Chess DNAHow chess games are actually lost
Not by erosion — by collapse. In 3,540 engine-analyzed rapid games, the average game's single worst move costs about a rook, at every rating level. Improvement doesn't remove the collapse. It shrinks it, and spaces it out.
68 players · five rating bands (11–16 each) · rapid time control only, opponents within ±250 Elo.
Every game contains a collapse
Average size of each game's single worst move, in centipawns, with piece values for scale. A pawn is 100.
This is the number that reframes how amateur chess is decided. Take every game, find its single worst move by engine eval, and average those: players under 600 give up 754 centipawns in that one moment — most of a queen. And the number barely deflates as you climb: at 1500+, the average game's worst move still costs 506 centipawns, a full rook. Not the worst game of the month. The average game.
A blunder of that size doesn't nudge a game — it decides it, or at minimum hands the decision to whoever collapses second. That's the honest description of amateur chess our data supports: two players trading small imprecisions that barely matter, until one of them produces the moment. It's the same conclusion we reached from a different angle in what each Elo level struggles with — the themes change with rating, but the shape of the loss doesn't.
Fewer collapses, never none
Blunders per game by rating band — a blunder being a move that loses at least 20% of win chance. The dashed line is one per game.
The count falls steadily — 1.66 per game under 600, 0.90 at 1500+ — and the per-move fall is steeper than it looks, because stronger players' games are also longer (61.5 → 74.5 moves): per 100 moves, blunders drop 55%, from 2.69 to 1.21. That's real, earned improvement. But look at the floor. Even at 1500+ — stronger than the vast majority of online players — the average game still contains 0.9 game-changing mistakes. Chess improvement, measured honestly, is not the road to clean chess. It's the same collapse, made smaller and rarer.
Which mistakes shrink, and which stubbornly don't, is its own story — we broke that down by theme in which chess mistakes go away as you improve. Short version: the loud ones fade; the quiet ones survive.
What actually predicts rating
Share of rating variance explained (r²) by each metric, across 68 players. The headline number chess sites show you is near the bottom.
If games are decided by collapses, then metrics dominated by big mistakes should predict strength best — and they do. Average centipawn loss explains 72% of rating variance, blunder rate 58%, the size of the worst move 48%. Meanwhile accuracy — the percentage every platform headlines after each game — explains 31%, and inaccuracy rate a nearly-irrelevant 12%. Accuracy compresses everyone into a flattering band and dilutes exactly the signal that separates players. Two players can post the same accuracy while one blunders twice as often; the blunder count is the one that shows up in their ratings.
The practical read: if you review games by staring at the accuracy number, you're studying the weakest signal on the board. The strongest one is the move where you keep blundering — the phase-by-phase version of this argument is in which phase of chess decides your rating.
What to do about your own collapses
Study the decisive mistake, not the average. Every loss contains one move that cost more than all the others combined — find it, and ask why it happened: missed opponent threat? Calculation stopped a move short? A pattern you've missed before? One understood collapse is worth more than fifty polished inaccuracies, because the collapse is what the rating ladder actually prices.
Then shrink the moment. A pre-move blunder check (checks, captures, threats against you) attacks the count. Reviewing your own decisive mistakes attacks the size — most players' collapses cluster in a small number of recurring patterns, which means the fix is narrower than it feels. That's the entire design premise of Chess DNA: engine-analyze your games, isolate the moves that decided them, and drill those exact positions until the collapse stops repeating.
Blunders per game at 1500+. Even strong club players hand over nearly one game-changing move per game.
Average centipawn loss vs rating — the best predictor we measured, because it's dominated by the big mistakes.
Of rating variance explained by inaccuracy rate. Small imprecisions barely separate players — stop sweating them.
Average game length in moves, under-600 → 1500+. Fewer early collapses = longer, tenser games.
Questions people actually ask
How are most chess games actually lost?
In one moment. The average game's single worst move costs 754 centipawns under 600 (nearly a queen) and 506 at 1500+ (a rook) — large enough to decide the result on its own. Improvement shrinks the collapse and makes it rarer; it doesn't eliminate it.
How many blunders do players make per game?
By band: 1.66 under 600 → 1.33 → 1.14 → 1.09 → 0.90 at 1500+ (blunder = losing ≥20% of win chance). Per 100 moves it falls 55%, from 2.69 to 1.21, since stronger players' games run 21% longer. The floor never approaches zero.
Does accuracy predict chess rating?
Weakly. Across 68 players: average centipawn loss r² = 0.72, blunder rate 0.58, worst-move size 0.48 — accuracy just 0.31 and inaccuracy rate 0.12. Accuracy compresses everyone into a narrow band; the big mistakes carry the signal that separates players.
What's the fastest way to lose fewer games?
Attack the collapse: a blunder check before every move, and a review habit that finds each loss's single decisive mistake and names why it happened. Polishing already-decent moves improves inaccuracy rate — which explains almost none of the rating gap between players.
How was this measured?
3,540 rapid games from 68 Chess DNA players, Stockfish 17 at depth 18, five bands by average game rating (opponents within ±250 Elo). Worst-move size = each game's largest centipawn loss, averaged per player, then per band. r² from single-variable fits across players.
Want to see your own collapse pattern? Chess DNA finds the exact moves that decided your games — and turns the recurring ones into drills.
Method. 3,540 rapid games from 68 Chess DNA players, analyzed with Stockfish 17 at depth 18 and restricted to rapid time control so time pressure doesn't confound the comparison. Bands by average game rating: <600 (n=15), 600–899 (13), 900–1199 (13), 1200–1499 (11), 1500+ (16); opponents within ±250 Elo. A blunder is a move losing ≥20% of win chance; worst-move size is each game's single largest centipawn loss (piece values: pawn 100, minor ~300, rook ~500, queen ~900), averaged per player then per band. Predictive power is r² from single-variable linear fits of each per-player metric against rating. Part of a series with What Each Elo Level Struggles With and Which Phase of Chess Decides Your Rating.