How to Model UFC Fights for Expected Value
Modelling UFC fights for expected value means estimating each fighter's win and finish probabilities from fight data, comparing them with devigged market prices, and acting only where the gap clears the margin. The hard parts are small samples per fighter, style interactions, judging, and the much larger margins in method of victory markets.
What the question really involves
Software that promises to find expected value in UFC fights is doing, or claiming to do, three things: estimating fair probabilities for each outcome, removing the margin from available prices, and flagging gaps. None of that requires a particular product. It requires data, a model that respects how little data each fighter has, and an honest validation loop. This page covers the method, not any vendor.
The data that exists
UFCStats.com is the standard public source for UFC fight statistics. Fight pages break down totals and round by round figures for each fighter: knockdowns, significant strikes landed and attempted, significant strikes by target (head, body, leg) and position (distance, clinch, ground), total strikes, takedowns landed and attempted, submission attempts, reversals and control time, plus the method, round and time of the finish. Many open source scrapers and public datasets are built from it. When checked on 2026-11-05, the site returned a JavaScript browser check to non-browser requests, so scripted collection may need a different approach; check the site's terms before automating anything.
Other useful sources cover what UFCStats does not. Fight record databases list bouts outside the UFC, which matter for newer fighters. MMADecisions.com archives judges' scorecards along with media scores, which supports modelling decisions and judge tendencies. BestFightOdds shows prices from multiple books, including method of victory and round props, with an archive of past events. Physical attributes such as height, reach and age are widely published but vary between sources and should be reconciled.
Small samples are the central problem
A typical UFC fighter has a handful of bouts in the promotion, each lasting at most fifteen or twenty five minutes. Per-minute rates such as significant strikes landed per minute swing wildly over three or four fights, and a single short finish can distort them. Treat raw rates as noisy measurements, not as traits.
The standard fix is shrinkage. Pull each fighter's rate toward a prior based on the division, the stance or the fighter's pre-UFC level, weighted by minutes observed. A hierarchical Bayesian model does this formally; a simple weighted average of the fighter's rate and the division mean does much of the same job. Opponent adjustment matters too: landing well against weak opposition says less than landing modestly against strong opposition. Rating systems in the Elo family handle that implicitly and make a reasonable baseline.
Style matchups without overfitting
Fight fans think in matchups: a strong wrestler against a striker with poor takedown defence, a southpaw against an orthodox fighter, a pressure fighter against a counter striker. These interactions plausibly matter. The problem is that each specific combination appears in very few fights, so a model with many interaction terms will find patterns that are noise.
Keep interactions few and structural. Takedown offence against opponent takedown defence is a defensible pairing, because one measures the other directly. Grappling control against opponent get-up rate is another. Anything more specific should earn its place through out of sample improvement, not because it explains a memorable upset.
Other factors the stats miss
Short notice replacements, missed weight, long layoffs, moves between weight classes, age, and fights at altitude all change outcomes in ways box scores do not capture. Most have small samples of their own. Encode them as simple adjustments with wide uncertainty, and check whether the market already prices them before assuming an edge. Late changes are also where prices move fastest, so data timestamps matter.
Method of victory markets
A full method market splits each fighter's win into knockout or technical knockout, submission and decision. A model must produce these as a consistent set: the three method probabilities for each fighter should sum to that fighter's win probability. Building the win probability first and then splitting it by finish rates, shrunk toward division averages, keeps the numbers coherent.
Margins are much larger here. Take an illustrative card where the moneyline is -160 and +135, an overround of 4.1 percent, and a multiplicative fair favourite probability of 59.1 percent. Suppose the method market prices the favourite by knockout +160, by submission +700, by decision +275, and the underdog by knockout +550, by submission +1400, by decision +400. Those implied probabilities sum to 119.7 percent, an overround of 19.7 percent.
The favourite by decision at +275 breaks even at 26.7 percent, while the multiplicative fair estimate is 22.3 percent. A model must sit 4.4 points above fair to break even, against 2.4 points on the moneyline. The two markets also disagree with each other: devigged multiplicatively, the method market implies the favourite wins 64.9 percent of the time, not 59.1. Longshot outcomes absorb most of the margin, so the method used to devig matters a great deal in these markets.
Judging and decision modelling
Decisions are scored round by round by three judges on the ten point must system. Round-level data from UFCStats lets a model estimate who wins each round, and scorecard archives show how individual judges have scored close rounds. Split decisions are common enough that decision props deserve a model of judging noise, not just a model of fighter performance.
Validating the model
Score the model's probabilities with log loss against devigged closing prices, and compare its calibration with the market's on the same fights. Track closing line value on every flagged price. The UFC holds several hundred bouts a year, so a season of results is a thin sample for judging profit, while closing line value accumulates faster. Machine learning methods, including the newer models often labelled as AI, do not remove the small sample problem; they make overfitting easier. This page describes modelling methods and is not betting advice.
Frequently asked questions
- Where can I get UFC fight data for modelling?
- UFCStats.com publishes fight and round level statistics for UFC bouts, including significant strikes by target and position, takedowns, submission attempts and control time. Fight record databases cover bouts outside the UFC, MMADecisions.com archives judges' scorecards, and BestFightOdds shows multi-book prices including method of victory props. Check each site's terms before scraping.
- Why is MMA harder to model than team sports?
- Each fighter has few bouts, each bout is short, and a single strike can end it, so per-fighter statistics are noisy. Matchup effects are strong but rarely repeat, and decisions add judging noise. Team sports offer many more repeated observations of the same players and teams, which makes their rates far more stable.
- Are method of victory bets better value than moneylines?
- Not by default. In the illustrative card on this page, the method market's overround is 19.7 percent against 4.1 percent on the moneyline, so a model needs a much larger disagreement with the market to break even. Method markets can be mispriced, especially on finish rates, but the margin to overcome is several times larger.
- Can AI find expected value in UFC betting?
- Machine learning can estimate probabilities from fight data, but it cannot create information the data lacks. With few bouts per fighter, flexible models overfit easily. Any claimed edge should be validated out of sample against devigged closing prices and closing line value. A model that cannot beat the closing market's calibration has no demonstrated edge, whatever it is called.