Response to discussion Paper on the future of the IRB Approach
Go back
We agree that the three pronged approach set out in the discussion paper will help improve the transparency of IRB models and their outcomes via developing technical standards and guidelines on key aspects of the IRB approach, monitoring of supervisory practices which in particular will include the use of benchmarking exercises and measures to enhance transparency around IRB models (with areas of priorities including the definition of default, LGD and conversion factor estimation, PD estimation, treatment of defaulted assets and Credit Risk Mitigation).
However as has been acknowledged by the EBA, although some of the non-risk based differences will be addressed by the above measures, legitimate reasons exist as to why the estimates of the drivers of risk for each exposure (PD, LGD, EAD and M) can be different across financial institutions including the inherent risk in the different asset classes, expert judgement, firms own historical loss experience, approach to managing and mitigating risk.
If it were possible to introduce the changes as part of a structured approach aligned to the annual validation and model rebuild timetable that firms maintain this would help mitigate the risk of untimely implementation.
Other than for CRM, where the possible changes appear to be relatively modest, the implementation timelines appear to be inadequate. For example if the default definition is changed, default histories would need to be re-stated to ensure that model calibrations are appropriate and in those cases a two year implementation window is unlikely to be sufficient.
If the definition of default, definition of an economic downturn and treatment of defaulted assets are modified, default or data histories would need to be refreshed, model parameters varied and/or calibrations updated across the model suite. In some cases, it might not be possible to re-state data.
With regard to PD estimation, the proposed requirements for calculating long-run default rates as basis for TTC PDs are defined by rating bucket/grade, hence would only be applicable to institutions that empirically calibrate their master scales to observed defaults. Where a bank’s modelling philosophy is fundamentally different, as allowed for under the CRR, and based on direct estimation of PDs without any bucketing. If enforced, such a change would call for revision and redevelopment of all PD models across the bank, hence the proposed timelines would be insufficient.
In respect to the treatment of defaulted assets, a harmonised definition of default would be challenging to apply to models based on external data for example. This could delay implementation considerably for those portfolios.
• The most substantive adjustment took place for the preparation of the first IRB applications during the implementation of Basel II in order to meet the inputs required. This was especially the case for the estimation of LGD which required a level of data that banks had not been previously been required to collate and thus was not generally available.
• Historical data (EDF time series) have been adjusted for some regions during a local downturn period to align the PD estimates with the observed default rates.
• External historical data (e.g. S&P CIQ) are periodically backward-adjusted by the data vendor affecting mainly PD models which have been calibrated to reflect the external data.
• Historical ratings data and list of defaulted counterparties are adjusted by Moody’s and S&P periodically, affecting models that incorporate such data.
The consultation paper refers to defaulted exposures that cure to non-default status and then return to defaulted status in a short period of time being treated as a constant default event from the first occasion. It would be helpful if clarity could be provided as to the meaning of a short period of time (does this mean one year, six months, 90 days, etc?). There is a potential risk of performing names being marked as defaults. Estimation of transition matrices also requires default to be a terminal state.
With regard to downturn adjustments of LGD and conversion factors, a downturn period could vary by industry sector or by region/geography. Any move towards consistency should recognise that a common downturn period might not be desirable. How the duration is defined is also of relevance (e.g. market indices from peak to trough).
Inclusion of qualitative factors or indicators in models is often questioned so guidance on the use of such factors would be welcome, along with the preferred approach for embedding qualitative factors.
Our experience of benchmarking is that whilst well intentioned, requires an extremely detailed and granular set of assumptions such that external disclosure without the full set of assumptions would be misleading and likely to have market impacts.
2. What would you consider the areas of priorities?
We consider that it is essential that the prioritisation of changes should be closely aligned to the Basel changes to avoid uneven global implementation.We agree that the three pronged approach set out in the discussion paper will help improve the transparency of IRB models and their outcomes via developing technical standards and guidelines on key aspects of the IRB approach, monitoring of supervisory practices which in particular will include the use of benchmarking exercises and measures to enhance transparency around IRB models (with areas of priorities including the definition of default, LGD and conversion factor estimation, PD estimation, treatment of defaulted assets and Credit Risk Mitigation).
However as has been acknowledged by the EBA, although some of the non-risk based differences will be addressed by the above measures, legitimate reasons exist as to why the estimates of the drivers of risk for each exposure (PD, LGD, EAD and M) can be different across financial institutions including the inherent risk in the different asset classes, expert judgement, firms own historical loss experience, approach to managing and mitigating risk.
3. Do you consider the proposed timeframe reasonable? In particular do you consider reasonable the proposed timeline for the implementation of the changes in the area of: a. definition of default; b. LGD and conversion factor estimation; c. PD estimation; d. treatment of defaulted assets; e. CRM?
In isolation the timelines are extremely challenging if the presumption is that all regulatory models will need to be rebuilt from first principles. However if the rules were flexible to allow for the use of high level post model adjustments to capture the essence of the change this would be achievable. However the challenge will be if the regulatory interpretation is such that a full and complete rebuild is required for which skilled resources are likely to be in extremely short supply.If it were possible to introduce the changes as part of a structured approach aligned to the annual validation and model rebuild timetable that firms maintain this would help mitigate the risk of untimely implementation.
Other than for CRM, where the possible changes appear to be relatively modest, the implementation timelines appear to be inadequate. For example if the default definition is changed, default histories would need to be re-stated to ensure that model calibrations are appropriate and in those cases a two year implementation window is unlikely to be sufficient.
If the definition of default, definition of an economic downturn and treatment of defaulted assets are modified, default or data histories would need to be refreshed, model parameters varied and/or calibrations updated across the model suite. In some cases, it might not be possible to re-state data.
With regard to PD estimation, the proposed requirements for calculating long-run default rates as basis for TTC PDs are defined by rating bucket/grade, hence would only be applicable to institutions that empirically calibrate their master scales to observed defaults. Where a bank’s modelling philosophy is fundamentally different, as allowed for under the CRR, and based on direct estimation of PDs without any bucketing. If enforced, such a change would call for revision and redevelopment of all PD models across the bank, hence the proposed timelines would be insufficient.
In respect to the treatment of defaulted assets, a harmonised definition of default would be challenging to apply to models based on external data for example. This could delay implementation considerably for those portfolios.
4. Are there any other aspects related with the application of the definition of default that should be clarified in the GL?
Where the definition of default is based upon the unlikeliness to pay indicators contained within CRR Art. 178.3 these will require a subjective assessment as they are prior to actual default. In these cases it will be difficult to achieve perfect harmonisation in the absence of overly prescriptive rules which are unlikely to ever be sufficient to cover all eventualities.5. Do you have experience with adjustments of historical data? What are the methods that you used to adjust historical data, including both internal and external data?
The following are examples of occasions where we have adjusted historical data:• The most substantive adjustment took place for the preparation of the first IRB applications during the implementation of Basel II in order to meet the inputs required. This was especially the case for the estimation of LGD which required a level of data that banks had not been previously been required to collate and thus was not generally available.
• Historical data (EDF time series) have been adjusted for some regions during a local downturn period to align the PD estimates with the observed default rates.
• External historical data (e.g. S&P CIQ) are periodically backward-adjusted by the data vendor affecting mainly PD models which have been calibrated to reflect the external data.
• Historical ratings data and list of defaulted counterparties are adjusted by Moody’s and S&P periodically, affecting models that incorporate such data.
6. To what extent is it possible to adjust your historical data to the proposed concept of materiality threshold for the purpose of calibration of risk estimates?
Adjusting retrospective data to fit a materiality threshold is complex and it may be difficult to create wholly reliable time-series of default data without a high margin of error. Incorporating unreliable data into models may lead to significant bias and will not improve RWA comparability. There would also be practical problems as business processes are aligned to the definition of default which were in force at the time. Assessing past performance on a definition that was not used at the time is unlikely to be a reliable indicator of future performances. We would therefore recommend that banks should be permitted to build up data on a go forward basis. Alternatively a trade off between the length of the historical period required and the reliability of the data would need to be reached.7. What is the expected materiality of the changes in your IRB models that will result from the proposed clarifications as described in section 4.3.2?
The expected materiality of the changes to IRB models is difficult to assess, pending fuller information about the proposals. This includes, among others, details on the nature, severity and duration of an economic downturn, details on PD computation (including guidance on how to combine different data sources and calibrate margins of conservatism), the identification of stressed years and how to cope with the absence of the time series of adequate stress conditions to capture a downturn, downturn adjustment of LGD and conversion factor estimates (including the treatment of incomplete workouts, treatment of collateral and guarantees, discounting factor, margin of conservatism and costs etc) and the treatment of multiple defaults.8. Do you consider the direction of the proposed changes adequate to address the weaknesses and divergences in the models across institutions?
We consider that differences in long-run default rates for similar groups of obligors are likely to converge across institutions based on the proposed changes. However, individual obligor PDs will depend less on the proposed changes and more on which factors are (or are not) used in the calibration of the model – e.g. use of different financial ratios or different calculation of financial inputs can produce large differences between institutions’ PD estimates.The consultation paper refers to defaulted exposures that cure to non-default status and then return to defaulted status in a short period of time being treated as a constant default event from the first occasion. It would be helpful if clarity could be provided as to the meaning of a short period of time (does this mean one year, six months, 90 days, etc?). There is a potential risk of performing names being marked as defaults. Estimation of transition matrices also requires default to be a terminal state.
With regard to downturn adjustments of LGD and conversion factors, a downturn period could vary by industry sector or by region/geography. Any move towards consistency should recognise that a common downturn period might not be desirable. How the duration is defined is also of relevance (e.g. market indices from peak to trough).
9. Are there any other aspects related with the estimation of risk parameters that should be clarified in the EBA guidelines?
Rules related to use of external data for PD and LGD/EAD calibration. In particular, the definition of default when an exact definition cannot be applied because it is not known how an obligor that was classified as a defaulted case by the data vendor actually defaulted, i.e. cases with unknown materiality threshold.Inclusion of qualitative factors or indicators in models is often questioned so guidance on the use of such factors would be welcome, along with the preferred approach for embedding qualitative factors.
10. Do you have dedicated LGD models for exposures in default that fulfil the requirements specified in section 4.3.4.(ii)?
For some portfolios where there is sufficient data to construct bespoke LGD models for defaulted assets, these have been applied. These are typically for Retail where sufficient default history exists. However in the majority of cases there is insufficient default data to develop statistical models and so “expert judgement” based methods are used to supplement the calculation of ELBE and LGD in default estimate.11. Do you consider the direction of the proposed changes adequate to address the weaknesses and divergences in the treatment of defaulted assets across institutions?
Yes. Our initial assessment suggests that the proposed changes will go a long way to eliminating firm specific differences in the treatment of defaulted assets.12. What else should be covered by the GL on the treatment of defaulted assets?
Guidance on the LGD approach for forbearance cases would be welcome.13. What are the impacts for the institutions that should be considered when specifying the conditions for PPU and roll-out?
A balance needs to be struck between the roll-out period given to banks and any quantitative RWA / exposure thresholds imposed. The time horizon to achieve roll-out should not be too short such that banks are unable to meet minimum risk management standards by ensuring appropriate quality of data, methodology and output. Additionally, it needs to be recognised that banks can only meet their time defined roll-out plans if the regulator is able to review models and provide feedback at a corresponding pace.14. Do you expect that your organisational structure and/or allocation of responsibilities will have to be changed as a result of the rules described in section 4.3.5?
We do not anticipate any direct organizational structural change would be required due to these proposals.15. Do you agree that CRM is a low priority area as regards the regulatory developments?
The proposed grouping and prioritisation of regulatory products is logical, however it should be recognized that any changes to the CRM framework also need to flow into LGD estimations. The omission of such guidance in the beginning of the process could lead to undesirable multiple changes to IT systems and LGD models to address LGD calibration and further revisions to LGD at a later date due to changes in CRM rules.16. Are there any other significant intra-EU or global discrepancies?
The main initiative that has the potential to diverge is the IFRS9 accounting rule changes that introduce the concept of significant deterioration, which is an earlier state to default as required under the regulatory rules. In addition, where exposures, either on an individual or collective basis, meet the significant deterioration criteria and assessment of lifetime ECL is required.17. Do you agree that the area of disclosures needs to be strengthened, in particular with regard to disclosures related with the benchmarking exercise, for instance by publishing them on the EBA website?
Fundamentally we think that disclosure is a key topic and that adequate disclosure will enhance transparency and reinforce internal models credibility. However, we disagree with the publication of benchmarking exercise results on the EBA website. Currently, the final version of the BCBS disclosures already include numerous disclosure templates for external stakeholders which would benefit from a greater harmonisation. Disclosure over and above the current requirements as this may present customer and firm confidentiality issues.Our experience of benchmarking is that whilst well intentioned, requires an extremely detailed and granular set of assumptions such that external disclosure without the full set of assumptions would be misleading and likely to have market impacts.