Improving Macroeconomic Forecasting by Accounting for Parameter Instability in Large-N Factor Models

Dimensionality reduction techniques have been widely researched in the literature and applied in a diversity of contexts. Dynamic approaches to specific Principal Components (PC) and Partial Least Squares (PLS) models have been developed to better exploit the properties of time series data. However, the time invariance of the loadings is continuously challenged in macroeconomic frameworks dealing with large sets of variables. In this paper, we test for the significance of this time instability modelled as a structural break and two-state Markov regime switches, and we design extended PC and PLS models that account for such non-constancy of loadings in the factor construction. Our results support existing findings on the presence of time variance and bring evidence of improved forecasting potential, particularly in the short term with Markov regime-switching models and over longer time horizons through the inclusion of a structural break.