Over the past three years, machine learning has entered the published work on deposit behavior, from peer-reviewed journals to central bank and supervisory papers. Supervisors have begun to write about how it fits within model risk management, and vendors now present it as the natural next step for non-maturity deposit (NMD) models.
The stakes are high because the NMD model travels far. As I set out in why one assumption drives the balance sheet, the same behavioral estimate of how much of the balance is stable, and for how long, sets the liquidity buffer, the economic value of equity (EVE) and net interest income (NII) reported under interest rate risk in the banking book (IRRBB), the hedges treasury puts on and the price each deposit receives through funds transfer pricing (FTP). In this article I look at what machine learning could do for deposit models, why I believe it is still the wrong core model for a bank today, and where AI does earn its place.
Most of the techniques banks use for NMDs rest on a limited set of explanatory variables and on relationships whose shape the modeler specifies in advance. Deposit behavior depends on far more than that: the level of rates, competition, the channel through which customers bank, how concentrated and how insured the balances are. No regression was designed to carry all of those signals at once.
Machine learning addresses that limit directly. Classification models such as random forests and gradient boosted trees could sort accounts into behavioral segments from the data itself and predict which balances are likely to leave, rather than relying on boundaries drawn by product and customer type. Neural networks could correlate far more variables than any regression can hold and learn nonlinear relationships without the modeler fixing their form beforehand. In my view this is the most disruptive application AI has for deposit modeling, and for most banks today it is too disruptive. The rest of this article explains why.
The evidence is thin, but it points one way. A 2026 study in the Journal of Risk Model Validation compared ARIMA and GARCH time-series models with a long short-term memory neural network and a random forest for forecasting demand deposit balances and attrition curves under IRRBB, using a sample of balances from Mexican banks. The machine learning models improved accuracy, and the random forest gave the best fit, with the time-series models struggling most under negative rates.
A 2023 paper in Frontiers in Artificial Intelligence built "Deep ALM", a neural network framework developed with a Swiss retail bank that sets monthly investment and funding decisions, with non-maturity deposits as one of its central difficulties. In the authors' experiments, the dynamic strategies it found outperformed static benchmarks, although they list the issues still to resolve before deployment. Beyond those two studies, the evidence is adjacent or untested, and vendor claims about how far linear models misstate deposit value should be read as claims.
Even with those caveats, machine learning can describe deposit behavior more closely than the regressions and replicating portfolios most banks run. It does so because it lets the data choose the relationships, which is exactly the order of work I argue deposit models need in non-maturity deposit models built the wrong way around. A model that fits behavior more closely leaves less room for error in the assumptions it produces, so the common objection that new tooling adds model risk has the arithmetic backwards.
The case for machine learning is strong on its own terms, and it still fails the test that decides whether a model can run a bank's deposit book. The model that sets interest rate risk, the liquidity buffer and deposit pricing has to be understood by those who answer for it: the asset and liability committee (ALCO) and the board that approve its assumptions, the internal and external auditors who sign it off, and the supervisors who review it.
The Prudential Regulation Authority's SS1/23 on model risk management expects documentation sufficient for an independent third party to understand how a model operates, and it expects the board to understand the capabilities and limitations of its most material models. A model that its approvers cannot follow is a model they cannot challenge, and a model nobody can challenge has not really been validated.
SHAP (Shapley additive explanations) has become the standard answer to the black-box objection. It breaks each prediction down into the contribution of each input variable. In the European Banking Authority's 2023 follow-up report on machine learning for internal ratings-based models, Shapley values were the most common explainability technique among the institutions that responded, cited by 40%, ahead of enhanced documentation at 28%.
The difficulty is who has to read those explanations. A SHAP chart answers a question that a quantitative analyst knows how to ask. An ALCO member asks a different one: why the model treats most of the current account book as stable, and what would have to change for that to be wrong. A board member signing off the assumptions, or an examiner reviewing them, needs the answer in the terms of the balance sheet rather than in the terms of feature attributions. Having that conversation about a machine learning core takes a depth of model knowledge that the people accountable for it are rarely expected to have.
Supervisors have also started to question the tools themselves. A September 2025 paper from the BIS Financial Stability Institute found that post-hoc explanation techniques such as SHAP and LIME can be inaccurate and unstable, and that there is no agreed way to check whether an explanation is correct. It favors tiering explainability by how critical the use is, allowing complex models in lower-risk applications with compensating safeguards. A model that sets a bank's structural interest rate and liquidity position sits at the top of that scale.
No regulation prohibits a machine learning core model for deposits: SS1/23 applies to machine learning the same principles that govern every other model. The EBA does set a direction: human judgment remains part of machine learning models, documentation must allow senior management to understand them, and complexity should be avoided where simpler methods perform comparably.
Practice matches that reading. The research behind this analysis found no public source describing a machine learning NMD, deposit beta or IRRBB deposit model in production at a bank. Where machine learning does run around deposits, it sits on the commercial side, targeting deposit acquisition, cross-sell offers and retention campaigns. After the 2023 runs, covered in what 2023 taught banks about deposit behavior, the risk-side response of practitioners centered on reverse stress testing and sensitivity to idiosyncratic outflows, not on AI.
The barrier, then, lies in the distance between how the model works and what the people who approve it can follow. Machine learning is the mathematically better answer, but not the one the industry needs today.
This conclusion moves AI to a different place in the modeling process. The core model stays simple enough for its audience: a segmented core and non-core split, decay curves, betas that can depend on the rate level and a replicating portfolio, each with parameters a treasurer can read and an auditor can trace. AI then works around that core on three tasks: calibrating and challenging it, monitoring when behavior drifts away from it, and explaining what it says. From large language models to algorithms built for interpretability, AI can do a great deal while the underlying model stays simple.
In "Non-Maturing Deposits Under Pressure", I compare the two positions criterion by criterion, from explainability and validation to supervisory acceptance and maturity. The trade is clear: the bank gives up a margin of precision in the core and recovers much of it through the work AI does around it.
Calibration and challenge come first. Machine learning can test many segmentations, lags and functional forms at scale without becoming the model. A random forest trained on account-level data can show which variables actually separate stable from volatile balances, and those findings can then define the segments of the transparent model. Run alongside the core, the same flexible model becomes a challenger: where the two diverge, the divergence points to behavior the core misses, and the modeler decides whether to change a parameter, add a segment or document why the core stands.
A core model calibrated once a year can drift away from reality for months before anyone notices. Machine learning can watch the inputs the model depends on and flag when behavior moves away from calibration. The useful signals go beyond balances: the share of customers banking online, for example, shapes how fast deposits can leave under stress, and a monitoring model can track it daily.
None of these signals is a deposit model. Each is an early warning that sends a treasurer back to the core before the next scheduled review, and the response is a review of the parameters by the people who own them rather than an automatic change.
Explanation is where generative AI changes what is possible. A large language model (LLM) with access to the model, its documentation and its results can turn a recalibrated beta or a shift in EVE sensitivity into plain language for ALCO: what moved, which assumption drove it, and how it compares with the previous quarter and with the stress scenarios. This is the explanation SHAP could not provide for a machine learning core, applied to a model whose logic is simple enough to explain truthfully.
The most immediate use of LLMs around an NMD model is documentation, which goes stale as fast as the assumptions it describes. An LLM connected to the model's code, data and change log can keep it current, draft the first version of validation reports and answer staff questions about why an assumption was set the way it was.
Agents take the same idea further. An agent with access to the bank's systems, to modeling skills and to the institution's own knowledge, including past validation reports, supervisory findings and ALCO minutes, can act as a modeling expert available at any hour. In practice, it can:
What the agent does not do is decide. Its proposals go to the modelers who own the model, independent validation challenges them, and ALCO approves the assumptions that result. The agent widens the evidence and shortens the cycle, while accountability stays where supervisors already expect it, with people.
With a transparent core and AI around it, the data proposes, the expert judges, and the challenge runs continuously instead of when a supervisor asks. That is also how a bank keeps its deposit model from drifting back into the habits described in what the data says about non-maturity deposit assumptions. Machine learning's precision still reaches the balance sheet, through a model that management, auditors and supervisors can understand.
The institutions that gain most from AI in deposit modeling will be those that use it to keep a simple model right, rather than those that replace it with one nobody at ALCO can follow.
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.