Asset-Specific, State-Dependent, and Target-Sensitive Volatility Modeling: Evidence from GARCH-Family Forecasts Across Asset Classes

This paper studies whether the relative performance of volatility models differs across asset classes and whether the differences can help improve forecasting. The study uses daily data for U.S. equities (SPY), investment-grade bonds (AGG), broad commodities (DBC), gold (GLD), and listed real estate (IYR) and compares GARCH, EGARCH, FIGARCH, and component GARCH within a common recursive out-of-sample framework. Historical MS-GARCH models are used to identify high- and low-volatility periods, while VIX-based and asset-specific volatility states are used to examine changes in forecast performance across market conditions. Forecast routing, GAS Tree parameter adaptation, and machine-learning-assisted model selection are also examined. SPY shows recurring support for EGARCH, but the best-performing fixed specification differs across assets. Some model comparisons show state-dependent performance, but routing and machine-learning selectors do not consistently outperform strong fixed benchmarks. GAS Tree provides selective gains, especially for AGG, but its advantage over Standard GAS reverses under the Garman–Klass proxy. Model suitability therefore depends on the asset, market conditions, and evaluation target.