Title: Forecasting with large Bayesian VARs: On the importance of the prior
Authors: Jamie Cross - BI Norwegian Business School (Norway) [presenting]
Aubrey Poon - university of strathclyde (United Kingdom)
Chenghan Hou - Hunan University (China)
Abstract: Substantive empirical evidence has shown that large Bayesian VARs can provide better in- and out-of-sample fit compared to smaller scale models. When specifying such models, a multitude of hierarchical shrinkage priors on the autoregressive coefficients have been proposed. We provide a detailed comparison of six of these prior distributions: Minessota, lasso, SVSS, Dirichlet-Laplace, normal-gamma and horseshoe priors. Additionally, we show how each of these priors can be implemented in a stochastic volatility framework in a computationally efficient manner. At the beginning we compare the relative in-sample fit and out-of-sample forecast performance on a frequently used large data set of US macroeconomic and financial variables. The primary result is that the Horseshoe prior dominates all alternatives both in- and out-of-sample. We then conduct a simulation exercise which explores the possible reasons for this improvement.