Treat Stability as a Hypothesis
The Cleveland Fed auto-loan case answers the opening question with a qualified but consequential finding: a relationship that appeared useful for translating unemployment into default risk can weaken after a large economic shock. The final issue is not whether to keep or discard one coefficient. It is how to recognize when a conditional relationship has stopped traveling well, determine which part of the model or its use is responsible, and respond in proportion to the decision at stake. The joint 2026 guidance on model risk supplies principles for that judgment, while the auto-loan evidence supplies the concrete failure surface.
Why the unemployment-default relationship weakened
The most defensible explanation has three layers. First, the macroeconomic shock changed. Unemployment rose sharply during COVID-19, but auto-loan defaults declined in the first semiannual period rather than following the pattern observed in earlier recessions. Second, the institutions and behavior connecting unemployment to repayment were unusual. The period included very large federal transfers and historically high auto-loan forbearance, among other changes. Third, the estimated conditional relationship changed: the rolling current-to-default unemployment effect fell from roughly 16 percent in December 2006 to slightly below 10 percent after the global financial crisis and to about 3 percent after 2020, where it remained through 2025.
Those layers should not be collapsed into the sentence “the pandemic made defaults lower.” That describes an outcome, not the model problem. The problem is that the old mapping from an unemployment input to a transition probability produced a sharp increase in predicted current-to-default movement when the observed outcome moved in the opposite direction. After new observations entered the rolling estimation, the model learned a lower sensitivity. The learning process is evidence that the previous conditional relationship did not fit the later regime, even though it does not identify which real-world mechanism changed it.
Portfolio composition is part of the explanation that must be separated out. Risk score, lender channel, loan age, borrower characteristics, and other features produce large baseline differences in transition risk. The study therefore does not infer parameter instability simply from a changing aggregate default rate. Its decomposition of December 2019 versus December 2021 assigns roughly 60–70 percent of the predicted change to macroeconomic data, 20–30 percent to coefficient changes, and about 18 percent to borrower composition. The exact shares are counterfactual attribution terms, but their ordering supports the conclusion that the observed borrower mix was not the primary reason the pre-pandemic model overpredicted.
Forbearance also remains important without being the whole answer. Predictions for all loans and for loans that never entered observed forbearance almost perfectly overlap, including the large 2020 errors. Adding a dummy for the 2020–2021 period makes the unemployment coefficient decline more slowly and settle somewhat higher, but the decline and persistence remain. These tests make it difficult to defend the old relationship by pointing only to the observed forbearance group or to a common calendar-period effect. They do not show that transfers, payment relief, servicing changes, or borrower behavior were irrelevant. They show that the available robustness checks do not isolate one of them as a complete explanation.
The distinction between what changed and what did not is the core judgment. The evidence supports a persistent change in the fitted macroeconomic relationship after two large shocks. It does not support the stronger claims that unemployment no longer affects repayment, that the post-2020 coefficient is a new universal constant, or that one intervention caused the entire break. The paper is a preliminary reduced-form study. Its estimates are conditional associations in a specified sample and model, not a structural account of household finances, lender incentives, or public policy.
The state-transition architecture explains why the break matters beyond a coefficient plot. A current-to-default parameter sits inside a row of competing paths: a loan may remain current, become delinquent, default, or exit. A change in that parameter alters the distribution entering later periods.