December 2014 FIDE rating list

[MartinCarpenter]

Definitely. For anyone who's got an established grade based on years of playing of that level, there's a strong initial supposition that any big deviation from that is random fluctuations. Huge evidence in favour of that I'd think

Easiest way I can think to get a non trivial disparity between grade and actual playing level at a highish level would be to go from playing lots of the currently grade consuming juniors to very few of them. Nothing the system can do about that though.
(Beyond trying to fix the juniors as has happened.).

[Keith Arkell]

Keith, you want your FIDE rating to be a performance rating, which is understandable given that you have been doing well lately.

Not at all Nick.

Sorry, but that is exactly what you are arguing for in your post. I'm not saying it's daft, but it isn't how the system works.

Imagine a 2500 rated junior who is in reality 2700 strength

I can, but only just. If it was ever possible for such a beast to exist, under the new rules, with k=40 for juniors up to 2300 and then k=20 up to 2400, I don't see how anyone's 2500 rating could be so wrong. The rating system is devised to be as accurate as possible for all players, and cannot be tailored specifically for cases which are somewhere between once-a-decade and mythical.

but equally it would be nice if the rating reflected his performance over a longer time, such as 6 months of active chess

No, it wouldn't. Even over a six month period, if a player's performance varies from his rating, this may be assumed to reflect some combination of change in playing strength and random fluctuation. You wish to throw out the second input, but it has been shown ( and, frankly, is obvious ) that this does not produce the best results.

Nick, I don't understand your point about my wanting a performance rating. You referred to my own case. For a number of reasons, which I won't bore people with here, I am quite a lot stronger than I was a year ago, and at some point in my old age the reverse will no doubt become the case. In both cases I want my playing strength reflected, as quickly as possible in my rating. Who doesn't? When I'm on the slide, aged 70, I want to be able to grab rating prizes

Some of you guys are better qualified at maths than me, and if there really are too many destabilising issues when making k-factor 20, then so be it, but I want to make clear that all I'm arguing for, and have been for 5 years, is a rating system which reflects a players current playing strength as quickly as possible, as accurately as possible. I guess at the moment FIDE are tackling the problem in the case of juniors, but what about, for example, when a player retires and comes back to chess after a 30 year absence?

Incidentally, obviously I cited the extreme example of a 2500 rated junior who was in reality 2700 strength because it demonstrates my calculations more clearly, rather than because I'm idiot enough to think that there are 100s of real life cases like that.

[NickFaulks]

all I'm arguing for, and have been for 5 years, is a rating system which reflects a players current playing strength as quickly as possible, as accurately as possible.

That sounds a wonderful idea, and I would be interested to know who has been arguing against it! The trouble is, the quicker the adjustment, the less accurate it will be, and vice versa. Any system is a compromise. Now that we finally have a sufficient history of good data, it has been found that for top players, k=10 remains a pretty good compromise, and certainly not flawed enough to ram a change down the throats of professionals who don't want it. The only question in my mind is whether 2400 is maybe a bit low for the final cutoff, but we've made some radical changes lately and must give them time to bed in.

[Keith Arkell]

Well, it's clear that nobody is absolutely sure what the ideal k-factors are, hence the reason for trying different values for different rating bands and ages. I suggested earlier something like 10 for +2700, 15 for +2600 and so on. You now say that maybe 2400 is too low for 10, so clearly it is reasonable to debate this. There is no absolutely obvious solution, and my point about 10 being on the sluggish side has to be considered, and presumably has been.

My main point was to highlight the dramatic slowing effect a list a month has on k-factor 10, compared with a list every 6 months. Ovbviously if it is considered that once someone passes 2400 then they are unlikely to change much in strength then just give us all k-factor 1 or zero!

[NickFaulks]

Well, it's clear that nobody is absolutely sure what the ideal k-factors are, hence the reason for trying different values for different rating bands and ages.

No, that's not how it goes at all. The various bands have been introduced because, on the basis of statistical analysis of historical evidence, they are believed to be likely to produce good predictive results. This is not a finger in the air exercise to see what works.

[Keith Arkell]

Well, good luck to you - but I am still entitled to my opinion that a higher k-factor more readily shows current playing strength, and I have recently received a lot of support for this in private messages from some very intelligent people. I will leave this for now, but when FIDE do eventually adopt a minimum of K-factor of 20 I will refrain from saying 'I told you so'.

[NickFaulks]

You are entitled to any opinion, but I still do not know which part of the statistical analysis you believe to be suspect.

If the evidence were to support k=20 at some point in the future, then I hope it would happen. Based on the past eight years, there is no such case. My fear is that, as was proposed ( in fact, passed! ) in Dresden in 2008, it might get pushed through just to add a bit of excitement.

[Keith Arkell]

Well, Nick, I hope it eventually goes through.

I've presented my analysis here as to why I think the change to a list every month, from a list every 6 months requires a minimum of K-factor 20, but if it fails to happen then C'est la vie.

[NickFaulks]

I've presented my analysis

We'll just have to disagree on the meaning of the word "analysis". All I have seen are the arbitrary claims that k=20 would make ratings react more quickly to a string of good or bad results ( which is obviously true ) and that this would make them more accurate ( which isn't ).

[Keith Arkell]

We'll just have to disagree on the meaning of the word "analysis". All I have seen are the arbitrary claims that k=20 would make ratings react more quickly to a string of good or bad results ( which is obviously true ) and that this would make them more accurate ( which isn't ).

We'll just have to disagree on the meaning of the phrase 'string of good or bad results'. For me it is, for example, 70 or a 100 games in a 6 month period. For you it, apparently, an infinite regress that can go on for more than a year.

[Keith Arkell]

My fear is that, as was proposed ( in fact, passed! ) in Dresden in 2008, it might get pushed through just to add a bit of excitement.

My fear is that, despite being correctly passed in Dresden, it will delay being passed because of the influence of the top players, and their desire to hold on to the status quo. My belief, however, is that the common sense of k-20 will prevail in the end.

[Richard Bates]

Hard cases make bad law...

Obviously such decisions should ultimately be justified by solid statistical analysis, although i don't think the views of players should be ignored where it affects their livelihoods. And i agree that their is a danger that changes (upwards) might be made for 'excitement'.

However, on an 'instinctive' argument basis, I think where i would see Keith's argument as flawed is where he is effectively comparing two methods - 6 monthly lists and monthly lists and only considering how they compare at their common intersections - ie. the six monthly point. In other words he is considering two lists both published six monthly, but calculated differently (with i suppose unofficial, unpublished ratings for the monthly methodology).

The six monthly lists had the great danger that significantly 'mis-rated' individuals could play a lot of games, but massively "overshoot" their true level of strength (which could then be replicated in the other direction if they played a lot of games in the next period etc...). However against that they also offered the protection that isolated extreme performance would have less effect on those who's ratings were roughly accurate - the lengthy time between lists meant that the good performances would be balanced by the bad.

Monthly/very frequent lists offer almost completely the opposite danger. On the one hand there is little danger of "overshoot", but greater chance of 'noise' from natural variation in performance distorting the accuracy of ratings. For an obvious example at the top Caruana has shown that it is still possible to gain significant points with K10. K20 he might have been on the verge of 2900! What excitement!

So when comparing K-factors on 6 monthly vs 1 monthly lists leads to the following points on accuracy.
- Individuals with "accurate" ratings (if there is such a thing as an 'accurate' rating) arguably "should" have much lower K-factors on monthly lists compared to 6-monthly lists to control the increased distortions cause by natural variation.
- Individuals who are 'mis-rated' "should" have higher K-factors (although this needs to be given the proviso that they did not play so many games under 6 monthly lists that they were 'over-shooters' and needed a reduction any way).

Keith's argument seems to be based solely on the theoretical cases of the latter individuals (who are far less likely to exist at higher rating levels). Whereas the actual K-factor should be a balance to best align with the actual population of the rating pool (which is presumably why it is higher at lower levels - the influence of the 'mis-rated' individuals is far greater). If concerned about isolated examples at higher ratings it may make more sense to concentrate on other 'fixes' aside from K-factors which have to cater for all players of differing 'true' strength vs their ratings and playing widely varying numbers of games (i dunno, maybe if someone gains/doesn't lose rating points in 6 successive lists including a minimum number of games, then their rating could be 'reset' on the basis of their performance over those games or something)

[Roger de Coverly]

then their rating could be 'reset' on the basis of their performance over those games or something)

I'm unconvinced that the simple expedient of K=40 for juniors will have the effect of eliminating wildly inaccurate ratings. It's well established that players can become GMs in their early teens, so that means a rating of 2500. The rating histories of those that achieve this shows that they entered the system at a relative high initial rating. But all these players were beginners once, so unless their natural talent was that they never played at a 1000 standard no matter how inexperienced, the danger is that they could acquire an initial rating of around 1000. Is the system as constructed capable of adding 1500 points over perhaps five or six years?

[NickFaulks]

Is the system as constructed capable of adding 1500 points over perhaps five or six years?

Yes.

[Christopher Kreuzer]

Nick, do you think there needs to be a way to 'reset' ratings if someone comes in at a low or high grade that is inaccurate and the system is taking a long time to 'correct' that even though a player is getting consistent results over a long period of time?