DFCNet: A Dynamic Early Warning Model for Financial Crises of Listed Companies Based on Deep Learning

In the current complex and ever-changing market environment, financial crises faced by listed companies have become a common phenomenon. Such crises not only threaten the survival and development of enterprises but can also have a severe impact on investors, creditors, and the stability of the entire financial market. Traditional financial crisis early warning models mainly rely on statistical analysis methods, such as Logistic regression and discriminant analysis, but they have clear limitations when dealing with high-dimensional data, nonlinear relationships, and dynamic changes. With the development of information technology, especially the advent of the big data era, the massive data generated by both internal and external environments of enterprises provides the possibility to build more precise and effective warning models. This paper proposes a deep learning-based dynamic financial crisis early warning model, DFCNet. The model leverages techniques such as multi-level feature representation, bidirectional feature extraction (combining bidirectional LSTM and GRU), and mask optimization, which allow it to better mine hidden patterns in financial data and improve the accuracy and robustness of the warning model. The experimental results demonstrate that, in five-class and ten-class tasks conducted on two different public datasets, DFCNet outperforms advanced models such as TimesNet, Transformer, LSTM, GRU, SVM, and Conformer, achieving accuracies of 95.5% and 93.4%, respectively. This fully showcases its superior performance and strong potential in various application scenarios. This lays the foundation for providing enterprises with more scientific and accurate financial risk warning services and contributes to maintaining the healthy development of enterprises and the stability of financial markets.