Implementation of the paper "TimeRecipe: A Time-Series Forecasting Recipe via Benchmarking Module Level Effectiveness."

Authors: Zhiyuan Zhao, Juntong Ni, Haoxin Liu, Shangqing Xu, Wei Jin, B.Aditya Prakash

Placement: ICLR 2026

Paper + Appendix: https://arxiv.org/abs/2506.06482

Please follow the training scripts provided in TimeRecipeResults.

To train a single setup

python -u run.py --seed 2021 --task_name long_term_forecast --use_norm True --use_decomp True --fusion temporal --emb_type token --ff_type mlp --${Other Args}$

To train a batch of setup

bash scripts/ecl_96_m/2021.sh 

or a customized batch of experiments aross datasets

bash run_2021.sh

All raw and processes results can be found at TimeReciperesults.

1. ./notebook/error_rank.ipynb: Convert the raw forecasting results over different random seeds to ranked results with averaged error and std.
2. ./notebook/read_res_m.ipynb: Filter and combine the top 30 ranked results (top_k=30) from different datasets to a single csv file.
3. ./notebook/cor_ana_m.ipynb: Perform statistic testing for the correlation analysis, using the combined csv file (Paper Table 3).
4. ./notebook/lightgbm_m.ipynb: Perform the training-free model selection using a LightGBM model and pre-trained results (Paper Table 2).
5. ./notebook/count_surpass.ipynb: Count the number of setups that TimeRecipe outperforms SOTA (Paper Section 4.1.1).

For data properties calculation, please follow: [Code], [Setup(en)], [Setup(cn)].

If you have any questions about the code, please contact Zhiyuan Zhao at leozhao1997[at]gatech[dot]edu.

If you find our work useful, please cite our work:

@article{zhao2025timerecipe,
  title={TimeRecipe: A Time-Series Forecasting Recipe via Benchmarking Module Level Effectiveness},
  author={Zhao, Zhiyuan and Ni, Juntong and Xu, Shangqing and Liu, Haoxin and Jin, Wei and Prakash, B Aditya},
  journal={arXiv preprint arXiv:2506.06482},
  year={2025}
}

This work also builds on previous works, please consider cite these works properly.

Time Series Library (TSLib). [Code]

TFB: Towards Comprehensive and Fair Benchmarking of Time Series Forecasting Methods. [Paper][Code]