Fourth Down and Forward: Predictive Modeling and the Future of Football Analytics

Abstract

This thesis explores the analytics behind fourth-down decision-making in football, with the goal of building a data-driven framework that can recommend the best option—go for it, kick, or punt—based on game context. While debates about fourth-down strategy have grown in recent years, most conversations lack a structured model that links outcome probabilities, expected field position, and long-term value into a single decision-making pipeline. This project takes on that challenge, building out a modular system of predictive models designed to work together to simulate and evaluate real fourth-down scenarios. Two modeling pipelines were developed: one based on generalized linear models, and another using XGBoost. Each approach included multiple components: success probability classifiers, expected yard line regressors, expected points calculators, and a decision bot that combines outputs to recommend the optimal play. The generalized linear models provided a solid baseline, particularly in structured scenarios like field goal predictions. The XGBoost models, however, showed stronger performance in complex or nonlinear situations, thanks to more tailored feature engineering and specialized model design, especially the use of separate expected points calculators for first-and-10 versus fourth down. Results showed that many of the bots’ decisions aligned with standard coaching logic, while others revealed opportunities for more aggressive or unconventional play-calling. Still, both modeling pipelines struggled with extreme edge cases, highlighting the need for future models to better capture sharp drop-offs and rare outcomes. This thesis closes by outlining a roadmap toward reinforcement learning as a more dynamic solution—one capable of reasoning across entire drives, adapting to context, and optimizing for long-run reward. By connecting predictive modeling with real-world strategy, this work offers both a proof of concept and a foundation for future systems that aim to bring more clarity, consistency, and context to in-game decision-making.