A1289
Title: Using subspace algorithms to estimate the factor dynamics in generalized dynamic factor models
Authors: Dietmar Bauer - Bielefeld University (Germany) [presenting]
Abstract: Generalized dynamic factor models decompose a stationary, high-dimensional time series into an idiosyncratic part (specific to the different variables and only weakly correlated across variables) and a factor part (containing variables relevant for numerous variables). While the idiosyncratic part is typically considered noise and is filtered out, the factor part is often assumed to be stationary with rational spectral density spanning a lower-dimensional subspace. Estimating this subspace is usually done using principal component analysis, while for the estimation of the corresponding dynamics, the singularity of the process (the dynamic factors are typically fewer than the static factors) poses challenges. The corresponding tall rational transfer function does not have a unique left pseudo-inverse. Subspace methods like the canonical variate analysis (VA) of Larimore, which are based on consistent state sequence estimation, are shown to lead to consistent estimation of the dynamics in this situation.
