Three dates put the score into the rulebook. Since 14 September 2019, an EU provider exempting a payment from strong customer authentication on risk grounds must combine its risk factors into "a risk scoring for each individual transaction" (Delegated Regulation (EU) 2018/389, Article 18). On 17 April 2026, the Federal Reserve, OCC and FDIC issued revised model risk guidance as SR 26-2.
On 29 October 2026, the Council adopted its text of the Payment Services Regulation, which extends transaction monitoring to the payee's provider; Parliament has not yet voted on it.
What is a fraud score?
A fraud score ranks one event, such as a payment, a login, a card application or an address change, by how likely it is fraud; producing it is fraud scoring, and some systems call it a fraud risk score. It is not a verdict on the customer.
Each vendor sets its own scale, 0 to 1, 0 to 100 or 0 to 1,000. Picture one payment that a gateway rates 0.62 and a bank rates 410: neither number means anything outside its own system.
How does a fraud score differ from a credit score or a customer risk score?
A credit score predicts whether a borrower repays over months or years. A customer risk score rates a customer for due diligence, the core of a customer risk assessment. A fraud score rates one event, now: a customer with strong credit and a low risk rating can still send a transfer that scores near the top once someone else controls the account.
What is an IP fraud score?
An IP fraud score rates one internet address for signs of a proxy, a VPN, a Tor exit node, a botnet or past abuse, and public lookup tools show it. It is one input to a login or payment score, never the decision. A corporate VPN looks like a proxy too.
What signals and models build a fraud score?
Signals fall into five families: the transaction, the device, behavior, identity and account history, and the network. The EU standards name six findings that rule out the risk based exemption from strong customer authentication: abnormal spending or behavior by the payer, unusual device or software access, malware in the authentication session, a known fraud scenario, an abnormal payer location and a payee in a high risk location.
Points based rules, supervised models trained on confirmed cases (the core of AI fraud detection) and anomaly detection turn the signals into one number, alone or blended.
complead, formerly Sanction Scanner, runs all three in its fraud prevention product: ready rules cover the known patterns from the start, a model trained on the firm's own confirmed cases comes in as labels accumulate, and anomaly detection that needs no labels runs throughout.
How do thresholds turn a score into a decision?
On an illustrative 0 to 1,000 scale, a payment under 200 is approved, 200 to 699 goes to step up (such as a one time code), 700 to 899 waits for an analyst, and 900 and above is declined.
How does a fraud score decide the PSD2 exemption from strong customer authentication?
| Exemption threshold value | Reference fraud rate, remote card payments | Reference fraud rate, remote credit transfers | Illustrative firm at 0.08% on cards |
|---|---|---|---|
| EUR 500 | 0.01% | 0.005% | No |
| EUR 250 | 0.06% | 0.01% | No |
| EUR 100 | 0.13% | 0.015% | Yes |
When a remote payment may skip strong customer authentication under transaction risk analysis (Delegated Regulation (EU) 2018/389, Annex; the last column is illustrative).
Under Article 18, a remote payment the payer starts counts as low risk only when all three hold: the real time analysis found none of the six findings, the amount is within the band's threshold value, and the provider's fraud rate for that type is at or under the band's reference rate. Article 19 counts that rate over a rolling 90 days across all remote payments of the type, fraud recovered or not. A rate above the reference is reported at once, and two consecutive quarters above it close that band (Article 20).
Under PSD2 the model does not decide how much may pass without strong customer authentication; the quarterly fraud rate does. An illustrative firm with EUR 32,000 of fraud on EUR 40,000,000 of remote card payments runs at 0.08 percent, so it may exempt only payments up to EUR 100, however low they score. The EUR 250 band needs fraud of EUR 24,000 or less.
Which rules apply to a fraud score?
| Rule | Jurisdiction | What it says about the score | Status |
|---|---|---|---|
| Delegated Regulation (EU) 2018/389, Articles 18 to 20 | EU | One risk score per payment exempted from strong customer authentication on risk grounds | Applies since 14 September 2019 |
| Payment Services Regulation, Articles 87 and 97 (Council text) | EU | Payee's provider monitors too; EBA drafts transaction risk analysis standards | Council text of 29 October 2026, awaiting Parliament's second reading |
| GDPR Article 22 and recital 71; SCHUFA, C-634/21 | EU | Solely automated decisions with significant effects need an exception; a credit agency's score counts as one when a firm draws strongly on it to make, run or end a contract | In force; judgment of 7 December 2023 |
| AI Act Annex III point 5(b); Regulation (EU) 2026/1744 | EU | AI that detects financial fraud is excepted from the high risk class for credit scoring of natural persons | Annex III duties from 2 December 2027 |
| SR 26-2 and OCC Bulletin 2026-13 | US | Model risk guidance: validation and outcomes analysis, vendor models included | Guidance since 17 April 2026, not enforceable |
| Regulation B, 12 CFR 1002 | US | A declined credit application needs specific reasons, given or on request; a failed score is not one | In force; amendment effective 21 July 2026 |
| Nacha fraud monitoring rules | US (ACH) | ODFIs, non consumer originators and third parties need risk based processes to catch ACH entries made because of fraud; RDFIs, for the credit entries they receive | Phases from 20 March and 22 June 2026 |
Rules that reach a fraud score (as of 15 November 2026).
Which guidance superseded SR 11-7?
SR 26-2 supersedes SR 11-7 as US model risk guidance. It sets no enforceable standards and is expected to matter most for banks with over $30 billion in total assets. Its model definition excludes simple arithmetic and deterministic rule based processes with no statistical, economic or financial theory behind them, so a score of hand set rule points can fall outside, while a trained model's score usually falls inside.
Is a decline on a fraud score adverse action under ECOA?
Under Regulation B, it depends on what was declined. Refusing a purchase at the point of sale is not adverse action, unless the refusal ends the account or makes its terms worse in a way that does not reach all or substantially all accounts of its class, or denies a request for more credit. Refusing credit in substantially the amount or on substantially the terms requested is adverse action, unless the applicant takes up a counteroffer, and "failed to achieve a qualifying score" is an insufficient reason under Section 1002.9(b)(2).
In the EU, GDPR recital 71 lets Union or national law expressly authorize automated decisions for fraud prevention, with safeguards, and Dun & Bradstreet Austria gives a person subject to a solely automated decision with significant effects the right to an explanation of "the procedure and principles actually applied".
How do you read, tune and benchmark a fraud score?
- Drivers before the number. Two scores of 800 can rest on different signals.
- Benchmark by fraud rate per band. Comparing confirmed fraud per score band is one form of the outcomes analysis the revised guidance describes; a band that drifts may warrant recalibration.
- Past traffic first. Price a new cut on last quarter's events, in losses, stopped customers and analyst hours, before it goes live.
- Vendor scores tested on your outcomes. The guidance applies even without the vendor's code and data.
- A log for every change. What moved, when, and who approved it.
The payment hardest for a score to read is the one a customer makes after a spoofing call from "the bank": device, login and behavior are all the customer's own.
Where does a fraud score sit in fraud and AML work?
The fraud team owns the thresholds and the outcomes; AML reads the same events. A payee account taking money from many unrelated senders looks like fraud to one team and like layering to the other.
complead assesses fraud and AML risk from one set of customer data and events, so the two teams share a record, and each fraud decision (approve, decline, review or step up) arrives with its reason.
What changed for fraud scores in 2026?
- 20 March, then 22 June. Nacha's fraud monitoring rules started in two phases: first for every ODFI and for the larger non consumer originators, third parties and RDFIs by 2023 ACH volume, then for all the others.
- 21 July. Regulation B now states that ECOA does not provide for an effects test.
- 29 October. Council text of the Payment Services Regulation adopted: the payee's provider monitors before releasing funds (Article 87).
The procedure file lists Parliament's second reading for the plenary of 14 December 2026, indicatively; the EBA would then draft the standard on "the methodologies and models to implement transaction risk analysis".
Related terms
- Device fingerprinting. Identifying a device from its hardware, software and browser attributes.
- Transaction fraud. A payment the account holder never authorized, or one the holder was tricked into sending.
- Account takeover fraud. A criminal taking control of a real customer's account.
- Card not present fraud. Paying with stolen card details where the card is never shown.
- Customer risk assessment. Rating one customer's money laundering and terrorist financing risk to set the depth of due diligence.
- Spoofing. Faking a caller ID, an email address or a payee name to pass as a trusted sender.