A New Approach to Macroeconomic Time Series Modeling and Its Applications to Monetary Policy and Financial Markets

Abstract

Vector autoregression (VAR) is one of the most widely used multivariate time series in econometrics, particularly for identifying and measuring the effects of monetary policy innovations on key macroeconomic variables. A popular way to capture information related to the variables to be forest from a large amount of macroeconomic variables is to encode them by a few factors, which are currently constructed based on principal component analysis (PCA). We introduce a new approach using reduced-rank regression to construct factors, which have higher predictive power for forecasting. Previous work by Bai and Ng (2008, Journal of Econometrics) has shown that the factors that are constructed by PCA can be improved if variable selection methods are first used to filter out irrelevant variables. These high-dimensional variable selection methods have been developed under the assumption of i.i.d. observations, which does not apply to time series data. We propose the group orthogonal greedy algorithm (GOGA) to select variables before factor construction and also provide an asymptotic theory for GOGA for time series data. Simulation studies have shown that our approach outperforms the PCA-based approach and another approach recently proposed by Bunea, Wegkamp and She (2011, 2012, Annals of Statistics). Some applications of our approach to monetary policy and interest-rate markets are also discussed. This is a joint work with Tze Leung Lai (Stanford University).