Time Series

10/25/2020 Sunspot_Time_Series_Data
In [ ]: !pip install pmdarima # Package for AutoArima
In [3]: !pip install plotly
In [2]: import matplotlib as mpl

import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np
import os
from datetime import datetime
import pandas as pd
mpl.rcParams['figure.figsize'] = (18, 7)
mpl.rcParams['axes.grid'] = False
# @ Ramendra Kumar
In [13]: df = pd.read_csv('sunspots.csv')
print(df.head())
df.info()
Unnamed: 0 Date Monthly Mean Total Sunspot Number

0 0 1749-01-31 96.7
1 1 1749-02-28 104.3
2 2 1749-03-31 116.7
3 3 1749-04-30 92.8
4 4 1749-05-31 141.7
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 3235 entries, 0 to 3234
Data columns (total 3 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 Unnamed: 0 3235 non-null int64
1 Date 3235 non-null object
2 Monthly Mean Total Sunspot Number 3235 non-null float64
dtypes: float64(1), int64(1), object(1)
memory usage: 75.9+ KB
We can see from info, Date column is stored as object i.e. string data type . Date column must be converted into
datatime format which makes it easier for working with date and time data.There is an unnecessary column
named 'Unnamed :0' which has to be removed.
df['Date']=pd.to_datetime(df['Date']) <-- can be used to convert a column to into datetime data type column
pd.drop can be used for dropping the unnnecesary column
Here, I am using 'usecols' argument inside pd.read_csv for selecting only required column.
'parse_date', & 'date_parser' arguments for converting Date column into datetime data type.
inside 'parse_data, we have to pass the column to be conveted into datetime, here, it is 'Date' column.
'dateparse' function below is requied which is basically converting any argument passed to it into datetime
data type . This is given to 'data_parser' inside pd.read_csv.
check the documentation of pd.read_csv, there are more than 15 arguments, which can be used to perform
many operations while importing the data itself.
file:///C:/Users/hP/Downloads/Sunspot_Time_Series.html 1/31
In [5]: from dateutil.parser import parse

dateparse=lambda dates:parse(dates)
In [7]: df = pd.read_csv('sunspots.csv',usecols=['Date','Monthly Mean Total Sunspot Nu

mber'],parse_dates=['Date'],date_parser=dateparse)
df.head()
Out[7]:
Date Monthly Mean Total Sunspot Number
0 1749-01-31 96.7
1 1749-02-28 104.3
2 1749-03-31 116.7
3 1749-04-30 92.8
4 1749-05-31 141.7
In [8]: df.info() ## Checking the info again : data type of Date column --> has convet
ed into datetime
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 3235 entries, 0 to 3234
Data columns (total 2 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 Date 3235 non-null datetime64[ns]
1 Monthly Mean Total Sunspot Number 3235 non-null float64
dtypes: datetime64[ns](1), float64(1)
memory usage: 50.7 KB
1. Exploratory Data Analysis

In [9]: df_non_index=df.copy() # Making a copy of initial data.Both will be used as re
quired
# The 'df_non_index' dataframe is used for some exploratory data analysis
# Later we will convert Date colum as index in 'df' dataframe.
Profit of datetime formated data:

You can do lot of date and time related operations easily without doing string opeations.
Here month is seprated and kept in another column named month so easily
Afeter that year is seperated and used.
In [10]: df_non_index['Month']=df_non_index.Date.dt.month
df_non_index.head()
Out[10]:
Date Monthly Mean Total Sunspot Number Month
0 1749-01-31 96.7 1
1 1749-02-28 104.3 2
2 1749-03-31 116.7 3
3 1749-04-30 92.8 4
4 1749-05-31 141.7 5
The following code is extracting the each year of the decade, for example in string '1749' last character i.e.
(3rd positional) is year 9 of that decade, which has been extracded and kept in another column named
'nth_year.
for '1748' it wil be year 8.
But for '1750' it will be year '0' which has to be 10. Thus .replace('0','10') is applied and finally converted
back into intger by type casting.
In [11]: df_non_index['nth_year'] =[int(str(i)[3]) for i in (df_non_index.Date.dt.year

)] # Note this is list comprehension
df_non_index['nth_year'].replace(0,10,inplace=True)
df_non_index.head()
Out[11]:
Date Monthly Mean Total Sunspot Number Month nth_year
0 1749-01-31 96.7 1 9
1 1749-02-28 104.3 2 9
2 1749-03-31 116.7 3 9
3 1749-04-30 92.8 4 9
4 1749-05-31 141.7 5 9
Plotting the data using seaborn boxplot
In [12]: fig, axes = plt.subplots(3, 1, figsize=(20,15), dpi= 80)

sns.boxplot(x='Date', y='Monthly Mean Total Sunspot Number', data=df_non_index
, ax=axes[0])
sns.boxplot(x='Month', y='Monthly Mean Total Sunspot Number', data=df_non_inde
x,ax = axes[1])
sns.boxplot(x='nth_year', y='Monthly Mean Total Sunspot Number', data=df_non_i
ndex,ax = axes[2])
# Set Title
axes[0].set_title('Year-wise Box Plot\n(The Trend)', fontsize=14);
axes[1].set_title('Month-wise Box Plot\n(The Seasonality)', fontsize=14)
axes[2].set_title('nth_year_each_decade\n(The Seasonality)', fontsize=14)
fig.tight_layout()
plt.show()
Eplanation of above plot:

The distribution of data is almost same in each month with few outliers
The distribution of data among each year of the decades are not same .
Returing back to dataframe 'df' and Making Date column as index

Once we make Date column as index,it is very easy to slice the data based on index (i.e. date) and even
plotting in pandas with datetime column as index is easy.
In [13]: df = df.set_index('Date')
df.head()
Out[13]:
Monthly Mean Total Sunspot Number
Date
1749-01-31 96.7
1749-02-28 104.3
1749-03-31 116.7
1749-04-30 92.8
1749-05-31 141.7
In [14]: df.tail()
Out[14]:
Monthly Mean Total Sunspot Number
Date
2018-03-31 2.5
2018-04-30 8.9
2018-05-31 13.2
2018-06-30 15.9
2018-07-31 1.6
In [15]: df.plot(grid=True) # plots in pandas itself take index as x axis, here it is d

atetime and y axis is 'Monthly Mean Total Sunspot Number'
# This plot is same to that of previous first box plot (that was a scatter pl
ot, here it dots are joined )
Out[15]: <AxesSubplot:xlabel='Date'>
The data is too to large to see it in a one graph, there are 3235 monthly entries from date 1749-01-31 to
2018-07-31.
One way to slice the data and visualise any particular time zone.
Plotly express provide slider and button to select particular time zone.
Checking both:
In [16]: df_2018=df.loc['2000':'2010'] # Slicing all data from 2000 to 2010

df_2018.plot(figsize=(16,7),grid=True)
In [17]: df_2018=df.loc['1900':'1920'] # Slicing all data from 1900 to 1910

df_2018.plot(figsize=(16,7),grid=True)
plotly express
In [18]: import plotly.express as px

fig = px.line(df_non_index, x='Date', y='Monthly Mean Total Sunspot Number', t
itle='Mean_Sunspot_Slider')
fig.update_xaxes(rangeslider_visible=True)
fig.show()
## There is slider belwo the graph using which we can select any particular ti
me zone
Mean_Sunspot_Slider
hly Mean Total Sunspot Number
400
300
200
100
Buttons options in plotly
In [19]: fig = px.line(df_non_index, x='Date', y='Monthly Mean Total Sunspot Number', t

itle='Mean_Sunspot_Slider')
fig.update_xaxes(
rangeslider_visible=True,
rangeselector=dict(
buttons=list([
dict(count=10, label="10y", step="year", stepmode="backward"),
dict(step="all")
])
)
)
fig.show()
Mean_Sunspot_Slider
10y 20y 30y 40y 50y all

hly Mean Total Sunspot Number
400
300
200
100
In [ ]: df_non_index.head()
Out[ ]:
Date Monthly Mean Total Sunspot Number Month nth_year
0 1749-01-31 96.7 1 9
1 1749-02-28 104.3 2 9
2 1749-03-31 116.7 3 9
3 1749-04-30 92.8 4 9
4 1749-05-31 141.7 5 9
Comparison of two consecutive 11 year: How to choose from where to where?
In above graph we can see the pattern is repeating after 11 year approx, choose the time to match any two
reapeated pattern
In [ ]: df_11_1985=df_non_index[(df_non_index.Date.dt.year>=1985) & (df_non_index.Date

.dt.year<1996)]
df_11_1996=df_non_index[(df_non_index.Date.dt.year>=1996) &(df_non_index.Date.
dt.year<2007)]
x=np.arange(1,len(df_11_1996['Date'])+1)
plt.plot(x, df_11_1985['Monthly Mean Total Sunspot Number'],label='df_60_1998'

)
plt.plot(x, df_11_1996['Monthly Mean Total Sunspot Number'],label='df_60_1958'
)
plt.legend()
plt.xlabel('Month')
plt.ylabel('Monthly Mean Total Sunspot Number')
plt.title('Comparison of Two consecutive 11 year')
Out[ ]: Text(0.5, 1.0, 'Comparison of Two consecutive 11 year')
Lag plot
It helps to understand the autocorrelation lag, visualizing for few, normally lag greater than 4 is not useful.
As we increase the lag time, the correlation is decresing.
the data is correlated with its recet time lag upt 4/5 time lag.
In [ ]: fig=plt.figure(figsize=(18,6))
fig.subplots_adjust(hspace=0.4, wspace=0.2)
ax1=fig.add_subplot(2,2,1)
pd.plotting.lag_plot(df['Monthly Mean Total Sunspot Number'],lag=1)
plt.title('Lag_1')
plt.title('Lag_3')
plt.title('Lag_6')
plt.title('Lag_24')
plt.show()
Checking the distribution by making histogram and kde plot
In [ ]: fig=plt.figure(figsize=(18,6))
fig.subplots_adjust(hspace=0.4, wspace=0.2)
df['Monthly Mean Total Sunspot Number'].hist()
plt.title('Histogram')
df['Monthly Mean Total Sunspot Number'].plot(kind='density')# kernel density p
lot
plt.title('KDE')
Out[ ]: Text(0.5, 1.0, 'KDE')
In [ ]:
2. Checking Stationarity of Time Series Data

From the plot of data we can see that the it is stationary, though we have to check it statistically. ### Check
Stationarity of a Time Series
A TS is said to be stationary if its statistical properties such as mean, variance remain constant over time. But
why is it important? Most of the TS models work on the assumption that the TS is stationary. Intuitively, we can
state that if a TS has a particular behaviour over time, there is a very high probability that it will follow the same
in the future. Also, the theories related to stationary series are more mature and easier to implement as
compared to non-stationary series.
Stationarity is defined using very strict criterion. However, for practical purposes we can assume the series to be
stationary if it has constant statistical properties over time, ie. the following:
constant mean (For different time slots)

constant variance (For different time slots)
(Rolling mean/variance should be checked and should be constant)
an autocovariance that does not depend on time
Two test for stationarity: ADF & KPSS test
https://www.statsmodels.org/stable/examples/notebooks/generated/stationarity_detrending_adf_kpss.html
(https://www.statsmodels.org/stable/examples/notebooks/generated/stationarity_detrending_adf_kpss.html)
Perform Augumented Dickey-Fuller test:

Dickey-Fuller Test: This is one of the statistical tests for checking stationarity. Here the null hypothesis is that the
TS is non-stationary. The test results comprise of a Test Statistic and some Critical Values for difference
confidence levels. If the ‘Test Statistic’ is less than the ‘Critical Value’, we can reject the null hypothesis and say
that the series is stationary.
Null Hypothesis - Series is not stationary

Alternate Hypothesis - Series is stationary
In [ ]: from statsmodels.tsa.stattools import adfuller
In [ ]: data_series=df['Monthly Mean Total Sunspot Number']
In [ ]: print('Results of Dickey-Fuller Test:')

dftest = adfuller(data_series, autolag='AIC')
dfoutput = pd.Series(dftest[0:4], index=['Test Statistic','p-value','#Lags Use
d','Number of Observations Used'])
for key,value in dftest[4].items():
dfoutput['Critical Value (%s)'%key] = value
print(dfoutput)
if dfoutput['Test Statistic'] < dfoutput['Critical Value (5%)']: ## Comparing
with 5% significant Level
print('Series is stationary')
else:
print('Series is not Stationary')
## OR
if dfoutput[1] > 0.05 :
else:
print('Series is Stationary')
Results of Dickey-Fuller Test:

Test Statistic -1.049256e+01
p-value 1.137033e-18
#Lags Used 2.800000e+01
Number of Observations Used 3.206000e+03
Critical Value (1%) -3.432391e+00
dtype: float64
Series is stationary
Series is Stationary
KPSS test for stationary: This another test

Null hypothesis - Series is stationary
Alternate hypothesis - Series is not stationary
In [ ]: from statsmodels.tsa.stattools import kpss
In [ ]: stats, p, lags, critical_values = kpss(df['Monthly Mean Total Sunspot Number'

], 'c',nlags='legacy')
## pass --> 'ct' if there is trend component in data
## pass --> 'c' if there is no trend component in data. In this case there is
not trend in the data being stationary data.
/usr/local/lib/python3.6/dist-packages/statsmodels/tsa/stattools.py:1687: Int
erpolationWarning:
p-value is greater than the indicated p-value
In [ ]: print(f'Test Statistics: {stats}')

print(f'p-value: {p}')
print(f'Critial Values: {critical_values}')
if p < 0.05 :
else:
print('Series is Stationary')
Test Statistics: 0.1435419872086975

p-value: 0.1
Critial Values: {'10%': 0.347, '5%': 0.463, '2.5%': 0.574, '1%': 0.739}
Series is Stationary
Note: For Non-Stationary data: First make it stationary

Differencing, Taking log and Differencing, Decompostion in components and detrending are few techniques
are used.
3. Modelling Time Series
There are many ways to model a time series in order to make predictions.Few are
discussed here:
Different Moving Averages
Exponential Smoothing
ARIMA
SARIMA
Rolling Statistics:
We can plot the moving average or moving variance and see if it varies with time. By moving average/variance I
mean that at any instant ‘t’, we’ll take the average/variance of the last year, i.e. last 12 months. But again this is
more of a visual technique :
Rolling Average OR Simple moving average = (t + (t-1) + (t-2) + ... + (t-n)) / n
In [ ]: df['Monthly Mean Total Sunspot Number'][:200].plot() # Checking for only first

200 data set
df['Monthly Mean Total Sunspot Number'][:200].rolling(3).mean().plot(label='ro
lling mean') ## rolling average with 3 time step also known as window
#df['Monthly Mean Total Sunspot Number'][:200].rolling(3).std().plot(label='ro
lling std')
plt.legend()
plt.title('Rolling Mean & Standard Deviation')
## df['Monthly Mean Total Sunspot Number'].rolling(12).mean().shift(1) # Rolli
ng mean with shift
Out[ ]: Text(0.5, 1.0, 'Rolling Mean & Standard Deviation')
Weighted moving average
Weighted moving average = (tweighting factor) + ((t-1)weighting factor-1) + ((t-n) *

weighting factor-n)/n
This is similar as rolling average except, we multiply with weighting factor so that more weight is given to
recent data.
this function is not availbele, we have to make our own
In [ ]: ## Making a function for calculating weighted average which is passed through

.apply()
def wma(weights):
def calc(x):
return (weights*x).mean()
return calc

200 data set
df['Monthly Mean Total Sunspot Number'][:200].rolling(3).apply(wma(np.array([
0.5,1,1.5]))).plot(label='weighted mooving_averate')
# Here inside wma 3 weights are passed since we are taking 3 time step only a
s window.
plt.legend()
Out[ ]: <matplotlib.legend.Legend at 0x7f6ee75800b8>
Exponential moving average\Exponential Smoothing
https://en.wikipedia.org/wiki/Moving_average#Exponential_moving_average
(https://en.wikipedia.org/wiki/Moving_average#Exponential_moving_average) #### * Luckly there is a
function for this in pandas.

200 data set
df['Monthly Mean Total Sunspot Number'][:200].ewm(span=3, adjust=False, min_pe
riods=3).mean().plot(label='Exponential Weighted Average')
## Here span=3 is provide thus α=2/(span+1) automatically calculated and appli
ed
## https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFram
e.ewm.html
plt.title('Exponential Weighted M.A.')
plt.legend()
Out[ ]: <matplotlib.legend.Legend at 0x7f6ee7eb1860>
Providing alpha for Smoothing

200 data set
df['Monthly Mean Total Sunspot Number'][:200].ewm(alpha=0.7, adjust=False, min
_periods=3).mean().plot(label='Exponential Smooting M A')
Out[ ]: <matplotlib.axes._subplots.AxesSubplot at 0x7f6ee7f8efd0>
Plotting All together and comparing
In [ ]: df_with_diff_avg=df[:200].copy()
df_with_diff_avg['Rolling mean']=df['Monthly Mean Total Sunspot Number'][:200]
.rolling(3).mean()
df_with_diff_avg['W_M_A']= df['Monthly Mean Total Sunspot Number'][:200].rolli
ng(window=3).apply(wma(np.array([0.5,1,1.5])))
df_with_diff_avg['E_W_A']= df['Monthly Mean Total Sunspot Number'][:200].ewm(s
pan=3, adjust=False, min_periods=0).mean()
df_with_diff_avg['E_S_M_A']= df['Monthly Mean Total Sunspot Number'][:200].ewm
(alpha=0.7, adjust=False, min_periods=3).mean()
print(df_with_diff_avg.head())
#df_with_diff_avg.set_index('Date', inplace=True)
df_with_diff_avg.plot()
Monthly Mean Total Sunspot Number Rolling mean ... E_W_A E_

S_M_A
Date ...
1749-01-31 96.7 NaN ... 96.7
NaN
1749-02-28 104.3 NaN ... 100.5
NaN
1749-03-31 116.7 105.900000 ... 108.6 112.
29600
1749-04-30 92.8 104.600000 ... 100.7 98.
64880
1749-05-31 141.7 117.066667 ... 121.2 128.
78464
[5 rows x 5 columns]
Out[ ]: <matplotlib.axes._subplots.AxesSubplot at 0x7f6ee7732c18>
In [ ]: df_with_diff_avg.dropna(inplace=True)
In [ ]: df_with_diff_avg.head()
Out[ ]:
Monthly Mean Total Sunspot Number Rolling mean W_M_A E_W_A E_S_M_A
Date
1749-03-31 116.7 105.900000 109.233333 108.6 112.296000
1749-04-30 92.8 104.600000 102.683333 100.7 98.648800
1749-05-31 141.7 117.066667 121.233333 121.2 128.784640
1749-06-30 139.2 124.566667 132.300000 130.2 136.075392
1749-07-31 158.0 146.300000 149.016667 144.1 151.422618
Making a function for comparing RMSE in all above modelling

We can see exponential smoothing Moving average has lowest RMSE.
In [ ]: def RMSE_CAL(df):
Rolling_Mean_RMSE=np.sqrt(np.sum((df.iloc[:,0]-df.iloc[:,1])**2))
W_M_A_RMSE=np.sqrt(np.sum((df.iloc[:,0]-df.iloc[:,2])**2))
E_W_A_RMSE=np.sqrt(np.sum((df.iloc[:,0]-df.iloc[:,3])**2))
E_S_M_A_RMSE=np.sqrt(np.sum((df.iloc[:,0]-df.iloc[:,4])**2))
return {"Rolling_Mean_RMSE":Rolling_Mean_RMSE,"W_M_A_RMSE":W_M_A_RMSE,"E_W_A
_RMSE":E_W_A_RMSE,"E_S_M_A_RMSE":E_S_M_A_RMSE}
RMSE_CAL(df_with_diff_avg)
Out[ ]: {'E_S_M_A_RMSE': 76.92627388727384,

'E_W_A_RMSE': 123.99324056510582,
'Rolling_Mean_RMSE': 175.07366703444842,
'W_M_A_RMSE': 130.1346727646232}
4. Decomposing a Time_Series Data

NOTE: This operation is not required for this data as it is stationary but while working with non-
stationary data this step may required.
Systematic: Components of the time series that have consistency or reocurrence and can be described and
modeled as level,trend, seasonality.
Non-Systematic: Components of the time series that cannot be directly modeled is noise/residual.
These components are defined as follows:
Level: The average value in the series.

Trend: The increasing or decreasing value in the series.
Seasonality: The repeating short-term cycle in the series.
Noise: The random variation in the series.
So a time series is thought to be an aggregate or combination of these four components. All

series have a level and noise. The trend and seasonality components are optional. It is
helpful to think of the components as combining either additively or multiplicatively as given
by relation below:
y(t) = Level + Trend + Seasonality + Noise

y(t) = Level Trend Seasonality * Noise
Since our data is stationary we will use additive decomposition
In [ ]: # Additive decomposition
from statsmodels.tsa.seasonal import seasonal_decompose
result = seasonal_decompose(df['Monthly Mean Total Sunspot Number'], model="ad
ditive",freq=11*12) # Data Trend is repeated after every 11 year,freq=11*12
result.plot()
plt.show()
/usr/local/lib/python3.6/dist-packages/ipykernel_launcher.py:3: FutureWarnin
g:
the 'freq'' keyword is deprecated, use 'period' instead
checking the definition of decompostion for additive nature time series data
y(t) = Trend + Seasonality + Noise
In [ ]: total_sum=result.trend+result.seasonal+result.resid
total_sum[:100] # compare this result with original Sunspot data
Out[ ]: Date
1749-01-31 NaN
1749-02-28 NaN
1749-03-31 NaN
1749-04-30 NaN
1749-05-31 NaN
...
1756-12-31 15.7
1757-01-31 23.5
1757-02-28 35.3
1757-03-31 43.7
1757-04-30 50.0
Length: 100, dtype: float64
In [ ]: df['Monthly Mean Total Sunspot Number'][:100]
Out[ ]: Date
1749-01-31 96.7
1749-02-28 104.3
1749-03-31 116.7
1749-04-30 92.8
1749-05-31 141.7
...
1756-12-31 15.7
1757-01-31 23.5
1757-02-28 35.3
1757-03-31 43.7
1757-04-30 50.0
Name: Monthly Mean Total Sunspot Number, Length: 100, dtype: float64
Detrended Data :
Since our data is additive in nature we are going to subtract the trend from observed value and get the
detrended data:
In [ ]: pd.DataFrame(result.observed-result.trend).plot()
Out[ ]: <matplotlib.axes._subplots.AxesSubplot at 0x7f6ee16a7978>
5. Autocorrelation plot
We can assume the distribution of each variable fits a Gaussian (bell curve) distribution. If this is the case,
we can use the Pearson’s correlation coefficient to summarize the correlation between the variables.
The Pearson’s correlation coefficient is a number between -1 and 1 that describes a negative or positive
correlation respectively. A value of zero indicates no correlation.
We can calculate the correlation for time series observations with observations with previous time steps,
called lags. Because the correlation of the time series observations is calculated with values of the same
series at previous times, this is called a serial correlation, or an autocorrelation.
A plot of the autocorrelation of a time series by lag is called the AutoCorrelation Function, or the acronym
ACF. This plot is sometimes called a correlogram or an autocorrelation plot.
This helps us to find if current value depends on previous values. In the plot you can observe that current
value is dependent on previous 120-130 values. This can be around 10/11 years as it is monthly data.
In [ ]: pd.plotting.autocorrelation_plot(df['Monthly Mean Total Sunspot Number']) ## f

or each month
Out[ ]: <matplotlib.axes._subplots.AxesSubplot at 0x7f411802ea20>
In [ ]: df['Monthly Mean Total Sunspot Number'].resample("1y").mean() ## Resample base

d on 1 year
Out[ ]: Date
1749-12-31 134.875000
1750-12-31 139.000000
1751-12-31 79.441667
1752-12-31 79.666667
1753-12-31 51.125000
...
2014-12-31 113.608333
2015-12-31 69.783333
2016-12-31 39.825000
2017-12-31 21.816667
2018-12-31 8.514286
Freq: A-DEC, Name: Monthly Mean Total Sunspot Number, Length: 270, dtype: flo
at64
In [ ]: pd.plotting.autocorrelation_plot(df['Monthly Mean Total Sunspot Number'].resam

ple("1y").mean())
Out[ ]: <matplotlib.axes._subplots.AxesSubplot at 0x7f41416f6668>
ACF and PACF plots:

Running the example creates a 2D plot showing the lag value along the x-axis and the correlation on the y-
axis between -1 and 1.
Confidence intervals are drawn as a cone. By default, this is set to a 95% confidence interval, suggesting
that correlation values outside of this code are very likely a correlation and not a statistical fluke.
acf: By looking at the plot we can improvise our understanding from above plot and say that present value
depends on previous 25-30 values.
pacf plot further says that present value depends only on previous 5/6 values. All these plots help us narrow
down thinking and make our model efficient.
In [ ]: from statsmodels.tsa.stattools import acf, pacf

from statsmodels.graphics.tsaplots import plot_acf, plot_pacf
# Draw Plot
plot_acf(df['Monthly Mean Total Sunspot Number'].tolist(), lags=20, ax=axes[0
])
plot_pacf(df['Monthly Mean Total Sunspot Number'].tolist(), lags=20, ax=axes[1
])
Out[ ]:
Application of ACF and PCF

Seasonal autoregressive integraded moving average model (SARIMA) is actually the combination of simpler
models to make a complex model that can model time series exhibiting non-stationary properties and
seasonality.
At first, we have the autoregression model AR(p). This is basically a regression of the time series onto itself.
Here, we assume that the current value depends on its previous values with some lag. It takes a parameter
p which represents the maximum lag. To find it, we look at the partial autocorrelation plot and identify the lag
after which most lags are not significant. In above PCF it is 6 (actually we have to experiment suing-->4,5,6).
Count all the lag till it becomes zero.
Then, we add the moving average model MA(q). This takes a parameter q which represents the biggest lag
after which other lags are not significant on the autocorrelation plot.In the plot count all the lag before
confidence cone, here it is 1.
After, we add the order of integration I(d). The parameter d represents the number of differences required to
make the series stationary.
Finally, we add the final component: seasonality S(P, D, Q, s), where s is simply
the season’s length. Furthermore, this component requires the parameters P and
Q which are the same as p and q, but for the seasonal component. Finally, D is
the order of seasonal integration representing the number of differences required
to remove seasonality from the series.
Combining all, we get the SARIMA(p, d, q)(P, D, Q, s) model. The main takeaway is: before modelling with
SARIMA, we must apply transformations to our time series to remove seasonality and any non-stationary
behaviors.
In [ ]:
Read this for parameter selection detail
https://alkaline-ml.com/pmdarima/tips_and_tricks.html (https://alkaline-
ml.com/pmdarima/tips_and_tricks.html)
Auto ARIMA
In [ ]: import pmdarima as pm
model = pm.auto_arima(df['Monthly Mean Total Sunspot Number'],

m=11, seasonal=True,
start_p=0, start_q=0, max_order=4, test='adf',error_acti
on='ignore',
suppress_warnings=True,
stepwise=True, trace=True)
## actually we have to set m=11*12,but it take too much time and it doesnt ma
tter much. Source Given below:
# https://robjhyndman.com/hyndsight/longseasonality/
Performing stepwise search to minimize aic

ARIMA(0,0,0)(1,0,1)[11] intercept : AIC=33679.421, Time=7.62 sec
ARIMA(0,0,0)(0,0,0)[11] : AIC=39401.647, Time=0.05 sec
ARIMA(5,0,0)(2,0,2)[11] intercept : AIC=inf, Time=100.73 sec
ARIMA(4,0,1)(2,0,2)[11] : AIC=inf, Time=49.77 sec
Best model: ARIMA(4,0,1)(2,0,2)[11] intercept

Total fit time: 1704.195 seconds
In [ ]: model.summary()
Out[ ]:
SARIMAX Results
Dep. Variable: y No. Observations: 3235
Model: SARIMAX(4, 0, 1)x(2, 0, [1, 2], 11) Log Likelihood -15009.440
Date: Sat, 24 Oct 2020 AIC 30040.881
Time: 17:20:49 BIC 30107.780
Sample: 0 HQIC 30064.852
- 3235
Covariance Type: opg
coef std err z P>|z| [0.025 0.975]
intercept 0.3047 0.152 2.005 0.045 0.007 0.603
ar.L1 1.1317 0.055 20.455 0.000 1.023 1.240
ar.L2 -0.2101 0.038 -5.584 0.000 -0.284 -0.136
ar.L3 0.0259 0.021 1.218 0.223 -0.016 0.068
ar.L4 0.0380 0.021 1.798 0.072 -0.003 0.079
ma.L1 -0.5708 0.053 -10.721 0.000 -0.675 -0.466
ar.S.L11 1.3828 0.123 11.248 0.000 1.142 1.624
ar.S.L22 -0.6426 0.113 -5.704 0.000 -0.863 -0.422
ma.S.L11 -1.3113 0.129 -10.172 0.000 -1.564 -1.059
ma.S.L22 0.5617 0.118 4.752 0.000 0.330 0.793
sigma2 633.9330 10.608 59.760 0.000 613.142 654.724
Ljung-Box (Q): 144.13 Jarque-Bera (JB): 1569.63
Prob(Q): 0.00 Prob(JB): 0.00
Heteroskedasticity (H): 0.96 Skew: 0.58
Prob(H) (two-sided): 0.55 Kurtosis: 6.21
Warnings:
[1] Covariance matrix calculated using the outer product of gradients (complex-step).
Split tha data into train and test set

In [ ]: df.reset_index(inplace=True)
In [ ]: df.head()
Out[ ]:
Date Monthly Mean Total Sunspot Number
0 1749-01-31 96.7
1 1749-02-28 104.3
2 1749-03-31 116.7
3 1749-04-30 92.8
4 1749-05-31 141.7
In [ ]: train=df[(df.Date.dt.year<1958)]
test=df[(df.Date.dt.year>=1958)]
In [ ]: (df.Date.dt.year>=1958) & (df.Date.dt.year<1968)
Out[ ]: 0 False
1 False
2 False
3 False
4 False
...
3230 False
3231 False
3232 False
3233 False
3234 False
Name: Date, Length: 3235, dtype: bool
In [ ]: test1=df[(df.Date.dt.year>=1958) & (df.Date.dt.year<1968)]

n=len(test1)
In [ ]: ## Taking only 50 test data
In [ ]: model.fit(train['Monthly Mean Total Sunspot Number'])
Out[ ]: ARIMA(maxiter=50, method='lbfgs', order=(4, 0, 1), out_of_sample_size=0,

scoring='mse', scoring_args={}, seasonal_order=(2, 0, 2, 11),
start_params=None, suppress_warnings=True, trend=None,
with_intercept=True)
In [ ]: forecast=model.predict(n_periods=n, return_conf_int=True)
In [ ]: forecast_df = pd.DataFrame(forecast[0],index = test1.index,columns=['Predictio

n'])
In [ ]: pd.concat([df['Monthly Mean Total Sunspot Number'],forecast_df],axis=1).plot()
Out[ ]: <matplotlib.axes._subplots.AxesSubplot at 0x7f6ed50090f0>
The result may not seem accurate but, note that time series forecasting is not
reasonable for many time step ahead. It may be valid only for 1,2 or few more
time step ahead in future.
An Approach to check how correct is our model

Making a list for training data call it 'history'
The model will predict next day's sunspot value with this data.
Later the corresponding test value is appended to training data and next day's value is predicted again. This
is repeated for all the test data. and plotted
From AutoArima, we have already got the parameters--> p,d,q.Using it directly on ARIMA model
In [ ]: from statsmodels.tsa.arima_model import ARIMA
In [ ]: history = [x for x in train['Monthly Mean Total Sunspot Number']]
In [ ]: test=[x for x in test['Monthly Mean Total Sunspot Number']]
In [ ]: predictions = []
lower_list = []
upper_list = []
for t in range(len(test)):
model = ARIMA(history, order=(4,0,1))
model_fit = model.fit(disp=0)
output = model_fit.forecast() # The number of time step ahead prediction o
ut of sample from the end of the sample. Default it is 1
yhat = output[0] #
lower = output[2][0][0] # lower bound for 95 % confidence interval for pre
dicted value
upper = output[2][0][1] # upper bound for 95 % confidence interval for pre
dicted value
predictions.append(yhat) # appending predicted value in prediction
lower_list.append(lower)
upper_list.append(upper)
obs = test[t] # taking t_th data from test as 'obs' and appendin
g it in 'history' list which is used for training
history.append(obs)
#print('predicted=%f, expected=%f' % (yhat, obs))
In [ ]: from sklearn import metrics

error = metrics.mean_squared_error(test, predictions)
print('Test MSE: %.3f' % error)
Test MSE: 551.199
In [ ]: error =metrics.mean_absolute_error(test, predictions)

print('Test MAE: %.3f' % error)
Test MAE: 17.350
In [ ]: # plot
plt.plot(test,color='black',label='test set')
plt.plot(lower_list,color='red',label='lower bound of 95 % confidence interva
l')
plt.plot(upper_list,color='green',label='upper bound of 95% confidenc interva
l')
plt.plot(predictions,color='orange',label='predicted for test set')
plt.legend()
plt.show()
This much for this module.
Dont forget to follow me for more such stuff.
https://www.linkedin.com/in/ramendra-kumar-57334478/
(https://www.linkedin.com/in/ramendra-kumar-57334478/)
HAPPY LEARNING!!!

Time Series

Uploaded by

Copyright:

Available Formats

Time Series

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Time Series

Uploaded by

Copyright:

Available Formats

10/25/2020 Sunspot_Time_Series_Data

In [ ]: !pip install pmdarima # Package for AutoArima

In [3]: !pip install plotly

In [2]: import matplotlib as mpl

Unnamed: 0 Date Monthly Mean Total Sunspot Number

In [5]: from dateutil.parser import parse

In [7]: df = pd.read_csv('sunspots.csv',usecols=['Date','Monthly Mean Total Sunspot Nu

1. Exploratory Data Analysis

Profit of datetime formated data:

In [11]: df_non_index['nth_year'] =[int(str(i)[3]) for i in (df_non_index.Date.dt.year

Plotting the data using seaborn boxplot

In [12]: fig, axes = plt.subplots(3, 1, figsize=(20,15), dpi= 80)

Eplanation of above plot:

Returing back to dataframe 'df' and Making Date column as index

In [15]: df.plot(grid=True) # plots in pandas itself take index as x axis, here it is d

In [16]: df_2018=df.loc['2000':'2010'] # Slicing all data from 2000 to 2010

In [17]: df_2018=df.loc['1900':'1920'] # Slicing all data from 1900 to 1910

In [18]: import plotly.express as px

Buttons options in plotly

In [19]: fig = px.line(df_non_index, x='Date', y='Monthly Mean Total Sunspot Number', t

10y 20y 30y 40y 50y all

Comparison of two consecutive 11 year: How to choose from where to where?

In [ ]: df_11_1985=df_non_index[(df_non_index.Date.dt.year>=1985) & (df_non_index.Date

plt.plot(x, df_11_1985['Monthly Mean Total Sunspot Number'],label='df_60_1998'

Out[ ]: Text(0.5, 1.0, 'Comparison of Two consecutive 11 year')

Checking the distribution by making histogram and kde plot

Out[ ]: Text(0.5, 1.0, 'KDE')

2. Checking Stationarity of Time Series Data

constant mean (For different time slots)

Perform Augumented Dickey-Fuller test:

Null Hypothesis - Series is not stationary

In [ ]: from statsmodels.tsa.stattools import adfuller

In [ ]: data_series=df['Monthly Mean Total Sunspot Number']

In [ ]: print('Results of Dickey-Fuller Test:')

Results of Dickey-Fuller Test:

KPSS test for stationary: This another test

In [ ]: from statsmodels.tsa.stattools import kpss

In [ ]: stats, p, lags, critical_values = kpss(df['Monthly Mean Total Sunspot Number'

p-value is greater than the indicated p-value

In [ ]: print(f'Test Statistics: {stats}')

Test Statistics: 0.1435419872086975

Note: For Non-Stationary data: First make it stationary

3. Modelling Time Series

Rolling Average OR Simple moving average = (t + (t-1) + (t-2) + ... + (t-n)) / n

In [ ]: df['Monthly Mean Total Sunspot Number'][:200].plot() # Checking for only first

Out[ ]: Text(0.5, 1.0, 'Rolling Mean & Standard Deviation')

Weighted moving average

Weighted moving average = (tweighting factor) + ((t-1)weighting factor-1) + ((t-n) *

In [ ]: ## Making a function for calculating weighted average which is passed through

In [ ]: df['Monthly Mean Total Sunspot Number'][:200].plot() # Checking for only first

Out[ ]: <matplotlib.legend.Legend at 0x7f6ee75800b8>

Exponential moving average\Exponential Smoothing

In [ ]: df['Monthly Mean Total Sunspot Number'][:200].plot() # Checking for only first

Out[ ]: <matplotlib.legend.Legend at 0x7f6ee7eb1860>

Providing alpha for Smoothing

In [ ]: df['Monthly Mean Total Sunspot Number'][:200].plot() # Checking for only first

Out[ ]: <matplotlib.axes._subplots.AxesSubplot at 0x7f6ee7f8efd0>

Plotting All together and comparing

Monthly Mean Total Sunspot Number Rolling mean ... E_W_A E_