Accuracy Without Profit: A Statistical Evaluation of Machine Learning Profitability in the English Premier League

This study evaluates the limits of machine learning profitability in the English Premier League betting market (2000–2021). Using a rigorous Walk-Forward Validation approach to prevent look-ahead bias, we tested three standard algorithms (XGBoost, LightGBM, and Random Forest) against the efficient market consensus. While the models achieved statistical distinctness from bookmaker odds (confirmed via Diebold-Mariano tests), they failed to generate consistent risk-adjusted returns. Our analysis isolates two drivers for this failure: (1) ALPHA DECAY, where the predictive edge dissipated significantly post-2015, and (2) CALIBRATION ERROR (ECE \(\approx\) 0.11), where model overconfidence caused standard risk-management strategies like the Kelly Criterion to increase bankruptcy risk rather than wealth. These findings suggest that in mature prediction markets, ”Accuracy” is a misleading metric, and Probability Calibration is the primary barrier to profitability.