Abstract

The estimand in this paper is the source-screened component of the source-target covariance derivative after target-only movements have been projected out. A reduced local Gaussian block experiment turns that component into a projected covariance-score GMM statistic. The theory gives the projected information geometry, local Gaussian efficiency constants, a rough-regime rank condition, pilot-adaptive transfer under logarithmic roughness accuracy, and uniform minimax guarantees under the primitive Volterra reduction. Synthetic experiments check the implementation against closed-form information and noncentrality constants. In a balanced Oxford-Man realized-volatility panel of eight global equity indices, physical-measure roughness remains in the rough-volatility range, with SPX near $H_{\mathbb{P}}=0.071$. The full-sample directed map is dense, so the empirical output is screened intensity ranking and rolling stability rather than sparse edge recovery. Matched physical/risk-neutral classification remains a paired option-panel task.

Citation

Vidal Llauradó, Joan. “Detecting Latent Volatility Contagion: A Projected Score Estimator for Rough Volatility Models.” 2026. doi:10.2139/ssrn.6616118