Scaled PCA: A new approach to dimension reduction

• Main Authors: HUANG, Dashan, JIANG, Fuwei, TONG, Guoshi, ZHOU, Guofu
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2019
• Subjects: Bond Return Predictability; Real Time Macro Data; Vintage; PCA; Big Data; Machine Learning; Corporate Finance; Finance and Financial Management
• Online Access: https://ink.library.smu.edu.sg/lkcsb_research/6216; https://ink.library.smu.edu.sg/context/lkcsb_research/article/7215/viewcontent/SSRN_id3358911.pdf
• Institution: Singapore Management University

The notion that bond risk premium varies with business cycles is challenged once real time macro data are used. In this paper, we argue that the macro factors extracted by using the standard PCA are not the most relevant for forecasting bond risk premium, because the PCA factors are designed to explain the most variation of macro data instead of the variation of bond risk premium. With the latter objective in mind, we propose a scaled PCA (sPCA) approach, which incorporates the information in bond risk premium in the factor extraction procedure. The real time macro sPCA factors have much stronger predictive power than the PCA factors, both in- and out-of-sample, and generate sizeable utility gains. Alternative approaches, target PCA and PLS, obtain similar results. The sPCA factors also strongly nowcast macro data revision and forecast future macroeconomic conditions, consistent with implications of standard asset pricing theories, and the forecasting power appears countercyclical, with expected bond returns high in recessions and low in expansions.