A machine learning based approach to analyse player performance in T20 cricket internationals

Cricket is one of the most-followed sports around the World. T20 is a short version of the game, growing in popularity over the past 20 years due to high profile tournaments such as the Indian Premier League. There is much demand for analysis of player performance, but traditional measures of this - batting and bowling averages, strike and economy rates - have limitations. We created a novel role-based performance metric using machine learning, allowing comparisons between players with similar roles in different teams. Using ESPN Cricinfo data on T20 international matches, we calculated these new and the traditional performance metrics. Clustering was used to find ‟natural” classes of player types, a Random Forest classifier employed to identify the features most indicative of each cluster then PCA used to obtain the new performance indicator. Finally, we compared the results of our novel approach to the classical player performance metrics and player classifications provided by experts.