
Updated October 2026. This page began in 2020 as a two-part survey of machine learning and the companies that sell fraud scoring. The vendor list aged badly and is gone. What remains is the question worth answering: how does the model watching your card decide, and why does it sometimes get you wrong?
The short answer: every card payment is scored in well under a second against your own history, the merchant, the device and the pace of recent attempts. That stops most stolen-card fraud. It cannot stop a payment you were talked into making yourself, which is where losses now grow. Our guide to what your bank does to protect your data covers the other layers.
How big the problem is
| Measure | Figure | Source |
|---|---|---|
| Card fraud losses worldwide, 2024 | $33.41 billion, down 1.2% on 2023 | Nilson Report |
| United States share | 26% of card spending, 42% of the losses | Nilson Report |
| Fraud reported by US consumers, 2025 | $15.9 billion across 3 million reports, up about 25% | Federal Trade Commission |
| UK fraud losses, 2024 | £1.17 billion, of which £450 million was authorised push payment fraud | UK Finance |
The two trends point in opposite directions. Card fraud is flat because machine learning works on it. Reported scam losses rose by a quarter in a year because scams route around the model: the victim authorises the payment, so the transaction looks legitimate.
What the model checks in 50 milliseconds
- Your own baseline. Amount, time of day and merchant type are compared with months of your history. A small coffee at 8 am scores low; a large electronics order at 3 am scores high.
- Place and device. Whether the card was present, which phone or browser was used, and how far it is from the last purchase.
- Velocity. How many attempts, across how many cards, from this device or address in the last few minutes. Stolen card numbers are tested fast.
- Relationships. Whether the merchant, device or delivery address has appeared in confirmed fraud before. This is where the newer models earn their keep.
Mastercard says its Decision Intelligence Pro weighs relationships across about a trillion data points and returns a score in under 50 milliseconds, raising detection by 20% on average in early tests. Visa says its systems blocked 144% more suspected fraud over the 2025 Black Friday weekend than a year earlier. The output is a risk score. Your bank chooses the line above which it declines.
Why an honest purchase gets declined
That line is the whole trade-off. Set it low and fraud slips through; set it high and good customers are refused. A first purchase abroad, a new laptop at an unfamiliar shop, three quick online orders in a row: each looks exactly like the first minutes of a stolen card. The model cannot know you are on holiday unless something told it.
Three things cut false declines. Keep the bank's app installed with notifications on, so a decline arrives with a one-tap "this was me" and the retry goes through. Pay with the phone wallet where you can; a tokenised, device-bound payment scores lower risk than a typed card number. And let the app use your location if it asks, because a card used where the phone is rarely looks stolen.
The fraud the model cannot see
When you send money yourself to a "safe account", an investment platform or a seller who never ships, the bank's model sees a customer making a normal transfer. In the UK, investment scams took £144 million in 2024, up 34%, while remote purchase card fraud rose 11% to nearly £400 million. Since October 2024, UK banks must reimburse most authorised push payment victims up to £85,000, and 86% of losses were returned in the first three months.
That rule pushed banks to add signals the card model never had: confirmation that the payee name matches the account, warnings that pause a first payment to a new recipient, and behavioural checks that notice hesitant typing while someone is being coached on a call. None of it works if the code is read out to the caller, which is why our passkey guide matters more than any model.
Frequently asked questions
Does the fraud model see what I bought?
It sees the merchant, its category and the amount, not the basket. A supermarket purchase and a supermarket gift-card purchase look the same, which is one reason gift-card scams succeed.
Do I still need to tell the bank I am travelling?
With most app-based banks, no. The phone's location and wallet payments carry that information. With an older card issuer, a travel notice still prevents the first-day decline.
Can I find out why I was declined?
The merchant only sees "declined". Your bank sees the reason code and will usually say whether it was the fraud score, a limit or a technical fault. Ask; a wrongly tuned model learns from your answer.
Conversation
Have a question or correction? Contact Rvasus.