Forecast accuracy, coefficient bias and Bayesian vector autoregressions Original Research Article

A Bayesian vector autoregression (BVAR) can be thought of either as a method of alleviating the burden of the over-parameterisation usually associated with unrestricted VARs, or as a method of correcting coefficient bias when the time series are nonstationary. Monte Carlo evidence is provided to show that the latter appears to be a more important characteristic of BVARs in experiments using a 4-equation cointegrated system, and with that system embedded in a 10-equation model containing six extraneous random walks. It is found that the BVAR model generally performs much better than a VAR in levels and is a viable alternative to a vector error correction model. It is also found that estimating constant terms when there is no drift in the data causes a major deterioration in forecasting performance.