Abstract:
|
Mixed frequency (MF) vector autoregressive models have been traditionally used to analyze collections of mixed frequency time series with at most one or two slow frequency time series. However, many larger mixed frequency datasets have many slow frequency series. Recently, MF VAR models have been shown to be generically identifiable from the first two moments. We argue that as the number of slowly sampled series increases the required identifiability assumptions become more likely to break down. We show that recently introduced non-Gaussian structural VAR models remain identifiable when the required genericity assumptions break down. For estimation, we model the non-Gaussian errors using a finite mixture of Gaussians and we develop a variational EM algorithm for scalable inference. For higher dimensional applications we additionally develop a screening procedure that first tests for potential non-identifiability from the first two moments. Our procedure divides the data set into two groups - those that may be identifiable from the first two moments and those that may not be identifiable. We apply our methodology to analyze GDP and other monthly indicators from multiple countries.
|