A0860
Title: A bivariate extension of the regularized adjusted plus-minus model
Authors: Luca Grassetti - University of Udine (Italy) [presenting]
Valentina Mameli - University of Udine (Italy)
Michele Lambardi di San Miniato - University of Udine (Italy)
Abstract: From the game's results point of view, the predictive power of model-based player performance measures, such as regularized adjusted plus-minus (RAPM), is typically poor. To enhance this feature, one can consider a bivariate model formulation like in the double Poisson specification, typical of soccer game results prediction. First, the home and away team scores are computed separately, and each score is specified by considering the effects of the combination of offensive and defensive non-negative players. Second, the scores can be computed over equally spaced periods. This solution implies that, for each period, the presence of players on the field must be measured by considering their usage percentage. The resulting model specification is equivalent to the model based on dummies, but it eases the bivariate generalization of the model from a computational point of view. The 2022-2023 NBA play-by-play data are analyzed to evaluate the proposal. The results of the empirical analyses show that bivariate RAPM models inherit the advantages of model-based procedures from the player's performance point of view. Moreover, the novel model specification can be used to describe the game's progress and try to predict the results of game periods.
