Measurement
Precision and recall
Definition
Precision is how often the system is right when it flags something; recall is how much of what it should have flagged it actually caught. Improving one usually costs the other.
Nearly every operational tuning argument is a precision/recall argument in disguise, even when nobody in the room uses those words. A fraud model with high recall floods investigators with false alarms; one with high precision misses losses that a looser threshold would have caught. The right balance is a business decision about the relative cost of a miss and a false alarm, and it should be set by the people who bear those costs, not by whoever tuned the model.
The threshold is a policy choice
Where you set the cutoff decides how much work lands in the exception queue and how much risk passes through untouched, and that trade-off is fundamentally a management decision about risk appetite, not a data science preference about which number looks best on a validation chart. Leaving the threshold to whoever built the model, without an explicit conversation with the business owner who bears the consequences either way, is a common and easily avoidable governance gap.
Price both error types
Put a number on a false positive — reviewer minutes spent chasing something that turned out fine — and on a false negative — expected loss from something that should have been caught and was not. The optimal threshold then computes itself once both costs are stated explicitly, rather than being debated qualitatively in a room where everyone has a different intuitive sense of which error type feels worse to them personally.
The right threshold moves over time
The relative cost of a false positive and a false negative is not fixed; it changes as staffing levels change, as fraud patterns evolve, or as regulatory scrutiny on a particular decision type increases. A threshold set correctly two years ago based on the cost structure at the time may no longer be optimal today, which is why the threshold deserves periodic revisiting rather than being set once during the original model build and left untouched indefinitely.
In practice
- Agree the cost of each error type before tuning anything
- Report the operating point, not just an accuracy headline
- Revisit the threshold when volumes or staffing change
- Have the business owner, not just the model builder, approve the final threshold
