Analysis on score predicting in limited overs cricket matches

Abstract

The purpose of this research is to analyze the existing methods to predict score in one day international cricket matches and to suggest and implement a machine learning and big data based process to predict scores. Score predicting in cricket matches is a moderately researched and published area. But target score calculation in interrupted cricket matches is heavily researched and practically in use. Since both systems use similar models, the literature review includes target score calculation models as well as score predicting models. Some researchers have tried score predicting using statistical approaches; tools like “winning and scoring predictor” (WASP) are examples for that. But the work related to these tools are not published due to the commercial value of the researches. The literature review sections contain previous work on target score calculation techniques, score predicting models and a section on application of machine learning to similar problems from other domains. The process of preparing a dataset to build a machine learning model is discussed in detail. Match data are scraped from the web and preprocessed to build a master set of features. Then automatic feature selection algorithms are applied on the master dataset to identify the best set of features. Several representations of the same dataset with different feature set combinations are tried on a variety of machine learning algorithms. After going through several iterations, best feature set and the best machine learning model is identified. The scope of this research is limited to score predicting in completed first innings with all 50 overs bowled. As future enhancements, the model can be extended to support all first innings as well as win percentage predictions in the second innings. A fully completed predictive model can be used as a predicting engine in news web sites. Since the research and implementation closely followed target score calculation techniques, the model can also be suggested as an alternative for current target score calculation techniques such as Duckworth Lewis Stern Method.