<![CDATA[This study evaluates a geopolitical-regime-aware hybrid ensemble (GRAHE) for one-day-ahead forecasting of the Indonesia Composite Index (IHSG) using observations from 2015 to 2025. The framework combines long short-term memory, extreme gradient boosting, and random forest forecasts through a softmax gate conditioned on one-month-lagged geopolitical risk and rolling market volatility. Hyperparameters were selected with Bayesian optimization on the 2022 validation period, and the locked test used monthly expanding-window retraining from 2 January 2023 to 5 December 2025 (697 forecasts). GRAHE obtained a return RMSE of 0.009334, directional accuracy of 50.50%, and Theil’s U of 1.001. It did not outperform the strongest benchmark: Ridge regression achieved an RMSE of 0.009316, while XGBoost produced the highest directional accuracy of 54.23%. The RMSE difference between Ridge and GRAHE was not statistically significant according to the Diebold–Mariano test (p = 0.744). In high-GPR observations, GRAHE recorded an RMSE of 0.009626, 0.30% above the best baseline. Removing GPR features increased ensemble RMSE slightly to 0.009350, but market-only XGBoost remained competitive at 0.009326. These findings do not support a strong claim that geopolitical conditioning improves daily IHSG point forecasts, although the framework provides a transparent basis for regime-dependent risk analysis and further robustness testing.]]>
Copyrights © 2026
