Using Machine Learning to Forecast Football Shot Outcomes

Abstract

Machine learning algorithms are employed on a football data set to forecast whether a shot results in a goal. Results from existing models are improved upon by employing various additional algorithms. At the initial stage, the study gathered data from an existing Kaggle article published online by Usama Waheed. The data set includes data from various leagues and their matches. The research analysed shot events, filtered out non-relevant events, and created features based on shot coordinates, angles, player skill, and shot type. Nine machine learning algorithms were implemented: logistic regression, XGBoost, random forests, support vector machines, k-nearest neighbours, decision trees, LightGBM, CatBoost, and artificial neural networks. The research also focused on removing the redundancy, optimizing performance on imbalanced datasets, and fine-tuning model hyperparameters by employing nested cross-validation. Lastly, the models were evaluated on accuracy, precision, recall, and training time metrics. The results revealed insights into the strengths and weaknesses of each model in predicting goals, with specific emphasis on areas of improvement for shot-based football analytics. Decision trees produced the most accurate predictions on the test set.