A1261
Title: From spectral decomposition to projection: Rethinking common components in dynamic factor models
Authors: Jan Bruha - Czech National Bank (Czechia) [presenting]
Abstract: The purpose is to revisit the estimation of the common component in dynamic factor models. Standard frequency-domain approaches reconstruct the common part via the inverse Fourier transform of spectral eigenvectors, which yields a two-sided filter. One-sided alternatives reduce to filtering conditional on observations at one particular date point, which, for some processes, may result in not efficiently capturing the lead-lag structure among series. A projection-based method is proposed that estimates the common component directly, using the spectral density to recover its autocorrelation function and cross-correlations with observables. This allows filtering based on any available data, even at the sample boundaries or in the presence of missing values due to publication lags. In addition, it is shown how this projection approach permits economically meaningful restrictions on filter weights, with an application to underlying inflation indicators.
