To address the challenges of multi-source heterogeneous data fusion and dynamic association modeling in regional financial risk monitoring, this study proposes a unified framework integrating spatiotemporal graph neural networks with multi-task learning. Using monthly state-level panel data covering all 50 U.S. states from 2012 to 2024 and a rolling split of 2012–2021 for training, 2022 for validation, and 2023–2024 for testing, the model achieves a one-step-ahead MAE of 0.118 and an AUC of 0.836, outperforming the compared statistical, tree-based, recurrent, and graph-based baselines within the same protocol. Ablation studies indicate that augmenting geographic adjacency with economically derived relation channels improves the median F1 score from 0.65 to 0.92, while multi-task learning yields an additional 13-percentage-point F1 gain in this experimental setting. Runtime experiments conducted in a unified implementation environment show 14.2 h of training time, approximately 35 ms inference latency, and stable performance up to a 30% simulated missing-data rate. Feature-attribution analysis is used here as an interpretive description of model behavior rather than causal proof, supporting the model's potential utility for regional financial monitoring and decision support.
Xuan Lan (Fri,) studied this question.