See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/338168811

The Application of Machine Learning Techniques for Predicting Results in Team

Sport: A Review

Preprint · December 2019

DOI: 10.13140/RG.2.2.22427.62245

CITATIONS READS
3 5,123

2 authors, including:

Rory Bunker
Nagoya University
14 PUBLICATIONS   224 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Sport Result Prediction View project

All content following this page was uploaded by Rory Bunker on 26 December 2019.

The user has requested enhancement of the downloaded file.

 The Application of Machine Learning Techniques for
Predicting Results in Team Sport: A Review
Rory Bunkera,∗, Teo Susnjakb
a
Nagoya Institute of Technology. Gokisocho, Showa Ward, Nagoya, Aichi, 466-8555,
Japan
b
Massey University. Massey University East Precinct, Dairy Flat Highway (SH17),
0632, New Zealand

Abstract
Over the past two decades, Machine Learning (ML) techniques have been
increasingly utilized for the purpose of predicting outcomes in sport. In this
paper, we provide a review of studies that have used ML for predicting results
in team sport, covering studies from 1996 to 2019. We sought to answer five
key research questions while extensively surveying papers in this field. This
paper offers insights into which ML algorithms have tended to be used in
this field, as well as those that are beginning to emerge with successful out-
comes. Our research highlights defining characteristics of successful studies
and identifies robust strategies for evaluating accuracy results in this applica-
tion domain. Our study considers accuracies that have been achieved across
different sports and explores the notion that outcomes of some team sports
could be inherently more difficult to predict than others. Finally, our study
uncovers common themes of future research directions across all surveyed pa-
pers, looking for gaps and opportunities, while proposing recommendations
for future researchers in this domain.
Keywords: Machine Learning, Sport Result Prediction, Literature Review,
Expert System, Team Sport

∗
Corresponding author
Email addresses: [email protected] (Rory Bunker), [email protected]
(Teo Susnjak)

Preprint submitted to arXiv December 26, 2019

1. Introduction
Sport result prediction is an interesting and challenging problem due to
the inherently unpredictable nature of sport, and the seemingly endless num-
ber of potential factors that can affect results. Indeed, it is this unpredictable
nature that is one of the main reasons that people enjoy sport. Despite its
difficulty, attempting to predict the outcome of sport is something that is
of interest to many different stakeholders, and this interest has increased
with sport data becoming increasingly available online, as well as with the
advent of online sports betting. Sport experts or former players are often
asked their predictions for forthcoming matches, which then appear online
or in newspapers. Other stakeholders who are interested in predicting sport
results include bookmakers and sports betting platforms, as well as gamblers
who bet on match results, or for certain events occurring within a particular
match. Note that the prediction of a specific event occurring within a match
is a more general task, and may include events other than the final result,
such as the occurrence of one particular event in a match, e.g., a goal being
scored by a specific player in a match. The focus of this review is on the
prediction of final results. Sport result prediction is also potentially useful
to players, team management and performance analysts in identifying the
most important factors that help to achieve winning outcomes, upon which
appropriate tactics can be identified.
In the academic context, the problem of sport result prediction has been
considered in the Statistics and Operations Research literature for some time,
however the application of Machine Learning (ML) techniques to the problem
is a more recent area of research. The first study in this domain was published
in 1996, and the topic has attracted increased research interest since then.
The scope of this review is on papers that have made use of at least one
ML technique for the problem of sport result prediction. For the sake of
brevity, we focus this review on prediction of team sports. The prediction
of outcomes in individual or non-team sports has also been studied, e.g.,
in Horse Racing (Davoodi & Khanteymoori, 2010), Swimming (Edelmann-
Nusser et al., 2002), Golf (Wiseman, 2016), Tennis (Somboonphokkaphan
& Phimoltares, 2009) and Javelin (Maszczyk et al., 2014). Fuzzy logic and
rule-based methods, which are not the focus of this review, have also been
applied to sport result prediction, e.g., by Rotshtein et al. (2005), Tsakonas
et al. (2002) and Min et al. (2008). Deep Learning has shown potential very

2
recently, with studies by Chen et al. (2020) and Rudrapal et al. (2020), but
is not the focus of this review.
To our knowledge, there have been two prior review papers that consid-
ered ML for sport result prediction, by Haghighat et al. (2013) and, more
recently, by Keshtkar Langaroudi & Yamaghani (2019). The present review
is justified on the grounds that it draws out more critical insight obtained
from previous studies as covered by Haghighat et al. Haghighat et al. (2013),
and it expands on the recent review by Keshtkar Langaroudi & Yamaghani
(2019) by providing a more comprehensive coverage of surveyed papers.
Schumaker et al. (2010) stated that Artificial Neural Networks (ANNs)
are one of the most predominant ML techniques used in sports. Indeed, in
the course of our review, we found that a number of studies, particularly
early ones, made use of ANNs as their sole predictive model. This leads us
to investigate whether the evidence in the literature suggested that ANNs
actually perform better than other ML models in practice for problem of
sport result prediction. One of the aims of this review is to consider this, as
well as four additional key research questions, which are listed as follows:

1. Does the evidence from the literature suggest that ANNs produce better
performance in terms of predictive accuracy than other ML models?

2. What are some of the defining characteristics that can be drawn from
successful studies?

3. Are the results of some sports inherently more difficult to predict than
in others?

4. What are the best ways to evaluate accuracy results in this application
domain?

5. What are the common themes of future research directions that can be
identified from surveyed papers?

Our contribution lies in an extensive and up-to-date literature review

which collates key studies carried out in the emerging field of predictive
analytics in sports. We highlight which algorithms have been preferred by
practitioners and have been at the forefront of this field. This review identifies
what have been some of the key drivers of successful studies with respect to
how machine learning was applied, and how those elements have contributed

3
to high predictive accuracies. This review makes a meaningful contribution
in identifying future research trends and opportunities in this field, while
combining all the findings into a useful set of recommendations for other
practitioners in this domain.
The remainder of this paper is structured as follows. In section 2 we
outline our methodological approach and inclusion criteria for this study.
Then, in section 3, we review the literature in machine learning for sport
result prediction and also present tabular summaries of studies by sport. We
address the aforementioned key research questions, and in doing so, provide
a discussion of these results, as well as potential limitations of our review, in
section 4. We conclude in section 5.

2. Methodology
A total of 31 papers were considered in the present study. These were
collected from the period of 1996 to 2019. The distribution of papers by year
of publication can be seen in Figure 1, showing a growing level of interest
in this field from 2009 onward. These papers covered nine separate sports.
Given that some papers studied multiple sports, a total of 35 sport studies
were conducted. The specific sports and the distribution of the number of
studies they received can be seen in Figure 2.

Figure 1: Number of papers by year considered in this review.

Inclusion Criteria:. Our review includes papers that incorporated at least

one ML technique in their experimentation and therefore the review did

4
Figure 2: Number of sporting codes studied between 1996 - 2019 for the purposes of
predictive modeling which were considered in this paper.
not consider statistical, mathematical or operational research studies related
to sport result prediction that do not use these techniques. Our focus is
on the application of ML - fuzzy methods and Deep Learning methods are
not the focus of this review. We also restricted our review to team sports.
Papers generally needed to be published in peer-reviewed academic journals
or conference proceedings in order to be included.

Algorithm Grouping:. Machine Learning algorithms were normalized by group-

ing them into families of algorithms in order to identify trends in their us-
age patterns. All variants of ANN such as Back Propagation (BP), Feed-
Forward, as well as Self Organising Maps and Long Short Term Memory
ANNs were clustered under the same umbrella of algorithms. CART, C4.5
and WEKA’s J48 algorithms were grouped under the Decision Tree family of
methods. RIPPER, FURIA and ZeroR were merged into the Rule Sets cate-
gory, while the Local Weighted Learning (LWL) algorithm was merged with

5
k-Nearest-Neighbors (kNN). Additionally, AdaBoost, XGBoost, LogitBoost,
RobustBoost, and RDN-Boost were clustered in the Boosting category, while
Sequential Minimal Optimization (SMO) was combined with Support Vec-
tor Machines (SVMs). Methods which combined several different families of
machine learning algorithm into a single decision making architecture were
termed Ensemble. Though Naive Bayes and Bayesian Networks share com-
mon foundations, a decision was made to keep them separate since the latter
includes the ability to incorporate domain knowledge to a larger degree,
which was a motivating factor for its usage in several studies. A histogram
depicting the usage patterns of the above algorithms, sorted according to
frequency of use, can be seen in Figure 3.

Figure 3: Histogram of usage patterns of various algorithms covered in this research.

Accuracy Reporting. This review considered the accuracy measure defined as

the proportion of the sum of total true positive and true negative classifica-
tions divided by all samples in the dataset, as the primary metric for analysis,
thus ensuring that comparisons and interpretations were meaningful across
different studies.
Note that the charts will be presented in section 3 only include studies
that reported accuracy as their measure of performance, however all results
(including one paper that reported balanced accuracy, and three that re-

6
ported the average ranked probability score) are presented in the summary
tables. Moreover, results from binary-class and multi-class problems are not
directly comparable since a given level of accuracy accuracy becomes harder
to achieve as the number of classes increases. Thus, where a three-class for-
mulation was used instead of a two-class formulation, these accuracies have
also been excluded from the charts for comparative purposes, but again, are
reported in the summary tables.
Future Research Themes. In order to draw insights across all papers as to
what the general future research direction trends might be, a set of recurring
general themes were first extracted from all the papers. Subsequently, all the
text referring to future research directions across all the papers was coded
based on the extracted themes, and a histogram was rendered depicting the
frequency of each theme. Below is the list of themes used for this analysis:
• improve machine learning parameterization
• improve training methods
• engineer richer features
• improve feature selection
• benchmark predictions with experts
• investigate the causal relationships deeper
• increase dataset size
• apply findings to a different sport
• use other machine learning algorithms

3. Literature Review
In this section we review the literature in the field of ML for sport result
prediction, in chronological order. We also summarize the results of the
literature in tables in terms of the competition/league, the models used, the
number of matches available in their original dataset, as well as the best
performing model with the number of features used in that best model. The
tables presented are: American Football - table 1, Basketball - table 2, Soccer
- table 5, Ice Hockey - table 3, Rugby Union - table 4 and other sports - table
6, which only had one study associated with them.

7
 Paper Models used No. of No. of Accuracy
features Matches of best
model
Purucker (1996) ANN with BP*, 6 90 75% (week
Unsupervised: 14 and 15
Hamming, ART, SOM combined)
Kahn (2003) ANN with BP 10 208 75%
David et al. (2011) Committees of ANNs 11 unknown unknown
trained by LM
Delen et al. (2012) SVM, CART*, ANN 28 244 86%
Table 1: American Football Studies (* denotes the best performing model).
3.1. 1990s
Purucker (1996) used an ANN as well as unsupervised learning techniques
to predict the results of US National Football League (NFL) football matches
in 1994, using data from weeks 11 to 16 of the competition (90 matches).
Six features were considered: victories, yardage differential, rushing yardage
differential, turnover margin, time in possession, and betting odds (the in-
clusion of betting line odds were found to improve upon initial results). An
ANN trained with BP was used (BP provided the best performance among
the different network training methods). Unsupervised learning methods
were also applied, in particular: the Hamming, Adaptive Resonance Theory
(ART), and Self Organizing Map (SOM) methods. The SOM provided the
best performance among the unsupervised methods, but it generally could
not match the performance of the ANN. Matches from weeks 12 to 15 were
used to predict week 16, with the ANN correctly predicting 11 out of the 14
matches (78.6%) in week 16. Weeks 12 to 14 were also used to predict week
15, with the ANN correctly predicting 10 of the 14 games (71.4%). The au-
thor recognised that the dataset consisted of a small number of matches and
features, and stated that improvements could be gained through fine tuning
the encoding, the ANN architecture, and training methods.

3.2. 2000s
Kahn (2003) predicted matches in the NFL using data from 208 matches
in the 2003 season. An ANN with BP was used, and the features included in
the model were: total yardage differential, rushing yardage differential, time
in possession differential, turnover differential, a home or away indicator,
home team outcome and away team outcome. The numeric features were

8
 Paper Competition Models used No. of No. of Accuracy
features Matches of best
model
Loeffelholz et al. (2009) NBA ANN (types: FFNN*, 4 650 74.3%
RBG, PNN, GRNN,
fusions of these)
Zdravevski & Kulakov (2009) NBA All models in WEKA 10 1230 72.8%
(Logistic Regression*)
Ivanković et al. (2010) Serbian First ANN trained with BP 51 890 81%
B
Miljković et al. (2010) NBA kNN, Decision Tree, 32 778 67%
SVM, Naïve Bayes*
Cao (2012) NBA Simple Logistic 46 4000 67.8%
Regression*, Naïve
Bayes, SVM, ANN
Shi et al. (2013) NCAAB ANN*, C4.5 Decision 7 32236 74%
Tree, RIPPER,
Random Forest
Thabtah et al. (2019) NBA ANN, Naïve Bayes, 8 430 83%
LMT Decision Tree*

Table 2: Basketball Studies (*denotes the best performing model).

calculated as the 3-week historical average (the average of the statistic over
the past 3 weeks of the competition), as well as the season average (the
average of the statistic over the entire season). It was found that using
the season average provided better accuracy. Weeks 1 to 13 of the 2003
competition were used as training data, with weeks 14 and 15 as the test
set. Accuracy of 75% was achieved, which was slightly better than expert
predictions on the same games. It was suggested that in future work, betting
odds or team rankings could be included as model predictors, and also that
previous seasons could be used for training a model instead of only the current
season.
O’Donoghue & Williams (2004) were interested in comparing human ex-
perts with computer-based methods in terms of their ability to predict the
results of matches in the 2003 Rugby Union1 World Cup. Multiple Linear
Regression, Logistic Regression, an ANN and a Simulation Model were the
computer-based models. The authors made use of data from the previous
four World Cups to develop predictive models of match outcomes based on

1
Rugby Union will hereafter be referred to simply as Rugby (not to be confused with
Rugby League, which is a different sport).

9
 Paper Competition Models used No. of No. of Accuracy
features Matches of best
model
Weissbock et al. (2013) NHL Naive Bayes, SVM, ANN*, 11 517 59%
C4.5 Decision Tree
Weissbock & Inkpen (2014) NHL Weka: ANN, Naive Bayes, 6 708 60.3%
Complement Naive Bayes,
Multinomial Naive Bayes,
LibSVM, SMO, J48, JRip,
Logistic, Simple Logistic,
Simple Naive Bayes,
Cascading Ensemble*
Gu et al. (2019) NHL KNN, SVM, Naive Bayes, 19 1230 91.8%
Discriminant Analysis,
Decision Trees + ensembles
of these: Boosting*,
Bagging, AdaBoost,
RobustBoost

Table 3: Ice Hockey Studies (*denotes the best performing model).

three variables: team strength (determined by synthesising world rankings2 ),
distance travelled to the tournament as a measure of home advantage, and
the number of recovery days between matches. The computer-based mod-
els predicted between 39 and 44.5 of the 48 matches correctly, while the 42
human experts correctly predicted an average of 40.7 matches. However, it
was noted that there was a far greater variation in the accuracies from the
human experts, with the most successful person correctly predicting 46 of the
48 matches. The most successful computer-based model was the Simulation
Model.
Reed & O’Donoghue (2005) built seven predictive models (including mod-
els based on Multiple Linear Regression, an ANN and Discriminant Analysis)
using seven predictor variables, to predict results in English Premier League
(EPL) Soccer and Premiership Rugby in England. The dataset was derived
from three football teams and two Rugby teams, across three seasons. After a
review, the authors included the following predictive variables: match venue,
rest, the team’s position on league table, the opposition team’s position on
league table, distance travelled to the match (for both teams), and form.
These variables were common to both Rugby and Soccer, i.e., the study did
not incorporate variables specific to in-play match events in Rugby or Soccer.
Considering that a draw is a much more common outcome in Soccer than

2
World rankings in Rugby had not yet been introduced at the time of the 2003 Rugby
World Cup.

10
in Rugby, it was surprising that the percentage of Soccer matches (57.9%)
that were predicted correctly was higher than Rugby (46.1%). Despite the
best performing predictive model accuracies being rather low, the models still
outperformed the predictions of human experts with an ANN found to yield
the best accuracy. The authors stated that motivational, injury and other
variables were not included, and that future studies could concentrate on
developing complex pattern recognition technology rather than using simple
linear models which could lack flexibility.
Joseph et al. (2006) found that it may be beneficial to incorporate ex-
pert knowledge into a modeling process. Expert knowledge was incorporated
while constructing a Bayesian Network (BN) model, and it was found that
including such knowledge can result in strong performance, especially when
the sample size is small (which is often the case with sports). A decision
tree (MC4) and kNN model were also used to predict the results of Soccer
matches played by the EPL team Tottenham. Their dataset consisted of 76
matches. Four variables were included in the expert model, while 30 vari-
ables were used in their original general model. The expert BN was found
to provide the best performance, achieving 59.2% accuracy in predicting a
home win, away win or a draw (a 3-class problem). In the future, the au-
thors proposed developing a more symmetrical model using similar data but
for all the teams in the EPL, and also incorporating player quality features
(e.g., players that have performed at international level) into the model. In
addition, they mentioned that additional nodes could be added to the BN,
e.g., attack quality and defence quality.
McCabe & Trevathan (2008) considered four different sports: Rugby
League in Australia’s National Football League (NFL), Australian Rules
Football in the Australian Football League (AFL), Super Rugby, and EPL
soccer, using data from 2002 to 2007. An ANN trained with BP and an
ANN trained with Conjugative-Gradient Descent (CGD) were applied. It
was found that training the ANN with BP was slightly more accurate than
the CGD algorithm; however, the training time with BP was longer. The
variables used in the study were the same across all of the four sports, i.e.,
sport-specific variables derived from in-play events within a match were not
included. The average accuracy achieved by the ANN with BP was 67.5%,
higher than expert predictions which ranged from 60% to 65%. For future
work, the authors mentioned that other sports could be considered, as well
as including more variables, and predicting the points margin.

11
 Loeffelholz et al. (2009) attempted to predict National Basketball Asso-
ciation (NBA) matches from the 2007/2008 season. There were 620 matches
that were used for training and testing, and 30 that were used as the valida-
tion set - treated as “un-played” games. The set of predictive variables con-
sisted of: field goal percentage, three-point percentage, free-throw percent-
age, offensive rebounds, defensive rebounds, assists, steals, blocks, turnovers,
personal fouls, and points. In terms of predicting the un-played games, aver-
ages of variables in the current season were found to result in better perfor-
mance than averaging statistics across the past five matches. Fusion of ANNs
was an approach also investigated by the authors, using Bayesian Belief Net-
works and Neural Network (NN) fusion in particular. Four different types of
ANN were applied - a Feed-Forward NN (FFNN), a Radial Basis Function
NN (RBF-NN), a Probabilistic NN (PNN) and a Generalized NN (GRNN).
Their best model - an ANN with FFNN with four shooting variables, cor-
rectly predicted the winning team 74.3% of the time on average, which was
better than experts from USA Today 2008 who achieved 68.7%. The fea-
ture selection technique used was an iterative Signal-to-Noise Ratio (SNR)
method (Bauer Jr et al., 2000), which selected four out of the 22 original
variables. It was mentioned that although fusion techniques did not result in
higher accuracy on this occasion, they still could be worth investigating in
the future. Also, different features could be used in the baseline model, and
the models created could be adjusted to determine if the model can beat the
betting odds rather than simply predicting which team will win.
Zdravevski & Kulakov (2009) obtained two seasons of NBA data consist-
ing of 1230 matches. The first season was used as the training dataset, while
the second was used as the test dataset. All of the algorithms contained in
the WEKA machine learning workbench were used for experimentation, ap-
plied with their default parameter settings. A set of 10 variables was selected
by the authors. Classification accuracy of 72.8% was achieved with Logistic
Regression. It was stated that in future work it would be preferable to com-
pare their predictions to those of experts, and that it might be possible to
cluster training and test data to use different models on each cluster in order
to account for winning and losing streaks. Furthermore, it was mentioned
that aggregations or ensembles of classifiers (e.g., voting) could be investi-
gated, and that automatic feature selection methods should be used rather
than manual human selection.

12
3.3. 2010s
3.3.1. 2010 - 2014

Buursma (2010) used data from 15 years of Dutch Soccer to predict re-
sults. Buursma was specifically interested in: which variables were important
when predicting match outcomes, how the probabilities for the match out-
comes can be predicted from these variables, and what bets should be placed
to maximize profit. The author experimented with the following models im-
plemented in WEKA: Classification via Regression, a Multi-class Classifier,
Rotation Forest, Logit Boost, Bayesian Network, Naive Bayes, and ZeroR.
There were three match outcomes to be predicted: home win, draw, and
away win. The feature set consisted of 11 features, and all features were
either aggregated or averaged across a team’s past 20 matches (20 was found
by experimentation to be the best number of matches in which to average
across). It was determined that Classification via Regression and the Multi-
class Classifier had the best prediction accuracy of 55%. For future work
the author considered including more features such as: yellow/red cards, the
number of players each team has, their managers, their player budget and
their home ground capacity. Buursma was also interested in applying such a
model to other similar sports such as basketball, baseball and ice hockey in
the future.
Huang & Chang (2010) used an ANN with BP to predict match results
in the 2006 Soccer World Cup. A total of 64 matches were played in the
tournament. The output of the ANN was the relative ratio between two
teams for each game, which could then be converted into an odds ratio. Ac-
curacy of 76.9% accuracy was achieved; however, it should be noted that
this is higher than some other Soccer studies because their model did not in-
clude prediction of draws. Feature selection was performed based on domain
knowledge of the authors. Eight features were used in the model, obtained
from the match statistics of the FIFA World Cup website. These were: goals
for, shots, shots on goal, corner kicks, direct free kicks on goal, indirect free
kicks on goal, possession, and fouls conceded. The authors remarked that
the ANN generally had difficulty predicting a draw, which was a frequent
outcome in the group stages of the tournament.
Ivanković et al. (2010) utilized an ANN with BP for predicting the results
of the Serbian First B Basketball League using five seasons of data from

13
 Paper Competition Models used No. of No. of Accuracy
features Matches of best
model
O’Donoghue & Williams (2004) 2003 Rugby Multiple linear 3 48 93%
World Cup regression, Logistic
Regression, ANN,
Simulation model*
Reed & O’Donoghue (2005) English Multiple linear 7 498 46.1%
Premiership regression, ANN*,
Rugby discriminant function
analysis
McCabe & Trevathan (2008) Super Rugby ANN trained with 19 unknown 74.3%
BP* and CGD
O’Donoghue et al. (2016) 2015 Rugby Linear regression*, 1 280 74.4%
World Cup Simulation

Table 4: Rugby Union Studies (*denotes the best performing model).

2005/2006 to 2009/2010 - comprising 890 matches. The authors applied the
Cross-Industry Standard Process for Data Mining (CRISP-DM) framework
(Wirth & Hipp, 2000) as their experimental approach, and investigated how
the successful shot percentage in six different regions of the court affected
match results. The input dataset was divided into 75:25 ratios for training
and testing, and 66.4% accuracy was obtained on the test set. The authors
then cycled back to the data preparation phase of the CRISP-DM framework
to see if adding additional variables could improve results. In particular,
offensive rebounds, defensive rebounds, assists, steals, turnovers and blocks
were subsequently incorporated. This improved the accuracy achieved by the
model to just under 81%. It was concluded that actions in the zone directly
under the hoop - rebounds in defence, as well as scoring in this zone were
crucial in determining the outcome of the game. It was mentioned that a
richer data set and new software solutions may be helpful in the future, so
that all relevant events are included.
Miljković et al. (2010) used data from 778 games in the regular 2009/2010
NBA season to predict basketball match outcomes. The predictive features
were divided into game statistics which directly relate to events within the
match e.g., fouls per game and turnovers per game, and those that relate to
standings e.g., total wins and winning streak. Naive Bayes was found to be
the best performing model compared to kNN, a Decision Tree and a Support
Vector Machine (SVM) model. Accuracy of 67% was achieved with Naive
Bayes, using 10-fold cross validation. Their future research plans included
applying their system to other sports, and to experiment with other models
such as ANNs.

14
 David et al. (2011) used a committees-of-committees approach with ANNs
to predict NFL matches, whereby many networks were trained against dif-
ferent random partitions of the data. Under their approach, 500 ANNs were
used for training, and from this, the best 100 were used in each committee.
Subsequently, 50 such committees were used to produce the final predictive
results. The mean was used to determine the vote of each committee and
to combine the predictions of the committees. The features used were dif-
ferentials between home and away teams based on: passing yards, yards per
rush play, points, interceptions, and fumbles. The differentials between home
and away teams were to incorporate the well-known home advantage phe-
nomenon. The season-to-date average of these statistics were used describe
each team, apart from the first five rounds of the season where the weighted
average between the current season and the previous season statistics was
used3 . A total of 11 inputs were used in the ANN for each game, and the
Levenberg Marquadt (LM) routine was used to train the network. Principal
Components Analysis (PCA) and derivative-based analyses were applied in
order to determine which variables were most influential. It was found that
99.7% of the variance in their data was due to passing and rushing differen-
tials. The results were compared to bookmakers, as well as predictions from
the website thepredictiontracker.com. It was found that the results were com-
parable to the predictions of bookmakers, and were also better than most of
the predictive models on the website. The avenues for future experimentation
were listed as: using different types of ANN, e.g., RBF, including additional
statistics (e.g., possession, strength-of-schedule, kicking game and injuries),
investigating how to best predict games early in the season, and application
to other levels of football (e.g., NCAA) as well as to other sports.
In contrast to Joseph et al. (2006), Hucaljuk & Rakipović (2011) found
that incorporating expert opinion did not produce any improvement in pre-
dicting the outcome of soccer matches. Their dataset consisted of 96 matches
(6 rounds) in the regular part of the European Champions League competi-
tion. This data was split three different ways into training and test datasets:
a 3 round-3 round training-test split, a 4 round-2 round split, and a 5 round-1
round split. Feature selection resulted in 20 features in their basic feature

3
100% of the statistics from the previous season were used for week 1, then 80% of the
previous season and 20% of the current season in week 2. This process was continued with
a 20% increase each week until week 6, at which point only the current season statistics
were used.

15
set. An additional feature set consisted of the 20 basic features plus variables
selected by experts. Naive Bayes, Bayesian Networks, Logit Boost, kNN, a
Random Forest and an ANN with BP were used in the experiments. Over-
all, the ANN performed the best, achieving classification accuracy of 68%.
Perhaps surprisingly, the expert-selected features were not found to yield
any improvement. It was mentioned that further improvements could be
gained in refining feature selection, modeling the form of players in matches
throughout a season, as well as obtaining a larger dataset.
Cao (2012) built a system that automated data collection and applied
various data mining algorithms for basketball match prediction. Six years
of data from NBA matches was obtained, from the 2005/2006 season to the
2010/2011 season (comprising around 4,000 matches). The dataset was di-
vided into training sets, test sets and validation sets. Four classification
algorithms were used in the experiments: Simple Logistic Regression, Naive
Bayes, a SVM and an ANN with BP. The feature selection process was man-
ual, with 46 features being selected based on the domain knowledge of the
author. All models were found to produce reasonably similar rates of classifi-
cation accuracy, with simple Logistic Regression achieving the best accuracy
of 67.8%. The Simple Logistic Regression model also provided automatic
feature selection. The best expert predictions on teamrankings.com were
slightly better, achieving 69.6% accuracy. The author suggested that in fu-
ture clustering techniques could be used to group players by positional group,
or to identify outstanding players. Using outlier detection methods to iden-
tify outstanding players or team status, investigating the impact of player
performance match outcome, and comparing different feature sets derived
from box-score and team statistics were also mentioned as avenues for future
work.
Delen et al. (2012) used a SVM, CART Decision Tree and an ANN to
predict the outcome of NCAA Bowl American Football matches. The CRISP-
DM framework was applied as the experimental approach. The dataset was
made up of 28 features and 244 matches. Both a classification approach pre-
dicting home win/away win, as well as a numeric prediction approach, which
predicted the points margin (home team points less away team points), were
compared. The CART Decision Tree provided the best performance with
86% accuracy (using 10-fold CV), which was statistically significantly better
than the other models. The classification approach was found to produce
better results compared with numeric prediction. When models were trained
on the seasons 2002/2003 to 2009/2010, and then tested on the 2010-2011

16
 Paper Competition Models used No. of No. of Accuracy
features Matches of best
model
Reed & O’Donoghue (2005) English Premier Multiple Linear Regression, 7 498 57.9%
League ANN*, Discriminant (3-class)
Function Analysis
Joseph et al. (2006) English Premier Bayesian Network, Expert 4 76 59.2%
League - Bayesian Network*, (3-class)
Tottenham Decision Tree, kNN
Hotspur
McCabe & Trevathan (2008) English Premier ANN trained with BP* and 19 unknown 54.6%
League CGD (3-class)
Buursma (2010) Dutch Eredivisie WEKA: 11 4590 55%
League MultiClassClassifier with (3-class)
ClassificationViaRegres-
sion*, RotationForest,
LogitBoost, Bayesian
Network, Naïve Bayes,
ZeroR
Huang & Chang (2010) 2006 Soccer ANN trained with BP 8 64 62.5%
World Cup (3-class),
76.9%
(2-class)
Hucaljuk & Rakipović (2011) European naïve Bayes, Baysian 20 96 68%
Champions network, LogitBoost, kNN, (3-class)
League random forest, ANN*
Odachowski & Grekow (2012) Various leagues BayesNet*, SVM, LWL, 320 1,116 70.3%
Ensemble Selection, CART, (2-class),
Decision Table 46%
(3-class)
Tax & Joustra (2015) Dutch Eredivisie WEKA: NaiveBayes, 5 4284 56.1%
League LogitBoost*, ANN, (3-class)
RandomForest, CHIRP,
FURIA, DTNB,
J48,HyperPipes
Prasetio (2016) English Premier Logistic Regression 4 2280 69.5%
League (2-class)
Danisik et al. (2018) Various leagues LSTM NN classification, 139 1520 52.5%
LSTM NN regression*, (3-class),
Dense Model 70.2%
(2-class)
Hubáček et al. (2019b) 52 leagues XGBoost classification*, 66 216,743 52.4%
XGBoost regression, (3-class)
RDN-Boost
Constantinou (2019) 52 leagues Hybrid Bayesian Network 4 216,743 51.5%
(3-class)
Berrar et al. (2019) 52 leagues XGBoost*, kNN 8 216,743 51.9%**
(3-class)

Table 5: Soccer Studies (*denotes the best performing model). Accuracies for 2-class (winn,loss)
and 3-class (win,loss,draw) problems are denoted. **Berrar (2019)’s best performing model was kNN, but
post-competition they mention that they improved on this with XGBoost. The accuracy of 51.9% is for
their in-competition result - the XGBoost accuracy was not reported in their paper but it would have
been slightly higher than this.

17
season, CART again had the best performance with 82.9% accuracy. The
suggested directions for future research were: including more variables or
representing them in different forms, making use of other classification and
regression methods (e.g., rough sets, genetic algorithm based classifiers, en-
semble models), experimenting with seasonal game predictions by combining
static and time series variables, and applying the approach to other sports.
Odachowski & Grekow (2012) analysed fluctuations in betting odds and
used ML techniques to investigate the value in using such data for predicting
soccer matches. The odds for home win, away win and draw for the preceding
10 hours, measured at 10-minute intervals, were tracked over time (the data
having been obtained from the Betfair and Pinnacle Sports websites). A
total of 32 features were computed from this time-series - e.g., maximum and
minimum changes in betting odds, overall changes in odds, and standard
deviations. The 10 hour period was divided into thirds and the features were
also calculated based on these sampling periods. The authors balanced their
dataset such that there were an equal number of home wins, away wins, and
draws (372 matches for each class). Six classification algorithms from WEKA
were compared: Bayesian Network, SMO, LWL, ensemble selection, Simple
CART and a Decision Table, and feature selection methods contained in
WEKA were applied. It was found that draws in particular were difficult to
correctly predict, with only around 46% accuracy obtained when attempting
3-class classification. However, accuracy of around 70% was obtained when
ignoring draws. Discretization and feature selection methods were found to
improve results. The authors suggested that additional features describing
changes in betting odds could be included going forward.
Shi et al. (2013) investigated the viability of machine learning in the
context of predicting outcomes of individual NCAAB matches which up
to this point had been dominated by statistical methods. They used the
WEKA toolkit for their comparative study of accuracies between ANN, C4.5,
RIPPER and Random Forest algorithms. Experiments were conducted us-
ing data from six seasons together with an expanding window approach so
that initially training was performed on the 2008 season and the testing on
the 2009 season. Thereafter, the combination of all previous seasons com-
posed the training set until the 2013 season. The authors concluded that on
average, the ANN algorithm with default parameter settings provided the
best accuracies, though statistical tests to confirm this were not provided.
The top ranked features in terms of importance were location, the so-called
"four factors" (Oliver, 2002), and adjusted offensive and defensive efficiencies

18
(kenpom.com). The authors also remarked that they experienced an upper
limit of 74% accuracy which they could not improve beyond, while noting
that feature engineering and selection hold promise for an improvement in
results.
Weissbock et al. (2013) noted that Ice Hockey had not received much
attention in ML research with respect to predicting match outcomes. The
continuous nature of this sport makes it difficult to analyse due to a paucity
of quantifiable events such as goals. This characteristic was cited as a possible
reason for the lack of attention Hockey had received historically. Weissbock
et al. (2013) focused on exploring the role of different types of features in
predictive accuracies across several types of ML algorithms. They consid-
ered both traditional statistics as features, as well as performance metrics
used by bloggers and statisticians employed by teams. They used WEKA’s
implementations of ANN, Naive Bayes, SVM and C4.5 for training classi-
fiers on datasets describing National Hockey League (NHL) match outcomes
in the 2012/2013 season. The entire dataset amounted to 517 games. The
authors concluded that traditional statistics outperformed the newer perfor-
mance metrics in predicting outcomes of single games using 10-fold CV, while
the ANN displayed the best accuracy results of 59%. Research in extracting
more informative features and predicting winners of the NHL playoffs was
cited as future work as well as incorporating knowledge from similar sports
such as soccer.
Weissbock & Inkpen (2014) combined statistical features with features de-
rived from pre-game reports, in order to determine whether the sentiment of
the pre-game reports had any usefulness for the prediction of NHL Ice Hockey
matches. Data from 708 NHL games in the 2012/2013 season was collected,
and pre-game reports were acquired from the website NHL.com. Both nat-
ural language processing and sentiment analysis based features were used in
the experiments. The three statistical features that were used were identified
from their previous work (Weissbock et al., 2013), which were: cumulative
goals against and their differential, and the match location (home/away).
The following algorithms from the WEKA toolkit were applied with de-
fault parameters: ANN, Naive Bayes, Complement Naive Bayes, Multinomial
Naive Bayes, LibSVM, SMO, J48 Decision Tree, JRip, Logistic Regression
and Simple Logistic Regression. Three models were compared: models that
used statistical features only, models using pre-game report text only, and
models trained with sentiment analysis features. It was found that using only
pre-game reports features for prediction did not perform as well as models

19
trained with statistical features only. A meta-classifier with majority vot-
ing was implemented where the confidence and predicted output from initial
classifiers was fed into a second layer. This architecture provided the best
accuracy of 60.25%. This cascading ensemble approach on the statistical fea-
ture set provided superior performance to using both feature sets, suggesting
that the pre-game reports and statistical features provided somewhat differ-
ent perspectives. The authors noted that it was difficult to predict matches
with a model trained beyond one or two seasons due to player changes and
so on.

3.3.2. 2015 - 2019

Tax & Joustra (2015) focused on identifying features which have a high
utility for predicting football match outcomes, and which satisfy the crite-
rion of being easily retrievable. In particular, they sought to explore whether
or not there were different levels of utility between features broadly classi-
fied as public match-data sources and those derived from bookmaker odds.
Their experiments covered data from the 2000 to 2013 seasons of the Dutch
Eredivisie competition. The researchers conducted numerous experiments
comparing the Naive Bayes, LogitBoost, ANN, Random Forest, CHIRP, FU-
RIA, DTNB, J48 and Hyper Pipes algorithms from the WEKA toolkit. The
experiments were performed in conjunction with feature reduction methods
such as PCA, Sequential Forward Selection and ReliefF. The best results
on the public match-data sources were achieved by the Naive Bayes and
ANN classifiers in combination with PCA, achieving an accuracy of 54.7%.
FURIA achieved the highest accuracy using bookmaker odds features with
55.3%; however, this was ultimately not found to be statistically significant.
A marginal improvement in accuracy was realised with LogitBoost and Re-
liefF when both bookmaker odds and public match-data features were used,
producing an accuracy of 56.1%; however, this also was not at a statistically
significant level, but nonetheless pointed to the potential utility in combining
a broader variety of features for further investigation.
Pathak & Wadhwa (2016) attempted to predict the results of One-Day
International (ODI) Cricket. The authors included four predictive variables
based on the prior statistical work of Bandulasiri (2008): toss outcome,
whether the game was played at home or away, whether the game was played
during the day or at night, and whether the particular team in question bat-

20
 Paper Sport- Models used No. of No. of Accuracy
Competition features Matches of best
model
McCabe & Trevathan (2008) Rugby ANN trained with 19 unknown 63.2%
League - BP* and CGD
NRL
McCabe & Trevathan (2008) Australian ANN trained with 19 unknown 65.1%
Rules BP* and CGD
Football -
AFL
Pathak & Wadhwa (2016) Cricket - Naïve Bayes, Random 4 unknown 61.7%
International Forest, SVM* (balanced
ODIs accuracy)
Valero (2016) Baseball - Lazy Learners, ANN, 3 24,300 59%
MLB SVM*, Decision Tree

Table 6: Other Sport Studies (*denotes the best performing model).

ted first or second. Three classification models were applied - Naive Bayes,
Random Forest and SVM. Data for matches from 2001 - 2015 was collected
from cricinfo.com. A separate model was constructed for each team, which
analysed that team with respect to all other teams. An 80-20 training-test
split was used. To mitigate the effect of imbalanced data sets (some teams
had a high ratio of wins to losses), the three models were evaluated based on
balanced accuracy and the Kappa statistic. SVM was found to perform the
best across all teams with an average balanced accuracy of 61.67%. It was
suggested that using more novel methods of classification for outcome predic-
tion, including additional features, application to other forms of cricket (e.g.
test matches, T20), and application to other sports (e.g. baseball, football)
could be performed in future work.
Prasetio (2016) used Logistic Regression to predict EPL soccer results
in the 2015/2016 season. Their data set consisted of data from six seasons
from the 2010/2011 to the 2015/2016 seasons - a total of 2,280 matches.
Home offense, away offense, home defence and away defence were used as
the input features, and it was found that the significant variables were home
defence and away defence (it was not mentioned in the paper how these
offense and defence ratings were actually constructed). Despite this, the
model that included all four variables was found to yield a higher accuracy.
The split of training and testing data was varied into four different training-
test partitions which produced four different sets of model coefficients. The
best performing model achieved 69.5% accuracy (however since they used a
binary Logistic Regression model, this result was based on the omission of

21
draws). In the future, they remarked that the results could be used to assist
management with game strategy, or that the trained models could be turned
into a recommendation system for evaluating which players to purchase.
Valero (2016) performed a comparative study for prediction of Baseball
matches, using 10 years of Major League Baseball (MLB) data. Lazy Learn-
ers, ANNs, SVM and Decision Trees were the candidate models. Inspired by
Delen et al. (2012), the CRISP-DM framework was applied, and addition-
ally, the performance of a classification approach (win or loss for the home
team) as well as a numeric prediction approach (predicting the run differ-
ence between the home and away team) were compared. Feature selection
methods from WEKA were applied for dimensionality reduction and rank-
ing of the original set of 60 features. Their model used only the top three
ranked variables - home field advantage, Log5 and the Pythagorean Expec-
tation4 (the addition of more variables was not found to improve results).
SVM was found to produce the best accuracy under both the classification
and numeric prediction approaches. Consistent with Delen et al. (2012), the
classification approach performed significantly better than the numeric pre-
diction approach. The SVM classification model achieved accuracy of around
59%. However, when using the 2005-2013 seasons as training data and the
2014 season as test data, the model’s predictions were not significantly more
accurate than predictions derived from match betting odds. The authors
highlighted the difficulty in predicting outcomes in Baseball using statistical
data alone, but suggested that experiments using the Japanese or Korean
Baseball leagues could be useful. In the future, the authors also consid-
ered adjusting their model parameters, refining features and extending their
datasets. In addition, they planed to apply the model to other sports such
as Basketball, Football and Water Polo.
O’Donoghue et al. (2016) compared the accuracy of 12 models based
on Linear Regression to analyse the effectiveness of predicting team perfor-
mances in the 2015 Rugby World Cup. The models differed in (i) whether
or not the assumptions of the predictive modeling technique were satisfied
or violated, (ii) whether all (1987-2011) or only recent Rugby World Cup
tournaments (2003-2011) data were used, (iii) whether the models combined

4
The Pythagorean Expectation was developed by the Baseball statistician Bill James
(James, 1984), and represents the expected number of wins for a team given their runs
scored and runs allowed, while Log5 ratings are essentially the same as Elo (1978) Ratings
(Koseler & Stephan, 2017).

22
pool and knockout stage match data, and (iv) whether the models included
a variable that tried to capture a relative home advantage. The common in-
dependent variable in all models was the relative quality, which represented
the difference between the higher ranked team’s world ranking points. The
dependent variable was the point difference in the match outcome between
two teams. All models were executed 10,000 times within a simulation pack-
age that introduced random variability. The best model achieved accuracy of
74%, with the key finding being that match outcomes in international Rugby
appeared more difficult to predict than in previous years. The best model
used data from all previous Rugby World Cups in a way that violated the as-
sumptions of Linear Regression. The best model used only one independent
variable and ignored the relative home advantage, while generating separate
models for the pool and knockout stage matches.
Danisik et al. (2018) applied a Long-Short-Term-Memory (LSTM) NN
model (Hochreiter & Schmidhuber, 1997) for match result prediction in a
number of different Soccer leagues. Classification, numeric prediction and
dense approaches were compared, and were also contrasted with an average
random guess, bookmaker’s predictions, and the most common class outcome
(home win). Player-level data was included, and was obtained from FIFA
video games. Incorporating player-level and match history data, a total of
139 features were included (134 in the dense model). Four seasons of the
EPL were considered - 2011/2012 to 2015/2016, comprising a total of 1,520
matches. The average prediction accuracy obtained for a three-class classifi-
cation problem was 52.5%, achieved with the LSTM Regression model. This
result was achieved using 2011/2012, 2012/2013 and 2015/2016 seasons as
the training dataset and the 2013/2014 season as the validation dataset. The
accuracy obtained for a two-class problem, excluding draws, was 70.2%. It
was stated that betting odds and some additional match-specific, player and
team features could be included in the future. Additionally, the use of convo-
lution to transform input attributes during training could be investigated as
well as a deeper exploration the the ability of LSTM NN to leverage features
like specific tactics.
Thabtah et al. (2019) used Naive Bayes, an ANN, and a LMT Decision
Tree model to forecast the outcome of NBA Basketball matches. They fo-
cused on testing different feature sets in order to find the optimal subset of
predictors. The dataset was obtained from Kaggle.com, and contained 430
NBA finals matches from 1980 to 2017 and 21 features. The target variable
was a binary variable expressed as a win or a loss. It was found that defensive

23
rebounds was the most important factor influencing match results. The fea-
ture selection methods used were based on Multiple Regression, Correlation
Feature Subset - CFS (Hall, 1998), and RIPPER (Cohen, 1995). Defensive
rebounds were selected as important features by all three feature selection
methods. The best performing model in terms of accuracy was trained on a
feature set containing eight variables. These were selected from the RIPPER
decision rules and subsequently trained using the LMT model, which gen-
erated 83% accuracy. The authors suggested that larger datasets and more
attributes could be considered going forward, such as the team’s players and
coach. The use of other models such as function based techniques and deep
learning methods were also stated as potential avenues for further research.
Gu et al. (2019) reported that ensemble methods provided encouraging
results in the prediction of outcomes of NFL Hockey matches over multi-
ple seasons. The data extraction was automated and scraped from several
websites listing NHL matches from 2007/2008 to 2016/2017 seasons (1230
matches). Data was merged from several sources including historical results,
opposition information, player-level performance indicators, as well as player
ranks using PCA. A total of 26 team performance variables were also in-
cluded in their model. The kNN, SVM, Naive Bayes, Discriminant Analysis
and Decision Tree algorithms, together with several ensemble-based methods,
were applied. The ensemble-based methods (Boosting, Bagging, AdaBoost,
RobustBoost) achieved the highest accuracy on the test set, with 91.8%. In
terms of future research, the authors mentioned that additional data could
further improve predictions, features consisting of different or additional
predictors could be used, e.g., player ranking metrics, moving-averaged or
exponentially-smoothed team and player performance metrics, psychological
factors, expert/coach strategic or tactical judgements and player’s physical
or mental assessments could be included. They also mentioned that the ML
problem could be re-formulated so that a different outcome could be pre-
dicted, e.g., whether or not a team will make the playoffs or which team will
win the championship, by training a model on the regular season data.

The Open International Soccer Database: Prediction Challenge. In

2017, a significant development took place for ML researchers interested in
Soccer. A comprehensive open-source database called the Open International
Soccer Database was compiled and made public. The database contains over
216,000 matches from 52 leagues and 35 countries (Dubitzky et al., 2019).
The motivation behind this project was to ultimately encourage ML research

24
in Soccer by building an up-to-date knowledge base that can be used on an
ongoing basis for the prediction of real-world soccer match outcomes, as
well as a benchmark dataset which makes comparisons between experiments
more robust. In order to maximise the utility of this database, a deliberate
design choice was made to collect and integrate only data that are easily
available for most soccer leagues worldwide, including lower leagues. The
consequence of this is that the database lacks fields that are highly-specialized
and sophisticated.
Subsequent to its creation, the 2017 Soccer Prediction Challenge was
conducted, which hosted a competition based on this dataset. The results
arising from it were published in a special issue of the Machine Learning
(Springer) Journal. The challenge involved building a single model in order
to predict 206 future match outcomes from 26 different Soccer leagues, which
were to be played from March 31 to April 9 in 2017. Unlike most prior
studies that used accuracy as a performance metric, this competition used
the average Ranked Probability Score (RPS) (Epstein (1969), Constantinou
et al. (2012)). In the remainder of this section, we summarize three of the
papers from the competition that focused on ML.
Hubáček et al. (2019b) experimented with both relational and feature-
based methods to learn predictive models from the database. Pi-ratings
(Constantinou & Fenton, 2013), which capture both the current form and
historical strengths of teams, and a rating based on PageRank (Page et al.,
1999) were computed for each team. XGBoost gradient boosted tree al-
gorithms (both regression and classification) were employed as the feature-
based method, while RDN-Boost was used as the relational method. Their
feature-based classification method with the gradient boosted tree algorithm
performed best on both the validation set and the unseen challenge test set,
with the model achieving 52.4% accuracy on the test set. Possibilities for
future work were cited as: augmenting the feature set, weighting aggregated
data by recency, including expert guidance in forming relational concepts,
e.g., using active learning, and identification of features that are condition-
ally important given occurrences of certain features at higher levels of the
tree.
Constantinou (2019) created a model that combined dynamic ratings with
a Hybrid Bayesian Network. The rating system was based partly on the
Pi-rating system of Constantinou & Fenton (2013), computing a rating that
captures the strength of a team relative to others in a particular competition.
Pi-ratings consider the actual goal differences in a match, compared with

25
the expected result (based on existing ratings) in order to update the team
ratings. The rating calculation involves placing more emphasis on the result
rather than the margin, so the effect of large goal differences are dampened.
Unlike the original Pi-ratings, this version also incorporated a team form
factor. This ratings algorithm essentially searched for continuously for over
or under-performance - see the paper for further details. Four ratings features
(two each for the home and away team) were used as the inputs into the
Bayesian Network. The model provided empirical evidence that a model can
make good predictions for a match between two particular teams, even when
the prediction was based on historical match data that involved neither of
those two teams. The model provided accuracy of 51.5% on the challenge
test data set. The author recognized the limited nature of this data set, and
that incorporating other key factors, such as player transfers, the availability
of key players, participation in international competitions, a new coach, level
of injuries, attack and defence ratings, and even team motivation/psychology
in the form of expert knowledge could lead to improved results.
Berrar et al. (2019) developed two methods in order to create two fea-
ture sets for result prediction - recency features and rating features. Recency
feature extraction involved calculating the averages of statistics across the
past nine matches, based on four feature groups: attacking strength, defen-
sive strength, home advantage and opposition strength. Rating features were
based on performance ratings of each team, updated after each match accord-
ing to the expected and observed match outcomes, as well as the pre-match
ratings of each team. The XGBoost and kNN algorithms were applied to each
of these two feature sets. Both of these models were found to perform bet-
ter on the rating feature set, and the best performance overall was obtained
with XGBoost on the rating features - however this result was achieved after
the competition. The best model submitted for the competition itself was
a kNN applied to the rating features, which achieved accuracy of 51.9% on
the unseen competition test set. It was mentioned that the generally small
number of goals per match and narrow margins of victory in soccer meant it
is difficult to make predictions based on goals only. The authors concluded
that innovative approaches to feature engineering as well as how well domain
knowledge can be incorporated into modeling is the key to success for sport
result prediction, and are likely more important than the choice of ML algo-
rithm. It is again recognized by the authors that the competition dataset was
limited, and that data about various game events (e.g., yellow and red cards,
fouls, ball possession, passing and running rates, etc.), players (e.g., income,

26
age, physical condition) and teams or team components (e.g., average height,
attack running rate) would help to improve results.
Overall, the competition involving the Open International Soccer Database
produced some innovative methods and approaches. Notably, researchers in
this competition tended to combine some form of ratings-based method with
machine learning techniques. Despite having a very large number of matches
on-hand in the dataset, all of the studies seemed to obtain accuracies that
levelled off at a certain point. This perhaps reflects that having a broad
range of predictive features is very important to success in predicting results
in sport.

4. Results & Discussion

In this section we focus on addressing the key research questions that
were outlined in the introduction, and in doing so, also provide some critical
analysis and discussion of the obtained results.

4.1. Does the evidence from the literature suggest that ANNs produce better
performance in terms of predictive accuracy than other ML models?
Our research found that the majority of studies (65%) considered ANNs in
their experiments, as seen in Figure 3. In earlier studies especially, we found
that 23% of the papers considered solely ANNs as their predictive model.
The higher propensity for researchers to use ANNs in sports domains has also
resulted in the majority of sports papers attributing their highest accuracy
to ANNs (Figure 4). It should also be noted that predictive research on
two sports in this graph (Rugby League and Australian Rules Football) only
considered ANNs as well. Three sports (American Football, Ice Hockey and
Basketball), which found their best accuracy performances with alternative
algorithms, did in fact also experiment with ANNs; however, alternative
algorithms outperformed them.
Therefore, the evidence does not suggest that ANNs have consistently
performed better than other ML algorithms in sporting domains. Indeed, the
broader ML literature, as well as industry-based applied contexts and ML
competitions such as those on Kaggle, do not support blanket statements
which would give primacy to ANNs over all other algorithms.
It is unclear why historically researchers in the sporting domain have
tended to display an overwhelming preference for ANNs. ANNs are not
straightforward to parameterise optimally and often over-fit, especially in

27
Figure 4: Highest recorded accuracies in research by different sport codes covering the
review period. The highest accuracy score is reported with the associated algorithm.
the absence of large enough datasets which tend to be the norm rather than
the exception in sporting domains. A further disadvantage of ANNs is that
they are not easy to interpret, and are therefore less useful to analysts and
team coaches looking to draw out insights, than some other ML models.
Most recently, advanced ANNs such as the LSTM, used in the field of Deep
Learning, have started to enter the sporting domain (Danisik et al., 2018).
Despite low accuracy results from initial research using these algorithms, as
datasets increase in size, the accuracy and utilization of these algorithms in
this field can be expected to grow.
Decision Tree algorithms were the second most popular techniques (Fig-
ure 3) used across all the papers covered in this review. The appeal of these
algorithms is obvious in that they are fast to train and usually do not pos-
sess a cumbersome number of tunable parameters. Importantly, they do
not generate black-box models, but instead embody varying degrees of in-
terpretability depending on their implementation. This property can offer
utility to practitioners beyond just the ability to make predictions, but also
in providing insight to coaches, management and athletes. Their widespread
use has resulted in both American Football and Basketball reporting their

28
highest accuracy results using CART and Logistic Model Trees (LMT) re-
spectively, as the underlying tree induction algorithms. Other variants of
Decision Tree implementations that have been popular in literature include
C4.5 (Quinlan, 1993) and its corresponding J48 implementation in WEKA
(Witten et al., 1999).
Ensemble methods including various Boosting algorithms, as well as Ran-
dom Forests, together form a top-three category of techniques used in re-
search in this domain (Figure 3). Given differing degrees of resilience of this
family of algorithms to over-fitting, it is unsurprising that they have been
liberally used and have registered top accuracy results in Ice Hockey in a
recent study that highlighted the potential of ensemble-based solutions (Gu
et al., 2019).

Figure 5: Distribution of the number of features used for machine learning per sport across
all surveyed papers.
Bayesian-based algorithms such as Naive Bayes and Bayesian Networks
are one of the most popular sets of techniques used in reviewed research.
Though these algorithms are not listed as best performing for individual
sports (Figure 3), they have been found in some comparative studies to
offer better accuracies over alternative algorithms (Miljković et al., 2010;
Odachowski & Grekow, 2012; Joseph et al., 2006). The popularity of Naive
Bayes in particular can be attributed to its common usage as a benchmarking
algorithm when assessing the learnability of a new problem, which is tied to

29
its ability to generate classifiers that do not over-fit. Bayesian networks were
also shown to be useful in incorporating expert knowledge and being effective
with minimal data (Joseph et al., 2006).

4.2. What are some of the defining characteristics that can be drawn from
successful studies?
In early studies in particular, predictive features were often selected man-
ually by researchers based on their knowledge of the given sport(s). More
recently, data-driven methods using various filter-based techniques have be-
come commonplace in successful studies. These have ranged from correlation-
based feature subset selection methods (Hall, 1998), to algorithms such as
ReliefF (Kira & Rendell, 1992), which have a greater contextual awareness
and consider the existence of dependencies between features. Others have
used feature-importance outputs from machine learning algorithms such as
RIPPER (Cohen, 1995) in order to inform which features should be retained
in the training of final models, while some have used ANN-specific methods
including signal-to-noise ratios (Bauer Jr et al., 2000). Furthermore, sporting
experts have also been consulted to select what they consider to be the most
important predictive features.
Whilst generally still modest in size, the general expansion of the pool
of features used across the various sports over time can be seen in the box-
plot of Figure 5. Associated with the growth in the size of feature sets, as
well as their descriptive capacity, the degree to which accuracy has improved
for each sport over time can be observed in Figure 6. The sports that have
seen some of the largest improvements in accuracy over time (Ice Hockey,
Soccer and Basketball) are also the sports which have on average exhibited
the largest expansions in size of the feature subsets.
In addition to the application of feature selection techniques, studies with
a robust experimental process have generally compared a number of different
feature subsets. Such subsets could be betting odds, in-play or match-related
features, external features, player-level features, expert-selected features, fea-
tures extracted from pre-game reports, features constructed from ratings,
among others. The performance of the various subsets of features can then
be compared with each other, as well as with the full set of features.
The distribution of the size of the datasets and their evolution over time
can be seen in Figure 7. It can be observed that generally the datasets have
tended to be of a modest size. Sports that have seen an inflation in dataset
sizes have been Basketball and, more acutely, Soccer.

30
Figure 6: Difference in accuracy percentage points between the lowest and highest predic-
tive results per sport.
It appears however that having a large dataset in terms of number of
matches does not necessarily lead to high accuracy. This was particularly
evident in the 2017 Soccer Prediction Challenge Competition, where model
accuracies were modest despite having access to over 216,000 matches in
their dataset. This particular pattern is depicted in Figure 8, where accuracy
trends over time are rendered for all sports. Arguably, the growth in both
the quantity and quality of features used in each sport (Figure 5) has grown
more than the size of the datasets, and is thus suggestive of the fact that
feature engineering combined with robust feature selection are key drivers of
predictive accuracy improvements.
Successful studies have formulated experimental designs that tested a
number of different training and validation data splits. For instance, using a
certain number of seasons for training a model, while validating the model on
another (future) season. If only one season of data was available, some studies
trained models on a certain number of competition rounds and then validated

31
Figure 7: Distribution of the number of samples in datasets used for machine learning per
sport across all surveyed papers.

Figure 8: Trends over time in accuracy by each sport.

on a future round within the same season. Researchers have often varied the
split of training seasons/rounds to validation seasons/rounds, and compared
the accuracy of models using these different splits. Cross validation (CV) was
found to have been used in a number of prior studies. In the sporting domain
this can be problematic if the technique is not appropriately modified. CV

32
randomly shuffles instances, meaning that future games might be used to
predict past games, and this pitfall was encountered in a number of studies
whose predictive accuracy might have been compromised as a consequence.
Comparison of the performance of models predicting the points margin
(home team points minus away team points), against a classification approach
determining discrete match outcomes (win/loss/draw) was investigated (De-
len et al., 2012; Valero, 2016; Danisik et al., 2018). Both approaches form a
distinctive approach to formulating the ML task. On one hand, Delen et al.
(2012) and Valero (2016) found that the classification approach performed
better, while Danisik et al. (2018) found that a numeric prediction approach
was superior on their dataset. Given these mixed results, we would recom-
mend that, where possible, future researchers perform such a comparison as
part of their experimental design.
Researchers have often proposed to apply their predictive models to other
sports. However we would emphasize that given each sport has specific fea-
tures that are unique to it and are associated with outcomes, it is generally
not possible to directly apply a trained model to a dataset for a different
sport. Rather, it is necessary to go through an new experimental process
on a new dataset for the other sport. It is for this reason that we recom-
mend that, in order to approach prediction problems in a structured way, re-
searchers follow an experimental framework such as the CRISP-DM (Wirth
& Hipp, 2000) framework, the Knowledge Discovery in Databases (KDD)
framework (Fayyad et al., 1996) or the Sport Result Prediction CRISP-
DM (SRP-CRISM-DM) framework (Bunker & Thabtah, 2017). The SRP-
CRISM-DM framework is an extension of CRISP-DM that was developed in
consideration of the characteristics that are specific to the application of ML
to sport result prediction.

4.3. Are the results of some sports inherently more difficult to predict than
in others?
Determining whether predicting outcomes for certain sports is inherently
more challenging than in others is difficult to answer conclusively given the
existing data. One reason is that sports have received highly imbalanced
amounts of attention in the ML literature. This makes it impossible to simply
attribute reasons for lower accuracies in sports like Cricket, Rugby League,
Baseball and Australian Rules Football to something that is intrinsically
more non-deterministic in them, compared to others, when these sports have
received minimal research focus (Figure 2) thus far. However, a little more

33
can arguably be inferred from sports such as Soccer, Basketball, American
Football, Rugby Union and Ice Hockey, which have garnered more attention.
Soccer has received the largest share of research, yet its highest recorded
predictive accuracy is 78% (Figure 4) and it comes fifth with respect to ac-
curacy comparisons across all the sports. Rugby Union on the other hand
has received limited research focus, yet it ranks the highest of all sports re-
viewed here, with an accuracy of 93% (O’Donoghue & Williams, 2004) being
recorded as its best outcome. Interpreting this requires caution. Low scoring
sports tend to embody a higher degree of random chance as a determinant of
outcomes and this in part explains some of the variation. However, this differ-
ential in accuracies can to a large degree be attributed to the characteristics
of the sports and the competitive contexts to which the predictive experi-
ments were applied. For instance, Rugby Union is a much smaller global
sport than Soccer, being played by fewer nations, while historically being
dominated by a handful of them. Understandably then, the best recorded
results for Rugby Union originated from a study O’Donoghue & Williams
(2004) on the 2003 Rugby World Cup, which exemplified a context with a
high degree of predictability5 . The context for Soccer and competitions from
which the results were collected, were diametrically opposite. The results
were drawn from both national and international events, where the depth of
competition was greater, and which ultimately created conditions for which
accurate prediction of outcomes was less deterministic.
Evidence that supports the notion that a great deal of the predictability
of sports is naturally determined by the inherent depth of the competitions
under observation, rather than on the sports themselves, is supported by
Figure 6, which shows a much higher variability in predictive accuracy for
Rugby Union than for Soccer. We can see that in Rugby Union studies that
considered national premiership (Reed & O’Donoghue, 2005) and interna-
tional franchise competitions (McCabe & Trevathan, 2008), which exhibit a
greater degree of equality between teams, the accuracy was correspondingly
much lower at 46% and 74% respectively.
Additionally, Soccer is a low-scoring sport compared to Rugby Union and
this is also a contributing factor to generally lower accuracies in its research.

5
It should be noted that Rugby Union is becoming less predictable as the gap between
low ranked and high ranked nations is narrowing over time, and consequently, more upsets
are occurring. This was particularly evident in the reduced predictability of the 2015
Rugby World Cup compared to previous tournaments (O’Donoghue et al., 2016).

34
In lower scoring sports, there is a larger element of randomness in outcomes
which decreases the performance of predictive models. Associated with this
is the higher likelihood of games ending in draws. Given that draws are
not improbable, many Soccer researchers have formulated the ML task as
a three-class problem (win/loss/draw), rather than a binary-class problem.
As the number of classes increase, the learning problem also becomes more
difficult and thus accuracies tend to reduce.
Ice Hockey on the other hand is a counter example even though a binary-
class approach was used throughout. The sport is relatively low-scoring al-
though higher scoring than Soccer, yet its most recent and best result (92%
accuracy) has demonstrated considerably higher accuracies in comparison, as
well as in comparison to American Football (86% accuracy) and Basketball
(83% accuracy), which are both high-scoring sports. The research from Ice
Hockey, American Football and Basketball, has overwhelmingly been derived
from American national league competitions like the NHL, NFL and NBA,
which generally exhibit high degrees of competitiveness between the various
teams. To some degree this had the effect of controlling for the lopsidedness
in expectations of outcomes that can exist in some sports and global com-
petitions. When examining the trends over time in these three sports, some
potential patterns emerged to help explain Ice Hockey’s unexpected higher
accuracies compared to the other two sports. We see that the best perform-
ing result in Basketball doubled the number of features that were used in
its research over the previous work (74% accuracy, Loeffelholz et al. (2009))
on NBA games. However, the size of the dataset decreased and comparable
ML algorithms were used, with standard Decision Tree based algorithms per-
forming the best out of the suite of methods explored. The best performing
result in American Football saw an 11 percentage point improvement over
the next best result in the NFL (Kahn, 2003), which can be attributed mostly
to a threefold increase in the total number of features used in its research
(28) from the prior work, since both the total size of the dataset remained
essentially the same and the ML algorithms and training procedures did not
represent a considerable increase in innovation. On the other hand, the leap
from 60% accuracy (Weissbock & Inkpen, 2014) to 92% accuracy for Ice
Hockey can be attributed to multiple factors. The authors used three times
as many features, doubled the size of the previous dataset and used more
sophisticated algorithms and training approaches in the form of ensembling.
The points-scoring systems and the manner in which points are attributed
to scoring events differ depending on the sport, and this can affect predictabil-

35
Figure 9: Classification of formal games according to Read & Edwards (1992), retrieved
from figure 3 in Hughes & Bartlett (2002).
ity. For instance, goals in Soccer increment the score by one, and there are
no other possibilities for a team to increment the score by more than this
from any one scoring play. Ice Hockey is similar in this respect. On the other
hand, in Basketball, it is possible to increment the score by different amounts
from different scoring actions - in particular, via three-point throws, in-circle
shots worth two points, and free-throws worth one point. Such differences
in score increments according to the scoring action can also be observed in
Rugby Union, Rugby League, American Football and Australian Rules Foot-
ball. Sports that have different possibilities for increments in score have more
possible permutations in the final match scores by each team, and therefore
final result.
It should be noted that the preceding discussion largely relates to so-
called invasion sports (figure 9), which are characterised by a dependence on
time, and usually consist of time periods such as halves or quarters. Strik-
ing/fielding innings-dependent sports such as Baseball and Cricket have been
more lacking in quality research related to sport result prediction, despite be-
ing very popular sports. Given its popularity and the fact that the highest
predictive accuracy achieved for predicting outcomes is 60%, Cricket presents
a rich ground for further research and accuracy improvements. Baseball could
also benefit from additional research given both the small amount of papers
that have covered it, as well as the relatively low accuracy achieved (as can
be observed in Figure 4).
In summary, though some invasion sports do embody characteristics such
as being low-scoring, which makes outcomes harder to predict accurately
especially if a multi-class formulation is used, the competitive depth of the
types of competitions in which the games take place is also an important
factor. In general, matches within sports, competitions or tournaments that
are highly competitive, are low-scoring, and have less possible increments in
score, and will generally be more difficult to predict. However, the evidence

36
does suggest that these difficulties can to be overcome to a degree when
large as well as rich feature sets are used, in conjunction with datasets that
have a substantive number of samples, together with the use of cutting-edge
training procedures that combine multiple algorithms into robust ensemble-
based solutions.

4.4. What are the best ways to evaluate accuracy results in this application
domain?
Authors in a previous review paper (Haghighat et al., 2013) highlighted
the inherent difficulties in comparing the results of studies in this application
domain. This difficulty arises due to studies usually differing in at least
one of the following dimensions: the sport considered, the input data set, the
model predictors, the class label, and/or the matches, seasons or competitions
that were considered. Part of the challenge lies in the scarcity of available
benchmark datasets. The Open International Soccer Database (Dubitzky
et al., 2019) was compiled recently; however, as was mentioned previously,
this particular dataset is limited because it does not include sport-specific
features derived from events within matches. Sport related datasets are also
becoming increasingly available on websites such as Kaggle.com, which could
produce benchmark datasets in future.
Alternatively, there are a number of other ways for researchers to evaluate
their experimental results, namely, by comparing results to some baseline
measure. Common approaches that have been used in the literature include
comparing the predicted outcomes with:
• Outcomes derived from the match betting odds - the outcome with the
lowest betting odds acts as the class label.6
• Outcomes made by experts on the same matches, e.g., those published
online or in newspapers.
• A rule that always selects the majority class. In most cases this will
predict a home team victory due to the existence of the well-known
home-advantage phenomenon.

6
Betting odds have also been shown to be useful for match result prediction even when
included as the sole model predictor (Tax & Joustra, 2015). However, Hubáček et al.
(2019a) pointed out that if the purpose of the model is to generate profit through betting
strategies, the betting odds should not be included as a model predictor if one wants to
"beat the house".

37
 • A randomly selected outcome.

Given the specific nature of most datasets in sporting domains, compar-

isons such as these are useful for researchers to make when reporting their
experimental results.

4.5. What are the common themes of future research directions that can be
identified from surveyed papers?
Figure 10 depicts the percentage of research themes cited as a future
direction by papers covered in this study. Only research themes that were
cited more than once are shown on the graph. Nearly 90% of papers listed
engineering additional richer features as one of their primary endeavours for
future work. This should come as no great surprise, since generating more
descriptive and therefore discriminatory features with the help of domain
knowledge, is generally accepted as the best strategy for improving predictive
accuracy using ML (Domingos, 2012). The ability to generate more effective
decision boundaries between instances of different classes is to a larger degree
determined by richer features then by algorithms of increasing sophistication.
To that end, domain expertise will play an important role in crafting more
descriptive features. Such domain expertise could be obtained in consultation
with coaches or athletes, or potentially from academic literature related to
the field of sport performance analysis. Much of the research in this field
considers so-called Performance Indicators (PIs), which can be defined as
a selection, or combination of action variables that aims to define some or
all aspects of a performance, which can be used to assess the performance
of an individual, team, or parts of a team, over time (Hughes & Bartlett,
2002). PIs essentially form the core of variables that make up sport result
prediction models, but are also augmented with features external to sport
matches that may have an influence on outcome, e.g., weather, venue, travel,
player availability. Sport performance analysts meanwhile are not usually
concerned with such external variables because they are usually outside the
control of coaches and players. We suggest that there is an opportunity
for knowledge transfer from this field, which is likely to be found in the
identification and development of new features that act as relevant model
predictors. Given that feature engineering has been identified in this review
as an area of highest priority for future directions, an increased collaboration
between these disciplines may result in meaningful advances. However, the
two disciplines do not appear to have reached the point where a significant

38
level of interchange is currently taking place, and researchers are encouraged
to expand their collaborations in this respect.

Figure 10: Percent of future direction research themes cited by all papers cited in this
study.
Experimenting with alternative machine learning algorithms was cited as
a future undertaking by nearly 40% of the surveyed studies. Each ML algo-
rithm embodies within it assumptions about characteristics of the problem
dataset which may or may not hold, and the extent of the disconnect between
the two will also affect the generalizability of different algorithms to varying
degrees. A reasonable course of action when an improvement in accuracy is
sought, is to investigate different families of algorithms, which is the inten-
tion that Figure 10 appears to indicate that researchers are heading towards.
What is more, non-parametric algorithms have a tendency to overfit, and this
is exacerbated on smaller datasets. A quarter of the studies signaled their
pursuit to increase the size of their datasets, which in this instance would be
the correct course of action for studies that have experienced this difficulty.
Meanwhile, improving training methods by reformulating a ML problem
can also have a significant effect on accuracy, which nearly 40% of the studies
intend to explore. The types of training modifications that were cited for
future work have included using different combinations of previous season
results (where appropriate) and giving greater weight to datasets that are
more recent. Some have proposed applying clustering to the datasets before
ML, and using alternative algorithms on different datasets.
Given that each sport has very unique characteristics and potential in-
flection points that can act as markers for winning outcomes, it is surprising
that a quarter of the studies plan to apply their ML methodology, including

39
features, to other sporting codes. This seems rather counter-intuitive, given
that 90% of these papers intend to generate more custom-designed features
that are more tailored for their respective sporting codes.
Furthermore, it was also unanticipated to see only 10% of the studies cite
improvements in feature selection as an area of pursuit for future research.
The problem of over-fitting is amplified when the pool of features is too large
with respect to the size of the datasets. Given that most of the available
datasets for ML in this field do not tend to be vast, coupled with a strong
consensus across all studies to pursue engineering and generation of more
features, the problem of over-fitting is likely to plague many studies unless
efforts in employing feature selection or dimensionality reduction are afforded
equal attention.

4.6. Limitations
In this review we have focused on studies that have made use of at least
one ML technique. There may be other studies from the statistics, operations
research or mathematics literature that may also be relevant or useful for
researchers in this field in improving their experimental approach and results
which were overlooked here. We encourage researchers to also consult the
literature in such fields, which may contain useful insights, particularly if
they are studies which relate to the same sport.
The scope of this review also considered only papers related to team
sports. Papers related to individual sports (some of which were mentioned
in the introduction) may provide useful insights to researchers, particularly
if they are looking at incorporating player-level data into their team sport
prediction model.
Comparisons between accuracies in studies within the same sport (e.g., in
our tabular summaries in section 3) were made despite these studies having
generally used different time-frames/seasons as well as different predictive
features. Given the large variability that exists even between studies covering
the same sport, there is a limit to the reliability of directly comparing and
drawing conclusions from these differences.
While we have discussed match outcome prediction in depth, we have
only briefly touched on how prediction of sporting outcomes can be applied
for the purpose of coming up with profitable betting strategies. Of course,
it is one thing to predict the outcome, but another to come up with a prof-
itable strategy when betting on matches. This includes decisions in terms

40
of whether the betting odd spread is sufficient to warrant the risk of actu-
ally placing a bet. Such considerations should be taken into account if one’s
purpose is to use a predictive model for betting on sport outcomes or events.
Finally, from the perspective of a coach, it is often important how a team
plays, not simply the final match result. For instance, a given team could
win but still play poorly against a weak opposition or, a team could play well
but still lose against a strong opposition. Future work could consider such
alternative labels to predict instead of, or in addition to simply win/loss/draw
outcomes, e.g., good win, average win, poor win, good loss, average loss, poor
loss and so on (as determined by a coach).

5. Conclusions
This paper has provided an in-depth and up to date review of studies that
have applied machine learning techniques for the purpose of predicting results
in team sports. We analyzed various characteristics of past studies including:
the types of algorithms used together with the best performing techniques,
the number of features included, and the total number of instances (matches)
that authors had available in their dataset.
Our findings suggest that just like in any other field, a wide set of can-
didate algorithms and ensembles should be used in experimentation in sport
result prediction, and that the movement away from just employing one al-
gorithm such as ANN, should continue to take place. Engineering rich sets
of features is important, more so than having a large number of instances.
It is worthwhile experimenting with different subsets of features and making
robust comparisons between them and their performance against full feature
supersets. Exploring the comparison of numeric prediction models (predict-
ing points margin), classification models (win/loss/draw), and even labels
based on a coaches’ evaluation of performance (e.g. excellent/average/poor)
warrants further investigation. Given the characteristics of specific sports
and their datasets, it is recommended that experimental process frameworks
are followed (e.g., KDD, CRISP-DM, SRP-CRISP-DM). Since k-fold cross-
validation shuffles instances randomly, it should be altered in a manner that
generates order-preserved folds, so as to prevent future matches being used
to predict past matches, e.g., using a season-by-season or round-by-round
training-validation split. Benchmark datasets for particular sports are still
limited in their availability. However, model predictions can be compared
to predictions from betting odds, expert predictions, and baselines such as

41
a rule that always select a home-side victory, or a randomly selected match
outcome.
Sports and competitions have varying degrees of inherent competitiveness
and thus predictability which are determined by whether they are high or low
scoring, whether there are different score increments for each scoring play as
well as different probabilities of draws occurring.
Lastly, there is an opportunity for knowledge transfer from sport sci-
ence and sport performance analysis to improve sports result prediction,
particularly in the identification of match-related features, while Baseball
and Cricket are popular team sports that could benefit from more quality
research.

Bandulasiri, A. (2008). Predicting the winner in one day international cricket.

Journal of Mathematical Sciences & Mathematics Education, 3 , 6–17.

Bauer Jr, K. W., Alsing, S. G., & Greene, K. A. (2000). Feature screening
using signal-to-noise ratios. Neurocomputing, 31 , 29–44.

Berrar, D., Lopes, P., & Dubitzky, W. (2019). Incorporating domain knowl-
edge in machine learning for soccer outcome prediction. Machine Learning,
108 , 97–126.

Bunker, R. P., & Thabtah, F. (2017). A machine learning framework for

sport result prediction. Applied computing and informatics, .

Buursma, D. (2010). Predicting sports events from past results. In 14th

Twente Student Conference on IT .

Cao, C. (2012). Sports data mining technology used in basketball outcome

prediction, .

Chen, M.-Y., Chen, T.-H., & Lin, S.-H. (2020). Using convolutional neural
networks to forecast sporting event results. In Deep Learning: Concepts
and Architectures (pp. 269–285). Springer.

Cohen, W. W. (1995). Fast effective rule induction. In Machine learning

proceedings 1995 (pp. 115–123). Elsevier.

Constantinou, A. C. (2019). Dolores: a model that predicts football match

outcomes from all over the world. Machine Learning, 108 , 49–75.

42
Constantinou, A. C., & Fenton, N. E. (2013). Determining the level of ability
of football teams by dynamic ratings based on the relative discrepancies
in scores between adversaries. Journal of Quantitative Analysis in Sports,
9 , 37–50.

Constantinou, A. C., Fenton, N. E., & Neil, M. (2012). pi-football: A

bayesian network model for forecasting association football match out-
comes. Knowledge-Based Systems, 36 , 322–339.

Danisik, N., Lacko, P., & Farkas, M. (2018). Football match prediction using
players attributes. In 2018 World Symposium on Digital Intelligence for
Systems and Machines (DISA) (pp. 201–206). IEEE.

David, J. A., Pasteur, R. D., Ahmad, M. S., & Janning, M. C. (2011).

Nfl prediction using committees of artificial neural networks. Journal of
Quantitative Analysis in Sports, 7 .

Davoodi, E., & Khanteymoori, A. R. (2010). Horse racing prediction using

artificial neural networks. Recent Advances in Neural Networks, Fuzzy
Systems & Evolutionary Computing, 2010 , 155–160.

Delen, D., Cogdell, D., & Kasap, N. (2012). A comparative analysis of data
mining methods in predicting ncaa bowl outcomes. International Journal
of Forecasting, 28 , 543–552.

Domingos, P. M. (2012). A few useful things to know about machine learning.

Commun. acm, 55 , 78–87.

Dubitzky, W., Lopes, P., Davis, J., & Berrar, D. (2019). The open inter-
national soccer database for machine learning. Machine Learning, 108 ,
9–28.

Edelmann-Nusser, J., Hohmann, A., & Henneberg, B. (2002). Modeling and

prediction of competitive performance in swimming upon neural networks.
European Journal of Sport Science, 2 , 1–10.

Elo, A. E. (1978). The rating of chessplayers, past and present. Arco Pub.

Epstein, E. S. (1969). A scoring system for probability forecasts of ranked

categories. Journal of Applied Meteorology, 8 , 985–987.

43
Fayyad, U. M., Piatetsky-Shapiro, G., Smyth, P., Uthurusamy, R. et al.
(1996). Advances in knowledge discovery and data mining volume 21.
AAAI press Menlo Park.
Gu, W., Foster, K., Shang, J., & Wei, L. (2019). A game-predicting ex-
pert system using big data and machine learning. Expert Systems with
Applications, 130 , 293–305.
Haghighat, M., Rastegari, H., & Nourafza, N. (2013). A review of data
mining techniques for result prediction in sports. Advances in Computer
Science: an International Journal , 2 , 7–12.
Hall, M. A. (1998). Correlation-based feature subset selection for machine
learning. Thesis submitted in partial fulfillment of the requirements of the
degree of Doctor of Philosophy at the University of Waikato, .
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural
computation, 9 , 1735–1780.
Huang, K.-Y., & Chang, W.-L. (2010). A neural network method for pre-
diction of 2006 world cup football game. In The 2010 international joint
conference on neural networks (IJCNN) (pp. 1–8). IEEE.
Hubáček, O., Šourek, G., & Železnỳ, F. (2019a). Exploiting sports-betting
market using machine learning. International Journal of Forecasting, 35 ,
783–796.
Hubáček, O., Šourek, G., & Železnỳ, F. (2019b). Learning to predict soccer
results from relational data with gradient boosted trees. Machine Learning,
108 , 29–47.
Hucaljuk, J., & Rakipović, A. (2011). Predicting football scores using ma-
chine learning techniques. In 2011 Proceedings of the 34th International
Convention MIPRO (pp. 1623–1627). IEEE.
Hughes, M. D., & Bartlett, R. M. (2002). The use of performance indicators
in performance analysis. Journal of sports sciences, 20 , 739–754.
Ivanković, Z., Racković, M., Markoski, B., Radosav, D., & Ivković, M. (2010).
Analysis of basketball games using neural networks. In 2010 11th Interna-
tional Symposium on Computational Intelligence and Informatics (CINTI)
(pp. 251–256). IEEE.

44
James, B. (1984). The Bill James Baseball Abstract, 1984 . Ballantine Books
New York.

Joseph, A., Fenton, N. E., & Neil, M. (2006). Predicting football results
using bayesian nets and other machine learning techniques. Knowledge-
Based Systems, 19 , 544–553.

Kahn, J. (2003). Neural network prediction of nfl football games. World

Wide Web electronic publication, (pp. 9–15).

Keshtkar Langaroudi, M., & Yamaghani, M. (2019). Sports result prediction

based on machine learning and computational intelligence approaches: A
survey. Journal of Advances in Computer Engineering and Technology, 5 ,
27–36.

Kira, K., & Rendell, L. A. (1992). A practical approach to feature selection.

In Machine Learning Proceedings 1992 (pp. 249–256). Elsevier.

Koseler, K., & Stephan, M. (2017). Machine learning applications in baseball:

A systematic literature review. Applied Artificial Intelligence, 31 , 745–763.

Loeffelholz, B., Bednar, E., & Bauer, K. W. (2009). Predicting nba games
using neural networks. Journal of Quantitative Analysis in Sports, 5 .

Maszczyk, A., Gołaś, A., Pietraszewski, P., Roczniok, R., Zając, A., & Stan-
ula, A. (2014). Application of neural and regression models in sports results
prediction. Procedia-Social and Behavioral Sciences, 117 , 482–487.

McCabe, A., & Trevathan, J. (2008). Artificial intelligence in sports predic-

tion. In Fifth International Conference on Information Technology: New
Generations (itng 2008) (pp. 1194–1197). IEEE.

Miljković, D., Gajić, L., Kovačević, A., & Konjović, Z. (2010). The use
of data mining for basketball matches outcomes prediction. In IEEE 8th
International Symposium on Intelligent Systems and Informatics (pp. 309–
312). IEEE.

Min, B., Kim, J., Choe, C., Eom, H., & McKay, R. B. (2008). A compound
framework for sports results prediction: A football case study. Knowledge-
Based Systems, 21 , 551–562.

45
Odachowski, K., & Grekow, J. (2012). Using bookmaker odds to predict the
final result of football matches. In International Conference on Knowledge-
Based and Intelligent Information and Engineering Systems (pp. 196–205).
Springer.

O’Donoghue, P., & Williams, J. (2004). An evaluation of human and

computer-based predictions of the 2003 rugby union world cup, .

Oliver, D. (2002). Basketball on paper. brassey’s.

O’Donoghue, P., Ball, D., Eustace, J., McFarlan, B., & Nisotaki, M. (2016).
Predictive models of the 2015 rugby world cup: accuracy and application.
International Journal of Computer Science in Sport, 15 , 37–58.

Page, L., Brin, S., Motwani, R., & Winograd, T. (1999). The PageRank
citation ranking: Bringing order to the web.. Technical Report Stanford
InfoLab.

Pathak, N., & Wadhwa, H. (2016). Applications of modern classification

techniques to predict the outcome of odi cricket. Procedia Computer Sci-
ence, 87 , 55–60.

Prasetio, D. (2016). Predicting football match results with logistic regression.

In 2016 International Conference On Advanced Informatics: Concepts,
Theory And Application (ICAICTA) (pp. 1–5). IEEE.

Purucker, M. C. (1996). Neural network quarterbacking. IEEE Potentials,

15 , 9–15.

Quinlan, J. R. (1993). C4. 5: programs for machine learning, .

Read, B., & Edwards, P. (1992). Teaching children to play games. Leeds:
White Line Publishing, .

Reed, D., & O’Donoghue, P. (2005). Development and application of

computer-based prediction methods. International Journal of Performance
Analysis in Sport, 5 , 12–28.

Rotshtein, A. P., Posner, M., & Rakityanskaya, A. (2005). Football predic-

tions based on a fuzzy model with genetic and neural tuning. Cybernetics
and Systems Analysis, 41 , 619–630.

46
Rudrapal, D., Boro, S., Srivastava, J., & Singh, S. (2020). A deep learning
approach to predict football match result. In Computational Intelligence
in Data Mining (pp. 93–99). Springer.

Schumaker, R. P., Solieman, O. K., & Chen, H. (2010). Sports data mining
volume 26. Springer Science & Business Media.

Shi, Z., Moorthy, S., & Zimmermann, A. (2013). Predicting ncaab match
outcomes using ml techniques-some results and lessons learned. In
ECML/PKDD 2013 Workshop on Machine Learning and Data Mining for
Sports Analytics.

Somboonphokkaphan, A., & Phimoltares, S. (2009). Tennis winner predic-

tion based on time-series history with neural modeling. In Proceedings of
the International MultiConference of Engineers and Computer Scientists.
volume 1.

Tax, N., & Joustra, Y. (2015). Predicting the dutch football competition us-
ing public data: A machine learning approach. Transactions on Knowledge
and Data Engineering, 10 , 1–13.

Thabtah, F., Zhang, L., & Abdelhamid, N. (2019). Nba game result predic-
tion using feature analysis and machine learning. Annals of Data Science,
6 , 103–116.

Tsakonas, A., Dounias, G., Shtovba, S., & Vivdyuk, V. (2002). Soft
computing-based result prediction of football games. In The First Inter-
national Conference on Inductive Modelling (ICIM’2002). Lviv, Ukraine.
Citeseer.

Valero, C. S. (2016). Predicting win-loss outcomes in mlb regular season

games–a comparative study using data mining methods. International
Journal of Computer Science in Sport, 15 , 91–112.

Weissbock, J., & Inkpen, D. (2014). Combining textual pre-game reports

and statistical data for predicting success in the national hockey league.
In Canadian Conference on Artificial Intelligence (pp. 251–262). Springer.

Weissbock, J., Viktor, H., & Inkpen, D. (2013). Use of performance metrics to
forecast success in the national hockey league. In MLSA@ PKDD/ECML
(pp. 39–48).

47
 Wirth, R., & Hipp, J. (2000). Crisp-dm: Towards a standard process model
for data mining. In Proceedings of the 4th international conference on the
practical applications of knowledge discovery and data mining (pp. 29–39).
Citeseer.

Wiseman, O. (2016). Using Machine Learning to Predict the Winning Score

of Professional Golf Events on the PGA Tour . Ph.D. thesis Dublin, Na-
tional College of Ireland.

Witten, I. H., Frank, E., Trigg, L. E., Hall, M. A., Holmes, G., & Cunning-
ham, S. J. (1999). Weka: Practical machine learning tools and techniques
with java implementations, .

Zdravevski, E., & Kulakov, A. (2009). System for prediction of the winner
in a sports game. In International Conference on ICT Innovations (pp.
55–63). Springer.

48

View publication stats