M*-Lbvar: Large Bayesian Vector Autoregression With Macroeconomic Stars

Abstract

Large Bayesian vector autoregressions (LBVARs) are  becoming increasingly popular for macroeconomic forecasting, as they can incorporate large datasets, while avoiding overfitting to sample data. Motivated by the interest rate and inflation modeling literature, this study presents an extension of the LBVAR incorporating \textit{macroeconomic stars}, that is, the gradual trends in macroeconomic variables. For each variable, our prior specification enables to automatically determine whether a trend exists and, if it does, estimate the trend. The endpoints of the estimated trends are then used for forecasting. Marginal likelihood results provide evidence in favor of the proposed approach over standard LBVARs. Furthermore, an out-of-sample forecasting exercise shows that our model significantly improves the predictive accuracy for variables characterized by high persistence and for long-horizon predictions.

Keywords: Forecasting, shrinkage prior, trend-cycle decomposition