A Time-Varying Bayesian Compressed Vector Autoregression for Macroeconomic Forecasting

This paper presents macroeconomic forecasting by using a time-varying Bayesian compressed vector autoregression approach. We apply a random compression by using projection matrix to randomly select predictive variables in vector autoregression (VAR), and then perform true out-of-sample forecast where the forecast values are averaged across all estimated models, containing different in both explanatory variables and number of those variables by using Bayesian model averaging (BMA). In addition to this, we allow the parameters in Bayesian compressed VAR to be time-varying by implementing dynamic model averaging (DMA) algorithm that is applicable with VAR using forgetting factor to control the degree of time-varying in the estimating parameters. We validate the performance of the proposed method via real macroeconomic data including up to 53 variables. The empirical results demonstrate that the predictive performance of time-varying Bayesian compressed VAR can beat traditional VAR types which are considered to have a potentiality to deal with large size variables.