# Does elo rating have a limit?

The highest Elo rating to date is 2882, which belongs to Carlsen, while the last Elo rating of Kasparov is 2851. Could I have an Elo rating of 2950 or maybe 3000? Does the Elo rating have a limit?

• There is no imposed limit on the rating, but to reach 2950 or 3000 you have to be significantly better than Carlsen. No one but a computer can do it. Aug 11 '15 at 22:39
• @overtheboard, If I never lose a game for 25 years, could I have an elo rating of 3100 ? :-D Aug 11 '15 at 22:41
• It depends who you play. A 2000 who keeps beating 800s won't gain rating points. And to beat high rated players near 3000 (e.g. Carlsen), you have to be better than all the world championship contenders. Aug 11 '15 at 22:44

No, Elo ratings have no theoretical limit. If Alice scored 76% against Carlsen (rating 2853) consistently, she would stabilize at a rating of 3053*, and if Bob scored 76% against Alice, he would stabilize at a rating of 3253**, etc. There is no theoretical end to this sequence. However, consensus seems to be that in practice a perfect player would have a rating somewhere in the 3000s.

*Assuming Carlsen performed well enough against other people that his rating stayed at 2853

**Assuming Alice performed well enough against other people that her rating stayed at 3053

• I upvoted this answer because I believe it to be 90% correct but actually ELO is a closed rating system that can be employed for any 1 on 1 game and the theoretical limit actually is limited by the amount of players and the starting points they are given, often 1500 Aug 13 '15 at 3:00
• It depends how you implement Elo (capitalized like this; the inventor was Arpad Elo). FIDE actually lets ratings grow arbitrarily even for a fixed pool because in its update step it never treats ratings as being farther apart than 400, so there is a lower bound on the number of points the winner gains. It is true that some implementations stop giving you credit after some large difference. If you had 101 players who were each given a starting rating of 1500 and no one could be farther apart than 400 points, the highest possible rating would indeed be only around 21,500.
– dfan
Aug 13 '15 at 11:41
• I am wondering whether the requirements (players having constant rating) are actually achievable considering that you have a finite number of players. (Apart from practical considerations as players would have to play an awful lot of games.) Dec 23 '18 at 7:48
• 1500 points is not the starting rating May 10 '20 at 11:58

Yes for chess, anyway; as ratings go higher , the proportion of perfect games will increase and therefore if chess is a theoretical draw, the rate of draws will also increase (assuming most games are played against players of similar rating.) Once 100% draws are reached there is no better rating. This is borne out by the graph here which if extrapolated indicates that 100% draws would occur for engines at a rating of about 5200. Similar reasoning obtains if chess is a win or loss for white, namely that once perfect play is reached in 100% of games played, all white games are won and all black are lost or vice versa, again assuming play against similar beings. Even if one plays against dissimilar beings, an asymptote is reached in the calculation of elo points won per game; as ones rating rises , it is harder and harder to win points against lower rated rivals, again enforcing an upper bound.

Elo is a relative scale that measures the winning probability between two players. The scale itself is arbitrary and there are many variations but here is no upper limit as such if the game is complicated enough and has enough room for improvements in game play.

In the game of Go a top professional has Elo rating around 6800 (KGS system), but a good AI engine can reach far beyond that. For example the Leela Zero Go engine has a current rating of 11,000+ (May 2018, after training on 7.5 million self-played games), and it is still improving. The game itself has enough complexities so how much further we can push up the Elo rating will be limited by the speed of the computer used and how much computer resources we use to train the AI.