An enhanced Bayesian Network prediction model for football matches based on player performance

In sports analytics, existing researches have showed that the Bayesian networks (BN) approach has greatly contributed to predicting football match results with considerably high accuracy as compared to other classical statistical and machine learning approaches. However, existing prediction models rely solely on historical team features including the match statistical data as well as team statistical data, together with the historical features of team achievement such as ranking in FIFA, ranking in league and total number of points gained at the end of a season. There is no known work to date that has analysed individual player performance data as part of the parameters used to predict football match results. To address this gap, this research proposes a BN model for match prediction based on player performance data called the Player Performance (PP) model. To validate the performance of the proposed PP model, three existing prediction models were re-implemented and measured for prediction accuracy. The existing models are the General Individual (GI) model, Match Statistical (MS) model, and Team Statistical (TS) model. All BN models were constructed using the Tree Augmented Naive Bayes (TAN) for structural learning. The dataset used was data for the Arsenal Football Club in the English Premier League (EPL) for seasons 2014-2015 and 2015-2016. Apart from the proposed individual player performance data, the dataset includes individual player rating, absence or presence of players in a match, match statistics, and team statistics. Then, the PP model were re-constructed using other machine learning techniques such as k-Nearest Neighbour (kNN) and Decision Tree (DT) in order to compare with BN for prediction accuracy. The experimental results showed two fold; the proposed PP model using BN achieved a higher accuracy in predicting the outcomes for football matches with an overall average predictive accuracy of 63.76% compare to GI model, MS model and TS model as well as higher than PP model using kNN and DT by 1.64% and 6.02%.

Name dc:type.qualificationname

mphil

Level dc:type.qualificationlevel

masters

Grantor dc:publisher.institution

Universiti Tun Hussein Onn Malaysia

Year dc:date.issued

2017

Author dc:creator

• Razali, Muhammad Nazim

• QA273-280 Probabilities. Mathematical statistics

Language dc:language

en