Overview
Built a predictive model to identify customers at risk of churning, enabling proactive retention strategies.
Key Features
- Feature engineering from customer transaction data
- Ensemble ML models (Random Forest, XGBoost, LightGBM)
- SHAP values for model interpretability
- Real-time scoring API
Technical Stack
- Data Processing: Pandas, NumPy, Scikit-learn
- Modeling: XGBoost, LightGBM, Random Forest
- Visualization: Matplotlib, Seaborn, Plotly
- API: FastAPI for model serving
Results
- 92% accuracy in predicting churn
- 85% precision for high-risk customers
- Reduced churn rate by 18% after implementation
- Identified top 5 churn indicators
Implementation
The model processes customer behavioral data including:
- Transaction frequency and recency
- Support ticket history
- Product usage patterns
- Demographic information
Deployed as a REST API that provides real-time churn probability scores for any customer.