The September rating list, first published on the 5th, has been revised and republished today (18th). 47 ratings out of 773 have changed with 44 increases (the largest of which was 37) and 3 decreases (the largest of which was 43). The full list of changes is at the end of the article. The reasons for the revision are explained below.

Why the Revision was Necessary

Updates to ratings of foreign players

The main problem was out-of-date ratings of foreign opponents who already existed in the database from previous tournaments. Their existing ICU rating (based mainly on an older FIDE rating) was used instead of their lastest FIDE rating. For established players, whose ratings don't vary much, this made little difference but for juniors with rapidly increasing ratings it meant that their strength was underestimated. Consequently, the players who benefited most from the correction were Irish juniors in competition against foreign juniors, i.e. the Glorney Cup and it's associated tournaments.

Misreported results

The other problem was a simple misreporting of two games from the Rathmines A versus Elm Mount C match in the Branagan Cup. Instead of the correct Burns 0-1 Dunne and Kennedy 1-0 Coyle, the opposite results had been sent in. Consequently, after the correction, Dunne and Kennedy had their ratings adjusted upwards while Burns and Coyle had theirs adjusted in the other direction.

The Curious Case of David Fitzsimons

Looking at the list of changes below, you'll see that there was a third player, David Fitzsimons, whose correction was negative, in addition to the other cases (Burns and Coyle, explained above). Since David was one of the Glorney players expected to have benefited from the FIDE rating corrections, such a large downward correction is counter intuitive at first sight. However, it is correct and the explanation is quite illuminating about how rating systems work.

The following table lists David's last two tournaments in the September list, giving before and after ratings, K-factor, expected and actual scores. Because David's Glorney result was underestimated in the first version of the list, his rating dipped after that tournament and, crucially, went below the 2100 threshold used to calculate the K-factor for his next tournament. Thus David's fabulous performance in the City of Dublin Masters (he came clear first) was further enhanced by a K-factor of 40 (instead of his usual 16).

Compare that with what happened in the revised list. After his opponents' FIDE rating corrections in the Glorney, his ICU rating went up (instead of down) and his K-factor remained at 16. Subsequently, he still gained in the City of Dublin, but far less, and his overall gain was not as great.

Step Functions

Why did a set of relatively small corrections which, for most players, meant either no change or a modest gain, produce such a large change in one isolated case? The answer lies in the way K-factors are calculated with rules of the following form:

if R < 2100 then K = 40 else K = 16

Such rules are easy for humans to comprehend but, mathematically, they are step functions. Unlike other functions used in Elo rating calculations (see [ART::3]), step functions are not smooth and are capable of large changes in their output (in this case K) given only small changes in their input (in this case R). This is what happened to David - a small adjustment to his rating (after the Glorney) made a big difference to what happened in his next tournament.

The current implementation of the ICU rating system is not very revision friendly, given the amount of work involved for the ratings officer. However, in a potential future reimplementation of the system, where revisions may be easier and therefore more common than they are now, it might be worth considering ways to dampen down the effects of sudden changes in K-factor.

Full List of Changes