Approximately 7.1 million U.S. households are unbanked, according to the most recent data from the FDIC, and 1.7 billion adults worldwide lack a bank account of any kind, according to global estimates from 2017.

While some major U.S. banks have plans to serve the unbanked, including the $2.9 trillion JPMorgan Chase, $1.96 trillion Wells Fargo and $498 billion U.S. Bancorp, about 53 million people in the U.S. don’t currently have traditional credit scores, according to published reports. This raises the question of how banks will extend credit while managing risk.

Alternative data sources may provide a path forward by allowing banks to create their own, internal credit scoring systems, said Rajiv Garg, associate professor of information systems and operations management at Emory University’s Goizueta Business School. Garg studies advanced technologies, including bots and artificial intelligence (AI).

A simple linear model could create a credit rating, Garg said, and the model could be improved over time by adding feedback data and leveraging AI tools such as machine learning. Banks could also use AI to analyze the data for behaviors that might indicate risk — or a lack of it — in ways traditional credit scores do not.

“These large banks, JPMorgan and Bank of America, they’re huge — they have all of these data on their consumers all across the world,” Garg told Bank Automation News. “They can see the financial behavior of every consumer and they can find similarities, using AI, unsupervised learning.”

AI could allow financial institutions to reach the unbanked or underbanked and keep them as customers before alternative financial technologies — cryptocurrencies, tech cards and investment apps — do so, Garg pointed out.

“If I was an executive at JPMorgan, I would have done this five years ago; every bank should be doing it,” Garg said. “From a bank’s point of view, if I’m giving you an internal credit score, and giving you a loan or credit cards … you’re going to be locked into my banking system because you got a loan from me.”

Banks have similar levels of data — for example, who makes car payments on time and who regularly contributes to a savings account — and could gain access to other data, by buying it or trading data with one another.

The challenge is how to do this ethically.

The challenge to making AI ethical

Public relations gaffes surrounding AI may become more frequent as financial institutions turn to the technology because financial institutions are often the last to know if their AI is biased, according to a recent report by the Model Risk Manager’s International Association (MRMIA). The professional organization submitted its report to the Office of the Comptroller of the Currency on May 25, in response to the request for information issued by the OCC and four other leading regulatory agencies about the use of AI in financial institutions.

“In our experience, we have never met a lender that intended to be biased in lending decisions,” the MRMIA stated. “The risk is of unintended bias, where a machine learning technique finds some unexpected and hard to explain patterns that give the lender an edge in pricing.”

Whether bias is intended is difficult to prove. But regardless of intent, its result can be discrimination. Outside of financial services, researchers have identified bias in facial recognition software and hiring software, for example. Within financial services, a November 2019 study titled “Consumer-Lending Discrimination in the FinTech Era” by University of California, Berkeley, found that “lenders charge otherwise-equivalent Latinx/African-American borrowers 7.9 (3.6) bps higher rates for purchase (refinance) mortgages, costing $765M yearly.”

The report goes on to state: “FinTechs fail to eliminate impermissible discrimination, possibly because algorithms extract rents in weaker competitive environments and/or profile borrowers on low-shopping behavior. Yet algorithmic lenders do reduce rate disparities by more than a third and show no discrimination in rejection rates.”

Ironically, what may hamstring some financial organizations from identifying bias are regulations that prohibit the collection of some protected class identifiers when making loans, the MRMIA said. This keeps lenders from performing tests to prove their models are unbiased or rebuilding them if they fail such a test, the organization claimed.

“How is a lender to know that use of a certain device [operating systems] combined with purchasing patterns highly correlates to a protected class? In fact, such things are generally discovered in the news, when borrowers discovered the bias that lenders could not see,” the MRMIA noted.

Creating ethical AI is largely about the data, said Svetlana Sicular, an AI analyst with the IT research firm, Gartner. Data can be biased despite the best intention. Ideally, Garg said, banks should be able to de-bias the AI by weighting or de-weighting certain issues in data that might make it biased. For instance, if only 2% of a bank’s customers are Hispanic because it ran a specific promotion, the bank might want to de-bias its data so decisions are not based on such a slim sliver of the Hispanic population. Ethical AI requires looking at existing data as well as the data that does not exist, experts said.

Sicular described an example of creating unintentional bias by omission: A group in Europe had created a product for altering the light in video. She asked how it worked on dark skin. “They looked at each other,” Sicular said. The group hadn’t considered this because they’d only thought about users within their own, mostly Caucasian, country.

Using AI for the unbanked

If the hurdles to ethical AI can be overcome, banks could use AI to create credit scores by grouping customers according to shared parameters – such as shared income level or housing stability. If most meet the description default on their payments, then the bank might assign a higher credit score. Basically, a bank could create an in-house scoring system that it keeps building based on feedback from the data about how the scoring performed, Garg said.

“It’s very easy when we talk about AI, like using a very simple machine — any model is possible,” Garg said. “When you go to more sophisticated models, it becomes very, very easy, and it all could be automated in the backend. … You don’t need any human underwriters anymore. Your AI is the underwriter now for any loan application.”

The ‘huge societal impact’ of alternative scoring

Banks are skittish about discussing how AI could be used for alternative credit scoring and decisioning.

Creating a way to help people get loans who might not typically be eligible is “a goal” said Stephen Thomas, executive director at the Analytics and AI Ecosystem in the Smith School of Business at Queen’s University in Ontario, Canada. Thomas offers an ethical AI certification for professionals, and works with banks such as the $926 billion Scotiabank to create ethical AI frameworks.

Banks have not yet deployed any such solutions, but “there are a number of exciting and promising research activities underway, at various stages,” Thomas told BAN.

Using simulations in the lab, the team at Queen’s University has been able to identify “tens of thousands of loans that were denied, but should not have been,” Thomas said. Alternative credit decisioning could have a huge societal impact, he added.

“There’s many ways you can look at this. There’s, from the bank’s point of view, lost revenue that they’re making or that they’re missing out on; from the individual’s point of view, it’s customer lifetime value,” Thomas said. “Then from a society point of view, this is spurring the economy, it’s giving solid, responsible people the boosts they need for that loan.”

Options for assessing the unbanked population

There are already some options available for reaching the unbanked, said Bhavani Palukuri, the San Francisco, Calif.-based Patelco Credit Union’s enterprise architect. The $8.8 billion credit union last month filed a statement in reply to the request for information on the usage of AI in credit decisioning issued by five federal regulators in March.

For example, alternative data, such as utility bills can be utilized for decisioning, Palukuri pointed out. Another AI solution even allows customers to opt into allowing a financial institution to leverage social media data for assessing risk, Palukuri said, adding, “Those are all the interesting things to look for. We are still in the research mode, we haven’t considered it — just in the research.”

Alternative data has worked for assessing credit risk in other countries, Garg noted. In Kenya, the $255k Sidian Bank uses Uber data on booking and customer satisfaction to decide whether to offer car loans to Uber drivers, he pointed out.

Such alternative data may open doors for the underbanked in the U.S. going forward. Pankaj Kulshreshtha, founder and CEO of the New York City-based credit decisioning platform Scienaptic, knows from experience how traditional credit scoring can work against customers. He was once denied credit because of card “overutilization.” Utilization tends to be one of the big drivers of risk, but Kulshreshtha’s financial institution did not consider that he only had one card and, every few months, he would book a vacation or buy something expensive with it, spiking his utilization.

“I’m actually a loyal customer, I have never defaulted,” Kulshreshtha said.

Scienaptic allows lenders to bring more data, such as utility bill data and LexisNexis data on liens and related judgments, into the credit decisioning process. It also uses different variables in its algorithms to score risk, Kulshreshtha said.

Most financial institutions have 40% to 50% approval ratings for loans, even though “only about 30% of Americans actually defaulted at any point in time, late on a particular payment,” Kulshreshth added. His customers are interested in boosting these approval rates, he said.

Considering the size of the unbanked and underbanked population in the U.S., those numbers beg a question.

“One of the things that banks and lenders should be asking is, ‘How are we actually increasing the size of the population that we are reaching?’” Kulshreshtha said.

Bank Automation News will host a webinar on automation technology for better risk management and security on Tuesday, June 15, at 11:30 a.m. ET. Register here.