Strategy selection and outcome prediction in sport using dynamic learning for stochastic processes

Abstract

Stochastic processes are natural models for the progression of many individual and team sports. Such models have been applied successfully to select strategies and to predict outcomes in the context of games, tournaments and leagues. This information is useful to participants and gamblers, who often need to make decisions while the sports are in progress. In order to apply these models, much of the published research uses parameters estimated from historical data, thereby ignoring the uncertainty of the parameter values and the most relevant information that arises during competition. In this paper, we investigate candidate stochastic processes for familiar sporting applications that include cricket, football and badminton, reviewing existing models and offering some new suggestions. We then consider how to model parameter uncertainty with prior and posterior distributions, how to update these distributions dynamically during competition and how to use these results to make optimal decisions. Finally, we combine these ideas in a case study aimed at predicting the winners of next year's University Boat Race.