http://db.chessmetrics.com/CM2/Play...

According to Sonas' page, the earlier games get weighted - linearly, for 4 years. That means, a game played a month ago should have a weight of 47/48. Not too much of a difference to 1 to get 30 points less. Also note, th fluctuations in the next months are nowhere near 30 points - despite the fact that all formulae applied by Sonas are linear.

<Had this tournament been rated for this new player for October, which using the chessmetrics methodology it should, the player probably would not have had a 2500 rating since additional games would lower weight of the padding of the opponents' ratings.> The 2500 number <is> already a padded performance. So he needs only to weight it. For October, the weight would be 1. <Additionally, chessmetrics ratings are geared toward ranking the higher rated players>

It still doesn't answer how h coms to the ratings. I had same lack of success trying to reconstruct ratings of world champions.

<I suspect, the chaotic variation is due to additional inactivity and variation in the rating of the player's opponents.>

That would make no sense as the games have (even though sinking) influence <four years ahead>. After 4 years, the ratings are totally different, and it would be somewhat senseless to rate a tournament played 4 years ago with ratings of today... Especially with the tournament being a youth championship...

<Did you check if such 30 point variations are common?>

Some players get their first rating after some 5-6 months after their first tournament (with no tournaments in between!) which is also unexplained. This player plays three tournaments in January and gets a rating in May: http://db.chessmetrics.com/CM2/Play....

For this one the ratings are not shown at all but we see that his highest rating is <86> points below his initial performance: http://db.chessmetrics.com/CM2/Play...

Other possible reasons for a delay in ratings entering the system is the criterion of how "connected" a player is to the top players.

There is obviously no way to check these hypotheses without reconstructing the input data.

http://web.aanet.com.au/memorysport...

In my experience looking at the data for super-GMs, it's circa 50 points, although with big differences between individual players. I would also expect it to depend strongly on playing strength. It's also been shown to depend on the time control.

Jeff Sonas seems to imply that it's about 35 points here http://www.chessbase.com/newsdetail...

It also answers a follow-up question that I had in the back of my mind. Is there discussion about correcting performance ratings for the white/black bias? (Jeff Sonas actually advocates this in the text you linked.)

In tournaments for individuals, this is not that much of an issue. (The number of games with white and black will hardly ever deviate by more than one). However, in team events, the white/black ratio can be strongly biased. E.g. at the 2008 Olympiad, Grischuk had 6 times white and black only twice:

" The process of rating players can be compared to the measurement of the position of a cork bobbing up and down on the surface of agitated water with a yard stick tied to a rope and which is swaying in the wind. "

Yes. And the two seem to be quite different. For example, I think a good prediction method should be very sensitive to players' current form, which is generally pretty volatile.

For example, player A is 2650 rated; he starts a tournament poorly, so prediction algorithm should predict his results as if "A" was considerably weaker player (let's say 2500). Therefore, the algorithm would have that player's rating at 2500 for that tournament (unless the player turns his performance around). However, for the next tournament the the algorithm should treat the player as somewhere closer to 2600, and then adjust to his performance during the tournament.

The algorithm would probably do pretty well if treated Grischuk as 2550-2600 player during Tata-Steel, but it should weight him higher than that for the next event.

<and that Fide's <Elo> is also based on the normal distribution (they don't use logistic curves)>

True, but I think that USCF uses logistical distributions because gaussians underestimate the rate of upsets at large rating differences. FIDE ratings attempt to make the gaussians heavy-tailed, by having a cutoff threshold which never asymptotically approaches zero. But these cause distortions that are worse than the supposed problem.