Model performance monitoring is a well-established component of model risk management in financial institutions; however, a critical gap remains in translating monitoring outputs into consistent and defensible risk decisions. While existing practices generate signals such as performance degradation, data drift, and override activity, responses are often fragmented, judgment-driven, and inconsistently applied. Current approaches also remain largely model-centric and do not adequately address aggregate model risk, where multiple models respond simultaneously to shared macroeconomic conditions.
This paper proposes a structured decision framework for converting model performance signals into consistent governance actions at both model and aggregate levels. At the model level, the framework defines a rule-based decision layer comprising predefined thresholds, escalation logic, and action-mapping rules that translate observed signal patterns into actions such as continued monitoring, targeted validation, recalibration, or model redevelopment, reducing reliance on subjective judgment. Machine learning techniques, including supervised classification, unsupervised anomaly detection, and explainable attribution methods, are applied to support signal interpretation and root-cause identification, including changes arising from data distribution shifts, population variations, or macroeconomic transitions.
At the aggregate level, the framework introduces a multi-model signal aggregation approach that integrates internal performance indicators with macroeconomic and event-driven context to identify systemic risk drivers affecting multiple models. Given the complexity of cross-model dependencies, machine learning techniques support detection of correlated deterioration patterns and interpretation of aggregate model risk, allowing institutions to distinguish isolated issues from broader regime shifts across key financial risk domains, including credit risk, IFRS 9 impairment, stress testing, market risk, and broader balance sheet and regulatory capital models.
The key contribution is the formalization of a decision layer within model risk governance that bridges signal detection and governance response. The framework incorporates explainable decision pathways, human-in-the-loop oversight, and an audit structure to ensure traceability and regulatory defensibility under SR 26-2 and PRA SS1/23.
