CauKer: Classification Time Series Foundation Models Can Be Pretrained on Synthetic Data Only

This repository contains synthetic time series data generated using the CauKer framework, as presented in the paper CauKer: Classification Time Series Foundation Models Can Be Pretrained on Synthetic Data.

CauKer is a synthetic data generation framework for pre-training classification Time Series Foundation Models (TSFMs) without relying on real-world data. It combines Gaussian Process (GP) kernel composition with Structural Causal Models (SCM) to produce diverse and causally coherent sequences with realistic trends and seasonality.

• Paper: CauKer: Classification Time Series Foundation Models Can Be Pretrained on Synthetic Data
• GitHub Repository: https://github.com/ShifengXIE/CauKer

Generation Script Usage

If you wish to generate your own synthetic data using the CauKer framework, you can use the following command from the official repository:

# Generate 200,000 synthetic time series (default: 512-length, 4-dimensional)
python CauKer.py -N 200000 -L 512 -F 4 -P 6 -M 18 -O CauKer200K.arrow

Citation

If you find this dataset or the CauKer framework useful, please cite:

@inproceedings{cauker2025,
  title={CauKer: Classification Time Series Foundation Models Can Be Pretrained on Synthetic Data Only},
  author={Shifeng Xie, Vasilii Feofanov, Marius Alonso, Ambroise Odonnat, Jianfeng Zhang, Ievgen Redko},
  booktitle={ICML Workshop on Foundation Models for Structured Data (FMSD)},
  year={2025}
}