Have there been any studies done on the correlation between average accuracy per game and ELO? • page 1/1 • General Chess Discussion • lichess.org

Compare it to soccer/football.
3 goals against Milrose School of the Blind is not equal to 3 goals against Arsenal.

You getting 90% against me is not the same as getting 90% against Magnus.

But sure you should always aim to keep your mistakes limited and playing accurate against higher rated players over time.

Compare it to soccer/football. 3 goals against Milrose School of the Blind is not equal to 3 goals against Arsenal. You getting 90% against me is not the same as getting 90% against Magnus. But sure you should always aim to keep your mistakes limited and playing accurate against higher rated players over time.

mkubecek

Not only that, accuracy also depends a lot on playing style a lot. In general, accuracy says much less about the quality of play than most people believe. Sorry, there is no one simple number telling you how well you played in a game.

IM zwenna

I am not aware of any studies, but https://lichess.org/page/accuracy has some useful information.

petri999

I did post article links in similar discussion about 2 years ago. Answer is yes assuming.

Lichess accuracy has a problem it also takes into account totally useless moves hence it has low correlatin
Here is is the original academic paper:
https://ailab.si/matej/doc/Using_Heuristic-Search_Based_Engines_for_Estimating_Human_Skill_at_Chess.pdf
And tiny practical research on matter from now defunct site:

https://web.archive.org/web/20151112153559/http://chess-db.com:80/public/research/qualityofplay.html

obviously all those analysis are from classical timelimits. Internet is about way faster limits so correlations may or may not be as strong-

I did post article links in similar discussion about 2 years ago. Answer is yes assuming. Lichess accuracy has a problem it also takes into account totally useless moves hence it has low correlatin Here is is the original academic paper: https://ailab.si/matej/doc/Using_Heuristic-Search_Based_Engines_for_Estimating_Human_Skill_at_Chess.pdf And tiny practical research on matter from now defunct site: https://web.archive.org/web/20151112153559/http://chess-db.com:80/public/research/qualityofplay.html obviously all those analysis are from classical timelimits. Internet is about way faster limits so correlations may or may not be as strong-

DoctorFuu

More importantly, rating is a number that relates your playing strength to the whole population of players. Accuracy is a number that show how well your moves were on average against the one specific player you just played. This means accuracy is only relative to your opponent when rating is relative to a whole bunch of players. Understanding that, it's very easy to break any hope of finding a relationship between the two:
if I, 1500, play against a 800 beginner, I will probably have a pretty good accuracy because my opponent will not manage to give me a difficult position, and to the contrary will probably give me a very easy position by blundering pieces or something. If instead I play a 2200 play, my accuracy will probably be terrible because my position will very quickly become very difficult to play. Right off the bat, it's very clear that accuracy is very dependent on the relative difficulty of the game, therefore it can't be a good predictor of your current rating.

Secondly, if I have an attacking allout style, playing only sharp sicilians and gambits, my games will probably have an average of 25/30 moves with many mistakes on both sides because the positions will be very tactical and difficult, and ending up in mate pretty fast on the side that blundered last. Those games will tend to have a low accuracy, because the positions will be very difficult to play for both players.
But if I have a slow positionnal style, and I like to drag the game into tehcnical endgames ? Maybe my games will instead have an average of 45-55 moves, with many positions in which no player makes big blunders but only many small postiional concessions. Because it's more difficult to make big blunders in quiet positions, the accuracy will tend to be higher. also, the games being longer means that even if there's a mistake somewhere, that loss in accuracy will be smoothed out more because more moves were played.
So even if we somehow managed to control for the difference in rating between the players to compute the accuracy, it would still be heavily biased by the playing styles of the players.

All of that being said, on average better players simply play better moves and therefore have better accuracy. This is because on average, players play agianst players of their levels. On average, their opponent have an average playing style (which means nothing in temrs of chess of course). And on average, players also have an average playing style.
Essentially, if we go in the realms of statistics, you would probably find a relation between the two. However, that does NOT mean you can guess from the accuracies of a player his playing strength. The relation between the two is only valid because we are averaging all the effects that throw off the predictive power of that model. But if you are reasoning at the scale of one player, you are not reasoning at a scale in which anything was averaged and therefore you can't draw conclusions.

An analogy for that would be that the average religion of a person on earth is muslim, but that doesn't mean that you can conclude that I am muslim just because I'm a person. As a single person, I do not possess all the attributes of the whole population that was averaged to give the average.

(First of all, a correlation between numbers that live in different worlds is probably incorrect anyways. Accuracy is a percentage, so lives between 0 and 100. Rating is a real number and can go from negative infinity to positive infinity. correlation is the linear relationship between two variables. If you set up a correlation between the two, someone with a surprisingly low or high elo could be "correlated" to play with an accuracy > 100% or <0%.) More importantly, rating is a number that relates your playing strength to the whole population of players. Accuracy is a number that show how well your moves were on average against the one specific player you just played. This means accuracy is only relative to your opponent when rating is relative to a whole bunch of players. Understanding that, it's very easy to break any hope of finding a relationship between the two: if I, 1500, play against a 800 beginner, I will probably have a pretty good accuracy because my opponent will not manage to give me a difficult position, and to the contrary will probably give me a very easy position by blundering pieces or something. If instead I play a 2200 play, my accuracy will probably be terrible because my position will very quickly become very difficult to play. Right off the bat, it's very clear that accuracy is very dependent on the relative difficulty of the game, therefore it can't be a good predictor of your current rating. Secondly, if I have an attacking allout style, playing only sharp sicilians and gambits, my games will probably have an average of 25/30 moves with many mistakes on both sides because the positions will be very tactical and difficult, and ending up in mate pretty fast on the side that blundered last. Those games will tend to have a low accuracy, because the positions will be very difficult to play for both players. But if I have a slow positionnal style, and I like to drag the game into tehcnical endgames ? Maybe my games will instead have an average of 45-55 moves, with many positions in which no player makes big blunders but only many small postiional concessions. Because it's more difficult to make big blunders in quiet positions, the accuracy will tend to be higher. also, the games being longer means that even if there's a mistake somewhere, that loss in accuracy will be smoothed out more because more moves were played. So even if we somehow managed to control for the difference in rating between the players to compute the accuracy, it would still be heavily biased by the playing styles of the players. All of that being said, on average better players simply play better moves and therefore have better accuracy. This is because on average, players play agianst players of their levels. On average, their opponent have an average playing style (which means nothing in temrs of chess of course). And on average, players also have an average playing style. Essentially, if we go in the realms of statistics, you would probably find a relation between the two. However, that does NOT mean you can guess from the accuracies of a player his playing strength. The relation between the two is only valid because we are averaging all the effects that throw off the predictive power of that model. But if you are reasoning at the scale of one player, you are not reasoning at a scale in which anything was averaged and therefore you can't draw conclusions. An analogy for that would be that the average religion of a person on earth is muslim, but that doesn't mean that you can conclude that I am muslim just because I'm a person. As a single person, I do not possess all the attributes of the whole population that was averaged to give the average.