Leveraging Diffusion Transformers for Stock Factor Augmentation in Financial Markets
Keywords: Financial Data Augmentation, Diffusion Models, Transformer
Abstract: Data scarcity poses a significant challenge in training machine learning models for stock forecasting, often leading to low signal-to-noise ratio (SNR) and data homogeneity that degrade model performance. To address these issues, we introduce DiffsFormer, a novel approach utilizing artificial intelligence-generated samples (AIGS) with a Transformer-based Diffusion Model. Initially trained on a large-scale source domain with conditional guidance to capture global joint distribution, DiffsFormer augments training by editing existing samples for specific downstream tasks, allowing control over the deviation of generated data from the target domain. We evaluate DiffsFormer on the CSI300 and CSI800 datasets using eight commonly used machine learning models, achieving relative improvements of 7.3\% and 22.1\% in annualized return ratio, respectively. Extensive experiments provide insights into DiffsFormer's functionality and its components, illustrating their role in mitigating data scarcity and enhancing model performance. Our findings demonstrate the potential of AIGS and DiffsFormer in addressing data limitations in stock forecasting, with the ability to generate realistic stock factors and control the editing process. These results validate our approach and contribute to a deeper understanding of its underlying mechanisms.
Primary Area: other topics in machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8851
Loading