Model Risk Management in Finance: The Role of Agentic Systems in Risk Mitigation

Abstract

Model risk management (MRM) has become increasingly important in machine learning due to the reliance on complex models for decision-making in financial applications. This paper explores the integration of an agentic system capable of independent decision-making into machine learning model risk management (MRM) practices. More precisely, we integrated human-in-the-loop to the agentic crew, enabling them to work together efficiently to complete challenging MRM tasks. The paper outlines the architecture of the agentic crew, which consist of a judge agent along with role-specific agents to facilitate comprehensive model evaluations, ensuring adherence to regulatory standards and enhancing the conceptual soundness of the evaluated procedure. The application of agentic systems in MRM is illustrated through case studies in fraud detection and portfolio credit risk assessment. Our findings suggest that adopting human-centered agentic systems can lead to more informed decision-making processes, identifying whether the model development followed an established procedure, assessing reproducibility, bias and fairness, and monitoring changes in behavior when unknown inputs are introduced.