ReNF: Rethinking the Design of Neural Long-Term Time Series Forecasters.

July 16, 2026 ยท View on GitHub

https://arxiv.org/html/2509.25914v6 https://img.shields.io/github/license/Luoauoa/ReNF

This is the official implementation of ReNF: Rethinking the Design of Neural Long-Term Time Series Forecasters.

ReNF has been accepted by ICML2026.

(The model was upgraded a little bit after the version in the paper, feel free to contact me if any problems reproducing it.)

Quick Start

  1. Download datasets from Google Drive or Baidu Cloud

  2. Update dataset paths in the scripts/ directory to match your local setup

  3. Run experiments using the provided scripts:

    bash ./scripts/traffic.sh      # Traffic dataset
bash ./scripts/electricity.sh  # Electricity dataset
bash ./scripts/weather.sh     # Weather dataset
# ... other datasets available

Note on adapting for other backbones.

To adapt BDO to other backbones, such as a transformer-based forecaster. It should identify the minimum but useful forecasting block as the sub-forecaster.

That means the sub-forecaster is at least already capable of generating moderately good forecasts. Then, applying BDO to connect these sub-forecasters can probably lead to better results when compared with the original version.

Cite

@article{lu2025renf,
  title={ReNF: Rethinking the Design Space of Neural Long-Term Time Series Forecasters},
  author={Lu, Yihang and Meng, Xianwei and Chen, Enhong},
  journal={arXiv preprint arXiv:2509.25914},
  year={2025}
}