Multivariate Probabilistic Time Series Forecasting with Correlated Errors (NeurIPS 2024)

This repository contains the PyTorch implementation for the paper Multivariate Probabilistic Time Series Forecasting with Correlated Errors by Vincent Z. Zheng and Lijun Sun. This work has been accepted at NeurIPS 2024.

Requirements

The code requires Python 3.10 or later. The file requirements.txt contains the full list of required Python modules. To install the dependencies, run:

pip install -r requirements.txt

Training and Evaluation

1. Obtain the Dataset

Use the generate_datasets.ipynb notebook located in the exps directory to transform datasets from GluonTS and .h5 files into the TimeSeriesDataSet format used by PyTorch Forecasting.

2. Train the Model

To train a baseline model on the m4_hourly dataset, execute the following command:

python .\src\train_baseline.py --model deepar --dataset m4_hourly

To train the proposed model on the m4_hourly dataset, execute the following command:

python .\src\train_batch.py --model deepar --dataset m4_hourly --loss kernel --num_mixture_r 4

Reference

@article{zheng2024multivariate,
  title={Multivariate Probabilistic Time Series Forecasting with Correlated Errors},
  author={Zheng, Vincent Zhihao and Sun, Lijun},
  journal={arXiv preprint arXiv:2402.01000},
  year={2024}
}