Response to discussion on the simplification and assessment of the credit risk framework
Q1. For the purpose of reporting under CRR Article 430a, which definition of loss should be used?
For the purposes of reporting under CRR Article 430a, the definition of loss should be harmonised as an observed (realised) annual loss measure. For avoidance of doubt, ‘observed (realised) annual loss’ should be defined as the net amount of credit loss recognised through objective loss realisation events within the reporting year (including write-offs, realised losses on collateral disposal or loan sale, and recoveries netted against such losses), excluding purely model-based re-estimation of expected losses for unresolved recovery processes
Losses should reflect the amounts actually realised and recognised during the reporting period across all outstanding real estate exposures, irrespective of when the default event occurred, and should not rely on estimated losses for incomplete recovery processes or cohort-based default imputations. To preserve cross-institution comparability, the reporting template should also prescribe the exposure value basis used for scaling the annual loss measure (e.g., gross carrying amount/EAD), and require consistent netting rules for recoveries
The objective of Article 430a reporting is supervisory convergence and cross-institutional comparability. A definition of loss that is conceptually aligned with IRB-LGD measurement – and therefore relies on model-dependent estimates and incomplete recovery assumptions – inevitably introduces structural heterogeneity across institutions and undermines this objective.
By contrast, a realised-loss definition anchored in observable accounting outcomes provides a common and auditable reference point and avoids embedding institution-specific modelling practices into a supervisory reporting metric.
It is important to acknowledge that a realised-loss metric is inherently backward-looking and may not capture emerging risk early, in particular where institutions delay the disposal of collateral during market downturns. However, this limitation reflects the intended nature of Article 430a data as an ex-post loss indicator and should not be addressed by re-introducing model-based or estimated loss concepts into the reporting definition.
Accordingly, the loss definition for CRR Article 430a should be standardised as an observed annual loss measure and should not be designed to approximate, or substitute for, prudential LGD concepts.
Q2. Should the loss data (CRR Article 430a) be used for the assessment of RWs of real estate exposures under CRR Article 126(4) and CRR Article 465(11)?
The loss data reported under CRR Article 430a should not be used as a primary or binding input for the assessment of risk weights under CRR Article 126(4) and CRR Article 465(11).
Article 430a loss data is, by construction, an ex-post outcome metric. It reflects realised losses driven by market timing, recovery processes and institutional workout practices, and not the structural and forward-looking risk characteristics that the risk-weight framework is intended to capture.
In particular, the CRR Article 430a loss metric:
– aggregates losses over defaulted and non-defaulted exposures and therefore cannot be linked to default-driven risk mechanisms,
– mixes heterogeneous collateral types, markets and contractual structures within a single outcome ratio,
– is materially distorted by the timing of collateral disposal and recovery strategies,
– is sensitive to internal forbearance and restructuring practices,
– reflects jurisdiction-specific insolvency and enforcement regimes rather than asset-level risk,
– embeds accounting recognition and provisioning policies into a prudential signal,
– is affected by valuation lag and appraisal practices rather than by actual collateral risk at origination,
– cannot be aligned consistently to LTV buckets used in the risk-weight framework,
– does not isolate income-dependency risk relevant for IPRE exposures,
– is backward-looking and inherently pro-cyclical when used for supervisory calibration purposes,
– is contaminated by legacy portfolio composition and historical underwriting standards,
– is driven by realised market liquidity at exit rather than by structural market depth and resilience,
– is not comparable between SA and IRB institutions due to different exposure value definitions, and
– cannot be reconciled with the conceptual objective of risk weights as ex-ante capital differentiation tools.
The metric is not structurally mappable to the RW architecture because it is neither conditioned on default events nor aligned to the segmentation logic (e.g., LTV buckets and income dependency) underpinning the real estate framework.
For these reasons, the informational content of CRR Article 430a loss data is not sufficiently robust to support supervisory decisions on the calibration or adjustment of real estate risk weights.
This limitation is particularly relevant for the assessment under CRR Article 465(11), where the objective is to evaluate the appropriateness of the real estate framework in the presence of the output floor. Using realised loss outcomes as a driver in this context would introduce a structurally backward-looking and potentially pro-cyclical element into a framework whose purpose is to remain risk-sensitive while being stable over the cycle.
CRR Article 430a loss data may nevertheless be used in a strictly supplementary and diagnostic capacity, for example as a high-level benchmarking input across jurisdictions and as a consistency check against other market evidence.
However, it should only be interpreted alongside structurally grounded and forward-looking indicators, including, inter alia, LTV distributions and migration, income-dependency characteristics of IPRE exposures, refinancing and maturity profiles, valuation practices, and real estate market liquidity and stress indicators.
In conclusion, CRR Article 430a loss data should not be used as a decisive or mechanistic input for the assessment of risk weights under CRR Article 126(4) and CRR Article 465(11).
The prudential interpretation underlying this response is further articulated in the attached conceptual note (“Capital as Forward Constraint”), which sets out the distinction between outcome-based description and forward-looking constraint as a design principle for capital frameworks.
Q3. Which elements of the real estate framework should be further simplified?
Further simplification of the real estate framework should focus on elements where regulatory complexity no longer delivers proportionate gains in risk sensitivity and where national discretions materially undermine supervisory convergence.
In particular, simplification should target the following areas:
– the fragmentation and proliferation of national derogations and discretions for real estate exposures, which materially weakens comparability across jurisdictions and complicates supervisory benchmarking,
– the current parallel layering of macro-prudential and micro-prudential real estate tools within the CRR framework, where similar risk signals are addressed through multiple partially overlapping mechanisms,
– the multiplicity of legal hooks used to adjust real estate risk weights (including separate triggers under different CRR articles), which creates a non-transparent and difficult-to-audit “stacking” of measures,
– the operational and interpretative complexity surrounding the interaction between national designated-authority measures and EU-level real estate risk weight calibration,
– the definition and operationalisation of IPRE versus non-IPRE exposures, where the current criteria remain difficult to apply consistently across business models and national lending structures,
– the reliance on heterogeneous valuation practices and appraisal concepts in the determination of collateral values used for LTV calculations, without a sufficiently standardised prudential valuation anchor,
– the treatment and reporting of LTV buckets, including bucket boundaries and their interaction with national measures, which materially increases reporting and implementation complexity without improving risk differentiation,
– the use of CRR Article 430a loss data within several real estate policy and calibration discussions, despite its limited predictive content and heterogeneous construction across institutions,
– the current separation of real estate exposure categories across several parts of the framework, requiring institutions to implement parallel classification logics for closely related exposure types,
– the lack of a unified and simplified framework for incorporating forward-looking real estate market signals into supervisory assessment, leading to ad-hoc and nationally divergent practices,
– the interaction between real estate risk weights and the output floor assessment, which introduces additional layers of complexity without a clear and stable linkage to the underlying real estate risk drivers,
– the reporting and data requirements associated with real estate exposures, which are currently not sufficiently aligned with the actual supervisory use of the data and therefore generate avoidable operational burden,
– the current framework for the application of real estate-specific measures to mixed-use and hybrid properties, where classification and treatment rules remain unnecessarily complex and jurisdiction-dependent,
– the governance and escalation mechanisms for activating, adjusting and withdrawing real estate-specific prudential measures, which are fragmented across several legal bases and supervisory processes.
In substance, simplification should prioritise consolidation of real estate-specific measures, reduction of national discretionary overlays, and a clearer separation between outcome-based indicators and structural risk drivers, while preserving the central role of property type, income dependency and LTV as the core risk differentiators in the framework.
Q4. Which other clarifications do you consider necessary to apply the new ECAI framework?
Beyond the temporary acceptance of ratings without government support under C3, further clarifications are necessary to ensure that the new ECAI framework can be applied consistently, prudently and without creating new sources of unwarranted variability.
In particular, clarification is needed on the following points:
– the precise operational meaning of “compatibility of rating scales”, including whether compatibility refers exclusively to the ordinal structure of the scale or also to its implied long-run default profile,
– the minimum quantitative evidence required from ECAIs to demonstrate that the removal of government support does not materially alter the calibration of the underlying scale used in previous EBA mapping exercises,
– the treatment of situations in which an ECAI maintains a common symbolic scale but applies materially different methodologies for ratings with and without government support,
– the extent to which banks are expected to perform their own due-diligence or validation of the methodological differences between supported and unsupported ratings, -Clarification should ensure that any bank-level due diligence remains proportionate and governance-focused, and does not de facto replicate the EBA’s mapping/validation function
– the supervisory expectations regarding the governance and documentation of the decision to use unsupported ratings during the transitional period,
– the conditions under which a previously mapped scale must be considered no longer valid due to a structural change in the underlying rating methodology,
– the treatment of mixed portfolios where supported and unsupported ratings coexist within the same exposure class and counterparty type,
– the handling of rating transitions and historical default data when unsupported ratings have only a short time series,
– the interpretation of “default” and performance data used for future remapping exercises, in particular where unsupported ratings are designed to represent hypothetical default behaviour absent public support,
– the treatment of jurisdictions where government support is already largely excluded from existing ratings, and how this interacts with the temporary regime under C3,
– the supervisory approach to potential cliff effects at the point when a new mapping exercise becomes available and replaces the temporary regime,
– the disclosure and transparency expectations vis-à-vis institutions’ reliance on unsupported ratings in regulatory reporting and Pillar 3 disclosures,
– the alignment of the ECAI framework with internal credit assessment processes, in order to avoid regulatory reliance on ratings that are not used, or not credible, within banks’ own credit and risk governance,
– the expectations regarding consistency between the use of unsupported ratings for risk-weight purposes and their use in internal risk management, limit setting and concentration monitoring,
– the supervisory treatment of cases where different ECAIs adopt divergent interpretations of “removal of government support” for otherwise comparable counterparties.
These clarifications are necessary to ensure that the temporary pragmatic solution proposed under C3 does not evolve into a structurally ambiguous regime and that the use of ratings without government support remains transparent, auditable and prudentially credible throughout the transition period.
Q5. Should the consolidation of regulatory products for credit risk be a priority or should the regulatory stability be preferable instead? Have you identified any redundancies in IRB products?
Consolidation of IRB regulatory products should be treated as a priority, provided that it is executed as a strictly editorial and structural exercise: consolidation that reduces fragmentation, removes duplication, and improves navigability strengthens regulatory stability rather than undermining it. Stability is not the number of PDFs in the rulebook; stability is the predictability of interpretation, the traceability of requirements, and the reduction of avoidable ambiguity in supervisory dialogue. Under that definition, C4–C5 are pro-stability interventions.
The practical problem is that the current IRB “single rulebook layer” is not single. Requirements are distributed across CRR Level 1, multiple RTS, multiple Guidelines, the validation handbook, and Q&As, with overlapping phrasing, partially redefined terms, and repeated expectations expressed with different legal force. That structure creates instability through interpretation drift, not through formal rule change.
Redundancies that are material in practice and should be removed or reconciled include the following.
– repeated articulation of representativeness requirements across PD/LGD guidance and later CCF workstreams, with similar concepts expressed under different structures and granularity, creating inconsistent implementation evidence packs,
– parallel definitions and expectations around “data deficiencies”, “bias”, and the remediation logic (adjustments vs MoC), with overlapping requirements dispersed across estimation guidelines and downturn guidance, making it difficult to maintain a single MoC governance model without duplication,
– overlapping and sometimes differently-scoped requirements for long-run average calibration, downturn calibration, and the role of reference values / add-ons, which results in institutions documenting the same calibration narrative multiple times for different supervisory artifacts,
– duplicated expectations around model use test and application controls (including override governance) appearing across multiple products, with minor wording differences that create unnecessary compliance layers without adding prudential content,
– repeated requirements on grade-level homogeneity / heterogeneity, minimum observation standards, and concentration considerations, which are reintroduced in different places and then further re-clarified via Q&As, creating a moving target in validation documentation,
– overlapping guidance on default definition linkage and estimation unit (obligor vs facility) across CRR, RTS and guidelines, while the CRR3 facility definition introduces new degrees of freedom that can only be applied consistently if the texts are reconciled into one coherent operational definition,
– repeated statements regarding the need for conservative use of expert judgment, overrides, and overrides monitoring, expressed across multiple documents without a single consolidated control standard and audit trail template,
– duplicated treatment of external and pooled data and their acceptable use cases, scattered across representativeness sections and model development expectations, generating inconsistent supervisory interpretations across jurisdictions,
– overlapping guidance on discounting, recovery cash-flow construction, and cost inclusion (direct/indirect), with parts embedded in PD/LGD estimation guidance and parts resurfacing in later calibration discussions, resulting in duplicated policy artifacts within institutions,
– replicated back-testing expectations (especially the articulation of what is being back-tested: raw estimates vs final estimates including MoC), where the same conceptual point reappears in multiple texts with different framing and therefore different evidentiary burdens,
– multiple references to the same core prudential principle—comparability and reduction of unwarranted RWA variability—implemented via separate micro-requirements rather than a single consolidated “comparability spine” that institutions and supervisors can audit consistently,
– duplicative “scope logic” across products (what is in model scope, what is excluded, what segmentation is permitted), which is often restated when new products are issued, instead of being maintained as a single authoritative scope layer.
Accordingly, consolidation should be prioritised, but with a hard constraint: no hidden re-calibration through editorial consolidation. The goal is a single coherent IRB corpus where each requirement exists once, where cross-references are explicit, where definitions are centralised, and where the legal hierarchy is unambiguous.
If executed in that manner, consolidation reduces model-risk governance cost, reduces interpretative dispersion, and improves supervisory convergence. If not executed in that manner—if consolidation becomes a vehicle for incremental rule change—then stability concerns become valid. The priority therefore is consolidation with strict change-control: structural simplification, not substantive drift. Any consolidation should be accompanied by an explicit redline mapping demonstrating that obligations are not substantively altered, and that any genuine policy change is processed through a separate, transparent consultation track.
The prudential interpretation underlying this response is further articulated in the attached conceptual note (“Capital as Forward Constraint”), which sets out the distinction between outcome-based description and forward-looking constraint as a design principle for capital frameworks.
Q6. Do you consider that the integration of environmental and social risks into the credit risk framework could be further enhanced without undermining its simplicity? Which areas, if any, would you prioritise for further work or clarification?
Integration of environmental and social risks into the credit risk framework can be enhanced materially without undermining simplicity, provided the integration is implemented as a governance and materiality architecture, not as a proliferation of new granular model requirements. Complexity grows when E&S is treated as an additional parallel modelling universe; simplicity is preserved when E&S is treated as a disciplined set of triggers that determine when, where, and how existing IRB mechanisms must incorporate new risk drivers, overrides, and stress calibration.
The priority therefore is not “more variables” but clear rules of admissibility: when E&S is relevant, what evidentiary standard applies, and what conservative treatment is required when data is insufficient. That approach increases prudential credibility and comparability while keeping the framework operationally coherent.
Priority areas for further work and clarification are as follows.
– a unified definition of E&S materiality for credit risk purposes that is explicitly tied to PD, LGD and CCF channels and that can be evidenced in model governance, rather than remaining a qualitative narrative layer,
– an explicit taxonomy of transmission mechanisms from E&S risk to obligor cash flows and collateral values, separating physical risk versus transition risk, and mapping each to the relevant risk parameter(s) and modelling phase (differentiation vs quantification vs application),
– supervisory expectations on when E&S risk drivers must be embedded in risk differentiation models versus when they may remain in overlays, expert adjustments or policy-based overrides, with clear “threshold conditions” and documentation standards,
– a standard approach to conservative treatment under data scarcity: when E&S drivers are plausible and material but empirical series are short, the framework should specify acceptable conservative add-ons, MoC treatment, or stress-based overlays rather than forcing institutions into pseudo-precision,
– harmonised guidance on the use of scenario analysis and stress testing as the primary bridge between forward-looking E&S uncertainty and point-in-time or through-the-cycle calibration, including how scenario outputs translate into parameter adjustments or decision constraints,
– clarity on the interaction between E&S integration and the representativeness framework, including how to treat non-representatative historical samples when the risk environment is structurally changing, and how to evidence that models remain valid under new regimes,
– explicit guidance on collateral valuation implications, particularly for real estate collateral, where physical risk can affect liquidity, insurance availability, capex requirements and exitability, which feed directly into LGD and recovery timing rather than only PD,
– a consistent approach to overrides and expert judgment for E&S considerations: what constitutes a justified override, how it is monitored, when it must be embedded into the model rather than repeated ad-hoc, and how override practices are constrained to avoid capital arbitrage,
– alignment of E&S integration with downturn and cyclical calibration logic, ensuring E&S is not treated as a separate “non-cyclical” overlay but is incorporated into downturn identification, recovery stress assumptions, and structural LGD shifts where relevant,
– a minimum disclosure and audit trail standard for E&S integration within IRB governance, enabling supervisors to assess whether E&S is decision-relevant and prudentially binding, as opposed to being a narrative compliance exercise,
– guidance on the acceptable use of external data sources and pooled datasets for E&S drivers, including standards for quality, comparability, linkage to obligor/facility characteristics, and the treatment of model risk when external indicators are used,
– clarification on the distinction between “risk differentiation” and “risk quantification” for E&S: many institutions can differentiate risk using E&S signals before they can quantify them; the framework should explicitly permit that sequencing while imposing conservative quantification where necessary,
– rules for consistency between Pillar 1 IRB integration and Pillar 2 / ICAAP practices, to prevent a dual system where E&S is treated as decision-binding in ICAAP but ignored in IRB, or vice versa, thereby reducing governance fragmentation,
– treatment of procyclicality and cliff effects: the framework should prevent abrupt capital impacts when E&S ratings or hazard assessments are updated, via clear transition mechanics and conservative but stable parameterisation rules,
– interaction with model change governance: E&S integration will often require iterative model evolution; supervisors and institutions need a clear, proportionate pathway for material model change classification, validation scope, and implementation timelines.
Under these priorities, E&S integration becomes simpler rather than more complex: it is embedded into existing IRB structures as a controlled set of admissibility conditions, conservative treatment rules, and stress-to-parameter translation mechanics. The alternative—adding granular mandatory E&S variables across models without a common materiality and governance spine—would produce exactly the outcome that should be avoided: inconsistent implementation, unverifiable comparability, and a structurally cosmetic framework.
Accordingly, further enhancement is both feasible and desirable, but it should be executed via a single, harmonised governance architecture that makes E&S risk decision-binding where material, and conservatively treated where uncertain, without multiplying parallel modelling requirements. Where quantification is not empirically defensible, Pillar 1 should require conservative governance-triggered treatments, while Pillar 2/ICAAP should remain the primary channel for forward-looking scenario severity assessment—without allowing dual, inconsistent risk stories.
Q7. Which requirements should apply in relation to the measurement of the performance of continuous models (e.g. Back-testing)? How could testing requirements be facilitated and enhanced for continuous models that are compliant with CRR, Part three, Title II, Chapter 3, Section 6 (Requirements for the IRB approach)?
Performance assessment of continuous rating models should be subject to the same prudential objectives as discrete grade-based models, but the testing architecture must explicitly recognise that continuous assignment of risk parameters is structurally incompatible with several grade-level validation requirements embedded in the IRB framework. The appropriate response is therefore not to dilute validation standards, but to require a controlled and auditable discretisation layer for validation purposes.
Accordingly, a mandatory validation discretisation should be introduced for continuous models, constructed solely for testing and monitoring and not for capital calculation. Supervisors may consider prescribing a minimum standard bucket schema (range and minimum counts) to reduce discretion in discretisation design.
The following requirements should apply.
– institutions should construct a stable, monotonic and risk-ordered validation scale by discretising the continuous output into a predefined and governance-approved set of buckets, designed to be homogeneous in risk and sufficiently populated to support statistical testing,
– the discretisation must be independent from outcome realisations and must be fixed ex ante for a defined validation cycle in order to avoid implicit outcome-driven optimisation,
– the number and structure of buckets should be justified based on portfolio size, default scarcity and model use, and should be sufficiently granular to preserve discriminatory information while ensuring minimum observation thresholds per bucket,
– the same discretised scale should be used consistently for all grade-level validation requirements prescribed by the IRB framework, including homogeneity, heterogeneity, concentration and minimum observation checks,
– back-testing of PD should be performed on the discretised validation scale using realised default rates per bucket, with explicit confidence intervals and statistical tests consistent with those used for discrete grade systems,
– calibration testing should distinguish clearly between testing of raw model output and testing of final applied parameters after conservatism and adjustments, and both layers should be monitored on the discretised validation scale,
– for LGD and CCF continuous estimates, realised outcomes should be compared against predicted values both at individual exposure level and on the discretised validation scale, in order to test bias, stability and dispersion simultaneously,
– discriminatory power should be assessed using continuous metrics such as AUC or Gini at the individual level, but these metrics should be complemented by discretised bucket-level stability and ordering tests to ensure consistency with the IRB grade philosophy,
– temporal stability and migration behaviour should be assessed on the discretised validation scale to ensure that the continuous model does not generate excessive short-term volatility that is inconsistent with the intended risk horizon,
– override and expert adjustment impact should be monitored on the discretised validation scale, including directional consistency, frequency by bucket and impact on observed performance,
– institutions should be required to demonstrate that the discretisation itself does not materially distort the risk ordering of the underlying continuous model output,
– supervisory comparability should be supported by requiring that core validation outputs are produced on the discretised validation scale in a standardised format across institutions.
Testing requirements for continuous models can be materially facilitated and enhanced by explicitly separating three layers.
The first layer is continuous-level diagnostics, used to assess ranking quality and marginal predictive power.
The second layer is discretised validation testing, used to satisfy all grade-based IRB requirements.
The third layer is application-level monitoring, used to assess stability of final parameters and capital outcomes.
This layered architecture allows continuous models to remain technically flexible while ensuring that the IRB validation logic remains fully applicable, comparable across institutions, and auditable under CRR Part Three, Title II, Chapter 3, Section 6.
Q8. Which requirements should apply in the application phase of continuous models (e.g. overrides)?
The application phase of continuous IRB models should be governed by requirements that preserve the same control objectives as for discrete grade-based systems, while explicitly addressing the specific risks created by point-estimate assignment and the absence of natural grade boundaries.
In particular, the following requirements should apply.
– application of continuous model outputs must be based on a clearly defined and governance-approved transformation from raw model output to final regulatory risk parameters, including any smoothing, truncation or monotonicity constraints applied prior to use in capital calculation,
– overrides and expert adjustments must be defined as explicit deviations from the continuous model output and must be applied through a controlled overlay mechanism, rather than by direct manual modification of the underlying continuous score or parameter,
– override policies must specify admissible override triggers, evidentiary standards and decision authority, with explicit differentiation between data quality corrections and risk judgement-driven adjustments,
– overrides must be directionally constrained and monotonic with respect to risk, such that they cannot produce a lower risk outcome when new adverse information is introduced,
– the magnitude of overrides must be bounded and subject to predefined escalation thresholds, including mandatory senior review and independent validation involvement for material impacts,
– override frequency and impact must be monitored on the discretised validation scale, enabling comparability with discrete models and detection of structural bias or capital optimisation behaviour,
– overrides must be explicitly linked to observable obligor- or facility-specific information and must not be used to compensate for known structural model deficiencies,
– systematic or recurring override patterns must trigger a mandatory model review or redevelopment process, rather than being treated as an acceptable permanent overlay,
– application of overrides must be fully auditable, including traceability from raw model output to final applied parameter and capital impact,
– consistency must be ensured between override practices in regulatory capital models and those used in internal credit decisioning and limit management, in order to avoid a dual governance regime,
– overrides and adjustments must be subject to ex-post performance monitoring, including analysis of realised outcomes relative to overridden and non-overridden estimates,
– any use of discretisation or bucket structures in the application phase (for example for policy rules or limits) must remain consistent with the discretised validation scale used for model performance testing,
– application processes must include controls to prevent implicit outcome-driven tuning of continuous outputs, such as frequent recalibration or ad-hoc transformation changes outside formal model change governance,
– final applied parameters must respect all regulatory floors and conservatism requirements in a transparent and mechanically verifiable manner.
These requirements ensure that continuous models can be applied operationally without introducing additional degrees of freedom that would weaken comparability, governance discipline and supervisory control relative to traditional discrete IRB systems.
Q9. Which challenges have you encountered in implementing the new CRR definition of facility?
The new CRR definition of facility introduces material implementation and governance challenges that go well beyond technical data alignment. The core difficulty is that the definition permits multiple valid aggregation levels across contracts, while IRB estimation, calibration and application processes require a single, stable unit of observation in order to remain internally consistent and auditable.
In practice, the main challenges are the following.
– the ability to define a facility either as a single contract or as a set of contracts creates ambiguity in the statistical unit used for PD, LGD and CCF estimation, which directly affects default counting, realised outcome construction and long-run average calibration,
– the same exposure structure can be legitimately represented at different aggregation levels for different parameters, creating a material risk of internal inconsistency across PD, LGD and CCF without an explicit single-facility policy,
– alignment of the new facility concept with existing data architectures and historical time series is non-trivial, as legacy systems and datasets are typically contract-based and not designed to support dynamic regrouping into facility sets,
– back-mapping historical defaults and recoveries to a redefined facility perimeter introduces non-observable assumptions and compromises the integrity of realised outcome series,
– implementation of a consistent facility definition interacts directly with the default definition, especially for retail and revolving exposures, and may materially change default event attribution compared with legacy practice,
– for off-balance-sheet and revolving structures, facility aggregation interacts with limit management and behavioural drawdown patterns, creating structural inconsistencies between estimation and application phases,
– for LGD estimation, the facility definition directly affects recovery aggregation, collateral attribution and cash-flow construction, especially where multiple contracts share collateral or enforcement processes,
– for CCF estimation, facility aggregation affects undrawn amount definition, reference dates and the construction of realised CCFs, and therefore materially alters estimated conversion behaviour,
– mixed collateral and mixed product structures within a single obligor complicate the identification of a single facility perimeter that remains meaningful for both risk differentiation and recovery modelling,
– implementation across multiple business lines and booking platforms requires extensive re-engineering of reporting and risk data pipelines, with material operational and control costs,
– changes to the facility definition are model-material and typically trigger model redevelopment and supervisory approval processes, even when the underlying risk drivers remain unchanged,
– parallel operation of legacy and CRR3-compliant facility definitions during transition periods creates significant governance and reconciliation complexity,
– the lack of explicit regulatory hierarchy between business-driven aggregation and modelling-driven aggregation makes it difficult to design a defensible and consistent internal policy,
– facility aggregation choices have measurable capital impact, which introduces model risk and supervisory sensitivity around perceived optimisation versus genuine alignment with business practices.
Overall, the principal challenge is not the concept of aggregation itself, but the absence of a single, mandatory and operationally anchored facility definition across parameters. Without additional guidance enforcing a consistent level of aggregation, the new definition materially increases model risk, governance burden and the risk of divergent supervisory interpretations.
Q10. Should a consistent and single facility definition be applied across all risk parameters?
A consistent and single facility definition should be applied across all risk parameters.
A unified facility perimeter is a necessary precondition for internal coherence of the IRB framework. Without a common unit of observation, PD, LGD and CCF estimates cease to represent the same underlying risk object and the parameter set becomes conceptually and operationally fragmented.
From a prudential perspective, a single facility definition is required to ensure that:
– default counting, realised outcome construction and calibration samples are aligned across parameters,
– the estimation unit used for PD, LGD and CCF reflects the same contractual and economic exposure structure,
– risk differentiation and risk quantification are applied to the same object in both the modelling and application phases,
– capital outcomes are not influenced by parameter-specific aggregation choices that are unrelated to underlying risk,
– governance and validation processes can be performed on a single, traceable and auditable exposure perimeter,
– back-testing and performance monitoring remain interpretable and comparable across parameters and institutions.
Allowing different facility aggregation levels for different parameters introduces degrees of freedom that are not linked to risk but to modelling convenience or legacy system constraints. This weakens supervisory convergence and creates scope for structural inconsistency and unintended capital effects.
At the same time, the requirement for a single facility definition must be accompanied by proportionate and pragmatic guidance on how contracts should be grouped into a facility in practice, in particular for multi-contract, multi-collateral and revolving structures. The objective is not to force a specific commercial product taxonomy, but to enforce a stable and defensible aggregation logic that is applied consistently throughout the IRB lifecycle.
Accordingly, a single and consistent facility definition across all risk parameters should be mandated as a core element of supervisory convergence. During transition, a minimum standard for parallel-run reconciliation and lineage controls should be specified to prevent facility-perimeter drift from becoming a hidden capital driver.
Q11. Are adjustments proposed in the representativeness requirement for the CCF parameter also suited for PD and LGD risk parameters? Which amendments would be needed to accommodate PD and LGD specificities?
Yes at the level of principle, the representativeness adjustments proposed for CCF are suited for PD and LGD, but only if they are translated into parameter specific requirements that reflect fundamental differences in observability, horizon, and regime behaviour. The main virtue of the CCF representativeness logic is that it separates the role of data samples by purpose and preserves strictness where it matters most: testing and performance assessment. That architecture can be extended to PD and LGD to enhance supervisory convergence without forcing artificial precision in model development.
However, PD and LGD require targeted amendments because their estimation and validation problems are not isomorphic to CCF.
Amendments that are necessary for PD:
– explicit distinction between calibration sample and development sample must be formalised, with a minimum requirement that calibration is anchored to a default-rate window demonstrably representative of the current portfolio and the intended calibration philosophy (TTC vs PIT), while allowing broader development samples for discrimination,
– representativeness tests must be defined along dimensions that are PD-relevant: obligor segment mix, rating philosophy, underwriting vintage, geographic and sectoral composition, and default definition consistency; CCF-style generic dimensions are insufficient without PD-specific mapping,
– the framework must specify minimum treatment for low-default portfolios, where statistical representativeness cannot be demonstrated via default counts alone; alternative evidence standards should be accepted, but only with explicit conservatism requirements,
– it must require explicit handling of structural breaks in default behaviour driven by underwriting changes, covenant tightening/loosening, macro regime shifts, or portfolio rebalancing; representativeness cannot be a static checklist,
– for external or pooled data, the representativeness framework must require a formal linkage between external default experience and the institution’s obligor population, including demonstrable comparability of default drivers and a conservative adjustment when comparability is incomplete,
– the consequence function for non-representativeness must be parameterised: what triggers recalibration, what triggers segmentation changes, what triggers MoC, and what triggers model redevelopment must be explicit and not left to narrative discretion,
– the interaction between representativeness and long-run average requirements must be clarified: a long historical window may be statistically rich but economically non-representative; the framework must specify how to reconcile historical breadth with current population relevance without creating hidden discretion.
Amendments that are necessary for LGD:
– representativeness must be defined separately for defaulted population and for recovery process conditions, because LGD is driven by collateral type, enforcement regime, recovery timing, and market liquidity; a PD-like representativeness lens is not sufficient,
– the framework must explicitly address the dependency of realised LGD on workout strategy and recovery timing, including sale delays and collateral management; representativeness testing must control for institutional process shifts that mechanically change realised LGD,
– for collateralised portfolios, representativeness must include valuation standards, haircuts, and collateral reappraisal frequency, because these change both observed recovery rates and discounting assumptions; these are not peripheral—they are structural LGD drivers,
– treatment of incomplete recoveries must be standardised within the representativeness logic: where workout is not complete, the acceptable use of estimates must be coupled with explicit bias controls and MoC linkage to estimation uncertainty,
– downturn representativeness must be parameter-specific: for LGD, representativeness of downturn conditions is not optional, because downturn LGD is explicitly mandated; the framework must state what evidence constitutes “downturn capture” and what fallback conservatism applies when it is absent,
– external and pooled data rules must include enforceable comparability conditions for legal regime, seniority structure, collateral enforceability, and time-to-recovery distributions; without these, pooled LGD data creates false confidence,
– representativeness consequences must be differentiated from PD: for LGD, non-representativeness often requires segmentation redesign and recovery-process controls, not merely recalibration.
Cross-parameter amendments required to preserve simplicity and convergence:
– a single unified representativeness framework should be adopted, but with parameter modules (PD-module, LGD-module, CCF-module) that define: required dimensions, minimum evidence standards, acceptable proxies under data scarcity, and mandated conservatism responses,
– the framework should hard-separate “development flexibility” from “testing strictness”: flexibility in development is acceptable if and only if the testing sample used for performance assessment is demonstrably representative or conservatively adjusted,
– supervisory reporting of representativeness assessments should be standardised into a small set of auditable outputs to prevent narrative arbitrage and to improve cross-bank comparability.
Under these amendments, extending the CCF representativeness approach to PD and LGD would be appropriate and would improve convergence. Without these amendments, a direct transplant would either fail to address PD/LGD-specific risks or would push institutions into excessive narrative discretion, which is the opposite of supervisory convergence.
Q12: Do you consider further simplification of the representativeness requirement, as proposed for the CCF parameter, as necessary for PD and LGD and if so, what kind of simplification?
Further simplification is necessary for PD and LGD, but only in a specific sense: simplification should reduce interpretative degrees of freedom and duplicate documentation, not reduce prudential content. The representativeness requirement is currently complex primarily because it is applied as an open-ended narrative obligation with too many dimensions, too many exception pathways, and no standardised consequence function. That creates heterogeneity across institutions and supervisors, which is the opposite of convergence. Representativeness should be treated as a controlled condition with a predefined consequence function, not as a narrative obligation.
A further simplification analogous to the CCF approach is therefore appropriate for PD and LGD, provided it is implemented through a single modular framework with explicit minimum standards, clear differentiation by sample purpose, and pre-defined remediation actions.
The simplifications that are necessary and proportionate are the following.
– a formal separation of representativeness by sample purpose, with three mandatory sample types defined and treated differently: model development, calibration, and performance testing; development may tolerate controlled non-representativeness, calibration must be anchored to the current population and rating philosophy, and testing must be strictly representative or subject to conservative adjustments,
– a reduction to a small set of mandatory representativeness dimensions for each parameter, with a closed list and a clear mapping to risk drivers; institutions can analyse additional dimensions, but supervisory assessment should focus on the mandatory set to support comparability,
– a standardised evidence hierarchy for low-default or data-scarce portfolios, defining what is acceptable evidence when statistical proof is impossible, and linking that explicitly to mandatory conservatism; this removes the current situation where “lack of data” is handled idiosyncratically,
– an explicit, standardised consequence function: for each material representativeness deficiency, the framework should specify the permitted remediation options and the minimum prudential response, distinguishing between recalibration, segmentation change, MoC increase, use of external/pooled data, or model redevelopment,
– a defined treatment of structural breaks and regime shifts, with a standard trigger set and minimum monitoring outputs, so representativeness is not treated as a static annual checklist but as a controlled ongoing condition,
– a single, standard template for representativeness reporting to supervisors, limiting narrative variability and enabling benchmarkability across institutions; the template should force disclosure of deficiencies, the chosen remediation path, and the quantitative capital impact of conservatism,
– a parameter-specific but simplified rule for external and pooled data, defined through a short list of mandatory comparability tests and conservative add-ons when comparability is incomplete; this prevents pseudo-precision while enabling proportionate use of external data,
– for LGD specifically, a simplified and standard approach to controlling workout-process representativeness, by requiring a limited set of process stability indicators and explicit treatment of changes in recovery strategy that mechanically alter realised LGD.
This type of simplification is necessary because it converts representativeness from a broad qualitative principle into a reproducible, auditable and comparable control system. It reduces operational burden by eliminating repeated bespoke narratives while increasing prudential robustness by constraining discretion.
A simplification that merely removes requirements or reduces the scope of representativeness analysis would be counterproductive. The required simplification is therefore a simplification of structure, evidence standards and remediation logic, not a simplification of prudential expectations.
Q13. Should these simplifications be pursued? Do you have any preferred approaches with respect to these simplifications?
These simplifications should be pursued, but only under a strict prudential design constraint: each fallback must be opt-in, conservative by construction, back-testable, and governed by explicit eligibility criteria that prevent the fallback from becoming a capital arbitrage route. The objective is not to “delete modelling”; it is to remove modelling activity whose marginal risk-sensitivity gain is demonstrably dominated by model-risk dispersion, interpretative heterogeneity, and supervisory review cost.
Preferred approaches by item are as follows.
C10 – MoC A/B fallback + standardise MoC C
A fallback for MoC A and B is appropriate where decomposition into subcomponents creates double counting risk and produces non-comparable outcomes across institutions. The preferred design is a single conservative overlay that is parameter-specific and anchored to an auditable measure of estimation uncertainty, applied on top of the best estimate after adjustments. The eligibility condition should be explicit: the fallback is allowed only when the institution demonstrates that subcomponent quantification is not operationally stable or creates double counting that cannot be resolved through governance. MoC C standardisation is justified because heterogeneity in “general estimation error” methods is currently a driver of unwarranted variability. The preferred approach is to standardise MoC C around a limited set of permissible statistical constructions with explicit treatment for low-default portfolios, rather than prescribing one single estimator.
C11 – Indirect and direct cost treatment via [X%] LGD uplift
A fallback uplift is acceptable as a proportionality tool, but only if it is anchored to evidence and constrained by floors and periodic recalibration. The preferred structure is a relative uplift on realised LGD, applied consistently across relevant segments, with a minimum and maximum bound and with a requirement to reassess the uplift at defined intervals based on observed cost experience where available. The uplift must be conditional on demonstrable inability to capture costs with sufficient data quality, and should not be available where cost capture is feasible. The key is to prevent permanent ignorance of costs when the institution could in fact measure them.
C12 – Downturn LGD/CCF fallback using reference value or fixed add-on
A fallback should be pursued, because downturn estimation is one of the highest-burden areas relative to incremental risk sensitivity, particularly when data is limited. The preferred approach is a tiered structure: if sufficient downturn observation years exist, allow a simplified reference-value-based downturn calibration; if not, require a fixed conservative add-on with a transparent mapping to portfolio type and collateral characteristics. The fallback must remain bounded below by long-run average floors and must be subject to outcome monitoring to prevent systematic underestimation. The reference value concept is directionally correct because it ties downturn calibration to observable high-loss years, but it must be framed as a conservative minimum rather than an optimisation target.
C13 – In-default LGD fallback using an SA-like approach
A simplified SA-like approach for in-default LGD can be justified where materiality is low and where minimum provisioning backstops materially mitigate undercapitalisation risk. The preferred design is eligibility-gated: only for portfolios where recoveries are operationally stable and where back-testing shows conservatism relative to realised recoveries. The SA-like approach must not eliminate monitoring; it must require a periodic comparison of realised loss outcomes against the simplified parameter to ensure that conservatism is maintained. This is a governance simplification, not an abandonment of recovery discipline.
C14 – Fixed CCF for entire exposure types
A fixed CCF fallback should be pursued, because CCF models often exhibit weak discriminatory power and generate high model-risk dispersion relative to their marginal risk differentiation. The preferred approach is a conservative fixed CCF with a mandatory MoC overlay and hard back-testing requirements, applied at exposure-type level only when the institution demonstrates persistent instability or non-predictiveness of model-based CCF. The design must prevent cherry-picking: opt-in should be portfolio-wide within the eligible exposure type, and supervisors should have the ability to revoke the fallback where performance deteriorates.
C15 – Flexibility around the 12-month fixed horizon approach
More flexibility should be pursued, because the 12-month fixed horizon approach can introduce technical bias and misalignment between estimation and application phases. The preferred solution is a controlled hybrid: permit cohort-based elements for realised CCF construction and risk driver alignment, but require reconciliation to the 12-month metric via a standard reporting bridge, with mandatory explanation and governance triggers when deviations from long-run average behaviour are significant. The purpose is to reduce estimation bias and operational complexity without weakening comparability. The existence of input floors provides an additional prudential safeguard, but it should not be used as a justification to loosen governance.
Net position
Yes, the simplifications should be pursued. The preferred implementation is a disciplined opt-in fallback regime with explicit eligibility, conservative calibration rules, standardised documentation, mandatory back-testing, and clear supervisory revocation rights. If those conditions are met, simplification increases comparability and reduces unwarranted variability without weakening prudential integrity. Eligibility should be applied consistently within the defined portfolio scope to prevent selective adoption driven by capital outcomes.
Q14. Do you have any comments and suggestions with reference to the calibration of the fall back approaches?
Calibration is the point where “proportionality” either remains a supervisory efficiency tool or becomes a hidden channel for capital underestimation. The calibration of fallback approaches should therefore be governed by a single prudential principle: fallback calibration must be conservative-by-design, empirically anchored where feasible, and automatically tightened under data scarcity or weak back-testing performance. Calibration should be standardised enough to ensure comparability, but not so rigid that it ignores portfolio structure and collateral realities.
The following calibration design features are recommended across all fallbacks.
– calibration parameters should be specified as functions of observable portfolio attributes (segment, collateral type, seniority, jurisdiction, recovery horizon) rather than as single universal constants, but the functional form should be standardised to avoid bespoke optimisation,
– every fallback parameter should be calibrated to a high quantile of observed outcomes rather than to the mean, with the quantile level standardised by exposure type and data quality tier,
– calibration should explicitly incorporate a data quality / representativeness tiering, where lower data quality triggers higher conservatism through pre-defined multipliers, not through narrative discretion,
– each fallback must have a hard floor linked to the relevant SA parameter (or to a conservative benchmark), and where applicable also a linkage to downturn add-ons and long-run average floors,
– back-testing should be embedded into calibration through an automatic escalation rule, where persistent underperformance triggers mandatory parameter uplift and potential withdrawal of fallback eligibility,
– fallback calibration should be reviewed on a fixed cycle and also event-driven (material process change, legal regime change, collateral valuation methodology change, macro regime shift), with the triggers defined ex ante.
Calibration comments by fallback area are as follows.
MoC fallback (C10)
MoC A/B fallback calibration should be tied to a conservative proxy for incremental estimation uncertainty, not to subjective management judgement. The fallback should be expressed as an add-on to the parameter estimate derived from either the dispersion of the estimator, or from a conservative bound on bias arising from the identified deficiency class. MoC C standardisation should not prescribe one single estimator, but it should prescribe a limited family of estimators and a common confidence level. For low-default portfolios, the standard should require conservative treatment that scales with default scarcity, to avoid pathological outcomes where “statistical precision” is assumed without data.
Cost uplift fallback (C11)
If a relative uplift [X%] is introduced, it should be calibrated to observed cost experience where available and otherwise to conservative supervisory benchmarks by exposure type and jurisdiction. The calibration should be segment-specific: foreclosure-heavy secured portfolios should have a structurally different cost uplift than unsecured recoveries. The uplift should be bounded by a minimum floor and a maximum cap only for operational reasonability, but the design should bias toward conservatism when data is weak. The uplift should not be static: if the institution later develops the capability to measure costs, the fallback should be revoked and replaced by measured inclusion.
Downturn fallback (C12)
Reference-value-based calibration should be anchored to a conservative interpretation: the reference value is a minimum downturn anchor, not the target. The fixed add-on option should be calibrated by exposure type and collateral characteristics, not as a single uniform add-on. Eligibility for the “reference value as basis” should require a minimum number of years that capture downturn-like conditions; where that is not met, the fixed add-on should become mandatory. Calibration should incorporate a mechanism to avoid procyclicality: once activated, the downturn uplift should not mechanically fall immediately when short-term conditions improve; a smoothing rule should be embedded to preserve prudential stability.
In-default LGD fallback (C13)
An SA-like approach in default must be calibrated to be conservative relative to realised recovery distributions. A simple requirement should apply: the fallback LGD must exceed a high quantile of realised in-default loss outcomes over a defined window, or else be uplifted. Back-testing should be conducted at least at segment level, and eligibility should be withdrawn if systematic underestimation is detected. Calibration should incorporate recovery horizon: in-default exposures with longer legal timelines should have structurally higher conservative parameters due to discounting and cost accumulation.
Fixed CCF fallback (C14)
Fixed CCF calibration should start from a conservative benchmark, and then be stress-uplifted. Calibration should not be to long-run average alone; it should incorporate tail drawdown behaviour and the “region of instability” near full utilisation. A prudent calibration rule is: fixed CCF equals at least the higher of (i) a conservative quantile of realised CCF outcomes, (ii) a floor tied to SA-CCF, and (iii) a supervisory benchmark for the product type. Back-testing must be mandatory and must be linked to an automatic uplift rule if realised outcomes exceed the fixed level beyond a defined tolerance.
12-month horizon flexibility (C15)
If cohort elements are permitted, calibration must preserve comparability through a standard “bridge” metric: institutions should be required to compute both the cohort-based estimate and the implied 12-month-horizon estimate and demonstrate that differences are explained by definitional effects rather than optimisation. Calibration should be constrained so that the flexibility cannot be used to reduce CCF materially versus the standard approach without strong empirical justification. Where a cohort approach produces lower estimates, a conservative overlay should be mandated unless evidence shows improved predictive validity and stable back-testing.
Implementation governance suggestion
Fallback calibration should be treated as a controlled supervisory regime with explicit eligibility, parameter setting rules, back-testing thresholds, and revocation mechanics. The calibration framework must be designed so that the operational relief comes from reduced modelling complexity, not from reduced capital. Fallback use should be time-bounded unless reaffirmed through periodic evidence, preventing permanent reliance where measurement becomes feasible.
Under these conditions, fallbacks can reduce burden and increase comparability without weakening prudential outcomes.
Q15. Do you see other potential simplification areas where the modelling burden is not commensurate to the gain in risk sensitivity?
Yes. There are additional areas where the current IRB modelling and supervisory review burden is structurally high while the incremental gain in risk sensitivity is either marginal, unstable, or dominated by model-risk dispersion and interpretative heterogeneity. The common pattern is not “modelling is bad”; the pattern is that some requirements force institutions to spend disproportionate effort on artefacts that do not materially improve capital adequacy, comparability, or forward-looking risk control.
The following areas merit consideration for simplification.
– Standardised validation output pack: require a single harmonised supervisory validation pack with fixed core metrics, confidence levels, and reporting templates, reducing bespoke narrative and reformatting costs without changing prudential substance.
– Unified and tiered treatment of low-default portfolios (LDP): introduce explicit tiers and pre-defined conservative options for discrimination and calibration under default scarcity, rather than forcing institutions into complex statistical workarounds with limited evidentiary value.
– Streamlined treatment of external and pooled data: codify a short list of admissibility tests and mandatory conservative overlays when comparability is incomplete, replacing the current case-by-case documentation burden and inconsistent supervisory interpretations.
– Simplified and standardised override governance: standardise override taxonomy, thresholds, escalation triggers, and monitoring outputs across PD/LGD/CCF models, eliminating duplicated policy artefacts and reducing supervisory friction while strengthening control comparability.
– Facility/obligor aggregation consistency controls: once a single facility definition is enforced, simplify downstream requirements by removing repeated cross-parameter reconciliation narratives and replace them with a single traceability control test.
– Calibration sample governance: formalise a small number of permissible calibration window constructs (TTC/PIT variants) with clear requirements and conservative safeguards, reducing bespoke calibration narratives that rarely improve actual calibration quality.
– MoC documentation rationalisation: retain MoC prudence, but reduce the burden of decomposing MoC into excessively granular subcomponents where the decomposition is not decision-relevant; enforce a standard MoC register and evidence schema instead.
– Discretisation layer for continuous models: mandate a single validation discretisation mechanism and then simplify duplicate grade-level testing requirements by applying them consistently to the discretised scale, avoiding parallel bespoke approaches.
– Harmonised downturn identification mechanics: standardise core downturn identification triggers and minimum evidence rather than allowing highly bespoke downturn period definitions whose marginal precision is weak relative to cost and variability.
– Recovery process stability indicators for LGD: replace expansive qualitative recovery narratives with a limited, standard set of process stability metrics (time-to-recovery distribution, collateral sale lag, enforcement route mix, cost ratios) that are directly linked to LGD behaviour.
– Model change classification and supervisory workflow: simplify the model change taxonomy and required evidence by introducing clearer thresholds and a standardised evidentiary package, reducing iterative supervisory dialogue that often adds process cost without improving risk measurement.
– Back-testing hierarchy and pass/fail logic: introduce a standard hierarchy of back-testing tests with clear interpretation rules and escalation actions, reducing heterogeneity in how “test failure” is defined and remediated across institutions.
– Simplification of concentration and granularity requirements: where requirements force extensive micro-segmentation with negligible incremental discrimination, introduce thresholds beyond which additional segmentation is discouraged unless it demonstrates material predictive improvement.
– Separation of Pillar 1 parameter estimation from Pillar 2 forward-looking overlays: clarify that certain forward-looking elements, including parts of E&S risk, are prudentially captured through stress testing and governance overlays when empirical parameterisation is not credible; this avoids forcing pseudo-quantification into Pillar 1 models.
The direction is consistent across these areas: reduce bespoke narratives, remove duplicated evidence requirements, standardise core governance artefacts, and introduce tiered conservative options where data scarcity makes “full sophistication” performative rather than risk-sensitive.
These simplifications would reduce modelling and supervisory review burden while improving comparability and auditability—without weakening prudential outcomes.
Q16. What do you perceive as challenges in your capacity to collect appropriate data, in particular in relation to indirect costs?
The challenge is not the absence of cost data per se, but the absence of defensible case-level attribution and lineage suitable for LGD parameterisation.The primary challenges in collecting appropriate data for indirect costs are structural rather than technical. Indirect recovery costs are generated by operational processes that were never designed as risk-parameter data sources, and the current regulatory expectations implicitly assume a level of cost attribution precision that does not exist in most institutions’ accounting and recovery infrastructures.
The main challenges are as follows.
– indirect costs are recorded at organisational or functional cost-centre level and are not natively attributable to individual obligors, facilities or recovery cases,
– recovery-related staff, legal and asset-management resources are typically shared across portfolios, products and legal entities, making allocation to single default events inherently model-based,
– time registration and activity-based costing systems are incomplete, inconsistent or unavailable for historical periods, particularly for legacy default and recovery cases,
– outsourcing and third-party servicer arrangements bundle multiple services into contractual fees that cannot be decomposed reliably into recovery-specific cost components,
– indirect costs evolve materially over time due to organisational restructuring, automation and changes in servicing strategies, which breaks the stability of historical cost relationships,
– indirect cost levels are sensitive to internal operating model choices rather than to exposure risk characteristics, introducing institution-specific artefacts into LGD estimation,
– recovery processes often span several years, while cost accounting cycles are annual and organisationally restructured, creating temporal misalignment between cost accrual and recovery outcomes,
– significant portions of indirect costs relate to portfolio-level activities (strategy, governance, litigation management, collateral management frameworks) rather than to case-level actions,
– legal and enforcement environments drive recovery effort and duration, but these drivers are not captured in accounting systems in a way that can be linked to cost consumption,
– changes in recovery strategy (for example, bulk sales versus work-out driven recovery) fundamentally alter the indirect cost structure, making historical data weakly representative for current processes,
– indirect costs associated with failed recoveries, discontinued actions or write-off strategies are difficult to associate consistently with realised recoveries and are often treated inconsistently across business lines,
– internal transfer pricing, recharges and overhead allocations are frequently revised and restated, compromising the auditability and comparability of historical indirect cost data,
– data lineage and control frameworks are typically designed for financial reporting, not for regulatory LGD modelling, creating gaps in traceability and reconciliation,
– consolidation across legal entities and booking platforms introduces inconsistencies in cost definitions and accounting treatments that cannot be resolved ex post through modelling assumptions.
As a result, indirect cost estimation is driven less by data quality and more by allocation conventions and management accounting design choices. This fundamentally limits comparability across institutions and creates model risk that is not related to credit risk itself.
This is precisely why a conservative and standardised fallback treatment for indirect costs, subject to periodic reassessment and back-testing, is operationally and prudentially preferable to attempting case-level attribution where the underlying data infrastructure does not support it.
Q17. Do you agree with the approach proposed by EBA? Do you see further measures as necessary?
Yes. The proposed analytical framework is appropriate and should be applied to the mandated L1 assessment reports. The six criteria jointly cover the core prudential trade-offs between risk sensitivity, operational complexity, transition costs and international consistency, and they provide a structured basis for avoiding ad-hoc or politically driven simplification initiatives. The framework should also require that any identified deficiency is mapped to the appropriate legal layer (L1 vs RTS/GL/Q&A) to avoid addressing interpretative issues through unnecessary L1 amendments
However, to make the framework operationally effective and to avoid it remaining a high-level narrative device, several additional measures are necessary.
First, the framework should explicitly distinguish between diagnostic use and decision use. The six criteria are suitable for diagnosing potential weaknesses of the framework, but the reporting mandates require decision-relevant conclusions. A formal rule should therefore be introduced that specifies which combinations of the criteria are sufficient to justify a recommendation for regulatory change. Without a defined decision threshold, the framework risks producing descriptive assessments without a clear supervisory position.
Second, the notion of materiality of miscalibration should be operationalised through a limited set of standard quantitative indicators. At present, the concept remains open-ended. A small and fixed set of benchmark measures—such as capital impact concentration by exposure segment, tail miscalibration indicators and cross-jurisdiction dispersion metrics—should be mandated to support comparability across different L1 reviews.
Third, the framework should explicitly incorporate a procyclicality and regime-dependence lens. A refinement that improves average calibration but materially increases capital volatility across the cycle should be treated differently from one that improves calibration in a structurally stable manner. This dimension is not fully captured by the current criteria and is particularly relevant for real estate and long-dated secured exposures.
Fourth, the framework should require an explicit assessment of model-risk and governance risk introduced by a refinement. Increased risk sensitivity that relies on fragile data, complex modelling layers or extensive expert judgment should be penalised in the evaluation, even if it improves theoretical calibration. This would ensure consistency with the objective of reducing unwarranted variability and supervisory burden.
Fifth, the interaction with Pillar 2 and ICAAP practices should be formally incorporated. A change to L1 that merely shifts risk sensitivity into Pillar 2 overlays or institution-specific capital add-ons does not reduce system-wide complexity and may weaken transparency. The framework should therefore require an assessment of whether an L1 refinement substitutes, duplicates or displaces existing Pillar 2 mechanisms.
Sixth, the framework should require a standardised transition and implementation risk assessment, including data readiness, system dependencies and supervisory approval capacity. Transition risk is currently treated mainly as a cost consideration, but for internal model frameworks it is also a prudential risk, as prolonged parallel runs and partial implementations can weaken control environments.
Finally, the framework should mandate a short, standardised supervisory impact summary for each report, explicitly stating whether the proposed refinement is expected to increase or decrease supervisory intervention intensity, model approval workload and ongoing monitoring effort. This would ensure that simplicity gains are assessed from a supervisory operations perspective, not only from an institutional compliance perspective.
In summary, the proposed framework is sound and should be applied. To make it decision-effective and comparable across L1 reports, it should be complemented by explicit decision thresholds, standard quantitative indicators for miscalibration, a procyclicality lens, a model-risk and governance-risk dimension, a Pillar-2 interaction assessment, a structured transition-risk analysis and a supervisory impact summary.