FreDF: Learning to Forecast in the Frequency Domain

Welcome to FreDF

Enhancing Time-series forecasting performance with one-line code.

The repo is the official implementation for the paper: FreDF: Learning to Forecast in the Frequency Domain.

We provide the running scripts to reproduce experiments in /scripts, which covers three mainstream tasks: long-term forecasting, short-term forecasting, and imputation. We also provide the scripts to reproduce the baselines, which mostly inherit from the comprehensive benchmark.

🤗 Please star this repo to help others notice FreDF if you think it is a useful toolkit. Please kindly cite FreDF in your publications if it helps with your research. This really means a lot to our open-source research. Thank you!

🚩News (2024.12) FreDF has been accepted as a poster in ICLR-25: [paper] [slide] [Video]

🚩News (2024.2) A blog in Chinese to introduce this work is available.

🚩News (2023.12) The implementation of FreDF is released, with scripts on three tasks.

Leaderboard

We maintain an updated leaderboard for time series analysis models, with a special focus on learning objectives. As of December 2024, the top-performing models across different tasks are:

Note: We will keep updating this leaderboard. If you have proposed advanced and awesome models, you can send us your paper/code link or raise a pull request. We will add them to this repo and update the leaderboard as soon as possible.

Compared models of this leaderboard. ☑ means that their codes have already been included in this repo.

• ☑ iTransformer - iTransformer: Inverted Transformers Are Effective for Time Series Forecasting [arXiv 2023] [Code].
• ☑ PatchTST - A Time Series is Worth 64 Words: Long-term Forecasting with Transformers [ICLR 2023] [Code].
• ☑ TimesNet - TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis [ICLR 2023] [Code].
• ☑ DLinear - Are Transformers Effective for Time Series Forecasting? [AAAI 2023] [Code].
• ☑ FEDformer - FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting [ICML 2022] [Code].
• ☑ Autoformer - Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting [NeurIPS 2021] [Code].
• ☑ Transformer - Attention is All You Need [NeurIPS 2017] [Code].
• ☑ TiDE - Long-term Forecasting with TiDE: Time-series Dense Encoder [arXiv 2023] [Code].
• ☑ Crossformer - Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting [ICLR 2023][Code].

Usage

0. Implement FreDF by adapting the following script in your pipeline

# The canonical temporal loss
loss_tmp = ((outputs-batch_y)**2).mean()
# The proposed frequency loss
loss_feq = (torch.fft.rfft(outputs, dim=1) - torch.fft.rfft(batch_y, dim=1)).abs().mean() 
# Note. The frequency loss can be used individually or fused with the temporal loss using finetuned relative weights. Both witness performance gains, see the ablation study in our paper.

1. Install Python 3.8 and pytorch 1.12.1. For convenience, execute the following command.

pip install -r requirements.txt

2. Prepare Data. You can obtain the well pre-processed datasets from [Google Drive] or [Baidu Drive], Then place the downloaded data in the folder ./dataset. Here is a summary of supported datasets.

3. Train and evaluate model. We provide the experiment scripts for all benchmarks under the folder ./scripts/. You can reproduce the experiment results as the following examples:

# long-term forecast
bash ./scripts/ltf_overall/ETTh1_script/iTransformer.sh
# short-term forecast
bash ./scripts/stf_overall/FreTS_M4.sh
# imputation
bash ./scripts/imp_autoencoder/ETTh1_script/iTransformer.sh

4. Apply FreDF to your own model.

• Add the model file to the folder ./models. You can follow the ./models/iTransformer.py.
• Include the newly added model in the Exp_Basic.model_dict of ./exp/exp_basic.py.
• Create the corresponding scripts under the folder ./scripts. You can follow ./scripts/ltf_overall/ETTh1_script/iTransformer.sh.

Citing FreDF

The paper introducing FreDF is available in ICLR-25. If you use FreDF in your work, please consider citing it as below and 🌟staring this repository to make others notice this library. 🤗

@inproceedings{wang2025fredf,
    title = {FreDF: Learning to Forecast in the Frequency Domain},
    author = {Wang, Hao and Pan, Licheng and Chen, Zhichao and Yang, Degui and Zhang, Sen and Yang, Yifei and Liu, Xinggao and Li, Haoxuan and Tao, Dacheng},
    booktitle = {ICLR},
    year = {2025},
}

Acknowledgement

This library is mainly constructed based on the following repos, following the training-evaluation pipelines and the implementation of baseline models:

• Time-Series-Library: https://github.com/thuml/Time-Series-Library.

All the experiment datasets are public, and we obtain them from the following links:

• Long-term Forecasting and Imputation: https://github.com/thuml/Autoformer.
• Short-term Forecasting: https://github.com/ServiceNow/N-BEATS.