(ICLR 2025) TimeKAN: KAN-based Frequency Decomposition Learning Architecture for Long-term Time Series Forecasting🚀

This is an offical implementation of "TimeKAN: KAN-based Frequency Decomposition Learning Architecture for Long-term Time Series Forecasting"

Overall Architecture

Results

Getting Started

1. Install requirements. pip install -r requirements.txt

2. Download data. You can download all the datasets from Autoformer.

3. Training. All the scripts are in the directory .scripts. If you want to obtain the results of input-96-predict-96 on the Weather dataset, you can run the following command:

sh scripts/long_term_forecast/Weather/weather_96.sh

Acknowledgement

We sincerely appreciate the following github repo very much for the valuable code base and datasets:

https://github.com/cure-lab/LTSF-Linear

https://github.com/kwuking/TimeMixer

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

https://github.com/ts-kim/RevIN

https://github.com/SynodicMonth/ChebyKAN

Citation

If you find this repository useful for your work, please consider citing it as follows:

@inproceedings{
  huang2025timekan,
  title={Time{KAN}: {KAN}-based Frequency Decomposition Learning Architecture for Long-term Time Series Forecasting},
  author={Songtao Huang and Zhen Zhao and Can Li and LEI BAI},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025},
  url={https://openreview.net/forum?id=wTLc79YNbh}
}