EBA REPORT ON THE IRB MODELLING PRACTICES

20 November 2017

EBA Report on IRB modelling practices

Impact assessment for the GLs on PD, LGD and the treatment of
defaulted exposures based on the IRB survey results

1
EBA REPORT ON THE IRB MODELLING PRACTICES

Contents
List of figures 4
List of tables 7
Abbreviations 11
Executive summary 12
1. Background and rationale 18
2. Introduction 20
2.1 Sample of institutions and models 20
2.2 PD and LGD estimates 23
2.3 Coverage of IRB survey 25
2.4 Data quality 26
3. General estimation requirements 29
3.1 Principles for specifying the range of application of the rating systems 29
3.2 Data requirements 32
3.3 Margin of conservatism (MoC) 36
4. PD models 43
4.1 Characteristics of the survey sample 43
4.2 Data requirements for model development 46
4.3 Default rates and PD assignment at obligor or facility level (retail exposures) 47
4.4 Rating philosophy 52
4.5 Data requirements for observed DRs 55
4.6 Calculation of the one-year DR 57
4.7 Calculation of the observed average DR 60
4.8 Long-run average DR 64
4.9 Calibration to the long-run average DR 69
4.10 Summary of model changes in PD estimation 76
5. LGD models 79
5.1 Characteristics of the survey sample 79
5.2 Recoveries from collaterals 81
5.3 Eligibility of collaterals 87
5.4 Inclusion of collaterals in the LGD estimation 91
5.5 Calculation of economic loss and realised LGD 91
5.5.1 Definition of economic loss and realised LGD 92
5.5.2 Unpaid late fees and capitalised interest 95
5.5.3 Additional drawings 98

2
EBA REPORT ON THE IRB MODELLING PRACTICES

5.5.4 Discounting rate 99

5.5.5 Direct and indirect costs 108
5.6 Long-run average LGD 109
5.6.1 Historical observation period 109
5.6.2 Calculation of long-run average LGD 115
5.6.3 Treatment of incomplete recovery processes 116
5.6.4 Treatment of cases with no loss or positive outcome 123
5.7 Downturn adjustment 125
5.8 Summary of model changes in LGD estimation 130
6. Estimation of risk parameters for defaulted exposures 134
6.1 General requirements specific to LGD in-default and ELBE estimation and risk drivers 134
6.2 Reference dates 136
6.3 The requirement to reflect current economic circumstances in ELBE estimates 136
6.4 Relation of LGD in-default and ELBE to credit risk adjustments 138
6.5 Specific requirements for LGD in-default estimation 139
6.6 Summary of model changes in LGD in-default and ELBE estimation 140
7. Application of risk parameters 144
8. Review of estimates 146

3
EBA REPORT ON THE IRB MODELLING PRACTICES

List of figures
Figure 1: Number of banks participating in the IRB survey, by country 20
Figure 2: Share of PD models in the IRB survey sample, by country of origin 21
Figure 3: Share of LGD models in the IRB survey sample, by country of origin 21
Figure 4: Mean exposure value covered by PD and LGD models across exposure values (in
EUR millions) 26
Figure 5: Do you apply adjustments to the observed average of DRs for the purpose of PD
estimation? Retail, corporate, institutions, and central governments and central banks 34
Figure 6: How do you treat non-representativeness of data (LGD non-defaulted) 36
Figure 7: How is MoC included in your PD estimates? 37
Figure 8: Do you include an MoC in your LGD estimates? How? 38
Figure 9: Other triggers for using MoC in PD estimates 40
Figure 10: Other triggers for using MoC in LGD estimates 41
Figure 11: Types of PD models — retail, corporate, institutions, and central governments and
central banks 45
Figure 12: Number of living grades or pools (if discrete rating scale is used) 46
Figure 13: At what level does the institution recognise default? Retail exposures only 48
Figure 14: What type of records are considered in the one-year DR calculation? By (retail) COREP
exposure class 49
Figure 15: Level of PD assignment, by COREP exposure class 51
Figure 16: Are there any obligors who are in the scope of application that do not receive an
individual PD estimation? 57
Figure 17: Frequency at which one-year DRs are calculated, by COREP exposure class 61
Figure 18: Use of overlapping versus non-overlapping windows in calculation of observed average
DR 62
Figure 19: What method (simple average or weighted average) is used to determine the long-run
average DR? By COREP exposure class 63
Figure 20: Start and end date of the historical observation period by PD model (internal data,
retail mortgages — non-SME) 65
Figure 21: Start and end date of the historical observation period by PD model (internal data,
corporate — SME) 66
Figure 22: Start and end date of the historical observation period by PD model (internal data,
corporate — specialised lending) 66

4
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 23: If you apply adjustments to the observed average DR, what is the direction of the
adjustment? 67
Figure 24: Grade versus portfolio calibration 71
Figure 25: Do you conduct calibration before or after the application of MoC? 76
Figure 26: Do you conduct calibration before or after the application of the PD floor? 76
Figure 27: Do you repossess collateral in the course of the recovery process? Retail, corporate,
institutions, and central governments and central banks 82
Figure 28: Which recovery value is recognised in the calculation of the realised LGD? 83
Figure 29: Are certain types of collateral not taken into account in the LGD estimates? Retail,
corporate, institutions, and central governments and central banks 87
Figure 30: What are the reasons for not recognising certain types of collateral in the LGD
estimates? 89
Figure 31: How do you include in the LGD estimates protection in the form of guarantees and
credit derivatives? Retail, corporate, institutions, and central governments and central banks 90
Figure 32: How is economic loss of a cured case measured? 92
Figure 33: Methodologies used to determine the discounting rate (LGD non-defaulted) — retail,
corporate, institutions, and central governments and central banks 103
Figure 34: Methodologies used to determine the discounting rate (LGD in-default) 104
Figure 35: Methodologies used to determine the discounting rate (ELBE) 104
Figure 36: Level of granularity at which the discounting rate is specified (LGD non-defaulted) —
retail, corporate, institutions, and central governments and central banks 107
Figure 37: Level of granularity at which the discounting rate is specified (LGD in-default) 107
Figure 38: Level of granularity at which the discounting rate is specified (ELBE) 108
Figure 39: Are direct costs incurred before default included in the calculation of the realised (non-
defaulted) LGD? 109
Figure 40: Are indirect costs incurred before default included in the calculation of the realised
(non-defaulted) LGD? 109
Figure 41: Historical observation period for LGD non-defaulted — retail exposures secured by
immovable property SME (internal data) 111
Figure 42: Historical observation period for LGD non-defaulted — retail exposures secured by
immovable property non-SME (internal data) 112
Figure 43: Historical observation period for LGD non-defaulted — corporate exposures (SME,
specialised lending and corporate other) (internal data) 113
Figure 44: Historical observation period for LGD non-defaulted — retail other non-SME and
qualifying revolving (internal data) 113
Figure 45: Level at which the long-run average LGD is calculated 115

5
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 46: Type of weighting used in the calculation of the long-run average LGD — retail,
corporate, institutions, and central governments and central banks 116
Figure 47: Is a maximum period defined after which incomplete recovery processes are treated as
closed for the purpose of the average realised LGD? By COREP exposure class 120
Figure 48: Where a maximum period for the recovery process is specified, how is this defined? 121
Figure 49: Method used to calculate the long-run average LGD 123
Figure 50: How are cases with no loss or positive outcome treated? 124
Figure 51: How are data selected used in downturn estimation? 127
Figure 52: At which level is the downturn adjustment specified? 129
Figure 53: What is the main methodology used to determine LGD estimates that are appropriate
for an economic downturn? 130
Figure 54: What is your approach to the estimation of LGD in-default? 134
Figure 55: What approach is used for ELBE estimation? 135
Figure 56: Which economic conditions are reflected in ELBE estimates? 137
Figure 57: If you incorporate current economic conditions in ELBE, how are these incorporated?
138
Figure 58: Which economic conditions are reflected in LGD in-default? 140
Figure 59: Do you have a pre-established frequency for developing a (re)calibration of the PD
model? If yes, what is that frequency? 146
Figure 60: What is the frequency at which the observed average DRs are calculated? 147
Figure 61: Frequency for calculating the observed average DRs, by exposure class 148
Figure 62: Do you have a pre-established frequency for redeveloping or re-estimating the LGD
model? If yes, what is that frequency? 149

6
EBA REPORT ON THE IRB MODELLING PRACTICES

List of tables
Table 1: Number of PD and LGD models for which survey was completed..................................... 20
Table 2: Distribution of models across COREP exposure classes ..................................................... 22
Table 3: Minimum, median and maximum of the observed average DR and final PD estimate (%)
across COREP exposure classes ........................................................................................................ 23
Table 4: Magnitude of the absolute (in percentage points) and relative (%) differences between
the observed average DR and the final PD estimate ....................................................................... 24
Table 5: Average realised LGD, LGD non-defaulted, LGD in-default and ELBE across COREP exposure
classes (%) ........................................................................................................................................ 24
Table 6: Total number of PD and LGD models ................................................................................. 25
Table 7: Overlap between COREP exposure classes (PD) ................................................................ 30
Table 8: Overlap between COREP exposure classes (LGD non-defaulted) ...................................... 30
Table 9: Is the default definition used during model development for risk differentiation the same
as that defined by the CRR? By exposure class ................................................................................ 32
Table 10: What are the main reasons for applying adjustments to the observed average DR? ..... 34
Table 11: What are the main triggers for using an MoC in your LGD estimates?............................ 40
Table 12: Level of governance of PD models ................................................................................... 43
Table 13: Advanced or foundation IRB approach, by exposure class .............................................. 43
Table 14: Use of continuous or discrete rating scale ....................................................................... 45
Table 15: Number of living grades or pools (if a discrete rating scale is used) ................................ 45
Table 16: Length of the RDS used for model development for risk differentiation (in years), by
exposure class .................................................................................................................................. 47
Table 17: What type of records are considered in the one-year DR calculation? ........................... 50
Table 18: Level of PD assignment..................................................................................................... 50
Table 19: Average DR and PD estimate for different levels of PD assignment, retail exposures only
.......................................................................................................................................................... 51
Table 20: How would the rating assignment process capture changes in the economic conditions?
By COREP exposure class.................................................................................................................. 52
Table 21: Descriptions related to ranking method, calibration method or both............................. 54
Table 22: PIT-TTC description of the rating philosophy ................................................................... 55
Table 23: PIT-TTC description of the calibration philosophy ........................................................... 55

7
EBA REPORT ON THE IRB MODELLING PRACTICES

Table 24: What are the reasons for applying adjustments or data exclusions to overcome issues in
the calculation of the observed average DR? .................................................................................. 58
Table 25: Was any specific analysis undertaken to justify the choice of overlapping versus non-
overlapping windows for the calculation of the observed average DR? ......................................... 61
Table 26: Is there a significant share of short-term or terminated contracts within the period over
which the observed average DR is calculated? ................................................................................ 62
Table 27: Length of the historical observation period for PD estimation (internal data), by
exposure class .................................................................................................................................. 64
Table 28: Use of different calibration types, by COREP exposure class ........................................... 70
Table 29: Observed average DR and final PD estimate across calibration types (retail exposures
secured by immovable property) ..................................................................................................... 72
Table 30: If you use type 2 or 4 calibration, how many points in time were reflected in the
calibration sample? .......................................................................................................................... 73
Table 31: If you use type 1 or 3 calibration, which method do you apply when calculating the long-
run average PD per grade? By COREP exposure class...................................................................... 75
Table 32: Summary of selected policy choices for PD estimation and the number of model changes
.......................................................................................................................................................... 77
Table 33: Summary of number of aspects to be changed in PD estimation .................................... 78
Table 34: Types of LGD models used within the institutions ........................................................... 79
Table 35: Different types of LGD models for which the survey was completed .............................. 79
Table 36: Types of scales used in LGD and ELBE estimation ............................................................. 80
Table 37: Assignment of LGD or ELBE estimate to the whole exposure or only the (un)secured part
of the exposure ................................................................................................................................ 80
Table 38: Average realised LGD and final LGD estimate, depending on whether the LGD is assigned
to the secured part of the exposure, the unsecured part of the exposure or the whole exposure 81
Table 39: Model components used in the estimation of LGD non-defaulted ................................. 81
Table 40: Use of time in-default and recoveries realised so far as model components in estimation
of LGD in-default and ELBE ................................................................................................................ 81
Table 41: Which recovery value is recognised in the calculation of the realised LGD? By COREP
exposure class .................................................................................................................................. 86
Table 42: How is collateral included in the LGD estimation?........................................................... 91
Table 43: Treatment of unpaid late fees and capitalised interest in the calculation of realised LGD
.......................................................................................................................................................... 95
Table 44: Treatment of additional drawings after default in the calculation of realised LGD......... 98
Table 45: Are additional drawings after default included in the estimation of the CCF? ................ 99

8
EBA REPORT ON THE IRB MODELLING PRACTICES

Table 46: Are additional drawings after default included in the calculation of realised LGD (non-
defaulted)? ....................................................................................................................................... 99
Table 47: Average level of the discounting rate (%) in the RDS ..................................................... 100
Table 48: Summary statistics of the discounting rate, differentiated by chosen methodology .... 100
Table 49: Average level of the discounting rate (%), by COREP exposure class ............................ 105
Table 50: Average level of the discounting rate (%), for models with exposures only to corporates,
retail and non-retail ....................................................................................................................... 105
Table 51: Average level of the discounting rate (%), for models with exposures only to corporates,
retail and non-retail and where the discounting rate is specified as funding rate or risk-free rate
plus add-on ..................................................................................................................................... 105
Table 52: Length of the historical observation period (years) ....................................................... 109
Table 53: Length of the historical observation period for LGD non-defaulted, expressed in years
(internal data) and by exposure class ............................................................................................ 111
Table 54: Did you exclude some of the available historical data from the specification of the
historical observation period? ....................................................................................................... 114
Table 55: How are incomplete recovery processes incorporated into the LGD estimation? ........ 118
Table 56: Average time of the recovery process in the RDS (expressed in months) and average
share of incomplete recovery processes (calculated in terms of the number of defaulted
exposures) regarding all defaults occurring during the historical observation period (LGD non-
defaulted, internal data) ................................................................................................................ 121
Table 57: How is a downturn period defined? ............................................................................... 125
Table 58: Summary of selected policy choices for LGD (non-defaulted) estimation and the number
of model changes ........................................................................................................................... 131
Table 59: Summary of number of aspects to be changed in LGD estimation ................................ 133
Table 60: What is the reference date for estimation? ................................................................... 136
Table 61: Do you use the information on SCRA in the ELBE estimation?........................................ 139
Table 62: Do you use the information on SCRA in the LGD in-default estimation?....................... 140
Table 63: Summary of selected policy choices for LGD (in-default) estimation and the number of
model changes ............................................................................................................................... 141
Table 64: Summary of number of aspects to be changed in LGD in-default estimation ............... 142
Table 65: Summary of selected policy choices for ELBE estimation and the number of model
changes .......................................................................................................................................... 142
Table 66: Summary of number of aspects to be changed in ELBE estimation.................................143
Table 67: What are the main triggers for including additional conservatism in the application of
the PD model? ................................................................................................................................ 145

9
EBA REPORT ON THE IRB MODELLING PRACTICES

Table 68: Classification of answers from the survey with respect to the policy chosen in the GLs –
PD models ...................................................................................................................................... 150
Table 69: Classification of answers from the survey with respect to the policy chosen in the GLs –
LGD (non-defaulted) models .......................................................................................................... 150
Table 70: Classification of answers from the survey with respect to the policy chosen in the GLs –
LGD (in-default) models ................................................................................................................. 154
Table 71: Classification of answers from the survey with respect to the policy chosen in the GLs —
ELBE models ..................................................................................................................................... 155

10
EBA REPORT ON THE IRB MODELLING PRACTICES

Abbreviations

A-IRB Advanced internal ratings-based approach

BCBS Basel Committee on Banking Supervision

CCF credit conversion factor

CEBS Committee of European Banking Supervisors

CET1 Common Equity Tier 1

Common Reporting standards (Commission Implementing

Regulation (EU) No 680/2014 of 16 April 2014 laying down
implementing technical standards with regard to supervisory
COREP
reporting of institutions according to Regulation (EU)
No 575/2013 of the European Parliament and of the Council
(Text with EEA relevance)

CP consultation paper

CRR Capital Requirements Regulation

DR default rate

EAD exposure at default

ELBE expected loss best estimate

F-IRB Foundation internal ratings-based approach

GLs Guidelines

IRB internal ratings-based approach

LGD loss given default

MoC Margin of conservatism

PD probability of default

RDS reference dataset

11
EBA REPORT ON THE IRB MODELLING PRACTICES

RTS regulatory technical standards

RWA risk-weighted asset(s)

SA standardised approach

SCRA specific credit risk adjustment(s)

SME small and medium-sized enterprises

12
EBA REPORT ON THE IRB MODELLING PRACTICES

Executive summary
This report provides an overview of the modelling techniques used in the estimation of risk
parameters for both non-defaulted and defaulted exposures, i.e. PD, LGD non-defaulted, LGD in-
default and ELBE, and provides an impact assessment for the GLs on PD, LGD and the treatment of
defaulted exposures. The information on these modelling practices is based on the responses that
the EBA received on its survey on internal models (the IRB survey), which was conducted in the
context of the GLs on PD, LGD and defaulted assets. The responses reflect the modelling practices
at the time of completion of the survey, i.e. January 2017, and only the information on approved
models is included in this report.

In total, 102 institutions from 22 Member States participated in the IRB survey. The 102
institutions considered in the sample for the quantitative analysis account for 64% of EU
institutions’ total credit risk-weighted exposures. Those 102 institutions completed the survey for
a total of 252 PD models, and 95 of these institutions completed the survey for a total of 202 LGD
models. The median bank completed the survey for 3 PD and 2 LGD models. In relation to the
total number of PD and LGD models that institutions currently use, coverage of the PD and LGD
models in the IRB survey is 17% and 20% for the PD and LGD models respectively.

In line with the scope of the GLs, which apply to both high-default and low-default portfolios, the
survey (and this report) covers both portfolio types. More specifically, the models in the survey
cover all exposure classes, although some are better represented than others. The COREP
exposure class ‘retail — secured by immovable property non-SME)’ is the best represented:
around 50% of PD and LGD models apply to this COREP exposure class. In contrast, the share of
the low-default exposure classes is much lower: central governments and central banks (7% of PD
and 4% of LGD models), institutions (11% of PD and 8% of LGD models) and specialised lending
(3% of PD and LGD models).

Because the number of institutions and the number of models in the sample is not evenly
distributed across countries (in some countries the number of institutions participating in the
survey is much higher than in others), the results of this survey are summarised as the share of PD
or LGD models applying a specific practice, as well as the share of exposures covered under these
PD or LGD models. This presentation should also ensure that the exposure amounts covered by
these models (the sizes of the models) are reflected in the results.

It should be emphasised that the results of this survey are dependent on the quality of the
submitted responses, and are therefore subject to data quality issues, which are unavoidable in
any survey context. In particular, it can be seen from the comments of some respondents that
some questions have not been understood as intended. This caveat should be kept in mind when
drawing conclusions on the results and/or extrapolating from them.

13
EBA REPORT ON THE IRB MODELLING PRACTICES

Furthermore, it should be acknowledged that the survey did not cover, and this report does not
provide, a quantification of the potential impact of the GLs on capital requirements. Whereas
such an exercise (i.e. a quantitative impact study) has been considered, it would have required
substantially more resources from institutions to completely re-estimate (some of the) current
models to determine the capital effect of implementing them. In addition to this resource
requirement, the study’s results would have been subjective given the absence of supervisory
guidance. Therefore, the IRB survey (i.e. a qualitative assessment of current modelling practices)
has been chosen as a compromise solution that minimises the burden for institutions while
obtaining the best possible qualitative picture of surveyed institutions’ current practices. As a
result of this choice, this IRB survey provides an assessment of the number of model changes
necessary to comply with the GLs, and does not quantify the impact on capital requirements. The
quantitative capital impact of implementing the GLs will depend on the extent to which they
require institutions to re-estimate existing models in practice and the effect of those re-
estimations on individual capital requirements.

That being said, the distribution of current modelling practices for the firms and models surveyed
has been duly taken into account in deciding on the policy choices made in these GLs. For most
policy choices, the policy chosen in the GLs represents the most common approach observed. On
an aggregate basis, we expect the impact of the proposal to be neutral for the models surveyed,
as the specification of the GLs takes into account current practices for those models.
Furthermore, it would be impossible to predict the impact on capital requirements on the basis of
the responses to the IRB survey, because internal models feature many possible modelling
choices. As a result, the final impact of these GLs will be known only after a redevelopment and
recalibration of the models. This aspect supports the need for monitoring the impact of the
implementation of these GLs.

One area where the survey results provided additional evidence to justify the chosen policy is the
frequency of calculating the one-year default rate (DR). The CP on the GLs specified that
institutions should calculate one-year DRs at least quarterly. The other options that were
considered are (i) at least a monthly frequency for all retail exposures and at least a quarterly
frequency for all other exposures and (ii) at least a quarterly frequency for all retail exposures and
at least a semi-annual frequency for all other exposures. The survey responses, however, showed
that a frequency of at least quarterly is already applied in 45% of all PD models, whereas this
percentage is between 52% and 84% for the COREP retail exposure classes. Based on the fact that
a quarterly frequency or higher is already quite common, the final GLs also require that
institutions should evaluate the observed one-year DRs at least quarterly. This will entail a change
in practice for around 54% of PD models.

Furthermore, the results on the specification of the historical observation period for the purpose
of PD estimation showed a considerable heterogeneity of approaches, due to the variability of
one-year DRs, differences in the availability of DRs from good and bad years, and changes in the
economic, legal or business environment within the historical observation period. Although a
precise quantification of these differences is difficult in this area, the responses to the survey

14
EBA REPORT ON THE IRB MODELLING PRACTICES

confirmed the feedback to the CP with respect to the difficulty of assessing a historical
observation period in which bad years are over-represented.

Therefore, the GLs clarify that the long-run average DR should be calculated as the average of
observed one-year DRs if the historical observation period is representative of the likely range of
variability of one-year DRs. Whenever insufficient bad years are included in the historical
observation period, this average of observed one-year DRs should be adjusted upwards, whereas
it may be adjusted downward, under strict conditions, where bad years are over-represented in
the historical observation period. To limit possible variability stemming from the application of
this concept a benchmark is proposed, namely the maximum of the average of one-year DRs over
the most recent five years and the average of one-year DRs over the whole available observation
period. Institutions may still estimate long-run average DRs below this benchmark, but this should
be duly justified and trigger an additional margin of conservatism.

For PD estimation, the survey also provided supporting evidence that contributed to the chapter
on calibration. The survey contained a list of possible calibration methods, and respondents were
asked to indicate which method they use. These responses and the comments showed that
additional clarity on the various calibration methods is necessary, and this guidance has therefore
been included in the final GLs, in the form of a list of the calibration types that are allowed under
the CRR. In addition, a definition of the term ‘calibration’ is included to (i) clarify the distinction
from model development (calibration is the process that leads to appropriate risk quantification)
and (ii) highlight that calibration ensures that, for a calibration segment, PD estimates in a
calibration sample correspond to the long-run average DR at the level relevant for the applied
calibration method. Regardless of the chosen level of calibration, the objective is to obtain PDs at
grade level that are representative of the long-run average DR. Furthermore, these responses
made it possible to identify whether institutions apply a portfolio calibration or a calibration at
grade or pool level. Given the consequences such a decision may have for the cyclicality of capital
requirements, the GLs specify that institutions should provide additional calibration tests at the
level of the relevant calibration segment if calibration is performed at grade or pool level, or
perform additional calibration tests at the level of the grade or pool if calibration is performed at
portfolio level. To take account of these different practices with respect to the level of calibration
and to enhance understanding of its consequences, these GLs require institutions to assess the
potential effect of the chosen calibration method on the behaviour of PD estimates over time.

For LGD estimation, one of the areas where the survey provided useful guidance is the treatment
of economic loss for a cured case. The responses showed that the most common approach is to
assume that the economic loss for a cured case is zero, which, however, is not prudent.
Furthermore, the results showed that the approach proposed in the CP on the GLs (to apply the
same methodology as for other defaulted exposures without discounting additional recovery cash
flows) is applied in only around 4% of the LGD models, whereas the approach where such
additional recovery cash flows are discounted is applied in around 32% of LGD models. Based on a
review of the pros and cons of both approaches, i.e. discounting or not discounting the artificial
cash flows (i.e. the amount that was still outstanding at the moment of return to non-defaulted

15
EBA REPORT ON THE IRB MODELLING PRACTICES

status (principal, interest and/or fees)), it was decided to favour the discounting of these artificial
cash flows, hence to change the approach proposed in the CP.

For the treatment of unpaid late fees and capitalised interest, the survey revealed significant
variation in practices: in most models (52% for unpaid late fees and 44% for capitalised interest),
these are included in the economic loss only (numerator of the realised LGD), whereas they are
not included in 20% and 26% respectively, are added to both the nominator and denominator in
8% and 10% of models, and are added to the denominator only in 5% and 8% of models
respectively. Whereas the approach proposed in the CP on the GLs was the most commonly
applied based on the survey results, this approach was also criticised by industry respondents to
the CP, among others, because this approach would be overly conservative, and does not take
into account the fact that interest and fees are not related to real cash flow from banks and are
hence different from costs in that sense. After a review of alternative policy options and their
pros and cons, an approach was chosen that is operationally the easiest to implement: unpaid
late fees and capitalised interest after default should not increase the economic loss or amount
outstanding at the moment of default, i.e. only fees and interest before default should be
included. This approach does not require data on values of fees and interest capitalised after
default.

Regarding the inclusion of additional drawings in the realised LGD, the survey demonstrated that
the approach proposed in the CP was also the most commonly applied, and was retained in the
final GLs. In particular, the GLs specify that the treatment of additional drawings in the realised
LGD should be consistent with that treatment in the CCF estimation. Therefore, the GLs specify
that additional drawings should be included in the denominator of the realised LGD whenever
they are included in the CCF, and should not be included in the denominator whenever there are
not included in the CCF. The responses to the survey allowed the EBA to differentiate the
treatment of the additional drawings in the realised LGD, depending on their treatment in the
CCF, and the results confirmed that the above policy choice was also the approach that is
currently most commonly applied. Nevertheless, this policy choice will require 36% of LGD models
to be changed to comply with the GLs.

The discounting rate in LGD estimation has been identified as one of the major drivers of undue
risk weighted assets (RWA) variability across institutions. The survey shows that at the time it was
carried out (January 2017) an average discounting rate of 6% was used across LGD models, but it
confirms that practices are highly heterogeneous. In addition, the economic arguments that
indicate which approach is most correct from a theoretical perspective have also been taken into
account. Three main options have been considered: (i) the Euribor or a comparable interbank
rate plus add-on; (ii) funding cost plus add-on; and (iii) the original effective interest rate. The
results of the survey suggest that a risk-free rate plus add-on is applied most often, i.e. in 30% of
models and 37% of exposures covered, whereas the funding rate (with or without add-on) and
the effective interest rate (original or current) are used only in 19% and 22% of models
respectively. Based on these results as well as the pros and cons of these options, the GLs specify
that the discounting rate should be composed of a primary interbank offered rate plus a fixed
add-on. The level of the add-on has been fixed at 5% as proposed in the CP. Given the current

16
EBA REPORT ON THE IRB MODELLING PRACTICES

average level of the discounting rate identified for the models surveyed (6%) and the current low
interest rate environment, we expect that this approach would, across institutions, not cause
major cliff effects in LGD calculations.

Another area where the IRB survey provided relevant evidence for the finalisation of the GLs is
the treatment of incomplete recovery processes in LGD estimation. On this aspect, the CRR
specifies that all defaults within the data sources should be included in the LGD estimates, which
could be interpreted as referring to (i) including the information on all closed defaults; (ii)
including the information on all defaults, as well as those for which the recovery process is still
open; or (iii) including the information on closed defaults and an estimate of costs and recoveries
on exposures with incomplete recovery processes. The IRB survey responses made clear that the
third approach (which was also proposed in the CP) is the most common approach; it is used in
39% of LGD models and 44% of exposures covered by these models. While other arguments have
also been taken into account in this policy decision, the prevalence of this approach has
contributed to this decision. Although the chosen option represents the most common approach,
this policy will require a model change in 49% of LGD models and 40% of exposures covered.

For the estimation of LGD in-default and ELBE, the GLs clarify that all provisions applicable to LGD
(non-defaulted) also apply to LGD in-default and ELBE, unless otherwise specified. This approach
was chosen to minimise cliff effects as much as possible. Consequently, the policies described
above are also relevant for these estimates, although the shares of LGD in-default and ELBE
models in the IRB survey vary between questions. Two areas where the IRB survey provided
relevant input to the finalisation of the GLs for LGD in-default and ELBE are the estimation
approaches permitted and the approach to setting reference dates to be used for grouping
defaulted exposures in accordance with the recovery patterns observed.

For LGD in-default estimation, the GLs specify that, for the purpose of incorporating the
information on time in-default and recoveries realised so far, institutions may include this either
directly as a risk driver or indirectly, by setting the reference dates for estimation. From the
survey it is evident that 45% of LGD in-default models are similar to the LGD model for non-
defaulted exposures, and that only 11% of such models for LGD non-defaulted exposures are
enriched with additional risk drivers. In 25% of models, LGD in-default is estimated as
ELBE plus add-on. For the latter, it is hard to say whether or not these models will need to be
changed to comply with the GLs, since this depends on whether or not the add-on reflects the
additional unexpected loss during the recovery period.

For ELBE estimation, it is currently common (26% of models) to use accounting provisions as ELBE
estimates. Since the GLs specify that institutions should estimate ELBE based on an LGD model as
for non-defaulted exposures calibrated to current economic conditions and taking into account all
relevant post-default information, it will no longer be permitted to assess ELBE on the basis of
accounting provisions, unless these stem from a model that complies with the specified
conditions. Although it is not possible to assess accurately for all survey responses whether or not
a model change will be necessary, it is expected that around 63% of ELBE models will need to be
changed to comply with this policy choice.

17
EBA REPORT ON THE IRB MODELLING PRACTICES

Finally, the requirement to set discrete reference dates at which the realised LGDs should be
computed should ensure that parameters for defaulted exposures are appropriate for their
current status. To ensure the adequacy of the estimates, institutions should set the reference
dates according to the recovery pattern observed on a specific type of exposures, where such
reference dates may either be event based, e.g. linked with the realisation of collateral, or reflect
certain time periods during which exposures have been in-default. Given that this approach is
currently applied in only around 20% of LGD in-default and ELBE models, it is expected that a
significant share of these models will need to be changed to reflect this policy.

18
EBA REPORT ON THE IRB MODELLING PRACTICES

1. Background and rationale

1. This report provides an overview of the findings and the responses that the EBA received on
its survey on internal models (the IRB survey), which was conducted in the context of the GLs
on PD, LGD and defaulted assets 1, and presents an impact assessment for the major policy
choices made in these GLs. These GLs are published on the EBA’s own initiative to reduce
unjustified variability in RWA and as part of the broader review of the IRB approach that is
carried out by the EBA. This plan is outlined in the Report on the regulatory review of the IRB
approach published in February 2016 2.

2. Since these GLs are focused on the definitions and modelling techniques used in the
estimation of risk parameters for both non-defaulted and defaulted exposures, the IRB
survey and this report covers the modelling practices of institutions applying the IRB
approach. Related to these GLs is the EBA’s work for the draft regulatory technical standards
(RTS) on the specification of the nature, severity and duration of an economic downturn in
accordance with Articles 181(3)(a) and 182(4)(a) of Regulation (EU) No 575/2013 (the CRR).
Selected aspects of the survey and of this report cover the modelling practices that relate to
the specification of an economic downturn.

3. This report shows that these GLs and these RTS will have a significant impact on modelling
practices in some institutions. This report seeks to outline the current IRB modelling practices
based on the IRB survey, to help inform policymaking in the GLs and the RTS. It should be
mentioned, however, that the industry feedback that respondents provided to the CP on the
GLs was another source of information that has been used to revise the GLs. As a result, both
the evidence on current practices provided by this survey and the industry feedback and the
rationale for the various policy alternatives have driven the final policy decisions in the GLs.

4. In this context, this report includes a cost-benefit analysis for the key policy decisions, where
it is explained which options have been considered, and which pros and cons have been
taken into account in steering the final policy direction.

5. Overall, the results confirm a very diverse set of modelling practices, which justifies the
harmonisation that the GLs on PD, LGD and defaulted assets will bring, in order to reduce the
undue variability in RWA.

6. It should be emphasised that the results of this survey are dependent on the quality of the
submitted responses, and are therefore subject to data quality issues, which are unavoidable
in any survey context. In addition, it can be seen from the comments of some respondents

1
https://www.eba.europa.eu/regulation-and-policy/model-validation/guidelines-on-pd-lgd-estimation-and-treatment-
of-defaulted-assets
2
https://www.eba.europa.eu/-/eba-sets-out-roadmap-for-the-implementation-of-the-regulatory-review-of-internal-
models

19
EBA REPORT ON THE IRB MODELLING PRACTICES

that some questions have not been understood as intended. This caveat should be kept in
mind when drawing conclusions on the results and/or extrapolating from them.

7. Finally, it should be acknowledged that the survey did not cover, and this report does not
provide, a quantification of the potential impact of the GLs on capital requirements. Whereas
such an exercise (i.e. a quantitative impact study) has been considered, it would have
required substantially more resources from institutions to completely re-estimate (some of
the) current models to determine the capital effect of implementing them. In addition to this
resource requirement, the study’s results would have been subjective given the absence of
supervisory guidance. Therefore, the IRB survey (i.e. a qualitative assessment of current
modelling practices) has been chosen as a compromise solution that minimises the burden
for institutions while obtaining the best possible qualitative picture of surveyed institutions’
current practices. As a result of this choice, this IRB survey provides an assessment of the
number of model changes necessary to comply with the GLs, and does not quantify the
impact on capital requirements. The quantitative capital impact of implementing the GLs will
depend on the extent to which they require institutions to re-estimate existing models in
practice and the effect of those re-estimations on individual capital requirements.

20
EBA REPORT ON THE IRB MODELLING PRACTICES

2. Introduction

2.1 Sample of institutions and models

8. In total, 102 institutions from 22 Member States 3 submitted responses to the IRB survey; see
Figure 1 for an overview of banks participating by country. The responses reflect the
modelling practices at the time of completion of the survey, i.e. January 2017 4, and only the
information on approved models is included.
Figure 1: Number of banks participating in the IRB survey, by country

35
31
30
25
20
15
10 9
10 7
5 5 5
5 3 3 2 2 2 2 2 2 2 2 2 2 1 1 1 1
0

9. Those 102 institutions completed the survey for a total of 252 PD models, and 95 of these
institutions completed the survey for a total of 202 LGD models, as shown in Table 1. The
median bank completed the survey for 3 PD and 2 LGD models. In Figure 2 and Figure 3, the
total number of PD and LGD models from institutions across countries is reported.
Table 1: Number of PD and LGD models for which survey was completed

N mean min max p50 sum

PD 102 2.47 1 7 3 252
LGD 95 2.13 1 5 2 202

10.Taking into account the exposure values covered by these PD and LGD models, the share of
institutions from each country in the total sample looks quite different. Figure 2 shows, for
instance, that the share of PD models from German banks is 29%, whereas it is only 16% if
the exposure values covered by these PD models are taken into account. For France and the

3
Institutions from 22 EU Member States plus one institution from Norway constitute the sample.
4
The deadline for submitting responses was 31 January 2017.

21
EBA REPORT ON THE IRB MODELLING PRACTICES

United Kingdom (UK), the opposite pattern can be observed: whereas the share of PD models
from banks under French jurisdiction amounts to 7% (7% for the UK), these models account
for 17% (22% for the UK) of the exposure values covered. Figure 3, on the other hand, shows
the shares of LGD models by country (equally weighted and exposure weighted). The share
of PD models from UK institutions is the highest in the sample (22%), followed by France
(17%), Germany (16%), Sweden (13%), Spain (12%) and Italy (8%).

11.Similarly, the share of LGD models from UK institutions is the highest in the sample (22%),
followed by France (18%), Spain (13%), Germany (12%), Sweden (11%) and Italy (9%).
Figure 2: Share of PD models in the IRB survey sample, by country of origin

30%
25%
20%
15%
10%
5%
71 22 19 18 17 16 11 10 9 6 6 5 5 4 4 3 3 3 3 3 2 1 1
0%
France

Norway
Greece
Germany

Italy

Ireland
Poland

Croatia
United Kingdom

Czech Republic
Belgium
Denmark

Portugal

Estonia

Lithuania

Luxembourg
Sweden

Spain

Hungary

Netherlands

Latvia

Finland
Austria

Share of PD models Share ofexposures values covered by PD models

Note: the numbers within the figure refer to the number of PD models for all institutions within each country.

Figure 3: Share of LGD models in the IRB survey sample, by country of origin

30%
25%
20%
15%
10%
5%
48 19 18 17 17 15 9 8 7 6 6 4 3 3 3 2 2 2 2 2 1 1 1
0%
France
Germany

Norway

Greece
Poland

Ireland

Croatia
United Kingdom
Italy

Denmark
Czech Republic
Belgium

Lithuania
Portugal

Finland
Luxembourg
Sweden

Spain

Netherlands
Austria

Hungary

Estonia

Latvia

Share of LGD models Share ofexposures values covered by LGD models

Note: the numbers within the figure refer to the number of LGD models for all institutions within each country.

22
EBA REPORT ON THE IRB MODELLING PRACTICES

12.Table 2 provides an overview on the coverage of models across COREP exposure classes.
Throughout the report, whenever the row or column header mentions ‘%’, this refers to share
of PD or LGD models, whereas ‘% EAD’ refers to the share of exposures covered by these PD
or LGD models.

13.For the PD models, around 48% of models apply to the COREP exposure class ‘retail —
secured by immovable property non-SME)’. Around 28% of PD models apply to ‘corporate –
SME’ and ‘retail – other non-SME’ 5. For the LGD models, a distinction is made in the LGD
model between LGD non-defaulted, LGD in-default and ELBE. The exposure class ‘retail
secured by immovable property — non-SME’ is also well represented; more than 50% of LGD
models apply to this type of exposures. Table 2 shows not only the share of the models across
COREP exposure classes, but also the share of the exposure values covered by these models
(column with heading ‘% (EAD)’). Based on the exposure values, the share of the ‘retail
exposures secured by immovable property — non-SME’ is even higher than that based on the
count of the models. Overall, there appears to be a fair representation of the models used
across exposure classes.
Table 2: Distribution of models across COREP exposure classes

PD LGD non-defaulted LGD in-default ELBE

% % % %
No. % No. % No. % No. %
EAD EAD EAD EAD
Central
governments and 17 7 8 8 4 6 6 3 7 5 3 7
central banks
Institutions 27 11 14 16 8 9 12 7 11 11 7 10
Corporate — SME 70 28 23 54 27 25 49 28 30 44 28 28
Corporate —
8 3 8 7 3 5 8 5 10 7 4 8
specialised lending
Corporate — other 82 33 31 59 29 31 52 29 34 47 30 31
Retail — secured
by immovable 40 16 14 50 25 24 47 27 26 40 26 24
property SME
Retail — secured
by immovable 120 48 59 109 54 63 99 56 63 84 54 66
property non-SME
Retail — qualifying
27 11 9 30 15 14 23 13 13 22 14 15
revolving
Retail —other SME 41 16 15 49 24 20 40 23 19 36 23 20
Retail — other
71 28 21 69 34 21 60 34 22 54 35 20
non-SME
Total 251 201 177 156

14.The number of PD models in Table 2 (251) is lower than the total number mentioned in Table
1 (252) because the information in Table 2 contains only those PD models for which the
institution selected at least one of all COREP exposure classes. The same holds for the LGD
models (201 instead of 202). Similarly, the number of LGD in-default and ELBE models (177 and
5
Note that the percentages do not add up to one since this question was a ‘tick box’ question, where respondents
could select multiple COREP exposure classes that are covered by their model.

23
EBA REPORT ON THE IRB MODELLING PRACTICES

156) is lower than the number of LGD non-defaulted models, because several institutions only
completed the information on the LGD non-defaulted model.

2.2 PD and LGD estimates

15.For each PD model, the institutions have been asked to specify both the observed average DR
during the historical observation period and the final PD estimate corresponding to the PD
model at the chosen reporting date 6. These are visualised across COREP exposure classes in
Table 3 7. Institutions were asked to specify these values as the obligor-weighted average 8
across the PD model. The observed DRs and PD estimates shown in Table 3 are the equally
weighted average of those values across all PD models applicable to a certain COREP
exposure class. There is a wide variation in observed average DRs and PD estimates across
exposure classes and across institutions. This would not indicate any divergence in practices
per se, however, as these differences may stem from differences in the risk characteristics of
the underlying portfolios.

Table 3: Minimum, median and maximum of the observed average DR and final PD estimate (%) across COREP
exposure classes

Observed average DR Final PD estimate

% %
N % min max p50 N % min max p50
EAD EAD
Total 206 2.20 1.42 0.02 30.31 1.53 215 2.52 1.60 0.01 27.24 1.67
Central
governments
9 2.47 1.74 0.02 13.78 1.45 12 1.97 1.11 0.10 13.78 0.16
and central
banks
Institutions 15 1.15 1.13 0.06 3.23 0.88 19 1.17 1.40 0.11 2.82 1.08
Corporate —
54 2.58 2.01 0.45 6.79 2.41 55 2.77 2.11 0.38 15.60 2.42
SME
Corporate —
specialised 6 2.05 1.24 0.59 3.16 2.30 6 2.24 1.82 1.44 2.75 2.29
lending
Corporate —
63 2.20 1.53 0.06 6.79 1.97 65 2.74 1.99 0.10 15.60 2.37
other
Retail —
secured by
immovable 33 2.58 2.19 0.06 6.79 2.10 33 3.04 2.18 0.09 25.99 2.22
property
SME
Retail —
secured by
103 2.12 1.35 0.06 30.31 1.15 105 2.10 1.40 0.01 27.24 1.30
immovable
property

6
The reporting dates correspond to 30 June 2016 (as requested in the survey) for around 88% of PD and LGD models.
For a few models, the respondents chose to report this information for 31 December 2015 or 30 September 2016
(around 5-8% of PD and LGD models), 31 December 2016 or 31 January 2017 (around 1% of PD models).
7
Note that zero values have been excluded in this figure.
8
In particular, institutions were asked to specify the observed average DR and PD, weighted by obligor, obligor by
product type, facility or single exposure, depending on what kind of records the institution includes in its one-year DR.

24
EBA REPORT ON THE IRB MODELLING PRACTICES

Observed average DR Final PD estimate

non-SME
Retail —
qualifying 22 3.52 1.36 0.32 30.31 1.98 22 3.52 1.57 0.37 27.24 2.56
revolving
Retail —
31 2.77 2.31 0.32 6.79 2.26 31 3.38 2.43 0.24 25.99 2.53
other SME
Retail —
other non- 64 2.57 1.41 0.29 30.31 1.38 64 2.97 1.61 0.01 27.24 1.55
SME

16.Table 4 focuses on the absolute and relative difference between the observed DR and the
final PD estimate. Although on average the absolute difference between the observed
average DR and the final PD estimate is less than one percentage point, a wide variation is
observed across the PD models. Among those models where the final PD estimate is higher
than the observed DR, a difference as high as 19.69 percentage points is observed in the
sample. In addition, cases where the final PD is higher than the observed average DR (140)
occur more often than cases where the final PD is lower than the observed average DR (60) in
this sample of PD models.
Table 4: Magnitude of the absolute (in percentage points) and relative (%) differences between the observed average
DR and the final PD estimate

Exposur
Equally
e
weighte
N weighte min p10 p50 p90 max
d
d
average
average
PD > DR Absolute 140 0.98 0.40 0.01 0.05 0.31 1.61 19.69
Relative 135 78.79 84.05 0.25 4.76 29.81 218.90 1 211.69
PD < DR Absolute 60 0.88 0.40 0.01 0.06 0.34 2.12 9.20
Relative 60 25.89 21.63 0.63 2.27 20.37 55.71 98.25

17.In Table 5, a summary of the average realised LGD, LGD non-defaulted, LGD in-default and
ELBE is presented across COREP exposure classes. It should be noted that the highest average
LGD and ELBE values are for the exposure class ‘retail — qualifying revolving’. The lowest
average realised LGD and LGD non-defaulted is observed in the exposure class ‘retail —
secured by immovable property non-SME’ (values of 22% and 25% respectively). As expected,
the average LGD values are higher for LGD in-default than for the LGD non-defaulted,
although the difference is small for central governments and central banks.
Table 5: Average realised LGD, LGD non-defaulted, LGD in-default and ELBE across COREP exposure classes (%)

Average realised LGD LGD non-defaulted LGD in-default ELBE

% % % %
N % N % N % N %
EAD EAD EAD EAD
Total 177 29 26 194 33 27 151 42 35 115 43 36
Central 6 32 38 7 37 41 5 37 39 2 42 42

25
EBA REPORT ON THE IRB MODELLING PRACTICES

Average realised LGD LGD non-defaulted LGD in-default ELBE

governments
and central
banks
Institutions 13 28 29 16 31 33 10 47 48 7 43 54
Corporate —
53 33 31 53 36 34 42 50 50 32 46 49
SME
Corporate —
specialised 7 27 23 7 37 31 2 56 62 2 58 63
lending
Corporate —
54 32 36 57 35 37 44 47 49 35 45 49
other
Retail —
secured by
45 30 28 47 32 31 37 44 41 31 46 46
immovable
property SME
Retail —
secured by
immovable 94 22 19 105 25 21 85 34 27 64 35 29
property non-
SME
Retail —
qualifying 28 44 36 29 48 41 21 60 54 18 65 58
revolving
Retail — other
46 34 30 47 39 34 35 54 52 29 59 56
SME
Retail — other
67 38 31 68 43 36 54 55 50 41 59 51
non-SME
Note: the number of observations refers to the number of models for which the parameter estimate was provided. The
number of models for which the exposure-weighted average parameter estimate is calculated is not reported and is
slightly lower, since the exposure value was not provided for all models.

2.3 Coverage of IRB survey

18.Institutions were asked to indicate how many PD and LGD models they have within their
institution. With a total of 1 493 PD and 1 000 LGD models, the coverage of PD and LGD
models in the IRB survey is 17% and 20% respectively (see Table 6). This should be taken into
account when generalising the conclusions obtained from this survey.
Table 6: Total number of PD and LGD models

N sum mean min max p50

Total PD models within the institution 102 1 493 14.64 1 100 9.5
Total LGD models within the institution 97 1 000 10.31 1 91 3

19.The exposure amounts covered by these models are heterogeneous within COREP exposure
classes (Figure 4) 9. By exposure classes, specialised lending models seem to cover higher
exposure amounts, EUR 90 billion (PD) and EUR 70 billion (LGD) on average, although this
average is based on a smaller sample of models (6 PD and 5 LGD models). In comparison, the

9
Note that the exposure values covered by the models refer to the reporting dates, see footnote 6 for more
information.

26
EBA REPORT ON THE IRB MODELLING PRACTICES

average size of a model for retail exposures secured by immovable property (non-SME) is
around EUR 32 billion (PD) and EUR 38 billion (LGD).

Figure 4: Mean exposure value covered by PD and LGD models across exposure values (in EUR millions)

100000
90000
80000
70000
60000
50000
40000
30000
20000
10000 16 8 25 15 68 53 6 5 78 58 40 50 118107 25 29 40 49 70 70
0

Retail - Other non-SME

Central governments

Corporate - Specialised
Institutions

Corporate - Other

immovable property

immovable property

Retail - Qualifying
Corporate - SME

Retail - Other SME

Retail - secured by

Retail - secured by
and central banks

revolving
non-SME
lending

SME

PD LGD

Note: the numbers within the figure indicate the number of PD and LGD models within each COREP exposure class.

20.The 102 institutions considered in the sample for the quantitative analysis account for 64% of
the total EU institutions’ credit risk-weighted exposures 10.

2.4 Data quality

21.It should be noted that the results of this survey are dependent on the quality of the
submitted responses, and are therefore subject to data quality issues, which are unavoidable
in any survey context. This caveat should be kept in mind when drawing conclusions on the
results and/or extrapolating from them.

22.In general, the quality of the responses varies from question to question, i.e. even if the
results of one question are less satisfactory, this does not mean that the responses to other
questions are also of a lesser quality. Several of the data quality issues have been addressed
by data cleaning, meaning that some responses where either removed from the sample, or
were re-classified as different answers based on the comments. Data cleaning was not
possible for all questions, however, in particular for questions where it was not possible to re-
classify answers based on the comments. For other questions, while the response seemed

10
Obtained as the ratio of credit RWA from COREP reporting (adding up RWA in SA, F-IRB and A-IRB, i.e. IDs 77905,
82429 and 84560) to risk-weighted exposure amounts for credit risk in EU Member States (European Central Bank (ECB)
statistics on consolidated banking data, code CBD2.Q.B0.W0.11._Z._Z.A.A.ECR00._X.ALL.RW._Z.LE._T.EUR).

27
EBA REPORT ON THE IRB MODELLING PRACTICES

acceptable a priori, it can be seen from the comments of some respondents that some
questions have not been understood as intended.

23.Because not all respondents answered all questions for all models and because of the data
cleaning described above, it should be noted that the sample of models for each individual
question is smaller or equal to the total sample of models. In particular, whereas the total
number of PD and LGD models is respectively 252 and 202, this sample may be smaller for
some individual questions. Whenever the IRB survey contained a list of possible responses
where respondents could tick multiple answers, the sample consists only of those models for
which the question has been answered (i.e. models for which the particular question has
been left unanswered are omitted from the sample).

24.One area where the quality of the responses may affect the results in PD estimation is where
not all respondents calculate an observed average DR, i.e. institutions may only calculate the
long-run average DR, using various approaches. Therefore when the question is asked of
whether or not adjustments are applied to the observed average of DRs for the purpose of PD
estimation, the responses may not accurately reflect whether or not the long-run average DR
reflects the equally weighted average of all one-year DRs, or whether or not any adjustments
have been made during the process. This also applies to the question where respondents
have been asked about the reasons for applying adjustments to the observed average DR.

25.In addition, this list of reasons distinguished use of the term ‘representativeness’ for model
development (i.e. risk differentiation) from that of the term ‘comparability’ with respect to
risk quantification (consistent with the wording in Article 179(1)(d) of the CRR). The
comments made in response to this question, however, showed that not all respondents
shared this understanding, and that they used these terms interchangeably.

26.Furthermore, the instructions to the survey asked respondents to consider the notion of
‘margin of conservatism’ (MoC) to refer to the correction for the expected range of
estimation error, versus the notion of ‘appropriate adjustment’ to refer to the adjustment
that is made to correct the identified error. Some of the comments, however, showed that
this distinction was not used when responding to the questions.

27.In addition, the distinction in meaning between risk differentiation and risk quantification was
not shared among all respondents. This was made clear not only by the survey responses but
also by the feedback to the CP, and has led to a re-structuring of the GLs, where these terms
have been defined, and where all requirements for risk differentiation and risk quantification
have been clearly differentiated and clarified. However, this lack of understanding implied
that the questions on the length of the reference dataset (RDS) for risk differentiation and on
the length of the historical observation period (for risk quantification) were not clear for all
respondents, and that the relevant results should be treated with caution.

28.Finally, for PD estimation, it should be noted that the list of calibration types was not precise,
since the reference to Article 180(2)(a) of the CRR (for retail exposures) was missing for type 1

28
EBA REPORT ON THE IRB MODELLING PRACTICES

calibrations (i.e. only the reference to Article 180(1)(a) of the CRR was included, which refers
to all exposure classes except retail). However, the total list of calibration types was still
exhaustive, since respondents could select the category ‘other’.

29.For LGD estimation, the same caveat about the historical observation period applies as for PD
estimation, i.e. whereas the IRB survey enquired about the length of the historical
observation period (the length of the RDS was not requested), some respondents may have
understood this as referring to the length of the RDS.

30.As regards downturn LGD estimation, the IRB survey first asked how the downturn period is
defined, and, secondly, how the data to be used in downturn estimation are selected (see
paragraph 240). Whereas the quality of the responses to the first question was satisfactory,
the responses to the second question were quite confused, and nearly 50% of the
respondents selected the category ‘other’. Several of them provided a wide range of
explanations not answering the question. For this question, the final sample of models for
which the response was retained was much smaller (148 instead of 202 in total) due to the
extensive data cleaning.

29
EBA REPORT ON THE IRB MODELLING PRACTICES

3. General estimation requirements

3.1 Principles for specifying the range of application of the rating

systems
31.The GLs clarify (in paragraph 12) that a rating system should cover all those exposures where
the obligors or facilities show common drivers of risk and credit worthiness, and
fundamentally comparable availability of credit-related information.

32.Furthermore, the GLs include (in paragraph 8) a common definition of a PD and LGD model,
and in particular:

 all data and methods used as part of a rating system within the meaning of Article 142(1)
point (1) of Regulation (EU) No 575/2013, which relate to the differentiation and
quantification of own estimates of PD and which are used to assess the default risk for each
obligor or exposure covered by that model;

 all data and methods used as part of a rating system within the meaning of Article 142(1)
point (1) of Regulation (EU) No 575/2013, which relate to the differentiation and
quantification of own estimates of LGD, LGD in-default and ELBE and which are used to
assess the level of loss in the case of default for each facility covered by that model.

33.In relation to that requirement and these definitions, it is insightful to consider the sample of
PD and LGD models in this survey, with respect to the COREP exposure classes to which they
apply. Table 7 and Table 8 show the overlap between PD and LGD models across those COREP
exposure classes. Note that the percentages refer to the share of the number of models; for
instance, 41% of the models that apply to central governments and central banks also apply
to institutions. The total number of models in each exposure class is reported in the last
columns of Table 7 and Table 8; for instance, a total of eight models applies to central
governments and central banks. Institutions were asked to use the definitions of PD and LGD
models proposed above when completing the survey. As a result, the PD and LGD models may
encompass several methods for both risk differentiation and risk quantification. It can be seen
from Table 7 and Table 8 that there is considerable overlap between some COREP exposure
classes. Whereas such overlap may be more common among retail and corporate models, it is
more significant between retail and corporate exposures. This holds in particular for the share
of ‘corporate — SME’ models in to ‘retail secured by immovable property — SME’ and vice
versa (29% for PD and 39% for LGD models, and 50% and 42% vice versa), and between ‘retail
— other SME’ (39% for PD and 37% for LGD models) and ‘corporate — other’ (20% for PD and
31% for LGD).

30
EBA REPORT ON THE IRB MODELLING PRACTICES
Table 7: Overlap between COREP exposure classes (PD)

Retail —
Retail —
secured
Corporat secured
Central by Retail — Retail —
Corporat e— Corporat by Retail —
governments Institutio immovab qualifying other Total
e — SME specialise e— immovab other
and central ns (%) le revolving non-SME models
(%) d lending other (%) le SME (%)
banks (%) property (%) (%)
(%) property
non-SME
SME (%)
(%)
Central governments and central
100 41 18 6 35 0 6 0 6 6 17
banks
Institutions 26 100 41 11 59 4 0 0 4 0 27
Corporate — SME 4 16 100 7 86 29 13 7 29 10 70
Corporate — specialised lending 13 38 63 100 88 13 0 0% 13 0 8
Corporate — other 7 20 73 9 100 20 9 2 20 5 82
Retail — secured by immovable
0 3 50 3 40 100 58 23 70 40 40
property SME
Retail — secured by immovable
1 0 8 0 6 19 100 18 14 43 120
property non-SME
Retail — qualifying revolving 0 0 19 0 7 33 81 100 26 85 27
Retail — other SME 2 2 49 2 39 68 41 17 100 41 41
Retail — other non-SME 1 0% 10 0 6 23 73 32 24 100 71

Table 8: Overlap between COREP exposure classes (LGD non-defaulted)

Retail —
Retail —
secured
secured
Central Corporat by Retail —
Corporat Corporat by Retail — Retail —
governments Institutio e— immovab qualifying Total
e — SME e— immovab other other
and central ns (%) specialise le revolving models
(%) other (%) le SME (%) non-SME
banks (%) d lending property (%)
property
non-SME
SME (%)
(%)
Central governments and central
100 25 25 0 25 0 13 0 13 0 8
banks
Institutions 13 100 63 13 69 13 19 19 25 25 16
Corporate — SME 4 19 100 11 87 39 30 15 35 28 54
Corporate — specialised lending 0 29 86 100 86 14 0 0 14 0 7
Corporate — other 3 19 80 10 100 32 24 12 31 25 59
31
EBA REPORT ON THE IRB MODELLING PRACTICES

Retail — secured by immovable

0 4 42 2 38 100 80 36 66 58 50
property SME
Retail — secured by immovable
1 3 15 0 13 37 100 22 28 42 109
property non-SME
Retail — qualifying revolving 0 10 27 0 23 60 80 100 67 87 30
Retail — other SME 2 8 39 2 37 67 63 41 100 82 49
Retail — other non-SME 0 6 22 0 22 42 67 38 58 100 69

32
EBA REPORT ON THE IRB MODELLING PRACTICES

3.2 Data requirements

34.One of the most important aspects of the data requirements included in the GLs is to ensure
that the data used for the purpose of estimation of risk parameters, including model
development and risk quantification (calibration), is representative of the current portfolio
covered by the model under consideration.

35.From the responses to the IRB survey as well as the feedback to the GLs it emerged that
additional clarity is needed on the distinction between model development and calibration.
Definitions of these concepts have therefore been included in the GLs (in paragraph 8), where
it is specified that model development is the process that leads to risk differentiation, and
calibration is the process that leads to appropriate risk quantification. In this regard, the
requirements on representativeness are specified separately for risk differentiation and for
risk quantification in the final GLs. Whereas the requirements on data representativeness
were included in separate sections for PD and LGD in the CP on the GLs, they have been
merged in the final GLs and merged into the data requirements in the section on general
estimation requirements, i.e. they have been redrafted so as to apply to both PD and LGD
estimation.

36.As regards model development, the GLs (in paragraph 23) do not require that the definition
of default that is used for model development is identical to that used for the purpose of
Article 178 of the CRR. Instead, the GLs contain requirements to ensure consistency in the
definition of default during the observation period, and to ensure that the default definition
used for the purposes of model development does not have a negative impact on the
structure and performance of the rating model, in terms of risk differentiation and predictive
power. On the contrary, the GLs do require (in paragraph 30) the definition of default
underlying the data used for risk quantification to be consistent with the definition of default
specified in the CRR.

37.The IRB survey contained an explicit question on whether or not the definition of default
used for model development in PD estimation is the same as that in the CRR. The responses
indicate that a different definition of default was used in 18.18% of PD models (see Table 9).
Across exposure classes, the share of PD models that use a different definition of default is
higher for retail and sovereign exposures (above 20% on average) and lower for exposures to
institutions and corporates.
Table 9: Is the default definition used during model development for risk differentiation the same as that defined by
the CRR? By exposure class

No, the definition of default is Yes, the definition of default

Total
different from the CRR definition corresponds to the CRR definition
No. % % EAD No. % % EAD No.
Total 44 18 19 198 82 81 242
Central governments and
5 31 4 11 69 96 16
central banks
Institutions 3 12 8 22 88 92 25

33
EBA REPORT ON THE IRB MODELLING PRACTICES

No, the definition of default is Yes, the definition of default

Total
different from the CRR definition corresponds to the CRR definition
Corporate — SME 12 17 23 57 83 77 69
Corporate — specialised
0 0 0 7 100 100 7
lending
Corporate — other 11 14 19 69 86 81 80
Retail — secured by
10 25 40 30 75 60 40
immovable property SME
Retail — secured by
24 21 23 89 79 77 113
immovable property non-SME
Retail — qualifying revolving 7 26 13 20 74 87 27
Retail — other SME 8 20 33 33 80 67 41
Retail — other non-SME 19 28 27 50 72 73 69

38.The questions in the IRB survey for PD estimation did use the distinct wording of
representativeness with respect to model development and comparability with respect to
risk quantification. In particular, the term ‘comparability’ was used only with respect to risk
quantification, which is consistent with the wording in Article 179(1)(d) of the CRR 11 (as well
as with the RTS on assessment methodology Article 45(2)(a), referring to comparability to the
required degree). From the responses to the survey, however, it can be seen that not all
respondents had this same understanding, and that they may have understood the two
concepts interchangeably. This caveat should be kept in mind when interpreting the results
in Figure 5 and Table 10 12.

39.Figure 5 shows for how many models (%) and how many exposure values (% EAD) the
institutions apply adjustments to the observed average DR for the purpose of PD estimation.
From the responses it can be seen that adjustments are applied in around half of the models
(exposures). The shares of models (exposures) for which adjustments are applied is also
higher for retail models than for corporate models 13.

11
‘The population of exposures represented in the data used for estimation, the lending standards used when the data
was generated and other relevant characteristics shall be comparable with those of the institution's exposures and
standards. The economic or market conditions that underlie the data shall be relevant to current and foreseeable
conditions.’
12
This caveat applies also to Figure 23 on page 68 in section 4.8 on Long-run average DR.
13
This may be related to the requirement in Article 180(2)(e) of the CRR for retail exposures, i.e. ‘An institution need
not give equal importance to historic data if more recent data is a better predictor of loss rates’ (see also paragraph 95
and Figure 19).

34
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 5: Do you apply adjustments to the observed average of DRs for the purpose of PD estimation? Retail,
corporate, institutions, and central governments and central banks

% EAD 39% 61%

Retail

% 56% 44%
central Institutio Corporat

% EAD 57% 43%

e

% 61% 39%
% EAD 52% 48%
ns

% 42% 58%
governm
ents and

% EAD 54% 46%

Central

banks

% 53% 47%
% EAD 43% 57%
Total

% 55% 45%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

No Yes

40.Institutions were also asked to indicate the reasons for which they make adjustments to the
observed average DR 14. These results are shown in Table 10. It is noteworthy that the
category ‘other’ was selected in 55% of the models under consideration, and these other
triggers are therefore analysed in more detail below. The most common reasons for applying
adjustments to the observed average DR is to ensure representativeness of the observed
average DR for the long-run average (25% of PD models), to adjust for insufficient one-year
DR from economic downturn periods (14%) and because the observed average DR is not
representative of the default behaviour of the current portfolio (13%).
Table 10: What are the main reasons for applying adjustments to the observed average DR?

%
No. Total %
EAD
The observed average DR is not representative of the long-run average DR (is not
29 116 25 29
composed of an appropriate mix of good and bad years)
The observed average DR is not representative of the default behaviour of the
15 116 13 16
current portfolio
There are insufficient one-year DRs available from economic downturn periods 16 116 14 6
Non-comparability due to different structure of the portfolio in terms of risk
0 116 0 0
drivers
Non-comparability due to different lending standards 4 116 3 0
Non-comparability due to different legal environment 0 116 0 0
Non-comparability due to different definition of default 10 116 9 20
Non-comparability due to different market or economic conditions 2 116 2 0
Non-comparability due to modified scope of application of the model 4 116 3 1

14
Note that in Table 10 only those responses are shown where the institution indicated that they make adjustments to
the observed average DR (as shown in Figure 5).

35
EBA REPORT ON THE IRB MODELLING PRACTICES

%
No. Total %
EAD
Other 64 116 55 38
Note: this was a ‘tick box’ question, hence respondents could select several of the above answers.

41.Many of the reasons for making adjustments are unclassified (i.e. ‘other’ in more than half of
the PD models in Table 10). Based on the comments, however, it is possible to obtain more
insight into these other reasons. Caution is necessary in interpreting these results, because it
appeared that the distinction between appropriate adjustment and MoC has not been
properly understood by the respondents. In particular, the instructions for the survey
required respondents to consider the notions of MoC and ‘appropriate adjustment’ as
specified in the GLs (in paragraphs 38 and 41), i.e. MoC relates to the expected range of
estimation error, which is distinct from the appropriate adjustment 15, which in turn relates to
the adjustment made to correct the identified error. Nevertheless, it appeared that these
concepts had not been understood as intended, since several respondents mentioned that
additional conservatism was a reason for adjusting the observed average DR. The most
common other reasons for adjusting the observed average DR were additional conservatism
in general (more than one third of the other reasons), conservatism to account for
uncertainty in the data, and a need to address shortcomings of external data. Other reasons
were the low number of defaults in certain portfolios, lack of one-year DRs in certain time
periods, IT errors, and a lack of risk drivers during certain periods.

42.To provide clarity as to which situations and deficiencies should (at least) be corrected by
means of an appropriate adjustment, and for which an MoC related to the expected range of
estimation error should be added, the GLs clarify the meaning and intended use of these
concepts (i.e. ‘appropriate adjustment’ and ‘MoC’) (in section 4.4.2. and 4.4.3.).

43.When it comes to LGD estimation, the survey enquired about the various triggers for using
an MoC in the estimates, including in particular the various dimensions of representativeness
(changes in the definition of default, changes in lending standards and recovery policies,
changes in the current and foreseeable economic and market conditions, and other reasons
of (non)-representativeness).

44.Non-representativeness for risk quantification should lead not to data exclusion but to
appropriate adjustments, where possible, and additional MoC. This is particularly important
and is consistent with Article 181(1)(a) of the CRR, which requires the use of all observed
defaults for LGD estimation.

45.Figure 6 shows how institutions currently deal with issues related to data representativeness
for LGD (non-defaulted) estimation along several dimensions. It can be seen that for around
half of the models or exposures covered, institutions do not apply any treatment (i.e. no

15
Strictly speaking, the instructions for the survey as well as the CP on the GLs used the term ‘data adjustment’, which
has been renamed as ‘appropriate adjustment’ in the final GLs.

36
EBA REPORT ON THE IRB MODELLING PRACTICES

MoC, data adjustment or data exclusions). Depending on the origin of the non-
representativeness, MoC, data adjustments and data exclusions are among the most
common options. In particular, data adjustments and data exclusions are most often applied
in cases of non-representativeness due to a change in the scope of application of the model,
different structure of the portfolio in terms of risk drivers, and different lending or recovery
policies. For non-representativeness due to a different definition of default, it is most
common to apply data adjustments. For non-representativeness due to a different legal
environment, or different market or economic conditions, it is most common to apply an
MoC.

Figure 6: How do you treat non-representativeness of data (LGD non-defaulted)?

Different market or applicatio
Modified

% EAD 2% 63% 0%
2% 29% 3%
1%
Different scope of

recovery environm definition economic n of the

of default conditions model

% 8% 66% 0%
2% 11% 11% 2%

% EAD 19% 67% 3%3%

2%5%3%

% 24% 54% 6%2%

2%6% 6%

% EAD 7% 50% 1%
4% 27% 4% 8%

% 4% 62% 3%4% 16% 7% 4%

Different Different

% EAD 14% 65% 11% 4%

0%
5%2%
lending or legal

ent

% 5% 75% 5%3%
0%5% 7%

% EAD 8% 44% 6%3% 13% 14% 13%

policies

% 7% 54% 3%3% 8% 16% 9%

structure
Different

portfolio

% EAD 4% 49% 7%1% 9% 13% 17%

in terms

drivers
of risk
of the

% 5% 62% 7%1% 8% 8% 9%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Additional margin of conservatism

All data are used without any adjustments
Both data adjustments and margin of conservatism
Both data exclusions and margin of conservatism
Data adjustments
Data exclusions
Other

3.3 Margin of conservatism

37
EBA REPORT ON THE IRB MODELLING PRACTICES

46.The GLs require (in paragraph 41) institutions to implement a framework for the
quantification, documentation and monitoring of estimation errors, in relation to the
requirement in Article 179(1)(f) of the CRR that requires institutions to add an MoC, which in
turn is related to the expected range of estimation errors. As a general concept institutions
are required to address the identified deficiencies in data or methods via appropriate
adjustments and MoC. An appropriate adjustment consists of rectifying the identified errors;
for instance, missing data points are filled in with the most probable information, or the
inaccuracies in data are corrected. The objective of the appropriate adjustment is to achieve
the most accurate estimates possible. However, as the appropriate adjustment is performed
to estimate and correct the bias due to the data deficiency, additional MoC has to be added
to address the uncertainty related to this estimation. Moreover, MoC aims at addressing all
errors that cannot be rectified through appropriate adjustment and any other uncertainties
related to the estimation of risk parameters. Furthermore, the GLs (in paragraphs 43-52)
contain requirements on the framework for quantifying, documenting and monitoring MoC.
In particular, it is required that an MoC is quantified for each of three specified categories:
Category A, identified data and methodological deficiencies; Category B, relevant changes to
underwriting standards, risk appetite, collection and recovery policies, and any other source
of additional uncertainty; and Category C, general estimation error.

47.Figure 7 shows the variety of industry practices regarding the inclusion of MoC in the PD
estimates. In general, 20% of PD models do not contain an MoC. In more than half of the PD
models, an MoC is included in the calibration target (i.e. long-run average DR). In 10% of PD
models, an MoC is included only in the ranking models, and less than 4% of PD models
contain an MoC in both the ranking and the calibration part of the model.

Figure 7: How is MoC included in your PD estimates?

MoC is incorporated both in the

1% ranking model and in the
11%
calibration target (long-run)
10% 4% MoC is incorporated into the
calibration target (long-run)
14%
15%
No MoC

52% 53%
Other
19%
21%

Yes, MoC is incorporated into

the ranking model

38
EBA REPORT ON THE IRB MODELLING PRACTICES

Note: the inner circle shows the share of each option where all PD models are weighted equally, whereas the outer
circle shows the share of each option where all PD models are weighted by their corresponding exposure value.

48.Therefore, it is common practice to add MoC to the long-run average DR or the central
tendency (which is quite often specified based on the long-run average DR of the portfolio
under consideration).

49.Around 11% of the LGD and LGD in-default models do not contain an MoC (Figure 8), which is
about half the relevant number of PD models. However, in 33% of ELBE models, no MoC is
included. In more than half the LGD (non-defaulted and in-default) models, an MoC is
incorporated into the development process, and in 29% (LGD non-defaulted) and 26% (LGD
in-default), an MoC is added on top of the estimates.

Figure 8: Do you include an MoC in your LGD estimates? How?

LGD non-defaulted LGD in-default ELBE

9% 11% 12%
9% 9% 10% 11%
11% 10%
11% 39%
46%
51%50% 53%53% 33%
29% 26% 39%
31% 26%
10%
9%

The CRR does not contain any guidance on the quantification of MoC for certain triggers or in
general. Therefore, the following options were considered:
(a) non-exhaustive longlist of triggers for MoC as part of the GLs (including
recommendations for the according appropriate adjustments):
 pros: provides a more harmonised approach towards the triggers that require MoC
and the method to quantify the impact;
 cons: could lead to less suitable approaches for individual models;
(b) MoC categories with minimum list of triggers in the GLs:
 pros: provides a base for a more harmonised reporting on the level of MoC, but also
leaves room for distinct approaches;
 cons: different approaches could still lead to different levels of MoC.

39
EBA REPORT ON THE IRB MODELLING PRACTICES

50.In the GLs, option (b) has been chosen. However, the various categories of MoC have been
revised vis-à-vis those included in the CP on the GLs on the basis of the industry feedback, as
well as the responses to the IRB survey, which indicated that institutions use a wide range of
triggers for applying MoC, and signalled that sufficient clarity would be needed in the GLs to
clearly differentiate the various triggers from one another.

51.In the IRB survey, institutions were asked to indicate which aspects in model estimation
trigger the inclusion of an MoC. For each PD model, institutions could tick several boxes,
including (i) missing default trigger in historical observations; (ii) changed default trigger in
historical observations; (iii) changed underwriting standards (i.e. not reflected in the historical
observations); (iv) changed rating system (e.g. newly relevant risk driver or change in scope of
application); and (v) other. However, the majority of respondents did not make use of the
predefined answers (86% or 175 models) and selected the category ‘other’. This indicates the
importance of providing sufficient clarify with respect to the various categories of triggers for
MoC, as proposed in the GLs (in section 4.4.3).

52.On the basis of the comments, however, it was possible to obtain a more accurate picture of
the other sources of triggers for using an MoC in the PD estimates 16. The comments revealed,
as shown in Figure 9, that the most frequent other trigger is general estimation errors
(around 50% of the other triggers), and many respondents made explicit reference to the
legal requirement given in Article 179(1)(f) of the CRR. The second most common trigger for
MoC concerns data issues (i.e. lack of data and data quality issues), present in around 15% of
the other triggers. The third most mentioned other trigger for MoC in the survey is based on a
supervisory measure (9% of other triggers).

53.The anticipation of potential future events represents around 5% of the other triggers. In
these cases, institutions mention they that have observed negative trends in the recent years
and therefore apply conservatism to prepare for this potential downturn. Note that the
percentages are simple sums by triggers, not weighted by the amount of MoC.

16
Wherever the comment was not clear, the response for that model was discarded and omitted from the results
shown in the table.

40
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 9: Other triggers for using MoC in PD estimates

General estimation error

11% Lack of data
4% 11%
0%
3% Lack of Representativeness
9%
7% 1% Data quality problems
5% 45%
49% Anticipating future events
6%
11% Cyclicality
7%
Methodological shortcomings
17%
15% Supervisory measure
Other trigger

Note: the inner circle shows the share of each option where all PD models are weighted equally, whereas the outer
circle shows the share of each option where all PD models are weighted by their corresponding exposure value.

54.With respect to the triggers for using MoC in the LGD estimates, a variety of triggers can be
observed. However, in more than half of the LGD and ELBE models, data deficiencies and
missing data, as well as general estimation errors, caused the application of MoC in the
estimates. To a smaller extent, aspects related to the representativeness (due to changes in
the definition of default, changes in lending standards, and changes in the current and
foreseeable economic conditions) are a driver for using MoC in the estimates.

Table 11: What are the main triggers for using an MoC in your LGD estimates?

LGD non-defaulted LGD in-default ELBE

% % %
No. % No. % No. %
EAD EAD EAD
Data errors/deficiencies and missing data 96 52 56 88 56 59 58 53 58
Diminished representativeness due to
9 5 10 8 5 11 9 8 13
changes in the definition of default
Diminished representativeness due to
changes in the lending standards and 12 6 10 12 8 11 5 5 4
recovery policies
Diminished representativeness due to
changes in the current and foreseeable 32 17 14 28 18 15 14 13 10
economic or market conditions
Diminished representativeness due to
13 7 10 10 6 9 9 8 13
other reasons
General estimation errors including errors
stemming from methodological 99 54 61 88 56 61 57 52 59
deficiencies
Others 74 40 31 58 37 33 38 35 38
Note: this was a ‘tick box’ question, hence respondents could select several of the above answers.

41
EBA REPORT ON THE IRB MODELLING PRACTICES

55.It can be seen from Table 11 that many respondents selected the category ‘other’ to describe
the triggers for MoC that are used in LGD estimates (40% for LGD non-defaulted, 37% for LGD
in-default and 35% for ELBE). In the case of LGD non-defaulted, the comments have been
analysed to produce a more granular picture of the other sources of triggers for applying MoC
in LGD estimates. For 66 of the 74 models for which the option ‘other’ was selected, it was
possible to identify a more precise reason (shown in Figure 10). To some extent, the answers
show patterns similar to the corresponding answers relating to PD, although other LGD
specific triggers have also been identified. These include downturn implementation, haircut to
repossession valuation and uncertainty stemming from the discount rate, as shown in Figure
10.
Figure 10: Other triggers for using MoC in LGD estimates

21% 20% General estimation error

Lack of data
20%
24%
Lack of Representativeness
Data quality problems
Anticipating future events
7% 11%
11% Haircut to repossession valuation
11%
Uncertainty of discount rate
7% 6%
Other trigger
0% 12% 6%
Supervisory measures
2% 5% 5%
16% Downturn implementation
13%
2%2%

Note: the inner circle shows the share of each option where all LG models are weighted equally, whereas the outer
circle shows the share of each option where all LGD models are weighted by their corresponding exposure value.

56.The key finding from this analysis is that 24% of these other triggers of MoC stems from banks
that consider their downturn add-on as an MoC (16 answers). This points to the need to
provide clarity on the downturn concept in the draft RTS stemming from the mandate in
Article 181(1)(c) of the CRR. Apart from that, the general estimation error is the most
common answer (13 answers). Aligned with the corresponding analyses for PD, data issues
are another common trigger for using MoC in the LGD estimates. More granularly, lack of
data in general (low default portfolios, lack of representativeness (external or internal data),
and data quality problems have been mentioned by the respondents of the survey. Note that

42
EBA REPORT ON THE IRB MODELLING PRACTICES

larger deviations between the equally weighted and exposure weighted responses can be
observed, which is due to the smaller sample size in Figure 10 (66 models).

57.Institutions are required to quantify the final MoC to be added to the best estimate of the risk
parameter as the sum of the MoC for categories A, B and C. On this aspect, the following
options have been considered:

(a) the total MoC should be computed as the sum of the MoCs of the categories A, B and C,
and institutions may decide how to determine the MoC at the level of the individual category:

 pro: clear and uniform methodology is used by all institutions;

 pro: institutions retain discretion on how to aggregate the MoCs per category that stem
from different deficiencies;

 con: the sum of the MoCs of categories A, B and C is not mathematically correct if A, B
and C are not mutually exclusive (i.e. in case of overlaps between the categories);

(b) institutions should specify how the MoCs of categories A, B and C are aggregated to
determine the total MoC;

 pro: institutions may develop innovative methods to ensure that the total MoC is
computed in a mathematically correct way;

 con: given that the three categories are already relatively distinct, overlap between
them may be limited, and therefore institutions may sum up the individual MoCs in any
event;

 con: could lead to unnecessary complexity in modelling

58.Based on these pros and cons, option (a) has been chosen in the final GLs, given that it was
deemed that institutions are allowed sufficient flexibility to quantify the MoC at the level of
the individual category.

43
EBA REPORT ON THE IRB MODELLING PRACTICES

4. PD models

4.1 Characteristics of the survey sample

Table 12: Level of governance of PD models

All institutions Only consolidated institutions Only individual institutions

No. % % EAD No. % % EAD No. % % EAD
Central model 110 45 59 87 53 62 20 26 27
Central model
partly developed
17 7 4 9 5 4% 8 10 4
on an external
pool
Local model 84 34 26 49 30 27 35 45 30
Local model but
centrally 29 12 8 16 10 4 13 17 31
developed
Local model
partly developed
3 1 2 3 2 3 — — —
on an external
pool
Other 1 0 1 — — — 1 1 8
Total 244 100 100 164 100 100 77 100 100

59.Table 12 shows that there is a fair balance between central models (52%) and local models
(47%) in the sample. The results are also differentiated along the level of consolidation of the
institution, which further shows that consolidated institutions chose to report central models
(58%), whereas the prevalence of local models is higher among individual institutions (62%).
Table 13: Advanced or foundation IRB approach, by exposure class

Advanced IRB Foundation IRB Total

No. % % EAD No. % % EAD No.
Total 219 87 96 32 13 4 251
Central governments and central banks 11 65 85 6 35 15 17
Institutions 17 63 88 10 37 12 27
Corporate — SME 57 81 93 13 19 7 70
Corporate — specialised lending 6 75 98 2 25 2 8
Corporate — other 62 76 88 20 24 12 82
Retail — secured by immovable property SME 36 90 99 4 10 1 40
Retail — secured by immovable property non-SME 118 98 100 2 2 0 120
Retail — qualifying revolving 26 96 100 1 4 0 27
Retail — other SME 38 93 100 3 7 0 41
Retail — other non-SME 69 99 100 1 1 0 70
Other 5 71 74 2 29 26 7

44
EBA REPORT ON THE IRB MODELLING PRACTICES

60.As shown in Table 13, the majority of the PD models in the sample of the survey use the
advanced IRB approach (87% of PD models and 96% of exposures covered), where advanced
IRB refers to where the institution has received permission from the competent authority to
use both PD and own LGD estimates, and foundation IRB refers to where the institution has
received permission only to use own PD estimates. It is further interesting to analyse this
distribution across exposure classes, since the CRR specifies that the use of the advanced IRB
approach is mandatory for retail exposures (Article 151(7) of the CRR). However, for some of
the reported PD models that apply only to retail exposures (four models altogether), the
institutions indicated that they use the foundation IRB approach, which seems to contradict
this CRR requirement. For sovereign exposures, exposures to institutions and corporates, the
split between the advanced and foundation IRB approaches is generally around 70% and 30%
respectively. When expressed in terms of exposure shares covered by the models, 96% of
exposures fall under the advanced IRB approach. This is due to the relatively high weight of
retail models, where generally 100% of exposures are in the advanced IRB approach.

61.Figure 11 shows the distribution of types of PD models in the IRB survey, in total and across
groups of exposures, i.e. central governments and central banks, exposures to institutions,
corporates and retail exposures 17. The sample of PD models included in the IRB survey
consists mostly of PD models based on scorecards (based on expert judgement or
quantitative data), which account for almost 90% of all models. Scorecards based on
quantitative data are the best represented, accounting for 63% of all PD models, and 65% of
all exposures covered by PD models in the sample. For the retail exposure classes, such
scorecard models are even used exclusively.

62.For sovereign exposures, exposures to institutions and corporates, some of the PD models
make use of external ratings by means of a mapping scale, in accordance with
Article 180(1)(f) of the CRR (five PD models altogether). Simulation models are seldom used
(only two models, or 1% of all reported PD models), and in none of the retail models for
which the survey was completed is the PD derived from total losses (EL and LGD estimates).

17
Note that these groups of exposures are not mutually exclusive, because institutions could indicate to which of the
COREP exposure classes the PD model applies. Therefore, there are, for instance, PD models that apply both to
corporate and retail exposures, such that there is a partial overlap in the information shown in Figure 11.

45
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 11: Types of PD models — retail, corporate, institutions, and central governments and central banks

Retail % EAD 10%0% 90% 0%

% 15% 0%
2% 83% 0%
central Instituti Corpora

% EAD 59% 3% 15% 23% 0%

te

% 46% 2% 14% 35% 2%

% EAD 67% 2% 17% 14% 0%
Total banks ons

% 44% 7% 30% 19% 0%

% EAD
Central

74% 5% 13% 8%0%

govern
ments
and

% 25% 13% 25% 38% 0%

% EAD 29% 1%5% 65% 0%
% 26% 2% 8% 63% 1%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Scorecards based on expert judgement

Mapping of ECAI (eligible credit assessment institution) scale according to CRR 180(1)(f) CRR)
Other or any combination of the above
Scorecards based on quantitative data
Simulation model

63.As shown in Table 14, the majority of the PD models in the sample (82% of models and 73%
of exposures) make use of a discrete rating scale, whereas only 17% of models (25% of
exposures covered) have a continuous rating scale. Table 15 and Figure 12 show that a PD
model has on average 15 living grades or pools, but a wide variation can be observed,
ranging from 4 to 67.
Table 14: Use of continuous or discrete rating scale

N % % EAD
Continuous rating scale 42 17 25
Discrete rating scale 203 82 73
Other 4 2 2
Total 249 100 100

Table 15: Number of living grades or pools (if a discrete rating scale is used)

N mean (%) mean (% EAD) min p10 p50 p90 max

196 15.32 15.87 4 9 14 23 67

46
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 12: Number of living grades or pools (if discrete rating scale is used)
30
20
Percent
10
0

0 20 40 60

4.2 Data requirements for model development

64.To obtain insight into the characteristics of the RDS for model development, i.e. for risk
differentiation, institutions have been asked to indicate the length of the RDS for risk
differentiation. However, based on the responses and the comments of institutions, it
became clear that this question was understood in various ways by the respondents. Some
institutions understood the question as intended, i.e. as referring only to the dataset used for
risk differentiation, whereas others understood it as referring to the total length of the
dataset for both risk differentiation and quantification, and others as the length of the time
period used for model calibration. Consequently, even though it was mentioned in the
instructions that these questions refer to the RDS for the purpose of risk differentiation, not
all respondents had that understanding when responding to this question 18. These apparent
differences in understanding imply that the statistics of the length of the RDS should be
treated with caution.

65.From Table 16 it can be seen that the average length of the RDS (8.73 years) is slightly longer
than the average length of the historical observation period (shown in Table 27). The average
length of the RDS is longest for sovereign exposures (23 years), institutions (12 years) and

(18) This became clear from respondents who indicated that they use, for instance, five years of data to calculate scores
but 12 years of data to calculate the central tendency, whereas the response to the length of the RDS was 12 years.

47
EBA REPORT ON THE IRB MODELLING PRACTICES

specialised lending (13 years). The RDS seems to be shorter for retail exposures, and for
qualifying revolving exposures in particular, the RDS is shortest (5.33 years).

Table 16: Length of the RDS used for model development for risk differentiation (in years), by exposure class

mean
mean
N (% min p10 p50 p90 max
(%)
EAD)
Total 213 8.73 8.43 1.00 2.17 6.75 15.00 45.00
Central governments and central banks 13 23.04 11.66 1.00 8.75 26.00 41.00 41.00
Institutions 20 12.45 8.00 3.00 6.00 8.46 10.50 45.00
Corporate — SME 54 8.50 9.21 2.00 5.00 7.75 11.00 33.00
Corporate — specialised lending 7 12.79 9.24 3.00 6.00 10.50 15.00 35.00
Corporate — other 65 9.09 9.16 2.75 6.00 8.00 10.50 35.00
Retail — secured by immovable property
33 7.45 7.21 1.00 4.00 6.00 11.00 19.00
SME
Retail — secured by immovable property
101 7.00 8.20 1.00 3.00 5.50 10.00 23.00
non-SME
Retail — qualifying revolving 21 5.33 4.05 2.00 3.00 5.00 5.92 12.42
Retail — other SME 32 6.86 7.30 1.00 4.00 5.50 10.00 19.00
Retail — other non-SME 59 6.84 6.18 1.00 5.00 5.50 10.00 19.00

4.3 DRs and PD assignment at obligor or facility level (retail

exposures)

Records
Default Level of final
included in the
identification PD assignment
one-year DR

66.The CRR allows, in Article 178, second subparagraph, that institutions apply the definition of
default at the level of an individual credit facility rather than in relation to the total
obligations of a borrower. To ensure that the recognition of default at facility level does not
bias the PD estimates, the GLs (in paragraph 61) require that, where there is a significant
number of customers carrying multiple facilities of the same type within the retail rating
system under consideration, and the institution identifies defaults at the level of an
individual credit facility, institutions should analyse the level of risk of such customers
compared with customers carrying only one facility of the relevant type and, where
necessary, reflect the difference in the level of risk in the model through appropriate risk
drivers. Furthermore, institutions should ensure that the estimates based on facility-level
default identification are not biased due to cases of customers carrying multiple facilities.

48
EBA REPORT ON THE IRB MODELLING PRACTICES

67.This bias could stem from a different increase in the denominator of the one-year DR
compared with the increase in the numerator, where obligors with multiple facilities have a
different DR to obligors with one facility. In particular, the bias in the one-year DR would
depend on the average number of facilities per obligor.

68.To assess current practices in every step of the process, institutions have been asked to
indicate the level at which they recognise default, the type of records included in the one-
year DR calculation, and the level of the final PD assignments. Institutions could choose
between single-exposure level, facility level, obligor level, obligor level by product type and a
residual category, ‘other’. Single-exposure level refers to where one PD is assigned to each
contract. The option ‘obligor level’ refers to where one PD is assigned to each obligor,
irrespective of the product type, whereas the option ‘obligor level by product type’ refers, for
instance, to where one PD is assigned to two mortgages related to two properties of one
obligor, but where a credit card of this obligor receives a different PD. ‘Facility level’
corresponds to cases that do not refer to the single exposure, obligor or obligor by product
type; for instance, several mortgages with different durations of one obligor related to the
same property could be seen as one facility.
Figure 13: At what level does the institution recognise default? Retail exposures only

18% 27%
32%
27%

1%
6%
0%
5% 47%

36%

Facility level Obligor level Obligor level by product type Other Single exposure level

Note: the inner circle shows the share of each option where PD models are weighted equally, whereas the outer circle
shows the share of each option where PD models are weighted by their corresponding exposure value

69.Given that the requirement to allow the application of the definition of default in points (a)
and (b) of Article 178(1) of the CRR at the level of an individual credit facility is allowed only
for retail exposures, the responses in Figure 13 are shown only for retail exposures. It can be
seen from Figure 13 that default is recognised at obligor (or obligor level by product type) in
more than half of the PD models, whereas the PD models where default is recognised at a

49
EBA REPORT ON THE IRB MODELLING PRACTICES

lower level (facility or exposure level) represent 45.22% of all models. Where default is
recognised at lower level, institutions were also asked to indicate whether or not they
perform an analysis to compare the DR at that lower level with the DR at obligor level. From
the responses it is clear that such an analysis is exceptional; it is performed only in 7% of the
PD models (5 of 72). Therefore, the requirement (in paragraph 61) of the GLs will result in a
change in practice for institutions.
Figure 14: What type of records are considered in the one-year DR calculation? By (retail) COREP exposure class
property property Qualifying Retail - Other non-

% EAD 18% 12% 65% 5%0%

Retail -

SME

% 21% 17% 49% 13% 0%

non-SME revolving Other SME

% EAD 9% 14% 72% 5%0%

% 11% 8% 68% 8% 5%

% EAD 0%
1% 92% 6%0%
immovable immovable Retail -

% 7% 7% 74% 11% 0%
secured by secured by

% EAD 36% 29% 30% 5%0%

Retail -

% 31% 22% 39% 8%0%

% EAD 17% 0% 77% 6%0%

Retail -

SME

% 11% 0% 79% 5% 5%

% EAD 32% 28% 34% 5%0%

Total

% 28% 21% 42% 9% 1%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Single exposures Facilities Obligors Obligors by product type Other

70.The next step in the process concerns the type of records (single exposures, facilities,
obligors or obligors by product type) that are included in the one-year DR calculation. Figure
14 provides a picture very similar to that in Figure 13. Across exposure classes, the inclusion
of single exposures or facilities in the one-year DR seems to be less common in the SME
exposure classes (immovable property SME and other SME). On average, the share of models
(exposures) where default is recognised at single-exposure or facility level in SME portfolios
is 14% (20%), whereas it is 46% (48%) for non-SME portfolios. Table 17 replicates the
statistics shown in Figure 14 when considering only those retail exposures for which the
default is recognised at facility or at single-exposure level. It should be noted that, among
those PD models, the institution considers obligors in the one-year DR calculation in 2.82% of
models.

50
EBA REPORT ON THE IRB MODELLING PRACTICES

71.A similar conclusion can be obtained from Table 18, which shows that, of the 70 PD models
where default is recognised at facility or single-exposure level, the final PD is assigned at
obligor level (or obligor by product type) in only three models (representing only 4.29%). This
means that, whenever default is recognised at single-exposure or facility level, the institution
continues to include those records at a level lower than obligor level in the calculation of the
one-year DR and in the final PD estimate.
Table 17: What type of records are considered in the one-year DR calculation?

Retail exposures for

which default is
All retail exposures recognised at facility
or at single-exposure
level (see Figure 13)
No. % % EAD No. % % EAD
Facilities 33 21 28 30 42 46
Obligors 67 42 34 2 3 1
Obligors by product type 14 9 5 0 0 0
Single exposures 45 28 32 39 55 52
Other 2 1 0 0 0 0
Total 161 100 100 71 100 100

Table 18: Level of PD assignment

Retail exposures for which default is

All retail exposures recognised at facility or at single-
exposure level (see Figure 14)
No. % % EAD No. % % EAD
Facility level 32 19 30 29 41 52
Obligor level 73 44 34 2 3 1
Obligor level by product type 15 9 9 1 1 4
Single exposure level 44 27 27 38 54 43
Other 2 1 0 0 0 0
Total 166 100 100 70 100 100

72.Figure 15 shows the level of PD assignment across all exposure classes. Across all exposure
classes, PDs are most often assigned at obligor level and obligor level by product type (68% in
terms of PD models and 62% in terms of exposures covered). In the exposure classes
‘corporates’ and ‘institutions’, PDs are assigned only at obligor level. For retail exposures
secured by immovable property non-SME, the split between lower level (single exposure or
facility) and obligor level (or obligor level by product type) is quite balanced, whereas for the
other retail exposure classes, the PD assignment is most often at obligor level (or obligor
level by product type) (between 63% and 88% of PD models, depending on the exposure
class, and between 70% and 90% in terms of retail exposure values).

51
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 15: Level of PD assignment, by COREP exposure class

non-SME % EAD 19% 11% 65% 0%

5%
Retail -
Other

% 23% 14% 53% 10%0%

% EAD 8% 13% 73% 5%0%

immovabl immovabl Retail - Retail -
central Institution Corporate Specialise Corporate e property e property Qualifying Other
SME

% 10% 7% 71% 7% 5%
non-SME revolving

% EAD 0%
1% 93% 6%0%

% 7% 7% 81% 0%
4%

% EAD 30% 33% 29% 9% 0%

secured secured
Retail - Retail -

by

% 28% 23% 41% 8%0%

% EAD 7%0% 77% 16% 0%

SME
by

% 8%0% 80% 8% 5%

% EAD 0% 99% 1%
0%
- SME d lending - Other

% 0% 98% 2%
0%
Corporate

% EAD 0% 100% 0%
-

% 0% 88% 0% 13%

% EAD 0% 99% 1%
0%

% 0% 99% 1%
0%

% EAD 0% 98% 2%
0%
s

% 0% 93% 0%7%
governme

% EAD 2%
0% 98% 0%
nts and
Central

banks

% 6%0% 88% 0%6%

% EAD 18% 20% 56% 6%0%

Total

% 18% 13% 62% 6%2%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Single exposure level Facility level Obligor level Obligor level by product type Other

Table 19: Average DR and PD estimate for different levels of PD assignment, retail exposures only

Level of PD assignment
Obligor level by Single-exposure
Facility level Obligor level
product type level
% % % %
No. % No. % No. % No. %
EAD EAD EAD EAD
All retail DR 25 2.12 1.69 63 2.57 1.68 13 1.33 1.55 40 2.51 1.14
exposures PD 27 2.04 1.68 63 2.94 1.82 13 1.39 1.36 40 2.37 1.26

52
EBA REPORT ON THE IRB MODELLING PRACTICES

Level of PD assignment
Retail — DR 21 1.75 1.56 44 2.25 1.32 8 0.90 1.39 30 2.53 1.14
exposures
secured by
immovable PD 23 1.63 1.56 44 2.41 1.45 8 0.92 1.08 30 2.33 1.24
property non-
SME
Retail — DR 11 2.47 2.91 46 2.96 1.73 8 1.57 1.79 17 2.79 1.41
qualifying
revolving and PD 11 2.71 3.09 46 3.53 1.93 8 1.66 2.05 17 2.86 1.49
other non-SME

73.Table 19 shows the average observed DRs and average final PD estimates for retail exposures
along the different levels of PD assignment. It should be noted that that the average
observed DR of PD models where the PD is assigned at facility level (2.12) is lower than when
the PD is assigned at obligor level (2.57), but they are similar when comparing the PD
assignment at single-exposure level (2.51) with that at obligor level (2.57). In addition, the DR
at single-exposure level is higher than the DR at obligor level when only retail exposures
secured by non-SME are considered. It should be noted that figures are not at grades or
pools level but at the portfolio level, and therefore subject to portfolio composition. A
Kolmogorov-Schmirnov test for difference in distribution between the facility assignments
(facility or single-exposure level) and obligor level (obligor level or obligor level by product
type) did not 19 show a significant difference. Therefore, it is not possible to establish
statistical evidence that PD assignments at facility level lead to a bias vis-à-vis models where
PD estimates are assigned at obligor level.

4.4 Rating philosophy

74.To assess how possible RWA variability stemming from different rating and calibration
philosophies could be addressed, several questions were included in the survey.
Table 20: How would the rating assignment process capture changes in the economic conditions? By COREP exposure
class

Fully Highly Low Not Not

Neutral Total
sensitive sensitive sensitive sensitive known
Total 8 63 72 83 7 11 244
% 3 26 30 34 3 5 100
% EAD 7 28 28 28 6 4 100
Central governments and
1 5 3 7 1 0 17
central banks
% 6 29 18 41 6 0 100
% EAD 31 40 22 7 0 0 100
Institutions 1 6 7 8 3 1 26

19
The test was performed for all subsamples listed in the three columns of Table 19, i.e. for retail exposures, for retail
secured by immovable property non-SME and for retail exposures secured by qualifying revolving and other retail, but
none of the tests showed statistical significance.

53
EBA REPORT ON THE IRB MODELLING PRACTICES

Fully Highly Low Not Not

Neutral Total
sensitive sensitive sensitive sensitive known
% 4 23 27 31 12 4 100
% EAD 8 22 39 29 3 0 100
Corporate — SME 4 11 27 24 0 2 68
% 6 16 40 35 0 3 100
% EAD 7 14 39 40 0 0 100
Corporate — specialised
0 3 2 2 1 0 8
lending
% 0 38 25 25 13 0 100
% EAD 0 22 37 41 0 0 100
Corporate — other 4 19 26 28 1 2 80
% 5 24 33 35 1 3 100
% EAD 5 17 44 35 0 0 100
Retail — secured by
4 10 9 13 0 4 40
immovable property SME
% 10 25 23 33 0 10 100
% EAD 10 40 20 14 0 15 100
Retail — secured by
immovable property non- 5 30 35 37 3 6 116
SME
% 4 26 30 32 3 5 100
% EAD 5 32 21 27 9 6 100
Retail — qualifying
3 7 7 8 0 2 27
revolving
% 11 26 26 30 0 7 100
% EAD 15 56 11 14 0 4 100
Retail — other SME 4 10 10 14 0 3 41
% 10 24 24 34 0 7 100
% EAD 10 36 22 27 0 5 100
Retail — other non-SME 4 23 18 18 1 5 69
% 6 33 26 26 1 7 100
% EAD 7 45 23 13 6 5 100
Other 1 2 4 0 0 0 7
% 14 29 57 0 0 0 100
% EAD 40 26 33 0 0 0 100

75.Taking an overall look to the figures in Table 20, it can be seen that:

• for approximately 3% of the models, the rating assignment process is described as fully
sensitive to economic conditions;
• for approximately 26% of the models, the rating assignment process is described as highly
sensitive to economic conditions;
• for approximately 33% of the models, the rating assignment process is described as low
sensitive to economic conditions;

54
EBA REPORT ON THE IRB MODELLING PRACTICES

• for approximately 29% of the models, the rating assignment process is described as
neutral sensitive to economic conditions;
• for approximately 3% of the models, the rating assignment process is described as not
sensitive to economic conditions;
• for approximately 5% of the models, the rating assignment process sensitiveness to
economic conditions is not known.

76.According to the comments provided, the option ‘neutral’ from the drop-down menu has
mostly been understood as ‘medium sensitive’ or ‘somewhat sensitive’. This being the case,
these results indicate that, for over 70% of models, institutions consider that their models do
not exhibit a significant sensitivity towards the economic cycle.

77.Institutions have also been asked to provide a description of the rating philosophy of the
model under consideration in an open manner (i.e. no drop-down menu was provided). All
responses to this question have been thoroughly analysed, and re-classified, according to
whether or not the answer is referring to the philosophy of the ranking part of the model
(rating philosophy), to the philosophy of the calibration (calibration philosophy), or both. For
some answers, however, it was not possible to assess whether the description relates to the
ranking part or the calibration part.

78.First, it can be seen that many different understandings of what the rating philosophy is
coexist among the institutions. Many answers described the philosophy underlying either the
ranking part of the models (30%) or the calibrations (19%), whereas 51% of the answers
described the philosophy underlying both parts of the model as a whole (see Table 21). In
40% of the answers, it was not mentioned if they were referring to the ranking method, to
the calibration or to both. Therefore, it can be concluded that there is no common
interpretation of what the rating philosophy is.
Table 21: Descriptions related to ranking method, calibration method or both

No. %
Ranking method 45 30
Calibration method 29 19
Ranking + calibration method 76 51
N/a 102 —
Total 252 100

79.Secondly, for those cases in which either the ranking method or the ranking plus calibration
method is described, the answers are summarised in Table 22. It can be seen that the
majority of the philosophies underlying the ranking part of the PD models (rating philosophy)
have a predominant point in time (PIT) component (PIT + HYBRID PIT = 58%
20
approximately) .

20
Note that for the remaining 131 answers no description of the rating philosophy was provided; therefore, the
percentages have been computed with respect to 121.

55
EBA REPORT ON THE IRB MODELLING PRACTICES

Table 22: PIT-TTC description of the rating philosophy

HYBRID-
TTC HYBRID HYBRID-PIT PIT N/a
TTC
TOTAL No. % No. % No. % No. % No. % No. %

Ranking method 45 14 31 6 13 8 18 1 2 10 22 6 13

Ranking + calibration
76 2 3 0 0 10 13 5 7 54 71 5 7
method

Total 121 16 13 6 5 18 15 6 5 64 53 11 9

80.On the other hand, the majority of the methods underlying the calibration part of the PD
models have a higher TTC component (TTC + HYBRID TTC = 70% approximately) 21 (see results
in Table 23). It has to be highlighted that 47 answers shared a PIT rating philosophy and a TTC
calibration philosophy. Some of the institutions described this as a hybrid approach, which
again raises the need to clarify the notion of rating philosophy.
Table 23: PIT-TTC description of the calibration philosophy

TTC + FORECAS HYBRID- HYBRID-

TTC HYBRID PIT N/a
T TTC PIT
TOTA No No No No No No
No. % % % % % % %
L . . . . . .

Ranking + calibratio 7 1
76 0 0 57 4 5 1 1 0 0 8 6 8
n method 5 1

4 2 1
Calibration method 29 5 17 13 0 0 6 3 2 7 0 0
5 1 0
6 1
Total 105 5 5 70 4 4 7 7 3 3 10 6 6
7 0

4.5 Data requirements for observed DRs

81.To ensure that all obligors in the scope of application of a PD model are assigned a PD, the
GLs specify (in paragraphs 53-54) that each and every natural or legal person or IRB exposure
is rated by the institution, with the model approved to be used on and appropriate to the
single original obligor, including where there is unfunded credit protection as referred to in
Article 161(3) of the CRR 22. This requirement ensures that all obligors or exposures within the

21
Note that for most of the remaining 147 answers no description of the calibration philosophy was provided;
therefore, the percentages have been computed with respect to 105.
22
Related to that, the CP on the GLs specify (in paragraph 49) that the denominator of the one-year DR should refer to
all obligors assigned to a rating grade or pool at the beginning of the observation period, taking over-rides into account,
but excluding any substitution effects due to credit risk mitigation, as well as any ex post conservative adjustments. The
latter provision has been included, because it was deemed necessary to clarify in the GLs that obligors should be
included in the calculation of the one-year DR of the grade they are assigned to, before taking into account any
substitution effects due to credit risk mitigation. This means that the pool or grade assignment of the obligor should be
based on the obligor’s creditworthiness, and not that of the protection provider, since the latter would only protect
against potential losses in case of default of the obligor.

56
EBA REPORT ON THE IRB MODELLING PRACTICES

scope of application of the rating system have a rating in line with the applicable model, and
that potential existence of any guarantees does not change that requirement. The GLs
further specify in the section on data requirements for PD calibration (in paragraph 71) that
exclusion of observations from the DR calculation should be performed only (i) where an
obligor did not default in accordance with the definition of default specified in Article 178 of
the CRR, or (ii) where an obligor has been wrongly assigned to the considered rating model.

82.To enquire about current practices and obtain information on whether or not there are
justifications to exclude obligors from the rating assignment, a question was included in the
IRB survey on whether or not there are any obligors or exposures that are in the scope of
application but which do not receive an individual PD estimation.

83.From the responses shown in Figure 16 it can be seen that all obligors are assigned a rating in
around 80% of the PD models, whereas for around 10% of PD models it is mentioned that a
specific treatment is envisaged for obligors for which essential information from the obligor
is missing or ineligible. In most cases where the mentioned reason is missing or ineligible
essential information from the obligor, this refers to cases where balance sheet information
is missing, or to relatively new clients. Some respondents mentioned that certain obligors
with missing, insufficient or ineligible information are excluded from the calculation of the
one-year DR, but that those obligors are treated in the standardised approach (SA). Some
others mention that those obligors are assigned to the worst living grade of the discrete
rating scale, to the obligor grade of the protection provider (if any), or to a predefined rating
grade for unrated clients.

84.Some other respondents mentioned that a specific treatment is applied for obligors that
belong to a group of connected clients (in line with the requirement in Article 172(1)(d) of
the CRR), or where the obligor benefits from unfunded credit protection. Some respondents
mention that an obligor can only remain unrated in exceptional circumstances. Such obligors
include VIP customers, customers whose age is lower or equal to 18 years or above
100 years, inactive customers, customers under guardianship with or without removal of
legal capacity, staff of the institution (bank), obligors without a current account, residents
outside the euro zone, and legally incapacitated adults.

85.It should be noted that an individual assessment of these practices is needed, but that
several of the above practices are not compliant with the CRR and/or with the requirements
in the GLs.

57
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 16: Are there any obligors who are in the scope of application that do not receive an individual PD estimation?

No

8%

7% Yes, a specific treatment is

10%
applied for obligors which
2% 5% belong to a group of connected
2% 2% clients (Article 172(1)(d) CRR)
2%
Yes, but only obligors benefiting
from unfunded credit protection
are excluded

Yes, for other reasons

81%

80%
Yes, in case essential
information from the obligor is
missing or ineligible

Note: the inner circle shows the share of each option where all PD models are weighted equally, whereas the outer
circle shows the share of each option where all PD models are weighted by their corresponding exposure value.

4.6 Calculation of the one-year DR

86.The GLs contain (in paragraphs 73-77) specific guidance on which obligors should be included
in the numerator and denominator of the one-year DR. In particular, it is specified that
obligors who migrated to a different rating grade, pool or rating model, rating system, or
approach to calculation of capital requirements, should also be included in the denominator
and numerator. Obligors whose ratings are based on missing or partly missing information,
or ratings based on outdated information, should also be included.

87.Table 24 gives an overview on current practices with respect to the reasons for excluding
obligors or adjusting the observed average DR. It should be noted that the most common
reason for excluding observations relates to IT errors (mentioned in around 21% of PD
models), whereas the presence of short-term, new or terminated contracts (15% of PD
models) and the exclusion of obligors without payment obligation (15% of PD models), for
instance when no amount has been withdrawn) are also reported often. It should be
mentioned that for all the above reasons, institutions almost always reported that they

58
EBA REPORT ON THE IRB MODELLING PRACTICES

excluded observations from the sample, rather than applying adjustments, which was only
mentioned in few exceptional cases.
Table 24: What are the reasons for applying adjustments or data exclusions to overcome issues in the calculation of
the observed average DR?

N % % EAD
Short-term, new or terminated contracts 209 15 16
IT errors 209 21 15
Exposures without payment obligation 209 14 9
Diminished representativeness of older exposures or exposures
209 6 7
for which certain risk drivers are not available
Exposures that do not meet the definition of default or materiality
209 11 13
threshold at the beginning of the observation period are excluded
Loans for which no amount has been withdrawn or loans of non-
209 1 1
active clients are excluded
Migrated obligors or obligors of a connected client that is assigned
209 5 6
to a different rating model are excluded
Diverse other reasons 209 7 4
Note: the respondents could select several of the above answers.

88.Consequently, the aspects specified in the GLs (in paragraphs 73-7 and 80) will entail some
change in practice for institutions. In particular for the treatment of short-term, new or
terminated contracts, the GLs do not allow any data exclusions, but rather prescribe that the
institution should analyse the potential bias, choose between overlapping and non-
overlapping windows, and apply an economic adjustment and appropriate MoC if necessary.
The provision to also include migrated obligors will entail a change in policy for those 5% of
PD models.

89.Furthermore, a credit obligation refers to any amount of principal, interest and fees as well
as to any off-balance-sheet items including guarantees (as prescribed in paragraph 73(a) of
the GLs). Excluding obligors that have not withdrawn any amount (as was mentioned by
some respondents) is therefore not in line with the GLs.

90.The GLs prescribe (in paragraph 78) that institutions evaluate one-year DRs at least quarterly,
to monitor the appropriateness of PD estimates. On this aspect, the following policy options
have been considered:
(a) the one-year DR should be calculated at least at a monthly frequency for all retail
exposures, and at least quarterly for all other exposures:
 pro: ensures up-to-date information for the purpose of internal risk management, and
allows identification of changes in the trends in a timely manner;
 con: low-default portfolios probably contain no new information;
 con: could be overly burdensome;
(b) the one-year DR should be evaluated at least quarterly for all exposures:

59
EBA REPORT ON THE IRB MODELLING PRACTICES

 pro: less burdensome but at the same time ensures minimum frequency of
monitoring, allowing identification of any seasonal effects;
 con: obligors that default in less than three months after credit origination could be
omitted in the DR;
(c) the one-year DR should be calculated at least quarterly for all retail exposures, and at
least semi-annually for all other exposures:
 con: obligors that default in less than half a year (for non-retail) after credit
origination could be omitted in the DR.

91.Option (b) has been chosen in the GLs as the option that balances the burden on the
institutions and ensures a base for comparability of analysis and reporting of DRs. The results
of the IRB survey (Figure 17) show that, in 45% of all PD models, the one-year DR is
calculated at least quarterly (i.e. quarterly, monthly or daily). However, for retail exposures
secured by immovable property SME and non-SME, these shares are higher (respectively
65% and 53%), and this also holds for qualifying revolving retail exposures (63%) and other
retail exposures (around 56%). Based on the fact that a quarterly frequency or higher is
already quite common, this requirement has been maintained in the final GLs. Nevertheless,
it should be acknowledged that this will entail a change in practice for, on average, around
54% of PD models.

60
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 17: Frequency at which one-year DRs are calculated, by COREP exposure class

% EAD 67% 0% 33% 0%

e - SME lending e - Other SME non-SME revolving SME non-SME Other % 57% 0% 43% 0%
% EAD 47% 0% 12% 40% 1%
0%
Qualifyin Retail - Retail -
Other

% 41% 0% 14% 39% 3%3%

% EAD 48% 0% 16% 36% 0%
Other

% 44% 0% 20% 34% 0%

2%
% EAD 16% 0% 13% 70% 0%
immovab immovab Retail -

% 37% 0% 26% 37% 0%

Corporat property property

% EAD
secured secured

35% 0% 12% 52% 0%

Retail - Retail -

by

le

% 43% 3% 13% 38% 2%

2%
% EAD 30% 0% 32% 38% 1%
0%
by

le

% 35% 0% 28% 33% 5%

0%
% EAD 63% 1% 18% 18% 0%
% 62% 1% 18% 18% 0%
Corporat

Specialis

% EAD 78% 0%
1% 21% 0%
e-

ed

% 75% 0% 13% 13% 0%

central Institutio Corporat

% EAD 51% 1% 25% 23% 0%

% 49% 1% 26% 24% 0%
% EAD 48% 0% 24% 28% 0%
ns

% 67% 0% 15% 19% 0%

governm
ents and

% EAD 57% 0% 22% 21% 0%

Central

banks

% 76% 0%6% 18% 0%

% EAD 46% 0% 16% 37% 0%
Total

% 52% 2% 15% 29% 1%

1%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Annually Semi-Annually Quarterly Monthly Daily Other

4.7 Calculation of the observed average DR

92.With respect to the calculation of the observed average DR, the GLs allow calculation using
both overlapping and non-overlapping windows, but require (in paragraph 80) that this
choice should be based on a documented analysis reflecting certain aspects. To obtain a view
on common practices across institutions, they were asked to indicate whether they use
overlapping or non-overlapping windows for the PD models under consideration, and

61
EBA REPORT ON THE IRB MODELLING PRACTICES

whether or not a specific analysis is undertaken to justify this choice. From the responses
shown in Figure 18 it is evident that non-overlapping windows are the most common
approach across PD models, i.e. the split between non-overlapping windows and overlapping
windows is roughly 60:40.
Figure 18: Use of overlapping versus non-overlapping windows in calculation of observed average DR

1%

2%

38%
36%
Overlapping windows
Non-overlapping windows
Other

61% 62%

Note: the inner circle shows the share of each option where all PD models are weighted equally, whereas the outer
circle shows the share of each option where PD models are weighted by their corresponding exposure value.

93.In most PD models (85%), no specific analysis was performed to justify this choice, although
this percentage is lower (70%) in those models where the calculation is performed based on
overlapping windows (see Table 25). Therefore, the requirement (in paragraph 80 of the GLs)
to analyse and document the considerations for choosing one or the other approach will
entail a change in practice in roughly 85% of PD models.
Table 25: Was any specific analysis undertaken to justify the choice of overlapping versus non-overlapping windows
for the calculation of the observed average DR?

PD models with non- PD models with

All PD models
overlapping windows overlapping windows

No. % % EAD No. % % EAD No. % % EAD

No 192 85 81 133 94 95 58 71 60
Yes 35 15 19 9 6 5 24 29 40
Total 227 100 100 142 100 100 82 100 100

62
EBA REPORT ON THE IRB MODELLING PRACTICES

94.From Table 26 it can be seen that the share of short-term or terminated contracts is not
significant in the majority of PD models (91%), although this share is lower (86%) in models
with overlapping windows and higher in models with non-overlapping windows (94%). This
lends support to the hypothesis that the presence of a short-term or terminated contract
could be one of the drivers for choosing to calculate the observed average DR using
overlapping windows. For 16 of 19 PD models for which a significant share of short-term or
terminated contracts was reported, the share of those contracts was specified. The average
share is around 11.5%, but the share ranges between 3% and 32%.
Table 26: Is there a significant share of short-term or terminated contracts within the period over which the observed
average DR is calculated?

PD models with non-overlapping PD models with overlapping

All PD models
windows windows
No. % % EAD No. % % EAD No. % % EAD
No 202 91 92 128 94 94 69 86 88
Yes 19 9 8 8 6 6 11 14 12
Total 221 100 100 136 100 100 80 100 100

95.With respect to the calculation of the observed average DR, the GLs specify (in paragraph 81)
that institutions should calculate the observed average DRs as the arithmetic average of all
one-year DRs, i.e. a simple average should be computed instead of a weighted average. With
respect to the weighting scheme, it appears that for around 67% of PD models (75% of
exposure values), a simple average is used to compute the long-run average, whereas a
weighted average is computed in about 33% of PD models (25% of exposure values). Some
small differences can be observed across COREP exposure classes, as shown in Figure 19. In
particular, the share of models using weighted averages to calculate the long-run average DR
is slightly higher (around 37% on average) for retail exposures than for exposures to
corporates (around 32% on average). This is probably related to the fact that the CRR
mentions, for retail exposures (see Article 180(2)(e) of the CRR), that an institution does not
need to give equal importance to historic data if more recent data are a better predictor of
loss rates.

63
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 19: What method (simple average or weighted average) is used to determine the long-run average DR? By
COREP exposure class

% EAD 83% 17%

e - SME lending e - Other SME non-SMErevolving SME non-SME Other
% 83% 17%
% EAD 69% 31%
ble Qualifyin Retail - Retail -
Other Other

% 67% 33%
% EAD 52% 48%
% 64% 36%
% EAD 60% 40%
immova immova Retail -

% 56% 44%
central Institutio Corporat ed Corporat property property

% EAD
secured secured

75% 25%
Retail - Retail -

by

% 66% 34%
% EAD 44% 56%
ble
by

% 50% 50%
% EAD 71% 29%
% 69% 31%
Corporat

Specialis

% EAD 79% 21%

e-

% 75% 25%
% EAD 69% 31%
% 67% 33%
% EAD 66% 34%
ns

% 50% 50%
governm
ents and

% EAD 80% 20%

Central

banks

% 41% 59%
% EAD 75% 25%
Total

% 67% 33%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Simple average Weighted average

96.For those models where a weighted average is used, several respondents mentioned how
these weights had been determined. In around 74% of those cases, it was mentioned that
the weight is determined by the number of observations in each one-year window. In around
25% of cases, a higher weight is given to more recent observations. For the latter, the most
common reason mentioned is that more recent observations reflect in a more adequate way
the credit policy of the institution, or because of improved processes and stricter credit
policies (i.e. the restricted availability of detailed information of the obligors in more distant

64
EBA REPORT ON THE IRB MODELLING PRACTICES

time periods). The determination of these weights is most often based on expert judgement.
For one model, it was mentioned that an appropriate weight is assigned to downturn years,
in line with specific guidance given by the competent authority.

4.8 Long-run average DR

97.Section 5.3.4. of the GLs describes provisions for estimating the long-run average DR.

98.Table 27 shows that the average length of the historical observation period from internal
data is 9.33 years, but there is a significant variation across exposure classes. Figures 20, 21
and 22 visualise the start and end dates of the reported historical observation period from
internal data for PD models applied to selected COREP exposure classes.

99.In these figures, wide variation can be observed among institutions, with respect not only to
duration but also to the start and end date. However, such differences may stem from the
availability of data and may be related to the variability of observed DRs, the existence, lack
or prevalence of one-year DRs relating to bad years, and to changes in the economic, legal or
business environment. None of the institutions uses non-consecutive periods.

100. It can also be seen from Table 27 that the length of the historical observation period is
shorter than five years for some models, which would contradict the CRR requirements in
Article180(1)(h) and 180(2)(e) of the CRR. Based on the comments, it became clear that
several of these institutions specified the length of the development sample (for risk
differentiation) instead of the length of the historical observation period. Other institutions
mentioned that that the model has been estimated on that (shorter than five-year) historical
observation period, but that a re-estimation of the model on the full historical observation
period leads to the same results.

101. It should be emphasised, however, that explanations justifying the length of the historical
observation period shorter than five years are not in line with the CRR.

Table 27: Length of the historical observation period for PD estimation (internal data), by exposure class

mean
mean
N (% min p10 p50 p90 max
(%)
EAD)
Total 222 8.72 9.75 0.83 4.92 8.00 14.01 21.85
Central governments and central banks 10 7.56 10.60 5.00 5.25 5.75 12.92 12.92
Institutions 18 7.50 8.93 4.00 5.00 6.67 12.92 12.92
Corporate — SME 64 8.84 9.33 1.50 5.00 8.00 13.92 17.51
Corporate — specialised lending 6 7.63 8.09 1.50 1.50 6.13 15.01 15.01
Corporate — other 65 8.85 8.97 1.50 5.00 8.00 13.92 19.01
Retail — secured by immovable property SME 38 8.69 8.66 3.50 4.84 8.96 13.92 17.51
Retail — secured by immovable property non-
117 8.90 10.39 0.83 4.00 8.92 15.01 21.85
SME
Retail — qualifying revolving 26 6.27 6.77 3.02 3.50 5.00 10.01 12.35

65
EBA REPORT ON THE IRB MODELLING PRACTICES

mean
mean
N (% min p10 p50 p90 max
(%)
EAD)
Retail — other SME 39 8.43 8.59 3.50 4.84 8.92 13.01 17.51
Retail — other non-SME 70 7.68 7.66 2.00 4.00 6.00 12.18 20.01

Figure 20: Start and end date of the historical observation period by PD model (internal data, retail mortgages —
non-SME)
31/01/1987

23/07/1992

13/01/1998

06/07/2003

26/12/2008

18/06/2014

09/12/2019

66
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 21: Start and end date of the historical observation period by PD model (internal data, corporate — SME)
01/01/1995

27/09/1997

23/06/2000

20/03/2003

14/12/2005

09/09/2008

06/06/2011

02/03/2014

26/11/2016

23/08/2019

Figure 22: Start and end date of the historical observation period by PD model (internal data, corporate — specialised
lending)
01/01/2002

27/09/2004

24/06/2007

20/03/2010

14/12/2012

10/09/2015

06/06/2018

102. Figure 5 shows for how many models (%) and how many exposure values (% EAD) the
institutions apply adjustments to the observed average DR for the purpose of PD estimation,
where such adjustments are applied in around half of the models (exposures). Figure 23

67
EBA REPORT ON THE IRB MODELLING PRACTICES

further shows that the direction of the adjustment is upward, i.e. conservative, in 87% of PD
models (71% of exposure values covered), and downward in 13% of models (29% of
exposures), corresponding to only 12 models in the sample.
Figure 23: If you apply adjustments to the observed average DR, what is the direction of the adjustment?

13% 29%
Downward (decrease the long-
run average default rate)
Upward (increase the long-run
average default rate)
71% 87%

Note: the inner circle shows the share of each option where all PD models are weighted equally, whereas the outer
circle shows the share of each option where all PD models are weighted by their corresponding exposure value.

103. Table 10 showed the distribution of reasons for applying (upward and downward)
adjustments to the observed average DR. When analysing the reasons for applying those
downward adjustments, the most common reason (applied in five of these 12 models,
covering 41% of models or 31% of exposures) is that the observed average DR is not
representative of the long-run average DR (i.e. it is not composed of an appropriate mix of
good and bad years), or the observed average DR is not representative of the default
behaviour of the current portfolio. In three models, it was mentioned that external data have
been used, which have been adjusted. However, it is necessary to exercise caution when
drawing conclusions from a small sample size (12 models). In addition, it cannot be ruled out
that several respondents took a literal reading of the question 23, such that institutions that
apply downward adjustments to the long-run average DR responded ‘no’ to this question.

104. To ensure harmonisation in the determination of the historical observation period, and to
ensure that the historical observation period is representative of the likely range of variability
of DRs, the GLs specify how to assess the representativeness of the likely range of variability
of DRs (in paragraph 83).

105. The GLs further specify that when the historical observation period is representative of
the likely range of variability of DRs, the long-run average DR should be computed as the
observed average of the one-year DRs in that period. Where no or insufficient bad years are
included, the average of the observed one-year DRs should be adjusted. Where bad years are
over-represented, the average of the observed one-year DRs may be adjusted, if institutions
23
Do you apply any adjustments to the observed average of DRs for the purpose of PD estimation?

68
EBA REPORT ON THE IRB MODELLING PRACTICES

can demonstrate a significant correlation between economic indicators and the available
one-year DRs. Finally, a benchmark is included in the GLs (in paragraph 86), to avoid
unjustified downward adjustments.

106. In relation to the over-representation of downturn years in the historical observation

period, institutions were asked whether or not they exclude some of the one-year DRs from
the historical observation period, i.e. whether or not the length of the period for which one-
year DRs is available is different from the length of the historical observation period. The
responses show that in 60% of reported PD models, all available periods for which one-year
DRs are available are included in the historical observation period, whereas some periods are
excluded in 40% of the PD models.

107. Many respondents explained these differences by mentioning in particular that only
certain periods are included, because the aim is to cover exactly one economic cycle. Some
respondents mentioned that older data are excluded because they are no longer
representative, but some other respondents also mentioned that the most recent data have
not been included because this would decrease the weight of the downturn periods. Some
mentioned that in older periods internal data were available, but that these data are not
granular enough (e.g. not available at segment level), and therefore that they include these
data after linking them with pooled data. Other respondents mentioned that only the most
recent data are included because these allow better predictions of loss rates.

108. The IRB survey further verified whether or not inclusion of years/periods for which no
one-year DRs are available in the historical observation period is common practice. Such a
situation may arise in particular in jurisdictions where no recent crisis has been experienced,
such that the most recent (available) DRs are not fully representative of the likely range of
variability of one-year DRs. Alternatively, institutions may include one-year DRs from older
periods, because only DRs from recent crisis years are available, and hence older data are
used to cover a full economic cycle. The responses show that the historical observation
period spans years for which one-year DRs are not available in only 20% of PD models.

109. These aspects have been taken into account in the GLs in the section on the long-run
average DR (section 5.3.4.), where there are criteria and conditions for assessing whether or
not the historical observation period is representative of the likely range of variability. In
particular, it is specified (i) that where no or insufficient bad years are included in the
historical observation period, the average of observed one-year DRs should be adjusted to
estimate a long-run average DR and (ii) that where bad years are over-represented in the
historical observation period, the average of observed one-year DRs may be adjusted to
estimate a long-run average DR if institutions can demonstrate the significance of the
correlation between economic indicators and the available one-year DRs. In addition, a
benchmark is included in the GLs (in paragraph 86), i.e. the maximum of the observed
average DR of the most recent five years, and the observed average DR of all available one-
year DRs, to include a backstop for unjustified downward adjustments of the PD estimates.

69
EBA REPORT ON THE IRB MODELLING PRACTICES

4.9 Calibration to the long-run average DR

110. One of the goals of the survey was to analyse the extent to which institutions apply
portfolio calibration (types 2 and 4) or grade calibration (type 1 and 3), whether or not this
involves an explicit function or methodology, and whether or not the PD is the long-run
average rate of that pool or grade. The explanation of the various calibration types, provided
in the instructions for the survey, is included here again for the convenience of the reader.

Type 1: Implicit calibration in accordance with Article 180(1)(a) of the CRR at grade or pool level.
The PD estimate per grade or pool is achieved by calculating the long-run average DR of the
relevant grade or pool. Therefore, the according grade or pool may have to be reconstructed if a
newly relevant risk driver has been incorporated into the model. The calibration sample equals
the sample where the long-run average DR is calculated.

Type 2: Explicit calibration in accordance with Article 169(3) of the CRR at portfolio level. The
rating model produces individual PDs per obligor or facility (‘raw PDs’). The underlying definition
of default may deviate from the regulatory notion of default. The raw PDs are adjusted such that
the average of the raw PDs is equal to the long-run average DR (usually on the level of the whole
portfolio, or of a calibration segment where relevant). The calibration sample equals the sample
on which the raw PDs are assessed for this purpose.

Type 3: Explicit calibration in accordance with Article 180(1)(a) of the CRR at grade or pool level,
based on individual estimates (which may also be scores) in accordance with Article 169(3) of the
CRR, and to a predefined master scale. The PD estimate is achieved by determining intervals of
estimates (or score) values such that the long-run average DR of obligors or facilities carrying
these values is equal to a predefined DR on a master scale (which will then be the PD estimate for
obligors with score values within the interval under consideration). The calibration sample equals
the sample on which the individual estimates are assessed.

Type 4: Explicit calibration in accordance with Article 180(1)(g) of the CRR at portfolio level, based
on averages of individual estimates in accordance with Article 169(3) of the CRR, and to a
predefined master scale. PD estimates are achieved by first adjusting the individual estimates
such that the average of the individual estimates equals the long-run average DR at portfolio
level, and then averaging the individual estimates that fall into a predefined interval on the
master scale. However, this type should also be chosen if the shifted individual estimates are
sorted into predefined PD intervals on the master scale. The calibration sample equals the sample
on which the individual estimates are assessed.

Type 5: Explicit calibration in accordance with Article 180(2)(b) of the CRR at portfolio level. PD
estimates are derived from an estimate of total losses and appropriate estimates of LGD. These
PD estimates are adjusted such that the average of these estimates equals the long-run average
DR at portfolio level.

70
EBA REPORT ON THE IRB MODELLING PRACTICES

Type 6: Implicit calibration in accordance with Article 180(2)(b) of the CRR at portfolio level. PD
estimates are derived from an estimate of total losses and appropriate estimates of LGD over the
relevant observation period.
Table 28: Use of different calibration types, by COREP exposure class

Type 1 Type 2 Type 3 Type 4 Other

% % % % %
No. % EA No. % EA No. % EA No. % EA No. % EA
D D D D D
Total 67 27 28 61 24 24 47 19 25 57 23 16 20 8 6
Central
governments and 3 18 5 1 6 3 8 47 60 3 18 13 2 12 19
central banks
Institutions 4 15 26 3 11 14 6 22 15 9 33 26 5 19 19
Corporate — SME 13 19 25 23 33 30 7 10 4 22 31 26 5 7 16
Corporate —
3 38 76 2 25 21 2 25 1 1 13 1 0 0 0
specialised lending
Corporate — other 17 21 28 20 24 20 11 13 8 27 33 28 7 9 17
Retail — secured
by immovable 6 15 10 16 40 47 5 13 22 12 30 18 1 3 4
property SME
Retail — secured
by immovable 30 25 29 35 29 31 19 16 24 24 20 13 12 10 2
property non-SME
Retail — qualifying
5 19 39 8 30 24 3 11 7 10 37 25 1 4 6
revolving
Retail — other
6 15 16 15 37 31 6 15 33 13 32 17 1 2 3
SME
Retail — other
21 30 37 17 24 22 9 13 19 20 28 19 4 6 3
non-SME

111. From Table 28 it can be seen that the only calibration types chosen by institutions are (i)
(implicit or explicit) calibration in accordance with Article 180(1)(a) of the CRR by grade or
pool level (i.e. types 1 and 3) and (ii) (implicit or explicit) calibration in accordance with
Article 169(3) of the CRR at portfolio level (i.e. types 2 and 4). The other types provided in the
drop-down menu (i.e. types 5 and 6, calibration based on total losses (implicit or explicit) in
accordance with Article 180(2)(b) of the CRR) have not been reported by institutions.

71
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 24: Grade versus portfolio calibration

6%

8%
Type 1 + 3 (grade or pool
calibration)
45% Type 2 + 4 (portfolio
40% 53% calibration)
47% Other

112. Overall, type 1 calibration is used the most (in 27% of models), and is closely followed by
type 2 (3 percentage points less), type 4 (23% of models), type 3 (19% of models), and other
calibration types. The survey shows a balanced range of practices in terms of level of
calibration. In terms of number of models, overall a slight majority calibrates to portfolio
central tendency (118 PD models, representing 47% of all PD models), while the remainder
calibrate by pool or grade (114 PD models, representing 45%) (see Figure 24). In terms of
share of EAD, a slight majority calibrates by pool or grade (53%), while the remainder
calibrate to portfolio central tendency (40%). Differentiating between retail and non-retail
does not reveal significantly different results.

113. Calibration type usage of single-exposure classes such as central governments or retail
secured by immovable property show more clear usage of grade calibration (with central
governments) and portfolio calibration (with retail secured by immovable property). This
may, however, be due to a low number of models analysed in these classes overall.

114. Regarding calibration types that were not covered by the given selection, the following
were mentioned, among others:

• explicit calibration in accordance with Article 180(2)(a) of the CRR — grade or pool level
(other institutions, however, claimed this to be type 1, and unfortunately the reference to
Article 180(2)(a) of the CRR was missing in the type 1 description above);
• mapping of ECAI (eligible credit assessment institution) scale in accordance with
Article 180(1)(f) of the CRR.

115. The GLs do not intend to restrict the various calibration types, but aim at clarifying the
them by providing a list of the references included in the CRR (in paragraph 91).
Furthermore, a definition of the term ‘calibration’ is included in the GLs in the definition
sections (in paragraph 8), to (i) clarify the distinction from model development (calibration is

72
EBA REPORT ON THE IRB MODELLING PRACTICES

the process that leads to appropriate risk quantification) and (ii) highlight that calibration
ensures that, for a calibration segment, PD estimates in a calibration sample correspond to
the long-run average DR at the level relevant for the applied calibration method.

116. The IRB survey results allow an assessment of the shares of models where calibration is
performed at portfolio level versus where calibration is performed at grade or pool level. In
the following, if the term ‘portfolio calibration’ is used, it refers to a process where banks
estimate one PD for each grade, with the objective (probably among other objectives) that
the average PD of the portfolio equals the long-run average DR at portfolio level. To avoid
misunderstandings, it should be noted that the objective of portfolio calibration and of grade
calibration is to estimate one PD for each grade. Therefore the term ‘portfolio calibration’
refers to the process of how the PD estimates per grade are achieved, rather than the result
of the calibration.

117. When the calibration of the current estimated average PD to the long-run average DR is
performed at portfolio level (types 2 or 4), the obligors will migrate across grades when
recalibrating. On the other hand, calibration to long-run average DR can also be checked and
assured at rating-grade level, taking into account the grades over the historical observation
period (types 1 or 3). Under this approach, and where this is combined with PIT rating
assignments, rating migrations induced by economic changes are transmitted to the average
portfolio PD, leading to cyclical capital requirements.

118. With respect to the level of calibration (grade or pool level, or at portfolio level), Table 29
shows (for the exposure class ‘retail exposures secured by immovable property’) that the
adjustment from the observed average DR to the final PD estimate is higher when calibration
is performed at grade or pool level (i.e. types 1 and 3) than when it is performed at portfolio
level.
Table 29: Observed average DR and final PD estimate across calibration types (retail exposures secured by
immovable property)

Calibration type
Type 1 or 3 Type 2 or 4 Other Total
N % % EAD N % % EAD N % % EAD N
Observed average DR 50 2.14 1.50 57 2.33 1.47 12 2.53 0.89 119
Average final PD 52 2.74 1.49 57 2.22 1.62 12 1.62 0.77 121
Difference (PD – observed average DR) 50 0.68 0.11 57 -0.11 0.14 12 -0.92 -0.13 119
Absolute difference if PD > DR 33 1.24 0.26 27 0.30 0.47 8 0.52 0.22 68
Relative difference if PD > DR 33 51.38 22.50 27 45.19 51.61 8 64.11 45.25 68
Absolute difference if PD < DR 13 0.56 0.34 25 0.59 0.32 4 3.80 3.95 42
Relative difference if PD < DR 13 29.23 25.77 25 15.36 17.08 4 64.39 62.48 42

119. To take account of these different practices with respect to calibration at portfolio level
versus at grade or pool level, the GLs specify (in paragraph 92) that institutions should
provide additional calibration tests at the level of the relevant calibration segment (which
corresponds to the portfolio level if there is only one calibration segment) where case

73
EBA REPORT ON THE IRB MODELLING PRACTICES

calibration is performed at grade or pool level, or perform additional calibration tests at the
level of the pool or grade where the calibration is performed at portfolio level.

120. For portfolio calibration (types 2 or 4), the number of time slices used (only one, or all
points in time available) determines whether the calibration sample is comparable with the
current portfolio in terms of obligor and transaction characteristics (usually only one point in
time), or whether the calibration sample is representative of the likely range of variability (all
points in time available). To prevent misunderstandings, ‘calibration sample’ here refers to
the dataset on which the ranking or pooling method is applied to perform the calibration;
this definition is also included in the GLs (in paragraph 8).

121. Using all time slices contained in the development sample refers to a situation where all
time slices that underlie the long-run average default rate calculation, are used in the
calibration. The calibration sample would then usually reflect the likely range of variability of
DR contained in the development sample. Therefore, the calibration sample would be equal
to the RDS used for the long-run average DR and the model development sample. Using this
approach, the current average PD at portfolio level may significantly deviate from the long-
run average DR at portfolio level, and this may lead to cyclical capital requirements,
reflecting the variability observed at the various points in time covered by the calibration
sample.

122. Using only one (the most recent) point in time for calibration would, on the other hand,
ensure that the current average PD at portfolio level is close to the long-run average PD at
portfolio level, and therefore that the calibration sample is most representative of the
current portfolio. This ensures that the requirement in Article 179(1)(d) of the CRR (referring
to comparability for the purpose of risk quantification) is met: ‘the calibration sample should
be comparable to the current portfolio in terms of obligor and transaction characteristics’.
This approach would lead to more stable capital requirements.

123. To assess how many points in time are used where institutions apply portfolio calibration,
a question was included in the survey. The results of the survey are as follows:
Table 30: If you use type 2 or 4 calibration, how many points in time were reflected in the calibration sample?

All time slices

contained in Otherwise:
Otherwise: 1 Otherwise: 2-
the more than 5 Other
time slice 5 time slices
development time slices
sample
% % % % %
N N N N N
% EA % EA % EA % EA % EA
o. o. o. o. o.
D D D D D
Total 56 48 45 27 23 20 12 10 12 20 17 23 2 2 0
Central governments and central
3 75 80 1 25 20 0 0 0 0 0 0 0 0 0
banks
Institutions 9 75 62 1 8 9 0 0 0 1 8 30 1 8 0
Corporate — SME 25 56 48 11 24 23 5 11 13 4 9 16 0 0 0
Corporate — specialised lending 1 33 6 1 33 3 0 0 0 1 33 91 0 0 0

74
EBA REPORT ON THE IRB MODELLING PRACTICES

All time slices

contained in Otherwise:
Otherwise: 1 Otherwise: 2-
the more than 5 Other
time slice 5 time slices
development time slices
sample
Corporate — other 28 60 55 11 23 20 5 11 9 3 6 15 0 0 0
Retail — secured by immovable
13 46 29 8 29 34 3 11 10 4 14 27 0 0 0
property SME
Retail — secured by immovable
22 38 36 13 22 21 8 14 15 14 24 28 1 2 0
property non-SME
Retail — qualifying revolving 9 50 27 7 39 38 2 11 35 0 0 0 0 0 0
Retail — other SME 8 29 8 11 39 43 3 11 12 6 21 37 0 0 0
Retail — other non-SME 13 35 18 13 35 33 4 11 27 7 19 22 0 0 0

124. Using all time slices contained in the development sample was the most common answer
(48% of all PD models) (see Table 30). However, a significant number of models (representing
23%) are also calibrated to only one time slice. Among those institutions that answered that
they took into account all time slices of the sample, some also mentioned the use of external
data.

125. To foster understanding of the impact of this aspect of calibration, and to achieve
convergence in practices, the GLs specify (in paragraph 88) that institutions should find an
appropriate balance between the calibration sample being comparable with the current
portfolio, in terms of obligor and transaction characteristics, and it reflecting the likely range
of variability of DRs.

126. As with portfolio calibration, where the number of time slices used determines whether
the calibration sample is comparable with the current portfolio in terms of obligor and
transaction characteristics (usually only one point in time), or whether the calibration sample
is representative of the likely range of variability (all points in time available), the timing of
the information used for grade calibration (type 1 or 3) affects the characteristics of the
calibration sample.

127. When the model is applied backwards to calculate the observed average DR per grade or
pool, the calibration sample is more representative of the likely range of variability, and less
of the current portfolio.

128. The analysis for type 1 and 3 calibrations reveals that a majority (65%) apply the model
backwards to assess the long-run average DR per grade and pool (see Table 31). However, a
significant share (35%) applies a variety of other methods. Most respondents were silent on
how they apply the model backwards in time, and none of them explained how they apply it
with qualitative components. On the basis of some comments, however, it is clear that the
model is applied backwards only for those risk drivers that are available at more distant
points in time.

75
EBA REPORT ON THE IRB MODELLING PRACTICES

Table 31: If you use type 1 or 3 calibration, which method do you apply when calculating the long-run average PD per
grade? By COREP exposure class

Apply the model backwards

to be able to calculate the
Other method
observed average DR per
grade or pool

No. % % EAD No. % % EAD

Total 69 65 75 37 35 25
Central governments and central banks 2 25 83 6 75 17
Institutions 2 25 81 6 75 19
Corporate — SME 13 65 78 7 35 22
Corporate — specialised lending 3 75 53 1 25 47
Corporate — other 18 69 54 8 31 46
Retail — secured by immovable property SME 6 55 83 5 45 17
Retail — secured by immovable property non-SME 35 74 82 12 26 18
Retail — qualifying revolving 5 63 92 3 38 8
Retail — other SME 9 75 93 3 25 7
Retail — other non-SME 24 80 87 6 20 13

129. The GLs specify (in paragraph 89) that institutions should conduct calibration before the
application of MoC or PD floors. This requirement has been inserted because, if an institution
first adds MoC to its long-run average DR and then calibrates (i.e. calibrates to the long-run
average plus MoC), then the effect of the calibration is smaller than when the PD estimates
are calibrated to the long-run average without inclusion of the MoC. Figure 25 shows that, in
around half of the PD models or exposures under PD models, the calibration is conducted
before the application of MoC, which is in line with the requirement of the GLs. This
requirement will, however, lead to a change in practice for around 41% of PD models, either
where no MoC is applied, or where calibration is currently conducted after the application of
MoC (where MoC is applied during calibration, the result is difficult to predict; whether or
not it will lead to a change in practice depends on the precise current practice).

76
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 25: Do you conduct calibration before or after the application of MoC?

5%
11% After application of MoC
7%
5%
34%
34% Before application of MoC

MoC is applied only during

calibration
55%
No MoC is applied
50%

Note: the inner circle shows the share of each option where all PD models are weighted equally, whereas the outer
circle shows the share of each option where PD models are weighted by their corresponding exposure value

130. The requirement in the GLs that calibration should be conducted before the application of
the PD floor follows the same logic as that for MoC, i.e. the application of the PD floor before
calibration leads to less conservative PD estimates. In the case of the PD floor, practices are
more consistent with the requirement in the GLs, i.e. in around 75% of PD models (79% of
exposures) calibration is already conducted before the application of the PD floor (see Figure
26). A change in practice is needed in around 20% of PD models.

Figure 26: Do you conduct calibration before or after the application of the PD floor?

1%
19%
3%
22%
After application of the PD floor
Before application of the PD floor
No PD floor is applied
74%
79%

Note: the inner circle shows the share of each option where all PD models are weighted equally, whereas the outer
circle shows the share of each option where PD models are weighted by their corresponding exposure value

4.10 Summary of model changes in PD estimation

77
EBA REPORT ON THE IRB MODELLING PRACTICES

131. From the above results of the survey it was possible, for selected questions, to directly
assess the shares of models (in the IRB survey sample) and the shares of exposure amounts
that would be affected by the chosen policy, and for which a change in modelling practice
will therefore be required once the GLs enter into force. For those selected questions for
which it is possible to directly assess whether or not there will be an impact, Table 32
provides an overview of the resulting model changes, in terms of both the share of PD
models (%) and the share of exposure values covered by these PD models (% EAD).

132. The appendix of this report shows how the various possible answers from the drop-down
menus in the survey have been classified to assess whether or not there would be a model
change, or if this is unknown.
Table 32: Summary of selected policy choices for PD estimation and the number of model changes

No model change Model change Not known

Paragraph
Policy choice No. % % EAD % % EAD % % EAD
in the GLs

All obligors who are in the

scope of application of the PD
53-54 252 81 80 19 20 0 0
model should receive an
individual PD estimation

Quarterly frequency to
78 252 46 54 54 46 1 0
calculate one-year DRs

Conduct a specific analysis to

justify the choice for
overlapping versus non-
80 227 15 19 85 81 0 0
overlapping windows for the
calculation of the observed
average DR

Conduct calibration before the

89 229 54 50 42 42 4 8
application of MoC 24
Conduct calibration before the
89 203 73 78 23 20 3 2
application of the PD floor 25

133. The requirement to conduct a specific analysis to justify the choice of overlapping versus
non-overlapping windows for the calculation of the observed average DR, and the
requirement to calculate one-year DRs at least at a quarterly frequency, will lead to a change
in practice in, respectively, 85% and 54% of the models. For the other policy aspects, the
share of affected models is smaller than the share of unaffected models.

24
Where the option ‘MoC is applied during calibration’ was selected by the respondent, it is difficult to assess whether
or not the GLs will entail a change in practice, since the precise practice is unknown and since the GLs introduce also a
definition of the concept of calibration. Those cases represent the category ‘not known’ in the last two columns of
Table 32.
25
Where the option ‘no PD floor is applied’ was selected by the respondent, it is difficult to assess whether or not the
GLs will entail a change in practice. Those cases represent the category ‘not known’ in the last two columns of Table 32.

78
EBA REPORT ON THE IRB MODELLING PRACTICES

134. Table 33 further shows how many aspects will at least need to be changed in the PD
models. It can be seen, for instance, that the share of unaffected models is only 8%, although
it should be mentioned that this relates only to the five aspects of PD models on which
explicit questions were included in the survey, described above. The areas where guidance is
given on PD models is very broad, however, and is likely to affect many more modelling
aspects. In addition, these calculations include only the models for which a model change is
expected, and not those for which it is unknown whether or not there a model change will be
necessary. The share of models for which at least two aspects will have to be changed is 33%.

135. Given that the list of policy aspects specified in the GLs is much longer than the selected
aspects included in Table 32 and Table 33, the estimates below are a lower bound to the true
number of affected models. In practice, one may assume that all models will probably have
to be changed in one or more aspect.

Table 33: Summary of number of aspects to be changed in PD estimation

Number of aspects
No. % % EAD
to be changed

0 21 8 21
1 62 25 20
2 82 33 29
3 59 23 16
4 25 10 10
5 3 1 4
Total 252 100 100

79
EBA REPORT ON THE IRB MODELLING PRACTICES

5. LGD models

5.1 Characteristics of the survey sample

136. Institutions have been asked to indicate which types of LGD models are used within their
institution. The results are presented in Table 34. Overall, 82% of institutions make use of
work-out LGD and almost 40% of institutions use multivariate regression analysis. Note that
the shares of types of LGD models do not add up to 100% because this as a ‘tick box’
question, where institutions could select multiple answers.
Table 34: Types of LGD models used within the institutions

No. N %
Work-out LGD 78 95 82
Market-implied LGD (market based) 8 95 8
LGD based on total losses and PD estimates 2 95 2
Multivariate regression analysis 36 95 38
Other 18 95 19
Note: this was a ‘tick box’ question, hence respondents could select several of the above answers.

137. The same question was also asked for the LGD models for which the survey was
completed, in the format of a drop-down menu. The results are presented in Table 35, for
the LGD non-defaulted, LGD in-default and ELBE models. It should be noted that the sample
size of the LGD in-default and ELBE models is usually smaller than that of the non-defaulted
LGD models, since several institutions did not always complete these additional columns. The
drop in the sample size of the non-defaulted (190) versus the total number of LGD models
(202; see Table 1) is due to data quality issues, or because some banks left this question
open. More than half of the non-defaulted LGD, LGD in-default and ELBE models tested use a
work-out LGD approach, while the second most common approach is multivariate regression
analysis, consistent with the reporting in Table 34. Overall, the types of models for which the
survey was completed are representative of the types of models within the institution.
Table 35: Different types of LGD models for which the survey was completed

LGD non-defaulted LGD in-default ELBE

% % %
No. % No. % No. %
EAD EAD EAD
LGD based on total losses and PD estimates 1 1 1 1 1 1 1 1 1
Market-implied LGD (based on market data) 3 2 4 4 2 5 1 1 0
Multivariate regression analysis/sophisticated
45 23 25 38 22 25 25 1% 15
statistical model
Work-out LGD 115 58 53 95 56 52 81 55 52
Other or any combination of the above 33 17 17 32 19 18 40 27 32
Total 197 100 100 170 100 100 148 100 100

80
EBA REPORT ON THE IRB MODELLING PRACTICES

138. The GLs specify (in paragraph 102) that LGD estimates should be based on the
institution’s own loss and recovery experience as it is reflected in historical data on defaulted
exposures, and that LGD estimates should not be derived only from the market prices of
financial instruments. Consequently, the GLs will entail a change for around 1.52%, 2.35%
and 0.38% of LGD non-defaulted, LGD in-default and ELBE models respectively.

139. A variety of types of scales are being used for LGD models. Among all options, a
continuous LGD scale is most popular (representing 48%, 55% and 61% of LGD non-
defaulted, LGD in-default and ELBE models), followed by facility pools (representing around
20-25% of models) (see Table 36).

Table 36: Types of scales used in LGD and ELBE estimation

LGD non-defaulted LGD in-default ELBE

% % %
No. % No. % No. %
EAD EAD EAD
Continuous LGD scale 95 48 55 95 55 61 92 61 60
Discrete scale of facility grades 36 18 11 27 16 10 17 11 13
Facility pools 50 25 19 41 24 13 31 20 10
Other 17 9 15 11 6 15 12 8 16
Total 198 100 100 174 100 100 152 100 100

140. Table 37 shows whether the LGDs are assigned to the secured part of the exposure, the
unsecured part of the exposure or the whole exposure. In the vast majority of LGD and ELBE
models, the final parameter estimate is assigned to the whole exposure (77% of models and
79% of exposure values for the LGD non-defaulted).

Table 37: Assignment of LGD or ELBE estimate to the whole exposure or only the (un)secured part of the exposure

LGD non-defaulted LGD non-defaulted LGD non-defaulted

No. % % EAD No. % % EAD No. % % EAD
Whole exposure 154 77 74 139 79 74 128 82 81
Secured part of the exposure 8 4 4 7 4 4 7 4 5
Unsecured part of the exposure 21 11 11 13 7 8 9 6 4
Other 17 9 11 16 9 14 12 8 10
Total 200 100 100 175 100 100 156 100 100

141. In relation to the average realised LGD and the average final LGD estimate, Table 38
reveals that the LGDs for the secured part of the exposures are significantly below the LGDs
that are assigned to the unsecured part of the exposure, which is in line with expectations.
The LGDs assigned to the whole exposure are between the average LGDs of the secured and
unsecured parts.

81
EBA REPORT ON THE IRB MODELLING PRACTICES

Table 38: Average realised LGD and final LGD estimate, depending on whether the LGD is assigned to the secured
part of the exposure, the unsecured part of the exposure or the whole exposure

Average realised LGD Average final LGD estimate

mean mean
mean mean
N (% min max N (% min max
(%) (%)
EAD) EAD)
Secured part of the
8 13.82 18.99 1.94 30.65 8 16.10 18.79 5.20 33.43
exposure
Unsecured part of the 100.0
18 39.54 35.90 6.30 90.80 20 46.18 41.31 5.40
exposure 0
Whole exposure 136 27.10 24.01 0.79 81.13 150 30.25 23.61 0.42 99.27

142. Table 39 and Table 40 show which model components are used most often in LGD non-
defaulted, LGD in-default and ELBE models. The cure rate (i.e. the rating of return to
performing portfolio) and the recovery rate with respect to the loan amount are used in
more than half of the LGD models. In LGD in-default and ELBE models, the time in-default and
recoveries realised so far are already used in around half of the models.

Table 39: Model components used in the estimation of LGD non-defaulted

N % % EAD
Rate of return to performing portfolio 197 58 58
Recovery rate conditional on returning to the living portfolio 197 22 17
Recovery rate with respect to the collateral value 197 47 58
Recovery rate with respect to the loan amount 197 57 48
Other 197 29 38
Note: this was a ‘tick box’ question, hence respondents could select several of the above answers.

Table 40: Use of time in-default and recoveries realised so far as model components in estimation of LGD in-default
and ELBE

LGD in-default ELBE

N % % EAD N % % EAD
Time in-default 177 49 54 156 51 62
Recoveries realised so far 177 47 37 156 52 40
Note: this was a ‘tick box’ question, hence respondents could select several of the above answers.

5.2 Recoveries from collaterals

143. For the purpose of the GLs, repossession of collateral is understood as a situation where
an institution realises collateral by taking it over and recording it on the balance sheet of the
institution, and at the same time the amount of credit obligation is diminished by the value
of the asset.

82
EBA REPORT ON THE IRB MODELLING PRACTICES

144. Figure 27 shows the distribution of practices when it comes to the repossession of
collateral in the course of the recovery process. It can be seen that banks do not repossess in
42% of LGD models, whereas in 58% of LGD models institutions do repossess at least
occasionally. In 29% of LGD models this practice is standard in the recovery process, whereas
in the other 29% it occurs only occasionally.
Figure 27: Do you repossess collateral in the course of the recovery process? Retail, corporate, institutions, and
central governments and central banks
Retail Corporate

% EAD 42% 44% 14%

% 38% 29% 33%

% EAD 27% 32% 41%

% 45% 28% 27%

Total # central Institutio

% EAD 48% 24% 28%

ns

% 43% 29% 29%

governme

% EAD 54% 6% 40%

nts and
Central

banks

% 43% 29% 29%

% EAD 28% 36% 36%

models

% 42% 29% 29%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

No Yes - but only ocassionally Yes - often, it's a part of standard recovery process

145. The GLs specify (in paragraph 116) that institutions should assess the value of the
repossession at the date of repossession as the lower of (i) the value by which the credit
obligation of the obligor has been diminished as a result of the repossession of the collateral
and (ii) the value of the repossessed collateral as recorded on the balance sheet of the
institution. In addition, it is specified (in paragraph 117) that institutions should apply an
appropriate haircut to this value where there is significant uncertainty over whether or not
the value of repossession adequately reflects the value of the repossessed collateral. Where
sufficient past experience with regard to repossession of collaterals exists, the haircuts
should be supported by historical observations and regularly back-tested (paragraph 117(c)).
In the absence of such experience, the assessment will have to be performed on a case-by-
case basis, but this will require more conservatism as such an assessment will be less reliable.

146. The IRB survey shows the institution’s current practices with respect to (i) whether or not
the value of the sale or the value of the repossession is included in LGD estimation and (ii)
whether or not a haircut is applied to this value. From Figure 28 it can be seen that taking the
value of the sale (after repossession), but a null recovery where the collateral has not yet
been sold on the LGD calculation date, is the most common approach, applied in 36% of the
LGD models. Overall, the value of sale is used in 59% of LGD models in the sample

83
EBA REPORT ON THE IRB MODELLING PRACTICES

(representing 65% of the exposure values), whereas the value of repossession is used in 25%
of models (representing 20% of exposures). When it comes to the application of a haircut (to
the value of sale or the value or repossession), no haircut is applied in around 53% of LGD
models, and a haircut is applied in only around 30% of LGD models.

Figure 28: Which recovery value is recognised in the calculation of the realised LGD?

Value of repossession after a hair-cut, regardless of whether it

has been sold or not on the LGD calculation date

Value of repossession without a hair-cut, regardless of whether

8% it has been sold or not on the LGD calculation date
15%
18% 12% 12%
1% Value of the sale (after reposession) but a null recovery in case
13% the collateral has not been sold yet on the LGD calculation date
4%
Value of the sale (after reposession) but the value of reposession
24% 19% after a hair-cut in case the collateral has not been sold yet on the
LGD calculation date
36% Value of the sale (after reposession) but the value of reposession
40% without a hair-cut in case the collateral has not been sold yet on
the LGD calculation date
Other

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

147. The approach included in the GLs (in paragraphs 116-117) corresponds to the option
‘value of repossession after a haircut, regardless of whether or not it has been sold on the
LGD calculation date’, and represents the current approach in 12% of the LGD models in the
sample. The most common approach in the IRB survey sample of LGD models is to take ‘the
sale (after repossession), but a null recovery where the collateral has not yet been sold on
the LGD calculation date’.

148. In this regard, two distinct options have been considered, each with their pros and cons:

(a) value of repossession: repossession of a collateral by an institution should be

considered a recovery;

(b) value of sale: the recovery should be associated with cash payments only, and hence
in this case only the final price for which the institution sells the repossessed
collateral should be taken into account.

84
EBA REPORT ON THE IRB MODELLING PRACTICES

149. The latter approach would address the situations where institutions take over collaterals
and then sell them with significant loss. In this case, LGD estimates based on the value of
repossession would be underestimated, as they would not include the loss that was incurred
by the institution on a sale. Furthermore, if the repossessed collateral is sold only after a long
period of time, the discounting effects are not included in the LGD estimates. Another
argument that was taken into account for this approach is that, where an institution
repossesses illiquid collateral, the value cannot be established in a reliable manner.
Therefore the real value can only be verified at the moment of sale.

150. However, in addition to the above argument, the following aspects have been taken into
account:

(a) The value of the collateral that is subject to repossession decreases the obligation of
the debtor at the moment of repossession; hence, even if the institution eventually
realises a loss on a sale of this object or property, it can no longer be claimed from
the obligor.

(b) Repossession is usually the choice of an institution, not an obligation. This could be
understood as an investment decision on the part of an institution. In such a context,
repossession of collateral is a situation equivalent to receiving cash recovery and then
investing this money in a certain non-credit asset. In this situation, the results of an
investment decision on the part of an institution should not influence the LGD
estimates.

(c) The value of repossession is under the scrutiny of the financial auditor, as it affects
the balance sheet of an institution and has to be set in accordance with the applicable
legal framework. While the institution has an incentive to keep the value as low as
possible, the obligor will not accept a value that is too low. This should ensure that
the value of repossession is reasonable.

(d) Repossession is expected to be used more often in bad times than in good times, and
therefore the value of collateral at the moment of repossession will most probably be
relatively low. The argument that the value of the sale will usually be lower than the
value of repossession, and lead to more conservative LGD estimates, may therefore
not always be true.

(e) After the repossession the asset is recorded on the balance sheet of the institution
and receives a risk weight that is adequate for non-credit assets. From that moment,
the risk for the bank is not credit risk, and therefore it should not be included in the
estimation of risk parameters for credit risk.

(f) Institutions that decide to repossess collateral may not have an intention to sell it
subsequently. In this case, the effect of realising collateral would never be reflected in
the estimates.

85
EBA REPORT ON THE IRB MODELLING PRACTICES

(g) In some cases, the sale of the repossessed collateral could only take place many years
after the repossession. In this case also, the effect of realising collateral could never
be reflected in the estimates, or would be reflected only with a significant delay.

(h) If institutions had to include in LGD estimates the loss or profit realised on collateral
after the repossession, then the proper calculation would have to include not only the
final sale price but also other cash flows related to this asset. For instance, in the case
of repossessed immovable property the bank would have to include the maintenance
costs (e.g. energy, security, insurance, etc.), possibly the rent received from tenants if
the property is rented, discounting effects on any cash flows, and finally the sale price
diminished by the cost of sale (e.g. intermediary, taxes, valuation, etc.). This would be
a very complicated calculation but it would have no relation to the performance of
the defaulted exposure.

151. After considering all these arguments, it has been decided that repossession should be
treated as a recovery and that the value of repossession should be treated as an amount of
recovery.
Table 41: Which recovery value is recognised in the calculation of the realised LGD? By COREP exposure class

Value of
Value of
the sale
the sale
Value of (after
Value of (after
Value of the sale repossessi
repossessi repossessi
repossessi (after on), but
on on), but
on after a repossessi the value
without a the value
haircut, on), but a of
haircut, of
regardless null repossessi
regardless repossessi
of recovery on
of on after a
whether where the without a Other Total
whether haircut
or not it collateral haircut
or not it where the
has been has not where the
has been collateral
sold on been sold collateral
sold on has not
the LGD on the has not
the LGD been sold
calculatio LGD been sold
calculatio on the
n date calculatio on the
n date LGD
n date LGD
calculatio
calculatio
n date
n date

Total no. of models 13 14 40 21 4 20 112

% 12 13 36 19 4 18 100
% EAD 8 12 40 24 1 15 100
Central governments
1 0 2 0 0 1 4
and central banks
% 25 0 50 0 0 25 100
% EAD 6 0 77 0 0 17 100
Institutions 1 1 3 0 1 2 8
% 13 13 38 0 13 25 100
% EAD 4 29 54 0 11 3 100

86
EBA REPORT ON THE IRB MODELLING PRACTICES

Value of
Value of
the sale
the sale
Value of (after
Value of (after
Value of the sale repossessi
repossessi repossessi
repossessi (after on), but
on on), but
on after a repossessi the value
without a the value
haircut, on), but a of
haircut, of
regardless null repossessi
regardless repossessi
of recovery on
of on after a
whether where the without a Other Total
whether haircut
or not it collateral haircut
or not it where the
has been has not where the
has been collateral
sold on been sold collateral
sold on has not
the LGD on the has not
the LGD been sold
calculatio LGD been sold
calculatio on the
n date calculatio on the
n date LGD
n date LGD
calculatio
calculatio
n date
n date

Corporate — SME 5 5 7 3 4 6 30
% 17 17 23 10 13 20 100
% EAD 18 39 12 6 8 16 100
Corporate —
0 1 0 1 1 1 4
specialised lending
% 0 25 0 25 25 25 100
% EAD 0 71 0 14 15 0 100
Corporate — other 4 4 11 3 4 10 36
% 11 11 31 8 11 28 100
% EAD 11 27 35 5 6 17 100
Retail — secured by
immovable property 2 6 9 4 1 3 25
SME
% 8 24 36 16 4 12 100
% EAD 1 22 52 11 1 14 100
Retail — secured by
immovable property 6 8 23 15 0 9 61
non-SME
% 10 13 38 25 0 15 100
% EAD 7 8 35 34 0 16 100
Retail — qualifying
0 2 2 1 0 0 5
revolving
% 0 40 40 20 0 0 100
% EAD 0 19 76 4 0 0 100
Retail — other SME 0 4 7 4 1 2 18
% 0 22 39 22 6 11 100
% EAD 0 27 60 6 1 6 100
Retail — other non-
2 7 10 4 0 1 24
SME
% 8 29 42 17 0 4 100%
% EAD 14 29 52 3 0 1 100%

87
EBA REPORT ON THE IRB MODELLING PRACTICES

152. Table 41 shows further that the chosen approach also differs according to COREP
exposure class. It can be seen, for instance, that the value of repossession is most common
(representing 60% of exposures covered) for LGD models covering corporate SME exposures,
whereas the value of sale is the most common approach for LGD models covering retail
exposures secured by immovable property SME and non-SME (representing respectively 64%
and 69% of exposures covered).

5.3 Eligibility of collaterals

153. The GLs clarify (in paragraph 124) that institutions may take into account any type of
collateral in the LGD estimation as long as the requirement of Article 181(1)(f) of the CRR is
met. This comes in addition to the clarification included in the RTS on IRB assessment
methodology (Article 55), i.e. that to meet this requirement the institution’s internal policies
should be at least fully consistent with the requirements of section 3 of Chapter 4 of the CRR
with regard to legal certainty and regular valuation of collateral.

154. Furthermore, it is also envisaged that, for the purpose of LGD estimation, institutions may
use specific types of collaterals that are not explicitly described in Chapter 4 of the CRR. In
these cases the policies and procedures relating to internal requirements for valuation and
legal certainty should be appropriate to the respective type of collateral.

155. With a view to identifying which types of collateral (if any) are not included in the LGD
estimation and the potential reasons for this, a question has been included in the IRB survey.
The results are shown in Figures 29 and 30.
Figure 29: Are certain types of collateral not taken into account in the LGD estimates? Retail, corporate, institutions,
and central governments and central banks
Corpora

% EAD 32% 68%

te

% 51% 49%
% EAD 69% 31%
Retail

% 65% 35%
% EAD
central Instituti

27% 73%
ons

% 43% 57%
% EAD 68% 32%
Central
govern
ments

banks
and

% 75% 25%
% EAD 60% 40%
Total

% 61% 39%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

No Yes

88
EBA REPORT ON THE IRB MODELLING PRACTICES

156. In 60% of the LGD models, on average, all collaterals are taken into account in the LGD
estimation. For some of those models where all collaterals are included in the LGD
estimation (i.e. where there are no exceptions), the respondents indicated that this is
because (i) there is no collateral in the portfolio (because the LGD model applies only to the
unsecured part of the exposure); (ii) the LGD model covers residential mortgages and
therefore only residential real estate (houses) are accepted as collateral, with only a few
immaterial exceptions; or (iii) the collateral is reflected in a form of recovery in the
calculation, and therefore all collateral types that generate recovery for the given exposure
in-default are considered.

157. In around 40% of the LGD models, there is at least one type of collateral that is not
(always) taken into account. The respondents mentioned a variety of types of collaterals, but
given the multitude of types and the lack of a specific question on the amount, and/or a
structured list of collateral types, it is only possible to obtain a broad view of collaterals that
were stated to be excluded in LGD estimation. In general, one can notice a significant
heterogeneity in practices. Some institutions, for instance, consider only residential real
estate as collateral (and discard commercial immovable property, cash and guarantees),
whereas others take all immovable property collateral, cash and guarantees into account,
discarding anything else. To give an overview of the comments, the following responses were
submitted: (i) only property, cash and guarantees are included, and everything else is
discarded; (ii) specialised equipment, computer equipment, intangible assets, money in other
bank accounts, stocks of companies with low rating, stocks of non-public companies, shares
in private limited companies, non-financial deposits and some types of receivables are not
included in LGD estimation; (iii) guarantees are excluded from LGD estimation; (iv) insurances
are excluded from LGD estimation; (v) motor vehicles, other collateral assignments with a
nominal value < EUR 500 000 (e.g. non-domestic other collateral assignment, precious
metals, computer software), other collaterals and other pledges are excluded from LGD
estimation; (vi) only real estate collateral is included in the LGD estimation, and all other
securities (e. g. financial collateral) are excluded; (vii) machinery and equipment, specific
vessels, oil service-related collaterals, other aircraft, trains, unspecified and unlisted shares,
specific financial collaterals and non-specified deposits, and specific inventories are excluded
from LGD estimation.

158. From Figure 29 it can further be seen that it is more common to exclude certain types of
collaterals from LGD estimation for exposures to institutions and corporates (57% and 49% of
LGD models).Whereas this may be due to the lower number of LGD models for exposures to
institutions (eight LGD models), which makes the results less reliable, it may be explained by
the wider range of possible collaterals for exposures to corporates.

159. From the above list of collaterals stated to be excluded in LGD estimation it can be
inferred that respondents understand that ‘collateral’ (as referred to in Article 181(1)(f) of
the CRR) refers to the broad range of possible funded as well as unfunded credit protection,
i.e. ‘collateral’ has not been understood as referring solely to physical or financial collateral.

89
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 30: What are the reasons for not recognising certain types of collateral in the LGD estimates?

17%
18% Limited historical observations
for that type of collateral
39% Limited use of this type of
40%
collateral
16% Regulatory non-eligibility (please
19%
explain the legal grounds)
Other
26%

24%

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value

160. Figure 30 shows that data-driven reasons (limited historical observations or limited use of
a certain type of collateral) account for 37% of LGD models (34% of exposures covered),
whereas around 25% stems from regulatory non-eligibility, and around 40% stems from
other reasons.

161. Among the reasons for regulatory non-eligibility, the respondents mentioned (i) non-
enforceability of the collateral; (ii) non-eligibility of certain guarantees; (iii) the assessment of
the collateral values not being in line with the CRR; and (iv) that closed funds without daily
market values cannot be included as collateral in the LGD estimation 26.

162. Among the many ‘other’ reasons mentioned are (i) that all of the other three reasons
apply (limited observations, limited use and regulatory non-eligibility); (ii) that collateral is
only used for the segmentation of debt, and in particular for senior secured securities; (iii)
that personal guarantees linked to states and institutions are treated under the SA because
there is no rating model validated for these types of counterparts (thereby making use of
Article 183(4) of the CRR). Funded collateral not linked to real estate is treated under the SA
because there is no market model validated to internally calculate the haircuts.

163. Figure 31 shows the current practices with respect to the inclusion of protection in the
form of guarantees and credit derivatives in the LGD estimates. The responses therefore
26
According to Article 197(5)(c) of the CRR, institutions may use units or shares in CIUs as eligible collateral only where
the units or shares have a daily public price quote.

90
EBA REPORT ON THE IRB MODELLING PRACTICES

provide an overview of the practices that are set out in Article 60 of the RTS on assessment
methodology, which specifies the eligibility and assessment of guarantees and credit
derivatives in risk parameters.

164. Whereas this question did not apply to many LGD models (because there are no
guarantees or credit derivatives), the results show that, for those remaining (91) LGD models,
the most common approach (used in around 45% of LGD models and exposures) is to include
the guarantee or credit derivative in the LGD estimation. However, this approach is more
common in retail portfolios than in corporate portfolios (53% versus 31% of LGD models). In
contrast, the substitution of the PD of the obligor with the PD of the protection provider is
more common in corporate portfolios than in retail portfolios (48% versus 25% of LGD
models), which can probably be explained by the fact that corporate obligors more often
have regulatory PDs available than retail obligors. Finally, it should be noted that the
treatment in accordance with Article 153(3) of the CRR (double default effect) is seldom
applied; only three LGD models in the IRB survey make use of this approach.

Figure 31: How do you include in the LGD estimates protection in the form of guarantees and credit derivatives?
Retail, corporate, institutions, and central governments and central banks
Corporat

% EAD 0% 27% 59% 0% 13%

e

% 0% 31% 48% 2% 19%

% EAD 1% 57% 28% 1% 14%
Retail

% 2% 53% 25% 5% 16%

Total # central Institutio

% EAD 0% 28% 72% 0%

ns

% 0% 25% 67% 0% 8%
governm
ents and

% EAD 0%
4% 21% 0% 74%
Central

models banks

% 0% 20% 40% 0% 40%

% EAD 1% 46% 36% 0% 17%
% 1% 45% 32% 3% 19%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Included in PD estimation
Included in LGD estimation
PD of obligor substituted with PD of the protection provider
Treatment in accordance with Article 153(3) CRR (double default effect)
Other

165. In relation to the GLs, clarification is included with respect to the requirements that apply
to treating ‘third parties’ in PD estimation (section 5.2.3.). In particular, it is specified that
institutions should have clear policies stating the conditions under which the rating of a third

91
EBA REPORT ON THE IRB MODELLING PRACTICES

party that has a contractual or organisational relation with an obligor of the institution may
be taken into account in the assessment of risk of the considered obligor. This is relevant in
particular to a rating transfer, the use of a rating of a third party serving as input to the PD
model reflecting potential support for the obligor, or the use of the rating of a third party as
an indication for an over-ride of the PD of the obligor.

5.4 Inclusion of collaterals in the LGD estimation

Table 42: How is collateral included in the LGD estimation?

N No. % % EAD
As a segmentation criterion (to an LGD model) 68 17 25 21
As a pooling criterion 68 10 15 13
As a risk driver 68 27 40 25
As a model component (separate recovery rates) 68 45 66 72
Other 68 2 3 0

166. Table 42 shows how institutions incorporate collaterals in their LGD estimates. It should
be noted that the IRB survey was designed such that institutions could indicate multiple
options: (i) as a segmentation criterion; (ii) as a pooling criterion; (iii) as a risk driver; and (iv)
as a model component (separate recovery rates). These are not mutually exclusive 27. It
should be noted that collaterals are included as a separate model component in 66% of the
LGD models, i.e. this is the most common approach, followed by including collateral as a risk
driver (40%) and as a segmentation criterion (25%). Regarding the inclusion of collateral
through separate recovery rates, the GLs include additional clarification on how this should
be undertaken (in section 6.2.3.). In particular, it is specified that (i) institutions should avoid
a bias that may stem from including in the estimation sample the observations where the
exposure was secured by only a part of the value of the collateral; (ii) they should recognise
and include in this estimation direct costs related to the collection on these types of
collaterals; and (iii) they should include all recoveries realised, including those where the
realisation of the collateral has been completed but the overall recovery process is not yet
closed.

5.5 Calculation of economic loss and realised LGD

167. One of the main aspects included in the GLs with regard to LGD estimation is a detailed
specification of the definition of economic loss and realised LGD. As these are the main
concepts underlying the estimation process, the harmonisation of these definitions is a
prerequisite for comparable LGD estimates. The GLs therefore contain specific provisions

27
It should be noted that there were also another options, i.e. ‘not applicable — there are no collaterals’. The results in
Table 42 cover only those models where the option ‘not applicable’ was not chosen by the institution, and where the
institution chose at least one of the other options (covering 68 models in total).

92
EBA REPORT ON THE IRB MODELLING PRACTICES

with regard to the treatment of additional drawings after default, unpaid late fees and
interest after default, direct and indirect costs, and discounting factor.

5.5.1 Definition of economic loss and realised LGD

168. The question here is how additional recovery cash flows stemming from exposures that
return to non-defaulted status should be treated in the realised LGD. In the IRB survey,
respondents were asked to indicate how economic loss for a cured case is measured.
Whenever this report refers to ‘cures’ it means exposures that returned to non-defaulted
status (this specification was also included in the instructions to the survey). In practice,
institutions use various definitions of cures, and sometimes additional criteria are specified
relating, for instance, to the length of time in-default or to the level of loss.

169. The CP on the GLs specified that the economic loss for exposures that return to non-
defaulted status should be calculated as for all other defaulted exposures, with the only
difference being that the additional recovery cash flow is added to the calculation as if the
payment was made by the obligor in the amount that was outstanding at the date of the
return to non-defaulted status, including any principal, interest and fees.
Figure 32: How is economic loss of a cured case measured?

Assume that the economic loss for cured cases is zero

5%
4%
Other
29%32%
43%
52%
Using the same methodology as for other defaulted
exposures including the discounted additional
22% recovery cash flow at the date of the return to non-
14% defaulted status
Using the same methodology as for other defaulted
exposures including, but not discounting, additional
recovery cash flow at the date of the return to non-
defaulted status

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

170. The responses shown in Figure 32 illustrate that such guidance is necessary to eliminate
undue RWA variability stemming from the treatment of cures.

171. Whereas the respondents to the CP asked for clarity on these provisions in the GLs, it was
also suggested that the amount that was still outstanding at the moment of return to non-

93
EBA REPORT ON THE IRB MODELLING PRACTICES

defaulted status (principal, interest and/or fees), i.e. the additional artificial cashflow, should
also be discounted.

172. Hence, the following policy options have been considered:

(a) approach proposed in CP – no discounting of artificial cash flow;

Pros Cons

Creates possibility for regulatory arbitrage.

Where an obligor is planning to repay the
Avoids potential overestimation of realised loss
obligation, institutions could change the status
where all obligations were repaid, including any
to non-defaulted before the payment, to avoid
penalty fees and interest.
discounting of the cash flow, and decrease the
calculated realised loss.

Would require a change in currently applied

Avoids potential discounting of the part of
practices to a larger extent than the alternative.
exposure that was never past due (this
It may not be appropriate to introduce a less
exposure may have been past due before,
conservative approach and force institutions to
however, especially in the case of
change the current approach to a less prudent
restructuring).
one.

The IRB survey results show that this approach

is applied in only around 4% of LGD models.

(b) artificial cash flow should be discounted as all other cash flows.

Pros Cons

Consistent with the understanding of the

concept of discounting factor (i.e. that it
reflects uncertainty around the cash flows on If the loan was not terminated at or after the
defaulted exposures at the moment of default). moment of default, the part of the exposure
At the moment of default, the whole exposure that returned to non-defaulted status may
was in-default and there was uncertainty never have been past due. Discounting this part
regarding the payments and potential of the exposure may therefore be considered
subsequent cure. The artificial cash flow should overly conservative.
therefore also be discounted, as at the moment
of default it was uncertain whether or not it
was going to return to non-defaulted status. All

94
EBA REPORT ON THE IRB MODELLING PRACTICES

cash flows are comparable as they are all

discounted at the same moment – the moment
of default.

The impact may be significant in particular in

the following cases:
(a) exposures that have been in defaulted
status for a considerable time, especially in the
case of exposures subject to distressed
Ensures equal treatment of cases that return to restructuring where the probation period
non-defaulted status with those that are paid in before the return to non-defaulted status, as
full by the obligor (and which do not return to specified in the GLs on default definition, is at
non-defaulted status because no outstanding least 1 year;
obligation remains).
(b) exposures with long maturities, and in
particular mortgage loans, as the value of
exposure that returns to non-defaulted status
would be a large part of the exposure at the
moment of default.

May create an incentive to repossess early

More commonly used in practice by the
instead of cooperating with the obligor to allow
industry than the alternative option (around
the cure (if the effect of discounting of the
32% of models compared with 4% for the
artificial cash flow would lead to loss higher
alternative, according to the results of IRB
than that expected as a result of collection
survey).
process).

173. The results shown in Figure 32 indicate that the proposal included in the CP on the GLs is
applied in only 4% of the LGD models (i.e. using the same methodology as for other
defaulted exposures but not discounting additional recovery cash flows at the date of the
return to non-defaulted status). The most common approach (applied in 43% of LGD models,
accounting for 52% of exposures under LGD models in the sample) is to assume that the
economic loss for cured cases is zero. This, however, may result from the different definitions
of cure that are currently used by institutions. This approach may also be imprudent, since
there may still be a considerable loss associated with some cured cases, in particular when
significant costs have been incurred to recover the collaterals. The application of the same
methodology as for other defaulted exposures where the additional cash flows are also
discounted is applied in 32% of LGD models (29% of exposures).

174. Based on a review of the pros and cons of both options, option (b) was chosen as a
compromise proposal. The GLs specify (in paragraph 135) that for exposures that return to
non-defaulted status, institutions should calculate economic loss as for all other defaulted
exposures, with the only difference being that additional recovery cash flows are added to

95
EBA REPORT ON THE IRB MODELLING PRACTICES

the calculation as if a payment was made by the obligor for the amount that was outstanding
at the date of the return to non-defaulted status, including any principal, interest and fees
(‘artificial cash flow’). The artificial cash flow should be discounted at the moment of default,
in the same manner as all observed cash flows.

5.5.2 Unpaid late fees and capitalised interest

175. The approach specified in the GLs (in paragraphs 137 and 138) is that unpaid late (after
default) fees and capitalised interest (after default) should not increase the amount of
economic loss or the amount outstanding at the moment of default 28.

176. However, in the CP on the GLs it was proposed that capitalised fees and interest after
default should be included only in the numerator of the realised LGD, i.e. they should be
added to the economic loss but the value of the outstanding obligation at default should
remain unchanged. Although the option proposed in the CP appears to be the most common
approach based on the survey results (used in 51% of models for late fees, and in 44% of
models for capitalised interest) (see Table 43), the feedback to the CP on the GLs showed
considerable disagreement with this proposal.
Table 43: Treatment of unpaid late fees and capitalised interest in the calculation of realised LGD

Capitalised interest (after

Unpaid late fees (after default)
default)
No. % % EAD No. % % EAD
Add to the outstanding amount at default
11 5 4 16 8 4
(denominator of realised LGD)
Both include in economic loss and add to
16 8 8 21 10 9
outstanding amount at default
Include only in the economic loss (numerator
104 51 50 88 44 49
of realised LGD)
Do not include 41 20 22 52 26 28
Other 30 15 15 25 12 10
Total 202 100 100 202 100 100

177. The following policy options have been considered:

(a) Capitalised fees and interest after default included only in the numerator of realised LGD
(proposal included in the CP): capitalised fees and interest after default are added to the
economic loss in the numerator of the realised LGD, but the value of outstanding obligation at
the moment of default remains unchanged. The underlying assumption is that fees and
interest after default have economic meaning similar to costs, and hence they are treated
similarly. Fees that are meant to cover costs already incurred by an institution are only
included in the calculation once (as costs).

28
Note that, for LGD in-default and ELBE estimation, the GLs specify (in paragraph 178) that institutions should include
all fees and interest capitalised before the reference date in the calculation of the realised LGD, and that they should
discount all subsequent cash flows and drawings at the reference date.

96
EBA REPORT ON THE IRB MODELLING PRACTICES

Pros Cons

Does not take into account the fact that

Does not require separate definitions of costs,
interest and fees are not related to real cash
fees and interest, as all these items are
flow from banks and are hence different from
included in the economic loss in the same way.
costs in this sense.

More prudent approach, which prevents

May be overly conservative, as the value of
underestimation of risk where the discounting
money in time is reflected twice, through
rate is lower than the actual charges by the
discounting effect and through including
institution, and negative realised LGDs are
interest and fees in economic loss.
largely prevented.

The most popular of the approaches currently

Approach frequently criticised by the industry
in use (43-51% of models, according to the IRB
as incorrect in the consultation process.
survey).

May be operationally burdensome to

implement, as information on fees and interest
capitalised after default would be required, and
may not be available for historical
observations.

May lead to different results depending on

applicable accounting standards (if different
rules on recognising profit from fees and
interest after default are applied).

(b) Only fees and interest before default included: fees and interest after default are not added
to the economic loss, but the recovery cash flows are still included. The underlying
assumption is that fees and interest after default have a different economic meaning to costs,
because they are not related to outgoing cash flow and hence should not increase economic
loss. In addition, it is assumed that the unrealised gains from fees and interest should not be
considered losses if they are unrelated to any expenses incurred by the institution. However,
recoveries on those items are profits realised by the institution, and hence can decrease the
economic loss.

Pros Cons

97
EBA REPORT ON THE IRB MODELLING PRACTICES

Takes into account non-cash character of fees Does not take into account the increases of the
and interest, and accurately reflects the exposure value based on accrued interest and
exposure value at the moment of default. fees.

Recognises additional gains related to

Does not take into account that capitalised fees
recoveries of interest and fees, in accordance
and interest increase CET1.
with final effect on CET1.

Negative realised LGDs are not a problem, as May lead to underestimation of risk where the
these will be floored to zero in accordance with discounting rate is lower than actual charges,
the relevant provision proposed in the GLs. and may lead to negative realised LGDs.

Independent of the accounting framework and

Consistency with Article 181(1)(i) of the CRR is
the applicable rules for capitalising fees and
based on the interpretation that unpaid late
interest, and of the principles for the order of
fees referred to in this article only refer to fees
allocation of payments (to fees, interest and
before default.
principal).

Operationally the easiest to implement, and

Approach not broadly used at present (20- 25%
does not require data on values of fees and
of models, according to the IRB survey).
interest capitalised after default.

(c) Capitalised fees and interest after default included in both numerator and denominator of
realised LGD: capitalised fees and interest after default are added to the economic loss and to
the exposure value at the moment of default; recovery cash flows are also still included. The
underlying assumption is that unrealised gains are considered losses. The treatment is
different from the treatment of costs, however, as costs do not increase the value of
exposure at the moment of default.

Pros Cons

May be operationally burdensome to

implement, as information on fees and interest
Full consistency with Article 181(1)(i) of the
capitalised after default would be required, and
CRR
may not be available for historical
observations.

Takes into account the increases of the May lead to different results depending on
exposure value based on accrued interest and applicable accounting standards (if different

98
EBA REPORT ON THE IRB MODELLING PRACTICES

fees. rules on recognising profit from fees and

interest after default are applied).

May lead to underestimation of risk where the

Negative realised LGDs are not a problem, as
discounting rate is lower than the actual
these will be floored to zero in accordance with
charges, and may lead to negative realised
the relevant provision proposed in the GLs.
LGDs.

Consistent with the treatment of interest and The least popular of the considered approaches
fees before default (which are also included in in the current practices (only 8-10% of models,
the exposure value at the moment of default). according to IRB survey).

178. Based on a review of the pros and cons of these options, option (b) was retained in the
final GLs, i.e. not including late (i.e. after default) fees and capitalised interest in the
numerator and denominator. This approach is currently applied in 20% (late fees) and 26%
(capitalised interest) of LGD models.

5.5.3 Additional drawings

179. The GLs specify (in paragraph 142) that additional drawings after default should be
included in the calculation of economic loss, and at the same time that all recoveries should
be taken into account, including those that relate to additional drawings. In addition, it is
specified that the treatment of additional drawings after default in the calculation of realised
LGD should be consistent with their treatment in the CCF estimation, to ensure meaningful
calculation of RWA. Where the estimation of CCF takes into account additional drawings
after default, therefore, these are also to be included in the denominator of the realised LGD,
i.e. institutions should increase the amount outstanding at the moment of default by the
amount of additional drawings by the obligor after the moment of default, discounted at the
moment of default (paragraph 140 of the GLs). Where additional drawings are not included
in the estimation of CCF, those additional drawings should not be included in the
denominator of the realised LGD (paragraph 141 of the GLs).
Table 44: Treatment of additional drawings after default in the calculation of realised LGD

No. % % EAD
Add to the outstanding amount at default (denominator of realised LGD) 22 11 10
Both include in economic loss and add to outstanding amount at default 30 15 16
Include only in the economic loss (numerator of realised LGD) 85 42 43
Do not include 29 14 15
Other 35 17 15
Total 201 100 100

99
EBA REPORT ON THE IRB MODELLING PRACTICES

180. As shown in Table 44, the results of the IRB survey confirm that, in the case of additional
drawings after default, inclusion in the economic loss without correcting the outstanding
amount at default in the denominator is the most popular solution. As described at the
beginning of this section, and in accordance with the GLs, this approach is proposed where
additional drawings after default are not included in CCF estimation. If CCF estimation
includes such drawings, it is proposed that the denominator of realised LGD should also be
corrected. The results of the survey should therefore be analysed taking into account the
approach to CCF calculation. This analysis is presented in Table 45.

Table 45: Are additional drawings after default included in the estimation of the CCF?

No. % % EAD
Yes 33 17 12
No 134 67 72
Other 32 16 16
Total 199 100 100

Table 46: Are additional drawings after default included in the calculation of realised LGD (non-defaulted)?

Additional drawings are included Additional drawings are not

in the CCF estimation included in the CCF estimation
No. % % EAD No. % % EAD
Add to the outstanding amount at
4 12 5 12 9 11
default (denominator of realised LGD)
Both include in economic loss and add to
20 61 92 9 7 7
outstanding amount at default
Include only in the economic loss
7 21 2 68 51 52
(numerator of realised LGD)
Do not include 0 0 0 28 21 21
Other 2 6 1 17 13 9
Total 33 100 100 134 100 100

181. These results further confirm that the approach presented in the GLs is the most common
approach, with the majority of institutions correcting both economic loss and outstanding
amount at default where CCF estimates include additional drawings after default (see Table
46). The sample of models where corresponding CCF estimates do not include additional
drawings after default is much larger, and therefore this is mostly reflected in the overall
results. The detailed split here confirms that the approach proposed in the GLs is already
incorporated into the majority of models.

5.5.4 Discounting rate

100
EBA REPORT ON THE IRB MODELLING PRACTICES

182. The discounting rate has been recognised as one of the major drivers of the lack of RWA
comparability across institutions, and for this reason specific guidance on this topic has been
included in the GLs.

183. As shown in Table 47, the average level of discounting rate in the RDS for the non-
defaulted LGD model is 6.54%. The average level of the discounting rate for LGD in-default
and ELBE estimation is similar, although the sample of models for which this variable has been
reported is much smaller. In general, great variability is observed across and within countries.
Whereas part of this variability is probably warranted because it is risk driven, the variance in
practices and the level of granularity with which the discount rate is specified logically leads
to non-risk-driven RWA variability.
Table 47: Average level of the discounting rate (%) in the RDS

N mean (%) mean (% EAD) min max

LGD non-defaulted 161 6.54 6.19 0.45 25.00
LGD in-default 133 6.58 6.41 0.45 25.00
ELBE 102 5.78 4.88 0.45 25.00

184. Table 48 29 further shows that part of these differences in the level of discounting rate are
driven by the different methodologies chosen to determine the discounting rate. In
particular, it can be seen that the average discounting rate is higher when the original and
current effective interest rate is used. For these approaches, one can infer from the 90th
percentile and the maximum values that these higher averages are driven by several high
discounting rates. When the risk-free rate plus add-on approach is used, one notices fewer
outlier values.
Table 48: Summary statistics of the discounting rate, differentiated by chosen methodology

mean
mean
N (% min p10 p50 p90 max
(%)
EAD)
Total 156 6.46 5.91 0.45 2.56 5.35 10.00 25.00
Risk-free rate + add-on 51 5.71 5.36 0.45 3.05 5.34 9.55 15.39
Original effective interest rate 20 8.98 5.36 2.46 4.38 5.60 18.19 22.29
Current effective interest rate 14 7.19 5.46 3.06 3.32 5.02 13.82 25.00
Funding rate 9 5.25 2.86 1.40 1.40 2.77 10.45 10.45
Funding rate + add-on 23 5.66 6.16 2.92 3.22 6.00 9.00 10.00
Other 39 6.62 7.56 0.67 2.30 7.30 10.00 17.23

29
Note that Table 51 on page 109 shows the levels of the discounting rate, where this rate is specified as the funding
rate or risk-free rate (plus add-on), separately for exposures to corporates, retail and non-retail.

101
EBA REPORT ON THE IRB MODELLING PRACTICES

185. The broad variety of practices therefore requires clear guidance on what should be
reflected by the discounting factor, and how it should be applied. In this regard, the following
options have been considered:
(a) Euribor (or comparable interbank rate in countries outside the eurozone) plus fixed add-on:
the add-on reflects the risk premium for the uncertainty related to the recoveries;

Pros Cons

Consistent with the current guidance Not risk sensitive — may not be accurate for
provided by CEBS and BCBS. some portfolios.
A proxy for the average funding cost of the
investor, in practice Euribor is available, as
May be understood as reflecting funding cost funding cost for prime institutions may
of potential investor (average of the underestimate the discounting rate (unless
market) + risk premium for uncertainty of addressed in an add-on); however, if
cash flows at the moment of default. understood as risk-free rate + add-on then
Euribor may be considered a proxy for risk-
free rate.
Ensures simplicity, comparability and Three-month Euribor rate not liquid in small
independence from the own-credit standing countries outside eurozone; other or shorter
of the institution. term rates have to be used instead.
Independent from the funding structure and
credit standing of the institution, which may
Fixed add-on may require future revisions.
be seen as favourable since it is an anti-
cyclical model component.

(b) funding cost plus add-on: discounting factor reflects the funding costs of the institution and an
appropriate risk premium reflecting the uncertainties associated with the receipt of recoveries
with respect to a defaulted exposure;

Pros Cons

Depends on the funding structure of an

institution, and may therefore be more prone
Consistent with the current guidance
to heterogeneity in the assessment of the
provided by CEBS and BCBS.
correct funding cost + add-on, which could
generate additional RWA variability.
Reflects own-credit standing of the
institution, and may therefore be overly
Aims at reflecting the shareholders’ loss
penalising for some institutions and
experienced by the institution.
jurisdictions, and may induce pro-cyclical
capital requirements.

102
EBA REPORT ON THE IRB MODELLING PRACTICES

Pros Cons

Fixed add-on may require future revisions.

(c) original effective interest rate: discounting factor is derived from facility-specific interest rates.

Pros Cons

Alignment with accounting standards not

Consistency with international accounting
necessary, as the calculation of IRB
standards – comparability between expected
shortfall/excess aims at eliminating the effect
loss and provisions.
of provisions.
If the aim is to reflect the finding cost of a
Reflects funding cost (of a performing and not
defaulted asset, an add-on would be
defaulted exposure) + risk premium + profit
necessary.
Lack of comparability of losses within and
across institutions — rates may vary
Highly risk sensitive — exposure specific.
significantly between the types of obligor and
types of products.
Depends on pricing policy and marketing
strategies of an institution, which should not
necessarily affect capital requirements, and
which may lead to non-risk-driven RWA
variability.
May be overly conservative, as it includes a
margin for profit.
Substantial complexity.
May not be adequate to the market and
economic conditions at the time of default.

186. In the GLs, option (a) was chosen as a compromise solution. In particular, the GLs specify
(in paragraph 143) that the annual discounting rate should be composed of a primary
interbank offered rate, applicable at the moment of default and increased by an add-on of
five percentage points. The primary interbank offered rate should be considered as the
three-month Euribor, or a comparable liquid interest rate in the currency of the exposure.
Given the current average level of the discounting rate identified for the models surveyed
(6%), and the current low interest rate environment, the add-on of 5% is not expected,
across institutions, to cause major cliff effects in LGD calculations.

103
EBA REPORT ON THE IRB MODELLING PRACTICES

187. The responses to the IRB survey confirm, furthermore, that the risk-free rate plus add-on
is the most common approach, and this approach is closest to that specified in the GLs. As
shown in Figure 33, risk-free rate plus add-on is currently applied in around 30% of LGD
models (37% of exposures under LGD models). For models with retail exposures or exposures
to corporates, these shares are even higher: risk-free rate plus add-on is used in 39% of
exposures to retail models, and 48% of exposures to corporates.

188. Therefore, the option of specifying the discounting rate as the Euribor (or a comparable
interbank rate in countries outside the eurozone) plus fixed add-on will rule out a significant
share of variability in practices, and it is also the policy that will lead to the fewest number of
model changes. It will, however, require a change in practice in around 70% of LGD models.
Figure 33: Methodologies used to determine the discounting rate (LGD non-defaulted) — retail, corporate,
institutions, and central governments and central banks
Corporate

% EAD 48% 8% 11% 7% 3% 24%

% 38% 14% 6% 5% 5% 32%

% EAD 39% 7% 12% 13% 3% 27%

Retail

% 32% 14% 12% 9% 6% 27%

banks Institutions

% EAD 24% 9% 8% 18% 12% 30%

% 7% 14% 29% 7% 14% 29%

governmen

% EAD 0% 24% 10%0% 66%

Central

central
ts and

% 0% 17% 17% 0% 67%

% EAD 37% 7% 13% 12% 2% 28%

models
Total #

% 30% 14% 13% 9% 5% 31%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Risk free rate + add-on Funding rate + add-on

Original effective interest rate Current effective interest rate
Funding rate Other

189. These results are generally consistent with those for the LGD in-default models (shown in
Figure 34), where the use of risk-free rate plus add-on represents 27% of LGD models, or 37%
of exposure values. In ELBE models, the use of current effective interest rates is more popular,
and is applied in 17% of ELBE models, or 27% of exposures under ELBE models.

104
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 34: Methodologies used to determine the discounting rate (LGD in-default)

15%
Current effective interest rate
12%
37% 27% 3% 1% Funding rate
7%
14% Funding rate + add-on
8% Original effective interest rate
11% Other
33%
Risk free rate + add-on
33%

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

Figure 35: Methodologies used to determine the discounting rate (ELBE)

Current effective interest rate

17% 27%
33% 26% Funding rate
2%
Funding rate + add-on
10%
1% Original effective interest rate
4%
13% Other
32% 13% Risk free rate + add-on
23%

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

190. Whereas the draft GLs proposed a level of 5% for the add-on, possible increased
granularity of an add-on between retail and non-retail exposures has been considered, as
suggested by some respondents to the draft GLs. In particular, respondents in the
consultation process suggested increasing the discount factor for retail (higher uncertainty)
and decreasing it for non-retail.

191. Table 49 shows the distribution of the levels of the discounting rate across the exposure
classes. Overall, it is difficult to identify a significantly different level of the discounting rate
between retail and non-retail exposure classes. It should be kept in mind, however, that the
classification across COREP exposure classes in Table 49 is not mutually exclusive, i.e.
institutions could indicate the applicable exposure classes in their LGD models. This being the

105
EBA REPORT ON THE IRB MODELLING PRACTICES

case, some of the LGD models in the IRB survey sample apply to both retail and non-retail
(for more details, see Table 8).

Table 49: Average level of the discounting rate (%), by COREP exposure class

N mean (%) mean (% EAD) min max median

Total 161 6.54 6.19 0.45 25.00 5.40
Central governments and central banks 3 7.33 9.61 5.00 11.00 6.00
Institutions 11 5.65 6.86 2.02 11.00 5.00
Corporate — SME 49 5.43 4.84 0.45 15.39 5.00
Corporate — specialised lending 7 6.08 4.36 3.84 10.00 5.00
Corporate – other 48 5.43 5.15 0.45 15.39 5.00
Retail — secured by immovable property SME 41 6.00 4.85 0.45 15.85 5.00
Retail — secured by immovable property non-SME 92 5.84 6.07 0.45 15.85 5.14
Retail — qualifying revolving 26 6.50 4.33 0.45 22.29 5.00
Retail — other SME 45 6.33 4.54 0.45 25.00 5.00
Retail — other non-SME 61 5.91 4.08 0.45 25.00 4.48

192. Table 50 therefore shows the average discounting rate for those models that cover only
retail, only corporate or only non-retail exposures (i.e. exposures to corporates, institutions,
or central governments and central banks). These data indicate levels of discounting rate for
retail exposures higher on average than those for all other exposure classes. However, the
difference between the mean and the median already suggests that this may be driven by
some outlier observations. Further analysis on the approach used to determine these
discounting rates suggests that these higher discounting rates stem largely from the
application of (current or original) effective interest rate as a discounting rate.
Table 50: Average level of the discounting rate (%), for models with exposures only to corporates, retail and non-
retail

N mean (%) mean (% EAD) min max p50

Corporate 28 5.68 5.43 2.09 11.00 5.11
Retail 101 7.05 6.46 0.67 25.00 6.00
Non-retail 36 6.11 6.53 2.09 11.00 5.27

193. This effect is presented also in Table 51, where the models using original or current
effective interest rate were eliminated. As a result, the extreme observations were
eliminated and the large difference in discounting rate between retail and non-retail
exposures disappears.
Table 51: Average level of the discounting rate (%), for models with exposures only to corporates, retail and non-
retail and where the discounting rate is specified as funding rate or risk-free rate plus add-on

mean mean
N min max p50
(%) (% EAD)

106
EBA REPORT ON THE IRB MODELLING PRACTICES

mean mean
N min max p50
(%) (% EAD)
Corporate if funding rate or risk-free rate (+ add-on) 15 5.42 5.46 2.09 10.00 5.01
Retail if funding rate or risk-free rate (+ add-on) 47 6.00 5.99 1.80 10.45 5.42
Non-retail if funding rate or risk-free rate (+ add-
16 5.40 5.44 2.09 10.00 5.01
on)

194. Although the results of the survey show discounting rates higher on average for retail
exposures, the difference may not be significant enough to justify differentiation of the add-
on.

In addition to the methodology used to specify the discounting rate, the IRB survey further
revealed significant variation in practices with respect to the level of granularity of the
discounting rate, i.e. whether the discounting rate is specified at institutional level, at portfolio
level, by product type or at single-exposure level (see Figure 36). Given the decision to retain the
specification of the discounting rate as the risk-free rate plus add-on at the moment of default,
this will entail a change in practice for the majority of models, since currently only 30% of
models use a specific discounting rate at the level of the single exposure, and this even includes
models where the discounting rate is specified as the original or current effective interest rate,
where the discounting rate is by definition exposure specific. These results are broadly
consistent for the LGD in-default models (as shown in Figure 37).

107
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 36: Level of granularity at which the discounting rate is specified (LGD non-defaulted) — retail, corporate,
institutions, and central governments and central banks
Corporate

% EAD 22% 20% 4% 26% 27%

% 24% 21% 3% 30% 22%

% EAD 16% 37% 4% 30% 13%

Retail

% 26% 23% 8% 30% 14%

banks Institutions

% EAD 28% 5%2% 20% 45%

% 25% 6% 6% 31% 31%

governmen

% EAD 50% 18% 9% 0% 23%

Central

central
ts and

% 14% 29% 29% 0% 29%

% EAD 19% 32% 3% 30% 15%

Total

% 25% 22% 6% 32% 15%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

On institution level On portfolio level By product type On single exposure level Other

Figure 37: Level of granularity at which the discounting rate is specified (LGD in-default)

13% 3%
12% 6% 22% By product type

27% On institution level

27%33% On portfolio level

On single exposure level
22% Other
36%

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

As shown in Figure 38, it is most common to specify the discounting rate at the single-exposure
level in ELBE. As reported in Figure 35, the majority of the ELBE models use the original or current
effective interest rate, which is by definition exposure specific.

108
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 38: Level of granularity at which the discounting rate is specified (ELBE)

13% 3%
14%
13% 7%
By product type
24% On institution level
On portfolio level

39% On single exposure level

39% 31%
18% Other

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

5.5.5 Direct and indirect costs

195. The GLs specify (in paragraph 144) that all material direct and indirect costs should be
taken into account in the calculation of realised LGD. In addition, a clear description of direct
and indirect costs is provided in the GLs (paragraphs 145 and 146). It is specified that all
direct costs should be considered as material and, for the indirect costs, a wide definition is
provided.

196. The responses to the IRB survey confirm the need for guidance on this aspect, in
particular for the inclusion of direct costs, where these costs are included in less than half of
the models (see Figure 39). Indirect costs are not included in around 75% of LGD models
(85% of exposures). However, the CRR definition in Article 5(2) specifies loss as ‘economic
loss, including material discount effects, and material direct and indirect costs associated
with collecting on the instrument’. Therefore, this requirement has been strengthened in the
GLs, since the GLs specify that all direct costs should be included, rather than material direct
costs only. A change in practice will be required for 50% (inclusion of direct costs) and 75%
(inclusion of indirect costs) of LGD models.

109
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 39: Are direct costs incurred before default included in the calculation of the realised (non-defaulted) LGD?

No
51%
48%48% Partially
51%
Yes
1%
0%

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

Figure 40: Are indirect costs incurred before default included in the calculation of the realised (non-defaulted) LGD?

15%
0%25%
No
1% Partially
Yes
74%
85%

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

5.6 Long-run average LGD

5.6.1 Historical observation period
Table 52: Length of the historical observation period (years)

mean mean (%
N min p10 p50 p90 max
(%) EAD)
Long-run average
158 11.35 10.88 2.33 6.00 10.51 18.91 32.77
LGD
LGD non-defaulted 165 10.54 10.99 2.00 5.00 9.75 17.51 32.77
LGD in-default 149 10.64 11.16 1.08 5.00 10.01 17.01 32.77
ELBE 123 9.92 9.13 0.00 4.33 9.02 16.09 32.77

197. It should be noted that the length of the historical observation period is not available for
all LGD models (202). Several institutions mentioned that the reason for this is that the long-
run average LGD is not calculated, although this would appear to contradict the CRR

110
EBA REPORT ON THE IRB MODELLING PRACTICES

requirements. Some of the other explanations are (i) that no defaults have occurred in the
historical database and that therefore the long-run average LGD is not estimated, and (ii)
that no internal data are available for the estimation of unsecured LGD.

198. It should also be noted that the length of the historical observation period for the
estimation of the long-run average, LGD, LGD in-default and ELBE estimates is shorter than
five years for some models, which would appear to contradict the CRR requirements in
Article 180(1)(j) and 181(2) subparagraph 2 of the CRR. For some institutions, explanations
could be found in the comments. The most common explanation is that the institutions
specified the length of the development sample instead of the length of the historical
observation period (this misunderstanding also applies to the specification of the historical
observation period for PD estimation as mentioned in paragraph 100). Another explanation
given is that the institution reviews the performance of each component of the model over
time, and picks the most conservative period for that component.

199. It should be emphasised, however, that explanations justifying a length of the historical
observation period shorter than five years are not in line with the CRR.

200. The use of non-consecutive time periods in the historical observation period is
exceptional; only one institution uses non-consecutive periods for the calculation of the long-
run average LGD, LGD and LGD in-default estimation 30. In this regard, the GLs specify (in
paragraph 147(c)) that the historical observation period should be composed of consecutive
periods, hence this would require a change in practice for this institution.

201. Few institutions specified the length of the historical observation period stemming from
external data (16-17 models for the long-run average and the LGD non-defaulted models),
but its average length (around 15 years) is longer than that based on internal data (on
average, around 10-11 years). For most models where the use of external data is specified,
these external data are used in addition to the internal data. This is not the case in a few
instances, where (i) the model applies only to sovereign exposures, or (ii) the institution
specified that the long-run average LGD is not calculated. In this regard, the GLs specify (in
paragraph 102) that LGD estimates should not be exclusively based on external data. This will
therefore require a change in practice in the models mentioned above, which rely solely on
external data in the historical observation period.

202. The use of pooled data seldom occurs; only around 6-7 models specified the length of the
historical observation period for data stemming from pooled data for the long-run average
LGD and LGD in-default estimation. The length of the historical observation period based on
these pooled data is on average around 15 years.

203. Table 53 shows the length of the historical observation period of LGD non-defaulted
(based on internal data), expressed in years and across exposure classes.

30
This institution is not included in the figures below visualising the historical observation period.

111
EBA REPORT ON THE IRB MODELLING PRACTICES

Table 53: Length of the historical observation period for LGD non-defaulted, expressed in years (internal data) and by
exposure class

mean
mean
N (% min p10 p50 p90 max
(%)
EAD)
Total 165 10.54 10.99 2.00 5.00 9.75 17.51 32.77
Central governments and central banks 1 5.00 5.00 5.00 5.00 5.00 5.00 5.00
Institutions 10 10.45 12.98 5.00 5.00 10.51 16.22 19.01
Corporate — SME 42 12.04 12.35 2.00 5.00 10.67 19.01 32.77
Corporate — specialised lending 5 10.87 14.14 4.00 4.00 11.34 19.01 19.01
Corporate — other 44 12.47 12.84 2.00 5.00 10.92 19.01 32.77
Retail — secured by immovable property
43 10.93 10.27 2.00 4.59 10.84 19.01 32.77
SME
Retail — secured by immovable property
91 11.10 10.79 2.00 5.00 10.50 17.01 32.77
non-SME
Retail — qualifying revolving 25 10.51 9.71 6.00 7.01 10.59 15.88 19.01
Retail — other SME 40 10.77 10.12 3.00 5.00 10.30 18.01 32.77
Retail — other non-SME 62 10.46 9.91 2.00 5.00 9.00 16.09 32.77

Figure 41: Historical observation period for LGD non-defaulted — retail exposures secured by immovable property
SME (internal data)
01/04/1984

22/09/1989

15/03/1995

04/09/2000

25/02/2006

18/08/2011

07/02/2017

31/07/2022

112
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 42: Historical observation period for LGD non-defaulted — retail exposures secured by immovable property
non-SME (internal data)
01/04/1984

22/09/1989

15/03/1995

04/09/2000

25/02/2006

18/08/2011

07/02/2017

31/07/2022

113
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 43: Historical observation period for LGD non-defaulted — corporate exposures (SME, specialised lending and
corporate other) (internal data)
01/01/1979

23/06/1984

14/12/1989

06/06/1995

26/11/2000

19/05/2006

09/11/2011

01/05/2017

22/10/2022
Figure 44: Historical observation period for LGD non-defaulted — retail other non-SME and qualifying revolving
(internal data)

114
EBA REPORT ON THE IRB MODELLING PRACTICES

01/04/1984

22/09/1989

15/03/1995

04/09/2000

25/02/2006

18/08/2011

07/02/2017

204. With respect to the specification of the historical observation period, the GLs specify (in 31/07/2022
paragraph 147(e)) that all available internal data should be considered ‘relevant’ (in relation
to Article 181(1)(j) and 181(2) subparagraph 2 of the CRR) and should be included in the
historical observation period. On this aspect, the survey enquired whether or not institutions
discard some of the available historical data for the historical observation period. The results
show that some historical data is discarded in around 46% of LGD models. This would appear
to contradict the CRR, which specifies that all defaults should be used for the purpose of
estimating the LGDs, in Article 181(1)(a) of the CRR. However, as also noted in
paragraphs 201 and 202, some institutions understood the historical observation period as
the period used for model development. This may explain the large share of data exclusions
from the historical observation period.

Table 54: Did you exclude some of the available historical data from the specification of the historical observation
period?

LGD non-defaulted LGD in-default ELBE

No. % % EAD No. % % EAD No. % % EAD
No 107 54 62 84 52 57 73 53 61

115
EBA REPORT ON THE IRB MODELLING PRACTICES

LGD non-defaulted LGD in-default ELBE

Yes 91 46 38 77 48 43 64 47 39
Total 198 100 100 161 100 100 137 100 100

5.6.2 Calculation of long-run average LGD

205. With respect to the level at which the long-run average LGD should be calculated, the GLs
specify (in paragraph 149) that the long-run average LGD should be calculated separately for
each grade or pool, as well as at the level of the portfolio covered by the LGD model. On this
aspect, the responses to the IRB survey show that this is currently the case only in 11% of all
LGD models (covering 14% of exposures). Most often (in 43% of LGD models and 42% of
exposures), the institution calculates the long-run average LGD only at grade or pool level, in
line with the requirement in Article 181(1)(a) of the CRR. In 32% of LGD models, the long-run
average LGD is calculated only at portfolio level, which would not appear to be in line with
the CRR requirement. It follows that a change in practice will be required for the majority of
LGD models, although these changes stem not only from the policy prescribed in the GLs
(which accounts for a change in practice in around 43% of LGD models), but also from a
correct interpretation of the CRR (accounting for a change in practice in around 32% of LGD
models).
Figure 45: Level at which the long-run average LGD is calculated

14%
Both at portfolio level and at
11%
30% grade/pool level
32% Level of grade or pool

Other
43%
14% 42% Portfolio level (scope of LGD
14%
model)

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value

206. With respect to the calculation of the long-run average LGD, it is clarified in the GLs (in
paragraph 150) that the long-run average LGD should be calculated as an arithmetic average
of realised LGDs over an historical observation period weighted by the number of defaults. In
line with the CRR requirement in Article 181(2), institutions are allowed to give a higher
weight to more recent data for retail exposures, if they demonstrate in a documented
manner that the use of higher weights for more recent data is justified by better prediction

116
EBA REPORT ON THE IRB MODELLING PRACTICES

of loss rates (paragraph 151 of the GLs). Based on the responses from the IRB survey (Figure
46), it is known that the option to weight all defaults equally is most common, and is used in
72% of models across all exposure classes. In 23% of LGD models, the long-run average
calculation is weighted based on the exposure value. In only one (of out 195 LGD models), it
was indicated that a higher weight is given to more recent data.

207. Across exposure classes, however, some differences may be noted. In particular,
weighting by exposure value is more popular in the retail models, where this method is
applied on average in 28% of all LGD models, as compared with only 8% in LGD models for
exposures to corporates.
Figure 46: Type of weighting used in the calculation of the long-run average LGD — retail, corporate, institutions, and
central governments and central banks
Corporat

% EAD 93% 5%
0%
3%
e

% 86% 8%0%6%
% EAD 69% 27% 3%
0%
Retail

% 67% 28% 1%5%

Total # central Institutio

% EAD 88% 0% 11%

1%
ns

% 87% 7%0%7%
governm
ents and

% EAD 90% 10% 0%

Central

models banks

% 71% 29% 0%
% EAD 76% 20% 4%
0%
% 72% 23% 1%5%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

All defaults are weighted equally Based on the exposure value

More recent defaults have higher weights Other

208. The GLs (in paragraph 150) further disallow that the long-run average LGD is calculated on
a subset of observations, for instance as the average of yearly LGDs. This policy is necessary
and warranted as explained in paragraph 227 of this report in section 5.6.3, ‘Treatment of
incomplete recovery processes’.

5.6.3 Treatment of incomplete recovery processes

209. In the context of the calculation of long-run average LGD, differences in the treatment of
incomplete recovery processes have been identified as one of the main sources of unjustified

117
EBA REPORT ON THE IRB MODELLING PRACTICES

variability of LGD estimates 31. Detailed requirements in that regard are therefore included in
the GLs (paragraphs 153-159).

210. In particular, the GLs specify (in paragraph 156) that institutions should define the
maximum average expected period of the recovery process for a given type of exposures,
and during which the institution realises the vast majority of the recoveries. All exposures
that remain in defaulted status for a period of time longer than the maximum period of the
recovery process should be treated as closed for the purpose of the calculation of the
observed average LGD, and institutions should calculate the ‘observed average LGD’ taking
into account realised LGDs only on those defaults that are related to closed recovery
processes and those that reached a certain threshold in terms of the time in-default, both at
grade or pool, and at portfolio level.

211. In addition, the GLs specify (in paragraph 158) that institutions should obtain the long-run
average LGD by adjusting the observed average LGD, taking into account the information
related to incomplete recovery processes, and the estimated future costs and recoveries on
these exposures 32. The adjustment can be estimated at the level of the single exposure,
grade or pool, or LGD model (paragraph 159(e) of the GLs).

212. The following pros and cons have been identified for this policy option:
 pro: addresses the problem that institutions cannot present reliable estimates for
further periods due to insufficient data;
 pro: where the collateral has not been realised within the specified period, it may
indicate some problems with the collateral that could prevent its realisation;
 con: less flexible and hence in some cases less accurate, as it is not possible to include
future expected cash flows even if there is high probability they will be realised (but if
the maximum period is defined appropriately this inaccuracy should not be
significant);
 con: cash flows from collaterals, if they exist, are usually more significant than other
cash flows, and at advance stages of recovery processes can be predicted on an
individual basis.

213. As an alternative to this policy option, the EBA has considered allowing institutions to
estimate future recoveries for periods beyond the maximum length of the recovery process
only if these recoveries will stem from the realisation of the existing collaterals. The following
pros and cons have been identified for this policy option:

31
EBA, Third interim report on the consistency of risk-weighted assets, 2013, p. 91
(https://www.eba.europa.eu/documents/10180/15947/20131217+Third+interim+report+on+the+consistency+of+risk-
weighted+assets+-+SME+and+residential+mortgages.pdf).
32
Note that any recoveries realised after the moment of default should be included in the calculation of the economic
loss for the purpose of obtaining the realised LGD for each exposure (as specified in paragraph 133 of the GLs under the
section on the definition of economic loss).

118
EBA REPORT ON THE IRB MODELLING PRACTICES

 pro: more accurate in some cases, especially where individual case-by-case

assessment is applied and there is high probability that the recovery on a given
exposure will be realised;
 con: often not enough data to estimate recoveries for further periods;
 con: less strict approach that allows more subjective and less comparable estimates,
and may lead to disregarding in practice the effect of the maximum length of the
recovery process.

214. To assess current practices on incomplete recovery processes, institutions were asked to
indicate for each LGD model how incomplete recovery processes are incorporated into the
LGD estimates: (i) incomplete recovery processes are not included; (ii) only with recoveries
realised so far; (iii) with recoveries realised so far and estimated future recoveries; (iv) as an
adjustment at grade or pool level; (v) as an adjustment at portfolio level; or (vi) any other
treatment. The results of the IRB survey confirm the existence of various practices with
regard to the treatment of incomplete recovery processes in the LGD estimation.
Table 55: How are incomplete recovery processes incorporated into the LGD estimation?

LGD non-defaulted LGD in-default ELBE

% % %
No. % No. % No. %
EAD EAD EAD
As an adjustment at portfolio level 7 3 5 7 4 6 6 4 6
As an adjustment at grade or pool level 1 1 1 0 0 0 0 0 0
Incomplete recovery processes are not
59 29 28 49 30 27 36 25 20
included
Only with recoveries realised so far 38 19 10 35 21 11 34 24 12
With recoveries realised so far and
78 39 44 52 32 39 51 36 49
estimated future recoveries
Other 16 8 8 19 12 13 15 11 13
Not applicable 3 1 4 2 1 4 0 0 0
Total 202 100 100 164 100 100 142 100 100

215. The most frequent approach is consistent with the GLs, i.e. that incomplete recovery
processes should be included in the estimation with recoveries realised until the moment of
estimation, as well as estimated future recoveries (39% of LGD non-defaulted models and
covering 44% of exposure values). Models where only recoveries realised so far are included
in the estimation, however, represent around 20% of all LGD models, and models where
incomplete recovery processes are not included represent around 30% of all LGD models. For
these models, therefore, the requirement to estimate also future recoveries will require a
change in modelling practices.

216. The response ‘other’ was selected in several models. The main reasons mentioned were
(i) that a different approach is used for secured versus unsecured exposures; (ii) that
incomplete cases are forced as cure or repossession; and (iii) that whether or not incomplete
recovery processes are included is dependent on the estimation and calibration step. One

119
EBA REPORT ON THE IRB MODELLING PRACTICES

institution mentioned that incomplete recovery processes are included in the estimation only
when these are provisioned. After reviewing the responses, the additional category ‘not
applicable’ has been created, because some respondents mentioned that there are no
incomplete recovery processes (because they use external data or because all historical
recovery processes are closed).

217. However, across countries and institutions a wide variety of practices can be observed,
which indicates the need for harmonisation. Balancing the pros and cons of the policy
included in the GLs, and given the confirmation that this policy option is the most common
approach currently used by institutions, this policy decision has been included in the final GLs
(in paragraphs 158(b) and 159).

218. The survey further investigated approaches with regard to the specification of the
maximum length of the recovery process. According to the results presented in Figure 47,
less than half of the models currently incorporate a specification of the maximum length of
the recovery process, with a much smaller share in the case of models for low default
portfolios such as central governments and institutions (only six models). For other portfolios
the split is more or less equal, indicating the need for harmonisation. The requirement in the
GLs to set a maximum period after which incomplete recovery processes are closed will
therefore entail a change in modelling practices for more than half of the LGD models.

120
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 47: Is a maximum period defined after which incomplete recovery processes are treated as closed for the
purpose of the average realised LGD? By COREP exposure class

% EAD
Retail -
67% 33%
Corpor Speciali Corpor able proper ing Retail - Other
central Institut ate - sed ate - proper ty non- revolvi Other non-
SME SME % 52% 48%
% EAD 68% 32%
% 55% 45%
immov able Qualify

% EAD
d by immov Retail -

76% 24%
ng

% 52% 48%
% EAD
Retail -

54% 46%
Retail - secure

Total banks ions SME lending Other ty SME SME

secure d by

% 59% 41%
% EAD 61% 39%
% 52% 48%
% EAD 69% 31%
% 64% 36%
% EAD
Corpor

81% 19%
ate -

% 43% 57%
% EAD 64% 36%
% 58% 42%
% EAD 94% 6%
% 87% 13%
Central

% EAD
govern

100% 0%
ments
and

% 100% 0%
% EAD 57.75% 42.25%
% 57.73% 42.27%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

No Yes

219. Figure 48 shows, for those models where a maximum period of the recovery process is
specified, how this maximum is determined. For around 85% of models (70 LGD models), a
maximum time period is set, whereas in 7% of models a more complex rule is used. The
following rules were mentioned: (i) incomplete work-outs that are treated as closed are
those positions fully provisioned, or whose vintage is greater than 10 years without mortgage
or bankruptcy procedures, or whose vintage is greater than 10 years, with mortgage or
bankruptcy but a coverage > 90%; (ii) incomplete work-outs that are treated as closed are
those where the recovery process takes more than five years and the last payment was more
than two years ago; (iii) incomplete work-out processes are treated as closed if the vintage is
greater than eight years and they have a provision coverage > 80%; and (iv) recovery
processes are treated as closed at the point where, on average, 90% of the exposure value is
recovered.

121
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 48: Where a maximum period for the recovery process is specified, how is this defined?

7%
7% 7%
7%

85%
86%

Time Complex rule Not specified

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

220. Based on those models where the maximum period is specified on a time dimension, the
maximum period specified for the recovery process is five years and eight months (average
based on 70 models). However, a wide variety can be observed across countries.

221. Regarding the average time of recovery processes and, related to it, the share of
incomplete recovery processes in the observations included in the RDS, significant
differences are observed between countries, which may be related to different legal
environments, among other things. The average time of the recovery process is two years
and eight months across all LGD models in the sample, but this average hides conservable
heterogeneity, with a maximum of 20 years and 5 months. It should be noted, however, that
the definition of a closed case was not harmonised at the time the survey was conducted, in
particular with respect to cases where the collateral has been recollected but not sold. The
share of incomplete recovery process in the LGD model is 20% on average across all EU
countries.
Table 56: Average time of the recovery process in the RDS (expressed in months) and average share of incomplete
recovery processes (calculated in terms of the number of defaulted exposures) regarding all defaults occurring during
the historical observation period (LGD non-defaulted, internal data)

Average time of the recovery process Share of incomplete recovery processes

mean mean
mean mean
N (% min max N (% min max
(%) (%)
EAD) EAD)
Total 164 32.76 29.20 2.00 245.00 156 20.51 21.43 0.00 100.00

222. The GLs specify, in relation to the treatment of incomplete recovery processes, how the
long-run average LGD should be calculated, i.e. based on all defaults observed during the
historical observation period, weighted by the number of defaults and by adjusting the
observed average LGD, taking into account the estimated future recoveries (paragraphs 150
and 158-159). The responses to the IRB survey show a variety of practices, but confirm that

122
EBA REPORT ON THE IRB MODELLING PRACTICES

the most common approach is to calculate the long-run average LGD based on the realised
LGD of all defaults (applied in 45% of all models and 43% of all exposures).

223. The calculation based on the average of the realised LGD of all defaults is used most often
(45% of all LGD models), followed by the average of the realised LGD of all closed defaults
(27%). The calculation based on the average of the annual average of the realised LGD of all
defaults is applied in nearly 3% of all LGD models. The option to consider all defaults (not
only closed but also some open cases) is more common (48%) than considering closed
defaults only (28%).

224. Furthermore, the methodology used to calculate the average of the annual average of the
realised LGD (based on all cases or on closed cases only) is rare (4%) in comparison with the
methodology used to calculate the long-run average as the average of the realised LGDs
(based on all cases or on closed cases only) (72%). This finding supports the policy in the GLs
(in paragraph 150) that specifies that institutions should not use any averages of LGDs
calculated on a subset of observations, and in particular any yearly average LGDs, unless they
use this method to reflect higher weights of more recent data on retail exposures, in
accordance with Article 181(2) of the CRR.

225. The option to calculate the long-run average LGD based on intermediate averages is only
used in 13 LGD models (7%), where institutions have given a variety of explanations: (i) one
institution mentions that long-run average LGD is calculated as the quarterly average of the
realised LGD of all defaults that occurred in each quarter; (ii) another institution mentions
that an exposure-weighted average LGD is calculated based on all defaults that occurred in
that month, and that the long-run average LGD is calculated as the default-weighted average
of these monthly averages of LGD; (iii) another institution mentions that long-run averages
are calculated at model component level; and (iv) another institution mentioned that the
long-run average LGD is calculated with reference to the underlying scenario.

123
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 49: Method used to calculate the long-run average LGD

Average of the annual average

of the r-LGD of all defaults
5% 0%

23% 3% Average of the annual average

1%
18% of the r-LGD of the closed
defaults
Average of the realised LGD of
7% all defaults
45% 43%
8%
Average of the realised LGD of
the closed defaults
27%
Long-run average LGD based on
22% intermediate averages (please
specify)
Other

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

226. The significant share of models for which the category ‘other’ was selected (18% of
models or 36 LGD models) indicates the general confusion around the different possible
approaches to calculating the long-run average LGD. The most common explanation given for
those ‘other’ models (in 9 of these 36) is that no long-run average LGD is calculated, which
would not be in line with the CRR and the GLs.

5.6.4 Treatment of cases with no loss or positive outcome

227. The GLs prescribe (in paragraph 160) that wherever institutions realise profit on their
observations of defaults, the realised LGD on these observations should equal zero for the
purpose of calculating the observed average LGD and the estimation of long-run average
LGD. The evidence collected in the IRB survey supports this requirement, and shows that a
zero-floor at the level of the individual default observation is applied in 73% of LGD models.

124
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 50: How are cases with no loss or positive outcome treated?

Zero-floor at the level of

17% individual default observation
12%
Zero-floor at the level of risk
6% parameter
9% 2%
8% Zero-floor at the level of pool or
1% grade
7% No floor
73% 66%

Other

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

228. When this requirement was considered, the following two options have been considered:

(a) Option 1: 0% floor on LGD estimates. This would be the continuation of policy applied
in GL 10 i.e. no obligatory floor at the level of individual realised LGD. Netting of gains
and losses would be allowed, but only where this is consistent with the business model
of the institution (subject to the assessment of the competent authority).

 Pro: no change in modelling practices.

 Con: application of netting between gains and losses on various exposures would only
have an effect on models based on grades or pools, whereas for many LGD models
that are estimated based on a continuous rating scale, this effect would not be
reflected. This would contribute to non-comparability of the estimates.

 Con: according to the definition of LGD included in point (55) of Article 4(1) of the
CRR, this parameter should measure potential losses but not gains on defaulted
exposures, and therefore this asymmetry is already included in the LGD definition.
The concept of capital requirements is to cover potential losses but not gains, and
therefore the calculation should be focused on losses.

 Con: gains achieved in a particular period of time may not be available to cover losses
experienced in a different period, and hence the netting over the observation period
would not be appropriate. In particular in the case of estimates sensitive to economic
conditions, the losses over the bad years would be compensated by gains over the
good years.

125
EBA REPORT ON THE IRB MODELLING PRACTICES

 Con: the EBA GLs will be subject to a ‘complain or explain’ mechanism (which was not
the case for GL 10). The repetition of requirements of GL 10 in a stronger legal tool is
expected to lead to a decrease of LGD estimates in the majority of cases, which would
not be prudent.

(b) Option 2: 0% floor on realised LGD for the computation of the long-run average. This
was the option specified in the CP on the GLs, and which broadly reflects current
practice. In addition, it would be clarified that while realised LGDs floored to zero
would have to be used in the calculation of long-run average LGD, institutions would
be allowed to use any information available in the model development (risk
differentiation). MoC for general estimation error would also be calculated on the
basis of full distribution before the application of floor.

 Pro: broadly consistent with current practices.

 Pro: harmonisation of practices and hence reduction in undue RWA variability.

 Con: the estimation of the risk parameters at pool or grade level is key in the IRB
approach; therefore the risk parameter estimates should reflect the actual economic
loss of these grades or pools. Where gain is achieved on some exposures and loss on
others, this effect is netted at the portfolio level. Elimination of these netting effects
would not be justified if that netting reflects real outcomes. Under this argument, real
netting effects should be allowed.

 Con: if the realised LGDs are floored to zero, the realised distribution of losses is
truncated. This increases the methodological challenges in LGD estimation where
data is already scarce.

 Con: for estimation based on total losses, netting of profits and losses within a pool is
allowed, indirectly. If netting effects in LGD estimation were not allowed, it might
incentivise banks to use the approach based on total losses, if they wanted to reflect
their actual economic loss in the IRB parameters.

229. Based on these considerations, option 2 has been chosen in the final GLs, i.e. the zero-
floor applies for the calculation of the observed average LGD and the estimation of long-run
average LGD. Irrespective of this requirement, however, all relevant information may be
used in the model development for the purpose of risk differentiation.

5.7 Downturn adjustment

Table 57: How is a downturn period defined?

No. % % EAD
Based on historical macroeconomic and credit factors 95 47 41
The year(s) with the highest observed realised LGD 34 17 15

126
EBA REPORT ON THE IRB MODELLING PRACTICES

No. % % EAD
The year(s) with the highest observed DR 17 8 12
Based on macroeconomic and credit factors, both historical and
16 8 6
forward-looking
Expert judgement 6 3 2
Not applicable (downturn adjustment is not necessary because
6 3 3
downturn is already reflected in the data)
Based on supervisory guidance 5 2 4
Based on a correlation analysis between PD and LGD 4 2 1
Not applicable (downturn is not reflected in the estimates) 3 1 1
Other 16 8 15
Total 202 100 100

230. Table 57 shows how institutions define downturn periods across all LGD models. In 47%
of all LGD models, the downturn period is defined on the basis of historical macroeconomic
and credit factors, and in an additional 8% of LGD models the downturn is defined based on a
combination of historical and forward-looking macroeconomic and credit factors. Several
respondents specified which credit factors are used: based on the years/months with the
highest litigation rates, based on the years/months with the highest loss rates (some banks
mention that they calculate these as the multiplication of observed DRs and observed LGDs),
or based on insolvency rates. Some of the macroeconomic factors are time series of real
estate prices, interest rates, GDP and unemployment rates.

231. In 8.5% of LGD models, the downturn period is defined based on the year(s) with the
highest DR. This approach is somewhat similar to that based on macroeconomic and credit
factors, where the period is defined based on loss rates.

232. Several other respondents (16.83%) indicated that the downturn period is defined on the
basis of the year(s) with the highest observed realised LGD. A few institutions also mentioned
that they then selected defaults to obtain an annual average realised LGD: by vintage on a
three-year window, or in accordance with the complete recovery processes.

233. In almost 3% of models, the downturn adjustment is reflected based on supervisory

guidance given by the competent authority (in one case, it was mentioned that a stressed
scenario is applied to the loan-to-value risk driver and the discount factor).

234. The answer ‘not applicable (downturn adjustment is not necessary because downturn is
already reflected in the data)’ was chosen in a few cases, for instance for sovereign
exposures, where it was argued that loss data always stem from downturn periods; for
municipalities, where it was mentioned that a downturn adjustment is not applicable; and
for a shipping portfolio and a portfolio of insurance products, where it was mentioned that
this segment has no risk of lower recovery rate during downturn periods.

235. Around 8% of responses could not be grouped in a specific category and are therefore
represented in the category ‘other’. While not all comments were entirely clear, the

127
EBA REPORT ON THE IRB MODELLING PRACTICES

following methods were mentioned: selecting the most conservative periods for each model
component over time; using the distance from each annual LGD from the long-run average;
using the volatility of loss rates over a seven-year period; and selecting the worst month-on-
month recoveries observed during the 2009 recession. In several cases, the approach is a
combination of several aspects. In one model, for instance, it was mentioned that the
downturn period was defined as the period with the maximum LGD selected from a PIT LGD
with buffer, long-run LGD (default-weighted average across five years), and stressed default
LGD (highest LGD at time when default peaked, +/- 9 months).

236. In four models, the downturn period is defined based on a correlation analysis between
PD and LGD estimates. The principle of downturn is seen as the correlation between PD and
LGD, which is lacking in the regulatory formula, as the unexpected aspect is only taken
through the PD. Therefore a stressed LGD was computed based on the correlation notion
between PD and LGD (Tasche approach).

Figure 51: How are data selected used in downturn estimation?

According to all observed defaults during the whole observation

period to which an adjustment is made to take into account
downturn conditions

15% According to defaulted exposures which recovery process closes

7% during the downturn period
2% 3%
38%
17% 39%
According to the recoveries occurred in the downturn period
19% (e.g. assigning exposures to a downturn period if the majority of
the realised recoveries are observed during the downturn
19% period)
15% All defaults which occurred during the downturn period are
included.
16% 11%

Not applicable

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

237. The IRB survey then enquired how institutions select data once the downturn period is
established to compute the long-run average LGD. Based on the responses, however, it
appears that this question was not properly understood, since nearly 50% of original
responses were for the category ‘other’ and provided a wide range of explanations not
answering the question. As a result, many of the responses have been discarded because the
explanations given responded to a different question. This was the case when it was
mentioned, for instance, that the data used in downturn estimation are selected based on

128
EBA REPORT ON THE IRB MODELLING PRACTICES

expert judgement, or based on historical time series, etc. The results shown in Figure 51 are
therefore based on a much smaller sample of LGD models than those represented in Table
57 (148 instead of 202).

238. In nearly 40% of LGD models, the data used in downturn estimation are selected based on
all observed defaults during the whole observation period to which an adjustment is made,
to take into account downturn conditions, whereas in 17% of models all defaults that
occurred during the downturn period are included.

239. In 19% of models, those exposures for which the recoveries occurred in the downturn
period are selected (e.g. assigning exposures to a downturn period if the majority of the
realised recoveries are observed during the downturn period). In two institutions, the data
are selected according to defaulted exposures for which the recovery process starts during
the downturn period. However, in around 15% of models defaulted exposures are selected
for which the recovery process closes during the downturn period.

240. Among the responses in the category ‘other’, one institution mentioned that it selects the
data used in downturn estimation according to exposures that default during the downturn
period. One institution mentioned a three-step approach: (1) downturn periods are
identified if the house price index has decreased; (2) the average house price decline during
the downturn period is calculated; and (3) the recovery rate under downturn periods is
computed by subtracting the average house price decline from the usual recovery rate.
Other institutions mentioned a combination of selecting all exposures that defaulted during
the downturn period for the unsecured part of the exposure, and selecting all exposures for
which the recovery process ends during the downturn period for the secured part of the
exposure.

241. In some cases, the respondent mentioned that the question is not applicable. This was
the case for a sovereign portfolio and an aviation portfolio, and in one case it was mentioned
that no downturn period could be identified.

129
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 52: At which level is the downturn adjustment specified?

Level of scope of application of

2% 3% the LGD model
4%
Level of grade or pool
4% 2%2%
12% 5% Level of type of collateral
6%

3% Level of model component

9% 46% 45%

6% Other

Level of an institution (the same

for all LGD models)
27%
Not applicable
25%
Differentiated according to
product type

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

242. In nearly half of the LGD models, the downturn adjustment is specified at the level of the
LGD model, whereas in smaller shares of models, the downturn adjustment is specified at a
lower level: at the level of the grade or pool (in 27% of models), differentiated according to
the type of collateral (9%) or differentiated by product type (2%) (see Figure 52). In around
4% of models, the downturn adjustment is specified uniformly in the institution. Some
respondents (around 6%) mentioned that the downturn adjustment is applied at model
component level, in which case it is not entirely clear whether this leads to a different
adjustment by grade or pool, collateral, or product type, or whether this leads to a uniform
adjustment for all exposures under the scope of application of the LGD model.

243. Some of the responses mentioned in the category ‘other’ are (i) that a different downturn
adjustment is performed for each individual value, or for each asset class, at the level of the
pool or at the level of the collateral (depending on whether or not significant downturn
effects are observable) and (ii) that adjustments are differentiated according to the level of
the collateral and applied at the level of the model component. In one model, the
adjustment is determined at portfolio level and then applied to each obligor.

130
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 53: What is the main methodology used to determine LGD estimates that are appropriate for an economic
downturn?

Apply the LGD estimation

methodology based on data
from the downturn period
(without using model
components)
Apply the LGD estimation
9% methodology using the
20% downturn period value for all
14% 9% model components

20% Apply the LGD estimation

22% methodology using the
downturn period value just
23% for the more relevant model
components
21% Downturn considered within
16% the conservatism applied in
17% the model development
18% process
11%
Fixed downturn adjustment

Other

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

244. When it comes to the methodologies that institutions use to determine downturn LGD
estimates, a wide variety of practices can be observed (see Figure 53). However, in 38% of
LGD models, the downturn period value is used for all model components (22%), or for the
most relevant components (16%). In 23% of LGD models, a fixed downturn adjustment is
applied, and in 9% of models the LGD estimation is based on data from the downturn period
without using model components.

245. Around 17% of respondents indicate that they use conservatism in the model
development process to reflect downturn LGD estimates.

5.8 Summary of model changes in LGD estimation

131
EBA REPORT ON THE IRB MODELLING PRACTICES

246. From the above results of the survey it was possible, for selected questions, to directly
assess the number of models (in the IRB survey sample) and the exposure amounts that
would be affected by the chosen policy, and for which therefore a change in modelling
practice will be required once the GLs enter into force. For those questions for which it is
possible to directly assess whether or not there will be an impact, Table 58 provides an
overview of the resulting model changes, both in terms of the share of LGD models (%) in
terms of the share of exposure values covered by these LGD models (% EAD).
Table 58: Summary of selected policy choices for LGD (non-defaulted) estimation and the number of model changes

No model Model
Not known
change change
Paragraph
% % %
(s) in the Policy choice No. % % %
EAD EAD EAD
GLs
LGD estimates should be based on institution's
102 own loss and recovery experience, and market- 197 59 53 2 4 40 42
implied LGD estimates are not allowed.

The recovery value that should be recognised in

the calculation of the realised LGD should be the
116-117 value of repossession after a haircut, regardless of 112 12 8 71 77 18 15
whether or not it has been sold on the LGD
calculation date.

Institutions should calculate economic loss for

cured cases (i.e. cases where the exposure returns
to non-defaulted status) as for all other defaulted
exposures, with the only difference that additional
135 recovery cash flows are added to the calculation at 200 32 29 46 57 22 14
the date of return to non-defaulted status in the
amount that was outstanding, and should apply
any discounting effects until the moment of
default.

Unpaid late (i.e. after default) fees should not be

137 included in the outstanding amount in the 202 20 22 65 63 15 15
denominator of the realised LGD.

Capitalised interest (i.e. after default) should not

138 be included in the outstanding amount in the 202 26 28 62 62 12 10
denominator of the realised LGD.
If additional drawings are included in the CCF, they
should also be included in the outstanding amount
140-141 in the denominator of the realised LGD. 167 53 58 36 35 11 8
If additional drawings are not included in the CCF,
they should not be included in the denominator.
The annual discounting rate should be composed
of a primary interbank offered rate applicable at
the moment of default, increased by an add-on of
143 five percentage points. The primary interbank 196 30 37 40 34 31 28
offered rate should be considered the three-month
Euribor or a comparable liquid interest rate in the
currency of the exposure.

132
EBA REPORT ON THE IRB MODELLING PRACTICES

No model Model
Not known
change change

Direct costs incurred before default should be

144-145 202 48 48 52 52 0 0
included in the calculation of the realised LGD.

Indirect costs incurred before default should be

144, 146 202 25 15 75 85 0 0
included in the calculation of the realised LGD.

All available data should be considered as relevant

(in relation to Article 181(1)(j) and 181(2) of the
147(e) 198 54 62 46 38 0 0
CRR) and should be included in the historical
observation period.
The long-run average LGD should be calculated
149 separately for each grade or pool, and at the level 195 11 14 43 42 46 44
of the portfolio covered by the LGD model.

The long-run average LGD should be calculated as

an arithmetic average of realised LGDs over an
historical observation period weighted by the
number of defaults, and for retail exposures,
150-151 institutions are allowed to give a higher weight to 195 72 77 23 20 5 4
more recent data in case of retail exposures, if
they demonstrate in a documented manner that
the use of higher weights to more recent data is
justified by better prediction of loss rates.

Institutions should obtain the long-run average

LGD by adjusting the observed average LGD, taking
158 into account the information related to incomplete 199 39 45 49 40 12 15
recovery processes and the estimated future costs
and recoveries on these exposures.

Institutions should define the maximum average

expected period of the recovery process for a
156 given type of exposures, during which the 194 42 42 58 58 0 0
institution realises the vast majority of the
recoveries.

The long-run average LGD should be calculated

based on all defaults observed during the historical
observation period, weighted by the number of
150, 158 195 45 43 30 27 25 30
defaults and by adjusting the observed average
LGD, taking into account estimated future
recoveries.

Wherever institutions realise profit on their

observations of defaults, the realised LGD on these
160 observations should equal zero for the purpose of 200 73 66 16 17 12 17
calculation of the observed average LGD and the
estimation of long-run average LGD.

247. Table 59 further shows how many aspects will at least need to be changed in the LGD
models. It can be seen, for instance, that none of the models in the IRB survey would be

133
EBA REPORT ON THE IRB MODELLING PRACTICES

unaffected by all of the policy aspects listed above. In 29% of the models, at least seven
policy aspects would need to be changed.

248. It should again be mentioned, however, that these statements concern only the aspects
of LGD models on which explicit questions were included in the survey, described above,
whereas the area where guidance is given on LGD models is very broad and likely to affect
many more modelling aspects. In addition, these calculations include only the models for
which a model change is expected, and not those for which it is unknown whether or not a
model change will be necessary. Therefore, these estimates below are a lower bound to the
true number of affected models. In practice, one may assume that all models will probably
have to be changed in one or more dimension.

Table 59: Summary of number of aspects to be changed in LGD estimation

No. of
aspects
No. % % EAD
to be
changed
2 3 1 1
3 6 3 3
4 18 9 10
5 27 13 14
6 27 13 13
7 58 29 20
8 33 16 27
9 19 9 10
10 11 5 2
Total 202 100 100

134
EBA REPORT ON THE IRB MODELLING PRACTICES

6. Estimation of risk parameters for

defaulted exposures

6.1 General requirements specific to LGD in-default and ELBE

estimation and risk drivers
249. The treatment of defaulted assets was identified as one of the drivers of variability of the
own funds requirements across institutions. Clarification has already been provided in the
RTS on IRB assessment methodology, in particular Article 54(2)(c), that the direct estimation
of LGD in-default should be consistent with the LGD for non-defaulted exposures, to avoid
potential cliff effects. Following this approach, it has been further clarified in the GLs (in
paragraph 167) that, for the purpose of estimating ELBE and LGD in-default, institutions
should use the same estimation methods as for estimating LGD on non-defaulted exposures,
as they are in fact part of the LGD model, unless otherwise specified. Paragraph 168 then
further clarifies that institutions should take into account all relevant post-default
information in their LGD in-default and ELBE estimates in a timely manner, and paragraph 176
specifies that the information on the time in-default and recoveries realised so far may be
taken into account directly, as a risk driver, or indirectly, by setting the reference dates for
estimation.

Figure 54: What is your approach to the estimation of LGD in-default?

Distributional approach
11% 1%
6%1% 22% ELBE + add-on
11% 11% 25%
LGD model ~ ND exposures
11%
12% LGD model ~ ND exp with add
RD
Specific model for D exposures
45%
43%
Other

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

250. Based on the IRB survey, 56% of the current LGD in-default models are similar to the LGD
model for non-defaulted exposures, either with or without additional risk drivers (see Figure

135
EBA REPORT ON THE IRB MODELLING PRACTICES

54). It can reasonably be assumed that if changes need to be made to these models, these
will be limited. Where LGD in-default is estimated as ELBE plus add-on 33, whether or not the
approach will be in line with the CRR depends on how the add-on is estimated, since the CRR
requires the add-on to reflect additional unexpected loss during the recovery period
(Article 181(1)(h) of the CRR). For the models where a distributional approach 34 is currently
applied, it is expected that some changes will be needed to comply with the requirement
specified in these GLs.

251. In the IRB survey another option was also provided in the drop-down menu, i.e. where
LGD in-default and ELBE are ‘not specified — risk weight is derived directly’, which referred to
where institutions determine the risk weight as a fixed percentage of the exposure at
default. This option was not chosen in any of the models in the sample.

252. For ELBE estimation, it should be noted that the use of accounting provisions for the ELBE
estimate represents 26% of the models and exposures. For these models, these GLs will
require a change in modelling practices. It should further be noted that in 50% of models
(52% of exposures), the ELBE estimate is obtained based on a similar model to that used for
the LGD non-defaulted, calibrated to current economic conditions (11% of models), or not
calibrated to current economic conditions (39% of models). However, it could be argued that
the requirement for ELBE estimates to reflect current economic conditions (see also Figure 56,
which focuses on this aspect) is already included in the CRR, in Article 181(1)(h). Therefore,
any change in practice on this aspect would not stem from the entry into force of these GLs.

Figure 55: What approach is used for ELBE estimation?

Use accounting provisions

10%
9% 26% LGD model ~ ND exposures
12% 26%
15%
LGD model ~ ND exposures cal.
to current econ. cond.
15%11%
Use of a specific model for ELBE
39%
37% Other

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

33
The ‘ELBE plus add-on’ approach is where the add-on is estimated in a different way to under the ‘distributional
approach’, and reflects adjustment in economic conditions considered (e.g. downturn conditions rather than current
economic conditions) or any other possible sources of unexpected loss or margin of conservatism.
34
The ‘distributional approach’ is where the LGD in-default is estimated as ELBE plus an add-on, where the add-on
reflects the uncertainty (for a given confidence interval) around the ELBE as a function of the distribution of past errors
(i.e. differences between estimated ELBE and the observed losses at the end of the recovery period).

136
EBA REPORT ON THE IRB MODELLING PRACTICES

6.2 Reference dates

Table 60: What is the reference date for estimation?

LGD in-default ELBE

No. % % EAD No. % % EAD
Current date for a defaulted exposure 48 28 12 58 38 25
Date of default 71 42 38 44 29 22
Reference date (as specified in the GLs) 32 19 33 32 21 37
Other/use multiple reference dates 19 11 18 20 13 16
Total 170 100 100 154 100 100

253. The difference between the LGD in-default and the ELBE is used for computing the risk
weight in accordance with Article 153(1)(ii) of the CRR, which is then applied to the current
outstanding exposure amount to obtain the risk-weighted exposure amount. Moreover, the
ELBE is compared with credit risk adjustments for IRB shortfall/excess purposes, where credit
risk adjustments are again computed with respect to the current value of exposures. Thus,
for the purpose of computing realised LGDs for defaulted exposures, institutions should use
reference points in time that will be relevant for the current outstanding obligations of
defaulted exposures.

254. The concept of current outstanding exposure is clearly defined in Article 166(1) of the CRR
and should also be used for defaulted exposures in the application of the ELBE and LGD in-
default. However, given data limitations, the continuous concept of current exposure
amount may not be suitable for estimation purposes. The GLs therefore specify (in
paragraph 171) that institutions should set discrete relevant reference dates to be used for
grouping defaulted exposures in accordance with the recovery patterns observed. These
reference dates should be used instead of the date of default in the estimation of ELBE and
LGD in-default. In this way, it should be feasible to estimate the parameters for defaulted
exposures that are appropriate for their current status. To ensure the adequacy of the
estimates, institutions should set reference dates in accordance with the recovery pattern
observed on a specific type of exposures, where such reference dates may either be event
based, e.g. linked with the realisation of collateral, or reflect certain time periods during
which exposures have been in-default.

255. From the responses to the IRB survey it can be seen that the reference date approach
specified in the GLs is used only in around 20% of LGD in-default and ELBE models (see Table
60). For the other models, the entry into force of the GLs will entail a change in practice.

6.3 The requirement to reflect current economic circumstances

in ELBE estimates

137
EBA REPORT ON THE IRB MODELLING PRACTICES

256. The ELBE estimates already reflect current economic conditions in 38% of models (49% of
exposures) (see Figure 56), and for the other approaches, a change in practice will be
required, unless the downturn or stressed economic conditions coincide with the current
economic conditions. However, it could be argued that such change would not stem from the
introduction of the GLs, since the requirement for ELBE estimates to reflect current economic
conditions is already included in the CRR (Article 181(1)(h) of the CRR).

Figure 56: Which economic conditions are reflected in ELBE estimates?

0%6% Current economic conditions

1%5%
Downturn economic conditions
38%
34% Long-run average economic
36% 49%
conditions
Stressed conditions

20%
Other
12%

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

257. The GLs specify (in paragraphs 184-185) how institutions should reflect the current
economic conditions in their ELBE estimates. In particular, institutions are allowed to estimate
ELBE on the basis of the long-run average LGD if certain conditions are met (among others,
that the model should directly include at least one macroeconomic factor as a risk factor, as
well as one material risk driver that is sensitive to economic conditions). Institutions may also
adjust the long-run average LGD for defaulted exposures to reflect current economic
conditions, if this adjustment is documented.

138
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 57: If you incorporate current economic conditions in ELBE, how are these incorporated?

Based on current provisions

12%
18% 9% Expert judgement is used to give higher weight to recent data, to
36% exclude downturn periods or to select historical observations which
reflect the current situation
Information from macroeconomic and credit factors is directly
25% included in the model
48%38%

Other
14%

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

258. Figure 57 shows the distribution of the methodologies used to reflect the current
economic conditions in the ELBE estimates. This information was available for 56 ELBE
models 35. In almost half of the models, the approach used to reflect current economic
conditions is based on expert judgement either to give higher weight to recent observations,
to exclude certain downturn periods, or to select the historical periods that are deemed to
reflect current economic conditions. In 27% of those ELBE models, the approach is based on
macroeconomic and credit factors in the model. The category ‘other’ was selected in nine
ELBE models, and based on the comments, these refer to approaches that rely on the current
exposure value and a calibration based on a PIT LGD, or a situation in which ELBE is calculated
based on long-run average LGD but calibrated to a specific point in time.

259. Whereas those ELBE models that are currently based on provisions (9% of models, or 12%
of exposures) will have to be changed after the introduction of the GLs, it is difficult to assess
ex ante whether or not those models that rely expert judgement or directly include
information from macroeconomic or credit factors in the model will have to be changed to
comply with the GLs.

6.4 Relation of LGD in-default and ELBE to credit risk adjustments

260. The GLs specify (in paragraphs 186-188) that the use of specific credit risk adjustments
(SCRA) as ELBE estimates should be limited to those cases where provisions models meet, or
could be adjusted to meet (e.g. by modifying the discounting rate in use), the prudential

35 This information was not available for all 63 models where current economic conditions are reflected in ELBE
estimates (representing the 38% in), because the answer was not available for all models.

139
EBA REPORT ON THE IRB MODELLING PRACTICES

requirements for own LGD estimates and the requirements specified in these GLs. An
exception to this rule is made allowing institutions to use individually assessed provisions as
a possible reason for over-ride where they are able to prove that they provide a more
accurate estimation than the ELBE estimated by facility grade or pool. For this purpose,
individually assessed provisions should be adjusted to be consistent with the requirements
on economic loss set out in the GLs.

261. The responses in Table 61 show that, in most of the ELBE models (more than 80%), the
SCRA calculated on portfolio basis and SCRA assessed individually (in almost 70% of ELBE
models) are not used for the purpose of ELBE estimation. For those ELBE models where SCRA is
automatically used as ELBE (15% for SCRA at portfolio basis and 25% if assessed individually),
the policy prescribed in the GLs would entail a change in practice, because it should be
verified that those models meet the prudential requirements for LGD estimates.
Table 61: Do you use the information on SCRA in the ELBE estimation?

SCRA calculated on a portfolio

SCRA assessed individually
basis
No. % % EAD No. % % EAD
No, these are not used for the purpose of ELBE 134 82 87 111 68 63
Yes, these are automatically recognised as ELBE 25 15 12 42 26 29
Yes, these are used as a possible reason for over-
2 1 1 5 3 2
ride
Other 3 2 0 5 3 5
Total 164 100 100 163 100 100

6.5 Specific requirements for LGD in-default estimation

262. From the GLs (paragraph 189) it should be clear that LGD in-default estimates should also
reflect economic downturn conditions if these are more conservative than the long-run
average LGD estimates. The survey confirms (see Figure 58) that this understanding is
already applied in most of the models (76%), and in an additional 3% of models these LGD in-
default estimates reflect stressed conditions.

140
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 58: Which economic conditions are reflected in LGD in-default?

Current economic conditions

1% 10%
3%2%
3%5% 8% Downturn economic conditions
7%

Long-run average economic

conditions
Stressed conditions
76%
83% Other

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

263. For LGD in-default estimations, the GLs specify (in paragraphs 191-192) that institutions
should analyse and correct the LGD in-default in those situations where the ELBE obtained
using SCRA is above the LGD in-default obtained through direct estimation. Furthermore, the
GLs also clarify that individually assessed SCRA may also be used to over-ride the LGD in-
default, but it should be ensured that the add-on to the ELBE covers for any increase of loss
rate due to potential additional losses during the recovery period.

264. The results on these aspects of the survey are presented in Table 62. In line with the
findings for ELBE, SCRA is not used for the purpose of LGD in-default in the majority of LGD
models (96% for SCRA at portfolio basis and 84% for SCRA assessed individually). The use of
SCRA assessed individually to over-ride LGD in-default estimations happens only in around
6% of models.
Table 62: Do you use the information on SCRA in the LGD in-default estimation?

SCRA calculated on a
SCRA assessed individually
portfolio basis
No. % % EAD No. % % EAD
No, these are not used for the purpose of LGD in-
167 97 99 142 84 77
default
Yes, these are automatically recognised as LGD in-
1 1 0 10 6 8
default
Yes, these are used as a possible reason for over-ride 2 1 1 4 2 1
Other 3 2 0 14 8 14
Total 173 100 100 170 100 100

6.6 Summary of model changes in LGD in-default and ELBE

estimation

141
EBA REPORT ON THE IRB MODELLING PRACTICES

Table 63: Summary of selected policy choices for LGD (in-default) estimation and the number of model changes

No model
Model change Not known
change

Paragrap
h(s) in Policy choice No. % % EAD % % EAD % % EAD
the GLs

LGD estimates should be based on

institution's own loss and recovery
102 170 56 52 2 5 41 43
experience, and market-implied LGD
estimates are not allowed.
The annual discounting rate should be
composed of a primary interbank
offered rate applicable at the moment
of default, increased by an add-on of
143 five percentage points. The primary 169 27 37 40 31 33 33
interbank offered rate should be
considered the three-month Euribor or
a comparable liquid interest rate in the
currency of the exposure.
All available data should be considered
as relevant (as referred to in
147(e) Article 181(1)(j) and 181(2) of the CRR) 161 52 57 48 43 0 0
and should be included in the historical
observation period.
Institutions should obtain the long-run
average LGD by adjusting the observed
average LGD, taking into account the
158 information related to incomplete 162 32 41 52 39 16 20
recovery processes, and the estimated
future costs and recoveries on these
exposures.
Institutions should set the reference
dates to be used for grouping
171 170 19 33 70 50 11 18
defaulted exposures in accordance
with the recovery patterns observed.
For the purposes of taking into account
the information on the time in-default
and recoveries realised so far,
institutions may take into account this
176 160 11 12 46 44 43 44
information either directly as risk
drivers or indirectly, for instance by
setting the reference date for
estimation.

LGD in-default estimates should be

189 appropriate for an economic 178 76 83 19 16 5 1
downturn.

142
EBA REPORT ON THE IRB MODELLING PRACTICES

Table 64: Summary of number of aspects to be changed in LGD in-default estimation

No. of aspects to be changed No. % % EAD

0 7 4 7
1 35 20 32
2 46 26 27
3 44 25 19
4 31 18 11
5 13 7 4
6 1 1 1
Total 177 100 100

Table 65: Summary of selected policy choices for ELBE estimation and the number of model changes

No model Model
Not known
change change

Paragrap
% % %
h in the Policy choice No. % % %
EAD EAD EAD
GLs

LGD estimates should be based on institution's

102 own loss and recovery experience, and market- 148 55 53 1 0 44 47
implied LGD estimates are not allowed.
The annual discounting rate should be
composed of a primary interbank offered rate
applicable at the moment of default, increased
by an add-on of five percentage points. The
143 152 26 33 41 44 32 23
primary interbank offered rate should be
considered the three-month Euribor or a
comparable liquid interest rate in the currency
of the exposure.
All available data should be considered as
relevant (as referred to in Article 181(1)(j) and
147(e) 137 53 61 47 39 0 0
181(2) of the CRR) and should be included in the
historical observation period.
Institutions should obtain the long-run average
LGD by adjusting the observed average LGD,
taking into account the information related to
158 142 36 49 49 32 15 19
incomplete recovery processes and the
estimated future costs and recoveries on these
exposures.
Institutions should set the reference dates to be
used for grouping defaulted exposures in
171 154 21 37 66 47 13 16
accordance with the recovery patterns
observed.
Institutions should use an LGD model as for non-
defaulted exposures, calibrated to current
167-168 149 11 15 64 63 24 22
economic conditions and taking into all relevant
post-default information.

143
EBA REPORT ON THE IRB MODELLING PRACTICES

No model Model
Not known
change change

ELBE estimates should reflect current economic

183-184 166 38 49 57 46 5 6
conditions.

Institutions should estimate ELBE on the basis of

the long-run average LGD if certain conditions
are met (among others, the model should
directly include at least one macroeconomic
factor as a risk factor, as well as one material
184-185 56 25 14 9 12 66 74
risk driver that is sensitive to economic
conditions), or institutions may adjust the long-
run average LGD for defaulted exposures to
reflect current economic conditions, if this
adjustment is documented.

Table 66: Summary of number of aspects to be changed in ELBE estimation

No. of aspects to be changed No. % % EAD

0 9 6 8
1 14 9 16
2 31 20 23
3 44 28 28
4 28 18 16
5 26 17 6
6 4 3 2
Total 156 100 100

144
EBA REPORT ON THE IRB MODELLING PRACTICES

7. Application of risk parameters

265. In the section on application of risk parameters, a requirement is included (in
paragraph 194) that specifies that where institutions receive new information with respect to
a relevant risk driver or rating criterion, they should take this information into account in the
rating assignment in a timely manner. The same section also requires that the relevant IT
systems and the corresponding rating or LGD assignment be updated and reviewed as soon
as possible, and where the new information relates to the default of an obligor, that the PD
of the obligor be set to 1 in all relevant IT systems in a timely manner.

266. In the IRB survey, institutions were asked (for their PD models only) whether or not they
already have a policy or practice regarding the inclusion of newly available rating-relevant
information to be incorporated into the rating assignment.

267. In around 90% of the PD models, the institutions reported having a specific policy to
include new available information, but for the remaining 10% there is no such policy. While
not all respondents mentioned the frequency at which new information is incorporated into
the rating assignment, for those who did, the rerating including the most recent information
is stated to be performed at least monthly in 73% of PD models; in 21% of the relevant PD
models this is performed at least annually or at least quarterly (8%). However, most
respondents also indicated that besides these regular calculations, the ratings are updated
more often where relevant information becomes available immediately, for instance based
on available delinquency data, based on an online link to databases of external credit
bureaus, because there are indications of a significant improvement or deterioration in the
obligor’s risk situation, when a trigger of default is identified or no longer applicable, or when
new appraisals are obtained.

268. In the CP on the GLs, it was proposed specifying that such new information should be
incorporated within three months. However, the industry argued that this requirement
would be particularly burdensome for corporate portfolios and would lead to distorting
seasonal effects (as a result of information updated only once a year). It was further
mentioned that qualitative components of a rating cannot be reviewed in such a short time.
Based on this feedback, and the overview of the current practices, it has been decided to
relax the wording to ‘as soon as possible’.

269. The GLs specify (in paragraphs 195-196) that institutions should apply additional
conservatism to the outcomes of the assignment of exposures to grades or pools, in case of
any identified deficiencies related to the implementation of the model in the IT system, or in
the process of assignment of risk parameters to obligors of facilities in the current portfolio
(application of risk parameters). The GLs mention in more detail that institutions should
consider at least the triggers mentioned in Table 67.

145
EBA REPORT ON THE IRB MODELLING PRACTICES

Table 67: What are the main triggers for including additional conservatism in the application of the PD model?

%
sum No. %
EAD
Missing data in the current portfolio 84 252 33 37
Missing updates of financial statements or credit bureau data as referred to in
42 252 17 13
paragraph 66(h) of the CP of the GLs
Outdated ratings in the current portfolio, where outdated rating should be
understood as specified in Article 25(2)(b) of Regulation (EU) No xxx/xxxx [RTS on 63 252 25 20
Assessment methodology]
Missing ratings, whereby an exposure is considered as being within the scope of
89 252 35 36
application of the IRB model but is not rated by it
Other 82 252 33 37

Note: this was a ‘tick box’ question, hence respondents could select several of the above answers.

270. The results in Table 67 show that these reasons are a common justification for applying
additional MoC in the application. There is a share of 33% of PD models where the category
‘other’ was selected for the question above. In most cases, the respondents indicated that no
additional MoC was applied or that such additional MoC is ‘not applicable’. Others referred
to a supervisory imposed additional MoC, additional MoC for statistical uncertainty, general
estimation errors, low-default portfolios, or reflection of uncertainty around the calibration
target/long-run average measurement.

271. Based on this feedback, the non-exhaustive list of triggers for additional MoC in
application has been maintained in the GLs.

146
EBA REPORT ON THE IRB MODELLING PRACTICES

8. Review of estimates
272. The GLs specify (in paragraphs 218-219) that institutions should perform an annual review
of the estimates in accordance with Article 179(1)(c) of the CRR, by establishing a framework
that includes the minimum scope of the analyses, predefined metrics and standards, and
predefined actions in case deviations from those standards are observed.

273. For PD estimates, it is specifically required (in paragraph 218(c)(i)) to analyse whether or
not including the most recent data leads to a significant change in the long-run average DR,
including an assessment of the period of likely range of DR and the mix of good and bad
years.

274. Based on the responses from the IRB survey shown in Figure 59, there is no pre-
established frequency for redeveloping or recalibrating the PD model in more than half of
the PD models. In 34% of PD models, the model recalibration is developed annually, and in
3% of PD models this is done quarterly. The option ‘other’ was selected in around 10% of PD
models, in which case the most common explanation is that an annual review of the
estimates is performed, without this review necessarily leading to a (re)calibration of the
model, i.e. the model is only (re)calibrated where the monitoring of the model indicates that
an earlier redevelopment/recalibration is necessary. Some respondents mentioned that the
model is only (re)calibrated where the metrics exceed fixed tolerance thresholds, where
back-testing indicates that (re)calibration is necessary, or where changes in the economic
environment or credit policies have been observed.

Figure 59: Do you have a pre-established frequency for developing a (re)calibration of the PD model? If yes, what is
that frequency?

3% Annually
14%
11% 3% Every 2 years
33%
34% Every 3 years
No pre-established frequency
Other
0%
52% 2% Quarterly
0%
47%

Note: the inner circle shows the share of each option where all PD models are weighted equally, whereas the outer
circle shows the share of each option where PD models are weighted by their corresponding exposure value.

147
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 60: What is the frequency at which the observed average DRs are calculated?

1%
14%
4% Annually
19% Monthly
Quarterly
24% Semi-annually
15% 62%61%

Note: the inner circle shows the share of each option where all LGD models are weighted equally, whereas the outer
circle shows the share of each option where LGD models are weighted by their corresponding exposure value.

275. Related to this requirement, the IRB survey enquired about the usual frequency at which
the observed average DRs are calculated, because these calculations would be the starting
point to indicate whether or not the PD estimates should be re-estimated or recalibrated. For
the majority of PD models (61.54%), the observed average DR is calculated annually, whereas
it is calculated monthly and quarterly in 15.38% and 18.80% of the PD models respectively
(see Figure 60). Figure 61 shows the split of these statistics across exposure classes. Some
significant deviations can be observed. More specifically, it can be seen that the share of
corporate models that recalculate the observed DR annually is higher (75%) than for the
retail models (58%). In retail models, the option to recalculate the observed average DR
monthly is more common (15% on average) than for the corporate models (only 6%).

148
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 61: Frequency for calculating the observed average DRs, by exposure class

% EAD 54% 2% 12% 33%

Retail

% 51% 6% 23% 20%

central Institutio Corporat

% EAD 87% 3% 8% 1%
e

% 75% 4% 14% 7%
% EAD 84% 0% 14% 2%
ns

% 84% 0% 12% 4%
governme

% EAD 65% 0% 32% 3%

nts and
Central

banks

% 88% 0%6% 6%

% EAD 61% 1% 14% 24%

Total

% 62% 4% 19% 15%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Annually Semi-annually Quarterly Monthly

276. Similarly to the requirements for the annual review of PD estimates, institutions are also
required to perform an annual review of the LGD estimates, and should in particular verify
whether or not the inclusion of the most recent data leads to a significant change in the long-
run average LGD or downturn LGD (paragraph 218(c)(i) of the GLs). On that aspect,
institutions have been asked to indicate whether or not they have a pre-established
frequency for the redevelopment or re-estimation of the LGD model. Figure 62 shows the
responses, which are in line with the findings in Figure 559. However, as compared with the
PD models, it seems that there is a higher share of LGD (and ELBE) models (around 40% as
compared with 34% for PD models) that are redeveloped or re-estimated on an annual basis,
and a slightly lower share of LGD models (also around 40% as compared with around 50% for
PD models) for which there is no pre-established frequency for redevelopment or re-
estimation.

149
EBA REPORT ON THE IRB MODELLING PRACTICES

Figure 62: Do you have a pre-established frequency for redeveloping or re-estimating the LGD model? If yes, what is
that frequency?

60%

40%
LGD non-defaulted
LGD in-default
ELBE
20%

0%
No pre- Annually Semi-annually Quarterly Other (please
established specify)
frequency

150
EBA REPORT ON THE IRB MODELLING PRACTICES

Appendix
Table 68: Classification of answers from the survey with respect to the policy chosen in the GLs – PD models

Paragra
No model change if Model change if answer
ph(s) in Question Not known if answer is
answer is is
the GLs
Yes, but only obligors
Are there any obligors
benefiting from
who are in the scope of
unfunded credit
53-54 application that do not No
protection are excluded
receive an individual PD
Yes, for other reasons
estimation?
(please specify)
Quarterly
Frequency at which one- Annually
78 Monthly
year DRs are evaluated Semi-annually
Daily
Was any specific analysis
done to justify the
choice for overlapping
80 versus non-overlapping Yes No
windows for the
calculation of the
observed average DR?
Do you conduct
calibration before or Before application of No MoC is applied MoC is applied during
89
after the application of MoC After application of MoC calibration
MoC?
Do you conduct
calibration before or Before application of the After application of the
89 No PD floor is applied
after the application of PD floor PD floor
the PD floor?

Table 69: Classification of answers from the survey with respect to the policy chosen in the GLs – LGD (non-defaulted)
models

Paragrap
No model change if Model change if answer
h(s) in Question Not known if answer is
answer is is
the GLs
Multivariate regression
analysis/sophisticated
LGD based on total
Which type of LGD Market-implied LGD statistical model
102 losses and PD estimates
model is this? (based on market data) Other or any
Work-out LGD
combination of the
above

151
EBA REPORT ON THE IRB MODELLING PRACTICES

Paragrap
No model change if Model change if answer
h(s) in Question Not known if answer is
answer is is
the GLs
Value of the sale (after
repossession), but a null
recovery in case the
collateral has not yet
been sold on the LGD
calculation date
Value of the sale (after
repossession), but the
value of repossession
without a haircut where
If you repossess
Value of repossession the collateral has not
collaterals at least
after a haircut, yet been sold on the
occasionally, which
regardless of whether LGD calculation date
116-117 value of recovery do Other (please specify)
or not it has been sold Value of repossession
you recognise in the
on the LGD calculation without a hair-cut,
calculation of realised
date regardless of whether
LGD?
or not it has been sold
on the LGD calculation
date
Value of the sale (after
repossession), but the
value of repossession
after a haircut where
the collateral has not
been sold yet on the
LGD calculation date
Assume that the
economic loss for cured
Using the same cases is zero
methodology as for Using the same
other defaulted methodology as for
How is economic loss of exposures, including the other defaulted Other
135
a cured case measured? discounted additional exposures including, but
recovery cash flow at not discounting,
the date of the return to additional recovery cash
non-defaulted status flow at the date of the
return to non-defaulted
status
Add to the outstanding
amount at default
(denominator of
realised LGD)
How do you
Both include in
incorporate unpaid late
economic loss and add
fees (late meaning after
137 Do not include to outstanding amount Other (please specify)
default) into the
at default
calculation of realised
Include only in the
LGD?
economic loss
(numerator of realised
LGD)

152
EBA REPORT ON THE IRB MODELLING PRACTICES

Paragrap
No model change if Model change if answer
h(s) in Question Not known if answer is
answer is is
the GLs
Add to the outstanding
amount at default
(denominator of
realised LGD)
How do you
Both include in
incorporate capitalised
economic loss and add
interest (meaning
138 Do not include to outstanding amount Other (please specify)
interest after default)
at default
into your calculation of
Include only in the
realised LGD?
economic loss
(numerator of realised
LGD)

If additional drawings
are not included in the
CCF:
Add to the outstanding
amount at default
(denominator of
Include only in the realised LGD)
economic loss Both include in
(numerator of realised economic loss and add
LGD) if additional to outstanding amount
Are additional drawings drawings are not at default
after default included in included in the CCF Don't include
140-141 Other (please specify)
the calculation of Both include in
realised LGD? economic loss and add If additional drawings
to outstanding amount are included in the CCF:
at default if additional Include only in the
drawings are included in economic loss
the CCF (numerator of realised
LGD)
Add to the outstanding
amount at default
(denominator of
realised LGD)
Do not include
Funding rate + add-on
What methodology do Original effective
you use to determine interest rate
143 Risk-free rate + add-on Other
the discounting rate Current effective
(LGD non-defaulted)? interest rate
Funding rate
Do you include direct
costs incurred before
No
144-145 default in the Yes
Partially
calculation of realised
LGD?
Do you include indirect
costs incurred before
No
144, 146 default in the Yes
Partially
calculation of realised
LGD?

153
EBA REPORT ON THE IRB MODELLING PRACTICES

Paragrap
No model change if Model change if answer
h(s) in Question Not known if answer is
answer is is
the GLs
Did you exclude some
of the available
historical data from the
Yes (please provide an
147(e) specification of the No
explanation)
historical observation
period (LGD non-
defaulted)?

At which level do you Portfolio level (scope of

Both at portfolio level
149 calculate long-run Level of grade or pool LGD model)
and at grade/pool level
average LGD? Other (please specify)

For retail exposures: For retail exposures:

All defaults are Based on the exposure
What type of weighting weighted equally value
do you use in the More recent defaults For all other exposures:
150-151 Other (please specify)
calculation of long-run have higher weights Based on the exposure
average LGD? For all other exposures: value
All defaults are More recent defaults
weighted equally have higher weights

Incomplete recovery As an adjustment at

How do you
With recoveries realised processes are not grade or pool level
incorporate incomplete
158 so far and estimated included As an adjustment at
recovery processes in
future recoveries Only with recoveries portfolio level
your LGD estimates?
realised so far Other (please specify)

Do you define a
maximum period after
which incomplete
recovery processes are Yes (please specify the
156 No
treated as closed for length of this period)
the purpose of
calculating the average
realised LGD?

Average of the annual

average of the realised
LGD of all defaults
occurred in a year
Average of the realised Average of the annual
Long-run average LGD
How do you calculate LGD of all defaults average of the realised
based on intermediate
150, 158 the long-run average occurred in the LGD of the closed
averages (please
LGD? historical observation defaults occurred in a
specify)
period year
Other (please specify)
Average of the realised
LGD of the closed
defaults occurred in the
historical observation

154
EBA REPORT ON THE IRB MODELLING PRACTICES

Paragrap
No model change if Model change if answer
h(s) in Question Not known if answer is
answer is is
the GLs
period

No floor
How do you treat cases Zero-floor at the level of Zero-floor at the level of
160 with no loss or positive individual default pool or grade Other (please specify)
outcome? observation Zero-floor at the level of
risk parameter

Table 70: Classification of answers from the survey with respect to the policy chosen in the GLs – LGD (in-default)
models

No model change if Model change if

Paragraph(s) in the GLs Question Not known if answer is
answer is answer is
Multivariate regression
LGD based on total analysis/sophisticated
Which type of LGD Market-implied
losses and PD statistical model
102 in-default model is LGD (based on
estimates Other or any
this? market data)
Work-out LGD combination of the
above
Funding rate + add-
What methodology on
do you use to Original effective
Risk-free
143 determine the interest rate Other
rate + add-on
discounting rate Current effective
(LGD in-default)? interest rate
Funding rate
Did you exclude
some of the
available historical
data from the Yes (please provide
147(e) No
specification of the an explanation)
historical
observation period
(LGD in-default)?
How do you Incomplete
As an adjustment at
incorporate With recoveries recovery processes
grade or pool level
incomplete realised so far and are not included
158 As an adjustment at
recovery processes estimated future Only with
portfolio level
in your LGD recoveries recoveries realised
Other (please specify)
estimates? so far

155
EBA REPORT ON THE IRB MODELLING PRACTICES

No model change if Model change if

Paragraph(s) in the GLs Question Not known if answer is
answer is answer is

What is the Reference date (as Current date for a Other/if you use
171 reference date for specified in the defaulted exposure multiple reference
estimation? GLs) Date of default dates, please specify

Distributional
approach
Use of an LGD
Use of an LGD ELBE + add-on (please
Please indicate model as for non-
model as for non- specify how the add-on
your approach to defaulted
defaulted is estimated)
176 the estimation of exposures
exposures with Use of a specific model
LGD for defaulted LGD in-default not
additional risk for defaulted exposures
exposures? specified — risk
drivers Other (please specify)
weight derived
directly (please
specify how)
Current economic
Which economic conditions
Downturn
conditions are Long-run average
189 economic Other (please specify)
reflected in LGD in- economic
conditions
default? conditions
Stressed conditions

Table 71: Classification of answers from the survey with respect to the policy chosen in the GLs — ELBE models

Paragrap
h(s) in No model change if answer is Model change if answer is Not known if answer is
the GLs
Multivariate regression
LGD based on total losses and analysis/sophisticated statistical
Market-implied LGD (based on
102 PD estimates model
market data)
Work-out LGD Other or any combination of the
above

Funding rate + add-on

Original effective interest rate
143 Risk-free rate + add-on Other
Current effective interest rate
Funding rate

147(e) No Yes (please provide an explanation)

156
EBA REPORT ON THE IRB MODELLING PRACTICES

Paragrap
h(s) in No model change if answer is Model change if answer is Not known if answer is
the GLs

As an adjustment at grade or pool

With recoveries realised so Incomplete recovery processes are level
158 far and estimated future not included As an adjustment at portfolio
recoveries Only with recoveries realised so far level
Other (please specify)

Current date for a defaulted

Reference date (as specified Other/if you use multiple
171 exposure
in the GLs) reference dates, please specify
Date of default

Use of accounting provisions

Use of an LGD model as for Use of an LGD model as for non-
non-defaulted exposures, defaulted exposures Use of a specific model for ELBE
167-168
calibrated to current ELBE not specified — risk weight Other (please specify)
economic conditions derived directly (please specify
how)

Long-run average economic

conditions
183-184 Current economic conditions Other (please specify)
Downturn economic conditions
Stressed conditions

Expert judgement is used to give

Information from higher weight to recent data, to
macroeconomic and credit exclude downturn periods or to
184-185 Based on current provisions
factors is directly included in select historical observations that
the model reflect the current situation
Other (please specify)

157