The expensive surprises
A fraud spike, a broken sensor, a pipeline that silently stopped—anomalies are cheap to catch early and costly to discover late.
Ways to spot them
- Statistical thresholds for simple, well-behaved metrics.
- Seasonal models that know Monday isn't Sunday.
- Unsupervised ML for high-dimensional data.
Tune for trust
Too many false alarms and people stop looking. We tune sensitivity to the cost of a miss versus the cost of noise, and route alerts to someone who can act.
