Data-Driven Inference on Sign Restrictions in Bayesian Structural Vector Autoregression


Sign-identified structural vector autoregressive (SVAR) models have recently become popular. However, the conventional approach to sign restrictions only yields set identification, and implicitly assumes an informative prior distribution of the impulse responses whose influence does not vanish asymptotically. In other words, within the set the impulse responses are driven by the implicit prior, and the likelihood has no significance. In this paper, we introduce a Bayesian SVAR model where unique identification is achieved by statistical properties of the data. Our setup facilitates assuming a genuinely noninformative prior and thus learning from the data about the impulse responses. While the shocks are statistically identified, they carry no economic meaning as such, and we propose a procedure for labeling them by their probabilities of satisfying each of the given sign restrictions. The impulse responses of the identified economic shocks can subsequently be computed in a straightforward manner. Our approach is quite flexible in that it facilitates labeling only a subset of the sign-restricted shocks, and also concluding that none of the sign restrictions is plausible. We illustrate the methods by two empirical applications to U.S. macroeconomic data.

5 Figures and Tables

Cite this paper

@inproceedings{Lanne2016DataDrivenIO, title={Data-Driven Inference on Sign Restrictions in Bayesian Structural Vector Autoregression}, author={Markku Lanne and Jani Luoto}, year={2016} }