Deep learning for football match outcome diagnosis using in-match statistics

Abstract

Statistical efficiency determines match outcomes in professional football, yet traditional linear forecasting methods fail to capture the complex, structural interactions of game data. This thesis presents a comparative study of deep learning algorithms, including 1D-Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, using in-match statistical data to analyze the determinants of match results. Using a dataset of over 25,000 matches from the European Soccer Database, the study demonstrates that the 1D-CNN architecture significantly outperforms temporal models, achieving a three-way match outcome classification accuracy (Home Win, Draw, Away Win) of 62.45\% compared to 58.30\% for the LSTM.

Feature importance analysis reveals that efficiency metrics, specifically xG and shot quality, are superior predictors of success compared to volume-based metrics like possession. Furthermore, this research proposes a ``Video-Assisted Feedback Loop,'' a scalable framework designed to translate these algorithmic probabilities into actionable video-based insights for non-elite coaching environments. The findings suggest that high-fidelity optical tracking is not a prerequisite for actionable tactical diagnosis, helping to broaden access to advanced analytics.