What question did this study set out to answer?

This research aims to improve customer churn prediction in the banking sector using behavioral analytics.

May 6, 2026Open Access

"Behavioral Analytics–Driven Customer Churn Prediction Using CatBoost"

Key Points

This research aims to improve customer churn prediction in the banking sector using behavioral analytics.
Analyzed a dataset of 10,000 European bank customers from the European Banking Dataset.
Developed a machine learning pipeline with over 40 engineered features, using CatBoost as the champion model.
Incorporated techniques like SMOTETomek for imbalance correction and custom threshold optimization.
CatBoost achieved an ROC-AUC of 0.866 and an overall recall of 77.89%.
Inactive, single-product customers had a churn rate over 38%, while active multi-product holders had rates as low as 8.2%.
The model's accuracy reached 79% on a test set of 2,000 customers.

Abstract

Customer churn in the banking sector is represents a significant and quantifiable threat to revenue stability, institutional trust, and long-term sustainability. While conventional retention strategies predominantly rely on demographic profiling and account-level financial indicators, this study research advances a behavioral analytics paradigm that situates customer engagement depth and product utilization breadth at the center of churn prediction. Using a dataset of 10,000 European bank customers sourced from the European Banking Dataset (Simulated/Public Benchmark Dataset), we developed a comprehensive machine learning pipeline featuring 40+ engineered behavioral and financial features, multi-model evaluation across six classifiers, and a production-grade prediction system. Our champion model, CatBoost with an optimized decision threshold of 0.20, achieved an ROC-AUC of 0.866, an overall recall of 77.89%, a Macro F1 of 0.7295, and an accuracy of 79% on a held-out test set of 2,000 customers. The pipeline incorporates SMOTETomek for within-fold imbalance correction, isotonic probability calibration, and a custom threshold optimization protocol designed to prioritize churn recall under economic constraints in the banking sector . Key findings confirm that inactive, single-product, high-balance customers constitute the highest-risk segment (churn rate > 38%), whereas active multi-product holders exhibit churn rates as low as 8.2%. This study provides a deployable framework for engagement-driven retention strategy, cross-sell optimization, and loyalty program design. Keywords: Customer Churn Prediction, CatBoost, Feature Engineering, SMOTETomek, Behavioral Analytics, Banking Retention, Class Imbalance, Threshold Optimization, ROC-AUC, Engagement Segmentation

AIに質問

Bookmark

View Full Paper