International Society for Bayesian Analysis

For many years, the only serious algorithm for tracking player strength over time in organized competitive games such as tournament chess was a system developed by Arpad Elo in the early 1960s. This algorithm was adopted with minor modifications by various national and international gaming organizations throughout and 1960s and 1970s, and is still in widespread use today. Despite its popularity, one major drawback of the Elo system is that its connections to probability models is ad hoc at best, so the resulting ratings from the system are difficult to interpret. In the mid-1990s, I developed an approximate Bayesian algorithm that was a linearization of a reasonably standard Bayesian time series model for binary outcomes. This new algorithm, conventionally known as the Glicko system, overcomes many of the difficulties with the Elo system. In fact, the Elo system can be seen as a special case of the Glicko system in which parameter estimates have no posterior uncertainty. The Glicko system is now used by many gaming organizations, such as the International Wargames Federation, the Italian Othello Federation, and the Australian Chess Federation, along with many online gaming organizations. In this talk, I describe the basic development of the Glicko system, its advantages over the Elo system, and its adoption by various gaming organizations over the years.