Using Machine Learning to Optimize NBA Lineups

Abstract

While research into lineups is a growing field within basketball analytics, considerations of the opposing team’s strategy are virtually nonexistent. This paper aims to correct for that, analyzing substitution decisions by taking a more realistic and two-sided view of how lineups interact. Using 13 seasons worth of NBA play-by-play data from 2003-2017, this paper uses machine learning to select the optimal lineup given a team’s roster and the opposing players. Taking both a regression approach and a clustering approach, the aim is to start by producing accurate predictions based on lineup matchups, and then to use that model to inform lineup recommendations. As such, the evaluation of these final recommendations will be based on an evaluation of the component prediction models. Finally, an interesting use case is explored in order to evaluate coaches by the percentage of time that their lineup decisions match the suggestions made by the model.