Modelling the NBA to make better predictions

Abstract

Unexpected events often occur in the world of sports. In my thesis, I present work that models the NBA. My goal was to build a model of the NBA Machine Learning and other statistical tools in order to better make predictions and quantify unexpected events. In my thesis, I first review other quantitative models of the NBA. Second, I present novel features extracted from NBA play-by-play data that I use in building my predictive models. Third, I propose predictive models that use team-level statistics. In the team models, I show that team strength relations might not be transitive in these models. Fourth, I propose predictive models that use player-level statistics. In these player-level models, I demonstrate that taking the context of a play into account is important in making useful prediction. Finally, I analyze the effectiveness of the different models I created, and propose suggestions for future lines of inquiry.