Estimating creditworthiness for consumers with limited credit history
When a consumer applies for credit at any financial institution—bank, credit card company or individual lender—the institution has to answer at least three important questions before deciding if and under what terms to provide credit to the consumer.
Is this consumer creditworthy ie How likely is the user to repay the loan?
Should we require the consumer to submit a down payment, ie a partial approval?
What interest rate and repayment period should we offer the consumer?
Historically, answering these questions relied on strict rules and were usually made unilaterally based on cut off requirements of a few numbers, like the FICO score. As technology has improved, credit decisions have become increasingly automated without requiring any human in the loop. The algorithms that drive the automation take into consideration inputs about the consumer’s credit history to estimate their likelihood of repaying the loan.
Automating credit approvals has enabled consumers to easily and quickly obtain a credit decision and improved the overall experience. However, because these algorithms mainly use an applicants credit history to make an approval decision, those who are new to credit or have shorter credit histories have trouble accessing credit. Consumer Financial Protection Bureau (CFPB) estimates that roughly 18 percent of the entire American population is either unscorable by credit algorithms or do not have a credit record at all.
Traditional credit models only look at information on the credit report to estimate creditworthiness. These largely include:
Recent payment history
Number of credit accounts owned (and their current standing)
Length of credit history
Total outstanding debt
When a consumer has a long credit history, these attributes tend to be excellent predictors of creditworthiness and in turn the likelihood of repaying the loan. But for the sizeable American population with shorter credit histories, these attributes represent an incomplete credit profile. Traditional credit models either cannot provide any score for such a profile or provide a score that doesn’t accurately reflect their creditworthiness resulting in inaccurate credit decisions for these consumers. Despite this knowledge, traditional credit models continue to dominate most of the financial industry, leading to a vicious cycle for those with shorter credit histories.
In a major shift from popular practice, the CFPB announced they would be looking into alternative ways of determining a consumers creditworthiness. In a similar move, the makers of the FICO score announced that they too are starting to use the way people manage the cash in their checking, savings, and money market accounts to inform the score in the hopes of giving more people access to credit.
At Affirm, we strongly believe that a frequently updated model that looks at more and different sources is able to approve more accurately by better estimating consumer creditworthiness. When models use more data points (and more context), they can create a complete credit profile even for a consumer with a limited credit history. With more data, our models get a better picture of the overall stability and credit trajectory and make better approval decisions.
Here are a few alternative data sources we believe will become important factors in modern underwriting.
1. Transaction data from bank accounts, credit cards or debit cards
Information from checking and savings accounts can help explain the current state of the consumer’s financial situation, but data on a consumer’s spending patterns and behavior better indicates their financial future. The ebbs and flows of cash can indicate if a consumer is able to afford debt in the future. Traditional credit reports do not get this view. Presently, financial institutions have to make decisions off a single snapshot in time.
2. Type of purchase
When a customer has a limited credit profile, understanding what item the loan is financing can be a good indicator of the riskiness. For example, data shows that an applicant with limited credit history that buys a Rolex is less likely to repay than a similar consumer buying a mattress. Therefore, while both consumers might have been rejected under traditional underwriting due to their limited credit history, a model that takes type of purchase into account could extend credit to the customer interested in buying the mattress because the data indicates they are a lower risk consumer. Context matters and traditional underwriting models completely ignore that aspect.
3. Rent and other utilities
Credit history isn’t the only way to prove an applicant will make good on a promise. Utilities are an example of a payment made after a consumer has used the service. Similarly, with rent, an occupant signs a lease to pay a certain amount and makes good on that promise every month—demonstrating both intent and ability to pay at regular intervals of time. These data sources can provide more information and help evaluate the risk of an applicant who might have avoided credit cards and has a limited credit history.
4. Credit history from other countries
When an immigrant or a US citizen living in another country for many years arrives in the United States, they have no credit history in our systems, restricting their access to credit. This is despite the fact that they might have built up a strong credit profile in another country. While this data isn’t readily available just yet, there are notable companies devising technology to incorporate their foreign credit profiles once them immigrate.
5. User reported education, employment, and income
Consumers who are new to credit generally have education, employment, and income information that can help estimate their future creditworthiness. The information can help establish whether new to credit professionals can afford what they are requesting. Even for consumers who have longer credit histories, employment and income information will help point to further stability and affordability in their future than their past credit score or profile indicates. While this data is not reliably available today, there are potential technology solutions that can integrate user supplied data seamlessly with the current underwriting process.
Potential for bias
As with any model or algorithm, it’s important to be aware of unintended potential biases. Having more and alternative data points doesn’t always lead to a less biased system. In fact, more data sources leads to a more complex system that might have an inherent disparate impact that adversely affects a particular group of people even though the model doesn’t intend to do so. As with any predictive model, it’s important to have the appropriate controls in place to both discover and address any unintended bias or disparate impact quickly. Independent third parties should regularly conduct disparate impact analyses on model outcomes to ensure the soundness of the predictive models.
Most of the above aren’t just the future, it’s the present as both government entities and private financial institutions alike start to shift their basic underwriting philosophy. But like most things in the financial world, these programs have been slow moving and have had limited implementation. Hopefully, the success of Fintech startups will quicken the pace of innovation in the rest of the industry.