Leading with Context: Credit in the Digital Age
Leading with Context: Credit in the Digital Age
Fooling FICO, Contextual Credit, and a Call for a Better Bureau
At Nyca, we get quite excited about credit. Extension of credit and the magic of leverage helped catapult the American economy to be the largest in the world, and the U.S. system has been quite effective for decades. However, like many areas of financial services, credit is undergoing a disruptive period. The availability of new and diverse data coupled with evolving modeling techniques and new forms of distribution have opened up an exciting new era. The following is our second installment of a three-part credit-focused series, “Fooling FICO, Contextual Credit, and a Call for a Better Bureau.” The first post can be found here.
Thank you to Max Liebeskind for providing insightful commentary and editorial skill!
As detailed in the first installment of this series, FICO Score 8 has seen better days. An entire industry has been built around either coaching how to improve one's FICO score or providing data that, while potentially misleading, is favorably interpreted by the Score 8 algorithm. The degradation of some bureau data combined with the rise in access to credit outside the bank branch has led us to propose a permanent change to how consumer credit is assessed. Rather than relying on a single score, lenders should utilize contextual, less manipulable data to more accurately assess a borrower for a given loan.
The rise of contextual underwriting
Product-specific underwriting is not a new idea for lenders. Even Fair Isaacs (the company that created the FICO Score) saw so much value in the concept, it launched an “auto FICO” score years ago, as car financing is quite distinct from applying for a credit card. The reason one rarely sees a “lending supermarket” is that specialization brings better results. Exceptional subprime credit card underwriters like Capital One, CardWorks, etc. have no interest in being a mortgage lender—in fact, Capital One sends their customers to Loan Depot for home loans. Top-performing banks often specialize in certain credit products because they are experienced practitioners in a given offering.
What if underwriting were less product-specific and more context-specific? And what if rather than relying on FICO-related data, a lender used more diverse data sets depending on context?
Context can be interpreted broadly. While the following is far from exhaustive, here’s how I’ve thought about the scope of lending with context, in no particular order:
Purpose: why does the consumer need credit?
Magnitude: how much is being borrowed relative to the individual’s income and ability to pay?
Term: Short-term (days to months) and long-term (years) credit are fundamentally different risks.
Features: does the loan amortize or revolve?
Security: is the loan secured by an asset?
Source: how and why did the borrower apply for credit from the lender? (in-store, bank branch, auto dealership, online, etc.)
Behavior: is the request for credit a considered action versus an emotional reaction?
Hierarchy: Where does repayment fit into the priority of payment of a borrower’s monthly financial needs?
For each aspect of context, one could focus on specific data that would be relevant and useful. The combination of this data would provide the building blocks of modern, forward-looking consumer underwriting. A borrower can be scored on a contextual basis and be offered products s/he may not qualify for on a traditional FICO basis, thereby increasing credit availability on an intended use basis.
Contextual Usage
Contextual underwriting is in widespread use today by sophisticated lenders, principally as supplementary data to traditional bureau data analysis. Below I’ve described some active use cases and dynamics I find interesting and impactful to providing thoughtful credit to consumers.
Purpose: As I laid out in my commentary on BNPL, a lender could make more informed credit decisions by knowing what is being purchased from which merchant. Mattresses perform better than electronics. In the personal loan space, home improvement has historically outperformed other use cases, especially wedding financing.
Magnitude: Ability-to-pay (ATP) has become an important metric for lenders, especially due to the CFPB’s emphasis on affordability. However, ATP can be manipulated by extending loan term. Debt-to-income ratios have historically been used to assess indebtedness but leave much to be desired. An example of contextual extrapolation could be a lender who looks at the payment amount relative to the consumer’s monthly excess cash flow (after rent, utilities, debt, food, and other mandatory monthly costs).
Term: More credit should be available on a short-term basis. If I lend a dollar to anyone (meaning the US adult population) today, that individual is extremely likely to be able to repay a dollar tomorrow. If I lend that same dollar for a week, there is a sliver of the population that may not be able to pay on that timeline, but it’s once again close to the whole population. A month out, the likelihood of repayment hasn’t really diminished. Six months incorporates incremental risk, 12 months even more. As we go out 2, 3, 5 years, the probability of repayment of that dollar diminishes. Thus, over time, the likelihood of default increases in a non-linear fashion.
Features: Loan flexibility can increase or decrease risk depending on use case and magnitude. Short repayment schedules and frequent availability permit greater access to credit. Revolving instruments aren’t all bad; lenders like Upgrade have incorporated stricter amortization of purchases on top of a revolving card product. How a borrower interacts with the card may guide the lender’s ability to expand credit access to certain individuals or profiles.
Security: Secured lending is the biggest opportunity in consumer lending today. Traditional secured lending - mortgages and auto loans - are relatively cheap sources of debt for individuals. What if low-rate debt were available on a broader basis? Many wealth platforms offer loans backed by public securities. Wealthy individuals borrow against land, planes, boats, and more. I was a co-founder of Aven, a credit card secured by one's home. For those with excess equity in homes, cars, or other stable assets, stronger repayment brings material cost of debt savings.
Source: In the early days of Lending Club, there were three main channels to source loan applicants: direct mail, Credit Karma, and SEO/SEM. For equivalent bureau files, we would see material performance differences between targeted acquisition (direct mail, Credit Karma) and web search-driven traffic. The applicants who googled “need money now” never performed well as a group! Moral of the story: sourcing and interaction really matters. Targeted, proactive offers traditionally are superior performers.
Behavior: How does an applicant interact with the application or process? How many times does the applicant revise his/her stated income? Are there intentional or unintentional inaccuracies in the application? One data verification vendor to lenders has uncovered a high correlation between loan performance and the quality of scans for uploaded documents (bank statements, W-2s). For equivalent bureau files, a borrower who takes the time and effort to deliver clean and complete application documentation simply performs better. Behavior is a very broad category and critical to expanded contextual underwriting.
Hierarchy: If one were to start with the hierarchy of financial payments, the 2008 financial crisis taught us that auto loans may be paid off even before mortgages. A lender must have a good idea on where it stands relative to consumer priorities. As we inevitably enter more difficult economic times due to the Fed’s aggressive rate hikes, if the consumer is short on his/her financial obligations for a month, which types of debt are consistently deprioritized?
Data!
Most services and lenders furnishing new credit data to the bureaus are providing information that could be compelling for underwriting purposes, but the existing data schema is ineffective for extrapolation. So, what happens?
Non-conforming data is being housed separately from traditional consumer credit data. Rental payment data, for example, while only furnished for positive behavior (no delinquency data available) is available by request from some bureaus, but is not incorporated into FICO Score 8, the most widely used version of the FICO score. BNPL Pay-in-4 data is also being assembled separately, if available at all. A number of Pay-in-4 providers are not currently furnishing. The same is true of utility payments. Bureaus have most of the data, but FICO might not be able to accommodate the unique information.
New data furnishing like rent, utilities, and BNPL is only scratching the surface on potential relevant data sources for consumer underwriting decisions. Three readily available data sources, thanks to fintech innovation, are likely to be very influential in modern consumer underwriting:
Bank account data has been seen as incredibly promising for quite some time. Use of bank account data and cash flow analysis is an adaptation of standard SMB underwriting. However, widespread use has been delayed due to consumer permissioning challenges and inconsistent data labeling. Consumers continue to be concerned about privacy when asked to grant account access. New, precise permissioning interfaces that Plaid and others have launched with major banks enhance consumer confidence. Additionally, while categorizing transactional data is improving at a rapid clip, the risk of improperly interpreting transactional data, thereby unfairly rejecting a credit application, looms large. Some companies like Brigit have done an exceptional job tackling data labeling and demonstrating the predictiveness of bank account data.
Payroll data has come into the mainstream over the past two years. Now that a lender doesn’t have to pay $50 to The Work Number for an automated pull from ADP, verification of income and employment is much more attainable. Two specific data points have been shown to be strong signals in credit models. The obvious one, income, can be a dynamic figure, looking at historical increases in income at the employer. Tenure has shown to be an incredibly useful characteristic as well. Those who stick around have more steady income characteristics and also may be more generally employable.
Platform data is a relatively new concept, mostly utilized in the SMB context. A user’s (consumer or SMB) behavior within an ecosystem can provide signals that are incredibly useful for underwriting. Platforms like Square, Stripe, Amazon, and PayPal have access to unique POS data that are often more predictive than conventional underwriting signals. Also, captive ecosystems have the advantage of loan/advance repayment in the flow of funds, a true advantage for these platforms. On the consumer side, one could assume trusted, scaled platforms like Chime, Albert, and CashApp should be able to execute on the same thesis.
These modern data sources can incorporate a forward-looking view of the consumer or SMB. FICO can be manipulated in part because it is principally a backward looking score. Lenders assume historical behavior is highly predictive of future performance. Every investment product today states that past performance is not indicative of future results, so why does FICO emphasize the past so strongly?
Below is a schematic of how modern underwriting assessment could work. Starting at the top, identity and fraud are critical inputs in the credit decisioning process. A clean pool of applicants will greatly improve model development and results. I continue to believe fraud is the most ignored feature of credit underwriting, with many lenders outsourcing too much of the process and having limited understanding of how they are being defrauded. Moving down, bureau data continues to be incredibly helpful in assessing credit and should continue to serve as the foundation for any underwriting process. Now comes the new stuff: using contextual data helps keep underwriting dynamic and fresh. Depending on the use case, a number of data sources and specific signals could be incorporated. The objective is to be able to detect and (if appropriate) underemphasize manipulated or misleading bureau data while incorporating context to better assess the applicant’s financial health. While there is no universal solution to better underwriting, contextual variables consistently improve performance if properly incorporated into underwriting models.
A word of caution
One hot button issue in new data usage is the potential for disparate impact. I’m excited about the opportunity for contextual underwriting to build a more inclusive credit system, removing barriers on how an individual initially establishes his/her credit profile. The effort to game FICO is specifically aimed at this dynamic, demonstrating creditworthiness in a Metro2-accepted format. Contextual credit provides a more sensible and sustainable path to inclusion.
However, even if unintentional, new data could introduce new biases to underwriting data. Risk teams must work in partnership with their Legal/Compliance colleagues to ensure variables are not directly biased. Upstart went so far as to seek a no-action letter from the CFPB for its use of employment data in making underwriting decisions. While the letter lapsed this year, it served as an acknowledged protection for introducing non-bureau variables into underwriting decisions. My hope is the CFPB increases its openness generally to new types of data, enabled partly through more modern guardrails.
Modern disparate impact services like Fairplay allow for model review and governance to detect and remove any unintended bias from models. Model discrimination is a dynamic problem, and the traditional quarterly disparate impact assessment by a consultant does not cut it in this day and age. By integrating compliance review into risk and data science efforts, lenders can push into contextual underwriting with governance and controls to expand the addressable population.
Expanded credit access through context and data
As stated above, lenders should be decreasing reliance on certain bureau data that has degraded in quality. Openness to new data sources is critical to filling gaps in our current system. I would argue contextual underwriting is a foundational approach to modern credit assessment, introducing forward-looking variables and expanding fair access to credit to un- and under-served individuals and SMBs.
A critical challenge today is the lack of infrastructure and tools to tackle this new phase of contextual credit. In our final installment on credit innovation, I will cover two major opportunities still in need of great startups to make contextual credit viable for mass adoption:
The fourth bureau: the establishment of a trusted third party to provide clean data from new sources is long overdue. Lender data science teams spend months with a new data source before the data is sufficiently reliable (clean, accurate, scaled) for usage in underwriting models. Complete coverage of new data sources is impossible for one business to achieve. The overlap in work is tremendously inefficient, and the US lending industry would greatly benefit from the establishment of a centralized source of modern credit data.
Modern modeling platform: The addition of contextual data to underwriting requires a new generation of flexible models to be built. Swapping in and out credit signals depending on context is very, very challenging for both traditional regression and machine-learning modeling techniques. How can underwriting or data science systems evolve to support faster iteration while maintaining integrity?