Predictive Analytics for Customer Retention: Designing and Evaluating a Machine Learning–Based Churn Prediction System

Abstract

Customer churn represents a critical challenge for organizations in highly competitive sectors, particularly banking, where retaining existing clients is more effective than acquiring new ones. This study develops a machine learning–based system for churn prediction in the banking sector. The system was designed in Python, integrating modules for data preprocessing, feature engineering, and model training and evaluation. Within this system, five models are trained using five different algorithms (one training algorithm produces one trained model): Logistic Regression, Linear Discriminant Analysis (LDA), Decision Tree, Random Forest, and Gradient Boosting, based on an initial dataset of 10,000 clients, stratified prior to use, obtained from a public Kaggle dataset. For the evaluation of model performance, metrics such as accuracy in churner identification, precision, recall, and F1-score were employed, along with supplementary procedures such as ROC–AUC curve analysis and bootstrapping for model stability assessment. Results demonstrate that ensemble models, particularly those trained using the Random Forest and Gradient Boosting algorithms, outperform baseline approaches across all selected evaluation metrics. Moreover, the 95% confidence intervals, which are narrower than those of the other trained models, along with the ROC–AUC curve analysis, indicate stable predictive performance for these two models. As a result, these algorithms are recommended for identifying potential churners across various business and industrial use cases. Their reliable performance indicates that they can effectively support customer retention campaigns.