A deep learning-based sentiment flow analysis model for predicting financial risk of listed companies

With the advancement of natural language processing technology, more studies are applying deep learning models to extract features from unstructured data and predict corporate financial risk through horizontal comparisons. This paper proposes the Sentiment Flow Analysis (SFA) model designed to capture the nuanced emotional dynamics present in corporate annual reports and analyst reports from a longitudinal perspective. By leveraging advanced contextual embeddings from a Transformer architecture and employing a sophisticated recurrent neural network, the model effectively processes sequential data, allowing for a comprehensive understanding of sentiment trends over a five-year period. The effectiveness of our model was validated through experiments predicting the removal of special treatment (ST) status for 344 listed companies under special treatment in 2022 and 2023, achieving a prediction accuracy of up to 82.27%. Compared to stateof-the-art models in the same field, our model demonstrates improvements in prediction accuracy and recall, showcasing the application of Artificial Intelligence (AI) in financial risk assessment.