Optimal Error Rate
The rational target for a harmful outcome is sometimes above zero: prevention should continue only while the damage avoided exceeds the value destroyed by preventing it.
A fraud system that achieves zero fraud may be worse than one that misses some fraud—because perfect blocking can also reject legitimate transactions.
E1The cost hidden inside prevention
Every additional safeguard produces two effects: it catches more harmful activity and obstructs more valuable activity. The last increment of prevention can therefore have negative expected-value even though it reduces the headline error rate. The optimum lies where the expected damage from another tolerated error is lower than the cost of eliminating it. That cost may be contained when the system has a bounded-downside, but perfect prevention can become its own failure mode—the overprotection-trap.
E1Not every error is safely tolerable
An average error rate can conceal a catastrophic class of mistakes. When one failure can spoil everything, or when a small number of cases causes most of the damage, treating all errors as interchangeable is reckless. The model applies only after distinguishing tolerable losses from ruinous ones.
Price the next unit of prevention
For one rule meant to stop a bad outcome, record both what it blocks correctly and what legitimate activity it rejects. Then evaluate the next tightening of that rule: keep it only if the additional harm prevented is worth more than the additional value lost.
E1Episodes that teach this
-
Cost Benefit Analysis Real Life Examples - Sensitivity vs Selectivity
1,168 views
'The ideal amount of fraud is never zero' because stopping all fraud would also stop valid transactions.