A new study from FDIC staff challenges the predictive models on which much mortgage underwriting increasingly depends, discrediting both workhorse predictive models used across the spectrum of consumer credit as well as the ML-based mortgage models that fare still worse as macroeconomic, policy, and borrower conditions change. The paper concludes that the analytical methodologies in default-risk models are ill-designed for the post-2008 mortgage market, empirically demonstrating that predictive-error ranges reach striking proportions from 2004 through 2016.

The full report is available to subscription clients. To find out how you can sign up for the service, click here.