Research on corporate financial risk prediction based on transformer models

초록

Corporate financial risk prediction is crucial for economic stability and sustainable business operations, yet existing methods struggle with complex interdependencies among financial indicators and temporal dynamics. This study addresses the limitations of conventional approaches in handling high-dimensional financial data, weak temporal pattern recognition, and distributional rigidity across sectors by proposing a novel deep learning framework integrating graph transformer networks (GTN), time-series contrastive learning (TSCL) and meta-learning-based adaptive optimisation (Meta-AO). The hybrid architecture combines graph neural networks with transformer attention mechanisms to model enterprise relationships, employs contrastive learning for robust temporal pattern extraction, and utilises meta-learning for cross-domain adaptation. The findings establish that the proposed framework effectively captures relational structures, temporal dynamics, and distributional shifts in financial data, offering a comprehensive solution for accurate risk prediction across diverse market conditions and enterprise types through its hierarchical feature learning paradigm.

키워드