Model Registry
Active champion models. Challengers are trained weekly and promoted if they pass the logloss + ECE gate.
Stack Architecture
Base Learners
XGBoost + LightGBM + CatBoost trained independently on walk-forward 5-fold CV. Each model gets Optuna hyperparameter search.
Meta-Learner
Logistic regression trained on out-of-fold base model probabilities. Learns optimal weighting from data — no hardcoded 60/40 split.
Calibration
Isotonic regression maps raw ensemble output to calibrated probabilities. Fitted on held-out calibration set (last 20% of training window).
NBAWINNER
NBA Moneyline
v2.4.1 · trained
Champion
0.6412
Log-Loss
0.2189
Brier
0.0312
ECE
57.3%
Accuracy
3,842
Samples
Ensemblexgb+lgb+cat+logistic meta+isotonic cal
Feature speca3f9c21e8b0d
MLBWINNER
MLB Moneyline
v2.3.8 · trained
Champion
0.6721
Log-Loss
0.2344
Brier
0.0418
ECE
54.1%
Accuracy
5,123
Samples
Ensemblexgb+lgb+cat+logistic meta+isotonic cal
Feature spec7d1e44af902b
NBAPROPS
NBA Player Props
v1.9.2 · trained
Champion
0.6891
Log-Loss
0.2401
Brier
0.0512
ECE
52.9%
Accuracy
28,441
Samples
Ensemblelgb+logistic meta+isotonic cal
Feature spec5b62f0d1c4a7
MLBPROPS
MLB Player Props
v1.7.4 · trained
Champion
0.6943
Log-Loss
0.2471
Brier
0.0589
ECE
52.1%
Accuracy
41,230
Samples
Ensemblelgb+logistic meta+isotonic cal
Feature spece91a3c7fb610
In Development
NFLFeature engineering in progress
ETA: August 2026 (preseason)
QB pressure rate, air yards, EPA/play, DVOA proxies. Will train on 10+ years of play-by-play.
NHLData pipeline planned
ETA: October 2026 (season start)
Corsi/Fenwick, expected goals, goalie save pct, PP efficiency. Full RTSS box score ingestion.
Models retrain weekly. Champion/challenger promotion gate: ≥0.5% logloss improvement with ECE and Brier sanity checks. All inference runs on the same GPU used for training.
Read the methodology →