News

Ratings Central Update

Published Tue 04 Jun 2019

From the Ratings Central Website:

On June 1, 2019, we modified the temporal-update algorithm and reprocessed all events. Everyone’s rating changed. All historical ratings changed.

Previously, if the rating system had not seen a player for a year, the standard deviation of the player’s law increased by at most 70 points.

With the new algorithm, if the rating system has not seen a player for a year, the standard deviation of the player’s law will increase by at most 79.4 points and the mean will increase by 7 points. However, the increase in mean will not be credited until the player plays in an event. (The new temporal-update is the sum of a mean-zero normal random walk with a variance of 70² per year and a Poisson jump process with jumps of 200 points, a mean of 7 points per year, and a variance of 200 × 7 per year.)

With the new model, players sometimes get better in jumps, while before they only got better in steps.

When we reprocessed all events, the ratings of 99.99% of the players increased or stayed the same. The average gain was 42 points. Only 0.02% of players gained more than 300 points.

We were motivated to consider improvements to the algorithm by a major event director who felt that the ratings for their players were deflating. After a year of experimentation and investigation, we concluded that part of the deflation was due to their choice of priors. However, even after using better priors, there still appeared to be some deflation. A related problem is that sometimes the system can be slow to react when a player improves rapidly. We think the new temporal-update algorithm better models how player playing strengths actually change and produces more accurate ratings.


Gallery