FIDE Rating Consultation

[Ian Thompson]

To Ian's query about the 2200 cut off, I presume that it is to do with titles. Given some qualifications start below 2200, I'm surprised it's set so high.

If it is to do with titles it's not a good reason. A better alternative would be to say that a rating qualifying for a title needs to be based on at least X games to count. That used to be the case for the FM title where X was either 25 or 30. (I can't remember which it was.)

[Kevin Thurlow]

"I'm pretty sure the rating team would get a good kicking from a different group if p-ratings were not published."

However, 9 games is not many, and it's an ideal opportunity to teach basic statistics to juniors, not that adults understand the subject much better in many cases...

[NickFaulks]

To Ian's query about the 2200 cut off, I presume that it is to do with titles.

No, those titles have to be based on at least 30 games so it can't be that. I don't recall this being discussed and, like some other more important things in this document, I think it is just wrong.

[Roger de Coverly]

No, those titles have to be based on at least 30 games so it can't be that.

The spread of international ratings is such that it is increasingly unlikely for a player to reach titled standard by playing in OTB national chess not rated by FIDE. Perhaps in the UK you could do it in weekend tournaments or leagues at faster move rates. Although also playimg exclusively online could get a player to a higher standard out of sight of the FIDE rating system. Perhaps it's just a nudge to encourage such players to establish a rating to get their improvement baked in as they approach title qualification.

[NickFaulks]

Perhaps it's just a nudge to encourage such players to establish a rating to get their improvement baked in as they approach title qualification.

There's no point guessing, and I would be surprised if that level of thought were involved. More likely, just a general desire to fiddle. It makes even less sense when combined with the two 1800 draws.

[SeanCoffey]

On the 2200 limit for initial rating: this seems to arise in connection with the proposed change in the calculation of initial ratings (page 12, item (4)). I suppose the idea is that otherwise someone could register a lopsided score against low-rated opposition and end up with an unrealistic rating. With the addition of the two virtual draws against 1800s, that possibility seems a corner case anyway. Otherwise, scoring 5/5 against 1700 opposition would give a rating of 2500.

I'm curious about what Nick Faulks thinks are the important things in the document that are wrong.

I found some of the discussion and justification to be inaccurate. But the simulation results on page 19 seemed to make a strong case for the main changes, if the results are believable. Are they? The match seems almost too good to be true.

[Brian Valentine]

...But the simulation results on page 19 seemed to make a strong case for the main changes, if the results are believable. Are they? The match seems almost too good to be true.

The simulation would be more convincing if it ran forward to the 2021-3 without deteriorating, to compare with the status quo 2021-3. The bottom right quadrant (above 2000, that should not change much)in the first three years 2017-9 is worse than 2008-12 .

[SeanCoffey]

Amongst the many puzzling assertions, I spotted this (page 15):

"With much of the Elo system being a zero-sum game where rating points are merely exchanged between players, there is simply no other significant means by which large amounts of excess rating points, or conversely large deficiencies of rating points, can be introduced into the rating system."

The zero-sum exchange of points applies only when both players have the same K-factor. When an established, high-rated player (with K = 10, say) underperforms against a lower- (and under-)rated newer player (with K = 40, say), the lower-rated player gains four times as many rating points as the higher-rated player loses, thus introducing excess rating points into the system.

[NickFaulks]

The zero-sum exchange of points applies only when both players have the same K-factor. When an established, high-rated player (with K = 10, say) underperforms against a lower- (and under-)rated newer player (with K = 40, say), the lower-rated player gains four times as many rating points as the higher-rated player loses, thus introducing excess rating points into the system.

When rating inflation was the obsession du jour, that was considered very important. Today, not so much. In a few years time, who knows?

[Paul Cooksey]

Has FIDE ever considered lowering the K of an established player playing someone with a higher K? Or adjusting K when there is a big rating difference?

I guess these go some way to fixing the issue where established players want to avoid playing underrated players, and underrated players stay underrated as a result.

[Roger de Coverly]

When an established, high-rated player (with K = 10, say) underperforms against a lower- (and under-)rated newer player (with K = 40, say), the lower-rated player gains four times as many rating points as the higher-rated player loses, thus introducing excess rating points into the system.

I am not convinced at an aggregate level that it has much effect since k=40 players can underperform as well. . After all suppose two players with the same rating meet but one is k=10 and the other k=40. If the k=40 player wins, they gain 20 points and the k=10 player loses 5. However if the k=10 player wins, they gain 5, but the k=40 player loses 20. Also if two k=40 players with the same rating meet, the winner takes away 20 points, but the loser concedes 20.

So k=40 penalises losses heavily as well as rewarding wins.

Here's the recent Kingston International
https://ratings.fide.com/report.phtml?event=310022&t=0 in which overall the rating system lost 47.5 points, the k=40 player finishing last losing 77.6.

[NickFaulks]

I am not convinced at an aggregate level that it has much effect since k=40 players can underperform as well.

The point here is the widespread assumption that "k=40" and "underrated" are equivalent terms. In the wake of the lockdowns there is more truth to that than there used to be but, as was shown in Kingston, it doesn't always go that way..

[Roger de Coverly]

Has FIDE ever considered lowering the K of an established player playing someone with a higher K? Or adjusting K when there is a big rating difference?

FIDE recently introduce a change that higher rated players could only take advantage of the "400 point" rule once a tournament. Beyond that they started using the extremities of the Elo values beyond a rating difference of 400.

It may be worth noting that the ECF inplementation of the Elo method gives some downside protection to k=40 players if they have a bad month by cutting their k back to 20. FIDE could consider the same.

= 40 if age < 18 and New Rating > Old Rating
= 700 / No. of Games This Month if (20 x No. of Games this month) > 700
= 20 otherwise

[SeanCoffey]

When an established, high-rated player (with K = 10, say) underperforms against a lower- (and under-)rated newer player (with K = 40, say), the lower-rated player gains four times as many rating points as the higher-rated player loses, thus introducing excess rating points into the system.

I am not convinced at an aggregate level that it has much effect since k=40 players can underperform as well. . After all suppose two players with the same rating meet but one is k=10 and the other k=40. If the k=40 player wins, they gain 20 points and the k=10 player loses 5. However if the k=10 player wins, they gain 5, but the k=40 player loses 20. Also if two k=40 players with the same rating meet, the winner takes away 20 points, but the loser concedes 20.

So k=40 penalises losses heavily as well as rewarding wins.

Here's the recent Kingston International
https://ratings.fide.com/report.phtml?event=310022&t=0 in which overall the rating system lost 47.5 points, the k=40 player finishing last losing 77.6.

If each player's rating is accurate, it shouldn't make a difference on average if they have different K's. But if one of the players is under-rated, it does. E.g., suppose both players are playing at true strength 2200. One is actually rated 2200 and has K = 10. The other is rated 2000 and also has K = 10. If they play each other, and only each other, they trade points, and eventually, with enough games, they should both have ratings that oscillate around 2100. I.e., both players have deflated ratings compared to their true strength. There is an initial deficit of rating points, and (with identical K's) no new points are created.

Sonas's assertion is that there is "simply" no way that new points can be created. However, with different K's, this is not true. For the first game, the rating system expects player A to score 0.74, whereas his true expected score is 0.5. Player A has an expected loss of 2.4. The rating system expects player B to score 0.26, whereas his true expected score is 0.5. Player B has an expected gain of 9.6. Thus, rating points are created on average. This should mitigate the deflation in the example above.

Not all K = 40 players are under-rated, certainly, and even if they are, they can still hit an unexpectedly bad patch like anyone else.

[Kevin Thurlow]

"If each player's rating is accurate"

There are different measures of accuracy. A player's rating should accurately reflect results. But a rating is not an accurate predictor of results. If it were, there would not be much point in playing. As Sir Richard Clarke (who created the BCF grading system) said, a grading/rating is not a measure of strength.

When Elo did his original work, the lower end was 2200 and the upper end about 2600, and as 2600s rarely played 2200s, a huge disparity in ratings was rare. Now ratings run from 1000 to 2800+, so there's more scope for weirdness. Also, Fide decided to spite the Polgars some years ago and award all other female players 100 points, which must have messed the system up a bit.