Skip to content
ChessOpticAnalyze your last 100 games for free

What is chess accuracy, and how do we calculate it?

At a glance.

Chess accuracy summarises the quality of your decisions as a score based on engine analysis. In a ChessOptic test using 461 games with no repeated player accounts, the average difference from Chess.com accuracy was 3.4 points. This is experimental validation, with worse results below 1000 and in very short or early-decided games, as well as differences by time control.

What does a game’s accuracy score mean?

Accuracy tries to summarise how close your decisions were to the options the analysis considers best. It offers a quick overview, but the same score can describe very different games: one with easy decisions and another full of difficult threats.

Read the score alongside the board. High accuracy can accompany a comfortable win or a loss in which you found many good moves but missed one decisive moment. Finding that moment and understanding your alternative gives you something to practise.

How do we calculate our accuracy estimate?

The method we studied starts with Stockfish. The engine evaluates positions so we can compare your move with other possibilities. We then summarise how much the position worsens after your decisions and how those changes are spread through the game.

The estimate considers features of the game such as the largest losses, its length and positions with few choices. A previously fitted formula turns those summaries into a score between 0 and 100. We kept the formula fixed in the validation test to check it on new games.

This describes research: the experimental score is not yet part of the free report. The report focuses on recurring mistakes, examples from your games and alternatives you can understand.

What does a 3.4-point average difference from Chess.com mean?

We compared our estimate with Chess.com’s reference scores in 461 games. For each game, we measured the distance between the two scores without letting overestimates cancel out underestimates. Those distances averaged 3.4 points on the 0-to-100 scale.

For example, scores of 80 and 84 are four points apart. That illustrates the measure; it is not a result from a study game. The average also does not mean every game was within 3.4 points of the reference.

Results were less favourable below a rating of 1000 and in very short or early-decided games. Coverage also varied by time control and rating. The full sample and figures are in “Study details” below.

Why might the score differ from Chess.com or Lichess?

Each method chooses how to analyse and summarise a game. It may consider different positions, engine analysis or weights for decisions. Our test compares with Chess.com; it does not officially reproduce its formula, and the result cannot be carried over to a comparison with Lichess.

Interpret a measure in its own context. Comparing two tools’ scores without looking at how they are calculated can lead to the wrong conclusion about a game or your playing strength.

Why do you need more than a score to improve?

A score answers “how can this game be summarised?” Practice needs another question: “which decision keeps recurring, and what can I check before moving?” A number alone does not show an undefended queen or explain why a pinned pawn could not recapture.

ChessOptic analyses up to 100 public games from Chess.com or Lichess to find recurring mistakes and show your own examples. Estimated impact helps you prioritise; the explanation and alternative give you a position to practise. This adds a view across games to individual game review.

How can you turn a score into a useful study session?

  1. Choose an important decision instead of trying to raise the score in the abstract.
  2. Set up the board before your move and find the enemy reply that changes it.
  3. Compare an alternative and explain which threat it avoids or opportunity it uses.
  4. Replay without the solution and carry that check into your next reviews.

Start with how to stop hanging pieces, missed captures or evaluation drops. The ChessOptic methodology explains how we select your game examples.

Common questions about chess accuracy

Does high accuracy mean I played like a grandmaster?

The score depends on the decisions that game demanded. One number does not establish an Elo equivalent or a title-level performance.

Can I lose with higher accuracy than my opponent?

Yes. One decisive mistake can change the result even if the rest of your moves were well assessed. Review the important moments as well as the overall score.

Does a 3.4-point difference mean 96.6% accuracy?

No. It is an average distance between two scores, not a percentage of correct moves or a guarantee for each game.

How can I use the report with my child or coach?

Pick a position and explain your idea when you played it. Then compare the enemy reply with the alternative. The aim is to understand a decision rather than judge a player by a number.

Study details

The test used 461 games from October 2026 and 922 distinct accounts, with no account repeated on either side and no detected overlap with historical study accounts. Independence was checked between identifiable accounts; one person may hold several accounts. One side per game was chosen by a rule fixed before analysis, so the main result has 461 observations.

The target was 600 games spread across time controls and ratings. We obtained 461, with uneven coverage: 23 of 24 groups were represented, and 15 reached the target of 25 games. Every game had at least ten moves per side and Chess.com reference scores available for both players. Selection covered only one month; the results describe that sample.

The mean absolute difference was 3.381 points, rounded to 3.4 in the main explanation. The 95% confidence interval for that mean was [3.077; 3.735], or [3.08; 3.74] rounded. The median error was 2.599 points and the 90th percentile was 6.671. The interval estimates the mean; it is not a maximum error for an individual game.

Each cell shows number of games / mean absolute difference in points. Rating bands follow the study selection and do not establish equivalences between platform ratings or FIDE ratings.

Protocol rating band Bullet Blitz Rapid Daily Total
<1000 25 / 4.62 25 / 3.56 25 / 6.90 21 / 3.06 96 / 4.60
1000–1399 16 / 2.22 25 / 3.18 25 / 3.36 18 / 3.19 84 / 3.06
1400–1799 14 / 3.44 25 / 3.56 25 / 4.02 12 / 3.08 76 / 3.62
1800–2199 25 / 3.08 25 / 2.61 25 / 3.12 2 / 1.87 77 / 2.91
2200–2500 25 / 3.35 25 / 2.85 25 / 2.45 2 / 2.00 77 / 2.86
>2500 25 / 3.16 25 / 2.51 0 / — 1 / 0.02 51 / 2.78
Total 130 / 3.38 150 / 3.05 125 / 3.97 56 / 2.97 461 / 3.38
Bullet and daily games were part of the study; they do not expand the free offer’s time controls described on this page. In rapid, the mean difference was 3.970 points; in rapid below 1000 it was 6.897, from 25 games. There were no rapid games above 2500, and the higher-rated daily groups contained only one or two games each.
Diagnostic subset n ridge12 MAE, points Absolute error P90, points
Games of up to 40 plies 81 4.343 7.273
Decided early under the PREC-04 definition 31 7.780 18.030
With mate in the evaluations 225 4.002 8.899
Rating below 1000 96 4.595 10.306
These subsets overlap: they show associations, not isolated causes. “Up to 40 plies” means up to 40 individual moves by either side, equivalent to 20 full moves. Early-decided games used the study’s previously established definition.

The experimental formula stayed fixed: ridge12 / simple12 / alpha=10, using twelve summaries of one side’s evaluations and decisions. It was not refitted after seeing these 461 games. The study used Stockfish 19 with cheap50k: 50,000 requested nodes per position, one thread, 256 MiB Hash and one principal variation. These are experimental settings, not a promise of free-report speed or performance.

The twelve summaries cover four raw scores; mean, 90th percentile and maximum loss; mean evaluation loss; length; proportion of decided positions; forced moves; and volatility. A calibrated sum produces a score between 0 and 100. Resampling for the intervals used 2,000 replicates at game level, with seed 14112026. The secondary analysis of both sides gave a 3.292-point mean difference across 922 observations grouped by the 461 games; it does not replace the main measure.

The published results are aggregates. Private study data were deleted at the close on November 17, 2026; the independent review used the evidence retained at that time.

Analyze your last 100 games for free.