What will we unlearn next? The implications of Lopez (2020)

Abstract

Lopez (2020) demonstrates clearly how the lack of precise, high-quality data can lead to imprecise results or analyses. In particular, this paper shows that once you know the precise distance to the first down line (“yards to go”) rather than just the integer-valued distances provided in the NFL’s play-by-play data, the decisions made by coaches are more closely in line with what we would expect from rational, data-driven decision-makers in their situation. However, from anNFL team’s perspective, it is unclear if player-tracking data was necessary to help individual coaches in this particular case. Could NFL teams and coaches make approximately the same decisions from a model trained on only play-by-play data, but evaluated in real-time with more precise inputs for yards to go? Fourth-down decisions are typically analyzed with expected points models and/or win probability models (Romer 2006). When making fourth-down decisions, analysts contend that NFL teams should input their current game situation into one of these models (including information such as the down, distance, yard line, score differential, time remaining, etc), and analyze the output. If the model’s computed win probability for a given situation is maximized by “going for it,” the coach should leave the offense on the field; if win probability is maximized by punting, the coach should elect to punt; and if it is maximized by attempting a field goal, the coach should put his field goal unit on the field. Yurko, Horowitz andVentura (2019) provide a detailed explanation of how to build expected points and win probability models, but briefly, the expected points model is a linear model (specifically, a multinomial logistic regression model), and the win probability model is a generalized additive model. Importantly, although only integer-valueddistances (“yards to go”) areprovided in the