This paper compares the forecasting performance of random walk, frequentist vector autoregression (VAR), and Bayesian vector autoregression with Minnesota prior (BVAR) models on quarterly Russian data sample running from 1995 to 2014. Maximal number of variables included in the model is equal to 14 that requires an endogenous search of optimal shrinkage hyperparameter. The search procedure follows [Bańbura et al., 2010; Bloor, Matheson, 2011].According to the selection method the shrinkage hyperparameter equates the forecasting quality of the frequentist VAR and BVAR for the minimal considered dimension of the model (three variables). For any dimension of the BVAR model the optimal shrinkage hyperparameter is robust to considered functions of relative forecasting accuracy.

We show that the BVAR provides a more accurate forecast than the frequentist VAR on the studied sample. For key macro indicators (the industrial production index, consumer priceindex and the interbank interest rate), forecasting horizons, and all model sizes, the mean squared error of the BVAR is lower than that of the frequentist VAR. Moreover, the results show that the forecast made using the BVAR is more precise than the forecast made using random walk model for the CPI and using white noise model for the interbank rate. However, the BVAR cannot beat the random walk while forecasting the industrial production index.