94%Real electricity data from an open research dataset, and synthetic bank transfers
Measuring what gets caught
of planted 30% drops caught in 318 real electricity histories, at 2.6% false alerts
- What was observed
- Drops and bursts of known size planted into histories, so what should be caught is known.
- What a generic approach says
- Report how many alerts fired, not how many problems were caught.
- What the engine read
- 94% of 30% drops were caught at 2.6% false alerts, and 51% of 10% drops. On synthetic bank transfers, 88% of $5,000 bursts were caught at 2.4% false alerts, but gradual falls on dormant accounts were not.
- What it meant for the decision
- Every deployment can state its detection floor: the size of change it catches and the size it misses.