Prescriptive Analytics for Sustainable Financial Systems: A Causal–Machine Learning Framework for Credit Risk Management and Targeted Marketing

Financial institutions increasingly rely on data-driven decision systems; however, many operational models remain purely predictive, failing to account for confounding biases inherent in observational data. In credit settings characterized by selective treatment assignment, this limitation can lead to erroneous policy assessments and the accumulation of “methodological debt”. To address this issue, we propose an “Estimate → Predict & Evaluate” framework that integrates Double Machine Learning (DML) with practical MLOps strategies. The framework first employs DML to mitigate selection bias and estimate unbiased Conditional Average Treatment Effects (CATEs), which are then distilled into a lightweight Target Model for real-time decision-making. This architecture further supports Off-Policy Evaluation (OPE), creating a “Causal Sandbox” for simulating alternative policies without risky experimentation. We validated the framework using two real-world datasets: a low-confounding marketing dataset and a high-confounding credit risk dataset. While uplift-based segmentation successfully identified responsive customers in the marketing context, our DML-based approach proved indispensable in high-risk credit environments. It explicitly identified “Sleeping Dogs”—customers for whom intervention paradoxically increased delinquency risk—whereas conventional heuristic models failed to detect these adverse dynamics. The distilled model demonstrated superior stability and provided consistent inputs for OPE. These findings suggest that the proposed framework offers a systematic pathway for integrating causal inference into financial decision-making, supporting transparent, evidence-based, and sustainable policy design.