Forecasting economic activity with mixed frequency BVARs

Abstract

Mixed frequency Bayesian vector autoregressions (MF-BVARs) allow forecasters to incor- porate a large number of time series observed at dierent intervals into forecasts of economic activity. This paper benchmarks the performance of MF-BVARs in forecasting U.S. real Gross Domestic Product growth relative to surveys of professional forecasters and documents the in uence of certain specication choices. We nd that a medium-large MF-BVAR provides an attractive alternative to surveys at the medium term forecast horizons of interest to central bankers and private sector analysts. Furthermore, we demonstrate that certain specication choices such as model size, prior selection mechanisms, and modeling in levels versus growth rates strongly in uence its performance.