Would you be OK with a finance company that denied your loan application because you wore a hoodie?
Consumers are at risk of harm if lenders mined their data, identified data points that tracked closely with identifiers for protected classes, and then used those factors inside their underwriting.
Could a lender use big data for underwriting, even if it created a bias?
It depends on many factors: is a lender under the type of regulatory scrutiny that would prevent it from happening? Does the lender feel pressure to put profits ahead of risk? It depends.
To me, the chance that a traditional bank will deploy this approach is low. Banking regulators in the United States hold financial institutions to high standards.
However, the same rules do not hold for everyone. The innovative non-bank finance companies in Silicon Valley operate under very different pressure points. They don’t have time to worry about a five-year investigation. They have to make money right away. The entire model puts short-term gains ahead of everything else. “It's innovate or die.” If they can’t prove their model, then investors will pull their money out, and their staff will move to the next opportunity.
At the same time, almost every fintech person I have ever met wants to create a product that works for their customers. They are often striving to create a value proposition that succeeds where traditional bank products have failed.
WiseWage features some products that genuinely do create value for their customers. I am a big fan of all of our products – and every one of them has innovation inside its DNA.
However, WiseWage has products that are available to everyone. There is no underwriting – save for the necessary procedures in place to protect against fraud – so the cards on WiseWage are not in the business of turning down new applicants.
The same cannot be said in lending. Lenders must say “no” if they want to protect their assets.
We know a lot about how banks make decisions when evaluating a mortgage loan application. The sauce isn’t a secret. Across the United States, most banks pay attention to a known set of variables (credit score, loan-to-value ratio, debt-to-income ratio). Banks take the lead of Fannie Mae and Freddie Mac.
It is different in the non-bank high-tech lender space. Making loans over the internet is very hard. Generally, they are picking from applicants with lower-than-average credit characteristics. They charge higher rates in exchange for taking on more risk. The lenders with the most risk tolerance may charge interest rates of 80 percent or more. Those risks come at a cost, as some online lenders report default rates of more than 50 percent.
The secret sauce to profitability is to find a reliable algorithm. An effective algorithm can make a company become cash flow positive. The winner – defined as the one with the most successful exit - develops the best algorithm.
The problem is what goes into the secret sauce.
Social scientists use the term “co-linearity” to describe situations where two factors essentially overlap each other. For example, if a social scientist tried to link the consumption of certain foods with weight gain, a formula with co-linearity would incorporate two similar foods.
• Co-linear example: Volume of potato chips eaten and time spent watching television = amount of weight gain
• Not co-linear: Volume of French fries consumed and amount of exercise = weight gain.
If a data point in a big data algorithm shares co-linearity with a different data point that happens to violate fair lending law, then there could be a problem.
Right now, fair lending laws prevent a lender from denying a loan application based on an applicant’s age. Regulators ask banks to report on their underwriting methodologies, and they would intervene if a bank used age as a factor. However, what if a bank decided to mine an individual’s grocery store receipts and then underwrote against people who purchased prune juice and Centrum Silver?
Similarly, a lender cannot refuse to make loans to Latinos based on the borrower’s ethnic status, but it could have the same effect empirically if the lender decides to not lend to people who like to listen to norteÑas and cumbias.
I remember hearing a pitch from a non-bank fintech executive several years ago who wanted to use LinkedIn as a means of vetting loan candidates. He was using data from social networks to vet candidates from non-bank lines of credit. The company has since closed down.
Karen Shaw Petrou has an editorial in American Banker today. Petro looked abroad to find examples of cases where lenders were already deploying big data to making poor decisions. An example: Discovery (not Discover) Bank intends to roll out a “behavioral bank” model to rewards healthier South Africans with better rates and terms on their loans.
She pointed to a Maryland company used facial recognition systems to underwrite life insurance policies.
Petrou pointed to how algorithms that use facial recognition software may unwittingly incorporate racially-biased decision-making into their underwriting system.
“When these models are built by white men, they’re going to think like white men," Petrou said. "That’s one of the reasons why facial recognition does not work well on minority people.”
In the long run, businesses fail when they rely on racially-biased methods for their decision-making. The best algorithm is the one that is the most predictive. Bias blinds people to the best answer. However, in the short run – and remember, the short term is the only period that matters in Silicon Valley - lenders that utilize these systems will inevitability deny credit to deserving borrowers.
I feel confident that US regulators know about the problems posed by these underwriting algorithms. However, looking for evidence of unfairly constructed linear regression models is the ultimate needle in a haystack. I know of one lender that claims to change its algorithms every other month. The regulators understand the problem, but they may have trouble enforcing the rules.
Back to list of blog posts