Predicting the winning team in basketball: A novel approach

Abstract

Predicting the winner of a basketball game is a difficult task, due to the inherent complexity of team sports. All 10 players on the court interact with each other and this intricate web of relationships makes the prediction task difficult, especially if the prediction model aims to account for how different players amplify or inhibit other players. Building our approach on complex systems and prototype heuristics, we identify player types through clustering and use cluster memberships to train prediction models. We achieve a prediction accuracy of ∼76% over a period of five NBA seasons and a prediction accuracy of ∼71% over a season not used for model training. Our best models outperform human experts on prediction accuracy. Our research contributes to the literature by showing that player stereotypes extracted from individual statistics are a valid approach to predict game winners.