2024-12-18

Replication repository for “Forecasting macroeconomic data with Bayesian VARs: Sparse or dense? It depends!”

Luis Gruber (luis.gruber@aau.at) and Gregor Kastner.

Preprint available at Arxiv. Working paper is conditionally accepted at International Journal of Forecasting.

Overview

This repository provides the replication material for the paper “Forecasting macroeconomic data with Bayesian VARs: Sparse or dense? It depends!”. Each figure and table is generated separately by its corresponding script file figure_[x].R or table_[x].R, respectively.

The main contents of the repository are the following:

• tables/: Folder containing the generated tables as .txt files.
• figures/: Folder containing the generated figures as .pdf or .png files.
• data/: Folder containing the data used in the empirical application.
• results/: Folder containing intermediary results which were generated using a computer cluster.
• computer_cluster_scripts/: Folder containing the computer cluster scripts.
• figure_[x].R: R scripts to generate the respective figures.
• table_[x].R: R scripts to generate the respective tables.
• bayesianVARs_0.1.3.tar.gz: Source files of the R package bayesianVARs (0.1.3).
• bayesianVARs_0.1.3.zip: A .zip file containing the binary builds of the R package bayesianVARs (0.1.3).

Instruction & computational requirements

All file paths are relative to the root of the replication repository. Please set your working directory accordingly.

All the estimation and analysis is done in R. The main package used for estimation and forecasting is bayesianVARs (0.1.3) with major dependency stochvol (3.2.0), where the values in parenthesis indicate the package versions we used. In addition, the following add-on packages are loaded and attached in the replication scripts (in alphabetical order): coda (0.19-4), colorspace (2.1-1), dplyr (1.1.4), ggplot2 (3.5.1), gsl (2.1-8), knitr (1.49), lubridate (1.9.3), mvtnorm (1.3-2), purrr (1.0.2), readr (2.1.5), reticulate (1.40.0), RhpcBLASctl (0.23-42), scatterplot3d (0.3-44), stringr (1.5.1), tibble (3.2.1), tidyr (1.3.1) and xts (0.13.0).

The files figure_[x].R and table_[x].R can be run individually in any order. On a standard computer the execution of each those files should take less than one minute, except figure_3.R, which should take less than five minutes.

We provide the source code of bayesianVARs in the file bayesianVARs_0.1.3.tar.gz. A great deal of the package is written in C++. Therefore, suitable compilers and related tools (e.g. Rtools on Windows) need to be installed in order to install the package from source. Additionally, we provide the file bayesianVARs_0.1.3.zip containing the binary builds. E.g., on Windows you can install the package in R with install.packages("bayesianVARs_0.1.3.zip").

The Python library mpmath is required in order to fully reproduce figures/figure_1.pdf. In R it is possible to call Python via the R package reticulate (see the instructions in figure_1.R).

The computations of the main results, more concretely the simulation study in Section 4.2 and the empirical application in Section 5, were carried out on a computer cluster with 3328 CPUs managed with the Slurm Workload Manager. The corresponding R scripts are located in the folder computer_cluster_scripts/. Those scripts should not be executed on standard computers. The script computer_cluster_scripts/simulation_study_clusterscript.R estimates 3240 models. The script computer_cluster_scripts/empirical_application_clusterscript.R estimates and evaluates forecasts of 284970 models. Even on a computer cluster it can take several days or even weeks to run those scripts, depending on the available resources.

Data

The data of the empirical application consists of 21 selected quarterly time-series obtained from the FRED-QD data base (McCracken and Ng, 2021, https://doi.org/10.20955/r.103.1-44). Release date 2021-07 (https://files.stlouisfed.org/files/htdocs/fred-md/quarterly/2021-07.csv). The data is transformed to be interpreted as growth-rates (first log-differences with the exception of interest rates, which are already growth rates). The prepared data is located at data/data_growth.RData.

In Figures 6 through 8 the NBER recession are highlighted among other things. We obtained the relevant data from the R package tis and provide it as .rds file at data/nber_recessions.rds.