Abstract

Quarterback decision-making is one of the most critical determinants of success in the National Football League (NFL), yet traditional statistics fail to account for the spatial, temporal, and contextual factors influencing each play. This thesis proposes a data-driven methodology for evaluating quarterback decisions using NFL player tracking data from the 2022 season. We replicate and extend the DeepQB framework by training deep learning models to estimate expected completion probabilities and expected yards gained for each receiver at the moment of the throw. Our extension applies the approach to a newer dataset with additional contextual features, enabling a more detailed evaluation of both throw execution and the quality of a quarterback's decision within each play's context. We validate and critique these models using calibration and sharpness analysis, revealing that while the predictions are well-calibrated, they often lack sharpness, highlighting the ambiguity of decision-making in the NFL. To visualize model behavior and support interpretability, we also develop a public animation tool that overlays model predictions onto play footage. From model outputs, we derive several quarterback evaluation metrics--including Completion Over Expectation (CPOE), Yards Over Expectation (YOE), Optimal Target Rate, and Decision Efficiency--that isolate decision-making quality from raw outcomes. By comparing individual quarterbacks to the model's "average quarterback" baseline, we uncover meaningful differences in performance not captured by conventional statistics.

Degree

College and Department

Computational, Mathematical, and Physical Sciences; Computer Science

Rights

https://lib.byu.edu/about/copyright/