METHODOLOGY

THE MODEL

For each matchup the engine runs 100,000 round-by-round Monte Carlo simulations using each fighter's career stats, recent form, official rank, and strength of schedule. Aggregating across the 100k sims gives the win probability, predicted method, predicted round, and prop predictions you see on every fight page.

CALIBRATION

Current production engine: v2-iter9. Measured against 500 historical fights:

• Brier 0.1855 — better than typical sharp closing lines (~0.21)
• 78.2% winner accuracy
• 46.8% of picks at 60%+ conviction
• Method distribution within 1.6pp of actual KO/Sub/Dec rates

Caveat: the audit uses current fighter stats against past fights, which biases the Brier slightly optimistic. The unbiased measure is /track-record — every pick logged before the fight, graded after.

EDGE vs MARKET

On every priced fight we compare the model's win probability to the de-vigged sportsbook line. Gaps of 5+ percentage points get surfaced on /picks as “Value Side” — the side the model rates higher than the market is pricing.

DATA & LIMITS

Fighter stats from UFCStats.com, odds from The Odds API across 15+ US sportsbooks. Refreshed automatically after every event. The model doesn't see injuries, weight cut quality, gym changes, or motivation — predictions for fighters with fewer than three UFC fights are inherently noisier. For research and entertainment; we don't accept bets.