A COMPARATIVE STUDY OF FOOTBALL MATCH OUTCOME PREDICTIONS USING DEEP LEARNING BY COMBINING STATISTICAL DATA AND PLAYER HEATMAPS

Abstract

The objective of this research was to compare the performance of MLP, CNN, and Multimodal Neural Network (MMNN) models in predicting the outcomes of English Premier League football matches. The study used statistical data, player heatmap data, and multimodal data integration through Early Fusion, Late Fusion, and heatmap conversion into structured numerical data based on field segmentation. The results revealed that the CNN model using statistical data combined with heatmap-based positional data divided into four field zones achieved the highest test accuracy at 63.50%, and consistently outperformed other models in terms of precision, recall, and F1 score. Furthermore, feature importance was analyzed using the SHAP (SHapley Additive exPlanations) technique, which indicated that defensive features of the home team, such as blocks, clearances, and aerial duel won, had the greatest influence on model prediction. In contrast, the positional features of the away team were found to have the least impact. These findings can be applied to the development of tactical analytics systems to support strategic decision-making and match preparation for football teams.