Uncertainty-Aware Deep Ensembles for Confident Customer Churn Prediction with Rejection Option

Customer churn prediction is an important problem in business intelligence, especially in industries where keeping customers is both difficult and expensive. Many existing models can predict churn accurately but do not show how confident they are in their results, which can cause costly or incorrect decisions. To solve this issue, this paper introduces an Uncertainty-Aware Ensemble (UA-Ensemble) framework that predicts churn while also estimating confidence levels. The model combines five neural network types, including attention-based LSTMs, Bayesian networks, and Monte Carlo Dropout, to measure both aleatoric and epistemic uncertainty. A cost-aware rejection rule is used to avoid unreliable predictions. Experiments on large banking, e-commerce, and telecom datasets reached 94.2% accuracy and reduced intervention costs by 18.3% compared to baseline models. The proposed approach outperforms traditional machine learning and deep learning methods, proving effective and trustworthy for business decision-making across various industries.