This project addresses the critical business challenge of customer attrition in the E-commerce sector. Using a dataset of ~5,600 customers, I developed a machine learning pipeline that identifies potential churners with 94% accuracy and 77% recall.
- Tenure is King: New customers (under 5 months) are significantly more likely to churn, indicating a need for better "Onboarding" engagement.
- Price Sensitivity: High cashback amounts are strongly correlated with retention, suggesting customers are highly reward-motivated.
- Pain Points: Customer complaints are a "smoking gun" for churn, requiring faster resolution times to save the account.
- Cleaning: KNN Imputer for missing data, Winsorization for outlier control.
- Balancing: SMOTE (Synthetic Minority Over-sampling) to handle class imbalance.
- Modeling: XGBoost Classifier with Stratified Cross-Validation.
- Explainability: Feature Importance (Weights) to derive business strategy.
| Metric | Score |
|---|---|
| Accuracy | 93.8% |
| Recall (Churn) | 77.0% |
| Precision (Churn) | 85.0% |
The business should implement a "First-90-Day" Loyalty Program. By targeting low-tenure customers with personalized cashback incentives, the model suggests we can proactively retain up to 77% of at-risk users.