From credit risk to fraud risk: what changes and how to prepare

If you've spent your career in credit risk, you've mastered the art of the long game. You understand predictive modelling, risk thresholds, macroeconomics and portfolios. You're used to assessing a borrower's ability and willingness to pay over months or years.

But if you're looking to make the leap into fraud risk, you need to prepare yourself. While both departments sit under the same risk umbrella, transitioning from credit to fraud calls for a complete change in mindset, and that goes much deeper than a new job description.

UK payment fraud losses reached £1.28 billion in 2025, and Cifas recorded more than 220,000 fraud-risk cases in the first half of 2026, the highest January to June total on record. Meanwhile in Thailand, 78% of bank C-suite leaders in a 2026 BioCatch survey reported rising fraud losses.

If you try to manage fraud with a credit risk mindset, you're going to run into some painful pitfalls. Here's what you need to know to make the transition successfully.

Key takeaways from this article: You're facing an active adversary. Fraudsters test your systems and look for blind spots, so the backward-looking logic of credit risk breaks down quickly. Investigation runs on narrative. Catching a fraud ring means reconstructing how they got in, what device they used, where the money went and who else they link to. GINI is the wrong yardstick. Fraud models are judged on false positive rate, review rate and how much fraud lands in the worst-scoring slice your team can investigate. Credit habits create blind spots. Trusting static identity data, over-applying friction and recording fraud as credit losses are the classic traps. Fight a network with a network. Relationship-first tools such as graph databases are central to detecting organised fraud

 

Credit risk and fraud risk at a glance

	Credit risk	Fraud risk
The adversary	Macroeconomics, life events and bad financial planning	An active, intelligent and organised fraudster
Time horizon	Months or years, often a 12-month outcome window	Milliseconds at the point of decision
Unit of analysis	The portfolio cohort	The attack and the network behind it
Typical models	Scorecards, logistic regression, gradient-boosted trees	Neural networks finding non-linear combinations in real time
Success metric	GINI and KS	False positive rate, review rate and capture in the bottom x%
Typical decision	Accept or decline	Refer for investigation, apply smart friction or let it flow

The headspace change: predictable maths versus active adversaries

In credit risk, your "adversary" is largely macroeconomics, life events or bad financial planning. Borrowers don't usually take out a mortgage with the explicit, malicious intent of defaulting on day one. Because of this, credit risk relies heavily on historic, linear behaviours. You look backward to predict forward.

In fraud risk, you are fighting an active, intelligent adversary.

• The fraudster's headspace. Fraudsters are creative, tech-savvy and highly organised. They aren't struggling with cash flow. They are actively testing your systems, finding your blind spots and trying to steal from you.
• The pivot. You have to stop thinking like an actuary and start thinking like a detective. Instead of calculating probability, you are trying to anticipate their next move.

Cifas data shows how quickly fraud tactics move. Unauthorised SIM swap cases rose 402% year on year in the first half of 2026.

The storytelling approach: moving beyond the spreadsheet

One of the hardest adjustments for credit risk veterans is moving from aggregate data to narrative storytelling.

In credit risk, a single customer who defaults is a statistic. You analyse them as part of a wider portfolio cohort.

In fraud risk, every attack has its own narrative. To catch a fraud ring, you have to reconstruct and understand it by working through four lines of enquiry:

•     How did they get in?
•     What device did they use?
•     Where did the money go?
•     Who else did they link to?

This is where data storytelling becomes your primary weapon. Instead of staring at static credit scores, you have to map out connections. Fraudsters operate in networks, sharing IP addresses, phone numbers and bank accounts.

To explain a fraud risk to your board or your team, a loss percentage on its own won't land. You have to walk them through the network, showing how seemingly unrelated accounts are in fact puppets controlled by the same master.

The model development change: why neural nets win and why GINI is a distraction

If you are a credit risk modeller, you have likely spent your career worshipping at the altar of the GINI coefficient or Kolmogorov-Smirnov (KS) metrics. You build scorecard models, optimise for the absolute cleanest cumulative distribution curves, and use GINI to prove your model's predictive power over a 12-month horizon.

But in fraud modelling, you need to put the GINI textbook away. Here's how the machine learning priorities change when you develop models to fight active adversaries.

Why neural networks are the choice

In credit, explainable logistic regression or gradient-boosted trees (like XGBoost) are the gold standards because you must explain to regulators exactly why an individual was declined. Fraud is a different beast.

• Non-linear, complex interactions. Fraudsters do not trigger a single red flag. They trigger complex, fleeting combinations of indicators, like a rapid sequence of device swaps paired with mismatched network latency and micro-deposits.
• The neural net advantage. Deep learning and neural networks are exceptionally good at finding these highly non-linear, multi-dimensional relationships across vast datasets in real time. They learn their own representations of the data. Because speed is everything, a well-optimised neural net can ingest thousands of raw inputs and output a probability score in milliseconds, long before the transaction window closes.

The metric trap: why GINI is the wrong focus

In credit, a model with a high GINI is a success because you are predicting a gradual, steady trend of defaults across a portfolio. In fraud, a high GINI can be a dangerous illusion, because fraud is highly imbalanced, volatile and time-sensitive. A model can have a spectacular GINI on historical data but fail completely in production because the fraud network changed its tactics yesterday.

Instead of GINI, the focus must be on operational, real-world metrics.

• The false positive rate (FPR). In credit, a decline is a decline. In fraud, if your model flags fifty genuine transactions to catch one fraudster, you are destroying the customer experience. Your core constraint during model development is minimising the FPR at your required detection threshold.
• The review rate. Start with how many cases your fraud operations team can realistically handle. If your model flags 5,000 "suspicious" accounts a day, but your analyst team only has the capacity to review 500, the rest are either auto-approved (allowing fraud to slip through) or auto-declined (creating massive customer friction). Your model must be optimised to fit your operational bottleneck.
• Maximising volume into the bottom x%. Fraud models are generally used to refer suspicious activity for investigation instead of declining it outright. Therefore, the metric that counts is capture rate in the top review tier. Your goal is to pack as much of the true fraud volume as possible into the bottom x% of your scored population (where x% represents the maximum capacity your company can physically investigate). If you can squeeze 85% to 90% of all fraudulent transactions into the worst-scoring 2% of accounts, your model is an operational masterpiece, regardless of what its overall GINI curve looks like.

And review capacity is already stretched. Fewer than a quarter of Thai banking leaders in the same BioCatch survey say they can fully investigate a fraud case within a day, while UK Finance counted 3.81 million unauthorised fraud cases in 2025. It's one reason we've asked whether fraud controls are becoming unsustainable.

Classic pitfalls for credit risk grads

When credit risk professionals cross over, they usually trip up on three key areas.

Pitfall A: over-relying on static identity data

In credit, if an applicant's name, address and date of birth match the credit bureau, you're usually good to go. In fraud, this is a massive vulnerability. With the rise of synthetic identity fraud, where criminals mix real and fake information to build credit profiles, an applicant can look pristine at application while being completely fabricated.

So instead of trusting single data points, you have to learn to look for "hyper-connectivity" (for example, ten different "pristine" applicants sharing the same phone number).

Identity fraud made up 59% of cases recorded to the UK's National Fraud Database in the first half of 2026, with Cifas flagging growing concern about synthetic identities. Our Fraudscape 2026 update looks at what that means for fraud teams.

Pitfall B: misunderstanding friction and false positives

Credit risk decisions are slow and deliberate. Fraud risk happens in milliseconds. If you apply too much friction to stop every single possible fraud attempt, you will alienate your genuine customers.

A key part of fraud risk is calibrating to false-positive rates as well as fraud losses. You have to get comfortable with "smart friction", knowing exactly when to step in and when to let a transaction flow.

Pitfall C: recording fraud as credit losses

Because sophisticated fraud (like synthetic identities or first-party fraud) often starts by looking like normal customer behaviour, these accounts will make their first few payments and then disappear. If you aren't careful, your systems will record these as standard credit defaults.

As a transitional leader, you must bridge this divide, ensuring your team can separate true credit default from malicious, organised fraud. We've looked at what happens when it goes unchecked in our piece on fraud contamination in credit risk models.

Fight a fraud network with a network

The ultimate lesson of fraud risk is that you cannot fight a highly connected network of criminals using siloed, tabular databases.

To survive the transition, you must embrace relationship-first technologies, like graph databases, and develop a storytelling mindset. Learn the headspace of your adversary, master the narrative of the transaction, and stop looking at data in separate rows and columns.

Your credit experience still counts. Modelling discipline and experience of explaining decisions to regulators carry straight across, and they're valuable in a fraud team that needs models it can defend. Jaywing's Archetype platform was built around that same need for explainable models.

Start with your own book. A practical first step is reviewing early-life defaults in your credit portfolio for signs of first-party or synthetic fraud. It builds the investigative habits described above and gives you a cleaner view of true credit risk at the same time.

See the network in action. Our companion piece, How a graph database caught 23% more fraud from a synthetic identity ring, shows how one lender put the network approach into production.

Jaywing's fraud specialists work with lenders on both sides of the line, from fraud model development to credit decisioning strategy. If you'd like to test these ideas on your own data, you can request a proof of concept or get in touch with the team.