Coming specifically to Keres:
Firstly, one needs to know that Sonas' source data was just the games included in 'Chessbase' at the time (so not full crosstables, so he misses out a huge amount of data). But in addition, if day and month dates for the games in Chessbase were not given, Sonas simply lumped all that years' games and tournaments at the beginning of the year (January). He could not be bothered obtaining easily available dates for even very important events such as USSR Championships or Moscow 1956.

So coming now to the year 1956 (under Keres' "career details"), we see the Alekhine Memorial, Moscow 1956 (which actually took place October - November 1956) as "January 1956" lumped together with 3 other events. So it cannot be separated out as to its effect on subsequent ratings. But we can estimate Keres' average "performance rating" over the 4 events to be about 2710. Now, here's the weird thing - although Keres' January 1956 rating is given as 2782, he loses only 2 rating points for this performance rated well under this. Perhaps this is a quirk in the way chessmetrics operates, but if so, you would expect it to be consistent. But now come to Botvinnik's "career details": Here Botvinnik has a January 1956 rating of 2756 and a performance rating for Moscow 1956 of 2778. Yet not only does Botvinnik not gain any rating points for this great performance, he actually loses one. Go figure!! So basically, the effect in chessmetrics is that Moscow 1956, in which Botvinnik came 2.5 points ahead of Keres, wasn't played. So, yeah, if chessmetrics is going to in effect not include events like Moscow 1956, it is perhaps not too surprising that Keres shows up in the chessmetrics ratings ahead of Botvinnik.

Interesting to note, too, that the 1952 USSR Championship, in which Keres performed truly disastrously, and which was held at the end of that year, was lumped in at the beginning of that year. Had it been put where it should have been, then Keres would certainly have been number 1 in the middle of that year (ie after his great performance at Budapest 1952). This was a genuine peak period of Keres' career.

Also to be noted is that after AVRO 1938 plus all Keres' great performances in 1937, he is still given as merely fifth. Behind, for example, Fine and Reshevsky, both of whom he had outperformed in the preceding two years.

Chessmetrics is complete rubbish and is effectively misinformation - it shouldn't be used for anything. It most certainly is not an "objective comparative assessment of Keres tournament results".

It is to be noted that in the period 1955-1957, Keres won no major event. Whereas in 1937, he won Semmering-Baden; in 1938, AVRO; in 1947, Parnu and the 1947 USSR Championship; in 1950, the USSR Championship; in 1951, the USSR Championship; in 1952 Budapest. Keres' peak periods were 1937-8, 1947, and 1950-1952. In my opinion, 1955-1957 was decidedly not a peak period. Same applies to 1958-1959.

I've shied away from using chessmetrics much myself, somewhat due to my work mostly being on older historical games / players.

I'm just curious then, what's your opinion on EDOchess - a site I do rely on quite a bit:

http://www.edochess.ca/

I haven't really dug down to the bare metal on the technical details of his (i.e. Rod Edwards') approach, but since he's an academic, my cursory reading, and his results, all this has given me a minimum level of confidence in it.

I hope that he continues his slow trend to including more and more modern play.

<A considerably better system is Edo ratings...>

<There is another point of similarity between the new Chessmetrics system and the Edo system: the use of hypothetical reference games (like the earlier version of the Edo system, but not the current one) to provide continuity in player ratings over time. As well as the actual match and tournament results over the 4-year period leading up to a given month, his formula includes two factors (like additional hypothetical game results), one which pulls a player's rating towards the average opponent's rating (during the simultaneous calculation) and another which pulls the rating toward a rating of 2300. The implementation of these last factors, however, gives them a much stronger effect in the Chessmetrics system than the similar factor did in the old Edo system. >

http://www.edochess.ca/Edo.explanat...

This is almost like reading a PhD thesis, requiring quite a bit of concentration to even get a grasp of what's going on.

But it seems that EDOchess isn't using them anymore ("but not the current one").

For older period ratings (pre-1890), I think chessmetrics is absolutely absurd. Sonas does not have much material available for this period and his utterly idiotic formula of losing rating points after only one month's inactivity completely distorts the ratings for a period when they necessarily played much less than they do now. That is a large part of his problem with Lasker as well.

Re Edo, I think Rod Edwards is doing a great job in collecting data. I, too, hope he can extend his work to the advent of Elo ratings, but I can see this would be an enormous amount of work.

I haven't looked at Edo in the same depth and it's some time since I did look.

I hadn't realised he had removed the hypothetical reference games. It's good to see he realised the folly of having these. They are still in chessmetrics of course.

I don't like that he includes odds games. Particularly after about 1878 when I think there are sufficient *real* games to stand alone. Perhaps odds games are necessary to flesh things out before 1878, but I'm not sure about this and dislike the concept of including them.

I believe his historical ratings are some sort of weighted average of events either side of the date in question. Thus they differ from the type of ratings that Elo are - if we applied his ratings today, we would have to wait about 3 years before we knew what today's ratings are. Thus they are not suitable for putting in as current ratings (ie without foreknowledge of what happens after) for players if covering a historical tournament, eg New York 1924.

"This is almost like reading a PhD thesis, requiring quite a bit of concentration to even get a grasp of what's going on". I agree, and I don't really understand his system or his justifications for it. I have to say I don't really trust these "weird" systems and much prefer Elo.

In general, as regards whether a system looks reliable or not, I check for absurd results. In this respect, Edo ratings do not appear to be too bad; nothing like the absurd rubbish chessmetrics produces. Nevertheless, there are one or two strange looking things:

(1) Lasker's rating is 2757 in 1899 and 2741 in 1900. Thus in the period when he had a huge victory at Paris 1900 and loses one casual exhibition game to Pillsbury, he nevertheless loses 16 points. Hard to imagine this happening if it was Elo ratings.

(2) Lasker loses a further 27(!) points to go to 2714 in 1901. This is in a period of playing exhibition games in which he scored +6-0=4. You lose 27 points for scoring 8/10?

(3) Lasker's highest rating is given as being in 1914. Personally, I don't believe Lasker was stronger in 1914 than he was in the period 1896-1900.

(4) Capablanca's rating in 1919 is given as 2832, while in 1921, it is given as 2802, with the only event Capa playing in 1920-1921 being the WC match with Lasker. You lose 30 points for beating a still high rated Lasker +4-0=10?

(5) Capablanca's peak rating seems extraordinarily high. Was he really that much better around 1919 than Lasker was in 1896-1900, or Alekhine in 1930-1931?

(6) Botvinnik is given a higher rating than Flohr for the whole period 1929-1932. But Botvinnik's drawn match with Flohr in 1933 was regarded at the time as a great achievement on Botvinnik's part - Flohr was fully expected to win. So this looks rather weird.

So in general, the Edo ratings look reasonably reliable, and I think would be more so if the odds games were taken out. But not really understanding how they work, I'm not sure I would trust them completely, and there are some strange results.

Obviously a lot to chew on, and it appears a few observations worthy of mulling over.

I'm not sure how much weight is put on odds games in the EDO system. I had always assumed (perhaps through laziness) not much.

I'm almost afraid to comment on your last idea, as this is just off the top of my head, but how would you put error estimates into a pure ELO type system?

Game Collection: Lasker - Janowski Exhibition Match

for a vote to become an official game collection.

Also awaiting review are:

Game Collection: Ostend 1906, a completion of one of the late User: Phony Benoni 's projects.

and

Game Collection: Charousek - Maroczy

Please visit:

"The guys at Ostend did things in the grand style."

Funny you should mention that.

As in not either, and not that hard to find:

https://books.google.com/books?id=F...

Karl Lepge = <Carl Christian F. Lepge>