Customer Churn Prediction

Problem Statement

Customer churn costs telecom companies billions annually. Identifying at-risk customers before they leave enables targeted retention campaigns. The challenge: churn is rare (~15-20%), requiring models optimized for recall over precision to catch as many churners as possible.

Technical Approach

Model Development

• XGBoost: Gradient-boosted trees with class weighting to handle imbalance — achieved 85% recall, correctly flagging 4 out of 5 churning customers
• Random Forest: Ensemble baseline for feature importance analysis and model interpretability
• Logistic Regression: Simple, interpretable baseline for threshold calibration

Feature Engineering

Engineered 40+ behavioral features from raw telecom subscription data: usage pattern velocity (week-over-week changes in call minutes and data usage), customer service interaction frequency, payment delay indicators, and contract lifecycle stage. The top three predictors: contract type (month-to-month vs annual), tenure length, and monthly charge amount.

Threshold Optimization

Tuned classification threshold to maximize recall while keeping precision above 50%. This ensured marketing teams could act on predictions without an overwhelming false-positive rate. Delivered actionable customer segments ranked by churn probability for phased retention outreach.

Key Results

• 85% recall — identified 4 out of 5 churning customers before they left
• Contract type, tenure, and monthly charges emerged as top churn predictors
• Delivered retention segments enabling marketing team to prioritize high-risk accounts

Tech Stack