Counterfactual Explanations for AML Risk Scores Under Causal Constraints

Abstract

This study introduces a counterfactual-explanation framework for anti-money-laundering (AML) risk scoring. The method distinguishes actionable from non-actionable variables within a causal graph and generates minimal valid changes needed to shift a case below a risk threshold. Experiments used 42.7 million transactions and 3.1 million customer profiles from a nationwide financial dataset. The system generated valid counterfactuals for 91.5% of high-risk cases, requiring an average of 2.4 feature changes per case. In a user evaluation with 26 compliance analysts, counterfactual explanations reduced review time by 22.6% and increased decision consistency from 0.62 to 0.75 (Cohen’s kappa). Causal constraints eliminated unrealistic recommendations in 98.6% of generated explanations. The method enhances model transparency while maintaining detection performance.