Simulating a Basketball Game with HDP-Based Models and Forecasting the Outcome

Abstract

We used HDP-based models to model the progression of a basketball game. As known to all, the hidden Markov model can be used for analyzing sequences of the game's content. By introducing Hierarchical Dirichlet Processes on feature extraction and HMMs, we can tackle down the challenges of unknown numbers of mixtures in both models by resorting to nonparametric approach. We employ variational inference for model calculation and cluster the extracted rounds of a basketball match in the form of HMM parameters to forecast the overcome. The proposed scheme is then verified by comparing with other commonly used forecasting approaches: logit regression of the outcome, Naive Bayes method, and Neural Networks. We found that HDP-based models are appropriate for modeling a basketball match and produces more accurate predictions.