A Large non-Gaussian Structural VAR with Application to Monetary Policy

We develop a large structural vector autoregression (SVAR) that is statistically identified using higher-order moments under non-Gaussian, mutually independent shocks, without imposing economically motivated restrictions. Estimation is performed via an efficient Gibbs sampler that scales well with dimensionality, and we introduce an estimator of the Deviance Information Criterion to facilitate model comparison, including the evaluation of over-identifying economic restrictions. The model handles high-dimensional systems through a factor structure in the reduced-form errors, allowing the number of structural shocks to be substantially smaller than the number of observed variables while explicitly accounting for idiosyncratic noise. Experiments with simulated data show that ignoring this noise can lead to biased estimates. An empirical application to U.S. monetary policy demonstrates that expanding the information set can materially affect structural inference.