Prior Parameter Uncertainty: Some Implications for Forecasting and Policy Analysis with VAR Models
John C. Robertson and Ellis W. Tallman
Federal Reserve Bank of Atlanta
Working Paper 99-13
Models used for policy analysis should generate reliable unconditional forecasts as well as policy simulations (conditional forecasts) that are based on a structural model of the economy. Vector autoregression (VAR) models have been criticized for having inaccurate forecasts as well as being difficult to interpret in the context of an underlying economic model. In this paper, we examine how the treatment of prior uncertainty about parameter values can affect forecasting accuracy and the interpretation of identified structural VAR models.
Typically, VAR models are specified with long lag orders and a diffuse prior about the unrestricted coefficients. We find evidence that alternatives that emphasize nonstationary aspects of the data as well as parsimony in parameterization have better out-of-sample forecast performance and smoother and more persistent responses to a given exogenous monetary policy change than do unrestricted VARs.
JEL classification: E44, C53
Key words: Bayesian inference, prior distributions, out-of-sample forecasting, structural VAR models, impulse responses
The authors thank Tao Zha for helpful discussions and the participants of the Atlanta Fed brown bag lunch series for comments. The views expressed here are the authors' and not necessarily those of the Federal Reserve Bank of Atlanta or the Federal Reserve System. Any remaining errors are the authors' responsibility.
Please address questions regarding content to John C. Robertson or Ellis W. Tallman, Research Department, Federal Reserve Bank of Atlanta, 104 Marietta Street, NW, Atlanta, Georgia 30303-2713, 404/498-8782 (Robertson), firstname.lastname@example.org, 404/498-8915 (Tallman), email@example.com.