A Practical Approach to CECL Model Back-Testing

I’ve had several conversations with CECL validation clients over the past few months about model back-testing. Institutions often bring it up because an examiner or auditor asked about it, and I always try to shift the focus to why it matters beyond just checking a regulatory box, and how to build a practical approach that actually adds value.

For a while, CECL back-testing didn’t feel all that relevant because losses were so low. But as certain portfolios start to show a little more stress, it’s becoming important again to make sure the model is still lining up with reality.

The good news is that back-testing doesn’t need to be complicated. At its core, it’s just comparing what the model forecasted for losses with actual charge-offs experienced. 

The key is to do this at the segment level. While portfolio-level results offer a high-level reasonableness check, they can easily hide what’s going on underneath. One segment can be under-reserved while another is over-reserved, and the offset masks potential weaknesses, biases, or calibration needs within specific CECL segments.

Just to be clear, everything I’m describing here refers to back-testing the model’s overall reserve forecast at the segment level and not the more complex back-testing of underlying statistical models like PD or other predictive components. 

A clean, practical approach to back-testing your CECL model reserve forecast is to compare each segment’s one-year forecasted reserve to its actual charge-offs for that same year. Some institutions do this quarterly, which provides an even better read on how assumptions are holding up over time.

A common mistake I often see is comparing a segment’s lifetime reserve forecast (the number the model reports) to one year of actual losses. That will always distort the picture because you’re comparing a lifetime estimate to a single-year outcome.

A simple way to deal with this is to take the lifetime reserve forecast for each segment and divide it by the segment’s average life. That gives you a clean, annualized forecast you can directly compare to actual charge-offs. Suddenly, the back-test becomes accurate, fair, and actionable.

At the end of the day, the goal of these models is not perfect prediction as almost no model can do that given how many variables and uncertainties are at play. What matters most is understanding where the balance sheet may be vulnerable and how risks might evolve.

Back-testing is one of the best ways to keep that perspective sharp. It helps you monitor performance, spot emerging weaknesses, and refine assumptions over time so the allowance stays aligned with the actual risk in the portfolio.