Label Unbalance in High Frequency Trading

Report: Label Unbalance in High-frequency Trading

高频交易中的标签不平衡问题研究 -- QuantML

Method

In this project, we mainly focus on the high frequency trading prediction in the scenario of label imbalance, based on machine learning methods. We provide four networks including: MLP, LSTM, BERT, and Mamba as the backbone. We use two simple methods to reduce the influence of label imbalance, including: resampling and class weighting.

Dataset

Due to copyright restrictions, we do not provide the original data. However, we provide the data structure and class distribution in our experiment. As a result, you can replace the dataset.py file with your own data and change some necessary parameters in train_classification.py to run the code on your dataset.

How to Run

To run the model, you can use the following command:

python train_classification.py --input_dim <input dim> --output_dim <class numbers> --model <please select from 'mlp', 'lstm', 'bert', and 'mamba'>

To change the model scale or other parameters related to training, please refer to the get_args function.

We provide the following methods for addressing label imbalance:

• --class_weight: You can specify the weight for each class in the loss function.

• --data_balance: We randomly delete 7/8 of the data in the '0' class to balance the dataset.

Reference

@misc{zhao2025labelunbalancehighfrequencytrading,
      title={Label Unbalance in High-frequency Trading}, 
      author={Zijian Zhao and Xuming Zhang and Jiayu Wen and Mingwen Liu and Xiaoteng Ma},
      year={2025},
      eprint={2503.09988},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2503.09988}, 
}