Glicko Rating Deviation Growth While Idle

Mark Glickman's contribution, developed at Harvard and Boston University in the 1990s and given away free on his own website, starts from an obvious observation that Elo's system cannot express: some ratings are much better known than others. A 1750 built from three hundred games over ten years and a 1750 assigned after four games are not the same claim, and treating them identically is how Elo goes wrong. Glicko attaches a second number to every player — the rating deviation, RD — which is the standard deviation of the system's belief about their true strength.

Games narrow it. Time widens it, and this relation is how. The variance of the belief grows linearly with the time since the last game, so the deviation itself grows as the square root, and a player who stops playing gradually reverts to being unknown. The rating itself does not move; only the confidence in it decays. That is exactly right — someone who has not played for eight years has not necessarily got worse, but you certainly know less about them than you did.

The constant c is unit-bound, and that is this page's one honesty problem. In the original papers c is defined per RATING PERIOD, and a rating period is whatever length the operator chose — Glickman's own advice is to pick a span in which the typical active player gets five to ten games, which might be a week for a busy server and a year for a small federation. This page counts the layoff in real days, so the c entered here is the deviation gained per square root of one day. If your c was fitted with a rating period of a week, divide it by the square root of seven before entering it. The number by itself does not say which convention it came from, exactly as with any unit-bound constant, and no calculator can guess.

The real system also caps RD at 350, the deviation of a player about whom nothing whatever is known, and pairs it with a floor of around 30, since no finite number of games makes a rating perfect. This page applies neither, because a capped curve cannot be run backwards and the inverse forms are useful. Read any answer above 350 as "unknown" and any below 30 as "as well known as this system ever gets".

The wider point is worth stating plainly. Glicko does not measure your strength more accurately than Elo does. It measures how much it knows, and it is honest about the answer — which is a different and rarer virtue in a statistical system. And it does not change the fundamental limitation: the rating is still a position within one pool, and the deviation is still uncertainty about that position, not uncertainty about anything transferable.