The Forecasting Accuracy of Regional Bayesian VAR Models with Alternative National Variable Choices

Abstract

The forecasting accuracy of vector autoregressive (VAR) and Bayesian VAR (BVAR) versions of five different regional models are compared for three states. Two base models for each state consist of three national variables and three state variables. Three misspecified models are then constructed by substituting or adding a simulated (random) national variable to the base models. Forecasting accuracy is measured by each model's ability to forecast the three state variables one, four, and eight quarters ahead. Inclusion of the random variable generally reduces the forecasting accuracy of unrestricted VAR models. The BVAR technique is superior to the VAR approach in terms of forecasting accuracy but, more important, is largely insensitive to the choice of national variables, even when the misspecified national variable is included.