Methodology
How Fight Forecast makes, locks and grades its forecasts — in plain language first, with the technical detail underneath.
What Fight Forecast predicts
For each UFC bout with enough data, Fight Forecast estimates the probability that each fighter wins. Some bouts also get separate estimates of whether the fight goes the distance and how it ends (KO/TKO, submission or decision).
What a probability means
A 65% estimate means that, across many similar fights, the favourite should win about 65 times in 100 — and lose about 35. A single upset does not make a forecast wrong, and a single win does not make it right. That is why the record tracks accuracy, log loss and Brier score over many fights.
How forecasts are locked
Before each event, every official forecast is saved with a timestamp and the exact model version that produced it. Locked forecasts are never edited. If a bout changes after locking, the old forecast is not scored and no replacement is created after the fact.
Who gets an official forecast
Both fighters need at least 3 prior detailed UFC fights. Bouts below that line get no official forecast. Some of them get an experimental limited-data estimate, labelled as such and tracked separately.
Data and model limitations
- The models use recorded UFC statistics only: no injuries, weight cuts, camps or short-notice context.
- Fighters with few UFC fights are described with more uncertainty.
- Live records are still small; early numbers will move.
- Method and finish estimates are unavailable when the scheduled fight length has not been verified pre-event (by an explicit official listing or a verified pre-event record). Fight length is never inferred from card position or title status.
- Ratings are descriptive and not opponent-adjusted.
Evidence and reliability
Each matchup shows how many prior detailed UFC fights each fighter has. Fewer fights means less evidence. Fight Forecast does not attach "confidence" badges to forecasts; the probability itself is the statement of uncertainty.
Winner versus method and distance
The winner estimate is the official forecast. Method and distance are separate components with their own, smaller records. The six-way outcome breakdown (who wins × how) is an approximation combining the two.
Outcome simulation
The Lab's Outcome Simulator takes the model's probabilities and draws thousands of random fights from them to show the spread of outcomes. Running more simulations does not make the model more accurate.
Generative fight simulation
The Lab's Fight Engine is a different, experimental approach: it builds fights action by action from fighters' past tendencies. Its results can disagree with the official forecast and with the Outcome Simulator. It is not used for official forecasts.
How results are tracked
After each event, every official forecast is graded against the official result. Excluded bouts, draws, no contests and bouts that changed after locking are listed but not scored. Records from different model versions are kept separate.
Post-fight reviews
Some bouts get a written review: what happened, what the model got right, what it missed and the takeaway. Reviews are written after the result and never change the locked forecast or the record.
What "experimental" means
Experimental features — limited-data estimates, method/distance components and everything in the Lab — are shown for transparency and research. They are clearly labelled and kept out of the official record.
Fighter ratings
Fighter ratings are descriptive 0–100 scores for seven specific areas, built from recorded UFC statistics. They are not rankings, not win probabilities and not inputs to the winner model. Areas are never combined into one number, Ratings are domain-specific descriptive scales; equal point differences across different domains do not imply equal absolute skill differences, and missing values stay missing.
- Ratings V2: striking (offense), power, striking defense, takedown defense and wrestling. Fighters with fewer than 5 detailed UFC fights are marked provisional. Wrestling is shown as a provisional measure for every fighter because its public scale has not passed review.
- Chin and Cardio (R1 research): frozen research candidates published exactly as frozen on 17 Oct 2026, on a separate research scale, each with its exact evidence label. Fights after October 17, 2026 are not reflected, and they are never filled in from other ratings. Chin is recorded exposure-adjusted knockdown resistance, not literal biological chin or recovery. Cardio is later-round striking activity and effectiveness retention, not direct physiological fitness.
Model history
| Model | Role | Status | Live record |
|---|---|---|---|
| Winner C-Primary v1 | Primary winner model | Live (prospective) | 5-2 |
| Winner V5 | Backup for bouts the primary model cannot cover; previous primary model | Retired as primary on 2026-10-27 | 17-18 |
| Limited-data estimate | Experimental estimate when a fighter has fewer than 3 detailed UFC fights | Experimental | ?-? |