Time-Series Machine Learning for Assessing Depositor-Risk Exposure

Abstract

This study proposes a time-series machine learning model to assess depositor risk by creating a composite Depositor Risk Index (DRI) that combines credit risk (NPLR, LLPR), liquidity risk (LIQR), and profitability (ROA). Lagged features are used to incorporate temporal dependencies, and XGBoost is applied in a temporally consistent train-test design that guarantees predictive performance without information leakage, which aligns with the real-world forecasting environment. SHAP-based explainability is provided to enhance transparency and enable both general and case-specific understanding of model behavior. Findings indicate that the outcomes are highly predictive, with close correspondence between the anticipated and actual DRI values and a constant residual trend. The most dominant factor is liquidity, which exhibits a strong inverse correlation with depositor risk, while credit risk indicators contribute positively but to a smaller extent. Lagged variables are key to improving model performance and hence establish the existence of temporal persistence of banking risk. The analysis also indicates significant heterogeneity among banks, with different volatility trends and structural changes over time. Non-linear relationships and threshold effects are well represented, thereby enhancing the model's explanatory power. In sum, the framework is an effective early-warning tool that can be used to proactively monitor depositor vulnerability and promote informed, data-driven regulatory decision-making in the changing banking sector in Vietnam. To further enhance the prediction of depositor risk and the relevance of policies, future research can integrate deep learning models, real-time streaming data, and macro-financial variables.