Fraud-contaminated credit models face £1.28bn UK fraud backdrop

Fraud contaminated credit models are drawing closer scrutiny in the UK after payment fraud losses reached £1.28bn in 2025, according to UK Finance’s 2026 Annual Fraud Report. Ben O’Brien, managing director at Jaywing, said fraud wrongly treated as ordinary credit loss can distort lenders’ scorecards and affect future underwriting decisions.

UK Finance reported that total payment fraud losses rose 4% year on year in 2025, while authorised push payment fraud increased 19%. O’Brien also cited Cifas’s Fraudscape 2026, which described organised and networked fraud, including synthetic identities created from bought data and money mule activity, as a persistent feature of the UK credit market.

His central warning is that some accounts recorded as credit defaults may not have failed because of affordability, income pressure or overextension. If those cases were fraudulent from the outset, a credit model trained on them may learn the wrong relationship between borrower characteristics and default outcomes.

How does fraud contaminate credit models?

A credit scorecard learns from historical accounts labelled as good or bad. When fraud is hidden inside the bad population, the model treats fraudulent profiles as if they were genuine borrowers who later suffered credit stress.

O’Brien said this can lead lenders to keep approving similar profiles because the model has no independent way to tell whether the original loss came from borrower behaviour or an organised fraud scheme. He argued that a retrospective review can help, using graph analytics and network checks to identify suspected organised fraud or synthetic identities in past portfolios before retraining models on cleaner data.

Graph analytics examines links between accounts, identities, devices, addresses and other shared attributes. In fraud work, it can expose clusters that appear unrelated in a conventional account-by-account review.

Organised fraud adapts differently from credit risk

O’Brien distinguished organised fraud from conventional macroeconomic credit risks such as inflation or unemployment. In his view, fraud networks actively test controls and alter their behaviour when a lender changes its screening process.

Cifas data cited by O’Brien showed SIM swap fraud rising 38% in 2025 and account takeovers linked to mobile, credit card and online retail products rising 90%. He said such movements indicate fraudsters shifting across channels where controls appear weaker.

He also pointed to slower-moving fraud already inside portfolios. Cifas recorded a 43% rise in misuse-of-facility cases in 2025, a category O’Brien linked to accounts that may behave normally for a period before a later bust-out event.

What should lenders monitor?

O’Brien said conventional backtesting may fail to reveal the problem because a model can appear accurate when it is tested against the same contaminated outcomes on which it was trained. He identified three monitoring signals: first instalment defaults, failed contact in collections and defaults among applicants with unusually clean credit profiles at origination.

He also set out three broad ways firms can combine fraud and credit risk work. A shared feature pool can give models richer inputs but may make decision processes more complex. A single outcome model can treat fraud and credit losses together, but O’Brien said that approach creates regulatory difficulties because firms must explain model decisions and loss attribution.

He described pipeline decisioning as the common compromise: screening clear fraud cases before credit assessment, then applying further fraud checks after credit and bureau data have been gathered. O’Brien said the UK’s failure to prevent fraud offence under the Economic Crime and Corporate Transparency Act, which came into force in September 2025, makes the choice of integration approach part of what firms may need to justify as reasonable procedures.

This story draws on original reporting from Finextra Research.