Methodology
How it works
The winner model
Fights are predicted by an ensemble trained on pre-fight features only — Elo ratings (overall and by discipline), recent form, layoff, physical edges, and career striking/grappling rates. Every probability shown is out-of-sample: the model is evaluated walk-forward, only ever predicting fights it was not trained on, so the numbers reflect genuine forecasting, not hindsight.
Ratings & rankings
Skill ratings (0–99) and Elo are recomputed after every event. The “Resume” view shrinks a fighter’s rating toward the mean by strength of schedule (quality of opposition); “Raw” shows the pre-adjustment number. Power rankings blend Elo, schedule quality, and momentum.
Model vs market
Odds are BestFightOdds closing moneylines, de-vigged to a fair probability. The closing line is the sharpest public price in the sport, and across our history the model trails it on average. Backtesting the model’s disagreements against the close found no reliable betting edge — its apparent “value” was the favorite–longshot bias, not the model. So this site compares the two as calibration insight, and is explicitly not betting advice.
Limits
Debutants have no UFC history, so their predictions lean on pre-UFC Elo and bio alone and are flagged as such. Odds coverage thins out for older and smaller cards. Prop markets (method, rounds, totals) are predicted by the model but market prop odds aren’t ingested yet.