Seeing inside the black box: ALM, deposit and CECL models brought in-house at a mid-size US bank

For a bank of this size the asset-liability model is one of the most consequential pieces of software it uses, and many banks rent it. This one sent data tapes to a specialist vendor each quarter and received PDF reports about three weeks later. Treasury could not see the assumptions behind the numbers, could not run its own scenarios between quarter-ends, and paid for every custom analysis. After the March 2023 failures, the committee wanted to know what a deposit beta fifteen points above the default would do to EVE, and nobody could answer quickly.

The supervisory side added pressure. Under SR 11-7 and OCC Bulletin 2011-12, a model has to be developed, validated and governed with documentation an examiner can follow, including sensitivity testing of the assumptions that matter. Vendor defaults for deposit decay and beta were industry averages, not the bank's own behavior, and CECL used a second vendor's prepayment speeds. Our team rebuilt the analytical stack on published methods and open-source libraries and fed it the bank's own instrument-level data, without using any vendor's proprietary code or report layouts.

The engine is a Ho-Lee term-structure model calibrated to today's Treasury and SOFR curves, with a Hull-White volatility fit for option-sensitive instruments, on QuantLib and the Open Source Risk Engine. Behavioral models sit on top: prepayment and default hazards fitted on public GSE loan-level data and recalibrated to the bank's portfolio, and non-maturity deposit decay and asymmetric betas fitted on fourteen years of account-level history. Peer data from FFIEC call reports steadies the thin products. EVE, net interest income, key rate durations, liquidity runoff and CECL all run on one cash-flow engine.

Deposit behavior was the finding that changed the answer. The bank's money-market beta in the 2022 to 2023 cycle ran well above the vendor default, and its savings balances decayed more slowly. Together these changed the answer. For a 200‑basis-point rise, EVE fell about 9.4 percent in the in-house model against 4.3 percent in the vendor report, and the 12‑month NII gain from the same rise shrank from 3.2 percent to 0.9 percent. In a 200‑basis-point fall the gap was smaller: EVE rose 2.6 percent against 1.1. The bank took it to its examiners on its own initiative. The numbers in this illustrative case are modeled, not measured, but the lesson is general: pricing math is published and takes weeks, while behavior takes months and decides the result.

Trust came from a parallel run. For two quarter-ends the new models ran beside the vendor, and base-case EVE agreed within 1.9 percent, every difference above one percent of EVE got a written explanation, and an independent validator reviewed the lot. Only then did the bank's staff run a close unaided and let the vendor contract lapse. Everything runs inside the bank's own tenancy with role-based access, version control and reproducible outputs, so any past committee number can be recreated for an auditor. Public data supplies nearly everything except option volatility, which needs one modest market-data license.