Detection performance
Understand alert quality, review outcomes and algorithm test results.
Loading performance data…
Operational alert quality
Each finding contributes its latest assessment once. Invalid findings count as reviewer-confirmed false positives. These results depend on which alerts analysts reviewed.
Compare individual detectors and versions
Algorithm accuracy and error rates
Accuracy and recall require labelled cases, including legitimate activity and suspicious patterns that were missed. The benchmark below tests detector logic in isolation. It does not describe live detector deployments.
The case mix deliberately includes missed suspicious patterns and legitimate lookalikes. The current methods have identical rates because those proportions are designed into the corpus. Use representative labelled data to compare real detection effectiveness.
How to interpret these metrics
- Precision
- True positives ÷ (true positives + false positives)
- Accuracy
- (true positives + true negatives) ÷ all cases
- Recall
- True positives ÷ all labelled suspicious cases
- False-positive rate
- False positives ÷ all labelled legitimate cases
The false-positive share among reviewed alerts is a different measure: invalid ÷ reviewed. A balanced synthetic test corpus has a very different base rate from real bank traffic. High accuracy here does not establish production accuracy.