A Supervised Learning Based Approach for Predicting the Final Score in Limited Overs Cricket Matches

Abstract

Cricket is a popular game in many countries across the world and the most popular sport in Sri Lanka. Knowing the outcome of a cricket match well before it happens has been valued by many parties involved with the game. In this study we propose a supervised learning based approach for making ball-by-ball predictions on the final score in fifty-over cricket matches. We propose an ensemble model of random forest regression to be applied for an inning segmented based on the resources remaining. Segment-wise modeling approach we followed made the overall model capable of adapting well to the different stages of an inning by locally optimizing its parameters. Our model out-performed the existing methodologies used in practice such as Duckworth-Lewis method and the run-rate method, which are also capable of making ball-by-ball predictions on the final score. We also make ball-by-ball predictions on the winner of one-day international (ODI) matches during the second inning using a random-forest classifier. Incorporation of the prediction results from our first model as an input to the classifier proved beneficial in improving the winner prediction performance. Empirical results show that our model performs significantly better compared to the state of the art. Our model seemed unbiased regardless of the amount of target to be chased.