Common Thread #6: Heads, fraud. Tails, credit.

The worst thing lenders can do for their portfolio quality is judge credit and fraud separately.

Common Thread #6:                 Heads, fraud. Tails, credit.

Why lenders lose when fraud risk and credit risk are judged in silos.


Hello everyone,

In the time it takes you to read this, India's digital NBFCs will approve well over a thousand personal loans. At that pace, even a small blind spot in how lenders decide gets repeated thousands of times a day.

One of those blind spots sits between two teams. At many lenders, the fraud team and the credit team look at the same borrower through different prisms, on separate systems, and neither sees what the other found. So a lender can turn away someone it already knows is good, or approve someone it had every reason to stop.

Take Shashi, who is applying for her first loan late on a weeknight.

She draws a steady salary but has never borrowed before, so the bureau has almost nothing on her. When the app asks for permission to read her phone data and bank statement, she taps yes, and a couple of minutes later she has her answer.

In those two minutes, two systems inside the lender looked at her.

The first was the fraud check. It saw the same phone, SIM and home location going back more than a year, found nothing that didn't add up, and cleared her as a real person.

 The second was the credit model. It went to the bureau, found almost no history and declined her. Her bank statement, with a salary landing on the same date every month, never came into the decision.

 Both systems worked exactly as they were designed to. One side of the lender knew she was genuine, while the other never thought to ask. Shashi closed the app and most likely opened a competitor's instead.

 The lender won't record Shashi as a loss anywhere. She wasn't fraud, so she never shows up in the fraud team's numbers, and a declined application costs nothing on paper, so she doesn't show up in credit's numbers either. That is exactly why the same thing keeps happening to the next applicant like her.

 Shashi isn't a rare case, and the cost of the split doesn't stop at customers who walk away. Southeast Asia's digital loan books are growing by about $57 million a day, and defaults are rising with them. In Indonesia, online loans in default have climbed to about $325 million, some $180 million more than a year ago. In the Philippines, bad salary loans at banks are up by around $115 million, to roughly $590 million. Not all of that comes from the split between fraud and credit, but some of it does, and the clearest sign of it is a loss that neither team wants to claim.

The loss nobody wants to own

Now take the opposite case. A different applicant has a decent credit history and is approved without trouble. The fraud checks find nothing unusual on the day, the money goes out, and he never pays the first instalment.

 A few weeks later, the argument starts. Credit says it was fraud, because nobody who intends to repay misses the very first payment. Fraud says it was credit, because the identity was real and the documents checked out. Both have a point, but the loss can only be booked to one side, and only that side's model learns from it. The other team goes on approving, or clearing, the same profile the following month.

 This isn't only a problem in our markets. Experian has written about how banks often write off early fraud losses as ordinary credit losses, which distorts their credit models and hides how much fraud they are actually carrying.

In my conversations with risk teams at banks and NBFCs over the past year, in India and across Southeast Asia, one question comes up again and again. It is rarely about whether the scorecard is accurate or whether the fraud rules are tight enough.

It is about ownership: whose loss is this?

 

Approved, then forgotten

There's a second problem, and it starts once a loan has been approved.

Most lenders score a borrower once, on the day they apply, and that score then sits in the file for the life of the loan, whether it runs for one year or two.

 People's lives don't stand still for that long. Someone who looked solid on approval day might take on two more loans six months later or start getting paid late. The warning signs usually sit in the same data the lender used to approve them, yet in most setups nobody looks again until the payments stop and collections picks up the phone.

 The case for re-reading borrowers as their cash flow and repayment behaviour change isn't new. Rajat made it on The Big Picture back in 2023, in a piece on continuous underwriting. 

What changes when both sides decide together

Fixing both problems doesn't take a better fraud model or a sharper scorecard. It takes a single decision that sees what both teams see and keeps looking after the loan goes out. Three NBFCs we work with in India made that switch, each starting from a different problem.

The first had a Shashi problem: it was turning away thin-file applicants because the bureau had little to say about them. Once fraud and credit signals fed one decision, approvals rose 10% while risk held flat, which works out to an estimated $3 million a year in revenue the lender had been walking past.

 The second lends personal loans to people with no credit history at all and underwrites around 15,000 borrowers a day, so for a lender like that the bureau is close to silent. With signals beyond the bureau in the decision, it has brought more than 50,000 thin-file borrowers into credit, earned $5.3 million in profit and cut its risk cost by 12%.

 The third had the opposite problem, because it was approving people it shouldn't have. On its book, the bureau score alone was a weak predictor of default, with a Gini of 13.4. Combined with device data, that rose to 37.8, nearly three times the predictive power, and bad loans fell 24% while approvals held.

 That last result is worth dwelling on. Most risk teams accept, almost as a law of lending, that cutting bad loans means cutting approvals, but here it didn't. Much of that trade-off is really the cost of keeping the two sides apart.

 That is what we built Sentinel AI to do: check whether someone is real, whether they can repay and at what price, all in a single flow.

The common thread

Every lender I speak to has a fraud team that is good at its job and a credit team that is good at its job. The problem is that they work in separate systems and rarely share what they know.

We're presenting this work at the Global FinTech Hackcelerator Demo Day at the Singapore FinTech Festival on 18 November, so if you're around, come and say hello. And if this sounds a lot like your last risk meeting, drop me a note, because I'd love to hear how your team handles it.

Cheers,

Bindesh Pandey

Still exploring this topic?
Get instant, cited answers from the FinBox lending knowledge base

Stay current

Get research like this in your inbox.

Join 5,000+ lending professionals who read FinBox's research on credit infrastructure, underwriting, and embedded finance.

Subscribe free