The recent banking crisis adds even more pressure on buy now/pay later lenders, which are taking steps to better manage underwriting and credit decisions in the months ahead.
Many lenders want to limit damage by using machine learning, which became more widely used during the past year as the BNPL industry faced financial and regulatory challenges.
"We have found it to be a better predictor than [traditional credit scoring]. We also use another machine learning model for fraud to determine if consumers are real or not," said Lee Brading, senior vice president of corporate development for Sezzle.
Sezzle's risk team uses machine learning to create a model it calls Prophet, which it launched in 2022.
"Another key area is consumer support. We use [artificial intelligence] in chat with consumers. … [It] saves us on personnel costs but also gives the consumer a quicker more timely response," Brading said.
The valuations of many
The recent closure of
"We're looking at different aspects: delinquencies, pricing, capital and any exposure to troubled firms," said Libor Michalek, president of Affirm.
One area that Affirm is considering is the company's own ability to originate credit. "We feel good about that. We have a robust set of backups," Michalek said. "We are keeping an eye on developments and can make adjustments."
Other concerns are more related to consumers. Affirm uses AI alongside other data modeling to vet consumers' ability to repay, or for Affirm to adjust the size of the loan, interest or time to repay. These can be changed quickly based on a consumer's personal economic circumstance or that consumer's ties to an external risk.
"A lot of these decisions aren't binary, or yes or no. It's more around the terms, the payment schedule, down payment or pricing," Michalek said.
Affirm uses what Michalek calls an ensemble of models that look at different sources of potential risk. Since BNPL borrowers tend to be repeat customers, each purchase feeds more data into the modeling, which could change the terms of future loans. Affirm communicates that to consumers when they apply for a new BNPL loan, Michalek said.
As economic conditions started to shift in 2022, Affirm made more than 100 changes to its credit program that are embedded in the company's decision making. Additionally, the company's duration of BNPL loans is about five months, which allows Affirm's books to "turn over" quickly.
"We have all recently learned that
A BNPL lender has three major concerns when making credit decisions: fraud, nonpayment and late payments, according to Nathan Hilt, payments and fintech solutions lead at Protiviti.
"It's difficult to catch fraudsters to collect stolen goods or recoup cash," Hilt said. "They were never a client. They targeted the BNPL because they could successfully perform a fraudulent transaction."
In the case of nonpayment, the borrower is known, which can help the BNPL lender collect on the debt or sell it off. "The purchaser may be able to purchase from the lender again using direct debit or a credit card," Hilt said. "There is a return-on-collection expense cutoff where a lender will write off the loss."
Additionally, a borrower's ability to pay back may change after the purchase, such as a sudden event for the borrower that could be linked to an economic crisis.
"Although the client pays, the ability for the BNPL provider to turn over capital begins to decrease which will erode margins," Hilt said.
Machine learning can contribute to risk management in this area by pulling banking data to analyze the income and expense types for the borrower's bank account(s) via application programming interfaces, according to Hilt.
Accounting data is being used via APIs that can provide revenue, expense and debt or liabilities to generate common cash and liquidity ratios typically used for underwriting, Hilt said, adding that this can determine the probability of default.
"Additionally, the type of company the borrower may be exposed to, liens, taxes and standing of the company can all be pulled using APIs into state and federal offices," Hilt said.
For BNPL lenders that originate new loans to subprime consumers, AI powered by machine learning could help reduce write-offs, said Nandan Sheth, CEO of Splitit, which supports installment lending by drawing on users' existing credit card lines, rather than creating new debt. Splitit uses AI to match consumers with products or categories.
"If we get this right we can deliver incremental sales or high [average order value] without disintermediating the merchant-shopper relationship," Sheth said.
The important role AI can play in aiding credit risk is drawing some attention from investors.
Bud Financial, a London-based open-banking and data intelligence firm, recently expanded to the U.S., buoyed by financial backing from TransUnion, which made a strategic investment in Bud in February. HSBC and Goldman Sachs are also among Bud's investors. "If you're making real-time decisions on lending, and the data is not real-time, the BNPL loan's terms can be based on performance that's four or five loans behind," said Ed Maslaveckas, CEO and co-founder of Bud Financial.
The amount of the TransUnion investment was not disclosed. Bud's core product is Transaction AI, which enables lenders to gain insight into a borrower's income, expenditure and creditworthiness in nearly real time.
"A lot can change in 24 hours, so the credit picture today may be very different from what it was four months ago," Maslaveckas said.
