At Zest.AI, we’ve spent years applying machine learning techniques to auto loan underwriting. Machine learning brings in a lot more data and more diverse credit signals to produce more nuanced risk analysis. We’ve dived deep into auto loan data across the credit spectrum to find the best way to differentiate borrowers who are creditworthy “thin/no-file” -- in other words, have little to no credit history -- from those who are likely to default.
The ability to parse and price non-prime credit risk more finely is becoming critical for auto lenders with subprime auto loan delinquencies hitting historic highs.
Based on our experience, here are three credit signals every auto lender should look out for when underwriting a thin/no-file borrower:
Discrepancies between the amount of time a co-borrower reports living at an address, and the verified amount of time they've been at that address.
Because thin/no-file borrowers can be hard to underwrite, some lenders will require the addition of a co-borrower. If a co-borrower claims that he or she has lived at an address for a long time but the public records tell a different story, that discrepancy can be a sign of fraud or other misrepresentation.
NOTE: We always advise being thoughtful in how you’re using length-of-residency signals since they can raise fairness concerns. Ensure your data is accurate and avoid over-emphasizing how long an applicant has lived at a location in ways that might unnecessarily exclude creditworthy applicants such as military service members, who move around a lot.
Whether the borrower maintains a positive minimum bank balance consistently over several months
When a low-income person manages to keep a positive bank balance, that signal is an indication of financial prudence and the ability to responsibly manage a budget. Incorporating this signal is a great way to help identify creditworthy borrowers that might otherwise be denied based on traditional metrics like income alone.
Whether the borrower has taken out an installment loan in an amount greater than 50% of their down-payment for the car in the last 30 days.
This signal helps identify what is sometimes referred to as loan-stacking, in which a customer borrows the money for a down payment for a car resulting in a 100%+ loan-to-value ratio. That situation can be a serious risk factor in the individual’s ability to pay back the loan.
How important are these signals? Zest data shows that thin-file borrowers who rank well on the above three variables default on average 4.7% of the time, while thin-file borrowers who rate poorly by these metrics default on average 30% of the time.
The non-prime auto lending business is in for some choppy waters and lenders will need to innovate in order to stay ahead of what’s to come. Machine learning and big data are powerful tools that, when used responsibly and fairly, can help auto lenders weather any turns in the economy.
For more information about using AI for auto loan underwriting and how to get the maximum impact from variables like these, email us at email@example.com.