Previous MSc Theses

2019 | 2017 | 2016 | 2015 | 2014
2013 | 2012

"Machine Learning Algorithms for Sports' Results Prediction." A. Igea. K. Liaskos. Department of Computer and Information Sciences, University of Strathclyde. 2019. Download PDF (BibTeX) SD

The forecasting of football matches’ results has proved to be a difficult task for which many attempts have been made. The use of traditional statistical models and their results produced until now have been relatively poor. However, better predictive accuracies have been claimed by some authors using machine learning models built with informative features derived from previous matches.
Is it possible to predict the outcome of football-matches using machine learning just like some other authors claim to do? Is it possible to outperform those models?
An ambitious aim of this dissertation is to produce a model able to surpass the accuracies achieved by the sports betting operators. The positivism research paradigm is adopted in this study. There are a myriad of factors influencing the results of football games and this project has defined some of these important features and assessed them. It has been found that using the author’s models, football players’ ratings obtained from the EA Sports video game FIFA (Borjigin, 2019) show a higher forecasting ability than other more ‘sport related’ features derived from previous games such as the number of shots on target, corners, yellow cards, goals, etc.

Using pipelines that combine pre-processing of data, engineering and selection of features, as well as the selection of the best hyper-parameters for several machine learning algorithms, the model designed has been able to outperform the book-makers during the 2011/2012 and 2012/2013 seasons of the German Bundesliga. For the remaining seasons (2010/2011, 2013/2014, 2014/2015), the project’s model was able to obtain an equal performance in the 2013/2014 season and a slightly inferior performance in the 2010/2011 and 2014/2015 to that of the sports betting operators.

The project also discusses that different predictabilities apply to different football leagues and seasons, and that for the season 2015/2016 of the German Bundesliga the author obtains a prediction accuracy of 51%. This performance is slightly lower than the five betting companies studied.