A comparative evaluation of Elo ratings- and machine learning-based methods for tennis match result prediction

Abstract

Elo ratings-based methods, including the recently proposed Weighted Elo method, have been found to perform well when forecasting tennis match results, however, whether they can outperform machine learning (ML) has not been established. In this study, a comparative evaluation of the two types of methods is conducted using the Sports Result Prediction CRISP-DM experimental framework. The first full year of mens ATP tennis data (2006), in a dataset containing matches from 2005 to 2020, was set to be the initial training set and 1 year of data was incrementally added to this set to predict 14 test years, from 2007 to 2020. Features were ranked based on their average rank across five feature selection techniques. It was found that, of the five ML models, Alternating Decision Trees (ADTrees) and Logistic Regression achieved higher accuracies than Elo ratings and similar accuracies to predictions derived from betting odds. Furthermore, ADTrees show potential in this domain, with solid performance achieved with an interpretable decision tree that allows for variation in the average betting odds difference threshold.