Why ALM Back-Testing Matters (and How to Do It Right)

After my post on CECL back-testing sparked a lot of good discussion, I wanted to follow up with a similar look at how institutions can approach back-testing for their ALM models.

Here is a simple, practical way to think about ALM back-testing as well.

ALM models are heavily assumption-driven. Prepayments, deposit betas and decay rates, reinvestment strategies, pricing assumptions, etc. Any one of these can meaningfully change the risk profile of the balance sheet. And because so many decisions flow directly from these assumptions, even small inaccuracies can add up fast.

That’s why back-testing your ALM model is so important. It’s how you make sure the assumptions still make sense and the model is behaving the way you expect.

A clean approach starts with comparing what the model projected for interest income and interest expense for a future period, say, one year out, to what actually happened (many institutions do this quarterly too). It is important to do this at the balance-sheet category level: major loan categories (residential mortgage, CRE, C&I), the key investment portfolio types, non-maturity deposits, CDs, and borrowings.

Doing this at the category level is critical because net interest income can look fine on the surface while large offsetting variances occur between assets and liabilities. Breaking the analysis down by major account categories lets you see where the material misses actually occurred, rather than letting them cancel each other out.

From there, a rate/volume analysis helps you pinpoint what drove the variances:

 • Were actual balances different than the model assumed?

 • Were actual asset yields or deposit costs higher or lower than projected?

This helps you isolate the drivers of material variances so you can tweak key assumptions as necessary.

It also helps to establish clear variance thresholds so you’re focusing on what actually matters. Not every difference is meaningful, and setting materiality limits keeps you from spending time researching small variances that have no real impact on the model or the decisions it informs.

One enhanced method I like is to go back to the prior modeling period and re-run the model using the actual monthly balances and actual yield curves for the period you’re evaluating. Since balance sheets are never truly “flat” and nobody can forecast the path of interest rates perfectly, this isolates the variances that were truly assumption-driven.

At the end of the day, ALM back-testing isn’t about judging the model on how well it predicted the future. It’s about understanding why results differed and which assumptions need to be refined so the model continues to support sound decision-making.

And just like with CECL, that feedback loop is where the real value is.