Testing monitoring rules on 370 real laundering attempts
Detection scenarios look for shapes: one account paying many, many paying one, money passing straight through. This tool runs those rules against 370 labelled laundering attempts from IBM's public anti-money laundering dataset and shows which typologies they catch.
- Problem
- Suspected mule and rapid pass-through activity spread across linked accounts, inside a queue where most alerts were noise.
- What I did
- Traced money movement across linked accounts, wrote STRs and investigative narratives setting out typologies, red flags and escalation rationale, and recommended restrictions, freezes or exits where findings exceeded risk appetite.
- Test the rules
- Turn rules on and off and move the thresholds. Watch the bipartite pattern: money moves between two groups of accounts, no single account stands out, and single-account rules almost never catch it. That gap is why network analytics and investigator judgment still matter.
Monitoring rule tester
IBM AML benchmark · 370 labelled attemptsData: HI-Small laundering patterns from IBM's synthetic AML dataset (Kaggle), 3,209 transactions dated 1–18 September 2022, extracted from a printed copy with character recognition. The file contains only laundering, so this measures what each rule catches, not how many false alerts it would raise on normal customers.