How much of a UFC result is the matchup? What 194 rematches show

Beating someone once says less than you’d think

I found 194 UFC pairings that met more than once and had a winner both times. Whoever won the first meeting won the next one 59% of the time. Include every professional fight on the records of UFC fighters, regional cards included, back to 1980, and you get 2,499 rematches and a repeat rate of 65%. The wider sample runs higher because regional matchmaking produces more mismatches.

How the first fight ended matters more than who won it:

After a split or majority decision, the first-fight winner won the rematch 48% of the time. The first result carried no information about the second. Time wears the rest of it down too: rematches within 18 months repeated 69% of the time, later ones 61%.

The betting market already treats a past win as thin evidence. In the 55 UFC rematches since 2021 that have closing prices, it made the first-fight winner a 54% favorite on average. Here’s a useful benchmark: if every UFC fight since 2021 were fought twice and the closing prices were exactly right, the same fighter would win both times about 57% of the time. UFC rematches repeat at close to that rate. So having watched two men fight each other already leaves you about as sure of the next result as the closing line is about an ordinary fight. No surer.

Pulling the pairing apart from the fighters

A repeat rate mixes three things: who is better in general, whatever is specific to the pairing, and what happened on the night. To get at the pairing, I predicted each rematch fight from how each man performed against everyone else around the same time, adjusted for who those opponents were. Then I looked at the leftover: how far the actual fight landed from that prediction. The measure is significant-strike margin per minute. It’s recorded throughout a fight rather than once at the end, so it carries more information per fight than a win or a loss.

If the pairing matters, the leftover should repeat. A fighter who did better than expected against this particular opponent the first time should tend to do better than expected the second time too. For a control, I ran the same calculation on the same fighters against two different opponents at the same points in their careers. If the leftover only repeats because a fighter happened to be in good form, the control picks that up.

Across 182 usable rematches, the leftover in the strike margin repeated from the first meeting to the second (correlation 0.17). For the same fighters against different opponents it didn’t repeat at all (−0.04 and 0.00). The gap between those two numbers is what belongs to the pairing. It held when I measured “form” with a nine-month half-life and with a three-year one. Split into shares, the fight-to-fight variation in strike margin comes from here:

Two things stand out. The pairing is not small: its best estimate is about two-thirds the size of the general skill gap. And both are small next to what happens on the night. The pairing estimate is also loose. With 182 rematches it could be anywhere from about zero to 30%. For the win itself the best estimate comes out in almost the same ratio (17% skill gap, 10% pairing), but a single win or loss is too blunt a measurement for this few rematches to tell the pairing apart from form.

Why no fighter’s record can show it

Suppose the pairing effect really is that big. Could you ever measure it for a specific pair? With the night contributing six times as much variation as the pairing, the same two men would have to fight about six times before their head-to-head record was more signal than noise. Only one pairing in UFC history, Brandon Moreno and Deiveson Figueiredo, has met even four times.

The usual workaround is to learn from similar opponents: how does he do against pressure wrestlers, or against long kickers? That hits a second limit, which is how long it takes any number about a fighter to mean something. I took the 614 fighters with at least ten UFC fights, split their first ten into odd and even fights, and checked how well each half predicted the other. From that you can work out how many fights each stat needs before it’s half signal and half noise:

The split is clean. Numbers that describe what a fighter does (how often he shoots, how much he throws, how long he holds position) settle within about three to eight fights. Numbers that describe how well it works take 11 to 27. For fighters who debuted between 2005 and 2016, the median UFC career lasted five fights. Only 27% reached 11, and only 10% reached 19. Raw win rate is the worst of the lot: a fighter’s record in his first five UFC fights has almost no relationship to his record in his next five (correlation 0.04), because each win gets him a harder opponent.

Now apply that to style matchups. A fighter’s results against one kind of opponent are a slice of results that already need 11 to 20 fights to settle. Almost nobody has enough fights against any one kind.

Where a matchup read has to come from

Put together, this rules some things out. The part of a fight that belongs to the pairing is real, and plausibly close in size to the skill gap. But it can’t be read off either man’s record, including the record between them. The only data about a fighter that settles in time describes his tools. It says nothing by itself about what happens when those tools meet a particular opponent’s tools. So a real matchup read has to come from a model of how styles interact, learned across the whole population and then applied to two style profiles. That is roughly what a sharp analyst is doing when he says one fighter’s defense was built for a certain kind of attacker. The data says there is something there to find.

It also says how such a read should be checked: against the part of rematches that repeats, and across many fights. A single fight can’t confirm it. A 21% shot like Gaethje’s comes in about one time in five even when there is no matchup edge at all.