The full methodology behind Fight Detective's AI prediction engine
Pull fighter stats, fight results, and odds from multiple sources. Every data point is time-stamped to prevent future information leaking into past predictions.
Transform raw data into 80+ predictive features. Style matchup scores, rolling form averages, Elo trajectories, and physical advantage composites — all computed point-in-time.
Candidate features, model recipes, and calibration methods are compared on later fights using only earlier information. The selected Engine 7 recipe is then frozen into a reproducible release artifact.
Each prediction is served as the de-vigged live-market probability. The model's own estimate is kept alongside as an internal research signal — it is not a betting edge or a bet signal.
New event predictions are timestamped and sealed before results are known. Historical pages are point-in-time replays and are explicitly labelled as reconstructions, never passed off as old live picks.
Every UFC fight in the modern era. Outcomes, methods, rounds, referee stoppages — all structured and time-stamped.
Significant strikes landed/absorbed per minute, takedown accuracy, submission attempts, striking defense, control time, and more.
Dynamic Elo ratings that update after every fight. Momentum tracking across win/loss streaks and finish sequences.
Wrestler vs striker scores, grappling exploit metrics, takedown offense vs defense matchups, striking accuracy differentials, and head-hunting tendencies.
Raw stats alone don't predict fights. The model uses engineered features that capture how fighters match up against each other — not just how good they are in isolation.
Prospective predictions are sealed before the fight happens. Historical event pages are retrospective point-in-time replays: the engine is asked what it would have predicted using the information available by that date. Both wins and misses remain visible.
The current published historical record contains 63.4% replay accuracy across 6,955 settled fights in the published record. It is evidence about repeatable methodology, not proof that every one of those calls was published before the event. We do not claim a betting edge or selective-betting ROI.
Every fight shows the de-vigged market probability, with the model as a research signal.
