The Proposed New Basel Capital Accord:

Preliminaries of The Internal Ratings-Based Approach

OCTOBER • 2001
Moody’s RiskCalcTM and the
Quantitative Credit Tools
Several factors are fueling rapid
Proposed IRB Approach
growth in appetite among holders The documents referred to herein are The New Basel Capital Accord and The Internal
or originators of credit sensitive Ratings-Based Approach, both by the Basel Committee on Banking Supervision and both
financial assets for independent, dated January 2001. Section number references are to the first document.
standardized quantitative tools to
assess corporate credut risk:
• A global surge in credit The New Basel Capital Accord about a borrower’s management,
operations and competitive position to
awareness, driven in part by In January 2001, the Basel Committee generate credit scores and diagnostics.
Basel and underperforming on Banking Supervision issued a second
credit assets; round of consultative documents This paper deals with the role Moody’s
• Anticipation of capital allocation proposing changes to the capital RiskCalc can play in addressing issues
reforms coming out of Basel; requirements for banks. The Committee raised by the Basel Committee.
is comprised of senior representatives
• Increased liquidity of credit
assets, made possible by
from banking supervisory authorities Probability of Default (PD)
securitization and advances and central banks in Belgium, Canada,
France, Germany, Italy, Japan, The core of the IRB approach is to
in portfolio analytics and made encourage banks to meaningfully
desirable by increased pressure Luxembourg, the Netherlands, Sweden,
Switzerland, the United Kingdom and differentiate borrowers based on risk.
from regulators and shareholders.
One indication of the impact of the United States. The proposals are Under the IRB approach, banks would
these and like factors is the directed towards replacing the 1988 categorize their book exposures into
explosive growth in credit Accord and have as a key element various classes: corporates, banks,
default swaps, that allow credit greater reliance on banks’ internal sovereigns, retail, project finance and
to be traded as a commodity. rating based systems (“IRB”) in the equity. For corporate, bank and
calculation of regulatory capital sovereign exposures, the approach
The Basel Committee has identified
three broad categories of rating charges. These proposals could be involves estimation of the PD, loss
systems in use by banks, two of adopted by banking supervisors in 2005. given default and expected exposure at
which use quantitative tools: default. In addition, there are standards
Irrespective of their adoption, we for treating guarantees and credit
• Model-based: In most banks believe the credit policy issues and derivatives. The PD component of the
where statistical or options-
recommendations raised by the analysis is defined in the proposal
theoretic models are used,
Moody’s RiskCalc models for Committee merit careful consideration as the greater of the one-year PD
public and private companies by banks, in the context of both risk associated with the internal
are part of a credit process management and risk-adjusted borrower grade and 3 basis points.
also involving analyst input. profitability measurement of customer
profitability. In response to the pace Moody’s RiskCalc produces one- and
• Constrained Expert Judgment:
of change in loan origination and five-year PDs for public and private
These systems allow an analyst corporations and banks. The private
to adjust a model-generated secondary-trading markets as well
as rapid innovation in financial firm methodology uses selected financial
score within certain parameters. statement data in a non-linear statistical
Moody’s Risk Advisor is an technologies, most financial institutions
are considering quantitative tools as one model to produce a credit score that is
example of an expert system
tool that integrates financial means to assist in this process. Moody’s then mapped into a PD at one- and
statement data with subjective offers several products in this category: five-year horizons. A final mapping to
inputs about a borrower’s Moody’s RiskCalc, which produces the historical Moody’s bond default rate
management, operations and probabilities of default; Moody’s is based on the five-year PD. For public
competitive position. LossCalc™, which produces estimates companies, essentially the same
of loss in the event of default; Moody’s methodology is used, but additional
• Full Expert Judgment: Moody’s
traditional public ratings are an Risk Advisor™, an expert system inputs include ratings (when they exist)
example of full expert judgment product which integrates financial and a variant of Merton’s contingent
models. statement data with judgmental inputs claims model expressed as a probability
of crossing a default barrier. As of
2
mid-Summer 2001, Moody’s RiskCalc 240
models have been released for public Overly broad risk categories are
firms in North America and Europe, unacceptable. Specifically, there must
for banks in North America and for be at least six to nine borrower grades
private companies in the U.S., Canada, for performing loans and at least two
Australia, Japan, Spain and Germany. grades for non-performing loans. The
Plans exist to extend geographical criterion for a non-performing grade is
coverage and capture industry-specific related to provisioning or loss.
risk where necessary.
241-243
Each grade must be clearly described
Granularity Minimum Requirements for to permit consistent and reliable
Corporate Exposures application throughout the bank and
More granularity—finer distinctions across lines of business and should
of risk, especially among riskier The Committee has proposed nine
intuitively match the average PD
obligors—can improve the quality of broad requirements for banks to be
associated with the category.
a bank’s portfolio risk analysis as eligible for the IRB approach, each
well as the meaningfulness with relevant to a different aspect of the 242
which it considers the profitability of rating and risk measurement process: There should be meaningful
its individual borrowers. One study* distribution of exposure across grades,
1. Meaningful differentiation of
of the fifty largest U.S. banks found with no more than 30 percent of the
credit risk
that over one third lumped half or gross exposures in any single borrower
more of their rated loans into a 2. Completeness and integrity of
grade.
single risk grade. Such systems rating assignment
appear to contribute little to the 3. Oversight of the rating system Moody’s RiskCalc produces a
understanding and monitoring of risk and processes probability of default, which can vary
posture. The authors note that such 4. Criteria and orientation of rating continuously from 0.01 percent to over
failure to distinguish degrees of risk system 15 percent for risk over a one-year
had been cited in Federal Reserve 5. Minimum requirements for period or over 25 percent for a five-year
examination guidance as a potentially estimation of PD period. For ease in administering a
significant shortcoming in a large 6. Data collection and IT systems credit policy, these PDs can be mapped
institution’s credit risk management 7. Use of internal ratings back to an institution’s own rating
process. 8. Internal validation system at fine levels of differentiation.
*William F. Treacy and Mark S. Carey,
9. Disclosure Moody’s also maps them into its own
“Credit Risk Rating at Large U.S. 21-grade rating scale, based on the
Banks,” Federal Reserve Bulletin: historical one- and five-year default
November 1998. Requirement 1 rates associated with each rating level.
Meaningful Differentiation of
Credit Risk
Requirement 2
237-238-239 Completeness and Integrity of
The rating system must have two Rating Assignments
dimensions. The first is an estimate
of the risk of borrower default. The 244
second is an estimate of transaction- or Each legal entity to which a loan is
facility-specific risk. made must be first rated.
Moody’s RiskCalc assigns PDs at the 245
borrower level. Transaction- or facility- Each rating must be subject to review
level risk should be separately assessed. or approval by an independent entity,
For this, institutions have historically such as a separate credit risk
used rules of thumb, look-up tables management unit, that does not stand
such as those provided in Moody’s to benefit from the rating assigned. The
bond default research. Moody’s now review process must be documented.
offers Moody’s LossCalc, which is an
Moody’s RiskCalc PDs constitute an
enhancement of those look-up tables
independent auditable assessment of
and incorporates a statistical model
risk and can be used to benchmark or
designed to estimate loss given default
corroborate the analyst rating.
more accurately.
246-247
Borrowers should be re-rated at least
annually and within 90 days of receipt
of new financial information. The bank
should have an effective process to
 3
obtain and update relevant information • Provide a detailed outline of the
on the borrower’s financial condition. theory, assumptions and/or
Higher risk or problem credits should mathematical and empirical basis of
be reviewed more frequently and with the assignment of PD estimates to
more urgency, particularly when new grades or individual obligors and the
information comes to light. data source(s) used to estimate the
model;
Moody’s RiskCalc PDs can be updated
as new financial statement, equity or • Establish a rigorous statistical
macro-economic data becomes available. process (including out-of-time and
They can be used as cost-effective and out-of-sample performance tests) for
timely “early warning systems” for validating the selection of explanatory Use of Quantitative Models
changes in the credit quality of corporate variables; and
obligors and to monitor developments It is Moody’s experience over the past
• Indicate circumstances under which ten years that quantitative models do
much more continuously than would be
the model does not work effectively not replace credit analysts. They
practical for analysts. This seems
such that the bank is fully aware of can, however, make credit analysts
particularly pertinent for middle
the limitations of the model. more effective. Quantitative models
market loan portfolios, where the large
number of relatively small exposures Moody’s publishes extensive and analysts have offsetting strengths
can render uniformly detailed analyst that can make them particularly
documentation on Moody’s RiskCalc,
powerful when used together.
attention prohibitively expensive or can including the theories and assumptions
result in costly delays in attending to underlying the models, the predictive • Models produce assesments
those credits at particularly high risk. variables considered and chosen, the using algorithms and are data
Moody’s RiskCalc can be used as a statistical power of the resulting model, dependent. An analyst uses
screening tool to focus analyst attention its calibration to probability of default, judgment, which includes at least
on outliers or credits with declining the validation metrics and data used some part intuition, and is able to
fundamentals. and discussions of the limitations of process information not easily
each model. Moody’s validation results codified for use in a formula.
are based on an extensive database of • The output of analysts throughout
Requirement 3 tens of thousands of default events a geographically-dispersed bank,
Oversight Over the Rating System and hundreds of thousands of public or one with heavy turnover or
and private firms. The documents that is growing rapidly through
and Processes describing our methodologies and mergers, can be idiosyncratic and
248 models are publicly and freely available inconsistent. Models are
All material aspects of the rating and at www.MoodysRMS.com. Client consistent across time and users.
PD estimation process must be feedback will guide subsequent model They constitute a safety net to
understood and approved by the board releases. help ensure that minimum
of directors, management committee, standards are maintained.
253-254
and senior management. • Models are closed systems
The bank must internally audit and
249 document its rating system at least until a revision is released.
Senior management should receive annually, including its quantification of Analysts are flexible and able to
the internal ratings. In some cases, an consider new types of information
monthly reports including risk profile
as they become available.
by grade, migration across grades, external audit may also be required.
quantification of loss estimates per 255
• Analysts can be expensive to
grade and comparison of realized support. Models are data-driven
The bank should have an independent and usually much less expensive.
default rates against expectations. credit review function responsible for Some banks use models to screen
250-252 the design, implementation and the best and worst credits, where
Management must ensure that the performance of the internal rating the outcome is relatively certain,
rating process, criteria and outcome system. The unit must: and leave less clear-cut credits
are comprehensively documented in • Assign and/or review and monitor for analysts.
paper or electronic form in sufficient internal ratings;
detail to facilitate audit and must, in
general, ensure the overall proper • Produce and analyze reports on
operation of the risk rating system. the outputs of the bank’s internal
rating system, historical data on
251
the performance of credits by
Where statistical models are used in internal grade, migration analyses,
the rating process, management must
ensure that the bank has in place a
comprehensive methodology document
for the model. The methodology
document must:
4
comparison of assigned grades to power, are both plausible and intuitive,
external ratings or default and, in short, are used to distinguish
prediction models and aggregate risk rather than to minimize regulatory
monitoring of credits in each grade capital requirements.
by key rating criteria;
265
• Ensure that procedures are in place Banks should take all relevant
to regularly check whether ratings information into account in assigning
are consistently assigned according ratings. This information should be
to established policies and criteria; current and the methodologies and
and data used clearly specified and
documented. At a minimum, a bank
• Review and document any changes
should consider:
to the rating process, including the
reasons for the changes. • Historical and projected capacity to
generate cash to repay debt and
256
support other cash requirements,
The credit risk unit must assume
such as needed capital expenditures;
responsibility for and control of any
model used in the rating process. • Capital structure and the likelihood
that unforeseen circumstances
could exhaust the capital cushion
Requirement 4 and result in insolvency;
Criteria and Orientation of • Quality of earnings, that is, the
Rating System degree to which revenue and cash
flow emanate from core business
258-259-260 operations as opposed to unique and
A bank must have a specific non-recurring sources;
documented system for rating
corporate exposures, with criteria • Quality and timeliness of information
specific enough to enable third-party about the borrower, including the
audit and reflecting historical availability of audited financial
experience with comparable borrowers. statements in conformity with
Rating criteria should be regularly applicable accounting standards;
reviewed and any changes • Degree of operating leverage and
documented.
susceptibility to interruption in
261 revenues;
Risk assessments should be conservative • Financial flexibility from access to
and, where indicated, extend beyond
debt and equity markets;
accounting information. The depth of
credit analysis should increase as a • Depth, skill, and prudence of
borrower’s financial condition management and its ability to
deteriorates and default becomes effectively respond to changing
more likely. conditions and deploy resources;
262 • Position within the industry and
The risk grade should incorporate the future prospects; and
bank’s conservative expectations as to
• Risk characteristics of the country
the borrower’s ability to meet future
in which it operates, including
contractual obligations and withstand
cross-border transfer risk.
normal business stresses.
266-267
263
All variables used in a model must
The risk grade should be mapped to a
have statistical power. Any critical
one-year PD. The PD should be
information not captured in the model
reviewed at least annually and more
must instead be considered by the
frequently for higher-risk borrowers.
bank’s analysts. Model-based rating
264 assignments must be subject to review
The bank must demonstrate that its and approval by personnel in the credit
criteria cover all factors that are risk control unit. The integrity of the
relevant to the analysis of borrower rating model must be assured by the
risk, have predictive and discriminatory credit risk control unit which must
have responsibility and control of
inputs and outputs.
 5
The best models are those that combine must be based on a well-developed and
theory and science with experience and well-documented thought process and
intuition. Even with perfect data, analysis.
statistical analysis uninformed by
271-272-273
common sense can lead to spurious
To ensure consistent estimation of risk
conclusions. On the other hand,
across banks and data sources, a
sophisticated financial theory without a
“default” shall always be defined as
grounding in substantial bodies of
involving one or more of the following:
empirical data is inefficient. Moody’s
modeling approach starts with a single • It is determined that the obligor is
factor analysis of the relationship unlikely to pay its debt obligations
of individual variables to default. (principal, interest, or fees) in full. Assumptions about
Although the variables selected may Population PD
vary by country or industry, they are • There is a charge-off, specific
provision, or distressed restructuring The estimate of long-run aggregate
all intuitively related to credit risk and probabilities of default for an
can be classified broadly as relating to: involving the forgiveness or
postponement of principal, interest, economy is the anchor for a model’s
profitability (e.g., return on assets or validity. Changing it affects all
return on net worth), liquidity (e.g., or fees.
predicted probabilities of default.
working capital to total assets), capital • The obligor is past due more than Generally, a variety of data sources is
structure (e.g., equity to assets) or 90 days on any credit obligation. used to triangulate a central tendency
market position (e.g., asset size). In estimate. In the long run, as better
addition, for companies with traded • The obligor has filed for bankruptcy
data accumulates, the estimates
equity, we use market information and or similar protection from creditors.
should improve.
the output of a structural model based The standard definition of default One approach to determining the
on Merton’s options-theoretic view of against which Moody’s RiskCalc population default rate is to use
firms. These inputs are then optimized models have been developed is for banks’ loan loss provisions, which,
in a multi-factor model. private debt: over time, will tend to equal actual
268 • Placement on internal non-accrual losses and hence, after backing out
The bank must have in place clear list loss given default (LGD), reflect the
guidelines and processes for monitoring underlying default rate.
cases where human judgment has • The credit is written down (e.g., in
For instance, the Bank of Spain recently
overridden an output of the model the U.S. this means placed in the
reported net loan loss provisions of
and have procedures, documentation regulatory classifications of
Spanish financial institutions as a
and sign-off for exceptions to input substandard, doubtful or loss) percentage of total lending to be
parameters. • 90 days past due 0.8-0.9 percent between 1986-1998.
Assuming an LGD of 50 percent, this
269 • Declared bankruptcy gives an estimated average default
For rating assignments based on expert rate of 1.6-1.8 percent. This estimate
judgment, banks must clearly articulate The standard definition is modified in
each market as needed to reflect unique might be further adjusted to reflect
the situations in which bank officers peculiarities of the sample time period
may override the outputs of the rating practices or characteristics. For
and to factor in other data sources.
process, including how and to what example, in some countries a 90-day
extent such overrides can be used and late payment is not abnormal and may
by whom. Instances of overrides must be a weak indicator of expected loss.
be clearly documented. Banks must 274-275
track separately the performance of At least annually, the bank should
overridden grades. review its estimate of the average PD
per grade, based on internal default
experience, mapping to external
Requirement 5 ratings, and statistical default models.
Minimum Requirements for Judgmental considerations are critical
Estimation of PD to this process to ensure a forward-
looking PD estimate, but must be
270 applied with a consistent and
A bank must estimate a one-year PD conservative bias.
for each of its internal rating grades,
representing a conservative long-run • The data sets used in estimating the
view and grounded in historical average PD should closely match or
experience and empirical evidence. be clearly comparable to the bank’s
At the same time, these estimates must loan portfolio.
be forward looking. Any adjustments
6
• The lending or underwriting which is the result of gathering
standards used to generate the defaults in various countries/regions
exposures in the data source should corresponding to different time spans
be comparable to those used by and sample sizes, contains—as of
the bank in building its current mid-Summer 2001—3,679 defaults for
portfolio of exposures. North America, 3,278 for Australia,
1,918 for Mexico, 3,347 for Japan,
• The economic or market conditions
1,485 for Germany, 2,265 for Spain,
for the period from which the
approximately 4,000 for the United
historical data was taken should be
Kingdom and 7,000 for France. In all
relevant to current and foreseeable
cases, the sampling of defaults is
conditions.
largest in more recent years. This data
Validation Testing collection effort is on-going, with 32
• There must be an adequate number
Because default events are rare and of loans in the sample, the data actively participating institutions.
risk scores for consecutive years period used must be relevant and
highly correlated, it is difficult to
The collection and pooling of data is
the underlying statistical analysis
design appropriate tests for susceptible to numerous potential
must be robust.
benchmarking credit risk models. biases, which we strive to identify and
It is often impractical to create a 276-277-283-164 correct. At the same time, there are
model using one data set and then A bank may use its own experience advantages to the pooling of data
test it on a separate “hold-out” data to estimate average PDs per grade, across institutions, such as Moody’s
set composed of completely adopting a conservative bias where has done with respect to those sectors
independent cross-sectional data. data is sparse or underwriting of a bank’s portfolio that are relatively
Such out-of-sample and out-of-time standards have changed, or it may use under-represented. Moody’s will publish
tests would be the best way to pooled data from other institutions if it descriptive statistics on its data to permit
compare models’ performance, if only can demonstrate that the internal users of Moody’s RiskCalc to determine
default data were less scarce. rating systems and criteria of the other the fit to their own book of loans.
Statistical tests for samples with low banks are comparable with its own.
default rates are susceptible to 278-279
The estimation of the average PD for a
random sample variations and can A bank may map its internal grades to
grade must be based on at least five
create the illusion of superiority of an external credit scale, such as those
years of historical data or, at the start
one model over another where no published by the rating agencies. Care
of the transition period, at least two
real difference exists. Moody’s has must be demonstrated to ensure that
years.
attempted to deal with the problem the mapping is valid and robust.
of sparse data while at the same 280-281-282
Moody’s RiskCalc may act as an
time avoiding “overfitting” through A bank may use an average PD for
a testing technique known as a
external credit scale into which internal
borrowers assigned by a statistical
“walk forward” analysis, where bank grades might be mapped and
model to a given grade. It must have
successive versions of the model are mechanisms exist within the product
in place a robust process for ensuring
built and tested over incrementally for banks to define such mappings. In
the accuracy, completeness and
longer periods of time. Moody’s also addition, Moody’s RiskCalc PDs have
appropriateness of the data inputs.
uses other forms of out-of-sample been mapped in the Moody’s traditional
Moreover, it must demonstrate that
testing to ensure model performance rating scale categories Aaa to C on the
the model was based on a population
stability. basis of the historical one- and five-year
not dissimilar to its own actual
bond default rate for each level.
Despite their limitations, it is borrowers.
necessary that such tests be carried
out, properly designed and accurately
Moody’s uses large databases of
reported, together with disclosure of financial statement, equity price and Requirement 6
their limitations. Transparency is a payment data to estimate and test Data Collection and IT Systems
prerequisite for credibility in this Moody’s RiskCalc. For public
area. Argument by anecdote is companies, Moody’s had access to 284-285
misleading and injurious. information on over 3,500 public Banks using the IRB approach must
companies that have defaulted since collect and store data on rating
1980. For private companies, Moody’s decisions, the rating histories of
has worked with banks and other borrowers and the probabilities of
financial institutions in each market default associated with rating grades
to accumulate databases of local and ratings migration in order to track
financial statement information and the predictive power of the rating
matching loan accounting data for system. For each borrower, a complete
mostly middle market companies for rating history must be retained,
the period 1989 to 2000. This database, including the dates ratings were
assigned, the methodology and key
data used and the person/model who
assigned the rating.
 7
286-287 297
For each broad grade, the bank must In conjunction with assessing its
keep a history of the estimated PD and capital adequacy, a bank must stress
the realized default rate associated with test its portfolio to determine the
that grade, as well as key pertinent potential impact of future events such as:
borrower characteristics and facility • economic or industry downturns
information. • market-risk events
• liquidity conditions
288
Banks must demonstrate the integrity 298
and robustness of their information Stress testing should include specific
technology systems to support exposure scenarios that quantitatively assess the
aggregation, data collection and impact of broad rating migration of
management reporting. exposures to lower rating grades and
should also examine the impact of
Moody’s calculates and stores PDs
higher default rates and lower recovery
and the underlying data for all public
rates.
firms it follows. For private firms in
certain markets, including North 299-300
America and Australia, we have An independent unit must conduct
assembled peer data, including PDs, stress tests at least every six months.
which can be used for benchmarking a These tests must be documented and
bank’s individual portfolio. In addition, reported to senior management and
Moody’s Financial Analyst software action taken as required.
can assist a bank in storing PDs and
301
related key data for private and public
A bank must have a credible track
firms.
record of at least three years in the use
of internal ratings.
Requirement 7
Use of Internal Ratings Requirement 8
289-290-291-292 Validation
Internal ratings must be an integral
302
part of daily credit risk measurement
Banks must have a robust system in
and management. Internal ratings
place to validate the accuracy and
must play an essential role in the
consistency of rating systems,
credit approval process. The PDs
processes and the estimation of PDs.
associated with internal ratings must
be used in the pricing of credit risk. 303-304
The setting of internal lending limits The bank must have in place adequate
and authority must be linked to processes to ensure the accuracy,
internal ratings. completeness and appropriateness of
data inputs to ratings. Detailed
293
documentation of exceptions to data
The distribution of exposures across
input parameters must be maintained
internal rating grades and the
and reviewed.
associated PD must be reported to
senior management. The data used in calculating PDs for
public firms are taken from sources
294-295-296
believed to be of high quality and
Internal ratings must be explicitly
then subjected to statistical anomaly
linked with the bank’s internal
detection and manual error
assessment of capital adequacy and
identification, research and correction.
considered in the process of reserving.
One recent sampling turned up
The bank must use the PD associated
errors—some minor—on a half percent
with an internal grade as input to its
of the equity data records provided by a
profitability analysis for management
leading vendor and on the order of
processes such as strategic resource
about 5 percent of the financial
allocation decisions or incentive
statement data sets. Changes made by
compensation plans.
Moody’s to the data are documented
and stored.
8
305 307
Model validation must include: Banks should quantitatively validate
their rating systems based on long data
• Ongoing periodic monitoring of
histories and covering a range of
model performance, including
economic environments and ideally a
evaluation and rigorous statistical
complete business cycle.
testing of the dynamic stability of
the model and its key coefficients; 308
Banks must demonstrate that their
• Identifying and documenting
quantitative testing methods and data
individual fixed relationships in the
Contact Us are consistent through time: changes in
model that are no longer appropriate;
methods and data (both data sources
Europe • Periodic testing of model outputs and the periods covered) must be
Rathfarnham Gate against outcomes on an annual clearly and thoroughly documented.
Rathfarnham basis, at a minimum; and
Dublin 14 Ireland
Phone: +353 1 492 4259 • A rigorous change control process,
which stipulates the procedures that Requirement 9
Fax: +353 1 492 4307 Disclosure
must be followed prior to making
North America changes in the model in response to 309
1350 Treat Boulevard, Suite 400 validation outcomes. Banks must meet minimum disclosure
Walnut Creek, CA 94596 requirements, including:
Phone: In the U.S. & Canada, dial Moody’s conducts and publishes
toll free 1 800 321 3278, ongoing validation tests on its models • General information on
otherwise call + 1 925 945 1005 against the largest data sets available. methodology and key inputs;
Fax: +1 925 945 1162 One way of presenting the results of • Quantitative descriptors of its
those tests is by constructing a power portfolio; and
Asia Pacific curve or CAP plot of the firms in the • Ex post performance of the
Room 2510-2514 sample ranked by PD plotted against system.
One International Finance Centre the cumulative proportion of defaults
1 Harbour View Street captured to enable a visual qualitative Moody’s publishes detailed research
Central assessment of model performance. concerning theory behind the
Hong Kong It is also helpful to aggregate the power development of its PD models, the
Phone: +852 2 916 1116 curve information into a single number methodologies used to develop them
Fax: +852 2 509 0165 to allow more precise comparison and their performance on large, pooled
across models. One such measure is the data sets of defaults. Being public, this
Middle East, Africa and India research is subject to scrutiny from
area under the power curve relative to
Rathfarnham Gate practitioners and academics alike,
the area for an “ideal” model that
Rathfarnham affording Moody’s a continuing source
correctly ranked all firms in the
Dublin 14 Ireland of feedback for improvement.
sample. Such an “accuracy ratio”
Phone: +353 1 492 4259
calculated for Moody’s RiskCalc
Fax: +353 1 492 4307
Germany came to 70.9 percent; that
Latin America number can be roughly understood in
204 Domenech absolute terms but is best interpreted in
San Juan, Puerto Rico 00918 comparison with other models run off
Phone: +1 787 756 6523 the same test set.
Fax: +1 787 274 0129
306
Banks must regularly compare realized
default rates with estimated PDs for
each grade to demonstrate their
correspondence. Such comparisons
should use historical data periods that
are as long as possible. The methods
and data used in such comparisons
must be clearly documented and
understood by the bank. These
comparisons must be conducted at
least annually.