This site publishes a win probability for every bout. Seeing a number near even, a reader takes it to mean the two are matched. The same number also appears when the site has seen very little of either man. Not a close call, but nothing to call it with.

The population is 66,915 bouts, folded from 133,830 corner rows. How a fight ended and in which round belong to the fight, so every count below counts bouts. Bouts whose published probability sits within five points of even number 8,603, and those were cut into three equal parts by how much the site has seen of the two men.

The three groups are the same size. What differed was how the fights ended.

The counts in the table are only the bouts whose finish could be classified, so three equal groups do not show equal numbers (see the qualifier).

The barely observed side finishes more often. The rounds are shorter too: 2.44 where the site has seen plenty, 1.58 where it has not. The two ends are 31.90 points apart, and two random halves of this same corpus come apart by 3.20, so the difference is the larger of the two.

Taking every drawn bout out of the population leaves the gap at 31.80 points, which is barely a move.

The same three groups, now by how often the favoured man won.

Only the barely observed side falls below a half. The two ends are 10.30 points apart against a random-split gap of 3.88. Across all 5 settings tried, varying the earliest year and whether the promotion is known, the thin side stayed under a half.

One thing can be read from that. A probability being close to even does not by itself make the pick more likely to be right.

Of the 2,868 bouts on the barely observed side, the published probability was exactly even in 1,376. Neither man is called the favourite, so there is nothing to be right or wrong about. That is 48.0% of the group.

On the well observed side the same thing happened in 4 bouts out of 2,867.

So an even number on the screen is sometimes a judgement that the two are level, and sometimes no judgement at all. One figure, pointing at two different things.

Holding how much the site has seen and moving only the distance from even. The table is printed; nothing is concluded from it.

The middle third sits between the two throughout: 64.0% in the closest band (1,400) and 70.5% in the furthest (14,389).

This table is description. Which of the two axes matters more, and by how much, is not measured in this article.

How a fight ended is recorded as free text and the classes are written in English words. No rule matched 2,929 bouts (4.4%), and those are kept out of every rate. Of them, 2,753 are recorded outside the Latin alphabet, naming in another language the submissions and stoppages the English rules would have caught. The classification is therefore selective by source language, and every rate above carries that.

That is not what made the gap. Narrowed to English-language records alone, the two ends are 33.20 points apart, which is wider than the whole corpus gives.

No-contests do not enter this count at all; the source holds 560 corner rows carrying one.

Of the 5 surfaces looked at, 2 publish a win probability. Of those, the number that also show the reader how much the site has seen of the two men is 0.

The distinction this article has spent its length drawing is not on the screen at the moment.

When a number near even appears, weigh it against how much of a record the two men leave behind. An even call between two well documented fighters and an even call between two who are barely recorded say different things through the same display.

In the second kind the fight tends to end early and the pick tends to be wrong.

Figures as of 27 September 2026.