False Positives Are a Cost Centre: The Arithmetic of Noise
Every false positive burns an analyst hour, delays a real review, and trains your team to distrust the queue. Here's the full-cost model — and where it breaks.
Most compliance budget conversations start with the cost of the screening tool. Ours should start with the cost of the noise the screening tool produces — because that number is usually ten times bigger, and it is entirely optional.
The line-item model
Take a mid-sized exchange doing 250,000 customer and transaction screens a year with a naive matcher. A false-positive rate of 1% sounds respectable. It is not. Here is the ledger:
| Line item | Estimate |
|---|---|
| Screens per year | 250,000 |
| False positives at 1% | 2,500 alerts/year — 208/month |
| Analyst minutes per false positive (open, read, check, close) | 12–25 minutes |
| Annual analyst time lost | ≈ 1,000 hours — half a full-time analyst |
| Fully loaded cost | ≈ $35k–$60k/year, before anything else |
That is the visible cost. The invisible ones are worse:
- Latency on real risk. Every false positive sits ahead of a true hit in the queue. Structuring patterns wait while "Falcon Trading LLC vs FZE" gets its fifteen minutes.
- Threshold inflation. Teams drowning in noise raise thresholds to survive. Real hits below the new threshold now never see a human. This is how the Salem Al Farsi cases of the world go four years unnoticed.
- Analyst distrust. After the 200th identical false positive, analysts stop opening alerts and start scanning them. The queue becomes a checkbox. That is not a control; it is a ceremony.
- Examiner optics. A queue with a 95% false-positive rate reads as an untuned control. Examiners notice — and it colours every other finding in the report.
Where the noise comes from
In our analysis of Gulf-region screening data, three causes account for most false positives:
- Corporate-suffix collisions — "Falcon Trading LLC" vs "Falcon Trading FZE" match on everything but the suffix, which carries no identity. The fix is suffix-aware normalisation, not a lower threshold.
- Generic-vocabulary drift — "Delta Marine Supplies" and "Meridian Shell Holdings" score respectably on fuzzy metrics while sharing not a single distinctive token. The fix is a distinctiveness guard that suppresses vocabulary artefacts.
- Non-party screening — cash tills, payroll files, card settlements and own-account transfers are booking metadata. Screening them is noise by construction; they should be filtered at ingestion.
The honest caveat
No system removes false positives entirely, and any vendor who promises zero is selling you a threshold so high it will also miss the real thing. The goal is different: make every alert earn its place in the queue, carry its evidence visibly, and let the analyst clear it in seconds with a trace they can defend. A smaller queue, where every row matters, is faster, cheaper and — when the examiner asks — easier to stand behind.