Please use this identifier to cite or link to this item: https://hdl.handle.net/10419/251551 
Authors: 
Year of Publication: 
2022
Publisher: 
ZBW - Leibniz Information Centre for Economics, Kiel, Hamburg
Abstract: 
Factor models feature prominently in the macroeconomic nowcasting literature, yet no clear consensus has emerged regarding the question of how many and which variables to select in such applications. Examples of both large-scale models, estimated with data sets consisting of over 100 time series as well as small-scale models based on only a few, pre-selected variables can be found in the literature. To adress the issue of variable selection in factor models, in this paper we employ sparse priors on the loadings matrix. These priors concentrate more mass at zero than those conventionally used in the literature while retaining fat tails to capture signals. As a result, variable selection and estimation can be performed simultaneously in a Bayesian framework. Using large data sets consisting of over 100 variables, we evaluate the performance of sparse factor models in real-time for US and German GDP point and density nowcasts. We find that sparse priors lead to relatively small gains in nowcast accuracy compared to a benchmark Normal prior. Moreover, different types of sparse priors discussed in the literature yield very similar results. Our findings are compatible with the hypothesis that large macroeconomic data sets typically used in now- or forecasting applications are not sparse but dense.
Subjects: 
factor models
sparsity
nowcasting
variable selection
JEL: 
C11
C53
C55
E37
Document Type: 
Working Paper

Files in This Item:





Items in EconStor are protected by copyright, with all rights reserved, unless otherwise indicated.