· The ongoing review of the continuing appropriateness of the models in use shall be conducted within the Pillar 2 framework The framework for supervisory cooperation should follow the steps outlined above.
3. Supervisor’s assessment of the application concerning the minimum requirements of the CRD – Credit Risk
3.1. Permanent Partial use and rollout
3.3.5. External vendor models
330. Annex VII, Part 4, Paragraph 36 of the CRD states that the use of a model obtained from a thirdparty vendor that claims proprietary technology is not a justification for exemption from documentation or any other of the requirements for rating systems. Thus these models generally have to fulfil the same requirements as models produced inhouse.
331. In particular, Annex VII, Part 4, Paragraph 31 of the CRD requires institutions to prove that all models used (including external models) have good predictive power, that the data used to build the model is representative of the actual portfolio of the institution, and that a regular cycle of model validation is in place. The burden is on the institution to satisfy competent authorities that it complies with these requirements. In other words, supervisors will not validate external vendors.
332. Under the IRB approach, institutions can use statistical models in the rating process and/or to estimate the risk parameters required for the approach. Although it should be emphasised that the rating process in total should be an internal rating process, it is not necessary that all parts of the process be developed internally. Most institutions use externally developed statistical models to some extent in their rating process.
333. In the context of these guidelines, an external vendor model is a model or parts of a model used by an institution and developed by an independent
external third party that uses certain inputs to assign exposures to certain rating grades or to estimate certain risk parameters.
334. In addition to the general requirements mentioned above, the transparency of the vendor model and of its linkage to the internal information used in the rating process will be examined closely by supervisors.
Transparency of the vendor model
335. To fulfil the CRD’s requirements regarding internal governance (Annex VII, Part 4, Paragraphs 123 to 126) and the requirements regarding the responsibilities of the Credit Risk control unit (Paragraphs 127 to 129) and the Internal Audit unit (Paragraph 130) for external vendor models, the institution has to prove that it understands the external model in all its aspects. The supervisor’s assessment of the institution’s use of external models will therefore place special emphasis on the inhouse knowledge concerning the development and the appropriate use of external vendors’ models.
336. This means, for example, that external vendors have to document the development and fundamentals of the validation process of their models in a way that permits third parties to gain a detailed understanding of the methodology applied and to assess whether the model is still performing adequately on their own customer bases. Moreover, the institution has to prove that the inhouse knowledge to do this is available. In particular, institutions should be aware of all the limitations of the model and the circumstances in which the model does not perform as expected.
337. The institution has to ensure that users will be adequately trained in the use of the model, and that inhouse instructors will be available. The institution also has to present plans for guaranteeing the validation and, if necessary, further development of the model in the future. Institutions have to ensure that the performance of the model can be assessed, and adjusted if necessary, even in the event that the vendor discontinues support, or similar events.
Linkage to internal information used in the rating process
338. The institution has to know what information (data) is processed in the external model and how this information is linked to information that is processed inhouse: for example, if the vendor’s definition of input factors such as sales and debt are consistent with those used inhouse. The institution has to make sure that the aggregation of the different parts of the model does not result in an inconsistent rating method, particularly in cases where parts of the model developed externally are used simultaneously with parts developed inhouse. It may also be necessary to check whether there is ’double counting’ of information in the internal and external parts of the rating model. Finally, the combination of separately developed parts of a rating system requires an extra riskquantification exercise: i.e., the risk parameters have to be estimated on an appropriate data set as in the case of a purely internally developed rating system.
An example:
339. Institution A estimates the rating class for corporate customers with a model that consists of two separately developed parts. Part A uses balance sheet data to estimate the probability of default with a statistical model purchased from an external vendor. Part B, using soft facts, was developed inhouse, and also uses statistical methods.
340. The institution has to prove that the data used to build the external model is representative of the actual portfolio. If the model is used only for assigning exposures (i.e., for ranking borrowers in relative terms) the proof of consistency with the CRD’s definition of default (Annex VII, Part 4, Paragraph 44) can be omitted.
341. If the institution also uses the model for PD estimation of rating classes (i.e., the PDs from model A are used as inputs in PD estimations for the entire model), the institution also has to prove that the default definition used in Part A is the same as the one used for Part B, and that both are consistent with the definition in the CRD. If this is not the case, the institution has to prove that it is able to map the default definition of Part A into Part B, and both definitions to the definition used in the CRD.
342. Furthermore, the institution has to be aware of how different information is processed in Part A. For validation purposes, the institution has to treat the combined outcome of the two parts of the model like the outcome of a single internally developed model. The institution should also compare the predictive power of the combined model with the predictive powers of the individual parts. In this way, the institution could determine if one of the two parts is no longer predictive, or if the combination of the two parts is the cause for a decline in predictive power.
3.4. Data
343. The CRD requires data to be held and stored for several different purposes. 344. Institutions’ physical databases need not be built to address each of these purposes separately, but may contain data relating to a mix of purposes. 345. There may also be different sources of data: · For creating the model for assignments: Internal, external, and pooled data might be used to determine the weight of the input variables. · For calibrating the model for estimates: Internal, external, and pooled
data (Annex VII, Part 4, Paragraphs 46, 66, 69, 71, 81 and 85 of the CRD) (external and pooled data with restrictions (Annex VII, Part 4, Paragraphs 57, 58, 64, 69, 111, 120, 129 and 130 of the CRD)).
· For outcome and performance data: Internal data generated during model development and use.
· For calculating the current minimum capital requirements: Internal data.