Primary tabs


Please refer to the attached pdf file for the full response including the covering letter
Key messages:

We recognise the challenges in defining rules which have an “aim at reducing unjustified variability in capital requirements by ensuring consistency in model outputs with regard to downturn LGD and CF estimation and thus comparability of risk weighted exposure amounts. We also note the EBA is aware the proposed model component approach in some ways is a rather complex method, where additional guidance may be necessary, just as it may require substantial resources both from institutions and supervisors when it comes to its implementation.

Although the aims and nature of the component based approach are understood and not considered overly complex, we would describe the approach as convoluted and overly complex from a practical perspective. Our opinion is the proposed approach does not achieve the EBA’s objectives of reducing undue variation in internal models or aid transparency. To the contrary, members believe the component based approach will add to the complexity and reduce transparency of IRB estimates. We have strong reservations over its suitability for use on other portfolios, particularly wholesale models.

The consultation paper states “the proposed approach however appears relevant in order to ensure harmonised approaches across institutions and limit the quite divergent industry practices surrounding the modelling of downturn LGDs and CFs. In this light, on balance, providing a fairly prescriptive and methodological approach appears necessary”.

Whilst the logic behind the component based approach has merit, there are a number of elements to its application that could unintentionally result in greater variability and less transparency as opposed to the stated aims of the consultation.

For example, undue variation would be introduced through:

1. Reliance on an expert panel where no strong or meaningful correlation exists between the model components and economic factors. For Retail, especially unsecured portfolios, we believe it would be difficult to form these relationships and the role of the expert panel may in practice be greater than anticipated,

2. Lack of transparency on how a final downturn LGD estimate would be constructed. Without any specific guidance, firms may adopt different approaches resulting in materially different impacts on final LGD estimates,

3. The 20 year horizon theoretically allows the use of different downturn periods for the same market,

4. The use of a downturn with lower losses due to a higher correlation with economic factors may be difficult to justify to competent authorities if the analysis identifies a downturn period that is different to their expectations.

There will be difficulties in implementation of the model component approach if there is paucity of default observations which is particularly applicable to some Wholesale portfolios.

Further difficulties in arise from the necessity to create understandable risk metrics, that are built from components covering different time periods and which will further the cause of risk management.

This leads to a key question

a) Should the determination of the economic downturn be based upon an assessment of each model component and its underlying risk drivers?


b) Would an economic downturn be better assessed holistically taking into account a more prescriptive direction by each competent authority to focus on the prior periods to be given prominence in the assessment? The analysis of the individual component values at a more granular level can then be used to understand and validate the downturn using expert judgment.


Our view is the proposals have a number of shortcomings which lead us to recommend the EBA review and revise the proposal. Our members support a simpler approach to determining the duration of the economic downturn.

We recommend the published RTS would benefit from setting out its principles by which the outcome can be assessed.

• Aimed at reducing unjustified risk weight variability whilst retaining risk sensitivity.

• Applicable to all exposure classes which LGD and EAD can be estimated (i.e. retail, corporate, low-default).

• Applicable to any type of model estimation methodology.

• Clear how the downturn add-on is to be included within model estimates – or state if banks are free to choose how to do so.

• Balanced in terms of the cost to banks (development cost and RWA impact) and supervisors (approval and oversight)

Our recommendation is the EBA should mandate competent authorities to define appropriate time periods applicable in each jurisdiction and exposure class. Specific guidance on the time period to be considered for each downturn, at least in key markets would result in more comparable approaches across institutions.

The consequence would be a consistent and comparable approach for each exposure class across institutions within jurisdiction. This approach would enable work-out periods to be specified within each jurisdiction thereby also contributing to determining the appropriate overall level of the capital in that jurisdiction.

This would allow institutions to focus on estimating parameters to an equivalent severity and enable expert panels to focus on the qualitative aspects and challenges to model developers.

For institutions that do not have sufficient data, we support the Bank of England’s approach to allow institutions to combine internal and external data in order to derive historical synthetic data to be used as the basis for modelling. . Our recommendations are aligned to the Bank of England’s proposals defined in CP 29|16 .

There are significant benefits to this approach which would simplify modelling choices and obviate dilemmas and any possible bias faced by expert panels. This would allow a more consistent framework for competent authorities and allow easier comparison of models. This is the approach set out with respect to the modelling of the UK residential mortgage portfolios.

Should the EBA conclude not to proceed with the model component approach we would not object to the use of either the reference value approach or supervisory add-on approach. However, we think both alternative options under the supervisory add-on approach may lack detail in the consultation paper and request the EBA provide further guidance.

We encourage the EBA to review their proposal.

Response to Q1

We are concerned the preferred approach has been designed with residential mortgage portfolios in mind and the rationale for selecting this approach (p.4-5 in the consultation paper) do not hold true for all credit portfolios, in particular wholesale portfolios:

Rationale (a) states recovery rates and realised drawings cannot be only explained by credit factors (default rates), but in the case of the wholesale lending, especially corporate specialised lending, collateral values are generally highly correlated with the obligor asset values. It would be completely appropriate to consider only default conditions for these portfolios.

In the case of rationale (b) we agree LGD modelling needs to take into account the fact the LGD is a probability distribution and LGD modelling cannot directly model the expectation. Although this is relevant to LGD model generally and is not specific to modelling downturn sensitivity, the resolution process for defaulted corporate exposures differs to residential mortgages as the aim is to restructure and maintain the corporate as a going concern, as opposed to liquidation of the collateral (for example). The bi-modal nature of losses is driven by different seniorities of debt, which are not dependent on economic factors, and not from different resolution paths (cure vs. liquidation).

The consultation seems to assume firms will have data readily available over a 20-year period for model components. This is not likely to be the case for all firms and across all EU jurisdictions and where it is available, it is likely it will not be reflective of the portfolio today.

Therefore some degree of ‘back casting’ and data manipulation would be required to create the data to meet this specific requirement. This could lead to unintended variability as firms are likely to adopt different approaches.

In the UK there is a requirement to go back to the late 1980s or early 1990s to incorporate 2 recessionary periods for residential mortgages.

In order to minimise variation and aid comparability we believe it would be beneficial for a consistent EU approach to calibrate estimates to a downturn equivalent to the early 1990s.

A consistent approach should be adopted where internal data is not available over this time period and the recommended approach to calculation of long run averages (LRA). Is the LRA the average of all the data or the average of 20 years’? Either way, this may result in the LRA being based on incomplete cycles and therefore a number that is variable over time. This leads members to question how often we would be expected to update our LRA parameters.

The use of economic variables to define the precise period of the downturn may have merits but it can be argued this should be carried out by the competent authorities in order to prescribe the downturn periods for their jurisdiction and to achieve consistency.

Where institutions do not have the granular historic data required for the model component approach they should be allowed to use the alternative approaches (reference value or supervisory add on), until such time that sufficient data history will allow the DT LG/CF development using the model component approach.

Another suggested alternative approach for consideration would be to accept a stress testing-based approach based on extrapolation of a reduced amount of historical data (e.g. as used in IFRS 9 models).
The consultation paper provides little detail on expectations on how the component based approach might be applied for CFs.

For LGD, we can see how model components can be identified (e.g. cure rate or propensity to repossess for residential mortgages) but for CF, we believe the only meaningful component is CF itself.

Also from a Retail perspective, we believe there is less likelihood of developing any meaningful relationship between CF and economic factors. For example, for credit cards, the proportion of available limit that is drawn down immediately prior to default is more likely to be driven by customer behaviour and management actions and not economic factors.

Therefore we believe the model component based approach is overly burdensome for CF and is unlikely to result in a meaningful outcome.

We recommend the potential use of the model component based approach is limited to LGD and firms are allowed flexibility to determine how they assess the impact of a downturn on CF estimation.
The draft proposal appears to indicate a one-level LGD model would not be expected, except in rare circumstances. It is common practice to model LGD based upon risk drivers, especially for residential mortgages; however for unsecured retail portfolios, LGD is often estimated directly.

Although it is clear what is required for residential mortgage portfolios, we are concerned the consequence of the draft proposals will be to introduce an onerous mandatory minimum modelling standard for all exposure types.

For Corporate exposures it is also unclear. If the proposal intends cure rates and recovery rates etc. by each collateral type, loan type, industry type or combination thereof as an individual model component, then this is going to be extremely difficult to calibrate and substantiate. It could result in a reduced ability to develop or maintain IRB models for corporate exposures.

The model component section could be clearer. Where a component is segmented and different parameter values are applied (i.e. LTV segmented PWD - partial write down - model) would each segment be treated as a different component? (Structurally high LTV exposures may have a different fundamental reaction to house price changes than in low LTV cases).

The end-use of downturn parameters (LGD or CF) is not clear in the RTS.

It could be implied downturn estimates are expected to serve as floors in the estimation. Yet, as per the November 2016 EBA consultation on PD estimation, LGD estimation and the treatment of defaulted assets , unresolved accounts are to be incorporated in LGD estimation process.

The paper is unclear on how to use such unresolved accounts in final downturn parameter estimation process. This leads to uncertainty regarding the application of consistent approaches.
The dependency approach may be difficult to operationalise in smaller firms where the relevant knowledge around linkages between economics and credit risk are focussed in the IRB modelling team (this could constitute an external cost).

LGD, and to an extent, CF, is so strongly driven by company and portfolio strategy, as well as political intervention, that we are not confident there will be significant harmonisation by requiring explicit links to the economy component by component. For these reasons, our view is that clear and unambiguous linkage with a handful of economic effects is unlikely and economic effect, magnitudes and timings.

Paragraph 3 lists 6 economic factors to be used at a minimum. We think this may become a box ticking exercise running all 6 of these plus any other potentially relevant economics for every single model component. For consumer credit for example, even if we saw close correlation with GDP growth we are concerned we may not be able to rationalise its use as a clear driver.

Members have developed and used stress testing and scenario models, and appreciate the challenges to obtain a strongly predictive multi-factor model, let alone a single factor one for any component.

Correlating movements in economic statistics to the LGD for unsecured portfolios can be very challenging. Models can miss turning points and the magnitude of peak values, and so the ‘back-casting’ approach where full data isn’t available can underestimate the downturn effect and require an increased amount of judgement overlay. This can be to the extent that renders the detailed economic analysis rather less significant than the expert judgement layer.

Industry studies have shown LGD by average year of recovery better captures the downturn effect than the LGD by year of default.

So to take into account the time lag effect and because LGDs are not only impacted by economic conditions at the time of default, downturn drivers should include forward looking elements such as interest rates in addition to measures such as changes in GDP (which are more backward looking).

The CP RTS requires “a uniform length of 20 years of historical data” for macroeconomic factors including factors such as default rates and credit losses. For some Wholesale segments this would not be possible jurisdictionally or otherwise. For interest rates, unemployment and GDP this requirement can be complied with but not for default rates and credit losses in all wholesale portfolios.

We conclude these factors may not have been considered, but should be.

We recommend the following should also be taken into account:

• Identifying the downturn and assessing the possible impact are separate steps which the CP RTS does differentiate. However the expectancy for the downturn to be correlated with the model component, which is determined as factors driving the bimodal shape of the LGD distribution, is not well explained.

• Yes, the bi-modal distribution will probably be more skewed in a downturn. The effect of a downturn period is impacted by the lag between default events and the outcome of work-out processes in each institution.

• Downturn impacts are driven by the choice and combination of downturn identifiers used to capture the state of the economy, exposure segment and an institutions’ recovery strategy.

Furthermore the model component approach would require a magnitude of data points. The more model components the more data points, which can be uncommon for wholesale portfolios. This means statistical significance will be difficult to demonstrate. Furthermore, if there is enough data and the identified several model components were impacted by the same economic factors at different time periods, there is a lack of clarity as to how should this be considered.
Yes, broadly we agree with this. Time series based around the period of the crystallisation of any loss is generally appropriate. This is because it focuses on the recovery rather than the default event which may have occurred a number of years before the loss was realised. This is about alignment between PD and LGD.

It may be beneficial to allow some flexibility in the time series to be used rather than narrowly focussing only on timing of the realisation however.

We think there are circumstances - particularly in a residential mortgage or other secured portfolio - in which there may be prolonged period between defaults and realisation of losses. This can result in the time series used for realisations encompassing different default periods. Treating realisations as a homogenous group from a time series point of view may obscure relationships with the underlying economics prevailing at the point of default.

For example in a forced sale discount (FSD) for residential mortgages, although the house price index (HPI) and the health of the housing market at time of sale will play an important role in determining the value of this model component, so will the circumstances of the customer at the original date of the default.

The EBA may be aware, FSD is considered by the PRA to be an important component of modelling a residential mortgage portfolio. Some of our members find grouping FSD by default date is far more meaningful to obtain a clear cyclical pattern.

So, for a “Secured loss given Write off model”, it may be more appropriate to sample based on point of default rather than point of realisation. The economic factors and collections policy at the point of default could be more likely to influence the outcome than the economic factors at the point of cure/write off.

We think in taking an approach whereby the overall LGD realisation governs the downturn period, supporting model component analysis can be grouped either by realisation or by default date, with the choice justified by the institution conducting the analysis.
Taking into our members’ experience of building forecasting and scenario analysis models, there are often compounding effects of economic metrics such as HPI and unemployment.

A short sharp shock may be moderately painful, but multiple years of poor annual performance of these factors can be far worse in their impact. Forecasting models often combine the change in the level of an economic factor with the rate of change as a modifying factor.

We are not convinced a fixed one-year downturn window is appropriate for all econometric parameters. For example for collateral values - where prevailing prices can fall for years at a time – it would be more logical to use the greatest peak to trough fall in valuation, rather than the largest 12 month fall. This is particularly important given the potentially long workout periods for some assets.

So in many cases using a simple ‘one-year’ value could result in missing the true effect of the economics on the outcome.

We recommend the downturn period should align with the most common workout period which would usually be greater than one year.
Economic downturns are likely to be structurally different and there is no certainty a past downturn will reflect the structure of the next downturn. Structurally different downturns are likely to exhibit different levels of severity.

Article 5 asks banks to consider a plausible variability in economic factors for the future. This is akin to stress testing which the CP states is not the point of the RTS. However, stress testing is likely to inform any expert judgement in assessing the appropriateness of the severity of the last two economic downturns.

The PRA publication CP29/16 uses the recession of the late 1980s / early 1990s as the downturn reference point for residential mortgages. However, this is not a plausible future scenario (on the grounds of interest rates, changes in recoveries practices, differences in the BTL market etc).

The proposal focuses on economic effects, and ignores business effects and changes in strategies which impact risk appetite and hence appropriate downturn values.

Internal variables may be more predictive than external economic variables and maximum loss may not have occurred when economic variables are at their worst. The proposal assumes the converse.

Realistically, being able to disregard the worst observed incidence of a model component due to it being idiosyncratic rather than economically driven (as proposed in the EBA’s presentation of the CP) is likely to be extremely difficult to rationalise to the satisfaction of any competent authority. Consequently, we agree there must be a forward-looking element to the proposal.

We recommend more guidance is provided on how to determine the ‘representativeness’ of data and how and what adjustments may be allowed based on statistical analysis or expert judgement in order to ensure the historical data is relevant for today’s business.
In this context the term ‘sufficiently severe’ requires further clarification and we would also welcome a definition of the term ‘plausible variability’.

Where the competent authority has specific requirements (e.g. reference to the downturn of the late 1980s / early 1990s for residential mortgages in the UK) we recommend they publish these requirements within their jurisdiction. We think this would be preferable to each firm reaching independent conclusions.

Guidance from each competent authority would be valuable, reduce the burden of analysis in this approach and enhance comparability within the EU.

Furthermore, we will need to rely on official EU government statistic sources (e.g. the UK ONS) for the economic factors. However it needs to be borne in mind the availability of certain time series data as well as definitions and calculations of some metrics have changed in the last 20-30 years. There seems no clear way in the draft to take this fact into account.
Yes we do. The example is useful however we do not understand scenario B. Specifically we do not understand how we could match a loss rate from an earlier period with a cure rate from a later period and result in a meaningful LGD. Picking and combining the very worst case for each factor is likely to result in values well in excess of stress scenario analysis.
Yes we have some concerns. Although selecting the scenario with the highest overall LGD is prudent, it may not be intuitive. If we consider LRA values are generally less severe than downturn, then the differences between downturn scenarios A and B could be significant (model redevelopments to align more closely with this CP could result in unintended capital shocks). Should the firm be allowed to argue the case for one downturn over another, or would this simply lead to increased risk variability across firms? We seek clarification of the EBA’s intention.

A further concern is the approach to calculate and compare multiple downturns seems unnecessarily convoluted. It may result with a choice of potential downturns. For example one potential downturn period may impute an LGD of 20% and another 21%. And yet choosing the 21% may not be as statistically robust as the 20%.

The CP assumes 20 years’ data will cleanly represent a ‘long run average’ by covering ‘at least two economic cycles’. We dispute this and think the concept of defining the long run itself will introduce variability in estimates.

This is a further example of why it may be more appropriate to pick the overall downturn point and derive the expected value of each LGD component from there, even if individual components are values more positive than their long run average. By looking for the worst economic LGD point this would ensure an internally consistent metric because the other factors will have a compensating effect.

The other challenge is when the portfolio composition impacts the value of the factors in a way unrelated to the economic factor chosen. In this case a simple historical analysis may not be sufficient.

We encourage the EBA to allow firms to use expert judgement in these scenarios.
And to this end this may be where the use of the expert panel may be of great benefit.
Yes. If a model component value has to be inferred based upon more recent relations between the component and the economic variable in question, then an element of uncertainty enters into the process for a quantitative perspective. Although this can be mitigated to a certain extent by the use of a Margin of Conservatism (MOC), any adjustment would contain an arbitrary element. From a qualitative perspective it might be difficult for the 'expert panel' to offer any firm view.

In the case of the UK, EAD/LGD is required to be calculated back to the late 1980s / early 1990s for residential mortgages. But it is unlikely most institutions will have suitable information available that far back to be able to apply their models or default definitions.

The Bank of England recently issued guidance on how to combine observed internal values and historical external data to derive proxy internal data for periods. Although this may be acceptable for the estimation of risk from a regulatory perspective it may present challenges to comply with the preferred approach requirements for it to be back-casted.

We encourage the EBA to review this potential conflict and to issue guidance on acceptable approaches, or the minimum expectations, and how to determine the ‘representativeness’ of data and what adjustments may be allowed based on statistical analysis or human judgement.

Otherwise if the matter is not addressed it could introduce a large amount of variability in LGD/EAD estimations for a UK institution, which would contradict one of the stated aims of the proposal.

Use of expert panel

The proposed approach expects institutions to also use an expert panel with increased levels of conservatism, using internal and external benchmarks to validate the estimate because back-casting of a single factor dependence relationship could not be relied upon to produce a suitable figure.

In our view this means in most cases, the role of the expert panel to derive a figure will be important. This will require competent authorities to accept the judgment of the expert panels in different firms could legitimately conclude differing outcomes.

The increased reliance on a panel of experts could lead to a greater subjective component to the economic downturn parameters which could lead to greater variation across banks.

Expert panels do already exist to a large extent. But membership is not independent from the modelling community for all our members.

The requirement for membership to be independent from the modelling community is challenging for smaller institutions. An additional panel may add complexity to existing governance processes. Institutions should have some flexibility on how they incorporate the expert panel within their governance and model development framework.
Please also see response to Question 2.

For certain portfolios, i.e. retail, we would expect to perform the CF analysis at the same time as we undertake the downturn LGD analysis. This is not necessarily the case for our members’ Wholesale portfolios.

However, should the final technical standard require the modelling of CFs downturn add-on we recommend it stresses they may have their own set of model components within the methodology.

For example in EAD models that use conversion factors, the approach could be adopted. For revolving products especially, it is likely a countercyclical relationship will be found to the economy as institutions retrench during an economic downturn.

However, when EAD is estimated directly, as allowed in Bank of England SS11/13, it is hard to see how the proposed approach will work.

Therefore, we think firms should be allowed to apply a more pragmatic approach.

For example, given the additional complexity of the proposed approach, there is unlikely to be a meaningful relationship for term products where EAD will be only marginally different to the current balance.
The text provided is adequate with following exceptions

• The explanation could be clearer if there were some diagrams, especially for the example given in Article 6 and Section 6.7 (as amended).

• We are keen to avoid the same text being repeated in multiple EBA publications.

• We are of the opinion the following two paragraphs (paragraphs 158 and 159 on pages 31 and 32) are ambiguous: the definition of long run average is problematic, as is deciding on a suitable metric to represent the ‘majority’ of model component realisations which are observed.


158 The downturn LGD should be computed according to the overall nature, severity and duration of an economic downturn determined in accordance with Article 6 of the [RTS on economic downturn]. This implies using the realised or estimated value for each model component in the respective economic downturn periods provided that those periods belong to the same downturn scenario, otherwise the long-run average value for each model component should be used.

159 For the purposes paragraph 158, institutions should compute the average realised value for each model component during the selected economic downturn period on a sample of exposures where the majority of the model component realisations are observed during that period but the cash flows should be discounted at the moment of default.
Yes, we think simpler alternative approaches for downturn adjustment should be considered in the spirit of proportionality.

We have no overall objection to the use of either the reference value approach or supervisory add-on approach. Although, we note our assessment is based on the limited information provided on the two approaches within the consultation.

Should the EBA opt to proceed with the model component approach institutions holding sufficient data should be required to develop their models accordingly. LGD outcomes for banks which are unable to meet the data and analytical requirements of the model component approach should always be more conservative than those banks with comparable portfolios which do meet the requirements.
Should the EBA choose not to proceed with the model component approach we do not object to the use of either the reference value approach or supervisory add-on approaches instead.

However, we request further guidance is provided on both of the options under the supervisory add-on approach as there in not sufficient detail for us to recommend a preference for one over the other.

Reference Value Approach

The reference value approach is simpler than the component based approach, however would still require some degree of analysis and judgment.

However, we are concerned it is unnecessarily complicated by the requirement to use as a minimum the list of economic factors set out in Article 3, the need to demonstrate single factor dependence relationships and the need to calculate and compare multiple downturns are an unnecessary complication.

The approach could be simplified and proportionality introduced by defining a downturn at LGD level based on an inspection of individual component time series and / or allowing the use of a regulatory prescribed downturn scenario provided by the competent authority.

Whilst this may not be the intention, in practice this approach could lead to the application of a virtual floor. Whilst the proposals indicate a hard floor is not the intention, it may be challenging to justify this when the reference value is higher than the downturn LGD, but it is not linked with downturn conditions. The lack of risk sensitivity may result in riskier lending where the floor is less punitive and an increased variability in risk parameters.

We also draw to the EBA’s attention; the reference value approach is already partially in use within the UK, with certain residential mortgage model component reference values either published in PRA supervisory statement SS11/13 or under consultation.

But as proposed we think the reference value approach may be overly simplistic as it doesn’t allow for differentiation due to portfolio composition factors. The method places a high burden on competent authorities, as opposed to the more intuitive approach of the competent authority prescribing an economic scenario per asset class which represents a downturn situation for each asset class.

The supervisory add on approach

The supervisory add on approach is simple but less risk sensitive than the other proposals. However, there are advantages to its simplicity as it allows banks to focus their resources on the estimation of the long-run average LGD or CCF values.

The approach would be consistent with the treatment of PD in the capital requirements calculation, where the regulation applies the economic stress to the long-run average PD estimates in the asset correlation component of the RWA formula. For retail exposures this factor is set by product type and a similar approach could be adopted for each jurisdiction within the EU.

Finally, the supervisory assessment of banks approaches to estimating the effect of the economic downturn in LGD and EAD will likely involve the supervisor benchmarking each banks estimates within their jurisdiction. The outcome of this being a convergence of estimates between banks as the supervisor crystallises its expectation of the add-on for each portfolio type within their jurisdiction, for example, the downturn effect for UK Corporate portfolios.

Adopting the supervisory add-on approach would reduce unjustified variability in risk weights, enable competent authorities to undertake their benchmarking activity sooner and use this to set the supervisory add-ons, and save time and money for both banks and competent authorities by not developing and maintaining the complex model component approach.
None of the participating members use the model component approach or would select it for use to estimate downturn LGDs.

Some members use estimation methods which are close to the model component approach for retail and mid-size corporate models.