Business Intelligence, Business Woes: Why Machine Learning is Causing Regulatory Concern

Toys ‘R’ Us plans to shutter all 200 of its physical stores after filing for bankruptcy, an indication to retailers of the changing landscape of consumers increasingly opting for online shopping and digital platforms.

This shifting consumer mindset is not just affecting e-commerce, but also the auto finance industry — which appears to be on the cusp of a revolution as it relates to digital car buying, vehicle ownership, and automation. It has long been said by auto finance executives, including at last month’s Auto Finance Innovation 2018, that if lenders are not rethinking their business models — or at least considering innovation — they could lose marketshare. This notion particularly held true with regard to machine learning, a hot-button topic at the event.

Machine learning — a type of artificial intelligence (AI) that gives computers the ability to learn without manual input — has captured the
industry’s attention not just for underwriting purposes, but for automating customer service.

However, because the technology is so new, some lenders are wary — particularly of its regulatory implications.

There are no international regulatory standards for machine learning, and data on the growing usage of AI is largely unavailable, leaving regulators unsure about its impact on the industry, according to a report by the global Financial Stability Board (FSB), an international body that monitors and makes recommendations about the global financial system.

The FSB left open in the report whether regulation is needed, but added that “developments in this area should be monitored closely,” an idea echoed by several auto lenders.

While machine learning can be utilized for credit decisioning to potentially reduce losses, if a lender doesn’t have the manpower — i.e., a team of data scientists — or the funding to outsource this technology to a third-party provider, it is a difficult beast to tame.

However, there are other areas within the auto finance sector where lenders can utilize more of these “deep-learning tools,” Sang Kim, vice president of risk at Consumer Portfolio Services, told Auto Finance News. “Machine learning doesn’t have to be used for [credit] decisions, but for more business process improvements — which is an area where we can utilize those techniques more.”

Exploring the Use-Cases

While much of the hype of machine learning has centered on the ability to assess credit quality and more accurately predict risk, there are many other use-cases for AI.

Daimler Financial Services, for example, is combining artificial and emotional intelligence (EI) to create a new mobility experience — the digital voice assistant called “Sarah,” the captive announced in March. Sarah is a proof of concept Daimler Financial has been developing since 2017 with its partner Soul Machines, an AI solutions provider based in Auckland, New Zealand.

“The project is in the early stage of development,” Udo Neumann, Daimler Financial’s global chief information officer, told AFN. “Emotional intelligence is one of many ways Daimler Financial Services is exploring technology to further improve customer experience and to tailor information customers are seeking for their specific needs.”

Sarah will be able to assist customers like a personal concierge, with the goal to offer the right information at the right time with an emotionally intelligent, digital touch point. Sarah can be programmed with knowledge about leasing options or the latest Mercedes models, for example. Sarah can see, hear, and recognize human emotions and non-verbal actions in an interaction, such as a head nod.

“Daimler Financial Services is the first captive automotive finance organization to attempt to redefine the customer experience with digital humans using data-driven insights, AI, and EI,” Neumann said. “Even though we are in the proof-of-concept stage of combining these technologies, it is clear that this powerful combination of technology will be a gamechanger in the industry in the near future.”

Other renditions of this customer service AI concept are being tested in the market, such as through voice assistants. SpringboardAuto, which has integrated its platform with Amazon’s Alexa, offered a demo at Auto Finance Innovation 2018 showing how consumers can shop for a car and obtain financing all through the use of the Amazon Echo Dot.

Another use-case includes A/B testing. Ally Financial Inc., for example, struck an interest in machine learning and is in discussions with an undisclosed company to deploy this form of artificial intelligence to optimize its web pages, Jennifer Heil, executive director of Ally unit Clearlane, said at Auto Finance Innovation 2018.

A/B testing is a method of using statistical analysis to compare two or more versions of an app or webpage to determine which one performs better for a given conversion goal.

Heil did not offer specifics on the machine learning initiatives at Ally. Typically, uses include running AI in the background to automatically recommend the best experience for each user.

“Our tech [team] has been looking at [machine learning], and we’re actually partnering with a company — though we don’t have a contract yet — to deploy that for our A/B testing so we [can go] online and look at what’s happening in the [direct lending] funnel,” she said. There is a “learning curve” to using machine learning techniques, but Ally is “exploring those types of processes when engaging with the customers,” she added.

Machine learning can also be used to automate daily tasks in an effort to streamline processes, which is an avenue Global Lending Services LLC is exploring.

“Automation is not really about [being] a cost-savings play, it’s about speed, consistency, scalability, and compliance,” Senior Vice President of Information Technology Andy Clements told AFN.

The company is looking to better leverage machine learning to take categorization tasks humans do and provide dealers faster turn times. For example, the lender already has automation in place to detect missing pieces in auto loan applications and to catch “certain types of weird paystubs,” Clements said. “We already have quite a bit of automation on the funding side, but we are looking to continue to push the boundaries — not with a goal of removing the human, but with the goal of focusing humans on where we need them most,” such as customer service and other areas of the operation.

“In 2019, what does machine learning mean for us?” Clements asked the attendees at Auto Finance Innovation. “One [use-case] is classification of the present. Meaning, how do we use artificial
intelligence to take away some of the work our teams do that are really routine like the things that aren’t fun or sexy? Are the contracts signed? Do the names match? Being able to use artificial intelligence to remove some of that tedious work from our workforce … can improve compliance and scalability.”

AI in Underwriting

Several lenders, including Ford Motor Credit Co., have used machine learning in the past couple years to bolster underwriting.

For underwriting, computer algorithms can use nontraditional data points to determine risk and creditworthiness. For example, machine learning can assess data points, such as whether applicants supplied the same cell phone number on previous loan applications and whether they have occupational licenses. Because the platform is capable of “learning” over time, it can propose changes to variables as patterns evolve or emerge.

Last August, Ford Credit explored changes to its underwriting approval process that included the incorporation of machine learning to look beyond credit scores. The change stemmed from a study of Ford Credit customers, conducted by fintech startup ZestFinance, that measured the effectiveness of machine learning to better predict risk.

In January, powersports lease provider MotoLease LLC improved the accuracy of its internal credit forecasting model, called M-Score 2.0, by using machine learning and alternative data.

“We used machine learning techniques so that our score is not a static score, but it changes and improves based on portfolio performance,” said MotoLease’s Managing Partner Emre Ucer. “For that reason, the importance of Fico scores went down because of everything else we put into this ‘secret sauce.’”

MotoLease has been reviewing its scoring in the past year, and conducted an “extensive” study with TransUnion analyzing almost 50,000 consumer files, Ucer said. “We came up with a very sophisticated algorithm” to underwrite subprime customers differently from prime or near-prime customers, he added. For thinfile or lower credit consumers, MotoLease reviews the consumer’s payment history, past delinquent accounts, and other payment patterns from the past 82 months.

However, there are several core challenges as it relates to using machine learning in underwriting.

“In terms of the originations scorecard, we have to be cautious due to regulatory reasons,” Consumer Portfolio Services’ Kim said, adding that while CPS does not use this AI, she has used it at several other companies. “We use the originations scorecard for every application decision we make, and we have to have a very clear, transparent method. When you utilize something like machine learning, it’s a black box.” Machine learning has been touted for its ability to allow lenders to better monitor disparate impact and change the model to quickly limit unintentional discrimination of protected classes. However, the uncertainty of the algorithm could pose as a roadblock, according to Kim.

“The more data I have, the more I learn about you — usually that’s the case,” Kim said. “What happens is that I really don’t know what this computer is doing. For example, let’s say I insert 100,000 data points into this machine learning system, and it gives me answer A. Then I give the machine 1 million data points, and … now it is answer Z. The answers are totally different because now the machine has more data points to make the decision. The ending result: I could tell the second model works so much better, but the thing is that I don’t know how that machine got to point Z.”

For regulatory reasons, it’s important to be “very careful using that type of machine learning in your decision making process,” Kim said.

That being said, regulators tend to move slower than technology, said Eric Hathaway, vice president of marketing at Zoot Enterprises, an automated credit decisioning platform.

“The benefits of machine learning are that there will be an electronic trail of the transaction,” Hathaway said. “The issue comes in whether the regulators will accept these changes [in] the reduction in human interaction. But more relevant is the speed in which changes can occur in the decision process by using machine learning and AI.”

Overall, knowing when a lender can and cannot use machine learning from a compliance and regulatory perspective is important, Kim said. “Then, if there is an area that you can be more research-oriented, you can play around with a new model or tool,” she added.

The Human Touch

Despite the “black box” effect that machine learning creates, lenders can effectively navigate this hurdle with a team of data scientists monitoring and knowing how the system works, some said.

“We are entering a big discussion in this country about algorithmic accountability and algorithmic transparency, because these algorithms are proliferating at higher and higher stake use-cases,” Kareem Saleh, executive vice president at ZestFinance, told AFN. ZestFinance, founded in 2009, applies big data and machine learning to credit underwriting.

“The math is just one part of [what we’ve built], but it’s also about opening up these black-box algorithms,” he said. “This technology that we have built is not only super-predictive, but its reasoning is really transparent. We can tell you what variables the model is taking into account and to what extent.”

However, “there still needs to be a human in the loop evaluating those variables from a compliance perspective to get everybody comfortable,” he added. Notwithstanding all the leadership and regulatory changes at the Consumer Financial Protection Bureau, “there are a bunch of other regulators who care about fair lending issues … so I think anybody who is popping champagne corks because they think fair lending enforcement is done for, is premature.”

Currently, 99% of machine learning is based on human input, Zoot’s Hathaway said. “We aren’t at the point of true AI in the enterprise or society quite yet. Even in the advanced stages of AI today we continue to see errors, biases, and indecision and/or lack of ability.”

Data scientists and many other personnel are going to be needed to implement a true AI model within large organizations. “AI doesn’t just create opportunity and benefits, it also opens up new channels for fraudulent activities as well,” Hathaway added. “They will need to have strategy, manpower, expertise, and teams to implement correctly.”

World Omni Financial Corp., for example, is partnering with universities in Florida and utilizing Ph.D. candidates as part of its overall focus on data integration and utilizing analytics, Alfredo Cateriano, vice president of risk management, data, and analytics, told AFN in May 2017.

“We’re partnering with scientists, statisticians, econometricians, and operations research,” he said at the time. “But, also, we are really using this to bring individuals with expertise in machine learning.”

Through the program, students will utilize World Omni’s data to explore the use of machine learning as well as data mining, though it’s unclear in what ways the lender will utilize machine learning techniques. World Omni declined to comment for this story.

“Outsourcing is a viable option if you don’t have the expertise in-house,” Global Lending Services’ Clements said. “One of the pitfalls of machine learning is you can throw data at it, but if you don’t understand what you are throwing at it, then you will not understand the output either. The debate we have is, ‘How do we push the boundaries of machine learning but make sure we hold true to our expertise and learnings developed over lifetimes of data analysis and stats?’ I think we are starting to find ways to leverage more and more machine learning, but you have to be careful how you do it because of the pitfalls of understanding your data.”