How to extend Elo: a Bayesian perspective

The Elo rating system, originally designed for rating chess players, has since become a popular way to estimate competitors’ time-varying skills in many sports. Though the self-correcting Elo algorithm is simple and intuitive, it lacks a probabilistic justification which can make it hard to extend. In this paper, we present a simple connection between approximate Bayesian posterior mode estimation and Elo. We provide a novel justification of the approximations made by linking Elo to steady-state Kalman filtering. Our second key contribution is to observe that the derivation suggests a straightforward procedure for extending Elo. We use the procedure to derive versions of Elo incorporating margins of victory, correlated skills across different playing surfaces, and differing skills by tournament level in tennis. Combining all these extensions results in the most complete version of Elo presented for the sport yet. We evaluate the derived models on two seasons of men’s professional tennis matches (2018 and 2019). The best-performing model was able to predict matches with higher accuracy than both Elo and Glicko (65.8% compared to 63.7 and 63.5%, respectively) and a higher mean log-likelihood (−0.615 compared to −0.632 and −0.633, respectively), demonstrating the proposed model’s ability to improve predictions.