The Fraud Score

2026-08-14

Almost every online transaction gets a number attached to it before it's approved, and almost no consumer ever sees it. Processors call it a risk score, a trust score, a fraud score depending on the vendor -- Stripe Radar, Sift, card-network tools like Visa Advanced Authorization all produce some version of the same thing: a 0-to-100-ish estimate of how likely this specific transaction is to get charged back.

What actually feeds it

The score isn't one signal, it's a blend, and the blend is the whole value of the product:

Each signal alone is weak and full of false positives. A VPN user, a shared family card, someone shopping while traveling -- all trip individual flags constantly. The score's entire job is combining weak, noisy signals into something with better precision than any one of them.

How it gets used

Most systems reduce the score to three buckets: allow, block, or send to manual/step-up review (OTP, 3DS challenge, human look). Merchants tune the thresholds against their own risk tolerance -- tighter thresholds cut fraud losses but also decline legitimate customers, and that false-decline cost is often larger than the fraud it prevents, which is why thresholds are a business decision as much as a security one.

Where fraud tries to hide under the line

None of this is secret from the people trying to beat it, which is why the more sophisticated fraud rings optimize for staying under the threshold rather than avoiding detection outright: smaller transaction amounts, warmed-up accounts aged for weeks before use, one card per session instead of testing many, deliberately slow and human-paced checkout flows. This is the actual arms race -- not "detect fraud" vs. "hide from detection," but a threshold-tuning contest on both sides.

The practical takeaway for anyone building or buying this kind of system: a fraud score is a probability estimate, not a verdict, and treating it as binary (block everything under 40, allow everything above) throws away the information in why a transaction scored the way it did. The reason code behind the score is usually more useful for a human reviewer than the number itself.