Comebacks by rating band: how often the side that fell behind did not lose

We took Lichess rapid games where one side fell well behind on the engine evaluation and checked how often that side still did not lose, split into 100-point bands by that player's rating. The short version: after falling 1.5 pawns behind, the share that did not lose stayed between 33.1% and 43.6% in every band with at least 30 games, and it did not move in one direction with rating. At 5.0 pawns the 2000 band came out low at 8.3%, but that cell has 36 games, too few to call a trend.

The game diagnosis explanation gives one share per time control. You may wonder whether that number holds at your own rating, so this page carries the table split by band.

The table by band

All games are rated Lichess rapid. The band is the rating of the side that fell behind first, in 100-point steps: 900+ means 900 to 999. Each cell is the share that did not lose after falling that far behind, counting wins and draws together, with the number of games in parentheses. A * marks a cell under 30 games, where the share is not reliable. Those cells are kept rather than removed, because removing them would hide where the sample runs out.

Taken together without bands, the 4,454 games give 37.5% at 1.5 pawns, 22.5% at 5.0 pawns and 20.8% at 12.0 pawns.

What we read from it

The 1.5 pawn column looks similar across bands. Every cell with at least 30 games, from the 600 band to the 2200 band, sits between 33.1% and 43.6%. More than one game in three did not end in a loss, whatever the band.

The 5.0 pawn column moves between 16.5% and 26.7% across the 900 to 1900 bands, where the sample is largest. It does not fall steadily as rating rises.

The 5.0 pawn cell for the 2000 band is low at 8.3%. The 2100 band above it is also 8.3%, but with 24 games it is a * cell. That fits the idea that stronger players close out games they lead by a lot, but one cell of 36 games cannot show it.

The 12.0 pawn column is hard to read. Even the cells over 30 games range from 10.0% to 38.5%, and neighbouring bands differ widely. The sample is too small to split this column by band.

Next to the first-behind move, the diagnosis also lists the tactics you allowed in those games. If one of them was a fork, its learn page walks through the same shape.

If you use this table next to your own diagnosis, the reading the sample supports is roughly this: in your band too, a game where you fell 1.5 pawns behind was not a finished game. The table does not predict the result of your next game.

Method

From the Lichess standard rated games of July 2026 we collected 5,000 rapid games that already carried engine evaluations, and read those evaluations instead of running an engine. Only the side that fell behind first is counted, so one game never enters twice. The 476 games lost on time are dropped, and a case counts only if at least 20 half-moves followed the first time the player fell behind. Pawn units describe the evaluation, not the actual material balance. The rating is the Lichess rating at the start of the game as recorded in the game file, so the same number on chess.com does not mean the same strength.

Limits

Two limits come with these numbers. The sample is games Lichess had already evaluated, and analysis there is requested or switched on rather than applied at random, so closely fought games may be overrepresented and the comeback shares may read a little high; that is the weakest point in the measurement. And splitting by band cuts each cell from thousands of games to tens or a few hundred. In the 400 and 500 bands and from 2300 up, every cell is under 30 games, and no cell in the 12.0 pawn column reaches 50.

Sources

• Lichess Open Database, the source games for the sample (CC0).
• Game diagnosis, the tool that finds the move where you first fell behind in your own games. The overall shares by time control and the method are in its explanation.