Two records. Clearly separated.

For every historical event, CageMetrix rebuilds the fighter ratings from earlier events only. Opponent baselines, division populations, Elo, CMR components and sample strength cannot see the fight being predicted or anything after it.

Predictor 0.1 uses 24 matchup features built from CMR components and Elo. Hyperparameter selection used 2022 only; the final coefficients were fitted on 2018–2022 and then frozen. Results from 2023 onward form the untouched evaluation period.

The historical test covered 1,516 matched fights through August 29, 2026. Predictor 0.1 reached 63.2% accuracy with a 0.2282 Brier score; calibrated Elo reached 54.7% with a 0.2459 Brier score. This is retrospective validation using today's corrected source data, not a claim that those predictions were published before those historical fights.

Historical comparisons require both fighters to have a prior rated UFC bout. In production, an unrated UFC debutant uses a neutral-start Elo fallback and carries a limited-history label. Betting odds, injury reports, camp changes and other news inputs are not part of Predictor 0.1.

Explanations on the prediction page come from the frozen model's actual feature contributions, grouped into readable skill areas. The live record starts fresh with Predictor 0.1; the previous Elo forecasts remain stored separately as an auditable baseline.

Download the Predictor 0.1 benchmark · Download the earlier CMR/Elo benchmark · Read the methodology and limitations