Enhancing Stock Prediction with Sentimental and Relational Information Distilled from Large Models

Zhipeng Liu, Mingyang Geng, Ying Li, Peibo Duan, Xiaosha Xue, Bin zhang

Abstract: An increasing number of investors are dedicating their efforts to enhancing the accuracy of stock market predictions, a crucial factor in financial markets that aids in making well-informed decisions and optimizing their investment strategies. The development of deep learning methods has evolved from solely considering fundamental price information to a broader spectrum of information, such as sentiment features from news articles. However, these methods often struggle to effectively extract such features when multiple companies are mentioned in a single news article (more details can be referred to Section 1). With the advancement of large language models (LLMs), which can more precisely extract the knowledge people require from text. To overcome the shortcomings of previous work, we propose a novel distillation method for precise knowledge extraction, including sentimental and fine-grained relational knowledge. After that, we incorporate the knowledge as augmented features to assist in the predictions of stock forecasting models. Through experiments on the real Chinese stock market, CSI300 and CSI500, our proposed method can indeed lead to significant improvements in the performance of stock prediction models, demonstrating a performance improvement of baseline models up to 3.23% in terms of accuracy.