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

net/publication/335230415

A Deep Learning Approach to Predict Football Match Result

Chapter · January 2020

DOI: 10.1007/978-981-13-8676-3_9

CITATIONS READS
16 11,993

4 authors, including:

Dwijen Rudrapal
National Institute of Technology, Agartala
37 PUBLICATIONS 262 CITATIONS

SEE PROFILE

All content following this page was uploaded by Dwijen Rudrapal on 27 January 2021.

The user has requested enhancement of the downloaded file.

 A Deep Learning Approach to Predict Football
Match Result

Dwijen Rudrapal*, Sasank Boro, Jatin Srivastava and Shyamu Singh

National Institute of Technology, Agartala, Tripura-799046, India

dwijen.rudrapal@gmail.com,sambsasank@gmail.com,srijatin1997@gmail.com,
shyamsingh188@gmail.com

Abstract. Predicting a match result is a very challenging task and has

its own features. Automatic prediction of a football match result is ex-
tensively studied in last two decades and provided the probabilities of
outcomes of a scheduled match. In this paper we proposed a deep neural
network based model to automatically predict result of a football match.
The model is trained on selective features and evaluated through experi-
ment results. We compared our proposed approach with the performance
of feature based classical machine learning algorithms. We also reported
the challenges and situations where proposed system could not predict
the outcome of a match.

Keywords: Predictive Modelling, Match result, Machine Learning, Multi-

layer Perceptron, Deep Learning

1 Introduction

Football is the most played and most watched sport in the world. Footballs
governing body, the International Federation of Association Football (FIFA),
estimated that at the turn of the 21st century there were approximately 250
million football players and over 1.3 billion people interested in football 1 . Pop-
ularity of this game created huge number of people’s association. Therefore,
prediction of match result in advance is very attractive to the experts and re-
searchers. But it is very difficult to guess the result of a football match by experts
or past statistics. There are many factors who can influence the match outcome
like, skills, player combination, key players forms, teamwork, home advantage
and many others. Prediction becomes more challenging when match played with
extra time or substitute players or injuries.
Football sports related data is now publicly available and grown interest in
developing intelligent system to forecast the outcomes of matches. In last two
decades, researchers proposed several state-of-the-art methodology to predict
outcome using past and available data. Most of the previous research devise the
prediction problem as a classification problem [1, 13]. The classifier predicts the
result with one class among win, loss or draw class [9].
1
https://www.britannica.com/sports/football-soccer
 In this paper we focus on deep neural network for this problem. Deep neural
network technique has shown to be effective in various classification problems in
many domains [4] including in this domain [5, 7].
The main contribution of this paper are as follows:
– We propose Multi Layer Perceptron based prediction model to predict foot-
ball match result. We experimented our model with classical machine learn-
ing based models and report our evaluation result.
– We propose new set of features to train our predictive model by detail anal-
ysis of features performance and their role on match outcome and from
head-to-head match result.
The remainder of this paper is organized as follows. A summary of previous
promising work on football predictions is presented in Section 2. In Section 3, the
proposed approach including the dataset, is presented followed by experiment
result in section 4. We discuss error analysis of our approach in section 5 and
finally, the conclusion is in Section 6.

2 Related Work
In this section we have discussed promising research work related to predictive
model for football match result. The research for predicting the results of football
matches outcome as started as early 1977 by [11]. proposed a model to predict
the match outcome using matrix of goal scoring distribution. Purucker proposed
[10] an ANN model to predict results in the National Football League (NFL)
using hamming, adaptive resonance theory (ART), self-organizing map (SOM)
and back-propagation (BP). The approach uses 5 features and experimented
on data of 8 round matches in the competition. The work extended in the ap-
proach proposed by Kahn [5] achieve greater accuracy by using large dataset of
208 matches in the 2003 season of National Football League (NFL). The work
includes features like home team and away team indicator indicator. The ap-
proach [8] presents a framework for football match result prediction based on
two components: rules based component and Bayesian network component. The
performance of the system depends on sufficient expert knowledge of past data.
McCabe and Trevathan predict result in their approach [7] for 4 different kind
of sports. Proposed approach targeted Rugby League, Australian Rules football,
Rugby Union and EPL. A multi-layer perceptron model trained using BP and
conjugative-gradient algorithms for predicting task.
The system in the work [3] proposed a prediction model to generate fore-
casts of football match outcome for EPL matches using objective and subjective
information of teams like team strength, team form, psychological impact and
fatigue. Ben Ulmer and Matthew Fernandez used features like game day data
and current team performance in the proposed work [14]. Albina Yezus[15] used
both static and dynamic features in proposed approach. Static features includes
forms of the players and teams, concentration and motivation, while dynamic
features includes goal difference, score difference and history.
 Tax and Joustra proposed a prediction system [12] for Dutch football com-
petition based on past 13 years data to predict the match results. Koopman and
Lit developed a statistical model in their work [6] for predicting football match
results. The work used Bayesian Networks and assumes a bivariate Poisson dis-
tribution with intensity coefficients that change randomly over time. Recently
Bunker and Thabtah [2] investigated machine learning approaches for sport re-
sult prediction and focused on the neural network for prediction problem. The
work identifies the learning methodologies, data sources, evaluation and chal-
lenges proposed a prediction framework using machine learning approach as a
learning strategy.

3 Proposed Methods for Football Match Prediction

3.1 Data Collection
English Premier League (EPL) is the most watched football league in the world
with almost 4.7 billion viewers. Its competitiveness as well as its outcomes are
random in nature. For example, in the season of 2010-11, the distribution of
wins, losses and draws was 35.5%, 35.5% and 29% respectively. For our current
research work, we concentrated on EPL matches. We collected different data like
team statistics and squad, player and result from various online sources 2,3,4,5 .
The dataset includes information of 11,400 matches from the season 2000-01 to
the season 2015-2016. We analyzed what parameters or features are influential
or highly influential in determining the outcome of a match. Detail statistics of
the dataset is reported in table 1.

Table 1. Statistics of dataset

Season count Match count Team count Player count

15 11400 42 462

3.2 Feature Selection

From literature survey of previous promising research work, we have drawn 40
features from our prepared dataset. Features are drawn from different class of
statistics like team, player and head-to-head match statistics. Each match is
classified into two classes, home match and away match. Through our analysis we
found that this classification plays crucial role to predict match result due to the
advantages of venue, crowd support and known environment to the home team.
2
football-data.co.uk
3
fifaindex.com
4
skysports.com
5
kaggle.com
So, all the features are calculated for home as well as away matches. Distribution
of features are reported in table 2. Details explanation of the features are discuss
in following subsections.

Table 2. Features distribution details

Features count
Features class
Home team Away team Total
Team 9 9 18
Player 7 7 14
Head-to-head
4 4 8
match
Total 20 20 40

Team Related Features

1. Attack Point: It describes the attacking ability of a team.
2. Mid Field Point: It describes the mid field ability of a team.
3. Defence Point: It describes the defensive ability of a team.
4. Team Rating: Calculated considering the attack, mid field and defence
point.
5. EPL Times Point: It is the times that a team played EPL during the 18
seasons. If a team played all the 18 seasons they are given 4 points.
6. EPL Streak Point: It describes how consistent a team is playing in EPL.
7. EPL Performance Point: It describes the best performance of a team in
EPL during these 18 seasons. A champion is given maximum of 4 points.
8. EPL Point: It is the average of EPL times Point, EPL Streak Point and
EPL Performance Point.
9. Standings Point: It is the standing point of each team in whichever league
they are. For example point can vary from 1-4 according to their position in
the points table.

Player Related Features

1. Palyer Rating: It describes the overall rating of a player considering pa-
rameters like aggression, tackle, shooting accuracy, dribbling and many more.
This gives an overview of the players performance. 94 was the highest overall
rating of a player in our dataset.
2. Player Potential: Player potential describes the maximum ability of a
player. The player rating can have a maximum value equal to player potential
when the player is at his best form.
3. Market Value: Market Value gives an idea of the form of player, how
dependable he is and how many seasons still left in his career. The maximum
value of this feature is 123 million in our corpus.
4. Goal Keeper Point: It describes the ability of the Goal keeper of each
team.
5. Defence Point: It includes total of 11 defence parameters of a player like
aggression, marking, positioning, sliding tackle, standing tackle. The average
value of these parameters is taken and calculated with respect to 4.
6. Mid Field Point: It includes total of 13 mid field parameters of a player
like acceleration, agility, ball control, crossing, dribbling.
7. Striking Point: It gives an idea of the striking or scoring ability of each
player. Includes total of 19 striking parameters of a player like acceleration,
curve, finishing, long shots, power shot.

Head to Head Match Related Features

1. Team Points: Team point is the average of the points achieved by a home
or away team. Every win conceded 3 points, draw and lost conceded 1 and
0 points respectively.
2. Form Points: It describes the form of a team during head to head matches.
Last 5 home and 5 away matches prior to the match is considered.
3. Goal Difference: It is the difference between goals scored and goals con-
ceded divided by the total number of matches.
4. Form Points Difference: It is the difference between the home form point
and away form point.

3.3 Classifier Selection

In this paper we have proposed Multi-Layer Perceptron (MLP) for football match
result prediction. In earlier similar research work [5] [10] MLP has shown compa-
rable performance. We have also experimented with popular statistical machine
learning algorithms like Support vector machine (SVM), Gaussian Naive Bayes
and Random Forest to evaluate our approach.

4 Experiment and Result

Proposed MLP classifier having 10 hidden states trained on our dataset. We used
holdout method to split our dataset into training set and testing set. So, 70%
of the dataset is tagged as training dataset and rest 30% tagged as test dataset.
In table 3, we report the accuracy as 73.57% on the test dataset. We also report
sensitivity, specificity, precision, recall and F1-score of our proposed approach.

Table 3. Performance evaluation of proposed approach

Parameter Accuracy Sensitivity Specificity F1-score

Score 73.57 71.02 75.48 71.45
 As sensitivity and specificity of our model does not differ so much, therefore
we can conclude that our model has a decent accuracy. We have compared the
accuracy of our model with promising statistical machine learning algorithms in
this domain research like SVM, Gaussian Naive Bayes and Random Forest.

Table 4. Analysis of proposed approach

Accuracy F1-score
Algorithm
(in %) (in %)
MLP 73.57 71.45
SVM 58.77 50.07
Gaussian Naive Bayes 65.84 64.26
Random Forrest 72.92 66.07

5 Error Analysis

In this section we make an in depth analysis of our model and report the situa-
tions where the model failed to predict match outcome.

1. In our current research work, we are considering the features of only the
starting 11 players. But sometimes substitute of a player also determines
very important role in the match. Sometimes, playing 11 may also change in
another match. So, these challenges can be resolved by considering all the
players playing 11 and substitutes.
2. There are many hidden features like crowd support, unfairness of a referee,
unpredictable nature of the game etc. In our current work we could not
include those features. Inclusion of those features will improve our model.
3. As the features which we have considered are dynamic in nature, the value
changes slowly and gradually with the passage of seasons. So, performance
of our approach is mostly dependant on recent past matches.

6 Conclusion

In this work we have achieved a decent accuracy of predicting the outcome

of a football match. Our model works very well considering recent past match
statistics. We made detail analysis of features and drawn effective features to
train our model. We have also shown that match wise features like home and
away match features are effective to the model.
Our future scope of the work is to study the importance of the features like
team’s reserve player strength, coach rating and inclusion of those to make our
model more robust.
 References
1. Abdelhamid, N., Ayesh, A., Thabtah, F., Ahmadi, S., Hadi, W.: Mac: A multiclass
associative classification algorithm. Journal of Information & Knowledge Manage-
ment 11(02), 1250011 (2012)
2. Bunker, R.P., Thabtah, F.: A machine learning framework for sport result predic-
tion. Applied Computing and Informatics (2017)
3. Constantinou, A.C., Fenton, N.E., Neil, M.: pi-football: A bayesian network model
for forecasting association football match outcomes. Knowledge-Based Systems 36,
322–339 (2012)
4. Davoodi, E., Khanteymoori, A.R.: Horse racing prediction using artificial neural
networks. Recent Advances in Neural Networks, Fuzzy Systems & Evolutionary
Computing 2010, 155–160 (2010)
5. Kahn, J.: Neural network prediction of nfl football games. pp. 9–15 (2003)
6. Koopman, S.J., Lit, R.: A dynamic bivariate poisson model for analysing and fore-
casting match results in the english premier league. Journal of the Royal Statistical
Society: Series A (Statistics in Society) 178(1), 167–186 (2015)
7. McCabe, A., Trevathan, J.: Artificial intelligence in sports prediction. In: Informa-
tion Technology: New Generations, 2008. ITNG 2008. Fifth International Confer-
ence on. pp. 1194–1197. IEEE (2008)
8. Min, B., Kim, J., Choe, C., Eom, H., McKay, R.B.: A compound framework for
sports results prediction: A football case study. Knowledge-Based Systems 21(7),
551–562 (2008)
9. Prasetio, D., et al.: Predicting football match results with logistic regression. In:
Advanced Informatics: Concepts, Theory And Application (ICAICTA), 2016 In-
ternational Conference On. pp. 1–5. IEEE (2016)
10. Purucker, M.C.: Neural network quarterbacking. IEEE Potentials 15(3), 9–15 (Aug
1996)
11. Stefani, R.T.: Football and basketball predictions using least squares. IEEE Trans-
actions on systems, man, and cybernetics 7, 117–121 (1977)
12. Tax, N., Joustra, Y.: Predicting the dutch football competition using public data:
A machine learning approach. Transactions on Knowledge and Data Engineering
10(10), 1–13 (2015)
13. Thabtah, F., Hammoud, S., Abdel-Jaber, H.: Parallel associative classification data
mining frameworks based mapreduce. Parallel Processing Letters 25(02), 1550002
(2015)
14. Ulmer, B., Fernandez, M., Peterson, M.: Predicting Soccer Match Results in the
English Premier League. Ph.D. thesis, Doctoral dissertation, Ph. D. dissertation,
Stanford (2013)
15. Yezus, A.: Predicting outcome of soccer matches using machine learning. Saint-
Petersburg University (2014)

View publication stats