Finding Product-Market-Fit for Underwriting (#47)
Industry continues to struggle with alternative forms of credit assessment
Welcome to the 47th issue of Unit Economics. For today’s write-up, I share a short note on the importance and challenges of underwriting in the lending market today. Dive in!
Everyone wants loans. For homes, vehicles, businesses, or even everyday purchases. Some feel the need to take credit to fund their expenses, and others see gratification in delaying their bank account drawdowns.
But if you give loans to everyone, you are not going to get enough back to remain in business. So you pick and choose the people you want to offer the loans to. You hope that your underwriting policy makes the right choices as often as required and that you constantly iterate until you get it right or, unfortunately, you run out of time.
For any lending product, the underwriting process then really defines its strength. And in a market full of lemons, underwriting is tough. You can build models and predict with accuracy all steps until the loan is in the hands of the consumer. Then it gets unpredictable. It is not too different from the last-mile delivery problem that most e-commerce companies deal with.
But why is it so tough to predict the credit behavior of consumers?
When underwriting, you need to (1) define the parameters you want to profile the consumers with, (2) find ways to access that information with acceptable confidence. And you need to ensure that you adjust your underwriting to the requirements of the adventurous business goals, and innovative product constructs of lending instruments.
Each of these steps comes with its challenges, and in an environment that demands high growth – the onus on being agile and driving product-market fit lies heavily on underwriting. So, where does it get challenging?
Defining your risk policy
Think of each consumer as an animated set of demographics, physiographic, or behavioral characteristics. The combination of all these traits is assumed to reliably predict a consumer’s future repayment behavior.
But you can never know all the characteristics that make that combination, and even if you do, accessing the information is hardly possible. So, when defining a risk policy, out of hundreds of such traits, you pick out the ones that – at the median – would be the strongest predictors of consumer credit behavior.
Once you have made these choices, you tie them in a decision tree. Each branch of the tree is crucial in taking the end decision on whether (1) consumer is eligible for credit, and (2) if yes, then how much risk can we handle for the consumer? or in simpler words, how much credit can we sanction given this profile characteristic?
But this is where it starts becoming challenging. When you are going out in the market with a new risk policy, the decision trees are driven by previous experiences and a host of hypotheses. Often, given the incentive of keeping credit defaults within the quarterly targets, the risk policy leans towards a little conservatism. This makes sense when your consumers have had little time on the books. But it hardly ever meets the expectations of the product or business teams. Why?
The policies introduce variables that add friction to the consumer onboarding. They do this by seeking information from consumers that consumers might not be comfortable in sharing (example: asking users to input income or employer information) or by introducing flows that require effort and almost always lead to significant drop-offs (example: mandating users to register for auto-debit, or to log in to upload bank statements, payroll slips).
Despite the many ways in which policies attempt to use alternative data, the lack of confidence in data quality and limited market options lead to almost a traditional treatment of new-to-credit users, which drives an approval rate that tends towards low double digits for most modern fintech teams (~10-30%).
With the fintech lenders competing for the same population today, the high friction in onboarding and low approval rates are challenges that go against the business goals and introduce conflicts. Often, today, where “tech” in fin-tech overrules the “fin” – the product wins until market results show otherwise. Reparations are made in the risk policy to allow for more seamless onboardings, and targets for portfolio risk are kept more liberal in anticipation that - over time – the risk policy will improve, and the faster user acquisition will allow more funding, and consequently, more time for the risk policy to get it right.
So, how does a risk policy accurately predict credit behavior without introducing elements that lead to onboarding friction or low approval rates? By accessing variables that are strong credit behavior predictors (1) for a higher proportion of users, and (2) with reliable accuracy.
This leads us to the root cause of the crisis: unreliable and, often, unavailable data.
Finding reliable data
The risk policies today continue to be strongly tied to the limited and often outdated credit bureau reports offered by the few. So much so that consumers are often bracketed into those with bureau data, or outcasted as “new-to-credit”. No denying that it makes sense for the risk teams to seek credit bureau data as their first source of truth, given the reliability of information for those with good credit scores and the comfort of knowing that others in the industry are doing the same.
But the coverage of credit bureau data in the country remains low at 50-60%, and even within those seen as credit active – hardly half have the scores worth underwriting. In reality, a low score is often an indicator of a lack of consumer data to determine his/her credit risk profile accurately.
So, what do we do if credit bureaus are not sufficient to underwrite a high proportion of users with accuracy? You seek alternate methods. Easier said than done, however. In the market presently, alternative data sources continue to be far and few – primarily focused on income and employer verification, and all come with their perils.
Bank statement analysis
Bank statements, especially for salary-linked accounts, offer a glance at the money-ins and money-outs for the formal population. The data is captured here in the form of bank-to-fintech sharing, wherein the credit seekers are required to perform net banking login to their salary account during onboarding and to allow consent for sharing bank statements. The method scores high on reliability, but low on ease-of-access.
Given the market experiences, you would be lucky to get even 60% of the users in your onboarding to go through with the login. And so fintech companies miss out on a large segment of potential good credit users.
An alternative flow allows users to instead upload their downloaded bank statements, which is then consumed – often with the help of a technology provider that provides the ability to OCR and parse the data. However, this method is relatively less reliable due to the variable quality of file uploads and comes across the same issues as the net banking logins during onboarding. Moreover, both methods neglect a large section of the informal economy that does not get salary deposits or does not have access to net banking logins.
SMS mining
Another popular alternative that most risk teams rely on today is the financial information captured in the SMSes of individuals. The data parsed through SMSes is relatively friction-free, as the consumer is only required to provide consent for the same – which is often achieved subtly by product team and doesn’t impact the onboarding flow as much.
However, the SMS data is renowned for poor reliability, wherein the accuracy of variables hovers not too high above 60% - and is often tested to have a high standard deviation. The variables parsed from SMS can help risk teams underwrite new-to-credit users, but the low reliability makes this a step-process wherein the underwriting is often of low-ticket size initially, and the sanction amounts are raised over time on the condition of good behavior.
Payroll data: pay slips, taxes, form 16, etc.
The most direct method of employer and income verification involves submitting proof of payslips, ITRs, or Form 16s that require similar flows to bank statement file uploads and come with equal hesitancy from the consumers. In practice, due to the relatively lower frequency with which users access these documents – there is a higher chance of consumer drop off given the effort required to log in or to upload the document. You would also notice then that not many credit teams choose to rely on these sources for assessing creditworthiness.
There are other methods, including credit card analysis and proprietary payment data evaluation, that are at the early stages of their lifecycle and promise higher reliability. But to reach the status of maturity, these solutions will require some years in the market.
Regardless, the bottom-line of reading into the credit underwriting is that the value chain of accessing and evaluating users’ financial information remains broken – with limited reliable solutions of alternative data assessment that continue to tie the hands of the risk teams. There is little to suggest that the conflicts against the product and business teams will lessen any time soon.
Moreover, given the clear market challenges, you can also treat this write-up as an investment thesis for anyone willing to solve for employer and income verification in the industry. And with this, there should be little doubt as to why certain initiatives, including the Account Aggregator framework, are received with so much promise in the industry.
For lending products to stand tall, however, there is much to look forward to in the next ten years of the underwriting revolution.
If you have any views or feedback to share on the topic, feel free to add a response below or to share your thoughts with me over Linkedin. In case you feel your friends or family would be interested in reading about payments, feel free to share the blog with them as well. See you in a couple of weeks!