NOTE
Disclaimer ⚠
Since we worked on this project in 2021, Patrick has continued to work on the deepvar package in his spare time. Among other things, he has found a bug in the original code, which has produced erroneous results for the test set in Table 1 of the paper that introduced the the approach (see here). We are also not convinced that the empirical results presented in the paper are robust, after looking at replicating the findings for another dataset. Nonethess, we believe that the proposed methodological framework is interesting and this package should help facilitate future researchers interested in exploring the applicability of deep learning to macroeconomic data. Since Patrick is caught up in his PhD for now, this project is officially idle. Do feel free to salvage what you can though.
deepvars
The deepvars package provides a framework for Deep Vector
Autoregression in R. The methodology is based on (Altmeyer, Agusti, and
Vidal-Quadras Costa 2021), a working paper initially prepared as part of
the Masters Degree in Data
Science
at Barcelona School of Economics. For a summary of the
first version of the working paper see
here.
Installation
Prerequisites
As one of its dependencies the deepvars uses tensorflow, which is an
R interface to the popular TensorFlow
library. We have tried to automate the TensorFlow configuration as
explained
here.
install.packages("tensorflow")
tensorflow::install_tensorflow()For uncertainty quantification we use tensorflow_probability for
Bayesian inference.
install.packages("tfprobability")
tfprobability::install_tfprobability()Should you run into issues you may have to manually install the TensorFlow dependencies. Detailed instructions to this end can be found here.
Install
You can either clone this repository and install from source or simply run the below in R:
devtools::install_github("pat-alt/deepvars", build_vignettes=TRUE)
library(deepvars)Getting started
Full documentation of the package is still a work-in-progress. In the
meantime, detailed guidance on different topics and estimation methods
covered by deepvars, can be found in the vignettes. Simply type the
following command once you have completed the steps above:
utils::browseVignettes('deepvars')Disclaimer
Date: 2 December, 2021.
This package was developed in tandem with the initial research for my masters thesis. Documentation is incomplete and it should at this point not be regarded as a fully-fledged, tested and production-ready piece of software, so please bear this in mind. That being said, I’m quite confident about the basic functionality of training and predicting from a Deep VAR as well as various plotting methods that can be used for visualizing the results. I encourage you to try it out yourself and should you encounter any problem, please just open an issue.