A1358
Title: Robust volatility forecasting with score-driven MIDAS models: Evidence from commodity futures
Authors: Azra May Kabiri - Macquarie University (Australia) [presenting]
Yanlin Shi - Macquarie University (Australia)
Zhuo Jin - Macquarie University (Australia)
Abstract: Mixed-frequency volatility models combine high-frequency financial returns with lower-frequency macroeconomic, policy, and environmental variables. However, their forecasting performance can deteriorate when either returns or explanatory variables contain extreme observations. A robust score-driven MIDAS volatility model, GAS-RMIDAS, is developed for mixed-frequency forecasting under data contamination. The model integrates Student-t score-driven dynamics for high-frequency conditional distribution with a robust MIDAS component that transforms low-frequency covariates through a bounded Student-t score before entering the long-run volatility equation. This limits the influence of extreme observations while preserving the information content of the low-frequency signal. Monte Carlo experiments compare GAS-RMIDAS with GARCH-MIDAS and standard GAS-MIDAS models under clean samples, return outliers, low-frequency covariate outliers, and joint contamination. Results show that GAS-type models are more robust to return outliers, while the RMIDAS extension stabilizes long-run parameter estimation and improves volatility forecasts when low-frequency variables are contaminated. An empirical application to gold, WTI crude oil, and lumber futures demonstrates that score-driven MIDAS models provide more reliable point and density forecasts around extreme commodity market events than conventional GARCH-MIDAS models.
