01-12',
'2011-01-14', '2011-01-17', '2011-01-19', '2011-01-21',
'2011-01-24', '2011-01-26',
...
'2011-12-09', '2011-12-12', '2011-12-14', '2011-12-16',
'2011-12-19', '2011-12-21', '2011-12-23', '2011-12-26',
'2011-12-28', '2011-12-30'],
dtype='datetime64[ns]', length=154, freq='C')

In [91]: pd.bdate_range(start, end, freq='CBMS', weekmask=weekmask)

Out[91]:
DatetimeIndex(['2011-01-03', '2011-02-02', '2011-03-02', '2011-04-01',
'2011-05-02', '2011-06-01', '2011-07-01', '2011-08-01',
'2011-09-02', '2011-10-03', '2011-11-02', '2011-12-02'],
dtype='datetime64[ns]', freq='CBMS')

686 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

See also:
Custom business days

4.13.5 Timestamp limitations

Since pandas represents timestamps in nanosecond resolution, the time span that can be represented using a 64-bit
integer is limited to approximately 584 years:
In [92]: pd.Timestamp.min
Out[92]: Timestamp('1677-09-21 00:12:43.145225')

In [93]: pd.Timestamp.max
Out[93]: Timestamp('2262-04-11 23:47:16.854775807')
See also:
Representing out-of-bounds spans

4.13.6 Indexing

One of the main uses for DatetimeIndex is as an index for pandas objects. The DatetimeIndex class contains
many time series related optimizations:
• A large range of dates for various offsets are pre-computed and cached under the hood in order to make gener-
ating subsequent date ranges very fast (just have to grab a slice).
• Fast shifting using the shift and tshift method on pandas objects.
• Unioning of overlapping DatetimeIndex objects with the same frequency is very fast (important for fast
data alignment).
• Quick access to date fields via properties such as year, month, etc.
• Regularization functions like snap and very fast asof logic.
DatetimeIndex objects have all the basic functionality of regular Index objects, and a smorgasbord of advanced
time series specific methods for easy frequency processing.
See also:
Reindexing methods

Note: While pandas does not force you to have a sorted date index, some of these methods may have unexpected or
incorrect behavior if the dates are unsorted.

DatetimeIndex can be used like a regular index and offers all of its intelligent functionality like selection, slicing,
etc.
In [94]: rng = pd.date_range(start, end, freq='BM')

In [95]: ts = pd.Series(np.random.randn(len(rng)), index=rng)

In [96]: ts.index
Out[96]:
DatetimeIndex(['2011-01-31', '2011-02-28', '2011-03-31', '2011-04-29',
'2011-05-31', '2011-06-30', '2011-07-29', '2011-08-31',
'2011-09-30', '2011-10-31', '2011-11-30', '2011-12-30'],

4.13. Time series / date functionality 687

pandas: powerful Python data analysis toolkit, Release 0.25.3

dtype='datetime64[ns]', freq='BM')

In [97]: ts[:5].index
Out[97]:
DatetimeIndex(['2011-01-31', '2011-02-28', '2011-03-31', '2011-04-29',
'2011-05-31'],
dtype='datetime64[ns]', freq='BM')

In [98]: ts[::2].index
Out[98]:
DatetimeIndex(['2011-01-31', '2011-03-31', '2011-05-31', '2011-07-29',
'2011-09-30', '2011-11-30'],
dtype='datetime64[ns]', freq='2BM')

Partial string indexing

Dates and strings that parse to timestamps can be passed as indexing parameters:
In [99]: ts['1/31/2011']
Out[99]: 0.3135034497740378

In [100]: ts[datetime.datetime(2011, 12, 25):]

Out[100]:
2011-12-30 -0.62564
Freq: BM, dtype: float64

In [101]: ts['10/31/2011':'12/31/2011']
Out[101]:
2011-10-31 -0.042894
2011-11-30 1.321441
2011-12-30 -0.625640
Freq: BM, dtype: float64
To provide convenience for accessing longer time series, you can also pass in the year or year and month as strings:
In [102]: ts['2011']
Out[102]:
2011-01-31 0.313503
2011-02-28 1.055889
2011-03-31 0.714651
2011-04-29 0.612442
2011-05-31 -0.170038
2011-06-30 -0.828809
2011-07-29 0.458806
2011-08-31 0.764740
2011-09-30 1.917429
2011-10-31 -0.042894
2011-11-30 1.321441
2011-12-30 -0.625640
Freq: BM, dtype: float64

In [103]: ts['2011-6']
Out[103]:
2011-06-30 -0.828809
Freq: BM, dtype: float64

688 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

This type of slicing will work on a DataFrame with a DatetimeIndex as well. Since the partial string selection
is a form of label slicing, the endpoints will be included. This would include matching times on an included date:
In [104]: dft = pd.DataFrame(np.random.randn(100000, 1), columns=['A'],
.....: index=pd.date_range('20130101', periods=100000,
,→freq='T'))

.....:

In [105]: dft
Out[105]:
A
2013-01-01 00:00:00 0.480217
2013-01-01 00:01:00 0.791875
2013-01-01 00:02:00 1.414514
2013-01-01 00:03:00 -0.801147
2013-01-01 00:04:00 0.462656
... ...
2013-03-11 10:35:00 0.486569
2013-03-11 10:36:00 -0.586561
2013-03-11 10:37:00 -0.403601
2013-03-11 10:38:00 -0.228277
2013-03-11 10:39:00 -0.267201

[100000 rows x 1 columns]

In [106]: dft['2013']
Out[106]:
A
2013-01-01 00:00:00 0.480217
2013-01-01 00:01:00 0.791875
2013-01-01 00:02:00 1.414514
2013-01-01 00:03:00 -0.801147
2013-01-01 00:04:00 0.462656
... ...
2013-03-11 10:35:00 0.486569
2013-03-11 10:36:00 -0.586561
2013-03-11 10:37:00 -0.403601
2013-03-11 10:38:00 -0.228277
2013-03-11 10:39:00 -0.267201

[100000 rows x 1 columns]

This starts on the very first time in the month, and includes the last date and time for the month:

In [107]: dft['2013-1':'2013-2']
Out[107]:
A
2013-01-01 00:00:00 0.480217
2013-01-01 00:01:00 0.791875
2013-01-01 00:02:00 1.414514
2013-01-01 00:03:00 -0.801147
2013-01-01 00:04:00 0.462656
... ...
2013-02-28 23:55:00 1.819527
2013-02-28 23:56:00 0.891281
(continues on next page)

4.13. Time series / date functionality 689

pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2013-02-28 23:57:00 -0.516058
2013-02-28 23:58:00 -1.350302
2013-02-28 23:59:00 1.475049

[84960 rows x 1 columns]

This specifies a stop time that includes all of the times on the last day:
In [108]: dft['2013-1':'2013-2-28']
Out[108]:
A
2013-01-01 00:00:00 0.480217
2013-01-01 00:01:00 0.791875
2013-01-01 00:02:00 1.414514
2013-01-01 00:03:00 -0.801147
2013-01-01 00:04:00 0.462656
... ...
2013-02-28 23:55:00 1.819527
2013-02-28 23:56:00 0.891281
2013-02-28 23:57:00 -0.516058
2013-02-28 23:58:00 -1.350302
2013-02-28 23:59:00 1.475049

[84960 rows x 1 columns]

This specifies an exact stop time (and is not the same as the above):
In [109]: dft['2013-1':'2013-2-28 00:00:00']
Out[109]:
A
2013-01-01 00:00:00 0.480217
2013-01-01 00:01:00 0.791875
2013-01-01 00:02:00 1.414514
2013-01-01 00:03:00 -0.801147
2013-01-01 00:04:00 0.462656
... ...
2013-02-27 23:56:00 0.032319
2013-02-27 23:57:00 0.773067
2013-02-27 23:58:00 -1.433695
2013-02-27 23:59:00 -0.759026
2013-02-28 00:00:00 0.659128

[83521 rows x 1 columns]

We are stopping on the included end-point as it is part of the index:

In [110]: dft['2013-1-15':'2013-1-15 12:30:00']
Out[110]:
A
2013-01-15 00:00:00 -2.553103
2013-01-15 00:01:00 1.310783
2013-01-15 00:02:00 0.851770
2013-01-15 00:03:00 -0.534226
2013-01-15 00:04:00 -1.030374
... ...
2013-01-15 12:26:00 1.679647
2013-01-15 12:27:00 0.049848
(continues on next page)

690 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

(continued from previous page)

2013-01-15 12:28:00 -0.722124
2013-01-15 12:29:00 0.870311
2013-01-15 12:30:00 0.360159

[751 rows x 1 columns]

New in version 0.18.0.

DatetimeIndex partial string indexing also works on a DataFrame with a MultiIndex:
In [111]: dft2 = pd.DataFrame(np.random.randn(20, 1),
.....: columns=['A'],
.....: index=pd.MultiIndex.from_product(
.....: [pd.date_range('20130101', periods=10,
,→freq='12H'),

.....: ['a', 'b']]))

.....:

In [112]: dft2
Out[112]:
A
2013-01-01 00:00:00 a 0.030508
b 0.201088
2013-01-01 12:00:00 a -0.822650
b -0.159673
2013-01-02 00:00:00 a 0.715939
b 0.635435
2013-01-02 12:00:00 a 0.071542
b 0.539646
2013-01-03 00:00:00 a -0.743837
b 1.319587
2013-01-03 12:00:00 a -0.501123
b -0.492347
2013-01-04 00:00:00 a -0.357006
b -0.252463
2013-01-04 12:00:00 a -1.140700
b -1.367172
2013-01-05 00:00:00 a 0.048871
b -0.400048
2013-01-05 12:00:00 a 1.325801
b 0.651751

In [113]: dft2.loc['2013-01-05']
Out[113]:
A
2013-01-05 00:00:00 a 0.048871
b -0.400048
2013-01-05 12:00:00 a 1.325801
b 0.651751

In [114]: idx = pd.IndexSlice

In [115]: dft2 = dft2.swaplevel(0, 1).sort_index()

4.13. Time series / date functionality 691

pandas: powerful Python data analysis toolkit, Release 0.25.3

In [116]: dft2.loc[idx[:, '2013-01-05'], :]

Out[116]:
A
a 2013-01-05 00:00:00 0.048871
2013-01-05 12:00:00 1.325801
b 2013-01-05 00:00:00 -0.400048
2013-01-05 12:00:00 0.651751
New in version 0.25.0.
Slicing with string indexing also honors UTC offset.
In [117]: df = pd.DataFrame([0], index=pd.DatetimeIndex(['2019-01-01'],
,→tz='US/Pacific'))

In [118]: df
Out[118]:
0
2019-01-01 00:00:00-08:00 0

In [119]: df['2019-01-01 12:00:00+04:00':'2019-01-01 13:00:00+04:00']

Out[119]:
0
2019-01-01 00:00:00-08:00 0

Slice vs. exact match

Changed in version 0.20.0.

The same string used as an indexing parameter can be treated either as a slice or as an exact match depending on the
resolution of the index. If the string is less accurate than the index, it will be treated as a slice, otherwise as an exact
match.
Consider a Series object with a minute resolution index:

In [120]: series_minute = pd.Series([1, 2, 3],

.....: pd.DatetimeIndex(['2011-12-31 23:59:00',
.....: '2012-01-01 00:00:00',
.....: '2012-01-01 00:02:00']))
.....:

In [121]: series_minute.index.resolution
Out[121]: 'minute'

A timestamp string less accurate than a minute gives a Series object.

In [122]: series_minute['2011-12-31 23']

Out[122]:
2011-12-31 23:59:00 1
dtype: int64

A timestamp string with minute resolution (or more accurate), gives a scalar instead, i.e. it is not casted to a slice.
In [123]: series_minute['2011-12-31 23:59']
Out[123]: 1

In [124]: series_minute['2011-12-31 23:59:00']

Out[124]: 1

692 Chapter 4. User Guide

 pandas: powerful Python data analysis toolkit, Release 0.25.3

If index resolution is second, then the minute-accurate timestamp gives a Series.

In [125]: series_second = pd.Series([1, 2, 3],
.....: pd.DatetimeIndex(['2011-12-31 23:59:59',
.....: '2012-01-01 00:00:00',
.....: '2012-01-01 00:00:01']))
.....:

In [126]: series_second.index.resolution
Out[126]: 'second'

In [127]: series_second['2011-12-31 23:59']

Out[127]:
2011-12-31 23:59:59 1
dtype: int64
If the timestamp string is treated as a slice, it can be used to index DataFrame with [] as well.

In [128]: dft_minute = pd.DataFrame({'a': [1, 2, 3], 'b': [4, 5, 6]},

.....: index=series_minute.index)
.....:

In [129]: dft_minute['2011-12-31 23']

Out[129]:
a b
2011-12-31 23:59:00 1 4

Warning: However, if the string is treated as an exact match, the selection in DataFrames [] will be column-
wise and not row-wise, see Indexing Basics. For example dft_minute['2011-12-31 23:59'] will raise
KeyError as '2012-12-31 23:59' has the same resolution as the index and there is no column with such
name:
To always have unambiguous selection, whether the row is treated as a slice or a single selection, use .loc.
In [130]: dft_minute.loc['2011-12-31 23:59']
Out[130]:
a 1
b 4
Name: 2011-12-31 23:59:00, dtype: int64

Note also that DatetimeIndex resolution cannot be less precise than day.
In [131]: series_monthly = pd.Series([1, 2, 3],
.....: pd.DatetimeIndex(['2011-12', '2012-01',
,→'2012-02']))

.....:

In [132]: series_monthly.index.resolution
Out[132]: 'day'

In [133]: series_monthly['2011-12'] # returns Series

Out[133]:
2011-12-01 1
dtype: int64

4.13. Time series / date functionality 693

pandas: powerful Python data analysis toolkit, Release 0.25.3

Exact indexing

As discussed in previous section, indexing a DatetimeIndex with a partial string depends on the accuracy of the
period, in other words how specific the interval is in relation to the resolution of the index. In contrast, indexing with
Timestamp or datetime objects is exact, because the objects have exact meaning. These also follow the semantics
of including both endpoints.
These Timestamp and datetime objects have exact hours, minutes, and seconds, even though they were
not explicitly specified (they are 0).

In [134]: dft[datetime.datetime(2013, 1, 1):datetime.datetime(2013, 2, 28)]

Out[134]:
A
2013-01-01 00:00:00 0.480217
2013-01-01 00:01:00 0.791875
2013-01-01 00:02:00 1.414514
2013-01-01 00:03:00 -0.801147
2013-01-01 00:04:00 0.462656
... ...
2013-02-27 23:56:00 0.032319
2013-02-27 23:57:00 0.773067
2013-02-27 23:58:00 -1.433695
2013-02-27 23:59:00 -0.759026
2013-02-28 00:00:00 0.659128

[83521 rows x 1 columns]

With no defaults.

In [135]: dft[datetime.datetime(2013, 1, 1, 10, 12, 0):

.....: datetime.datetime(2013, 2, 28, 10, 12, 0)]
.....:
Out[135]:
A
2013-01-01 10:12:00 0.827464
2013-01-01 10:13:00 -2.923827
2013-01-01 10:14:00 -2.109236
2013-01-01 10:15:00 -1.088176
2013-01-01 10:16:00 0.788969
... ...
2013-02-28 10:08:00 -1.473233
2013-02-28 10:09:00 0.090886
2013-02-28 10:10:00 0.211966
2013-02-28 10:11:00 -1.521579
2013-02-28 10:12:00 -0.318755

[83521 rows x 1 columns]

Truncating & fancy indexing