DeepRRTime: Robust Time-series Forecasting with a Regularized INR Basis (TMLR 2025)

Figure 1. Overall approach of DeepRRTime.

Official PyTorch code repository for the DeepRRTime paper. DeepRRTime advances state-of-the-art in time-series forecasting amongst deep time-index models, a recent modeling paradigm for time-series forecasting.

Requirements

Dependencies for this project can be installed by:

pip install -r requirements.txt

Experiments

Steps to reproduce the results in Tables 1 and 2:

1. Download datasets

   • Pre-processed datasets can be downloaded from the following links, Tsinghua Cloud or Google Drive, as obtained from Autoformer's GitHub repository.
   • Place the downloaded datasets into the storage/datasets/ folder, e.g. storage/datasets/ETT-small/ETTm2.csv.

2. Generate experiments for various combinations of forecast-horizons (e.g., 96, 192, 336 or 720), lookback multipliers (e.g., 1, 3, 5, 7 or 9) and regularization options (e.g., none/orth1.0).

   • To generate all experiments for a single dataset, you can run:make build-all path=experiments/configs/Exchange/
   • Likewise, to generate all experiments for all datasets, you can run:make build-all path=experiments/configs/*

3. Run all experiments: sh run.sh

4. Finally, you can observe the results on tensorboard tensorboard --logdir storage/experiments/ or view the storage/experiments/**/metrics.npy file. The hyperparameters were chosen based on the validation MSE.

Acknowledgements

The implementation of DeepRRTime heavily relies on the original DeepTime implementation (https://github.com/salesforce/DeepTime). We thank the original authors for open-sourcing their work. Compared to the original implementation, only the following python files were updated:

• experiments/base.py
• experiments/forecast.py
• models/DeepTIMe.py
• models/modules/regressors.py

Citation

To cite our work please use the following reference:

@article{
    sastry2025deeprrtime,
    title={Deep{RRT}ime: Robust Time-series Forecasting with a Regularized {INR} Basis},
    author={Chandramouli Shama Sastry and Mahdi Gilany and Kry Yik-Chau Lui and Martin Magill and Alexander Pashevich},
    journal={Transactions on Machine Learning Research},
    issn={2835-8856},
    year={2025},
    url={https://openreview.net/forum?id=uDRzORdPT7},
    note={}
}