This site puts every boxer on one rating. A bout from 1905 and a bout from 2026 sit on the same scale. So how well does that scale predict? We measured it ourselves.
Only the rating the site held just before the bout is used. The two ratings give an expected win probability, and that is checked against what happened. The formula is the same one the site displays.
90,405 decided bouts. The 2,718 draws are left out, because a win-or-lose prediction cannot be checked against them. So are the 2,189 bouts in which the two ratings are exactly equal: with neither man the favourite, there is no favourite to have won.
Unless stated otherwise, the tables below cover the 31,881 bouts in which both men had five or more bouts already counted. Why, further down.
Rating gap against how often the favourite wins (both sides 5+ bouts, 31,881 bouts)
The wider the gap, the more often the favourite wins. The order holds across all eight bands. Even at a gap of 200 to 300, the lower-rated man wins one time in four.
Predicted win rate against the actual one (both sides 5+ bouts)
In all five bands the miss is within 1.5 points. The Brier score is 0.1425. (Always saying 50% gives 0.25; perfect is 0.)
Build the same table over all 90,405 bouts and the actual win rate comes in 2 to 9 points above the prediction in every band. The rating is underselling the favourite.
The reason is the population. 62.1% of the decided bouts have a gap of 400 or more, and most of those opponents appear in the record only once. A man nothing is known about is given the starting value, so the site treats him as an average fighter. He is in fact far below average. That difference lifts the whole table.
As a number, the Brier score is better over all bouts (0.1037), because more of them are easy. It is the five-or-more table that shows what the rating itself can do.
Favourite's win rate at a gap of 200 or more, by era (both sides 5+ bouts)
From the 1940s to the 2010s the hit rate goes from 77.2% to 90.9% and the Brier score from 0.2015 to 0.1254. But this must not be read as “there were more upsets in the old days”. The further back you go the thinner the record is, and what survives leans towards famous bouts. This table cannot separate a difference in the rating's performance from a difference in how completely the era was recorded.
The 2020s fall for a different reason: the more current a fighter is, the fewer bouts the rating has behind him.
The usual line is that one punch can end anything, so prediction is hopeless. Measured, it is the other way round.
Bouts at a gap of 200 to 400, split by how they ended (whole population)
Same rating band, and almost the same number of bouts. Even so, the favourite wins 10 points more often in the bouts that ended inside the distance. The upsets are happening in the bouts that went to the cards. Why that is, these numbers do not say.
Of the bouts lost by the man with the highest expected win probability, these are the ones both fighters' records agree on.
Also Steve Little beating Michael Nunn in 1994 (expected 97.2%), and Vasiliy Lomachenko beating Gary Russell Jr. in 2014 (94.0%).
The same calculation is run on the MMA version of this site. At a gap of 200 to 300 the favourite wins 66.4% there against 76.4% here, both among fighters with five bouts or more.
At the same rating gap, the favourite wins more often in boxing. An upset that comes once in three in MMA comes once in four in boxing. These numbers do not say why.
Nothing at all was measured about predicting a draw. 2,718 bouts are left out. A draw is a real result in boxing, but this model only answers win or lose.
Nothing can be said about women's bouts. Only 255 bouts have a recorded sex on both sides, and no band reaches thirty.
The older figures are pulled around by how completely the era was recorded. As above.
The measurement is probably flattering. Identifying opponents still carries errors we have not caught. A bout where the opponent was wrongly treated as an unknown is one the favourite tends to win, and the hit rate is that much higher for it.
What was measured is the prediction the site displays. It is not the calculation used internally when a rating is updated.
The figures in this article are based on data as of 21 September 2026. The data is updated as it comes in, so the last digit may differ from the current tally.