Football Match Prediction Using

Machine Learning
Author: Fredrik Sjöberg - 42011

Supervisor: Sepinoud Azimi Rashti

Master’s thesis in Computer Engineering

Åbo Akademi University, Finland
Faculty of Science and Engineering (FNT)
Information Technology
2023
Abstract
Football is the most popular sport in the world. It has become so popular since it is
easy to play and affordable. Football can be played both inside and outside almost
at any place in the world and little equipment is needed to play it. The rules are
simple and easy to understand. Since football is so popular, matches and
tournaments exhibit a high level of availability for both men and women at any age
worldwide. Therefore, it is not surprising that professional football constitutes a
colossal market. The most interesting aspect of football is of course the results of
the matches and how they can be affected to go in favour of one’s team. This is
where predicting the results of the football matches is useful.

Machine learning is an excellent method for making predictions. In this thesis

multiple machine learning algorithms and data sets are compared and analysed to
find the best approach. The higher prediction accuracy a model achieves, the more
accurate its predictions are. When predicting football match results, there are three
possible outcomes: win for the home team, draw, or win for the away team. To
tackle such classification problems using machine learning, multiclass
classification can be used.

One of the most important aspects of building a successful machine learning model
is to find a suitable data set and data features. Feature selection is difficult and can
be made in many ways, but it is decisive for the results of the predictions. Data sets
containing match statistics, team related features, and player related features, were
used in this thesis. They were tested with different machine learning algorithms, to
find the best possible combination.

Keywords: machine learning, multiclass classification, predictions, football

Table of Contents
1. Introduction ................................................................................................................ 1

2. Machine Learning ....................................................................................................... 3

2.1 Overview ............................................................................................................. 3

2.2 Algorithms........................................................................................................... 9

2.2.1 Multiclass Classification ............................................................................ 17

2.3 Problems and Solutions .................................................................................... 18

3. Previous Research ..................................................................................................... 22

3.1 Sport Predictions............................................................................................... 22

3.2 Football Predictions .......................................................................................... 29

4. Implementation ........................................................................................................ 34

4.1 Overview ........................................................................................................... 34

4.2 Data Collection .................................................................................................. 36

4.3 Data Preparation............................................................................................... 40

5. Prediction Models ..................................................................................................... 48

5.1 Model Selections............................................................................................... 48

5.2 Team Attributes ................................................................................................ 49

5.3 Player Attributes ............................................................................................... 53

5.4 Match Statistics................................................................................................. 56

6. Discussion ................................................................................................................. 59

7. Conclusion................................................................................................................. 62

Svensk sammanfattning .................................................................................................... 63

Bibliography ...................................................................................................................... 68
Preface
I would like to thank my supervisor Sepinoud for her guidance and support
throughout the project. I would also like to thank my employer for letting me
prioritise my studies over work.
1. Introduction
Predicting the results of football matches is challenging. It is impossible to be done
perfectly even with the help of machine learning. If it would be possible, the
excitement of the sport would be lost and sports betting organisations would go
bankrupt. One of the primary factors that football has such a large popularity around
the world is that it is an unpredictable sport. In every match, it is possible for a
weaker team to win over a stronger team and no matter how superior the stronger
team is, it cannot be certain that it will win. Seemingly unpredictable events like
the weather and a lucky last-minute goal can be the decisive event that will help an
underdog defeat the giant. However, we know that it is possible to predict football
matches with an accuracy over 90% [1]. In this thesis we investigate which
approaches and strategies are the best to be able to achieve the most accurate
football match predictions using machine learning.

In the recent years, the amount of data available online about football and other
sports have increased massively. This has made it possible for researchers and
hobbyists to develop and improve football prediction methods themselves. The goal
of some people in this area is to find the best prediction methods for making as high
profit as possible in sports betting. However, in such cases the aim is also to find
the weak spots of the predictions of the sport betting organisations, rather than only
finding the best possible match predictions. The aim of this thesis is not to find the
best methods for making profit on sports betting sites, but purely to find out how
predictable the game of football really is and how the best possible predictions can
be made.

In this thesis multiple machine learning algorithms are compared, with the aim to
find the best possible model for football match predictions. Bayesian networks and
logistic regression are the most common algorithms to be used for football
predictions according to previous research papers on this topic. Except these
algorithms, the random forest algorithm was also chosen to be used in this
implementation. The wanted output of the created models are the overall result of
football matches and not the exact score. The three possible outcomes are win for

1
the home team, draw, or win for the away team. Multiclass classification is the most
suitable type of machine learning algorithms to solve such problems.

The second chapter of the thesis gives an overview of machine learning, machine
learning algorithms, and machine learning related problems and solutions. The
focus is on multiclass classification. The third chapter consists of an overview of
previous research in the topic. It also describes how football predictions and sport
predictions in general can be made. The fourth chapter describes the own
implementation done on football predictions. It consists of an overview and the data
collection and data preparation phases. In the fifth chapter the models of the
implementation are described, and the results of the models are visualised and
evaluated. The thesis ends with discussion and conclusion.

2
2. Machine Learning

Machine learning is a subset of artificial intelligence and computer science. It is a

technique that learns from data using algorithms, and tries to develop to improve
its accuracy. This chapter gives an overview of machine learning, machine learning
algorithms, and machine learning related problems and solutions.

2.1 Overview

Machine learning is an artificial intelligence technique that improves the

performance of computer systems of a task by letting them train and learn from
experience. Machine learning algorithms learn to make predictions from
observations and data instead of being programmed exactly what to do. However,
machine learning is still quite narrow, which means the problem must be defined
well so that the machine can focus on that desired activity [2]. A machine learning
model is the result of the training with a machine learning algorithm on a specific
data set. The machine learning algorithms must usually be modified to be used for
each case depending on the input data, which differ from case to case, and therefore
the machine learning models become unique. The goal of machine learning
algorithms and models is to maximize the accuracy of their predictions of the
desired output values [3].

The difference between problems in machine learning and manual programming is

that the algorithm is unknown in machine learning. An algorithm is like a recipe, it
is followed to perform tasks on the input data to generate the output data. In
problems that can be performed using manual programming, both the input- and
output data are known. In other problems the input data is also known, however,
the output data is unknown. Only the expected format of the output data is known,
such as numbers, yes or no, or categorical classifications. In other words, it is not
known how to transform the input data into output data. Machine learning can be
used to solve these types of problems. [4]

The first obvious step in machine learning is defining the problem. The data and
algorithm one needs, depend on what kind of problem is defined. The first practical

3
step in machine learning is data collection. Both the quantity and quality of the data
impact the results of machine learning models. The data set to use in machine
learning can be a mix of one or many data sources collected by oneself or already
existing data sources. The data sources can be, for example, databases, files, web
services, and other APIs. The data itself can contain numerical data, text data,
temporal data, and categorical data. Since the training of machine learning models
is dependent on the input data, some of the most important parts of machine learning
are the choice of what input data set to use and how to divide it into training- and
test data sets. Machine learning models rely on data to be able to know what to do,
the performance of previous models, and possible improvement points [2]. Machine
learning can also handle large volumes of data that is difficult to analyse using
regular data management tools, such data is called big data [2]. By leveraging large
data sets, learning algorithms can achieve higher accuracy, which have made the
use of big data popular in machine learning recently [4].

Since machine learning relies on data, the data set selection and feature selection
are very important parts of it. When the data set has been chosen, the next step is
data preparation, which consists of data cleaning, feature selection, and feature
extraction. Data cleaning is needed so that the values are represented correctly, and
noise and faulty data are removed, for example, missing data and duplicate values.
Feature selection is then done to make the data set more efficient in the machine
learning algorithm or model. Feature selection means that only the most important
features of the data set are selected for further use. This will reduce the training time
of the model, the model will be more comprehensible, and the result will have a
higher level of generalisation. The feature selection can be made based of how much
information the features give and what accuracy the features generate. However,
sometimes it can be difficult to determine whether a feature is unnecessary, it
depends on the other features in the data set and the dependencies between them.
Feature extraction reduces the number of features by combining them, without any
loss of information. A popular technique for feature extraction is Principal
Component Analysis (PCA), which reduces the number of features by combining
them linearly. This also makes the features uncorrelated, but all the important
information is still left. [3]

4
Machine learning works by manipulating data using mathematical functions. Both
the input and the generated output are represented mathematically. The input data
must therefore be converted to mathematical representations and the resulting
output data must be converted back to useful data again [2]. This is the case for text
data, which is often categorical data. This conversion process is called feature
encoding. The most known feature encoding technique is one-hot encoding. One
dummy variable is created for each categorical value in the data set. If an
observation belongs to a category, the dummy variable is set to 1, otherwise 0.
Another simpler feature encoding technique is label encoding. Every categorical
value is converted into a unique random number. However, this can cause problems
in machine learning algorithms which interpret the values as ordinal instead of
categorical.

The next step after the data preparation is to choose a machine learning algorithm
for the model. It is challenging to find the most suitable algorithm for a problem.
The performance of an algorithm depends on the type of data set used. The data set
exhibits varying degrees of compatibility with different machine learning
algorithms. However, the choice is not black and white, multiple algorithms can
perform equally well, and therefore it is usually recommended to try multiple
algorithms. Different types and examples of machine learning algorithms are
described in Chapter 2.2.

When the machine learning algorithm has been chosen, it is time to train the model
with the data set. The input data set must be split into two subsets: a training data
set and a test data set. The training data set should have more data than the test data
set, around 80% training data and 20% test data is a popular split ratio. The machine
learning algorithm is trained using the training data set. After the training episode,
the algorithm is supposed to make predictions on the new unseen data set, the test
data set, based on what it has learned from the training data. The machine learning
algorithms tries to find regularities in the training data to be able to make predictions
on the test data. The regularities are then generalised and built into the model. The
performance of the model on unseen data is called generalisation [5]. The model
must never be trained using test data, nor tested using training data [3]. Both
scenarios would generate a high test accuracy, however, the model has “cheated”

5
and will perform badly on new unseen data. The model has “cheated”, since it has
already seen the data which it is tested on, and therefore it has made a perfect fit for
it.

When the model has been tested, the output values and an overall result of the model
will be received. To be able to know how good the model is, it should be compared
with other models or previous research. The performance of machine learning
models can be evaluated to find its strengths and weaknesses. Usually, the result of
a machine learning model is termed as accuracy, however, it can only be calculated
for classification models. In regression an evaluation metric called R-squared is
used instead. R-squared measures the distance between the data points and the
regression line, which essentially is the error of the predictions. The output value
of the metric is a percentage between 0 and 100%. When the R-squared value is
100%, all the data point is predicted exactly right at the regression line. When the
R-squared value is 0%, the model has not been able to explain any of the variation
in the data around its mean [3]. Usually a higher R-squared value means a better
performance of the regression model. However, what percentage is considered as a
high R-squared value depends on the problem.

For classification models, the most common evaluation metrics are accuracy,
precision, recall and f1-score. A popular visualisation of the performance of
classification models is the so-called confusion matrix. It is a table that contains the
four different classification results: true positives, false positives, true negatives,
and false negatives. True positives are observations that are predicted and actually
positive, false positives are observations that are predicted positive but are actually
negative, true negatives are observations that are predicted and actually negative,
and false negatives are observations that are predicted negative but are actually
positive. In binary classification the size of the confusion matrix is 2x2 (two rows
and two columns), but confusion matrices work for multiclass classification too.
Then the size of the confusion matrix is n times n, where n is the number of classes.
[3, 6]

The accuracy of a classification model is all the correctly classified observations

divided by the total number of observations. The precision is how well the model
has made positive predictions. It is measured as the true positives divided by all

6
positives. The recall is how well the model have classified positives correctly. It is
measured as the true positives, divided by the true positives plus the false negatives.
The f1-score is a measure similar to accuracy. It is the harmonic mean of precision
and recall, and is measured as two times the precision times the recall, divided by,
the precision plus the recall. The f1-score is higher than the accuracy when the
number of observations in the classes are uneven. The output value of all these
metrics are percentages usually represented as decimals between 0 and 1. The
model performs better when the metrics has a value closer to 1. However, one must
be suspicious of a model that produce an accuracy of close to 100%, then there
might have been too little test data or the test data might have been leaked in the
training phase. [3]

One way to evaluate machine learning models is by using the so-called cross-
validation technique. In cross-validation the data set is randomly split into k subsets
of equal size [3]. One of the subsets is selected as test data, and all the other subsets
are used as training data. The test subset selection is repeated k times, so that every
subset has been used as test data. The result is then the average of all the tests. This
technique is called k-fold cross-validation and is the most common cross-validation
technique [3]. Cross-validation will both improve the performance of the model and
the generalisation of the results [6]. Another evaluation metric is the ROC (receiver
operating characteristic) curve [3]. It plots the connection between the true-positive
rate and the false-positive rate. The area under the curve is called the AUROC (area
under ROC curve). The greater AUROC, the better overall performance of the
model [6].

There are several types of machine learning approaches. The three main categories
are supervised learning, unsupervised learning and reinforcement learning.
Supervised learning is when the machine learns from labelled training data [7].
Labelled data is data that are tagged with some labels containing information such
as characteristics and classifications. The machine is given examples of input data
and their corresponding output data. By learning from these examples, the machine
tries to find patterns and using them it creates mathematical models that maps inputs
to outputs [8]. The model can then be used on new unseen data to predict the output
data. The two most common supervised learning model groups are classification

7
models and regression models. Classification models maps the input data into a
limited number of classes that have been predetermined [8]. Regression models
maps input data into a numerical domain with any value within the range.

In unsupervised learning, only the input data is known while the output is not
predetermined [4]. Since the supervisor does not tell the machine how the correct
output data should look like, the unsupervised learning algorithms must find out the
relations in the data themselves. They try to find similarities in the data and patterns
that occurs more often than others. By finding these patterns they can tell what
generally is going to happen with certain input data [4]. Unlike in supervised
learning, the training data used in unsupervised learning are not labelled. There are
not any predetermined classes either. The goal of unsupervised learning is therefore
to find classification labels for the data [8]. The models categorize the input data
into groups called clusters, by searching for commonalities in the data.
Unsupervised classification is called cluster analysis or clustering [9].

Semi-supervised learning is a machine learning approach that can be put in between

supervised learning and unsupervised learning. It is a mixture of them since it uses
both labelled and unlabelled training data. In semi-supervised learning the amount
of labelled data is usually much smaller than the amount of unlabelled data [8]. The
goal of semi-supervised learning is to learn a model to predict classes more
correctly with future test data than a model that has been learned using only labelled
data. Semi-supervised learning reminds of the way humans learn, for children the
environment is full of unlabelled data in the beginning but supervisors teach them
new labelled data gradually. [10]

Reinforcement learning is a machine learning approach which learns by receiving

a picture of the environment in which it operates. The machine has a specific goal
in the dynamic environment and any action that the machine does in it, impacts the
environment. The environment gives the machine feedback, which the machine can
use to learn from and improve itself. [8]

8
2.2 Algorithms

The core of the machine learning model is the machine learning algorithm itself.
There are multiple options of machine learning algorithms for every type of
machine learning approach. However, the choice of machine learning algorithm is
not black and white. Although a problem seems to fit for a particular machine
learning approach, there might be other types of machine learning approaches and
algorithms that works equally well. The tree main types of machine learning
algorithms are regression, classification and clustering [3]. Machine learning
algorithms might consist of a set of rules, a tree structure or a type of network [7].

Regression is a method that shows the relation between variables by creating a

function that fits the data best. The output value is the dependent variable, which is
the effect of the input variables called independent variables. When the independent
variables are manipulated, they will cause a change in the dependent variable that
is measured. Regression algorithms are based on supervised learning. They are
specialised to predict numerical values. An example of a regression problem is to
predict the price of a house. It may learn features of the house and prices of other
similar houses in the same area.

The most known regression algorithm is linear regression. Since linear regression
is a supervised algorithm, it uses labelled training data to train a model to predict
labels of new unlabelled test data. The goal of the algorithm is to find a line that
best fits the dataset by finding relationships and dependencies in the data. This
means that linear regression predicts a continuous target variable, which can be used
to predict the outcome of any unseen test data. The equation of simple linear
regression can be written as a linear algebraic equation (y = a*x + b). However, in
machine learning multiple independent variables are usually used to predict one
single dependent variable. Then the equation can be expressed as in Equation 1,
where y is the dependent variable, x are the independent variables, a determine the
slope and e is the y-intercept. [3, 8]

y = a0 + a1 ∗ x2 + ⋯ + an ∗ xn + e

Equation 1. Linear regression

9
Classification algorithms are also based on supervised learning. Classification
algorithms predict categorical values instead of numerical values as regression
algorithms do. The categorical values to which the input data should be classified
into, are called classes. The classes are finite and predetermined for each problem.
A classification with only two classes is called binary classification. A classification
with three or more classes is called multiclass classification. An example of a
classification problem is to classify images into which animal that are pictured on
them. The classes could, for example, be dog, cat and horse. For a human such a
problem might be simple, but for large quantities of images it would be a boring
and time-consuming task. This is of course why machine learning and computers
in general are useful.

There are multiple popular classification algorithms. Some of them are decision
tree, random forest, support vector machine, k-nearest neighbour, logistic
regression and Naïve Bayes [3]. Decision tree is a classification algorithm that
shows the results as a flowchart, which often looks like a treelike structure. The
decision tree classifies the input data by flowing through the tree [10]. The tree
consists of nodes with edges (branches) between them. The first node has no
incoming edges and is called the root. All the other nodes in the tree have exactly
one incoming edge. The last nodes in the tree have no outgoing edges and are called
leaves. Since each leaf has been assigned to one class, the leaves determine the
output values of the algorithm [8]. Most often the decision tree is a binary tree,
which means that each node has only two outgoing edges. Decision trees use
conditional logic to find the path of an input value to the output value [3]. In each
node there is a condition which determines to which node the data point should
continue to. In a binary tree the condition can be seen as an if-else statement, true
or false, or yes or no. Decision trees can also be used for regression. Then the leaves
are numerical values instead of classes.

Random forest is a classification algorithm which consists of decision trees.

However, the random forest algorithm uses multiple trees instead of just one as the
decision tree algorithm. A random forest trains the trees with random subsets of the
original data set, and the result is determined with a majority vote between the trees.
This makes random forests more generalised than individual decision trees [3].

10
Random forests can also be used for regression. Then the numeric prediction is
made by taking the average of the output of the trees [3]. An example of a random
forest structure and its decision trees is visualised in Figure 1.

Figure 1. A random forest structure where the green nodes create the decision path.

Support vector machine is a classification algorithm that can be used for both
classification and regression problems. It classifies the input data into two classes,
which makes it a binary linear classifier. The data points in a space are separated
into the two classes so that the gap between them is as wide as possible. New test
data is then mapped to a class based on which side of the gap they are on. Support
vector machine can also perform nonlinear classification. This can be done using
the so-called ‘kernel trick’, which projects the data on a higher dimension feature
space. [10]

K-nearest neighbour is a classification algorithm that can be used for both

classification and regression problems. It classifies data points that are close to each
other to the same class. It looks at the k nearest neighbours of a new point, and the
new point is then added to the class which most of its neighbours are belonging to.
K can be any number, but an odd value of k makes the occurrence of a tie less
probable. In case of a tie, the class can be chosen based on the class of the closest
neighbour, randomly, or after decreasing or increasing the value of k. K-nearest

11
neighbour is a simple algorithm, but it can be effective for problems like searching
for similar items. [3]

Logistic regression is a supervised algorithm related to linear regression. The most

significant difference between them is that linear regression is a regression
algorithm while logistic regression is a classification algorithm, despite its name.
Logistic regression predicts categorical dependent variables while linear regression
predicts continuous dependent variables. Logistic regression is usually used as a
binary classification algorithm and therefore the output can be one of two classes,
like true or false, and yes or no. The output value is determined using probabilities
of a data point being in one of the two classes. The probabilities are calculated using
a sigmoid function, which is an activation function. It multiplies each feature with
a weight and then add them up. The output values to which the sigmoid function
maps the input values, must be between 0 and 1. The data points that are lower than
a threshold value, are classified to one of the classes, and the data points that are
higher than the threshold value, are classified to the other class. The equation of
logistic regression can be expressed as in Equation 2, where y is equal to the
outcome probability, and z is a model parameter which includes the x-value, the
midpoint of the curve and the scale of the curve. The resulting probability will
always be in the range between 0 to 1, which is required in a sigmoid function. [3,
8]

1
𝑦=
1 + 𝑒 −𝑧

Equation 2. Logistic regression

Logistic regression can be seen as the cousin of the linear regression. In a sense,
logistic regression is linear regression on which a sigmoid function has been applied
[3]. In Figure 2. the difference between linear regression and logistic regression is
visualised. Logistic regression produces a S-shaped curve which is called sigmoid
curve or logistic curve. Since the output values must be between 0 and 1, input
values that go to negative or positive infinity will have an output value of 0 and 1
respectively. Linear regression on the other hand, simply produces a continuous
line. However, a common problem for both logistic regression and linear regression

12
is multicollinearity [3]. Multicollinearity is when an independent variable is
strongly correlated with another independent variable. It is a problem since it makes
it difficult for the model to find significant effects of changing the independent
variables, although they might be of great importance.

Figure 2. Linear regression on the left, logistic regression on the right.

Naïve Bayes is a classification algorithm and a special case of a so-called Bayesian

network. A Bayesian network is a graphical model based on probabilities. The
nodes in the graph are random variables and the edges are conditional probabilities
of the variables. The dependencies between the random variables are calculated
using Bayes theorem. It describes the conditional probability of an event, which
means the probability of an event happening given that another event has happened
before. The Bayes Theorem is based on previous knowledge of the conditions that
are related to the event. The equation of Bayes Theorem is expressed in Equation
3, where A and B are events. It is expressed as the conditional probability of A
happening when B is true, is equal to, the conditional probability of B happening
when A is true, times the probability of A, divided by the probability of B. [3, 11]

𝑃(𝐵|𝐴) ∗ 𝑃(𝐴)
𝑃(𝐴|𝐵) =
P(B)

Equation 3. Bayes Theorem

The graph of a Bayesian network is a directed acyclic graph (DAG), because of the
conditional probabilities and dependencies between the nodes in the graph. In such
a graph, there are no cycles or self-connections allowed, which means that each

13
edge goes from one node to another without creating a loop. Learning a Bayesian
network can be divided into two main steps: structure learning and parameter
learning. Structure learning is finding out the directed acyclic graph that best
represents the relationships in the data. This can be done by using algorithms like
the DAG search algorithm and the K2-algorithm. In the beginning they give all
DAG structures the same probability, and then they search for the structure that
gives the maximum probability for the data, which is called the Bayesian score.
Parameter learning means learning the distribution of conditional probabilities. The
maximum likelihood estimator is the method used for parameter learning. It
estimates the parameters of the probability distribution by maximising a likelihood
function so that the probability of the data is as high as possible. [11]

Naïve Bayes is a more restricted form of Bayesian network. Naïve Bayes assumes
that all the attributes are independent of each other [12]. The class nodes should
have no parents, and are therefore the root nodes. Each node can only have one
parent and the nodes corresponding to the attribute variables should have no
connections or edges between them [12]. An example of the Naïve Bayes structure
is visualised in Figure 3. The c-node is the class node and the a-nodes are the nodes
corresponding to attribute variables. Naïve Bayes can be used for both binary- and
multiclass classification and it works well with both small and large datasets [3]. It
usually performs better using categorical input data than using numerical input data.
Naïve Bayes is good at predicting data based on historical results [3].

Figure 3. Naïve Bayes Structure

Naïve Bayes is a popular classification algorithm because it is simple to construct

since it doesn´t need structure learning. The classification process is also efficient
since the attributes are assumed to be independent. However, it also creates a
problem when selecting features for the model. Features that are correlated to each

14
other, may not perform well in a Naïve Bayes model. This can be solved by
selecting feature subsets using, for example, the forward selecting method or the
best-first search algorithm [12]. The most significant difference between Naïve
Bayes and logistic regression is the classification type. Naïve Bayes is a generative
classifier, which means that it learns the probability distribution of the data and
understands how the data looks like in general. Logistic regression on the other
hand, is a discriminative classifier, which means that it learns decision boundaries
and divides the data using those boundaries [8].

Clustering is a type of unsupervised machine learning algorithms. Clustering is the

unsupervised equivalent to the supervised classification algorithms. The goal of
them both is to divide the input data into separate classes or clusters, however in
clustering there are no predetermined classes, like there are in classification [9].
Instead of dividing the input data set according to how the training data was divided,
clustering divides the data using statistical methods where data with common
features are grouped into the same clusters. Clustering can also be seen as dividing
the original data set into subsets with similar data [9, 13]. There are different types
of clustering methods, for example partitional clustering, density-based clustering
and hierarchical clustering.

Since the human intelligence is based in our brains, they are an interesting topic in
artificial intelligence and machine learning. Artificial neural networks are learning
algorithms that imitate the functions of biological neural networks in the brain [4].
The brain consists of billions of neurons which are connected to each other. The
neurons are information messengers. They use different frequencies that are
summed when going from neuron to neuron [4]. They can be seen as weights, which
can be represented in a computer as zeros and ones. A neural network consists of
multiple layers. The first is the input layer, in the middle are the hidden layers and
last comes the output layer [3]. The neurons receive their input from all neurons on
the previous layer, and their own output are forwarded to every neuron on the
following layer [4].

As machine learning is a subset of artificial intelligence, deep learning is a subset

of machine learning. Deep learning is based on artificial neural networks, and
‘deep’ means that there are multiple layers used (at least two hidden layers) [3].

15
Deep neural networks woks like artificial neural networks. Each layer combines the
values from the previous layer, and therefore they develop all the time by learning
more and more complicated functions [4]. The abstraction increases for every layer,
which means that unnecessary data that do not help to solve the problem are hidden
or not considered. Deep learning uses hyper parameters, chosen by oneself, to set
up the model. The most important hyper parameters are the number of hidden
layers, the number of neurons in the hidden layers and weight initialisations [3].
Deep learning has become interesting recently since it is not focused to do one
specific thing, but can be used for many different purposes. It can also make use of
large amounts of data, like big data, if the computer systems are powerful enough.
Deep learning also finds the best features from the data set by themselves, which
makes the work simpler [4]. In other types of machine learning, one must do the
feature extraction and feature selection by hand or using algorithms, but deep
learning does the feature extraction automatically. Neural networks are mostly used
for classification problems, but can also be used for regression problems.

Three popular neural network types in deep learning are artificial neural network
(ANN), recurrent neural network (RNN), and convolutional neural network (CNN).
In artificial neural networks, a biological neuron is represented as a mathematical
model called perceptron. It takes features as input and tries to classify them into
classes by separating them with a line or plane. ANN uses sigmoid functions for the
separation, and therefore the ANN algorithm is like logistic regression. However,
logistic regression only uses one perceptron, while ANN uses one for each neuron
in their networks. In addition, the output of ANN is only the class association, but
in logistic regression the probability of a variable belonging to a class is also
provided. RNN is a neural network architecture that can handle sequential data. It
is especially suitable for language translation, speech recognition and text
generation. In an RNN, the neurons can have connections between themselves in
the same layer in addition to connections between the layers [4]. CNN is a specific
type of an artificial neural network, that is especially useful in image recognition.
In ANN, each input is related to only one pixel in the image, but in CNN the spatial
relationship between the pixels is also considered. [14]

16
2.2.1 Multiclass Classification

The most basic classification algorithms are binary classification algorithms, which
only have two classes. Multiclass classification algorithms, which have more than
two classes, are much more complicated. Some multiclass classification problems
may have more than thousand different classes, for example in image recognition
and text translation [15]. Multilabel classification is similar to multiclass
classification, however, in addition to having many classes multilabel classification
can also assign one or more classes to each observation [15]. The difference
between the three types of classifications can be seen in, for example, image
classification. If the images only picture ‘a cat’ or ‘a dog’, it is binary classification.
If the images picture ‘a cat’, ‘a dog’ or ‘a human’, it is multiclass classification.
However, if the images pictures ‘cats’, ‘dogs’ and ‘humans’ and they can be mixed
in a single image, for example ‘a human’ and ‘a dog’ in the same image, it is
multilabel classification.

Many binary classification algorithms can naturally be used for multiclass

classification, while other can be transformed to be used for it. Algorithms that can
be used naturally for both binary- and multiclass classification are decision trees,
k-nearest neighbour, Naïve Bayes and neural networks. Decision trees classifies the
observations using a tree-like structure. Each leaf in the tree represents one class.
In binary classification the decision tree only has two outputs, and therefore two
leaves. However, more leaves can be added, thereby representing multiple classes,
which enables multiclass classification. Since the random forest classifier is built
out of decision trees, it will also work for multiclass classification. [16]

K-nearest neighbour can naturally be used for multiclass classification. New data
points will be added to the category which most of its neighbours belong to, also if
there are more than two classes. Naïve Bayes can also be used for multiclass
classification. The observations will be categorised to the class with the highest
probability of all classes, whether there are two or more classes. In neural networks
the number of output neurons can be increased to enable it for multiclass
classification, so that each output neuron represents one class. Each neuron tells the
probability of the observation belonging to that class. [16]

17
Other binary classification algorithms can be transformed to be used for multiclass
classification using either the One-vs-rest strategy or the One-vs-one strategy. The
One-vs-rest strategy (One-vs-all, OVA) uses one binary classification model for
each class, where the class is put against all other classes. The One-vs-one strategy
(All-vs-all, AVA) uses one binary classification model for each pair of classes. In
both strategies the observations are classified to the class with the highest score.
Logistic regression is a binary classification algorithm that can be transformed
using one of these strategies to be used for multiclass classification. [15, 16]

2.3 Problems and Solutions

To be able to receive improved and valid results, machine learning algorithms are
usually computed multiple times. A so-called epoch occurs each time the whole
training data set is used for training the machine learning model. When the problem
and input data is complex, more repetitions are needed [2]. However, the result
should not radically change even though the input data changes. If it does, the model
has not been good enough and so-called overfitting or underfitting has occurred.

Overfitting and underfitting are the most common causes of poor results in a
machine learning algorithm or model. Overfitting is when the ability of a model to
solve the problem suddenly is not increasing anymore [17]. It occurs in the training
process when the model starts to also learn noise. It means that a model cannot
generalise seen data to unseen data well enough [18]. Overfitting causes the model
to perform well on the training data, but badly on new testing data. Instead of
learning how the data behave, the model has learned all the data by rote. That means
it has also memorised noise and abnormalities in the training data [17]. When the
model encounters new type of information in the test data set, the model will
struggle to understand it according to the procedure learned using the train data set.
So, even if the model is a perfect fit to the training data, it is not guaranteed that the
prediction accuracy of the testing data will be high [7].

The causes of overfitting are usually a too small train data set, too much noise in
the train data set, or too many inputs in the algorithm. There are multiple strategies
for reducing and avoiding overfitting. A method that is good at avoiding overfitting

18
is the so-called early stopping method. Its purpose is to stop the training before the
accuracy of the model stops increasing or starts decreasing [17]. The training
accuracy increases always overtime, but the validation accuracy increases in the
beginning but slows down and eventually starts decreasing in the end. At the same
time the training error decreases overtime while the validation error first decreases
but after a time it starts to increase [17]. The difficult part is to find where the point
is that the accuracy and error rate shifts at. The point can be found by following the
validation accuracy while the training is going on. After each epoch the accuracy is
evaluated and when the accuracy stops improving or starts worsening the training
should be stopped [18]. Often validation is made between the training and testing
in the early stopping method. Validation helps to find the best values for the hyper-
parameters (for example size of neural network). This increases the generalisation
level and reduces the bias while increasing the variance [18]. When the model has
been finalised using the validation data set, finally the test data set is used to test
the final model.

Since noise learning is one cause of overfitting, one solution to reduce the
overfitting is to reduce the amount of noise. That can be done by removing unused
or unimportant data and features. Unimportant features can also be given less
weight than the other features so that they are not taken into consideration as much
as the other ones [18]. Often overfitting can be reduced by expanding the training
data set. Especially if the model is complicated, more parameters are needed to
achieve good results [18]. Small data sets are overall more vulnerable to overfitting
than large data sets [17].

Underfitting is the opposite of overfitting. Instead of having a too good model that
learns all the test data as in overfitting, underfitting occurs when the model is too
simple to represent the variability of the data [17]. Unlike overfitting, underfitting
means that a model performs badly on both training- and test data. The solutions
for reducing or avoiding underfitting is practically like the solutions for overfitting.
To achieve a higher accuracy of the model the important features must be increased
and therefore a larger data set must be used. Another solution is to complicate the
model or increase the training time or number of epochs and iterations.

19
In Figure 4. the different results of a supervised model are visualised. When
underfitting has occurred, the model has not been able to represent the data in an
optimal way. Such a model will have high error rate in both the training- and test
data. When the model is balanced, the accuracy is optimised and the model can be
used on new unseen data still with high accuracy. The model ignores outlier points
and noise. However, when overfitting occurs, the model has been fitted to the
training data too well. It considers every point in the training data, since it has not
been generalised enough. The model will not be able to produce high accuracy on
unseen data.

Figure 4. From the left: Underfitting, balanced and overfitting in supervised

learning.

Overfitting and underfitting occurs especially in supervised learning. However,

there are equivalents to them in unsupervised learning as well, which occurs in
clustering. In unsupervised learning the equivalent to overfitting and underfitting is
the problem of finding the optimal number of clusters to use [9]. The equivalent to
overfitting is when the number of clusters are too many in proportion to the data
points. This will lead to the clusters only including a few data points each, and that
they are not clustered according to their similarities. The equivalent to underfitting
is thus the opposite, which is when the number of clusters are too few so that the
data points are clustered together although they do not have many common traits
between them. The different results of a clustering model are visualised in Figure
5.

20
Figure 5. From the left: Underfitting, balanced and overfitting in unsupervised
learning.

Another cause of overfitting and underfitting are too many inputs in the algorithm.
When there are too many inputs, the model will have a high accuracy on average
but with a low consistency. For some data sets the model may work well, but for
other data sets it may work badly. To solve this, the so-called bias-variance trade-
off must be improved. Bias-variance trade-off means the balance between the bias
and the variance in a model [18]. Since we want the machine learning model to be
generalised and perform as good as possible on new unseen data, the so-called
generalisation error should be minimised [5]. This can be done by keeping the bias
and variance in balance. Bias in machine learning is how much a model learns
unnecessary things, like noise, and makes too many assumptions. Variance in
machine learning is how well the model can handle changes in the training data set
[5]. Bias can be seen as the training error rate and variance as the testing error rate.
It would be optimal to include all the patterns found in the input data, but still have
a high level of generalisation of the data, but they interfere with each other so they
cannot be achieved simultaneously. A model with a low bias and a high variance
will result in overfitting [5]. On the other hand, a too high bias means that the
training error rate has been high and this will result in underfitting. A model with a
low bias is usually more complex, and a model with high bias is usually less
complex [5]. Therefore, the bias-variance trade-off can also be seen as a trade-off
in complexity [18]. Although, these points can help in building a successful model,
everything depends on the data set and the model must be build different from case
to case. The key in building a machine learning model while avoiding overfitting
and underfitting, is to compromise between accuracy and consistency [18].

21
3. Previous Research

Predicting results in sports have always been popular, especially betting on the
winner or exact result. However, in the recent years the ability to predict things
have increased drastically thanks to development in techniques and data gathering.
Using machine learning we can now find relations between different factors that
influence the result of sports, that we were not aware of previously. The increase of
available data about sports, also helps to improve the ability to make predictions.
Nowadays the results of a game or competition can be found in seconds and
statistics of them can be analysed in detail. This chapter describes how sport
predictions can be made in general and how especially football predictions can be
made.

3.1 Sport Predictions

Predicting results in sports are in the interest of various groups of people. Managers
and agents want to predict the results of competitions to be able to change the
outcome to their advantage. The teams or athletes may want to know their status
compared to other competitors ahead of a competition. Fans want to predict the
results to know how well their athlete or team will do. Predicting sport results are
also of high importance in so-called fantasy leagues, which have grown in
popularity recently. They are based on the results of real-world sports, and the
participants receive points based on how well they have predicted the outcome. The
predictions are usually made by choosing a team of players from different teams,
that the participant think is going to do well in the near future. Fantasy leagues are
most often based on team sports, for example football, American football and ice
hockey.

Since the business of sports in general is enormous, predictions of results in sport

are of course also associated with money. Sports betting is a large market globally,
especially thanks to the growth of online betting. It is expected to grow even more
when more and more people around the world can watch sports [19]. The most
popular sports in betting differ from country to country, but overall, the most

22
common ones are team sports like football, basketball, cricket and American
football, racing sports like track cycling and horse racing, and other sports like
tennis, Esports and boxing [19].

Betting is when one predicts a result and then playing money on it. There are two
main types of sports betting, either one must predict the exact score of the game, or
one must predict the overall result of the game, which essentially is who will win.
When predicting the overall result, there are three possible outcomes; win for the
home team, draw, or win for the away team. Home and away are used to explain
which team are playing on their home ground. If the game is played on a neutral
ground, the home team are usually drawn randomly. The three different outcomes
are usually represented as ‘1’ for a home win, ‘X’ for a draw and ‘2’ for an away
win. If the right answer is predicted, one will retrieve more money than the stake
depending on the odds, however if the prediction is wrong the stake is lost.

Betting odds are becoming more and more popular due to online betting, and the
number of sports and competitions available on betting sites are increasing all the
time. Betting odds are said to be the most accurate probability forecast for sports
[20]. It is especially useful to use odds in sports where the probability of the
outcome varies from game to game. In sports or games where the probabilities of
the outcome are even, there are no direct use of odds. Betting odds can be seen as
the potential win relative to the stake.

The probability is how likely it is that an event will happen. The range is between
zero and one, where zero means that the event is impossible to happen and one
means that the event will happen with all certainty. Odds are related to the
probability and can be calculated using a simple formula. The odds are the ratio
between the probability of an event to happen and the probability of an event not to
happen. For example, if the odds of an event to happen are 1:2, it will happen in
one out of three times and the probability of it is 0.33 or 33%.

Betting odds however are usually represented a little bit different from normal odds.
A traditional way of displaying betting odds is using fractional odds. Fractional
odds are the win to be paid out to the bettor in case of a correct bet, relative to the
stake. In case of a win, one will retrieve the stake plus the fractional odds times the

23
stake. This means that only the potential win can be seen from the fractional odds
and not how much one will retrieve in the total with the stake included. For
example, if the fractional odds of a game are 1/2 and the bet is correct, one will
retrieve the stake plus half of the stake as a winning price. The probability of it to
happen is 67%. If the fractional odds of a game are 1/1 and the bet is correct, one
will retrieve the stake two times. The probability of it to happen is 50%. If the
fractional odds of a game are 2/1 and the bet is correct, one will retrieve the stake
three times. The probability of it to happen is 33%.

Betting odds can also be displayed as decimal odds. Decimal odds are more
comprehensible than fractional odds, since they directly tell how many times the
stake that one will retrieve if the bet is correct. Using the same examples as for
fractional odds; if the fractional odds are 1/2 the decimal odds are 1.5, which means
the money one will retrieve if the bet is correct will be 1.5 times the stake. If the
fractional odds are 1/1 the decimal odds are 2, which means the money one will
retrieve if the bet is correct will be 2 times the stake. If the fractional odds are 2/1
the decimal odds are 3, which means the money one will retrieve if the bet is correct
will be 3 times the stake. So fractional odds can be converted to decimal odds by
converting the fractions to decimals and add one.

A bookmaker is a person or organisation that organise sport betting events. For the
bookmakers to make a profit from the betting, they must manipulate the odds to not
match the true probabilities of the outcome. If the odds would be the same as the
probabilities of the outcome, all the money put as stake would be given back as
wins to the bettors, and there would be no profit left for the bookmakers. The
bookmakers manipulate the odds by increasing the probabilities of all the possible
outcomes so that the total probability goes over 100%. For example, the true
probabilities of the outcome of football games have been proven to be less than the
odds with a few percent [21]. This will cause the total wins paid out to the bettors
to be lower than the total stakes. The leftover is the profit of the bookmakers. This
also means that there are more people losing their money in sports betting than there
are people winning money.

Bookmakers takes help from data- scientists and analysts to make predictions of
sport results [21], but they don´t predict the results only by themselves. They use

24
the predictions of the public to balance the odds so that they will make more profit.
Still, it is quite surprising how well the odds match the result of games, which means
that the public can predict sport results quite accurately without the help of sport
experts, bookmarkers or machine learning models.

However, except that many people are addicted to and lose their money on sports
betting, match fixing and money laundering are also problems caused by sports
betting. When there are much money on the line, some people want to take risks
and illegal actions to be able to win. Match fixing can be done by bribing the referee
or competitors of the game, and means that they try to manipulate the game to have
the outcome of the match-fixers bets. According to a report made by Sportradar
[19], the occurrence of match fixing is increasing in sports worldwide, due to the
fast growing of the market. The most cases of suspicious games have occurred in
football, e-sports and basketball. In football most cases occur in the lower-level
competitions, where there are not as many organisations that governs or monitors
the betting. The players are more vulnerable targets for match-fixers in lower league
tiers since their athlete education and income is not at the same level as in higher-
level leagues.

According to the annual match fixing report by Sportradar [19], the most popular
sport in betting is by far football, with a market of over 700 billion Euro and over
50% of the sports betting made in 2021. Other popular sports in betting in 2021
were tennis and basketball. In these three sports, the greatest number of suspicious
matches also occurred. The most popular competitions in sports betting in 2021
were UEFA Champions league (football), English Premier League (football), and
Indian Premier League (cricket). According to Sportradar, match fixing has become
more widespread after the COVID-19 pandemic, spreading to new sports and
lower-level leagues.

There are significant differences between sports and competitions in how well their
results can be predicted. One of the primary factors is how much the strengths of
the competitors differ from each other. If the difference in strength is high, the result
will be more predictable. This can be the case in so-called closed leagues, where
the teams do not change depending on their performance after a season. This means
that bad teams still must play against teams that are considerably stronger than

25
them, instead of being relegated to a lower-level league with teams with more
matching strength. Leagues with one or a few teams that are relatively superior to
the other teams, will also be more predictable. Another factor that can improve the
predictability is if seeding is used in knockout competitions. Seeding is when
stronger competitors are drawn to compete against weaker ones, so that the strong
competitors have a fair chance to reach the final (if the games would be drawn
randomly, two strong competitors might have to face each other already in the first
rounds of the knockout stage). Sports and competitions that can end in a draw, is
usually more difficult to predict than sports with only ‘win’ or ‘loss’ as possible
outcomes [22].

One of the most effortless sports to predict are sports with high difference in the
strength of the competitors, like Formula 1. The team with the best cars and the best
driver of that team will win most of the races. Some of the most difficult sports to
predict are football and ice hockey. Since they are relatively low scoring sports, it
is difficult to predict the outcome although one of the teams may have a strength
advantage. Low scoring sports are more difficult to predict because the probability
that a weaker team can beat a stronger team is higher since one single goal is more
crucial and can be scored with luck. Sports like basketball and tennis are more
predictable since they are high scoring sports, in which luck and unpredictable
events have less importance on the outcome.

Bunker and Thabtah [23] have proposed a framework for sport result prediction. It
includes six main steps: domain understanding, data understanding, data
preparation, modelling, evaluation and deployment. It is like the procedure of most
machine learning problems. Domain understanding means to know the problem and
the characteristics of the sport in question. The problem could be to predict the
overall result, the exact numerical results of the matches or to predict which matches
it should pay off to bet on. The characteristics of the sport are of course important
to know to be able to find the factors that impacts the outcome of matches. The data
used in the model must be understood as well. The data can be match statistics,
team related, or player related. The amount of publicly available data online about
sports have been increasing, which makes it more straightforward for researchers
and hobbyists to collect the data needed for their sport predictions. If there are no

26
suitable data sets publicly available, it is of course possible to extract online data to
make an own database and data sets. [23]

In team-sports the most important feature types for predicting the outcome, are
previous match statistics, team related features and individual player related
features, of both the competing teams [24]. In individual sports the most important
feature types for predicting the outcome are player related features and previous
match statistics [25]. However, the prediction model will not perform well if the
features are not processed before use. Most researchers in sports predictions do
feature selection and feature extraction before using the machine learning algorithm
[26]. As for all machine learning problems, one of the most important aspects to
achieve a good result in sport predictions is the feature selection. The original data
set can be divided into feature subsets [23]. The different subsets can include
features like, for example, match related features, team related features, standings
features and betting odds features. The subsets can be tested individually combined
with a machine learning algorithm to select the best features. Feature selection
algorithms can also be used to find the most important features.

Data preparation is important to achieve valid results out of the machine learning
model. Feature extraction must be made on all the used features. Beyond that, match
related features must be handled differently to other external features [23]. Since
the match statistics are known only during or after the match has been played, they
must be treated differently. Usually, an average of each match related feature is
calculated to be used in the model. The best number of matches to be used for the
averaging process has been found to be the past 20 matches [27]. The target variable
must then be defined. For classification problems in sport predictions the target is
the class variable, which usually consists of three values (home win, draw, or away
win), but can also consists of only two values (home win or away win) [23].

The choice of machine learning algorithms according to the selected features is also
of great importance. The most common machine learning algorithms in sports
predictions are classification algorithms. The wanted output is usually the overall
result of the game as win, draw or away win. The other option is to predict the total
score of a game. Regression algorithms can be used for such numerical prediction
problems. The most used classification algorithms in sport predictions are different

27
types of artificial neural networks [23]. Other classification algorithms suitable for
sport predictions are Bayesian networks and logistic regression. Artificial neural
networks are suitable for sport predictions since they can be flexible on how to
interpret the class variables. They can also deal with dependencies between the
input variables, while Bayesian networks on the other hand see all input variables
as independent to each other.

In the modelling phase the chosen algorithms can be used for testing on the different
feature subsets, to find the best possible combination. Sport prediction data sets can
be split into training- and test data sets in multiple ways. Bunker and Thabtah [23]
argue that the best option is to do the training test split round-by-round, especially
for sports where each team plays exactly one match per round. The order of the
matches still must be considered; matches should only be predicted based on
previous matches. They argue that the best way is to use a few rounds as test data
equally spread within all the rounds. For each test round, the training data should
consist of all the previous rounds. The average of the accuracies of all the tests
could then be used as the overall model accuracy. Another way to do the training
test split round-by-round is by using a block of the first rounds as training data, and
a block consisting of a few of the last rounds as test data. This is slightly simpler
since only one accuracy is produced and an average of the tests is not needed. If the
amount of data in the original data set is higher and includes multiple seasons, the
training test split can be done by splitting season-by-season instead of round-by-
round. Although, in most sports the strength and player of the teams changes each
season and therefore it could be misleading to do the predictions based on previous
seasons. Another option is to predict the results match-by-match. In that case the
training data is all the previous matches and the match in question is the test data.
However, if the goal is to predict the outcome of a whole season, this option
demands the training data to be updated for every match. [23]

Machine learning based sport predictions most often use prediction accuracy as the
main evaluation metric. A classification matrix can be used to visualise the results
in classification problems. Since class values of sport prediction data sets usually
are quite balanced, classification accuracy can be used to measure the model
performance [23]. In many sports there are slightly more home wins than away wins

28
due to the home advantages. There are more wins than draws, however, since there
are two types of wins the class values are still quite balanced. When predicting the
overall result of a game (home win, draw or away win), the prediction accuracy of
the model must be over 33% to be higher than chance. A model with a prediction
accuracy below 50% can still be considered as unsuccessful. A model with a
prediction accuracy of between 70% and 90% is good, although it is subjective and
depends on the type of problem in question. If the prediction accuracy is close to
100%, it is usually a sign of overfitting rather than a perfect model. It is difficult to
find an average prediction accuracy of all sports predictions since they cannot all
be compared to each other. However, most research papers have claimed a
prediction accuracy of between 60% and 80%, with around 70% being the average
[26].

Since most sport prediction models use data of previous game results, cross-
validation is not appropriate to use for evaluating sport predictions [23, 28]. If
temporal data is used in cross-validation, the classification algorithm may see
training data that has happened later in time than the test data, which would not be
possible in a real-world scenario. If sport predictions are supposed to be made on
game to game, the chronological order of the data must be preserved, and therefore
cross-validation is not a suitable option for evaluation. However, it is not a problem
when match results are not considered or when only previous match results are used
for a whole season at a time.

3.2 Football Predictions

Football is one of the most difficult sports to predict. That a team has a high strength
advantage over the opponent does not automatically means that it will win. Since
football is a low scoring sport, the probability of a weaker team winning or drawing
against a stronger one, is higher than in many other sports. In football, the results
may depend on features of individual players, the tactics and motivation of the
teams, the recent form of players and teams, and external factors like weather and
home ground [24]. Which ground the match is played on, is an important factor due
to the home advantage phenomenon. The team playing on their home ground will
have more experience of the weather conditions and the pitch, which type and

29
condition may vary between grounds. The away team must travel to the ground of
the opponent and the travel distance might impact their energy. In addition, there
are usually more home fans than away fans in the stadiums, which will help the
home team since it receives more positive support while the away team might
receive disturbances and heckling.

The percentage of home wins, away wins, and draws, are different from league to
league and changes between the seasons. Overall football matches most often end
up in a home win, followed by away win, and least often in a draw. Around 45% of
matches end up in a home win, 30% in an away win, and 25% in a draw [29, 30].
During the 2020/2021 season many leagues played without fans due to the COVID-
19 pandemic. During that season the home win percentage decreased to around
40%, and the away win percentage increased to 35%. This proves that home
advantage, especially when fans are attending, has a great impact in football. [29]

Although football is difficult to predict in comparison to other sports, studies have

shown that it is possible to achieve a high prediction accuracy. According to the
review on sport predictions made by Horvat [26], most football match predictions
achieve a prediction accuracy between 55% and 70%, but there are a few studies
that claim they have achieved prediction accuracies up to 90%. In the last years
several studies have been made on making football match predictions, some more
determined to achieve the highest possible prediction accuracy and other more
focused on finding a successful model for future research serving as a benchmark.

Owramipur et al [1] used Bayesian networks in their football match predictions in

2013. They achieved an outstanding prediction accuracy of 92%, however, they
only focused on one team for one specific season (FC Barcelona, Spanish La Liga
2008-2009). They used multiple types of features for their predictions, both
psychological and non-psychological. The values of main features that affected the
result mostly in favour of one team, were when the average age of the players was
in the middle, the result of the five last games were good, there were not any injured
main players, the psychological state of the players was good, the performance of
all the players was high, and the weather was good.

30
Shin and Gasparyan [31] used virtual data collected from a video game to predict
football matches in 2014. They collected data of player attributes from ‘FIFA
2015’, which is a football video game made by EA Sports. The matches they
predicted were one season of Spanish La Liga. They used multiple models for their
prediction, with algorithms like, for example, the support vector machine, logistic
regression and Naïve Bayes. The results using virtual data were compared with
results using real data, which were collected based on average performances in
previous matches. Some models using virtual data achieved a higher prediction
accuracy than models using real data. The best model for both data categories were
the one using linear support vector machine. Using virtual data, the model achieved
an accuracy of around 79%, while using real data the model achieved an accuracy
of around 73%. Their research proved that virtual video game data is not always
totally artificial and imaginary, but instead it could be used for predicting real world
scenarios with good results.

Tax and Joustra [28] made football match predictions for the Dutch League
Eredivisie in 2015. They used different feature sets including match statistics and
betting odds from a period of 13 seasons. They used multiple classification
algorithms, which they compared with each other using the different feature sets.
When all match features were used, Naïve Bayes and artificial neural network
resulted in the highest classification accuracy of 54.7%. When a hybrid model of
both the match- and betting odds features were used, they achieved a classification
accuracy of 56.1%.

Prasetio and Harlili [32] used logistic regression to predict football match results in
2016. However, they only used two possible outcomes: home win or away win.
They used multiple seasons of the English Premier League for training, to predict
the 2015/2016 season of the same league. The main features they used were “home
offence”, “home defence”, “away offence”, and “away defence”, out of which the
defence-related features were the most important. The prediction accuracy of their
model was 69.5%.

Razali et al [33] used Bayesian networks in their research on football match

predictions in 2017. They predicted the results for three seasons of the English

31
Premier League (EPL), using match statistics as the main factors. They achieved an
average accuracy of 75.09%.

Chen [34] made football match predictions based on player features in 2019. The
predictions were made on multiple seasons of Spanish La Liga. The origin of the
player features was from the football video game FIFA. Three algorithms were
compared: convolutional neural network, random forest and support vector
machine. The best result was achieved using the convolutional neural network with
an accuracy of over 57%. Using the other two algorithms an accuracy of around
54% was achieved.

Rahman [35] predicted football match results using deep learning in 2020. The data
set used contained rankings and team performances based on previous football
match results of international teams. The matches predicted were the matches in the
FIFA World Cup 2018 (international football world championships). The
predictions were made using artificial neural networks and deep neural networks.
The model predicted the overall result of FIFA World Cup 2018 group stage
matches correctly with an accuracy of 63.3%.

Taspinar et al [30] made football match predictions for Italy Serie A League in
2021. They predicted the match results for multiple seasons using features based on
match statistics. When using a logistic regression model on all the features, they
achieved an average classification accuracy of 88.85%. However, they concluded
that the accuracy could be improved by selecting effective features. When they only
selected the most important features, the accuracy increased to 89.63%.

As for sport predictions in general, the most used machine learning algorithms for
football predictions are Bayesian networks, logistic regression and artificial neural
networks. It is recommended to test several of them to find the best method in
combination with a particular data set [22]. All three algorithms suit well for
classification problems like predicting the overall result in football matches.
However, the prediction accuracy could be improved by treating the match result
as only two classes: home win and away win [30]. Although this would only be
realistic when predicting knockout competitions where there are no draws, since in
football leagues draws are usually allowed.

32
Since the outcome of football matches are not evenly balanced between home wins,
away wins, and draws, the machine learning models will also tend to predict the
outcomes unevenly. It is most difficult to predict draws correctly since there are
least number of matches with such an outcome. Moreover, the draw is ‘in between’
home win and away win, which means that the statistics of a match ending in a
draw are similar to other situations (home win and away win on each side) [30]. On
the other hand, home wins and away wins are the extremes and therefore simpler to
classify. The difficulty of predicting draws correctly can be seen in many research
papers [28, 30, 34]. An important direction to improve the overall prediction
accuracy of one’s model, is to improve the predictions of matches ending in a draw
[30].

As we have seen, football match prediction using machine learning has been studied
in multiple research papers during the recent years. However, many of them have
not extensively compared different methods with each other, but instead focused on
a specific method or algorithm. Many research papers also focused on a specific
type of football related data for the predictions, instead of comparing the
performance of, for example, team attributes and match statistics. To contribute to
the research on how to find the best frameworks and methods for football
predictions using machine learning, I decided to make my own implementation
where I compare multiple methods and data types.

33
4. Implementation

In the same pace as artificial intelligence and machine learning are developing, so
does also football predictions. The amount of available statistical data of matches
and players in the game are increasing all the time. Every move and touch on the
ball that the players do can be tracked on the field. However, football is a difficult
sport to predict since luck and surprises are inherent factors. As we have seen in the
previous chapter, machine learning models can still achieve high prediction
accuracy on football matches. This chapter describes the own implementation that
I have done on football predictions, to contribute to the research on how to find
optimal frameworks and methods for football predictions using machine learning.

4.1 Overview

This implementation includes the whole process of machine learning, starting with
data collection, followed by data preparation, model development, and model
evaluation. The steps in the implementation process also largely follow the
proposed framework for sport result predictions made by Bunker and Thabtah [23],
in which the main steps are domain understanding, data understanding, data
preparation, modelling, evaluation, and deployment. The goal of the
implementation is to test and compare different approaches for football predictions
using machine learning and to achieve a prediction accuracy that is higher than the
average prediction accuracies of previous research in football predictions.
According to Horvat [26], the average prediction accuracy for football predictions
ranges from 55% to 70%. A prediction accuracy of above 60% was set up as an
initial goal for this implementation.

This implementation predicts the overall result of football matches in a particular

season of a particular football league of high professional standard. The overall
results signify whether a football match ends in a win for the home team, a draw,
or a win for the away team. The exact numerical result of the matches, that is the
number of goals scored by each team, is not predicted. Since the overall result can

34
have three different outcomes (or classes), the machine learning problem of this
implementation is a multiclass classification problem.

The data used in this implementation were collected from free publicly available
sources online. The data used in the implementation consist of match statistics, team
related attributes and player related attributes. The different types of data are first
tested separately and then merged to test if this will improve the model
performances. Multiple machine learning algorithms are used in the
implementation, not only to try to achieve the highest possible prediction accuracy,
but also to compare how the algorithms perform in football prediction problems
overall. Since this implementation is a multiclass classification problem, the
machine learning algorithms chosen are algorithms that support multiclass
classification. All the different data types are tested on all the algorithms, to find
the best possible combination of data set and algorithm. Although this requires the
creation of many different models, the work is still reasonable since one can use the
same data sets in all the algorithms with just small adjustments. Moreover, the basic
structure of the machine learning algorithms can be the same regardless of the input
data set used.

The predictions are done on the whole season at once. In models where the match
statistics are used, the data must be kept as temporal data. In such cases, the test and
training data are split so that the test data follow the training data in chronological
order. Then the predictions are done match by match, although the goal of the
predictions is still to achieve high prediction accuracy for the complete season. The
implementation is not developed for the purpose of predicting single matches.

The programming languages used in the implementation are Python and SQL.
Python is considered as one of the best programming languages for machine
learning. SQL is used to query data from an SQLite database. Data preparation is
made in the code editor Visual Studio Code. The machine learning model
development is made with Python in Jupyter Notebook. Jupyter Notebook is an
optimal tool for machine learning, especially since the code can be divided into
cells so that loading data multiple times is not needed when experimenting. It is
also a suitable tool for making visualisations.

35
4.2 Data Collection

Data collection is the first practical step in machine learning. However, before the
data sets for the machine learning models can be created, one must know what data
to use. Knowing what impacts the outcome of the sport in real-life is needed to
achieve a successful sport prediction model. To know the main characteristics of
the sport that will impact the machine learning problem, is called domain
understanding [23]. In football the results depend on the performance of individual
players, the performance of the teams overall, and external factors like weather. The
data used in this implementation include player attributes, team attributes, and
match statistics. Some combination of these three categories of football data have
usually been used in previous football match prediction research. When testing all
three of them, as well as combinations of them, one will understand how they
impact the results of football matches and which combinations are the best for
predicting the results. Since the predictions were supposed to be made on a whole
season at a time, external factors like weather data were not considered in this
implementation. The weather on a particular match day obviously cannot be known
beforehand.

The quantity and quality of football data available online vary much from source to
source. Both the quantity and quality of the data sets used in machine learning
models influence the results. Match statistics of high standard leagues can
effortlessly be found online, not only live scores and results of the current season,
but also statistics from years back. The statistics can be found in data source types
like CSV files and databases. The most effortless way to collect data for machine
learning developers, is of course to directly find these kinds of ready data sets that
suits their problem. However, if there are no suitable data sets for one’s problem,
online data can be extracted. The match statistics can be extracted from websites
and API:s with live scores and results from many previous seasons. Most data
sources about football match statistics include basic statistics like goals, possession,
corners, cards, and fouls. Data sources with more in-depth match statistics like
player performances during the game can be more difficult to find, especially ones
that are free and publicly available. Some well-known websites with football
statistics are Sofascore.com, Flashscore.com and Transfermarkt.com.

36
The data sets used in this implementation are a mix of already existing data sources
and data collected from websites. The base data set used in the implementation is a
data set named ‘European Soccer Database’ published by Hugo Mathien on Kaggle
[36]. Kaggle is a community of data scientists and machine learning developers,
where data sets are published for others to find and use in, for example, machine
learning development. The ‘European Soccer Database’ consists of football data
from eleven of the greatest football leagues in Europe. The database is published
on Kaggle as an SQLite database. The football data in the database consists of
match data, team attributes, and player attributes from season 2008 to 2016.
Overall, it sums up to around 25 000 matches, 11 000 players, and 300 teams. The
match statistics include statistics like goals, shots, fouls, cards, and possession. All
the players who played in a match are listed in the match data as keys, which can
be used to join the players with their player attributes. Besides that, the match data
also includes betting odds from several betting sites and metadata like date, teams,
season, and stage of the match. [36]

The team- and player attributes in the database originate from the football video
game EA Sports FIFA. It is a video games series developed by the video game
company Electronic Arts, labelled as EA Sports. They release a new game for every
season, which have made it the best-selling sports video game franchise in the
world. Every new game is named so that it reflects on which year the season will
end, for example the season 2015-2016 is covered in the game called ‘FIFA16’. EA
Sports FIFA is based on authentic real world football leagues and players, thanks
to its partnership with the international football association FIFA. The player
attributes in the game are based on real life performances and are updated many
times during the season [37]. The player attributes and ratings are on a scale from
1-99, where 99 is the best rating. The team attributes are based on the players in the
team and overall team performance. Although the team- and player attributes in EA
Sports FIFA are regarded as one of the best ratings of footballers in the world, they
are still subjective and cannot be totally accurate. For example, players in high
standard leagues often have higher ratings automatically. Even though the team and
player attributes would be totally accurate according to the real world, player
performances will still fluctuate from match to match, and therefore they cannot
alone give perfect football match predictions. Virtual football data from video

37
games have been proven to be able to achieve accurate football match predictions
[31] and were therefore chosen to be tested and compared with other types of data
in this implementation.

As described in the overview of the implementation, one season of a particular

league had to be chosen. For the implementation the German Bundesliga was
chosen, which is the highest football league division in Germany and one of the
best leagues in the world. Any of the other leagues in the ‘European Soccer
Database’ would have worked equally well, but Bundesliga was chosen since it has
not been used as much as other top European leagues in previous football prediction
research papers, and because it is an interesting league being one of the greatest
leagues with close connections to the Nordic countries. The season 2015-2016 was
chosen for the implementation since it is the most recent season available in the
‘European Soccer Database’. The German Bundesliga has 18 teams and during one
season there are totally 34 stages and 306 matches (one home match and one away
match against every other team in the league).

The team- and player attributes in the ‘European Soccer Database’ were sourced by
Mathien from the website Sofifa.com, which is a database for statistics in the EA
Sports FIFA game [36]. The team attributes in the database include team ratings of
their build-up play, chance creation, and defence types. However, these attributes
are quite detailed and for some reason the overall team ratings were missing in the
database. That is why some additional team attributes were collected, which then
were added to the data set. These team attributes were the overall rating of the
teams, and their defence, midfield and attack ratings. These team attributes were
sourced from Fifaindex.com [38], which is a database for statistics in EA Sports
FIFA like Sofifa.com. Only the team attributes for the selected league and season
(German Bundesliga, 2015-2016) were collected and added to the data set.

The match statistics of the German Bundesliga season 2015-2016 in the ‘European
Soccer Database’ lacked clean values for some attributes. For example, the goals
were embedded in a piece of code and it would have required much cleaning to
extract the relevant values. However, the most significant limitation of the data was
that there was no attribute for directly determining the overall result of the match,
that is ‘home win’, ‘draw’ or ‘away win’. This could of course be calculated by

38
looking at the scored goals for the teams in each match, but that would have required
some extra coding. Instead, a second data source including match statistics from the
chosen league and season was used. The data source is a CSV file sourced from
DataHub.io, which is a data hub for open data (free to use for everyone) [39]. This
data file includes similar match statistics and metadata as the ‘European Soccer
Database’, but with cleaner and more complete data for all attributes and with the
important addition of the missing ‘overall result’ attribute. This attribute is named
‘FTR’, shortened from ‘full time result’. The three possible values are ‘H’ for ‘home
win’, ‘D’ for ‘draw’ and ‘A’ for ‘away win’. These will be the output classes of the
football prediction classification.

A diagram of the data used in the implementation, is visualised in Figure 6. The

sources of the team attributes are the ‘European Soccer Database’ and the website
Fifaindex.com. The ‘European Soccer Database’ is the only source of the player
attributes, while the only source of the match statistics used in the implementation
is the data set from DataHub.io. All three different data categories are joined with
match data to form the final data sets. The source of the match data is the same data
set from DataHub.io as the one used for match statistics. The match data consists
of the date, home team, away team and overall result of the match. The matches are
sorted in chronological order.

Figure 6. Diagram of the data sources used for the data sets in the implementation.

39
4.3 Data Preparation

In machine learning, the next step after data collection is data preparation. It
consists of data cleaning, feature selection and feature extraction [3]. In this step
one need to know what input data to use and what output data one wants. To know
what data to use, the data sets one work with need to be known. In this
implementation the goal was to use three different main data sets: one data set with
team attributes, one with player attributes, and one with match statistics. The data
preparation was made in Visual Studio Code using SQL and Python. To be able to
explore and query the SQLite database European Soccer Database in Visual Studio
Code, the SQLite extension was installed.

The base of every data set contains the same attributes, the match data. The date,
the teams, and the overall result of every match, were taken from the data set
sourced from DataHub.io. The overall result is the attribute called FTR, and it
serves as the output data in the football prediction models. This attribute is the most
important in the data sets, since the accuracy of the classification predictions is
determined on how correctly this attribute can be predicted. The teams in each
match consist of the name of the home team and the away team. However, the team
names and the FTR attribute are represented as text data. The text data must be
converted into mathematical representations using feature encoding. The three
possible classes of the FTR attribute were dummy encoded to the numbers 0, 1, and
2. The team names were also encoded using dummy encoding, where every class is
represented as a number instead of the text. There are 18 teams in the German
Bundesliga, and therefore 18 possible classes for the home- and away team
attribute.

The team attributes table in the main database, the European Soccer Database,
includes team ratings on build-up play, chance creation, and defence types. The
table includes a total of 25 attributes, of which three are different kinds of ids and
one is the date when the team attributes were updated. However, of the remaining
21 attributes, only nine include numerical values. The rest of the attributes are called
“classes of the attributes” and consists only of simplifications of the numerical
values. The numerical values have been grouped according to some specific ranges

40
into a few classes in text form. For example, the values of the build-up play speed
of the teams have been grouped into either slow, balanced, or fast. It is unnecessary
to use both the numerical and classificational approach. Since the numerical values
are more precise, only those nine attributes were included in the final team attributes
data set.

To make an efficient data set for machine learning, only data that is relevant to
one’s problem must be used. Since the European Soccer Database includes around
300 teams, it was clear that only the 18 relevant teams from the German Bundesliga
season 2015-2016, had to be queried from the team attributes table. The team
attributes table does not have a country-id, which meant the relevant teams had to
be queried using the team-ids instead. The team-ids for the 18 teams could be found
by querying one stage of the 2015-2016 German Bundesliga season from the match
table, since every team will only play one match in each stage and therefore there
will only be 18 team-ids visible without duplicates or missing values.

All the team attributes for the relevant teams were queried together with their team
names. The team names were needed to be able to join the team attributes with the
match data sourced from DataHub.io. The match data does not use the same team-
ids as the European Soccer Database does, and therefore the data must be joined
using the team names instead. However, the team names are also different in the
two data sources. Football teams usually have a long official name, but it is common
to use a shortened version. This can obviously lead to different versions of the team
names in different places, especially when the team names are not in English. In
this case the European Soccer Database uses a longer version of the German
football teams, than the match data from DataHub.io uses. For example, the
European Soccer Database uses the team names FC Schalke 04 and Borussia
Mönchengladbach, while the match data uses Schalke 04 and M’gladbach. To be
able to join the two data sources using the team names, the team names used in the
match data was mapped to the team-ids in the team attributes query. The final query
result was exported to a CSV file, which can effortlessly be done in Visual Studio
Code after an SQLite database have been queried using the SQLite extension.

In the FIFA video game, the main ratings of teams are the overall rating and the
overall ratings in different fields: defence, midfield, and attack. The overall ratings

41
are calculated based on more detailed team attributes, which also are found in the
European Soccer database. However, for some reason the overall team ratings were
missing in the database, and therefore the overall ratings were added to the team
attribute data set manually from a website. In situations where one need to retrieve
data from websites, a normal approach is data scraping. In data scraping data can
be imported from websites using computer programs, to be saved in files on one’s
own computer. In this case, the additional data needed included just four attributes
for 18 teams. Therefore, this data was added to the team attributes data set manually.
It would have been more work to create a data scraping program to retrieve the
specific data automatically, than simply copying the 72 values manually.

The next task was to join the team attributes with the match data. This was done
using Python in Visual Studio Code. The match data from DataHub.io was first
joined with only the team attributes from the European Soccer Database. However,
when figuring out that the overall ratings were missing from the data set, the match
data were joined with the team attributes including the overall ratings instead. The
match data was first extracted from the match statistics from DataHub.io. The
original data source is a CSV file, and was read to a Pandas data frame. Pandas is a
common Python library used for data analysis. The CSV file including the team
attributes and overall ratings, was also read to a Pandas data frame. The two data
frames were joined on the team names. The final data set was to include all matches
of the selected season, with the team attributes of the teams facing each other in
every match. Therefore, the match data was the base data frame, and joined with a
left join with the team attributes two times, ones for the home team and ones for the
away team. An ‘h’ and an underscore were added in front of the attribute names of
the home team, and an ‘a’ and an underscore were added in front of the attribute
names of the away team, to be able to distinguish between the attributes of each
team.

The player attributes table in the main database, the European Soccer Database,
includes player ratings on attributes like shooting, passing, physicality, pace,
movement, defence, and goalkeeping. It is by far the largest table in the database,
since it includes attributes of over 10 000 players for each season and version. The
player ratings change from season to season, but can also be updated during a

42
season. The player attributes table includes a total of 42 attributes, of which three
are different kinds of ids and one is the date when the player attributes were updated.
Of the remaining 38 attributes, three include text data: a player’s preferred foot,
attacking rate and defence rate. These attributes were also removed from the data
set. The 35 remaining player attributes all include numerical ratings between 1 and
99, which makes feature scaling unnecessary in the modelling phase.

To retrieve the player attributes for every team in the 2015-2016 season of the
German Bundesliga, the match table must be used like for the team attributes.
However, it does not make sense to add all the players of a team to every match,
since there are only eleven players in the start of a match and a few players may
come in later in the match as substitute players. The players sitting at the bench will
not impact the result of the match, and therefore their attributes are unnecessary to
use. Luckily, the match table includes the starting eleven of every team for every
match. The substitute players are not included in the table and even though they
would be, the use of substitute players would be very complex, since then one would
have to add weights to the data according to how long they have been on the pitch
to make the predictions realistic. In the match table the players are represented as
player-ids, which can be used to join it with the player attributes table. The ‘match’
table also includes coordinates of the players on the field, which makes it possible
to determine what formation a team uses and where each player is positioned.
However, the players in the match table are ordered with the goalkeeper first, then
the defenders, the midfielders and lastly the attackers, so the position of players can
roughly be determined without using the coordinates.

The joining of the player attributes and the match data was more difficult than
joining the team attributes with the match data. For every match, all the player
attributes of the 22 players involved in the match are to be used. This means that
the player attribute table had to be joined with the match data 22 times. To be able
to do the joins more effortlessly, the player attributes were first joined with the
match data in the European Soccer Database, and after that joined with the FTR
column of the match data from DataHub.io. The joins could therefore be done using
SQL only within the European Soccer Database. The select of the query consisted
of the date, the home team-id and the away team-id from the match table, and all

43
the player attributes of the 22 players from the player attributes table. The match
table was joined with the player attributes table using left-joins on the player-ids.
The match data had to be filtered on the 2015-2016 season of the German
Bundesliga and the player attributes had to be filtered on one specific version. The
version of the player attributes was chosen as the first official release of the FIFA16
player ratings. However, some players move clubs within a season and some player
ratings may have been incomplete during the first release, so the query resulted in
246 matches instead of the original 306 matches. In addition, 27 matches included
player attributes with null values, which had to be removed, and therefore reduced
the final match total to only 219 matches. The final query result was exported to a
CSV file.

Since the teams in each match were represented as ids, the team names in the match
data from DataHub.io had to be converted to the team-ids used in the European
Soccer Database. This was done using Python in Visual Studio Code. A conversion
table was created, that included all the 18 teams with their team names and their
respective team-ids in the European Soccer Database. The table was left joined with
the match data two times, ones for the home team and ones for the away team.
When the team names were integrated in the data set, it could be joined with the
‘FTR’ column of the match data from DataHub.io. The join was done using a left
join on the team names of each match.

The match statistics data used in this implementation was sourced from DataHub.io.
The data source is a CSV file with match statistics of the German Bundesliga season
2015-2016. It includes information about the matches, such as date, participating
teams, statistics, and betting odds. The statistics of the two teams are separated in
the names of the attributes, so that the statistics of the home team are attributes
including an ‘H’, and statistics of the away team are attributes including an ‘A’.
The match statistic attributes are the number of goals, shots, shots on target, fouls,
corners, yellow cards, and red cards, that have occurred during the match.
Furthermore, the overall result of the match (the FTR attribute that is also used in
the match data combined with the team attributes and player attributes) and the
goals scored are divided into both the fulltime- and halftime results. The betting
odds have been gathered from over ten different betting sites, and consist of the

44
odds for home win, draw or away win. However, these statistics are not used in the
implementation, since the most accurate betting odds are finalised exactly before
the start of a match and therefore cannot be gathered for all matches of a whole
season before the season has even started.

Although the match statistics data set from DataHub.io includes clean data and most
of the attributes needed to make predictions, there was a major obstacle. The match
statistics cannot be used directly to predict the results of the football matches, since
the statistics obviously have been gathered after the matches have ended. It does
not make sense to make predictions on matches that have already occurred, instead
previous matches must be used for the predictions. The statistics of every match
had to be converted into averages of previous matches. This makes the predictions
more logical, but in a real-world problem the result of upcoming matches cannot be
predicted using match statistics, except for the next upcoming match of a team.
However, in this implementation the whole season was predicted at a time, but the
chronological order of the matches was kept so that the results of matches later in
time are not “leaked” in the learning process of the machine learning models.

The data preparation of the match statistics was done with Python in Visual Studio
Code using the Pandas library. The first step was to read the data from the original
data source file into a Pandas data frame. Only relevant attributes were read into the
data frame, the betting odds were not considered. The type of every numerical
attribute was changed to float instead of integer, so that they would be simpler to
handle when calculating averages. The columns representing goals scored for each
team in a match were copied, so that there would be one column representing goals
scored and one column representing goals conceded for each team in a match.

In many football league tables online, the results of the last few matches for each
team are visible. This gives a hint of the current form of a team. A team that has
won all their recent matches is in form, but a team that has not won any of their
recent matches is in poor form. The FTR attribute was used in the implementation
to utilize the form of the teams. The FTR column was copied so that the form of
each team can be seen in every match. The values were converted to numerical
values to be able to calculate the averages. For each team, a win was set to two, a

45
draw to one, and a loss to zero. Thus, teams with a higher average exhibit superior
form.

Since the collected data only includes statistics from one season, the first matches
of the season will not have any data to make the averages on. Therefore, the
statistics of the last match of the 2014-2015 season of every team were added to the
data frame manually. The next step was to calculate the averages of the match
statistics. For each match, the statistics of the specific match had to be changed into
the averages of the past matches. For football match predictions, it has been found
that the best number of matches to be used for the averaging process, is the past 20
matches [27]. This was also chosen as the maximum number of past matches
considered in the averaging process of this implementation. In the 19 first stages of
the season, only the number of available past matches were considered.

The averages were calculated using a temporary table to keep track of the statistics
of past matches. A for-loop was used to iterate over the matches. In the beginning
of every iteration, the match statistics were saved to temporary variables together
with the stage number. The stage number was the same as the iteration order, which
was kept track of using a variable. The statistics were to be saved in the temporary
table, however, first the averages had to be calculated. Otherwise, the statistics of
the current match would also be considered in the averages and thus leaking
statistics that only should be seen after the match has ended.

Next the average of every attribute was calculated using the past match statistics in
the temporary table. The temporary table was filtered on the team names and a sum
of every attribute was calculated. The sums were then divided by the number of
past matches used, to receive the final averages. The averages were rounded to three
decimal places and then added to match statistics replacing the match statistics for
the current match. Next the match statistics of the current match, which were saved
to temporary variables, were added to the temporary table. The statistics of the two
teams in every match were separated in the temporary table so that it would be
simpler to find past statistics for a specific team. Lastly the stage number was
increased by one, and if it had reached 20 the match statistics of the teams in the
current match that had the lowest stage number in the temporary table, were
removed from the table so that the maximum number of past matches would not

46
exceed 20. When all the matches had been iterated, the final data frame was saved
to a CSV file.

The final data sets are shown in Figure 7. The player attributes data set consists of
774 columns, the team attributes data set consists of 30 columns, and the match
statistics data set consists of 22 columns.

Figure 7. The attributes of the three final data sets.

47
5. Prediction Models

When the data collection and data preparation are done, data sets that can be used
in the machine learning models are possessed. This chapter describes the prediction
models created in the implementation and the evaluation of their results.

5.1 Model Selections

The machine learning algorithms used in this implementation were chosen based
on their type, popularity, and use in previous research in football predictions. Since
multiclass classification is needed, the chosen algorithms must be able to support
it. The Naïve Bayes algorithm was chosen because it is one of the most known
machine learning algorithms and it has been used in previous research in football
predictions. Logistic regression and random forest were chosen since their approach
is very different, and therefore more possible types of algorithms for making
football predictions could be tested.

Every machine learning algorithm was implemented using Scikit-learn, which is a

Python library specialised in machine learning operations. The target in the models
is the FTR attribute, which consists of match results as categorical values.
Therefore, the CategoricalNB method was used in the Naïve Bayes implementation.
The random forest algorithm was implemented using the RandomForestClassifier
method, with 100 decision trees as the ‘n_estimators’ parameter. The logistic
regression algorithm was implemented using the LogisticRegression method. So
that the algorithm could handle multiclass classification, the ‘multi_class’
parameter was set to ‘ovr’ (One-vs-rest). To balance the number of predictions per
class, the classes were given weights in some logistic regression models. The
weights were based on the average distribution of match results in football, which
is around 45% for home wins, 30% for away wins, and 25% for draws [29]. The
frequency distribution of the results in the 2015-2016 German Bundesliga season,
is visualised in Figure 8. They are similar to the average distribution. However, the
actual distribution was not used, since it should not be known before the matches
have been played. The weights were switched around so that the weight for the

48
draws was the greatest. It was calculated with an equation where the proportions
between the values was considered. The final weights were 42% for draws, 35%
for away wins, and 23% for home wins.

Figure 8. The frequency distribution of the match results (the FTR attribute).

5.2 Team Attributes

The team attributes data sets were used with each of the three selected machine
learning algorithms. By combining the algorithms with different subsets of the
original data sets, multiple models were created. A subset in this context is a part
of the original data set, for example, to test the performance using just a few of the
original features. New subsets were tested with all the three algorithms to receive
an overview of their performance, and not only the result of one single model.

The data set was split randomly into training data and test data, so that 70% was
training data and 30% was test data. Before the overall ratings of the teams were
added to the team attributes data set, the data set was tested in the models. The
original team attributes data set included nine team attributes and the team name of
both the home team and the away team. These 20 attributes were used in the first
models. The prediction accuracy of the Naïve Bayes model was 51.1%. The logistic
regression model achieved a prediction accuracy of 53.3% and the random forest
model a prediction accuracy of only 43.5%. The achieved accuracies were decent
compared to previous research, bearing in mind that these were the first tests. To
validate the models more accurately, the k-fold cross-validation technique was
used. Using 10-fold cross-validation, the results were similar. Logistic regression
achieved an average accuracy of 52%, Naïve Bayes an average of 51%, and random
forest an average of 45%.
49
The most interesting aspect of the first tests, was to receive an estimate of the feature
importance. The feature importance can be used for feature selection in the model
testing phase. The feature importance using the random forest algorithm with the
original team attributes data set, is visualised in Figure 9.

Figure 9. The feature importance of the original team attributes data set using the
random forest algorithm.

According to the feature importance in Figure 9, the dribbling and passing in the
build-up play, the defence pressure, and the chance creation crossing, were the most
important attributes in this model. Another interesting finding is that the away team
seems to be more important than the home team in predicting the outcome of a
match. The next tests were done using a subset of the original team attributes,
including the four most important attributes from the first test and the team name
of each team. The models using these ten attributes performed better than the first
models. All the three new models improved by a few percentages, the best one still
being the logistic regression model, now with an accuracy of 59.4%.

When the team names were removed from the subset, the models performed more
balanced. The prediction accuracy of the logistic regression model decreased to
57.6%, but the accuracy of the random forest model increased to 54.3% and the
Naïve Bayes model achieved an accuracy of 52.2%. The confusion matrices of the
three models using the subset with eight attributes, are visualised in Figure 10. We

50
can see that the best precision occurred for the home win class, but the highest recall
occurred for the away win class. Overall, the f1-score was highest for home wins
and worst for draws. The most common faulty predictions by the models, were that
a match would end up in an away win instead of a draw or home win.

Figure 10. Prediction accuracy and confusion matrix of each model. From the left:
Naïve Bayes, logistic regression and random forest.

The second version of the team attributes data set also included the overall ratings
of the teams. When the four new attributes were added to the data set, the number
of attributes per team was 14, which brings the total number of attributes in the data
set to 28. The feature importance of this data set, according to the random forest
model, is visualised in Figure 11.

Figure 11. The feature importance of the team attributes data set including overall
ratings using the random forest algorithm.

51
Using the whole data set, the prediction accuracy of the models was 52.2% for
Naïve Bayes, 54.3% for logistic regression and 51.1% for random forest. According
to the 10-fold cross-validation scores, logistic regression achieved an average
accuracy of 53%, Naïve Bayes an average of 48%, and random forest an average of
46%. The same four team attributes as in the original data set, can be found among
the top attributes for feature importance of this data set. We can also see that the
midfield rating of the teams is the most important overall field rating.

Since the same four team attributes were again among the top attributes for feature
importance using the additional overall ratings, they were again chosen to be used
in subsets for more tests. The overall rating of the teams was removed and the
midfield rating was added to the subset. According to the feature importance the
defence rating and the attack rating have quite similar importance, however, in the
most tests the models performed better using the defence rating instead of the attack
rating. The logistic regression model achieved a prediction accuracy of 63.0% using
the four original attributes, the midfield rating and the defence rating. The result of
the model is visualised in Figure 12. However, we can see that barely any draws
were predicted. When running the same algorithm and data with weights, the
accuracy decreases a few percentages, but the recall of the draws was improved.
For the Naïve Bayes and random forest models, the prediction accuracy did not
change much regardless of combination of overall ratings together with the four
original attributes, but according to the 10-fold cross-validation scores all three
models improved by a few percentages using the six mentioned attributes.

Figure 12. The best single prediction result using team attributes data was achieved
using logistic regression with a data set of six attributes.

52
Although the data set that achieved the best single result among the team attributes
data sets, included some overall ratings, the addition of the overall ratings did not
improve the predictions significantly. In most of the tests, the logistic regression
algorithm performed the best according to the prediction accuracy. However, it
predicted only a few draws, and the f1-score of the draw class was very low. The
Naïve Bayes algorithm performed slightly better than the random forest algorithm
using team attributes data.

5.3 Player Attributes

The player attributes data sets were used with each of the three selected machine
learning algorithms. By combining the algorithms with different subsets of the
original data sets, multiple models were created. The data set was split randomly
into training data and test data, so that 70% was training data and 30% was test data.
The original player attributes data set included 35 attributes per player in every
match and the team name of both the home team and the away team, which brings
the total number of attributes to 772. These attributes were used in the first models.
The prediction accuracy of the Naïve Bayes model was only 42.4%, however, it
was done using the method GaussianNB instead of CategoricalNB. The
CategoricalNB method did not function properly due to different set of values in
the features. The logistic regression model achieved a prediction accuracy of 50.0%
and the random forest model a prediction accuracy of 56.1%. The logistic
regression model was even able to predict draws without weights, which the
original data sets using team attributes and match statistics were not able to do.
Using 10-fold cross-validation, Naïve Bayes achieved an average accuracy of 50%,
logistic regression an average of 42%, and random forest an average of 53%.

Since the number of attributes in the first models was so high, the feature
importance of an attribute was calculated by taking the mean of all players. This
resulted in 35 attribute groups plus the team name attributes. The estimates of their
feature importance are visualised in Figure 13. The away team attribute is clearly
standing out of the rest attributes. All the goalkeeper related attributes had low
feature importance, which was to be expected because the goalkeeper ratings of all
players in a match were included in the data set. However, the attributes of the

53
players are ordered in the data set according to their position on the field during
matches. By removing every goalkeeper related attribute, except those for the first
and the twelfth player, only the attributes of the goalkeepers were left. Every normal
player attribute of the goalkeepers was also removed. These changes in the data set
improved the results of the Naïve Bayes and logistic regression models with a few
percentages, but the random forest model performed slightly worse than in the first
test.

Figure 13. The feature importance of the original team attributes data set using the
random forest algorithm.

For the next models, some attributes with a low feature importance were removed
to test if it would improve the model predictions. When only a few attributes were
removed, the predictions worsened. However, when half of the attributes were
removed, the Naïve Bayes model achieved a prediction accuracy of 53.0% and the
random forest model a prediction accuracy of 56.1%. This time the logistic
regression model performed worst, achieving a prediction accuracy of only 39.4%.
The confusion matrices of the three models are visualised in Figure 14. Using 10-
fold cross-validation, the results were similar. Random forest achieved an average
accuracy of 54%, Naïve Bayes an average of 51%, and logistic regression an
average of 43%.

54
Figure 14. Prediction accuracy and confusion matrix of each model. From the left:
Naïve Bayes, logistic regression and random forest.

The feature importance of the attributes used in this data set was balanced with
small differences, which makes it difficult to make further feature selections. The
feature importance is visualised in Figure 15. The 19 attributes used per player were
on average more important than the 18 removed attributes.

Figure 15. The feature importance of the data set including half of the original
player attributes, using the random forest algorithm.

Since the number of attributes in the original data set was so high, the number of
possible combinations of them are also very high. However, all further tests with
player attributes data sets resulted in worse prediction accuracies than in the
mentioned tests. In some models only one attribute of each player was tested. The
overall rating and the potential of a player are examples of attributes that describe

55
the overall quality of a player more accurately. However, only a few of these
attributes achieved a prediction accuracy of over 50% on their own. Since the
playing position of the player attributes are known, we can look at the feature
importance of the overall rating attribute for each player to compare their
importance. It seems like the order from most to least important are attackers,
midfielders, goalkeepers, and defenders. A higher player position also has a higher
number, for example, the home attackers of the home team are number 8, 9, and 10
and the attackers of the away team are number 19, 20, and 21. Their feature
importance is visualised in Figure 16.

Figure 16. The feature importance of the players’ overall rating.

5.4 Match Statistics

The match statistics data sets were used with each of the three selected machine
learning algorithms. By combining the algorithms with different subsets of the
original data sets, multiple models were created. The data set was split into training
data and test data chronologically, so that the first 70% of all the matches was
training data and the last 30% was test data. The original match statistics data set
included nine team attributes and the team name of both the home team and the
away team. These 20 attributes were used in the first models. The prediction
accuracy of the Naïve Bayes model was 52.2%. The logistic regression model

56
achieved a prediction accuracy of 51.1% and the random forest model a prediction
accuracy of 55.4%.

The feature importance of the data set could be analysed after the first tests. The
feature selection in future models is based on the feature importance. The feature
importance using the random forest algorithm with the original match statistics data
set, is visualised in Figure 17. We can see that the red card attribute clearly has the
least importance. Red cards usually have a significant impact on match results, but
red cards from previous matches have a low impact on the results, except when a
key player receives suspension from the next matches due to a red card. The form
attribute also seems to be of low importance, although it usually indicates which
team will win. Surprisingly, the number of fouls made by the home team has
especially high feature importance.

Figure 17. The feature importance of the original match statistics data set using
the random forest algorithm.

The models performed more balanced when the red card columns were removed,
but they did not achieve a higher accuracy than using the whole data set. When
other attributes were removed, the models always performed worse. This suggests
that the whole original data set was the best data set to use, and that if more match
statistics would have been added the models, they may have performed even better.
The random forest algorithm achieved the highest prediction accuracy in almost all
57
the tests. The worst algorithm for the match statistics was logistic regression. Again,
it was mostly due to its inability to predict draws without the use of weights. When
weights were added to the models, the logistic regression achieved a lower
prediction accuracy but a higher f1-score for draws. The results of the three models
using the whole match statistics data set, are visualised as confusion matrices in
Figure 18. Since the values of the attributes are in different scales, feature scaling
was tested on the logistic regression models. However, it did not improve the
results.

Figure 18. Prediction accuracy and confusion matrix of each model. From the left:
Naïve Bayes, logistic regression and random forest.

The training accuracies of the Naïve Bayes and logistic regression models were
slightly higher than their test accuracies, however, the training accuracies of the
random forest models exceeded 90%. This is a sign of overfitting, which in this
case may be due to a too small data set. To validate the models more accurately, the
k-fold cross-validation technique could have been used. However, since the match
statistics data consists of previous match results, cross-validation is not appropriate
to use [23, 28]. If temporal data is used in cross-validation, the models would see
training data that has happened later in time than the test data, which would not be
possible in a real-world scenario. The model that achieved the highest prediction
accuracy using match statistics, was the random forest model that used the whole
original data set, achieving a prediction accuracy of 55.4%.

Link to the GitHub repository which includes the notebook used for making the
predictions: https://github.com/fredriksjoberg17/MasterThesis

58
6. Discussion

Many different combinations of data sets and algorithms were tested in the
implementation. Most of the models achieved a prediction accuracy of around 50%.
A few of the best models reached over 60%, but the worst models barely reached
40%. The best model used six team attributes with the logistic regression algorithm,
and it achieved a prediction accuracy of 63.0%. It is in the range between 55% and
70%, which is the average prediction accuracy achieved by football predictions
[26]. The initial goal of this implementation was to achieve a prediction accuracy
of above 60%, which was reached.

The logistic regression algorithm performed the best using the team attributes data
sets. The Naïve Bayes algorithm performed slightly better than the random forest
algorithm with the same data sets. However, using the player attributes data sets,
the results were completely switched. Now random forest performed the best,
followed by Naïve Bayes, while logistic regression performed very badly. The order
was the same using match statistics data sets, but the differences between the
algorithms were smaller. One of the reasons why the performance of a particular
algorithm changes when the data set is changed, is the number of features in the
data set. The random forest algorithm generally performs well when the number of
features is high, but may struggle when the number of features is low. This can also
be seen in this implementation, when the random forest algorithm performed best
on the player attributes data set, which by far had the most features out of the three
original data sets.

The Naïve Bayes algorithm performed most balanced out of the tested algorithms.
It did not perform best with any of the original data sets, but some Naïve Bayes
models performed the best by a little margin using smaller subsets. It is common
for the Naïve Bayes algorithm to perform well generally, but it can still be less
accurate compared to other classification algorithms [40]. Naïve Bayes assumes
that all the attributes are independent of each other, which was not the case for many
attributes used in this implementation [12]. The random forest algorithm performed
best using both player attributes and match statistics, but worst using team
attributes. The logistic regression algorithm was responsible for the best single

59
model performance, and was the best algorithm for team attributes overall.
However, it was the worst algorithm for both player attributes and match statistics.
The average prediction accuracy of all models was around 52% for every tested
algorithm. Therefore, it is difficult to determine which algorithm is the best for
predicting football matches according to this implementation. Logistic regression
using team attributes was the best combination of algorithm and data set, random
forest was the best algorithm using match statistics and player attributes, but Naïve
Bayes was the most balanced algorithm.

The performance of the three types of data sets is more distinguishable. The team
attributes data set performed best on average and produced the best single model.
On average the team attributes models achieved a prediction accuracy of around
54%, and its best model achieved an accuracy of 63.0%, The match statistics models
achieved an average accuracy of around 53% and its best model achieved an
accuracy of 55.4%. The player attributes models achieved an average accuracy of
only around 51% and its best model achieved an accuracy of 56.1%. The team
attributes data set clearly was the best for football predictions according to this
implementation. Although the number of features in the data set was low compared
to the player attributes data set, it performed well also after feature selections and
removals. Some tests with mixed data types were also done in the implementation.
The most important team attributes and player attributes were joined and tested on
the algorithms. The highest accuracy was achieved by a logistic regression model
with a prediction accuracy of around 52%, but the average was around 50%.

One reason why the match statistics data set did not perform as well as the team
attributes data set, may be that the values of some match statistics did not differ
from each other enough, and therefore did not make any significant impact in the
generalisation of the data. Since the values were averages, many attributes with low
values, like scored goals, did not differ from each other as much as for other
attributes, like fouls. It can also be one reason why fouls had the greatest feature
importance according to Figure 17. However, feature scaling was tested in some
logistic regression models, but it did not improve the results. The primary reason
why the player attributes data set performed worst, was the size of the data set. The
team attributes data set and the match statistics data set both included 306 rows, one

60
for every match played in a German Bundesliga season. However, the player
attributes data set only included 219 rows due to incomplete data or null values.
Since the number of features in the original data set included over 700 features, it
means that the number of features was larger than the number of observations. This
can lead to overfitting, especially for logistic regression. This is one reason why
logistic regression performed so badly with player attributes in this implementation.

Although some models achieved a high prediction accuracy of over 60%, all the
models shared a common issue. They were unable to accurately predict draws. It
can be seen in the confusion matrices in Chapter 5, for example on the result of the
best model in Figure 12. We can see that the recall is very low, and the precision is
not over 50%. The confusion matrix of a well performing model should show a
clear diagonal line of true predictions, which cannot be seen in many models in this
implementation. It can be seen in Figure 19, which is a confusion matrix of the
training results of a Naïve Bayes model. Weights were used to increase the number
of predicted draws for models with a low f1-score of the draws. This improved the
f1-score of the draws, but the overall prediction accuracy usually decreased. The
difficulty of predicting draws correctly have been noticed in many research papers
[28, 30, 34]. Draws are difficult to predict because least number of matches end up
in a draw, and because the class in the middle is the most difficult to predict [30].

Figure 19. Training results of a Naïve Bayes model.

61
7. Conclusion

Football match prediction using machine learning can be performed using various
techniques and methods. According to previous research [26], most football match
predictions achieve an accuracy of between 55% and 70%, and so did the
predictions in this implementation. The highest prediction accuracy achieved in this
implementation was 63.0%, using six team related attributes with the logistic
regression algorithm. To achieve a higher accuracy, the predictions could have been
made using multiple seasons of data so that the models would have had more data
to train on. The primary factor why a greater accuracy was not achieved in this
implementation, was that the models were not able to predict enough draws
correctly in relation to home wins and away wins.

Many previous studies about football match predictions have used Bayesian
networks, logistic regression, and artificial neural networks. In this implementation,
the Naïve Bayes algorithm did not perform best with any type of data, but it was
the most balanced algorithm. Logistic regression performed best using team
attributes, and random forest performed best using player attributes and match
statistics. In future research, artificial neural network could be compared with these
algorithms and tested on the three different data sets used in this implementation.
The framework for sport result predictions made by Bunker and Thabtah [23]
worked well for football predictions in this implementation.

This thesis supports the argument that virtual video game data could be used for
predicting real football matches with great accuracy. Although it is impossible to
predict football match results perfectly, it is possible to achieve accuracies that are
much greater than for randomly made predictions. By comparing multiple different
combinations of machine learning algorithms and data sets, surprisingly accurate
football match predictions can be achieved.

62
Svensk sammanfattning

Förutsägelse av fotbollsmatcher med hjälp av

maskininlärning
Fotboll är världens populäraste sport. Den har blivit så populär tack vare sin
enkelhet och förmånlighet. Fotboll spelas världen över av personer tillhörande alla
kön och i alla åldrar. Därför är det ingen överraskning att det rör sig mycket pengar
inom professionell fotboll. Det intressantaste inom fotboll är förstås själva
resultaten av matcherna och vad som påverkar dem. För att kunna påverka
resultaten måste man kunna förutsäga dem med hjälp av data. Maskininlärning är
en utmärkt metod för att göra förutsägelser. I den här avhandlingen analyseras och
jämförs olika metoder och strategier för att hitta det bästa tillvägagångssättet. Ju
bättre noggrannhet som en metod uppnår i sina förutsägelser, desto bättre anses den
vara. Fotbollsmatcher kan sluta på tre olika vis: vinst för hemmalaget, oavgjort eller
vinst för bortalaget. För att hantera sådana förutsägelser kan man använda flerklass
klassificering.

Förutsägelse av fotbollsmatcher är inte en enkel uppgift. Det är omöjligt att

förutsäga dem exakt, även med hjälp av maskininlärning. Om det skulle vara
möjligt, skulle spänningen i sporten förloras och vadslagningsbyråerna skulle gå i
konkurs. En orsak att fotbollen har fått en så stor popularitet är att den är så
oförutsägbar. I varje match är det möjligt för ett svagare lag att vinna över ett
starkare, inte bara med hjälp av taktiska drag, utan också med hjälp av tur. Faktorer
så som vädret och hemmafavör kan leda till oväntade resultat.

Under de senaste åren har mängden tillgängliga data om fotboll och andra sporter
ökat enormt. Det här har möjliggjort för forskare och privatpersoner att själva
utveckla och förbättra metoder för att förutsäga fotbollsmatcher. För vissa personer
är målet med förutsägelserna att uppnå en så hög vinst som möjligt i vadslagning.
Däremot är då också syftet att hitta svaga punkter i vadslagningsbyråernas
förutsägelser, och inte bara att förutsäga de rätta resultaten på matcherna. I denna
avhandling är målet endast att analysera hur bra man kan förutsäga fotbollsmatcher
med hjälp av maskininlärning.

63
Översikt av maskininlärning

Maskininlärning är ett delområde inom artificiell intelligens och datavetenskap. Det

är en teknik som lär sig av data med hjälp av algoritmer och försöker uppnå bättre
och bättre resultat. En maskininlärningsmodell är resultatet av tränandet med en
specifik maskininlärningsalgoritm och specifika indata. Maskininlärning går ut på
att en maskin tränas att göra förutsägelser från observationer och tidigare
erfarenhet. Det enda som vi behöver berätta för maskinen är vad vi vill att ska
förutsägas och med hjälp av vilka data. Skillnaden mellan manuell programmering
och maskininlärning är att omvandlingen av indata till utdata är okänd i
maskininlärning. I manuell programmering är både indata och utdata känd, och vi
vet hur algoritmen ska skapas för att omvandlingen ska fungera korrekt.
Omvandlingsprocessen i maskininlärning är däremot okänd. Utseendet på utdata
kan vara känt, men inte vilken indata som resulterar i vilken utdata.

Det första steget i maskininlärning är att definiera problemet. Valet av indata och
maskininlärningsalgoritm beror på hurudant problem som är definierat. Det första
praktiska steget i maskininlärning är datainsamling. Både kvaliteten och kvantiteten
på data påverkar maskininlärningsmodellernas resultat. Följande steg är
databehandling, vilken består av datarensning och val av attribut. Data som
innehåller felaktiga värden eller saknar värden måste tas bort eller korrigeras. Data
i textformat måste dessutom konverteras till numeriska värden, eftersom
maskininlärning använder sig av matematiska funktioner. Attributen kan väljas
enligt hur relevanta de är eller hur bra resultat de ger i testmodeller.

Följande steg efter databehandlingen är valet av maskininlärningsalgoritm.

Maskininlärning kan delas in i tre huvudtyper: övervakat lärande, oövervakat
lärande och semi-övervakat lärande. Övervakat lärande använder indata som är
klassificerade, medan oövervakat lärande använder indata som inte är klassificerade
från början, utan det är helt och hållet maskinens uppgift att hitta någon slags
struktur. De vanligaste typerna av maskininlärningsalgoritmer är regression och
klassificering, som båda hör till övervakat lärande. Regression är en metod som
visar relationen mellan variabler genom att skapa en funktion som passar data.
Regression är speciellt bra på att förutsäga numeriska värden. Klassificering
förutsäger kategoriska värden i stället för numeriska värden. Alla möjliga värden

64
på utdata, som indata ska klassificeras i, kallas för klasser. Om det finns fler än två
klasser, kallas det för flerklass klassificering.

Efter att en maskininlärningsalgoritm har valts, kan modellen börja tränas med
indata. Indata måste delas upp i träningsdata och testdata så att delen träningsdata
är större. Testdata används för att testa hur bra maskininlärningsmodellen har lärt
sig av träningsdata. Det sista steget i maskininlärning är utvärdering av resultaten.
Den vanligaste måttet i utvärderingen är noggrannhet, alltså hur många procent av
observationerna som förutsades rätt.

Projekt

Det finns stora skillnader mellan olika sporter och tävlingar, med tanke på hur bra
man kan förutsäga dem. En av de viktigaste faktorerna är hur stor nivåskillnaden är
mellan de tävlande. En annan viktig faktor är hur många poäng eller mål som de
tävlande vanligtvis får i en tävling. Fotboll är en sport där lagen vanligtvis gör
relativt få mål, vilket försvårar förutsägelsen av resultatet. Ändå har vissa studier
inom förutsägelse av fotbollsmatcher lyckats uppnå noggrannheter närmare 90
procent [1]. De flesta studier har dock uppnått en noggrannhet mellan 55 och 70
procent [26]. De vanligaste maskininlärningsalgoritmerna som används för att
förutsäga fotbollsmatcher är naiv bayesiansk klassificerare, logistisk regression och
artificiella neuronnät.

Målet med detta projekt var att jämföra olika maskininlärningsalgoritmer och
metoder, och analysera hur bra man kan förutsäga fotbollsmatcher med hjälp av
dem. Projektet innefattar hela maskininlärningsprocessen, från datainsamling till
utvärdering av resultaten. I projektet förutsades resultaten på fotbollsmatcher i den
tyska högsta divisionen Bundesliga, säsong 2015–2016. Endast matchernas
slutresultat förutsades, alltså inte exakt hur många mål som lagen gjorde.
Förutsägelserna utfördes med hjälp av data som är relaterad till lagen som helhet,
de individuella spelarna och lagens tidigare matchresultat.

Flera olika datakällor användes i projektet. Data om lagen och spelarna hämtades
från en SQL-databas vid namn European Soccer Database. Den publicerades online
på Kaggle av Hugo Mathien. Databasen innehåller attribut, vars ursprungliga källa
är fotbollspelserien EA Sports FIFA. Lagens och spelarnas skicklighet inom olika

65
kategorier har betygssatts mellan 1–99, där 99 är det bästa möjliga betyget.
Databasen innehåller också matchresultat, men vissa attribut var ofullständiga eller
saknade värden. I stället hämtades matchstatistik på den valda ligan och säsongen
från en CSV-fil på Datahub.io.

Nästa steg efter datainsamling är databehandling. Databehandlingen utfördes i

programmet Visual Studio Code med programmeringsspråken Python och SQL.
För att kunna förutsäga matcherna med hjälp av de tre olika typerna av data, var det
nödvändigt att förena dem med information om varje match. Informationen som
behövs är vilka lag som spelade mot varandra och matchresultatet. Efter att lag- och
spelarattributen hade förenats med matcherna, togs onödiga attribut bort från
tabellerna. Tabellerna konverterades sedan till CSV-filer, för att enkelt kunna
hanteras i miljön för själva maskininlärningen.

Maskininlärningsmodellerna skapades med Python i Jupyter Notebook. Tre

maskininlärningsalgoritmer testades i projektet: naiv bayesiansk klassificerare,
logistisk regression och slumpmässig skog. Naiv bayesiansk klassificerare är
baserat på sannolikheter och är bra på att beakta tidigare händelser. Logistisk
regression är specialiserad på binär klassificering, men genom att använda flera
olika kombinationer av klasser fungerar den också för flerklass klassificering. En
slumpmässig skog består av flera beslutsträd, vilka är trädformade strukturer som
använder villkorlig logik.

Indata delades upp i träningsdata och testdata, så att 70 procent användes som
träningsdata och 30 procent som testdata. I den ursprungliga indata med lagattribut,
fanns det 20 attribut. Maskininlärningsmodellerna uppnådde noggrannheter runt 50
procent då alla attribut användes. De attribut som hade minst betydelse på resultatet
togs inte i beaktande i följande test. Det bästa resultatet med lagattribut, uppnåddes
av en modell som använde sex attribut med algoritmen logistisk regression.
Modellen lyckades uppnå en noggrannhet på 63 procent. I den ursprungliga indata
med spelarattribut fanns det 35 attribut för alla de 22 spelarna i en match. Eftersom
det totala antalet attribut blev så stort, testades kombinationer av endast några
attribut per spelare. Den bästa modellen som använde spelarattribut, använde
attribut över spelarnas totala skicklighet med algoritmen slumpmässig skog.
Modellen lyckades uppnå en noggrannhet på 54 procent. För att kunna använda
tidigare matchstatistik i förutsägelsen av kommande matcher måste man räkna ut

66
medeltal av attributen. Ett medeltal av statistiken, från de 20 senaste matcherna för
varje lag i varje match, räknades ut. Vid användning av tidigare matchstatistik är
det viktigt att beakta den kronologiska ordningen på matcherna, så att matcherna i
träningsdata har inträffat före matcherna i testdata. Den bästa modellen som
använde matchstatistik, använde algoritmen slumpmässig skog. Den lyckades
uppnå en noggrannhet på 55 procent.

Slutsats

Förutsägelse av fotbollsmatcher kan utföras på många olika sätt med hjälp av

maskininlärning. Enligt tidigare forskning uppnår de flesta förutsägelser av
fotbollsmatcher en noggrannhet mellan 55 och 70 procent [26]. Det stämmer
överens med det bästa resultatet som uppnåddes i mitt projekt. Den högsta
noggrannheten uppnådd var 63 procent, i en modell som använde lagattribut med
algoritmen logistisk regression. För att få ökad noggrannhet i projektet, kunde man
ha använt data från flera säsonger, så att modellerna skulle få mera data att träna på.
Den största orsaken till att inte en större noggrannhet uppnåddes i projektet, var att
modellerna inte gjorde tillräckligt många korrekta förutsägelser av oavgjorda
resultat, i relation till hemmavinster och bortavinster.

Många studier inom förutsägelse av fotbollsmatcher har använt naiv bayesiansk

klassificerare, fast i detta projekt presterade den sämre än logistisk regression och
slumpmässig skog i de flesta test som gjordes. En nackdel med naiv bayesiansk
klassificerare är att den antar att alla attribut är oberoende från varandra, vilket inte
är fallet i en stor del av den data som användes i detta projekt. Den bästa algoritmen
för lagattribut var logistisk regression, medan den bästa algoritmen för
spelarattribut och matchstatistik var slumpmässig skog. Naiv bayesiansk
klassificerare presterade dock mest balanserat utav de tre algoritmerna.

Fastän det är omöjligt att förutsäga resultaten i fotbollsmatcher perfekt, går det att
uppnå noggrannheter som är mycket större än slumpmässiga resultat. Genom att
jämföra flera olika kombinationer av maskininlärningsalgoritmer och indata kan
man uppnå förvånansvärt bra resultat.

67
 Bibliography

[1] F. Owramipur, P. Eskandarian and F. S. Mozneb, “Football Result Prediction with

Bayesian Network in Spanish League-Barcelona Team,” International Journal of
Computer Theory and Engineering, vol. 5, no. 5, 2013.

[2] W. Rahman, in AI and machine learning, New Delhi, India, SAGE Publications Pvt
Ltd, 2020, pp. 19-20, 38-41.

[3] O. Campesato, in Artificial intelligence, machine learning, and deep learning,

Dulles, Virginia ; Boston, Massachusetts ; New Delhi, Mercury Learning and
Information, 2020, pp. 26-31, 39, 67-73, 87-89, 105.

[4] E. Alpaydin, in Machine Learning, Massachusetts, MIT Press, 2021, pp. 16-18, 105-
107, 127-129, 143.

[5] J. Rogel-Salazar, in Data Science and Analytics with Python, New York, Chapman
and Hall/CRC, 2017, pp. 102-104.

[6] Handelman et al, “Peering Into the Black Box of Artificial Intelligence: Evaluation
Metrics of Machine Learning Methods,” American Journal of Roentgenology, vol.
212, no. 1, pp. 38-43, 2019.

[7] T. Dietterich, “Overfitting and Undercomputing in Machine Learning,” ACM

Computing Surveys, vol. 27, no. 3, 1995.

[8] V. Nasteski, “An overview of the supervised machine learning methods,” Bitola,
Macedonia, 2017.

[9] N. Grira, M. Crucianu and N. Boujemaa, “Unsupervised and Semi-supervised

Clustering: a Brief Survey,” INRIA Rocquencourt, Le Chesnay Cedex, France, 2015.

[10] M. Mohammed, M. B. Khan and E. B. M. Bashier, Machine Learning: Algorithms

and Applications, Auerbach Publications, 2017.

[11] Liu et al, Computational and Statistical Methods for Analysing Big Data with
Applications, Australia: Academic Press, 2016.

68
[12] J. Cheng and R. Greiner, “Comparing Bayesian Network Classifiers,” University of
Alberta, Edmonton, Alberta, Canada, 1999.

[13] T. Soni Madhulatha, “AN OVERVIEW ON CLUSTERING METHODS,” IOSR Journal of

Engineering, vol. 2, no. 4, pp. 719-725, 2012.

[14] Choi et al, “Introduction to Machine Learning, Neural Networks, and Deep
Learning,” Translational Vision Science & Technology, vol. 9, no. 2, 2020.

[15] J. Brownlee, “4 Types of Classification Tasks in Machine Learning,”

https://machinelearningmastery.com/types-of-classification-in-machine-learning/,
Accessed 5 March 2023.

[16] M. Aly, “Survey on Multiclass Classification Methods,” 2005.

[17] H. K. Jabbar and R. Z. Khan, “METHODS TO AVOID OVER-FITTING AND UNDER-

FITTING IN SUPERVISED MACHINE LEARNING (COMPARATIVE STUDY),” Aligarh
Muslim University, Aligarh, India, 2015.

[18] X. Ying, “An Overview of Overfitting and its Solutions,” Journal of Physics:
Conference Series, vol. 1168, no. 2, 2019.

[19] Sportradar Integrity Services, “BETTING CORRUPTION AND MATCH-FIXING IN

2021,” 2022.

[20] E. Štrumbelj, “On determining probability forecasts from betting odds,”

International Journal of Forecasting, vol. 30, no. 4, pp. 934-943, 2014.

[21] L. Kaunitz, S. Zhong and J. Kreiner, “Beating the bookies with their own numbers -
and how the online sports betting market is rigged,” Cornell University, 2017.

[22] G. Fialho, A. Manhães and J. P. Teixeira, “Predicting Sports Results with Artificial
Intelligence - A Proposal Framework for Soccer Games,” Procedia Computer
Science, vol. 164, pp. 131-136, 2019.

[23] R. P. Bunker and F. Thabtah, “A machine learning framework for sport result
prediction,” Applied Computing and Informatics, vol. 15, no. 1, pp. 27-33, 2019.

69
[24] A. Almazov, “The main factor in predicting football bets,” https://alvin-
almazov.com/theory/the-main-factor-in-predicting-football-bets/, Accessed 10
April 2023.

[25] S. Wilkens, “Sports prediction and betting models in the machine learning age: The
case of tennis,” Journal of Sports Analytics, vol. 7, no. 2, pp. 99-117, 2021.

[26] T. Horvat, “The use of machine learning in sport outcome prediction: A review,”
WIREs Data Mining and Knowledge Discovery, vol. 10, no. 5, 2020.

[27] D. Buursma, “Predicting sports events from past results “Towards effective betting
on football matches”,” in 14th Twente Student Conference on IT, Twente, Holland,
2011.

[28] N. Tax and Y. Joustra, “Predicting The Dutch Football Competition Using Public
Data: A Machine Learning Approach,” TRANSACTIONS ON KNOWLEDGE AND DATA
ENGINEERING, 2015.

[29] Betting Offers, “Home Advantage in Football: How Often Does the Home Team
Win?,” https://www.bettingoffers.org.uk/how-big-is-home-advantage-in-football/,
Accessed 7 February 2023.

[30] Y. S. Taspinar, I. Cinar and M. Koklu, “Improvement of Football Match Score

Prediction by Selecting Effective Features for Italy Serie A League,” MANAS Journal
of Engineering, vol. 9, no. 1, pp. 1-9, 2021.

[31] J. Shin and R. Gasparyan, “A novel way to Soccer Match Prediction,” Stanford
University, 2014.

[32] D. Prasetio and D. Harlili, “Predicting football match results with logistic
regression,” in International Conference On Advanced Informatics: Concepts,
Theory And Application (ICAICTA), Penang, Malaysia, 2016.

[33] Razali et al, “Predicting Football Matches Results using Bayesian Networks for
English Premier League (EPL),” in IOP Conf. Ser.: Mater. Sci. Eng., Malaysia, 2017.

[34] H. Chen, “Neural Network Algorithm in Predicting Football Match Outcome Based
on Player Ability Index,” Advances in Physical Education, vol. 9, no. 4, 2019.

70
[35] M. A. Rahman, “A deep learning framework for football match prediction,” SN
Applied Sciences, vol. 2, no. 165, 2020.

[36] H. Mathien, “European Soccer Database,”

https://www.kaggle.com/datasets/hugomathien/soccer, United Kingdom, 2016.

[37] R. Murphy, “FIFA player ratings explained: How are the card number & stats
decided?,” https://www.goal.com/en/news/fifa-player-ratings-explained-how-are-
the-card-number--stats-decided/1hszd2fgr7wgf1n2b2yjdpgynu, Accessed 13
March 2023.

[38] FIFA Index, “EA Sports FIFA16,”

https://www.fifaindex.com/teams/fifa16_73/?league=19&order=desc, Accessed
January 2023.

[39] DataHub.io, “season-1516,” https://datahub.io/sports-data/german-bundesliga,

Accessed January 2023.

[40] H. Bhavsar and A. Ganatra, “A Comparative Study of Training Algorithms for

Supervised Machine Learning,” International Journal of Soft Computing and
Engineering (IJSCE), vol. 2, no. 4, 2012.

71