Calibrated MMA Forecast

Most public UFC models report 75%+ accuracy that vanishes out of sample, because they leak career totals into historical rows. This one is built to survive contact with fights it has never seen.

Trained on 6,453 UFC bouts. Measured on 664 fights from 2025–2026 held out by date, never used in training or calibration: 65.7% accuracy, 0.630 log loss, 0.219 Brier, 0.704 AUC. An Elo-only baseline scores 57.8% on the same fights.

What makes it different:

• Point-in-time features. Every feature for a bout is computed only from fights that finished before that bout's date. Two fights on the same card cannot leak into each other.

• Corner bias removed. ufcstats lists the eventual winner first 64% of the time. Corners are randomised, dropping the base rate to 50.2%, so the model gets no free signal from bout ordering.

• Exactly symmetric. Swap the two fighters and the probabilities are the exact complement — enforced at inference and through the calibration map.

• Calibration is the product. Reliability curves, confidence-band hit rates and baseline comparisons ship with every answer, measured on a time-held-out period rather than a random split.

• Honest about its ceiling. The closing betting market lands around 68–72% on moneyline favourites. This model is below that, and says so rather than implying an edge.

Ask it for a single matchup and you get the probability, the historical hit rate of that confidence band, the features that moved the number in plain language, and a point-in-time tale of the tape. Ask it for a full card and you get a self-contained interactive dashboard.

Research and calibration tool. It does not give betting advice, picks, stake sizes, bankroll guidance or parlays, and will decline if asked.

Tags
#Sports #Analysis #Research #Statistics