Hierarchical PCA and Modeling Asset Correlations

Abstract

Modeling cross-sectional correlations between thousands of stocks, across countries and industries, can be challenging. In this paper, we demonstrate the advantages of using Hierarchical Principal Component Analysis (HPCA) over the classic PCA. We also introduce a statistical clustering algorithm to identify homogeneous clusters of stocks or “synthetic sectors”. We apply these methods to study the cross-sectional correlations in the US, Europe, China, and Emerging Markets.

Keywords: correlations, factor models, hierarchical PCA, statistical clusters

JEL Classification: C31, C32, C58, C01, C61