Model Catalog

Production Model Registry — All models tracked via MLflow (v3.0.0, Python 3.11.14, sklearn 1.8.0)

Overview

Project	Model ID	Algorithm	Size	Primary Metric	Status
BankChurn	bankchurn-stacking-v3.0.0	StackingClassifier (RF+GB+XGB+LGB→LR)	4.1 MB	AUC 0.87	Production
NLPInsight	nlpinsight-tfidf-v3.5.0	TF-IDF + LogReg (prod) / FinBERT (GPU)	~5 MB / ~440 MB	Acc 80.6%	Production
ChicagoTaxi	chicagotaxi-rf-v3.5.0	RandomForest (lag features)	~2 MB	R² 0.96	Production

BankChurn Predictor

Metric	Value
AUC-ROC	0.87
F1 (Churn)	0.62
Coverage	90% (199 tests)

Pipeline: SimpleImputer → StandardScaler/OneHotEncoder → StackingClassifier(RF+GB+XGB+LGB→LR) Features: SHAP explainability, drift detection (KS+PSI+Evidently), fairness audits (disparate impact, equal opportunity)

NLPInsight Analyzer

Metric	Value
Accuracy	80.6%
F1 (macro)	0.748
Coverage	98% (74 tests)

Pipeline: TF-IDF + LogisticRegression (production) or FinBERT Tokenizer → FinBERT (ProsusAI) (GPU) Features: Dual backend auto-detection, Twitter Financial News trained, fairness audits (F1 parity), batch support (500 texts)

ChicagoTaxi Demand Pipeline

Metric	Value
R²	0.9649
RMSE	7.87 trips
Coverage	91% (122 tests)

Pipeline: PySpark ETL → Lag Features → RandomForest (temporal split) Features: 6.3M trips processed, leak-free lag features, Dask batch prediction (19K rows/sec)

Promotion Criteria

1. Primary metric exceeds baseline by ≥5%
2. Cross-validation std dev <3%
3. P95 inference <200ms
4. Test coverage >70%
5. Clean security scans
6. Model card updated

Retraining Triggers (ADR-006)

Automated via K8s CronJob → GitHub Actions workflow_dispatch.

Trigger	BankChurn	ChicagoTaxi	NLPInsight
Performance	AUC <0.75	R² <0.90	F1-macro <0.60
Drift (PSI)	PSI >0.20	PSI >0.20	Distribution shift >25%
Scheduled	Quarterly	Quarterly	Quarterly

Links

• BankChurn Model Card
• NLPInsight Model Card
• ChicagoTaxi Model Card

---

Last Updated: March 2026 — v3.5.3