Designing a disciplined approach for capturing emerging risk, model limitations and judgment outside the core Expected Credit Loss engine.
No Expected Credit Loss framework, however sophisticated, captures every relevant aspect of credit risk at all times. Models are built on data. Data reflects history. History, however rich, is always incomplete. New risks emerge before they are visible in default experience. Sector-specific disruptions may affect only a narrow part of the book. Legal or policy changes may alter recovery behaviour before models can be recalibrated. Data lags may delay recognition of deterioration. Borrower vulnerability may be understood by management before it is fully represented in model variables. In such situations, the institution confronts one of the most sensitive questions in ECL: how should it adjust the allowance when the base model is directionally useful, but not fully sufficient?
Overlay Framework and Management Adjustments explain how institutions should address emerging risks, model limitations and residual uncertainties not fully captured in base ECL models. A strong overlay framework uses targeted, evidence-based and temporary adjustments with clear quantification, robust governance, double-counting controls and defined release criteria, ensuring that judgment supplements the model without overwhelming it.
No Expected Credit Loss framework, however sophisticated, captures every relevant aspect of credit risk at all times. Models are built on data. Data reflects history. History, however rich, is always incomplete. New risks emerge before they are visible in default experience. Sector-specific disruptions may affect only a narrow part of the book. Legal or policy changes may alter recovery behaviour before models can be recalibrated. Data lags may delay recognition of deterioration. Borrower vulnerability may be understood by management before it is fully represented in model variables. In such situations, the institution confronts one of the most sensitive questions in ECL: how should it adjust the allowance when the base model is directionally useful, but not fully sufficient?
This is where overlays and management adjustments enter the framework.
An overlay is not a substitute for the model. It is not a tool for smoothing earnings. It is not a quiet corridor through which preferred outcomes are inserted into the allowance. At its best, an overlay is a disciplined response to a clearly identified limitation, gap or emerging risk that the base ECL architecture does not yet capture adequately. It exists because the institution recognises that models, however rigorous, are approximations of credit reality rather than complete replicas of it.
This is why overlay design deserves a full pillar article. Overlays are among the most misunderstood parts of the entire ECL framework. Some institutions resist them so strongly that obvious blind spots remain unaddressed. Others use them so freely that the reported allowance becomes increasingly judgment-driven and hard to validate. Some apply overlays once and never unwind them. Others keep creating new ones because they have not corrected recurring model weaknesses. A mature institution takes a more disciplined path. It accepts that overlays may be necessary, but only within a robust governance framework that defines when they are justified, how they are quantified, how they are documented, how they are reviewed and how they are eventually released or embedded into the core model if they prove persistent.
This article explores overlays and management adjustments in depth: what they are, when they are appropriate, what kinds of limitations they address, how they should be quantified, how they should interact with model outputs, how governance should work, and what failures most often undermine trust in this critical but delicate part of the ECL framework.
Every ECL model reflects a set of assumptions about the world.
A PD model assumes that the chosen variables, segmentation and calibration capture default likelihood with reasonable adequacy.
An LGD model assumes that recovery history, collateral treatment, timing and cost assumptions provide a realistic picture of severity.
An EAD model assumes that the projected exposure path reflects contractual and behavioural reality.
Scenario models assume that the chosen macro variables and transmission channels capture future conditions appropriately.
All of these assumptions can be reasonable and still incomplete. Models can lag emerging conditions. Data may not yet reflect a newly developing stress. Historical relationships may break under structural change. The portfolio may have evolved faster than the model design. Operational or legal changes may alter recovery economics. Concentration risk may be known qualitatively but not sufficiently expressed quantitatively in the base framework.
Overlays exist because the institution sometimes knows that the model is directionally incomplete before it is able to redesign the model itself.
A disciplined overlay framework therefore is a recognition of reality, not a confession of failure.
An overlay is a post-model adjustment, or in some frameworks a parallel management adjustment, applied to the allowance to reflect identified risks, limitations or emerging conditions not fully captured in the base ECL estimate.
The important words here are identified, not fully captured, and adjustment.
Identified means the institution can articulate the issue clearly.
Not fully captured means the model already reflects some of the risk, but not enough, or perhaps in the wrong place or timing.
Adjustment means the overlay modifies the allowance in a controlled and explainable way.
A strong overlay is therefore specific. It has a rationale, a defined scope, a quantification basis, a governance path and a review cycle. A weak overlay is vague, persistent, unsupported or outcome-driven.
One of the most important principles in this pillar is that overlays should sit on top of a credible ECL model, not in place of one.
If segmentation is poor, default definitions are weak, scenario design is absent, recovery data is stale and stage logic is underdeveloped, then applying overlays cannot solve the problem. It can only conceal it temporarily.
A mature institution therefore asks a preliminary question before applying any management adjustment:
Is this issue truly residual to an otherwise reasonable model, or is it evidence that the model itself needs redevelopment?
This distinction matters because recurring structural overlays are often signs that the core framework is mis-specified. Overlays are most defensible when they address temporary, emerging or specifically identified gaps. They are least defensible when they become permanent crutches for unresolved model design weaknesses.
There are several recurring situations in which overlays may be appropriate.
Emerging risk not yet visible in historical data
A sector begins weakening, but defaults have not yet risen enough to move model parameters materially. Management has supportable evidence that risk is building faster than the model can see.
Model lag
The core model updates quarterly or annually, but conditions have changed significantly during the current period.
Data limitations
Critical fields are incomplete or delayed, preventing full capture of a known risk.
Portfolio change
A product, segment or origination channel has changed materially, but the historical model still reflects an earlier portfolio structure.
Concentration risk
A portfolio has built meaningful exposure to a sector, geography or customer class whose vulnerability is not sufficiently captured in pooled model outputs.
Legal or policy changes
Recovery behaviour, enforcement effectiveness or borrower relief regimes have changed in ways not yet reflected in LGD or staging assumptions.
Exceptional events
A shock, disruption or event creates portfolio stress that cannot yet be captured reliably through existing model variables.
These examples show why overlays exist. They are often responses to timing and model-boundary issues rather than to general discomfort with the allowance.
Among all overlay rationales, emerging risk is often the most compelling and the most difficult.
This is because ECL is forward-looking, yet models still depend partly on evidence. Sometimes management becomes aware of weakening conditions before those conditions have produced enough observable data to alter PD, LGD or stage migration materially. For example:
If the institution waits for model history to catch up completely, the allowance may lag the risk. In such cases, an overlay can provide interim recognition of emerging deterioration.
But because this rationale is forward-looking and judgment-heavy, it must be especially well documented and governed.
A strong overlay is scoped carefully. It should answer:
This matters because vague overlays are hard to challenge and harder to unwind. An institution that says "economic uncertainty remains elevated" and therefore adds a broad reserve without specifying the affected segments is not operating with the precision ECL requires.
A better overlay says, in substance: "This adjustment applies to SME borrowers in sectors A and B in region X because current margins, refinancing pressure and arrears-leading indicators suggest deterioration not yet captured in the model."
Precision creates discipline.
Almost every institution can describe why an overlay feels reasonable. The more demanding task is quantifying it in a way that is supportable.
Good overlay quantification may use:
What matters is not that the overlay be mathematically elegant. It matters that the institution can explain how the number was derived and why it is proportionate to the identified gap.
A weak overlay arrives as a rounded reserve amount selected for comfort.
A strong overlay arrives through a chain of reasoning that can be followed, challenged and revisited.
A mature overlay framework tries, wherever possible, to connect the adjustment back to the logic of the base model.
For example:
Even if the adjustment is applied at aggregate allowance level, the institution should still understand which model component or portfolio mechanism it is standing in for. This is crucial for two reasons.
First, it improves conceptual clarity and avoids arbitrary adjustment.
Second, it helps determine whether the issue should later be embedded into the model rather than left as a recurring manual addition.
One of the greatest risks in ECL overlays is double counting.
This can happen when the base model already captures part of the risk through macro scenarios, staging logic, recent recalibration or component sensitivities, and the overlay is added without properly identifying what remains uncaptured.
For example:
A strong overlay framework therefore requires explicit assessment of what is already captured and what is not. The question is never "is this risk present?" The question is "how much of this risk is already in the model, and what residual remains?"
Not all overlays are alike in duration.
A temporary overlay addresses a short-term gap, emerging event or transitory lag between observed risk and model capture.
A structural overlay persists because the model repeatedly fails to represent a known feature of the portfolio.
This distinction is important because structural overlays should trigger a stronger question: why is the model not being redeveloped?
A mature institution tolerates temporary overlays as part of prudent governance. It becomes increasingly uncomfortable with overlays that survive quarter after quarter for the same reason. Such persistence usually signals that the adjustment should either be embedded into the model, replaced by better segmentation or otherwise addressed in framework redesign.
An overlay that never dies is often no longer an overlay. It is an undeclared part of the model.
Most discussions of overlays focus on upward adjustments to the allowance, but in principle overlays can also reduce ECL if the model is known to be overstating loss in a specific and supportable way.
In practice, downward overlays are more sensitive and usually attract greater scrutiny, because they risk being used to release reserves prematurely. Yet the conceptual point still matters: overlay logic should be symmetrical if the rationale is strong.
The key is that whether an overlay increases or decreases the allowance, the standards of justification, quantification and governance must remain equally demanding. The institution should not make downward overlays easier to approve simply because they are convenient.
Overlays are one of the clearest places where management judgment enters the ECL framework directly. This is not inherently a weakness. In fact, one of the virtues of a well-designed ECL process is that it does not pretend judgment can be eliminated.
But judgment must be made governable.
That means the institution should document:
This is how management judgment becomes a controlled part of the framework rather than a discretionary override process.
Because overlays are inherently judgment-sensitive, governance must be especially strong.
A good governance framework typically addresses:
The objective is not bureaucracy for its own sake. It is control. An overlay changes the allowance outside the normal automated model pathway. That makes governance indispensable.
A very common weakness in practice is that institutions focus intensely on creating overlays but much less on removing them.
This is a major problem because overlays are often introduced in periods of uncertainty and then allowed to remain through inertia even after the original rationale weakens. Over time, the allowance becomes burdened with layers of historical judgment that no longer correspond clearly to current conditions.
A mature framework therefore requires explicit release criteria.
For each overlay, the institution should ask:
This release discipline is one of the clearest signs of overlay maturity.
It is often useful to maintain an overlay inventory or tracking register that records:
This gives management a clear view of how much of the allowance is model-driven and how much reflects residual judgment. It also prevents overlays from becoming invisible background adjustments.
An institution that cannot list and explain its overlays is unlikely to be governing them properly.
Another important distinction is that overlays should not be used to convert ECL into a general stress-testing reserve.
Stress testing and ECL are related but different. Stress testing explores adverse resilience under severe conditions, often beyond central expected paths. ECL estimates expected credit loss using reasonable and supportable information and probability-weighted outcomes.
An overlay becomes problematic if it effectively imports a stress-testing mindset into the allowance without clear justification. This may happen when severe tail risks are added broadly even though scenario weighting and current portfolio conditions do not support them as expected outcomes.
A mature institution therefore keeps the distinction clear. Overlays should address model capture gaps in expected loss, not serve as a hidden stress capital buffer.
It is helpful to illustrate the kind of overlay use that is often defensible.
Sector-specific emerging weakness
A certain industry segment shows sharply rising cancellations, delayed payments and covenant pressure, but the PD model has not yet incorporated enough history to reflect the change. A targeted overlay is applied to that segment based on observed leading indicators and reviewed monthly.
Recovery environment deterioration
A secured asset market has weakened materially, and recent disposals indicate lower net recovery than the LGD model still assumes. A temporary overlay is applied to affected collateral classes until the model is recalibrated.
New product immaturity
A recently originated product has insufficient loss history, and the provisional model appears optimistic relative to comparable portfolios. A conservative overlay is applied with a documented plan to replace it once sufficient data develops.
Concentration vulnerability
A large single sector exposure has become more fragile under new policy changes, and the portfolio-average model does not fully reflect concentration risk. A focused overlay is applied to the relevant segment rather than across the entire book.
These examples share an important feature: they are specific, evidenced and reviewable.
Several failures recur repeatedly.
One is using overlays to compensate for obviously weak base models rather than fixing the model.
Another is applying broad macro overlays without specifying which portfolios are actually affected.
A third is double counting risks already captured in scenarios, staging or component assumptions.
A fourth is choosing overlay amounts heuristically or round-numbered without supportable quantification.
A fifth is allowing overlays to persist for years without model remediation or release analysis.
A sixth is weak documentation, so management remembers generally why the overlay exists, but not exactly how it was derived.
A seventh is using overlays as an earnings management tool, which quickly destroys credibility.
These failures are serious because overlays sit very close to the question of trust. If users begin to feel that the allowance is shaped more by unexplained judgment than by disciplined modelling, confidence in the whole ECL framework deteriorates.
Consider two institutions facing deterioration in the same commercial real estate segment.
The first institution identifies that office-leasing weakness is affecting a specific urban portfolio with longer vacancy risk and slower collateral sale times. Its LGD model still uses pre-shift recovery assumptions. It applies a targeted overlay only to that segment, based on updated disposal evidence and time-to-sale stress, documents the rationale, quantifies the impact through scenario comparison and sets a timetable for model recalibration.
The second institution states broadly that "the market is uncertain" and adds a general reserve across all secured lending without distinguishing office from other property types, without testing overlap with scenario assumptions and without a release condition.
Both institutions say they are using overlays prudently. Only one is operating with discipline.
A strong institutional overlay framework usually includes:
The strength of this framework lies in transparency. It allows judgment to exist without allowing it to become uncontrolled.
Overlay frameworks and management adjustments are among the most important safeguards in Expected Credit Loss, precisely because models are not perfect. They allow the institution to recognise emerging risks, temporary blind spots and residual weaknesses before those issues become visible enough to force model change from the outside. But overlays are also one of the greatest tests of discipline in the entire framework. They are where credit judgment, accounting outcome and management pressure can meet most directly.
A strong institution handles this tension well. It uses overlays sparingly but not fearfully. It documents them clearly, quantifies them thoughtfully, governs them rigorously and removes them when their rationale fades or their logic is embedded into the model. It knows that overlays should supplement the framework, not quietly replace it. It understands that a good overlay is not a shadow model or a management comfort reserve. It is a temporary bridge between known risk and imperfect model capture.
In that sense, this pillar teaches a vital lesson about ECL: prudence is not achieved by mistrusting models in general. It is achieved by knowing exactly where the model ends, and where disciplined judgment must begin.
