Far from the Gidget-esque California of bathing suits and Santa Monica beaches - specifically a two-and-a-half-hour drive north with one hour spent driving on a two-lane road surrounded by crops - is the small city of Porterville.
There, truckers honk in a friendly way at unknown passersby, locals chance upon the town judge at the Mexican cafe, and a farm-heavy landscape demands people to drive to work. Porterville is the type of place rural Americans know well and city dwellers might find confusing.
On a recent drizzly day off of Porterville's North Main Street, Robert (Bob) Hughes worked in a Finance and Thrift office surrounded by books on big data, an iMac, photographs of horses and family, artwork created by his daughter (more of which is showcased in the parking lot) and a glass wall for collaborative doodling. More curious items on display in the chief executive's office include an electric dollar sign, a mug that reads, "I find your math disturbing" and Australian throwing sticks.
Here, Hughes makes decisions about how the $128 million-assets bank will serve people in his region who may have accounts at department stores, pawn shops or cash checking, necessities like indirect auto loans. At the same time, decisions he makes need to make shareholders money.
When Hughes joined the community bank in 2007, used car inventory was low, unemployment rates were high and Beyonce's "Irreplaceable" was the top song on Billboard. F&T operated 22 branches -- each location procuring and originating loans with presumably different styles.
Realizing that some branches had four times the loss rates of others, Hughes sought out a computer-run model that could quicken underwriting decisions for subprime loans that require higher capital of the bank. "If you originate a $500 loan, you can't spend $500 making it," Hughes points out. F&T offers deposit accounts and loans for expenses like furniture, funerals and birthday parties, but indirect auto loans account for most of its sales.
Hughes hankered for a loan decisioning model that made the same decisions every time to improve its operational challenge. Because unlike humans, an algorithm wouldn't get tired or fickle.
"You can teach a [credit underwriting] model to eventually get to the right answers," says Hughes, who wrote on regression analysis as a student. "It's all about consistency."
Where a human can see you're wearing purple and smile large, data shows you work for Kmart, pay your bills on time and are likely having an affair. You cannot hide from an algorithm unless you pay in giftcards, bitcoin, cash or the data is flawed, modelers believe.
To execute his vision for the bank, which is regulated by the California Department of Financial Institutions and the FDIC, Hughes sought out Patrick Reily, chief executive officer of Verde International in Atlanta. Reily, a vintage car enthusiast who has used transactional data for predictive modeling since the late 1980s, can conjure up a person's character through math the way Meryl Streep portrays them in movies. "For me, [modeling] is like a symphony. It's a beautiful expression of how the world is," says Reily. "I understand people through mathematics."
Expressing his modeling techniques and Hughes' business rules, Reily built a web-based loan origination system (LOS) and decision engine called Aurora to create consistency for a borrower base of largely low-income laborers.
"You don't want decisions based on a bad day," says Reily. "The computer doesn't get bored."
Updated in September and widely available for purchase to banks in December, Verde's Aurora can scale up to make 100,000 loan decisions per hour for F&T. Currently, the bank originates about 380 loans monthly that fund $11,000 cars. Interest rates average 16% and contracts run up to 60 months.
Aurora, which has been in place at F&T since 2008, calculates creditworthiness based on estimates of the size, timing and probability of events such as an early repayment, the expected economic impact of a lending decision and anticipated monthly cash flows. A farmer will wait to hire a picker if the lemons come in late, which will delay the borrower's auto repayment. But later, he could catch up.
"The person with little respect for the [loan] obligation, living beyond his means, has a pattern distinct from that of a person living responsibly but confronted with an income disruption he can't navigate around," Reily says. "Each pattern is expressed before, during and after the actual failure and is distinguishable just as easily as one can distinguish a Van Gogh from a Matisse. And most importantly, these two have very different future default risks."
Aurora finds clues even in the way applicants spell. "I would rather lend to a person who spells poorly than a person who doesn't," Reily says. Likewise — and somewhat counter intuitively -- healthcare debt and couch surfing do not imply riskier auto borrower behavior. "If a pattern shows the borrower is relocating for a better job opportunity, that's a good indicator of a willingness to repay debt," says Reily.