CauKer: Classification Time Series Foundation Models Can Be Pretrained on Synthetic Data Only
March 9, 2026 · View on GitHub
🚨 NEW: 2 million CauKer dataset is now available on Hugging Face! 🚨
CauKer is a new synthetic data generation framework for pretraining classification Time Series Foundation Models (TSFMs) without using any real data.
🚀 Oral Presentation and Paper @ ICLR 2026 📅 April 23-27, 2026 | 📍 Rio de Janeiro
🚀 Best Time Series Paper and Spotlight Presentation @ ICML 2025 FMSD Workshop
📅 July 18, 2025 | 📍 Vancouver
Overview
Recent work in TSFMs has focused on training with large-scale real-world corpora, which are expensive and hard to collect. CauKer tackles this challenge by generating synthetic time series data that are both:
- Temporally realistic (trends, seasonality, non-stationarity)
- Causally structured (DAG-based propagation of Gaussian Process root nodes)
Our method combines:
- Gaussian Process Kernel Composition
- Structural Causal Graphs
- Rich activation and mean function libraries
With CauKer, we show that state-of-the-art classification models like Mantis and MOMENT can be pretrained entirely on synthetic data and still outperform or match real-data baselines in zero-shot classification tasks.
A more efficient version and training code will be released soon.
Quick Start
You can try CauKer directly in your browser via Google Colab:
# 1. Clone the repo
git clone https://github.com/ShifengXIE/CauKer.git
cd CauKer
# 2. Install dependencies
# 3. 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
Example Use Case
Once generated, the synthetic .arrow dataset can be used to pretrain your own TSFM (e.g., Mantis or MOMENT) and then evaluated in a zero-shot setting on UCR, UEA benchmarks.
Citation
@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}
}
❤️ Acknowledgements
This work is the result of a great collaboration—thanks to all my amazing co-authors: Vasilii Feofanov, Marius Alonso, Ambroise Odonnat, Jianfeng Zhang and Ievgen Redko, for their guidance and support throughout this project. Thanks MOAKHER Yessin for propose some ideas to be faster and safe.
If you have questions, collaboration ideas, or just want to discuss TSFMs, feel free to reach out: