We construct Bayesian vector autoregressive (BVAR) models optimized by the Posterior Information Criterion (PIC), in which hyper-parameters are data-determined in the same way as the lag length and trend order. We also assess the performance of the selected models by one-step ahead forecasts...