ORIGINAL

Abstract

The modern National Basketball Association (NBA) has seen enormous growth in its statistical analysis and research in recent years. A metric which has risen in popularity as a result is live win probability which tracks the estimated chances of either team winning the current game. While there has been academic research on this topic, there are still many unexplored ideas. Specifically, few academic works exist which focus on the relationship between win probability and the box score of an NBA game. Among other issues is the lack of accessible research data for win probability modelling. In this thesis, we aim to fix both of these problems by building a robust play-by-play data set for the 2008-09 to 2018-19 NBA seasons which incorporates running box score information and will be publicly released. We then train various supervised learning methods both to understand the impact of box score statistics on win probability and to see if we can improve upon current win probability modelling implementations. We find that box score statistics have a negligible impact on win probability predictions as the current score difference is the dominant factor in determining win probability. Our best model performs comparably to ESPN's proprietary win probability model and to the market-implied probabilities from live sports betting.