Virtue & Voltage

A few products are getting rewired right now. Full voltage resumes shortly.

Field notes

When preventing one harm creates another, what actually decides where the line sits

A bank’s fraud model has a threshold, and every threshold is a trade. Tighten it and more fraud gets caught, but more legitimate customers get locked out of their own money at the worst possible moment. Loosen it and legitimate customers sail through, but more fraud gets through with them. There is no setting that avoids harm. There is only a setting that decides whose harm it will be…

The instinct is to solve this by optimising for the threshold that minimises total dollar loss across both kinds of error. This is not the whole answer, because it treats a false decline and a fraud loss as commensurable purely in cost terms, and they are not.

A blocked transaction is usually reversible within hours, a phone call, a retry.

A fraud loss that empties an account can cascade for weeks, a missed rent payment becomes an eviction filing is a much larger problem than the original dollar figure suggests. Matching dollar amounts does not mean matching severity, and a model tuned purely on expected loss has no way to see that difference.

Once you look past the loss-minimisation number, it is who is being asked to absorb the risk of error, and whether they can…

False declines are note evenly shared, they land hardest on customers whose spending patterns look statistically unusual to the model:

  • people who have just moved,
  • people making a first large purchase,
  • people who work cross-border.

(the same population the fairness question keeps returning to)

That is not a coincidence.

Whichever population is hardest for a model to read tends to absorb both kinds of error, wrongly flagged and wrongly waved through, because the model has the least reliable signal about them either way.

Which means the honest answer to “where should the line sit” – is about an judgement which population is least able to absorb which kind of mistake, made explicitly, revisited as the mix of customers and fraud patterns shifts, and owned by someone who can explain why the line sits where it does, not derived once from a loss-minimisation exercise and left untouched!

Where this connects

Who absorbs the error is exactly the question a materiality threshold is built to answer, rather than leaving it implicit in a single loss-minimisation number.

← Previous noteWhat Article 4 actually requires, and what the market is selling instead Next note →How we actually decide something is right, not just legal

© Virtue & Voltage. Quote a passage with a link back; please do not republish this note in full without asking.

← Back to Field notes