OMG: How Texting Grammar Could Impact Your Creditworthiness
BankMobile looks to machine learning and alternative data sources to boost the bottom line.
Can consumers and regulators be convinced?
Imagine being approved for a loan based on which college you attended, your online shopping patterns, or even the punctuation quality of your instant messages. While seemingly far-fetched, credit approval enabled by artificial intelligence-machine learning (AI-ML) is all but upon us, with digital-first lenders like BankMobile at the forefront. At a time when AI-ML is hotly touted but far from ubiquitous within credit lending, the question remains whether lenders can satisfy customers, shareholders and regulators to achieve sustained competitive advantage.
With consumer debt in the US rising 5.1% to a record $3.83T in 2017 (and the mortgage crisis still ringing in the ears of creditors), finding ways to efficiently make sound lending decisions is key to success [1]. However, the traditional lending process leaves much to be desired. Lenders rely heavily on credit bureau data which systemically disadvantages so-called ‘thin file’ borrowers, such as young people, new immigrants and the self-employed. Other information is often left incomplete, incorrect, or even intentionally falsified by the borrower. Lending decisions require significant human judgement and can take upwards of weeks to process. And yet, default rates still exceed 9% in key categories such as student loans [2]. What’s more, a recent study showed that while 83% of borrowers have never defaulted on a loan, only 45% receive access to prime lending rates [3].
For BankMobile, an online-only retail bank whose median customer age is 28, the ability to serve a target market of borrowers with limited credit history has clear appeal [4]. Furthermore, with their parent company, Customers Bank, reporting a loan loss allowance of $38M on a net income of $79M in 2017, selecting the right candidates to lend to could drive significant impact on the bottom line [5].
Their solution? Automate the lending process using machine-learning algorithms and non-traditional, personal customer data.
In 2017, BankMobile became the first bank to partner with AI-ML peer lending service, Upstart. The service leverages proprietary algorithms to determine creditworthiness and pricing based not only on an applicants’ credit score, but other factors including their highest level of education, school name, area of study and even how they interact with the application [6]. The methodology, created in collaboration with the Consumer Financial Protection Bureau (CFPB) allows BankMobile to test changes to the algorithms, and analyze their ability to maintain fair lending standards as defined by the regulator [3].
So far, it looks to be working.
Presently, customers can receive personal loans in as little 24 hours, with more than half of all applications assessed fully automatically. Furthermore, a recent study showed that, for a given loan approval rate, the Upstart service experienced 75% fewer defaults versus a peer group of large US banks [3].
While BankMobile’s medium term lending strategy remains largely undisclosed, plans do exist to expand the model beyond personal loans to credit cards, student loan refinancing, home equity and car loans in the coming years [6]. In parallel, partner Upstart continues to refine its underlying algorithms with the ultimate goal of achieving full loan process automation for 80-90% of deals [7].
While their FinTech partnership has enabled BankMobile to swiftly storm the market, addressing key risks in the near term is critical. Firstly, the bank should look to best practices to control possible algorithmic biases, namely by working to closely understand the limitations of the algorithms and assisting Upstart data scientists to shape data samples accordingly [8]. BankMobile employees should be trained in the lending process at a high-level and capable of facilitating warm handovers to the servicer as necessary. Further, at minimum, the bank must reduce its reliance on Upstart, implementing systems to back-up crediting arrangements in the event of partner issues [9]. The bank should seek to remove any variables with ‘gameable’ influence on the algorithms predictions, and closely monitor the system’s performance, which is yet to be tested through a full economic cycle.
Longer term, BankMobile may consider bringing or building its AM-ML lending capability in-house to enable rapid prototyping and to own the customer relationship end-to-end [10]. Such a move would likely command the creation of cross-functional teams anchored by top developer talent. Additionally, the bank may seek to build a more nuanced view of creditworthiness by incorporating other alternative data sources such as customer bill payment history or online shopping patterns. Lastly, given the sensitive nature of customer information and the fact that 1 in 3 focused cyber-attacks results in a security breach, investing in cyber security governance, cyber resilience and cyber response readiness is imperative [11].
Beyond these operational challenges, key questions still loom. For consumers, how palatable is a ‘Netflix-esque’ loan pricing and approval process, and what might lenders need to provide to earn their trust? For BankMobile specifically, should they insource their strategic AI-ML lending capability, and if so, when?
(Word count: 782)
Sources:
[1] Federal Reserve. “Consumer credit increased at a seasonally adjusted annual rate of 5-1/4 percent during the third quarter.”, press release, November 7, 2018. Federal Reserve website, https://www.federalreserve.gov/releases/g19/current/, accessed November 2018.
[2] Vasily Souzdenkov, “States That Don’t Pay Their Bills on Time”, Credit Sesame Blog, May 31, 2018, https://www.creditsesame.com/blog/debt/states-that-dont-pay-bill-on-time/, accessed November 2018.
[3] Upstart. “Results to date” https://www.upstart.com/about#result-to-date-1b, accessed November 2018
[4] Bank on It, “Episode 128 (2 of 4) Recorded live from Lendit USA 2018”, April 24, 2018, podcast, http://bankonitpodcast.com/episode-128-2-of-4-recorded-live-from-lendit-usa-2018, accessed November 2018.
[5] Global Newswire. “Customers Bancorp Reports First Quarter 2018 Net Income of $20.5 Million; Diluted EPS of $0.64”, press release, April 30, 2018. Global Newswire website, https://globenewswire.com/news-release/2018/04/30/1489797/0/en/Customers-Bancorp-Reports-First-Quarter-2018-Net-Income-of-20-5-Million-Diluted-EPS-of-0-64.html, accessed November 2018.
[6] American Banker. “BankMobile Deploys AI Alternative Data to Lend to FICO Poor Students”, https://www.americanbanker.com/news/bankmobile-deploys-ai-alternative-data-to-lend-to-fico-poor-students, accessed November 2018.
[7] Benzinga. “Upstart, the brainchild of Google execs is using Fintech to make smarter loans”, https://www.benzinga.com/fintech/17/09/10040990/upstart-the-brainchild-of-google-execs-is-using-fintech-to-make-smarter-loans, accessed November 2018.
[8] McKinsey. “Controlling machine learning algorithms and their biases”, https://www.mckinsey.com/business-functions/risk/our-insights/controlling-machine-learning-algorithms-and-their-biases, accessed November 2018.
[9] U.S. Department of the Treasury, “Opportunities and Challenges in Online Marketplace Lending”, Treasury Connect Blog, May 10, 2016, https://www.treasury.gov/connect/blog/Documents/Opportunities_and_Challenges_in_Online_Marketplace_Lending_white_paper.pdf, accessed November 2018.
[10] The Boston Consulting Group. “Digital Transformation”, https://www.bcg.com/en-us/capabilities/technology-digital/digital.aspx, accessed November 2018.
[11] Accenture. “Building Confidence Facing the Cybersecurity Conundrum”, https://www.accenture.com/t20170406T052041Z__w__/sa-en/_acnmedia/PDF-35/Accenture-Building-Confidence-Facing-Cybersecurity-Conundrum-Transcript.pdf, accessed November 2018.
Very interesting article N. Fleming. Having gone through the borrowing process I find it incredible that this company can process loans in 24 hours while still reducing defaults by almost 75%. I’m curious as to what data sources have the highest correlation with default rates/ ability and willingness to repay. It will be interesting to see what data consumers are willing to share in order for lending decisions to be made.
I agree with your assertion that the algorithm will not fully be tested until we see how it performs through each phase of the economic cycle. On the other hand, I’m unsure about whether BankMobile should replicate Upstarts core competency in house. Why not leverage Upstart’s expertise while focusing on growing your business and extending your product offerings?
In response to the author’s question about what it will take for consumers to find BankMobile’s model palatable, my answer is very little. I believe that the company has already put itself in an attractive position simply by targeting younger borrowers with limited credit histories. This group is one with limited alternatives and less concerns about giving company’s access to more of their information, making them prime for acquisition.
I think the real stakeholder BankMobile should be concerned about is regulatory bodies. I worry that the variables they have added into their credit formula (e.g. highest level of education, school name) could expose them to discrimination claims and regulatory action. We saw in the Aspiring Minds case that ML had systematically discriminated against female candidates, and I worry that something like that could happen here. As such, BankMobile should continue to work very closely with the CFPB to assess their algorithm.
N, this is a great article and was well-worth the read. As I read, I couldn’t help but think how machine learning is enabling modern lenders to be more like the lenders of the past in making loan decisions. From the Handelsbanken case, we learned how the local banker would only make positive lending decisions to people that he/she personally knew would pay the loan back. I would wager that much of the lending choice boiled down to where the lessee went to school, whether or not they were upstanding citizens in the town, and the personal relationship between both parties. As banks continue to scale, and the competition to reduce costs drives them to reduce their brick and mortar footprint in towns (thus reducing their ability to personally know their customer), I can see how the banks would use the data outlined above to make more informed lending decisions.
Thank you for such an interesting article! I was very shocked at first to learn that an algorithm to enable a fully-automated (for over 50% of loan applicants) loan application process was created in partnership with the CFPB, a group formed specifically to protect retail investors and consumers from risky business practices, given my initial reaction that removing all face-to-face interaction between lender and potential borrower would possibly lead down a dark path towards 2008. But when I read the statistic you included that “for a given loan approval rate, the Upstart service experienced 75% fewer defaults versus a peer group of large US banks,” my initial hypothesis was debunked as this suggests that the machine learning tool they have developed is more successful at evaluating business risks based on the wealth of available information and the algorithm’s associations. I would be very interested to know exactly what information is included in the scope of the algorithm’s analysis and whether the person applying for the loan is aware themselves of the full scope, or if data is scraped from public areas of the internet without being supplied through the applicant directly.