A0768
Title: Hockey is a cruel game: Empirical Bayes woes in predicting productivity of hockey players
Authors: Ivan Mizera - University of Alberta (Canada) [presenting]
Abstract: A prominent domain of illustrations/applications of the potential of empirical Bayes methodology is the prediction of the overall performance of players in sports from rather scant data, achievable by borrowing strength from their peers. The existing instances of such data-analytic ventures, from illuminating examples to comprehensive studies, considered rather binomial responses of hits-and-misses, like bats in baseball or penalties in basketball. The Poisson modelling of point events, also featured in very early empirical Bayes pursuits mainly in the context of actuarial data, was not that much exposure in the sports context; however, the scheme of evaluating productivity in hockey via goals and assists leads exactly to this setting. Relevant parametric and nonparametric approaches of the empirical Bayes methodology therein, adapted to and comparing the specifics of the game of ice hockey to the actuarial situations, are confronted; among other aspects, ramifications like stratification and including covariates are considered. Some preliminary evaluations of the methodology on the data from the NHL season 2018/2019 are presented, and some directions for future reflection and exploration are discussed.