Identification and Estimation of Causal Factor Models of Stationary Time Series

This paper makes three contributions to the literature on dynamic factor analysis. Firstly, we investigate the identification problem for a general class of causal dynamic factor model and provide conditions under which the model is identified. Secondly, we present an analytical expression for the information matrix of an autoregressive factor model which can be computed far more efficiently than the standard numerical expression, and thirdly we propose an accelerated EM algorithm which has the same convergence properties as the traditional scoring algorithm but has the same storage and CPU-time requirements per iteration as the standard EM algorithm. We illustrate the very significant computational gains over the standard approach with simulations.