Policy Uncertainty and Asset Volatility: A Reproducible Review and Controlled GARCH-Family Demonstration

Abstract

Where an uncertainty variable enters a volatility model determines the estimand, yet this distinction is often blurred. We combine a reproducible review of 36 GARCH-family applications with a controlled expanding-window demonstration. All primary S&P 500 comparisons use the same 1,193 forecast dates. Rankings vary across QLIKE, MSE, one- and five-day targets, transformations, and benchmarks; realized-volatility models lead point rankings, but 10% model-confidence sets retain all ten candidates. The observed one-day QLIKE gain is far below the design’s 80%-power minimum detectable effect. Official geopolitical-risk revisions are directionally positive in 52 of 58 months, showing that first releases systematically understate the later vintage, yet final-versus-real-time losses remain statistically indistinguishable after realized volatility is included. A 1% loss-trimming audit changes the QLIKE leader, revealing stress-day dependence. Applying the same model ladder to Japan, EUR/USD, and BRL/USD produces different winners across markets and losses. Thus uncertainty measures can be conditional signals, but neither coefficients nor rankings are portable across architecture, information set, target, benchmark, and market; the designs do not identify causal effects.

Author abstract from the research manuscript. Typography and line breaks have been normalized; the linked source may show a different manuscript version.