Machine Learning-Based Volatility Forecasting and Systemic Risk Dynamics in Indonesian State-Owned Banks

Abstract

This study evaluates volatility forecasts and systemic-risk indicators for four Indonesian state-owned banks (BBRI, BBTN, BMRI, and BBNI) from January 2010 to December 2025. Random Forest (RF) and Gradient Boosting (GB) models use information available at each forecast origin and are tuned by expanding-window validation. We compare them on a common 774-day test sample with a Random Walk, historical mean, moving average, HAR, GARCH(1,1), EGARCH, and GJR-GARCH. The GARCH-family models produce the most accurate 20-day rolling-volatility forecasts, with out-of-sample R^2 values of 0.9793–0.9838. RF attains R^2 values of 0.8929–0.9459 and does not outperform persistence. HAC-corrected Diebold–Mariano tests favor the Random Walk over RF and GB, but favor the GARCH-family models over the Random Walk. The first principal component of the Systemic Risk Index (SRI) explains 76.7% of component variation and correlates 0.999 with the equal-weight index; ∆CoVaR estimates indicate downside dependence. After 10-basis-point switching costs, an ex-ante RF volatility-timing rule improves Sharpe ratios for BBRI, BMRI, and BBNI, but not BBTN. RF and GB revealbank-specific nonlinear associations, whereas the econometric models provide the most accurate point forecasts.