Solutions to specification errors in stress testing models

The regulatory and business need to expand the use of macroeconomic‐scenario‐based forecasting and stress testing in retail lending has led to a rapid expansion in the types and complexity of models being applied. As these models become more sophisticated and include lifecycle, credit quality, and macroeconomic effects, model specification errors become a common, but rarely identified feature of many of these models. This problem was discovered decades ago in demography with Age‐Period‐Cohort (APC) models, and we bring those insights to the retail lending context with a detailed discussion of the implications here. Although the APC literature proves that no universal, data‐driven solution is possible, we propose a domain‐specific solution that is appropriate to lending. This solution is demonstrated with an auto loan portfolio.