Forecasting and Explaining Non-Performing Assets in Indian Banking Groups Using Panel Econometrics and Explainable Machine Learning Models

This study examines the determinants and predictive dynamics of Non-Performing Assets (NPAs) in Indian banking groups using an integrated framework combining panel econometric techniques, machine learning models, and explainable artificial intelligence. The analysis is based on a balanced panel dataset comprising 108 observations covering the period 1998-2024, incorporating key macroeconomic variables such as credit growth, GDP growth, inflation, and lending interest rates. Panel estimations using Pooled OLS, Fixed Effects, and Random Effects models are conducted, with the Hausman test supporting the Random Effects specification. The results indicate that lending interest rates have a positive and statistically significant impact on NPAs, whereas inflation exhibits a significant negative effect. Diagnostic tests reveal the presence of heteroskedasticity, serial correlation, and cross-sectional dependence, highlighting the persistent and systemic nature of NPAs. Robustness checks using an IV-based dynamic panel model and time-aware machine learning validation confirm the persistence of asset quality trends. To enhance predictive performance, the study employs Linear Regression, Random Forest, and XGBoost models, with XGBoost demonstrating the highest accuracy (R2 = 0.84). SHAP analysis identifies lagged NPAs and interest rates as key drivers. Policy-wise, the findings suggest that regulators should prioritize the monitoring of interest rate cycles and historical credit persistence over aggregate GDP trends to mitigate systemic asset quality risks. The study underscores the importance of integrating econometric and data-driven approaches for effective risk assessment and supervisory policy formulation in the banking sector.