How AI helped double volume at this auto lender

Register now

Until recently, Prestige Financial Services was struggling to cope with a down market in which the delinquency rate for subprime auto loans is higher than it was even during the financial crisis. Like other subprime auto lenders, it tightened its credit scoring.

But this year has seen something of a turnaround. Prestige, which targets borrowers with sub-550 FICO scores, 40% of whom are in bankruptcy, has doubled its lending volume since January. It’s had a 36% increase in new applicants and a 14% increase in borrower approvals. The company says the new loans are performing as well as, if not better than, those Prestige had previously issued.

The difference, according to Steven Warnick, chief credit and analytics officer for Prestige Financial Services, was that this year it started using machine learning software from ZestFinance to make loan decisions.

“The new software let us price loans more competitively because of the ability to see risk in a new way across customer applications,” Warnick said.

The more competitive pricing brought in new borrowers from dealerships Prestige hadn’t dealt with before, with loan rates that range from 11.99% to 24.99%.

More data, better decisions

Previously, the company used a logistic regression model to score loan applications. The model could accept only 23 creditworthiness attributes.

The machine learning model Prestige created with ZestFinance considers 2,700 data elements derived from credit bureau data, commercially available alternative bureau data and loan application data. It's not so much that it's looking at different data, but that it's looking more deeply into the data the lender receives.

“Each one of those variables we add in, in and of itself, isn’t that significant, but each one adds an incremental amount of power,” said Mike Armstrong, president of ZestFinance.

One example: A conventional model may consider whether or not a potential borrower has ever had a bankruptcy. The machine learning model will consider what type of bankruptcy the person had (Chapter 11 or 7), when it happened and how many bankruptcies the person has had.

The new model also analyzes the interactions of those bankruptcies with the other 2,000-odd variables in the person’s record. For instance, a potential borrower who had a Chapter 7 bankruptcy and a payment delinquency four months ago might be a higher risk than one who had a delinquency 12 months ago.

In the subprime industry, FICO has never been a good predictor of risk, Warnick said.

“It’s been proven over time that a more sophisticated model like a neural network is better at predicting the end result than a logistic regression model,” he said.

Black-box worries

Like all lenders, Prestige worried that using artificial intelligence to make loan decisions could prompt regulators to claim it was using a black box, within which loan decisions couldn’t be adequately explained.

“We were skeptical at the onset,” Warnick said. “The black box is a major concern from a compliance level and an executive level because we want to understand what’s going on.”

Armstrong acknowledged that this comes up a lot with financial institutions.

“When you put a model in production, you have to make sure the models isn’t discriminating and there’s no disparate impact,” Armstrong said. “If a lender declines an applicant, it needs to flesh out what were the key variables that triggered the decline decision and furnish those adverse-action notes to the consumer in the form of an adverse-action letter.”

Warnick found that the Zest software provides rank-ordering attributes and the impact that those attributes have on the model, so it has just as much explainability as the old model.

“That allowed us to gain comfort in those primary attributes that were impacting the score,” Warnick said. The software also generates the adverse-action codes that must be given to every potential borrower who is declined.

Armstrong said Zest has talked to the main bank regulators like the Consumer Financial Protection Bureau and the Office of the Comptroller of the Currency and has gotten the sense its technology is aligned with regulators’ goals.

“We want to make sure we’re able to offer fair and transparent credit to everybody,” Armstrong said. “Through this technology, more Americans who have been overlooked by the incumbent technology will be able to get credit from lenders.”

Editor at Large Penny Crosman welcomes feedback at

For reprint and licensing requests for this article, click here.
Machine learning Artificial intelligence Auto lending RegTech Conference