A Bayesian Approach to Modelling Graphical Vector Autoregressions

We introduce a Bayesian approach to model assessment in the class of graphical vector autoregressive (VAR) processes. Due to the very large number of model structures that may be considered, simulation based inference, such as Markov chain Monte Carlo, is not feasible. Therefore, we derive an approximate joint posterior distribution of the number of lags in the autoregression and the causality structure represented by graphs using a fractional Bayes approach. Some properties of the approximation are derived and our approach is illustrated on a four-dimensional macroeconomic system and five-dimensional air pollution data.Causality; Fractional Bayes; graphical models; lag length selection; vector autoregression