The Amateur Chess Observatory
47% of hanging pieces lost could have been saved
Every rating band is affected, in the same proportion. Defense is the most under-trained skill in amateur chess — the proof from 52,005 real games.
Published September 22, 2026 · figures frozen for this edition
52,005
games analyzed
302,391
mistakes detected and named
38
mistake mechanisms tracked
5
rating bands, from 400 to 2399
The result
Of 13,754 pieces lost after being left hanging, 6,486 could still be saved at the moment of the mistake: a move existed that put the piece out of danger. The move was there; the player played something else.
That proportion — 47% — is the same in every rating band measured, from 400 to 2399. That is the result: the absence of a slope. Between 600 and 2200 rating, the frequency of hanging pieces is divided by three; the share of them that were savable does not shrink.
As players improve, they create fewer losing situations. Faced with a given losing situation, they do not solve it any better.
Of 13,754 pieces lost after being left hanging in amateur games, 47% could still be saved at the moment of the mistake — and that proportion is the same from 400 to 2399 rating.
Share of hanging pieces a quiet move could save, by rating band (%)
Pieces lost after being left hanging, restricted to cases where the following position is evaluated; a piece counts as "savable" when the engine’s best move put it out of danger at no cost.
How this number is established
Every game in the corpus is analyzed move by move with an engine. When a player loses material after leaving a piece hanging, we look at what the engine would have played instead at the exact moment of the mistake. If that best move is a QUIET one — it captures nothing, gives no check, and puts the piece out of danger — the piece was savable: the resource required no combination and no counter-attack, only to be looked for.
One exclusion, decided by re-reading positions one by one: a piece given up by a player whose position stays winning after the loss (at least three pawns ahead) is not counted as a mistake. Trading down into a simpler, still-winning endgame is a simplification, not an error. Those cases — 6.6% of the total — are excluded, and excluding them RAISES the number: they were mostly cases without a save.
Every rate is computed player by player (player × band pairs with at least 20 games), then averaged: one player = one vote — a large game volume never outweighs ten players. Full definitions are at the bottom of the page, and every example position below comes from a real game in the corpus.
Five real positions, one per band
In each position, the player to move — at the bottom of the board — is about to lose a piece they could still save. The red arrow is the move they played, the green one the move that saved it, the blue one the punishment that followed.
The five positions are the most costly cases of their band, one distinct game per case — no position is invented or retouched.
Why this skill does not improve: it is barely trained anywhere
A skill that does not improve across 1600 rating points is not inevitable: it is a skill nobody trains. The most used tactical-motif taxonomy in the world — the one behind Lichess puzzles — counts 25 motifs. One is defensive.
Our own exercise banks, built before this study, carry the same imbalance: 59,629 positions ask you to find the winning move, 5,122 ask you to hold. Twelve to one. Chess training, as it exists, teaches taking — almost never saving.
The imbalance has a practical cause: an attacking exercise has one answer — the winning move. A defensive exercise requires checking that every alternative loses, which makes it costlier to build and more ambiguous to grade. It is easier to produce a thousand attacking puzzles than a hundred defensive positions — and that is what the market did.
1 / 25
defensive motif in the Lichess puzzle taxonomy
12: 1
attacking positions for every defensive one in our own banks
Of the 25 tactical motifs behind Lichess puzzles, one is defensive. Chess training teaches taking — almost never saving.
The second fact: the weakened king, the mistake that holds on best
Of the 26 mistake mechanisms frequent enough to track, the one that best resists improvement is weakening your own castled king — h3 "for air", g3 "to breathe", the pawn pushed in front of the king for reassurance. 13.6 occurrences per 100 games in the 400-799 band, 14.6 in the 2000-2399 band.
And it costs: for the same player, games where it appears score 7 to 10 points less per 100 games than their other games (a win is worth 1 point, a draw ½), in every band. A measured correlation, not a proven cause — but it points the same way everywhere.
The contrast with attack says it all: over the same climb, the missed forced mate is divided by 6. Attacking gets learned by playing. Protecting your own king, far less easily.
Of every mistake mechanism measured, weakening your own castled king is the one that best resists improvement: it is more frequent at 2000-2399 rating than at 400-799 — and the games where it appears score 7 to 10 points less.
Two curves, two fates (occurrences per 100 games)
Means of per-player rates, by 400-point rating band.
Three habits these numbers point to
If defense is barely trained, the least we can do is say what to train. Each habit targets a mechanism improvement does not erase.
1. Piece attacked: look for the move that puts it out of danger first.
Nearly one lost hanging piece in two could have been saved, often by a simple retreat. Before looking for a counter-attack or accepting the loss, exhaust the simplest question: is there a quiet move that saves this piece?
2. After every opponent move, ask what it attacks now.
The missed defense is the most frequent mistake in the corpus, at every level. The habit that reduces it is re-reading the position from your opponent’s point of view — every move, before looking for your own.
3. Before h3 or h6, ask which square that pawn stops defending.
The weakened king is the mistake that best resists improvement, and it is almost always a pawn move played for reassurance. The check fits in one question: after this move, which square in front of my king is no longer defended, and which enemy piece is eyeing it?
What about YOUR games?
These numbers describe a level, not a player. ChessPivot runs the same 38 detectors on your own games and tells you which of these mistakes keep coming back in yours — named, counted, and linked to the exercise that trains them.
Citing the Observatory
Observatory data is freely reusable with attribution.
Amateur Chess Observatory, Fall 2026 edition, ChessPivot. chesspivot.com
The headline fact by band, the 38 mechanisms × 5 rating bands, the barometer, the survival index and the race exposure. Every row carries the definition of what it measures. Separator: comma; decimal: point; encoding: UTF-8.
Journalists, creators, coaches: for a precise figure, unpublished data or an interview — contact@chesspivot.com.
The systematic review
The Observatory’s reference indicators, re-measured identically in every edition — same definitions, same estimator — so that editions stay comparable. Means of per-player rates, bullet excluded.
The games
As the rating climbs, games get longer, end by checkmate less and less, and swing less often — but one winning position in four still fails to be converted.
| Rating band | Moves (median) | Ending in mate (%) | By resignation (%) | On time (%) | Advantage changed sides (%) | Winning position not converted (%) |
|---|---|---|---|---|---|---|
| 400-799 | 28 | 35.8 | 43.0 | 18.4 | 51.9 | 32.5 |
| 800-1199 | 30 | 32.7 | 52.6 | 11.0 | 51.0 | 30.7 |
| 1200-1599 | 32 | 24.0 | 57.9 | 14.2 | 49.0 | 28.5 |
| 1600-1999 | 33 | 18.3 | 59.2 | 17.9 | 45.1 | 27.2 |
| 2000-2399 | 35 | 13.7 | 61.1 | 18.2 | 40.9 | 25.8 |
The moves
Big mistakes — at least two pawns lost in one move, game still open — are nearly halved between the extreme bands.
| Rating band | Big mistakes / 100 moves | Games with a hanging piece (%) |
|---|---|---|
| 400-799 | 7.64 | 18.8 |
| 800-1199 | 6.58 | 15.9 |
| 1200-1599 | 5.24 | 13.3 |
| 1600-1999 | 4.31 | 9.1 |
| 2000-2399 | 3.41 | 6.6 |
The survival index: what fades, what holds on
Base 100: the rate in the 400-799 band. The bar shows what is left of it in the 2000-2399 band. A short bar is a mistake improvement erases; a long bar, a mistake that holds on.
What fades fastest: the early queen, the missed mate, the hanging piece — seeing material, yours and your opponent’s, gets learned by playing. What holds on: the missed defense, conceded material, activity, centre and space — and, at the top of the ranking, the weakened king, more frequent at the end than at the start. Reading what your opponent is preparing barely gets learned at all.
What is left of each mistake in the 2000-2399 band (base 100: the 400-799 band)
Means of per-player rates. Only mechanisms with meaningful volume appear: at least 2 occurrences per 100 games in the 400-799 band. The pawn race is measured against the races actually contested.
Zooms: a few mechanisms up close
Three pairs of mechanisms, chosen for this edition — they may change in the next one —, followed from band to band. Each pair sets a mistake improvement erases next to one it leaves in place. Occurrences per 100 games, means of per-player rates.
The hanging piece fades, the bad trade holds on
Two material mistakes, two fates. The piece left hanging drops from 22.8 to 7.0 occurrences per 100 games between 400-799 and 2000-2399 (divided by 3.3). The bad trade — giving up an active piece for a passive one, or simplifying in a position that punishes it — only drops from 19.7 to 12.2 (divided by 1.6). Seeing a hanging piece gets learned by playing; judging a trade, far less.
Hanging piece and bad trade (occurrences per 100 games)
Means of per-player rates, by 400-point rating band. Primary label only.
In the opening: the queen gets fixed, castling lags
The early queen is the mistake that fades fastest in the corpus: 3.3 occurrences per 100 games at 400-799, 0.5 at 2000-2399 (divided by 6.6). Late castling only drops from 4.5 to 2.5 (divided by 1.8): at 2000, players still castle too late almost as often as at 1200. The first is a reflex that gets corrected; the second, a judgment call players keep getting wrong.
Early queen and late castling (occurrences per 100 games)
Means of per-player rates, by 400-point rating band. Primary label only.
The two most frequent mistakes are also the ones that hold on
The missed defense and conceded material are, by far, the two most frequent mechanisms in the corpus — and the two that recede the least. The missed defense drops from 123.8 to 75.8 occurrences per 100 games (divided by 1.6), conceded material from 80.7 to 54.2 (divided by 1.5). At 2000-2399, the two together still add up to more than one mistake per game. In two curves, that is the result of this edition: what you hand to your opponent barely gets learned at all.
Missed defense and material conceded (occurrences per 100 games)
Means of per-player rates, by 400-point rating band. Primary label only.
The endgames
One game in four reaches an endgame, and most enter it already decided — more than three pawns apart. The contested endgame is rare; when it exists, it decides.
| Rating band | Games reaching an endgame (%) | Endgames entered already decided (%) | Contested endgames (% of games) |
|---|---|---|---|
| 400-799 | 19.7 | 72.8 | 2.7 |
| 800-1199 | 23.5 | 70.8 | 3.9 |
| 1200-1599 | 24.8 | 65.9 | 4.7 |
| 1600-1999 | 27.7 | 57.8 | 6.3 |
| 2000-2399 | 29.3 | 49.9 | 9.0 |
The cost of the weakened king
A player’s score counts 1 point per win, ½ per draw, 0 per loss, per 100 games. For each player, we compare their score in games WITH a weakening of their castled king to their score in games WITHOUT (at least 5 games on each side), then average those gaps: −10 means a player scores on average 10 points less per 100 games when they weaken their king — as if 10 wins in 100 turned into losses. A measured correlation, not a proven cause.
| Rating band | Weakened king (/100 games) | Score gap (points per 100 games) |
|---|---|---|
| 400-799 | 13.6 | -8 |
| 800-1199 | 16.2 | -10 |
| 1200-1599 | 16.9 | -7 |
| 1600-1999 | 13.8 | -10 |
| 2000-2399 | 14.6 | -10 |
Methodology and limits
The corpus
52,005 finished, analyzed games (chess.com and Lichess), played by real players rated 400 to 2399. Detection: 38 named mistake mechanisms, engine-verified, with anti-false-positive guards re-read on the board.
Why bullet is excluded
Under an estimated three minutes (base + 40 × increment), a share of the recorded mistakes owes more to the clock than to understanding — and bullet is not played equally at every level. 12,149 games are excluded on that ground. Unreadable time controls (correspondence, over-the-board games) are kept: they are anything but bullet.
Weighting
The statistical unit is the player × band pair (the player’s rating in the game, 400-point bands), kept from 20 games. Aggregates are means of PER-PLAYER rates — never a pool of games, which would let one heavy player weigh like ten.
One estimator, checked under two
Every published figure is a mean of per-player rates. While preparing this edition, every claim was also recomputed under per-game counting: whatever did not hold under both calculations was not published.
The primary label
The analysis assigns each faulty move one PRIMARY mechanism, and sometimes secondary ones. All published frequencies count the primary label only: the same move never counts twice.
The "still winning" rule
A piece given up by a player whose position stays winning after the loss (at least +3) is counted as a simplification, not a mistake. The rule lives inside the detectors themselves; it excludes 6.6% of cases, and applying it raises the savable share — the excluded cases mostly had no save.
Normalizing by opportunity
The pawn race is the only mechanism whose denominator moves with the level: the higher you climb, the more you reach the endgame. Per 100 games the mistake seems to rise; per 100 games where a race actually happened, it recedes. The second figure is the published one (measured August 22, 2026).
What we cannot name
On a share of faulty moves, none of our named mechanisms applies: they are classified "undiagnosed" and enter no figure on this page. Publishing that limit is part of the method.
Correlation, not causation
The cost of the weakened king compares the same player’s scores with and without a weakening (at least 5 games on each side). The gap points the same way in every band; we do not claim to isolate the cause.
Frozen figures
The measurements in this edition were taken on August 22, 2026, on the corpus as it stood that day. The numbers on this page belong to the Fall 2026 edition and will not change. Future editions will re-measure the same indicators, with the same definitions, on the corpus of their own date; every edition’s dataset stays downloadable.