A Comparative Analysis of Stacked Ensemble and Neural Network Models in Enhancing Customer Retention in Banking

Abstract

Customer churn is a major problem for the banking sector, as it affects the profitability and sustainability of the business. Therefore, it is important to identify the customers who are likely to leave to take appropriate actions towards retaining them. The study aims to apply machine learning models to predict churn and identify the most effective model based on performance metrics such as log loss, ROC AUC, and accuracy using a dataset of 10,000 customers with 14 features. The study also utilizes domain feature engineering to get better predictive signals. we performed data analysis, and tested models including a neural network and stacked ensemble models comprising Random Forest, XGBoost, LightGBM, and Gradient Boosting with logistic regression as the meta-learner. The study performed a comparative analysis of the performance of a neural network with stacked ensemble model of all four tree models. The stacked ensemble model outperformed others, achieving a log loss of 0.2497, ROC AUC of 0.9609, and accuracy of 89.11% on the test set. This indicates that the ensemble model, by combining the strengths of individual learners, effectively handles complex patterns in the data, thereby providing a tool for reducing churn and improving customer engagement strategies in the banking sector.