Abstract: We study a reinforcement learning framework for reducing systemic risk in financial networks under fairness and explainability constraints. The problem is motivated by lender-of-last-resort interventions, where institutions with similar attributes should be treated equally. 

Modeling the financial system as a network, we design policies using message-passing neural networks that enforce fairness by construction. We further derive convergence bounds that depend on the graphs characteristics and empirically analyze the performance trade-off induced by the regulatory fairness constraints on synthetic networks with Eisenberg–Noe-type contagion.