Using convolutional neural networks to value game positions in professional basketball

Abstract

Many professional sports are currently midway through their own “Data Revolutions”, with advances in analytics being driven by the intense competition for even the most minor advantages, although typically being focused on offensive actions within a game, rather than statistics or metrics pertaining to a player’s defensive abilities. A significant portion of defensive duties is positioning one’s self to prevent opponent actions from occurring at all, and often defensive actions such as tackles or steals are indicative of defensive mistakes rather than an example of good defending. In other areas, Deep Learning models have shown an uncanny ability to identify trends and perform tasks typically considered too complex for software, and better left to humans. Various Deep Learning models have been designed and developed to estimate the value of positions in multiple sports, and shown to be very promising. This thesis attempts to contribute to this by investigating the use of Convolutional Neural Networks to estimate the value of basketball possessions. The model predicts both the terminal actions and the expected value of a possession, and achieves good results relative to the closest published literature in similar tests. Finally, we show how such a model may be utilised to produce metrics describing a player’s positioning ability as a potential tool for player scouting or coaching.