Department of Computer Science

TDT4259 - Applied Data Science

Group Assignment

Decreasing Criminal Activity Across the

City of Chicago Through Data Science

Group:
Group 6
Author:
Lars Gundersen
Janna Salte Hovstad
Sebastian Sundet Flaen
Mons Erling Mathiesen
Mathilde Skeide Ruth
Mathis Tolnæs

November 20, 2020

Table of Contents

List of Figures ii

List of Tables ii

1 Introduction and Problem Definition 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 The Team . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 Background 3
2.1 Definition of Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2.2 Decrease criminal activity across the city of Chicago . . . . . . . . . . . . . . . . . 3

2.3 Choosing the Data Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Description of Data Strategy - CRISP-DM . . . . . . . . . . . . . . . . . . . . . . . 5

3 Method 7

3.1 Description of Data Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3.2 Methods and Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.3 Demographics and relevant statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 8

4 Analysis 9
4.1 Analysis on Decrease number of occurrences within the most common crimes across
Chicago . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4.2 Analysis on Decrease the fastest growing crimes in the areas of Chicago with the
highest crime rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

5 Limitations 19
5.1 Dataset Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

5.2 Method limitation - Impact of Significant Abnormalities . . . . . . . . . . . . . . . 19

5.3 Method limitation - Adjust for Seasonalities . . . . . . . . . . . . . . . . . . . . . . 20

6 Interpretations and Recommendations 21

6.1 Recommendations for Decreasing number of occurrences within the most common
crimes across Chicago . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

6.2 Recommendations for Decreasing the fastest growing crimes in the areas of Chicago
with the highest crime rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
6.3 Further analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

i
Bibliography 25

Appendix 26

List of Figures

1 Impact of Covid-19 and protests on crime rates . . . . . . . . . . . . . . . . . . . . 9

2 Narcotics in Chicago in 2020 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3 Most common crimes in 2020 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

4 Weekly and monthly trends for most of common crimes . . . . . . . . . . . . . . . 11
5 Long term forecast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

6 Short term forecast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

7 Confidence intervals of crimes forecast . . . . . . . . . . . . . . . . . . . . . . . . . 13
8 Accuracy of forecast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
9 Criminal Damage 2020 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

10 Most increasing crimes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

11 Change in Narcotics in largest areas from 2017 to 2019 . . . . . . . . . . . . . . . . 16
12 Devolopment of Narcotics in Humboldt Park . . . . . . . . . . . . . . . . . . . . . 16

13 Heatmap of Narcotics in Humboldt Park . . . . . . . . . . . . . . . . . . . . . . . . 17

14 Location distribution of Narcotics crimes in Humboldt Park from 2017 to 2019 . . 17
15 Descriptions of Narcotics crimes in Humboldt Park from 2017 to 2019 . . . . . . . 18

List of Tables

1 Monthly total crime count maximum and minimum since 2010 . . . . . . . . . . . 8

2 Implementation plan for Decreasing number of occurrences within the most common
crimes across Chicago . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3 Implementation plan for Decreasing the fastest growing crimes in the areas of
Chicago with the highest crime rates . . . . . . . . . . . . . . . . . . . . . . . . . . 23

ii
1 Introduction and Problem Definition

1.1 Introduction

Many American cities have high crime rates which impose negative effects on the public’s health,
safety and economy1 . The city of Chicago is regularly listed as one of the most violent cities in
the United States[Bleakley, 2019]. Chicago was assigned a safety index of 7, by The Neighborhood
Scout, on a scale where 100 was considered the safest2 . This result implies that Chicago is safer
than only 7% of all U.S. cities2 . As a consequence, the Chicago Police Department is faced with
a number of challenges fighting crime in a city characterized by internal socio-economic and racial
divisions[Bleakley, 2019]. David O. Brown, the Superintendent of the Chicago Police, stated in
the 2019 annual report that a data driven decision methodology will be further implemented in
the force3 . Brown points out the necessity of a strategic approach for improved decision making
relying on verifiable information3 .

An optimized resource allocation strategy within the Chicago Police Department has never been
more important. According to an article in the Chicago Tribune, mayor Lori Lightfoot proposed
a plan to remove over 600 unoccupied positions in the department. The goal of this plan was to
contribute in reducing or closing the city’s $1,2 billion budget hole. During the same City Council
Budget Hearing, police officials revealed a 41% case clearance rate up against 50% at the end
of the previous year. The Chief of Detectives, Brendan Deenihan, expressed his wish to see the
clearance rate trend upwards4 . Enhanced and informed planning procedures combined with data
driven decision making can be an essential resource for the department in order to achieve the goal
of increasing the clearance rate with more limited staff.
Chicago is the third largest city in the United States and consists of 77 different community
areas5 . The frequency of crime incidents varies within the different areas1 . According to a study
conducted by Troy Hernandez, five community areas had an increase in crime incidents in 20196 .
The decision making process concerning the distribution of the police department’s resources needs
to consider the varying criminal activities and their development throughout the city. The scope
of this report will therefore take advantage of more area specific data in order to better address
specific improvements within the different community areas.

1.2 Problem Definition

This report will focus on providing analytical results that can contribute to optimized resource
allocation for the Chicago Police Department (CPD). The following part will present the questions
to be answered using data analytics. These aspects will give a clear picture of where and which
improvements should be made. The goal of this report is to create actionable, specific insights
which can help reduce crimes in Chicago. Insights are made more actionable by partitioning the
city’s crime situation into more specific cases, areas and types. An area-level analysis will provide
specific enough results that can be directly addressed by the department.

Data analytics can contribute to answer the following questions:

1. What are the most common crime types in Chicago?
2. Are there any annual, monthly or diurnal trends for the number of occurrences of the different
crimes types?
3. Which community areas of Chicago are the most exposed to different crime types?
4. At which locations do the different incidents of each crime type most commonly occur?
1 https://www.cityofchicago.org/- Accessed 8th of November (13:15)
2 https://www.neighborhoodscout.com/- Accessed 8th of November (12:45)
3 https://home.chicagopolice.org/- Accessed 8th of November (11:10)
4 https://www.chicagotribune.com/- Accessed 9th of November (11:00)
5 https://www.chicago.gov/- Accessed 8th of November (09:00)
6 https://troyhernandez.com/- Accessed 9th of November (14:00)

1
1.3 Outline

Section 2 of this report describes the background and the objectives for this assignment. The data
strategy and methodology will be further explained with a specific focus on the Cross-Industry-
Standard-Process-of-Data-Mining(CRISP-DM) methodology. Section 3 explains the origin and
content of the data set for the analysis and elaborates on any relevant methods and tools used.
Section 4 contains the resulting plots from the data analysis in addition to a description of the
findings. Section 5 discusses the limitations of the dataset and the analysis performed. Finally,
Section 6 summarizes the findings, presents an implementation plan and recommendations for the
Chicago Police Department. The recommendations are based on the results obtained from the data
analysis and proposes improvement measures for the department’s future work. It also presents
specific ideas for future analysis in order to further reduce crime rates across the city of Chicago.

1.4 The Team

The team that produced this report consists of six students that shared several roles and respon-
sibilities. All the team members showed interest in all aspects of the report and the tasks were
distributed thereafter. Initially, descriptive analytic ideas were brainstormed within the team. The
team was subsequently divided into two separate groups, focusing on the actual implementation of
the data analysis. The larger data analysis tasks were separated into smaller, more specific tasks,
with the purpose of producing plots to possibly unveil interesting aspects of the data. Further,
the team decided which parts of the analysis that would be the main focus of the report. The
team members taking charge of the data analysis were Janna Salte Hovstad, Sebastian Flaen and
Lars Gundersen. Janna focused on exploring data and description features to understand pat-
terns and create insights regarding how criminal activity is distributed across the city of Chicago.
Sebastian assisted Janna with visualizations and created a general forecasting model. Lars used
this information to map how different crimes were trending over time. Prior to the analysis, the
problem definition and business initiatives were defined by Mathilde Ruth, Mons Mathiesen and
Mathis Tolnæs. The necessary parts of the report and its content were then strategically planned
and produced. Results were assessed, material produced, and findings evaluated. The fact that
these processes were conducted in parallel with the analysis facilitated efficient work as it provided
required information to the data scientists.

2
2 Background

In order to effectively tackle challenging business objectives the use of data driven decision making
is emerging as the go-to strategy. With an ever increasing data production it is essential for
companies to be able to use data to their advantage. As Provost and Fawcett [2013] describes,
the data that companies are producing is of such a volume and complexity that manual analysis
is not viable. The labor intensity of manual analysis outweighs possible returns, resulting in an
increasing interest in the concept of data science.
Despite the fact that there is a monetary incentive for many commercial companies to advocate
data science in order to reach objectives, an increasing amount of public institutes are following
suit. The public sector is now recognizing data as a strategic asset, and in turn using it to create
public value[OECD, 2019]. This is done by taking advantage of an already existing routine of
extensive data recording.

2.1 Definition of Objective

In recent years the CPD has initiated a department that leverages data analytics in order to ”deploy
resources across the city” and ”conduct crime analysis for strategic planning”7 . The concept of
predictive policing is preferable in comparison to the act of reacting to criminal acts[Perry et al.,
2013]. The city of Chicago is subject to a large variety of crimes happening across the city. With
a reported forty thousand crimes annually8 , the CPD allocate large resources to decreasing the
amount of criminal activity across the city.
Predictive policing originated in 2008 by Police Chief William Bratton, from the Los Angeles Police
Department (LAPD), presented to acting directors of the Bereau of Justice Assistance (BJA) and
the National Institute of Justice (NJI)[Perry et al., 2013]. The concept was an initiative in order to
benefit from mathematical, predictive analytics in law enforcement with the purpose of identifying
potential criminal activity[Rienks, 2015]. The Federal Bureau of Investigation (FBI) reports that
the crime in Santa Cruz, California was reduced by 19% over a 6-month period10 , after implement-
ing the practice of predictive policing. This experiment used data driven algorithms to increase
effectiveness in predictive policing, with no additional variables, meaning neither any increase in
hires nor any increase in shift lengths. Further, the LAPD performed their own experiments using
the same algorithm. The standard procedure, which this experiment was compared to, include
standard crime analyses performed by the LAPD through hot spot methods9 . Reports from the
experiments show that the algorithm turned out to be twice as effective compared to the LAPDs
standard hot spot methods. 10 .

2.2 Decrease criminal activity across the city of Chicago

These facts form the basis for the overall objective of this project namely to help the CPD decrease
criminal activity across the city of Chicago. Even though the amount of crimes in Chicago has
been decreasing the last couple of years, the CPD’s budget is higher than ever. After adjusting
for inflation the CPD is spending almost three times as much, per capita, on policing compared to
numbers from 196411 . This means that even though the city of Chicago is experiencing a decline in
number of crimes, they still need to improve efficiency of resources. This is particularly important
in recent times due to the Covid-19 situation creating large deficits that will force the CPD spend
their resources more effectively. Through the analyses in Section 4 it will become evident that
there is sufficient data signaling an opportunity to develop predictive policing in Chicago.
The overall objective of this project is to decrease criminal activity across the city of Chicago.
7 https://www.tableau.com/
- Accessed 5th of November (15:11)
8 https://home.chicagopolice.org/- Accessed 5th of November (15:34)
9 https://whatworks.college.police.uk/ - Accessed 15th of November (19:48)
10 https://leb.fbi.gov/ - Accessed 9th of November (15:50)
11 https://www.injusticewatch.org/ - Accessed 9th of November (16:22)

3
Considering the complex scope of this overall objective, it is preferable to produce objectives with
a narrower scope. These will be central to the success of the overall objective. The first objective
is to decrease the number of occurrences of the most common crimes across Chicago. Combining
this with the other objective of decreasing the fastest growing crimes in the areas of Chicago with
the highest crime rates will lead to a significant decrease in the overall criminal activity across
Chicago. These two objectives will be the focus of this report.

2.2.1 Decrease number of occurrences of the most common crimes across Chicago

As presented in Section 2.2 there is a substantial amount of crimes happening in Chicago on a

yearly basis. By accurately tracking the seasonalities of the top most common crimes, the CPD
could take advantage of how some crimes have a tendency to increase in certain months of the
year, and even days of the week. Further, the CPD could recognize some periods of the year
as being more prone to certain crimes, and produce training and education strategies to address
these. By taking advantage of the time stamps in the data set, the police department could build
models showing how the most common crime in Chicago is happening in a certain time interval,
and strategize accordingly.
In order to decrease the number of common crimes, a trend analysis and forecasting model based
on the data features’ timestamp and type can be created. The outcome of the objective is to
deliver specific recommendations for today’s situation.

2.2.2 Decrease the fastest growing crimes in the areas of Chicago with the highest
crime rates

Preventive work in order to reduce the number of crimes occurring is a well established practice
across the police industry. With insights into the historical data of criminal activity in Chicago,
it is possible to map out the areas with the highest crime rates. Using this insight to focus police
strategies on high activity areas, the police force have the opportunity to both effectively catch
criminal acts and work in a preventive manner. As mentioned in Section 2.1 predictive policing is a
highly effective strategy. The CPD should take advantage of data through data science techniques
and create insights regarding what crimes are increasing, and in which areas. Strategy development
could use these insights to prevent the growth of trending crimes in areas with already high crime
rates contributing to a significant decrease in criminal activity.
Naturally, if the CPD is able to curb the development of quickly growing criminal types, the crime
scene will become more controllable and predictable. Time, place, and criminal type are all features
utilized to achieve the objective. The goal of the objective is to give the CPD continuous updates
on trend developments of certain crimes, with their associated locations. With this information,
the police department can better plan their preventive work. The trend analysis can be used as an
estimate for the future crime scene, if no changes are implemented. This prediction can be used
as a measure for success in the future.

2.3 Choosing the Data Strategy

Choosing the right data strategy for a given project is essential in the process of finding valuable
results. Since different industries have different objectives, values and opportunities, it is essential
to choose suitable tools and frameworks for the project12 . Since businesses are becoming more
reliant on data driven business development, as mentioned in Section 2, the focus on choosing the
right data strategy has increased. The most popular strategies are CRISP-DM, Sample-Explore-
Modify-Model-Assess(SEMMA) and the Business Analytics Methodology(BAM)13 . The strenghts
and weaknesses of these three strategies were considered in the process of finding the right data
strategy for this project. The three data strategies mentioned are all strategies which aim at
12 https://www.tiempodev.com/ - Accessed 11th of November (14:31)
13 https://www.datascience-pm.com/ - Accessed 11th of November (14:41)

4
providing guidance in all types of data science activities. CRISP-DM was quickly favored as
one of the preferred alternatives in the selection process, since it is well documented and has
a good reputation. The strategy is known to have weaknesses like being somewhat rigid and
documentation heavy. However, with a close-knit team and a thorough understanding of the
framework, the project was designed to counteract these possible downsides. An example of a
counter-measure was to have rapid iterations with sufficient, but not extensive, documentation.
SEMMA, similarly, has an even more narrow focused technicality. The steps used for business
understanding in CRISP-DM are replaced with data sampling in SEMMA. As the project aim
to support the CPD, a business understanding was considered necessary and BAM, on the other
hand, was recognized as being quite the opposite of SEMMA. Where SEMMA has a very technical
strategy, BAM was considered not technical enough. As Hindle and Vidgen [2018] explains, BAM
is not a prescription but rather a logical structure for analytical activities. CRISP-DM emerged the
preferred method as it reflects many of the positive aspects of both SEMMA and BAM. Considering
that SEMMA is quite technical, and BAM is not technical enough, the perfect middle ground was
CRISP-DM - the data strategy chosen for this project.

2.4 Description of Data Strategy - CRISP-DM

CRISP-DM14 is a framework that supports data scientists in realizing value in data mining projects.
The CRISP-DM framework consists of 6 main steps, which cover everything from project initial-
ization to recommendation delivery.

This strategy complements methods used by the CPD to reduce criminal activity. The strategy
department of the CPD will be able to allocate resources more efficiently due to new insights
concerning which crimes are most probable to happen in the future, and in which areas. Using data
insight to patrol areas that are more probable of being the location of a criminal act, the strategy
team can assign more frequent patrols accordingly. Another method that would be enhanced
through the insights from this analysis would be how the training of police officers would be
planned. As the Bureau of Justice Assistant15 describes, it is very effective to review lessons prior
to the occurring of an event. Similarly, the CPD could use information of seasonality of a given
crime to review lessons on the handling of these situations - as the analysis of data gives them an
indication of the increase in a given crime.

Business Understanding
In order to acquire valuable insights into the situation of a business, concrete objectives are es-
tablished. Well-defined objectives add a purpose and direction for data scientists. In business
terms, an objective is not always directly translatable to a data mining problem. Therefore the
data science team relies on reformulating the objective to a problem with listed data sources and
a defined result or outcome for the analysis.

The background and plan to achieve the business objectives are described in Section 2.2.
Data Understanding
After establishing concrete objectives it is necessary to explore and familiarize the data. The data is
typically acquired by the data science team, but as the CPD already has substantial data reporting
systems in place this process was not necessary for this project. Regardless, all data collected from
the real world have some flaws and inaccuracies. The familiarization-phase should identify data
quality problems and give initial insights of potential indications of trends. If possible, the team
can construct a set of hypotheses based on initial indications in the data.
The data set produced by the CPD is a collection of seven million rows consisting of 22 columns
describing all recorded criminal activity in the city of Chicago the last 22 years. Each row sum-
marizes a specific crime, with location, time and type of crime being the most relevant to this
analysis.
Data Preparation
14 https://thinkinsights.net/ - Accessed 10th of November (15:18)
15 https://bja.ojp.gov/ - Accessed 15th of November (20:14)

5
Data should be appropriately prepared to facilitate an effective data analysis. Most data science
tasks rely on data from multiple sources requiring extensive cleaning before the analysis can begin.
The Chicago Crime data set provides all necessary data to produce recommendations for the
objectives set. Other data sources were considered, but data sets from other sources like the FBI
were found to be of little relevance since these crimes are usually interstate and would not be
addressed by the CPD.

Inconsistent labeling practices have been one of the biggest setbacks for this project. As long as
the team was aware of the inconsistency, it is possible to make reasonable judgments. The real
challenge lies in the possible inconsistencies that was not discovered.
Modeling
The analysts selected the relevant techniques for the defined data mining problems. The selection
of models depended on the data that was available. It is not unusual to revisit the step of data
preparation as team members realize errors and missing data.
Trend modeling and analysis is the foundation of this project for the CPD. The weekly and monthly
trends contribute to giving valuable insights into how the citywide crime scene is changing. Fore-
casting is used as a tool for future prediction of criminal activity.
Evaluation
Evaluate the models built in the previous step and select the best-performing ones. Best performing
models can be measured by how they answer the business objectives. A key aspect of the CRISP-
DM framework is how it encourages analysts to revisit the business objectives multiple times during
the project. A common issue in data science is an analysis’ failure to be of value to the initial
objective. In addition to model evaluation, the results are scrutinized and the process as a whole
is reviewed.
The ongoing pandemic is one of the main discussion points when it comes to the evaluation of the
used methods. Forecasting and trend analysis is challenging when the data suddenly experiences
irregular changes. Even more so, 2020 has been a year impacted by the ”Black Lives Matter”-
movement which resulted in surging burglary rates due to looting. Both Covid-19 and large
demonstrations are considered unusual circumstances and can explain irregularities in the data
set.
Deployment
Finally, the deployment, has two main objectives. Firstly, the interpretation of the results from
the analysis needs to be translated into concrete recommendations for the business. For example,
a step-by-step plan containing action points with success measures can be presented. Secondly, the
deployment phase handles the arrival of new data. Securing successful data flow from collection,
through preparation and analysis, is a crucial step to ensure long term value.

6
3 Method

3.1 Description of Data Set

According to Kaggle (2020): ”The data set reflects reported incidents of crime (with the exception
of murders where data exists for each victim) that occurred in the City of Chicago from 2001 to
present, minus the most recent seven days. Data is extracted from the Chicago Police Department’s
CLEAR (Citizen Law Enforcement Analysis and Reporting) system. In order to protect the privacy
of crime victims, addresses are shown at the block level only and specific locations are not identified.
This data includes unverified reports supplied to the Police Department. The preliminary crime
classifications may be changed at a later date based upon additional investigation and there is
always the possibility of mechanical or human error. Therefore, the Chicago Police Department
does not guarantee (either expressed or implied) the accuracy, completeness, timeliness, or correct
sequencing of the information and the information should not be used for comparison purposes
over time”16 .

Attributes and features

The data set consists of 22 columns, whereas five are used in the analysis. Some columns contain
overlapping and redundant information. For example, there are 9 different features (Latitude,
Longitude, Location(lat, long), Northing and Easting in local coordinate system, Block id, District
ID, Ward ID, Area ID, Location Description) related to the location of the crime. Of those, latitude
and longitude are used for the creation of heat maps and Area ID is used in area-specific analysis.
All features used in this project are listed below:

• Time
• Location Description (i.e. School, parking lot, ...)

• Area (Area ID in Chicago)

• Location (latitude, longitude)
• Primary Type (i.e. Battery, Theft, ...)

• Description (i.e. Possesion, Aggravated)

Quality, Criticism and Scrutiny

Even though the CPD have made efforts to generalize the data with proper labeling, inconsistencies
occur throughout the data set. A common inconsistency is the use of abbreviations. For example,
Poss and Possession both indicate a possession type crime, but the labeling is inconsistent. In-
consistencies are not of concern, as long as the analysts manage to address them accordingly. The
data set quality is further discussed in Section 5.1.

3.2 Methods and Tools

Analysis
All the data was analyzed using Pandas and Prophet in Python. Pandas is a fast and popular
open source data analysis and manipulation tool, tailor-made for processing and analyzing large
amounts of data, like the data set used. Prophet is Facebook’s cutting-edge open-source forecasting
library[Taylor and Letham, 2017]. The main benefits of Prophet are evident when forecasting
human behavior. Facebook utilize the forecasting library themselves, to predict user activity on
the platform.
Both Pandas and Prophet were chosen by the analysts in the team based on previous experiences
with similar projects.
16 https://www.kaggle.com/ - Accessed 7th of November (10:56)

7
The most basic implementation of Prophet is used for this project. The forecasting model is
composed of:

• ”A piecewise linear or logistic growth curve trend. Prophet automatically detects changes in
trends by selecting changepoints from the data”17 .

• ”A yearly seasonal component modeled using Fourier series”17 .

• ”A weekly seasonal component using dummy variables”17 .

It is sufficient for the reader to know that Facebook Prophet is an effective method of forecasting
future behavior based on statistics from the past.

Link to code written is attached in the Appendix.

Visualization
The majority of visualizations in this report are line graphs or bar charts. Visualization theory
suggests limiting the number of variables[Mikalef, 2020], or lines in a line graph, to avoid overloading
short term memory. Line plots in this report are limited to five different lines, allowing readers to
get intuitive impressions of trends.
As Mikalef [2020] suggests the colors and plots have been carefully chosen based on best practices.
All quantitative data follows the same sequential colors, and categorical data have qualitative
schemes. The colors were picked from ColorBrewer 2.018 as recommended. The colors are also
chosen in order to take into account people that are color blind, as this affects approximately 10%
of men19 .

3.3 Demographics and relevant statistics

Demographic data has purposely been left out of the Chicago Crime data set to protect the
anonymity of individuals. Racial profiling in criminal statistics have previously led to civil unrest
and received criticism20 . The general trend in Chicago is a decline in crime rates and a slight
decline in population21 .

As crime rates have declined since 2010, the most active criminal month, July of 2011, is seen at
the start of the period of interest(2010-2020). On the other hand, the month which experienced
the fewest amount of reported crimes was April of 2020, with 25% less reported crimes than any
other month. The strict regulations and recommendations implemented because of the Covid-19
situation explains the unusually low amount of criminal activity experienced in April.

Extreme Month Count

Minimum 2020-04 12754
Maximum 2011-07 33256

Table 1: Monthly total crime count maximum and minimum since 2010

17 https://https://research.fb.com/ - Accessed 12th of November (14:09)

18 https://colorbrewer2.org/ - Accessed 15th of November (16:21)
19 https://www.colourblindawareness.org/ - Accessed 15th of November (17:38)
20 https://en.wikipedia.org/wiki/Stop-and-frisk_in_New_York_City - Accessed 14th of November (15:14)
21 https://en.wikipedia.org/wiki/Chicago - Accessed 14th of November (15:30)

8
4 Analysis

This section will present the analysis of crimes in Chicago for each of the two objectives. The
analysis follows the approach of CRISP-DM where results are presented through a logical structure
and interpreted step-by-step.
Noticeably, when examining the data, it was quickly discovered how different 2020 had been com-
pared to the previous years, as a direct consequence of the Covid-19 situation. Figure 1a shows
how the crime rates declined in March 2020. While most crimes are decreasing, there are some
specific crimes in specific areas that increase significantly and which deserves a mention. The two
peaks of crimes incidents in Figure 1a are consequences of the George Floyd Protest and the Black
Lives Matter Protest by a high degree of certainty. Figure 1b shows how burglaries significantly
increased in both June and August 2020, as looting became a serious issue22 .

(a) Total crimes in Chicago since 2017 (b) Burglaries in Chicago in 2020

Figure 1: Impact of Covid-19 and protests on crime rates

Most crimes have declined due to the pandemic and Figure 2 shows the drastic decrease in Narcotics
in March, when the city was forced into lockdown.

Figure 2: Narcotics in Chicago in 2020

Analyzing the impact of the Covid-19 situation and the protests should be a task for further anal-
ysis. This and other limitations of the model are described in Section 5. Faced with a challenging
year in terms of analyzing trends, each precondition made will be described throughout. The
predictions incorporate the year of 2020 as the model perform well in adjusting for seasonalitites
like protests and adjusting to the possible impacts and restrictions set as a result of the Covid-19
situation. However, when defining the most increasing trends it was deemed necessary to expect
2021 to be more comparable to 2019 than 2020.
22 https://www.chicagotribune.com/ - Accessed 15th of November (21:06)

9
4.1 Analysis on Decrease number of occurrences within the most com-
mon crimes across Chicago

Figure 3: Most common crimes in 2020

Reaching the objective of decreasing the occurrences of the most common crimes first required a
process of identifying them. Figure 3 shows the crimes with the most occurrences in 2020. Battery,
Theft, Criminal Damage, Assault and Deceptive Practice was observed to be the most common.
These will be the crimes of interest for the rest of this objective.
The next task was to analyze these crimes individually to see if there existed characteristics in
”when” and ”where” these incidents happen, the first to be addressed was ”when”.

4.1.1 When do crimes occur?

Figure 4 was created with Facebook Prophet to present trends for the most common crimes. From
Figure 4a it is evident that different crimes occur more frequently at different days of the week.
Both Criminal Damage and Battery are most common during the weekends, while Assaults happen
more frequently on Wednesdays. All of this makes up a strong argument for where to distribute
resources.

10
 (a) Weekly

(b) Monthly

Figure 4: Weekly and monthly trends for most of common crimes

The same can be said about the monthly trends, which are presented in Figure 4b. Deceptive
Practice is the criminal action least affected by season, as it only has small variations, while other
crimes has a clear peak during the summer months.
Continuing the analysis on trends, a time series model by Facebook Prophet was used to forecast
when each type of crime were likely to occur in the future. This model is tailor-made to adjust for
seasonalities like protests and is adjusted to potential restrictions due to the Covid-19 situation.
Because of this, Figure 5 presents the crime rate predictions for the next months, with a high degree
of accuracy. The short ”pointy” waves have periods of four or five per month and represents weekly
trends, while the longer waves are monthly. The reader is encouraged to validate that the shapes
of the weekly trend in the forecast in Figure 4a are similar to those in Figure 5.

11
 Figure 5: Long term forecast

With a scope of the next two weeks Figure 6 shows that Batteries have a high prediction on 7th
and 8th November with over 120 predicted incidents. This comes as no surprise as people are more
likely to end up in a fight late at night during the weekend. An interesting aspect is that the CPD
now has a tool for quantifying likely amounts of each crime for each day, through this analysis.

Figure 6: Short term forecast

Figure 7 shows the expected minimal and maximum crimes for the next two weeks plotted with its
confidence interval of 95%. This means that the model is 95% confident that the amount would
not exceed these values. For November 8th it can be 95% assumed that the number of Batteries

12
will be above 95 and below 150, while Criminal Damage will not exceed 55 by the same degree of
certainty.

Figure 7: Confidence intervals of crimes forecast

By dividing the data into two separate data frames, training frame and test frame, it is possible
to measure how well the predictions fit the real observations. The forecast in Figure 8 is only
based on the training data frame no later than October 31, this is in order to predict the test data
on Batteries of the following week. As this week has passed the observed values were plotted, as
the red line, and predictions can be observed for accuracy by how well they lay inside the grey
confidence interval.

13
 Figure 8: Accuracy of forecast

With some insights on when crimes could occur, the focus is shifted to the ”where” of the criminal
activities.

4.1.2 Where do crimes occur?

By using Pythons library Folium, heat maps of specified crimes in all community zones and neigh-
borhoods across the whole city can be visualized. Figure 9 is an example of Criminal Damage for
the whole city of Chicago in 2020. It is clear to see that some parts of the city are especially ex-
posed, those being the clusters of the darkest shades of red. These clusters are especially valuable
in regard to predictive policing.

Figure 9: Criminal Damage 2020

14
4.2 Analysis on Decrease the fastest growing crimes in the areas of
Chicago with the highest crime rates

Figure 10: Most increasing crimes

This analysis was initiated by identifying which crimes are increasing. Figure 10 shows the five
most increasing crimes from 2017 to 2019 and it is observed that Narcotics has been increasing
with the largest amount with a 30% increase from 2017 to 2019. A probable explanation to the
decrease in 2020 is the Covid-19 situation, however there are still two months left of the year at
the time of the production of this report. As people stay more at home there will be a natural
decrease, but there might be a correction if the situation transitions back to how it used to be.
Next, a search through areas was necessary to see which contribute the most to this trend.
An interesting aspect was Narcotics in Humboldt Park, as observed from Figure 11 to be the
one large area with the most increasing cases of Narcotics. Figure 12 shows the development of
Narcotics in Humboldt Park since 2017. We observed narcotics to increase by 80% in two years
time.

15
 Figure 11: Change in Narcotics in largest areas from 2017 to 2019

Figure 12: Devolopment of Narcotics in Humboldt Park

Below are all crimes involving Narcotics in Humboldt Park plotted in a heatmap. It can be observed
that Narcotics is more increasingly frequent in some areas, with areas close to West Chicago Ave
being the most prominent. Note that it is possible to perform similar analyses on all other areas
and crimes as well.

16
 (a) 2018 (b) 2019

Figure 13: Heatmap of Narcotics in Humboldt Park

So far Narcotics in Humboldt Park has been identified to be the most drastically increasing crime,
and plots show in which areas they do occur. Followingly, there were performed a deeper analysis
of the locations within these areas. What types of locations that were most incidental in having
Narcotics related crimes in Humboldt Park were explored. This analysis aims to provide insights
regarding the need for specialized measures based on the most common locations for the different
crime types. Figure 14 shows the distribution of narcotic crimes.

Figure 14: Location distribution of Narcotics crimes in Humboldt Park from 2017 to 2019

From Figure 14 it can be observed that the main contributor to the increasing amount of Narcotics
is incidents in the streets, which increases a lot from 2017 to 2019.
Further, the more detailed description of the crimes is addressed to better understand the devel-
opment of the different types of crimes within the Narcotics category.

Figure 15 shows that the most frequent type of Narcotics related crimes are possession of white
heroin and manufacturing and delivering of white heroin. These results are more evenly distributed
with more than 100 incidents in 2019 for all descriptions. All the plotted types have experienced
a large increase from 2017 to 2018. Between 2018 and 2019, four out of the seven types in total
experience an increase.

17
Figure 15: Descriptions of Narcotics crimes in Humboldt Park from 2017 to 2019

18
5 Limitations

Some limitations can be identified for the dataset, the analysis methods and the overall data
analysis. This section further address and discuss these limitations which involve inconsistencies
in attributes and trends. As introduced in Section 2.4 (as part of the CRISP-DM method), it
is important to evaluate the method used in the analysis by identifying aspects that could have
altered results.

5.1 Dataset Limitations

The Chicago Crime dataset contains all reported incidents of crime and includes several relevant
features. Nevertheless, it could be further improved to increase associated potential and analytical
possibilities. One limitation is the lack of standardization which is present in some features. Some
descriptions are labeled as ”OTHER” and there exists more than one label with identical meanings
in both the Crime Description attribute and the Location Description attribute. These aspects
can make some results deviate from reality. Results would be misleading if the crimes which are
labeled as ”OTHER” were supposed to be labeled within the same category. An equal spread
of the original categories of the ”OTHER” labeled crimes could signify that results are more or
less representative for reality. Although somewhat imprecise, the results can still provide useful
insights. Another limitation involves the possibility of human errors. Crime events are reported by
individuals which may mislabel events by either creating wrong formatted labels or by mistyping
when recording an incident.
The quality of data could be improved by implementing some relatively low effort changes. The
reporting of events in the database should be more regulated, with stricter restrictions. One
suggestion is to implement a new set of fixed labels. To make sure the old crime data will remain
applicable, the new labels could be based on the most common, existing ones. An officer should
have to choose from these existing labels when reporting new crimes and the instantiation of new
labels should be centralized and limited to certain personnel with the required authority. In the
event of an officer not being able to find an appropriate label, the reporting should be flagged for
evaluation.
These changes would make sure all data features are clear and representative, and thereby facilitate
precise, reliable and trustworthy analyses. The improvements would be noticeable a couple of years
after implementing the standardized, fixed formats, that is, when sufficient new data points have
been reported. Specific addresses are not present in the data set for privacy reasons, which makes
it harder to improve this limitation.

5.2 Method limitation - Impact of Significant Abnormalities

The method of the conducted analysis is also subject to improvement and further potential is
identified for future analysis. The Facebook Prophet forecasts and predictions were based on data
from 2010 to the present. One limitation is related to the fact that 2020 was subject to the Covid-19
pandemic which involved nation-wide quarantines and caused businesses to shut down. This year
may be seen as an exception compared to previous trends and patterns, and could have affected
the results. The Covid-19 situation can be classified as an unnatural event which could disturb
the predictions made by Facebook Prophet.
The year of 2020 is excluded in the rest of the analyses due to these special circumstances. This
decision is based on a consideration of the Covid-19 situation as a single, one-time event, where
society is expected to return to pre-Covid-19 operation in the near future. If 2020 were included
in the analysis of specific crime trends in an area, the results would be misleading. An interesting
possibility would be to analyze the reported crimes from March to December 2020 to predict
scenarios and create insights which could be used in the event of similar pandemics in the future.
Further analysis could create three different data frames including, excluding or focusing on the
period of the Covid-19 pandemic to investigate the given time period’s impact on results. Trends

19
in crimes are equally important in 2020. This unique time, with unique trends and patterns, should
be investigated in more detail.

5.3 Method limitation - Adjust for Seasonalities

The predictions and forecasts in the report were conducted on a weekly, monthly and yearly
basis. The accuracy of the predictions could have been improved by including seasonalities like
holidays and weather forecasts. Resource allocation would, for instance, benefit from knowing if
less crimes happen on Christmas Day. Similarly, including weather forecast data could provide
insights regarding which crimes are most likely to happen. One hypothesis could be that less
Narcotics is sold on the streets when it is raining. Another seasonality considered was including
the effects of large sports events, with the hypothesis that large sporting events on television would
mean less people out on the streets doing crimes, as presented in [Copus and Laqueur, 2018].
It is important to admit to the limitations of the results to help the CPD use and interpret the
results in a suitable way. The presented results provide limited, but useful insights which can serve
as an example of what can be achieved. These results can contribute in convincing the CPD to
invest time and resources in improving data quality and analyzing data. Recommendations based
on these insights are presented in the following section.

20
6 Interpretations and Recommendations

Similar to Section 4 this section is initially divided into the two objectives and handled separately.
For both objectives all recommendations were constructed by interpreting insights from the anal-
ysis. Implementation plans were then created for each objective and includes relevant stakeholder,
time-frame and success criterion’s for each recommendation. Lastly, more general recommenda-
tions are described in the future analysis section where both a monitoring system and utilizing
route optimization are suggested as future work.

6.1 Recommendations for Decreasing number of occurrences within the

most common crimes across Chicago

The analysis made in Subsection 4.1 can be summarized into four main findings on the given
objective:

1. Theft, Battery, Criminal Damage, Assault and Deceptive Practice were the most common
crime types in Chicago.
2. Four out of the five most common crime types saw an increase in occurrences during the
summer months from May to September. The exception was Deceptive Practices.
3. Battery and Criminal Damage experienced an increase in number of incidents during the
weekends.
4. A tool for forecasting future crimes was developed, with thresholds for different crimes for
each day, week and month.

6.1.1 Recommendations

Four out of five crime types were subject to an increase during the summer months. The first
recommendation therefore suggests increasing police officers presence in the streets during these
months. Increasing the CPD’s visibility in the streets could reduce the numbers of crimes commit-
ted as described in [Braga et al., 2019].

The forecasts presented in Figure 6, predict the daily variation in occurrences of the different crime
types. The forecasting model will give accurate indicators of number of expected incidents. As
previously mentioned, Battery and Theft were the top two crime types in Chicago in 2020. These
should therefore be monitored most carefully. The second suggested recommendation is therefore
to allocate the number of police officers ready to respond to Theft and Battery incidents based on
forecasts.

Theft is discovered to be the most common crime type in Chicago in 2020. A specific recommen-
dation to be implemented by the CPD is to create campaigns that raise awareness of how to make
residences less prone to theft. This could be found beneficial in order to properly educate the
public about the issue and avoid future incidents.

Lastly, since Battery is the second most common crime type in Chicago specialized measures focus-
ing on reducing the number of incidents should also be implemented. It is therefore recommended
that police officers make school visits to educate teenagers on the dangers of Battery. This is seen
as preventive in order to possibly reduce the frequency of occurrences for the future.

6.1.2 Implementation plan

Table 2 contains all the previously mentioned recommendations as well as relevant stakeholders,
time-frame and success criteria. The recommendations are sorted based on their time frame and
importance.

21
 Recommendation Stakeholders Time-frame Success criteria
Increase number of Decreasing reported
police officers in the Jan 2021- incidents by
Patrolling officers
streets during the summer Jan 2023 15% from
months 2021-2022
Allocate number of police officers
Decrease reported
ready to respond to theft Staffing, Jan 2021-
incidents by 10%
and battery incidents patrolling officers Jan 2023
from 2021-2022
based on daily forecasts
Strategy department/
Initiate school visits to educate program development, Decrease reported
Jan 2021-
teenagers on the dangers school resource officers, incidents by 5%
Jan 2023
of assault and battery police chief, from 2021-2022
school principal
Create campaigns to raise awareness Decrease reported
Jan 2021-
of how to make residents less prone Public relations department incidents by 10%
Jan 2024
to theft from 2021-2023

Table 2: Implementation plan for Decreasing number of occurrences within the most common
crimes across Chicago

The first two recommendations will likely give results that can be observed on the short-term.
Recommendation number three and four on the other hand represent more long term efforts and
therefore have a longer time-frame.
Another focus for the recommendations was to involve different stakeholders in order to distribute
the workload evenly and effectively achieve the success criteria. In this way it is easier for the
CPD to delegate tasks in the nearest future since the recommendations can be handled by many
different departments and employees.

6.2 Recommendations for Decreasing the fastest growing crimes in the

areas of Chicago with the highest crime rates

The analysis made in Subsection 4.2 can be summarized into five main findings on the given
objective:

1. Narcotics, with 30% increase was the one crime increasing the most in Chicago in the past
three years.
2. Humboldt Park was the main contributor to these trend, as occurrences increased by over
600 a year or by 80% from 2017 to 2019.
3. Heatmaps presented what areas where most affected, and clear trends where observed towards
more concentrated rates of Narcotics in the areas close to West Chicago Avenue.
4. Incidents reported at the streets is the main driver behind the large increase in reports.
5. Incidents involving heroine and crack are the drug types increasing the most.

6.2.1 Recommendations

These findings were used to prepare concrete recommendations of actions to the CPD.

Firstly, as heatmaps show the occurrence of narcotic-related crimes, the CPD is recommended to
implement more systematic police patrols in Humboldt Park. Spending more time being present
in areas most prone to criminal activity involving Narcotics would have a preventive effect[Braga
et al., 2019].

22
Secondly, the CPD is recommended to prioritize more resources into Narcotics investigations in
Humboldt Park. Narcotics considering heroin and crack were observed to be the main forces
driving the Narcotics crime rates and focusing on stopping manufacturing and delivering these
drugs would decrease the total Narcotics crime numbers.
Lastly, the CPD is recommended to create an expert group coordinating with other police de-
partments with successful history in dealing with heroin and crack crimes. This group should
back up the work done by local Narcotics investigators and has the responsibility of exploring and
implementing innovative techniques for investigation.

6.2.2 Implementation plan

Recommendation Stakeholders Time-frame Success criteria

Systematic police patrols in Decreasing reported
Jan 2021 -
concentrated areas of Narcotics Patrolling officers incidents from 2021
Jan 2023
heatmap in Humboldt Park to 2022 by 10%
Increase number of Narcotics Decreasing reported
Staffing, training, Jan 2021 -
investigators working with heroin incidents from 2022
police investigators Jan 2024
and crack in Humboldt Park to 2023 by 10%
Create an expert group Staffing, other Decreasing reported
Jan 2021-
on heroin and crack police departments, incidents from 2022
Jan 2024
investigation police investigators to 2023 by 10%

Table 3: Implementation plan for Decreasing the fastest growing crimes in the areas of Chicago
with the highest crime rates

Table 3 summarizes the recommendations with relevant stakeholders, time-frame and success cri-
teria. Note that more Narcotics crimes are expected to be reported as a result of more patrolling
of certain areas. Therefore the success criteria is not defined as simply measure the amount of
reports in 2021 and observe if this number decreased from 2019. Remember the assumption made
in the Section 4 that 2021 will be more similar to 2019 than 2020 because of the pandemic. Success
for the first recommendation should therefore be defined by a decrease in Narcotics reported from
2021 to 2022 as it will take some time to measure the actual impact.
The second and third recommendation takes longer to implement and the time-frame for the
implementation is set to be from January 2021 to 2024. Heroin and crack are shown to be a
trending problem and investigations should step up quickly. The second recommendation is of
course partially dependent on the third as the expert group is meant to consult the investigators
into being more efficient, but the investigations are recommended to start of independently, while
the expert group is formed.

6.3 Further analysis

The Chicago Crime data set used in predictive policing has potential to increase the efficiency
of the CPD[OECD, 2019]. A suggestion for future work and further analysis is to explore the
alternatives within an automatically updated monitoring system. The monitoring system could
implement the models used in this project. Accurate and informative insights could be given in
”real-time” as models are updated with new data daily. The purpose of implementing these models
in a dashboard is to give everyone in the CPD access to recommendations and basis for decision
making. As discussed in Section 4, example models have been applied to one specific area and
recommendations are given to that area. The general model is applicable to all areas of Chicago,
with little to no modification.
The monitoring system can be a dashboard, used by police officers that have planning and strategic
responsibilities. One functionality of the dashboard will be workforce size recommendations. A

23
key dashboard feature is customizability. Depending on the user, the dashboard should display
different information. For example, the officer with staffing responsibility for Humboldt Park would
be presented the recommended workforce-size for that area. In addition, the dashboard could
display forecasts and trend developments in areas and criminal types, also dependent on the users
requirements or interests. The resulting product will, in desired intervals, produce reports that
compare current trends and crime rates to historical data. An implemented dashboard could for
example be used by the strategy department. Guidelines on how to incorporate recommendations
in current plans and routines will also be a part of the platform. This will give the CPD an
opportunity to verify how a change of strategies could yield the desired result of decreasing criminal
activity across the city of Chicago.
Another analysis able to complement the recommendations would be using optimization to suggest
optimal route planning for vehicles. The required level of detail for route planning optimization
is higher than what has been examined in this report, and is therefore recommended for future
analysis. Route planning would optimize time spent patrolling the most exposed areas, and would
increase effectiveness of resource allocation, complementing the other actions recommended.

24
Bibliography

Paul Bleakley. A thin-slice of institutionalised policebrutality: A tradition of excessive force

in the chicago police department. Criminal Law Forum, 30(4):425–449, 2019. doi: 10.1007/
s10609-019-09378-6.
Anthony A. Braga, Brandon Turchan, Andrew V. Papachristos, and David M. Hureau. Hot spots
policing of small geographic areas effects on crime, Sep 2019. URL https://onlinelibrary.
wiley.com/doi/full/10.1002/cl2.1046.

Ryan Copus and Hannah Laqueur. Entertainment as crime prevention, 2018.

Giles A. Hindle and Richard Vidgen. Developing a business analytics methodology: A case study
in the foodbank sector. European Journal of Operational Research, 268(3):835–851, Aug 2018.
doi: 10.1016/j.ejor.2017.06.031.
Patrick Mikalef. Data visualization - design principles (powerpoint slides), 2020.

OECD. The application of data in the public sector to generate public value. 2019. doi: https://doi.
org/https://doi.org/10.1787/1ab27217-en. URL https://www.oecd-ilibrary.org/content/
component/1ab27217-en.
Walter Perry, Brian Mcinnis, Carter Price, Susan Smith, and John Hollywood. Predictive policing:
The role of crime forecasting in law enforcement operations. 2013. doi: 10.7249/rr233.
Foster Provost and Tom Fawcett. Data Science for Business:. OReilly, 2013.
Rutger Rienks. Predictive policing: Taking a chance for a safer future, 2015. URL https://
issuu.com/rutgerrienks/docs/predictive_policing_rienks_uk.

Sean J. Taylor and Benjamin Letham. Forecasting at scale. 2017. doi: 10.7287/peerj.preprints.
3190v2.

25
Appendix

See https://github.com/larshgu/Chicago Crime for code used in this project.

26