A meta-learning enhanced dynamic graph convolutional network for cross-region financial risk propagation prediction

The increasing interconnectedness of global financial systems has amplified the risk of cross-regional financial contagion, posing significant challenges to economic stability. Traditional models often struggle to capture the dynamic spatio-temporal dependencies in financial networks, particularly under heterogeneous and sparse data conditions. To address this, we propose a novel framework, Meta-Dynamic Graph Convolutional Network, which integrates meta-learning with dynamic graph convolutional networks for cross-regional financial risk propagation prediction—the first such integration to enhance adaptability in sparse and heterogeneous financial scenarios. Our approach employs dynamic graph convolutional networks to model the evolving financial network’s spatio-temporal dynamics, incorporating graph convolution, temporal attention mechanisms, and dynamic edge updates. Furthermore, meta-learning optimizes model initialization, enhancing generalization across regions with limited data. Experiments on public financial datasets and simulated networks demonstrate that our framework outperforms baselines, achieving a statistically significant (p < 0.05 via t-tests) 25 %–49 % reduction in mean absolute error and root mean square error, and a 20 %–34 % improvement in F1 score. It predicts both regression-based risk values, such as economic recession indices, and classification-based risk categories, such as high or low risk.