Response to consultation on Draft Guidelines on the methodology to estimate and apply credit conversion factors under the Capital Requirements Regulation
1. Question 1: How material are the cases for your institution where you would have to assign an SA-CCF to exposures arising from undrawn revolving commitments and thus restrict the use of own estimates of LGDs within the scope of application for IRB-CCF in the CRR3? For which cases would you not have enough data to estimate CCFs but have enough data to es-timate own estimates of LGDs?
There are instances of sub-portfolios with low materiality in which parameters cannot be estimated in a robust manner, which is why we appreciate the introduction of the minimum value and agree with the simplified quantification and newly introduced back-testing requirements. Otherwise, the quantification would be overly complex and unduly burdensome, particularly when considering the mandatory exclusion of such sub-portfolios from the LGD estimates that would lead to fragmentation of the model landscape. For our institution, this also corresponds to the current practice and approved models under applicable law, which is why there is a strong preference to keep the minimum fixed value of 100% in the current draft.
8. Question 8: Are there cases for your institution where the calibration samples should be shorter than the sample used to calculate the long run average (LRA) CCF
There are several instances, for which the calibration sample could be shorter than the sample used to calculate the LRA CCF. The most relevant ones for our institution are listed below.
- First, when a material driver of risk is available only from a specific point in time of the history, while target variables are available for a longer period, calibration on grade level is possible only with large uncertainty, since any imputation method introduces such uncertainty to the estimates. Therefore, this stipulation incentivizes institutions to base risk differentiation functions only on such drivers that exist for the full history and thus accept lower discriminatory power of the model for the sole purpose of avoiding such situations in the calibration.
- Second, when values of observed target variables are biased in the historical data due to significantly different early collection processes in different phases of the data history, the relationship between target variables and risk drivers is not representative over time. One extreme, yet common, example for this would be the introduction of blocking withdrawals from commitments based on the deterioration of the borrower’s creditworthiness: in the period before the introduction of such rule, the relationship between the realised CCF and the borrower’s creditworthiness is likely negative (i.e., higher CCF for lower creditworthiness), while after its introduction, the relationship would likely be positive (i.e., lower CCF for lower creditworthiness).
In both such instances, allowing for shortening of the calibration sample and introducing appropriate adjustments and MoC through the risk quantification sample would lead to models based on material drivers of risk, and increase their discriminatory and predictive power. Furthermore, in our understanding, the current regulatory framework for PD&LGD allows for shortening of the calibration sample as an appropriate adjustment to the estimates, which is why we strongly propose to allow for such possibility also for the purpose of the quantification of the CCF.
9. Question 9: Do you have any concerns with the requirements introduced to analyse and mitigate a lack of representativeness for CCF? Do the requirements on the different data samples when observing a lack of representativeness impede your ability to model CCF portfolios?
No structural concerns. However, since no such exists, introduction of a precise definition of representativeness (and lack thereof) would be beneficial to avoid misunderstandings of the relevant regulations. Furthermore, guidance on how to consider structural differences between the application portfolio and the samples would be beneficial, because we see deferring interpretations among different supervisors, in particular with respect to the treatment of instances where different distributions persist between the application portfolio and the development / calibration samples.
Specifically, where different grade distributions exist between the two samples, regulators may falsely conclude persisting underestimation of parameters or lack of representativeness of development & quantification samples solely based on average estimates being lower on the application portfolio compared to average realisations on the samples. We would appreciate a clarification that such differences are inherent to parameters estimated on defaulted portfolios and applied to performing ones, and outline guiding principles which point towards instances in which issues do persist that require remediation.
11. Question 11: Are there any concerns with requiring consistency in the analysis of changes in the product mix with the institution’s definition of facility? Are institutions able to identify and link contracts (partially) replacing other contracts where the closing or repayment of one contract is related to the origination of a new contract? Are institutions able to link new contracts that are originated after the reference date to related contracts existing at refer-ence date? In particular, is it possible in the case contracts that are revolving commitments are replaced by contracts that are non-revolving commitments (e.g. by a term loan)?
No concerns to note.
12. Question 12: Do institutions consider it proportionate to the risks of underestimation of CCF to perform the identification analysis and allocation procedure? If it is deemed not propor-tional, what would be an alternative approach that is still compliant with Article 182(1b) CRR?
No concerns to note.
13. Question 13: Do you have any concerns on the proposed approach for the treatment of so-called ‘fast defaults’? In case you already apply a 12-month fixed-horizon approach, do you apply a different treatment for ‘fast defaults’ in practice, (and if so, which one)? Is the ‘fast default’ phenomenon material according to your experience? If yes, for which exposures, exposure classes or types of facilities?
While we agree in principle with the 12-month approach and the treatment of fast defaults, please clarify in the GL which information is required to be stemming from precisely 12 months (i.e., 365 days) prior to the default, since this requires storage of daily data. In particular for the definition of risk drivers, we consider storage of all potential information on a daily basis unduly burdensome and likely not available historically, therefore, we propose to introduce the selection of an appropriate reference date for risk drivers (e.g., 12th month-end prior to default).
Fast defaults are a material phenomenon in the portfolio especially in the revolving portfolio, since they also include newly opened facilities within 12 months of the customer’s default.
14. Question 14: Do you have any concerns on the multiple default treatment? To what extent are your current models impacted by the application of a multiple default treatment?
No concerns. In fact, the multiple default treatment is vital in ensuring consistency between LGD and CCF data samples.
Some members of ESBG have indicated that, due to the mitigation of limited representativeness, it is often necessary to specify different concatenation periods for different parts of the history. It would be beneficial if the EBA highlights this possibility explicitly given the diverging supervisory messages on this topic in the past.
15. Question 15: Do you agree with the three principles for the calculation for realised CCF in the context of consumer product mix, and their implications for the cases mentioned as ex-amples? In case of disagreement, what is the materiality of the cases with unwarranted re-sults, in particular in relation with the definition of facility applied in your institution? In case of material unwarranted results, can you describe your alternative practice to this CP?
ESBG sees a significant misalignment with the core principle of capital adequacy as defined in Regulation (EU) 575/2013 (CRR), where capital held by an institution is to reflect the assets and off-balance sheet items currently on its books.
The proposed requirement expands this definition to also include future exposures, which adds a hypothetical element to the capital requirement that is not present in the CRR. Notably, in the case where an increase of a limit requires a renewed credit decision, the bank makes a credit assessment of the customer where the repayment capacity is evaluated prior to increasing the exposure on its books. Incorporating potential increased limits between the reference date and default date as described in the consultation paper would severely impact the CCF estimates and lead to CCF estimates taking future exposures, which banks have the possibility to reject, into consideration.
Hence, if there is a credit decision before the limit is increased, then the increase should not fall under the proposed requirement.
The effects of the requirement can be shown in a theoretical example where observed CCF values would be highly skewed due to an increase of the limit between the reference date and default date. The following example is based on the limits set out in case 1, example box 3 of the consultation paper. When the drawn amount at the reference date is altered, the effect on the realised CCF gets disproportionately large, ceteris paribus.
Case Limit at reference date On balance at reference date Limit at default date On balance at default date Realised CCF
1 100 50 150 140 180%
2 100 75 150 140 260%
3 100 95 150 140 900%
When it comes to the definition of facility, the bank agrees that there should be no skewing of estimates due to different definitions of facility. An alternative practise would be to allow the reference date to be modified to the last limit increase or incorporating an adjustment to the limit in the formula when a credit decision precedes a limit increase.
Moreover, in a similar question in EBA Q&A 2013_6871 regarding the risk parameter maturity (M), the EBA has clarified that risk parameters should not consider future exposures where the customer is not in a position to increase the limit without the bank's pre-approval through a credit decision.
1 2013_687 Maturity under Article 162 of Regulation (EU) No 575/2013 | European Banking Authority
16. Question 16: Are there any concerns related to the allocation mechanism described in these GL?
No concerns to note.
19. Question 19: Do you see any unwarranted consequences of the proposed approach for in-corporating additional drawings after default? In particular, in order to maintain consistency between the realised CCF calculation and the calculation of the denominator of the realised LGD as described in paragraph 140 of the GL PD and LGD, would this require a redevelop-ment of your LGD models?
It is unclear what the definition of the target variable for LGD estimation would be. Unwarranted effects may emerge in case of repeated payments & re-drawings.
20. Question 20: Do you think that the relative threshold is an appropriate approach to restrict the use of the alternative CCF approach for those facilities in the region of instability? Do you think it is appropriate to define a single relative threshold per rating system or are there circumstances where multiple relative thresholds would be warranted? Do you see a need to use an absolute threshold in addition to the relative thresholds?
Different thresholds should be allowed for different calibration segments. For example, for the case where a calibration segment consists of facilities with lower limits, a higher relative threshold may be necessary. Also, different product characteristics may necessitate different relative thresholds. We believe that absolute thresholds shall also be allowed, especially since the CCF becomes unstable when the denominator is small in absolute terms, not necessarily when it is compared to the limit.
21. Question 21: Do you consider the guidance sufficiently clear in relation to the requirement for institutions to set up a policy to define a threshold value?
Yes, it is sufficiently clear for this purpose.
22. Question 22: Do you consider it appropriate to set a prescribed level or range for the de-fined threshold, and if so, what would be an appropriate level for the threshold? In case an absolute threshold is warranted, what would be an appropriate prescribed level for an abso-lute threshold?
A pre-defined threshold may not be appropriate for all portfolios; therefore, we agree that banks shall set their own thresholds.
23. Question 23: Do you think that, for the facilities in the region of instability, and/or for fully drawn revolving commitments, a single approach should be prescribed (e.g. one of the ap-proaches above defined in the Basel III framework), or that more flexibility is necessary for institutions to use different approaches they deem most appropriate for these facilities?
Prescribing option a) as mandatory single approach would not be appropriate, due to the following reasons: CRR3 requires to replace negative realised CCFs with zero in the risk quantification phase. This essentially mandates to floor realised EADs at the value of drawn amount at the reference date for each individual default. Option a), however, does not enable to ensure consistency with such principle, neither for fully drawn cases, nor for the region of instability, as flooring the limit factor at the value of 1 would lead to inconsistent methods between the direct EAD estimation compared to EAD estimation through the CCF parameter. For option b), such consistency could be ensured both for the RoI and fully drawn cases through flooring at the value of 1, while for option c), flooring realisations at zero could be applied.
To summarise, a single approach would be acceptable in case option b) or option c) is chosen, otherwise flexibility should be provided to the institutions to use different approaches that they deem most appropriate).
25. Question 25: Which of the three approaches described in the Basel III framework is pre-ferred in case a single approach would be prescribed?
Option b) would be preferred, as it is already in use for our institution. However, option c) would be ok as well from the perspective of consistency in the risk quantification of the CCF (as described in the answer to Question 23).
29. Question 29: Do you have any comments on the modelling approach to estimate additional drawings after default for unresolved cases?
While the framework is self-consistent, it creates dependencies between LGD and CCF parameters, since changing the adjustment for unresolved cases in LGD leads to a re-estimation of CCF.
31. Question 31: For CCF estimation, do you use estimation methods that incorporate portfolio-level-calibration of the estimates? What are the main reasons to use a calibration at a level that is higher than the grade-level calibration?
Yes, it should be made possible due to the following reasons:
There are several instances, for which the calibration sample could be shorter than the sample used to calculate the LRA CCF. The most relevant ones for our institution are listed below.
- First, when a material driver of risk is available only from a specific point in time of the history, while target variables are available for a longer period, calibration on grade level is possible only with large uncertainty, since any imputation method introduces such uncertainty to the estimates. Therefore, this regulation incentivizes institutions to neglect that driver already for the purpose of model development and accept lower discriminatory power of the model to avoid such situations in the calibration
- Second, when values of observed target variables are biased in the historical data due to significantly different early collection processes in different phases of the data history, the relationship between target variables and risk drivers is not representative to the current by the historical data series. One example for this would be introduction of blocking withdrawals from commitments based on the deterioration of the borrower’s creditworthiness: in the period before the introduction of such rule, the relationship between the realised CCF and the borrower’s creditworthiness is likely negative (i.e., higher CCF for lower creditworthiness), while after its introduction, the relationship would likely be positive (i.e., lower CCF for lower creditworthiness).
In both such instances, allowing for portfolio-level calibration and introducing MoCs through the risk quantification sample would lead to models based on material drivers of risk, and increase their discriminatory and predictive power.
33. Question 33: Do you have any comments on the determination of the low share of ob-served additional drawings after default in the historical observation period relative to the observed undrawn amount at default date? Do you consider it appropriate to set a pre-scribed threshold to determine what constitutes this low share? If so, what would be an ap-propriate value for such a materiality threshold?
ESBG believes that is correct for institutions to define their own thresholds.
35. Question 35: Do you think the add-on of 15 percentage points is adequately calibrated when the downturn impact cannot be observed nor estimated? Could you provide clear examples or reasons why this add-on should be higher or lower than 15 percentage points?
CCF is not sensitive to the economic cycle: in many cases we observed very low (or even zero) DT add-ons. This is due to tightening of lending standards and early collection policies in downturn periods. Therefore, a lower value could be set for CCF.
37. Question 37: The possibility to have no downturn effect on CCF estimates is restricted to the case where observations are available during a downturn period. Which alternative meth-odologies could be used to prove the non-existence of a downturn effect on CCF estimates, in the case where no observation is available during a downturn period?
No.