We consider the following problem: We are given a multivariate time series X. We assume that X together with a hidden multivariate time series Z forms a structural vector autoregressive (SVAR) process W with structural matrix A. The goal is to identify as much of A as possible, based on X alone. We show that under certain assumptions, using only X we can fully identify that part of A that captures the interaction between the components of X. The assumptions are: (1) at least half the components are observed, (2) the noise is non-Gaussian, and (3) two certain parameter matrices have full rank. This identifiability result may help to improve causal analysis of time series in certain cases.
View on arXiv