A Framework for Accessing Individual Defensive Performance in Coverage

1. Introduction

Current defensive coverage metrics are only calculated on plays where that defender’s assigned offensive player is the one targeted. This approach is flawed and incomplete. On any given play, there are around five available receivers, and the best coverage may prevent the quarterback from even throwing the ball to a given receiver. Further, these metrics identify coverage assignment simply by considering the responsible defender to be the nearest one at the time the ball arrives. Thus, this methodology only gives us coverage assignment for one player per play, and even that assignment is questionable; the closest defender at one particular moment isn’t necessarily the assigned defender.

The work presented here illustrates a framework we’ve developed to address these shortcomings. Our approach begins by identifying the responsibilities of each defender on a given play. If a player is in man (everywhere) coverage, he is assigned a specific offensive player whom he is to cover until the ball is thrown. On the other hand, if a player is in zone coverage, he is instead responsible for an area and any receivers who enter that area. Because the responsibilities of a player differ in each coverage type, the evaluation of performance should be different for each. Once we identify the type of coverage being played by each defender, we then consider two separate time points for evaluation. For one, we consider the time frame after the pass is thrown. This is the most straightforward but limits us to only the targeted receivers, as players tend to abandon their assignments and converge to the ball once it is thrown. To include all defenders in our analysis, we also consider the time frame between the snap and the pass. Figure 1 below illustrates our proposed framework…….

Read our full paper on Kagge here: https://www.kaggle.com/model284/a-defensive-player-coverage-evaluation-framework

The following table contains our spread and total predictions for each NFL game. We will update this post with new predictions/matchups each week and keep track of how our models perform against the closing Vegas lines for each game. Our models are built using a combination of variables, including our Elo Ratings, Vegas spreads/totals, situational factors, team statistics (from a team’s last 7 games played), historical QB rating of the starting QBs, and more. Our predictions are generated as a composite (i.e., the average of multiple models). We are working on a more detailed write up on our modeling methodology that we will eventually link in here.
Continue reading 2018 NFL Game Predictions →