This is an official implementation of TimePFN: Effective Multivariate Time Series Forecast 6320 ing 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.
@inproceedings{taga2025timepfn,
title={TimePFN: Effective multivariate time series forecasting with synthetic data},
author={Taga, Ege Onur and Ildiz, Muhammed Emrullah and Oymak, Samet},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={39},
number={19},
pages={20761--20769},
year={2025}
}
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)