Prediction of Customer Time Deposit Subscriptions Based on Bootstrap Uncertainty Quantification

Abstract

Traditional retail bank telemarketing for time deposit subscriptions has a low conversion rate, suffers from data leakage, and fails to account for individual uncertainty in point predictions. To address these issues, this paper constructs a risk-aware analytical system that integrates high-performance ensemble learning with non-parametric uncertainty quantification. First, a strict leakage-prevention protocol is introduced by excluding post-dialing variables, such as call duration, and chronological data splits are used to avoid look-ahead bias. Second, to address macroeconomic non-stationarity over time and space, the absolute account balance is converted into a quintile-based relative wealth ranking. Third, four machine learning models are tested, and XGBoost performs best, with an AUC-ROC of 0.9100, an F1-score of 0.8504, and a total accuracy of 85.49%. Additionally, a parallelized Bootstrap engine is employed to obtain individual predictive interval widths and, from the raw output probabilities, to construct an operational dual-tier decision system. The framework sets up micro-level risk alarms for uncertain predictions and macro-level manual overrides. Based on the above experiments, the system will reduce operating costs in the low-uncertainty marketing area by approximately 80%. Finally, conformal prediction is proposed as a future low-latency path for real-time deployment.