Mike, do you have actual data to support those numbers? I agree with the statement that lower rated players would profit, while higher rated players would actually lose ELO. Also the numbers you give seem reasonable.

Disagree with your last point. If you play, a4, breaking the queenside later with c3 or c4 will be much worse because (in addition to d3 also the pawn on b3 will be weak), which is not the case with a pawn on a2.

Arguably, the kingside attack plan would be very slow. In ten or fifteen moves black would have plenty of time to escape with the king and/or move some pieces to the kingside. Also the pawn structure on the kingside is quite hard to break.

My concern is not that it is different for different GM, but different depending on the chosen set of problems. For instance, GMs would have a hard time to guess moves in patzer games or in endgames with many equivalent moves.

So you'd have a value for this set of positions (and the value might very well be different for another set of positions). I doubt that any comparison like this has been done, but I think it would be very interesting to do. Still not too happy with the tiitle of this question which suggests that there is some universal accuracy value.

OK, but I still believe that there is no such thing as a typical move prediction accuracy in general as it would depend on the games you base it on (even with averaging). Instead the following questions could make more sense: 1. What is the move prediction accuracy of a GM for [specified set of games]? or 2. How much better is the move prediction accuracy of a GM relative to an amateur if given the same set of games. The latter question might not have a universal answer and might also depend on the set of games.

I split off the two knights and bishop part here: chess.stackexchange.com/questions/29739/…

White does not have any isolated pawns.

Are the names of the players visible to the guessing person?