What centipawn loss actually measures (and when it lies)
A centipawn is one hundredth of a pawn. Average centipawn loss — ACPL — is the mean of how far each of your moves drifted from the engine’s favourite, in those units. It is the number most reviewers still reach for when they want a single figure for “how well you played.”
It is also a bad figure for that job. The failure is structural, not a rounding error.
What a single centipawn drop means
Take a position the engine evaluates at +0.40, and play a move that leaves it at +0.10. That move cost 30 centipawns. In a level position, 30 centipawns is a real concession: a slightly worse pawn structure, a misplaced piece, a tempo. You can feel it.
Now take a position at +8.00 — you are up a queen — and play a move that leaves it at +5.00. That cost 300 centipawns. On any raw average it dwarfs the first move. In the game, you are still completely winning. The extra queens and rooks you “lost” on paper were never going to be needed.
ACPL treats those two events as if the second one were ten times more important. The opposite is true.
Why the average is dominated by decided games
Most club games are not a long fight around equality. They are a stretch of ordinary chess, then one or two large errors, then a mopping-up operation in which the winner can drop a piece and still be winning, and the loser can drop a piece and still be lost.
The mopping-up is where the centipawns live. A 400-centipawn hang in a +9 position is cheap in win-probability terms and expensive in ACPL. Play enough of those — and anyone who converts a won game will — and your average is a record of how sloppy you were after the result was already settled.
That is why a 40 ACPL in a game you won from move twelve can look worse than a 25 ACPL in a game you lost by a staircase of 20-centipawn concessions. The first number is mostly the conversion. The second number is the game.
What we do instead
Classification — best, good, inaccuracy, mistake, blunder — is still a centipawn measurement. It answers a different question: how far from the engine’s best move was this one? That is a useful thing to know about a single ply, especially when you are comparing candidate moves on the board. The thresholds we use are in the classification post.
The number we put on the game as a whole is not ACPL. It is accuracy, which converts every evaluation into a win probability first and then scores the drop in that. A swing from +0.2 to +1.4 near equality moves the needle. A swing from +6 to +9 barely does. The formula, and why it disagrees with chess.com, is in How accuracy is calculated.
Two numbers, two jobs:
- Centipawn loss per move tells you how far a move was from best. Read it on the move list.
- Accuracy tells you how much of your winning chance you kept. Read it on the game.
Do not average the first one and call it the second.
A number that is still worth looking at
There is one place centipawn loss is the right unit: comparing phases of the same player. If your average loss per move in the endgame is 1.4 times your average in the rest of the game, that ratio is doing real work. Both sides of it are inflated by decided positions in the same way, so they still compare.
That is the phase chart on a Chess Wizard game review. It is not a claim that 42 ACPL in the opening is “good.” It is a claim that the endgame is where your points are going, relative to you.
For everything else — “am I improving,” “was this a good game,” “how do I compare to a friend on another site” — use win probability, or do not use a single number at all.