A Multilevel Factor Model with Observation-Driven Dynamics
Authors: Mariia Artemova, Prof. Dr. Francisco Blasques, Prof. Dr. Siem Jan Koopman
Abstract: We propose a new parsimonious multilevel dynamic factor model where the unobserved factors are obtained using an observation-driven filter. Our model can describe rich factor dynamics and interdependence structures across multiple sectors of economic activity. We introduce a simple and fast estimation procedure for the loadings and factors based on the sequential least squares estimators. We establish the limit stationarity, ergodicity, bounded moments, and invertibility of our dynamic factor filter, and give conditions for the consistency of the least squares estimators. Additionally, we show how to produce forecasts and impulse response functions and discuss measures for group interconnectedness. Monte Carlo experiments demonstrate good finite sample performance in terms of extracting the factors and approximating the data. In application, we analyze the importance of the US industries in the US economic activity using a two-level model where the common factor is driven by all observations, while the idiosyncratic factors are driven only by the observations corresponding to its group. We find that aggregate co-movements are more related to the Industrial Production (IP) sectors than to the non-Industrial Production (non-IP) ones. Furthermore, we analyze the sector interdependence and how shocks spread through the network of interconnections using impulse response functions and network graphical measures.