This is a fork of GluonTS accompanying the paper "High-Dimensional Multivariate Forecasting with Low-Rank Gaussian Copula Processes" accepted to NeurIPS 2019 as a poster.
The code has been rewritten in GluonTS, we recommend installing it the following way
git clone https://github.com/mbohlkeschneider/gluon-ts/
cd gluon-ts
git checkout mv_release
pip install -e .
To run the model:
python src/gluonts/multivariate/train_and_plot_predictions.py
This will run the model on the selected dataset with the hyperparameters in the paper (see supplementary material for details on the GluonTS implementation).
The results obtained with this implementation are as follow for CRPS:
| estimator | exchange | solar | elec | traffic | taxi | wiki |
|---|---|---|---|---|---|---|
| @GPCOP | 0.009+/-0.000 | 0.416+/-0.007 | 0.054+/-0.000 | 0.106+/-0.002 | 0.339+/-0.003 | 0.244+/-0.003 |
The model will also be released in GluonTS, this fork is created to keep a version with results as close as possible as the one published in the paper. While being close to the submission, the results are not exactly the same as the code was rewritten in GluonTS. However this version should be very close if not better.
If the datasets, benchmark, or methods are useful for your research, you can reference the following paper:
@article{lowrank_gp_multivariate_neurips,
title={{High-Dimensional Multivariate Forecasting with Low-Rank Gaussian Copula Processes}},
author={Salinas, D. and Bolhke-Schneider M. and Callot L. and Medicco R. and Gasthaus J.},
journal={International Conference on Neural Information Processing Systems},
series = {NEURIPS'19},
year={2019}
}