Expected points (EP) and win probability (WP) are value functions fundamental to strategic in-game decision making in American football, particularly for fourth down decision making. The EP and WP functions which are widely used today are statistical models fit from historical data. These models, however, are subject to serious statistical flaws: selection bias, overfitting, ignoring autocorrelation, and ignoring uncertainty quantification. We develop a machine learning framework that accounts for these issues and extracts our analysis into a decision-making inference. Along the way, we introduce a novel methodological approach to mitigate overfitting in machine learning models. Specifically, we extend the catalytic prior, initially developed in the context of linear models, to smooth our tree machine learning models. Our final product is a major advance in fourth-down strategic decision making: far fewer fourth-down decisions are as obvious as analysts claim.

翻译：暂无翻译