TimePFN

This is an official implementation of TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data (AAAI 2025).

This repository contains the codebase of the TimePFN. We recommend using a conda virtual environment to load the dependencies listed in requirements.txt.

We provide the model checkpoint, testing, training, and fine-tuning scripts. Please check pfn_scripts. For the datasets, please refer to iTransformer's datasets.zip gdrive link.

Download them and put them under the directory ./datasets.

To generate synthetic datasets for the pretraining task, please refer to the directory synthetic_data_generation. Please read the comments and directives in the bash scripts.

Citation

@article{taga2025timepfn,
  title={TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data},
  author={Taga, Ege Onur and Ildiz, M Emrullah and Oymak, Samet},
  journal={arXiv preprint arXiv:2502.16294},
  year={2025}
}

Acknowledgement

We thank to the following repositories for their valuable code contributions, which helped immensely:

• iTransformer (https://github.com/thuml/iTransformer)
• PatchTST (https://github.com/yuqinie98/PatchTST)
• Reformer (https://github.com/lucidrains/reformer-pytorch)
• Informer (https://github.com/zhouhaoyi/Informer2020)
• Autoformer (https://github.com/thuml/Autoformer)