🚩 Updates (2025-02-27) Initial version available on arXiv FinTSB.
🚩 Updates (2025-02-24) First version of the FinTSB code released.
FinTSB is a comprehensive and practical financial time series benchmark. Our goal is to address the three systemic limitations in the evaluation of the Financial Time Series Forecasting field.
➊ Failure to account for the full spectrum of stock movement patterns observed in dynamic financial markets (Diversity Gap).
➋ The absence of unified assessment protocols undermines the validity of cross-study performance comparisons (Standardization Deficit).
➌ Neglect of critical market structure factors, resulting in inflated performance metrics that lack practical applicability (Real-World Mismatch).
Ensure you are using Python 3.10.16 and install the necessary dependencies by running:
pip install -r requirements.txt
Begin by downloading the required datasets. All datasets are conveniently available at https://quantapi.eastmoney.com/. After downloading the data, you should preprocess it as outlined in our paper (We will open source it as soon as the camera-ready version of the paper is published). Finally, create a separate folder named ./data and neatly organize all the CSV files as shown below:
FinTSB
└── extreme
└── dataset_1.pkl
└── dataset_2.pkl
└── dataset_3.pkl
└── dataset_4.pkl
└── dataset_5.pkl
└── fall
└── dataset_1.pkl
└── dataset_2.pkl
└── dataset_3.pkl
└── dataset_4.pkl
└── dataset_5.pkl
└── fluctuation
└── dataset_1.pkl
└── dataset_2.pkl
└── dataset_3.pkl
└── dataset_4.pkl
└── dataset_5.pkl
└── rise
└── dataset_1.pkl
└── dataset_2.pkl
└── dataset_3.pkl
└── dataset_4.pkl
└── dataset_5.pkl
The pipeline of FinTSB includes the Data layer, Training layer, Backtesting layer, and Feedback layer, as shown in the figure below.
All configs are located in ./configs. For instance, to train the LSTM model using the FinTSB dataset, simply run:
python train.py --config_file configs/config_lstm.yamlAfter training:
- Your trained model will be safely stored in
./output/model.bin. - Numerical results in .csv format and .pkl format can be found in
./output/backtest_result.csvand./output/pred.pkl. - A comprehensive summary of quantitative metrics is accessible in
./output/backtest_report.txt. - A visualization is available in
./output/xxx.jpg.
If you find this repo useful, please consider citing our paper as follows:
@article{hu2025fintsb,
title={FinTSB: A Comprehensive and Practical Benchmark for Financial Time Series Forecasting},
author={Yifan Hu and Yuante Li and Peiyuan Liu and Yuxia Zhu and Naiqi Li and Tao Dai and Shu-tao Xia and Dawei Cheng and Changjun Jiang},
journal={arXiv preprint arXiv:2502.18834},
year={2025},
}
@article{hu2025finmamba,
title={FinMamba: Market-Aware Graph Enhanced Multi-Level Mamba for Stock Movement Prediction},
author={Hu, Yifan and Liu, Peiyuan and Li, Yuante and Cheng, Dawei and Li, Naiqi and Dai, Tao and Bao, Jigang and Xia Shu-Tao},
journal={arXiv preprint arXiv:2502.06707},
year={2025}
}
@article{hu2025timefilter,
title={TimeFilter: Patch-Specific Spatial-Temporal Graph Filtration for Time Series Forecasting},
author={Hu, Yifan and Zhang, Guibin and Liu, Peiyuan and Lan, Disen and Li, Naiqi and Cheng, Dawei and Dai, Tao and Xia, Shu-Tao and Pan, Shirui},
journal={arXiv preprint arXiv:2501.13041},
year={2025}
}
@inproceedings{hu2025adaptive,
title={Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting},
author={Hu, Yifan and Liu, Peiyuan and Zhu, Peng and Cheng, Dawei and Dai, Tao},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
year={2025}
}
@inproceedings{bian2024multi,
title={Multi-patch prediction: adapting language models for time series representation learning},
author={Bian, Yuxuan and Ju, Xuan and Li, Jiangtong and Xu, Zhijian and Cheng, Dawei and Xu, Qiang},
booktitle={Forty-first International Conference on Machine Learning},
year={2024}
}Special thanks to the following repositories for their invaluable code and datasets:
If you have any questions, please contact huyf0122@gmail.com or submit an issue.
