2,000+
Fighters Tracked
600+
Events Analyzed
72%
Prediction Accuracy
Upcoming Events
The Official Calendar
Fight Archives
Rewatch history
The Official Calendar
Rewatch history
Every prediction comes from two independent systems — a statistical model and an AI fight analyst — that are reconciled into one read, and then graded in public after the fight. Here's exactly how that works, and just as importantly, what it is not.
Statistical model
81 features · XGBoost · calibrated
AI fight analyst
second opinion · never sees the model
Reconciled
where they agree — and, more usefully, where they don't
Graded in public
every pick, win or miss, no edits
The model learns from UFC fights going back to 2015. For each matchup it builds 81 features per bout — striking output and defense, grappling rates, durability, experience, layoff, age, reach, recent form and momentum — all computed strictly from information available before the fight, so no hindsight leaks into training. Recent fights are weighted more heavily than old ones, so it tracks how the sport evolves rather than how it looked a decade ago. Separate models call the winner, the method (KO/TKO, submission, decision), and the round.
Tested on fights it never saw. We evaluate forecasting the only honest way: train on the earlier fights, tune on a later window, then grade on the most recent 393 bouts the model was held out of entirely (Oct 2025 – Jun 2026). That held-out number — not a flattering in-sample statistic — is what we publish, and the live scoreboard then holds us to it.
Winner accuracy
65.8%
393 unseen bouts
Calibration error
2.3%
ship gate: under 8%
Features / bout
81
all pre-fight
Calibration over bravado. A model that says "70%" should be right about 70% of the time — that property is called calibration, and we treat it as a shipping requirement. A retrained winner model is promoted only if it clears a minimum-accuracy bar and a calibration gate (average error under 8%; the current model measures 2.3%). The win probabilities we publish for real fights are also clamped to a conservative band, so no pick is dressed up as a near-certainty — in this sport none is.
Where betting lines exist, we strip the bookmaker's margin and feed the market's implied probability to the model as one input among many. It's constrained to treat that signal sensibly — a fighter the market likes more can never be scored less likely to win because of it — but the model still forms its own view from the fight data, which is what lets it disagree with the market rather than echo it.
The market is also our toughest benchmark, and we hold our own with it rather than beat it. On the 170 held-out bouts that had lines, picking the market favorite is right 69.7% of the time and the model 71.5% — an edge within noise on a sample that small, though the model does price those bouts with lower error. When the two diverge we flag a "model edge," but be clear-eyed about what that is: on the bouts where the model disagreed with the market, it was right only about half the time. So a model edge is an interesting difference of opinion between two estimates — not a betting recommendation, and not a signal we'd bet on ourselves.
Alongside the statistical model, an AI analyst (Anthropic's Claude, grounded in our fighter database) writes an independent read of each matchup — styles, paths to victory, and risk factors. The analyst is deliberately not shown the model's output, so it's a genuine second opinion rather than a paraphrase.
A final reconciliation then explains where the two perspectives line up and where they split — often the most useful part of the whole prediction is the disagreement, and why it happens.
Every prediction is stamped with the model version that made it and frozen at prediction time. After each event an automated resolver grades the pick against the official result — winner, method, and round — with no retroactive edits. The track record page counts each bout once, breaks the record down per model version, and publishes the calibration table so you can check our stated confidence against what actually happened.
The models are retrained as new fights resolve, and material changes ship as new versions with their backtest published here.
Found and fixed a training-data bug that corrupted the betting-market feature on half the training examples, added recency weighting, and constrained the model to treat market and ranking signals sensibly. Held-out winner accuracy improved from 63.5% to 65.8%, with the clearest gains on fights the market doesn't price (about 58% to 61.4%) and in calibration; on bouts with lines it stays about even with the market favorite (71.5% vs 69.7%).
First production release: winner model at 63.5% held-out accuracy, method at 48.2% (a three-way call), round at 52.0%. The method and round models still run this version.
Fighting is chaotic, and honest probability estimates reflect that. A well-calibrated model picks winners in the mid-60s percent of the time — meaningfully better than a coin flip, about level with the betting market's favorite where the two can be compared, and still far from certainty. Anyone advertising near-perfect fight prediction accuracy is selling something else.
Our predictions are for analysis and entertainment. They are not betting advice, and no prediction — ours or anyone's — guarantees an outcome. We publish our full track record, misses included, because a record you can audit beats a claim you can't.
Fight statistics are ingested from publicly available sources, reconciled nightly, and updated live on event nights. The same database that powers the stats pages powers the models — when a fighter's record updates, their next prediction reflects it.
See it in action. The public track record grades every pick, and Champion members get the full predictions, the fighter-vs-fighter simulator, and the AI matchup chat.
For analysis & entertainment only. Predictions and model-edge figures are informational — they are not betting or financial advice, and past accuracy does not guarantee future results. We are not responsible for any losses from betting or gambling activity. You must be 18+ (21+ in some jurisdictions) to gamble. If gambling is a problem for you or someone you know, call 1-800-GAMBLER or visit ncpgambling.org. Full disclaimer