There must be at least a rough correlation between accuracy and playing strength/ rating?

1) Has anyone used the accuracy % to guide his improvement?  If so, how?

2) My own accuracy is in the 80s about half the time, but it ranges from the mid-60s to the mid-90s. Has anyone used variance of accuracy to improve reliability?

Thoughts on how to use this statistic to improve?

Thanks.

Too much of accuracy depends on the opponent.

For example if I blunder a piece on half my moves, then even a near beginner could have a high accuracy % playing against me.

But when you play a strong player, not only are they not blundering, but they're trying to set difficult problems for you. For example when they're losing, they'll choose your winning path to be the longest and most technically challenging.

So what tends to happen is as long as you're playing people near your rating, you can score 70, 80, 90% accuracy almost no matter what you rating is.

Other than early blunders (noted), I find accuracy to better correlate with gap between you and the opponent's rating. If you are rated much higher than them, then expect you to have a higher accuracy percentage as you exploit weaknesses. If the players are roughly equal, then I find the accuracy score to be too unreliable for representative game quality. It is just a "cool" statistic generated by a chess.com computer to motivate beginners (which offers it's merits), but I usually take this score lightly with my games, or games from stronger players.

I don't find it useful at all and I do agree with the previous posts. I had situations where I had a 90% accuracy even though I myself hand't found the game to be particularly good. Sometimes you get an incredibly high accuracy in short games just because you followed theory, I can get a 98% accuracy in a 20 move game just by knowing the opening. You also seem to get incredibly low accuracy on games with very long endgames, because the engine in endgames sometimes just derps out. I also had a game a couple of days ago where I won, but had over 20% lower accuracy than my opponent.

Point is don't rely on these things too much they are more gimmicks than useful tools, they just make things more interesting to look at. My advice is not to rely on engines period. Just study your games your self, see what moves were good and judge for your self.

to find out the correlation between average centipawn loss (also called accuracy) and rating, you would have to gather enough randomized data. Let's say if you and many other people played 1000 games against opponents ranged 1200-1400, 1000 games against 1400-1600 and so on, a function can be set up, showing the correlation between your average performance and ratings of other players with similar performance, thus deducting your aproximate strength. Rating is not an absolute number, it is merely a system we use to compare playing strength relative to other players. If worlds strongest 100000 chess players suddenly stopped playing there would be a big rating shift, and people that are in 2500s might become 27 600s or even 2700s, without their average centipawn loss against players of same strength ever actually changing.

*2600s

I greatly appreciate the thoughtful feedback.  Thank you.

yes

I bet to differ.

Like Go, there are moves that can be classified in the same category but affect the entire game more. Accuracies are only a measure of close you are to a certain chess engine's top moves but it's all about the amount of quantities of good moves you made, not how much each move affect the entire game. While it does signify some level of competence in certain chess variations, it does not say how much each move can affect the entire game as seen on the left bar of the analysis or in the analysis screen.

Thus, if you made a critical wrong move, you can end up losing the entire game while still having high accuracy or win a game with slightly lower accuracy. The problem is that some moves just affect more and can be considered as being critical. 

Though, with the new Game Review, the accuracy algorithms were changed, so comparisons between the old an new, to gauge improvement, may not be valid.