Smart Sports Predictions via Hybrid Simulation: NBA Case Study

Abstract

Increased data availability has stimulated the interest in studying sports prediction problems via analytical approaches; in particular, with machine learning and simulation. We characterize several models that have been proposed in the literature, all of which suffer from the same drawback: they cannot incorporate rational decision-making and strategies from teams/players effectively. We tackle this issue by proposing hybrid simulation logic that incorporates teams as agents, generalizing the models/methodologies that have been proposed in the past. We perform a case study on the NBA with two goals: i) study the quality of predictions when using only one predictive variable, and ii) study how much historical data should be kept to maximize prediction accuracy. Results indicate that there is an optimal range of data quantity and that studying what data and variables to include is of extreme importance.