We examine the challenges of economic forecasting and model misspecification errors faced by financial institutions implementing the Current Expected Credit Losses (CECL) allowance methodology and its impact on model risk and bias in CECL projections. We document the increased sensitivity to model and macroeconomic forecasting error of the CECL framework with respect to the incurred loss framework that it replaced. An empirical application illustrates how to leverage simple machine learning strategies and statistical principles in the design of a nimble and flexible CECL modeling framework. We show that, even in consumer loan portfolios with tens of millions of loans (eg, mortgage, auto or credit card portfolios), it is possible to develop, estimate and deploy an array of models quickly and efficiently, without a forecasting performance penalty. Drawing on more than 20 years of auto loans data and experience from the 2007–9 Great Recession and the 2020–21 Covid-19 pandemic, we leverage basic econometric principles to identify strategies to deal with biased model projections in times of high economic uncertainty. We advocate for a focus on the resiliency and adaptability of models and model infrastructures to novel shocks and uncertain economic conditions.