Time Series Forecasting Through the Lens of Dynamics - ICML 2026

Important

You can find the interactive version of this paper here, including 40+ TSF models analyzed through our dynamics-based nomenclature !

This repository contains the code developed for the paper "Time Series Forecasting Through the Lens of Dynamics" accepted at ICML 2026. It is based on this repository, developed for the TFB benchmark (Qiu et al. 2024).

Installation

Requirements

Please see the requirements.txt file.

• tested on Python 3.9.7 and 3.10.7
• tested on PyTorch 1.12.1 and 2.1.2

Please refer to the original repository detailed instruction guide to setup your environment, if needed.

Datasets

All the datasets are accessible with the following link: Google Drive. Place the downloaded data under the folder ./dataset/forecasting/. We provide it with the three lightest datasets ILI, NASDAQ and NYSE, with the metadata FORECAST_META.scv for the forecasting task, for convenience.

Paper files description

From the .\ts_benchmark\baselines\time_series_library\ folder, the models are in the models folder. Each model has a method for a specific time-series task. The time-series forecasting is identified by the methods forecast or short_term_forecast.

Get started

The train and test file is adapters_for_transformers.py, where default hyperparameters are set.

The scripts, where the hyperparameters are indicated, are under the .\scripts\multivariate_forecast\ folder. They are stored in .sh files.

To run a model from the shell, follow this example:

sh ./scripts/multivariate_forecast/ILI_script/Informer_DYN.sh

When running under pycharm, please escape the double quotes, remove the spaces, and remove the single quotes at the beginning and end.

Such as: '{"d_ff": 2048, "d_model": 512, "horizon": 24}' ---> {\"d_ff\":2048,\"d_model\":512,\"horizon\":24}

--config-path "rolling_forecast_config.json" --data-name-list "ILI.csv" --strategy-args {\"horizon\":24} --model-name "time_series_library.Informer_DYN" --model-hyper-params {\"d_ff\":2048,\"d_model\":512,\"factor\":3,\"horizon\":24,\"norm\":true,\"seq_len\":104} --adapter "transformer_adapter" --gpus 0 --num-workers 1 --timeout 60000 --save-path "ILI/Informer_DYN"

Once a model is run, results are stored in the .\result\arg(save_path) folder, where arg(save_path) is the path indicated in the --save_path argument when running the script. MSE and MAE are reported under mse_norm and mae_norm.

Citation

If you find this repo useful, please cite our paper.

@misc{brachet2026timeseriesforecastinglens,
      title={Time Series Forecasting Through the Lens of Dynamics}, 
      author={Alexis-Raja Brachet and Pierre-Yves Richard and Céline Hudelot},
      year={2026},
      eprint={2507.15774},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2507.15774}, 
}

References

Qiu, X., Hu, J., Zhou, L., Wu, X., Du, J., Zhang, B., Guo, C., Zhou, A., Jensen, C. S., Sheng, Z., & Yang, B. (2024). TFB : Towards Comprehensive and Fair Benchmarking of Time Series Forecasting Methods. arXiv.org. https://arxiv.org/abs/2403.20150