Pigskin Predictions: Estimating In-Game Win Probabilities for American Football

Abstract

Football is America's most popular sport, with the National Football League (NFL) boasting the highest attendance and TV viewership among all professional sports leagues. Estimating the probability that each team will win a game given its current state is of great interest to many people: not only to coaches and general managers who will do whatever it takes to maximize their team's chances of success, but also to sports bettors, who place millions of dollars' worth of wagers each week, and sportsbooks, who take on large amounts of money in risk. This thesis attempts to determine which factors and models best predict the likelihood that each team wins an NFL football game. We investigate various supervised machine learning techniques such as logistic regression, random forest regression and k-nearest neighbor regression. We find that a time-dependent logistic regression model most accurately predicts win probabilities, generalizing well across all possible game states, and outperforming previous win probability models.