Saint-Petersburg State University

Mathematics and Mechanics Faculty

Albina Yezus

Predicting outcome of soccer matches

using machine learning
Term paper

Scientific adviser:
Alexander Igoshkin, Yandex Mobile Department

2014
1 CONTENTS
2 Abstract .............................................................................................................................................................................. 3
3 Introduction ..................................................................................................................................................................... 4
4 Problem statement ........................................................................................................................................................ 4
5 Approach ........................................................................................................................................................................... 4
5.1 Choosing match set that is to be analyzed ................................................................................................. 4
5.2 Deciding on key features. .................................................................................................................................. 5
5.3 Data extraction ...................................................................................................................................................... 5
5.4 Testing various machine learning algorithms .......................................................................................... 5
5.5 Improving implemented algorithm .............................................................................................................. 5
6 Algorithms ........................................................................................................................................................................ 5
6.1 Match data ............................................................................................................................................................... 5
6.2 Key features ............................................................................................................................................................ 6
6.2.1 Feature formulas......................................................................................................................................... 6
6.2.2 Plots.................................................................................................................................................................. 7
6.3 Data extraction ...................................................................................................................................................... 8
6.4 Machine learning algorithms........................................................................................................................... 9
7 Technology ....................................................................................................................................................................... 9
7.1 Programming language...................................................................................................................................... 9
7.2 Environment ........................................................................................................................................................ 10
7.3 Libraries ................................................................................................................................................................. 10
8 Difficulties ....................................................................................................................................................................... 10
8.1 Difficult data extraction ................................................................................................................................... 10
8.2 Layout changes.................................................................................................................................................... 10
8.3 Feature evaluation ............................................................................................................................................. 11
8.4 Classifier evaluation .......................................................................................................................................... 11
9 Results .............................................................................................................................................................................. 11
10 Further research...................................................................................................................................................... 11
11 Related work ............................................................................................................................................................. 11
12 Conclusion .................................................................................................................................................................. 12
13 Links ............................................................................................................................................................................. 12
14 References .................................................................................................................................................................. 12
2 ABSTRACT
In this study the methods of machine learning are used in order to predict outcome of soccer
matches. Although it is difficult to take into account all features that influence the results of the
matches, an attempt to find the most significant features is made and various classifiers are tested
to solve the problem.
Keywords: machine learning, sporting prediction, soccer, data mining
3 INTRODUCTION
The aim of this work is to see if it is possible to predict the outcome of sport games with good
precision. It is to be done by analyzing soccer matches of various football leagues. Firstly, it is
crucial to choose features that seem to be significant carefully and analyze their influence on
matches outcome. Secondly, using machine learning methods, such as KNN, Random Forest, logistic
regression, SVM and others, the model is to produce an output representative of the probable
outcome of the match.
Several attempts were made to create a model that would be able to predict the outcome of the
games with good precision; however, it has appeared to be utterly difficult to succeed in this field.
Thus, the study is considered to be successful if it predicts the outcome with the precision of 70%.

4 PROBLEM STATEMENT
To create a model that predicts outcome of soccer matches with sufficient accuracy. Sufficient
accuracy means one of following:
1. 70% accuracy when predicting results of match set;
2. Making profit betting against bookmakers.

5 APPROACH
It is possible to divide all work into two steps:
1. Choosing match set that is to be analyzed;
2. Deciding on key features;
3. Data extraction;
4. Testing various machine learning algorithms;
5. Improving implemented algorithm.
These steps may be intermingled, as feature’s significance can be evaluated basing on the results of
algorithms that are applied later.

5.1 CHOOSING MATCH SET THAT IS TO BE ANALYZED

The problem of choosing between different leagues is not an easy one as the following problems
should be avoided:

1. Unfair refereeing;
2. Match-fixing;
3. Difficult or impossible data extraction.
5.2 DECIDING ON KEY FEATURES.
In order to create bedrock for future research the initial model is created. The features of this
model are chosen to fit the following conditions
1. What seems legit?
2. What is possible to extract?
3. What really matters?

5.3 DATA EXTRACTION

Data extraction task is solved by using data from websites that store information about league
which is chosen to be observed. The following information should be available:

1. Match information;
2. Season information;
3. Result table for each moment in the season.
The data is extracted by parsing html source of pages with necessary information.

5.4 TESTING VARIOUS MACHINE LEARNING ALGORITHMS

As the problem is one of classification problems where result equals one of 3 values {0, 0.5, 1}
which mean loss of the fist team, draw match and win of the first team respectfully, several
methods can be applied. The methods used in this term paper are:
1. K nearest neighbors;
2. Random forest;
3. Logistic regression;
4. Support vector machine.

5.5 IMPROVING IMPLEMENTED ALGORITHM

It may be useful to improve implemented algorithm in accordance to the stated problem. As it is not
yet decided which algorithm suits the problem best, this part of the research is not done yet.

6 ALGORITHMS

6.1 MATCH DATA

The subject of the research is English Premier League as it seems to lack problems stated before.
6.2 KEY FEATURES
6.2.1 Feature formulas
The initial set of features can be divided into two groups:

1. Static;
2. Dynamic;
Where static features are given for each team and do not depend on the rival and dynamic features
depend on both teams and represent their correlation.
All features are normalized, which means they are from the interval [0; 1].

6.2.1.1 Static features

1. Form:
5
1
∑ 𝑟𝑒𝑠𝑘
10
𝑘=1
where resk - result of the –kth match (value in {0, 1, 2});

2. Concentration:
1– 2 ∗ 𝑥
where x – the nearest match lost to the weak team (the difference between the current team
and that team >= 7);
3. Motivation:
𝑑𝑖𝑠𝑡 𝑡𝑜𝑢𝑟 + 𝑑𝑖𝑠𝑡
min(max (1 – , 𝑑𝑒𝑟𝑏𝑦, ) , 1)
3 ∗ 𝑙𝑒𝑓𝑡 2
where:
o derby – 1 if match is a derby and 0 otherwise;
o dist – distance to the nearest “key position”;
o left – tours left in the season;
o tour – 1 if left < 6 and 0 otherwise;
o key position – value in {1, 2, 3, 4, 5, 6, 17, 18}.

6.2.1.2 Dynamic formulas

4. Goal difference:
1 𝑑𝑖𝑓
+
2 2 ∗ max_𝑑𝑖𝑓
where:
o dif – difference between goals;
o max_dif – maximal difference between goals;
 5. Score difference:

1 𝑑𝑖𝑓
+
2 2 ∗ max_𝑑𝑖𝑓
where:

o dif – difference between scores;

o max_dif – maximal difference between goals;
6. History:
𝑝1 + 𝑝2
2
where pi – result of –ith match (value in {0, 0.5, 1}).

6.2.2 Plots
As it is difficult to estimate feature significance at this point, it may be possible to reveal
dependences based on plots. The results are not very encouraging. Still, they are interesting to look
at.
It is not difficult to see that the higher the score difference and goal difference, the more matches
are won. On the other hand, it is difficult to find correlation between form difference and match
outcome, so it is a good thing we have separated those two features.

6.3 DATA EXTRACTION

The information was extracted from the following websites:

1. http://www.championat.com/
2. http://www.statto.com/

From champinat.com it was easy to loop over matches for each season and get information about
form, concentration and history by parsing page with match information.

Statoo.com was useful as it has result table for each date of the season, so information based on
scores, positions etc was extracted from there.
The collected dataset has the following representation:

6.4 MACHINE LEARNING ALGORITHMS

For the algorithms described below implemented classifiers from python library were used.
The methods applied with the resulting precision:

1. K nearest neighbors: 0.558

2. Random Forest: 0.634
As these methods are academic and are not usually used in real researches, not much effort was put
into them. Considering this fact, the results are quite encouraging.

7 TECHNOLOGY

7.1 PROGRAMMING LANGUAGE

Python is the main programming language of the project as there are many useful libraries
included, that simplify data extraction and training/testing machine learning classifiers much
easier.
The following versions of Python were used:
1. Python 3.3;
2. Python 3.4.
Python 3.3 was used as it is the latest available version on Wakari.
Python 3.4 was used as it is the latest available version of Python.
7.2 ENVIRONMENT
There are quite few GUIs for Python. As this paperwork is the result of collaboration, there is a need
of something that allows several people to have access to the source code. Thus the following
environments were used:
1. Wakari;
2. PyCharm.
Wakari allows to create notebooks in which it is easy to launch various part of the code and write
report on the go, so it is handy for writing classifiers.

PyCharm simplifies git sync and library installation, supports TODO syntax and has many other
useful features.

7.3 LIBRARIES
The following libraries for Python were used:
1. For extraction:
a. Selenium;
b. Urllib;
c. BeautifulSoup;
2. For maching learning algorithms:
a. Sklearn;
3. For data analysis and everywhere else:
a. Numpy;
b. Pandas;
c. Matplotlib;
d. Etc.

8 DIFFICULTIES
During the work on the research several problems were encountered.

8.1 DIFFICULT DATA EXTRACTION

Although chosen features required simple information to be extracted, it was not always a
straightforward task. The main problem was with statto.com, where information is stored in
javascript and cannot be accessed simply by following the link. It was necessary to send a request
that changes the page layout so that it can be parsed. This problem has been solved by applying
selenium library that simplifies the process.

8.2 LAYOUT CHANGES

In the middle of the work championat.com layout changed. This difficulty was overcome by
adjusting code to new layout.
8.3 FEATURE EVALUATION
As only a couple of classifiers were applied, it is difficult yet to evaluate significance of each feature.
Moreover, no methods were applied to select the most significant once. This problem is to be solved
in the future by looking into classifiers’ results with less / other / modified features and applying
various methods, such as Lasso.

8.4 CLASSIFIER EVALUATION

It is not yet clear how to try the classifiers against the bookmakers. This problem is to be solved in
the future when better classifier is found.

9 RESULTS
The following result has been got:
1. Dataset with 9 features and 640 objects;
2. Scripts for simple data extraction;
3. Academic machine learning classifiers applied;
4. Results of classifier over 60%.

10 FURTHER RESEARCH
This work has recently been started and although there are already some results obtained, there is
still a long way ahead. The directions of the further investigation are:
1. Varying data. By this we understand varying size of the dataset, number of features and
formulas of features.
2. Training and testing classifiers. As only academic classifiers were yet tested, it is
necessary to apply other methods, such as SVM and Logistic regression.
3. Adjusting the best classifier. This means either rewriting it in accordance to the stated
problem or writing some kind of combination of different classifiers.
4. Beating bookmakers. This is the ultimate goal of this paperwork. The point is to find an
algorithm that would be able to make profit by betting.

11 RELATED WORK
As it was mentioned before, attempts to create a model that would successfully predict outcomes of
sport events were made before.
Concerning football one piece of research (Hamadani, 2006) offered approach similar to the current
study. In that study 3 seasons were tested and it was discovered that each season a different set of
features had more significance which means either that football is an unstable game or that optimal
features lie somewhere between those found by the study.
Another study (Blundell, 2009) has proved that American Football matches can be accurately
modelled using features within a regression model. It was also discovered that simple logistic
model could achieve just as accurate forecasts compared to some more complex alternatives.
However, all aforementioned studies researched long periods without taking into account a third
factor. This study, on the other hand, focused on a particular sport league without limit to matches
in this league only – results of the players in other leagues are also to be taken into account. Thus
the results of this study are expected to be more accurate.

12 CONCLUSION
Machine learning methods can be applied to different fields, including sports. On the example of
English Premier League it is shown that it is possible to find a classifier that predicts the outcome of
soccer matches with the precision of more than 60%. However, there is still a lot of work to be done
and the research will be proceeded.

13 LINKS
The source code may be found by the following links:

 Work related to machine learning: https://wakari.io/sharing/bundle/Albinutte/Project;

 Work related to data extraction: https://github.com/Albinutte/football-prediction.

14 REFERENCES
[1] Predicting the outcome of NFL games using machine learning, Babak Hamadani, cs229 -
Stanford University (2006)

[2] Numerical Algorithms for Predicting Sports Results by Jack David Blundell, School of Computing,
Faculty of Engineering (2009)
[3] Predicting football results using Bayesian nets and others, machine learning techniques, A.
Joseph ¤, N.E. Fenton, M. Neil (2006)
[4] Predicting Margin of Victory in NFL Games: Machine Learning vs. the Las Vegas Line, Jim
Warner (2010)