Forecasting Non-Performing Retail Loans During the Covid-19 Pandemic : the Effect of Forbearance on Model Error

The extreme economic conditions exhibited during the COVID-19 pandemic in 2020, in conjunction with the historic government policy response, posed an enormous challenge to the accuracy of credit-loss forecasting models trained on loan performance data prior to the pandemic. In this paper, we investigate the extent of “model risk” associated with the pandemic in two retail portfolios: mortgage and credit card. After constructing a simplified PD-EAD-LGD framework on pre-2019 loan-level data for each product, we produce a forward looking four-quarter forecast through 2020 from the perspective of portfolios observed in 2019Q4 using actual macroeconomic conditions during the pandemic. In both cases, we find massive over-prediction of credit losses relative to actuals similar to what many banks have observed. Although there are several potential explanations for this inaccuracy, we explore the role of forbearance policies that permitted borrowers to suspend payments without penalty. That is, we study whether the ability of consumers to delay payment on mortgage and credit card balances “artificially” created low rates of loan non-performance that, otherwise, would have been significantly higher. Our results show that the potential “impact” of forbearance on model performance is quite different between the two products, being much larger for mortgage than card. Such results illustrate important differences in consumer behavior with respect to these two products during times of economic stress that may further inform model risk management in the presence of future shocks