Abstract:
We develop a data-driven procedure to identify which correlations in high-dimensional dynamic systems should be time-varying, constant, or zero. The method integrates a vine-based multivariate partial correlation model with sequential penalized estimation. Applied to 50 US equities and systematic risk factors, results indicate that asset-level correlation dynamics are primarily induced by time-varying exposures to systematic factors. We further uncover persistent, non-zero, and occasionally time-varying partial correlations within industries, even after controlling for standard risk and industry factors. Finally, we show how the new methodology may be used to explore the relevance of systematic risk factors in an impartial way.