Macroeconomic Parameter Instability in Auto Loan Loss Models

We estimate a discrete-time Markov transition model of auto loan performance over the 2000-2025 period using multinomial logistic regressions. Using rolling pseudo out-of-sample forecasts, we document persistent declines in the sensitivity of the probability of default to an unemployment shock in both the global financial crisis and Covid recessions. The estimated effect of a 1 percentage point increase in unemployment on default risk declined from 16 percent in 2006 to 3 percent post-2020. Two-year cumulative default forecasts over 2020-2021 using pre-pandemic parameters overstate actual defaults by 100 basis points (25 percent), with forecast errors largest in absolute terms for subprime borrowers and largest in relative terms for prime borrowers. The instability persists after controlling for forbearance usage and pandemic-period dummy variables, and is driven primarily by changes in macroeconomic relationships rather than borrower composition. These findings have implications for stress testing models and loss forecasting practices that rely on stable unemployment-default relationships.