An end‑to‑end multi‑source heterogeneous data fusion and spatiotemporal graph convolutional network for enterprise financial crisis early warning

In the early warning of enterprise financial crisis, the difficulty of multi-source heterogeneous data fusion and the complexity of risk transmission effect modeling are the key challenges. To this end, this study proposes an end-to-end deep learning framework named MFGCN to achieve accurate early warning. This paper proposes an MFGCN model that fuses a multi-source heterogeneous data fusion module (using a GNN-based pattern alignment and attention weighting mechanism) and a spatiotemporal graph convolution early warning module (combining R-GCN and GRU to capture spatiotemporal dependencies). On a multi-source dataset covering 15,000 companies around the world, the model achieved an accuracy of 89.1%, an F1-score of 0.862, and an AUC value of 0.896, significantly outperforming the baseline models. Moreover, the model shows excellent robustness (performance degradation under noise is only 4.2%), long-term predictive capability (maintaining an AUC of 0.861 at a lead time of Δ=4 quarters) and domain generalization (no domain performance degradation < 4.52% is seen). The conclusion shows that the MFGCN framework provides an efficient and reliable solution for enterprise financial crisis early warning through end-to-end fusion, multi-relationship risk transmission modeling, and a dynamic adaptive mechanism. Its core contribution is to realize end-to-end learning from raw multi-source data to accurate early warning. Although the computational complexity is high, it points out the direction of optimizing efficiency and introducing federated learning for subsequent research.