BEST HOLDOUT2025–26 chronological holdout
BoutMetrics ML
62.4%accuracy
- TEST SAMPLE
- 591
- BRIER SCORE
- 0.223
Best leak-free public-stats winner pipeline tested on this holdout.
BOUTMETRICS / RESEARCH & VALIDATION
Model lab.
Measured performance. Visible limitations. Explore how our models hold up on future fights, with sample sizes and market comparisons alongside every claim.
BEST WINNER MODEL62.4%
591-fight
temporal holdout
WHAT WORKED
BoutMetrics ML beat the stacking ensemble on accuracy, Brier score, log loss, and AUC.
WHAT DIDN’T
Closing odds still beat our public-data model decisively; value bets and method props lost money.
HOW WE USE IT
For strike, duration, and matchup context—not as a claim that public data beats the market.
WINNER MODELS
Out-of-sample comparison
Accuracy is only one metric. Brier score measures probability quality; lower is better.
AUDITED MODELSame 2025–26 holdout
Stacking ensemble
62.1%accuracy
- TEST SAMPLE
- 591
- BRIER SCORE
- 0.2286
Faithfully retrained on the identical split for an apples-to-apples bar.
AUDITED MODELSeparate 2025 holdout
Logistic baseline
62.2%accuracy
- TEST SAMPLE
- 407
- BRIER SCORE
- 0.2379
Strong simple baseline, but measured on a different test window.
THE REALITY CHECK
The market knew more
BoutMetrics ML was the best public-data pipeline in the lab. On 332 fights with matched closing odds, the consensus market was still much sharper.
MARKET ACCURACY68.4%332 matched fights
MODEL ACCURACY60.5%Same real-odds subset
PRODUCT RULE
No “value bet” badges from public-stat winner probabilities. The evidence does not support them.
REPEATABLE SIGNALS
What mattered most
Recent formLast 3–5 fights
Age differenceYouth showed a consistent edge
Career win rateUseful, but market-aware
ExperienceSecondary context
Reach & heightModest on their own
Relative importance is summarized across the audited linear-model reports; it is not a universal causal ranking.
BOUTMETRICS ML
Current projection layer
Held-out fighter-rows since July 2025. Lower MAE is better.
Significant strikes
26.6Composed total: predicted rate × predicted duration
MAE · 40.3 naive · n=726Fight duration
295sA modest improvement over historical averages
MAE · 309s naive · n=726Over / under lean
66.9%Research result without owned totals odds
accuracy · 65.0% naive · n=726Takedowns
1.14Did not beat the baseline; display as experimental
MAE · 1.09 naive · n=726VALIDATION STANDARD
Past trains.
Future tests.
- 01Chronological splitModels train only on fights that happened before the test period.
- 02Probability qualityAccuracy sits beside Brier score and sample size.
- 03Market realityReal-odds backtests are disclosed—even when they lose.
- 04Uncertainty firstLow agreement becomes a warning, not false confidence.
FROM RESEARCH TO MATCHUP
Use the model where it belongs.
This page shows how the models perform. Matchup stories, model context, and projections are included with a free account.
Create free account →