HOW THE OUTPUTS ARE GOVERNED.
This page explains what enters the public prediction process, how outputs are judged and which details remain proprietary—without presenting the model as a guaranteed betting system.
PRE-FIGHT INFORMATION ONLY
Model features are fixed at the pre-fight boundary. Information created by the result itself is excluded from that fight’s prediction. Bookmaker odds are not used as model features; the model forms its own probabilities before any comparison with market prices.
THREE SEPARATE QUESTIONS
- Moneyline: the probability of each fighter winning.
- Goes the Distance: whether the scheduled final bell is reached.
- Method of Victory: knockout, submission and decision paths for each fighter.
These outputs must remain internally coherent, but one market is not relabelled as another. The market-derived fair prices on the upcoming-event page are also separate from model predictions.
HISTORICAL EVALUATION
Development models are evaluated on chronologically later fights that were not used to fit them. Assessment includes probability calibration and scoring measures such as Brier score, expected calibration error and log loss. Any public performance result must identify its model version, evaluation window, qualifying sample and pricing assumptions before it is treated as a published benchmark.
LIMITATIONS
MMA outcomes are noisy, datasets can be incomplete, new fighters have less evidence, and market availability varies. Probabilities can be wrong even when a model is well calibrated. Implementation details that would expose proprietary feature engineering or model construction are not published.
Review the data and correction standardsand the safer-gambling guidancebefore using any output.