How well do Elo-based ratings predict professional tennis matches?

Abstract

This paper examines the performance of five different measures for forecasting men's and women's professional tennis matches. We use data derived from every match played at the 2018 and 2019 Wimbledon tennis championships, the 2019 French Open, the 2019 US Open, and the 2020 Australian Open. We look at the betting odds, the official tennis rankings, the standard Elo ratings, surface-specific Elo ratings, and weighted composites of these ratings, including and excluding the betting odds. The performance indicators used are prediction accuracy, calibration, model discrimination, Brier score, and expected return. We find that the betting odds perform relatively well across these tournaments, while standard Elo (especially for women's tennis) and surface-adjusted Elo (especially for men's tennis) also perform well on a range of indicators. For all but the hard-court surfaces, a forecasting model which incorporates the betting odds tends also to perform well on some indicators. We find that the official ranking system proved to be a relatively poor measure of likely performance compared to betting odds and Elo-related methods. Our results add weight to the case for a wider use of Elo-based approaches within sports forecasting, as well as arguably within the player rankings methodologies.