How it works

Plain answers, no secret sauce.

The data

Every UFC fight since 1994, round by round: significant strikes landed and absorbed, takedowns, control time, knockdowns and submission attempts, plus each fighter's height, reach, stance and age.

The model

For each fighter we build a profile from their previous UFC fights only: a rating that rises and falls with every result, striking output and defence per minute, wrestling numbers, finishing rate, how they've been finished, current streak, recent form and time since their last fight. The model compares the two profiles and gives a win chance.

Current model: v4 (November 7, 2026). v2 added strength of schedule, five-round and title-fight experience and recent knockout losses; v4 adds each fighter's record outside the UFC (regional promotions, Contender Series) from Wikipedia, which matters most for newer fighters. (v3, a round-by-round cardio idea, didn't improve the test and was dropped.) Every pick on the record shows which model made it.

Fighters making their UFC debut have no UFC history yet, so those predictions are close to 50% and marked low confidence.

How it ends

For KO/TKO, submission and decision we multiply each fighter's win chance by how likely that win is to come each way. That part looks at the winner's tools (knockdowns, striking volume, submission attempts, takedowns) against the other fighter's durability (strikes absorbed, defence, how often they've been stopped), plus the division: heavyweights get knocked out far more often than flyweights.

Tested the same way as the win model on 2,506 fights (2022–2026), it was more accurate than plain division averages in 5 of 5 years. Over that period 33% of fights ended by KO/TKO, 18% by submission and 48% on the scorecards.

The market, and how it moves

Next to our number you'll see the sportsbooks' price with the bookmaker margin removed (Pinnacle where available), the range across books, and the prices on the prediction markets Kalshi and Polymarket when they're quoted tightly. We also show where the line opened when we first saw it, and which way it has moved since.

A check we run on 2022–2026 UFC fights: when our model picked a different winner than the opening line (659 fights), the line moved toward our pick 62% of the time and away 31%. That's not simply lines drifting toward 50/50: across all fights, lines moved away from 50/50 more often (49%) than toward it (38%).

It is not a betting edge. By fight night the price has usually caught up, and after the bookmaker's margin, betting our disagreements at the opening price still lost money in our test. It means our model tends to see early what the market works out later in the week.

How we test it

• No peeking. To test a year, the model is trained only on fights before it.
• Calibration. A 70% chance should win about 70% of the time. We publish that check.
• Against the market. We compare with the bookmakers' closing odds, the hardest benchmark there is, and publish the result even though the market currently wins.
• Locked picks. Live picks are logged with a timestamp before the fight and never edited.

What we don't do

We don't take bets, sell "locks", or promise profit. CageIQ is analysis. If you bet, bet only what you can afford to lose.