Since this work was published as our master's research project, Patrick has continued to work on the deepvar package. 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. Since that bug has been removed from the package, that part of Table 1 can no longer be reproduced with the current package version (thanks to Hannes Osterchrist for flagging). In any case that part of the empirical exercise was somewhat flawed and will be removed (look-ahead bias). We have since moved to a rolling-window framework to assess the forecasting performance of our models (currently Figure 3 in the paper).
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 the deepvar package should help facilitate future research. We all have very limited bandwidth to continue work on this project at the moment, so this is project is officially idle. Do feel free to salvage the existing code though.
This repository contains all the code for Altmeyer, Agusti, and Vidal-Quadras Costa (2021). This research project started off as a master’s thesis project, but has since been carried forward and accepted for a poster presentation at the NeurIPS 2021 MLECON Workshop. It is worth flagging that we still consider this very much a work-in-progress - both the research and the companion package. We therefore very much welcome any feedback, suggestions and comments.
For comments regarding the research and methodology, please open an issue in this repository. For any concerns regarding the companion package please open an issue here.
Vector Autoregression is a popular choice for forecasting time series data. Due to its simplicity and success at modelling monetary economic indicators VAR has become a standard tool for central bankers to construct economic forecasts. A crucial assumption underlying the conventional VAR is that interactions between variables through time can be modelled linearly. We propose Deep VAR: a novel approach towards VAR that leverages the power of deep learning in order to model non-linear relatonships. By modelling each equation of the VAR system as a deep neural network, our proposed extension outperforms its conventional benchmark in terms of in-sample fit, out-of-sample fit and point forecasting accuracy. In particular, we find that the Deep VAR is able to better capture the structural economic changes during periods of uncertainty and recession. By staying methodologically as close as possible to the original benchmark, we hope that our approach is more likely to find acceptance in the economics domain.
A few useful pointers:
- At a glance: NeurIPS 2021 poster and video presentation.
- More detailed poster.
- Even more detailed slides.
- Full detail: paper.
- Code: R Package.
Please cite our paper as follows:
@article{altmeyer2021deep,
author = {Altmeyer, Patrick and Agusti, Marc and Vidal-Quadras Costa, Ignacio},
date-added = {2021-09-23 13:33:59 +0200},
date-modified = {2021-11-30 16:33:49 +0100},
title = {Deep Vector Autoregression for Macroeconomic Data},
url = {https://thevoice.bse.eu/wp-content/uploads/2021/07/ds21-project-agusti-et-al.pdf},
year = {2021}
}
Please cite the companion package as follows:
@software{Altmeyer_deepvars_Deep_Vector_2021,
author = {Altmeyer, Patrick},
month = {12},
title = {{deepvars: Deep Vector Autoregression}},
url = {https://github.com/pat-alt/deepvars},
version = {0.1.0},
year = {2021}
}