I had no incurracies,blunders or mistakes then also my centipawn loss was 48.

What is the problem here or I have misunderstood average centipawn loss.

You can have a look at that game at lichess. Link to the game: https://lichess.org/fhvYi9SS/black

  • 3
    Do yourself a favour and forget about centipawns! – David Sep 30 '19 at 11:39

Average centipawn loss is the difference of your move to the best computer move averaged over all moves.

Inaccuracies/Mistakes/blunders as defined per lichess are moves that are at least 0.5=50 centipawns / 1=100 centipawns / 3=300 centipawns worse than the suggested computer move.

This rule is not strictly enforced in situations where you have a clear winning position and a choice of several good moves.

So if you have a choice between a move that leads to +10 advantage (best computer move) and +9 advantage this will count to your centipawn loss, but will not show up as mistake.

Even without it, it is possible to collect centipawn losses < 50 that don't show up as inaccuracy.

  • 1
    Yes Now I understand,thanks – Amar Shukla Sep 26 '19 at 18:29
  • 2
    I don't think this explanation is quite correct. Even if the player follows exactly the best moves according to engine, the evaluation curve will not be flat and there will be nonzero centipawn loss. The reason is that evaluation changes after a move has been committed (the engine can recurse deeper). – leftaroundabout Sep 27 '19 at 12:17
  • @leftaroundabout: I guess you'd have to read the source code to know for sure, but I would expect that the comparison is made with the computer evaluation at any particular move. Whether the evaluation changes after that should be irrelevant in that case. – user1583209 Sep 27 '19 at 14:27

I think the issue is that, when one side is winning by a large margin, the analysis will not always indicate an inaccuracy for a move which is not best but still clearly winning.

For example, after 20.Qg4, the analysis says that 20...dxe3 would have a score of -18.9, but your played move 20...d3 has a score of -10.9. That's an 800 centipawn difference. The analysis does not show an inaccuracy here, but that's a lot of centipawns.


I have thought (maybe too much, lol) a lot about this.

"Average centipawn loss (cpl) is the difference of your move to the best computer move averaged over all moves." Yes.

Like all averages this is pretty meaningless without some measure of variance. Very long games will tend to have a smaller cpl because of so many moves and there are not a lot of differences in the value of variations with only a few pieces left on the board. Simply put with lots of moves you are dividing by a larger number, there is little variance between move choices, and this yields a small cpl.

Also, there are games where the top 10 moves will vary in value very little e.g. closed positions like a stonewall. These games will show a low cpl.

I tend to play a very positional game where the positions end up with valuations that vary very little across any of the top 6-8 moves. These games show a low cpl. It does not mean I am playing well. It means the games are intrinsically calm, not sharp and tactical.

So if we had a measure of variance to go along with the cpl we would have a context to better understand what a cpl means. A 25 cpl with a high variance would mean you played rather well. A 25 cpl with a low variance would mean you played OK.

  • Depends what you define as "playing well". If you have a 25 average cpl with high variance, you could still have a game with a high number of moves and a very closed portion of the game and then a few blunders towards the end bringing the cpl up.A game with a 25 average cpl and low variance can consist of a series of sharp plays resulting in a winning position where sacrificing some cpl makes sense and that can still produce low variance. – Wins94 Aug 6 '20 at 19:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.