Model Performance
Training metrics for every model in the registry · last trained 2026-08-11 20:30
5
Models
0
Training fights
0
Upcoming fights
Model Comparison (temporal holdout)
| Model | Algorithm | Features | Holdout AUC (95% CI) | Holdout Acc | Log Loss | Brier | vs. Champion |
|---|---|---|---|---|---|---|---|
|
XGBoost
PRIMARY
Gradient-boosted trees on the full engineered feature set
|
xgboost | 108 (full) |
0.6988
[0.671, 0.727]
|
64.3% | 0.6322 | 0.2206 | Statistically tied |
|
LightGBM
LightGBM on the full engineered feature set
|
lightgbm | 108 (full) |
0.6965
[0.668, 0.726]
|
64.0% | 0.6357 | 0.2221 | Statistically tied |
|
Logistic (Core)
Regularized logistic regression on curated differential features
|
logistic | 18 (core) |
0.6929
[0.664, 0.721]
|
64.1% | 0.6379 | 0.2234 | Statistically tied |
|
CatBoost
CatBoost (ordered boosting) on the full engineered feature set
|
catboost | 108 (full) |
0.6975
[0.669, 0.726]
|
65.4% | 0.6383 | 0.2233 | Statistically tied |
|
Ensemble
Soft-vote of XGBoost, LightGBM and logistic regression
|
ensemble | 108 (full) |
0.7015
[0.673, 0.730]
|
66.1% | 0.6299 | 0.2196 | CHAMPION |
Holdout metrics are computed on the most recent ~20% of fights, which the models never saw
during training or tuning — an honest estimate of forward, real-world performance.
AUC and accuracy: higher is better. Log loss and Brier: lower is better (they punish
confidently wrong probabilities). The AUC range in brackets is a 95% bootstrap confidence
interval — how much that number could plausibly move on a different set of holdout fights.
"vs. Champion" is a paired significance test against the model with the best point-estimate
AUC: "Statistically tied" means the difference could easily be noise, not a real gap.
Outcome Model — Method & Round
| Sub-model | Classes | Holdout Accuracy | Holdout Log Loss | Holdout Size |
|---|---|---|---|---|
| Method (Decision / KO-TKO / Submission) | 3 | 54.1% | 0.9601 | 1267 |
| Round bucket (finishes only) | 3 | 50.0% | 1.0054 | 638 |
A separate pair of models answering "how" a fight ends rather than "who" wins — shown on every
fight's detail page. Method accuracy is measured against 3 classes (chance ≈ 33-48% depending on
class balance); round accuracy is measured only on fights that were predicted correctly to be finishes.
Feature Importance — CatBoost
Age edge (significant gap)
12.128
Net striking dominance
11.843
Elo rating edge
8.063
Experience edge (total fights)
6.910
Fighter 2 · Strike rate delta r1 to r3
6.088
Fighter 1 · Strike rate delta r1 to r3
4.902
Fighter 2 · Age
3.789
Defensive skill edge
3.750
Fighter 1 · Age
3.433
Fighter 2 · Avg takedowns attempted
2.732
Fighter 1 · Avg takedowns attempted
2.699
Fighter 1 · Strikes absorbed/min
1.468
Reach edge
1.459
Fighter 2 · Avg control time
1.458
Head-strike accuracy edge
1.457
Fighter 2 · Distance strike defense
1.276
Fighter 2 · Round 1 strike output
1.198
Fighter 1 · Distance strike defense
1.011
Fighter 2 · Strikes absorbed/min
0.977
Fighter 1 · Fights per year
0.907
Global importance: how much the model relies on each feature across all fights.
For a single fight's explanation, open that fight's detail page.