Graph convolutional network with dynamic gating LSTM improves corporate credit default prediction

Traditional corporate credit evaluation models often treat enterprises as isolated entities and rely on static financial data, making it difficult to capture dynamic risk evolution and inter-enterprise risk transmission. This study proposes a Graph-fused Multi-Modal Dynamic Gating Network (GMDG-Net) that integrates three information modalities: internal financial and market data, external online public opinion text, and an enterprise association graph constructed from equity pledge relationships. A graph convolutional network learns spatial risk transmission patterns from the association graph. A word vector model extracts sentiment features from public opinion text. A structural adaptive gating long short-term memory network dynamically fuses these multi-modal temporal features and adjusts contribution weights of different information sources in real time. Experiments using a corporate stock pledge default prediction dataset (2017–2022, 12,384 samples) show that GMDG-Net achieves an accuracy of 0.913 and F1-score of 0.851, significantly outperforming baseline models including support vector machine (0.842), standard LSTM (0.889), and graph convolutional network alone (0.872). Multi-modal feature fusion experiments confirm that adding public opinion features increases accuracy by 2.2%, and adding graph structural features further increases accuracy by 1.8%. The model maintains robust performance under imbalanced sample conditions (weighted F1-score 0.897). This framework provides financial institutions with a more accurate and forward-looking credit risk assessment tool.