Incorporating lifecycle and environment in loan-level forecasts and stress tests

The new FASB current expected credit loss (CECL) proposal, IASB's IFRS 9, and regulatory stress testing all require that the industry move toward forecasting probabilities of future events, rather than simply rank-ordering loans. Even more importantly, effective loan pricing requires this same forward-looking, loan-level forecasting. We created a loan-level version of Age-Period-Cohort (APC) models suitable for forecasting individual loan performance at a point-in-time or for the loan's lifetime. The APC literature explains that any model of loan performance must make either an explicit or implicit assumption around the embedded model specification error between age of the loan, vintage origination date, and performance date. We have made this assumption explicit and implemented a technique using augmented macroeconomic history to stabilize the analysis. The preceding steps provide robust estimates of lifecycle and environmental impacts. We then use a Generalized Linear Model (GLM) with a population odds offset for each age/time combination derived from the lifecycle and environment functions in order to estimate origination and behavior scores. Analyzing a small US auto loan portfolio, we demonstrate that this model is robust out-of-sample and out-of-time for predicting both rank-ordering and probabilities by inserting the odds offset appropriate for the environment being modeled. In addition to producing loan-level forecasts and stress tests, the scores produced have higher rank-order performance out-of-sample and out-of-time than standard scores. The scores prove to be robust years into the future with no measurable degradation in performance because of the stabilizing effect of the offset factor during model construction. (C) 2016 Elsevier B.V. All rights reserved.