SIA Retention Engine — Churn Risk Model (v1.0.0)

XGBoost classifier trained on the SIA Retention Engine's Phase 1 dataset (real IBM/Kaggle Telco Customer Churn data, mapped to SIA's Subscriber schema — see the dataset repo for exactly what's real vs. synthetic).

Metrics

See metrics.json in this repo for the full train/val/test breakdown, including a direct ROC-AUC comparison against the original hand-written heuristic formula it replaces (backend/data_generator.py), measured on the same test split.

Top features (mean |SHAP|, test split)

plan_Flexi: 0.8123
plan_Premium: 0.5059
avg_monthly_spend: 0.3718
data_usage_gb: 0.3561
signal_strength_score: 0.1611

Intended use

Churn risk scoring for SIA's Monitor agent, behind the USE_ML_MODEL feature flag (config/settings.py). Falls back to the original formula when this model isn't loaded — see backend/models/loader.py.

Files

model.joblib — the trained XGBoost classifier (load with joblib.load)
feature_columns.json — exact column order/names the model expects

Downloads last month: -; Downloads are not tracked for this model. How to track