AI Predicts Financial Firesales

A graph‑based deep learning model trained on security‑level holdings of nonbank intermediaries can substantially outperform traditional systemic risk metrics in forecasting trading behavior and asset returns during stress. Embedded into an optimal policy framework, these predictive gains translate into sharper, welfare‑improving macroprudential interventions. — If regulators adopt such models, supervision could become more forward‑looking and targeted, but it creates policy choices about data access, model transparency, and institutional reliance on opaque algorithms.