Comparison of Machine Learning Approaches Applied to Predicting Football Players Performance

Abstract

This thesis investigates three machine learning approaches: Support Vector Machine (SVM), Multi-Layer Perceptron (MLP) and Long Short-Term Memory (LSTM) on predicting the performance of an upcoming match for a football player in the English Premier League. Each approach is applied to two problems: regression and classification. The last four seasons of English Premier League is collected and analyzed. Each approach and problem is tested several times with different hyperparameters in order to find the best performance. We evaluate on five game weeks by picking a lineup for each model that is then measured by its collective score. The results indicate that regression outperforms classification, with LSTM being the best performing model. The score ends up outperforming the average of all managers during the evaluated period in the online football game, Fantasy Premier League. The findings could be used to assist in providing insight from historical data that might be too complex to find for humans.