Forecasting of player performance is one of the key components in the roster construction of baseball teams. Traditional methods to model seasonal plate appearance outcomes for batters have approached the problem marginally or considered the outcomes to be independent. They also rarely provide estimates of uncertainty. Most state of the art prediction models are the proprietary property of teams or industrial sports entities. This research introduces a mixed effects multinomial-logistic-normal hierarchical Bayes modeling approach to predict multinomial outcome vectors from longitudinal baseball data.
The approach accounts for any correlation between outcomes across player-season and within player-career, and provides a posterior predictive distribution from which uncertainty can be quantified. The purpose of this methodology is to provide a theoretically sound approach to predicting baseball player performance which can be applied to data consisting of players moving between the Japanese (NPB) and American (MLB) major leagues. The problem of accounting for correlations among baseball performance outcomes in prediction is a generally less researched topic.