Conditioning on a Volatility Proxy Compresses the Apparent Timescale of Collective Market Correlation

Battle Count: 64

The paper is primarily a methodological/attribution study rather than a direct trading strategy paper. However, it has significant indirect relevance: (1) Understanding that collective correlation persistence is largely inherited from VIX rather than intrinsic changes how one models correlation dynamics for portfolio risk; (2) The finding that conditioning on VIX compresses the effective timescale from 298 to 61 days has implications for how quickly correlation regimes shift; (3) The mechanical/informational decomposition clarifies what portion of VIX-correlation coupling is structural vs. informational; (4) The regime-switching hybrid model identifies a rare extreme-event channel (stress state lasting ~1.1 days) relevant for tail-risk management; (5) The negative result on orthogonal residual forecasting tempers expectations about additional predictive signals beyond VIX. The work is more relevant to risk management and systemic risk monitoring than to alpha generation or direct trading signals.

The core OU model fitting with exact Gaussian likelihood is straightforward. However, the full pipeline involves: (1) constructing 60-day rolling correlation matrices for 237 stocks; (2) computing leading eigenvalue fractions; (3) fitting multiple stochastic models (bare OU, field-coupled OU, quartic variants, regime-switching with Hamilton filter, state-dependent noise); (4) generating 100 AR(9) placebo fields; (5) mechanical/informational decomposition with multiple recipe variants; (6) quiet-regime ACF pooling with bootstrap; (7) disjoint weekly and 60-day block reconstructions; (8) exact 2D linear-Gaussian VAR fitting with matrix logarithm and Lyapunov equation; (9) orthogonal residual phase-space analysis; (10) six anchored chronological holdouts. The mathematical derivations (Appendix A) are non-trivial, and the layered experimental design requires careful implementation to avoid circularity. The BIC-based model comparison and pseudo-true interpretation add conceptual complexity.

The paper provides extremely detailed methodological descriptions, exact likelihood derivations in Appendix A, and full execution protocols in Appendix B. Data sources are specified (FRED for VIX, standard S&P 500 constituents). However, processed datasets and analysis code are only available from authors on reasonable request, not publicly hosted. The 100 placebo simulations use a fixed seed protocol but the seed is not specified. Window sweep and robustness experiments are fully specified. The main limitation for reproducibility is the lack of a public code/data repository.