Evaluation
Model results
Historical performance, data coverage and the limits of the estimates used in the research workspace.
Historical replay
These fight outcomes were already examined during development. Replaying them with corrected fighter records is useful for checking the model, but it is not independent validation.
- Winner accuracy
- — Across the replay sample
- Brier score
- — Probability error; lower is better
- Fights evaluated
- — Decisive, identified bouts
- Events
- — Represented in the sample
Did the probabilities match the outcomes?
Each row groups fights by the model’s probability for its chosen fighter. Compare the average estimate with how often that fighter actually won.
| Confidence | Fights | Mean estimate | Won |
|---|
How the data is divided
- Training
- Loading…
- Calibration
- Loading…
- Model inputs
- — features
Training fits the model. Calibration adjusts its probabilities on a separate period. The historical replay uses later fights whose outcomes have already been inspected.
Research coverage
- Records through
- Loading…
- Fighters in lookup
- —
Loading the current research coverage…
Coverage is different from evaluation
The research dataset can include newer records than a saved test. Each matchup report shows the fight history and sample available for those fighters.
What remains unproven
Historical accuracy does not establish an advantage over market prices. There is no demonstrated betting edge, and current injuries, camps and late changes are not fully captured by fight records.
- Future performance needs predictions recorded before the outcomes are known.
- Historical source snapshots may contain information revised after a fight.
- Small groups of high-confidence predictions cannot establish reliable high-confidence performance.
Earlier project models Archived diagnostics
These are stored scores from the original project classifiers. They do not evaluate the corrected model used above. Preprocessing was fitted before cross-validation, so these scores are not independent holdout estimates.
| Model | Random-split CV | Chronological CV |
|---|---|---|
| Loading archived scores… | ||