Question: Can a fourth-down decision model identify situations where going for it on 4th down (not punting/attemtping a field goal) would increase expected win probability compared to historical coaching decisions?
Why is this question worth answering? I wanted to answer this question because it connects my interest in sports analytics and my experience as an NFL fan. While watching games, I have noticed that teams often choose to punt or attempt a field goal on fourth down, even in situations where the offense appears capable of converting. These decisions frequently feel conservative, especially where a single possession can significantly affect the outcome.
Therefore, I want to explore whether a data-driven model can provide clearer guidance for fourth-down decision making. Specifically, focusing on whether a model can identify situations in which going for it on fourth down would increase a team's chances of winnign compared to a traditional coaching decision. Further, testing whether commonly observed decision patterns align with what historical data suggests is optimal.
Hypothesis: A data-driven 4th down decision model that incorporates game variables (e.g. score differentials, field positions, etc.) will identify situations where going for it on 4th down yields higher expected win probability than historical coaching decisions.
Why/What leads you towards that hypothesis? This hypothesis is motivated by previous observations in football analytics that suggest coaches tend to be a bit more risk-averse; especially, in early-game or short-yardage 4th down situations. With that in mind, a model trained on play-by-play data may reveal systematic patterns where conservative decisions reduce expected win probability. As a result, a model-based policy may recommend more aggressive fourth-down attempts in specific situations where the potential upside outweighs the risk.
The dataset was generated using R by compiling multiple seasons of NFL play-by-play data and preserving detailed game-state information for each play.