Identification and estimation of dynamic factor models



We propose a lag-augmented dynamic factor model which, compared existing models, allows for simple identification restrictions and in particular for fast estimation of the loadings as well as the factors. We establish consistency results for the loadings and the factors, and discuss model selection using e.g. suitable information criteria. The Hessian of the conditional (Gaussian) log-likelihood function is approximately block-diagonal, allowing us to replace each Newton-Raphson step by two steps along orthogonal directions in the parameter space. The optimization of the log-likelihood may thus be conducted as a sequence of simple Least-Squares regressions. The finite-sample statistical and numerical properties of the proposed procedure are found to be quite satisfactory in Monte Carlo simulations.


(joint work with Jörg Breitung (University of Cologne))