The Portfolio Is a Moving Target
The Cleveland Fed auto-loan model does not treat every loan as an interchangeable claim on the same economic future. Its current focus is the structure underneath an aggregate default number: how origination credit quality, lender channel, loan features, and regional conditions shape the transition probabilities for a particular portfolio. Understanding that structure lets us ask a sharper question about the unemployment relationship: if the borrower mix changes, could the portfolio look different even when the conditional model relationships do not?
Baseline risk is distributed across borrowers, loans, and channels
Origination risk score is the model’s clearest marker of baseline credit quality. In the 2000–2019 current-state estimates, borrowers with origination Equifax Risk Scores below 560 have a current-to-default relative risk of 22.227 compared with borrowers above 760. The same comparison for current-to-delinquent is much larger still: the paper describes it as roughly 70 times the reference-group level. These figures do not mean that a deep-subprime borrower has a 22-fold probability of default in every circumstance. They compare destinations with remaining current, holding the model’s other specified inputs fixed. They show how much starting risk differs before an unemployment shock is applied.
Lender channel adds another layer. Relative to a credit-union loan in the same current-state model, an auto-finance loan has a current-to-default relative risk of 1.551, or about 55 percent higher, while a bank loan has a relative risk of 0.886. The paper also finds that auto-finance companies originate the largest share of subprime loans. The channel is therefore related to the population it serves, but it is not reducible to the risk-score distribution. Even after observable characteristics are included, performance differs across lender types.
That last result needs careful wording. A lender indicator is not a laboratory treatment. Auto-finance companies, banks, and credit unions can differ in borrower selection, underwriting, loan terms, servicing, collateral, and incentives. The data and model do not isolate which of those features produces the conditional difference. The defensible point is that lender channel contains information about transition risk in this sample. A risk analyst who removes it may be discarding a meaningful pattern; an analyst who treats it as a causal label may be claiming more than the evidence supports.
Other variables help shape the same baseline. In the fitted current-state model, younger borrowers, borrowers without a mortgage, loans further from origination, loans without a joint account, and some balance levels have different relative risks from remaining current. Loan age is allowed to have a quadratic relationship with transitions rather than a forced straight-line pattern. These details matter because a macro coefficient is interpreted conditional on them. A portfolio of older loans, single-borrower accounts, and lower-score originations can have a different default level from a portfolio with the same unemployment path but a different distribution of those features.
The distinction between risk level and macro sensitivity is easy to lose. Risk level asks how likely a transition is for one group relative to another under the model’s conditions. Macro sensitivity asks how that transition changes when an economic input moves. A group can have high baseline default risk and a comparatively stable unemployment coefficient. Another group can have low baseline risk and a small absolute forecast error that is large relative to its realized defaults. Neither cross-sectional comparison answers the time-stability question by itself.
Relative-risk comparisons also have a denominator that must stay visible. The deep-subprime figure of 22.227 is a comparison with remaining current for the reference conditions, not a comparison with the default rate of the whole portfolio. The auto-finance figure of 1.551 is a comparison with a credit-union loan in the same fitted current-state framework. These ratios describe how the model ranks possible destinations conditional on a starting state. They cannot be multiplied together casually to produce a borrower’s final probability, because the model also allocates probability to delinquency, continued current status, and non-default exit.
That distinction changes how a risk analyst reads a portfolio report. A high relative risk can coexist with a small observed contribution to total defaults if the group is a small share of the book. A lower relative risk can matter greatly if the group is large. A segment’s macro sensitivity can also be statistically important while producing a small dollar effect when its baseline exposure is small. The model’s coefficients and the portfolio weights answer different questions; a decision-maker needs both before deciding whether an apparent change is a risk shift, a mix shift, or a calibration problem.
The interfaces among these components are what make the portfolio a system rather than a list of covariates. Origination score describes a starting risk position. Lender channel partly describes the institution and product through which the loan was made. Loan and borrower characteristics shape the state-specific baseline. Regional unemployment and house prices provide changing conditions. The transition model combines them, then aggregates the resulting probabilities across all loans. A change in the aggregate can occur at any interface: the inputs can change, the mapping from inputs to transitions can change, or both can change together.
The diagram is deliberately conditional. It does not say that each input has an independent or equally large effect. It shows the path by which a loan-level observation becomes a portfolio-level result. Risk score and lender channel can move the starting distribution of risk. Macro variables can move conditions over time. The estimated transition relationship determines how those inputs are translated into movements among states. Testing stability means checking the last connection, not merely counting how the first inputs changed.
Composition can change the picture without changing the relationship
The underlying market composition did change over the study period. Banks’ share of auto-loan originations declined, while credit unions’ share increased. Auto-finance companies played a particularly large role in subprime lending. The distribution of origination risk scores also shifted: the share of loans above a score of 760 rose from roughly 20 percent in 2000 to roughly 40 percent in 2025. Those movements can alter the aggregate default rate even if each score group and lender channel follows an unchanged conditional relationship.
Consider a simple aggregation problem. Imagine that high-score loans become a larger share of a portfolio while the transition probabilities within every score group remain constant. The portfolio’s average default rate may fall because the weights on lower-risk groups have fallen. Nothing in that change requires the unemployment coefficient to move. Conversely, a portfolio can become riskier because it contains more loans from a high-default channel even if the channel-specific relationship with unemployment is unchanged. Aggregate outcomes combine within-group risk with the weights assigned to each group.
The same arithmetic applies to a stress response. If a low-risk group expands, the portfolio may show a smaller average increase in defaults after an unemployment shock simply because more weight sits on a group with a lower baseline probability. That observation alone cannot tell us whether the group’s unemployment coefficient weakened. To answer that, the analyst must compare the response within comparable groups or hold the group weights fixed while changing the estimated relationship. Aggregate resilience and conditional resilience are not synonyms.
Conversely, a stable overall default rate can hide offsetting movements. A riskier lender channel may expand while a higher-score group expands at the same time. One pushes the weighted average up and the other pushes it down. The net result can look calm even though the portfolio’s internal exposures and transition paths have changed. This is one reason a single headline default rate is a poor diagnostic for parameter stability: it compresses the weights and the within-group relationships into one number.
This is why the paper includes origination score and lender type rather than relying on aggregate default rates. The controls do not make the composition problem disappear. They make it possible to estimate conditional relationships while observing how the composition itself is changing. Later, a decomposition can ask how much of a prediction gap is associated with changed borrower and loan data, changed macro data, and changed coefficients. The logic is important even before the result is known: a portfolio shift is a competing explanation that needs to be tested, not an assumption to accept or ignore.
Origination score is used instead of a score that moves with the loan’s later history. The paper’s reason is that movements in a contemporaneous risk score may correlate with changes in other model variables, including macroeconomic outcomes. A score measured at origination is not a perfect measure of present risk, but it offers a more stable reference for separating initial borrower quality from conditions observed later. This choice illustrates a broader trade-off. A variable can be imperfect yet useful if it avoids absorbing the very economic movement whose relationship the analyst is trying to estimate.
The study’s data design sets further boundaries on what that separation can accomplish. The Consumer Credit Panel is an anonymized 5 percent national sample of consumers with Equifax credit histories. The underlying observations are semiannual before 2017, quarterly from 2017 through 2019, and monthly beginning in 2020; the authors use semiannual samples throughout to maintain consistency. The analysis therefore gains a long, comparable time span but does not observe every monthly transition in the COVID period. A short-lived delinquency and an exit between observations may be represented differently from a continuously observed event.
Using a consistent semiannual interval is a trade-off between comparability and detail. It makes the long-run comparison less dependent on a change in reporting frequency, but it can compress the timing of a shock. An unemployment increase and a payment intervention may occur between two observations, while the model sees only the state recorded at the endpoints. That matters for the interpretation of a transition coefficient: the estimate captures the relationship visible at the chosen observation interval, not every short-term response that might occur inside it. A later validation exercise using a different frequency could therefore be informative, even if it should not be assumed to overturn this paper’s result.
The macro variables are also specified rather than universal. The model uses state-level year-over-year changes in seasonally adjusted unemployment and in the house price index. The authors omit state fixed effects because they find that those effects reduce prediction accuracy and are likely highly correlated with the state-level macro variables. They do not interact the unemployment change with credit-score groups, and they report that other candidates such as Treasury rates, the prime rate, and used-car price indices were not statistically significant or meaningful in their estimation. Those decisions make the baseline model tractable, but they also mean that the estimated sensitivity belongs to this particular information set.
The loan-age specification provides a useful example of why model form can matter before a macro variable enters. The authors compare exponential, Gompertz, quadratic, Weibull, and nonparametric hazard forms using fit and prediction criteria. They choose a quadratic specification because it balances fit with parsimony. A forced monotonic pattern could misrepresent periods in which loan risk rises or falls with age, and that misspecification could be mistaken for a macro effect. Choosing a functional form is therefore part of the stability problem: a coefficient can drift because the economic relationship changed, because another part of the model is misspecified, or because both are true.
What would evidence of changing conditional sensitivity look like after this baseline structure is included? It would not be enough to point to a lower aggregate default rate or a changing score distribution. The stronger test is to estimate the relationships over time, use the observed covariates to form out-of-time predictions, and examine whether the coefficient and forecast behavior change around major shocks. If composition were the sole explanation, holding the conditional relationships fixed while updating the observed mix should account for most of the difference. If the relationship itself changed, predictions based on earlier parameters would miss even after the model accounts for who is in the portfolio.
That test also has a limit. Including observed score, lender, age, balance, and other characteristics does not create a structural model of household finances or lender behavior. Unobserved product terms, servicing decisions, collateral conditions, and borrower responses can remain inside the error. The model can separate measured composition from measured sensitivity without proving that all relevant mechanisms have been observed. A persistent coefficient change is therefore informative evidence about the fitted model, not a complete explanation of why borrowers responded differently.
The practical payoff is a better diagnostic sequence. First ask whether the portfolio’s weights changed. Then ask whether the observed macro path changed. Finally ask whether the conditional transition mapping still works when those inputs are held in view. If the first question is answered and the third is never tested, a model can appear to explain a new period while simply carrying an old relationship into new conditions. The next evidence question is whether the Cleveland Fed’s rolling estimates show that kind of conditional drift through the global financial crisis and COVID-19.