How the model works

Every prediction comes out of two layers. The first — Model A — looks only at the two fighters themselves: their histories, their styles, their physical matchup. It never sees a betting line, which is what makes it usable for a hypothetical matchup that hasn't been booked yet — there's no market price to look at for a fight that doesn't officially exist. The second layer — Model B — only switches on once a fight is actually scheduled and priced by sportsbooks. It blends Model A's read with the market's own opening and closing lines, the same way a trading desk treats a competitor's price as one more input rather than ignoring it. That blended view is what powers the Model vs. Vegas page; the fighter-only view is what powers the Predict page, where no market price exists.

Every matchup is rebuilt as of the fight date, using only what was true at that moment. A fighter's win rate, striking numbers, and form are computed from their fights up to that point — never from their full, eventual career record. This matters more than it sounds: without it, a model could accidentally "know" that a fighter was about to go on a title run, which is exactly the kind of hindsight a real prediction never gets. Every matchup is also evaluated from both corners and averaged, so the order two names are typed in doesn't quietly tilt the read one way.

The model was given roughly 300 candidate signals and kept the ones that actually earned their place through repeated out-of-sample testing — not the ones that merely sounded plausible. In practice, a handful of themes carry most of the weight:

Age and experience gap. The single strongest signal in the whole model. Fight sports have a real athletic prime, and a meaningful age or experience mismatch between two fighters is one of the most reliable predictors of who wins.

Quality of competition, not just win/loss record. A perfect record against weak opposition reads very differently from a .500 record forged against ranked contenders. The model tracks each fighter's Elo rating — adjusted for how tough their opponents were — rather than a raw win percentage, so a hot streak against undercard-level competition doesn't get mistaken for genuine improvement.

Striking output and efficiency, in both directions. How much significant striking volume a fighter lands and absorbs per minute, and how efficiently — not just gross volume, but volume relative to what they take back. A fighter who both hits hard and has a track record of not getting hit rates very differently from one who wins a lot of close decisions while absorbing heavy damage.

Grappling control and finishing threat. Takedown rate and defense, submission attempt rate, and time spent in dominant positions all feed a composite read on how much a fighter can dictate where a fight takes place — on the feet or on the mat — and how likely they are to finish it there.

Physical matchup. Reach and height differentials, and how a fighter's physical tools interact with the specific opponent in front of them, not just in the abstract.

Ring rust and activity. Time since a fighter's last fight. Long layoffs — from injury, contract disputes, or matchmaking gaps — show up as a real, measurable drag on performance.

Style matchups. Rather than reducing each fighter to one overall skill number, the model keeps a separate offensive and defensive profile for every fighter — inspired by the “blade and chest” framing from competitive-game theory — so it can represent things a single rating can't, like a durable pressure-fighter who struggles specifically against elusive counter-strikers. It's a more formal version of what fight fans already do instinctively: styles make fights, and the better fighter on paper isn't always the better fighter on the night.

Once a fight is booked and sportsbooks post a line, a second, much simpler model blends three things: Model A's fighter-only read, the market's opening price, and the market's closing price. Closing lines in particular tend to be hard to beat — they reflect all the money and information that moved the market between open and close — so treating the market as a signal rather than ignoring it tends to produce a better combined estimate than either input alone. This blended number, not the fighter-only one, is what's shown on the Model vs. Vegas page.

A separate model estimates how a fight is likely to end — decision, KO/TKO, or submission — using the same as-of-date fighter data, plus each fighter's own finish and finish-rate history. It runs independently of the win-probability model, so it's possible (and common) for a heavy favorite to still carry a real finish risk either way.

Every version of this model is fit strictly on fights before a cutoff date and evaluated only on fights after it — a genuine holdout, never seen during training. The model actually running this site today is trained on fight data through 2026-07-11 and gets retrained as new results come in, which means it's always working from the most recent data available but no longer has an untouched holdout of its own to score itself against — that honest, never-touched evaluation happens in a separate internal copy. For the actual, current numbers — including where the model does and doesn't beat the closing line — see the Track Record page, which is transparent about exactly what is and isn't a blind test.