Probabilistic forecasting with modified N-BEATS networks
_{Jente Van Belle, Ruben Crevits, Wouter Verbeke [2024]}

This repository provides code for the paper "Probabilistic forecasting with modified N-BEATS networks" (IEEE TNNLS, 2024)

Installing

The requirements.txt provides the necessary packages. All code was written for python 3.10. Weights and Biases (W&B) is required to log the experiments.

Usage

To get the results for DeepAR, run DeepAR.ipynb for the different seeds as indicated in the notebooks.

To get the results for ETS, ETS-C1, and ETS-C2, run ETS_probabilistic.ipynb```. To get the results for ETS-C3, run ETS_C3.ipynb```.

To get the results for N-N-BEATS-C run the N-N-BEATS-C.ipynb notebook for the different seeds as indicated in the notebook.

To get the results for the different setups of N-N-BEATS-S, change the hyperparameters in the N-N-BEATS-S.ipynb notebook. Note that $\lambda$ in the paper has to be re-calculated to $\lambda$ in the code as follows: $\lambda_{code}=\lambda_{paper}/(\lambda_{paper}+1)$

The results are saved as .csv files. Save these files for all different setups in folders as indicated in the Evaluation.R file. Run this file to aggregate the results and generate the plots.

Citing

Please cite our paper and/or code as follows:

@article{van2024probabilistic,
  title={Probabilistic forecasting with modified N-BEATS networks},
  author={Van Belle, Jente and Crevits, Ruben and Caljon, Daan and Verbeke, Wouter},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2024},
  publisher={IEEE}
}