Dynamic EVE is held up as the gold standard in some regulatory views. This article asks whether dynamic modelling of non-maturing deposits actually improves an EVE metric. We use the EBA standardised RTS assumptions as the case study.
We do not dismiss dynamic modelling. On contractual items with real prepayment risk, mortgages above all, or on non-maturing deposits in an NII simulation, we would not argue against it: there migration and pass-through play out period by period and a bank can see them coming and manage them. Where we are less convinced is its application to those same deposits in EVE.
Background
The EBA’s standardised approach to IRRBB (RTS 2022/09) sets out how non-maturing deposits are treated when calculating EVE sensitivity to rate shocks. Article 7 builds the core figure in steps:
- Classify by counterparty — retail transactional, retail non-transactional, and wholesale. Wholesale deposits from financial customers are treated as entirely non-core.
- Split stable from non-stable, using ten years of how each book’s volume moved when rates rose and fell. The non-stable part is the money that does not reliably stay.
- Split the stable part into core and non-core. Non-core is the stable balance times its pass-through rate; core is the remainder, the stable heart of the book.
- Slot the result. Everything non-core, the non-stable part and the pass-through slice, goes to the overnight bucket at par. The core is spread over a behavioural life capped at four to five years, and the share that may be treated as core is itself capped, at 90% for retail transactional deposits down to 50% for wholesale.
The dynamic scalars then sit on top of this slotting:
Under an upward rate shock, the framework multiplies the core component by 0.8. That is equivalent to reclassifying 20% of stable core balance as non-core at t=0 of the shock. The reclassified portion is then treated as if it had become overnight money at par. No behavioural duration, no franchise spread.
Under a downward rate shock, the framework multiplies the core component by 1.2. That expands the modelled core portion by 20%, with the additional balance valued at the bank’s behavioural duration.
These scalars are a fallback, and the clearest way to see the logic: most banks use their own internal models for EVE and SOT reporting, with the standardised approach required only where a model is judged unsatisfactory. But the critique reaches beyond the RTS. The EBA has positioned dynamic, scenario-dependent modelling as the preferred approach, reinforced through its 2024 IRRBB Heatmap and its 2025 and 2026 Implementation Reports, and supervisors are pointed to test internal NMD models against that benchmark. Any bank moving toward more dynamic NMD assumptions, not only one on the standardised approach, meets the same problems.
Q1. Does the scalar measure anything the method hasn’t already removed?
Before the scalar is applied, the framework has taken the rate-sensitive money out of the book twice.
The first cut, the stable split, removes the deposits that walk out when rates move. The second cut, the pass-through split, removes the ones that reprice. What is left, the core, is the stable remainder of the book. Both kinds of rate-sensitive money are gone before the scalar is even reached.
The scalar then cuts the core anyway. Under an up-shock it removes a fifth of it, “to adequately reflect interest rate sensitivity of client behaviour” (the EBA’s explanation). But leaving and repricing are the whole of that sensitivity, and both are already out. The haircut sends the money to overnight at par, which is where a deposit lands whether it walks out or reprices in full. So whichever one the scalar has in mind, the method has charged for it already.
The instructions contradict themselves outright. Core is defined as the balance “unlikely to reprice even under significant changes in the interest rate environment”, yet the next step cuts a fifth of it because rates changed. One line says core does not move with rates; the next assumes it does.
Underneath the contradictions, the scalar reflects a tension over where EVE prudence sits. Deposit profiles are genuinely uncertain, and a bank has traditionally often handled that uncertainty by assuming a prudent core and a prudent duration: it puts the caution in the base, carrying a lower franchise value, the economic value the core holds above par, and a low sensitivity. The scalar appears to do the reverse. It pins the bank to a precise, fuller base-case value on stability and beta assumptions one can only assume the EBA treat as razor-thin, then stresses it hard: an extra fifth, which a prudent split would treat as non-stable, is left in core at the fuller value, then stripped under the shock as if it had run to par overnight.
Same caution either way, but the scalar makes it look like a much more dramatic loss, on top of a higher base. That loss exists only because the framework first recognised the value it then takes away: an artefact of where the conservatism was put, not a risk the bank runs. What the headline drops is that sensitivity seen relative to the base valuation it sits on. The down-shock scalar is sharper still: its 1.2 forces core a fifth wider than the bank chose, so prudence is overridden in both directions, core trimmed when rates rise and padded when they fall, each time pushing the number the wrong way.
Q2. Does deposit volume migrate fast and far enough to warrant a scalar this size?
Take volume first.
ECB Deposit Facility Rate (top) and cumulative migration of euro area household overnight deposits from the July 2022 peak (red), against the 20% core reclassification the EBA's 0.8 up-shock multiplier imposes instantaneously (black). The 450bp was delivered with an average timing of ~6 months; migration's average timing was ~11 months — a ~5-6 month lag, against a methodology that assumes the full move at t=0. The realised cycle was 1.8× the 250bp shock the multiplier is calibrated to, yet balances drifted only ~10% and have since recovered. Source: ECB Data Portal, series FM (DFR) and BSI L21.
The chart puts the two representations side by side. The 0.8 multiplier asserts a 20% reclassification that lands in full at t=0 and is held there, the flat line at the top of the plot. What the cycle actually delivered is the gradual descent below it: household overnight balances drifted down by about 10% to a low in early 2024 and have since recovered as rates came back down. Balances fell from a €5,574bn peak in July 2022 to a €4,993bn trough in February–March 2024, and were back to €5,481bn by March 2026. The shaded gap is the distance between what the methodology books and what the most aggressive EUR cycle in twenty years produced.
That 10.4% drift came on a 450bp cycle, 2.25 times the 200bp shock the multiplier is meant to represent, and arrived gradually rather than at t=0. Weighting each hike by its size, the rate move landed around six months into the cycle and the deposit response around eleven, a lag of five to six months, with the full drift taking the better part of two years. The methodology compresses both into a single instant, and on a shock more than twice the size still books roughly double what the cycle delivered.
The comparison is more lopsided than it first looks. The 0.8 multiplier reclassifies 20% of core, the balance the bank has already filtered for rate-insensitivity. The 10.4% is migration of the total household overnight book, core and non-core together. Non-core is by definition the rate-sensitive bucket; it is meant to run off when rates rise. So most of the observed drift is the non-stable balance doing exactly what makes it non-core, not the core the multiplier claims sheds a fifth. On any reasonable split the non-core bucket alone accounts for the bulk of the outflow, leaving the stable core close to flat. The multiplier reintroduces rate-sensitivity to the very balance the bank established as rate-insensitive, and sets its size above what the entire book, rate-sensitive part included, actually delivered.
Pace. The first six months of the cycle, DFR from 0.00% to 2.00%, produced almost nothing: monthly migration averaged 0.17% of the base. Through the most active phase, months 6 to 12, DFR from 2.00% to 4.00%, it ran at 0.74%, and the worst single month was 1.80%, in October 2023. Across the 17 outflow months observed, mean monthly migration was 0.75%. Reaching the methodology’s 20% even at that sustained peak velocity would take 27 consecutive months at peak intensity; the actual cycle held peak for six. The pace the scalar assumes bears no relationship to what the most aggressive EUR cycle in twenty years produced.
Q3. Was pass-through fast or large enough to warrant it?
We have established that volume does not migrate far or fast enough to warrant a 20% reclassification. That leaves pass-through. Let’s look at beta, then, and how it behaved through the cycle. Could beta, or convexity in pass-through, take a fifth out of core?
The ECB Deposit Facility Rate moved from -0.50% to a peak of 4.00% between July 2022 and September 2023, a cumulative 450bp over 14 months. Over the same window the weighted-average rate paid on euro area household overnight deposits rose from effectively zero to 0.33% at the rate peak, and to 0.39% by its own peak in January 2024, about four months behind the DFR. That is a cycle pass-through of about 7%. Corporate sight repriced faster, nearer 18%, but household overnight is the dominant balance for a retail-funded bank, and both sit far below what the scalar implies.
ECB Deposit Facility Rate against the average rate euro area banks paid on household overnight deposits (ECB MIR series L21). The deposit rate peaked at 0.39% as policy reached 4.00% — a cycle pass-through of about 7%. The shaded area is the franchise spread the bank earns on the balance; it widened as rates rose. Source: industry workbook (ECB DFR and MIR data).
The chart makes the point on its own. The policy rate climbs to 4%, and the rate banks actually paid on overnight deposits barely lifts off the floor, peaking at 0.39%. Because it is an average across the whole overnight book, the flat line also tells you customers did not shift in any size into higher-paying sight products. The widening gap is the franchise spread the bank earns, and it grew through the cycle. This is the value the dynamic multiplier asserts will erode. It did the opposite.
A higher assumed beta only cuts in the bank’s favour: the more of its balance the bank already places at par, the less core the multiplier has left to convert, so the overhead falls hardest on the bank that claimed the largest, lowest-beta core.
Of the two channels, pass-through is likely the more material residual risk for most banks. Volume stability has largely held. UK banks passed through closer to 40%, on our calculation from Bank of England data, well above the euro area’s single digits. Even then the hit is softened, lagging the rate move by months while the bank earns the wider spread and reprices its assets, and it comes down to whether the bank’s beta assumption is right. And beta is as much a pricing assumption the business sets as an empirical model output: the bank chooses how much of a rate rise to pass on, and the risk is that competition later forces its hand. Convexity in that beta, the way it rises as rates rise, distorts EVE so heavily precisely because EVE runs the core off: a rate-dependent beta is applied to the whole discounted franchise at t=0, not period by period as it actually occurs. The cleaner split is to hold beta static in EVE and capture its convexity in the earnings metrics, where it plays out gradually and the bank can manage it.
Q4. Where balances do move, is the franchise value actually lost?
Grant the migration the first question disputed. Suppose a fifth of the balance really does leave overnight deposits. The scalar treats that as franchise value gone: reclassified to non-core, valued at par, no behavioural duration, no spread. The cycle says otherwise, and there are two offsets the metric is built to miss.
Franchise rotation. The cycle data measures gross migration out of overnight deposits, not where the balance ends up. Much of what leaves a sight account in a rising-rate cycle rotates inside the same relationship: into time deposits, money market funds or wealth products at the same bank. The euro area aggregate shows exactly this. Over the cycle household term deposits rose by roughly €810bn while overnight balances fell about €580bn to their trough. More balance rotated into term than ran off overnight altogether. The outflow the methodology treats as franchise loss is, in the aggregate, franchise rotation.
Major banks disclose the same thing. Through the cycle JPMorgan Chase held its instant-access savings rate at effectively 0.01%, far below the rates online competitors offered, and the franchise still held. At its 2023 Investor Day it reported retaining 60% of yield-seeking flows internally and generating net new money through CDs and wealth products; of the roughly 5% of customers who did move money to online banks, it saw “no change in primary bank behavior.” The 0.8 multiplier treats the reclassified balance as if it had left the franchise entirely. The reality, from a major retail bank in the most aggressive US cycle in two decades, is that most of the yield-seeking money stayed inside the relationship and generated new franchise value.
Run-off measurement perimeter. Even where migration does happen, EVE cannot recognise the offsetting value. It is a run-off metric: it values the cash flows from the existing balance sheet under shock, with no recognition of new business written in response. When balance rotates from an overnight NMD into a term deposit at the same bank, the economic position changes, but the new term deposit is new business and sits outside the EVE perimeter. The bank may earn a lower spread on the term deposit than on the overnight balance, but the relationship still holds economic value, and a run-off lens cannot recognise it. The methodology writes off the full overnight franchise value and stays blind to where the balance landed.
The earnings offset. The second offset is the larger one. The multiplier books an economic-value loss on the core franchise under rising rates. That same franchise, under those same rates, earns the bank more net interest income, not less: deposit rates barely moved while asset yields repriced, so the spread widened, exactly as the pass-through chart showed. The scalar writes the franchise down at the moment it is paying the bank most. EVE on its run-off basis cannot net the two, because the spread income is future business outside the perimeter, so it records the write-down and stays silent on the earnings that are its economic counterpart.
Q5. Could a bank running a normal hedge ever hedge this stress away?
The base case is what the bank has built itself to operate against. Behavioural assumptions, hedge positioning and commercial strategy are calibrated either to a modelled view of how core deposits behave, or, where they are not modelled, to the bank’s strategic risk appetite for them. What is left as static EVE is the residual rate risk the bank has chosen to retain, or the bit it hasn’t been able to hedge away for operational reasons. The hedging programme typically allows for migration and rate moves happening at normal cycle pace by diversifying the maturity of the corresponding assets into a replicating portfolio.
Dynamic EVE breaks that picture instantly. The 0.8 multiplier forces 20% of core to reclassify as non-core at t=0; the 1.2 multiplier does the equivalent in reverse. The deposit base hasn’t changed. The hedges haven’t changed. The customers haven’t done anything. The framework has imposed a different representation of the same liability, and the difference between the two gets measured as additional EVE sensitivity.
Every shock direction pushes that representation away from the base the bank optimises against, so the optimisation point and the measurement point are no longer the same place. The result can be an EVE metric that shows negative rate sensitivity under every scenario.
Look at why the hedge can absorb a real migration but not an instantaneous reclassification. A retail bank’s replicating portfolio is a laddered set of fixed-rate instruments, its weighted-average life calibrated to the behavioural duration of the NMDs. Supervisory data on 67 euro area Significant Institutions (ECB Working Paper 3140, 2024) puts the weighted-average maturity assigned across the whole NMD book at around two years, with wide variation across institutions. Take an example from that blend. If the whole book averages two years and the core is around 80% of it, with the non-core fifth sitting at overnight, the core is modelled at a duration of roughly two and a half years. For a two-and-a-half-year core, a 5-year linear ladder amortising at 1/60 of notional a month, about 1.67%. That rate is not a free choice. It falls straight out of the duration assumption, so the hedge is calibrated by construction to absorb migration at the pace that duration implies.
Two buffers absorb migration, not one. First, before the ladder is built, the bank applies a stability haircut: it carves the rate-sensitive non-core out of the balance it hedges and holds it short, between 10% and 50% of the book under the Basel and EBA caps. Migration comes out of that bucket without the core ladder being touched. The cycle’s 10.4% of total migration sits inside even a conservative haircut, so the stable core the ladder hedges was never threatened: realised migration sat at or below the bottom of the cap-implied range.
Second, behind it, the ladder’s amortisation. At a two-and-a-half-year duration it rolls off about 1.67% of notional a month, and any migration that reaches the core can be met by reinvesting less of the maturing tranche. The worst single month of the cycle, 1.80% in October 2023, edged just above one month’s roll-off; the rest sat well inside it, and across the twenty-month descent cumulative roll-off of around 33% covered cumulative migration of 10.4% three times over.
So the combination, a stability haircut plus amortisation at roughly 1.7% a month, is built precisely to absorb a lagged run-off at the pace rate cycles produce. What it cannot absorb is a 20% reclassification of core in a single instant: freeing a fifth of the core through the ladder alone would take around twelve months of reinvesting nothing at all. Nothing in a structural hedge delivers a fifth of core at t=0, so the dynamic ΔEVE is a hit the bank has no operational path to hedge, whatever you assume about the underlying customer behaviour.
Q6. Is the combined stress still a plausible scenario?
The questions so far have tested the scalar piece by piece. Step back to the stack it sits on. Dynamic EVE is a fourth stress, layered on a static measure that is already three deep.
Static EVE is conservative on three counts, each long accepted. It values core deposits on a run-off basis, as if the franchise amortised away rather than persisting and earning spread on new business. It applies the rate shock instantaneously, the full 200bp at t=0, when the most aggressive cycles in twenty years took more than a year to deploy: the ECB 450bp over 14 months, the Fed 525bp over 17. And it rests on a behavioural duration that is itself a judgement call, set short, typically 2 to 3 years, against a stable franchise that in reality persists for decades.
The dynamic methodology adds a fourth: an instantaneous breakdown of the very behavioural assumptions that calibrate the static number. Each layer alone is a defensible upper bound on something real. But the framework now requires all four in the same instant: a 200bp curve shock at t=0, a run-off valuation, an already-conservative duration, and a fifth of core reclassifying at t=0. A stress stops being useful at the point it asks for more than the worst case on record ever delivered, and that is the line the scalar crosses.
Q7. Is the dynamic framework better than static, for NMDs?
A risk metric earns its place by guiding action. This one carries a second purpose on top, as a supervisory outlier test: to flag the banks carrying more rate risk than their peers. The EVE measure should be driven by the genuine gap risk between the bank’s target hedge position and what it has actually hedged at the measurement date. That gap is what ALM teams can close, what supervisors can sensibly engage on, and what represents real exposure to rate moves. It does neither.
Dynamic EVE buries that signal. The breach is generated by the liability-side multiplier. An overhead the bank cannot hedge against, that the empirical record does not support, and that the metric structurally cannot offset because EVE excludes the new business the bank captures from rotating balance. For our stylised €1tn deposit bank that overhead is around 11% of Tier 1 at +200bp. The underlying gap the bank actually manages is around 10% of Tier 1. The noise is larger than the signal.
Worse, that overhead is much the same for every deposit bank. It scales with the franchise, not with the hedging. The metric then cannot do the one thing a risk measure is for: tell a well-managed balance sheet from a badly-managed one. The genuine gap, the part that differs between banks and that an ALM team can act on, is dwarfed by a common, unhedgeable overhead standing in for an event that will almost never happen.
There is one lever, and it points the wrong way. The dynamic loss scales with the behavioural duration on core, so a bank under pressure can shrink it by shortening that duration. But the same duration calibrates the hedge ladder that stabilises NII. The only route to a smaller number runs through making real earnings less stable, trading genuine cycle protection for a smaller hypothetical loss. The number can be cut, but only by harming the business.
To the supervisor the breach says little more than “this bank has a deposit franchise,” and the conversation turns to exemptions and overlays rather than the hedge gap the bank is actually carrying. Dynamic EVE becomes a fixed cost of deposit funding, scaling with the very business it is meant to help manage.
What this means in numbers
How much of the reported number a bank will actually experience, and how much is run-off overhead it never will, depends on its own balance sheet, and the difference is large and bank-specific. So rather than lean on one stylised attribution, build it from your own, a step at a time. The going-concern figure, what the bank faces under the shock, stays within the threshold; the run-off overhead it will never realise is what drives the breach. Whether the threshold should sit on the going-concern figure or the total is for supervisors to decide; the point is that a single number hides the distinction.
One input per step, from your own balance sheet. The number builds as you go. Defaults are the €1tn retail bank from the opener.
-
1 Your deposit book and capital
The scale the rest is measured against.
-
2 The core of your deposits
Take the stable share, net of pass-through; what is left is the core. Set the behavioural life you value it over. This is the balance the hedge backs and the scalar acts on.
—
-
3 Your structural hedge
The assets backing the core run a little longer than the core itself. That residual gap, under the shock, is your static EVE: the rate risk you actually manage.
—
-
4 The scalar
The 0.8 multiplier reclassifies a fifth of the core to par at t=0, valued at the core duration under the shock. That is the overhead it adds.
—
—
—
Stylised, first-order (duration-based) model. Core = stable share × (1 − β) × NMD. Static EVE = (hedge asset duration − core duration) × shock × (core ÷ Tier 1), the residual after hedging the core. Scalar overhead = 0.2 × core × core duration × shock ÷ Tier 1. A small asset-side dynamic (prepayment slowing), held at 2.5% of Tier 1, completes the reported figure.
Conclusion: what a prudent framework would do
None of this argues for abandoning behavioural sensitivity. Pass-through faster than modelled, migration arriving before the ladder can absorb it, a duration that proves wrong: real risks, and a bank should measure them. The objection is to baking them into the headline at a magnitude and pace no cycle supports, where the metric’s own run-off design then forbids the offsetting value from ever showing up. The fix is not a more sophisticated overlay but getting the base right, in order of preference:
-
A prudent core. Set the core volume to allow some headroom and hold a duration the bank can stand behind, judged on how the balance actually behaves, not only on how it moves with rates. A prudent stable share and a short duration already pull the franchise towards par, the conservative direction: prudence in two assumptions a bank can justify and a supervisor can challenge head on, not a scalar bolted on that re-prices the same balance twice. Balance the two it serves, though. A short duration is conservative for EVE but the opposite for earnings, so the core has to sit between them, not at either extreme. GL 2022/14 frames the NMD decay assumption exactly this way, as “balancing the benefits to net interest incomes against the additional economic value risk.”
-
If uncertainty must be shown beyond that, anchor it on par. Start with the whole book valued at par, the overnight amount the customer can withdraw on demand, then walk it back to within the SOT limit: show the behavioural duration it takes to get there, with a qualitative assessment of why the bank is comfortable holding it. Par is observable and verifiable without second-guessing a behavioural model, so the judgement on display is the duration itself, stated and open to challenge, not a scalar buried in the headline. The behavioural view sits alongside as that disclosed walk from par.
-
Keep the one residual worth taking seriously where it belongs. That pass-through proves higher than assumed and compresses the spread is a real exposure. But it is an earnings sensitivity, an NII risk, not an economic-value one, and the framework confuses the two by expressing it as a haircut to EVE. Where the beta assumption itself is in doubt it is model risk. Either way it belongs in the NII measure and the ICAAP, where it can be seen coming and managed, not in the SOT as a fixed scalar that lands at t=0 and cannot be.
-
Failing all of that, attribute the number rather than report it whole. Show the structural EVE the bank manages to, separated clearly from the run-off sensitivity on non-maturing instruments it will not realise, because it is not in run-off.
That is the move Part I argued for on the static SOT, making the assumption visible instead of decisive. Dynamic modelling adds another layer and finds no risk the static number had missed; it reports the same risk through a scenario the bank will never face, then books the difference as a breach.
Part I of this series, Does the EVE SOT Really Tell Us Anything?, tests what a one-year shift in deposit duration does to the static SOT.
Sources
Regulatory framework
- EBA Guidelines on IRRBB and CSRBB (EBA/GL/2022/14)
- EBA RTS on IRRBB Supervisory Outlier Tests (EBA/RTS/2022/10)
- EBA RTS on the standardised methodology on IRRBB (EBA/RTS/2022/09): definitions 14 (stable) and 16 (core); Article 7 (NMD slotting — the ten-year stable/non-stable split, the pass-through core/non-core split, and the 0.8/1.2 scalars); explanatory paragraph 17(b)
- Commission Delegated Regulation (EU) 2024/856 (EVE SOT shock specifications)
- Commission Delegated Regulation (EU) 2024/857 (standardised approaches for IRRBB)
- BCBS, “Recalibration of shocks in the interest rate risk in the banking book standard” (July 2024)
EBA heatmap and implementation
- EBA IRRBB Heatmap (January 2024)
- EBA Report on IRRBB Heatmap Implementation Phase 1 (EBA/REP/2025/04, February 2025)
- EBA Report on IRRBB Heatmap Implementation Phase 2 (January 2026)
Empirical data and bank disclosures
- ECB BSI statistics, series L21 (overnight deposits, households), non-seasonally-adjusted: volume cycle observations 2020-2026. The seasonally-adjusted series gives the same ~10% cycle magnitude, with the peak and trough months shifted by one to two months.
- ECB MIR statistics, outstanding amount rates, household overnight deposits
- ECB Deposit Facility Rate trajectory, July 2022 to present
- Bank of England statistical database (Bankstats): UK household deposit rates against Bank Rate over the 2021-24 cycle; the ~40% pass-through figure is the author’s calculation from these series
- ECB Working Paper 3140, “Banking on assumptions? How banks model deposit maturities” (2024)
- JPMorgan Chase, 2023 Investor Day transcript, 22 May 2023
Further reading
- Drechsler, Savov, Schnabl & Wang, “Deposit Franchise Runs”, Journal of Finance (2026)
- San Francisco Federal Reserve, “Bank Franchise as a Stabilizing Force” (2024)
- New York Federal Reserve Liberty Street Economics, “How Do Interest Rates (and Depositors) Impact Measures of Bank Value?” (April 2023)
Stephen Harvey is the founder of irrbb.com and Neuro-XI. He spent 15 years implementing and overseeing IRRBB measurement systems at two G-SIBs and one D-SIB.