evaluating past champs by their moves

[lkaufman]

My article on comparing great players of the past and present by their "Accuracy" measured against chess engines was published on chess.com on Monday, and has received a lot of publicity and commentary. https://www.chess.com/article/view/ches ... tings-goat. The key innovation in my method was to exclude draws from the study, primarily because draws get much higher accuracy scores than wins, so including them unduly favors players with high draw percentages, especially short draws (which may even be "Perfect"). I also found a way to adjust for opponents' strength.

Although the fit with actual ratings of modern players was very good, much better than expected, a few of the historical players seem to have unreasonably low "absolute" ratings. I don't know the reason for this in general. Anyway, if anyone reading the article has ideas of how it might be improved in the future, let me know here. I warn you though, it's not easy to suggest improvements that do more good than harm in general, I have tried!

[Fritz 0]

It's hard to believe that la Bourdonnais was weaker than me(!) and Staunton just a bit stronger(!!). Maybe it would be somewhat different if draws were included, but I suppose there were not many draws back then anyways.

As for other things, it's amazing how little Carlsen's play drops from classical to rapid, and not even the official rapid (15'+10'') but much faster one.

[lkaufman]

It's hard to believe that la Bourdonnais was weaker than me(!) and Staunton just a bit stronger(!!). Maybe it would be somewhat different if draws were included, but I suppose there were not many draws back then anyways.

As for other things, it's amazing how little Carlsen's play drops from classical to rapid, and not even the official rapid (15'+10'') but much faster one.

Similarly for me it seems strange that in my early 60s I played better (according to this metric) than Morphy or Rubinstein or Euwe at their peaks! I cannot imagine winning a match from any of those great players at any time in my life. But the method predicts current ratings quite well, so I don't have an explanation. Regarding the dropoff from classical to Rapid, remember that Kasparov crushed the Israeli team (2600+) in a simul, which is sort of like classical to rapid odds, so that is easy for me to believe.

[Ras]

One side aspect: Fischer's conditions for matching Karpov. I don't see how that would have been fair, let alone advantageous for Karpov. Playing until 10 wins, but tie at 9:9 meant that Karpov could not have won 10:9. He would have had to win 10:8 at least. Which means that what Fischer was asking for was a full game handicap.

[carldaman]

My article on comparing great players of the past and present by their "Accuracy" measured against chess engines was published on chess.com on Monday, and has received a lot of publicity and commentary. https://www.chess.com/article/view/ches ... tings-goat. The key innovation in my method was to exclude draws from the study, primarily because draws get much higher accuracy scores than wins, so including them unduly favors players with high draw percentages, especially short draws (which may even be "Perfect"). I also found a way to adjust for opponents' strength.

Although the fit with actual ratings of modern players was very good, much better than expected, a few of the historical players seem to have unreasonably low "absolute" ratings. I don't know the reason for this in general. Anyway, if anyone reading the article has ideas of how it might be improved in the future, let me know here. I warn you though, it's not easy to suggest improvements that do more good than harm in general, I have tried!

It's an interesting concept, but the methodology seems to focus too much on World Championship events. World class players like Rubinstein, and even a recent challenger like Caruana [due to total draw exclusion, which is another potential problem] get to completely fall thru the cracks, which is a bit ridiculous. As are the ratings of players like Staunton and Anderssen - I wish I could play like them. :=P

I do agree with your conclusions on rating inflation (from 1986 to 2006) and deflation since. I remember how people vehemently denied the obvious inflation happening about 20-25 years ago, whenever I tried to point it out to them...

[lkaufman]

My article on comparing great players of the past and present by their "Accuracy" measured against chess engines was published on chess.com on Monday, and has received a lot of publicity and commentary. https://www.chess.com/article/view/ches ... tings-goat. The key innovation in my method was to exclude draws from the study, primarily because draws get much higher accuracy scores than wins, so including them unduly favors players with high draw percentages, especially short draws (which may even be "Perfect"). I also found a way to adjust for opponents' strength.

Although the fit with actual ratings of modern players was very good, much better than expected, a few of the historical players seem to have unreasonably low "absolute" ratings. I don't know the reason for this in general. Anyway, if anyone reading the article has ideas of how it might be improved in the future, let me know here. I warn you though, it's not easy to suggest improvements that do more good than harm in general, I have tried!

It's an interesting concept, but the methodology seems to focus too much on World Championship events. World class players like Rubinstein, and even a recent challenger like Caruana [due to total draw exclusion, which is another potential problem] get to completely fall thru the cracks, which is a bit ridiculous. As are the ratings of players like Staunton and Anderssen - I wish I could play like them. :=P

I do agree with your conclusions on rating inflation (from 1986 to 2006) and deflation since. I remember how people vehemently denied the obvious inflation happening about 20-25 years ago, whenever I tried to point it out to them...

I did include enough "super tournaments" to be able to rate past greats such as Rubinstein fairly, though with the current high draw percentage I couldn't easily get enough decisive games to rate the current stars (other than Carlsen and Nepo). Yes, I could rate them by including more events, but my focus was on the players of the past, since current ones can be compared by Elo rating. Including all draws would have produced nonsense results, so many quick draws look perfect but were 100% played from memory. Perhaps including draws of at least 40 moves would have be okay, but due to varying rules (sometimes all draws by agreement allowed, sometimes after move 30, or move 40, or with TD approval only) this is very problematic. Excluding draws helped the players of old (who had lower draw percentages), so this doesn't explain the seemingly low ratings of some past greats such as Rubinstein and Chigorin. Other than including more events, what else could I improve?

[Peter Berger]

Similarly for me it seems strange that in my early 60s I played better (according to this metric) than Morphy or Rubinstein or Euwe at their peaks! I cannot imagine winning a match from any of those great players at any time in my life.

Your model makes sense to me. But it probably just can't answer this question.

Let's assume you played a match against Rubinstein. Assuming you'd do your very best in your prime and be well-prepared, Rubinstein would be up to quite a lot of surprises at first. Some of the concepts in your game would be news to him. But he'd adapt - and probably very quickly.

This is something we can't test and have no data on. We have data on players facing the next generation, but not three or four generations later.

Let's say you play a Petrosian-like exchange sacrifice. It is not rocket science to +understand+ this concept, inventing it is something else.

My money would be on Rubinstein

[carldaman]

I was going to suggest including longer draws, too, but perhaps you may have to give them a lower weight so they don't overload and skew the results. But excluding them entirely can introduce serious distortions, too.

Carl Schlechter was very strong before WWI, but was known, almost notorious, as a drawing master. He even drew WCh Lasker in a match in 1910, when Lasker was at his peak and crushing lesser opponents.
Not counting draws does such players a great disservice. And from around 1907 to 1912 Rubinstein had established himself as someone in the same 'league' as Lasker and Capablanca, with talks advancing toward a WC match.

Before Lasker, Chigorin and Zukertort gave Steinitz a real 'run for his money' - the latter may have even won if his health had not collapsed.
Very weak ratings for such players doesn't make much sense. Also, Tarrasch was probably the strongest player in the world between 1889-1894,
and was easily Pillsbury's equal (behind Lasker) in the latter half of the decade.

[lkaufman]

I was going to suggest including longer draws, too, but perhaps you may have to give them a lower weight so they don't overload and skew the results. But excluding them entirely can introduce serious distortions, too.

Carl Schlechter was very strong before WWI, but was known, almost notorious, as a drawing master. He even drew WCh Lasker in a match in 1910, when Lasker was at his peak and crushing lesser opponents.
Not counting draws does such players a great disservice. And from around 1907 to 1912 Rubinstein had established himself as someone in the same 'league' as Lasker and Capablanca, with talks advancing toward a WC match.

Before Lasker, Chigorin and Zukertort gave Steinitz a real 'run for his money' - the latter may have even won if his health had not collapsed.
Very weak ratings for such players doesn't make much sense. Also, Tarrasch was probably the strongest player in the world between 1889-1894,
and was easily Pillsbury's equal (behind Lasker) in the latter half of the decade.

I did check both Zukertort and Tarrasch, but their move "Quality" was clearly below Steinitz and Pillsbury respectively by my measure. There is also the possibility that some of those old games have data errors that show up as "blunders" but were not actually played. I tried to look out for that, but of course I could only catch some obvious cases.

[carldaman]

Re: Pillsbury and Tarrasch, they really had comparably strong results, despite Pillsbury's accuracy being 'clearly' better, as you state.
Perhaps some of that can be accounted for by certain stylistic differences, but it could also be that the absence of drawn game data
is giving an incomplete picture.

But Pillsbury might not have reached his full potential in the 1890s, since illness eventually derailed his career, so the greater accuracy may be
an indication of that.