The Fraud Score
2026-08-14Almost 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:
- Identity age and stability -- how long has this email, phone number, device, or card been seen before, and has it been consistent. A card first seen ten minutes ago, paired with an email created yesterday, is a very different prior than a card with two years of clean history.
- Velocity -- how many transactions, on how many cards, from this device or IP or shipping address, in how short a window. Velocity across seemingly unrelated accounts is one of the strongest signals there is, because it's expensive for a fraudster to fully compartmentalize every session.
- Network reputation -- has this IP, device fingerprint, or card BIN shown up attached to confirmed fraud anywhere else in the processor's cross-customer data. This is why a single processor like Stripe or a card network can catch patterns an individual merchant never could: they see fraud reported across thousands of unrelated businesses.
- Behavioral and environmental noise -- mismatched timezone vs. billing address, browser languages that don't match geography, autofill patterns that look scripted rather than typed, checkout speed that's inhumanly fast.
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.