We’re back with another installment of IF:Then, a weekly newsletter about legal strategy at the intersection of law and technology. To get this sweet content straight to your inbox each week, go ahead and subscribe below:
My mailbox inspired this post. I received all of these offers on the same day, and I would say this is fairly typical. Either there’s money to be made in personal lending, or I’m just really great at getting pre-approved.
With so many options, how can one decide? What if anything differentiates these lenders from any other fintech, or god forbid an incumbent bank? Below, we’ll look at the business model for personal lenders, how new entrants approach coming into a very large, but crowded market, and take a deep look at a company with the Underwrite Stuff: space explorers of the universe beyond the credit score. But as usual dangers lie ahead, and users of alternative data in underwriting need to grapple with existing consumer protection regulations and have a clear understanding of their own models.
A lending business is about cost of capital and risk management. At a high level virtually every lending business fundamentally works the same way: the lender charges the borrower an interest rate that makes the revenue from the loan higher than it cost to acquire the capital and the customer. The key variable then is loss — what percentage of borrowers will default on the loans?
Lenders generally underwrite that risk on two high-level metrics: ability to pay and willingness to pay. Ability to pay is usually broadly measured by your income or your debt to income ratio, and is sort of a threshold decision -- does this borrower have or make enough money to afford these payments? Willingness to pay is fuzzier -- what is the borrower’s past history with credit? Do they tend to make payments late? What other factors can I use to evaluate the borrower? AM I GOING TO GET SCREWED HERE!?
Where this traditionally ends up is the FICO score - a byzantine mathematical formula generally designed to assess a borrower’s willingness to pay, geared towards the risk of default in the next 12 months. It is also tracked by three separate credit reporting agencies in similar, but slightly (and irritatingly) different ways.
The lenders ability to nail this risk evaluation in underwriting is key to the business — maximize your profit by making as many loans as possible within an acceptable, anticipated range of defaults. Prove out your model to lending partners and capital providers and your costs decrease as well.
Enjoying IF:Then? Why not spread the wealth and spread that joy to others in your network?
A Lender Has Entered the Chat
Consumer lenders come in a few shapes and sizes. Some make loans off their balance sheet, while others have raised and are lending debt capital. Retail banks use customer deposits to fund their loan books, and many of the fintechs are actually marketplaces or platforms that simply acquire the customer, process the underwriting, and service the loan for a partner bank actually doing the lending. A growing startup likely progresses from balance sheet —> to debt facility —> to loan servicing in order to prove out its model. But in all scenarios, a new lender needs a capital, state licensing, and some sort of hook.
Rohit Mittal, the founder and CEO of Stilt, a lending platform for immigrants, wrote about 5 key areas he believes make up the framework for building any credit startup. (Disclosure: I made a small investment in Stilt through a syndicate).
Insight
Data
Risk Model
Economics
Scale
Rohit makes the point that the insight is the most relevant step, and everything on some level flows from that:
This is the most important step before launching a credit startup. Clearly define what is missing or wrong with how things work right now. This insight will be the foundation for every decision you make and every action you take.
Examples: Stilt started with the insight that immigrants are mispriced by the traditional financial system. They are mispriced because not having a credit history is considered a negative. It shouldn’t be. Immigrants are financially responsible. That’s why they deserve affordable and low cost credit products.
Of course to deliver value based on any given insight, it needs to manifest itself in some form of data, you need to be able to leverage data, and then you need to feed that data into a risk model.
if we believe immigrants are lower risk, then we need alternative data to prove their low risk. Education, employment, and international data sources exist that can be used to underwrite immigrants.
Stilt bucks the traditional model because credit score is a broad brush based in large part around credit history in the US. That makes it a particularly poor evaluator of immigrants, and any lender focusing on that as a key metric is likely to leave the opportunity to make loans to good borrowers on the table. Thus, Stilt leans on alternative data like education and employment that it believes can more accurately underwrite this segment of borrowers.
Re-pricing a previously mis-priced segment isn’t new or unique to Stilt. SoFi got its start by recognizing that a certain segment of graduate school students were entering the workforce with massive debt and limited credit history, and were being priced essentially as high-risk borrowers despite stable jobs with six-figure salaries. Recognizing that the risk of default for this cohort was extremely low, and refinancing their loans accordingly seems obvious in retrospect — this was hardly democratization, these were essentially prime or super-prime borrowers — but no lender was underwriting that way at the time. Like with Stilt, the main drivers of their underwriting process were alternative data sources — education and employment. But alternative data can go far beyond such high-level metrics and can come from a variety of sources.
Any data that you can build into a model and evaluate at scale can theoretically serve as alternative data. Simply looking at a users cash flow or payroll data can be much more informative than a credit score. Looking into payment history with rent or utilities — typically not reported to credit bureaus — can be instructive. Getting an idea on the borrower’s use of funds is informative. Get sufficient data and put it into a machine learning model and get deeper insights on your customers.
Let’s make up some alternative data: iPhone users are less likely to default; cord-cutters are more likely to refinance early; anyone who baked a loaf of bread early in the pandemic has since lost their optimistic spirit. With a large enough data set you can probably glean some insights from any of these.
All that said, most lenders today still focus on credit score and income to build their portfolios. While more and more startups will take a different approach, there are only so many high-level insights to glean from simply using education and employment data along the lines of Stilt and SoFi. But as financial data and consumer insights become increasingly portable thanks open banking initiatives, APIs, and services like Plaid or Pinwheel, granular insights from alternative data becomes more likely to play a central role in underwriting across all providers. Using more sources of data and finding new ways to analyze that data opens the market to more consumers, and allows alternative lenders to reach underserved segments.
If you’re interested in joining IF:Then’s community of attorneys, regulatory experts, and legal strategists, and investors, please email me at david@ifthen.vc.
Regulatory Roundup
Whether their underwriting sources are traditional or alternative, lenders still need to follow regulatory parameters. To date, public conversation around use of alternative data has revolved around its adherence to federal lending laws, including the Truth in Lending Act (“TILA”), the Equal Credit Opportunity Act (“ECOA”), and the Fair Credit Reporting Act (“FCRA”) amongst other rules and regulations.
TILA generally defines what information must be disclosed to borrowers, while FCRA is focused on how credit information is collected and used, and ensuring its accuracy and availability to borrowers. Where users of alternative data can get tripped up is regarding fair lending and the Equal Credit Opportunity Act. The ECOA states that lenders cannot make credit decisions based on non-financial factors — particularly protected designations such as race, religion, age, or marital status.
Certainly we’ve seen banks and lenders get dinged for ECOA violations, and probably should see more. The US financial system’s extensive history of “redlining” is a fairly overt and intentional example. But uses of alternative data can take less nefarious, but equally harmful routes to inequality if providers aren’t careful. One company using AI in its underwriting model wisely got ahead of the conversation.
Start Me Up - Building and Iterating with Alternative Data
Upstart is an alternative-data focused platform formed in 2014 that offers personal loans to consumers and provides its “automated borrowing technology” to its lending partners. In 2017, Upstart applied for a no-action letter from the Consumer Financial Protection Bureau (“CFPB”) regarding its alternative data underwriting model that touted its “ability to to identify differences in risk between ‘thin file’ applicants.” Upstarts model purported to allow the company to “offer credit to segments of the population with limited credit or work history at more favorable rates.”
“By relying exclusively on the credit report and traditional modeling techniques, lenders ignore some of the most predictive information about potential borrowers. . . . By complementing (not replacing) traditional underwriting signals with others that are correlated with financial capacity as well as propensity to repay a loan, Upstart’s underwriting properly understands and quantifies risk associated with all borrowers—those with credit history, and those without.” - Upstart CFPB NAL Application 2017-09 (emphasis mine).
The application gives some limited insight into Upstart’s model at the time. A borrower had to pass a set of eligibility requirements, including a “near-prime” credit score above 620 (or if no credit score, an educational degree more advanced than a high-school diploma), no existing past due or delinquent debts, and a debt to income ratio above 45%. Upstart would then engage its model to make a decision based on some proprietary “financial indicators,” and like Stilt and SoFi, looked closely at education and employment.
Upstart may use an applicant’s educational information including, but not limited to, the school attended and degree obtained, and their current employment, to develop a statistical model of the borrower’s financial capacity and personal propensity to repay. This information, along with the applicant’s financial and credit-related variables, is used to determine the applicant’s loan terms.
Upstart was careful to note that its decisions are based on a proprietary combination and weighting of all inputs, and thus its model could only be evaluated with each variable in concert as opposed to individually. Perhaps a convenient way to avoid scrutiny for any given input, but a holistic view is probably in everyone’s best interest.
A key aspect of the no-action letter was its emphasis on ongoing monitoring and compliance with ECOA. Obviously, using protected class inputs like a “higher APR for borrowers under 30” would be an ECOA violation, but it’s not enough to avoid those direct discriminatory practices. As a general course, lenders need to avoid metrics that result in a disparate impact to protected communities. Underwriters building models from already-biased data sources derived from decades of systemic inequality are not going to produce equitable results. Understanding and interpreting the context around the data is key, and any responsible lender will need to put together a program to strip the data of preexisting bias, especially as their modeling techniques become more advanced and scalable.
By late 2020, Upstart had iterated and versioned its model several times. What was a proprietary “statistical model” in 2017 had become an AI and machine learning-driven underwriting platform. In 2017 all Upstart loans were funded by Cross-River Bank, but by 2020 Upstart loans were powered by 10 funding partners. In 2020, loans could flow through Upstart.com, or through a white-label program wherein a partner could use Upstarts AI underwriting under its own branding and marketing. But the biggest difference was clearly that Upstart collected enough data to upgrade and scale their model. Accordingly, they applied for and were granted another no-action letter from the CFPB:
“Upstart uses AI techniques and alternative data to improve underwriting accuracy and outcomes. By combining a more robust utilization of the data in credit files, alternative data points such as education and employment history, and modern AI techniques, Upstart believes it has developed a model that is more predictive of credit performance than a traditional model. As a result, the Platform enables Partners to offer credit to segments of the population with limited credit or work history at lower interest rates.” - Upstart CFPB no-action letter application 2020-11
The only clearly identified alternative data is once again education and employment — same sources used by the old model, and same sources used by Stilt and SoFi. No mention of cashflow data, payroll data, consumer behaviors, or other “financial indicators” other than the use of AI, which is clearly the focus this time around.
Risky Business
The 2020 no-action letter eventually lays its reason for being, identifying three key risks with the model, which remain fairly universal:
The model may deny protected class applicants at rates that are higher than non-protected class applicants
The model may result in the origination of loans to protected class applicants at prices that are higher than non-protected class applicants
The model may fail to adequately predict the creditworthiness of borrowers due to the design of the model or programming
This essentially brings us back to disparate impact. Upstart doesn’t really have a direct answer as to how it will ensure that its model does not inadvertently target protected classes, only that it will engage in extensive monitoring, institute an oversight procedure (I interpret this as manual review), and work with the CFPB on compliance protocols. In its S-1, Upstart lists “regulatory compliance” as a “competitive strength” and delivered a similar message:
We have worked with regulators since our inception to ensure we operate in compliance with applicable laws and regulations. AI lending expands access to affordable credit by constantly finding new ways to identify qualified borrowers, yet AI models must avoid unlawful disparate impact or statistical bias that would be harmful to protected groups. We have demonstrated to the CFPB that our platform does not introduce unlawful bias to the credit decision, and we have developed sophisticated reporting procedures to ensure future versions of the model remain fair. Upstart Form S-1 Registration Statement, Nov 2020.
Staying on the right side of the regulators is probably a good idea. The CFPB and FTC is clearly on the lookout for abuse of these types of systems, and while we haven’t seen a rash of enforcement, the FTC has in at least one instance, shown a willingness to require a business misusing AI to delete its algorithms and data.
In all likelihood, the impact around alternative data in underwriting remains to be seen, and the process needs to play itself out. It seems clear that simply relying on credit score and income isn’t enough. On aggregate, using alternative data seems to produce better outcomes for lenders, and more loans issued at better rates to borrowers. But will other sources of alternative data beyond education and employment make more inroads? Given increasing access to cash-flow and payroll data, will that play a more central role? Will other aspects of consumer behavior, either before or during the life of a loan start to affect access?
Upstart’s model still feels like the beginning. In the future it seems unlikely that any lender will make a decision based solely on the credit score. But which inputs prove to be effective remains to be seen. In meantime, I’ll keep sorting through my mail to stay up to date, and IF:Then will be here to help.
Until next week friends - David Ikenna Adams
Twitter | LinkedIn | Email | ifthen.vc
If you liked this weeks’ edition of IF:Then, go ahead share this with a friend! They’ll thank you later, and I’m thanking you now.