Quantifying Credit Risk I: Default Prediction

Author(s): Stephen Kealhofer

Source: Financial Analysts Journal, Vol. 59, No. 1 (Jan. - Feb., 2003), pp. 30-44
Published by: Taylor & Francis, Ltd.
Stable URL: https://www.jstor.org/stable/4480449
Accessed: 06-05-2020 21:36 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Taylor & Francis, Ltd. is collaborating with JSTOR to digitize, preserve and extend access to
Financial Analysts Journal

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
 Quantifying Credit Risk 1: Default Prediction
Stephen Kealhofer

Until the 1990s, corporate credit analysis was viewed as an art rather than
a science because analysts lacked a way to adequately quantify absolute
levels of default risk. In the past decade, however, a revolution in credit-risk
measurement has taken place. The evidence from this research presents a
compelling case that the conceptual approach pioneered by Fischer Black,
Robert Merton, and Myron Scholes provides a powerful practical basisfor
measuring credit risk.
~ ~ ~. : . . . . . . . . . .
O ne of the major frontiers in modern finance equity prices, it produces empirical estimates of
is the quantification of credit risk. More default probability that outperform well-accepted
than 25 years ago, Black and Scholes (1973) public benchmarks, such as agency debt ratings, in
proposed that one could view the equity of predicting default. Finally, in contrast to earlier
a company as a call option. This insight provided a findings, this study finds that the Black-Scholes-
coherent framework for the objective measurement Merton approach provides a better empirical fit to
of credit risk. As subsequently elaborated by Mer- the value of corporate bonds than more conven-
ton (1973,1974), Black and Cox (1976), and Ingersoll tional approaches that use agency bond ratings.
(1977a), this approach has come to be called "the
Merton model."
The KMV and Merton Models
Initially, empirical research to implement the
Merton model (notably, Jones, Mason, and Rosen- The approach Black and Scholes developed can be

feld 1984 and Ogden 1987) produced discouraging illustrated in a simplified case. Suppose a company

results, and even today, researchers continue to has a single asset consisting of 1 million shares of
reproduce the original negative results (e.g., Jarrow Microsoft stock. Furthermore, it has a single fixed
and Van Deventer 1999). In 1984, however, Vasicek liability-a one-year discount note with a par
took a novel approach to implementation of the amount of $100 million-and is otherwise funded
Merton model that has proven to have considerable by equity. In one year's time, the market value of
success in measuring credit risk. This version of the the company's business will either be sufficient to
Merton model has been extended by KMV Corpo- pay off the note or it will not, in which case the
ration to become a defacto standard for default-risk company will default. One can observe that the
measurement in the world of credit risk.1 equity of the company is logically equivalent to 1
This first of two companion articles on the million call options on Microsoft stock, each with
empirical research on the KMV model in the past an exercise price of $100 and a maturity of one year.
decade explains the differences in implementation The implication of this illustration is that the
between the KMV and Merton models and reviews equity of a company is like a call option on the
studies comparing the performance of the KMV company's underlying assets. The value of the
model against agency debt ratings in predicting equity thus depends on, among other things, the
default.2 Part II will present results on the use of market
the value of the company's assets, their volatil-
model to value corporate debt relative to alterna- ity, and the payment terms of the liabilities. Implicit
tive approaches. in the value of the option is a measure of the prob-
The thesis of the articles is that the Black- ability of the option being exercised; for equity, it
Scholes-Merton approach, appropriately executed, is the probability of not defaulting on the com-
represents the long-sought quantification of credit pany's liability.3
risk. It is an objective cause-and-effect model that Figure 1 provides an illustration of these con-
provides analytical insight into corporate behavior. cepts. The horizontal axis represents time, begin-
Driven by information in a company's public ning with the current period ("Today") and looking
into the future. The vertical axis depicts the market
value of the company's assets. As of the current
Stephen Kealhofer is a managing director at KMV Cor- period, the assets have a single, determinable
poration, San Francisco, California. value, as shown on the vertical axis, but one year

30 02003, AIMR?

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
 Quantifying Credit Risk I

Figure 1. Illustration: Frequency Distribution of Asset Value at Horizon and

Probability of Default

Expected Asset Value

Asset _ --
Value-

Default Point

EDF

Today 1 Year

Note: The distance from the expected asset value to default is three standard deviations.

from now, a range of asset values is possible, and increases, or if the volatility of the assets' market
their frequency distribution (shown in Figure 1 on value increases. These three variables are the main
its side) gives the likelihood of various asset values determinants of the company's default probability.
one year in the future. The most likely outcomes are What Black, Scholes, and Merton actually pro-
nearest to the starting value, with much larger or posed is a general framework for valuing contin-
smaller values less likely. The mean is shown by the gent claims. There is no single Merton model;
dashed line. The likelihood of extreme outcomes indeed, the KMV model is largely a variant of
depends on the volatility of the assets-the more Merton models. The KMV model differs from the
volatile, the greater the probability of extreme out- standard variants in some significant ways, how-
comes. The dotted horizontal line shows the par ever, and it also includes a scheme for empirical
amount of the liability due in one year. If the com- default-probability measurement that lies outside
pany's asset value in one year is less than the the standard approach. To appreciate these differ-
amount of the liability, the company will default. ences, one must begin with the canonical Merton
Note that this decision is an economic decision: The model.

equity owners could put additional money into the The classic exposition of the Merton model is

company, but that decision would be irrational Merton (1974). This model has the following char-
because the money would go to pay creditors; if the acteristics:

owners defaulted, they would not be required to * The company has a single debt liability, has
put in additional money and they could use this equity, and has no other obligations.

money for their own benefit rather than giving it to * The liability promises a continuous fixed cou-
the creditors. The probability of default is thus pon flow and has an infinite maturity.
given by the area under the frequency distribution * The company makes no other cash payouts
below the default point, which represents the like- (e.g., equity dividends).
lihood of the market value of the company's assets Under the assumption that the market value of
in one year being less than what the company owes. the company's assets evolves as a lognormal pro-
It is immediately obvious that the probability of cess, Merton showed that this model can be solved
default will increase if the company's market value for a closed-form solution for the value of the
of assets today decreases, if the amount of liabilities company's debt. The aim of the model, and much

January/February 2003 31

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
Financial Analysts Journal

subsequent work based on this model, is to obtain tA = expected market return to the assets
a valuation equation for the company's debt. per unit of time6
The KMV model, building on previous work cA = volatility of the market value of the
by Black and Cox and by Ingersoll (1977a) begins company's assets per unit of time7
with a somewhat more general characterization of From a certain standpoint, these model differ-
the company's capital structure. The most impor- ences are not significant. In terms of practical imple-
tant distinction between the models, however, is mentation, however, they are critical. Researchers
the KMV model's primary focus on the probability using the Merton model to investigate the pricing
of default of the company as a whole, rather than of corporate debt obtained poor results. Research
valuation of the debt. The KMV model has the based on the distance to default as the measure of
following characteristics:4 default risk has yielded excellent results.
* The company may have, in addition to com- Why focus on default-risk measurement rather
mon equity and possibly preferred stock, any than debt valuation? The answer is that debt valu-
number of debt and nondebt fixed liabilities.
ation should implicitly contain default-risk mea-
* The company may have warrants, convertible surement. Thus, if the default-risk measurement is
debt, and/or convertible preferred stock. verified as correct, the subsequent debt valuation
* Obligations may be short term, in which case should also be correct. Because earlier efforts at
they are treated as demandable by creditors, or valuation had failed, a focus on default-risk mea-
long term, in which case they are treated as surement made sense because it could be indepen-
perpetuities. dently tested. Any insights garnered from that
* Any and all classes of liability, including equity, testing could then be used to diagnose and, poten-
may make fixed cash payouts. tially, resolve the valuation problems.
* If the market value of the company's assets falls The KMV model focus on default-risk mea-
below a certain value (the default point), the surement leads to another important aspect of the
company will default on its obligations; this model. As noted previously, the distance-to-
default point depends on the nature and extent default measure is an ordinal measure. For valua-
of the company's fixed obligations. tion purposes, however, one needs an absolute
* Default is a company-wide event, not an measure, an explicit probability of default. To date,
obligation-specific event. other researchers have obtained default probabili-
Whereas the objective in the Merton model is ties by using the assumed lognormal asset-value
the valuation of the company's debt based on the distribution of the Merton approach.
company's asset value and volatility, the focus in The assumption of log-normality cannot be
the KMV model is on the relationship between the evaluated without determining the actual default
company's equity characteristics and its asset char- experience of a large, well-defined population of
acteristics. Given the asset characteristics (i.e., companies. KMV has been tracking the default
value and volatility) and given the company's experience of all publicly traded companies in the
default point, one can use the KMV model to imme- United States since 1973. These data made possible
diately calculate a simple, robust measure of the a comparison of the default probabilities calculated
company's default risk-the number of standard from the lognormal distribution with actual real-
deviation moves required to bring the company to ized default rates. This comparison established that
the default point within a specified time horizon.5 small but important differences exist between the
Distance to default, DD(h), or the number of theoretical and actual default rates. For instance,
standard deviations to the default point by horizon under the normality assumption, any company
h, is an ordinal measure of the company's default more than about four standard deviations from its
risk. As such, it provides a simple and robust mea- default point would have essentially zero probabil-
sure of default risk. Mathematically, it is calculated ity of default. In actuality, the default probability
as for such companies is meaningfully higher, about
0.5 percent. These small numeric deviations in tail
ln(A) - ln(DPT) + (11A - 1/2GA)h probabilities between the normal distribution and
DD(h) =cA 1
Ah1/ the empirical distribution translate into economi-
cally significant differences in terms of default risk.
where For instance, a company that is four standard devi-
A = current market value of the company's ations from the default point as measured by the
assets normal density would be better than AAA in qual-
DPT = the company's default point ity, whereas with an actual 0.5 percent default

32 ?2003, AIMR?

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
 Quantifying Credit Risk I

probability, it would not even be investment it subsequently defaults. A Type II error is identi-
grade.8 fying a company as bad (because it has a value at
With the use of the KMV default database, we or below the cutoff) but it subsequently does not
found that we could measure the empirical distri- default.
bution with sufficient accuracy that the empirical Three aspects of testing default-risk measures
probabilities could be substituted for the theoreti- deserve special attention. They concern arbitrary
cal probabilities. This measurement relies on the metrics, error trade-off, and sample dependence.
distance to default as a "sufficient statistic" for the * Arbitrary metrics. Almost by definition, differ-
default risk, so all the default data for companies ent default-risk measures are based on differ-
with similar DDs can be pooled. Put differently, the ent metrics. Thus, cutoff values are unrelated
differences between individual companies are from one metric to another. Comparing differ-
expected to be reflected in their asset values, their
ent indexes at arbitrary cutoff values is gener-
volatilities, and their capital structures, all of which
ally meaningless.
are accounted for in their DDs. The estimation need
* Error trade-off. The levels of Type I and Type II
not be performed on separate subsamples-for
errors are related to each other and to the level
instance, by industry or size. Where feasible, results
of the cutoff. Using a very high cutoff mini-
for subsamples have been compared with the over-
mizes Type I error, tl, but maximizes Type II
all pooled results, and no statistically significant
error, t2. Thus,
differences have been observed.9
tl(v) is decreasing in v;
The result of this process is the KMV EDFTM
(expected default frequency) credit measure. The t2(v) is increasing in v.
EDF is the probability of default within a given time * Sample dependence. The levels of error depend
period. It is a monotone function of the distance to on the particular sample tested. In general, one
default, so it preserves the ordinal properties of the cannot compare errors from one sample with
DD measure while providing a cardinal measure of errors from another sample. For instance, one
default probability. sample may consist of the larger companies or
the companies with longer histories. Differ-

KMV Model Default-Predictive ences of this type almost invariably translate

into different average default rates and differ-
Power ent levels of error.
Two types of tests on the predictive power of the These three observations have a number of
Merton approach (as implemented in the KMV implications. For instance, because of error trade-
model) have been carried out. The first type, off and sample dependence, stating the amount of
"power tests," characterizes the relative ability of a Type I error in isolation, as in "our internal risk
default-risk measure to correctly identify compa- ratings are 99 percent accurate," is meaningless. If
nies that subsequently default versus incorrectly
the underlying sample had only a 1 percent default
identifying companies as being likely defaulters
rate, then a measure that passed the entire sample
that do not default. The second type of test, "intra-
as good would be 99 percent accurate and also have
cohort analysis," provides a method for evaluating
zero Type II errors!
whether the differences between two measures are
A less obvious but equally fallacious result
the result of additional information or are simply
occurs when default-risk measures are tested on
noise.
different samples but the results are reported as

Power Tests. Default is a binary event; it comparable. For instance, suppose an analyst tests

either occurs or it does not. Most measures of two models on the same sample of companies; one
default risk-for instance, agency debt ratings- model calls for three years of financial data, and the
are indexes that indicate relative likelihood of other, for only the current financial data. Because
default. One can translate ordinal measures, such some companies will drop out of the database in a
as indexes, into binary choices for purposes of test- three-year period, fewer companies will have three
ing by interpreting them as an acceptance/rejection years of financial data. The data for these survivors
criterion. To do so, one determines a cutoff value, will be somewhat less volatile, on average, and the
v, and accepts as "good" all companies with an sample will have a lower average default rate.
index value above v.10 Rather than testing each model on the subset of
Given a cutoff value v, two types of errors can companies with data available for that particular
be made. A Type I error is identifying a company model, one must test the two models on exactly the
as good (because it has a value above the cutoff) but same set of companies-in this case, the subset of

January/February 2003 33

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
Financial Analysts Journal

companies with both historical and current data subsequent defaults divided by 20 percent of the
available. total number of subsequent nondefaults. For the
The problem of arbitrary metrics can be over- second measure, it is 70 percent of the total number
come by comparing the Type I error rates of the of subsequent defaults divided by 20 percent of the
models for which the cutoffs have been set to pro- total number of subsequent nondefaults. Because
duce the same levels of Type II error. (Equivalently, the total number of subsequent defaults and non-
one could compare Type II errors holding Type I defaults is the same for both measures by construc-
errors constant.) Because, in practice, one is inter- tion, the second measure, the more powerful one,
ested only in errors, not arbitrary cutoff values, the produces a higher default rate in the rejected group.
relevant issue is how much error of one type exists It follows that it must produce a lower default rate
for a given amount of the other type of error. in the "accepted" group.
Mathematically, the "power curve" for index i, In other words, a uniformly more powerful
pi(x), is graphed as measure results in a greater range of variation in
predicted default rates across the population. For

Pi(x) = 1 - t1iI[t21 (x)] (2) example, imagine a very poor measure that has no
predictive power. Regardless of the cutoff value
-that is, the percentage of defaults forecasted cor-
used, it will always predict the same default rate-
rectly given the percentage of nondefaults fore-
namely, the population average. In contrast, a per-
casted incorrectly (t2j1 denotes the inverse of
fect measure will predict a 100 percent default rate
function t2i).
for the companies that subsequently default and a
Power curves for different default predictors
0 percent default rate for the rest. The poor measure
can be compared with each other, which leads to
has no cross-sectional variation in predicted
the following definitions:
default rate; the perfect measure has maximal
* One measure is more powerful than another at
cross-sectional variation-0 to 100 percent.
x if it produces less Type I error than the other
Why is having more cross-sectional variation
when both produce Type II errors equal to x;
in predicted default rates good? For most applica-
Pi (X) > pj(x). (3a) tions, the issue is not setting a cutoff value for
acceptance but, rather, valuation and pricing. For
* One measure is uniformly more powerful than
any group of companies that has been determined
another if it is more powerful for all admissible
to have the same default rate by the less powerful
levels of Type II error;
measure, the more powerful measure can actually
pi(X) ? pj(x) for all values of x (3b) determine differences between their predicted
and default rates. In other words, one can determine
price or value more accurately by using the more
pi(X) > pj(x) for some x. (3c) powerful measure. It "sees" differences in future
The purpose of the power curve is to fully default rates that are invisible to the less powerful
characterize both types of errors that a default- measure.
prediction model makes-namely, failing to pre- N Empirical results of power tests. As noted,
dict a company that does default and predicting a only relative default-predictive power can be deter-
default that does not occur. One economic interpre- mined because absolute default-predictive power
tation of the power curve can be understood by is sample dependent. In evaluating the default-
thinking about the cutoff value as representing the predictive power of the KMV model, a natural
value a company must have to be approved for a point of comparison is agency debt ratings. Debt-
loan. A uniformly more powerful measure is the rating agencies, such as Standard & Poor's, Fitch
one that results in a lower default rate for approved IBCA, and Moody's, provide ordinal rankings of
loans for any given refusal rate. companies that embody their respective estimates
Another important economic interpretation is of default probability.1
that a uniformly more powerful predictor of Figure 2 shows power curves based on all
default results is one that results in more accurate identified defaults of nonfinancial companies with
predicted default rates. Consider the following public debt ratings from 1979 to 1990. (The power
example of two measures compared at any given curve values are indicated by the tops of the bars).
level of Type II error, say, 20 percent. Assume that To be included, the company had to have, as well
the first measure rejects 60 percent of the defaults as an expected default frequency, a rating 12
at that cutoff and the second measure rejects 70 months prior to the event of default. An event of
percent. For the first measure, the default rate of the default was defined as a failure to pay any sched-
rejected group is 60 percent of the total number of uled liability on time because of financial condition

34 ?2003, AIMR?

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
 Quantifying Credit Risk I

Figure 2. S&P versus KMV EDF, 1979-90 would be a power curve along the 45-degree line of
Defaults Excluded (%) Figure 2. Both measures clearly have default pre-

100 dictive power.

The point of greatest interest is that the KMV
90 2
model is uniformly more powerful than the agency
debt ratings. For instance, using a cutoff value for
70- each measure that results in equal Type II errors of

60-
20 percent, EDFs successfully identified 72 percent
of the defaults whereas the agency ratings identi-
50
fied about 61 percent. Such a difference is highly
40- significant, both economically and statistically. 14
30- Figure 3 shows power curves for S&P implied
ratings and "company" (i.e., unadjusted by the
20-
researchers) ratings, Moody's implied ratings, and

10
EDFs for publicly rated defaults from 1990 through
1999 (Lim 1999; Kealhofer and Kurbat 2001). This
10 20 30 40 50 60 70 80 90 100
sample was constructed in the same fashion as just
Population Excluded (%) described, with the adjustments to the S&P and
E EDF D S&P Rating Moody's ratings. The number of rated defaults is
121; the average number of nondefaults is 1,347 a
Source: KMV database.
year, or 1,579 unique companies.15 Consistent with
the results from the earlier period, the KMV model
substantially outperforms the rating agencies in
or an exchange of securities in which the new secu- terms of default-predictive power. As the graph
rities exacted a meaningfully reduced claim (a "dis- shows, the default-predictive power for implied
tressed exchange"). The defaults were generally Moody's and implied S&P ratings is virtually
but not invariably reflected in the agency debt rat- identical and is slightly below the power of S&P
ings changing to a D. company ratings.16
The S&P debt ratings were adjusted to reduce These findings reinforce the previous finding
the effect of issue specifics. Because the objective that EDFs are uniformly more powerful than
was to measure company default probability, agency debt ratings in predicting default. On the
rather than the treatment of the issue in bank- one hand, these results, given the widely held view
ruptcy, subordinated and secured issues ratings
had to be adjusted to create a uniform, non-issue-
specific measure. Both the written statements of
Figure 3. Power Curves for Four Measures,
S&P (see Standard & Poor's 1991) and a statistical
1990-99
analysis of their ratings suggested the following
Defaults Excluded (%)
adjustment. When the rating was BB or worse, the
100 ..... .
rating of subordinated debt of all types would be
raised two fine grades (for example, from B- to B+); 90 V

otherwise, the debt rating would be raised one fine 80 -

grade. The rating of secured debt would be reduced 70 -

one fine grade. The result of the adjustment was an 60
approximate measure of how the agency would
50
have rated the senior, unsecured debt of each com-
40 ./
pany.12 The final sample included 93 defaulted
companies. 30 :1

The nondefault sample was based on the rated, 20

nonfinancial, public companies with expected :10

10 *1
default frequencies from data from 30 June for each
year.13 For the entire time period, this sample con- 0 10 20 30 40 50 60 70 80 90 100

sisted of 1,066 unique companies, or approximately Population Excluded (%)

657 companies a year.
EDF S&P Implied
Note that if either the KMV EDF or the S&P
.... . .--S&P . - ----Moody's Implied
rating had no default-predictive power, the result

January/February 2003 35

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
Financial Analysts Journal

that the Merton approach does not work well cantly in the previous year to two years but whose
empirically, should perhaps be a surprise. Not only ratings have not changed. For Harnischfeger, its
does the Merton approach work well in predicting EDF was approximately consistent with its rating
default, it meaningfully outperforms commercial as late as mid-1997, two years before it defaulted.
benchmark measures that are well accepted by To understand what caused its EDF to change,
practitioners. On the other hand, because the Mer- Figure 5 presents the market value of Harnisch-
ton approach is harnessing the information in pub- feger's assets and equity and its total book liabilities
lic equity prices to predict default, perhaps the for the period shown in Figure 4. Figure 5 shows a
superiority of such an approach is not surprising.
significant decline in the value of the business
0 Example. A specific example should help
beginning in October 1997. Within a year, Harnisch-
illustrate these results. Figure 4 shows the EDF and
feger had lost almost half of the value in its business
the S&P debt rating for Harnischfeger Industries,
and three-fourths of the value of its equity. Over this
which defaulted in the summer of 1999. The S&P
same period, liabilities experienced little change,
rating is shown as the stepped line, with default
resulting in a substantial increase in Harnisch-
probability on the vertical axis.17 Although
feger's actual financial leverage and a substantial
Harnischfeger defaulted in June that year, it had
decrease in its distance to default.
been rated investment grade as late as March. The
EDF for Harnischfeger indicated that it was no
Intracohort Analysis. The previous example
longer investment grade as of October 1997, how-
illustrated a case in which the KMV approach did
ever, and in fact, by October 1998, there was an
a better job of predicting a single default than
enormous disparity between the agency rating and
agency debt ratings. The power curve analysis
the EDF. The rating was BBB-; the EDF was more
typical of companies then rated B- by the agencies. showed that the KMV approach consistently out-

If one looks at companies with a given letter performs the agency ratings. Power curve analysis,
rating, one can almost invariably find a significant however, does not rule out that in some cases, the
group of companies with EDFs that are both signif- agency debt ratings may do a better job of predict-
icantly higher and lower than what is typical for ing default than the KMV analysis. If so, the impli-
that rating. These companies are usually corpora- cation would be that some combination of the
tions whose credit quality has changed signifi- approaches might be superior to either alone, so

Figure 4. EDF versus S&P Rating for Harnischfeger Industries, November 1994-October 1999
Rating/Default Probability

D 20

CE 15
10
7
5

B 2

BB
0.50

0.20
BBB 0 .20._............. ..... . . . . . . . . . . .. . . . . . . . . . . .
0.15 \ / S&P Rating

A
A0.10 t \<
0.05
AA

AAA 0.02 _ E l l l l l l l l

11/94 4/95 10/95 4/96 10/96 4/97 10/97 4/98 10/98 4/99 10/99

36 ?2003, AIMR?

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
 Quantifying Credit Risk I

Figure 5. Market Value of Assets and Equity and Total Book Liabilities for
Harnischfeger Industries, November 1994-October 1999

Value ($ millions)
4,500

4,000 -

3,500 -Ast

3,000

2,500

2,000
*Liabilities

. . ... ..... _..

1,000_ *
Equity
500

0 I I

11/94 4/95 10/95 4/96 10/96 4/97 10/97 4/98 10/98 4/99 10/99

the KMV model (and, by implication, the Merton every company in the cohort is predicted to have
approach) would be only a partial or incomplete the same default probability by the first measure.
specification. A broad intracohort analysis of EDFs and
To discover whether the agency rating is right agency debt ratings was reported by Miller in 1998.
some of the time and the EDF is right some of the Using data on rated nonfinancial issuers for every
time, the use of intracohort analysis is called for.18 month end between June 1986 and November 1996
Intracohort analysis is a powerful technique for and all rating grades from AAA to B-, Miller stud-
evaluating the "overlap" in the information from ied the subsequent occurrences of default based on
two different measures. Its power results from its the original percentile ranking of the issuer within

generality; if one measure contains useful informa- its fine rating group. Panel A of Figure 6 reproduces
the findings from his study for the default results
tion not contained in the other measure, intracohort
one year later by KMV EDF percentile rank within
analysis will identify it. Coupled with power
the S&P rating cohort.
curves, intracohort analysis provides a complete
The evidence from the chart in Panel A is strik-
characterization of the relationship between two
ing: The defaults occur predominantly among com-
default-prediction measures. The only requirement
panies that were ranked in the lower half of the
for intracohort analysis is that one measure com-
rating grades by EDF; they are ordered approxi-
bine with the other measure in a consistent direc-
mately with their percentile ranking; and there is
tion (but not magnitude) across different credit-
an enormous range of difference in realized default
quality ranges.
rates between the low-ranked and the high-ranked
Given two default-prediction measures, intra-
companies. For instance, 41 defaults occurred in the
cohort analysis examines the subsequent default
worst quartile, and 2 defaults occurred in the best
experience of a cohort of companies constructed to quartile, a difference in default rate of 20:1.
have the same default probability when the first
The power of intracohort analysis is that it does
measure is used but different default probabilities
not rely on specifying anything restrictive about
as measured by the second measure. If the differ- the nature of the relationship between the two mea-
ences resulting from the second measure are infor- sures, such as a linear relationship. It simply tests
mation (not noise), then the subsequent default whether the ranking of one variable has informa-
experience of the cohort should align with the sec- tion that is not contained in the ranking of the other
ond measure. If the differences are noise, one variable. It is difficult to imagine any mathematical
should find no systematic differences in subse- specification of the two variables that would be
quent default experience within the cohort because practically better than the first variable alone if the

January/February 2003 37

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
Financial Analysts Journal

Figure 6. Intracohort Analysis: Twelve Months before Default

A. KMV EDF, June 1986-November 1996 Data
Number of Defaults
14

12

10

0
5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100

KMV Percentiles within S&P Rating Cohorts

B. S&P Rating, 1990-99 Data

Number of Defaults
14

12 -

10

0
10 20 30 40 50 60 70 80 90 100

S&P Percentiles within EDF Cohorts

second variable failed to have additional predictive that was used for the power test for the 1990-99
power by this test. For this reason, the test is con- period. In this case, companies were put into one of
sidered a nonparametric test (i.e., one that does not 60 cohorts based on their EDFs. Each cohort was
rely on a particular parameterization of the rela- sorted by agency debt rating from high to low and
tionship between the two variables but considers given a percentile score. The subsequent defaults
all reasonable alternatives simultaneously). were then determined and compared with their
Miller did not conduct the reverse compari- percentile scores.
son-that is, constructing EDF cohorts and testing Under the hypothesis that agency debt ratings
whether the agency rating differences corre- contain information that is not contained in EDFs,
sponded to different realized default rates. His the high and low percentile groups resulting from
results show that a large amount of the variation of this test should reflect a marked difference. The
EDFs within a given rating grade reflects real infor- results shown in Panel B of Figure 6 indicate, how-
mation. The results do not rule out, however, that ever, that there is virtually no difference in default
there may be some information in agency ratings rate between the high- and low-rated companies
that is not contained within EDFs. within EDF cohorts. For instance, comparing the
To address this point, Kurbat and I (2001) bottom half of the rated companies to the top half
conducted the reverse analysis using the same data of rated companies yields almost exactly the same

38 ?2003, AIMR?

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
 Quantifying Credit Risk I

default rate. Unlike the case of the EDF percentiles tified the overall median EDF for the five-year time
Miller found, in which the hypothesis of no addi- period and established the upper and lower bound
tional EDF information could be easily rejected at for each fine grade as the geometric average of the
high levels of statistical significance, the results median for the fine grade and the median of the
shown in Panel B are statistically and economically neighboring fine grade. They considered that the
insignificant.19 EDF had predicted an upgrade (or downgrade) if
What the power curve and the intracohort it was less than the lower bound (or greater than
analyses demonstrate is that not only is the KMV the upper bound) at a given horizon before the
model a more accurate predictor of default than are actual event of agency upgrade or downgrade. The
agency debt ratings (be they Moody's or S&P rat- patterns of upgrades and downgrades as predicted
ings) but also that EDFs contain all the information by EDF are shown in Figure 7.
in ratings and more. Put differently, if one tried to Figure 7 shows that the KMV model's EDF
combine agency debt ratings with EDFs to obtain a correctly predicted nearly 80 percent (averaging the
better predictor, one would find that the optimal upgrades and downgrades) of the sub-investment-
predictor puts a zero weight on ratings. The output grade rating changes six months in advance and
of the KMV model cannot be improved by mixing approximately 65 percent of the investment-grade
it with debt ratings. Although this result might rating changes. A careful examination of the cases
seem surprising, it is consistent with much empir- revealed that a high proportion of the investment-
ical research on the quality of the information con- grade downgrades were linked to "announcement"
tained in equity prices (see Fama 1970, 1991). events, mostly announcements of mergers, acquisi-
If the errors in agency debt ratings are the tions, or significant restructurings. If these events
result of staleness, one should be able to use the had been excluded (because they are not very pre-
KMV model to predict rating upgrades and down- dictable by any model), the fraction of correctly
grades. To test this hypothesis, Bohn and Korablev predicted investment-grade rating changes would
(2000) constructed a sample of companies using all have been closer to 70 percent and would be above
companies that had both EDFs and S&P debt rat- 75 percent at some time prior to the actual rating
ings at some time during the period September change.
1993 through August 1998. This sample contained This analysis provides additional evidence
1,639 observations distributed over essentially all that at least a portion of the error in agency ratings
industry groups. They used fine grades to put the is a result of lack of timeliness. EDFs, because they
agency debt ratings onto a linear scale with are based partly on equity prices, move in advance
AAA = 1 and D = 22. For each fine grade, they iden- of agency ratings.

Figure 7. Correct Predictions of Upgrades and Downgrades by EDF,

September 1993-August 1998
Correct (%)

90

85 Sub-Investment-Grade =
80 _ Downgrades.
0 Sub-Investment-Grade
75
75 .... Upgrades

70 '':Investrnent-Gradel
65 ~~~~~~~~ ~~~Upgrades --t
60
60 vestment-Grade
Downgrades
55

50 l
-12 -9 -6 -3 -1

Months before Upgrade/Downgrade

Note: North American companies.

January/February 2003 39

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
Financial Analysts Journal

Using Rating Changes to Measure Credit grades over this period. A company was assumed
Risk. Because EDFs provide a more accurate mea- to "migrate" from one grade to another when its
sure of current default probability than agency rat-EDF moved from one range to another. By con-
ings, they can be used to examine the properties of struction, these ranges were nonoverlapping. EDFs
measurement schemes based on agency ratings. For were bounded above by 20 percent. All companies
example, consider the following approach that has that had maximum EDF values but were not actu-
been popularized in recent years (Gupton, Finger, ally in default were treated as belonging to the CCC
and Bhatia 1997). One begins by assuming that com- range. Some companies disappeared within the
panies' default probabilities change through time measurement horizon without defaulting, but
according to a Markov process. One then makes the including them as an additional migration category
further assumption that the states of the process are
did not materially alter the results, so they were
represented by agency debt ratings (including
dropped in the reported results. The fraction of
default as a terminal state). On that basis, one then
companies migrating to each range within one year
substitutes an empirically measured rating transi-
was aggregated over each initial month date to
tion matrix to make the model operational.
produce the final transition matrix.
The key problem with using such matrixes is
The matrix based on agency ratings and the
the implicit assumption that ratings change when-
matrix based on nonoverlapping EDF ranges are
ever credit quality changes and credit quality
shown in Table 1. The differences between the two
changes whenever ratings change. As should be
matrixes are striking. With EDFs, the actual proba-
clear from the preceding section, this assumption
bility that a company will remain within a credit
is not even approximately satisfied.
quality range typical of a given rating is about 45
Kealhofer, Kwok, and Weng (1998) explored
percent; for rating transitions, the probability
the empirical implications of using rating transition
reported is 85 percent. The ratings approach over-
matrixes by comparing EDF-based migration prob-
abilities with rating transitions. Using as a selection estimates the probability of unchanged credit qual-

criterion all U.S. companies with at least two ity by a factor of about 2. The cause appears to be,
monthly EDFs available between February 1990 at least in part, the lack of timeliness in agency rating
and June 1997, they identified a sample of 11,397 changes. Ratings are "sticky"; they do not change
unique companies, averaging around 6,000 compa- enough and thus significantly overstate the proba-
nies a year. Next, they determined ranges for rating bility of credit quality not changing. This stickiness
grades by taking the geometric average of the is reflected also in the enormous range of EDFs that
median EDFs associated with S&P debt rating can be found within any given letter grade.

Table 1. Credit-Quality Transition Matrixes, February 1990-June 1997

Rating at Year-End

Initial Rating AAA AA A BBB BB B CCC Default

A. Based on agency ratings

AAA 90.81% 8.33% 0.68% 0.06% 0.12% 0.00% 0.00% 0.00%

AA 0.70 90.65 7.79 0.64 0.06 0.14 0.02 0.00
A 0.09 2.27 91.05 5.52 0.74 0.26 0.01 0.06
BBB 0.02 0.33 5.95 86.93 5.30 1.17 0.12 0.18
BB 0.03 0.14 0.67 7.73 80.53 8.84 1.00 1.06
B 0.00 0.11 0.24 0.43 6.48 83.46 4.07 5.20
CCC 0.22 0.00 0.22 1.30 2.38 11.24 64.86 19.79

B. Based on nonoverlapping EDF ranges

AAA 66.26% 22.22% 7.37% 2.45% 0.86% 0.67% 0.14% 0.02%

AA 21.66 43.04 25.83 6.56 1.99 0.68 0.20 0.04
A 2.76 20.34 44.19 22.94 7.42 1.97 0.28 0.10
BBB 0.30 2.80 22.63 42.54 23.52 6.95 1.00 0.26
BB 0.08 0.24 3.69 22.93 44.41 24.53 3.41 0.71
B 0.01 0.05 0.39 3.48 20.47 53.00 20.58 2.01
CCC 0.00 0.01 0.09 0.26 1.79 17.77 69.94 10.13

Sources: For Panel A, Standard & Poor's CreditWeek (15 April 1996); for Panel B, KMV Corporation (January 1990-September 1995).

40 ?2003, AIMR?

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
 Quantifying Credit Risk I

Not surprisingly, using rating transitions to Statistical Methods for Estimating

proxy for credit-quality changes will result in
substantial errors in any applications. Panel A in Default Risk
Table 2 shows a calculation from the CreditMetrics An alternative approach to Merton for estimating
documentation based on a rating transition matrix; default risk is the use of statistical methods to deter-
Panel B shows a recalculation based on applying mine the relationship between realized defaults
the EDF-based transition matrix. These calcula- and their characteristics prior to default. This article
tions are to determine the standard deviation of the does not have space to address the performance of
bond's value arising from credit risk at a horizon of Merton versus such a statistical scoring model, but
one year-that is, to measure the risk of the bond. two points are worth noting. First, in our research,
Because the bond's future value (and thus its risk)
we have not been able to identify statistical models
will have little variation if credit quality is believed
that perform as well as the Merton approach empir-
to have not changed, the use of the matrix based on
ically.20 Second, statistical models measure correla-
agency rating transitions results in a significant
tion, not causation. The confusion between these
understatement of risk. In this case, the EDF-based
concepts can be illustrated by considering the prob-
matrix results in a standard deviation of 4.86 per-
lem of predicting whether someone is overweight
cent versus the rating-transition-based value of
2.99 percent, a 60 percent error. This same criticism by examining the contents of that person's shop-

applies to transition-matrix-based valuation ping basket. In doing so, we might observe a corre-
approaches to credit risk, such as that of Jarrow, lation between people being overweight and
Lando, and Tumbull (1997), when they are imple- buying diet food, but we would not want to jump
mented by using ratings data to estimate the tran- to the (causation) conclusion that "eating diet food
sition matrix. makes you fat."

Table 2. Bond Standard Deviation Calculation: CreditMetrics Example

Deviation of Probability-
Probability New Bond Value Probability- Bond Value from Weighted Squared
Year-End Rating of State plus Coupon Weighted Value Mean Value Deviation

A. Based on rating changes

AAA 0.02% $109.37 $0.02 $2.28 0.0010
AA 0.33 109.19 0.36 2.10 0.0146

A 5.95 108.66 6.47 1.57 0.1474

BBB 86.93 107.55 93.49 0.46 0.1853

BB 5.30 102.02 5.41 -5.06 1.3592

B 1.17 98.10 1.15 -8.99 0.9446

CCC 0.12 83.64 0.10 -23.45 0.6598

Default 0.18 51.13 0.09 -55.96 5.6363

Mean = $107.09 Variance = 8.9477

Standard deviation = 2.99%

B. Based on nonoverlapping EDF ranges

AAA 0.30% $109.37 $0.33 $3.86 0.0447
AA 2.80 109.19 3.06 3.68 0.3793

A 22.63 108.66 24.59 3.15 2.2464

BBB 42.54 107.55 45.75 2.04 1.7715

BB 23.52 102.02 24.00 -3.49 2.8637

B 6.95 98.10 6.82 -7.41 3.8154

CCC 1.00 83.64 0.84 -21.87 4.7827

Default 0.26 51.13 0.13 -54.38 7.6885

Mean = $105.51 Variance = 23.5922

Standard deviation = 4.86%

Note: Calculations of volatility in value due to possible credit-quality changes for a single bond.
Sources: For Panel A, Gupton et al.; for Panel B, KMV Corporation.

January/February 2003 41

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
Financial Analysts Journal

This point is not a hypothetical criticism, as is default rate within agency rating groups is actually
illustrated by the effect of leverage on a company's highly diverse. Furthermore, these differences can
default risk. Leverage by itself has little default- be accurately detected in a timely way by the KMV
predictive power. The reason is straightforward. model.
Generally, lenders provide the most leverage to the Cumulatively, this research constitutes strong
safest companies-in other words, the ones with evidence in favor of the generalized Black-Scholes-
the least business risk. This fact can be seen in the Merton approach to credit risk, as exemplified by
cross-sectional relationship between company the KMV model. Contrary to previous studies that
asset volatility and leverage. On average, leverage were unsuccessful in implementing the Merton
goes up as volatility goes down. model, this body of work shows that a variant of
A statistical model cannot completely separate the Merton model does a superior job of predicting
these two effects. In essence, if a statistical model is and measuring default risk when compared with
asked to report the effect on default risk of raising conventional credit measures.
a given company's leverage, it will say that the risk
has gone up because of greater leverage but also
Conclusion
that the risk has fallen because the company is now
Twenty-five years ago, the work of Black, Scholes,
less volatile. This answer seems patently silly, but
and Merton held out the hope of objectively quan-
it arises because the model has gotten the causation
tifying credit risk. From a conceptual standpoint,
reversed: Lending more to a company does not
the Merton model has come to dominate the view-
make it less volatile; rather, on average, the less
point of academic finance on credit risk. But imple-
volatile companies use more leverage.
menting the model in practice has proven to be
Most model users are interested in causation
more difficult than originally anticipated.
rather than correlation because they need to deter-
Building on the original insights of Black-
mine the effect of their or other's actions on a com-
Scholes-Merton, Vasicek and Kealhofer developed
pany's capital structure. For example, if an analyst
a straightforward variant of the Merton model.
changes the inputs to a statistical model to deter-
During the past decade, the probabilities of default
mine the effect of, say, making a loan, the analyst is
determined from the KMV model, expected default
using the model to make a causal inference. The
frequencies or EDFs, have become a major tool of
inference will be correct only if the statistical model
credit-risk transactors (lenders, borrowers, traders,
happens to correspond to the true underlying
structural model and is complete, in the sense of etc.) and portfolio managers.21
correctly capturing the effect of all relevant vari- Cumulatively, the research described here pro-

ables. Few statistical modelers would be willing to vides strong support for the generalized Black-

make such an assertion. Scholes-Merton approach to credit risk, as exempli-

In contrast, the Merton approach can be fied by the KMV model. Because the Merton

adapted to include relevant variables within a approach provides a logically complete and coher-

causal framework. Thus, Merton models can be ent framework for understanding corporate risk, a

used to determine the effect of increasing divi- successful practical implementation of the
dends, shortening debt maturity, adding leverage, approach provides a powerful tool for examining
issuing convertible rather than straight debt, and so the valuation consequences of corporate behavior.
forth. In practice, the KMV model has been exten- Many more companies have equity that is
sively used by banks to analyze complex corporate traded than debt that is traded. Equity markets are
finance transactions, such as mergers, acquisitions, more continuous and more transparent than debt
and divestitures. markets. Therefore, by providing a blueprint for
In summary, empirical studies of the KMV using information from the price of a company's
model show that it works better than agency rat- equity to determine the default risk of any of the
ings as a predictor of default. Intracohort analysis company's liabilities, a practical Black-Scholes-
shows that EDFs contain default-predictive infor- Merton approach is a crucial analytical tool. The
mation that is not in agency ratings but that agency workable integration of corporate equity and fixed-
ratings appear to contain no default-predictive income valuation creates the potential for a deeper
information that is not in EDFs. Analysis of agency understanding of corporate finance, greater debt-
debt ratings suggests the source of these results, market liquidity, and new paradigms for managing
namely, that there is strong evidence that the investments in corporate obligations.

42 ?2003, AIMR?l

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
 Quantifying Credit Risk I

Notes

1. For instance, Moody's Investors Service has decided to use 12. The agencies themselves have used similar procedures
a Merton approach to predict default (see Cass 2000). when performing historical statistical studies (see, for
2. For a description of theoretical issues underlying the imple- example, Carty and Fons 1993).
mentation of the KMV model, see Crouhy, Galai, and Mark 13. Because the nondefaulted companies did not have a default
(2001). date, of course, there is no specific date to measure their
3. The implicit probability is the so-called risk-neutral proba- default risk. Several approaches were used, including
bility (i.e., the one that would obtain if the underlying asset matching dates and scoring percentiles by year. The various
value process had a mean return equal to the risk-free approaches yielded virtually identical results.
return rather than its actual mean return). This point is 14. If p is the probability of default, the variance of an observa-
discussed further in Part II. tion is given by p(l - p) and the standard deviation by
4. More complete treatment of the KMV model from a theo- /[p(l - p)]/n, where n is the number of total observations
retical perspective can be found in Caouette, Altman, and corresponding to p, under the assumption of independence
Narayanan (1998), Crouhy et al., and Vasicek (1984). of defaults. Independence is a reasonable assumption in this
5. Part II will provide a discussion of the "horizon" versus case because the defaults occur primarily at different times
"absorbing barrier" versions of this calculation. throughout the sample period. For the example here, with
6. The expected return or "asset drift rate" is a required input a Type II error rate of 20 percent in which EDF excludes 72
for the model. Part II will address the method used and the
percent of defaults versus only 61 percent for the agency
implications for empirical estimates of default probabilities
ratings (for a difference of 0.72 - 0.61 = 0.11), EDF is 2.4, so
and secondary-market credit spreads.
p = 0.024 and Jp(l -p) = 0.153. Given the large sample
7. For estimating asset volatility, one uses a procedure similar
sizes involved (respectively, 7,320 and 13,591), the standard
to that explained in Jones et al.
errors are, respectively, approximately 0.0018 and 0.0013.
8. From the standpoint of logical consistency, one would want
Thus, the difference in power between the EDFs and the
to have agreement between the distributional assumption
ratings is significant, with t-values of, respectively, approx-
used to estimate the asset value and the distributional form
used to obtain the default probabilities. KMV's practical imately 63 and 85.

experience has been that the empirical deviations, although 15. Note that the studies being described were carried out with
important for default estimation, do not have substantial the total population excluded on the x-axis rather than
effects on the estimated asset values or volatilities. As a nondefaults excluded on the x-axis. This distinction is of no
result, we have simply accepted the logical inconsistency as practical importance, however, because the number of com-
a practical expedient, and we look to future work to find its panies defaulting a year is approximately 1 percent of the
resolution. number of nondefaulting companies a year.
9. The comparison is not always feasible because some indus- 16. The slight superiority of the actual S&P company ratings to
tries have numerous defaults but many industries do not. the implied senior unsecured ratings is plausible, given that
For instance, utilities may have a different experience from the algorithm discussed previously for implying ratings is
that of other companies for any number of postulated ex ante presented by the rating agencies as an approximation to
reasons. The default data on utilities are statistically consis- what they use to imply ratings in practice, and it produces
tent with the pooled results, but the number of defaults is a mild degradation in the power curve results.
so low that the data would also be consistent with many 17. For purposes of graphing, the agency debt ratings are
other alternatives. depicted at the level of default risk consistent with the
10. A "higher" index value is defined as a lower probability of median EDF of all companies with the given rating over the
default. five-year period.

11. S&P states that its measures actually represent default prob- 18. The idea of intracohort analysis was suggested by Miller
ability, but it adjusts ratings for subordination and to some (1998).
extent for security, which is not consistent with this 19. In addition to agency ratings, we also tested various
premise. Moody's, presenting a similar rating approach, accounting variables for marginal default-predictive
interprets its ratings as representing expected loss (i.e., the power. In all cases, marginal default-predictive power was
product of default probability and expected loss given decisively rejected.
default). Thus, in both cases, the ratings need to be modified 20. See, for instance, Kealhofer and Kurbat.
by eliminating the effect of subordination and security. I do 21. Smithson (2001) reported a survey of banks' use of default
not address Fitch IBCA ratings in this article. models.

References
Altman, Edward I. 1968. "Financial Ratios, Discriminant Black, Fischer, and John Cox. 1976. "Valuing Corporate
Analysis and the Prediction of Corporate Bankruptcy." Journal Securities: Some Effects of Bond Indenture Provisions." Journal
of Finance, vol. 23, no. 4 (September):589-609. of Finance, vol. 31, no. 2 (May):351-367.
Altman, Edward I., and Vellore M. Kishore. 1996. "Almost Black, Fischer, and Myron Scholes. 1973. "The Pricing of Options
Everything You Wanted to Know about Recoveries on and Corporate Liabilities." Journal of Political Economy, vol. 81,
Defaulted Bonds." Financial Analysts Journal, vol. 52, no. 6 no. 3 (May/June):637-659.
(November/December):57-64. Bohn, Jeffrey R. 2000a. "A Survey of Contingent-Claims
Altman, Edward I., R. Haldeman, and P. Narayanan. 1977. Approaches to Risky Debt Valuation." Journal of Risk Finance,
"ZETATm Analysis: A New Model for Bankruptcy vol. 1, no. 3 (Spring):53-78.
Classification." Journal of Banking and Finance, vol. 1, no. 1
(June):29-54.

January/February 2003 43

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms
Financial Analysts Journal

. 2000b. "An Empirical Assessment of a Simple Jones, P., S. Mason, and E. Rosenfeld. 1984. "Contingent Claims
Contingent-Claims Model for the Valuation of Risky Debt." Analysis of Corporate Capital Structures: An Empirical
Journal of Risk Finance, vol. 1, no. 4 (Summer):55-77. Investigation." Journal of Finance, vol. 39, no. 3 (July):611-625.
Bohn, Jeffrey R., and Irina Korablev. 2000. "Predicting Agency Kao, Duen-Li. 2000. "Estimating and Pricing Credit Risk: An
Rating Upgrades and Downgrades with KMV's Expected Overview." Financial Analysts Journal, vol. 56, no. 4 (July/'
Default Frequency (EDF)." KMV Corporation. August):50-66.
Carty, Lea V. 1997. "Moody's Rating Migration and Credit Kealhofer, Stephen. 1991. "Credit Monitor Overview."
Quality Correlation, 1920-1996." Moody's Investors Service. Unpublished. KMV Corporation.
Carty, Lea V., and Jerome S. Fons. 1993. "Measuring Changes in Kealhofer, Stephen, and Matthew Kurbat. 2001. "The Default
Corporate Credit Quality." Moody's Special Report Prediction Power of the Merton Approach, Relative to Debt
(November). Ratings and Accounting Variables." Unpublished paper, KMV
Carty, Lea V., and Dana Lieberman. 1996. "Corporate Bond Corporation (www.kmv.com/Knowledge-Base/public/
Defaults and Default Rates 1938-1995." Moody's Investors general/white/Merton_Prediction_Power.PDF).
Service, Global Credit Research (January). Kealhofer, Stephen, Sherry Kwok, and Wenlong Weng. 1998.
Cass, Dwight. 2000. "Moody's Launches Risk Model." Risk, vol. "Uses and Abuses of Bond Default Rates." CreditMetrics Monitor
13, no. 6 (June):14-15. (First Quarter):37-55

Caouette, John B., Edward I. Altman, and Paul Narayanan. 1998. Lim, Francis. 1999. "Comparative Default Predictive Power of
Managing Credit Risk: The Next Great Financial Challenge. New EDFs and Agency Debt Ratings." Unpublished paper, KMV
York: Wiley. Corporation (December).
Collin-Dufresne, Pierre, and Bruno Solnik. 2001. "On the Term Merton, Robert C. 1973. "Theory of Rational Option Pricing."
Structure of Default Premia in the Swap and LIBOR Markets." Bell Journal of Economics, vol. 4, no. 1 (Spring):141-183.
Journal of Finance, vol. 56, no. 3 (June):1095-1115.
. 1974. "On the Pricing of Corporate Debt: The Risk
Crouhy, Michel, Dan Galai, and Robert Mark. 2001. Risk Structure of Interest Rates." Journal of Finance, vol. 29, no. 2
Management. New York: McGraw-Hill.
(May):449-470.
Delianedis, Gordon, and Robert Geske. 1998. "Credit Risk and Miller, Ross. 1998. "Refining Ratings." Risk, vol. 11, no. 8
Risk Neutral Default Probabilities: Information about Rating (August):97-99.
Migrations and Defaults." Working Paper #19-98, University of
Ogden, Joseph P. 1987. "Determinants of the Ratings and Yields
California at Los Angeles.
on Corporate Bonds: Tests of the Contingent Claims Model."
Demchak, Bill. 2000. "Modelling Credit Migration." Risk, vol. 13,
Journal of Financial Research, vol. 10, no. 4 (Winter):329-339.
no. 2 (February):99-102.
Pitts, C.G., and M.J. Selby. 1983. "The Pricing of Corporate Debt:
Eom, Young Ho, Jean Helwege, and Jing-Zhi Huang. 2000.
A Further Note." Journal of Finance, vol. 38, no. 4
"Structural Models of Corporate Bond Pricing: An Empirical
(September):1311-13.
Analysis." Working paper, Yonsei University.
Sarig, Oded, and Arthur Warga. 1989. "Some Empirical
Fama, Eugene F. 1970. "Efficient Capital Markets: A Review of
Estimates of the Risk Structure of Interest Rates." Joiurnal of
Theory and Empirical Work." Journal of Finance, vol. 25, no. 2
Finance, vol. 44, no. 4 (December):1351-60.
(May):383-417.
Smithson, Charles W. 2001. "Managing Loans: The State of Play."
. 1991. "Efficient Capital Markets: II." Journal of Finance,
Credit (December/January):26-31
vol. 46, no. 5 (December):1575-1617.
Sobehart, J., and S. Keenan. 1999. "Equity Market Value and Its
Gupton, Greg M., Christopher C. Finger, and Mickey Bhatia.
Importance for Credit Analysis: Facts and Fiction." Working
1997. "CreditMetricST-Technical Document." J.P. Morgan
paper, Moody's Investor Services.
(www.defaultrisk.com/pp-model_20.htm).
Standard & Poor's. 1991. "S&P's Corporate Finance Criteria."
Gurnaney, Anil. 1998. "Empirical Test of EDF, 1990-1996." KMV
Corporation. Vasicek, Oldrich A. 1984. "Credit Valuation." Unpublished
paper, KMV Corporation.
Ingersoll, Jonathon. 1977a. "A Contingent-Claims Valuation of
Convertible Securities." Journal of Financial Economics, vol. 4, no. .1995. "EDFs and Corporate Bond Pricing." Unpublished
3 (May):289-321. paper, KMV Corporation.

. 1977b. "An Examination of Corporate Call Policies on . 1997. "Credit Valuation." NetExposure, vol. 1 (October);
Convertible Securities." Journal of Finance, vol. 32, no. 2 reprinted in 1998 on CD in Derivatives: Theory and Practice of
(May):463-478. Financial Engineering. Edited by P. Wilmott. London: J. Wiley &
Sons.
Jarrow, Robert A., and Donald R. van Deventer. 1999. "Practical
Usage of Credit Risk Models in Loan Portfolio and Counterparty . 2001. "Actual and Risk-Neutral Probabilities of
Exposure Management." In Credit Risk: Models and Management. Default." Unpublished paper, KMV Corporation.
Edited by David Shimko. London: Risk Books. Wei, David Guoming, and Dajiang Guo. 1997. "Pricing Risky
Jarrow, Robert A., D. Lando, and S. Turnbull. 1997. "A Markov Debt: An Empirical Comparison of the Longstaff and Schwartz
Model for the Term Structure of Credit Risk Spreads." Reviezv of and Merton Models." Journal of Fixed Inconie, vol. 7, no. 2
Financial Studies, vol. 10, no. 2 (Summer):481-523. (September):8-28.

44 ?2003, AIMR?

This content downloaded from 52.13.166.141 on Wed, 06 May 2020 21:36:15 UTC
All use subject to https://about.jstor.org/terms