Lichess Blitz rating to FIDE rating

[lkaufman]

I discovered a real oddity on the topic of comparing Lichess blitz ratings with CCRL blitz ratings. There is an engine listed there called "Safrad 2.2.40". It has a CCRL blitz rating of 1009. However if you go to the link where you can download it, it is stated that it has a Lichess Blitz rating of 2050!! I've been claiming that CCRL blitz ratings were too low relative to human ratings, especially Lichess ratings, but a difference of well over a thousand elo?? Can this possibly be true? Does anyone know anything about this engine or have any idea of what could explain this? If this is legit then we really need to rethink how we quote human elos for engines.

The actual quote says that it has a ~2050 lichess Blitz GLICKO2 RATING, but the actual current Lichess blitz rating for this engine is 1705. So where does one see the "GLICK02" rating for it, and why would they be 345 elo different? I know that Glicko weights games by how recent they are, but this wouldn't explain any big discrepancy here. Could it have to do with removing games played vs. engines to leave only human games?

[MonteCarlo]

As far as I can see, the 2050 claim is made for 3.0, which does not seem to have played any games on lichess: https://lichess.org/@/Safrad-3/all

For 2.2, the claim is ~1750 lichess, which lines up with the bot: https://lichess.org/@/ChessChildren/all

Bot identities from "Play Online" column athttps://sx.rosada.cz/chess/.

Rating claim for 2.2: https://sx.rosada.cz/projects/safrad-2.2

Info page for version 3 doesn't have the version number in the url, which might be part of the confusion: https://sx.rosada.cz/projects/safrad, but note the "Other Versions" section on the page.

Cheers!

[lkaufman]

As far as I can see, the 2050 claim is made for 3.0, which does not seem to have played any games on lichess: https://lichess.org/@/Safrad-3/all

For 2.2, the claim is ~1750 lichess, which lines up with the bot: https://lichess.org/@/ChessChildren/all

Bot identities from "Play Online" column athttps://sx.rosada.cz/chess/.

Rating claim for 2.2: https://sx.rosada.cz/projects/safrad-2.2

Info page for version 3 doesn't have the version number in the url, which might be part of the confusion: https://sx.rosada.cz/projects/safrad, but note the "Other Versions" section on the page.

Cheers!

Thanks, this explains most of the discrepancy (although it doesn't explain how 3.0 has a lichess rating without playing any games there). Regarding the 1750 rating claim for 2.2, it says " lichess Blitz GLICKO2 RATING: ~1750". Is this Glick02 rating different from the normal lichess blitz rating, or just the full name of it? Would this account for the difference between 1750 and the current 1705, or is it just that it had a rating of 1750 at one point in time?

[MonteCarlo]

Glicko 2 is just the rating system lichess uses.

If I had to guess, 1750 is just the multiple of 50 it's stayed near to the longest. It was pretty solidly over 1750 until October 2021, when it dropped to its current level.

On the pages for the individual versions, there's a link to the bot account for the rating claim for 2.2, but not for 3, which means if I had to guess, I'd say the lichess estimate for 3 is just the author's estimate based on strength improvements tested otherwise.

Cheers!

[lkaufman]

Glicko 2 is just the rating system lichess uses.

If I had to guess, 1750 is just the multiple of 50 it's stayed near to the longest. It was pretty solidly over 1750 until October 2021, when it dropped to its current level.

On the pages for the individual versions, there's a link to the bot account for the rating claim for 2.2, but not for 3, which means if I had to guess, I'd say the lichess estimate for 3 is just the author's estimate based on strength improvements tested otherwise.

Cheers!

Thanks, now we have a good candidate to connect lichess (and indirectly FIDE) ratings to CCRL. Although Safrad 2.2.40 (a.k.a. "chesschildren" on lichess) is only rated 1705 Lichess blitz, looking at its results it seems to consistently gain elo vs. humans and lose elo to other engines. So I just looked at the last fifty rated blitz games against humans (the filter for human games didn't work, I had to just ignore BOTS) where the time limit was at least 3' + 2" or slower (5' + 3" maybe typical), and the performance rating was about 1800 (I didn't calculate it precisely, could be off 10 to 20 elo). This equates to 1683 FIDE based on the conversion formula in this thread. The CCRl blitz rating is 1008. So I think it is fair to say, based on this one engine, that a 1000 CCRL blitz rating for an engine means that it should be evenly matched at 5' + 3" (on the CCRL reference i7 hardware) with an average human with a FIDE rating around 1675 (say 1650 to 1700 to allow for various possible causes of error). There is enough data to do a similar calculation for Rapid, but we don't have a formula for converting Lichess Rapid to FIDE ratings. The results for Safrad were actually higher in Rapid, but as noted in this thread the human Rapid ratings on Lichess seem to run higher (for the same humans) than blitz ratings. At least it doesn't look like a major difference between slow blitz and Rapid. If anyone can suggest another engine with a CCRL blitz rating and lots of games with humans rated close to the engine in Lichess, that would be very helpful; at least we have one pretty solid data point now.