Gold, Bitcoin, and equity market stress: Evidence from Double Machine Learning

We reexamine whether Gold and Bitcoin protect U.S. equity portfolios during market stress. Using daily data from December 2014 to February 2026, we combine a contemporaneous Baur–Lucey safe-haven benchmark, horizon-specific local projections, and an orthogonalised Double Machine Learning (DML) stress-interaction estimator. The integrated design separates three questions that the literature usually conflates: whether an asset hedges equities on average, whether it protects on the day of a stress realisation, and whether any protection persists over the subsequent holding period. The headline finding is for Bitcoin: the contemporaneous stress interaction with the S&P 500 is δ0=0.124 with p≈1.1×10−6, positive (the opposite of the safe-haven prediction) and Bonferroni-significant at the family-wise 5% level. Under VIX-defined stress, the orthogonalised DML stress interaction for Bitcoin is positive and Bonferroni-significant at three of five horizons; the more demanding nonlinear conditioning strategy reinforces rather than overturns the rejection. Gold’s stress interactions are weak, unstable, and imprecisely estimated; we do not reject a defensive role for Gold but the data do not support a robust safe-haven role either. A multi-allocation portfolio grid and a spot-Gold robustness exercise corroborate the regression evidence. Economically, the results are more consistent with liquidity-driven liquidation and risk-on/risk-off dynamics than with either asset as reliable equity crash insurance.