An alarming headline from Dec. 2, 2016 read: “New auto loans to borrowers with credit scores below 660 have nearly tripled since the end of 2009.”
This reads like it could be cause for intervention. Is the next credit bubble already at our doorstep?
It might be. This headline, however, does very little to answer that question.
This industry’s most common yardsticks—averages or rough distributions of credit scores— are fundamentally flawed. In the context of a single pool, from a single issuer, at a single point in time, these metrics satisfy the need for a quick proxy of borrower credit risk. But every time we extend their use—to compare different deals, or an entire sector, or issuance over time— we make an implicit assumption. When we then stack these assumptions on top of each other, the status quo that we’re used to is at best misunderstood and at worst nearly meaningless.
The good news is, there is now a way. But first, let’s take a look at those assumptions.
Credit Score vs Credit Risk
Credit scoring models—whether from VantageScore, FICO, or any other developer— are designed to rank order consumers. Credits scores—the outputs of those models— are not designed to predict a fixed level of risk.
The relationship between credit score and probability of default is defined for each model when initially developed, but may then change over time. For models that perform well, rank order will hold across the credit cycle such that a score of 600 always suggests a higher level of risk than a 700 and a lower level than a 500. The specific probability of default at each of those scores, however, can and will change.
At any point in time we can translate a score into an estimate of risk based on recent performance.
Assuming a fixed relationship, however, can obscure meaningful changes.
A score of 660 indicated a PD of approximately 6.0 percent in 2006, of 8.1% at the height of the financial crisis, and of 3.8% today. Prime consumers demonstrated this same dynamic, with risk levels more than doubling and then halving again through the course of this most recent credit cycle. This is precisely the reason that sophisticated lenders will periodically adjust their credit policies based on their actual and projected credit experiences.
Clearly this suggests a practical limitation to the use of credit scores when comparing deals originated at different times.
Linear vs. Exponential
Using the average credit score to estimate the riskiness of a pool is mathematically flawed.
For a given pool of loans, each borrower’s credit score typically will fall somewhere on a linear scale between 300 and 850. As described above, each of those credit scores has an associated level of risk for any given point in time. Taken together, these risk levels can be represented as a probability of default curve. Unlike the score range, which is linear, the underlying risk curve is exponential. Taking an average of the credit scores (a linear function) and using it to represent risk (an exponential function), can obscure the true risk of a group of loans.
To illustrate, consider two hypothetical pools of loans in the attached chart. Both pools consist of 10 loans and both have an average credit score of 660. Using this measure, these pools represent ostensibly the same risk profile. As illustrated below, however, the average risk of Pool 2 is actually double that of Pool 1.
Default Risk Indexing
As we’ve shown, credit scores may be familiar but they are often misused. The expansion of loan-level security data, however, has opened up a better approach: convert the credit score on each loan into a probability of default, using the most recently-available performance charts, and then calculate the weighted average probability of default.
This value, which we call the Default Risk Index, is mathematically sound and reliable over time. A more detailed methodology is available free of charge at DefaultRiskIndex.com.
To illustrate, let’s reconsider the state of auto lending.
The headline quoted at the top of this article suggested a bubble, citing the growth in originations to sub-660 borrowers. As demonstrated above, however, 660 isn’t a reliable yardstick over time. Instead, let’s look at the Default Risk Index values for recent auto originations.
As we see in the chart below, the volume of originations has generally trended up over the past three years. The risk profile of those originations, however —as indexed to 2013— has modestly tightened.
Of course, borrower credit is only part of the picture. To fully assess a bubble, we need to also consider changes in loan terms, servicing technology, and channel mix. Looking purely at borrower credit, however, we can see how the auto industry began a modest tightening over a year ago. Transparency is increasing everywhere. It’s high time that the secondary market rethinks the ways in which it relies upon credit scores.
Mike Trapanese is senior vice president , strategy and alliances, at VantageScore Solutions.