Churn Analysis on Banking Dataset
Predicting customer churn in the banking sector using boosting algorithms, Optuna optimization, and interpretable ML techniques.
This project develops a customer churn prediction model for the banking sector, aiming to identify the top 10,000 clients most likely to close their accounts.
Methods
- Preprocessing: data cleaning, categorical encoding
- Sampling: cost-sensitive techniques to address class imbalance
- Boosting algorithms: LightGBM, XGBoost, CatBoost, with ensemble weighting
- Hyperparameter optimization: automated tuning with Optuna
- Interpretability: SHAP values for global and local feature importance
Key results
- Ensemble boosting models achieved the best performance using a custom Rank Probabilities metric, prioritizing recall of churners
- Synthetic data improved testing robustness on unseen distributions
- SHAP analysis revealed key socio-demographic and account features driving churn risk
Boosting models with interpretability for churn prediction