Book a Demo
Most non-maturity deposit models are not wrong because banks use the wrong techniques. They are wrong because they are built in the wrong order: the answer comes first, and the data are asked to agree with it. For a decade that did not matter much. Since the deposit runs of March 2023, it matters a great deal.
Before 2022, the question asked of a non-maturity deposit (NMD) model was whether its core share and repricing maturity fitted within the regulatory caps for interest rate risk in the banking book (IRRBB) and the run-off floors for liquidity. After 2023, supervisors and management started asking whether those assumptions are right. In this article I look at why so many NMD models struggle with that second test, what the published evidence says about their weaknesses, and why I believe the root cause lies in the order of work rather than the methods. I close on what building and validating a model data first looks like in practice.
What changed after March 2023?
Supervisors said it openly. Michael Barr, then the Federal Reserve's Vice Chair for Supervision, said in a February 2024 speech that uninsured deposits "proved flightier than previously assumed", with deposit flows "much higher than assumed under" standardized liquidity requirements such as the liquidity coverage ratio (LCR). The test applied to an NMD assumption has moved from permissible to correct.
How did the low-rate decade teach models the wrong lessons?
For most of the decade before 2022, policy rates sat near zero, balances grew year after year, and deposits cost almost nothing to keep. NMD models were built, validated and approved inside that environment, and they absorbed its lessons.
The pricing lesson was the first to mislead. When market rates are close to zero, deposit rates have nowhere to fall and little reason to rise, so a regression of deposit rates on market rates finds almost no relationship. Measured betas came out small because there was almost nothing to pass through, not because depositors had stopped caring about rates. The volume lesson pointed the same way: decay analysis run on a growing book shows money that stays and keeps arriving, which supports high core shares and long behavioral maturities.
This combination changed how institutions treated the model itself. An assumption that never seemed to move drifted from the center of asset-liability management (ALM) analysis to the edge of the compliance calendar. It was refreshed when a supervisor or an auditor asked for it, and approval narrowed to two checks: whether the parameters stayed inside the caps, and whether they matched last year's.
The rate cycle broke that picture quickly. A study of US listed banks from 1997 to early 2023, published in the Journal of Financial Intermediation, found that the average deposit beta was more than 40% lower when the fed funds rate sat below 2% than when it was above it. A model calibrated only on years below that threshold had measured the low-rate half of the relationship and applied it to both.
What does the published evidence say about NMD assumptions?
Measured against the standard of being right, the published evidence points to four weaknesses in how banks set NMD assumptions. They differ widely between banks, they are rarely updated, they bear little relation to actual pricing, and they are set separately for each risk. I look at the underlying figures in more detail in what the data says about non-maturity deposit assumptions.
The dispersion is the most visible. The European Banking Authority's February 2025 report on the implementation of its IRRBB heatmap drew on end-2023 data from a sample of 120 institutions. The core share reported for non-maturity deposits ran from 0% at the 5th percentile to 90% at the 95th, with a median of 52%; for retail NMDs alone the median was 59%. Deposit bases differ, and some spread is expected. A range this wide among institutions under the same rules, often in the same markets, is hard to explain by depositors alone. The EBA's follow-up work is covered in our piece on the EBA IRRBB heatmap.
The staleness is documented in a European Central Bank working paper that followed 67 significant institutions, representing around 72% of the euro area banking sector, from mid-2019 to the third quarter of 2023. Only about half changed their NMD modeling assumptions in that period, and the share of banks updating their models was no higher after monetary tightening began in July 2022 than before it.
The disconnect from pricing may be the most damaging. The same EBA report noted that deposit remuneration rose at most institutions in 2023 while behavioral models calibrated on long histories were left unchanged, so that "no clear relationship is observed between the NMD behavioral assumptions calibrated with historical data and the pass-through rates of the NMD in the current IR environment."
The fourth weakness is the least measured, and here I can only speak from practice. In my experience advising banks on ALM and structural risk, most institutions still run static, rarely recalibrated NMD assumptions, and they set the interest rate risk and liquidity versions separately, often inconsistently. The same deposits are treated as long-term money in one report and as a flight risk in another, with no single view that reconciles the two.
Why does the order of work explain every symptom?
Each of these weaknesses has a ready explanation. Dispersion looks like a data problem, staleness like a process problem, the pricing gap like a calibration problem, and siloed assumptions like an organizational one. None of those explanations accounts for why the four appear together.
The common cause is the order in which the work is done. In many institutions the process starts from a conclusion, and the data calibrate parameters toward an answer that is already known. My diagnosis is blunt: we are working the models the wrong way around. The bank's expert has a feeling about how their deposits behave and builds, or hires an external consulting firm to build, a model around it that gives the expected result and satisfies regulators and auditors. That carries the highest possible model risk.
This order of work explains every symptom in the evidence. Assumptions differ widely because they encode different experts' priors more than different depositors. They are rarely updated because a model that returns the expected answer gives nobody a reason to reopen it. They bear little relation to pricing because pricing was never what they were fitted to. And they differ between interest rate risk and liquidity because each team brings its own prior to its own metric.
The extreme version surfaced in the Federal Reserve's April 2023 review of Silicon Valley Bank. In April 2022 the bank lengthened its deposit duration assumptions, a change the review describes as made to address an interest rate risk limit breach rather than to manage the actual risk. In November 2022 supervisors formally found its interest rate risk simulations "not reliable". The measure moved. The risk did not. The full case is covered in SVB's IRRBB anatomy.
Why does validation rarely catch the problem?
Most institutions stop well short of that case, although the mechanism is the same in milder form. A model that is simply wrong can be caught by testing it against data; a model built to confirm a prior is judged by its agreement with that prior, so it cannot reveal that the prior is wrong.
Validation often reinforces the pattern rather than breaking it. The same expert frames the question, and the validator tests the model on the same history it was fitted to. Validation then confirms that the model reproduces what it was designed to reproduce, and the caps confirm that the answer is permissible. Neither tests whether deposits actually behave that way.
The Prudential Regulation Authority's supervisory statement SS1/23 on model risk management, in force since May 2024 and lightly amended in April 2026, sets expectations that run directly against three habits common in NMD practice:
- Validation that is not independent of development, when SS1/23 requires an independent validation function with the standing to challenge
- Boards that approve models without understanding their capabilities and limitations, which SS1/23 expects the board to understand for its most material models
- Recurring post-model adjustments for the same limitation, which SS1/23 treats as a possible sign of flawed model design that may call for recalibration or redevelopment
Together, these habits describe a model that nobody outside its builders is in a position to reject. Overlays deserve particular attention in deposit modeling, because an adjustment added each year to bring a stale beta closer to observed pricing leaves the model untouched while its outputs look current.
How Should Banks Model Non-Maturity Deposits After 2023?
What does data first building look like?
The remedy is to reverse the order, and it does not require new techniques. Segmentation, core estimation, decay and survival analysis, beta regression and replicating portfolios can confirm a prior or learn from data. The same five steps serve both purposes; the checkpoints between them decide which.
The whitepaper "Non-Maturing Deposits Under Pressure" follows each technique in turn and sets out what building it data first involves. Across all five, the discipline comes down to three habits:
- Each estimate is tested on data the fit did not use, preferably a period with a genuine rate move
- Parameters such as lags, confidence levels, segment boundaries and ladder maturities are chosen by the fit, or fixed in advance and justified
- Every override of the data is written down with its reason and its effect
Each habit closes a door through which a prior can enter unnoticed. Core estimation shows why. Its most important inputs, the confidence level and the horizon, are policy choices, and setting them after seeing the core they produce is the easiest way for a prior to return. A data first build fixes both before the estimate is run and lets the fit determine the core share.
Round numbers deserve suspicion for the same reason. Take a static decay of 1.67% a month on the stable balance. That rate is one sixtieth of the balance each month, a straight-line run-off over exactly five years. When a fitted number sits that close to a round convention, the model owner should be able to show that the data put it there.
How should NMD validation be designed so the model can fail?
Validation is the point at which the order of work is either enforced or quietly skipped. US supervisors set out the expectation in 2010: the interagency advisory on interest rate risk management (SR 10-1), still in effect, reminds institutions to "document, monitor, and regularly update key assumptions" and expects independent review to cover the reasonableness of assumptions and "the backtesting of assumptions and results". The controls are familiar; what makes them work is that each one is capable of rejecting the model.
Backtesting for NMDs is hard, because histories are short relative to the behavioral maturities claimed and the low-rate decade supplies almost no observations of a rising rate cycle. These difficulties explain weak backtesting; they do not excuse it. The direct question remains whether the model would have predicted what the bank's deposits did in 2023.
Sensitivity testing then shows where the evidence matters most. Lengthening the behavioral maturity of core balances lowers the reported economic value of equity (EVE) loss in a rising rate shock, justifies a longer hedge and, on an upward sloping curve, raises the transfer price credited to the units that gather deposits. A lower assumed beta improves projected net interest income (NII) in the same scenario while leaving the liquidity buffer untouched. (For how the two metrics pull against each other, see EVE vs NII explained.) These are the assumptions that deserve the most evidence.
Consistency across risks is the last control, and the one siloed practice undermines. Some difference between IRRBB and liquidity assumptions is legitimate, since a stress horizon of days asks a different question from a repricing horizon of years. The control is one set of behavioral assumptions, or explicitly reconciled ones, documented in one place and owned by one function, so that the asset and liability committee (ALCO) can see where the IRRBB, liquidity, funds transfer pricing (FTP) and planning views differ and why.
What changes for ALCO and the board?
For institutions, this changes what an NMD model has to demonstrate. Fitting within the caps and passing audit show only that an answer is allowed. After 2023, supervisors and boards want evidence that the answer came from the deposits themselves, and a model engineered around the expert's expectation cannot supply it.
The written record of a data first model also changes who can judge it. Management, internal and external auditors and supervisors can read where the data ends and judgment begins. A model built to confirm a prior cannot survive that reading, because its overrides carry the answer.
The same estimate sets the liquidity buffer, the reported EVE and NII, the hedge and the price paid for deposits. Institutions that keep treating it as a calibration to be refreshed on request will carry the largest judgment on their balance sheet with the least evidence behind it.
Key takeaways
- Since 2023, the test for an NMD assumption has moved from whether it fits within the caps to whether it is right.
- Published EBA and ECB evidence shows wide dispersion, infrequent updates and little link to actual deposit pricing.
- All four weaknesses share one cause: models built to reproduce an expert's expected answer.
- Data first building and validation that can reject the model reverse that order without new techniques.
For the full toolkit, from core balance estimation and deposit betas to replicating portfolios, and how Basel, EU, UK and US rules cap these assumptions, see Mirai RiskTech's whitepaper Non-Maturing Deposits Under Pressure: how banks model non-maturity deposits for IRRBB and liquidity.
