An applied study of multimodal financial data for financial early warning of listed companies

This paper focuses on the performance risk identification of listed companies, aiming to solve the challenge of secondary market investors to accurately judge investment risks and make reasonable decisions through financial report information. Most of the existing financial risk research methods are based on public financial statement data, but with the development of technology, it has become a new direction to mine the textual information of financial reports and fuse it with structured data for risk prediction. From the perspective of text feature extraction and feature fusion, the main contribution of this paper is to introduce natural language technology into the field of risk identification of listed companies, which proves that unstructured data also contains corporate risk information, and the fusion of structured data and text data can improve the accuracy of financial risk judgement. The related work, including BERT model and MLP, is presented in the paper, and the experimental results show that the bimodal (statement and text) fusion outperforms the data unimodal and the data modal outperforms the textual modal on the test dataset, which verifies the validity of the modal fusion in the identification of listed companies' risks.