Context & Constraint
A fast-growing bank in Oman had been expanding at exceptional pace, but on the ALM and liquidity side one part of the balance-sheet view had not evolved with the franchise. Deposit MAL and SAL reports were still driven by hardcoded, regulator-prescribed volatility assumptions. Those assumptions were prudentially conservative and operationally simple, but they treated the deposit book through standardized runoff logic rather than through the bank’s own observed customer behaviour. The issue was not whether the assumptions were permissible. The issue was whether they had become materially more punitive than the bank’s actual funding profile warranted.
For Risk, Finance, and Treasury, that created a familiar but important distortion. The reported deposit maturity profile was being shaped by a regulatory floor rather than a measured behavioural view. In practice, that affected short-bucket runoff, funding stability recognition, liquidity-gap representation, and internal confidence in the balance-sheet view. Without a formal behavioural study, however, the bank had no defensible basis for moving away from the prescribed treatment.
Franchise Signal & Analytical Mandate
Internally, there was a strong and consistent view that the bank’s customers were exhibiting materially more stickiness than the standardized assumptions suggested. Growth had not translated into unstable deposit behaviour. Relationship balances appeared durable, balance persistence looked stronger than the reporting implied, and the franchise seemed structurally more stable than the hardcoded maturity logic was recognizing. But internal belief is not a usable governance standard. Treasury cannot defend improved maturity treatment on instinct, and regulators do not accept “stickiness” as a narrative without measurement.
The bank therefore asked us to answer a narrower and more serious question: what portion of the deposit book was genuinely stable, how did that stability vary across customer and product regimes, and how much of that persistence survived under increasingly adverse behavioural conditions. The mandate was not to produce a more favourable result. It was to produce a mathematically rigorous, intuitively interpretable, and regulator-acceptable framework for internally derived maturity estimates.
Data Architecture
The study was built on approximately eight years of transaction- and instrument-level history. In monthly terms, that meant as many as 96 observation points for long-surviving balances before allowing for closures, migrations, vintage effects, account breakage, and missing-data treatment. That depth mattered because behavioural maturity cannot be estimated credibly from short windows or headline balances alone.
The unit of analysis was not the reported product balance. It was the observed balance process. A product series can appear stable in aggregate even while the underlying accounts turn over continuously. Equally, a portfolio can look volatile because of a handful of large balances while the granular base remains highly persistent. The modelling dataset therefore preserved the behavioural trail at the relevant level: opening and closing balances, inflow and outflow frequency, balance volatility, minimum balances, tenor, customer segment, product family, relationship age, repricing behaviour, operating-account indicators, and relevant external variables.
A substantial part of the work sat before formal modelling began. Deposit books are full of false signals: internal transfers, matured rollovers, product migrations, one-off corporate flows, dormant balances, temporary excess liquidity, promotional deposits, and rate-sensitive money that behaves nothing like structural funding. If these flows are left untreated, the model will produce elegant outputs with weak commercial meaning. The first analytical step was therefore to separate observed balances into behaviourally meaningful components, especially stable core balance versus excess or volatile balance, before any maturity inference was attempted.
Segmentation Logic
The most important modelling decision in this type of work is often not the choice of algorithm. It is the segmentation architecture. A single survival curve across the full deposit base may look neat in a presentation, but in practice it usually averages away the real behaviour. Retail salary-linked balances, SME operating accounts, rate-sensitive term deposits, transactional corporate money, and long-tenor relationship balances do not share the same economic life.
The framework therefore treated segmentation as a formal modelling layer rather than a reporting convenience. Candidate segmentations were tested across customer type, product family, operating versus non-operating usage, balance size, vintage, pricing sensitivity, transaction intensity, volatility bands, relationship depth, and response under stress. The goal was to isolate behavioural regimes rather than administrative labels: stable, growing, reducing, episodic, rate-sensitive, operating, non-operating, and stress-vulnerable.
This is where the work had to stay statistically disciplined. Not every segment deserves independent estimation. If a cell is thin on history, account count, or balance materiality, forcing a separate model usually produces unstable outputs and weak governance value. In those cases, segments were pooled with statistically adjacent groups or estimated through structures that borrow strength across related segments. The balance was always the same: enough granularity to capture real behaviour, enough stability to survive validation, and enough explainability for Treasury and Risk to trust the output.
Statistical Modelling Framework
At core, the problem was one of persistence estimation. For each eligible balance component, the bank needed to know three things: how much of it was structurally stable, how long that balance was likely to survive, and how that behavioural survival translated into usable maturity-bucket weights.
The framework needs to analyse three linked equations. The first expression to decompose the observed balance into stable and volatile components. The second is the survival function: the probability that a balance remains beyond a given horizon. The third converts the survival curve into maturity-bucket weights aligned to the bank’s reporting ladder. That last conversion is what makes the work operational. ALM systems do not consume theoretical persistence curves. They consume overnight, 1M, 3M, 6M, 1Y, 2Y+ and similar maturity buckets.
The modelling stack was deliberately layered rather than reliant on a single technique. Empirical survival views such as Kaplan–Meier style estimation provided a non-parametric anchor. Hazard-based structures were used to understand conditional runoff pressure and the timing of behavioural breakpoints. Parametric time-to-runoff formulations were useful where segment behaviour followed sufficiently regular decay structures. Hierarchical and mixed-effects approaches helped where segment-level estimation required partial pooling. State-space thinking was relevant in separating latent stable and volatile balance components. Challenger machine-learning models were used selectively to test whether non-linear driver interactions added signal beyond the more interpretable core framework.
In practical terms, the bank did not need a model that was merely statistically sophisticated. It needed one that could survive scrutiny from Treasury, Model Validation, Risk, and ultimately the regulator. That is why the framework was designed as a model stack: empirical evidence to anchor the behaviour, interpretable structures to explain it, and challenger layers to test robustness rather than replace judgment.
Calibration, Validation & Shock Design
A behavioural maturity model should not be accepted on in-sample fit. A curve can fit benign history well and still fail under production scrutiny. The validation programme therefore focused on four questions: did the estimates remain stable out of time, were the segment-level results economically reasonable, how sensitive were they to behavioural shock, and did small data changes create unstable maturity-bucket shifts.
The scenario design was especially important. The output was not a single optimistic maturity curve. It was a family of conditioned behavioural views under increasing degrees of stress. That allowed the bank to see how runoff accelerated under adverse assumptions, where residual core balances remained durable, and which segments required more conservative treatment despite benign base-case behaviour. In practitioner terms, the study was not trying to prove stickiness in calm periods. It was testing the survival of stickiness once conditions deteriorated.
A serious validation pack therefore included out-of-time stability checks, segment-level reasonableness testing, observed-versus-predicted runoff comparison, back-testing, challenger-model comparison, sensitivity to segmentation choices, confidence ranges, and maturity-weight stability across scenarios. The analytical standard was simple: if a result could not be explained to Treasury and defended under challenge, it was not yet ready for production use.
Regulatory Translation & Treasury Use
The work only became valuable when the behavioural outputs were translated into the bank’s MAL and SAL framework. Instead of applying hardcoded prescribed assumptions indiscriminately across the deposit base, the bank could now assign internally derived maturity weights to the relevant stable balance components and segments. That meant the modelling outputs had to be production-ready, not just analytically interesting.
The final deliverables included stable and volatile balance shares by segment, base and stressed maturity weights, runoff curves, bucket-level maturity allocations, confidence indicators, monitoring thresholds, override logic, and recalibration rules. That made the framework usable across MAL and SAL reporting, ALCO packs, liquidity stress testing, FTP discussions, IRRBB interpretation, and broader funding strategy.
The study was ultimately submitted to the regulator, and approval was granted for the bank to use internally derived estimates for the relevant maturity buckets. That changed the economics of the exercise immediately. Once the bank could support its behavioural assumptions with evidence, it was able to move away from overly punitive generic treatments and toward a more faithful representation of its actual funding behaviour. The result was meaningful optimization, improved balance-sheet efficiency, and tangible savings, but those outcomes were a consequence of measurement discipline, not assumption relaxation.
Conclusion
This case in Oman demonstrates the practical value of using long-period internal data to estimate bank-specific deposit behaviour. This study of eight years of behavioral history converted observed balance behaviour into internally derived maturity estimates, aligned those estimates to MAL and SAL reporting buckets, and supported regulatory approval for their use.
For a rapidly growing bank, the difference between prescribed assumptions and observed behaviour can be economically significant. But that gap only becomes actionable when it is translated into a rigorous analytical framework: granular data architecture, disciplined segmentation, survival-style persistence estimation, shock conditioning, validation, and maturity-bucket translation.
That is the point of the work. Not to make deposits look stable. To measure where they are stable, where they are not, and how that distinction changes liquidity decisions.
