A multilevel factor model with observation-driven factors

In this paper we introduce a new parsimonious multilevel factor model with observation-driven time-varying parameters that specify the factors' dynamics. The model incorporates several levels of factors. For example, in the two-level model there are common (first level) and group-specific (second level) factors, where the former is driven by all observations, while the latter is driven only by observations corresponding to its group. We propose a simple and fast estimation procedure for the loadings and factors, that does not explicitly require any identification restrictions. Moreover, all forecasts and impulse response functions follow straightforwardly from our model. We apply the proposed model to study the interconnectedness of the U.S. industries and their importance in economic activity. We find that the Industrial Production (IP) industries are related more to the aggregate shocks than the non-Industrial Production (non-IP) ones. On the other hand, the non-IP industries are strongly connected to other industries according to impulse responses, hence both IP and non-IP industries play a considerable role in the U.S. economy.