Predicted motion pressure-metricizing pressure created by pass rushers in the NFL and predicting their motions using weighted K-nearest neighbors machine learning models

Abstract

A successful pass rush has traditionally only been able to be measured by one of these three outcomes: a sack, a hit, or a hurry-up, which has resulted in pressure being a binary variable. In reality, pass rush is an intricate and rapid part of American football, which is why a more precise metric to evaluate pressure is desired, consequently allowing for more in-depth analysis of both players' and teams' performances, as not only the occurrence, but also the amount of pressure created during a play is of interest and can be vital for performance analytics. In this paper, a weighted k-nearest neighbors (wKNN) machine learning model is used to produce such a metric, returning a percentage of pressure created for every pass rusher at any given moment during a play, and is able to predict the binary occurrence of pressure on a play with over 91% accuracy. Additionally, this wKNN is also used to predict the motion of the pass rusher. The pressure created by the predicted motion is then directly compared with the true pressure, allowing for a concrete analysis of a pass rusher's decision-making compared to the league's average.