Machine Learning-Based Strategies to Optimize Salary Cap Allocation in the National Football League

Abstract

American football is the most popular sport in the United States, with each team in the National Football League (NFL) worth an average of $2.57 billion. The ultimate goal of NFL teams in a season is to reach the playoffs and win the Super Bowl. Successfully doing so, however, requires the General Manager (GM) of a team to evaluate countless potential team-building strategies and execute their chosen strategy to near perfection. This thesis investigates the optimal allocation of resources in the free agency and draft in the NFL from the perspective of the GM. The objective is two-fold: 1) to quantify the relative importance of different player positions to team success and 2) to identify the optimal player-selection strategies that GMs should adopt to improve their team. Pro Football Reference's Approximate Value (AV) metric, which measures the value a player contributes to their team during a season, was utilized in order to accomplish both objectives. Clustering techniques were successful in grouping teams into tiers of different performance, while supervised machine learning algorithms - namely linear regression, logistic regression, and support vector machines - yielded highly accurate results in establishing a predictive relationship between team success and positional AVs. The algorithms discovered that GMs should prioritize investing their resources in the offensive line and defensive pass-rushers. Finally, a knapsack problem weighted based on positional importance was formulated to select the optimal 53-man roster in two scenarios, after which the 2018 NFL season was simulated. When filling an initially empty roster, the optimization problem created an essentially undefeatable team. When attempting to improving an existing team that performed poorly during the 2017 season, the selected roster achieved double the win total and established a solid foundation for future success.