Loss Given Default Models
Contents
Definition
Loss Given Default Models (also LGD Models) are models and algorithms used for ex-ante (prior to a Default Event) loss given default estimation. A list of published LGD models is given in the Catalog of Loss Given Default Models. This entry describes the overall structure of such models
Model Usage
- For internal Credit Risk Management and Credit Portfolio Management
- As Risk Parameters for Basel III/CRD IV Regulation calculations of Regulatory Capital requirements
- For IFRS 9/ CECL purposes as inputs to Expected Credit Loss models
Pooled versus Individual Estimates
For retail portfolios it may be the case that the LGD models are estimated and used always at a portfolio level, with the implicit assumption that all credit exposures within the portfolio are homogeneous with respect to LGD risk factors. Collective (Portfolio-wide recovery rate) versus Individual facility models.
Regulatory Requirements
Risk Drivers
LGD Estimates must be based on the material drivers of risk. To comply with this requirement[1], [2], institutions should identify and analyse potential risk drivers under paragraphs 121-123 of the EBA GL on PD and LGD. When selecting the risk drivers, institutions should take into consideration any changes in product mix or characteristics between the reference and default dates.
Discriminatory Power
Institutions’ rating systems must provide for a meaningful assessment of obligor and transaction characteristics, a meaningful differentiation of risk and accurate and consistent quantitative estimates of risk. To comply with this requirement institutions should demonstrate that, in terms of the range of application of LGD models, the model performs adequately (in terms of discriminatory power and predictive power) on economically significant and material sub-ranges of application of the rating systems. The sub-ranges are identified by splitting the full range of application of the LGD model into different parts on the basis of potential drivers for risk differentiation, including the drivers referred to in paragraph 121 of the EBA GL on PD and LGD.
Number of LGD Classes
The number of grades and pools must be adequate for a meaningful risk differentiation and for the quantification of the LGD at the grade or pool level.To comply with this requirement, institutions should ensure the following:
- an adequate distribution of facilities across grades or pools in the datasets used for development and (initial and regular) validation. For this purpose:
- any unusually low number of facilities in a grade or pool is expected to be supported by empirical evidence of the adequacy of isolating those facilities in a specific grade or pool;
- any unusually high concentration of facilities in a grade or pool is expected to be supported by empirical evidence of homogeneity within these grades or pools (for example by analysing whether some potential risk drivers (e.g. exposure size) that could further differentiate between riskier and less risky facilities have not been considered).
- sufficient homogeneity of the risk within each grade or pool by providing empirical evidence that the grade-level LGD is adequate for all facilities in that grade. For this purpose, in cases where it is found (through the use of additional drivers or a different discretisation of the existing ones) that a material subset of facilities within a grade or pool yields a significantly different average realised LGD to that of the rest of the grade or pool, this is considered to indicate a lack of homogeneity.
- sufficient heterogeneity of the risk across grades or pools by providing empirical evidence that the average realised LGD is different across consecutive grades or pools, for subsets for which there is a meaningful order.
Continuous LGD Estimates
Where an institution uses direct estimates of risk parameters, these may be seen as estimates assigned to grades on a continuous rating scale. In this case, the same requirements apply when an institution uses direct estimates of risk parameters as apply to grade-based models. To comply with these requirements, institutions are expected to ensure risk differentiation in accordance with the following principles:
- If the LGD estimates used to calculated the risk-weighted exposure amounts are based on default weighted LRAs of realised LGDs for grades or pools, irrespective of whether at some point direct LGD estimates may have been used to define such grades or pools, this grade or pool level is the relevant one for the application of the principles set out in paragraph 104.
- When the situation described in point (a) above does not apply and, instead, several components are estimated separately and then combined in order to obtain the direct LGD estimates at facility level, institutions should provide empirical evidence that these components are independent. In the event of dependency, institutions are expected to adequately reflect this dependency in the models (for example using relevant risk drivers). The combination of components is expected to cover all possible losses relevant for the calculation of realised losses. For example, in cases where zero loss is assumed for some termination or stage during the recovery process, usually for cured processes or processes closed in the pre-litigation, this should be supported by empirical evidence.
- In the case of other direct LGD estimates (i.e. where no components are defined) the principles above are expected to be applied where relevant.
Facility Level Risk Differentiation
There is a risk that a meaningful differentiation of risk will not be achieved at facility level. To mitigate this risk, institutions should ensure that no bias is introduced in the risk differentiation when combining the different components in order to obtain the final LGD estimate at facility level. Specifically:
- the allocation of recovery flows to these components should be adequately documented and implemented in a consistent way;
- risk differentiation (analogous to risk quantification) should be ensured with respect to facility level.
Excessive Variability of RWA
In December 2017 the BCBS finalised the post-crisis Basel III reforms addressing the excessive variability of RWA. Certain LGD models have been identified as contributing to Model Risk uncertainty, which led to the introduction of LGD Input Floors[3]
