A comparison of vector autoregressive forecasting performance: spatial versus non-spatial Bayesian priors

Abstract

Forecasting performance of spatial versus non-spatial Bayesian priors applied to a large vector autoregressive model that includes the 48 lower US states plus and the District of Columbia is explored. Accuracy of one- to six-quarter-ahead personal income forecasts is compared for a model based on the Minnesota prior used in macroeconomic forecasting and a spatial prior proposed by Krivelyova and LeSage (J Reg Sci 39(2):297–317, 1999). While the Minnesota prior emphasizes time dependence taking the form of a random walk, the spatial prior relies on past values of neighboring state income growth rates while ignoring own-state past income growth. Our findings indicate that forecast accuracy for longer future time horizons is improved by the spatial prior, while that for shorter horizons is better for the non-spatial prior. This motivated a hybrid approach that combines both spatial and time dependence in the prior restrictions placed on the model parameters.