Definition
Degree to which the fraud model's current authorization threshold diverges from the threshold that would be optimal given the current transaction distribution.
The Measurement Gap
The Measurement Gap
No metric tells you whether the fraud model's score thresholds are still calibrated to the current transaction distribution. The model itself may score accurately (FSCI may look healthy) but the threshold that converts those scores into approve/decline/review decisions was tuned weeks or months ago against a different distribution of transaction amounts, merchant categories, and fraud patterns. As the distribution shifts, the threshold that was optimal at tuning time now either declines too many legitimate transactions or approves too many fraudulent ones.
This gap is invisible to standard fraud metrics because the threshold compensates in predictable ways. If the threshold is too aggressive, fraud basis points looks excellent while false positives climb silently. If the threshold is too permissive, approval rate looks strong while fraud losses accumulate in the chargeback pipeline. In both cases, the individual metrics that operations teams monitor appear acceptable: the problem only surfaces in the net economics (FDNEY), which most organizations do not yet measure.
The staleness problem is structural. Transaction distributions shift continuously due to seasonal patterns, marketing campaigns, new merchant onboarding, and evolving fraud tactics. A threshold tuned quarterly drifts within weeks. A threshold tuned monthly drifts within days during high-change periods. No standard metric tracks the divergence between what is in production and what should be in production.
Decision Supported
Does the current production fraud threshold need re-optimization, and how much economic value is the stale threshold destroying?
Formula
Plain Language
Compare the fraud score threshold currently in production to the threshold that would maximize FDNEY on the trailing 30-day transaction distribution. Express the gap as a percentage of the production threshold.
Notation
ATSI = |Threshold_production − Threshold_optimal_current| / Threshold_production × 100
Where: Threshold_optimal = argmax_t [Σ FraudPrevented(t) − Σ RevenueLost(t)]
Variables
| Variable | Definition | Unit |
|---|---|---|
| Threshold_production | Fraud score threshold currently deployed in the production authorization system | Score value |
| Threshold_optimal_current | Threshold that maximizes FDNEY on the trailing 30-day transaction distribution | Score value |
| FraudPrevented(t) | Dollar value of fraud prevented at threshold t | $ |
| RevenueLost(t) | Dollar value of legitimate revenue lost at threshold t | $ |
Proposed formula, pending validation pilots.