Explainable defense coverage classification in NFL games using deep neural networks

Machine learning (ML)-powered football analytics has received considerable interest in recent years, with majority of existing analytic measures centered around offense strategies and performances. In contrast, the defensive side of the game has received relatively less attention and development. At the core of understanding and analyzing any defensive strategy is the coverage scheme, i.e., the rules and responsibilities of each defender tasked with stopping the pass. Classifying the coverage scheme for every pass play provides insights and new understanding to the football game to teams, broadcasters and fans alike. The preferences of play callers become apparent through coverage scheme data, such as Bill Belichick using Cover 1 at a top 5 rate in five consecutive seasons. Coverage scheme classification also allows deeper understanding on how respective coaches and teams continuously adjust their strategies based on their opponent’s strengths. For example, the Packers and Chiefs both faced significantly more man coverage through the first 11 weeks of the 2022 season than they did in 2021 after both teams traded away their leading receivers during the offseason (Davante Adams & Tyreek Hill, respectively). Finally, coverage scheme classification enables the development of new defensive-oriented analytics such as uniqueness of coverages. In 2020, Brandon Staley designed the most unique set of coverages for the Rams while the fired Gregg William’s was the least unique.

For the full abstract, select the download button.