Telco Churn & Retention Intelligence

Problem Statement

Telecommunication providers suffer high customer turnover. Blanket 15% promotional discounts across all accounts cause massive margin erosion (>\$205K/quarter), while selecting customers on raw probability alone wastes call center capacity on low-tenure, low-value trialists.

Dataset

7,043 subscriber accounts from IBM Cognos Telco dataset across 21 demographic, service, and contract attributes with strict zero-leakage stratified splits and data hygiene.

Architecture

End-to-end ML pipeline with ColumnTransformer preprocessing, XGBoost classification, MLflow experiment tracking, SHAP TreeExplainer local attribution, an asynchronous FastAPI REST microservice, and an executive Streamlit decision dashboard.

Model Selection

Benchmarked Logistic Regression, Random Forest, LightGBM, and XGBoost. XGBoost achieved top holdout discrimination (ROC-AUC 0.84, PR-AUC 0.63). Calibrated threshold optimization at tau*=0.23 balances false negatives against discount costs.

Training Process

Stratified 80/20 train/test split with zero-leakage feature pipelines. Cost-sensitive threshold search tuned to capture 93.85% of churning accounts while targeting only high-CLV subscribers.

Evaluation Metrics

Results

A live Streamlit intelligence application and FastAPI service delivering real-time churn scoring, directional SHAP explanations, customer cohort pattern charts, and prioritized outreach queues.

Key Learnings

• 1Ranking outreach by Expected Loss (Probability * CLV) delivers far superior capital efficiency than raw churn probability ranking.
• 2Calibrated decision thresholds (tau*=0.23) are essential in asymmetric cost business problems where missing a churner costs far more than a retention offer.
• 3SHAP TreeExplainer transparency provides frontline account reps with clear, actionable talking points during customer retention calls.