Seasonal BVAR models: A search along some time domain priors

Many authors have stressed the importance of using seasonally unadjusted data in modelling time series. However, few seasonal multivariate methodologies are currently available. In this paper, we extend the Bayesian Vector Autoregressive (BVAR) methodology of Litterman (1979, 1984, 1986), Doan, Litterman, and Sims (1984), and Sims (1989) to the context of seasonal time series. We examine some possible modifications to the regular Litterman's priors along three time domain specifications often encountered in seasonal time series modelling. We assess the forecasting performance of the seasonal BVAR models in the context of a monthly model of the U.S. economy (1967:1–1989:9), comprising both seasonal and nonseasonal variables.