Chapter 4: General estimation requirements

Principles for specifying the range of application of the rating systems

The segmentation principles aim to provide guidance on the highest level of rating system design. These principles are particularly relevant to changes in the range of application of certain rating systems, for example where an existing rating system is rolled out to an acquired portfolio, or a portfolio that is otherwise not yet treated under the IRB Approach. Among other things, the GL require in this regard the availability of fundamentally comparable credit-related information, meaning that, with respect to the obligors or exposures to which the rating system is extended, the relevant information has a similar nature and is available for the purpose of rating assignment, or is at least possible to obtain. For instance, as the information available for business clients and for natural persons is fundamentally different, these should not be covered by the same rating system.

A rating system is a broad concept which includes both PD and LGD models. Obligors and exposures covered by a rating system should be homogeneously managed, including in particular the use of common obligor and facility rating scales. Although a rating system may comprise more than one PD and LGD model, and although these models may comprise multiple calibration segments, the results of all calibrations within a rating system have to be reflected in the single obligor rating scale and a single facility rating scale applicable within the rating system. It is also crucial that the definition of default is used identically for the purpose of calibration of all models within a rating system. Data requirements

Good quality of data is a fundamental condition for developing a robust rating system. The data requirements in this general part apply to model development, risk quantification and application of all risk parameters and contain clarifications regarding the assessment of accuracy, completeness and appropriateness of data. More detailed requirements, especially regarding the scope of required data specific to PD or LGD estimation, and to LGD-in-default or EL BE estimation, are described within the relevant chapters of these GL.

Representativeness of data

Representativeness of data may influence the accuracy of the estimates; where the underlying historical data is less representative of the current portfolio the estimates may be less adequate. To ensure good performance of the models and their good predictive power, institutions should have adequate policies, processes and methods for assessing the representativeness of data used for the purpose of estimation of risk parameters, and they should pay particular attention to situations where data from different sources are used. While the methods of assessment may in some cases differ for data from different sources, especially because of the structure and availability of information, the applied standards should be the same for all data used in the estimation process.

The dimensions of representativeness specified in the GL include the scope of application of the model, definition of default, distribution of the relevant risk drivers, current and foreseeable economic and market conditions, and lending standards and recovery policies. The requirements for data representativeness are split into two sub-sections, namely requirements for data used in model development and for data used in calibration of risk parameters (i.e. for the data used to calculate the long-run average default rate and the long-run average LGD). This split is consistent with the structure of the CRR, where data requirements regarding model development are treated in sub-section 1 of Section 6 in Chapter 3 of Part Three, Title II of the CRR, and data requirements for risk quantification (calibration) are treated in sub-section 2 of the same section. This split has also been consistently followed in Chapters 7 and 8 of the RTS on IRB assessment methodology respectively.

The data requirements for risk differentiation and risk quantification differ both in terms of the methods for analysis and with regard to required actions where the analysis reveals insufficient data representativeness. In particular, for the purpose of model development, institutions may, subject to certain requirements, use a definition of default other than that specified in Article 178 of the CRR. However, they should closely analyse the ranges and distributions of key risk characteristics within the historical observations as compared with the current portfolio, because, for the purpose of model development, material observed differences in these characteristics should be avoided by selecting an appropriate sample of data, for instance by using a shorter observation period. On the other hand, for the purpose of calculating long-run average default rate and long-run average LGD, the definition of default has to strictly reflect the requirements of Article 178 of the CRR, whereas the comparability of the ranges of the key risk characteristics is necessary only to a required degree. Furthermore, requirements regarding the representativeness of current and foreseeable market and economic conditions are included only for data used in calibration, and are related to the requirements for estimating long-run average default rate for PD models, and long-run average LGD as well as downturn considerations for LGD. The steps that should be taken where significant non-representativeness of data is identified are also different in terms of model development and calibration. Whereas in model development non-representativeness is normally addressed by the adequate choice of a sample, the non-comparability of the historical data underlying risk quantification should not lead to any data exclusions, but should trigger an appropriate adjustment and increased MoC. The same principle should be followed where adjustments to the observed average default rate are necessary to meet the requirements laid down in Section 5.4. Thus data exclusions may be a tool to quantify adjustments to the observed average default and loss rates where these are necessary; however, they should never be taken into account for the purpose of calculating the observed average default rate.

The principles described above also apply to the specific case of the disposal of the portfolio of non-performing exposures and to the way these cases are reflected in the LGD estimation. This means that, while all defaulted observations, including those subject to the sale, have to be included in the calculation of long-run average LGD for the purpose of LGD calibration, institutions may decide not to include these observations in the sample used for the purpose of risk differentiation, for instance in the design of relevant pools. As specified in the GL, the LGD estimation methodology should be consistent with the collection and recovery policies adopted by the institution and adequate to the type of exposures to which LGDs are applied. One of the possible risk estimation methods includes specification of possible recovery scenarios and their probabilities, where the probability of each scenario would be a component of the model and would be determined taking into account relevant risk drivers, which may influence the frequency of use of a certain scenario. Another possibility would be to discard (some of) the default observations of sold credit obligations in the model development, whereas all default observations, including those related to sold credit obligations, should be included directly in the LGD calibration. In other words, where (some of) the observations subject to the sale would not be included in the phase of model development, they should provide the basis for the adequate adjustment in the calibration phase in order to reflect the calibration target, i.e. the long-run average LGD, which should always be based on all observed defaults.

Human judgement in estimation of risk parameters

Development of a robust rating system cannot be a purely statistical process, but to some extent also has to involve human judgement, to make sure that the models are appropriate for current and foreseeable portfolios and conditions, and that the models are acceptable for business users. Expert judgement may be necessary in particular with respect to the verification of model assumptions, and whether or not these are in line with economic expectations, to the design of the model, to the choice of risk drivers, etc. However, to ensure high quality of the models, expert judgement has to be appropriately documented and justified. This way, the judgemental elements of the model can be appropriately challenged and verified both by the validation function and by competent authorities. Therefore, this part of the GL clarifies the requirements regarding human judgement, including its documentation, in the estimation of risk parameters.

Treatment of deficiencies and margin of conservatism

When estimating risk parameters, institutions should identify any deficiencies that may lead to a bias in the quantification of risk parameters, or to increased uncertainty that is not fully captured by the general estimation error. These deficiencies may be related to data and methodology issues (defined in the GL as category A), or to the changes in relevant processes or external environment which may lead to additional uncertainty in the quantification of risk parameters (specified as category B). Within category B, institutions should also consider whether or not any changes in legal environment may lead to changes in default or loss rates, in particular changes to bankruptcy law and any regulations related to legal collection processes. However, this does not refer to changes in legal regulations on internal models or definition of default. Any data deficiencies related to the definition of default or any weaknesses in methods used in internal models should be classified under category A. Categories A and B are expected to be non-overlapping, i.e. each identified deficiency should be classified in only one of the categories.

The same rules should also be applied to any deficiencies related to the representativeness of data. The classification will therefore depend on the nature of the identified deficiency. Any missing data, including missing information on certain risk drivers in historical observations should be classified under category A. However, any changes in the distribution of risk characteristics in the application portfolio due to, for instance, changes in the underwriting standards, will be classified under category B.

As a general principle, institutions are required to address the identified deficiencies via appropriate adjustments and MoC. As an example, adjusting the data where deficiencies have been identified may involve rectifying the identified errors, for instance where missing data points are filled in with the most probable information, or the inaccuracies in data are corrected. Appropriate adjustment in the case of changes in the underwriting standards may take the form of an adequate shift of the estimated risk parameters. In any case, the objective of the appropriate adjustment is to achieve the most accurate estimates possible. To avoid excessive use of adjustments, institutions should be able to demonstrate that this objective has actually been achieved by a certain adjustment. They should also document any adjustments that have been applied, and regularly monitor their adequacy.

MoC should be applied on top of the best estimate of the risk parameter (i.e. a parameter after applying all appropriate adjustments). This also applies to situations where a predefined master scale is used; in this case, MoC is applied on top of the parameters defined in the master scale. While for internal purposes institutions may use best estimates of risk parameters if this is more appropriate, for own funds requirements calculation the final risk parameters, including MoC, should be taken into account.

MoC should be quantified for all deficiencies that could not be rectified by appropriate adjustment. However, even if the appropriate adjustment has been applied, the estimates are subject to additional uncertainty resulting from the application of the adjustments and its potential inaccuracy. Institutions should quantify MoC related to the identified deficiencies in the same categories used to classify deficiencies (categories A and B). In addition, they should quantify a general estimation error and present it in a separate category (category C). In future, this categorisation will help increase transparency with regard to the levels and underlying reasons for MoC.

The quantification of MoC related to categories A and B should reflect the additional uncertainty resulting either from the application of the adjustments or, where no adjustments are possible, from the uncertainty driven by the deficiencies in the relevant category. The quantification of the MoC for the general estimation error should reflect the dispersion of the distribution of the statistical estimator. The reference to statistical variance has been avoided in order to not impose a fixed methodology, which might for example lead to disproportionate MoC for low default portfolios.

The GL require that the final MoC to be added to the best estimate of the risk parameter be the sum of MoC for categories A, B and C. This proposal is based on an expectation that the categories will in general be non-overlapping. Nevertheless, the assumption of independence between MoC related to these categories would not be correct, as the appropriate adjustments which are made in relation to deficiencies stemming from categories A and B may influence the general estimation error. Therefore, for the purpose of harmonisation and in order not to impose over sophistication, it has been decided to require the aggregation of MoC between the categories based on a simple sum. However, different aggregation techniques may be used within each of the categories. Here, flexibility is left for the institutions to address in an adequate manner different types of deficiencies and their potential interrelations. All methods used for quantification and aggregation of MoC should be documented and regularly monitored.

Example of appropriate adjustment and margin of conservatism

Assume a change in regulatory requirements regarding the materiality threshold for detecting defaults triggered by the 90 days past due criterion. Assume the institution under consideration has stored the information regarding outstanding exposures monthly rather than daily. Thus the date of default and the exposure value cannot be retrieved historically. Even though some calculations could be made based on the monthly data, the 90 days may have been reached during a particular month and the default may not be visible based on monthly data. Therefore the institution has decided to set up a parallel default detection according to the new trigger, and to estimate MoC starting from the difference of the amount of defaults detected according to the old trigger and the new trigger.

An appropriate adjustment could be derived as follows: calculating the relative change in the number of defaults triggered according to the old and the new 90 days past due criterion compared with all defaults (i.e. taking into account all triggers) on a monthly basis. Thus an average correction factor can be estimated and applied retrospectively. As a result, the number of defaults according to the new 90 days past due trigger can be estimated for the available historical data.

The additional MoC could be derived, for example, from the 90% confidence interval around the average of the new default rates. However, this method is only an example, and institutions may use other methods for deriving MoC if these are deemed more appropriate.

1. Forbhreathnú

1.1 Tagairtí

1.2 Ceanglais aitheanta

1.3 Caighdeáin ghaolmhara

2. Ceanglais aitheanta

3. Caighdeáin ghaolmhara

• EBA - Published by the European Banking Authority at https://www.eba.europa.eu/