[gamegoblin]

The "200 points difference means 25% chance to win" breaks down at the highest levels. It works fine near the middle of the bell curve -- i.e. 800-2000 Elo -- but once you get to 2200 Elo you are talking about the >99th percentile. For example, I don't know of a single 2400 player who can score an average of 0.25 against 2600 players.

Even at my own mediocre level of 1800, I definitely do not score 0.25 against 2000 rated players. More like 0.1 if I'm feeling sharp.

[stouset]

As an example of this, until the Niemann game Magnus had a 53-game win^H^H^Hunbeaten streak. Prior to this he had a 125-game unbeaten streak. Many (most?) of these games were played against competitors within 200 Elo. Many of these were played against the 10 next-best chess players in the world.

The back of the envelope percentage calculation absolutely does not apply at this level of chess. In reality if Niemann were to play Magnus in 100 games, he would be exceedingly lucky to win one game.

[dfan]

I disagree with the second paragraph but not enough to get into a public debate about it. But it is worth pointing out that Carlsen's 53-game streak was a non-loss streak, not a win streak. Many of those games were draws.

[stouset]

You are of course correct on that point and I have edited my comment.

[andrepd]

So hitting a 1/100 chance means he is cheating? 1% is slim, but far from impossible.

[stouset]

First, I said he would be lucky to have a 1 in 100 chance. Second, absolutely nobody is saying that's the only reason to be suspicious of this game. Regardless of whether or not you believe Niemann cheated, if you think the fact that he won is the only claim in this accusation you simply aren't paying attention.

[jefftk]

It's not just the winning, it's also how he played

[Tenoke]

>I don't know of a single 2400 player who can score an average of 0.25 against 2600 players

I mean, you can look at the stats. They play all the time and while it becomes less accurate at the highest ratings (more so at the 2800+ level), 2400 vs 2600 does still result in something in the general range of 0.25. However, if it's 0.1 (like in your example) then my point is even stronger since it would be even harder to turn that into a win consistently with just 1 hint.

>Even at my own mediocre level of 1800, I definitely do not score 0.25 against 2000 rated players.

If you are noticing that at your level, it is probably either selective memory or specific to your play as ELO-estimated winning chances hold up well enough at 1800-2000.

[dfan]

The Elo system is calibrated so that that the expected value from playing a player 200 points stronger than you is 0.24. This is true independent of the strength of the players. If you are scoring 0.10 against players 200 points stronger than you (that would mean, for example, 1 draw and 4 losses over 5 games) but maintaining a stable rating, then you must be crushing players that are weaker than you and/or doing very well against players at your level.

(FWIW, I am 2000 USCF and an expected value of 0.24 vs a 2200 and 0.76 vs an 1800 feels quite reasonable to me.)

[lupire]

Elo assumes that performance across n^2 players fits the logistic curve model of n players. There is no reason to believe that assumption is remotely accurate and that Elo would ever stabilize. Players often avoid playing lower rated players, for this reason.

[8note]

Is the current ELO accurate though? How many fide rated games were played over covid?

[lupire]

Elo is not an acronym.

[Bakary]

Indeed. It stands for Extra-acronymic Language Object.