Average Centipawn Loss, and What It Does Not Tell You

A centipawn is one hundredth of a pawn. When an engine says a move cost you 45 centipawns, it means the position was worth 0.45 pawns more before you played than after.

Average centipawn loss takes that figure for every move you made and averages it. It is the most widely used single number for how well someone played a game, and it is genuinely informative. It is also averaged, which is where the trouble starts.

How the number is produced

For each of your moves, the engine evaluates the position before you move and after. The difference is your loss on that move. Play the best move and you lose zero by definition, because the best move is what the evaluation assumed.

Two details change the result more than people expect. The first is depth: a game analysed at depth 12 produces different figures from the same game at depth 20, because the engine's opinion of what you should have played changes. The second is capping. Most implementations cap the loss on any single move, often at 300 or 1000 centipawns, so that one catastrophic move does not swamp the average. Whether a cap is applied, and where it sits, changes the final number substantially.

This is why the same game analysed on two sites gives two different numbers. Neither is wrong. They are measuring with different rulers.

Rough bands

Time control shifts every band. A 2000-rated player might average 25 in classical and 70 in three-minute blitz, and neither figure says anything different about their strength.

The three blind spots

Averages hide the decisive move. This is the important one. A game where you play thirty good moves and hang a rook once can produce a better ACPL than a game where you drifted slightly for forty moves and lost nothing outright. The first game is the one with a lesson in it.

Forced sequences flatter you. When there is only one legal move, or only one move that does not lose immediately, playing it costs zero centipawns. A game full of forced recaptures gives you a low average for making moves anybody would find. Positions where you had a genuine choice are the ones that measure anything.

Losing positions are cheap to play. Once you are three pawns down, the evaluation cannot fall much further per move, so your losses shrink. A game where you were crushed early and then defended a hopeless position can score better than a hard-fought draw. The number is not lying, it is measuring something other than what you assume.

What to do with it instead

Track it across many games rather than reading one. A month of games at 60 dropping to 45 is a real improvement signal, because the noise averages out over that many games in a way it does not within one.

For a single game, ignore the average and sort your moves by individual loss. The three worst moves are the whole content of the review. That is where the pattern you keep repeating lives, and no summary statistic will point at it for you.

A worked comparison

Two games, both by the same player, analysed at the same depth.

Game A has the better average and the worse play. One move threw a piece away and decided the game; the other forty-one were fine, which is exactly what drags the average down to a respectable 31. Game B never lost more than about two pawns at once but leaked value steadily across a dozen decisions.

The review that reports Game A as the better performance is arithmetically correct and practically backwards. The player needs to look at one move in Game A and at a recurring habit in Game B, and the averages point at neither.

Time control changes everything

The same player, analysed across a month:

• Classical: around 30
• Rapid: around 45
• Blitz: around 70
• Bullet: above 110

None of that variation is skill. It is thinking time, and comparing an ACPL from a three-minute game against one from a two-hour game tells you which clock was running, not who played better. Any figure quoted without its time control is missing the half of the story that moves it most.