At this time, the CECL accounting standard is required for non-public entities with assets of $10 million or more. As smaller lenders grapple with this, it reminds me of how predictive modeling concepts are now, even infiltrating departments like the Accounting Department.
Risk management is really about optimization. It’s about sharpening one’s operations to make better use of tools that drive efficiency. As the concepts of risk management continue to permeate deeper into the lending industry, I thought I’d provide some clarity on the subject as it pertains to CECL, and what’s expected of its use and implementation.
The big “change” is that now even smaller lenders are forced to use some form of predictive modeling in their operations. But don’t fret, that doesn’t mean you need to build complex ML scoring models or anything like that. But it does mean you need a way to predict a newly originated loan’s future cumulative net loss (CNL) rate at the end of its term.
The CNL rate is quite literally the amount of lost principal on a pool of loans after the pool has aged to completion. For example, an EOL CNL rate of 16% means 16% of the principal that was loaned out was never recovered. Does that mean the lender lost money? Not if they collected enough interest + fees to overcome that loss.
Here’s an (outdated) slide of the subprime auto industry’s CNL curves as published by S&P Global:
It’s customary to chart-out the CNL curves by annual vintage at each month-end. And here’s an example of 3 individual vintages from a recent consult we did here at Lendisoft:
Notice how these curves begin to take shape, and how you can visualize how to “draw-in” the remainder of each curve, taking influence from neighboring curves, and the persistent shape of what CNL curves typically look like. This is a monthly exercise you should be doing on each vintage of your own portfolio.
Well that’s great, we now have “forecasted” CNL curves at end of life for each vintage of our portfolio. Now what? Next, we need to do this same thing as we drill-down into smaller segments of each vintage, preferably using a highly predictive variable like FICO score, LTV, or something else (as an aside, at Lendisoft, we can perform a Characteristic Analysis to determine precisely your most powerful variables and their optimal break points).
For example you might produce CNLs based on 7 or more different ranges of FICO score. You’ll likely notice that lower FICO ranges have higher CNLs and vice-versa. Congratulations, you’ve just built your CECL model. From there, you can use that model to estimate future CNLs for all active loans according to the individual FICO scores.
Now with all model-building, there are more advanced steps that need to be considered; things like normalizing the model so it’s not overfit to the sample data and works well on out-of-time data; and things like accommodating “business” adjustments to the model when necessary. But I hope this article demystifies some of the task before you.
And if you’d like help with any of this, Lendisoft offers a very affordable “Valuations” service where we’ll build a sharp model indeed, and run these numbers for you each month, complete with total loss provision dollars for your active portfolio. The service also includes an estimate of your portfolio’s worth, based on forecasted future payments. Lendisoft is your one-stop shop for everything related to risk management for your lending institution. Contact us today!