Financial distress prediction based on deep learning and multisource data fusion

Abstract

With enterprises facing increasingly severe financial risks, the prediction of financial distress has become a critical research direction in academia and industry. Traditional methods primarily rely on structured financial data, but they face limitations when processing complex market dynamics and unstructured data. In recent years, the rapid development of deep learning technology has provided new insights for financial distress prediction. This paper proposes a novel financial distress prediction method based on the Long Short-Term Memory (LSTM) network in deep learning, integrating multi-source data such as stock prices, financial data, and annual report texts. By combining time series of stock prices, financial indicators, and annual report text information, this method overcomes the limitations of traditional approaches that rely solely on structured data. The LSTM model is used for prediction, and a multi-layer LSTM model is designed with an attention mechanism to further enhance the model’s predictive performance. Experimental results show that inputting multi-type feature data significantly improves the accuracy of financial distress prediction, especially when the annual report text data is included, which leads to a notable improvement in both prediction accuracy and recall rate.