A V12 dual-ensemble ML model trained on 8,500+ fights. 70.4% accuracy over a 7-year walk-forward backtest. No guesswork, no gut feelings — just data.
7-year walk-forward backtest (2020–2026). Not cherry-picked. Not overfit. Independently audited for data leakage.
Past performance does not guarantee future results.
From raw fight data to actionable predictions in four steps.
12+ years of UFC fight data, fighter stats, and historical odds scraped and cleaned.
47 engineered features with 5-year half-life time-weighting. Recent fights matter more.
4-model ensemble (LR + XGBoost + LightGBM + CatBoost) with odds-aware and odds-blind dual architecture.
Live scanning of Polymarket and sportsbook lines to find mispriced fights. Edge, not luck.
Two sub-models blended to avoid overfitting to betting lines while still leveraging market signal.
Stable baseline with interpretable coefficients
Gradient boosting for complex feature interactions
Fast leaf-wise training with categorical support
Ordered boosting to reduce prediction shift
Real model output for UFC 327. This is what you get before every fight night.
Every prediction comes with a confidence tier so you know exactly how much weight to give it.
We scan Polymarket and sportsbook lines in real time and compare them against our model probabilities. When the market is wrong, we flag it.
If our model says a fighter wins 76% of the time but the market prices them at 62%, that's a +14% edge. We surface these opportunities so you can act on them.
Join the waitlist and be the first to receive model output, edge alerts, and confidence-rated picks for every UFC event.