Multimodal deep learning framework for shadowbanking risk prediction - dynamic decisionoptimization integrating knowledge graph andreinforcement learning

Amid increasing digitization and globalization of financial systems, the detection and mitigation of systemic risk withinnon-traditional financial sectors has emerged as a critical research imperative in computer science. Traditional statistical andeconometric models for risk assessment often suffer from static assumptions, limited capacity to model interdependencies, andlack of regulatory interpretability—shortcomings that hinder real-time and scalable solutions in complex financial ecosystems.To overcome these limitations, we propose a multimodal deep learning framework that integrates a graph-theoretic neuralarchitecture, GFA-Net, with a policy-aware strategic module, PCS-Flow. GFA-Net encodes financial systems as dynamictransaction graphs enriched with semantic and regulatory features, enabling robust structural learning and forward simulationacross accounting periods. PCS-Flow further ensures that model outputs remain consistent under heterogeneous policyregimes and evolving fiscal scenarios incorporating differentiable scenario perturbations and compliance regularizers.Through these synergistic components, our approach delivers a unified solution for forecasting, anomaly detection, anddecision optimization in high-dimensional financial environments. Experimental results on simulated and real-world datasetsdemonstrate superior accuracy, compliance fidelity, and temporal stability, thus validating the utility of our method forpolicy-consistent risk prediction. This work contributes to the field by advancing interpretable, regulation-aware machinelearning frameworks capable of navigating the evolving landscape of financial technologies.