WEIGHTED PROBABILITY ENSEMBLE LEARNING FOR DATA-DRIVEN FOOTBALL MATCH OUTCOME PREDICTION

Abstract

Football is the most widely followed sport globally, and it generates copious statistics that can be effectively utilized in data-driven applications. We present an Optimized Weighted Ensemble Model for football outcome prediction by harmoniously combining five individually competitive base learners, LightGBM, Support Vector Machine (SVM), Random Forest, XGBoost, and CatBoost, each contributing its unique modelling prowess to the collective intelligence of the ensemble. The probabilistic outcomes from the ensemble-based learners are aggregated with optimized weights and an adaptively set decision threshold, searched with a grid, to enhance the ensemble decision-making process for final prediction. With the features of possession, shot accuracy, shot conversion, and defense, the proposed method outperformed all the baseline classifiers, reaching 0.952 accuracy and 0.993 AUC. The study indicates the competitive performance of weighted ensembles with careful design and optimization, harnessing the synergy among different learners for robust and high-precision football outcome predictions.