A New Approach to Basketball Game Predictions by Simulating Games Based on the Predicted Numbers and Types of Plays

In this paper, we aim to predict score outcomes for NBA and EuroLeague games from seasons 2000-2020 using simulations. For this purpose, we represent a basketball game as a sequence of plays, as opposed to a sequence of possessions which is standard in most advanced basketball analyses. We first attempt to predict different types of plays for each opposing team using penalized Poisson and penalized negative binomial regressions. These predictions are then used to simulate each game using a compound Poisson process to predict the game outcome. The prediction models give insights about similarities and differences between both leagues. Our simulations predicted game outcomes with an accuracy that is highly competitive to, and often better than, the literature. Our method also provides probability distributions of the number of points each opposing team is likely to score making the predicted outcomes easy to visualize and interpret in comparison with many black–box models.