CAN ELO RATINGS BE IMPROVED? A CASE STUDY WITH ELITE CHESS PLAYERS

Abstract

Originally designed as a way to reflect past performance, chess ratings are now widely used to reflect players strength with many important aspects in tournament scheduling, advertising and premium shares. The ELO system has been officially adopted by World Chess Federation (FIDE). We used Bayesian analysis of actual data from elite chess players to fit parametric statistical models that could subsidize proposals for rating system improvement. Although most of the considered options are not new, since based on well known preference models, the use of a weighed likelihood function to emulate dynamic rating systems via Bayesian  inference is novel. We compared descriptive ability using marginal likelihood based information criteria. Akaike information criterion was used to compare predictions. Many of the considered options improve on Elo ratings and there is strong evidence that dynamic models considering both white advantage and propensity to draws would result in more accurate systems.