Bayesian Compression for Mixed Frequency Vector Autoregressions: A Forecast Study for Germany

Abstract

Up until now, the concept of compression in single- or multivariate regressions has been limited to the common-frequency case. Having an application of macroeconomic forecasting in mind, one inevitably has to deal with variables sampled at various frequencies. Consequently, this work attempts to extend the concept of Bayesian Compressed Vector Autoregressions (BC-VAR) to the mixed-frequency case, leading to what could be labeled a compressed MF-BC-VAR. Starting off from a mixed-frequency VAR formulation, discussing alternative ways of incorporating mixed frequencies, this work demonstrates how to apply compression in this scenario. The empirical evaluation sketches the picture that not the entire variable set is necessary for GDP forecasting. However, the presented MF-BC-VAR model provides competitive results within the baseline evaluation, but suffers in a changing environment.

Keywords: Mixed Frequencies, VAR Models, Forecasting, Bayesian Methods, Random Compression

JEL Classification: C11, C32, C53