In-Game Win Prediction Models for Cricket

The growth of cricket as a global sport cannot be understated. Indeed, in recent years cricket has experienced the most rapid growth rates on many business and popularity measures, both domestic and global. However, cricket as sport is surprisingly understudied and thus presents exciting opportunities for advanced analytics. Building on the dynamic logistic regression framework of Asif and McHale (2016), this paper introduces a improved methodology for in-game win prediction in the context of Indian Premier League (IPL) T20-style cricket matches using ball-by-ball data. After transitioning the model to a Bayesian setting, the work makes two specific advances: integrating historical data via power priors and more transparent and flexible smoothing of the dynamic coefficients via Gaussian processes. To assess the performance of the proposed method relative to that of Asif and McHale, we use cross validation on data from hundreds of IPL matches. Metrics such as the Brier score and accuracy suggest that the proposed method exhibits considerably better predictive performance in addition to the increased optionality afforded to the modeler.