GM inaccuracies

Does anyone know what the range of inaccuracy rate (as assessed by computer) is for GMs? My understanding is that they are present in almost all games, but I have no idea whether they occur a few or many times in a typical game.

I am curious to know how the frequency of inaccuracies (and mistakes and blunders too) vary with FIDE rating.

Not sure what you implied by "rate" here. There are far too many parameters to tweak to determine a "what is an inaccuracy" via a chess engine.

For example, if you let Houdini spend 10 seconds per move on a decent i3 laptop analyzing a GM vs. GM game, you'd see inaccuracies as shown in the eval. graph below:

(where the Y-axis is the engine evaluation score, X being the moves)

One could argue that ALL of Capa's "down-slope" shifts are inaccuracies. In other words, any time the GM deviated from what an engine deemed best was inaccurate?

Is this what you implied?

If you compared this to an eval. graph of two intermediate (1400-1600) players playing each other, you'd have something like this:

Here  the # of big up/down slopes signify far more than just plain inaccuracies ... it is blunder city!

It would be cool if somebody statistically analyzed all the games in a GM megabase to see who were the most accurate/inaccurate were ... though I imagine that is a tough task to perform given that it would be hard to "standardize" the results for a given combination of  engine + hardware + analysis parameters.

Have you seen this?

Insightful post. I agree the precise evaluations would depend on the setup, and in any case each player will vary according to whether he or she is having a good day - but I think there would still be enough data to fit a trend line and estimate roughly how (and how well) accuracy correlates with FIDE rating.

Yes, that is very interesting! It's exactly the kind of analysis I had in mind. I wonder if the authors have continued their investigations. 

Also see this.

Shivsky's graph has more to do with an engine's progression of how it sees the game than it has to do with the accuracy of the actual moves that are played, right? After 1. e4 e5, a typical engine will already have spurted out evaluation numbers ranging from -0.10 to +0.30. Does that change the accuracy of the two halfmoves e2-e4 and e7-e5?

Heinzie made an excellent point ... those graphs just happened to be what my engine + hardware coughed up giving it an arbitrary amount of time / position.

If today's engine evaluation was infalliable, all engine vs. engine games would logically be  drawn (or so says Game Theory!)

At best,  one can measure "fuzzy" levels of "move" accuracy or inaccuracy with them....though it is valid to re-package these statistics as :

"According to Houdini on a Quad Core machine spending 10 minutes per move .... Grand Master A is more accurate than Grand Master B" and so on.

Nothing wrong with that and I'd love to see stats like this ... though what is truly "accurate" is something that requires chess to be solved.