A man who has been away loses his comeback fight. Fans say it, commentators say it, and at first the numbers agree.

The population is 17,963 bouts: both men already had enough rated fights behind them, and the bout produced a decision. Draws (608) and bouts where the two ratings were identical (14) are left out. The table below counts 35,594 rows seen from a fighter's side, so one bout appears twice, once in each corner. Rows are not fights.

With a short turnaround, 52.0% (8,710 rows). After a long lay-off, 33.9% (882 rows). A drop of 18.1 points, and it matches the story everybody tells.

Look at who the rested men are fighting. Below is the average of their own rating minus their opponent's.

After a short break, +27.3 -- a shade above. After a long one, -163.5. The sign does not merely flip, it travels a long way. A man coming back from two years out is, at that moment, rated far below the man across from him. He is not losing because he rested. He is losing because he is the weaker fighter on the night.

So compare like with like. Each row below is one bout: we take only the side that was rated higher going in, and split those by whether that side was the one coming off a lay-off.

From 64.2% (4,598 bouts) to 55.4% (289 bouts). The 18.1-point drop becomes 8.8. Most of the raw figure was who these men are and who they meet. Something is still left.

Where do the remaining 8.8 points show up? Split the same bouts into the close ones and the lopsided ones.

In close bouts it falls clearly, 53.5% to 41.1%. In lopsided ones, 74.6% to 72.0% -- barely a move. A heavy favourite wins anyway, ring-dull or not. The longest band on the close side holds only 56 bouts, so read that cell as untested.

If rust accumulated, a longer lay-off would be monotonically worse. Cut it finer, favourites only.

The first three bands fall. After that: 55.3%, 56.2%, 57.5%, 53.7% -- up and down, on 315 and 127 bouts. Thin, and not a slope.

So far this has all been win rates. There is another way to look: compare the outcome with the win probability this site itself was showing before the bout. If the prediction holds up, the rating already knows about the lay-off.

Short break, +0.8 points. Longer, -5.7, then -7.2. The prediction sits above what happened.

Is that just the model's ordinary wobble? Split the same rows at random, on something with nothing to do with rest, and the two halves come out at +0.2 (17,922 rows) and -0.1 (17,672 rows) -- a spread of 0.3 points. Across the lay-off bands the spread is 8.8 points, about 29 times larger. Vary the start year and the minimum experience six ways and the longest band stays between -7.2 and -12.0. The sign never turns.

This rating lets its confidence decay while a fighter is idle. A man long away drifts back toward a state where the rating will not commit either way, and as confidence falls the prediction moves toward an even split.

So the model was saying it no longer knew. Reality answered differently. It was not that he had become unknown. He had become worse.

Uncertainty and decline do not point the same way. The first pulls the prediction toward the middle; the second pulls the result to one side. This rating carries only the first, so it systematically overrates a fighter coming back. That is the point of this article.

They are easy to confuse, so plainly: a win rate and a gap from a prediction are different quantities.

The win rate looks like it stops falling in the long bands because we are looking only at favourites there. A man still rated above his opponent after two years away is already an unusual man. The gap, by contrast, is measured on every row, and it keeps widening. One flat and one falling is not a contradiction; they count different things.

The remainder cannot be called ring rust. Age is not measured here, nor injury, nor a change of weight class. A man who sits out for years is that much older, is often coming back from surgery, and has sometimes moved division. What is above is the sum of those, not a breakdown of them.

The longest band is thin in every table: 289 bouts on the favourite side, and 152 under the strictest of the six settings. Not thin enough to move the order of magnitude, not thick enough to trust a first decimal.

One thing to take away, then. When you read the rating of a fighter here who has been out a long time, treat it as generous. The model converts rest into uncertainty; reality answered with decline. Over the range measured here the prediction sits 7.2 points above what happened in the longest band, and the displayed number is safest read that much lower.

The figures in this article are based on data as of 23 September 2026. The data is updated weekly, so the last digit may differ from the current tally.