Bayesian Averaging of Classical Estimates in Asymmetric Vector Autoregressive (AVAR) Models

The estimated Vector AutoRegressive (VAR) model is sensitive to model misspecifications, such as omitted variables, incorrect lag-length, and excluded moving average terms, which results in biased and inconsistent parameter estimates. Furthermore, the symmetric VAR model is more likely misspecified due to the assumption that variables in the VAR have the same level of endogeneity. This paper extends the Bayesian Averaging of Classical Estimates, a robustness procedure in cross-section data, to a vector time-series that is estimated using a large number of Asymmetric VAR models, in order to achieve robust results. The combination of the two procedures is deemed to minimize the effects of misspecification errors by extracting and utilizing more information on the interaction of the variables, and cancelling out the effects of omitted variables and omitted MA terms through averaging. The proposed procedure is applied to simulated data from various forms of model misspecifications. The forecasting accuracy of the proposed procedure was compared to an automatically selected equal lag-length VAR. The results of the simulation suggest that, under misspecification problems, particularly if an important variable and MA terms are omitted, the proposed procedure is better in forecasting than the automatically selected equal lag-length VAR model.