Utilization of machine learning methods for predicting the outcomes of football matches

This bachelor’s thesis investigates the use of machine learning to predict football match outcomes, aiming to achieve high predictive accuracy and compare it with bookmakers’ accuracy. A secondary goal was to evaluate the potential profitability of the models. We utilized a novel comprehensive dataset with detailed statistics, including expected goals (xG), to enhance predictive accuracy. The dataset covered seven English Premier League (EPL) seasons, with the 2023/2024 season as the test set. Extensive feature engineering was conducted, resulting in 134 features, of which 52 were selected using recursive feature elimination with XGBoost. Hyperparameters for XGBoost and Random Forest models were optimized through time-based cross-validation and random search. The models achieved an accuracy of 58.4% (Random Forest) and 58.6% (XGBoost) on 370 matches from the 2023/2024 EPL season, closely matching the bookmakers’ accuracy of 58.9%. A significant shortcoming was the models’ inability to predict draws, likely due to class imbalance and the use of accuracy as the evaluation metric for feature selection and hyperparameter tuning. Theoretical profitability was assessed by betting a constant amount based on predictions, but it was not statistically significant at the 5% level. Future research could incorporate player-level features, dive deeper into the problem of draw predictions and explore more sophisticated betting strategies to achieve statistically significant profits.