Simulating a basketball match with a homogeneous Markov model and forecasting the outcome

We used a possession-based Markov model to model the progression of a basketball match. The model’s transition matrix was estimated directly from NBA play-by-play data and indirectly from the teams’ summary statistics. We evaluated both this approach and other commonly used forecasting approaches: logit regression of the outcome, a latent strength rating method, and bookmaker odds. We found that the Markov model approach is appropriate for modelling a basketball match and produces forecasts of a quality comparable to that of other statistical approaches, while giving more insight into basketball. Consistent with previous studies, bookmaker odds were the best probabilistic forecasts.