Segmentation
Also called customer segmentation, peer grouping
Segmentation is the grouping of customers or accounts with similar characteristics and expected behaviour so that transaction monitoring rules and thresholds can be set for each group. A retail current account and a corporate treasury account behave differently, so each segment gets scenarios and limits calibrated to its normal activity.
How it works
Analysts build segments from data such as customer type, product, business sector, turnover and geography. Statistical techniques like clustering help confirm that customers in each segment behave alike. Thresholds are then tuned per segment, often using percentiles of historical activity. Segments are reviewed at least annually and whenever products, customer bases or typologies change.
Good segmentation cuts false positives without hiding risk. A single threshold applied to all customers either floods analysts with alerts on large businesses or misses suspicious activity in small accounts. Because segmentation drives what the monitoring system detects, supervisors and validators treat it as part of the model. Firms document the logic, test it and show how it maps to their risk assessment.
The rules
Requires a risk-based transaction monitoring programme with documented, tested detection scenarios and thresholds, and annual board or senior officer certification.
New York-regulated banks, check cashers and money transmittersSets expectations for model development, validation and governance, superseding SR 11-7 and SR 21-8, the earlier statement on models supporting BSA/AML compliance.
Fed-supervised banking organisations, chiefly those above $30 billion in assetsCommon questions
Why does segmentation matter for transaction monitoring?
Customers behave differently, so one threshold cannot fit all. Segmentation lets firms set limits that reflect normal activity for each group, which reduces false positives and makes genuine anomalies easier to spot.
How often should segments be reviewed?
Many firms review segmentation at least annually, and also when they launch products, enter new markets or see shifts in customer behaviour. Reviews check that customers still sit in the right group and that thresholds remain calibrated.
Is segmentation treated as a model?
Often yes. Because segmentation shapes what a monitoring system detects, validators and supervisors commonly include it in model risk management, expecting documentation, testing and independent review of the logic and data behind it.
