Fighters who have never lost, and fighters who have, lined up under conditions that are identical except for that.
The most expensive number in boxing is not the win column and it is not the knockouts. It is the zero in the loss column. It is talked about as something to protect, and bouts get chosen to protect it.
So what does that zero actually forecast? The population is bouts from 1970 on, by fighters whose record table starts at their debut, against opponents with at least 5 prior bouts, decided one way or the other. That is 5,847 corner rows on the unbeaten side and 13,258 on the other.
Start with the plain win rate: bouts entered unbeaten, split by how long the unbeaten run had got by then.
A corner row is one side of one bout, so a bout makes two of them. It is not a bout count.
80.3% overall. It is a big number and it says nothing. They are unbeaten because they are good; whether being unbeaten makes them good is not written anywhere in this table.
The shortest band is a fighter a few bouts into his career. His rating has barely moved, so nothing after this rests on him.
This site publishes a win probability before the bout. Measured against that, the part explained by "he is simply better" has already been subtracted. What is left over is what the rating missed.
Unbeaten men come in +12.3 points above their prediction and beaten men at -1.4, a gap of 13.7 points. Splitting the same rows on something meaningless -- a hash of the date and the fighter -- moves the answer by 0.00 and 0.50.
Corner rows in brackets.
Raising or dropping the opponent floor does not remove it: the gap runs +10.8 to +14.9. Whatever else is arguable here, this much holds under every cut.
The first objection is the obvious one: he is unbeaten because somebody is picking his fights.
It runs the other way. The opponent's mean rating climbs from 1524.4 to 2153.4 and his mean prior bouts from 10.8 to 28.9. The longer the unbeaten run, the harder the night.
Corner rows in brackets.
Holding the rating gap fixed says the same thing. All five bands point the same way, +8.8 to +23.1, and the widest of them is the band where the two men are level: +23.1.
An unbeaten fighter is, by definition, on a run. So: does a fighter who has lost, but is on a run of the same length, beat his prediction by the same amount?
Corner rows in brackets.
Most of it, yes. The gap shrinks from +25.3 to +5.4 as the streaks lengthen. Most of the unbeaten premium is just being on a run. It does not reach zero, though: compare equal streaks and the man who has never lost is still ahead.
The next objection: he is not special, the rating simply does not know him yet. RD is how unsure the rating is, so hold it fixed and ask again. The bottom band's beaten cell stands on only 38 corner rows, barely over the threshold of 30; it is a thin cell.
Corner rows in brackets. A cell standing on fewer than 30 of them is treated as not measured.
With the opponent floor at 5 or above, all four bands point the same way. But at no floor at all, which is much the largest population (32,668 and 42,822 corner rows), 2 of the 4 bands turn over: +10.1, +5.1, -1.6, -7.7.
Nothing in this data decides which of those is right. So this article does not claim that the gap survives holding RD fixed. The gap between the groups holds under every cut; the RD-controlled version of it does not, and declining to write the second one is part of the result.
In the previous column, this site's rating was too kind to fighters coming back from a long lay-off: men who had been away lost more often than the number said they would.
Here it is the other way up. It is too hard on the unbeaten. Opposite direction, same defect.
A rating can only be built out of the past, so it cannot keep up with a fighter who is still changing. It is harsh on the ones on the way up and kind to the ones on the way down. The zero was never magic; it was a marker that someone was still climbing.
Age is not measured here. Nor is weight class, nor the quality of the card. What this gap would look like with those in it is not something this article knows. It is a finding about the rating, not about the fighters.
When you look at a predicted win probability for an unbeaten fighter on this site, read it as an underestimate. The rough size is 13.7 points -- win-rate points, not rating points.
But the same thing happens to a beaten fighter on a long run (+5.7). Do not read the zero. Read whether he is still climbing.
Measured 2026-09-26