Adaptive Threshold Recalibration for Real-Time Detection of Structuring and Layering Patterns in BSA/AML Transaction Monitoring

Abstract

Rule-based transaction-monitoring systems, still the dominant approach for detecting structuring and layering under the Bank Secrecy Act (BSA), rely on fixed thresholds that generate false-positive rates industry sources place at 90 to 95 percent, imposing enormous investigative cost while the U.S. Financial Crimes Enforcement Network (FinCEN) reports suspicious activity report (SAR) filings grew substantially in recent years, reaching roughly 2.5 million filings in fiscal year 2020 across all filer types. This article surveys published, cited evidence on adaptive threshold recalibration as an alternative to static rule-based thresholds for real-time detection of structuring and layering patterns, drawing on peer-reviewed and industry-benchmarked results rather than illustrative projections. Published studies report strong illicit-transaction classification performance when graph-based and ensemble machine-learning methods are combined with dynamic behavioural features on large, labelled AML transaction datasets, alongside published findings that novel graph-derived features can materially improve detection-model F1 scores over rule-based baselines, and vendor-reported reductions in the 50-to-80-percent range under production deployment conditions. The article reviews the architecture such systems require, the specific statistical and machine-learning techniques underlying published results, the regulatory context for structuring-related SAR filings, and the evidentiary limits of currently available research, concluding with a candid assessment of what remains unverified pending broader multi-institution validation.