Forecast Disagreement and Conditional Model Fragility: Evidence from Eight Equity Volatility Markets

Abstract

Forecast disagreement is observable before a volatility forecast is evaluated, but it is not a sufficient model-selection rule. Across eight equity indices from 2006 to 2018, a frozen rule replaces XGBoost with HAR when dispersion among EWMA, HAR, and XGBoost forecasts exceeds its expanding 90th percentile. In the S&P 500, the rule identifies unusually large relative XGBoost losses under equal-frequency and run-preserving randomization. That timing evidence survives deleting the five largest contribution days and excluding 2008, but not deleting the largest ten; 77.5% of the total gain comes from 2008. Formal Giacomini-White tests do not reject conditional predictive equality, and always using HAR has lower point-estimate loss. A synchronized eight-market bootstrap supports a positive equal-weight effect but not the joint claim that every market effect is positive. An independent SPY archive replicates at five-minute, not one-minute, sampling; 2019 is a calm-period null. A two-day confirmation rule improves gain per activation while reducing switching, although all implementation rules are post-referee diagnostics without monetized costs. Disagreement is therefore a crisis-sensitive marker of model fragility, not a generally profitable selector or evidence of a ranking reversal.

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