Research on BERT-LSTM-based multimodal financial risk early warning with attention fusion

Addressing the challenge that traditional financial risk assessment models struggle to effectively integrate structured sequential financial data with unstructured text, leading to insufficient capture of early warning signals for corporate financial distress, this study proposes a multimodal fusion method based on an attention mechanism. This method utilizes the Bidirectional Encoder Representations from Transformers (BERT) network to extract deep semantic features from the Management Discussion and Analysis (MD&A) text. It is combined with a Long Short-Term Memory (LSTM) network to capture the temporal dynamics of financial indicators. Finally, an attention module adaptively weights these two heterogeneous features to achieve precise risk discrimination. Experimental validation on a dataset of over 28,500 samples from the Chinese A-share market shows that the proposed method achieves a state-of-the-art F1-score and AUC (Area Under the ROC Curve) of 0.882 and 0.957, respectively. This conclusion demonstrates that, compared to a unimodal model relying solely on textual information (F1-score of 0.776) or financial data (F1-score of 0.743), the attention-based fusion method can significantly enhance warning performance, providing an effective new paradigm for risk assessment in complex financial environments.