Validating transaction-monitoring models

Summary

Transaction monitoring sits at the intersection of financial crime and model risk. Whether the system is rule-based or uses machine learning, validation follows the same structural discipline: conceptual soundness, data, methodology, performance and governance.

Conceptual soundness

Each scenario should be traceable to a typology in the EWRA. Scenarios without a documented typology rationale are candidates for retirement.

Data

Completeness and lineage of transaction, customer and reference data are the foundation. A validation that does not test data has not validated the model.

Methodology and thresholds

Threshold-setting should be evidence-based: above-the-line and below-the-line testing, productivity analysis, and disposition quality. Round-number thresholds without a rationale are a common finding.

Performance

Effectiveness is measured by alert quality and by coverage of known typologies. False-positive rates alone are misleading; a low false-positive rate can indicate over-tuning.

Tuning governance

Tuning changes should be documented, tested and approved through a defined governance path, with a clear audit trail.

Machine-learning approaches

ML-based monitoring introduces additional validation requirements: feature stability, drift monitoring, and explainability sufficient to defend alerts to investigators and to supervisors.

Limitations

Validation does not certify that no financial crime passes undetected. It provides assurance that the system is fit for its stated purpose and operates within known limits.

Related expertise

See AML and Financial Crime and Model Risk and Validation.