Predicting the outcomes of football matches

Abstract

Football, one of the world’s most popular sports, is regarded as a complex and dynamic game between two teams. Predicting the outcome of football matches is not an easy thing to do. Machine learning algorithms are able to build precise and accurate predicting models. Nowadays, machine learning is being used in many areas such as medicine and finance. Lately, the increasing ease of data availability, computer power and popularity of machine learning techniques gave rise to more sports analysts. The aim of this dissertation is to predict the full time match result (Home Win, Away Win or Draw) whilst using both match statistics and the line up strength as the main predictors. We utilise six machine learning techniques being, Artificial Neural Network, Support Vector Machine, Random Forest, Decision Tree, Logistic Regression and k-Nearest Neighbours. The machine learning models are trained and tested on the top five European Leagues which are the English Premier League, German Bundesliga, Spanish La Liga, Italian Serie A and French Ligue 1. We consider matches played from the 2016/2017 season to the 2020/2021 season. The necessary data, such as match statistics and player line up names, is collected through a custom built web-scraper. To derive the line up strength of the team we make use of the video game series FIFA. A base model which always predicts a home win was defined and obtained an average accuracy score of 0.450 across all the five leagues. The first predictions obtained from the machine learning models utilised the match statistics as the main predictors. The best accuracy score across all leagues was obtained by the Logistic Regression which recorded an accuracy score of 0.606. The next set of experiments involved utilising the strength of the line up of both teams as the main predictors. We use four player statistics from the FIFA video game, which are overall rating, potential rating, market value and wage bill as the main predictors. With these predictors, the models recorded a mean accuracy score of 0.486 across the five leagues. Finally, we combine both set of predictors to analyse whether the predictive performance of the machine learning algorithms would improve. Results show that the models achieved a mean accuracy score of 0.602. Hence, we concluded that the addition of line up strength predictors did not add value to that the machine learning techniques could learn from.