Predicting the outcome of soccer matches in

order to make money with betting

Master Thesis
Business Analytics and Quantitative Marketing

Author:
Tim van der Zaan (362006)

Supervisor:
Dr. Andreas Alfons

Econometrics & Operations Research

Erasmus School of Economics
Erasmus University Rotterdam

February 3, 2017
 Abstract
For several decades soccer is reckoned among the most beloved sports on the globe.
This manifests itself by continuous attention from millions of people from all over the
world. The money volume involved in soccer is exceptional and expands consistently.
Simultaneously, bookmakers profit acquired from soccer betting has increased over
the recent past. The urge on the construction of statistical betting strategies that
systematically generate profit can potentially be accomplished since nowadays more
soccer-related data is publicly accessible. This paper presents various betting strategies
that exploit the predicted likelihood of the outcome possibilities assigned by match
outcome prediction models. In this paper is shown that the extension of an existing
betting strategy generates an average out-of-sample profit share of 0.26 per bet euro.
Contents

1 Introduction 2

2 Related work 4
2.1 Clustering algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Match outcome models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3 Soccer betting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 Data 7

4 Method 8
4.1 Clustering of the competitions . . . . . . . . . . . . . . . . . . . . . . . . 8
4.2 Proposed models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4.2.1 Ordered Probit prediction model . . . . . . . . . . . . . . . . . . 9
4.2.2 Simplified prediction model . . . . . . . . . . . . . . . . . . . . . 13
4.2.3 Extended ordered probit models . . . . . . . . . . . . . . . . . . 14
4.3 Comparison of match predictions . . . . . . . . . . . . . . . . . . . . . . 18
4.3.1 Converting bookmakers odds . . . . . . . . . . . . . . . . . . . . 19
4.3.2 Comparing the models . . . . . . . . . . . . . . . . . . . . . . . . 19
4.4 Proposed betting strategies . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.5 Season predictability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.5.1 Season simulation . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.5.2 Comparison of season table vectors . . . . . . . . . . . . . . . . . 29

5 Evaluation 31
5.1 PCA and Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.2 Model predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.3 Model comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.4 Evaluation of betting strategies . . . . . . . . . . . . . . . . . . . . . . . 43
5.5 Predictability of the seasons . . . . . . . . . . . . . . . . . . . . . . . . . 50

6 Conclusion 54

Appendix 60
A Output PCA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
B Output cluster method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
C Extended AIC models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
D Extended BIC models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
E Evaluation methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
F Variable explanation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
1 | Introduction

Along with the start of the soccer season 2015-16, the bookmaker William Hill set
a fixed-odd equal to 5,000 for Leicester City to win the Premier League (Englands
highest division of soccer). Joe Crilly, Press-Officer of William Hill, reacted on Leicester
City leading the Premier League with only twelve matches left to play; If they hold
on to claim the title, it would be the biggest upset in sports history., quoted in [27].
When Leicester City unexpectedly became the champions of the 2015-16 Premier
League season it demonstrated once more that the real-life outcome of sport events
can be surprising. From the occurrence of this unforeseen denouement one can deduce
that the outcome of soccer matches is not always easy to predict for both bettors and
bookmakers. According to Dobson and Goddard in [7] the popularity of soccer, which
is currently the most watched and participated sport worldwide, was rising across the
globe in the last years of the 20th century. Historically the sport has been popular in
especially Europe and South-America, however [21] emphasizes the rising popularity
of soccer in the USA over the last decades. Furthermore, the recent expenses of Chi-
nese soccer clubs indicate an (expected) growth in Asian club soccer. With the growth
in popularity, the growth in online gambling on soccer matches has grown even more
rapidly over the last years. The possibility to bet online and on mobile phones con-
tributed largely in this enormous growth according to Hing [18]. Finnigan and Nordsted
even state in [10] that betting on soccer provide online bookmakers the largest turnover
compared to other sports. Preston and Smith [36] assert that bettors, which are present
among all welfare and racial classes in modern society [11], are fond of making deci-
sions depending on their own believes. It is also stated that bettors might think that
they personally posses the knowledge and cleverness to select winning bets. Besides,
various bettors believe their fate is well blessed. In [36] is asserted that bettors even
find pleasure in betting whilst losing money. Sauer endorses this in [35] by asserting
that gamblers mainly bet for pleasure and not in the first place to earn money.
I get nervous before the games because soccer is not mathematics, its completely
unpredictable. Thats what makes it so popular. - Arsne Wenger, manager of Premier
League club Arsenal, quoted in [29]. Just like Arsne Wengere indicates here, prediction
regarding the outcome of soccer matches (and competitions) is associated with lots of
uncertainty. The outcome of soccer matches is partly dependent on measurable factors,
nevertheless a large share cannot be explained by measurable factors. In papers of
Pollard and Pollard as well as Sosik and Vergin, in [32] and [40], respectively, it is
proposed that the home team experiences an advantage over the away team. In [5],
Clarke and Normand demonstrate that the distance between the stadium of the home
and away team is an influential factor. Asimakopoulos and Goddard [12] incorporated
these, and other, soccer-related variables in an ordered probit prediction model in an

2
attempt to accurately forecast the outcome of soccer matches. The predictions from
this model are able to compete with the bookmakers forecasts. In [37], prediction
results from various forecasting models are obtained and these assigned match outcome
probabilities are used as input for several betting strategies. The obtained profit varies
over the different betting strategies. Several of these strategies generate positive profit.
Although Arsne Wenger asserted that soccer is not mathematics, statistics will be
used in this paper to come up with models that predict the outcome of soccer matches.
The focus is on soccer in Europe, in particular the matches within the highest national
leagues of European countries. Cluster analysis is used to classify the competitions into
groups. The resulting clusters assist by the selection of the competitions that will be
checked more thoroughly. Because Hamadani found in [16] that for a certain compe-
tition the set of explanatory variables varies each season, models are build to forecast
the outcomes of soccer matches in several seasons and competitions. More precisely,
four different prediction models are constructed for three recent seasons of four differ-
ent competitions. These forecast models are evaluated by applying several measures
on out-of-sample outcome predictions, and compared to the outcome probabilities as-
signed by bookmakers. Subsequently, several betting strategies will be considered in an
attempt to systematically make profit. The probabilities for the outcome possibilities
estimated by the prediction models are input for these betting strategies. The link
between the obtained profit and the season predictability is studied over the considered
seasons and competitions. Three different measures are applied to indicate the season
predictability.
The remainder of this paper is structured as follows. In Ch. 2 existing related
literature is discussed. The data used during the research is described in Ch. 3. In
Ch. 4 the methods used during the research are described into detail. The results
of the performed methods are denoted and evaluated in Ch. 5, whereas in Ch. 6 a
conclusion is drawn and some suggestions for further research are provided.

3
2 | Related work

The goal of this research is to construct models that can accurately predict the outcome
of soccer matches, based on historic match results and other statistics. At first, cluster
analysis will be applied to split the competitions into groups. Existing literature on
clustering methods will be discussed in Sect. 2.1. Subsequently, Sect. 2.2 presents
an overview of existing literature regarding the prediction of soccer match outcomes.
Numerous papers exist that propose varying models in order to predict soccer match
outcomes. Sect. 2.3 elaborates on literature concerning strategies specialized in betting
on soccer matches.

2.1 | Clustering algorithms

Clustering methods provide the possibility to split a set of data points into several
groups based on the similarity of the attributes. Within a cluster the data points
contain similar characteristics whereas data points assigned to different clusters contain
smaller similarity of attributes. The K-means clustering algorithm, also called Lloyds
method [25], is a widely used clustering method. A requested number of clusters is the
only necessary input to perform this method. The amount of clusters, K, is acquired
applying the pseudo-F measure [41]. In [33] it is suggested to reduce the dimensionality
using principal component analysis (PCA) [30] before clustering is applied in case of
either a high-dimensional dataset or large correlation within the dataset. In [24] it
is showed that the approach of using PCA before K-means clustering provides good
cluster output on the original data.
In this paper a clustering method will be used to determine a selection of European
competitions that will be investigated. Because of correlation among the data attributes
prior to the K-means cluster analysis PCA will be performed.

2.2 | Match outcome models

All the proposed models in existing literature concerning the prediction of soccer match
outcomes can mainly be divided into two kinds of models; the first kind of models
forecasts the amount of goals scored per team and so, indirectly, the match outcome
is predicted. These models are called goal-models. Maher [26] proposes such a goal-
model that forecasts the amount of goals scored by both the home and away team using
poisson distributions. This prediction is based on attacking and defending parameters
for both the home and away team, as well as a parameter representing the home town
advantage. Plenty of literature is published in which several models are proposed to
indirectly forecast soccer match outcomes based on this model of Maher.

4
 More recently several papers are published in which the outcome of a soccer match
is predicted directly. These toto-models predict whether a soccer match ends in a vic-
tory for the home team, away team, or that a draw occurs. Koning [22] proposed to
use an ordered probit regression model in order to investigate the competition balance
of Dutch soccer. Based on this model of Koning, Graham and Scott [14] proposed a
comparable model. They compared the resulting probabilities of their dynamic pro-
bit regression model on English leagues, from August 2004 to November 2006, with
the outcome probabilities obtained from bookmaker William Hill. Although the pre-
dicted probabilities of William Hill did outperform the constructed model, it was not a
significant difference. In addition to current season match results all sort of other ex-
planatory variables are included in the prediction models proposed in [12]. The English
match outcomes of seasons 1998-99 through 2000-01 are predicted in Asimakopoulos
and Goddard [12] applying an ordered probit model as well. The resulting probabilities
can compete with the predictions of the bookmakers. A comparison of the betting
profit obtained from the probabilities assigned by several prediction models, containing
both toto-models and goal-models, is applied in [37]. Snyder shows in this paper that
the betting profit of various strategies is significantly higher for toto-models than for
goal-models.
During this research the accessible data combined with the results found in [37] made
us conclude to concentrate on the direct modelling of soccer match outcomes. Four
different toto-models will be constructed incorporating current season match results,
previous season match results and various other explanatory variables.

2.3 | Soccer betting

The amount of papers that elaborate upon strategies to bet on soccer matches is limited.
A naive betting strategy is proposed in [38]. The strategy systematically bets on the
outcome that historically occurred most frequently, a home win. For the soccer seasons
199900 through 2001-02 of the German highest national league the obtained betting
profit was not significantly different than the generated profit obtained from several
more sophisticated betting strategies. In [14], Graham and Scott studied betting odds
of bookmaker William Hill for the English Premier League seasons 2001-02 and 2002-
03. They found that the assigned odds of William Hill for home wins and draws are
better than for away wins. Betting on a home win, draw or away win for all matches
generated a profit share per bet euro of -0.111, -0.103 and -0.160, respectively.
Langseth suggests in [23] to solely bet on betting options with an expected positive
profit. Equal betting on these matches results in positive profit for the English Premier
League season 2011-12. Depending the weights on the assigned probability, as that
the expected profit is equal for all betting options, results in a similar profit profile.
Another betting strategy in which the expected value of the betting budget, expressed in

5
logarithm, is maximized, called Kellys betting strategy [20], is performed as well. In [34]
a betting strategy is introduced where the betting weights depend on the expected profit
and the variance of this profit. These two strategies gave approximately similar profit
as the aforementioned strategies for the seasons 2011-12 and 2012-13 in the English
Premier League. Not one of these four betting strategies outperforms the others
systematically based on the models build by Langseth in [23]. Snyder [37] proposes
twelve betting strategies, of which multiple are build upon the betting strategy of
Kelly. Ten of these twelve betting strategies generate a positive profit share per bet
euro for multiple seasons and competitions.
In this paper extensions on several existing betting strategies are applied on the
results of the constructed models. In addition, the naive prediction model is used as
a baseline indication. The profit share per bet euro obtained by the betting strate-
gies will be studied for three recent seasons of four different European competitions.
Furthermore, a potential relation between the obtained profit share and the season
predictability for the considered seasons is investigated

6
3 | Data

In this paper models are provided that predict the outcome of soccer matches in several
European national leagues. To be able to build such models it is essential to have
data regarding matches played in these leagues over the past years. From the sport
database of Infostrada data is obtained for the highest national leagues of 51 European
countries. This dataset covers club and score information regarding matches of the
seasons 1986-87 through 2015-16. Besides the data covering the matches in the national
leagues, data regarding European matches and national cup matches is included in this
dataset. Unfortunately, the data for several competitions contains a lot of missing
values for certain attributes, e.g. the number of spectators per match. Nevertheless,
for sixteen national leagues the dataset is complete. Two of these competitions are
summer competitions, the remaining fourteen competitions are winter competition.
Summer competitions start in the spring and end in the autumn as that no matches
are played in the winter period. On the other hand, winter competitions start in the
autumn and end before the summer. The data from Infostrada has been the input
for a tool of Hypercube Business Innovation B.V. This tool provides an up-to-date
strength indication for all clubs in the highest division of the 51 European countries,
called the European Club Index (ECI)1 . This model is built based on the popular chess
rating system called the ELO-rating, introduced in 1978 by Arpad Elo in [8]. The
ECI value of each club depends on the results of historic matches and is updated after
either a national or international round of matches for club teams. Clubs ECI strength
indication estimated by this tool is available from the start of the winter competition
season 2007-08 through 2015-16. Besides, a dataset containing betting odds for matches
of European competitions is obtained from www.oddsportal.com. These match odds
represent for all three outcome possibilities the average of odds assigned by multiple
global bookmakers, such as Bet365, Unibet and Bwin. The online retrieved match
odds cover the matches played in the highest divisions of European competitions since
the start of the winter season 2007-08.

1 http://www.euroclubindex.com/asp/Ranking.asp

7
4 | Method

In order to answer the main research questions several econometric methods will be
applied in this paper. In this chapter, these methods are introduced and discussed.
In Sect. 4.1, cluster analysis is introduced for the selection of a number of European
national competitions. Sect. 4.2 introduces four different ordered probit models to
predict the outcome of soccer matches. Multiple evaluation measures to compare the
results of the models are discussed in Sect. 4.3, whereas Sect. 4.4 introduces several
betting strategies. Sect. 4.5 elaborates on measures that indicate the predictability of
a certain competition.

4.1 | Clustering of the competitions

In Ch. 3 is stated that the dataset used during this research contains soccer-related infor-
mation about the national league of sixteen European countries. These leagues differ
in composition, strength and other characteristics. Based on season 2015-16, fifteen
different league-specific variables are considered to reflect the characteristics of these
competitions. The competition characteristics vary from the average stadium capacity
to the mean age of the players that have played during the season. A cluster analysis
will be applied to split these competitions into several groups. The competitions within
a cluster contain comparable characteristics whereas competitions in different clusters
vary relatively more in characteristics.
Because a number of league characteristics seem to be correlated the results of
normal clustering algorithms would be affected and the reliability of corresponding
results can be questioned. To overcome this problem, [1] suggests to perform PCA be-
fore the cluster analysis is applied. A normalized version of the initial dataset is used.
This mathematical procedure reduces the dimensionality by finding a number of new
uncorrelated underlying components that are a composition of the original (possibly)
correlated variables. This is achieved by maximizing the variance explained by these
new components. Components with the largest eigenvalues contain the most useful
information about the initial data. Decision on the number of components is an im-
portant step in PCA. A popular criterion, suggested in [42], is to use the components
with eigenvalues larger than one. This corresponds to using components explaining at
least one Qth of the total explained variance, where Q gives the number of variables.

8
 Subsequently, a K-means clustering algorithm will be applied on the set of orthogo-
nal components. The number of clusters, K, is the input for this clustering algorithm.
The pseudo-F measure [41] is seen as the rule of thumb to decide on the number of
clusters. Using the K-means clustering method the starting points, cluster means, are
randomly assigned at first. Thereafter, the method will edit these cluster means until
the optimal ones are obtained. Each of the data points is assigned to the closest cluster
mean.

4.2 | Proposed models

In this section four models will be introduced to predict match outcomes. Each of these
proposed models can be categorized as an ordered probit model [15]. An introduction of
the ordered probit models will be given in Subsect. 4.2.1. The first constructed model,
based on solely the ECI values of the competing clubs, is described in this subsection
as well. Subsect. 4.2.2 elaborates upon an ordered probit model where the parameters
are estimated based on data of the past three seasons. The third and fourth model,
incorporating several explanatory attributes, are introduced in Subsect. 4.2.3.

4.2.1 Ordered Probit prediction model

The prediction of soccer match outcomes in this paper is performed using ordered probit
models. These models generate probability estimations for a home win, draw and away
win for each match. The statistics behind the ordered probit model will be given as the
first model will be introduced. This first model is referred to as the basic ECI model
in the remainder of the paper. The ordered probit model incorporates an unobserved
latent variables. The latent variable is dependent on explanatory variables concerning
the competing clubs of match i, given by xi . This gives the following unobserved latent
variable yi = xi + i , where i N (0, 1). As i is standard normally distributed and
yi is linearly related to i , yi is also normally distributed under the assumption of fixed
predictors.
Now, imagine a match between two clubs with the exact same characteristics and
similar historic results, playing on neutral territory. Although the data suggests equal
chances for both teams to win the match several unexplained factors, such as luck, cause
that this match will not always end in a draw in reality. For each possible match this
ordered probit model will assign probabilities to all three outcome possibilities. Fig. 4.1
visualizes the normally distributed yi and points out that the value of yi might be dif-
ferent per match depending on club and match characteristics. This way, the values of
the latent variable yi varies per match whereas the values of 1 and 2 are fixed for a
season. This causes that outcome predictions differ per match.

9
 1 0 yi 2

Figure 4.1: This figure visualizes the relation between the latent variable yi , which
follows a normal distribution, and the treshold values 1 and 2 .

Irrespective of the value of the latent variable, yi , the treshold values, 1 and 2 ,
separate the density of the normal distribution into three parts. The values of 1 and
2 indirectly incorporate the division of chances assigned to a home win, draw and an
away win for a specific competition season, and so home advantage is incorporated in
this model. The specific estimation of these parameters is based on historic data and
will be discussed later in this subsection. Although the treshold values, 1 and 2 , are
fixed within a season, the chances for each possible outcome, home win, draw and away
win, vary per match. This is caused by different match and club characteristics, xi ,
resulting in different values of yi . To come back at the previous mentioned example of a
match between two clubs with equal characteristics. The corresponding latent variable,
yi , follows a normal distribution with equal chances assigned to the occurrence of a
home win and an away win. For the basic ECI model, the only explanatory variable
affecting the latent variable is the difference in ECI value between the home and away
club, denoted by ECIHomei and ECIAwayi , respectively. This results in the following
latent variable yi ;

yi = [ECIHomei ECIAwayi ] + i . (4.1)

10
 Fig. 4.1 visualizes the relation between the treshold values, 1 and 2 , and the unob-
served latent variable yi . The following equations illustrate how the latent variable and
the treshold values are linked to the ordinal response of the match outcome {1, 0, 1} ;

yi = 1 (home win) if yi + i > 2, (4.2)

yi = 0 (draw) if yi + i < 2 and yi + i > 1, (4.3)

yi = 1 (away win) if yi + i < 1. (4.4)

Ones the parameters, 1 , 2 and , are estimated the match-specific value of the
unobserved latent variable yi can be used to calculate the probability of occurrence for
the three match outcome possibilities. More specifically, the probability of a home win
can be calculated using Eq. 4.5, whereas Eq. 4.6 shows the calculation to obtain the
probability of an away win. The probability of a draw can be retrieved from Eq. 4.7,
which can also be derived ones the probability of a home win and an away win are
calculated.

P [yi = 1] = P (i > 2 yi ) = 1 i (2 yi ), (4.5)

P [yi = 1] = P (i < 1 yi ) = 1 i (1 yi ), (4.6)

P [yi = 0] = P (1 yi < i < 2 yi )

= i (2 yi ) i (1 yi ) (4.7)
= 1 P [yi = 1] P [yi = 1],

where i represents the cumulative distribution function (CDF) of a standard nor-

mal distribution for observation i.
Estimation of the parameters, 1 , 2 and the variable coefficient(s), , of the latent
variable, is performed using the method of maximum likelihood estimation(MLE). De-
fine M dummy variables for each possible outcome ({-1,0,1}), as that dummy resulti,m =
1 in case match i ends in outcome m M and resulti,m = 0 if not. Per match the
product over all M outcome probabilities raised by resulti,m provides the individual
likelihood. The product over the individual likelihood of all matches gives the likelihood
of the sample, resulting in

N
Y 1
Y
L= P [yi = m]resulti,m
i=n M =1
(4.8)
N
Y N
Y N
Y
resulti,1 resulti,0 resulti,1
= P [yi = 1] P [yi = 0] P [yi = 1] .
i=n i=n i=n

11
 In case Eq. 4.5, Eq. 4.6 and Eq. 4.7, defining the calculations of the outcome probail-
ities, are incorporated into the sample likelihood it can be rewritten to

N
Y
L = [1 i (1 yi )]resulti,1
i=n
N
Y
[i (2 yi ) i (1 yi )]resulti,0
i=n
N
Y
[1 i (2 yi )]resulti,1 . (4.9)
i=n

In case MLE will be performed it is convenient to use the log-likelihood function

to estimate the parameters. Transformation of the likelihood function provides the
following log-likelihood function;

N
X
l = log L = resulti,1 log [1 i (1 yi )]
i=n
N
X
+ resulti,0 log [i (2 yi ) i (1 yi )]
i=n
N
X
+ resulti,1 log [1 i (2 yi )]. (4.10)
i=n

The optimal parameters are obtained by maximizing the log-likelihood with respect
to the parameters, 1 , 2 and the variable coefficient(s), , of the latent variable using
the Quadratic Hill Climbing method [13].
This ordered probit model, the basic ECI model, is constructed to predict the out-
come of three consecutive seasons, 2013-14, 2014-15 and 2015-16 for winter competitions
(and 2013, 2014 and 2015 for summer competitions) for the selected European compe-
titions. In the remainder of the paper 2013(-14) refers to season 2013 for the summer
competitions and season 2013-14 for the considered winter competitions. The parame-
ters for the prediction model of a certain season are estimated based on historic data
about matches from 2007-08 through the season prior to the season under consideration.
So, imagine a model for the season 2013-14 of the English competition, the parameter
values are estimated based on matches from season 2007-08 through 2012-13. Using
this procedure the obtained match outcome predictions are out-of-sample forecasts.

12
4.2.2 Simplified prediction model
For the second model, an ordered probit regression model is constructed where the
difference in ECI values of the competing clubs remains the main input. The treshold
values of the basic ECI model, described in Subsect. 4.2.1, are estimated based on
matches since the start of the season 2007-08. The model introduced in this subsection
incorporates information of the last three seasons to predict the outcome of soccer
matches. The focus with this model is more on the recent past compared to the basic
ECI model. The main difference compared to the basic ECI model is in the estimation
of the parameters. Less parameters have to be estimated. In the remainder of the
paper this second model is therefore called the simplified model. The only parameters
that remain for estimation are the treshold values, 1 and 2 , as the coefficient for
1
ECIHomei - ECIAwayi is fixed at a value of = 1000 . This specific fixed value is
chosen as that the parameter values of 1 and 2 are of a similar magnitude as for the
basic ECI model. The calculation of the latent variable yi for the simplified model is
different and given by

[ECIHomei ECIAwayi ]
yi = + i , (4.11)
1000
where i N (0, 1) and so yi remains normally distributed.

The treshold values are estimated in a different way than for the basic ECI model,
where MLE provides the treshold estimations based on the sample of matches since
the start of the season 2007-08. Instead, this model makes use of the balance of home
wins, draws and away wins of the last three seasons to estimate the treshold values.
For each of these three seasons, the treshold values are set in such a way that the
amount of matches ending in a home win, draw and away win equals the amounts
in real-life. The three obtained values of 1 (2 ) estimated for the three preliminary
seasons are averaged resulting in the estimation of 1 (2 ) for the considered season.
So, the simplified prediction model incorporates the outcome balance of the last three
seasons instead of since the start of season 2007-08. The outcome balance over the
recent seasons seems more useful for the match outcome estimation of the successor
season.
Similar as for the basic ECI model, described in Sect. 4.2.1, this model is utilized
for the match outcome prediction of matches from season 2013(-14), 2014(-15) and
2015(-16). As the parameter values are obtained from the three preliminary seasons,
the obtained outcome predictions remain out-of-sample forecasts.

13
4.2.3 Extended ordered probit models
In this subsection a third and fourth model are introduced that both build upon the
ordered probit regression model described in Subsect. 4.2.1. The only difference is in
the calculation of the latent variable yi , which is more complicated for these extended
models than for the basic ECI model. For the basic ECI model the latent variable
solely depends on ECIHomei ECIAwayi , whereas for the extended models many
more other variables are considered and included in the estimation of the latent variable.
The obtained models vary per season and competition caused by both the difference in
coefficients for the variables incorporated in the latent variable as the difference in the
season-specific treshold parameters.
For these extended ordered probit models many match and club characteristics will
be considered alongside the difference in ECI value of the competing clubs. These char-
acteristics contain information about several factors that might influence the outcome
of soccer matches. The fields of information will be briefly covered in the remainder of
this subsection.

In 2014, Yezus [43] suggested to include a dummy for a match featured as a

regional derby in models that predict soccer match outcomes. Instead of only
regional derbies, dummies for all national club combinations are considered in
the models constructed in this paper. The only restriction is that the competing
clubs must have played against each other in the national competition at least six
times since the start of season 2007-08. This minimum of six matches, or three
seasons, is introduced to prevent a lack of observation.

Blundell [2] uses recent match results as shape indicator in his model, whereas
Dixon and Coles [6] specified recent club form based on the last couple of matches
played. In both papers, where models are proposed to predict soccer match out-
comes, is attempted to express the shape of clubs purely based on recent results.
In this paper several variables are constructed that can be categorized as shape-
related variables. The average number of goals scored and received in (home or
away) matches, the amount of points, the current rank and the percentage of pos-
sible points obtained for each club enclose a small selection of considered shape
indicators. Besides these direct shape indicators, more variables are constructed
that can be seen as shape indicators. Since the season of 2007-08 the changes
in ECI value of clubs after one, three and five consecutive matches are gathered.
From these ECI increment values five equally sized groups are formed using per-
centiles, for all three categories. These groups are formed for each competition
separately. Subsequently, these shape indicators after one, three and five matches
are assigned to the clubs of all matches the model will predict based on their
shape at that specific moment.

14
 Goddard and Asimakopoulos suggest in [12] to incorporate the effect of national
cup matches. They found varying effects for the successive match, both positive
and negative. The influence of national cup matches as well as European matches
will be considered in the models during this study. Dummies will be constructed
to indicate the national cup matches and European matches before national league
matches.

As Vergin and Sosik proposed in [40], the home team in sport obtains an advan-
tage over the away team. Besides the predicted season-specific outcome balance
covered by the values of 1 and 2 , the home advantage of each specific club is
considered. Dummies are constructed for each club in the national competition
based on data since the season 2007-08. In [2] it is suggested to incorporate the
stadium capacity of both the home and away team into the model. Furthermore,
the number of spectators and the distance between stadiums are factors that are
considered in the proposed extended ordered probit models.

In [28], Balmer and Williams found that referees respond different to crowd noise
and stadium size. All national combinations of club and referee since the sea-
son 2007-08 are considered to investigate club preferences of referees. The only
restriction is that the two parties must have crossed at least six times in the
national competition since the season 2007-08. This minimum of six matches
is introduced to prevent a lack of observation. In consultation with Hypercube
Business Innovation B.V. is decided to keep the specific referees anonymous. The
influence of club preferences from referees contributes to the model predictions
but the specific referee is not mentioned by name.

Statistics regarding the performance of Manchester City for each month in the En-
glish Premier League for season 2007-08 through 2015-16 are shown in Fig. 4.2.
Although these statistics do not correct for the strength of the opponents, it looks
like the performance of Manchester City is better in Augustus than their perfor-
mance in for example February or March over the past nine seasons. Although
the patterns of performance per month are not similar for all seasons, the perfor-
mance of Manchester City seems to differ over the months. Similarly, this could
be the case for the time a match is scheduled and the day of the week. In this
paper dummies are constructed for the day of the week and the month of the year.
For the time of the match, competition-specific dummies are constructed. These
dummies represent matches played in a certain time span based on equally sized
groups of historic matches since the season 2007-08, for example between 4PM
and 8PM. The combination of a club and a time span, day of the week or month
of the year is only considered in case this combination has occurred at least five
times in the national league since the start of season 2007-08.

15
 2007-08
20 2008-09
2009-10
2010-11
15 2011-12
2012-13
2013-14
10 2014-15
2015-16
5

Aug Sep Oct Nov Dec Jan Feb Mar Apr May

Figure 4.2: This figure shows the stacked plots of the average amount of gathered
points per match by Manchester City per month for the soccer seasons 2007-08
through 2015-16 in the English highest national league. Each of the plots represents a
certain season.

Available data regarding artificial turf is limited to the Dutch competition. In

the season 2015-16 six Dutch clubs played their home matches on artificial turf,
of whom Heracles Almelo and PEC Zwolle played on artificial turf for several
years. Both of these clubs ended somewhat surprising in the subtop of the league
table in the season 2015-16 of the Dutch national league. Fig. 4.3 presents the
average amount of goals scored in home and away matches by the clubs ended
at the positions four to nine in the season 2015-16. The clubs with the biggest
difference in the average amount of goals scored in home and away matches are
Heracles Almelo and PEC Zwolle. This statistic could be coincidence, however
the influence of artificial turf is investigated and considered in the models for the
Dutch competition. The influence of matches on artificial turf is considered for
each of the clubs in competition based on historic matches of different time spans.

16
 home matches

away matches

2.11
1.98 1.94
2 1.89 2
Average amount of goals

1.78
1.67
1.56 1.50
1.39
1.22 1.28

1 0.94 1

0 0

PEC Zwolle
FC Utrecht

Groningen
Heracles

Vitesse
AZ

Figure 4.3: This figure shows the average amount of goals scored in home and away
matches for the subtop (positions four to nine) of the Dutch highest division in the
season 2015-16 (including PEC Zwolle and Heracles Almelo who play on artificial turf
for several years).

The procedure to construct the models will be explained here. In case the con-
structed attributes in the group of angstgegner are considered, one can imagine that
the amount of attributes is large. This is also the case for the constructed variables
in other fields of information. In order to overcome the problem of a massive amount
of variables in the ordered probit models, a pre-selection step is performed. For this
filtering procedure a simple ordered probit model is formed for all variables, expressed
by Attributei , as is given in Eq. 4.12. The variable ECIHomei - ECIAwayi is included
to correct for the difference in strength of the competing clubs, in order to obtain an
accurate estimate of 2 .

yi = 1 + 2 Attributei + 3 [ECIHomei ECIAwayi ] + i , (4.12)

where i N (0, 1).

17
 Considering the coefficient 2 in Eq. 4.12, only the attributes with a p-value lower
than 0.075 are kept. The resulting selection of remaining variables is used to construct
the models. In order to do so, first all remaining variables are included in the model.
Subsequently, reduction steps are performed using backward elimination [17]. Another
model is constructed using the Akaike information criterion (AIC) [3]. This third
model is referred to as the extended AIC model from now on. Stepwise is checked
whether excluding the least significant variable, based on p-value, will improve the
model according to the model fit of AIC. The model fit of AIC is calculated by

2l 2k
AIC = + , (4.13)
T T
where l indicates the loglikelihood value, T represents the number of observation
and k equals the amount of parameters included in the model.

A fourth model is constructed using an almost similar approach. Instead of AIC

the inclusion of variables for the fourth model depends on the Bayesian information
criterion (BIC) [4]. AIC and BIC are both useful measures to estimate the fit of the
constructed model with a penalty assigned for each included parameter. Overall, the
AIC prefers more complex models, whereas BIC chooses the model that contains the
least number of parameters. The fourth model is referred to as the extended BIC model
in the remainder of the paper. The formula of this information criterion is the following

2l klog(T )
BIC = + . (4.14)
T T
Each season the composition of clubs playing in a certain competition is different
due to seasonal promotion and relegation. Besides, the selection of referees operating
in the highest national division changes over time. As some of the constructed variables
depend on the specific clubs and referees in a competition, the considered variables vary
per competition and even per season. This makes the extended AIC models and the
extended BIC models competition- and season-specific.

4.3 | Comparison of match predictions

In the previous section four prediction models are introduced. In this section meth-
ods are discussed to compare the outcome of these models with each other and to the
probabilities assigned by bookmakers. In Subsect. 4.3.1 a method to convert book-
makers odds to probabilities is given. Subsequently, Subsect. 4.3.2 enumerates several
evaluation measures to compare the prediction accuracy of the models.

18
4.3.1 Converting bookmakers odds
As is mentioned in Ch. 3 data from www.oddsportal.com is used to gather the average
assigned odds by multiple bookmakers for all matches. These bookmakers odds can be
converted to predicted outcome probabilities, also called bookmakers probabilities. The
conversion is performed by the following two steps of equations according to Graham
and Scott [14].

1 1 1
R= + + , (4.15)
Home Draw Away

(1/Home )
P [yi = 1] = Home = ,
R
(1/Draw )
P [yi = 0] = Draw = , (4.16)
R
(1/Away )
P [yi = 1] = Away = ,
R

where Home , Draw and Away represent the bookmakers odds for a home win, a
draw and an away win, respectively. Furthermore, Home , Draw and Away represent
the assigned probabilities for a home win, a draw and an away win, respectively.

Table 4.1 gives the bookmakers odds assigned for the match Arsenal - Tottenham
Hotspur in the season 2013-14. The aforementioned equations are applied on these odds
resulting in the bookmakers probabilities, given in the same table.

Table 4.1: The bookmakers odds and bookmakers probabilities assigned for the
match Arsenal - Tottenham Hotspur in the season 2013-14.

Arsenal vs. Tottenham

Home Draw Away
2.12 3.41 3.62
Home Draw Away
0.453 0.282 0.356

4.3.2 Comparing the models

In order to compare the outcome predictions of the models several evaluation meth-
ods are considered. The models have in common that a probability is assigned to all
three possible outcomes for each match. These assigned probabilities can be considered
in two ways. On the one hand one can see the outcome with the highest assigned
probability as the predicted outcome. On the other hand one can interpret all three
assigned probabilities together. Based on these two interpretations, six different eval-

19
uation methods are introduced. For some of the evaluation methods the calculation is
illustrated with an example regarding three matches of the English Premier League
in September 2013, given in Table 4.2. The assigned predicted probabilities for these
matches are obtained from the extended AIC model described in Subsect. 4.2.3.

Table 4.2: The outcome probabilities for three matches in the English Premier
League seasons 2013-14 calculated by the extended AIC model.

Match Away Draw Home Result

Arsenal - Aston Villa 0.101 0.230 0.669 1-3
Liverpool - Stoke City 0.143 0.266 0.591 1-0
Norwich - Everton 0.429 0.316 0.255 2-2

For each of the six evaluation methods the corresponding mathematical formulation
will be given in this subsection. Input for these formulas are matrices representing the
predicted match outcomes, real-life results and assigned probabilities. For the matches
in Table 4.2 these matrices are constructed and given by the following matrices

1 1

A= 1 (4.17) B= 1 (4.18)


1 0

0.101 0.230 0.669

= 0.143 0.266 0.591 (4.19)


0.429 0.316 0.255

where A represents the predicted match outcomes, B gives the real-life results and
shows the assigned probabilities. The columns of this matrix give the assigned
probabilities of an away win, draw and home win, respectively.

Direct outcome evaluation

For this method the outcome with the highest assigned probability is seen as the pre-
dicted outcome. This method counts the well predicted matches and divides this by the
total number of matches in the season per competition. The direct outcome evaluation
value (DO value) is calculated using the following formula

|A B|
DO(A, B) = . (4.20)
|A|

20
Direct probabilities evaluation

For this method the assigned probabilities are used directly. The method sums up the
assigned probability of the real-life outcome and divides this by the total number of
matches in the season per competition expressed. The direct probabilities evaluation
value (DP value) is calculated using this formula
P P3
i k=1 i,k I[Bi = k 2]
DP (B, ) = , (4.21)
|B|
where I[] represents an indicator function that equals 1 in case the statement be-
tween brackets holds, and 0 otherwise.

Jaccard index

The third evaluation measure to compare the predicted outcomes from the models is
performed based on the Jaccard similarity index [19]. This measure compares the pre-
dicted outcomes with the real-life results. In order to make this possible the predicted
outcome equals the match outcome possibility with the largest assigned probability.
The Jaccard similarity value (J value) is calculated using the following equation

|A B| |A B|
J(A, B) = = . (4.22)
|A B| |A|+|B||A B|

The obtained Jaccard value is bounded between 0 and 1. The closer the Jaccard
similarity index reaches 1, the more similar the vectors and the more accurate the
prediction. The matches in Table 4.2 result in the following Jaccard similarity values
|AB| |AB| 1
J(A, B) = |AB| = |A|+|B||AB| = 3+31 = 15 .

Ordered probit penalty index

The ordered probit penalty index evaluates the distribution of chances over the outcome
possibilities. This method depends on the division of the assigned probability and the
real-life outcome expressed in {-1,0,1}, and penalizes the amount of wrong assigned
probabilities. In case the real-life score equals a home win, the assigned probability of
an away win will be penalized twice as hard as the assigned probability for a draw. The
calculation of this evaluation method is the following:

i,3 i,2 2i,1 , if Bi = 1

X
OP P (B, ) = i,3 i,1 , if Bi = 0 , (4.23)
i,2
i
2 , if B = 1

i,1 i,2 i,3 i

21
 To illustrate this measure the matches from Table 4.2 are considered once more.
The OP P value of the first match equals 0.101 - 0.230 - 2 0.609 = 1.338. The
OP P value for the three matches combined equals the average OP P value, calculated
by 1.388+0.0380.368
3 = 0.327. The model containing the highest OP P value has
most accurately assigned probabilities to the possible match outcomes according to
this evaluation measure.

Rank probability score

In [39], Strumbelj and Sikonja employed the rank probability score (RPS) to evaluate
the accuracy of odds set by several bookmakers. Eq. 4.24 presents this method to
evaluate the probability forecasts of ranked categories that was introduced in 1969 by
Epstein [9].

J 3 i
!2
1 XX X
RP S(B, ) = j,k Bj , (4.24)
J j i=1
k=1

where J is the amount of matches under consideration.

In order to illustrate the method it will be applied on the matches given in Table 4.2.
2
For the first match, the calculation equals (0.101 - 1)2 +((0.101 + 0.230) 1) +(1 -
1)2 = 1.256. Similar calculation for the other matches results in a RP S value of
1.256+0.188+0.249
3 = 0.564 applying the rank probability score. The model containing the
lowest RP S value indicates the most accurate model according to the rank probability
score.

Mean squared error

The mean squared error method [31] is a widely used evaluation method that compares
the predicted outcome with the real-life outcome. The mean squared error value (M SE
value) is calculated using the following equation

P3  2 !
1X
J
i=1 j,i Bj
M SE(B, ) = , (4.25)
J j l

where l represents the number of ordered categories. An illustration of this method

is given by applying it on the matches given in Table 4.2. For the first match of this
(0.6690)2 +(0.2300)2 +(0.1011)2
table the individual M SE value equals 3 = 0.436.

22
 0.436+0.086+0.239
The M SE value of the three matches combined is given by 3 = 0.254.
For the mean squared error method the model with the lowest M SE value is the most
accurate.

The six aforementioned evaluation methods are applied to compare the models with
each other and the bookmakers predictions, obtained from the converted bookmakers
odds.

4.4 | Proposed betting strategies

In this section various betting strategies will be proposed depending on the bookmakers
odds and the match outcome probabilities assigned by the models. At first, a strategy is
considered where one bets that all matches from a season will end in a similar outcome,
independent on the assigned outcome probabilities. This naive strategy is studied for
home wins, draws or away wins in order to study whether one of the three outcome
possibilities is systematically under- or overpriced. Another considered betting strategy
is that bettors bet a unit of stake on the outcome possibility containing the largest
assigned probabilities per match. Statistically more advanced betting strategies seem
more likely to result in sustainable profit. In [23], Langseth suggests that with betting
on soccer matches it is more important to gain money than to be right about the
outcome. Imagine the following three possible independent betting options;

I Probability of an away win is (j =) 0.1 with and odd (j =) 12

II Probability of an away win is (j =) 0.4 with and odd (j =) 3

III Probability of a home win is (j =) 0.8 with and odd (j =) 1.5

The gross of bettors will choose different bets based on their personal believes and
risk attitude. Betting option I contains a small probability of occurrence compensated
by a relatively large potential payout, whereas the potential payout of option III, 1.5,
has a relatively large probability. Clearly, a bettor does not necessarily have to choose
between the betting options. The betting budget C can be divided over N > 1 betting
PN
options with varying weights ci for different matches i, where i ci C.
Each match contains three possible betting options, with corresponding bookmakers
odds j and assigned probability j resulting from one of the constructed models,
where j {Home , Draw , Away } and j {Home , Draw , Away } . The odds
or potential payouts can be seen as a compensation for the uncertainty of the outcome
and the risk taken by the bettor. The key-step is to approach the potential profit j
of betting option j per unit of stake as a random variable with E[j ] = (j j - 1)
and Var[j ] = 2j j (1 j ). Each playing round of league matches multiple betting
options might have an positive expected profit (E[j ] > 0). The aforementioned three

23
betting options, which vary in assigned odds and probabilities, have equal expected
payout. For the betting strategies introduced in the remainder of this section, the
set of betting options is restricted to the ones with an expected profit that is strictly
positive. The betting budget, C, for each of these betting strategies equals the amount
of betting options with a positive expected profit per playing round. This way the
budget is similar for all betting strategies per prediction model. The resulting weights
for the betting options will be used to split the betting budget, C, over the N betting
c
options per playing round as that PN j C equals the specific amount bet on betting
i=1 ci
option j. Using these notations several betting strategies will be introduced.

Equal weights

The N betting options with positive expected profit receive similar weights independent
of the corresponding odds and assigned probability, equal to ci = 1.

Equal payout

This betting strategy assigns weights to the betting options as that the potential output
is similar for all betting options in one specific playing round. Betting options containing
a relatively small assigned outcome probability will get smaller weights. On the other
hand larger weights will be assigned to betting options with larger assigned probability.
1
The weights equal ci = i .

Variance adjusted betting strategy

Rue and Salvesen [34] propose to take the variance of the betting options into account to
obtain optimal weights for the betting options. It is suggested to maximize the expected
profit whilst the variance of the profit is minimized. This equals a minimization of the
expected profit minus the variance and results in the following minimization

min ci i i i (1 i )(ci i )2 . (4.26)

ci

The weights resulting from this minimization equal

1
ci = . (4.27)
2i (1 i )

Kellys betting strategy

As a fourth betting strategy the proposed betting approach of Kelly [20] is implemented.
In this strategy the betting budget C is incorporated to obtain optimal weights. The
utility of having amount C after a bet is determined at ln C, whereas the utility of going
bankrupt equals -. The expected utility is calculated by maximizing the following
equation

24
 max i ln(C + ci i ) + (1 i ) ln(C ci ). (4.28)
ci

This equation generates the following weights

i i 1
ci = C . (4.29)
i 1

A drawback for all of these four betting strategies is that each betting option with
positive expected profit is bet on. Although for some strategies the weights depend on
the assigned probability, each betting option with a positive expected profit is bet on.
Imagine a betting option with an assigned outcome probability of 0.01 and an odd equal
to 101. Despite a positive expected profit one can imagine that a rational bettor is not
likely to bet on this betting option assuming that the predicted outcome probability is
assigned by an accurate prediction model. For this reason an extra restriction on the
set of betting options is performed prior to applying one of the four introduced betting
strategies. This restricts the set of betting options to the ones with j > , where
presents a treshold value between 0 and 0.80. The proposed betting strategies are
applied for various values of . Hence, each of the betting strategies is applied on a set
of betting options with positive expected profit and j > .

4.5 | Season predictability

This section proposes three methods that attempt to measure the predictability of a
certain season. It is investigated whether the obtained profit share per bet euro is
related to the season predictability. The predictability of a season is indicated using
three different measures. In Subsect. 4.5.1 a simulation algorithm is introduced in
order to predict the season outcome. Subsect. 4.5.2 describes a method to compare
the resulting season tables of two consecutive seasons. As a third indicator the profit
obtained from betting on the outcome possibility with the largest assigned probability
is considered.

4.5.1 Season simulation

In Sect. 4.2 is explained that the difference in strength of the competing clubs, embodied
by the ECI value of the clubs, is the main seed for each of the four constructed models.
In this subsection the ECI value for clubs is used to simulate the outcome of competition
seasons. For each team the ECI value at the start of the competition is input for
the simulation algorithm. The outcome probabilities for the considered matches are
assigned using the simplified model, described in Subsect. 4.2.2. For each match the
assigned outcome probabilities are used to calculate the expected result by

25
 ExpectedResulti = 1 Home + 0 Draw 1 Away
(4.30)
= Home Away .

Once ExpectedResulti is calculated, Resulti {1, 0, 1} is simulated from the

assigned outcome probabilities using a random number generator. Given the simulated
outcome the exact score is simulated based on the occurrence frequency since season
2000-01. The exact score indicates the goals scored by both teams. The simulated
Resulti and ExpectedResulti lead to an update of the ECI value for both clubs. This
change in ECI value, the increment value, is determined by the following equation

Incrementi = [Resulti ExpectedResulti ], (4.31)

where is called the update factor.

This update factor is calculated for each competition season and remains fixed for
the season. The value of depends on the ECI value of the clubs at the start of the
competitions; namely the difference of the maximum and minimum ECI value of clubs
within the competition. Table 4.3 shows the value of that belongs to specific value
ranges of max ECI min ECI. This parameter estimation is encountered from the
dataset.

Table 4.3: This table presents the parameter value of for different values of
max ECI min ECI.

max ECI - min ECI

2000+ 35
1500-2000 30
1000-1500 25
1000- 20

Subsequently, the ECI value of both clubs will be updated using the value of
Incrementi . Consider a match played at time t, the updated ECI value for both
the home and away club at time t + 1 is calculated using

ECIHomet+1,i = ECIHomet,i + Incrementi ,

(4.32)
ECIAwayt+1,i = ECIAwayt,i Incrementi ,

where ECIHomet,i and ECIAwayt,i represent the ECI values before the match and
ECIHomet+1,i and ECIAwayt+1,i represent the updated ECI values after the match.

26
 Eq. 4.32 indicates that [ECIHomet+1,i - ECIHomet,i ] = - [ECIAwayt+1,i - ECIAwayt,i ],
as that a similar amount of ECI points is maintained within the ranking system after
each match. Later in this subsection an example is given concerning the update of the
ECI value for several matches.
The described procedure is performed for all matches in the season schedule, pro-
viding a season simulation. This season simulation is performed 10,000 times which
results in a prediction of the resulting season table. Alg. 1 summarizes the performed
simulation algorithm.

Algorithm 1 Simulation
Initialize the match schedule of the competition season, containing K matches.
Initialize the ECI starting values for each club in the competition season.
Initialize the competition season parameters 1 , 2 and .
Initialize the exact result probabilities for each outcome possibility gathered from
season 2000-01 through the prior season.

for Run m 10,000 do

for (match i = 1 to K) do
Calculate the outcome probabilities for match i (Eq. 4.5, Eq. 4.7 and Eq. 4.6)
Calculate Expected Result (Eq. 4.30).

Simulate the outcome of match i using random number generation.

Simulate the exact result of match i using random number generation.

Calculate the increment of the ECI value based on the result (Eq. 4.31).
Update the ECI values of both clubs of match i (Eq. 4.32).

end for
Determine the winner of league simulation run m.

end for
Determine the probabilities to become champion for each club based on the m per-
formed simulation runs.

For simplicity, the following two assumptions are made for the proposed simulation
algorithm;

The results of European matches and national cup matches have no influence on
the performance in the national league, nor on the ECI value of the clubs in the
league. This causes a simplified representation of the reality because in reality
the results of clubs in other competitions most certainly can have influence on the
performance in the national league.

27
 The strength difference between the competing clubs is related to the outcome of
the match {home win, draw, away win} via the models. The exact score of the
matches does not depend on the strength difference of the competing clubs within
this algorithm. For the simulation algorithm, the exact score of matches has only
small impact on the outcome of the season. Solely in case two or more clubs lead
the season table with an equal amount of points at the end of the simulated season,
the goal difference gets determinative. In the end, goal difference is worth less
than one point. The simulation output points out that goal difference is crucial in
defining the champion for only two to nine percent of the simulation runs. This
percentage varies per competition and season, depending on the number of club
sin the competition and the starting ECI values of the corresponding clubs.

Next, examples are given regarding updating the ECI value of the competing clubs
for three matches of the English Premier League season 2013-14. The three considered
matches are given in Table 4.4 with the corresponding ECI values.

Table 4.4: This table presents three matches played in the Premier League in the
season 2013-14.

Home team ECI Away team ECI

Manchester City 3499 Norwich City 2244
Fulham 2419 Manchester United 3692
Arsenal 3474 Liverpool 3067

Applying the simplified ordered probit model, described in Subsect. 4.2.2, the esti-
mated outcome probabilities for the three matches are given in Table 4.5. The expected
results of the matches, calculated by Eq. 4.30, are given as well. For the English Pre-
mier League season 2013-14 the value of equals 30. Subsequently, the increment
value for the three possible outcomes for each match can be calculated using Eq. 4.31.

Table 4.5: This table presents the outcome probabilities and the expected result for
each of three matches accompanied with the increment value for each of the possible
outcomes.

Home team Away team Home Draw Away Expected Result

Manchester City Norwich City 0.776 0.267 0.560 0.720
Fulham Manchester United 0.148 0.266 0.586 -0.439
Arsenal Liverpool 0.561 0.275 0.163 0.398
ECI ECI ECI
Manchester City Norwich City 9.8 -25.2 -60.2
Fulham Manchester United 50.3 15.3 -19.7
Arsenal Liverpool 21.1 -13.9 -48.9

28
4.5.2 Comparison of season table vectors
In this subsection the predictability of the competition season tables is studied by
another measure. The similarity of resulting season tables of two consecutive seasons
is investigated. In order to do so, the resulting season tables of two consecutive seasons
are transformed into two vectors. The first vector represents the resulting season table
for the clubs in the successive season, whereas the second vector contains the resulting
rank for the clubs in the season table of the prior season. The included clubs per season
will be different caused by promotion and relegation. For this reason the comparison
is performed on the clubs in the highest league of the successive season. The clubs
that were relegated in the prior season are removed from the season table. These clubs
are replaced by the clubs that earned a promotion to the highest league during this
season. In case multiple clubs got promoted, they get ranked based on their results in
the second division. Imagine club p (1st in second division) and club q (2nd in second
division) got promoted to the English Premier League, that contains twenty club. As
two clubs got promoted obviously two clubs got relegated as well. For eighteen clubs the
ranking in both seasons can be retrieved directly. For the two remaining clubs, p and
q, determination of the rank in the successive season is straightforward. Furthermore,
for the the prior season club p gets rank nineteen and club q gets rank twenty assigned
based on the results in the second division.
To illustrate this method, assume an imaginary European competition containing
six clubs from the four considered competitions. Table 4.6 represents the resulting
competition table of the seasons 2019-20 and 2020-21. For this competition the club
that ended last will be replaced by the champion of the second division, presented with
an asterisk in the table.

Table 4.6: This table presents the competition table of the seasons 2019-20 and
2020-21 of the imaginary competition. 1* represents the champion of the second
division, which is promoted to the highest division for the consecutive.

Club 2019-20 Club 2020-21

Atltico Madrid 1 Arsenal 1
Barcelona 2 Barcelona 2
Feyenoord 3 Atltico Madrid 3
Arsenal 4 Feyenoord 4
Liverpool 5 Malm FF 5
Ajax 6 Liverpool 6
Malm FF 1* PSV Eindhoven 1*

29
 Performing the proposed method on this competition results in the following two
vectors;
4 1

2 2



1 3
C= (4.33) D = , (4.34)

3 4


6 5


5 6

where the clubs playing in the highest division in season 2020-21 are considered for
both vectors. Vec. 4.33 represents the resulting season table of the season 2019-10,
whereas Vec. 4.34 represents the resulting season table for the season 2020-21.
A similarity measure is performed on the two vectors to indicate the predictability
of the successive season. This predictability of a season is indicated by the cosine
similarity measure, given by
Pn
C D Ci Di
cos(C, D) = = pPn i=12 pPn , (4.35)
||C|| ||D|| C
i=1 i
2
i=1 Di

where Ci and Di are observations of the vectors C and D, respectively.

In case the cosine similarity equals 1, the resulting season tables are similar in the
two seasons. The minimum value of the cosine similarity for this competition compar-
ison is around 0.5. This occurs in case the resulting seasons table of the successive
season is in no way similar to the resulting season table of the prior season. Hence, in
case the champion of the prior season finishes as last in the successive seasons, whereas
the promoted club becomes champion in the successive season, and so on. The sea-
son end ranks are normalized before the cosine similarity measure is performed. This
measures provides a similarity compared to the previous season indicating the relative
predictability of the considered season.

30
5 | Evaluation

This chapter contains the evaluation and discussion of the results retrieved from the
proposed methods in Ch. 4. In Sect. 5.1 the results for the applied PCA and cluster
analysis are discussed. The constructed models will be evaluated in Sect. 5.2 previous
to the comparison of the models in Sect. 5.3 based on several evaluation methods. Con-
sequently, several betting strategies are performed using the results of the constructed
models. The profitability of these betting strategies will be discussed in Sect. 5.4.
Eventually, robustness checks of the betting strategies are performed and studied in
Sect. 5.5.

5.1 | PCA and Clustering

In order to select a number of competitions from a total of sixteen European competi-
tions, of which the data is complete, is decided to perform a cluster analysis. Caused by
high correlation between the attributes of competition characteristics most clustering
methods are not applicable. To overcome this issue a combination of PCA followed
by K-means clustering is performed. The first five components resulting from PCA on
a normalized dataset of competition characteristics are given in Table A.1. The first
component reflects the strength of the competitions, whereas the second component
seems to cover the difference in strength of clubs within the competitions. A biplot of
the first two components, that is given in Fig. A.1, shows this. The other three pre-
sented components obey a clear declaration. Table 5.1 shows the amount of variance
each of the principal components declare as well as the eigenvalue of these components.
The number of resulting components to use for the K-means clustering is three, as only
the first three components have an eigenvalue larger than one. Besides, each of these
three components explain more than 1/Q of the total variance, where Q indicates the
number of included variables. For this study Q equals 15. An overview of the exact in-
cluded variables can be found in Appendix B. The three suggested components explain
cumulatively 77.5% of the total variance.

Table 5.1: This table gives the amount of variance each of the principal components
(PCs) declare and the eigenvalues of the components.

PC1 PC2 PC3 PC4 PC5

Proportion of variance 0.545 0.129 0.101 0.059 0.056
Cumulative proportion 0.545 0.674 0.775 0.835 0.890
Eigenvalue 8.179 1.930 1.520 0.890 0.835

31
 Using the first three constructed principal components resulting from PCA, the K-
means clustering algorithm is performed to split the sixteen competitions into groups.
A pseudo-F measure points out that two clusters provide optimal clustering for this
selection of competitions. Table 5.2 presents the division of the sixteen competitions
over the two clusters. Cluster one contains seven European competitions that can be
seen as the most appealing competition. The other competitions are assigned to the
second cluster. It seems like the European rank of the competitions has large influence
on the cluster classification. The obtained cluster means, in case the number of clusters
(K) equals two, can be found in Appendix B.

Table 5.2: This table gives the competitions per cluster.

Cluster 1 Cluster 2
England Austria
France Belgium
Germany Denmark
Italy Netherlands
Portugal Norway
Russia Scotland
Spain Sweden
Switzerland
Ukraine

For the remainder of the paper a number of European competition has to be selected
in order to study the accuracy of soccer match outcome prediction. Multiple competi-
tions and seasons will be considered to ensure the prediction accuracy of the models for
several European competitions. When the selected competitions are different, based on
characteristics, the strength of the profitable proposed models is reinforced. From each
of the two obtained clusters two competitions will be selected that will be studied in
more detail. From the first cluster the competitions of England and Spain are selected,
which are the most watched and followed soccer competitions worldwide with the most
money involved. Also, these competitions vary a lot from each other as the Spanish
competition is yearly dominated by the same small amount of clubs, whereas the En-
glish competition has been less predictable over the seasons 2007-08 through 2015-16.
From the second cluster the Dutch competition is selected because extra data concern-
ing artificial turf is available for the Dutch competition. The Swedish competition is
chosen as fourth competition such that one of the considered competitions is a summer
competition.

32
5.2 | Model predictions
For each of the four selected European competitions models are constructed for three
consecutive seasons. These four different prediction models are proposed in Ch. 4. The
models are formed for the seasons 2013(-14), 2014(-15) and 2015(-16) for the competi-
tions of England, Netherlands, Spain and Sweden based on data since the start of the
season 2007-08.
The basic ECI model, described in Subsect. 4.2.1, can be seen as a straightforward
model. The latent variable for each match i depends solely on the difference in ECI value
of the competing clubs, ECIHomei - ECIAwayi . Table 5.3 contains the estimated
parameters for the three seasons of the four selected competitions. The table presents
that is positive for all competitions and seasons, meaning that if the difference in
ECI value (ECIHomei - ECIAwayi ) becomes larger, the probability of a home win
gets larger. The coefficients can not be compared among the seasons and competitions
because the other parameters, 1 and 2 , vary as well. The table indicates that all
three parameters (, 1 and 2 ) do not differ largely over the seasons, whereas the
difference between competitions is larger.

Table 5.3: The table gives the treshold values from the seasons 2013-14, 2014-15 and
2015-16 of The English, Dutch, Spanish and Swedish competition estimated by the
Basic ECI model. Between brackets the standard errors are given.

England 1 2
2013-14 6.55E-4 (3.24E-5) -0.696 (0.030) 0.102 (0.028)
2014-15 6.58E-4 (3.00E-5) -0.672 (0.028) 0.099 (0.026)
2015-16 6.52E-4 (2.80E-5) -0.659 (0.026) 0.104 (0.024)
Netherlands 1 2
2013-14 9.33E-4 (5.04E-5) -0.650 (0.033) 0.043 (0.030)
2014-15 8.95E-4 (4.67E-5) -0.656 (0.030) 0.042 (0.028)
2015-16 8.91E-4 (4.34E-5) -0.644 (0.028) 0.054 (0.026)
Spain 1 2
2013-14 6.29E-4 (3.52E-5) -0.671 (0.030) 0.003 (0.027)
2014-15 6.22E-4 (3.19E-5) -0.660 (0.027) 0.012 (0.025)
2015-16 6.36E-4 (2.91E-5) -0.653 (0.026) 0.028 (0.024)
Sweden 1 2
2013 6.83E-4 (6.78E-5) -0.595 (0.038) 0.100 (0.035)
2014 7.05E-4 (6.17E-5) -0.592 (0.035) 0.110 (0.033)
2015 7.20E-4 (5.78E-5) -0.594 (0.032) 0.110 (0.030)

33
 1
For the simplified model the coefficient is fixed at a value of 1000 . The other
parameters, 1 and 2 , are retrieved from the optimal treshold values of the three
preliminary seasons, depending on the division of match outcomes in these seasons.
Table 5.4 presents the resulting treshold values, 1 and 2 , for the considered seasons
of the four competitions.

Table 5.4: This table gives treshold values for the seasons 2013-14, 2014-15 and
2015-16 of The English, Dutch, Spanish and Swedish competition estimated by the
Simplified model.

Competition 2013(-14) 2014(-15) 2015(-16)

1 2 1 2 1 2
England -0.690 0.136 -0.596 0.138 -0.601 0.136
Netherlands -0.692 0.019 -0.669 0.072 -0.648 0.126
Spain -0.688 -0.011 -0.661 0.036 -0.613 0.080
Sweden -0.631 0.077 -0.609 0.126 -0.607 0.187

For the extended AIC model and the extended BIC model, the latent variable de-
pends on a lot of other variables alongside the difference in ECI value of the competing
clubs. The resulting extended models for three seasons of the four considered compe-
titions can be found in Appendix C and Appendix D, respectively. In those tables all
included variables are presented along with the corresponding coefficients and standard
errors. An explanation of the included variables can be found in Appendix F. The
models for a specific competition vary widely per season, indicating that the prediction
of match outcomes is not straightforward. As expected, the models where the incorpo-
rated variables are selected based on BIC contain significantly less variables compared
to the models where the variables are selected based on AIC.
In Table 5.5, Table 5.6, Table 5.7 and Table 5.8 the variables are presented that are
consistently significant, and so included, in the extended AIC models for the prediction
of the seasons 2013(-14) through 2015(-16) of the English, Dutch, Spanish and Swedish
competition, respectively. Explanation of these included variables can be found in
Appendix F. For each of the competitions the most striking significant coefficient(s)
will be discussed.

34
Table 5.5: The table contains the variables that are significant in all three extended
AIC models (2013-14, 2014-15 and 2015-16) for the English competition. + means a
positive coefficient, whereas - means a negative coefficient in the models for the three
considered seasons.

Variables Sign
5M atchesShapeHteam 3i +
5M atchesShapeHteam + 3i +
AwaymatchesHteami -
ECIHomei - ECIAwayi +
EU AwayDrawP ostW interi +
EvertonShape 1i +
HomeM anCityi +
M anCityEN GRef eree4i -
M anCityV S.N ewcastlei +
M anU nitedShape 3i +
Shape + 1V S.Shape + 1i +

The negative coefficient for AwaymatchesHteami in Table 5.5 shows that data
from preliminary seasons of the English highest league indicates that the probability
of a home win declines with the growth of the amount of away matches played by the
home team. In other words, the probability of a home win is smaller at the end of
the season than a home win of the similar match at the begin of the seasons, ceteris
paribus. Other notable coefficient signs are found for M anCityEN GRef eree4i and
M anU nitedShape 3i . The negative coefficient for M anCityEN GRef eree4i indi-
cates that Manchester City performs less in case English Referee 4 leads their league
match, ceteris paribus. The positive sign for M anU nitedShape 3i shows that in case
Manchester United performed bad in recent league matches, the data of preliminary
seasons assert a larger probability of a win for the next league match of Manchester
United.

35
Table 5.6: The table contains the variables that are significant in all three extended
AIC models (2013-14, 2014-15 and 2015-16) for the Dutch competition. + means a
positive coefficient, whereas - means a negative coefficient in the models for the three
considered seasons.

Variables Sign
ECIOnArtif icialT urf Ateami -
AjaxV S.U trechti -
AverageGoalsAgainstAteami +
AZN ovemberi +
ECIHomei - ECIAwayi +
F eyenoordJanuaryi -
GroningenShape 2i -
P SV Bef ore2P Mi -
Shape 3V S.Shape + 1i +
Shape + 1V S.Shape + 1i +
Shape + 1V S.Shape + 3i -
Spectatorsi +
U trechtBetween6P M And8P Mi -
V itesseN ET Ref eree2i -

Table 5.6 shows that the models built for the seasons 2013-14, 2014-15 and 2015-
16 of the Dutch soccer competition all find a negative coefficient for the variable
ECIOnArtif icialT urf Ateami . This indicates that data from preliminary seasons
finds that the larger the amount of gathered ECI points in away matches on artificial
turf, the larger the probability of an away win against another opponent playing home
matches on artificial turf. The negative sign for AjaxV S.U trechti shows that historic
data from seasons since 2007-08 indicates that FC Utrecht is a so-called angstgegner
for Ajax, meaning that the probability of a win for Ajax is smaller with FC Utrecht as
opponent compared to similarly performing other teams, ceteris paribus. The negative
signs for P SV Bef ore2P Mi and U trechtBetween6P M And8P Mi indicate a significant
worse performance for PSV and FC Utrecht to play before 2 PM and between 6 PM
and 8 PM, respectively. The probability of a win for FC Utrecht is smaller in case the
match is played before 2 PM and a win for PSV gets less likely in case they play a
match between 6 PM and 8 PM, ceteris paribus.

36
Table 5.7: The table contains the variables that are significant in all three extended
AIC models (2013-14, 2014-15 and 2015-16) for the Spanish competition. + means a
positive coefficient, whereas - means a negative coefficient in the models for the three
considered seasons.

Variables Sign
BilbaoSP ARef eree1i -
ECIHomei - ECIAwayi +
EspanyolShape 3i -
EU LossP reW interAteami -
EU AwayW inAteami -
Getaf eSP ARef eree5i +
Getaf eSP ARef eree6i -
HomeBetisi -
SevillaSP ARef eree4i +
SevillaSP ARef eree11i +
Shape + 1V S.Shape + 1i +
Shape + 2V S.Shape 1i +
V alenciaM archi -

The extended AIC models for the seasons 2013-14, 2014-15 and 2015-16 of the Span-
ish competition all find a negative sign for V alenciaM archi . This indicates that historic
data from Spanish seasons since 2007-08 show that the results of matches from Valencia
played in March are significantly worse than other combinations of equally strong clubs
and months, ceteris paribus. Furthermore, Table 5.7 presents that preliminary seasons
point out that an European away win followed by an away league match gives rise to
the probability of an away victory. This can be concluded from the negative coefficient
sign for EU AwayW inAteami .

Table 5.8: The table contains the variables that are significant in all three extended
AIC models (2013-14, 2014-15 and 2015-16) for the Swedish competition. + means a
positive coefficient, whereas - means a negative coefficient in the models for the three
considered seasons.

Variables Sign
3M atchesShapeAteam + 2i -
AIKShape2i +
Cupf ighterAteami -
DjurgardensSW ERef eree2i -
ECIHomei - ECIAwayi +
M atchesLef tHteami +

37
 Table 5.8 presents the variables that are consistently significant in the prediction
models for the seasons 2013, 2014 and 2015 of the Swedish competition. The nega-
tive coefficient sign for Cupf ighterAteami indicates that in case the away team of a
Swedish league match is still playing in the national cup, the probability of a home win
is lower than when the away team is already eliminated from the national cup. The
positive sign for M atchesLef tHteami shows that the probability of a home win gets
smaller as the home team has played more matches.

In Table 5.9, Table 5.10, Table 5.11 and Table 5.12 the variables are presented
that are consistently significant, and so included, in the extended BIC models for the
prediction of the seasons 2013(-14) through 2015(-16) of the English, Dutch, Spanish
and Swedish competition, respectively. An explanation of these included variables can
be found in Appendix F.

Table 5.9: The table contains the variables that are significant in all three extended
BIC models (2013-14, 2014-15 and 2015-16) for the English competition. + means a
positive coefficient, whereas - means a negative coefficient in the models for the three
considered seasons.

Variables Sign
ECIHomei - ECIAwayi +
HomeM anCityi +
M anCityV S.N ewcastlei +

The positive sign for HomeM anCityi in Table 5.9 indicates that, based on pre-
liminary seasons, Manchester City gets a larger probability for a home win than when
equally strong teams compete with each other under similar circumstances. Besides,
the positive sign for the variable M anCityV S.N ewcastlei indicates that a match be-
tween Manchester City and Newcastle United gets larger probability assigned for a win
of Manchester City than when other teams, of equal strength, play against each other,
ceteris paribus.

Table 5.10: The table contains the variables that are significant in all three extended
BIC models (2013-14, 2014-15 and 2015-16) for the Dutch competition. + means a
positive coefficient, whereas - means a negative coefficient in the models for the three
considered seasons.

Variables Sign
ECIOnArtif icialT urf Ateami -
ECIHomei - ECIAwayi +
Spectatorsi +
V itesseN ET Ref eree2i -

38
 Similar to the findings of the extended AIC models in Table 5.6 a positive signif-
icant coefficient is found for ECIOnArtif icialT urf Ateami for all three considered
Dutch seasons by the extended BIC models, given in Table 5.10. Furthermore, a larger
amount of spectators leads to a larger probability for a home victory, according to the
positive sign of Spectatorsi .

Table 5.11: The table contains the variables that are significant in all three extended
BIC models (2013-14, 2014-15 and 2015-16) for the Spanish competition. + means a
positive coefficient, whereas - means a negative coefficient in the models for the three
considered seasons.

Variables Sign
ECIHomei - ECIAwayi +
Getaf eSP ARef eree6i -

The constructed BIC models for the seasons 2013-14, 2014-15 and 2015-16 of the
Spanish competition show a negative sign for Getaf eSP ARef eree6i . This means that
the results of Getafe CF in a match leaded by Spanish Referee 6 are significantly worse
than for another combination of competing clubs and referee, ceteris paribus.

Table 5.12: The table contains the variables that are significant in all three extended
BIC models (2013-14, 2014-15 and 2015-16) for the Swedish competition. + means a
positive coefficient, whereas - means a negative coefficient in the models for the three
considered seasons.

Variables Sign
AIKShape2i +
ECIHomei - ECIAwayi +

In case the shape of AIK can be considered as Swedish shape 2, the probability of a
win for AIK is larger compared to the combination of other equally strong teams classi-
fied in Swedish shape group 2. This is concluded from the positive sign of AIKShape2i
in Table 5.12.

5.3 | Model comparison

In order to find out which of the four constructed models most accurately predicts
the outcome of soccer matches six different evaluation methods are introduced in Sub-
sect. 4.3.2. These evaluation methods all differ and evaluate the outcome predictions
from the models in another way, concentrating on different aspects of the prediction.
Mainly, one of two different approaches is followed by each of these comparison methods.

39
One approach is to take the outcome possibility with the highest assigned probability
as the predicted outcome, whereas with the other approach the division of probabilities
among all three outcome possibilities is evaluated. At first, the most straightforward
evaluation methods are studied. Table 5.13 presents the results from the evaluation
method introduced as the direct outcome evaluation method in Subsect 4.3.2. This
method follows the first approach and takes the outcome possibility with the highest
assigned probability as the predicted outcome. Table 5.14 presents the results from
the direct probabilities evaluation method introduced in Subsect 4.3.2. This method
follows the second approach and takes the division of probabilities assigned to all three
outcome possibilities into consideration.

Table 5.13: The table gives the fraction of correctly predicted match outcomes per
season for each of the four constructed ordered probit models. The score of the best
predicting proposed model per competition seasons is underlined and bold in this
table. An asterisk (*) is added in case the best predicting proposed models performs
better than the bookmakers probabilities.

Model Competition 2013(-14) 2014(-15) 2015(-16) Average

Basic ECI model England 0.588* 0.533 0.464 0.528*
Netherlands 0.466 0.556 0.550 0.524
Spain 0.528 0.570* 0.525 0.541
Sweden 0.510 0.473 0.556 0.513
Simplified model England 0.586 0.525 0.464 0.525
Netherlands 0.462 0.557* 0.551 0.523
Spain 0.525 0.570* 0.520 0.538
Sweden 0.515* 0.498 0.573* 0.529*
Extended AIC England 0.555 0.537* 0.450 0.514
model Netherlands 0.490* 0.546 0.572* 0.536*
Spain 0.524 0.555 0.516 0.532
Sweden 0.483 0.513* 0.538 0.511
Extended BIC England 0.568 0.529 0.458 0.518
model Netherlands 0.464 0.542 0.552 0.519
Spain 0.524 0.558 0.516 0.533
Sweden 0.492 0.500 0.550 0.514

Bookmakers England 0.558 0.529 0.468 0.518

probabilities Netherlands 0.484 0.510 0.487 0.494
Spain 0.547 0.568 0.553 0.556
Sweden 0.492 0.496 0.483 0.490

One can see in Table 5.13 that over the twelve considered seasons, divided over four
competitions, the direct outcome evaluation method finds varying models as the best
performing model. From the table can be concluded that the direct outcome evaluation
method does not clearly indicate a preferred model. The bottom part of the table
contains the results of the evaluation method applied on the bookmakers probabilities.

40
For only three of the twelve seasons, the direct outcome evaluation method prefers
the bookmakers probabilities over each four proposed models. The preferred proposed
model is underlined and bold in Table 5.13. Per competition the preferred model is
compared to the bookmakers odds based on the direct outcome evaluation method. In
case this proposed model is preferred over the bookmakers probabilities, an asterisk (*)
is placed by the specific obtained profit share.

Table 5.14: The table gives the mean of correct assigned outcome probabilities
averaged over the amount of matches per season for each of the four constructed
models. The profit of the best predicting proposed model per competition seasons is
underlined and bold in this table. An asterisk (*) is added in case the best predicting
proposed models performs better than the bookmakers probabilities.

Model Competition 2013(-14) 2014(-15) 2015(-16) Average

Basic ECI model England 0.433 0.419 0.387 0.413
Netherlands 0.405 0.426 0.429 0.420
Spain 0.437 0.457 0.444 0.446
Sweden 0.390 0.389 0.399 0.393
Simplified model England 0.482* 0.423 0.389 0.431*
Netherlands 0.409 0.431 0.443* 0.428
Spain 0.438 0.457 0.439 0.445
Sweden 0.410* 0.404 0.409 0.408*
Extended AIC England 0.441 0.437* 0.396* 0.425
model Netherlands 0.419* 0.435* 0.440 0.431*
Spain 0.448* 0.461* 0.441 0.450*
Sweden 0.386 0.400 0.414* 0.400
Extended BIC England 0.434 0.424 0.389 0.416
model Netherlands 0.415 0.425 0.430 0.423
Spain 0.437 0.456 0.445* 0.446
Sweden 0.386 0.387 0.400 0.391

Bookmakers England 0.429 0.420 0.392 0.414

probabilities Netherlands 0.413 0.421 0.411 0.415
Spain 0.442 0.460 0.443 0.448
Sweden 0.397 0.410 0.398 0.402

The extended AIC model can be seen as the preferred model by the direct prob-
abilities evaluation method in Table 5.14. From the twelve considered seasons, this
evaluation method ranks the extended AIC model as the best of the proposed models
for seven seasons. For the remaining five seasons this method is four times ranked as
second best. The simplified model can be seen as a reasonable alternative model. Ac-
cording to the direct probabilities evaluation method the extended AIC model performs
better in predicting the outcome of soccer matches than the bookmakers probabilities.
This extended AIC model is preferred over the bookmakers probabilities in nine of
twelve considered seasons. The alternative simplified model is preferred over the book-

41
makers probabilities for six of the twelve considered seasons. Also, the last column
shows that the preferred model, averaged over the three considered seasons, performs
better than the bookmakers probability for each of the four competitions.
The results from the other four proposed evaluation methods can be found in Ap-
pendix E. An overview of these results will be discussed here. Similar to the direct
outcome evaluation method the Jaccard index evaluation method sees the outcome pos-
sibility with the highest assigned probability as the predicted outcome. The outcome
of the Jaccard index evaluation method is given in Table E.1. The other evaluation
methods, the ordered probit penalty index, the rank probability score and the mean
squared error method, follow the other approach in which the division of probabilities
over the three outcome possibilities is evaluated. Their results can be found in Ta-
ble E.2, Table E.3 and Table E.4, respectively. The results of the Jaccard index, the
rank probability score and the mean squared error method are comparable to the re-
sults of the direct outcome evaluation method ; the evaluation method does not clearly
prefer one model. For almost all of the twelve considered seasons either the basic ECI,
the simplified or the extended AIC model is the preferred model according to each of
these three evaluation methods. The extended BIC model is in almost no situation the
preferred model. The results of the ordered probit penalty index, which can be found
in Table E.2, indicate a clear preference for the extended AIC model. For eight of the
twelve considered seasons the extended AIC model is preferred over the other three
proposed models by the ordered probit penalty index. Overall, one can conclude that
the preference of a certain model not only depends on the specific season, but also
on the evaluation method. Taken all the six evaluation methods into consideration,
the extended BIC model is the least preferred model. The direct probability evaluation
method and the ordered probit penalty index method indicate that the extended AIC
model is the best prediction model. The other four evaluation methods have no clear
preference for one of the following three models; the basic ECI, the simplified and the
extended AIC model.
Comparing the proposed models with the bookmakers probabilities using the same
evaluation methods leads to varying insights. The direct outcome evaluation method
and the direct probability evaluation method, given in Table 5.13 and Table 5.14, indicate
that the preferred proposed model performs better than the bookmakers probabilities
for nine and eleven of the twelve considered seasons, respectively. According to the rank
probability score method and the mean squared error method the bookmakers provide
better prediction probabilities for match outcomes. From the twelve seasons only three
and one season(s) the preferred model, for that specific season, is able to perform better
than the bookmakers probabilities. Once more the findings here depend largely on the
evaluation method. The specific results from the other four evaluation methods can be
found in Appendix E. Considering all evaluation methods one can conclude that none
of the proposed models is able to systematically beat the bookmakers probabilities.

42
Two of the six evaluation methods point out that the extended AIC model beats the
bookmakers in more than half of the considered seasons, but this finding is totally
different for other evaluation methods. One might conclude that the extended AIC
model is best able to compete with the bookmakers probabilities. The basic ECI model
and simplified model follow the extended AIC model, whereas the extended BIC model
is worst able to compete with the bookmakers probabilities. This suggests that the
extra included variables contain important imformation concerning the prediction of
soccer match outcomes.

5.4 | Evaluation of betting strategies

It is attempted in this paper to come up with a betting strategy that is able to sys-
tematically beat the bookmakers and generate a positive profit with betting on soccer
matches. At first a naive betting strategy is considered to check if a certain match
outcome (home win, draw, away win) is systematically under- or overpriced for one
of the considered competitions. The results of betting based on this strategy over the
period of three seasons, 2013(-14) through 2015(-16) for each competition, are given in
Table 5.15. One-sided students t-tests [17] are applied to check whether the average
profit share is significantly different than zero using a 5% significance level. These results
indicate that betting on one specific possible outcome (home win, draw or away win)
for all matches does not systematically generate a significant positive profit share for
one of the four considered competitions. In only three out of twelve considered betting
situation, given in the table, the profit share is (insignificantly) positive. This betting
strategy makes us conclude that none of the possible match outcomes is systematically
under- or overpriced for one of the four competitions.

Table 5.15: The table gives the portion of profit per bet euro of the naive betting
strategy over a period of three seasons; 2013(-14) through 2015(-16). An asterisk (*)
indicates an average profit that is significantly different than zero.

Home win Draw Away win

England -0.021 -0.080 0.023
Netherlands -0.066 0.048 -0.125*
Spain 0.000 -0.089* 0.045
Sweden -0.053* -0.064* -0.089*

Next, the predictions resulting from each of the four proposed models are utilized to
bet on the matches of the 12 considered soccer seasons. Table 5.16 presents the profit
share per bet euro of betting unit stake on the match outcome with the highest assigned
probability for each of the four proposed models and the converted bookmakers odds.
This table shows that the profit share can vary several percentages per model and not

43
one of the models consistently outperforms the other models. However, the profit share
per model seems to be correlated, suggesting that certain seasons can be seen as harder
to predict than other seasons. In a certain way this betting strategy can also be seen as
an evaluation method to compare the proposed models. For each considered season the
model with the highest profit is underlined. Although the preferred model varies largely
over the seasons, the average portion of profit per bet euro over the twelve considered
seasons points out a clear preference. Where the basic ECI model, the simplified model
and the extended BIC model generate an average profit share of -0.027, -0.038 and
-0.038 per bet euro over the twelve considered competitions, respectively, the extended
AIC model generates a positive average profit of 0.03 per bet euro. According to this
evaluation method the extended AIC model outperforms the other proposed models.
Also, the bookmakers probabilities are beaten by the extended AIC model, as the
bookmakers probabilities obtain an average profit share of even -0.039 per bet euro.
In fact, the average profit share of all four proposed models is higher than that of the
bookmakers probability.

Table 5.16: The table gives the portion of profit per bet euro for the seasons 2013-14
through 2015-16 in case a bettor bets on the highest assigned outcome possibility of
three models and the bookmakers predictions. The model that provides the highest
profit share using this strategy is underlined for each considered season.

Basic ECI Simplified Extended AIC Extended BIC Bookmakers

model model model model
England
2013-14 0.063 0.066 0.020 0.036 0.089
2014-15 -0.031 -0.011 0.043 -0.018 -0.035
2015-16 -0.105 -0.109 -0.081 -0.122 -0.111
Netherlands
2013-14 -0.200 -0.213 -0.073 -0.193 -0.175
2014-15 0.018 0.018 0.014 -0.006 -0.035
2015-16 0.004 0.023 0.108 0.012 -0.002
Spain
2013-14 -0.055 -0.062 -0.063 -0.064 -0.032
2014-15 0.008 0.001 0.007 -0.021 -0.017
2015-16 -0.054 -0.072 -0.021 -0.075 -0.004
Sweden
2013 -0.014 -0.026 -0.042 -0.030 -0.095
2014 -0.066 -0.150 0.039 -0.048 -0.109
2015 0.112 0.077 0.085 0.075 0.055

Even though the aforementioned betting strategy, where the outcome possibility
with the largest assigned probability is bet on, gives an average seasonal profit share
of 0.03 per bet euro for the extended AIC model, only eight of the twelve considered
seasons generate a positive portion of profit per bet euro. In Sect. 4.4 four more so-

44
 phisticated betting strategies are introduced. Table 5.3 shows the performance of these
strategies applied on the match outcome predictions of the four proposed models. The
table gives the average seasonal portion of profit per bet euro over the twelve considered
seasons for each combination of betting strategy and proposed model accompanied with
the restriction of a minimum assigned probability given by . Once more, one-sided
students t-tests are used to check whether the average profit share per bet euro is
significantly different than zero.

Table 5.17: The table average portion of profit per bet euro of the betting strategies
applied on the outcome of the four models. underline means positive profit in each
season and model. An asterisk (*) indicates that the average profit share over the
twelve considered seasons is significantly different than zero. An underlined average
profit indicates that the profit share of this specific is positive for each of the twelve
considered seasons.

Equal weights betting strategy

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Basic ECI model -0.006 -0.007 -0.008 -0.019 -0.019 0.020 0.030 0.073 0.171*
Simplified model -0.061* -0.058* -0.036* -0.028 -0.044 0.028 0.033 0.137* 0.168*
Extended AIC model 0.008 0.013 0.019 0.028 0.078* 0.114* 0.101* 0.180* 0.233*
Extended BIC model -0.035 -0.034 -0.023 -0.032* -0.039 0.009 0.047 0.090 0.044

Equal payout betting strategy

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Basic ECI model -0.003 -0.004 -0.009 -0.011 -0.009 0.026 0.030 0.073 0.171*
Simplified model -0.041* -0.040* -0.036* -0.028 -0.034 0.034 0.030 0.138* 0.168*
Extended AIC model 0.018 0.020 0.019 0.024 0.066* 0.099* 0.098* 0.166* 0.226*
Extended BIC model -0.011 -0.011 -0.011 -0.015 -0.015 0.024 0.061* 0.095* 0.047

Variance adjusted betting strategy

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Basic ECI model 0.018 0.018 0.015 0.018 0.023 0.046 0.046* 0.085 0.173*
Simplified model -0.014 -0.013 -0.011 -0.002 0.001 0.062* 0.048 0.154* 0.171*
Extended AIC model 0.055* 0.056* 0.055* 0.063* 0.097* 0.127* 0.118* 0.177* 0.226*
Extended BIC model 0.018 0.018 0.017 0.017 0.021 0.043* 0.081* 0.111* 0.045

Kellys betting strategy

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Basic ECI model 0.041 0.042 0.040 0.045 0.041 0.061* 0.061* 0.096* 0.205*
Simplified model 0.004 0.007 0.028 0.043 0.033 0.095* 0.079 0.164* 0.169*
Extended AIC model 0.183* 0.174* 0.163* 0.166* 0.178* 0.201* 0.151* 0.229* 0.229*
Extended BIC model 0.013 0.012 0.021 0.027 0.022 0.035 0.064 0.125* 0.026

45
 Table 5.17 indicates that both the variance adjusted betting strategy and Kellys
betting strategy applied on the outcome probabilities assigned by the extended AIC
model generate a significant positive profit share per bet euro independent of the tresh-
old value . Only six combinations of proposed model, proposed betting strategy and
minimum value of generate a significant positive profit share per bet euro among all
of the twelve considered seasons. These six betting situations, underlined in Table 5.17,
all use the probabilities assigned by the extended AIC model. Four of these six bet-
ting situations are generated by a combinations of the extended AIC model with either
the variance adjusted or Kellys betting strategy and a treshold value . Choosing
between the variance adjusted betting strategy and Kellys strategy accompanied with
the extended AIC model based on the results of Table 5.17 indicates clear preference
for Kellys betting strategy with the extended AIC model. The average portion of
profit per bet euro over the considered seasons is consistently higher for Kellys betting
strategy than for the variance adjusted betting strategy. However, an average obtained
profit share over four competitions is considered in this table. In Figure 5.1 the profit
share per bet euro for each competition averaged over the three considered seasons is
therefore compared for both the variance adjusted betting strategy as Kellys betting
strategy applied on the extended AIC model for two values of . The values of that
are considered are 0.4 and 0.7. The value 0.7 is chosen as Table 5.17 shows that four of
the six betting situations that generate a positive profit share per bet euro among all
twelve considered seasons contain a treshold value of equal to 0.7. As a second value
= 0.4 is chosen as the other two of the six situations obtain a value of equal to 0.4
and 0.5.

46
 Sweden
0.274 = 0.7 0.331

0.161 = 0.4 0.245

Spain
0.248 = 0.7 0.293

0.100 = 0.4 0.119

Netherlands
0.098 = 0.7 0.059

0.096 = 0.4 0.108

England
0.089 = 0.7 0.234

0.031 = 0.4 0.241

0.3 0.2 0.1 0 profit 0 0.1 0.2 0.3

share
variance adjusted betting strategy Kellys betting strategy

Figure 5.1: This figure shows a comparison of the average profit share per bet euro
using the variance adjusted betting strategy and the Kellys betting strategy
accompanied with a minimum assigned probability of equal to 0.4 and 0.7. This
average profit share over three seasons is compared for the four considered
competitions.

Figure 5.1 indicates a preference for the combination of Kellys betting strategy
applied on the extended AIC model over the variance adjusted betting strategy applied
on the probabilities assigned by the extended AIC model. In case the assigned proba-
bilities are obtained from the extended AIC model, only the Spanish competition with
a -value equal to 0.4 results in a larger profit share for the variance adjusted betting
strategy compared to Kellys betting strategy. The difference between the obtained
portion of profit per bet euro performing the variance adjusted betting strategy and
Kellys strategy on the match outcome probabilities retrieved from the extended AIC
model is not that large.

47
 Solely the obtained profit for the English competition is much larger in case Kellys
betting strategy is performed compared to the variance adjusted betting strategy. Over-
all, one can conclude from the figure that Kellys betting strategy provides a higher por-
tion of profit per bet euro than the variance adjusted betting strategy does for values
of equal to 0.4 and 0.7.

Table 5.18: This table presents the portion of profit per bet euro obtained by
performing the Kellys betting strategy with several values of for each of the twelve
considered seasons. The obtained profit share per bet euro for the optimal value of
is bold.

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

England
2013-14 0.103 0.105 0.112 0.121 0.192 0.250 0.260 0.258 0.452
2014-15 0.798 0.732 0.573 0.493 0.420 0.569 0.392 0.336 0.386
2015-16 0.450 0.398 0.273 0.244 0.112 0.109 0.014 0.109 0.039
Netherlands
2013-14 0.013 0.010 0.002 -0.017 0.019 0.038 0.073 0.07 0.002
2014-15 0.071 0.071 0.092 0.103 0.119 0.115 0.057 0.026 0.016
2015-16 0.139 0.145 0.158 0.202 0.186 0.136 0.115 0.146 0.151
Spain
2013-14 0.280 0.279 0.268 0.246 0.143 0.157 0.230 0.437 0.449
2014-15 0.090 0.089 0.099 0.108 0.080 0.103 0.119 0.148 0.101
2015-16 0.080 0.080 0.091 0.077 0.134 0.209 0.222 0.294 0.285
Sweden
2013 -0.060 -0.060 -0.046 -0.021 0.064 0.152 0.067 0.141 -0.223
2014 0.054 0.056 0.119 0.148 0.135 0.002 0.022 0.196 -0.037
2015 0.178 0.178 0.209 0.286 0.535 0.572 0.471 0.655 1.125

Table 5.18 shows the seasonal profit share per bet euro of Kellys betting strategy
applied on the outcome probabilities of the extended AIC model for the four consid-
ered competition with different values of . From these results one can conclude that
performing Kellys betting strategy on the assigned probabilities of the extended AIC
models with optimal values of lead to an positive average seasonal share of profit per
bet euro. This average seasonal portion of profit per bet euro for England, Netherlands,
Spain and Sweden equals 0.449, 0.131, 0.297 and 0.491, respectively. One can see from
the table that the optimal value of varies largely per season and competition. Al-
though the optimal treshold -value of the preliminary season is most of the time not
optimal for the successive season, it provides a significant positive portion of profit per
bet euro. Figure 5.2 shows the obtained share of profit per bet euro for the seasons
2014(-15) and 2015(-16) for the four considered competitions. The figure shows the
profit share obtained for the optimal value of and the obtained profit share in case
the value of is estimated out-of-sample. The optimal value of from the preliminary
seasons is used in this case. One can see that for only two of the eight considered sea-

48
sons the out-of-sample profit share equals the profit share obtained using the optimal
. Table 5.18 shows that for these two cases the out-of-sample value of is the optimal
value of . The average seasonal profit share of this method for England, Netherlands,
Spain an Sweden equals 0.393, 0.122, 0.198 and 0.329, respectively, over the seasons
2014-(15) and 2015(-16). An overall average seasonal out-of-sample profit share per
bet euro of 0.260 over the eight considered seasons for four different competitions is
generated with this strategy.

0.798 0.450 0.119 0.202 0.148 0.294 0.196 1.125

0.336 0.450 0.057 0.186 0.101 0.294 0.002 0.655

1.2 Optimal 1.2

Out-of-sample
1.0 1.0

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
2014-15 2015-16 2014-15 2015-16 2014-15 2015-16 2014 2015
England Netherlands Spain Sweden

Figure 5.2: This figure visualizes the difference in obtained portion of profit per bet
euro applying the variance adjusted betting strategy with the optimal value of and
the out-of-sample value of . This is performed for the seasons 2014(-15) and
2015(-16) of the English, Dutch, Spanish and Swedish competition.

Overall, one can conclude that applying Kellys betting strategy on the match out-
come probabilities retrieved from the extended AIC model systematically provides a
positive share of profit per bet euro. This positive profit is obtained for both optimal
values as out-of-sample values of . Table 5.18 and Fig. 5.2 show that the obtained
profit share per bet euro varies largely per season and competition, both for optimal
values of and out-of-sample values of .

49
5.5 | Predictability of the seasons
Previous sections elaborate upon the results of several prediction models and betting
strategies performed on the matches of three seasons from four European competitions.
Kellys betting strategy applied on the outcome predictions obtained from the extended
AIC models turned out to perform best and delivers the highest profit share per bet
euro. As this method is applied on the twelve different seasons, the usability of the
method, expressed in obtained profit share per bet euro, can be studied. Among these
twelve considered seasons, there are seasons in which a lot of outcome matches were
unforeseen. For other seasons the majority of matches ended in the expected outcome.
A link between obtained profit and the predictability of a specific season is studied.
Classification of seasons based on their predictability is not straightforward. Three
different measures are applied to indicate the predictability of a certain season.
At first, a simulation algorithm is applied in which a competition season is simulated
10,000 times based on the ECI value of the clubs at the start of the season. Fig. 5.3
shows the simulated probability for the real-life league winner to become the champion
according to Alg. 1 in the seasons 2013-(14) through 2015(-16) for each of the four
competitions. The dotted line presents the average amount of simulated probability
assigned to the league winner for the seasons 2007(-08) through 2015(-16). The figure
suggests that the champion of the Spanish league is better predictable than the Swedish
league winner, based on simulations, for the seasons 2007(-08) through 2015(-16). The
simulation algorithm suggests that season 2015 of the Swedish competition, season
2013-14 of the Spanish competition and season 2015-16 of the English competition were
relatively hard to predict. On the other hand, season 2014 of the Swedish competition
and season 2013-14 of the Dutch competition were relatively easy to predict according
to this measure. Table 5.18 and Fig. 5.2 find a higher portion of profit per bet euro for
the seasons that contain high prediction correctness than for the seasons that contain
small prediction correctness. This suggests that hard predictable seasons lead to higher
profit share. However, this link is not present for all considered seasons. Furthermore,
no clear link between season predictability and the optimal value of can be found.

50
 Sweden Netherlands
0.5 0.392 1.0
0.576
0.188 0.383
0.149 0.274
0.012 0.119
0.0 0.0
2013 2014 2015 2013-14 2014-15 2015-16

1.0 0.5

0.530 0.178 0.216 0.296

0.435 0.363
0.013 0.001
0.0 0.0
2013-14 2014-15 2015-16 2013-14 2014-15 2015-16

Spain England

Figure 5.3: This figure presents the simulated probability of the league winner for
the seasons 2013(-14) through 2015(-16) for the four considered seasons. The dotted
line presents the average amount of simulated probability assigned to the champion of
the seasons 2007(-08) through 2015(-16).

Taking this simulation algorithm as an indicator for the predictability of a certain

season contains a drawback. The simulation algorithm uses the ECI value of the clubs
at the start of the season as the only input. In case the real-life league winner is a club
that performed disappointing in the foregoing season, the simulation will predict bad.
An unexpected league winner does not necessarily imply that the total gradient of the
competition has to be unpredictable.
Another measure that is used as an indicator measure for season predictability is the
cosine similarity between the vectors of the resulting season tables of two consecutive
seasons. If the resulting season table varies a lot with the resulting season table of the
preliminary season this measure finds low similarity, whereas a resulting season table
that is relatively similar to the preliminary season will get high similarity. Table 5.19
shows for each of the twelve considered seasons the similarity of the resulting season
table compared to the resulting season table of the preliminary season using the cosine
similarity measure. Besides, the average cosine similarity for the seasons 2000(-01)
through 2015(-16) is given for each of the four competitions. The table suggests that
the resulting season table for season 2013-14 of the Spanish competition and season
2015-16 of the Dutch competition had an relatively unexpected outcome, whereas the
resulting season table for season 2014-15 of the English, seasons 2013-14 and 2014-15

51
of the Dutch competition, season 2014-15 of the Spanish competition and season 2013
of the Swedish competition were relatively easy to predict. No clear correlation can be
found between the similarity of the resulting season tables with that of the preliminary
season and the obtained profit share.

Table 5.19: This table gives the similarity of the considered season with the
preliminary season using cosine similarity on the vectors of the resulting season tables.
The last column contains the average value of the cosine similarity measure applied on
the seasons 2000(-01) through 2015(-16) for each of the four considered competitions.

Model 2013(-14) 2014(-15) 2015(-16) 2000(-01) - 2015(-16)

England 0.921 0.973 0.932 0.928
Netherlands 0.970 0.980 0.931 0.946
Spain 0.895 0.978 0.938 0.912
Sweden 0.936 0.915 0.902 0.901

A drawback of this method is that it only compares the season tables at the end of
the season. Besides, this method solely compares the similarity of the current season
with the preliminary season. A low similarity between these two season vectors only
suggests an unexpected resulting season table compared to the resulting season table
of the preliminary season.
As a third indicator measure the obtained profit from the betting strategies, given
in Table 5.16, is studied. This table shows the profit of betting unit stack on the
outcome possibility with the highest assigned probability for each of the four proposed
models and the bookmakers probabilities. Table 5.20 presents the average profit share
per bet euro of these four proposed models and the bookmakers probabilities for each
of the twelve considered seasons. This betting strategy bets per match on the outcome
possibility with the highest assigned probability according to the model at hand. This
indication measure evaluates the proportion of matches that have an expected outcome
according to the models. The measure suggests that season 2015-16 of the English
competition and season 2013-14 of the Dutch competition were the seasons with the
most unpredictable match outcomes. To a lesser extent season 2013-14 and 2015-16 from
the Spanish competition and season 2013 and 2014 from the Swedish competition can
be classified as unpredictable seasons. On the other hand, season 2013-14 of the English
competition and season 2015 of the Swedish competition are seasons containing a lot
of matches with an expected outcome. Also for this measure of seasons predictability
Table 5.18 and Fig. 5.2 find no clear link between the average profit of betting on the
outcome with the highest assigned probability, given in Table 5.20, and the obtained
profit share.

52
 Table 5.20: This table gives the average profit share per bet euro of the four
proposed models and the bookmakers odds betting unit stack on the outcome
possibility with the highest assigned probability.

Model England Netherlands Spain Sweden

2013(-14) 0.055 -0.171 -0.055 -0.041
2014(-15) -0.010 0.002 -0.004 -0.067
2015(-16) -0.106 0.029 -0.045 0.081

A drawback of this third indication measure is that it depends on the proposed

models. The results of this method rely largely on the assigned bookmakers odds and
the accuracy of the four proposed models.
Although each of the three measures consists of mentioned shortcomings, all these
measures focus on another aspect of season predictability. The results of the three
measures present that these indicator measures are not consistent in classifying seasons
as predictable or unpredictable. For only one of the measures a relation between the
predictability and the obtained profit share is discovered. The seasons for which the
simulation algorithm assigned low probability to the real-life league winner generated
a larger profit share per bet euro than seasons where the simulation algorithm assigned
high probability to the real-life league winner. However, this link is not applicable for
all twelve considered seasons. This makes us conclude that Kellys betting strategy
applied on outcome probabilities assigned by the extended AIC models can be seen
as a robust betting strategy. Both seasons in which the majority of matches ends as
expected and seasons in which a lot of unexpected match outcomes occur generate a
positive share of profit per bet euro. The size of the profit share and the value of the
optimal value of are neither correlated with the predictability of the season indicated
by one of the three considered indication measures.

53
6 | Conclusion

One of the main goals of this paper is to come up with prediction models that accurately
predict the outcome of soccer matches. In addition, it is attempted to construct a
betting strategy that is able to defeat the bookmakers and generate profit from betting
on soccer matches.
For this research, the focus is on European national competitions. PCA accom-
panied with a K-means clustering algorithm is applied on a dataset of competition
characteristics to select a number of European leagues to study in more detail. This
method led to selection of the Dutch, English, Spanish and Swedish competition. For
each of these four selected competitions various ordered probit forecasting models are
constructed to predict a probability of occurrence for all three outcome possibilities.
Two of the four prediction models solely depend on the strength of the competing clubs,
expressed in ECI value, and the division of home wins, draws and away wins in recent
seasons. Two other prediction models incorporate a widely range of club-specific ex-
planatory variables alongside the ECI value for clubs. The extended AIC model uses
the Akaike information criterion to decide on the inclusion of variables. The extended
BIC model uses Bayesian information criterion. The performance of the proposed mod-
els varies widely over six evaluation measures. Overall, the extended AIC model is the
preferred model for most competitions and seasons, whereas the extended BIC model is
clearly the least preferred. According to some of the evaluation measures the prediction
accuracy of the proposed models outperforms the bookmakers probabilities, whereas the
converted bookmakers odds predict more accurately than each of the proposed models
according to other evaluation methods.
Consequently, several betting strategies are implemented for which the predicted
probability of the possible soccer match outcomes assigned by the prediction models
is the input. A minimum assigned probability restriction is imposed to four existing
betting strategies. Applying Kellys betting strategy on the predicted outcome prob-
abilities obtained from the extended AIC model provides the overall best performing
betting strategy. This betting strategies generates an average out-of-sample profit share
per bet euro of 0.26 over the four considered competitions in 2014(-15) and 2015(-16).
It generates a positive profit share per bet euro for all considered seasons.
Eventually, it is investigated whether the obtained profit is related to the predictabil-
ity of a certain season. A season simulation algorithm, resulting season table vector
comparison of consecutive seasons and the obtained profit from betting on the match
outcomes with the highest assigned probability are three measures to indicate a sea-
sons predictability. Overall, the indicated predictability and the obtained profit shares
obey a clear link among the considered seasons. This indicates that the proposed strat-

54
egy is a robust betting strategy that systematically generates profit independent of the
predictability of the season.
We can conclude that the extended AIC model is the most accurate model in pre-
dicting soccer match outcomes. It is best able to compete with the bookmakers. The
preference for this model shows that the included soccer-related variables contain sig-
nificant information to predict match outcomes. Applying the betting strategy of Kelly
on the results of this model provides an average profit share of 0.26 per bet euro.
Future research on this content should investigate if inclusion of player-specific vari-
ables into the prediction models is worth it. Injuries and suspensions might have in-
fluence on the outcome probabilities and inclusion of these factors might improve the
constructed models. For this research the used betting odds are the average match
odds assigned by several bookmakers. As the match odds vary per bookmaker one
could attempt to construct a betting algorithm that benefits from these discrepancies
in odds by different bookmakers. As this paper finds a profitable betting strategy using
average odds an even higher retrieved profit could be found in case the most favorable
assigned odds per match can be chosen from several bookmakers.

55
Bibliography

[1] Abdi, H., Williams, L.J.: Principal component analysis. Wiley Interdisciplinary
Reviews: Computational Statistics 2(4), 433459 (2010)

[2] Blundell, J.: Numerical Algorithms for Predicting Sports Results. Ph.D. thesis,
University of Leeds, School of Computer Studies (2009)

[3] Bozdogan, H.: Model selection and akaikes information criterion (aic): The gen-
eral theory and its analytical extensions. Psychometrika 52(3), 345370 (1987)

[4] Burnham, K.P., Anderson, D.R.: Multimodel inference understanding aic and bic
in model selection. Sociological methods & research 33(2), 261304 (2004)

[5] Clarke, S.R., Norman, J.M.: Home ground advantage of individual clubs in english
soccer. The Statistician pp. 509521 (1995)

[6] Dixon, M.J., Coles, S.G.: Modelling association football scores and inefficiencies
in the football betting market. Journal of the Royal Statistical Society: Series C
(Applied Statistics) 46(2), 265280 (1997)

[7] Dobson, S., Goddard, J.A., Dobson, S.: The economics of football. Cambridge
University Press Cambridge (2001)

[8] Elo, A.: The rating of chess players, past and present (arco, new york) (1978)

[9] Epstein, E.S.: A scoring system for probability forecasts of ranked categories.
Journal of Applied Meteorology 8(6), 985987 (1969)

[10] Finnigan, M., Nordsted, P.: The premier football betting handbook 2010/11.
Hampshire, Great Britain: Harriman House (2010)

[11] Frey, J.H.: Gambling: A sociological review. The Annals of the American Academy
of Political and Social Science 474(1), 107121 (1984)

[12] Goddard, J., Asimakopoulos, I.: Modelling football match results and the effi-
ciency of fixed-odds betting. Tech. rep., Working Paper, Department of Economics,
Swansea University (2003)

56
[13] Goldfeld, S.M., Quandt, R.E., Trotter, H.F.: Maximization by quadratic hill-
climbing. Econometrica: Journal of the Econometric Society pp. 541551 (1966)

[14] Graham, I., Stott, H.: Predicting bookmaker odds and efficiency for uk football.
Applied Economics 40(1), 99109 (2008)

[15] Greene, W.H.: Econometric analysis (international edition) (2000)

[16] Hamadani, B.: Predicting the outcome of nfl games using machine learning. URL
http://cs229. stanford. edu/proj2006/BabakHamadani-PredictingNFLGames. pdf
(2005)

[17] Heij, C., De Boer, P., Franses, P.H., Kloek, T., Van Dijk, H.K., et al.: Econometric
methods with applications in business and economics. OUP Oxford (2004)

[18] Hing, N.: Sports betting and advertising. Australian Gambling Research Centre,
Australian Institute of Family Studies (2014)

[19] Jaccard, P.: The distribution of the flora in the alpine zone. New phytologist 11(2),
3750 (1912)

[20] Kelly, J.: A new interpretation of information rate. IRE Transactions on Informa-
tion Theory 2(3), 185189 (1956)

[21] King, B., Reynolds, J., Botta, C., Wardle, S.a.: Dynamic modelling and prediction
of english football league matches for betting. Street and Smiths SportsBusiness
Journal (www.sportsbusinessjournal.com) 17(ISSUE B) (2014, June)

[22] Koning, R.H.: Balance in competition in dutch soccer. Journal of the Royal Sta-
tistical Society: Series D (The Statistician) 49(3), 419431 (2000)

[23] Langseth, H.: Beating the bookie: A look at statistical models for prediction of
football matches. In: SCAI. pp. 165174 (2013)

[24] Liang, Y., Balcan, M.F., Kanchanapally, V.: Distributed pca and k-means clus-
tering. In: The Big Learning Workshop at NIPS. vol. 2013. Citeseer (2013)

[25] Lloyd, S.P.: Least squares quantization in pcm. Information Theory, IEEE Trans-
actions on 28(2), 129137 (1982)

[26] Maher, M.J.: Modelling association football scores. Statistica Neerlandica 36(3),
109118 (1982)

[27] Markazi, A.: Why leicester city become biggest long shot champion sports his-
tory. In: http://espn.go.com/espn/feature/story/_/id/14759409/why-leicester-
city-become-biggest-long-shot-champion-sports-history (2016)

57
[28] Nevill, A.M., Balmer, N.J., Williams, A.M.: The influence of crowd noise and
experience upon refereeing decisions in football. Psychology of Sport and Exercise
3(4), 261272 (2002)

[29] Norris, D.: Footballs beauty is its unpredictability. In:

%http://www.irishexaminer.com/sport/soccer/footballs-beauty-is-its-
unpredictability-says-arsene-wenger-321805.html (2015)

[30] Peason, K.: On lines and planes of closest fit to systems of point in space. Philo-
sophical Magazine 2, 559572 (1901)

[31] Pohlman, J.T., Leitner, D.W.: A comparison of ordinary least squares and logistic
regression (2003)

[32] Pollard, R., Pollard, G.: Home advantage in soccer: A review of its existence and
causes. International Journal of Soccer and Science Journal 3(1), 3144 (2005)

[33] Prabhu, P., Anbazhagan, N.: Improving the performance of k-means clustering for
high dimensional data set. International journal on computer science and engineer-
ing 3(6), 23172322 (2011)

[34] Rue, H., Salvesen, O.: Prediction and retrospective analysis of soccer matches in a
league. Journal of the Royal Statistical Society: Series D (The Statistician) 49(3),
399418 (2000)

[35] Sauer, R.D.: The economics of wagering markets. Journal of economic Literature
36(4), 20212064 (1998)

[36] Smith, R.W., Preston, F.W.: Vocabularies of motives for gambling behavior. So-
ciological Perspectives 27(3), 325348 (1984)

[37] Snyder, J.A.L.: What actually wins soccer matches: Prediction of the 2011-2012
premier league for fun and profit (2013)

[38] Spann, M., Skiera, B.: Sports forecasting: a comparison of the forecast accuracy of
prediction markets, betting odds and tipsters. Journal of Forecasting 28(1), 5572
(2009)

[39] trumbelj, E., ikonja, M.R.: Online bookmakers odds as forecasts: The case
of european soccer leagues. International Journal of Forecasting 26(3), 482488
(2010)

[40] Vergin, R.C., Sosik, J.J.: No place like home: an examination of the home field
advantage in gambling strategies in nfl football. Journal of Economics and Business
51(1), 2131 (1999)

58
[41] Wilkinson, L., Engelman, L., Corter, J., Mark, C.: Systat 10 statistics i. 4. Cluster
Analysis pp. I77 (2000)

[42] Wold, S., Esbensen, K., Geladi, P.: Principal component analysis. Chemometrics
and intelligent laboratory systems 2(1-3), 3752 (1987)

[43] Yezus, A.: Predicting outcome of soccer matches using machine learning. Ph.D.
thesis, Saint-Petersburg State University, Mathematics and Mechanics Faculty
(2014)

59
Appendices

60
A | Output PCA

Table A.1: The table gives the values of the principal component for each
competition.

Country PC1 PC2 PC3 PC4 PC5

Netherlands -0.616 -2.211 -0.852 -0.885 1.451
Belgium -1.665 1.698 0.606 0.519 0.874
England 3.993 0.617 1.220 0.309 -0.272
Italy 2.885 0.076 -0.452 0.966 -0.854
France 1.475 -0.064 1.316 0.890 0.951
Scotland -3.108 1.042 -1.765 1.971 0.655
Spain 4.263 0.608 1.139 -0.866 1.120
Germany 5.444 -1.202 -0.804 -0.256 -0.204
Portugal 1.254 -0.463 -1.574 -0.348 0.0817
Denmark -3.202 -0.458 -0.059 -0.114 -0.457
Norway -2.736 -2.163 1.012 -0.380 -1.432
Ukraine -0.266 -0.375 -2.647 -0.164 -0.350
Austria -3.400 1.560 0.535 -0.880 1.032
Russia 0.491 1.569 0.601 0.785 -1.637
Sweden -3.120 -2.313 1.887 0.590 0.008
Switzerland -1.692 2.081 -0.164 -2.147 -0.966

61
 2

Switzerland

Russia Austria
Belgium
PC2 - 12.9% of total variation

1
N
Scotland
England
Germany J
H
EA F
Italy D B
0
France
G
Ukraine
O
L Denmark
C K
Portugal M I

Spain
-1

Norway
Netherlands
Sweden

-2
-2 -1 0 1 2
PC1 - 54.5% of total variation

A| Average stadium capacity I| Different champions last 10 seasons

B| Average number of spectators J| Average age of players
C| Competition size K| Average amount of goals per match
D| UEFA competition points rank L| Number of relegated teams
E| Number of CL tickets M| Perc. of possible points champion
F| Number of teams in Europe at 1 Dec. N| points first and last club
G| Number of teams in Europe at 1 Apr. O| ECI value first and last club
H| Average ECI value

Figure A.1: This figure shows the biplot of the first two principal components using
scaled data. The 16 observations are visualized by dots. The variables are given by
vectors. In the legend is specified which variable belongs to each of three vectors. The
remaining 13 variables are not visualized as the length of these vectors was
insignificant compared to the three given in the figure.

62
B | Output cluster method

Table B.1: The table gives the coefficients of the cluster means per competition
characteristic.

Variable Cluster 1 Cluster 2

Average stadium capacity 36250.14 18015.06
Average number of spectators 24516.02 9618.38
League size 18.86 13.89
UEFA competition points rank 70658.71 28132.78
CL tickets 3.29 1.44
European active teams 1 dec 4.86 1.11
European active teams 1 apr 2 0.11
Average value of ECI 2504.20 1786.03
Different winners last 10 seasons 3.57 4.11
Average age of players 27.68 26.66
Average goals per match 2.62 2.69
Number of relegation per year 2.5 1.5
% of points from the winner 0.78 0.76
Point difference first and last 0.31 0.28
Difference ECI value first and last 1855.80 1266.03

63
C | Extended AIC models

Table C.1: The table gives the extended ordered probit model based on AIC for
England in the season 2013-14.

Variables Coefficient Std. Error

5M atchesShapeHteam 3i 0.192 (0.090)
5M atchesShapeHteam + 3i 0.293 (0.091)
ArsenalEN GRef eree1i -1.594 (0.602)
AverageHomeAgainstHteami -0.200 (0.075)
AwaymatchesAteami 0.102 (0.037)
AwaymatchesHteami -0.084 (0.036)
CapacityHteami 8.09E-6 (2.56E-6)
CapacityAteami -5.50E-6 (2.46E-6)
ChelseaAprili 0.691 (0.271)
ECIHomei - ECIAwayi 0.50E-3 (5.20E-5)
EU AwayDrawP ostW interi 0.680 (0.337)
EU AwayW inP ostW interAteami -0.783 (0.379)
EvertonShape 1i 0.591 (0.207)
EvertonV S.Arsenali -0.967 (0.387)
HomeM anCityi 0.432 (0.136)
LiverpoolEN GRef eree2i 1.098 (0.515)
M anCityAugusti 0.752 (0.327)
M anCityEN GRef eree4i -1.006 (0.363)
M anCityV S.N ewcastlei 1.592 (0.602)
M anU nitedShape 3i 0.626 (0.295)
P ercHomepointsAteami 0.430 (0.203)
P ercP ointsAteami -0.595 (0.247)
SouthamptonShape1i 1.377 (0.491)
Shape + 1V S.Shape + 1i 0.449 (0.169)
Shape + 3V S.Shape + 3i 0.346 (0.157)
StokeEN GRef eree6i 0.768 (0.285)
SunderlandV S.W BAi -1.262 (0.483)
T ottenhamF ebruaryi 0.618 (0.295)
V illaM archi -0.774 (0.248)

1 -0.833 (0.148)
2 0.004 (0.147)

64
Table C.2: The table gives the extended ordered probit model based on AIC for
England in the season 2014-15.

Variables Coefficient Std. Error

3M atchesShapeHteam 2i -0.169 (0.081)
5M atchesShapeHteam 3i 0.209 (0.083)
5M atchesShapeHteam + 3i 0.210 (0.088)
ArsenalEN GRef eree1i -0.932 (0.444)
AwaymatchesHteami -0.111 (0.049)
ChelseaN ovemberi -0.526 (0.226)
CupmatchHteami 0.172 (0.082)
ECIHomei - ECIAwayi 0.64E-3 (4.16E-5)
EU AwayDrawP ostW interi 0.979 (0.317)
EU P ostW interHteami 0.169 (0.086)
EU In6W inP ostW interHteami 0.664 (0.275)
EvertonShape 1i 0.562 (0.186)
HomeV illai -0.252 (0.104)
HomeM anCityi 0.494 (0.125)
HomeStokei 0.307 (0.113)
LiverpoolM archi 0.493 (0.247)
LiverpoolEN GRef eree2i 1.123 (0.535)
LiverpoolEN GRef eree3i -1.458 (0.741)
M anCityEN GRef eree4i -0.936 (0.363)
M anCityV S.N ewcastlei 1.541 (0.587)
M anU nitedShape 3i 0.519 (0.242)
M atchesLef tAteami -0.050 (0.025)
N ewcastleEN GRef eree5i 1.134 (0.539)
N ewcastleV S.Swanseai -1.482 (0.646)
P ercP ointsAteami -0.844 (0.257)
RankAteami -0.019 (0.007)
Shape + 1V S.Shape + 1i 0.460 (0.153)
Shape + 2V S.Shape + 2i 0.343 (0.152)
SouthamptonSaturdayi 0.478 (0.170)
SunderlandJanuaryi 0.635 (0.247)
SwanseaEN GRef eree7i 1.114 (0.463)
SwanseaEN GRef eree8i -2.053 (0.814)
SwanseaShape 1i 0.693 (0.303)
T ottenhamEN GRef eree9i -0.637 (0.081)
T ottenhamV S.Swanseai 1.289 (0.647)
W BAEN GRef eree9i -0.847 (0.386)
W BASeptemberi 0.666 (0.291)
W estHamEN GRef eree5i 1.398 (0.573)

1 -2.827 (0.874)
2 -2.014 (0.874)

65
Table C.3: The table gives the extended ordered probit model based on AIC for
England in the season 2015-16.

Variables Coefficient Std. Error

3M atchesShapeHteam 2i -0.228 (0.075)
5M atchesShapeAteam 3i -0.177 (0.076)
5M atchesShapeHteam 3i 0.190 (0.078)
5M atchesShapeHteam + 3i 0.240 (0.080)
AverageHomeAgainstHteami -0.187 (0.069)
AverageGoalsAwayHteami 0.272 (0.094)
ArsenalV S.Swanseai -0.849 (0.421)
AwaymatchesHteami -0.099 (0.046)
CapacityAteami -4.78E-6 (2.09E-6)
ChelseaAprili 0.583 (0.239)
ECIHomei - ECIAwayi 0.58E-3 (4.38E-5)
EvertonShape 1i 0.570 (0.182)
EU AwayDrawP ostW interi 0.829 (0.317)
EU HomeIn6W inP reW interi -0.240 (0.115)
HomeM anCityi 0.468 (0.119)
HomeStokei 0.281 (0.108)
HomepointsHteami -0.012 (0.006)
LiverpoolM archi 0.491 (0.229)
M anCityEN GRef eree4i -1.000 (0.362)
M anCityV S.N ewcastlei 1.641 (0.571)
M anU nitedShape 3i 0.461 (0.218)
M atchesLef tAteami -0.062 (0.024)
N ewcastleV S.Swanseai -1.220 (0.500)
P ercAwaypointsHteami -0.595 (0.250)
Shape + 1V S.Shape + 1i 0.365 (0.147)
SouthamptonJanuaryi 0.777 (0.342)
SouthamptonSaturdayi 0.395 (0.144)
Spectatorsi 5.92E-6 (2.15E-6)
StokeEN GRef eree6i 0.554 ( 0.254)
SwanseaAugusti 0.801 (0.403)
SwanseaEN GRef eree8i -1.754 (0.723)
SwanseaShape 1i 0.649 (0.249)
T ottenhamEN GRef eree9i -0.631 (-0.301)
T ottenhamV S.Swanseai 1.508 (0.623)
W BAEN GRef eree9i -0.929 (0.379)
W BASeptemberi 0.804 (0.268)

1 -2.813 (0.834)
2 -2.013 (0.833)

66
Table C.4: The table gives the extended ordered probit model based on AIC for
Netherlands in the season 2013-14.

Variables Coefficient Std. Error

ECIOnArtif icialT urf Ateami -0.010 (0.003)
AjaxM archi 0.832 (0.336)
AjaxV S.U trechti -1.074 (0.342)
AverageGoalsAgainstAteami 0.177 (0.065)
AwaymatchesAteami -0.019 (0.009)
AwaypointsHteami 0.018 (0.006)
AZN ovemberi 0.690 (0.271)
AZShape 2i -0.645 (0.242)
ECIHomei - ECIAwayi 0.73E-3 (7.41E-5)
EU LossAwayP ostW interi -1.057 (0.318)
F eyenoordJanuaryi -0.807 (0.309)
F eyenoordShape1i 0.457 (0.227)
GroningenN ET Ref eree2i 0.829 (0.354)
GroningenShape 2i -0.558 (0.210)
N ACN ET Ref eree5i -1.226 (0.618)
P SV Bef ore2P Mi -0.752 (0.299)
RKCRef eree6i -1.459 (0.733)
Shape 3V S.Shape + 1i 0.649 (0.217)
Shape + 1V S.Shape + 1i 0.541 (0.183)
Shape + 1V S.Shape + 3i -0.568 (0.199)
Spectatorsi 8.05E-6 (3.26E-6)
T wenteF ebruaryi -0.542 (0.245)
U trechtBetween6P M And8P Mi -0.449 (0.222)
V itesseN ET Ref eree2i -1.276 (0.342)

1 -0.212 (0.134)
2 0.523 (0.134)

67
Table C.5: The table gives the extended ordered probit model based on AIC for
Netherlands in the season 2014-15.

Variables Coefficient Std. Error

ECIOnArtif icialT urf Ateami -0.011 (0.002)
AjaxV S.U trechti -0.855 (0.313)
AverageGoalsAgainstAteami 0.151 (0.060)
AwaymatchesAteami -0.020 (0.008)
AwaypointsHteami 0.017 (0.006)
AZN ovemberi 0.595 (0.246)
CapacityAteami -7.88E-6 (2.66E-6)
ECIHomei - ECIAwayi 0.53E-3 (8.09E-5)
EU AwayLossP ostW interHteami -0.833 (0.347)
F eyenoordJanuaryi -0.775 (0.278)
GroningenN ET Ref eree2i 0.821 n (0.349)
GroningenShape 2i -0.499 (0.194)
P SV Bef ore2P Mi -0.808 (0.288)
Shape 3V S.Shape + 1i 0.435 (0.186)
Shape + 1V S.Shape + 1i 0.460 (0.170)
Shape + 1V S.Shape + 3i -0.494 (0.178)
Spectatorsi 1.30E-5 (3.22E-6)
U trechtBetween6P M And8P Mi -0.461 (0.199)
U trechtN ET Ref eree3i 0.723 (0.353)
V itesseM ayi -0.977 (0.415)
V itesseN ET Ref eree2i -1.017 (0.282)
W illemIIN ET Ref eree7i -0.882 0.385)

1 -0.416 (0.141)
2 0.328 (0.141)

68
Table C.6: The table gives the extended ordered probit model based on AIC for
Netherlands in the season 2015-16.

Variables Coefficient Std. Error

ECIOnArtif icialT urf Ateami -0.013 (0.002)
AjaxM archi 0.590 (0.262)
AjaxV S.U trechti -0.728 (0.293)
AverageGoalsAgainstAteami 0.128 (0.057)
AZN ovemberi 0.575 (0.227)
ECIHomei - ECIAwayi 0.70E-3 (6.15E-5)
EU AwayLossP ostW interHteami -0.791 (0.337)
F eyenoordJanuaryi -0.514 (0.257)
F eyenoordN ET Ref eree4i 0.959 (0.456)
Graaf schapShape1i -0.509 (0.241)
GroningenN ET Ref eree1i 0.870 (0.398)
GroningenN ET Ref eree2i -0.587 (0.263)
GroningenShape 2i -0.419 (0.174)
P SV Bef ore2P Mi -0.669 (0.263)
Shape 3V S.Shape + 1i 0.460 (0.174)
Shape 2V S.Shape + 1i -0.331 (0.140)
Shape 1V S.Shape + 2i -0.365 (0.167)
Shape + 1V S.Shape + 1i 0.429 (0.159)
Shape + 1V S.Shape + 3i -0.475 (0.165)
Shape + 2V S.Shape 3i 0.357 (0.175)
Spectatorsi 1.05E-5 (2.69E-6)
T wenteF ebruaryi -0.553 (0.200)
U trechtBetween6P M And8P Mi -0.399 (0.190)
U trechtN ET Ref eree3i 0.814 (0.320)
V itesseN ET Ref eree2i -0.732 (0.243)
W illemIIN ET Ref eree7i -0.903 (0.372)

1 -0.255 (0.117)
2 0.486 (0.117)

69
Table C.7: The table gives the extended ordered probit model based on AIC for
Spain in the season 2013-14.

Variables Coefficient Std. Error

BilbaoSP ARef eree1i -1.018 (0.431)
CapacityAteami -6.50E-6 (1.74E-6)
CapacityHteami 7.57E-6 (1.86E-6)
ECIHomei - ECIAwayi 0.45E-3 (6.30E-5)
EspanyolShape 3i -0.611 (0.199)
EU AwayW inAteami -0.519 (0.197)
EU In6P reW interi 0.260 (0.090)
EU LossP reW interAteami -0.870 (0.282)
Getaf eSP ARef eree5i 0.877 (0.419)
Getaf eSP ARef eree6i -0.885 (0.325)
HomeBetisi -0.360 (0.141)
LevanteSP ARef eree7i 0.916 (0.412)
RankAteami 0.012 (0.006)
RealM adridW eeki 0.808 (0.366)
SevillaSP ARef eree4i 1.085 (0.383)
SevillaSP ARef eree11i 1.222 (0.586)
Shape + 1V S.Shape + 1i 0.469 (0.164)
Shape + 2V S.Shape 1i 0.431 (0.174)
SociedadSP ARef eree10i 1.580 (0.721)
V alenciaM archi -0.651 (0.227)

1 -0.486 (0.128)
2 0.215 (0.127)

70
Table C.8: The table gives the extended ordered probit model based on AIC for
Spain in the season 2014-15.

Variables Coefficient Std. Error

ECI3M atchesAteami 0.001 (0.001)
ECI5M atchesAteami -0.001 (0.001)
AlmeraSP ARef eree12i 0.830 (0.412)
AtleticoSP ARef eree2i -0.633 (0.282)
BilbaoSP ARef eree1i -0.934 (0.384)
BilbaoV S.Espanyoli -0.826 (0.340)
CapacityAteami -5.74E-6 (1.66E-6)
CeltaSP ARef eree3i 1.322 (0.623)
DeportivoSP ARef eree4i -0.945 (0.455)
DeportivoN ovemberi 0.781 (0.286)
DeportivoV S.Atleticoi -1.090 (0.486)
DeportivoV S.Gij oni -1.109 (0.484)
ECIHomei - ECIAwayi 0.51E-3 (5.44E-5)
EspanyolShape 3i -0.604 (0.187)
EU In6P reW interi 0.230 (0.084)
EU LossP reW interAteami -0.957 (0.267)
EU P ostW interAteami -0.237 (0.069)
EU AwayW inAteami -0.589 (0.177)
Getaf eJanuaryi -0.567 (0.225)
Getaf eSP ARef eree5i 1.081 (0.434)
Getaf eSP ARef eree6i -0.869 (0.305)
HomeBetisi -0.413 (0.125)
HomepointsHteami -0.010 (0.003)
LevanteSP ARef eree7i 0.900 (0.410)
RealM adridSP ARef eree9i 1.180 (0.598)
SevillaAf ter10P Mi -0.322 (-0.412)
SevillaSP ARef eree4i 1.058 (0.383)
SevillaSP ARef eree11i 1.245 (0.583)
SevillaV S.Getaf ei -0.755 (0.347)
Shape + 1V S.Shape + 1i 0.466 (0.155)
Shape + 2V S.Shape 1i 0.418 (0.161)
SociedadV S.Barcelonai 1.016 (0.443)
Spectatorsi 8.10E-6 (1.93E-6)
V alenciaM archi -0.723 (0.206)
V alenciaV S.Sociedadi -1.260 (0.480)

1 -0.685 (0.069)
2 0.025 (0.068)

71
Table C.9: The table gives the extended ordered probit model based on AIC for
Spain in the season 2015-16.

Variables Coefficient Std. Error

AtleticoSP ARef eree2i -0.548 (0.267)
BarcelonaAf ter10P Mi 0.509 (0.190)
BarcelonaSundayi 0.321 (0.147)
BetisV S.Rayoi -1.501 (0.633)
BilbaoSP ARef eree1i -0.763 (0.340)
BilbaoN ovemberi 0.771 (0.240)
BilbaoV S.Deportivoi -0.840 (0.366)
BilbaoV S.Espanyoli -0.681 (0.326)
CapacityHteami 5.34E-6 (1.59E-6)
CeltaSP ARef eree3i 1.042 (0.497)
CeltaShape1i 0.630 (0.234)
DeportivoSP ARef eree4i -0.916 (0.451)
DeportivoN ovemberi 0.573 (0.267)
DeportivoV S.Atleticoi -1.092 (0.464)
ECIHomei - ECIAwayi 0.57E-3 (4.66E-5)
EspanyolShape 3i -0.539 (0.177)
EspanyolSP ARef eree8i 0.616 (0.302)
EU AwayW inAteami -0.398 (0.167)
EU LossP reW interAteami -0.849 (0.243)
EU P ostW interAteami -0.272 (0.066)
Getaf eJanuaryi -0.487 (0.213)
Getaf eSP ARef eree5i 1.004 (0.376)
Getaf eSP ARef eree6i -0.911 (0.308)
HomeBetisi -0.370 (0.128)
HomepointsHteami -0.010 (0.003)
RayoV S.Getaf ei 1.213 (0.541)
RealM adridAprili 0.569 (0.254)
RealM adridOctoberi 0.609 (0.297)
RealM adridSP ARef eree9i 1.202 (0.571)
SevillaSP ARef eree3i 1.110 (0.549)
SevillaSP ARef eree4i 1.054 (0.386)
SevillaSP ARef eree11i 1.193 (0.585)
Shape + 1V S.Shape + 1i 0.421 (0.143)
Shape + 2V S.Shape 1i 0.505 (0.154)
SociedadV S.Barcelonai 1.111 (0.414)
V alenciaM archi -0.481 (0.196)
V alenciaV S.Granadai 1.175 (0.592)
V alenciaV S.Sociedadi -0.777 (0.388)

1 -0.488 (0.067)
2 0.235 (0.067)

72
Table C.10: The table gives the extended ordered probit model based on AIC for
Sweden in the season 2013.

Variables Coefficient Std. Error

3M atchesShapeAteam + 2i -0.259 (0.117)
5M atchesShapeAteam 2i 0.271 (0.128)
5M atchesShapeAteam 5i 0.378 (0.117)
AIKShape2i 0.732 (0.252)
AIKSW ERef eree1i 0.822 (0.337)
AverageGoalsHomeAteami -0.258 (0.075)
BrommaAugusti -1.135 (0.401)
BrommaSW ERef eree5i 1.140 (0.410)
Cupf ighterAteami -0.223 (0.073)
DjurgardensSW ERef eree2i -0.640 (0.299)
ECIHomei - ECIAwayi 0.69E-3 (8.25E-5)
HomeHalmstadsi -0.343 (0.158)
HomepointsHteami 0.041 (0.011)
KalmarSW ERef eree7i -0.648 (0.261)
M atchesLef tHteami 0.028 (0.007)
M jalbySW ERef eree1i 1.046 (0.381)
M jallbyM archi 1.822 (0.778)
N orrkopingSW ERef eree5i 1.096 (0.437)
P ercAwaypointsHteami 0.706 (0.295)

1 0.042 (0.229)
2 0.780 (0.230)

73
Table C.11: The table gives the extended ordered probit model based on AIC for
Sweden in the season 2014.

Variables Coefficient Std. Error

ECI5M atchesHteami -0.003 (0.001)
3M atchesShapeAteam 3i 0.243 (0.110)
3M atchesShapeAteam + 2i -0.216 (0.109)
3M atchesShapeHteam + 4i 0.244 (0.113)
AIKSW ERef eree1i 0.690 (0.334)
AIKShape2i 0.824 (0.232)
BrommaSW ERef eree5i 0.811 (0.387)
BrommaSW ERef eree6i -1.048 (0.475)
Cupf ighterAteami -0.212 (0.063)
DjurgardensSW ERef eree2i -0.667 (0.303)
DjurgardensShape 3i -1.048 (0.475)
ECIHomei - ECIAwayi 0.65E-3 (2.88E-5)
Elf sborgM ayi 0.545 (0.251)
HomeHalmstadsi -0.336 (0.142)
HomepointsHteami 0.031 (0.010)
HalmstadsSW ERef eree4i 1.119 (0.497)
M atchesLef tHteami 0.026 (0.006)
M jalbySW ERef eree1i 1.057 (0.382)
M jallbyM archi 1.625 (0.815)
M jalbySW ERef eree7i 1.101 (0.510)
OrebroV S.Helsingborgsi 0.821 (0.379)
P ercAwaypointsHteami 0.715 (0.274)
P ercHomepointsAteami -0.600 (0.204)
P os.ECILast5M atchesHteami 0.304 (0.102)

1 0.087 (0.215)
2 0.828 (0.215)

74
Table C.12: The table gives the extended ordered probit model based on AIC for
Sweden in the season 2015.

Variables Coefficient Std. Error

3M atchesShapeAteam 3i 0.210 (0.102)
3M atchesShapeAteam + 2i -0.293 (0.100)
3M atchesShapeHteam + 4i 0.292 (0.100)
AIKM archi -1.089 (0.510)
AIKShape2i 0.703 (0.212)
Cupf ighterAteami -0.131 (0.056)
DjurgardensSW ERef eree2i -0.785 (0.299)
DjurgardensV S.Hammarbyi -1.407 (0.664)
ECIHomei - ECIAwayi 0.82E-3 (6.38E-5)
Gef leSW ERef eree1i -0.622 (0.308)
HalmstadsShape3i -0.606 (0.293)
HalmstadsSW ERef eree3i -0.590 (0.282)
HalmstadsSW ERef eree4i 1.072 (0.510)
HammarbyShape1i -1.594 (0.649)
HammarbySundayi -0.944 (0.344)
KalmarSW ERef eree7i -0.589 (0.235)
M atchesLef tHteami 0.026 (0.006)
N eg.ECILast5M atchesAteami 0.306 (0.102)
OrebroV S.Helsingborgsi 0.713 (0.333)
P os.ECILast5M atchesAteami 0.220 (0.101)

1 0.240 (0.217)
2 0.975 (0.217)

75
D | Extended BIC models

Table D.1: The table gives the extended ordered probit model based on BIC for
England in the season 2013-14.

Variables Coefficient Std. Error

5M atchesShapeHteam + 3i 0.247 (0.088)
ECIHomei - ECIAwayi 0.65E-3 (3.26E-5)
HomeM anCityi 0.408 (0.127)
M anCityV S.N ewcastlei 1.454 (0.597)
SouthamptonShape1i 1.427 (0.495)

1 -0.661 (0.314)
2 0.145 (0.029)

Table D.2: The table gives the extended ordered probit model based on BIC for
England in the season 2014-15.

Variables Coefficient Std. Error

ECIHomei - ECIAwayi 0.64E-3 (3.02E-5)
EU AwayDrawP ostW interi 0.898 (0.308)
HomeM anCityi 0.437 (0.120)
M anCityV S.N ewcastlei 1.508 (0.583)
N ewcastleV S.Swanseai -1.838 (0.649)
SwanseaEN GRef eree8i -2.129 (0.819)

1 -0.661 (0.028)
2 0.119 (0.026)

Table D.3: The table gives the extended ordered probit model based on BIC for
England in the season 2015-16.

Variables Coefficient Std. Error

3M atchesShapeHteam 2i -0.220 (0.073)
ECIHomei - ECIAwayi 0.62E-3 (2.86E-5)
EU AwayDrawP ostW interi 0.896 (0.308)
HomeM anCityi 0.417 (0.113)
M anCityV S.N ewcastlei 1.566 (0.570)
M anU nitedShape 3i 0.597 (0.214)

1 -0.669 (0.027)
2 0.102 (0.025)

76
Table D.4: The table gives the extended ordered probit model based on BIC for
Netherlands in the season 2013-14.

Variables Coefficient Std. Error

ECIOnArtif icialT urf Ateami -0.011 (0.002)
AjaxV S.U trechti -0.928 (0.335)
ECIHomei - ECIAwayi 0.84E-3 (6.07E-5)
EU LossAwayP ostW interi -0.946 (0.309)
Shape + 1V S.Shape + 3i -0.584 (0.197)
Spectatorsi 8.34E-6 (2.86E-6)
V itesseN ET Ref eree2i -1.094 (0.335)

1 -0.495 (0.065)
2 0.215 (0.064)

Table D.5: The table gives the extended ordered probit model based on BIC for
Netherlands in the season 2014-15.

Variables Coefficient Std. Error

ECIOnArtif icialT urf Ateami -0.012 (0.002)
ECIHomei - ECIAwayi 0.78E-3 (5.64E-5)
P SV Bef ore2P Mi -0.805 (0.285)
Shape + 1V S.Shape + 3i -0.494 (0.176)
Spectatorsi 7.82E-6 (2.64E-6)
V itesseN ET Ref eree2i -0.977 (0.273)

1 -0.501 (0.060)
2 0.220 (0.059)

Table D.6: The table gives the extended ordered probit model based on BIC for
Netherlands in the season 2015-16.

Variables Coefficient Std. Error

ECIOnArtif icialT urf Ateami -0.013 (0.002)
ECIHomei - ECIAwayi 0.76E-3 (5.25E-5)
Shape 3V S.Shape + 1i 0.492 (0.173)
Spectatorsi 7.51E-6 (2.46E-6)
V itesseN ET Ref eree2i -0.725 (0.244)

1 -0.474 (0.056)
2 0.242 (0.055)

77
Table D.7: The table gives the extended ordered probit model based on BIC for
Spain in the season 2013-14.

Variables Coefficient Std. Error

CapacityAteami -6.61E-6 (1.62E-6)
CapacityHteami 5.49E-6 (1.68E-6)
ECIHomei - ECIAwayi 0.43E-3 (1.68E-6)
EspanyolShape 3i -0.580 (0.197)
Getaf eSP ARef eree6i -0.895 (0.324)
SevillaSP ARef eree4i 1.025 (0.384)
Shape + 1V S.Shape + 1i 0.451 (0.161)

1 -0.710 (0.073)
2 -0.025 (0.072)

Table D.8: The table gives the extended ordered probit model based on BIC for
Spain in the season 2014-15.

Variables Coefficient Std. Error

HomeBetisi -0.376 (0.123)
CapacityAteami -6.55E-6 (1.60E-6)
ECIHomei - ECIAwayi 0.44E-3 (5.09E05)
Spectatorsi 7.69E-6 (1.82E-6)
Getaf eSP ARef eree6i -0.858 (0.302)
EspanyolShape 3i -0.528 (0.184)
V alenciaM archi -0.628 (0.203)

1 -0.722 (0.061)
2 -0.040 (0.060)

Table D.9: The table gives the extended ordered probit model based on BIC for
Spain in the season 2015-16.

Variables Coefficient Std. Error

CapacityHteami 5.60E-6 (1.47E-6)
ECIHomei - ECIAwayi 0.59E-3 (3.44E-5)
EU P ostW interAteami -0.199 (0.061)
Getaf eSP ARef eree6i -0.954 (0.298)
HomeBetisi -0.373 (0.125)
HomepointsHteami -0.009 (0.003)
Shape + 1V S.Shape + 1i 0.431 (0.141)
Shape + 2V S.Shape 1i 0.466 (0.151)

1 -0.472 (0.062)
2 0.217 (0.062)

78
Table D.10: The table gives the extended ordered probit model based on BIC for
Sweden in the season 2013.

Variables Coefficient Std. Error

5M atchesShapeAteam 5i 0.370 (0.115)
AIKShape2i 0.757 (0.253)
AIKSW ERef eree1i 0.869 (0.337)
AverageGoalsHomeAteami -0.283 (0.075)
AwaymatchesHteami 0.025 (0.007)
BrommaAugusti -1.110 (0.400)
BrommaSW ERef eree5i 1.110 (0.410)
Cupf ighterAteami -0.201 (0.072)
ECIHomei - ECIAwayi 0.61E-3 (8.05E-5)
HomepointsHteami 0.037 (0.011)
M jalbySW ERef eree1i 1.067 (0.379)
P ercAwaypointsHteami 0.899 (0.291)

1 0.023 (0.226)
2 0.747 (0.227)

Table D.11: The table gives the extended ordered probit model based on BIC for
Sweden in the season 2014.

Variables Coefficient Std. Error

AIKShape2i 0.800 (0.228)
AwaymatchesHteami 0.018 (0.006)
Cupf ighterAteami -0.188 (0.062)
ECIHomei - ECIAwayi 0.61E-3 (7.46E-5)
HomepointsHteami 0.026 (0.010)
P ercAwaypointsHteami 0.797 (0.262)
P ercHomepointsAteami -0.643 (0.119)

1 -0.197 (0.195)
2 0.516 (0.196)

79
Table D.12: The table gives the extended ordered probit model based on BIC for
Sweden in the season 2015.

Variables Coefficient Std. Error

3M atchesShapeAteam + 2i -0.319 (0.098)
3M atchesShapeHteam + 4i 0.283 (0.098)
AIKShape2i 0.697 (0.207)
ECIHomei - ECIAwayi 0.72E-3 (5.80E-5)

1 -0.605 (0.035)
2 0.107 (0.033)

80
E | Evaluation methods

Table E.1: The table gives the results from the Jaccard index evaluation method for
each of the four constructed models. The profit of the best predicting proposed model
per competition seasons is underlined and bold in this table. An asterisk (*) is added
in case the best predicting proposed models performs better than the bookmakers
probabilities.

Model Competition 2013(-14) 2014(-15) 2015(-16) Average

Basic ECI model England 0.413 0.355 0.301 0.356
Netherlands 0.299 0.385* 0.378 0.354
Spain 0.357 0.400* 0.357 0.371*
Sweden 0.345* 0.330 0.392 0.356*
Simplified model England 0.415 0.362 0.301 0.359
Netherlands 0.297 0.385* 0.385 0.356
Spain 0.355 0.397 0.350 0.367
Sweden 0.341 0.304 0.374 0.340
Extended AIC England 0.384 0.367 0.290 0.347
model Netherlands 0.325* 0.375 0.400* 0.367*
Spain 0.355 0.397 0.348 0.367
Sweden 0.319 0.345* 0.361 0.342
Extended BIC England 0.397 0.360 0.297 0.351
model Netherlands 0.302 0.372 0.381 0.352
Spain 0.355 0.387 0.348 0.363
Sweden 0.326 0.333 0.374 0.344

Bookmakers England 0.431 0.360 0.306 0.366

probabilities Netherlands 0.319 0.375 0.385 0.360
Spain 0.377 0.397 0.382 0.385
Sweden 0.326 0.330 0.396 0.351

81
Table E.2: The table gives the results from the ordered probit penalty index for each
of the four constructed models. The profit of the best predicting proposed model per
competition seasons is underlined and bold in this table. An asterisk (*) is added in
case the best predicting proposed models performs better than the bookmakers
probabilities.

Model Competition 2013(-14) 2014(-15) 2015(-16) Average

Basic ECI model England -0.329 -0.363 -0.440 -0.377
Netherlands -0.387 -0.342 -0.329 -0.353
Spain -0.346 -0.280 -0.319 -0.315
Sweden -0.395 -0.414 -0.389 -0.399
Simplified model England -0.342 -0.369 -0.442 -0.384
Netherlands -0.399 -0.353 -0.337 -0.363
Spain -0.345 -0.281 -0.327 -0.318
Sweden -0.441 -0.451 -0.430 -0.441
Extended AIC England -0.327 -0.323* -0.422* -0.357*
model Netherlands -0.365* -0.339* -0.319* -0.341*
Spain -0.317* -0.273 -0.325 -0.305
Sweden -0.462 -0.421 -0.394 -0.426
Extended BIC England -0.350 -0.365 -0.442 -0.386
model Netherlands -0.377 -0.367 -0.347 -0.364
Spain -0.348 -0.288 -0.319 -0.318
Sweden -0.462 -0.460 -0.425 -0.449

Bookmakers England -0.326 -0.375 -0.433 -0.378

probabilities Netherlands -0.383 -0.364 -0.356 -0.368
Spain -0.326 -0.268 -0.310 -0.301
Sweden -0.424 -0.396 -0.364 -0.395

82
Table E.3: The table gives the results from the rank probability score for each of the
four constructed models. The profit of the best predicting proposed model per
competition seasons is underlined and bold in this table. An asterisk (*) is added in
case the best predicting proposed models performs better than the bookmakers
probabilities.

Model Competition 2013(-14) 2014(-15) 2015(-16) Average

Basic ECI model England 0.395 0.403 0.432 0.410
Netherlands 0.415 0.401 0.381 0.399
Spain 0.402 0.367 0.389 0.386
Sweden 0.413 0.418 0.396 0.409
Simplified model England 0.396 0.403 0.431 0.410
Netherlands 0.413 0.400 0.381 0.398
Spain 0.402 0.366 0.390 0.386
Sweden 0.412* 0.417 0.402 0.410
Extended AIC England 0.412 0.401 0.449 0.421
model Netherlands 0.417 0.407 0.378* 0.401
Spain 0.398 0.376 0.398 0.391
Sweden 0.447 0.415 0.405 0.422
Extended BIC England 0.421 0.406 0.435 0.421
model Netherlands 0.407* 0.407 0.385 0.400
Spain 0.406 0.371 0.393 0.390
Sweden 0.435 0.422 0.407 0.421

Bookmakers England 0.381 0.394 0.419 0.398

probabilities Netherlands 0.408 0.398 0.385 0.397
Spain 0.389 0.356 0.370 0.372
Sweden 0.413 0.398 0.381 0.397

83
 Table E.4: The table gives the results from the mean squared error. The profit of
the best predicting proposed model per competition seasons is underlined and bold in
this table. An asterisk (*) is added in case the best predicting proposed models
performs better than the bookmakers probabilities.

Model Competition 2013(-14) 2014(-15) 2015(-16) Average

Basic ECI model England 0.186 0.195 0.211 0.197
Netherlands 0.201 0.194 0.188 0.194
Spain 0.191 0.181 0.189 0.187
Sweden 0.203 0.201 0.194 0.199
Simplified model England 0.187 0.195 0.210 0.197
Netherlands 0.205 0.194 0.188 0.196
Spain 0.191 0.181 0.190 0.187
Sweden 0.202 0.201 0.195 0.199
Extended AIC England 0.192 0.194 0.217 0.201
model Netherlands 0.207 0.196 0.186* 0.196
Spain 0.189 0.185 0.193 0.189
Sweden 0.213 0.201 0.197 0.204
Extended BIC England 0.195 0.196 0.211 0.201
model Netherlands 0.203 0.196 0.189 0.196
Spain 0.192 0.183 0.191 0.189
Sweden 0.209 0.203 0.197 0.203

Bookmakers England 0.181 0.192 0.207 0.193

probabilities Netherlands 0.201 0.191 0.188 0.193
Spain 0.186 0.176 0.183 0.182
Sweden 0.202 0.194 0.188 0.195

84
F | Variable explanation

Table F.1: The table provides an explanation for variables included in the extended
prediction models, the extended AIC models and the extended BIC models.

Variable Definition
ECI3M atchesAteami The difference in ECI value over the past three matches
for the away team
ECI5M atchesAteami The difference in ECI value over the past five matches
for the away team
ECI5M atchesHteami The difference in ECI value over the past five matches
for the home team
ECIOnArtif icialT urf Ateami The increment in ECI value in away matches on artificial
turf in current and last season away team
3M atchesShapeAteam + 2i Dummy variable that is 1 in case the away team is in 3
matches shape groups +2 and 0 in all other cases 2
3M atchesShapeAteam 3i Dummy variable that is 1 in case the away team is in 3
match shape groups -3 and 0 in all other cases 2
3M atchesShapeHteam + 4i Dummy variable that is 1 in case the home team is in 3
match shape groups +4 and 0 in all other cases 2
3M atchesShapeHteam 2i Dummy variable that is 1 in case the home team is in 3
match shape groups -2 and 0 in all other cases 2
5M atchesShapeAteam 2i Dummy variable that is 1 in case the away team is in 5
match shape groups -2 and 0 in all other cases 2
5M atchesShapeAteam 3i Dummy variable that is 1 in case the away team is in 5
match shape groups -3 and 0 in all other cases 2
5M atchesShapeAteam 5i Dummy variable that is 1 in case the away team is in 5
match shape groups -5 and 0 in all other cases 2
5M atchesShapeHteam + 3i Dummy variable that is 1 in case the home team is in 5
match shape groups +3 and 0 in all other cases 2
5M atchesShapeHteam 3i Dummy variable that is 1 in case the home team is in 5
match shape groups -3 and 0 in all other cases 2
AIKM archi Variable that is 1 in case AIK plays a home match in
March, -1 in case AIK plays an away match in March
and 0 in all other cases
AIKShape2i Variable that is 1 in case AIK plays a home match in
shape 2, -1 in case AIK plays an away match in shape 2
and 0 in all other cases
AIKSW ERef eree1i Variable that is 1 in case AIK plays a home match with
SWE referee 1, -1 in case AIK plays an away match with
SWE referee 1 and 0 in all other cases
AjaxM archi Variable that is 1 in case Ajax plays a home match in
March, -1 in case Ajax plays an away match in March
and 0 in all other cases

85
Variable Definition
AjaxV S.U trechti Variable that is 1 in case Ajax plays home against FC
Utrecht, -1 in case FC Utrecht plays home against Ajax
and 0 in all other cases
AlmeraSP ARef eree12i Variable that is 1 in case UD Almera plays a home match
with SPA referee 12, -1 in case UD Almera plays an away
match with SPA referee 12 and 0 in all other cases
ArsenalEN GRef eree1i Variable that is 1 in case Arsenal plays a home match
with ENG referee 1, -1 in case Arsenal plays an away
match with ENG referee 1 and 0 in all other cases
ArsenalV S.Swanseai Variable that is 1 in case Arsenal plays home against
Swansea City, -1 in case Swansea City plays home against
Arsenal and 0 in all other cases
AtleticoSP ARef eree2i Variable that is 1 in case Atltico Madrid plays a home
match with SPA referee 2, -1 in case Atltico Madrid
plays an away match with SPA referee 2 and 0 in all
other cases
AverageGoalsAgainstAteami The average amount of goals against by the away team
AverageGoalsAwayHteami The average amount of away goals by the home team
AverageHomeAgainstHteami The average amount of home goals against by the home
team
AverageGoalsHomeAteami The average amount of home goals by the away team
AwaymatchesAteami The amount of away matches played in the current season
by the away team
AwaypointsHteami The current amount of away points of the home team
AZN ovemberi Variable that is 1 in case AZ Alkmaar plays a home match
in November -1 in case AZ Alkmaar plays an away match
in November and 0 in all other cases
AZShape 2i Variable that is 1 in case AZ Alkmaar plays a home match
in shape -2, -1 in case AZ Alkmaar plays an away match
in shape -2 and 0 in all other cases
BarcelonaAf ter10P Mi Variable that is 1 in case FC Barcelona plays a home
match after 10 PM, -1 in case FC Barcelona plays an
away match after 10 PM and 0 in all other cases
BarcelonaSundayi Variable that is 1 in case FC Barcelona plays a home
match on Sunday, -1 in case FC Barcelona plays an away
match on Sunday and 0 in all other cases
BetisV S.Rayoi Variable that is 1 in case Real Betis plays home against
Rayo Vallecano, -1 in case Rayo Vallecano plays home
against Real Betis and 0 in all other cases
BilbaoN ovemberi Variable that is 1 in case Athletic Club de Bilbao plays
a home match in November -1 in case Athletic Club de
Bilbao plays an away match in November and 0 in all
other cases
BilbaoSP ARef eree1i Variable that is 1 in case Athletic Club de Bilbao plays a
home match with SPA referee 1, -1 in case Athletic Club
de Bilbao plays an away match with SPA referee 1 and 0
in all other cases

86
Variable Definition
BilbaoV S.Deportivoi Variable that is 1 in case Athletic Club de Bilbao plays
home against Deportivo La Corua, -1 in case Deportivo
La Corua plays home against Athletic Club de Bilbao
and 0 in all other cases
BilbaoV S.Espanyoli Variable that is 1 in case Athletic Club de Bilbao plays
home against RCD Espanyol, -1 in case RCD Espanyol
plays home against Athletic Club de Bilbao and 0 in all
other cases
BrommaAugusti Variable that is 1 in case IF Brommapojkarna plays a
home match in August, -1 in case IF Brommapojkarna
plays an away match in August and 0 in all other cases
BrommaSW ERef eree5i Variable that is 1 in case IF Brommapojkarna plays
a home match with SWE referee 5, -1 in case IF
Brommapojkarna plays an away match with SWE ref-
eree 5 and 0 in all other cases
BrommaSW ERef eree6i Variable that is 1 in case IF Brommapojkarna plays
a home match with SWE referee 6, -1 in case IF
Brommapojkarna plays an away match with SWE ref-
eree 6 and 0 in all other cases
CapacityAteami The stadium capacity of the away team
CapacityHteami The stadium capacity of the home team
CeltaShape1i Variable that is 1 in case Celta de Vigo plays a home
match in shape 1, -1 in case Celta de Vigo plays an away
match in shape 1 and 0 in all other cases
CeltaSP ARef eree3i Variable that is 1 in case Celta de Vigo plays a home
match with SPA referee 3, -1 in case Celta de Vigo plays
an away match with SPA referee 3 and 0 in all other cases
ChelseaAprili Variable that is 1 in case Chelsea plays a home match in
April, -1 in case Chelsea plays an away match in April
and 0 in all other cases
ChelseaN ovemberi Variable that is 1 in case Chelsea plays a home match
in November, -1 in case Chelsea plays an away match in
November and 0 in all other cases
Cupf ighterAteami Dummy variable that is 1 in case the away team is still
active in the national cup and 0 in all other cases
CupmatchHteami Dummy variable that is 1 in case the home team played
an national cup match within six days before the current
match and 0 in all other cases
DeportivoN ovemberi Variable that is 1 in case Deportivo La Corua plays
a home match in November, -1 in case Deportivo La
Corua plays an away match in November and 0 in all
other cases
DeportivoSP ARef eree4i Variable that is 1 in case Deportivo La Corua plays a
home match with SPA referee 4, -1 in case Deportivo La
Corua plays an away match with SPA referee 4 and 0
in all other cases

87
Variable Definition
DeportivoV S.Atleticoi Variable that is 1 in case Deportivo La Corua plays
home against Atltico Madrid, -1 in case Atltico Madrid
plays home against Deportivo La Corua and 0 in all
other cases
DeportivoV S.Gij oni Variable that is 1 in case Deportivo La Corua plays
home against Sporting Gijn, -1 in case Sporting Gijn
plays home against Deportivo La Corua and 0 in all
other cases
DjurgardensShape 3i Variable that is 1 in case Djurgrdens IF plays a home
match in shape -3, -1 in case Djurgrdens IF plays an
away match in shape -3 and 0 in all other cases
DjurgardensSW ERef eree2i Variable that is 1 in case Djurgrdens IF plays a home
match with SWE referee 2, -1 in case Djurgrdens IF
plays an away match with SWE referee 2 and 0 in all
other cases
DjurgardensV S.Hammarbyi Variable that is 1 in case Djurgrdens IF plays home
against Hammarby IF, -1 in case Hammarby IF plays
home against Djurgrdens IF and 0 in all other cases
ECIHome ECIAwayi Gives the difference in ECI value between the home team
and away team
Elf sborgM ayi Variable that is 1 in case IF Elfsborg plays a home match
in May, -1 in case IF Elfsborg plays an away match in
May and 0 in all other cases
EspanyolShape 3i Variable that is 1 in case RCD Espanyol plays a home
match in shape -3, -1 in case RCD Espanyol plays an
away match in shape -3 and 0 in all other cases
EspanyolSP ARef eree8i Variable that is 1 in case RCD Espanyol plays a home
match with SPA referee 8, -1 in case RCD Espanyol plays
an away match with SPA referee 8 and 0 in all other cases
EU AwayDrawP ostW interi Variable that is 1 in case the last match of the home team
was a European match after Newyears that ended in a
draw, -1 in case the last match of the away team was
a European match after Newyears that ended in a draw
and 0 in all other cases
EU AwayLossP ostW interHteami Dummy variable that is 1 in case the last match of the
home team was a European away match after Newyears
that ended in a loss and 0 in all other cases
EU AwayW inAteami Dummy variable that is 1 in case the last match of the
away team was a European away match that ended in a
win and 0 in all other cases
EU AwayW inP ostW interAteami Dummy variable that is 1 in case the last match of the
away team was a European away match after Newyears
that ended in a win and 0 in all other cases
EU HomeIn6W inP reW interi Variable that is 1 in case the last match of the home
team was a European home match before Newyears that
ended in a win and was within six days, -1 in case the
last match of the home team was a European home match
before Newyears that ended in a win and was within six
day and 0 in all other cases

88
Variable Definition
EU In6P reW interi Variable that is 1 in case the last match of the home team
was a European match before Newyears and was within
six days, -1 in case the last match of the home team was
a European match before Newyears and was within six
days and 0 in all other cases
EU In6W inP ostW interHteami Dummy variable that is 1 in case the last match of the
home team was a European match after Newyears that
ended in a win and was within 6 days and 0 in all other
cases
EU LossAwayP ostW interi Variable that is 1 in case the last match of the home team
was a European away match after Newyears that ended
in a loss, -1 in case the last match of the away team was
a European away match after Newyears that ended in a
loss and 0 in all other cases
EU LossP reW interAteami Dummy variable that is 1 in case the last match of the
away team was a European match before Newyears that
ended in a loss and 0 in all other cases
EU P ostW interAteami Dummy variable that is 1 in case the last match of the
away team was a European away match after Newyears
and 0 in all other cases
EU P ostW interHteami Dummy variable that is 1 in case the last match of the
home team was a European match after Newyears and 0
in all other cases
EvertonShape 1i Variable that is 1 in case Everton plays a home match
in shape -1, -1 in case Everton plays an away match in
shape -1 and 0 in all other cases
EvertonV S.Arsenali Variable that is 1 in case Everton plays home against
Arsenal, -1 in case Arsenal plays home against Everton
and 0 in all other cases
F eyenoordJanuaryi Variable that is 1 in case Feyenoord plays a home match
in January, -1 in case Feyenoord plays an away match in
January and 0 in all other cases
F eyenoordN ET Ref eree4i Variable that is 1 in case Feyenoord plays a home match
with NET referee 4, -1 in case Feyenoord plays an away
match with NET referee 4 and 0 in all other cases
F eyenoordShape 2i Variable that is 1 in case Feyenoord plays a home match
in shape -2, -1 in case Feyenoord plays an away match in
shape -2 and 0 in all other cases
Gef leSW ERef eree1i Variable that is 1 in case Gefle IF plays a home match
with SWE referee 1, -1 in case Gefle IF plays an away
match with SWE referee 1 and 0 in all other cases
Getaf eJanuaryi Variable that is 1 in case Getafe CF plays a home match
in January, -1 in case Getafe CF plays an away match in
January and 0 in all other cases
Getaf eSP ARef eree5i Variable that is 1 in case Getafe CF plays a home match
with SPA referee 5, -1 in case Getafe CF plays an away
match with SPA referee 5 and 0 in all other cases

89
Variable Definition
Getaf eSP ARef eree6i Variable that is 1 in case Getafe CF plays a home match
with SPA referee 6, -1 in case Getafe CF plays an away
match with SPA referee 6 and 0 in all other cases
Graaf schapShape1i Variable that is 1 in case De graafschap plays a home
match in shape 1, -1 in case De graafschap plays an away
match in shape 1 and 0 in all other cases
GroningenN ET Ref eree1i Variable that is 1 in case FC Groningen plays a home
match with NET referee 1, -1 in case FC Groningen plays
an away match with NET referee 1 and 0 in all other cases
GroningenN ET Ref eree2i Variable that is 1 in case FC Groningen plays a home
match with NET referee 2, -1 in case FC Groningen plays
an away match with NET referee 2 and 0 in all other cases
GroningenShape 2i Variable that is 1 in case FC Groningen plays a home
match in shape -2, -1 in case FC Groningen plays an
away match in shape -2 and 0 in all other cases
HalmstadsShape3i Variable that is 1 in case Halmstads BK plays a home
match in shape 3, -1 in case Halmstads BK plays an
away match in shape 3 and 0 in all other cases
HalmstadsSW ERef eree3i Variable that is 1 in case Halmstads BK plays a home
match with SWE referee 3, -1 in case Halmstads BK
plays an away match with SWE referee 3 and 0 in all
other cases
HalmstadsSW ERef eree4i Variable that is 1 in case Halmstads BK plays a home
match with SWE referee 4, -1 in case Halmstads BK
plays an away match with SWE referee 4 and 0 in all
other cases
HammarbyShape1i Variable that is 1 in case Hammarby IF plays a home
match in shape 1, -1 in case Hammarby IF plays an away
match in shape 1 and 0 in all other cases
HammarbySundayi Variable that is 1 in case Hammarby IF plays a home
match on Sunday, -1 in case Hammarby IF plays an away
match on Sunday and 0 in all other cases
HomeBetisi Dummy variable that is 1 in case Real Betis plays a home
match and 0 in all other cases
HomeHalmstadsi Dummy variable that is 1 in case Halmstads BK plays a
home match and 0 in all other cases
HomeM anCityi Dummy variable that is 1 in case Manchester City plays
a home match and 0 in all other cases
HomeStokei Dummy variable that is 1 in case Stoke City plays a home
match and 0 in all other cases
HomeV illai Dummy variable that is 1 in case Aston Villa plays a
home match and 0 in all other cases
HomepointsHteami The current amount of home points of the home team
KalmarSW ERef eree7i Variable that is 1 in case Kalmar FF plays a home match
with SWE referee 7, -1 in case Kalmar FF plays an away
match with SWE referee 7 and 0 in all other cases

90
Variable Definition
LevanteSP ARef eree7i Variable that is 1 in case Levante UD plays a home match
with SPA referee 7, -1 in case Levante UD plays an away
match with SPA referee 7 and 0 in all other cases
LiverpoolEN GRef eree2i Variable that is 1 in case Liverpool plays a home match
with ENG referee 2, -1 in case Liverpool plays an away
match with ENG referee 2 and 0 in all other cases
LiverpoolEN GRef eree3i Variable that is 1 in case Liverpool plays a home match
with ENG referee 3, -1 in case Liverpool plays an away
match with ENG referee 3 and 0 in all other cases
LiverpoolM archi Variable that is 1 in case Liverpool plays a home match
in March, -1 in case Liverpool plays an away match in
March and 0 in all other cases
M anCityAugusti Variable that is 1 in case Manchester City plays a home
match in August, -1 in case Manchester City plays an
away match in August and 0 in all other cases
M anCityEN GRef eree4i Variable that is 1 in case Manchester City plays a home
match with ENG referee 4, -1 in case Manchester City
plays an away match with ENG referee 4 and 0 in all
other cases
M anCityV S.N ewcastlei Variable that is 1 in case Manchester City plays home
against Newcastle United, -1 in case Newcastle United
plays home against Manchester City and 0 in all other
cases
M anU nitedShape 3i Variable that is 1 in case Manchester United plays a home
match in shape -3, -1 in case Manchester United plays an
away match in shape -3 and 0 in all other cases
M atchesLef tAteami The amount of matches left to play in the current seasons
by the away team
M atchesLef tHteami The amount of matches left to play in the current seasons
by the home team
M jalbySW ERef eree1i Variable that is 1 in case Mjllby AIF plays a home match
with SWE referee 1, -1 in case Mjllby AIF plays an away
match with SWE referee 1 and 0 in all other cases
M jalbySW ERef eree7i Variable that is 1 in case Mjllby AIF plays a home match
with SWE referee 7, -1 in case Mjllby AIF plays an away
match with SWE referee 7 and 0 in all other cases
M jallbyM archi Variable that is 1 in case Mjllby AIF plays a home match
in March, -1 in case Mjllby AIF plays an away match in
March and 0 in all other cases
N ACN ET Ref eree5i Variable that is 1 in case NAC Breda plays a home match
with NET referee 5, -1 in case NAC Breda plays an away
match with NET referee 5 and 0 in all other cases
N eg.ECILast5M atchesAteami Dummy variable that is 1 in case the increment in ECI
value of the away team in negative over the last five
matches

91
Variable Definition
N ewcastleEN GRef eree5i Variable that is 1 in case Newcastle United plays a home
match with ENG referee 5, -1 in case Newcastle United
plays an away match with ENG referee 5 and 0 in all
other cases
N ewcastleV S.Swanseai Variable that is 1 in case Newcastle United plays home
against Swansea City, -1 in case Swansea City plays home
against Newcastle United and 0 in all other cases
N orrkopingSW ERef eree4i Variable that is 1 in case IFK Norrkping plays a home
match with SWE referee 4, -1 in case IFK Norrkping
plays an away match with SWE referee 5 and 0 in all
other cases
OrebroV S.Helsingborgsi Variable that is 1 in case rebro SK plays home against
Helsingborgs IF, -1 in case Helsingborgs IF plays home
against rebro SK and 0 in all other cases
P ercAwaypointsHteami The percentage of possible awaypoints home team
P ercHomepointsAteami The percentage of possible homepoints away team
P ercP ointsAteami The percentage of possible points away team
P os.ECILast5M atchesAteami Dummy variable that is 1 in case the increment in ECI
value of the away team in positive over the last five
matches
P os.ECILast5M atchesHteami Dummy variable that is 1 in case the increment in ECI
value of the home team in positive over the last five
matches
P SV Bef ore2P Mi Variable that is 1 in case PSV plays a home match before
2 PM, -1 in case PSV plays an away match before 2 PM
and 0 in all other cases
RankAteami The current rank of the away team
RayoV S.Getaf ei Variable that is 1 in case Rayo Vallecano plays home
against Getafe CF, -1 in case Getafe CF plays home
against Rayo Vallecano and 0 in all other cases
RealM adridAprili Variable that is 1 in case Real Madrid plays a home match
in April, -1 in case Real Madrid plays an away match in
April and 0 in all other cases
RealM adridOctoberi Variable that is 1 in case Real Madrid plays a home match
in October, -1 in case Real Madrid plays an away match
in October and 0 in all other cases
RealM adridSP ARef eree9i Variable that is 1 in case Real Madrid plays a home match
with SPA referee 9, -1 in case Real Madrid plays an away
match with SPA referee 9 and 0 in all other cases
RealM adridW eeki Variable that is 1 in case Real Madrid plays a home match
outside the weekend, -1 in case Madrid plays an away
match outside the weekend and 0 in all other cases
RKCRef eree6i Variable that is 1 in case RKC Waalwijk plays a home
match with NET referee 6, -1 in case RKC Waalwijk
plays an away match with NET referee 6 and 0 in all
other cases

92
Variable Definition
SevillaAf ter10P Mi Variable that is 1 in case Sevilla FC plays a home match
after 10 PM, -1 in case Sevilla FC plays an away match
after 10 PM and 0 in all other cases
SevillaSP ARef eree11i Variable that is 1 in case Sevilla FC plays a home match
with SPA referee 11, -1 in case Sevilla FC plays an away
match with SPA referee 11 and 0 in all other cases
SevillaSP ARef eree3i Variable that is 1 in case Sevilla FC plays a home match
with SPA referee 3, -1 in case Sevilla FC plays an away
match with SPA referee 3 and 0 in all other cases
SevillaSP ARef eree4i Variable that is 1 in case Sevilla FC plays a home match
with SPA referee 4, -1 in case Sevilla FC plays an away
match with SPA referee 4 and 0 in all other cases
SevillaV S.Getaf ei Variable that is 1 in case Sevilla plays home against
Getafe CF, -1 in case Getafe CF plays home against
Sevilla and 0 in all other cases
Shape + 1V S.Shape + 1i Dummy variable that is 1 in case the home team is in
shape groups +1 and the awayteam is in shape group +1
and 0 in all other cases 3
Shape + 2V S.Shape 1i Dummy variable that is 1 in case the home team is in
shape groups +2 and the awayteam is in shape group -1
and 0 in all other cases 3
Shape + 2V S.Shape 3i Dummy variable that is 1 in case the home team is in
shape groups +2 and the awayteam is in shape group -3
and 0 in all other cases 3
Shape + 2V S.Shape + 2i Dummy variable that is 1 in case the home team is in
shape groups +2 and the awayteam is in shape group +2
and 0 in all other cases 3
Shape + 3V S.Shape + 3i Dummy variable that is 1 in case the home team is in
shape groups +3 and the awayteam is in shape group +3
and 0 in all other cases 3
Shape 2V S.Shape + 1i Dummy variable that is 1 in case the home team is in
shape groups -2 and the awayteam is in shape group +1
and 0 in all other cases 3
Shape 3V S.Shape + 1i Dummy variable that is 1 in case the home team is in
shape groups -3 and the awayteam is in shape group +1
and 0 in all other cases 3
Shape 1V S.Shape + 2i Dummy variable that is 1 in case the home team is in
shape groups -1 and the awayteam is in shape group +2
and 0 in all other cases 3
Shape + 1V S.Shape + 3i Dummy variable that is 1 in case the home team is in
shape groups +1 and the awayteam is in shape group +3
and 0 in all other cases 3
SociedadSP ARef eree10i Variable that is 1 in case Real Sociedad plays a home
match with SPA referee 10, -1 in case Real Sociedad plays
an away match with SPA referee 10 and 0 in all other
cases
SociedadV S.Barcelonai Variable that is 1 in case Real Sociedad plays home
against FC Barcelona, -1 in case FC Barcelona plays
home against Real Sociedad and 0 in all other cases

93
Variable Definition
SouthamptonJanuaryi Variable that is 1 in case Southampton plays a home
match in January, -1 in case Southampton plays an away
match in January and 0 in all other cases
SouthamptonSaturdayi Variable that is 1 in case Southampton plays a home
match on Saturday, -1 in case Southampton plays an
away match on Saturday and 0 in all other cases
SouthamptonShape1i Variable that is 1 in case Southampton plays a home
match in shape 1, -1 in case Southampton plays an away
match in shape 1 and 0 in all other cases
Spectatorsi The number of spectators
StokeEN GRef eree6i Variable that is 1 in case Stoke City plays a home match
with ENG referee 1, -1 in case Stoke City plays an away
match with ENG referee 1 and 0 in all other cases
SunderlandJanuaryi Variable that is 1 in case Sunderland plays a home match
in January, -1 in case Sunderland plays an away match
in January and 0 in all other cases
SunderlandV S.W BAi Variable that is 1 in case Sunderland plays home against
West Bromwich Albion, -1 in case West Bromwich Albion
plays home against Sunderland and 0 in all other cases
SwanseaAugusti Variable that is 1 in case Swansea City plays a home
match in August, -1 in case Swansea City plays an away
match in August and 0 in all other cases
SwanseaEN GRef eree7i Variable that is 1 in case Swansea City plays a home
match with ENG referee 7, -1 in case Swansea City plays
an away match with ENG referee 7 and 0 in all other
cases
SwanseaEN GRef eree8i Variable that is 1 in case Swansea City plays a home
match with ENG referee 8, -1 in case Swansea City plays
an away match with ENG referee 8 and 0 in all other
cases
SwanseaShape 1i Variable that is 1 in case Swansea City plays a home
match in shape -1, -1 in case Swansea City plays an away
match in shape -1 and 0 in all other cases
T ottenhamEN GRef eree9i Variable that is 1 in case Tottenham Hotpur plays a home
match with ENG referee 9, -1 in case Tottenham Hotpur
plays an away match with ENG referee 9 and 0 in all
other cases
T ottenhamF ebruaryi Variable that is 1 in case Tottenham Hotspur plays a
home match in February, -1 in case Tottenham Hotspur
plays an away match in February and 0 in all other cases
T ottenhamV S.Swanseai Variable that is 1 in case Tottenham Hotspur plays home
against Swansea City, -1 in case Swansea City plays home
against Tottenham Hotspur and 0 in all other cases
T wenteF ebruaryi Variable that is 1 in case FC Twente plays a home match
in February, -1 in case FC Twente plays an away match
in February and 0 in all other cases
U trechtBetween6P M And8P Mi Variable that is 1 in case FC Utrecht plays a home match
between 6 PM and 8 PM, -1 in case FC Utrecht plays an
away match between 6 PM and 8 PM and 0 in all other
cases

94
 Variable Definition
U trechtN ET Ref eree3i Variable that is 1 in case FC Utrecht plays a home match
with NET referee 3, -1 in case FC Utrecht plays an away
match with NET referee 3 and 0 in all other cases
V alenciaM archi Variable that is 1 in case Valencia CF plays a home match
in March, -1 in case Valencia CF plays an away match in
March and 0 in all other cases
V alenciaV S.Granadai Variable that is 1 in case Valencia CF plays home against
Granada CF, -1 in case Granada CF plays home against
Valencia CF and 0 in all other cases
V alenciaV S.Sociedadi Variable that is 1 in case Valencia CF plays home against
Real Sociedad, -1 in case Real Sociedad plays home
against Valencia CF and 0 in all other cases
V illaM archi Variable that is 1 in case Aston Villa plays a home match
in March, -1 in case Aston Villa plays an away match in
March and 0 in all other cases
V itesseM ayi Variable that is 1 in case Vitesse plays a home match in
May, -1 in case Vitesse plays an away match in May and
0 in all other cases
V itesseN ET Ref eree2i Variable that is 1 in case Vitesse plays a home match with
NET referee 2, -1 in case Vitesse plays an away match
with NET referee 2 and 0 in all other cases
W BAEN GRef eree9i Variable that is 1 in case West Bromwich Albion plays
a home match with ENG referee 9, -1 in case West
Bromwich Albion plays an away match with ENG ref-
eree 9 and 0 in all other cases
W BASeptemberi Variable that is 1 in case West Bromwich Albion plays
a home match in September, -1 in case West Bromwich
Albion plays an away match in September and 0 in all
other cases
W estHamEN GRef eree5i Variable that is 1 in case West Ham United plays a home
match with ENG referee 5, -1 in case West Ham United
plays an away match with ENG referee 5 and 0 in all
other cases
W illemIIN ET Ref eree7i Variable that is 1 in case FC Utrecht plays a home match
with NET referee 7, -1 in case FC Utrecht plays an away
match with NET referee 7 and 0 in all other cases

2 Ten approximately equally sized groups are formed based on percentiles. Five positive and Five

negative shape groups

3 Six approximately equally sized groups are formed based on percentiles. Three positive and three

negative shape groups

95