Published: Jul 2026
A model score becomes useful only after selecting a decision threshold. In fraud analytics, threshold selection should minimize expected business loss.
Let the threshold be \(t\). We can write expected cost as:
$$ C(t) = c_{FP} \cdot FP(t) + c_{FN} \cdot FN(t) $$The best threshold is:
$$ t^* = \arg\min_t C(t) $$When missed fraud is expensive, \(c_{FN} \gg c_{FP}\), so the operating point shifts toward higher recall.
If analysts can review at most \(K\) alerts/day, then choose the highest-value threshold satisfying:
$$ \text{Alerts}(t) \le K $$This links model tuning directly to team capacity and business value.