The Impact of Macroeconomic Shocks on Credit Risk in Indonesian Banking Using Machine Learning

This study analyzes how multivariate macroeconomic shocks transmit to credit risk in Indonesia’s banking sector using a hybrid framework that combines a factor augmented vector autoregression with machine learning techniques. In contrast to conventional stress testing approaches that rely on linear assumptions, the model extracts latent macro financial factors through Principal Component Analysis and forecasts non-performing loans using Elastic Net regression in order to address multicollinearity and enhance predictive stability across different phases of the business cycle. The empirical analysis employs monthly data covering the period from January 2010 to December 2024. Principal Component Analysis reduces the dimensionality of the macroeconomic dataset to six components that collectively explain 90.99 percent of total variance. The Elastic Net model is optimally tuned with an alpha value of 0.0668 and an L1 ratio of 0.1, placing greater emphasis on the ridge penalty to improve coefficient stability and out of sample reliability. The results indicate strong predictive performance, with an R squared of 61.91 percent and a root mean squared error of 0.0692. Liquidity related indicators, particularly money supply and loan growth, consistently emerge as the most influential determinants of non-performing loans and exhibit a stable negative relationship throughout the sample period. A key empirical contribution of this study is the identification of parameter instability, as the BI Rate, exchange rate, and Industrial Production Index display reversals in their estimated effects when comparing the pre pandemic period with the full sample. Most notably, the BI Rate shifts from a positive to a negative association with non-performing loans, providing empirical support for the Lucas Critique. This structural change is plausibly linked to regulatory forbearance under POJK 11/2020, which weakened the conventional transmission mechanism from higher interest rates to rising credit risk. Overall, the findings suggest that under conditions of extreme macroeconomic stress, policy interventions can fundamentally alter structural relationships and reduce the reliability of traditional linear forecasting models. The proposed framework therefore offers a robust alternative for regulatory stress testing by explicitly accommodating parameter instability and policy induced shifts in banking sector credit risk.