Large Bayesian vector autoregressive forecasting for regions: A comparison of methods based on alternative disturbance structures

Abstract

Bayesian vector autoregressions that allow for non-Gaussian, heteroscedastic and time-dependent disturbance structures for models involving a large number of variables have been recently introduced in the literature. A comparison of forecasting accuracy is carried out using regional time series for models based on the conventional assumption of homoscedastic Gaussian disturbances versus models that allow for non-Gaussian, heteroscedastic and time-dependent disturbance structures. The comparison was based on 12-month-ahead forecast for the 48 lower US states over two three-year time periods, 2012–2014 and 2015–2017. The non-Gaussian models were found to provide large improvements in both forecast accuracy and forecast precision for almost all states. In addition, we explored a (non-spatial) ridge-type (Minnesota) prior distribution for the parameters versus a spatial prior distribution that utilizes information on growth rates from neighboring regions. This comparison resulted in smaller improvements in forecast accuracy and precision than those associated with relaxing the assumption of homoscedastic Gaussian disturbances.