3 Winner prediction in cricket using machine learning

Abstract

The paper performs analysis of the Duckworth-Lewis (D/L) method [Duckworth and Lewis, JORS 1998;49(3):220–227] for One Day International (ODI) cricket matches. The paper compares the accuracy of the D/L method against various classification models (using machine learning techniques) to predict the outcome of a cricket match which is somehow interrupted during second innings of the match. We collected data for 3,450 ODIs and the runs scored by the chasing team after every over in those matches from CricInfo.com. We then applied D/L formula as well as our proposed learned classifiers on different overs of second innings of the match and found out that our classifiers were more accurate in predicting the outcome of the match at different overs as shown in Table 1. The paper also aims to optimise the D/L resource table [ICC-Cricket. com] (used in D/L formula) using Particle Swarm Optimisation (PSO) and compared its accuracy with the modified resource table. Finally, the paper develops an Unpredictability Index of cricket playing nations showing how unpredictable the teams are, based on their previous performances. The results show that developed classification models have better accuracy than D/L method to predict the match winner. The considerable achievement here is that the classification models predict the match result almost with the same accuracy at 10 overs that the D/L formula predicts after 20 overs have been bowled. This can be helpful if a match is abandoned before 20 overs and a winner needs to be decided (say after 10 overs). This early prediction can also be helpful when considering the betting aspect. The modified D/L resource table also yields better accuracy than the original D/L resource table. Finally, our unpredictability index shows interesting patterns that have often been pointed out by experienced critics and commentators of international cricket.