Year: 2025
Type: conference-paper
Abstract: In recent years, high returns on stock market investments have been accompanied by high risks. The financial condition of publicly listed companies can be reflected through stock data or financial statements, but the complexity of financial reports and the high variability of analysts' subjective judgments make it difficult to accurately identify financial crises in companies. Therefore, it is particularly important to adopt a more objective and systematic approach to assess the financial health of enterprises. This paper proposes a novel two-stage dynamic financial crisis early warning model that integrates domain transfer and adversarial training frameworks. In the first stage, a pre-trained Transformer encoder is used to fuse multi-source datasets for Q-guided, KV-guided, and consistency training. The second stage employs an adversarial training model, enhancing the model's robustness by adding noise to the input feature vectors. The experiments use diverse datasets, including financial statement data, macroeconomic indicators, stock market trading data, news reports and public opinion dynamics, changes in corporate credit ratings, and special events, totaling 6,171 time series data points. Comparative experiments show that, compared to advanced models such as GNN, SNN, CNN, RNN, TC-Net, and Transformer, our model achieves peak accuracies of 93.6% in the ten-class task and 95.3% in the five-class task, significantly improving prediction performance and demonstrating stronger robustness and stability. This study provides a new method for financial crisis early warning in companies and lays the foundation for further technological development.
Authors: Zhou Zhou, Yixuan Wu, Zejun Chen
Language: English
Subfield: Accounting
Domain: Social Sciences
Open Access status: gold
OpenAlex ID: W4412822329