Reproduce “HRFT: Mining High-Frequency Risk Factor Collections End-to-End via Transformer” on OpenResearch

About this paper

Wenyan Xu, Rundong Wang, Chen Li, Yonghong Hu, Zhonghua Lu

The Intraday Risk Factor Transformer (IRFT) presents an end-to-end Transformer-based methodology for directly generating interpretable, formulaic high-frequency risk factors from stock data. This approach significantly improves inference speed by up to 17 times and achieves higher predictive accuracy (R^2 up to 0.92) and investment returns in simulated trading on real-world datasets.

Key insights

• Framing symbolic mathematics as a language modeling problem enables advanced deep learning architectures, like Transformers, to directly generate interpretable mathematical formulas for risk factors.
• The combination of a specialized data generation process with an asymmetric Transformer architecture and post-prediction constant refinement significantly enhances both the accuracy and interpretability of generated financial models.
• Effective scaling and bagging strategies are crucial for applying symbolic regression methods to the high volume and complexity of real-world high-frequency trading datasets, overcoming traditional limitations.

Summary by alphaXiv.