Scoring with Data: An Ensemble of Machine Learning Models for Premier League Match Predictions

Predicting the outcomes of football matches remains a complex task due to the numerous dynamic and unpredictable factors that can influence the results. This study proposes a datadriven approach to classify English Premier League matches as wins, draws, or losses using machine learning and deep learning techniques applied to data from the last eight seasons. A web scraping tool was developed to systematically collect relevant match statistics and team information. The performance of Random Forest, XGBoost, and TabNet models was evaluated, along with an ensemble model that combines their complementary strengths. Results show that the ensemble model achieves higher predictive accuracy, especially when recent team performance metrics are included. A feature importance analysis highlights variables such as recent form, expected goals, and possession as critical for accurate prediction. Lastly, the ensemble model is benchmarked against external sources, including an AIbased predictor and a professional betting house, providing a comparative assessment of its practical applicability.