[Federal Register: August 4, 2003 (Volume 68, Number 149)]
[Notices]
[Page 45949-45988]
From the Federal Register Online via GPO Access [wais.access.gpo.gov]
[DOCID:fr04au03-137]
-----------------------------------------------------------------------
DEPARTMENT OF THE TREASURY
Office of the Comptroller of the Currency
[Docket No. 03-15]
FEDERAL RESERVE SYSTEM
[Docket No. OP-1153]
FEDERAL DEPOSIT INSURANCE CORPORATION
DEPARTMENT OF THE TREASURY
Office of Thrift Supervision
[No. 2003-28]
Internal Ratings-Based Systems for Corporate Credit and
Operational Risk Advanced Measurement Approaches for Regulatory Capital
AGENCIES: Office of the Comptroller of the Currency (OCC), Treasury;
Board of Governors of the Federal Reserve System (Board); Federal
Deposit Insurance Corporation (FDIC); and Office of Thrift Supervision
(OTS), Treasury.
ACTION: Draft supervisory guidance with request for comment.
-----------------------------------------------------------------------
SUMMARY: The OCC, Board, FDIC, and OTS (the Agencies) are publishing
for industry comment two documents that set forth draft supervisory
guidance for implementing proposed revisions to the risk-based capital
standards in the United States. These proposed revisions, which would
implement the New Basel Capital Accord in the United States, are
published as an advance notice of proposed rulemaking (ANPR) elsewhere
in today's Federal Register. Under the advanced approaches for credit
and operational risk described in the ANPR, banking organizations would
use internal estimates of certain risk components as key inputs in the
determination of their regulatory capital requirements. The Agencies
believe that supervisory guidance is necessary to balance the
flexibility inherent in the advanced approaches with high standards
that promote safety and soundness and encourage comparability across
institutions.
The first document sets forth Draft Supervisory Guidance on
Internal Ratings-Based Systems for Corporate Credit (corporate IRB
guidance). This document describes supervisory expectations for
institutions that intend to adopt the advanced internal ratings-based
approach (A-IRB) for credit risk as set forth in today's ANPR. The
corporate IRB guidance is intended to provide supervisors and
institutions with a clear description of the essential components and
characteristics of an acceptable A-IRB framework. The guidance focuses
specifically on corporate credit portfolios; further guidance is
expected at a later date on other credit portfolios (including, for
example, retail and commercial real estate portfolios).
The second document sets forth Draft Supervisory Guidance on
Operational Risk Advanced Measurement Approaches for Operational Risk
(AMA guidance). This document outlines supervisory expectations for
institutions that intend to adopt an advanced measurement approach
(AMA) for operational risk as set forth in today's ANPR.
The Agencies are seeking comments on the supervisory standards set
forth in both documents. In addition to seeking comment on specific
aspects of the supervisory guidance set forth in the documents, the
Agencies are seeking comment on the extent to which the supervisory
guidance strikes the appropriate balance between flexibility and
specificity. Likewise, the Agencies are seeking comment on whether an
appropriate balance has been struck between the regulatory requirements
set forth in the ANPR and the supervisory standards set forth in these
documents.
DATES: Comments must be received no later than November 3, 2003.
ADDRESSES: Comments should be directed to:
OCC: Please direct your comments to: Office of the Comptroller of
the Currency, 250 E Street, SW., Public Information Room, Mailstop 1-5,
Washington, DC 20219, Attention: Docket No. 03-15; fax number (202)
874-4448; or Internet address: regs.comments@occ.treas.gov. Due to
delays in paper mail delivery in the Washington area, we encourage the
submission of comments by fax or e-mail whenever possible. Comments may
be inspected and photocopied at the OCC's Public Information Room, 250
E Street, SW., Washington, DC. You may make an appointment to inspect
comments by calling (202) 874-5043.
Board: Comments should refer to Docket No. OP-1153 and may be
mailed to Ms. Jennifer J. Johnson, Secretary, Board of Governors of the
Federal Reserve System, 20th Street and Constitution Avenue, NW.,
Washington, DC, 20551. However, because paper mail in the Washington
area and at the Board of Governors is subject to delay, please consider
submitting your comments by e-mail to regs.comments@federalreserve.gov,
or faxing them to the Office of the Secretary at 202/452-3819 or 202/
452-3102. Members of the public may inspect comments in Room MP-500 of
the Martin Building between 9 a.m. and 5 p.m. on weekdays pursuant to
Sec. 261.12, except as provided in Sec. 261.14, of the Board's Rules
Regarding Availability of Information, 12 CFR 261.12 and 261.14.
FDIC: Written comments should be addressed to Robert E. Feldman,
Executive Secretary, Attention: Comments, Federal Deposit Insurance
Corporation, 550 17th Street, NW., Washington, DC, 20429. Commenters
are encouraged to submit comments by facsimile transmission to (202)
898-3838 or by electronic mail to Comments@FDIC.gov. Comments also may
be hand-delivered to the guard station at the rear of the 550 17th
Street Building (located on F Street) on business days between 8:30
a.m. and 5 p.m. Comments may be inspected and photocopied at the FDIC's
Public Information Center, Room 100, 801 17th Street, NW., Washington,
DC between 9 a.m. and 4:30 p.m. on business days.
OTS: Send comments to Regulation Comments, Chief Counsel's Office,
Office of Thrift Supervision, 1700 G Street, NW., Washington, DC 20552,
Attention: No. 2003-28. Delivery: Hand deliver comments to the Guard's
desk, east lobby entrance, 1700 G Street, NW., from 9 a.m. to 4 p.m. on
business days, Attention: Regulation Comments, Chief Counsel's Office,
Attention: No. 2003-28. Facsimiles: Send facsimile transmissions to FAX
Number (202) 906-6518, Attention: No 2003-28. e-mail: Send e-mails to
regs.comments@ots.treas.gov, Attention: No. 2003-28, and include your
name and telephone number. Due to temporary disruptions in mail service
in the Washington, DC area, commenters are encouraged to send comments
by fax or e-mail, if possible.
FOR FURTHER INFORMATION CONTACT:
OCC: Corporate IRB guidance: Jim Vesely, National Bank Examiner,
Large Bank Supervision (202/874-5170 or james.vesely@occ.treas.gov);
AMA guidance: Tanya Smith, Senior International Advisor, International
Banking & Finance (202/874-4735 or tanya.smith@occ.treas.gov).
Board: Corporate IRB guidance: David Palmer, Supervisory Financial
Analyst, Division of Banking Supervision and Regulation (202/452-2904
or david.e.palmer@frb.gov); AMA guidance: T. Kirk Odegard, Supervisory
Financial Analyst, Division of Banking Supervision and Regulation (202/
530-6225 or thomas.k.odegard@frb.gov). For users of Telecommunications
Device for
[[Page 45950]]
the Deaf (``TDD'') only, contact 202/263-4869.
FDIC: Corporate IRB guidance and AMA guidance: Pete D. Hirsch,
Basel Project Manager, Division of Supervision and Consumer Protection
(202/898-6751 or phirsch@fdic.gov).
OTS: Corporate IRB guidance and AMA guidance: Michael D. Solomon,
Senior Program Manager for Capital Policy (202/906-5654); David W.
Riley, Project Manager (202/906-6669), Supervision Policy; Teresa A.
Scott, Counsel (Banking and Finance) (202/906-6478); or Eric
Hirschhorn, Principal Financial Economist (202/906-7350), Regulations
and Legislation Division, Office of the Chief Counsel, Office of Thrift
Supervision, 1700 G Street, NW., Washington, DC 20552.
Document 1: Draft Supervisory Guidance on Internal Ratings-Based
Systems for Corporate Credit
Table of Contents
I. Introduction
A. Purpose
B. Overview of Supervisory Expectations
1. Ratings Assignment
2. Quantification
3. Data Maintenance
4. Control and Oversight Mechanisms
C. Scope of Guidance
D. Timing
II. Ratings for IRB Systems
A. Overview
B. Credit Ratings
1. Rating Assignment Techniques
a. Expert Judgment
b. Models
c. Constrained Judgment
C. IRB Ratings System Architecture
1. Two-Dimensional Rating System
a. Definition of Default
b. Obligor Ratings
c. Loss Severity Ratings
2. Other Considerations of IRB Rating System Architecture
a. Timeliness of Ratings
b. Multiple Ratings Systems
c. Recognition of the Risk Mitigation Benefits of Guarantees
3. Validation Process
a. Ratings System Developmental Evidence
b. Ratings System Ongoing Validation
c. Back Testing
III. Quantification of IRB Systems
A. Introduction
1. Stages of the Quantification Process
2. General Principles for Sound IRB Quantification
B. Probability of Default (PD)
1. Data
2. Estimation
3. Mapping
4. Application
C. Loss Given Default (LGD)
1. Data
2. Estimation
3. Mapping
4. Application
D. Exposure at Default (EAD)
1. Data
2. Estimation
3. Mapping
4. Application
E. Maturity (M)
F. Validation
Appendix to Part III: Illustrations of the Quantification Process
IV. Data Maintenance
A. Overview
B. Data Maintenance Framework
1. Life Cycle Tracking
2. Rating Assignment Data
3. Example Data Elements
C. Data Element Functions
1. Validation and Refinement
2. Developing Parameter Estimates
3. Applying Rating System Improvements Historically
4. Calculating Capital Ratios and Reporting to the Public
5. Supporting Risk Management
D. Managing data quality and integrity
1. Documentation and Definitions
2. Electronic Storage
3. Data Gaps
V. Control and Oversight Mechanisms
A. Overview
B. Independence in the Rating Approval Process
C. Transparency
D. Accountability
1. Responsibility for Assigning Ratings
2. Responsibility for Rating System Performance
E. Use of Ratings
F. Rating System Review (RSR)
G. Internal Audit
1. External Audit
H. Corporate Oversight
I. Introduction
A. Purpose
This document describes supervisory expectations for banking
organizations (institutions) adopting the advanced internal ratings-
based approach (IRB) for the determination of minimum regulatory risk-
based capital requirements. The focus of this guidance is corporate
credit portfolios. Retail, commercial real estate, securitizations, and
other portfolios will be the focus of later guidance. This draft
guidance should be considered with the advance notice of proposed
rulemaking (ANPR) on revisions to the risk-based capital standard
published elsewhere in today's Federal Register.
The primary objective of IRB is to enhance the sensitivity of
regulatory capital requirements to credit risk. To accomplish that
objective, IRB harnesses a bank's own risk rating and quantification
capabilities. In general, the IRB approach reflects and extends recent
developments in risk management and banking supervision. However, the
degree to which any individual bank will need to modify its own credit
risk management practices to deliver accurate and consistent IRB risk
parameters will vary from institution to institution.
This guidance is intended to provide supervisors and institutions
with a clear description of the essential components and
characteristics of an acceptable IRB framework. Toward that end, this
document sets forth IRB system supervisory standards that are
highlighted in bold and designated by the prefix ``S.'' Whenever
possible, these supervisory standards are principle-based to enable
institutions to implement the framework flexibly. However, when
prudential concerns or the need for standardization override the desire
for flexibility, the supervisory standards are more detailed.
Ultimately, institutions must have credit risk management practices
that are consistent with the substance and spirit of the standards in
this guidance.
The IRB conceptual framework outlined in this document is intended
neither to dictate the precise manner by which institutions should seek
to meet supervisory expectations, nor to provide technical guidance on
how to develop such a framework. As institutions develop their IRB
systems in anticipation of adopting them for regulatory capital
purposes, supervisors will be evaluating, on an individual bank basis,
the extent to which institutions meet the standards outlined in this
document. In evaluating institutions, supervisors will rely on this
supervisory guidance as well as examination procedures, which will be
developed separately. This document assumes that readers are familiar
with the proposed IRB approach to calculating minimum regulatory
capital articulated in the ANPR.
B. Overview of Supervisory Expectations
Rigorous credit risk measurement is a necessary element of advanced
risk management. Qualifying institutions will use their internal rating
systems to associate a probability of default (PD) with each obligor
grade, as well as a loss given default (LGD) with each credit facility.
In addition, institutions will estimate exposure at default (EAD) and
will calculate the effective remaining maturity (M) of credit
facilities.
Qualifying institutions will be expected to have an IRB system
consisting of four interdependent components:
[sbull] A system that assigns ratings and validates their accuracy
(Chapter 1),
[sbull] A quantification process that translates risk ratings into
IRB parameters (Chapter 2),
[sbull] A data maintenance system that supports the IRB system
(Chapter 3), and,
[[Page 45951]]
[sbull] Oversight and control mechanisms that ensure the system is
functioning as intended and producing accurate ratings (Chapter 4).
Together these rating, quantification, data, and oversight
mechanisms present a framework for defining and improving the
evaluation of credit risk.
It is expected that rating systems will operate dynamically. As
ratings are assigned, quantified and used, estimates will be compared
with actual results and data will be maintained and updated to support
oversight and validation efforts and to better inform future estimates.
The rating system review and internal audit functions will serve as
control mechanisms that ensure that the process of ratings assignment
and quantification function according to policy and design and that
noncompliance and weaknesses are identified, communicated to senior
management and the board, and addressed. Rating systems with
appropriate data and oversight feedback mechanisms foster a learning
environment that promotes integrity in the rating system and continuing
refinement.
IRB systems need the support and oversight of the board and senior
management to ensure that the various components fit together
seamlessly and that incentives to make the system rigorous extend
across line, risk management, and other control groups. Without strong
board and senior management support and involvement, rating systems are
unlikely to provide accurate and consistent risk estimates during both
good and bad times.
The new regulatory minimum capital requirement is predicated on an
institution's internal systems being sufficiently advanced to allow a
full and accurate assessment of its risk exposures. Under the new
framework, an institution could experience a considerable capital
shortfall in the most difficult of times if its risk estimates are
materially understated. Consequently, the IRB framework demands a
greater level of validation work and controls than supervisors have
required in the past. When properly implemented, the new framework
holds the potential for better aligning minimum capital requirements
with the risk taken, pushing capital requirements higher for
institutions that specialize in riskier types of lending, and lower for
those that specialize in safer risk exposures.
Supervisors will evaluate compliance with the supervisory standards
for each of the four components of an IRB system. However, evaluating
compliance with each of the standards individually will not be
sufficient to determine an institution's overall compliance. Rather,
supervisors and institutions must also evaluate how well the various
components of an institution's IRB system complement and reinforce one
another to achieve the overall objective of accurate measures of risk.
In performing their evaluation, supervisors will need to exercise
considerable supervisory judgment, both in evaluating the individual
components and the overall IRB framework. A summary of the key
supervisory expectations for each of the IRB components follows.
Ratings Assignment
The first component of an IRB system involves the assignment and
validation of ratings (see Chapter 1). Ratings must be accurately and
consistently applied to all corporate credit exposures and be subject
to initial and ongoing validation. Institutions will have latitude in
designing and operating IRB rating systems subject to five broad
standards:
Two-dimensional risk-rating system--IRB institutions must be able
to make meaningful and consistent differentiations among credit
exposures along two dimensions--obligor default risk and loss severity
in the event of a default.
Rank order risks--IRB institutions must rank obligors by their
likelihood of default, and facilities by the loss severity expected in
default.
Calibration--IRB obligor ratings must be calibrated to values of
the probability of default (PD) parameter and loss severity ratings
must be calibrated to values of the loss given default (LGD) parameter.
Accuracy--Actual long-run actual default frequencies for obligor
rating grades must closely approximate the PDs assigned to those grades
and realized loss rates on loss severity grades must closely
approximate the LGDs assigned to those grades.
Validation process--IRB institutions must have ongoing validation
processes for rating systems that include the evaluation of
developmental evidence, process verification, benchmarking, and the
comparison of predicted parameter values to actual outcomes (back-
testing).
Quantification
The second component of an IRB system is a quantification process
(see Chapter 2). Since obligor and facility ratings may be assigned
separately from the quantification of the associated PD and LGD
parameters, quantification is addressed as a separate process. The
quantification process must produce values not only for PD and LGD but
also for EAD and for the effective remaining maturity (M). The
quantification of those four parameters is expected to be the result of
a disciplined process. The key considerations for effective
quantification are as follows:
Process--IRB institutions must have a fully specified process
covering all aspects of quantification (reference data, estimation,
mapping, and application).
Documentation--The quantification process, including the role and
scope of expert judgment, must be fully documented and updated
periodically.
Updating--Parameter estimates and related documentation must be
updated regularly.
Review--A bank must subject all aspects of the quantification
process, including design and implementation, to an appropriate degree
of independent review and validation.
Constraints on Judgment--Judgmental adjustments may be an
appropriate part of the quantification process, but must not be biased
toward lower risk estimates.
Conservatism--Parameter estimates must incorporate a degree of
conservatism that is appropriate for the overall robustness of the
quantification process.
Data Maintenance
The third component of an IRB system is an advanced data management
system that produces credible and reliable risk estimates (see Chapter
3). The broad standard governing an IRB data maintenance system is that
it supports the requirements for the other IRB system components, as
well as the institution's broader risk management and reporting needs.
Institutions will have latitude in managing their data, subject to the
following key data maintenance standards:
Life Cycle Tracking--Institutions must collect, maintain, and
analyze essential data for obligors and facilities throughout the life
and disposition of the credit exposure.
Rating Assignment Data--Institutions must capture all significant
quantitative and qualitative factors used to assign the obligor and
loss severity rating.
Support of IRB System--Data collected by institutions must be of
sufficient depth, scope, and reliability to:
[sbull] Validate IRB system processes,
[sbull] Validate parameters,
[sbull] Refine the IRB system,
[sbull] Develop internal parameter estimates,
[sbull] Apply improvements historically,
[sbull] Calculate capital ratios,
[sbull] Produce internal and public reports, and
[[Page 45952]]
[sbull] Support risk management.
Control and Oversight Mechanisms
The fourth component of an IRB system is comprised of control and
oversight mechanisms that ensure that the various components of the IRB
system are functioning as intended (see Chapter 4). Given the various
uses of internal risk ratings, including their direct link to
regulatory capital requirements, there is enormous, sometimes
conflicting, pressure on banks' internal rating systems. Control
structures are subject to the following broad standards:
Interdependent System of Controls--IRB institutions must implement
a system of interdependent controls that include the following
elements:
[sbull] Independence,
[sbull] Transparency,
[sbull] Accountability,
[sbull] Use of ratings,
[sbull] Rating system review,
[sbull] Internal audit, and
[sbull] Board and senior management oversight.
Checks and Balances--Institutions must combine the various control
mechanisms in a way that provides checks and balances for ensuring IRB
system integrity.
The system of oversight and controls required for an effective IRB
system may operate in various ways within individual institutions. This
guidance does not prescribe any particular organizational structure for
IRB oversight and control mechanisms. Banks have broad latitude to
implement structures that are most effective for their individual
circumstances, as long as those structures support and enhance the
institution's ability to satisfy the supervisory standards expressed in
this document.
C. Scope of Guidance
This draft guidance reflects work performed by supervisors to
evaluate and compare current practices at institutions with the
concepts and requirements for an IRB framework. For instances in which
a range of practice was observable, examples are provided on how
certain practices may or may not qualify. However, in many other
instances, practices were at such an early stage of development that it
was not feasible to describe specific examples. In those cases,
requirements tend to be principle-based and without examples. Given
that institutions are still in the early stages of developing
qualifying IRB systems, it is expected that this guidance will evolve
over time to more explicitly take into account new and improving
practices.
D. Timing
S. An IRB system must be operating fully at least one year prior to
the institution's intended start date for the advanced approach.
As noted in the ANPR, the significant challenge of implementing a
fully complying IRB system requires that institutions and supervisors
have sufficient time to observe whether the IRB system is delivering
risk-based capital figures with a high level of integrity. The ability
to observe the institution's ratings architecture, validation, data
maintenance and control functions in a fully operating environment
prior to implementation will help identify how well the IRB system
design functions in practice. This will be particularly important given
that in the first year of implementation institutions will not only be
subject to the new minimum capital requirements, but will also be
disclosing risk-based capital ratios for the public to rely upon in the
assessment of the institution's financial health.
II. Ratings for IRB Systems
A. Overview
This chapter describes the design and operation of risk-rating
systems that will be acceptable in an internal ratings-based (IRB)
framework. Banks will have latitude in designing and operating IRB
rating systems, subject to five broad standards:
Two-dimensional risk-rating system--IRB institutions must be able
to make meaningful and consistent differentiations among credit
exposures along two dimensions--obligor default risk and loss severity
in the event of a default.
Rank order risks--IRB institutions must rank obligors by their
likelihood of default, and facilities by the loss severity expected in
default.
Calibration--IRB obligor ratings must be calibrated to values of
the probability of default (PD) parameter and loss severity ratings
must be calibrated to values of the loss given default (LGD) parameter.
Accuracy--Actual long-run actual default frequencies for obligor
rating grades must closely approximate the PDs assigned to those grades
and actual loss rates on loss severity grades must closely approximate
the LGDs assigned to those grades.
Validation process--IRB institutions must have ongoing validation
processes for rating systems that include the evaluation of
developmental evidence, process verification, benchmarking, and the
comparison of predicted parameter values to actual outcomes (back-
testing).
B. Credit Ratings
In general, a credit rating is a summary indicator of the relative
risk on a credit exposure. Credit ratings can take many forms. The most
widely known credit ratings are the public agency ratings, which are
expressed as letters; bank internal ratings tend to be expressed as
whole numbers--for example, 1 through 10. Some rating model outputs are
expressed in terms of probability of default or expected default
frequency, in which case they may be more than relative measures of
risk. Regardless of the form, meaningful credit ratings share two
characteristics:
[sbull] They group credits to discriminate among possible outcomes.
[sbull] They rank the perceived levels of credit risk.
Banks have used credit ratings of various types for a variety of
purposes. Some ratings are intended to rank obligors by risk of default
and some are intended to rank facilities\1\ by expected loss, which
incorporates risk of default and loss severity. Bank rating systems
that are geared solely to expected loss will need to be amended to meet
the two-dimensional requirements of the IRB approach.
Rating Assignment Techniques
Banks use different techniques, such as expert judgment and models,
to assign credit risk ratings. For banks using the IRB approach, how
ratings are assigned is important because different techniques will
require different validation processes and control mechanisms to ensure
the integrity of the rating system. To assist the discussion of rating
architecture requirements, described below are some of the current
rating assignment techniques. Any of these techniques--expert judgment,
models, constrained judgment, or a combination thereof--could be
acceptable within an IRB system, provided the bank meets the standards
outlined in this document.
---------------------------------------------------------------------------
\1\ Facilities--loans, lines, or other separate extensions of
credit to an obligor.
---------------------------------------------------------------------------
Expert Judgment
Historically, banks have used expert judgment to assign ratings to
commercial credits. With this technique, an individual weighs relevant
information and reaches a conclusion about the appropriate risk rating.
Presumably, the rater makes informed judgments based on knowledge
gained through experience and training.
[[Page 45953]]
The key feature of expert-judgment systems is flexibility. The
prevalence of judgmental rating systems reflects the view that the
determinants of default are too complicated to be captured by a single
quantitative model. The quality of management is often cited as an
example of a risk determinant that is difficult to assess through a
quantitative model. In order to foster internal consistency, banks
employing expert judgment rating systems typically provide narrative
guidelines that set out ratings criteria. However, the expert must
decide how narrative guidelines apply to a given set of circumstances.
The flexibility possible in the assignment of judgmental ratings
has implications for the types of ratings review that are feasible. As
part of the ratings validation process, banks will attempt to confirm
that raters follow bank policy. However, two individuals exercising
judgment can use the same information to support different ratings.
Thus, the review of an expert judgment rating system will require an
expert who can identify the impact of policy and the impact of judgment
on a rating.
Models
In recent years, models have been developed for use in rating
commercial credits. In a model-based approach, inputs are numeric and
provide quantitative and qualitative information about an obligor. The
inputs are combined using mathematical equations to produce a number
that is translated into a categorical rating. An important feature of
models is that the rating is perfectly replicable by another party,
given the same inputs.
The models used in credit rating can be distinguished by the
techniques used to develop them. Some models may rely on statistical
techniques while others rely on expert-judgment techniques.
Statistical models. Statistically developed models are the result
of statistical optimization, in which well-defined mathematical
criteria are used to choose the model that has the closest fit to the
observed data. Numerous techniques can be used to build statistical
models; regression is one widely recognized example. Regardless of the
specific statistical technique, a knowledgeable independent reviewer
will have to exercise judgment in evaluating the reasonableness of a
model's development, including its underlying logic, the techniques
used to handle the data, and the statistical model building techniques.
Expert-derived models.\2\ Several banks have built rating models by
asking their experts to decide what weights to assign to critical
variables in the models. Drawing on their experience, the experts first
identify the observable variables that affect the likelihood of
default. They then reach agreement on the weights to be assigned to
each of the variables. Unlike statistical optimization, the experts are
not necessarily using clear, consistent criteria to select the weights
attached to the variables. Indeed, expert-judgment model building is
often a practical choice when there is not enough data to support a
statistical model building. Despite its dependence on expert judgment,
this method can be called model-based as long as the result--the
equation, most likely with linear weights--is used as the basis to rate
the credits. Once the equation is set, the model shares the feature of
replicability with statistically derived models. Generally, independent
credit experts use judgment to evaluate the reasonableness of the
development of these models.
---------------------------------------------------------------------------
\2\ Some banks have developed credit rating models that they
refer to as ``scorecards,'' but they have used expert judgment to
derive the weights. While they are models, they are not scoring
models in the now conventional use of the term. In its conventional
use, the term scoring model is reserved for a rating model derived
using statistical techniques.
---------------------------------------------------------------------------
Constrained Judgment
The alternatives just described present the extremes, but in
practice, many banks use rating systems that combine models with
judgment. Two approaches are common.
Judgmental systems with quantitative guidelines or model results as
inputs. Historically, the most common approach to rating has involved
individuals exercising judgment about risks, subject to policy
guidelines containing quantitative criteria such as minimum values for
particular financial ratios. Banks develop quantitative criteria to
guide individuals in assigning ratings, but often believe that those
criteria do not adequately reflect the information needed to assign a
rating.
One version of this constrained judgment approach features a model
output as one among several criteria that an individual may consider in
assigning ratings. The individual assigning the rating is responsible
for prioritizing the criteria, reconciling conflicts between criteria,
and if warranted, overriding some criteria. Even if individuals
incorporate model results as one of the factors in their ratings, they
will exercise judgment in deciding what weight to attach to the model
result. The appeal of this approach is that the model combines many
pieces of information into a single output, which simplifies analysis,
while the rater retains flexibility regarding the use of the model
output.
Model-based ratings with judgmental overrides. When banks use
rating models, individuals are generally permitted to override the
results under certain conditions and within tolerance levels for
frequency. Credit-rating systems in which individuals can override
models raise many of the same issues presented separately by pure
judgment and model-based systems. If overrides are rare, the system can
be evaluated largely as if it is a model-based system. If, however,
overrides are prevalent, the system will be evaluated more like a
judgmental system.
Since constrained judgment systems combine features of both expert
judgment and model-based systems, their evaluation will require the
skills required to evaluate both of these other systems.
C. IRB Ratings System Architecture
Two-Dimensional Rating System
S. IRB risk rating systems must have two rating dimensions--obligor
and loss severity ratings.
S. IRB obligor and loss severity ratings must be calibrated to
values of the probability of default (PD) and the loss given default
(LGD), respectively.
Regardless of the type of rating system(s) used by an institution,
the IRB approach imposes some specific requirements. The first
requirement is that an IRB rating system must be two-dimensional. Banks
will assign obligor ratings, which will be associated with a PD. They
will also either assign a loss severity rating, which will be
associated with LGD values, or directly assign LGD values to each
facility. The process of assigning the obligor and loss severity
ratings--hereafter referred to as the rating system--is discussed
below, and the process of calibrating obligor and loss severity ratings
to PD and LGD parameters is discussed in Chapter 2.
S. Banks must record obligor defaults in accordance with the IRB
definition of default.
Definition of Default
The consistent identification of defaults is fundamental to any IRB
rating system. For IRB purposes, a default is considered to have
occurred with regard to a particular obligor when either or both of the
two following events have taken place:
[sbull] The obligor is past due more than 90 days on any material
credit
[[Page 45954]]
obligation to the banking group. Overdrafts will be considered as being
past due once the customer has breached an advised limit or been
advised of a limit smaller than current outstandings.
[sbull] The bank considers that the obligor is unlikely to pay its
credit obligations to the banking group in full, without recourse by
the bank to actions such as liquidating collateral (if held).
Any obligor (or its underlying credit facilities) that meets one or
more of the following conditions is considered unlikely to pay and
therefore in default:
[sbull] The bank puts the credit obligation on non-accrual status.
[sbull] The bank makes a charge-off or account-specific provision
resulting from a significant perceived decline in credit quality
subsequent to the bank taking on the exposure.
[sbull] The bank sells the credit obligation at a material credit-
related economic loss.
[sbull] The bank consents to a distressed restructuring of the
credit obligation where this is likely to result in a diminished
financial obligation caused by the material forgiveness, or
postponement, of principal, interest or (where relevant) fees.
[sbull] The bank has filed for the obligor's bankruptcy or a
similar order in respect of the obligor's credit obligation to the
banking group.
[sbull] The obligor has sought or has been placed in bankruptcy or
similar protection where this would avoid or delay repayment of the
credit obligation to the banking group.
While most conditions of default currently are identified by bank
reporting systems, institutions will need to augment data capture
systems to collect those default circumstances that may not have been
traditionally identified. These include facilities that are current and
still accruing but where the obligor declared or was placed in
bankruptcy. They must also capture so called ``silent defaults''--
defaults when the loss on a facility was avoided by liquidating
collateral.
Loan sales on which a bank experiences a material loss due to
credit deterioration are considered a default. Material credit related
losses are defined as XX. (The agencies seek comment on how to define
``material'' loss in the case of loans sold at a discount). Banks
should ensure that they have adequate systems to identify such
transactions and to maintain adequate records so that reviewers can
assess the adequacy of the institution's decision-making process in
this area.
Obligor Ratings
S. Banks must assign discrete obligor grades.
While banks may use models to estimate probabilities of default for
individual obligors, the IRB approach requires banks to group the
obligors into discrete grades. Each obligor grade, in turn, must be
associated with a single PD.
S. The obligor-rating system must result in a ranking of obligors
by likelihood of default.
The proper operation of the obligor-rating system will feature a
ranking of obligors by likelihood of default. For example, if a bank
uses a rating system based on a 10-point scale, with 1 representing
obligors of highest financial strength and 10 representing defaulted
obligors, grades 2 through 9 should represent groups of ever-increasing
risk. In a rating system in which risk increases with the grade, an
obligor with a grade 4 is riskier than an obligor with a grade 2, but
need not be twice as risky.
S. Separate exposures to the same obligor must be assigned to the
same obligor rating grade.
As noted above, the IRB framework requires that the obligor rating
be distinct from the loss severity rating, which is assigned to the
facility. Collateral and other facility characteristics should not
influence the obligor rating. For example, in a 1-to-10 rating system,
where risk increases with the number grade, a defaulted borrower with a
fully cash-secured transaction should be rated a 10--defaulted--
regardless of the remote expectation of loss. Likewise, a borrower
whose financial condition warrants the highest investment grade rating
should be rated a 1 even if the bank's transactions are subordinate to
other creditors and unsecured. Since the rating is assigned to the
obligor and not the facility, separate exposures to the same obligor
must be assigned to the same obligor rating grade.
At the bottom of any IRB system rating scale is a default grade.
Once an obligor is considered to be in default for IRB purposes, that
obligor must be assigned a default grade until such time as its
financial condition and performance improve sufficiently to clearly
meet the bank's internal rating definition for one of its non-default
grades. Once an obligor is in default on any material credit obligation
to the subject bank, all of its facilities at that institution are
considered to be in default.
S. In assigning an obligor to a rating category, the bank must
assess the risk of obligor default over a period of at least one year.
S. Obligor ratings must reflect the impact of financial distress.
In assigning an obligor to a rating category, the bank must assess
the risk of obligor default over a period of at least one year. This
use of a one-year assessment horizon does not mean that a bank should
limit its consideration to outcomes for that obligor that are most
likely over that year; the rating must take into account possible
adverse events that might increase an obligor's likelihood of default.
Rating Philosophy--Decisions Underlying Ratings Architecture
S. Banks must adopt a ratings philosophy. Policy guidelines should
describe the ratings philosophy, particularly how quickly ratings are
expected to migrate in response to economic cycles.
S. A bank's capital management policy must be consistent with its
ratings philosophy in order to avoid capital shortfalls in times of
systematic economic stress.
In the IRB framework, banks assign obligors to groups that are
expected to share common default frequencies. That general description,
however, still leaves open different possible implementations,
depending on how the bank defines the set of possible adverse events
that the obligor might face. A bank must decide whether obligors are
grouped by expected common default frequency over the next year (a so-
called point-in-time rating system) or by an expected common default
frequency over a wider range of possible stress outcomes (a so-called
through-the-cycle rating system). Choosing between a point-in-time
system and a through-the-cycle system yields a rating philosophy.
In point in time rating systems, obligors are assigned to groups
that are expected to share a common default frequency in a particular
year. Point-in-time ratings change from year to year as borrowers'
circumstances change, including changes due to the economic
possibilities faced by the borrowers. Since the economic circumstances
of many borrowers reflect the common impact of the general economic
environment, the transitions in point-in-time ratings will reflect that
systematic influence. A Merton-style probability of default prediction
model is commonly believed to be an example of a point-in-time approach
to rating (although that may depend on the specific implementation of
the model).
Through-the-cycle rating systems do not ask the question, what is
the probability of default over the next year.
[[Page 45955]]
Instead, they assign obligors to groups that would be expected to share
a common default frequency if the borrowers in them were to experience
distress, regardless of whether that distress is in the next year.
Thus, as the descriptive title suggests, this rating philosophy
abstracts from the near-term economic possibilities and considers a
richer assessment of the possibilities. Like point-in-time ratings,
through the cycle ratings will change from year to year due to changes
in borrower circumstance. However, since this rating philosophy
abstracts from the immediate economic circumstance and considers the
implications of hypothetical stress circumstances, year to year
transitions in ratings will be less influenced by changes in the actual
economic environment. The ratings agencies are commonly believed to use
through-the-cycle rating approaches.
Current practice in many banks in the U.S. is to rate obligors
using an approach that combines aspects of both point-in-time and
through the cycle approaches. The explanation provided by banks that
combine those approaches is that they want rating transitions to
reflect the directional impact of changes in the economic environment,
but that they do not want all of the volatility in ratings associated
with a point-in-time approach.
Regardless of which ratings philosophy a bank chooses, an IRB bank
must articulate clearly its approach and the implications of that
choice. As part of the choice of rating philosophy, the bank must
decide whether the same ratings philosophy will be employed for all of
the bank's portfolios. And management must articulate the implications
that the bank's ratings philosophy has on the bank's capital planning
process. If a bank chooses a ratings philosophy that is likely to
result in ratings transitions that reflect the impact of the economic
cycle, its capital management policy must be designed to avoid capital
shortfalls in times of systematic economic stress.
Obligor-Rating Granularity
S. An institution must have at least seven obligor grades that
contain only non-defaulted borrowers and at least one grade to which
only defaulted borrowers are assigned.
The number of grades used in a rating system should be sufficient
to reasonably ensure that management can meaningfully differentiate
risk in the portfolio, without being so large that it limits the
practical use of the rating system. To determine the appropriate number
of grades beyond the minimum seven non-default grades, each institution
must perform its own internal analysis.
S. An institution must justify the number of obligor grades used in
its rating system and the distribution of obligors across those grades.
The mere existence of an exposure concentration in a grade (or
grades) does not, by itself, reflect weakness in a rating system. For
example, banks may focus on a particular type of lending, such as
asset-based lending, in which the borrowers may have similar default
risk. Banks with such focused lending activities may use close to the
minimum number of obligor grades, while banks with a broad range of
lending activities should have more grades. However, banks with a high
concentration of obligors in a particular grade are expected to perform
a thorough analysis that supports such a concentration.
A significant concentration within an obligor grade may be
suspected if the financial strength of the borrowers within that grade
varies considerably. If obligors seem unduly concentrated, then
management should ask themselves the following questions:
[sbull] Are the criteria for each grade clear? Those rating
criteria may be too vague to allow raters to make clear distinctions.
Ambiguity may be an issue throughout the rating scale or it may be
limited to the most commonly used ratings.
[sbull] How diverse are the obligors? That is how many market
segments (for example, large commercial, middle market, private
banking, small business, geography, etc.) are significantly represented
in the bank's borrower population? If a bank's commercial loan
portfolio is not concentrated in one market segment, its risk rating
distribution is not likely to be concentrated.
[sbull] How broad are the bank's internal rating categories
compared to those of other lenders? The bank may be able to learn
enough from publicly available information to adjust its rating
criteria.
Some banks use ``modifiers'' to provide more risk differentiation
to a given rating system. A risk rating modified with a plus, minus or
other indicator does not constitute a separate grade unless the bank
has developed a distinct rating definition and criteria for the
modified grade. In the absence of such distinctions, grades such as 5,
5+, and 5- are viewed as a single grade for regulatory capital purposes
regardless of the existence of the modifiers.
Loss Severity Ratings
S. Banks must rank facilities by the expected severity of the loss
upon default.
The second dimension of an IRB system is the loss severity rating,
which is calibrated to LGD. A facility's LGD estimate is the loss the
bank is likely to incur in the event that the obligor defaults, and is
expressed as a percentage of exposure at the time of default. LGD
estimates can be assigned either through the use of a loss severity
rating system or they can be directly assigned to each facility.
LGD analysis is still in very early stages of development relative
to default risk modeling. Academic research in this area is relatively
sparse, data are not abundant, and industry practice is still widely
varying and evolving. Given the lack of data and the lack of research
into LGD modeling, some banks are likely, as a first step, to segment
their portfolios by a handful of available characteristics and
determine the appropriate LGDs for those segments. Over time, banks'
LGD methodologies are expected to evolve. Long-standing banking
experience and existing research on LGD, while preliminary, suggests
that collateral values, seniority, industry, etc. are predictive of
loss severity.
S. Banks must have empirical support for LGD rating systems
regardless of whether they use an LGD grading system or directly assign
LGD estimates.
Whether a bank chooses to assign LGD values directly or,
alternatively, to rate facilities and then quantify the LGD for the
rating grades, the key requirement is that it will need to identify
facility characteristics that influence LGD. Each of the loss severity
rating categories must be associated with an empirically supported LGD
estimate. In much the same way an obligor-rating system ranks exposures
by the probability of default, a facility rating system must rank
facilities by the likely loss severity.
Regardless of the method used to assign LGDs (loss severity grades
or direct LGD estimation), data used to support the methodology must be
gathered systematically. For many banks, the quality and quantity of
data available to support the LGD estimation process will have an
influence on the method they choose.
Stress Condition LGDs
S. Loss severity ratings must reflect losses expected during
periods with a relatively high number of defaults.
Like obligor ratings, which group obligors by expected default
frequency, loss severity ratings assign facilities to groups that are
expected to experience a common loss severity. However, the different
treatment accorded to PD and LGD in the model used to calculate IRB
capital requirements mandates an
[[Page 45956]]
asymmetric treatment of obligor and loss severity ratings. Obligor
ratings assign obligors to groups that are expected to experience
common default frequencies across a number of years, some of which are
years of general economic stress and some of which are not. In
contrast, loss severity ratings (or estimates) must pertain to losses
expected during periods with a high number of defaults--particular
years that can be called stress conditions. For cases in which loss
severities do not have a material degree of cyclical variability, use
of a long-run default weighted average is appropriate, although stress
condition LGD generally exceeds these averages.
Loss Severity Rating/LGD Granularity
S. Banks must have a sufficiently fine loss severity grading system
or prediction model to avoid grouping facilities with widely varying
LGDs together.
While there is no stated minimum number of loss severity grades,
the systems that provide LGD estimates must be flexible enough to
adequately segment facilities with significantly varying LGDs. Banks
should have a sufficiently fine LGD grading system or LGD prediction
model to avoid grouping facilities with widely varying LGDs together.
For example, a bank using a loss severity rating-scale approach that
has credit products with a variety of collateral packages or financing
structures would be expected to have more LGD grades than those
institutions with fewer options in their credit products.
Other Considerations of IRB Rating System Architecture
Timeliness of Ratings
S. All risk ratings must be updated whenever new relevant
information is received, but must be updated at least annually.
A bank must have a policy that requires a dynamic ratings approach
ensuring that obligor and loss severity ratings reflect current
information. That policy must also specify minimum financial reporting
and collateral valuation requirements. For example, at the time of
servicing events, banks typically receive updated financial information
on obligors. For cases in which loss severity grades or estimates are
dependent on collateral values or other factors that change
periodically, that policy must take into account the need to update
these factors.
Banks' policies may include an alternative rating update timetable
for exposures below a de minimus amount that is justified by the lack
of materiality of the potential impact on capital. For example, some
banks use triggering events to prompt an update of their ratings on de
minimus exposures rather than adhering to a specific timetable.
Multiple Ratings Systems
Some banks may develop one risk-rating system that can be used
across the entire commercial loan portfolio. However, a bank can choose
to deploy any number of rating systems as long as all exposures are
assigned PD and LGD values. A different rating system could be used for
each business line and each rating system could use a different rating
scale. A bank could also use a different rating system for each
business line with each system using a common rating scale. Rating
models could be used for some portfolios and expert judgment systems
for others. An institution's complexity and sophistication, as well as
the size and range of products offered, will affect the types and
numbers of rating systems employed.
While using a number of rating systems is feasible, such a practice
might make it more difficult to meet supervisory standards. Each rating
system must conform to the standards in this guidance and must be
validated for accuracy and consistency. The requirement that each
rating systems be calibrated to parameter values imposes the ultimate
constraint, which is that ratings be applied consistently.
Recognition of the Risk Mitigation Benefits of Guarantees
S. Banks reflecting the risk-mitigating effect of guarantees must
do so by either adjusting PDs or LGDs, but not both.
S. To recognize the risk-mitigating effects of guarantees,
institutions must ensure that the written guarantee is evidenced by an
unconditional and legally enforceable commitment to pay that remains in
force until the debt is satisfied in full.
Adjustments for guarantees must be made in accordance with specific
criteria contained in the bank's credit policy. The criteria should be
plausible and intuitive, and should address the guarantor's ability and
willingness to meet its obligations. Banks are expected to gather
evidence that confirms the risk-mitigating effect of guarantees.
Other forms of written third-party support (for example, comfort
letters or letters of awareness) that are not legally binding should
not be used to adjust PD or LGD unless a bank can demonstrate through
analysis of internal data the risk-mitigating effect of such support.
Banks may not adjust PDs or LGDs to reflect implied support or verbal
assurances.
Regardless of the method used to recognize the risk-mitigating
effects of guarantees, a bank must adopt an approach that is applied
consistently over time and across the portfolio. Moreover, the onus is
on the bank to demonstrate that its approach is supported by logic and
empirical results. While guarantees may provide grounds for adjusting
PD or LGD, they cannot result in a lower risk weight than that assigned
to a similar direct obligation of the guarantor.\3\
---------------------------------------------------------------------------
\3\ The probability that an obligor and a guarantor (who
supports the obligor's debt) will both default on a debt is lower
than the probability that either the obligor or the guarantor will
default. This favorable risk-mitigation effect is known as the
reduced likelihood of ``double default.'' In determining their
rating criteria and procedures, banks are not permitted to consider
possible favorable effects of imperfect expected correlation between
default events for the borrower and guarantor for purposes of
regulatory capital requirements. Thus, the adjusted risk weight
cannot reflect the risk mitigation of double default. The ANPR
solicits public comment on the double-default issues.
---------------------------------------------------------------------------
Validation Process
S. IRB rating system architecture must be designed to ensure rating
system accuracy.
As part of their IRB rating system architecture, banks must
implement a process to ensure the accuracy of their rating systems.
Rating system accuracy is defined as the combination of the following
outcomes:
[sbull] The actual long-run average default frequency for each
rating grade is not significantly greater than the PD assigned to that
grade.
[sbull] The actual stress-condition loss rates experienced on
defaulted facilities are not significantly greater than the LGD
estimates assigned to those facilities.
Some differences across individual grades between observed outcomes
and the estimated parameter inputs to the IRB equations can be
expected. But if systematic differences suggest a bias toward lowering
regulatory capital requirements, the integrity of the rating system (of
either the PD or LGD dimensions or of both) becomes suspect. Validation
is the set of activities designed to give the greatest possible
assurances of ratings system accuracy.
S. Banks must have ongoing validation processes that include the
review of developmental evidence, ongoing monitoring, and the
comparison of predicted parameter values to actual outcomes (back-
testing).
Validation is an integral part of the rating system architecture.
Banks must have processes designed to give
[[Page 45957]]
reasonable assurances of their rating systems' accuracy. The ongoing
process to confirm and ensure rating system accuracy consists of:
[sbull] The evaluation of developmental evidence,
[sbull] Ongoing monitoring of system implementation and
reasonableness (verification and benchmarking), and
[sbull] Back-testing (comparing actual to predicted outcomes).
IRB institutions are expected to employ all of the components of
this process. However, the data to perform comprehensive back-testing
will not be available in the early stages of implementing an IRB rating
system. Therefore, banks will have to rely more heavily on
developmental evidence, quality control tests, and benchmarking to
assure themselves and other interested parties that their rating
systems are likely to be accurate. Since the time delay before rating
systems can be back-tested is likely to be an important issue--because
of the rarity of defaults in most years and the bunching of defaults in
a few years--the other parts of the validation process will assume
greater importance. If rating processes are developed in a learning
environment in which banks attempt to change and improve ratings, back
testing may be delayed even further. Validation in its early stages
will depend on bank management's exercising informed judgment about the
likelihood of the rating system working--not simply on empirical tests.
Ratings System Developmental Evidence
The first source of support for the validity of a bank's rating
system is developmental evidence. Evaluating developmental evidence
involves making a reasonable assessment of the quality of the rating
system by analyzing its design and construction. Developmental evidence
is intended to answer the question, Could the rating system be expected
to work reasonably if it is implemented as designed? That evidence will
have to be revisited whenever the bank makes a change to its rating
system. If a bank adopts a rating system and does not make changes,
this step will not have to be revisited. However, since rating systems
are likely to change over time as the bank learns about the
effectiveness of the system and incorporates the results of those
analyses, the evaluation of developmental evidence is likely to be an
ongoing part of the process. The particular steps taken in evaluating
developmental evidence will depend on the type of rating system.
Generally, the evaluation of developmental evidence will include a
body of expert opinion. For example, developmental evidence in support
of a statistical rating model must include information on the logic
that supports the model and an analysis of the statistical model-
building techniques. In contrast, developmental evidence in support of
a constrained-judgment system that features guidance values of
financial ratios might include a description of the logic and evidence
relating the values of the ratios to past default and loss outcomes.
Regardless of the type of rating system, the developmental evidence
will be more persuasive when it includes empirical evidence on how well
the ratings might have worked in the past. This evidence should be
available for a statistical model since such models are chosen to
maximize the fit to outcomes in the development sample. In addition,
statistical models should be supported by evidence that they work well
outside the development sample. Use of ``holdout'' sample evidence is a
good model-building practice to ensure that the model is not merely a
statistical quirk of the particular data set used to build the model.
Empirical developmental evidence of rating effectiveness will be
more difficult to produce for a judgmental rating system. Such evidence
would require asking raters how they would have rated past credits for
which they did not know the outcomes. Those retrospective ratings could
then be compared to the outcomes to determine whether the ratings were
correct on average. Conducting such tests, however, will be difficult
because historical data sets may not include all of the information
that an individual would have actually used in making a judgment about
a rating.
The sufficiency of the developmental evidence will itself be a
matter of informed expert opinion. Even if the rating system is model-
based, an evaluation of developmental evidence will entail judging the
merits of the model-building technique. Although no bright line tests
are feasible because expert judgment is essential to the evaluation of
rating system development, experts will be able to draw conclusions
about whether a well-implemented system would be likely to perform
satisfactorily.
Ratings System Ongoing Validation
The second source of analytical support for the validity of a bank
rating system is the ongoing analysis intended to confirm that the
rating system is being implemented and continues to perform as
intended. Such analysis involves process verification and benchmarking.
Process Verification
Verification activities address the question, Are the ratings being
assigned as intended? Specific verification activities will depend on
the rating approach. If a model is used for rating, verification
analysis begins by confirming that the computer code used to deploy the
model is correct. The computer code can be verified in a number of
established ways. For example, a qualified expert can duplicate the
code or check the code line by line. Process verification for a model
will also include confirmation that the correct data are being used in
the model.
For expert-judgment and constrained-judgment systems, verification
requires other individual reviewers to evaluate whether the rater
followed rating policy. The primary requirements for verification of
ratings assigned by individuals are:
[sbull] A transparent rating process,
[sbull] A database with information used by the rater, and
[sbull] Documentation of how the decisions were made.
The specific steps will depend on how much the process incorporates
specific guidelines and how much the exercise of judgment is allowed.
As the dependence on specific guidelines increases, other individuals
can more easily confirm that guidelines were followed by reference to
sufficient documentation. As the dependence on judgment rises, the
ratings review function will have to be staffed increasingly by experts
with appropriate skills and knowledge about the rating policies of the
bank.
Ratings process verification also includes override monitoring. If
individuals have the ability to override either models or policies in a
constrained-judgment system, the bank should have both a policy stating
the tolerance for overrides and a monitoring system for identifying the
occurrence of overrides. A reporting system capturing data on reasons
for overrides will facilitate learning about whether overrides improve
accuracy.
Benchmarking
S. Banks must benchmark their internal ratings against internal,
market and other third-party ratings.
Benchmarking is the set of activities that uses alternative tools
to draw inferences about the correctness of ratings before outcomes are
actually
[[Page 45958]]
known. The most important type of benchmarking of a rating system is to
ask whether another rater or rating method attaches the same rating to
a particular obligor or facility. Regardless of the rating approach,
the benchmark can be either a judgmental or a model-based rating.
Examples of such benchmarking include:
[sbull] Ratings reviewers who completely re-rate a sample of
credits rated by individuals in a judgmental system.
[sbull] An internally developed model is used to rate credits rated
earlier in a judgmental system.
[sbull] Individuals rate a sample of credits rated by a model.
[sbull] Internal ratings are compared against results from external
agencies or external models.
Because it will take considerable time before outcomes will be
available, using alternative ratings as benchmarks will be a very
important validation device. Such benchmarking must be applied to all
rating approaches, and the benchmark can be either a model or judgment.
At a minimum, banks must establish a process in which a representative
sample of its internal ratings is compared to third-party ratings
(e.g., independent internal raters, external rating agencies, models,
or other market data sources) of the same credits.
Benchmarking also includes activities designed to draw broader
inferences about whether the rating system--as opposed to individual
ratings--is working as expected. The bank can look for consistency in
ranking or consistency in the values of rating characteristics for
similarly rated credits. Examples of such benchmarking activities
include:
[sbull] Analyzing the characteristics of obligors that have
received common ratings.
[sbull] Monitoring changes in the distribution of ratings over
time.
[sbull] Calculating a transition matrix calculated from changes in
ratings in a bank's portfolio and comparing it to historical transition
matrices from internal bank data or publicly available ratings.
While benchmarking activities allow for inferences about the
correctness of the ratings system, they are the not same thing as back-
testing. The benchmark itself is a prediction and may be in error. If
benchmarking evidence suggests a pattern of rating differences, it
should lead the bank to investigate the source of the differences.
Thus, the benchmarking process illustrates the possibility of feedback
from ongoing validation to model development, underscoring the
characterization of validation as a process.
Back Testing
S. Banks must develop statistical tests to back-test their IRB
rating systems.
S. Banks must establish internal tolerance limits for differences
between expected and actual outcomes.
S. Banks must have a policy that requires remedial actions be taken
when policy tolerances are exceeded.
The third component of a validation process is back-testing, which
is the comparison of predictions with actual outcomes. Back-testing of
IRB systems is the empirical test of the accuracy of the parameter
values, PD and LGD, associated with obligor and loss severity ratings,
respectively. For IRB rating systems, back-testing addresses the
combined effectiveness of the assignment of obligor and loss severity
ratings and the calibration of the parameters PD and LGD attached to
those ratings.
At this time, there is no generally agreed-upon statistical test of
the accuracy of IRB systems. Banks must develop statistical tests to
back-test their IRB rating systems. In addition, banks must have a
policy that specifies internal tolerance limits for comparing back-
testing results. Importantly, that policy must outline the actions that
would be taken whenever policy limits are exceeded.
As a combined test of ratings effectiveness, back-testing is a
conceptual bridge between the ratings system architecture discussed in
this chapter and the quantification of parameters, discussed in Chapter
2. The final section of Chapter 2 discusses back-testing as one type of
quantitative test required to validate the quantification of parameter
values.
III. Quantification of IRB Systems
Ratings quantification is the process of assigning numerical values
to the four key components for internal ratings-based assessments of
credit-risk capital: probability of default (PD), the expected loss
given default (LGD), the expected exposure at default (EAD), and
maturity (M). Section I establishes an organizing framework for
considering IRB quantification and develops general principles that
apply to the entire process. Sections II through IV cover specific
principles or supervisory standards that apply to PD, LGD, and EAD
respectively. The maturity component, which is much less dependent on
statistical estimates and the use of data, receives somewhat different
treatment in section V. Validation of the quantification process is
covered in section VI.
A. Introduction
Stages of the Quantification Process
With the exception of maturity, the risk components are
unobservable and must be estimated. The estimation must be consistent
with sound practice and supervisory standards. In addition, a bank must
have processes to ensure that these estimates remain valid.
Calculation of risk components for IRB involves two sets of data:
the bank's actual portfolio data, consisting of current credit
exposures assigned to internal grades, and a ``reference data set,''
consisting of a set of defaulted credits (in the case of LGD and EAD
estimation) or both defaulted and non-defaulted credits (in the case of
PD estimation). The bank estimates a relationship between the reference
data set and probability of default, loss severity, or exposure; then
this estimated relationship is applied to the actual portfolio data for
which capital is being assessed.
Quantification proceeds through four logical stages: obtaining
reference data; estimating the reference data's relationship to the
parameters; mapping the correspondence between the reference data and
the portfolio's data; and applying the relationship between reference
data and parameters to the portfolio's data. (Readers may find it
helpful to refer to the appendix to this chapter, which illustrates how
this four-stage framework can be applied to ratings quantification
approaches in practice.) An evaluation of any bank's IRB quantification
process focuses on understanding how the bank implements each stage for
each of the key parameters, and on assessing the adequacy of the bank's
approach.
Data--First, the bank constructs a reference data set, or source of
data, from which parameters can be estimated.
Reference data sets include internal data, external data, and
pooled internal/external data. Important considerations include the
comparability of the reference data to the current credit portfolio,
whether the sample period ``appropriately'' includes periods of stress,
and the definition of default used in the reference data. The reference
data must be described using a set of observed characteristics;
consequently, the data set must contain variables that can be used for
this characterization. Relevant characteristics might include external
debt ratings, financial measures, geographic regions, or any other
factors that are believed to be
[[Page 45959]]
related in some way to PD, LGD, or EAD. More than one reference data
set may be used.
Estimation--Second, the bank applies statistical techniques to the
reference data to determine a relationship between characteristics of
the reference data and the parameters (PD, LGD, or EAD).
The result of this step is a model that ties descriptive
characteristics of the obligor or facility in the reference data set to
PD, LGD, or EAD estimates. In this context, the term `models' is used
in the most general sense; a model may be simple, such as the
calculation of averages, or more complicated, such as an approach based
on advanced regression techniques. This step may include adjustments
for differences between the IRB definition of default and the default
definition in the reference data set, or adjustments for data
limitations. More than one estimation technique may be used to generate
estimates of the risk components, especially if there are multiple sets
of reference data or multiple sample periods.
Mapping--Third, the bank creates a link between its portfolio data
and the reference data based on common characteristics.
Variables or characteristics that are available for the current
portfolio must be mapped to the variables used in the default, loss-
severity, or exposure model. (In some cases, the bank constructs the
link for a representative exposure in each internal grade, and the
mapping is then applied to all credits within a grade.) An important
element of mapping is making adjustments for differences between
reference data sets and the bank's portfolio. The bank must create a
mapping for each reference data set and for each combination of
variables used in any estimation model.
Application--Fourth, the bank applies the relationship estimated
for the reference data to the actual portfolio data.
The ultimate aim of quantification is to attribute a PD, LGD, or
EAD to each exposure within the portfolio, or to each internal grade if
the mapping was done at the grade level. This step may include
adjustments to default frequencies or loss rates to ``smooth'' the
final parameter estimates. If the estimates are applied to individual
transactions, the bank must in some way aggregate the estimates at the
grade level. In addition, if multiple data sets or estimation methods
are used, the bank must adopt a means of combining the various
estimates.
A number of examples are given in this chapter to aid exposition
and interpretation. None of the examples is sufficiently detailed to
incorporate all the considerations discussed in this chapter. Moreover,
technical progress in the area of quantification is rapid. Thus, banks
should not interpret an example that is consistent with the standard
being discussed, and that resembles the bank's current practice, as
creation of a ``safe harbor'' or as an indication that the bank's
practice will be approved as-is. Banks should consider this guidance in
its entirety when determining whether systems and practices are
adequate.
General Principles for Sound IRB Quantification
Several core principles apply to all elements of the overall
ratings quantification process; those general principles are discussed
in this introductory section. Each of these principles is, in effect, a
supervisory standard for IRB systems. Other supervisory standards,
specific to particular elements or parameters, are discussed in the
relevant sections.
Supervisory evaluation of IRB quantification requires consideration
of all of these principles and standards, both general and specific.
Particular practical approaches to ratings quantification may be highly
consistent with some standards, and less so with others. In any
particular case, an ultimate assessment relies on the judgment of
supervisors to weigh the strengths and weaknesses of a bank's chosen
approach, using these supervisory standards as a guide.
S. IRB institutions must have a fully specified process covering
all aspects of quantification (reference data, estimation, mapping, and
application). The quantification process, including the role and scope
of expert judgment, must be fully documented and updated periodically.
A fully specified quantification process must describe how all four
stages (data, estimation, mapping, and application) are implemented for
each parameter. Documentation promotes consistency and allows third
parties to review and replicate the entire process. Examples of third
parties that might use the documentation include rating-system
reviewers, auditors, and bank supervisors. Periodic updates to the
process must be conducted to ensure that new data, analytical
techniques, and evolving industry practice are incorporated into the
quantification process.
S. Parameter estimates and related documentation must be updated
regularly.
The parameter estimates must be updated at least annually, and the
process for doing so must be documented in bank policy. The update
should also evaluate the judgmental adjustments embedded in the
estimates; new data or techniques may suggest a need to modify those
adjustments. Particular attention should be given to new business lines
or portfolios in which the mix of obligors is believed to have changed
substantially. A material merger, acquisition, divestiture, or exit
clearly raises questions about the continued applicability of the
process and should trigger an intensive review and updating.
The updating process is particularly relevant for the reference
data stage because new data become available all the time. New data
must be incorporated, into the PD, LGD, and EAD estimates, using a
well-defined process.
S. A bank must subject all aspects of the quantification process,
including design and implementation, to an appropriate degree of
independent review and validation.
An independent review is an assessment conducted by persons not
accountable for the work being reviewed. The reviewers may be either
internal or external parties. The review serves as a check that the
quantification process is sound and works as intended; it should be
broad-based, and must include all of the elements of the quantification
process that lead to the ultimate estimates of PD, LGD, and EAD. The
review must cover the full scope of validation: evaluation of the
integrity of data inputs, analysis of the internal logic and
consistency of the process, comparison with relevant benchmarks, and
appropriate back-testing based on actual outcomes.
S. Judgmental adjustments may be an appropriate part of the
quantification process, but must not be biased toward lower estimates
of risk.
Judgment will inevitably play a role in the quantification process
and may materially affect the estimates. Judgmental adjustments to
estimates are often necessary because of some limitations on available
reference data or because of inherent differences between the reference
data and the bank's portfolio data. The bank must ensure that
adjustments are not biased toward optimistically low parameter
estimates for PD, LGD, and EAD. Individual assumptions are less
important than broad patterns; consistent signs of judgmental decisions
that lower parameter estimates materially may be evidence of bias.
[[Page 45960]]
The reasoning and empirical support for any adjustments, as well as
the mechanics of the calculation, must be documented. The bank should
conduct sensitivity analysis to demonstrate that the adjustment
procedure is not biased toward reducing capital requirements. The
analysis must consider the impact of any judgmental adjustments on
estimates and risk weights, and must be fully documented.
S. Parameter estimates must incorporate a degree of conservatism
that is appropriate for the overall robustness of the quantification
process.
In estimating values of PD, LGD, and EAD should be as precise and
accurate as possible. However, estimates of PD, LGD and EAD are
statistics, and thus inherently subject to uncertainty and potential
error. It is often possible to be reasonably confident that a risk
component or other parameter lies within a particular range, but
greater precision is difficult to achieve. Aspects of the ratings
quantification process that are apt to introduce uncertainty and
potential error include the following:
The estimation of coefficients of particular variables in a
regression-based statistical default or severity model.
[sbull] The calculation of average default or loss rates for
particular categories of credits in external default databases.
[sbull] The mapping between portfolio obligors or facilities and
reference data when the set of common characteristics does not align
exactly.
A general principle of the IRB approach is that a bank must adjust
estimates conservatively in the presence of uncertainty or potential
error. In many cases this corresponds to assigning a final parameter
estimate that increases required capital relative to the best estimate
produced through sound-practice estimation techniques. The extent of
this conservative adjustment should be related to factors such as the
relevance of the reference data, the quality of the mapping, the
precision of the statistical estimates, and the amount of judgment used
throughout the process. Margins of conservatism need not be added at
each step; indeed, that could produce an excessively conservative
result. The overall margin of conservatism should adequately account
for all uncertainties and weaknesses; this is the general
interpretation of requirements to incorporate appropriate degrees of
conservatism. Improvements in the quantification process (use of better
data, estimation techniques, and so on) may reduce the appropriate
degree of conservatism over time.
Estimates of PD, LGD, EAD, or other parameters or coefficients
should be presented with an accompanying sense of the statistical
precision of the estimates; this facilitates an assessment of the
appropriate degree of conservatism.
B. Probability of Default (PD)
Data
To estimate PD accurately, a bank must have a comprehensive
reference data set with observations that are comparable to the bank's
current portfolio of obligors. Clearly, the data set used for
estimation should be similar to the portfolio to which such estimates
will be applied. The same comparability standard applies to both
internal and external data sets.
To ensure ongoing applicability of the reference data, a bank must
assess the characteristics of its current obligors relative to the
characteristics of obligors in the reference data. Such variables might
include qualitative and quantitative obligor information, internal and
external rating, rating dates, and line of business or geography. To
this end, a bank must maintain documentation that fully describes all
explanatory variables in the data set, including any changes to those
variables over time. A well-defined and documented process must be in
place to ensure that the reference data are updated as frequently as is
practical, as fresh data become available or portfolio changes make
necessary.
S. The sample for the reference data must be at least five years,
and must include periods of economic stress during which default rates
were relatively high.
To foster more robust estimation, banks should use longer time
series when more than five years of data are available. However, the
benefits of using a longer time series (longer than five years) may
have to be weighed against a possible loss of data comparability. The
older the reference data, the less similar they are likely to be to the
bank's current portfolio; striking the correct balance is a matter of
judgment. Reference obligors must not differ from the current portfolio
obligors systematically in ways that seem likely to be related to
obligor default risk. Otherwise, the derived PD estimates may not be
applicable to the current portfolio.
Note that this principle does not simply restate the requirement
for five years of data: periods of stress during which default rates
are relatively high must be included in the data sample. Exclusion of
such periods biases PD estimates downward and unjustifiably lowers
regulatory capital requirements.
Example. A bank's reference data set covers the years 1987
through 2001. Each year includes identical data elements, and each
year is similarly populated. For its grade PD estimates, the bank
relies upon data from a sub-sample covering 1992 through 2001. The
bank provides no justification for dropping the years from 1987
through 1991. The bank contends that it is not necessary to include
those data, as the reference sample they use for estimation
satisfies the five-year requirement. This practice is not consistent
with the standard because the bank has not supported its decision to
ignore available data. The fact that the excluded years include a
recession would raise particular concerns.
S. The definition of default within the reference data must be
reasonably consistent with the IRB definition of default.
Regardless of the source of the reference data, a bank must apply
the same default definition throughout the quantification processes.
This fosters consistent estimation across parameters and reduces the
potential for undesired bias. In addition, consistent application of
the same definition across banks will permit true horizontal analysis
by supervisors and engaged market participants.
This standard applies to both internal and external reference data.
For internal data, a bank's default definition is expected to be
consistent with the IRB definition going forward. Banks will be
expected to make appropriate adjustments to their data systems such
that all defaults as defined for IRB are captured by the time a bank
fully implements its IRB system. For any historical or external data
that do not fully comply with the IRB definition of default, a bank
must make conservative adjustments to reflect such discrepancies.
Larger discrepancies require larger adjustments for conservatism.
Example. To identify defaults in its historical data, a bank
applies a consistent definition of ``placed on nonaccrual.'' This
definition is used in the bank's quantification exercises to
estimate PD, LGD, and EAD. The bank recognizes that use of the
nonaccrual definition fails to capture certain defaults as
identified in the IRB rules. Specifically, the bank indicates that
the following kinds of defaulted facilities would not have been
placed on nonaccrual: (1) Credit obligations that were sold at a
material credit-related economic loss, and (2) distressed
restructurings. To be consistent with the standard, the bank must
make a well-supported adjustment to its grade PD estimates to
reflect the difference in the default definitions.
Estimation
Estimation of PD is the process by which characteristics of the
reference
[[Page 45961]]
data are related to default frequencies.\4\ The relevant
characteristics that help to determine the likelihood of default are
referred to as ``drivers of default''. Drivers might include variables
such as financial ratios, management expertise, industry, and
geography.
---------------------------------------------------------------------------
\4\ The New Basel Capital Accord produced by the Basel Committee
on Banking Supervision discusses three techniques for PD estimation.
IRB banks are not constrained to select from among these three
techniques; they have broad flexibility to implement appropriate
approaches to quantification. The three Basel techniques are best
regarded not as a complete taxonomy of the possible approaches to PD
estimation, but rather as illustrations of a few of the many
possible approaches.
---------------------------------------------------------------------------
S. Estimates of default rates must be empirically based and must
represent a long-run average.
Estimates must capture average default experience over a reasonable
mix of high-default and low-default years of the economic cycle. The
average is labeled ``long-run'' because a long observation period would
span both peaks and valleys of the economic cycle. The emphasis should
not be on time-span; the long-run average concept captures the breadth,
not the length, of experience.
If the reference data are characterized by internal or external
rating grades, one estimation approach is to calculate the mean of one-
year realized default rates for each grade, giving equal weight to each
year's realized default rate. PD estimates generally should be
calculated in this manner.
Another approach is to pool obligors in a given grade over a number
of years and then calculate the mean default rate. In this case, each
year's default rate is weighted by the number of obligors. This
approach may underestimate default rates. For example, if lending
declines in recessions so that obligors are fewer in those years than
in others, weighting by number of obligors would dilute the effect of
the recession year on the overall mean. The obligor-weighted
calculation, or another approach, will be allowed only if the bank can
demonstrate that this approach provides a better estimate of the long-
run average PD. At a minimum, this would involve comparing the results
of both methods.
Statistical default prediction models may also play a role in PD
estimation. For example, the characteristics of the reference data
might include financial ratios or a distance-to-default measure, as
defined by a specific implementation of a Merton-style structural
model.
For a model-based approach to meet the requirement that ultimate
grade PD estimates be long-run averages, the reference data used in the
default model must meet the long-run requirement. For example, a model
can be used to relate financial ratios to likelihood of default based
on the outcome for the firms--default or non-default. Such a model must
be calibrated to capture the default experience over a reasonable mix
of good and bad years of the economic cycle. The same requirement would
hold for a structural model; distance to default must be calibrated to
default frequency using long-run experience. This applies to both
internal and vendor models, and a bank must verify that this
requirement is met.
Example 1. A bank uses external data from a rating agency to
estimate PD. The PD estimate for each agency grade is calculated as
the mean of yearly realized default rates over a time period (1980
through 2001) that includes several recessions and high-default
years. The bank provides support that this time period adequately
represents long-run experience. This illustrates an estimation
method that is consistent with the standard.
Example 2a. Like the institution in example 1, a bank maps
internal ratings to agency grades. The estimates for the agency
grades are set indirectly, using the default probabilities from a
default prediction model. The bank does so because although it links
internal and agency grades, the bank views the default model's
results as more predictive than the historical agency default
experience. For each agency grade, the bank calculates a PD estimate
as the mean of the model-based default probabilities for the agency-
rated obligors. In order to meet the long-run requirement, the bank
calculates the estimates over the seven years from 1995 through
2001. The bank demonstrates that this time period includes a
reasonable mix of high-default and low-default experience. This
estimation method is consistent with the standard.
Example 2b. In a variant of example 2a, a bank uses the mean
default frequency per agency rating grade for a single year, such as
2001. Empirical evidence shows that the mean default frequency for
agency grades varies substantially from year to year. A single year
thus does not reflect the full range of experience, because a long-
run average should be relatively stable year to year. Such
instability makes this estimation method unacceptable.
Example 2c. Another bank calculates the agency grade PD
estimates as the median default probability of companies in that
grade. The bank does so without demonstrating that the median is a
better statistical estimator than the mean. This estimation method
is not consistent with the standard. A median gives less weight to
obligors with high estimated default probabilities than a simple
mean does. The difference between mean and median can be material
because distributions of credits within grades often are
substantially skewed toward higher default probabilities: the
riskier obligors within a grade tend to have individual default
probabilities that are substantially worse than the median, while
the least risky have default probabilities only somewhat better than
the median.
S. Judgmental adjustments may play an appropriate role in PD
estimation, but must not be biased toward lower estimates.
The following examples illustrate how supervisors will evaluate
adjustments:
Example 1. A bank uses the last five years of internal default
history to estimate grade PDs. However, they recognize that the
internal experience does not include any high-default years. In
order to remedy this and still take advantage of its experience, the
bank uses external agency data to adjust the estimates upward. Using
the agency data, the bank calculates the ratio between the long-run
average and the mean default rate per grade over the last five
years. The bank assumes that the relationship observed in the agency
data applies to its portfolio, and adjusts the estimates for the
internal data accordingly. This practice is consistent with the
standard.
Example 2. A bank uses internal default experience to estimate
grade PDs. However, the bank has historically failed to recognize
defaults when the loss on the default obligation was avoided by
seizing collateral. The bank makes no adjustment for such missing
defaults. The realized default rate using the more inclusive
definition would be higher than that observed by the bank (and loss
severity rates would be correspondingly lower). This practice would
not be consistent with the standard, unless the bank demonstrates
that the necessary adjustment is immaterial.
Mapping
Mapping is the process of establishing a correspondence between the
bank's current obligors and the reference obligor data used in the
default model. Hence, mapping involves identifying how default-related
characteristics of the current portfolio correspond to the
characteristics of reference obligors. Such characteristics might
include financial and nonfinancial variables, and assigned ratings or
grades.
Mapping can be thought of as taking each obligor in the bank's
portfolio and characterizing it as if it were part of the reference
data. There are two broad approaches to the mapping process:
Obligor mapping: Each portfolio obligor is mapped to the reference
data based on its individual characteristics. For example, if a bank
applies a default model, a default probability will be generated for
each obligor. That individual default probability is then used to
assign each obligor to a particular internal grade, based on the bank's
established criteria. To obtain a final estimate of the grade PD in the
subsequent application stage, the bank averages the default
probabilities of individual obligors within each grade.
Grade mapping: Characteristics of the obligors within an internal
grade are
[[Page 45962]]
averaged or otherwise summarized to construct a ``typical'' or
representative obligor for each grade. Then, the bank maps that
representative obligor to the reference data. For example, if the bank
uses a default model, the default probability associated with that
typical obligor will serve as the grade PD in the application stage.
Alternatively, the bank may map the typical obligor to a particular
external rating grade based on quantitative and qualitative
characteristics, and assign the long-run default rate for that rating
to the internal grade in the application stage.
Either grade mapping or obligor mapping can be part of the
quantification process; either method can produce a single PD estimate
for each grade in the application stage. However, in the absence of
other compelling considerations, banks should use obligor mapping for
two reasons:
[sbull] First, default probabilities are nonlinear under many
estimation approaches. As a result, the default probability of the
typical obligor--the result of a grade mapping approach--is often lower
than the mean of the individual obligor default probabilities from the
obligor mapping approach. For example, consider a bank that maps to the
S&P scale and uses historical S&P bond default rates. For ease of
illustration, suppose that one internal grade contains only three
obligors that individually map to BB, BB-, and B+. The historical
default rates for these three grades are 1.07, 1.76, and 3.24 percent,
respectively (based on 1981-2001 data). Using obligor mapping, those
rates would be assigned directly to the three obligors, yielding a mean
PD of 2.02 percent for the grade. Using grade mapping, the grade PD
would be only 1.76, because the grade's typical obligor is rated BB-.
[sbull] Second, a hypothetical obligor with a grade's average
characteristics may not represent well the risks presented by the
grade's typical obligor. For example, a bank might observe that
obligors with high leverage and low earnings variability have about the
same default risk as obligors with low leverage and high earnings
variability. These two types of obligors might both end up in the same
grade, for example, Grade 6. If so, the typical obligor in Grade 6
would have moderate leverage and moderate earnings variability--a
combination that might fail to reflect any of the individual obligors
in Grade 6, and that could easily result in a PD for the grade that is
too low.
A bank electing to use grade mapping instead of obligor mapping
should be especially careful in choosing a ``typical'' obligor for each
grade. Doing so typically requires that the bank examine the actual
distribution of obligors within each grade, as well as the
characteristics of those obligors. Banks should be aware that different
measures of central tendency (such as mean, median, or mode) will give
different results, and that these different results may have a material
effect on a grade's PD; they must be able to justify their choice of a
measure. Banks must have a clear and consistent policy toward the
calculation.
S. The mapping must be based on a robust comparison of available
data elements that are common to the portfolio and the reference data.
Sound mapping practice uses all common elements that are available
in the data as the basis for mapping. If a bank chooses to ignore
certain common variables or to weight some variables more heavily than
others, those choices must be supported. Mapping should also take into
account differences in rating philosophy (for example, point-in-time or
through-the-cycle) between any ratings embedded in the reference data
set and the bank's own rating regime.
A mapping should be plausible, and should be consistent with the
rating philosophy established by the bank as part of its obligor rating
policy. For a bank that uses grade mapping, levels and ranges of key
variables within each internal grade should be close to values of
similar variables for corresponding obligors within the reference data.
The standard allows for use of a limited set of common variables
that are predictive of default risk, in part to permit flexibility in
early years when data may be far from ideal. Nevertheless, banks will
eventually be expected to use variables that are widely recognized as
the most reliable predictors of default risk in mapping exercises. In
the meantime, banks relying on data elements that are weak predictors
must compensate by making their estimates more conservative. For
example, leverage and cash flow are widely recognized to be reliable
predictors of corporate default risk. Borrower size is also predictive,
but less so. A mapping based solely on size is by nature less reliable
than one based on leverage, cash flow, and size.
Example 1. In estimating PD, a bank relies on observed default
rates on bonds in various agency grades for PD quantification. To
map its internal grades to the agency grades, the bank identifies
variables that together explain much of the rating variation in the
bond sample. The bank then conducts a statistical analysis of those
same variables within its portfolio of obligors, using a
multivariate distance calculation to assign each portfolio obligor
to the external rating whose characteristics it matches most closely
(for example, assigning obligors to ratings so that the sum of
squared differences between the external grade averages and the
obligor's characteristics is minimized). This practice is broadly
consistent with the standard.
Example 2. A bank uses grade mapping to link portfolio obligors
to the reference data set described by agency ratings. The bank
looks at publicly rated portfolio obligors within an internal grade
to determine the most common external rating, does the same for all
grades, and creates a correspondence between internal and external
ratings. The strength of the correspondence is a function of the
number of externally rated obligors within each grade, the
distribution of those external ratings within each grade and the
similarity of externally rated obligors in the grade to those not
externally rated. This practice is broadly consistent with this
standard, but would require a comparison of rating philosophies and
may require adjustments and the addition of margins of conservatism.
S. A mapping process must be established for each reference data
set and for each estimation model.
Banks should never assume that a mapping is self-evident. Even a
rating system that has been explicitly designed to replicate external
agency ratings may or may not be effective in producing a replica;
formal mapping is still necessary. Indeed, in such a system the kind of
analysis involved in mapping may help identify inconsistencies in the
rating process itself.
A mapping process is needed even where the reference obligors come
from internal historical experience. Banks must not assume that
internal data do not require mapping, because changes in bank strategy
or external economic forces may alter the composition of internal
grades or the nature of the obligors in those grades over time.
Mappings must be reaffirmed regardless of whether rating criteria or
other aspects of the ratings system have undergone explicit changes
during the period covered by the reference data set.
Banks often use multiple reference data sets, and then combine the
resulting estimates to get a grade PD. A bank that does that must
conduct a rigorous mapping process for each data set.
Supervisors expect all meaningful characteristics of obligors to be
factored directly into the rating process; this should include
characteristics like the obligor's industry or physical location. But
in some circumstances, certain effects related to industry, geography,
or other factors are not reflected in rating assignments or default
estimates. In such cases, it may be appropriate for banks to capture
the impact of the
[[Page 45963]]
omissions by using different mappings for different business lines or
types of obligors. Supervisors expect this practice to be transitional;
banks will eventually be required to incorporate the omitted effects
into the rating system and the estimation process as they are uncovered
and documented, rather than adjusting the mapping.
Example 1. The bank maps its internal grades carefully to one
rating agency, and then assumes a correspondence to another agency's
scale despite known differences in the rating methods of the two
agencies. The bank then applies a mean of the grade default rates
from these two public debt-rating agencies to its internal grades.
This practice is not consistent with the standard, because the bank
should map to each agency's scale separately.
Example 2. A bank uses internal historical data as its reference
data. The bank computes a mean default rate for each grade as the
grade PD for capital purposes, and asserts that mapping is
unnecessary because ``its strong credit culture ensures that a 4 is
always a 4.'' This practice is not consistent with the standard,
because no mapping has been done; there is no assurance that a
representative obligor in a grade today is comparable to an obligor
in that same grade in the past.
S. The mapping must be updated and independently validated
regularly.
The appropriate mapping between a bank's portfolio and the
reference data may change over time. For example, relationships between
internal grades and external agency grades may change during the
economic cycle because of differences in rating philosophy. Similarly,
distance-to-default measures for obligors in a given grade may not be
constant over time. These likely changes make it imperative that the
bank update all mappings regularly.
Sound validation practices may include tests for internal
consistency such as ``reverse mapping.'' Using this technique, a bank
evaluates obligors from the reference data set as if they were subject
to the bank's rating system (that is, part of the bank's current
portfolio). The bank's mapping is then applied to these reverse-mapped
obligors to see whether the mapped characterization of the reference
obligor is consistent with that of the initial evaluation.\5\ Another
valuable technique is to apply different mapping methods and compare
the results. For example, mappings based on financial ratio comparisons
can be rechecked using mappings based on available external ratings.
---------------------------------------------------------------------------
\5\ For example, suppose a bank asserts that its Grade 3
corresponds to an S&P rating of A. Applying reverse mapping, the
bank would take a sample of A-rated obligors from the reference
data, run them through the bank's rating process (perhaps a
simplified version), and check to see that those obligors usually
receive a grade of 3 on the bank's internal scale.
Example. A bank mapped its internal grades to the rating scale
of one public debt-rating agency in 1992. Since then, the bank has
completed a major acquisition of another large bank and
significantly changed its business mix in other ways. The bank
continues to use the same mapping, without reassessing its validity.
This practice is not consistent with the standard.
Application
In the application stage, the bank applies the PD estimation method
to the current portfolio of obligors using the mapping process. It
obtains final PD estimates for each rating grade, which will be used to
calculate minimum regulatory capital. To arrive at those estimates, a
bank may adjust the raw results derived from the estimation stage. For
example, it might aggregate individual obligor default probabilities to
the rating grade level, or smooth results because a rating grade's PD
estimate was higher than a lower quality grade. The bank must explain
and support all adjustments when documenting its quantification
process.
Example. A bank uses external data to estimate long-run average
PDs for each grade. The resulting PD estimate for Grade 2 is
slightly higher than the estimate for Grade 3, even though Grade 2
is supposedly of higher credit quality. The bank uses statistics to
demonstrate that this anomaly occurred because defaults are rare in
the highest quality rating grades. The bank judgmentally adjusts the
PD estimates for grades 2 and 3 to preserve the expected
relationship between obligor grade and PD, but requires that total
risk-weighted assets across both grades using the adjusted PD
estimates be no less than total risk-weighted assets based on the
unadjusted estimates, using a typical distribution of obligors
across the two grades. Such an adjustment during the application
stage is consistent with this guidance.
S. IRB institutions that aggregate the default probabilities of
individual portfolio obligors when calculating PD estimates for
internal grades must have a clear policy governing the aggregation
process.
As noted above, mapping may be grade-based or obligor-based. Grade-
based mappings naturally provide a single PD per grade, because the
estimated default model is applied to the representative obligor for
each grade. In contrast, obligor-based mappings must aggregate in some
manner the individual PD estimates to the grade level. The expectation
is that the grade PD estimate will be calculated as the mean. The bank
will be allowed to calculate this estimate differently only if it can
demonstrate that the alternative method provides a better estimate of
the long-run average PD. To obtain this evidence, the bank must at
least compare the results of both methods.
S. IRB institutions that combine estimates from multiple sets of
reference data must have a clear policy governing the combination
process, and must examine the sensitivity of the results to alternative
combinations.
Because a bank should make use of as much information as possible
when mapping, it will usually use multiple data sets. The manner in
which the data or the estimates from those multiple data sets are
combined is extremely important. A bank must document its justification
for the particular combination methods selected. Those methods must be
subject to appropriate approval and oversight.
The data may come from the same basic data source but from
different time periods or from different data sources altogether. For
example, banks often combine internal data with external data, use
external data from different sample periods, or combine results from
corporate-bond default databases with results from equity-based models
of obligor default. Different combinations will produce different PD
estimates. The bank should investigate alternative combinations and
document the impact on the estimates. When ultimate results are highly
sensitive to how estimates from different data sources are combined,
the bank must choose among the alternatives conservatively.
C. Loss Given Default (LGD)
The LGD estimation process is similar to the PD estimation process.
The bank identifies a reference data set of defaulted credits and
relevant descriptive characteristics. Once the bank obtains these data
sets (with the facility characteristics), it must select a technique to
estimate the economic loss per dollar of exposure at default, for a
defaulted exposure with a given array of characteristics. The bank's
portfolio must then be mapped, so that the model can be applied to
generate an estimate of LGD for each portfolio transaction or severity
grade.
Data
Unlike reference data sets used for PD estimation, data sets for
severity estimation contain only exposures to defaulting obligors. At
least two broad categories of data are necessary to produce LGD
estimates.
First, data must be available to calculate the actual economic loss
experienced for each defaulted facility. Such data may include the
market value of the facility at default, which can be
[[Page 45964]]
used to proxy a recovery rate. Alternatively, economic loss may be
calculated using the exposure at the time of default, loss of
principal, interest, and fees, the present value of subsequent
recoveries and related expenses (or the costs as calculated using an
approved allocation method), and the appropriate discount rate.
Second, factors must be available to group the defaulted facilities
in meaningful ways. Characteristics that are likely to be important in
predicting loss rates include whether or not the facility is secured
and the type and coverage of collateral if the facility is secured,
seniority of the claim, general economic conditions, and obligor's
industry. Although these factors have been found to be significant in
existing academic and industry studies, a bank's quantification of LGD
certainly need not be limited to these variables. For example, a bank
might expand its loss severity research by examining many other
potential drivers of severity (characteristics of an obligor that might
help the bank predict the severity of a loss), including obligor size,
line of business, geographic location, facility type, obligor ratings
(internal or external), historical internal severity grade, or tenor of
the relationship.
A bank must ensure that the reference data remains applicable to
its current portfolio of facilities. It must implement established
processes to ensure that reference data sets are updated when new data
become available. All data sources, variables, and the overall
processes concerning data collection and maintenance must be fully
documented, and that documentation should be readily available for
review.
S. The sample period for the reference data must be at least seven
years, and must include periods of economic stress during which
defaults were relatively high.
Seven years is the minimum sample period for the LGD reference
data. A longer sample period is desirable, because more default
observations will be available for analysis and may serve to refine
severity estimates. In any case, a bank must select a sample period
that includes episodes of economic stress, which are defined as periods
with a relatively high number of defaults. Inclusion of stress periods
increases the size and potentially the breadth of the reference data
set. According to some empirical studies, the average loss rate is
higher during periods of stress.
Example. A bank intends to rely primarily on internal data when
quantifying all parameter estimates, including LGD. Its internal
data cover the period 1994 through 2000. The bank will continue to
extend its data set as time progresses. Its current policy mandates
that credits be resolved within two years of default, and the data
set contains the most recent data available. Although the current
data set satisfies the seven-year requirement, the bank is aware
that it does not include stress periods. In comparing its loss
estimates with rates published in external studies for similarly
stratified data, the bank observes that its estimates are
systematically lower. To be consistent with the standard, the bank
must take steps to include stress periods in its estimates.
S. The definition of default within the reference data must be
|
