Forecasting Multivariate Realized Stock Market Volatility: PCA or MFA?

• Activity Number: 356
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Business and Economic Statistics Section
• Author(s): Xiaohang Wang*+ and Jianhua Zhao and Philip L.H. Yu
• Companies: The University of Hong Kong and Yunnan University of Finance and Economics and The University of Hong Kong
• Keywords: Matrix factor model ; Principal component analysis ; Vector autoregressive model ; Covariance separable

Abstract

Forecasting multivariate realized volatility matrix typically involves the dimensionality reduction of the realized covariance matrices. Two methods are considered: the principal component analysis (PCA) and the matrix factor model (MFA) recently proposed by Tao et al. (2011). Unlike PCA, MFA is based on 2D data so that the volatility matrix does not need to be vectorized before dimension reduction. In this talk, we compare these two methods theoretically and empirically. We show that MFA can be viewed as PCA under certain condition on the covariance matrix. Our experiments on simulated and real-world stock market data sets reveal that MFA can only outperform PCA when the sample size is small, says less than 40.