Overview
Using the Cleveland Fed's 2000–2025 auto-loan study and its anonymized 5% national sample of consumers with Equifax credit histories, this series explains how state transitions, portfolio mix, rolling estimates, robustness tests, and forecast architecture shape stress results. It shows that unemployment sensitivity fell after the global financial crisis and COVID-19, while the preliminary evidence supports model-specific instability—not a universal causal law—and turns that boundary into a validation approach.
Learning outcomes
- A stable-looking unemployment coefficient is a hypothesis, not a permanent constant.
- In the Cleveland Fed model, the current-to-default relative-risk effect fell from about 16% in 2006 to about 3% after 2020 and stayed low through 2025.
- Portfolio mix, macro conditions, and conditional relationships can move aggregate defaults in different ways.
- Markov state transitions improve average forecasts here, but better architecture does not repair stale macro parameters.
- Robustness tests narrow explanations for the break without identifying one causal mechanism.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
Why can an unemployment coefficient stop traveling across recessions?
In the Cleveland Fed auto-loan model, it can stop traveling when the shock, institutions, borrower or lender behavior, or conditional mapping changes across regimes. The rolling evidence shows persistence, but not one identified cause.
What does the Cleveland Fed model's 16% to 3% pattern mean?
It means the estimated current-to-default relative-risk effect associated with a one-percentage-point annual unemployment change weakened from about 16% in 2006 to about 3% after 2020. These are conditional ratios, not percentage-point default-rate increases, and unemployment may still matter.
Can changing portfolio composition explain lower aggregate defaults?
It can, but the Cleveland Fed study's December 2019–2021 decomposition assigned about 18% of the predicted change to borrower composition, versus roughly 60–70% to macro data and 20–30% to coefficient changes. These are counterfactual attribution terms, not causal estimates.
Does a Markov model solve the auto-loan forecast problem?
No. In this evidence, it better represents cures, exits, and delinquent-state paths than a default-only model, but it still overpredicted COVID-period outcomes because the unemployment relationship changed.
How should lenders test whether a stress model still works after a shock?
They should combine rolling or expanding-window estimates, pseudo out-of-sample and out-of-time forecasts, segment-level outcomes, level and percentage error measures, differential diagnosis, and effective challenge. The response should match materiality and documented limitations.