Tuesday, September 22 @ 11:00 - 12:30 PM (ONLINE)
Hierarchical PCA and modeling asset correlations
Marco Avellaneda, NYU
ABSTRACT: Modeling return correlations between thousands of stocks poses great challenges, as empirical estimators tend to perform poorly when assets don’t share common risk factors, such as country or industry sector. In this paper, we show the advantages of using Hierarchical Principal Component Analysis (HPCA) for modeling correlations, as opposed to the classic PCA. Furthermore, we propose a statistical clustering approach using orthogonal decomposition, which seems to work outstandingly well in identifying “synthetic sectors”. To show the robustness of the results and make them fully reproducible, we conduct a comprehensive analysis of the stock markets of the US, Europe, China, and Emerging Markets.