ECF grades compared with FIDE ratings

[Alex Holowczak]

Would Alex care to publish his fit of the juniors only comparison to show whether or not it is significantly different from the non-juniors?

I got y = 6.97x + 703 for juniors only.

The R^2 factor dropped from 0.84 for adults to 0.79 for juniors.

The thing that can't be ignored is the link between ECF and FIDE that says the conversion must be y = 8x + n, where n is a constant to be determined.

[IM Jack Rudd]

The thing that can't be ignored is the link between ECF and FIDE that says the conversion must be y = 8x + n, where n is a constant to be determined.

I make it y = 7.72x + n, unless you're reading the tables here differently from me.

[Alex Holowczak]

The thing that can't be ignored is the link between ECF and FIDE that says the conversion must be y = 8x + n, where n is a constant to be determined.

I make it y = 7.72x + n, unless you're reading the tables here differently from me.

Sure, but if we're going for a whole number, then 8 is nearer than 7.

[Richard Bates]

The thing that can't be ignored is the link between ECF and FIDE that says the conversion must be y = 8x + n, where n is a constant to be determined.

I make it y = 7.72x + n, unless you're reading the tables here differently from me.

Sure, but if we're going for a whole number, then 8 is nearer than 7.

The 8 has always, logically, seemed a bit high seeing as the number 7 seems to play a more important role in calculating FIDE ratings (generally up unto about 100 pt difference each 0.01 in expected score is banded in groups of 7 with the odd exception). The correlation with expected score with ECF grades doesn't touch 8 until the difference reached 200pts, which is quite a large gap since most ECF graded chess is structured in such a way as to encourage playing opponents reasonably close in strength.

[Alex Holowczak]

The 8 has always, logically, seemed a bit high seeing as the number 7 seems to play a more important role in calculating FIDE ratings (generally up unto about 100 pt difference each 0.01 in expected score is banded in groups of 7 with the odd exception). The correlation with expected score with ECF grades doesn't touch 8 until the difference reached 200pts, which is quite a large gap since most ECF graded chess is structured in such a way as to encourage playing opponents reasonably close in strength.

I had completely overlooked this point. To me, this is the explanation of why we're seeing y = 7x + n, rather than y = 8x + n.

[Roger de Coverly]

The correlation with expected score with ECF grades doesn't touch 8 until the difference reached 200pts, which is quite a large gap since most ECF graded chess is structured in such a way as to encourage playing opponents reasonably close in strength.

I suppose you can argue that, historically at least, much FIDE chess was in large tournaments with seeded pairings, which as we know are liable to generate pairings between players 200 points or more apart. 4NCL could be an exception, but there can still be wide variances in the ratings of players, particularly in the top division.

In the revaluation exercise, the grading team made much of the issue that on ECF grades, players weren't making their expected scores and used this as an excuse to add 20 to 25 points to the middle of the grade distribution. Does non-linearity come into it. We're only guessing really, that the curve of expected results should be nearly straight (Elo) or exactly straight (ECF). So if player A can score 60% against a field 10 ECF points lower and B in that field can score 60% against a field 10 points below that, should you be able to deduce that A can score 70% against a field 20 points lower?

If anyone had the time and patience to set it up, you could run English grading on a historic reconstruct using Elo rules (copying the FIDE rules would be a start point) and see what the outcomes were. The idea being to eliminate data as a source of variance and just concentrate on how the rules reacted to the same set of results. At the start of the period, you would convert existing players using any arbitrary conversion formula. You might want an 8 in it, but you might try 7 as well. From then on, you apply the Elo expected score rules, but only bring in new players once they had scored at least 1/9 against existing rated opposition. How variant will the outcomes be at the end?

[Alex Holowczak]

Does non-linearity come into it. We're only guessing really, that the curve of expected results should be nearly straight (Elo) or exactly straight (ECF)?

I don't think that's a guess; I think that's right. We're using a linear approximation of a curve, so there will be the biggest errors for finding a conversion is where the gaps between the line and the curve are largest.

[Roger de Coverly]

I don't think that's a guess; I think that's right. We're using a linear approximation of a curve, so there will be the biggest errors for finding a conversion is where the gaps between the line and the curve are largest.

I think the point is that Elo and Clarke chose a near linear and linear formula respectively without having the detail results to test how well it modelled reality. The grading team applied a linear performance model to the ECF data and by virtue that players seemed to under perform against the expectation, inferred that the grades needed to be closer together. They could equally, in my view, have inferred that the results they got were because the underlying expectations were, or had become non-linear.

[Roger de Coverly]

There's a new paper on this at http://www.englishchess.org.uk/wp-conte ... 201401.pdf

The headline from the paper is a proposal (amongst others) that the conversion factor should be Elo = 7.5 * ECF + 700.

The 7.5 reflects the slightly non-linear nature of the International Elo when working out score expectations.


Some comparisons using this table

ECF FIDE
80 1200
100 1450
120 1600
140 1750
160 1900
180 2050
200 2200 ( that matches the traditional Elo = 8 * ECF + 600)
220 2350
240 2500
260 2650
280 2800

There is practical advantage to using a simple, if slightly approximate formula. It gives a table that can easily be remembered.

There remains the suspicion that deflation is now a feature of the International rating scale, in the 1800 to 2200 range anyway. Whenever e2e4 run an under 2050 tournament, it can attract a number of players usually only eligible for Opens.

[Christopher Kreuzer]

Well, those of use with FIDE ratings will all soon discover if our ECF gradings are getting closer to our FIDE ratings or not... (PS. Roger, when you say "Whenever e2e4 run an under 2050 tournament, it can attract a number of players usually only eligible for Opens", there was the tournament you won, what other examples are there?)

[MartinCarpenter]

Maybe quite a few - to be honest it would be amazing if the known, rather substantial, lagging junior FIDE grade effect wasn't having at least some deflationary effect around that level. It could easily be fairly large.

Even just random ECF/FIDE differences (quite large in general for obvious reasons) would create quite an overlap with the bottom ranges basically matches. (most opens >175 now iirc?).

[Jonathan Bryant]

Maybe quite a few - to be honest it would be amazing if the known, rather substantial, lagging junior FIDE grade effect wasn't having at least some deflationary effect around that level. It could easily be fairly large.

Even just random ECF/FIDE differences (quite large in general for obvious reasons) would create quite an overlap with the bottom ranges basically matches. (most opens >175 now iirc?).

Certainly there are players graded over 200 ECF who can get into Adam Raoof's under 2200 Hampstead/Golders Green tournaments.


Incidentally, I don't think that the 'lagging junior grade effect' is the only reason for deflation. I played at Golders Green this Saturday and one of my two wins was against an unrated opponent - and therefore it won't count for rating.

Of my most recent tournaments:-
Hampstead November 2013: 3 of 4 games counted, the fourth being a win for me
Hampstead October 2013: 3 of 4 games counted, the fourth being a win for me
Hampstead August 2013: 4 of 6 games counted, the fifth and sixth being a win and a draw for me.
Penarth July 2013: 7 of 9 games counted, the eighth and ninth being wins for me.


So in the last 6 months, 7 missing games from which I scored +6 =1 -0. That's not at all an untypical experience for me since I got my first elo rating in 2011. 7 games missing from 27 is a hefty chunk. Not surprising that elos and ECFs don't match up when a quarter of your games aren't counted.

Hence:

. my elo started out touching 2050 and plummeted to barely above 1900 after which I (slowly) reclaimed about 30 points.
. while at the same time (from mostly the same games) my ECF remained largely static (and will rise by about 10 points in the new list).

Which is why the conversion table mentioned above in no way reflects my rating nor that of the people that I play.



From previous conversations I gather this experience is not typical of people who play mostly in 4NCL. Maybe the conversion table works better there.

[Roger de Coverly]

(PS. Roger, when you say "Whenever e2e4 run an under 2050 tournament, it can attract a number of players usually only eligible for Opens", there was the tournament you won, what other examples are there?)

There was the Wycombe tournament won by Nick Burrows of Cowley (or even Coley as it's often styled). He had the short straw of having to duff up a 200 graded opponent to win and also contend with the top board for the second section being in the darkest corner of the room after sunset. Sitting on the next board or so, I had contrived to win before it got dark.

[MartinCarpenter]

I can see how all of the missing games will produce randomness for FIDE vs ECF grades. Some of the low/medium level FIDE grades in 4NCLN look very random in both directions.

What I can't really think is how that might produce systemic deflation - surely you're as likely to have clustered losses (or at least bad results) in those non counted games as wins? The juniors do of course spread their low grades because they tend to be quite active with it.....

To be honest, below quite a high level, I'd just ignore UK FIDE grades.

[Roger de Coverly]

What I can't really think is how that might produce systemic deflation - surely you're as likely to have clustered losses (or at least bad results) in those non counted games as wins?

You could also regard it as a correction of past inflation. For many years, the minimum for an International rating was 2000. A side effect of this was that players who were unlikely to get past 175, had ratings. They acquired these because in effect the rating system only ever recorded their better performances as if they didn't meet players above 2000, any bad results, such as losses to 140 players didn't count. Over time they became the field that would be encountered by new players, particularly in the 4NCL. This perpetuated the effect. With the ratings now going down to 1000, the floor effect isn't present, so bad results, always possible with improving juniors in the pool, now drag down the rating.

It may be recalled that a least squares fit to the plot of Elo v ECF came up with the formula of FIDE = ECF * 5 + 1250 some years back.