When an Economic Shock Changes the Model’s Meaning
The subject of this series is a practical question in auto-loan risk modeling: what happens when a relationship that once connected unemployment to default no longer describes what borrowers do? The Cleveland Fed’s Working Paper 26-18 studies that question with a loan-performance model covering 2000–2025. Its focus is not simply whether auto loans became safer or riskier, but whether the model’s estimate of how an unemployment shock changes loan performance remained stable across two very different recessions. That distinction gives readers a way to judge stress forecasts, rather than merely repeat their outputs.
The scale of the market makes the question more than a statistical curiosity. The paper notes $1.56 trillion in outstanding U.S. auto loans in the fourth quarter of 2025, while its analysis uses an anonymized 5 percent national sample of consumers with Equifax credit histories. The sample is large enough to examine transitions across borrower and lender groups, but it is still a particular measurement of a large market. Its result should therefore be read as a serious warning about model behavior, not as a claim that every lender’s portfolio has the same coefficient or the same break.
The stakes are concrete. Lenders and regulators use multi-period forecasts to estimate losses, assess capital needs, and compare risk across portfolios. If unemployment rises, a model trained on earlier experience may predict more delinquency and default. That prediction can be sensible when the old relationship still holds. But if payment relief, government transfers, servicing practices, or other conditions change how an economic shock reaches household cash flow, the same coefficient can turn a severe-looking macro scenario into an overstated loss forecast. A model can be mathematically correct within its design and still become unreliable when the world that supplied its relationships changes.
There is also a scale issue inside the statistic. A relative-risk estimate compares the chance of moving to a destination with the chance of remaining in the current state, holding the specified covariates fixed. A 16 percent change in that ratio is not a 16-percentage-point default rate, and its translation into an actual probability depends on the starting state and all the other transition choices in the model. That is why a coefficient can be economically important without determining a portfolio loss number on its own. The model’s meaning lies in how the estimate interacts with the state distribution and the forecast horizon.
The Cleveland Fed study reports a striking version of this problem. In its current-to-default transition model, the estimated relative-risk effect associated with a one-percentage-point increase in the annual change in unemployment was about 16 percent at the start of the rolling estimates in 2006. It fell to slightly below 10 percent after the global financial crisis and to about 3 percent after 2020, remaining low through 2025. These figures are not percentage-point increases in the default rate, and they are not proof that unemployment ceased to matter. They are conditional estimates in one reduced-form model. Their importance is that the estimated mapping from a macro shock to a loan transition changed, and stayed changed.
The real object of study is a relationship inside a model
It is tempting to describe the issue as “unemployment predicts default less well now.” That phrasing is close, but incomplete. The study is about a model of auto-loan performance and the relationship it estimates between macroeconomic conditions and movements through loan-performance states. The model includes borrower and loan characteristics, lender type, and state-level changes in unemployment and house prices. It then asks how well parameters estimated with past data forecast the next period as new observations arrive.
That setup separates three things that can produce the same visible change in aggregate defaults.
- The portfolio may change. The mix of credit scores, lender channels, loan ages, and borrower characteristics may shift. A portfolio with more prime borrowers could show fewer defaults even if each borrower faced the same conditional risks as before.
- The economic shock may change. A recession can be larger, faster, or accompanied by policies that were absent in earlier downturns. The observed path of unemployment, house prices, and household support can differ materially across episodes.
- The conditional relationship may change. Holding the model’s observed characteristics in view, the same measured unemployment movement may be associated with a different transition risk because the institutions and behavior connecting job loss to payment failure have changed.
Those explanations are related, but they are not interchangeable. A model that fails to distinguish them may “learn” the wrong lesson from a period of unusual performance. It might attribute a lower aggregate default rate to a safer borrower mix when the more important change was a different response to unemployment. Or it might keep an old coefficient in place and mistake an exceptional policy environment for an ordinary stress response.
The evidence does not make the third explanation a universal law. The paper is a preliminary working paper, and its regressions estimate conditional associations rather than causal effects of unemployment, forbearance, transfers, or lender behavior. The useful claim is narrower: in this national sample and specification, the unemployment coefficient changed substantially after large shocks, and the change had material forecasting consequences. That is already enough to challenge the habit of treating a familiar coefficient as permanent.
The diagram’s three inputs are not three competing names for “the economy.” Composition describes who and what is in the portfolio. Macroeconomic conditions describe the shock path applied to it. The conditional relationship describes how the model translates those conditions, given its other inputs, into transition probabilities. A forecast change can reflect any combination of the three. A credible diagnosis must therefore ask which input moved, which mapping moved, and which distinction matters for the decision at hand.
Why a stable-looking coefficient deserves scrutiny
Historical stability is attractive because it makes a model portable. If a one-percentage-point unemployment increase has repeatedly been associated with a particular increase in default risk, a lender can use that estimate to extend a forecast beyond the observations used to fit the model. Stability is not an unreasonable working assumption. It is a useful hypothesis about whether the same economic mechanism continues to operate in a similar way.
The danger is treating that hypothesis as a property of the coefficient itself. A coefficient is not a physical constant. It summarizes relationships observed under particular credit products, servicing practices, borrower expectations, policy regimes, and patterns of household income and expenditure. Large shocks can alter several of those conditions at once. The model may then need new data to learn that the old relationship has weakened—or it may continue to use the old relationship until its forecasts visibly miss.
The two recessions in the study make this distinction especially important. During the 2007–2009 recession, the model’s unemployment sensitivity fell from its earlier level. During the short but extraordinary COVID-19 recession, unemployment spiked while auto-loan defaults declined rather than rising as the pre-existing relationship would have suggested. The study’s rolling design allows the model to be re-estimated as each semiannual observation arrives, revealing that the post-shock coefficient did not quickly return to its previous level.
This is why stress testing cannot be reduced to applying a severe unemployment path to a fixed portfolio. The path may be severe, but the transmission from that path to default may not be the same as it was in the training period. At the same time, abandoning macroeconomic variables altogether can conceal useful information and worsen forecasts for other transitions. The relevant judgment is not whether macro variables are “good” or “bad.” It is whether their estimated relationships remain fit for the purpose, horizon, and portfolio to which they are being applied.
What this series will keep separate
The first separation concerns risk level versus sensitivity. Credit score and lender channel can explain large differences in baseline default risk. That does not establish that every channel responds to unemployment in the same way, or that a change in aggregate defaults is a change in macro sensitivity. The paper finds that auto-finance companies concentrate more subprime lending and that conditional risk differs by lender type. Those facts make composition important, but they also make aggregate results harder to interpret without segment-level analysis.
The second separation concerns one-period transitions versus cumulative outcomes. An unemployment coefficient enters a transition probability, not a final two-year loss number by itself. Loans can become delinquent, cure, remain delinquent, default, or exit without default. As those states change, the population exposed to later transitions changes as well. A modest one-period miss can therefore accumulate differently across a portfolio than a direct default-only calculation would suggest.
The third separation concerns evidence of instability versus evidence of cause. Rolling estimates can show that the relationship changed. Comparisons excluding observed forbearance, pandemic-period controls, and decompositions of borrower composition can narrow possible explanations. They do not by themselves prove which policy, behavior, or institutional change caused the break. Good model judgment depends on preserving that difference. It determines what a lender should test next and how confidently they should alter a model.
The practical question at the center of this series is therefore disciplined rather than dramatic: when a stress model’s familiar unemployment-default relationship weakens, can its users tell whether the problem lies in portfolio mix, state-transition architecture, macroeconomic parameters, or the conditions under which the model is being used? Answering that question requires a model of how auto loans move through risk states, evidence about what happened in the data, and a clear view of what the evidence cannot establish. Only then can a forecast be challenged in a way that improves decisions instead of replacing one untested assumption with another.