Rough Guide to new grades

[Mike Gunn]

Following on from Ian's comment I would like to reinforce the point that it is always dangerous to generalise from examples using a small number of players and/or games. (Lighthouse keepers again.) We need to remember that all this stuff is to do with statistics and the results of calculations can be inaccurate when they are based on a small number of data values (game results). Richard Haddrell posted some time ago on this forum about how new players are currently treated, does anyone have the link for that? As I recall the first pass of grading calculation doesn't include ungraded players, but on a second pass their grades are calculated from their results with graded players, and then a further pass is done using all results including those of the (newly graded) new players.

I think some inaccuracy could be introduced if we ever used a game result where one of the new players had a grade based on less than 9 games (but I'm not sure if that is done, or not). Arguably 9 is too low to get accurate results, but isn't the whole grading system constructed on the principle that 9 results is enough?

[Roger de Coverly]

First, its based on the assumption that in the days when graders had to estimate the grades of ungraded players and use them in their calculations, overall they overestimated the true strength of those players. What evidence is there to support this?

The absence of negative grades is one piece of evidence. Did any graders ever assign negative estimates?

Second, if we assume that graders did overestimate the strength of ungraded players in the past, that would have lead to inflation in the grading system while they were doing it. Now they no longer have to do this estimate, there is no longer inflation due to it. That's left us with an inflated grading system.

We are talking about the left hand side of the normal curve - players from "newcomer" to "average". I would expect there to be a number of destabilising forces in operation both inflationary and deflationary. Grading lag is probably a deflationary force in this area because of players improving faster than the 30 game rule and annual grading lists can keep track of. Over-estimation of new players is inflationary. I don't know how these various factors balance out and I don't think the central grading team do either. It remains my belief these issues should have been explored before trying to rebuild the grading system.

[E Michael White]

Richard Haddrell posted some time ago on this forum about how new players are currently treated, does anyone have the link for that?

This was the most popularly anticipated reply shown on the foot of the main index as .. (Most users ever online was 59 on Wed Nov 21, 2007 9:37 pm). If I remember correctly I quizzed RH over the previous 2 months, so if you log on and look at the member list and then query RHs posts around Sep - Nov 2007 you should be able to find the basis.

[E Michael White]

Resulting in an absurdity. Happily the real world is not capable of producing such. What intrigues me is the sort of throught process that would consider such an outcome desirable.

I cant help you with that one, you will have to ask those who set up the system or those who want to keep it unchanged. You seem to fall into the second category so you should explain your reasons yourself. The ECF grading system has very little to do with statistics as its more of an arithmetic scoring system with anomalies and cant usefully be employed predictively.

[Mike Gunn]

Thanks. This is the thread that dealt with how starting grades are found: http://www.ecforum.org.uk/viewtopic.php?f=4&t=140

(Nowhere is it said if there is a minimum number of games for a player's grade to be used for calculating the grades of others.)

[E Michael White]

Elo ratings would only converge the lighthouse keepers to each others' grades if they played few enough games. If they played enough, they'd end up with their grades getting further away from each other.

I am not especially a FIDE expert but are you sure about that Jack ? Wont both grades form alternately converging sequences if they played the same number of games each year ?

[Michele Clack]

I keep a spreadsheet of my grading performance. This year I have run old and new grades in tandem. Under the old grading system I estimate that my new grade would be 99. When I first started playing in the early 1980's, having only ever played at school at beginner level before, my very first grade was 101. Unlike Richard Bates I am confident that I am a far superior player now than I was then. I was so bad when I got that first grade that it was amazing that I ever beat anyone.

When I play now I expect to have a good game with anyone upto about 140. Above that I very rarely achieve even a draw. In the Birmingham League most opponents are no more than say 30 points above me. In the much smaller Worcestershire Leagues I often play people in the 140's, 150's or even 160's. Because of the 40 point rule I will score my grade -10 points when I lose to them. I can't see how this can be anything but deflationary.

It seems to me that this 40 point rule is the cause of a lot of the problem. I accept that rapidly improving juniors are bound to have an effect in a yearly review system. However, there are so few juniors rapidly improving or otherwise in the midlands that the effect here must be small.

[IM Jack Rudd]

Elo ratings would only converge the lighthouse keepers to each others' grades if they played few enough games. If they played enough, they'd end up with their grades getting further away from each other.

I am not especially a FIDE expert but are you sure about that Jack ? Wont both grades form alternately converging sequences if they played the same number of games each year ?

OK, let us use a worked example.

Suppose lighthouse keeper A has a rating of 2400 and lighthouse keeper B has a rating of 2510. Their rating difference is 110, for an expected score of 0.35. They actually have the same playing strength, so A will on average get +0.15 per game. Having little to do for a couple of months, they play each other twice a day, for a total of 120 games in a rating period.

+0.15 over 120 games is +18, and so with their K factors of 10, keeper A will gain 180 rating points and keeper B will lose 180. Their new ratings are 2580 and 2330, 250 apart rather than the previous 100.

(The only way to avoid this is to have an Elo system that updates after every game, which is not possible for practical purposes anywhere except an online server.)

[Paul McKeown]

I cant help you with that one, you will have to ask those who set up the system or those who want to keep it unchanged. You seem to fall into the second category so you should explain your reasons yourself. The ECF grading system has very little to do with statistics as its more of an arithmetic scoring system with anomalies and cant usefully be employed predictively.

Snort. Actually to start with, the ECF system does include the 40 point rule. So that would make your desired system have a range of 10 to 190. I don't you know why you think it should only apply to a multiround APA of all graded players. When and where did RWBC say that???? I agree that no rating system has genuinely deep theoretical groundings, but I'm fairly happy when Professors of Mathematics feel able to attempt predictions on the back of rating systems, including the Clarke system. Actually your postings on thi subject are so bizarre I'm not sure why I'm wasting my time responding to them. Flat earth society stuff.

[Ian Thompson]

When I play now I expect to have a good game with anyone upto about 140. Above that I very rarely achieve even a draw. In the Birmingham League most opponents are no more than say 30 points above me. In the much smaller Worcestershire Leagues I often play people in the 140's, 150's or even 160's. Because of the 40 point rule I will score my grade -10 points when I lose to them. I can't see how this can be anything but deflationary.

Because every time you lose 10, your opponent gains 10, so overall there's no change, so no deflation. What you have got is stretching. The lower graded player's grade gets lower and the higher graded player's grade gets higher without there being any change in either players' standard of play.

It seems to me that this 40 point rule is the cause of a lot of the problem. I accept that rapidly improving juniors are bound to have an effect in a yearly review system. However, there are so few juniors rapidly improving or otherwise in the midlands that the effect here must be small.

In France they have (or at least used to have when I played there regularly 10 years ago) a different '40 point rule'. If the difference was more than 350 Elo points the game wasn't rated if the stronger player won. It was rated if he didn't win. This stops stretching, but pretty well guarantees compression of the grades because there will be the occasional upset when the stronger player doesn't win.

Is there an argument for not grading games, whatever the result, if the difference between the two player's grades exceeds some amount (not necessarily 40). I don't know, but I think there might be.

[Paul McKeown]

(The only way to avoid this is to have an Elo system that updates after every game, which is not possible for practical purposes anywhere except an online server.)

Jack, I thought that FIDE, for example, had changed their Elő system a few years back to upgrade ratings game by game, even if results were delivered per tournament at a later date? Perhaps I'm wrong and it is something I read a while back and misunderstood or have now misremembered or it is something that some delusional spirit at the back of my mind has just invented... but...

Perhaps you could tell me whether I'm right or wrong?

[E Michael White]

Suppose lighthouse keeper A has a rating of 2400 and lighthouse keeper B has a rating of 2510. Their rating difference is 110, for an expected score of 0.35. They actually have the same playing strength, so A will on average get +0.15 per game. Having little to do for a couple of months, they play each other twice a day, for a total of 120 games in a rating period. +0.15 over 120 games is +18, and so with their K factors of 10, keeper A will gain 180 rating points and keeper B will lose 180. Their new ratings are 2580 and 2330, 250 apart rather than the previous 100.

OK - I see what you are getting at now. Is it not correct that after 2 years the difference is more than 400 points which is treated as 400, and causes the difference in later years to be limited. However you are correct it does not converge.

Is there not a FIDE rating rule which specifies a limit on the rating increase from one event ? If these games are treated as one event a different result is then obtained.

[E Michael White]

Actually to start with, the ECF system does include the 40 point rule. So that would make your desired system have a range of 10 to 190.

This would not be my desired system. It would be the most effective system for ranking players but would not be any use predictively. The 40 point rule leads to spread even in an all play all.

.... apply to a multiround APA of all graded players. When and where did RWBC say that????

In the 1960/70s RWBC stated that the system could be used for round robin and unbalanced chess such as league or swiss

'......but I'm fairly happy when Professors of Mathematics feel able to attempt predictions on the back of rating systems, including the Clarke system. Actually your postings on thi subject are so bizarre I'm not sure why I'm wasting my time responding to them. Flat earth society stuff.

Who are your professors, get them to join this forum? My postings are based on a knowledge of the relevant parts of mathematics, markov processes and statistics although I dont think statistics has much to offer in this area other than to explain anomalies and test for deflation/spread.

[Roger de Coverly]

Jack, I thought that FIDE, for example, had changed their Elő system a few years back to upgrade ratings game by game,

Not quite. The Fide site certainly now has individual results, not just event scores. However the base point of the calculation is the rating in the published list at the start of the event. This leads to the anomaly that at the 4NCL May weekend, performances were based on the October 2008 grading list. It also means that Jack's rating overshoot example is possible if you can combine being underrated with playing lots of games and getting good results.

[Alex Holowczak]

The system is simply too slow.

I am carrying games from three years ago in my grade of 90. Last season, I was a match for people graded 120 or so. Three years ago, I was barely competitive with people graded 70. So my grade averages out at about 90, even though if you only graded the last season, I'd be graded 110ish. So, whenever people graded 120 for ages play me, they're getting a roughly equal game, but if they draw, rather than getting equal points, they're getting a substantially lower number of points than they should be getting for the quality of opposition they're meeting. So their grades are getting lower. As more and more juniors go through the system, this will continue. I am aware that juniors get quotas to reduce this effect - but I don't think they're working.

For juniors, who improve rapidly, the current measures are inadequate. There is too much of a rating lag.

Compare the following three people:
A ECF grade ~ 150, online rating ~ 1800
B ECF grade ~ 170, online rating ~ 2000
C Ungraded, online rating ~ 1400

From that information, you'd assume that the ungraded person, who will be graded before the end of July, should be about 100.

I bet that this won't be the case, however. Counting in C's grades will be a score of 0.5/6 in the UKCC Terafinal last August, where the weakest opponent C faced was 75 on the "new" system. A was teaching C for much of the year though, and in February, C scored 2/5 in the Warwickshire Championship U90s, where C beat someone graded 117, drew against a 103, and lost against a 112 and a 108. Ignoring the 40 point rule, that tournament, C would get a grade of ~ 89, which is fair enough. A knows from experience that that is a good indicator of the strength of C, having played C so many times in practice. However, due to counting games from 8 months previously, when C was much worse than currently, C will have a grade which is much lower than it ought to be, if the grade is designed to be a true indicator of strength. Furthermore, C is not a junior, so the opponents of C will not get any grading benefit.

In the above scenario, I am A, and B and C are friends. So I'm going to be very interested to see what the grade of C will be.