Bivariate Poisson Models in Basketball Match Analysis and Prediction

We propose modifying the conventional Bivariate Poisson model utilised for football match prediction to reflect the characteristics of the basketball framework. Generally, the model specifications rely on bivariate distributions of home and away scored points depending on the opposing teams’ strengths. Due to continuous and unrestricted substitutions, these measures cannot be considered fixed during the match in the present framework. The proposed solution employs Regularised Adjusted Plus-Minus metrics, typically used in evaluating basketball players to define team strengths. The model was formulated on data aggregated by fixed-length rounds. Consequently, the standard specifications of the Regularised Adjusted Plus-Minus model must be adjusted to accommodate a pseudo-design matrix. Players’ contributions across the rounds can be measured as fractions of time spent on the court. This work explores the application of the players’ performance model within this modified framework and the predictive capability of the derived bivariate models.