In the fourth quarter of 2022, many auto lenders were surprised by elevated levels of delinquency and charge-off that have continued through the first quarter of this year. The deterioration did not occur due to loose lending or ineffective collections, but rather was driven by a predictable shift in the market. This is not an uncommon event, and I have even seen it happen to some very good, sophisticated operators over my 25 years in the industry.
Such surprises rarely sink an auto lender, except in cases of fraud or incompetent management; however, they can create numerous pain points for all involved. The most immediate impact is a decrease in the borrowing base. Most warehouse lines of credit have covenants that reduce available debt as delinquency increases, leading to a cash crunch. Lenders must either cut volume or come up with more equity to stack against the line in order to keep funding at the same level. In more extreme cases, companies must layoff originations staff, cut off dealers and tighten credit in order to adjust to a reduced plan. Add to that negative press, increased cost of funds and poor execution on bonds, and the lender’s problems pile up.
Changes in credit quality are an inevitability. The hallmark of good management is not that surprises don’t happen, but rather how a proficient management team recognizes and responds to them. Performance shifts occur due to one of three reasons. First, originations quality gets worse. Second, the effectiveness of the collections operation deteriorates. Third, and finally, the environment changes the outcomes lenders achieve with all other things being equal. This article deals with the latter, meaning the lender does not go wheels-off on operational execution, but experiences a decline regardless.
Lenders may effectively manage these shifts by focusing on the following risks:
• Static Models
• Macro Shifts
• Insufficient Tracking
Surprises cannot be completely eliminated, but lenders can thrive through turbulent environments by managing the basics.
A credit scoring model is not a loss forecast, it is an index. Simply put, it allows the lender to sort applications from most-risky to least-risky. A lifetime default rate (or forecast) is imposed on top of the index after the fact based on each lender’s historical performance. Whether an analyst uses a weight-of-evidence approach, or some other more exotic estimation method, models are static once deployed. They cannot be clairvoyant to future changes. For this reason, lenders typically update their origination scores every 24 to 36 months, as it takes that long to accumulate a significant number of seasoned accounts with which to re-estimate the model.
The limitations of predictive models are, in part, connected to the nature of the inputs that are used. Consider the case of artificial intelligence applied to the problem of driverless vehicles. During the model training process, the algorithms learn that a red octagon with the word “STOP” in the middle always represents a stop sign. It does not transform to a green triangle or a different word depending on the environment. This is not the case when it comes to estimating predictive credit models.
Many surveys of defaulting consumers have been conducted by lenders over the years. The primary reasons people give for why they default overwhelmingly relate to divorce, a catastrophic medical issue, a business failure or unemployment. Unfortunately, there is no credit bureau attribute that suggests you will get divorced, fail in your business or lose your job in the coming months. Credit origination models rely on the “odds” of you being more or less successful at paying your bills (among other things) based on your history.
The relationship is imperfect, and the correlations to the outcome change over time based on environmental factors. Many who might have defaulted during the pandemic received unprecedented forbearance from lenders, stemming the tide of negative performance. In the last year, the evaporation of that support combined with increased gas and food prices has pushed some into default that might have paid otherwise.
Addressing the static issue involves moving toward economically neutral models. This means integrating data from various periods of good, bad and neutral (i.e., matches the long-term baseline) in order to average down differences across the credit cycle. Many lenders will need to augment their data with statistically similar records from the credit reporting agencies in order to accomplish this, but the result will be models that produce less variability over time. The desired model is not one that perfectly fits yesterday, but one that is most likely to do well regardless of what the future environment brings.
Developing more robust models is an important step in working to mitigate the risk of performance shocks, but it is not enough. Credit scoring models typically have an intercept, or baseline level of credit losses, around which the score is centered. As with the previous section, this intercept is static – the market is not. The absolute level of defaults a lender will experience is integrally related to economic conditions, as well has how much lending capacity is available in the capital markets.
Consider the chart titled “Change in Consumer Credit vs. Default”. While we track hundreds of industry and economic factors, there are a few that stand out as strong leading indicators of where credit performance is headed. Please note, this is just one example and not the only thing lenders should consider. The G-19 report of the Federal Reserve is published monthly, and contains data on how many billions of dollars are owned and securitized over a long period of time.
The chart plots the percentage change from month to month in dollars outstanding compared to the annualized net loss rate for subprime auto asset-backed securities. It is important to note that every time there is a significant and protracted drop in the growth rate of credit, a spike in portfolio net loss occurs 12 to 24 months later. For my hardcore, quantitative friends that will be the point where the sign of the second derivative changes, indicating that the growth rate in new credit has started to decline.
The banks and rating agencies that review loan level performance detail for many lenders have a vantage point into where the market is headed that no individual lender has. When the debt providers see early signs of credit deterioration, they tend to limit their exposure by contracting access to capital. Again, this is not a perfect leading indicator, but is among several strong factors that analysts should monitor in order to adjust score-based loss expectations and credit programs ahead of a negative shift.
Mid-cycle adjustments must be made based on the lender’s own history of defaults across multiple and varied periods. For companies without a large enough performance sample, I recommend partnering with external data providers in order to establish a baseline and an expected range of variation. The process of managing forward looking loss expectations need not be an overly complex one, and may be accomplished by following these steps:
1. Establish a baseline and range. For example, a lender might observe cumulative default rates that vary from 17 to 23 percent, with a baseline of 20 over the preceding 10 years. This exercise should also be performed on various stages of delinquency, as that will be part of the early warning system that indicates performance trends are shifting.
2. A committee should convene no more than once a quarter to assess the environment. This group should include executives from risk management, operations and finance. In addition, there must be a strong executive champion leading such a meeting, so that when changes are observed the discussion does not devolve into a finger-pointing session.
3. If a decision is made to adjust the loss expectation, the lender should do so in a modest and measured fashion. For example, begin with 50 basis points and commit to monitor performance for a sufficient period of time before making further adjustments. There is typically much noise in credit performance data, and over-reacting to every up or down will create unnecessary disruption.
Responding in a timely manner to shifts in a lender’s own credit quality, as well as that of the broader market, is fundamental to heading off performance shocks; however, many are not tracking the key factors necessary to do so. Analytic-driven lenders typically track population stability and perform characteristic analysis on those attributes that comprise their credit scoring models to ensure that the current population is reasonably close to the one their models were developed on. I encourage all those using scoring models to do so, but it is not enough as it relates to predicting shifts in expected credit performance.
Models are built off a snapshot frozen in time. Unfortunately, both the weightings on the predictors and the size of the error (i.e., the performance that is not explained by the model) fluctuate across the credit cycle. For example, payment to income ratio (PTI) is a common and important scoring model input. Most lender’s try to keep their average in the 10 to 12 percent range, and typically have credit policies that cap that value at 18 to 20 percent. While a model builder in 2020 may have accurately measured the relationship between PTI and default at the time of the model build, they could not have accounted for what would happen when fuel, food and other staples in the household budget were hit with significant inflation. A 12 percent PTI in today’s economy may effectively behave like 20 percent in prior periods.
In the same way, many who booked loans at a 110 percent loan-to-value (LTV) ratio last year are in for a rude awakening when used vehicle values drop. There are many factors related to the economy, inventory, the state of competition and applicant credit histories that are strong leading indicators of performance, and most of them are not contained in the scoring model itself. This is not a deficiency in the scorecard development process, but is simply a byproduct of operating in a dynamic environment. The process I am suggesting is much more akin to the work performed by meteorologists, where they model how things move and interact over time in order to assess the most likely scenarios.
Effective tracking systems are not meant to model precise relationships that lead to automatic outcomes, but rather, are intended as a guide to management that may be used in evaluating forward-looking adjustments. For them to be used in this way, they must be easily monitored and understood. A common mistake lenders make in attempting to build such systems are to track too many trends, many of which are merely noise. Doing so will overwhelm managers and nullify the value of such an exercise.
Highly-skilled statistical analysts rely on data reduction techniques to take a vast array of information and boil it down to key factors that may be tracked alongside performance trends. Factor analysis, cluster analysis and other multivariate tools are used to determine how thousands of elements group and move together, and which dimensions are most important as it relates to shifts in performance. Many lenders subscribe to a variety of data attributes (credit, alternative and other types) which may be used for this purpose. When the above multivariate methods are employed on such data, analysts will observe that all attributes related to delinquency, auto credit, inquiries and other broader categories tend to group together. They are weighted according to which ones explain the greatest amount of variation in the total data set. Once collapsed into a single vector (linear trend), each factor may be tracked alongside historical shifts in credit performance, making it much easier for managers to digest. An example of one such report is included with this article.
I have spoken to a number of large, well-run auto lenders in the past few months. Many of them have expressed frustration with rising delinquency. Each one of these companies has solid management, a disciplined credit program, analytic sophistication and an effective collections organization – yet they were still surprised by performance shifts. There are a variety of reasons why this happens.
It is common for originations staff to cite poor collections effort for performance issues, while the collectors blame loose underwriting. Both of these groups also blame the risk team for having bad models. For executives, it is often difficult to pinpoint where the deterioration came from. The key takeaway in all of this is that there are many contributing aspects to credit outcomes that are completely outside the control of the company. By tracking these factors, and forming a team responsible for evaluating action, lenders can effectively plan and mitigate these risks before they create a crisis. On a final note, there are parts of this article that are necessarily technical and are difficult to fully detail within the confines of a single article. For those who would like more clarification, feel free to reach out to me at d[email protected], and I will do my best to elaborate.