Machine Learning Framework for Football Performance Prediction

This technical report presents a novel machine learning framework designed to forecast football results by analyzing player actions and their contributions to team performance. The research builds on existing models in the football analytics community, addressing challenges in valuing player contributions, particularly in defensive work and off-ball actions. The methodology involves a dataset from the Premier League seasons, utilizing a specifically designed neural network to predict subsequent actions during matches. The model learns to simulate games and estimate outcomes based on player lineups and game states. Results from simulations of the 2017-2018 Premier League season indicate that the model generates reasonably accurate final tables, although it underestimates certain statistics like expected goals per shot. The report also discusses validation methods for the model's predictions, highlighting its strengths and limitations in capturing player interactions and contributions in football analytics.