2

The Structural Approach to Default

Prediction and Valuation

Structural models of default risk are cause-and-effect models. From eco-

nomic reasoning, we identify conditions under which we expect borrow-
ers to default and then estimate the probability that these conditions
come about to obtain an estimate of the default probability.

For limited liability companies, default is expected to occur if the asset

value (i.e., the value of the firm) is not sufficient to cover the firm's liabili-
ties. Why should this be so? From the identity

Asset value = Value of equity + Value of liabilities

and the rule that equity holders receive the residual value of the firm, it
follows that the value of equity is negative if the asset value is smaller
than the value of liabilities. If you have something with negative value,
and you can give it away at no cost, you are more than willing to do so.
This is what equity holders are expected to do. They exercise the walk-
away option that they have because of limited liability and leave the firm
to the creditors. As the asset value is smaller than the value of liabilities,
creditors' claims are not fully covered, meaning that the firm is in default.
The walk-away option can be priced with standard approaches from op-
tion pricing theory.

This is why structural models are also called option-theoretic or contin-

gent-claim models. Another common name is Merton models, because
Robert C. Merton (1974) first applied option theory to the problem of
valuing a firm's liabilities in the presence of default and limited liability.
In this chapter, we first explain how structural models can be used for es-
timating default probabilities and valuing a firm's liabilities. We then
show how to implement structural models in the spirit of the original
Merton model. The focus is on the estimation of default probabilities
rather than valuation. Then we will implement a widely used alternative
model: the CreditGrades approach. With this model we conduct a case
study of Lehman Brothers prior to its default in September 2008.

DEFAULT AND VALUATION IN A STRUCTURAL

MODEL

The basic premise of structural models is that default occurs if the value
of the assets falls below a critical value associated with the firm's liabili-
ties. To clarify the issues, we consider the simple set-up examined by
Merton (1974): the firm's liabilities consist of just one zero-coupon bond
with principal L maturing in T. There are no payments up until T, and eq-
uity holders will wait until T before they decide whether to default or not.
(If they defaulted before T they would forego the chance of benefiting
from an increase of the asset value). Accordingly, the default probability
is then the probability that, at time T, the value of the assets is below the
value of the liabilities.

Figure 2.1 Default probability in the Merton model

What is required to determine this probability? In Figure 2.1 we get the

firm's liability from the balance sheet (hoping that it is not manipulated).
We then need to specify the probability distribution of the asset value at
maturity T. A common assumption is that the value of financial assets fol-
lows a lognormal distribution, i.e., that the logarithm of the asset value is
normally distributed. We denote the per annum variance of the log asset
value changes by σ2. The expected per annum change in log asset values
is denoted μ – σ2/2, where μ is the drift parameter.1 Let t denote today. The
log asset value in T thus follows a normal distribution with the following
parameters:

If we know L, At, μ and σ2, determining the default probability is an exer-

cise in elementary statistics. In general, the probability that a normally
distributed variable x falls below z is given by Φ[(z – E[x])/σ (x)], where Φ
denotes the cumulative standard normal distribution. Applying this result
to our case, we get

In the literature, one often uses the term distance to default (DD). It mea-
sures the number of standard deviations the expected asset value AT is
away from the default. We can therefore write

So far, we have not used any option pricing formula. In fact, there is no
theoretical reason why we need them to determine default probabilities,
but instead a practical one: for a typical firm, we cannot observe the mar-
ket value of assets. What we can observe are book values of assets, which
can diverge from market values for many reasons. If we do not observe
asset values, we do not know today's asset value At needed for formula
(2.2). In addition, we cannot use observed asset values to derive an esti-
mate of the asset volatility σ.

Option pricing theory can help because it implies a relationship between

the unobservable (At, σ) and observable variables. For publicly traded
firms, we observe the market value of equity, which is given by the share
price multiplied with the number of outstanding shares. At maturity T,
we can establish the following relationship between equity value and as-
set value (see Figure 2.2): as long as the asset value is below the value of
liabilities, the value of equity is zero because all assets are claimed by the
bond holders. If the asset value is higher than the principal of the zero-
coupon bond, however, equity holders receive the residual value, and
their pay-off increases linearly with the asset value.

Mathematically, the pay-off to equity holders can be described as

Figure 2.2 Payoff to equity and bond holders at maturity T

This is the pay-off of a European call option. The underlying of the call
are the firm's assets; the call's strike is L. The pay-off to bond holders cor-
responds to a portfolio composed of a risk-free zero-coupon bond with
principal L and a short put on the firm's assets, again with strike L.

If the firm pays no dividends, the equity value can be determined with
the standard Black-Scholes call option formula:

where r denotes the logarithmic risk-free rate of return.

Remember our problem of determining the asset value At and the asset
volatility σ. We now have an equation that links an observable value (the
equity value) to these two unknowns (σ enters Equation (2.5) via Equation
(2.6)). However, we have only one equation, but two unknown variables.
So where do we go from there? We can go back into the past to increase
the available information. There are several ways of using this informa-
tion, and we illustrate two different methods in the next two sections.

IMPLEMENTING THE MERTON MODEL WITH A

ONE-YEAR HORIZON

The iterative approach

Rearranging the Black-Scholes formula (2.5), we get

If we go back in time, say 260 trading days, we get a system of equations:

For simplicity, we have not added time subscripts to the d1s and d2s,
whereas we have added them to the other variables that can change over
time. Using time-varying interest rates and liabilities is somehow incon-
sistent with the Merton model, in which both are constant. However, we
can hope to come closer to market valuations with this approach, because
the latter will be based on the information the market has at a particular
date.

Equation system (2.8) is composed of 261 equations in 261 unknowns (the

asset values). Have we made any progress? Although it seems as if we
have an additional unknown variable, the asset volatility σ, this should
not bother us, because this variable can be estimated from a time series
of As. Therefore, the system of equations can be solved.

Before applying this procedure to an example firm, however, we have to

translate the stylized firm of the Merton model into the real world.
Typical firms have many different liabilities maturing at different points
in time – from one day to 30 years or more. One solution often found in
the literature is the following: assume that the firm has only liabilities
that mature in one year. The choice may appear to be ad hoc, and outra-
geously so. It is largely motivated by convenience. Structural models are
often used to produce one-year default probabilities. Had we assumed a
maturity of, say, three years, it would not have been obvious how to con-
vert the three-year default probability to a one-year probability.

If we make the ad hoc assumption that the maturity is one year, there is
no reason why we should not apply it to every day in the past. On the con-
trary, it seems natural because firms often have relatively stable maturity
structures, i.e., issue new debt when some part of the debt is retired.
Setting (T – t) to one for each day within the preceding twelve months,
(2.8) simplifies to

This system of equations can be solved through the following iterative

procedure:

Iteration 0: Set starting values At–a for each a = 0, 1,…, 260. A sensible choice is to set the At–
a equal to the sum of the market value of equity Et–a and the book value of liabilities Lt–a.
Set σ equal to the standard deviation of the log asset returns computed with the At–a.

For any further iteration k = 1,…, end

Iteration k: Insert At–a and σ from the previous iteration into the Black-Scholes formulae d1
and d2. Input these d1 and d2 into Equation (2.7) to compute the new At–a. Again use the At–
a to compute the asset volatility.

We go on until the procedure converges. One way of checking conver-

gence is to examine the change in the asset values from one iteration to
the next. If the sum of squared differences between consecutive asset val-
ues is below some small value (such as 10−10) we stop.

We will now implement this procedure for Enron, three months before its
default in December 2001. At that time, this default was the biggest corpo-
rate default ever. It also caught many investors by surprise because
Enron had decent agency ratings until a few days before default.

We collect quarterly data on Enron's liabilities from the SEC Edgar data
base. The one-year US treasury rate serves as the risk-free rate of return2
and the market value of equity can be obtained from various data
providers. When linking the daily data on equity value with the quarterly
liability data, we take the most recent, available data. The date of avail-
ability is taken to be the filing date stated in the SEC filings. On July 31,
2001, for example, the liability data is from the report for the first quarter
of 2001, filed on May 15, 2001. We therefore use only information actually
available to the market at our valuation date.

Table 2.1 Using the iterative approach to estimate asset values and asset
volatility

The data and calculations are contained in Table 2.1. We start by entering
the initial values for the asset value in column F. Our guess is that the as-
set value equals the market value of equity plus the (book) value of liabili-
ties. Cell F4, for example, would read = B4 + C4.

Column G contains the system of Equations (2.9). For each day, we com-
pute the asset value using the rearranged Black-Scholes formula. For con-
venience we write a VBA-function BSd to compute the d1 as given in
Equation (2.6):
 Function BSd(S, x, h, r, sigma)
‘S=value underlying, x=strike, h=time to maturity,
‘r=risk-free rate, sigma=volatility underlying
BSd = (Log(S / x) + (r + 0.5 * sigma ^ 2) * h) / (sigma * h ^ 0.5)
End Function

Note that the horizon (T – t) is here denoted by h.

In column H, we compute the log returns of the asset values from column
F. We use the function STDEV to determine their standard deviation and
multiply the result with the square-root of 260 (the number of trading
days within a year) to transform it into a per annum volatility (this is an
application of the root-T-rule explained in Box 2.1).

Box 2.1 Root-T-rule for scaling standard deviations of return

The percentage price change over T periods from t = 0 to t = T can be writ-

ten as

PT / P0 = R0,T = R1 × R2 × R3 × … × RT

where P denotes price and R the simple, gross return. With logarithmic
returns r = ln(R) we have (recall ln(x y) = ln(x) + ln(y))

r0,T = r1 + r2 + r3 + … + rT

If the returns are independent across periods, the T-period variance is

just the sum of the one-period variances

Var(r0,T)= Var(r1) + Var(r2) + Var(r3) + … + Var(rT)

If return variances are identical across time, Var(r1) = Var(r2) = … =

Var(rT) = Var(rt), we can then write

Var(r0,T) = T × Var(rt)

For the standard deviation of returns, it follows that

This is the root-T-rule. An example application is the following: we multi-
ply the standard deviation of monthly returns with the square root of 12
to get the annualized standard deviation of returns. The annualized stan-
dard deviation is usually called volatility.
The iterative procedure is implemented through the macro iterate. Its job
is very simple: just copy column G into column F as long as the sum of
squared differences in asset values (in G and F) is above 10−10. The sum of
squared differences is computed in cell I6 using the function SUMXMY2.

For the default probability formula, we need the expected change in asset
values. With the asset values obtained in Table 2.1, we can apply the
standard procedure for estimating expected returns with the Capital
Asset Pricing Model (CAPM). We obtain the beta of the assets with respect
to a market index, and then apply the CAPM formula for the return on an
asset i:

with R denoting the simple risk-free rate of return (R = exp(r) – 1). We

take the S&P 500 index return as a proxy for RM, the return on the market
portfolio. Computations are shown in Table 2.2. We copy the asset values
from column G of Table 2.1 into column B of Table 2.2 and then add the
S&P index values and the risk-free rate of return. In columns F and G, we
compute the excess return on the assets and the S&P 500 (excess return is
return minus risk-free rate).

Table 2.2 Using estimated asset values and the CAPM to derive an esti-
mate of the drift rate of asset returns
By regressing the asset value returns on S&P 500 returns, we obtain an es-
timate of the asset's beta. This is done in Cell I5 with the function SLOPE.
Assuming a standard value of 4% for the market risk premium E[RM] – R,
the expected asset return is then 4.6%. This, however, is not the drift rate
μ that we use in our formula 2.2. The drift rate μ is for logarithmic re-
turns. We determine μ as ln(1.046).

Now that we have estimates of the asset volatility, the asset value and the
drift rate, we can compute the default probability. This is done in Table
2.3. The estimated one-year default probability as of August 31, 2001, is
8.72%.

Table 2.3 Using the estimates to determine the implied default

probability

A solution using equity values and equity volatilities

The iterative solution of the last section used the Black-Scholes formula

and solved the problem of one equation with two unknowns by examin-
ing (2.11) for various dates t.

Another common approach is to use (2.11) for the current date t only, and
introduce another equation that also contains the two unknowns. Since
equity is a call on the asset value, its riskiness depends on the riskiness of
the asset value. Specifically, one can show that the equity volatility σE is
related to the asset value At and the asset volatility σ in the following way:
where d1 is the standard Black-Scholes d1 as given in Equation (2.6). If we
know the equity value Et and have an estimate of the equity volatility σE,
(2.11) and (2.12) are two equations with two unknowns. This system of
equations does not have a closed-form solution, but we can use numerical
routines to solve it.

In the following, we apply this approach to the case study from the previ-
ous section. We use the same data and assumptions, i.e., we set the hori-
zon T – t to one year, we take the equity value Et from the stock market,
set liabilities L equal to book liabilities, and use the one-year yield on US
treasuries as the risk-free rate of return. The only new parameter that we
need is an estimate of the equity volatility σE. We choose to base our esti-
mate on the historical volatility measured over the preceding 260 days.
Data and computations are shown in Table 2.4. Daily Enron stock prices
are in column B.3 They are converted to daily log returns in column C. For
example the formula reads =LN(B3/B2) for cell C3. By applying the STDEV
command to the range containing the returns, we get the standard devia-
tion of daily returns. Multiplying this figure with the square root of 260
gives us the annualized equity volatility (see Box 2.1). The whole formula
for cell E2 then reads =STDEV(C3:C263)*260^0.5.

Table 2.4 Estimating equity volatility from stock prices

Table 2.5 Calibrating the Merton model to equity value and equity
volatility
We now have all the data needed to solve the Black-Scholes equation sys-
tem. This is done in Table 2.5. Our input data is contained in the range
B2:B6.

The unknown parameters are in cells B9:B10. It is necessary to assign fea-

sible initial values to them, i.e., values larger than zero. To speed up the
numerical search procedure, it is also advisable to choose the initial val-
ues such that they are already close to the values that solve the system.

A good choice for the initial asset value in cell B9 is the market value of
equity plus the book value of liabilities. An approximation of the un-
known asset volatility in cell B10 can be based on Equation 2.12. Solving
this equation with respect to σ and assuming Φ(d1) = 1, we get the
approximation

To see why Equation (2.13) is useful, examine when the assumption Φ(d1)
= 1 holds. Through the properties of the normal distribution, Φ(d1) lies be-
tween 0 and 1. For large d1, Φ(d1) approaches 1. Comparing the equation
for d1 (2.6) with the equation for the distance to default (2.3), we see that
they have the same structure, and differ only in the drift rate and the sign
of the variance in the numerator. Thus, a large d1 goes along with a high
distance to default, and a low default probability. If this is true – and most
firms have default probabilities smaller than 5% – the approximation
(2.13) is reasonable.

The option pricing equations are entered in B13:B16. We could again use
our Bd1 function. For the sake of variation, we type the formulae for d1
and d2 in cells B13 and B14, respectively. The two Black-Scholes equations
(2.11) and (2.12) are in cells B15 and B16, respectively.

The equation system is solved if the difference between model values and
observed values is zero. That is, we would like to reach B15=B2 and
B16=B3 by changing B9 and B10. To arrive at a solution, we can minimize
the sum of squared differences between model values and observed val-
ues. Since equity value and equity volatility are of different order, it is ad-
visable to minimize the sum of squared percentage differences.
Otherwise, the numerical routine could be insensitive to errors in equity
volatility and stop short of a solution that sets both equations to zero.

The objective function that we are going to minimize thus reads

which we write in cell B19. We then use the Solver to minimize B19 by
changing B9 and B10 (see screenshot in Table 2.5). The precision option
of the Solver is set to 0.000001. We also tick the options ‘Assume non-nega-
tive’ and ‘Use Automatic Scaling’.

The Solver worked fine in this example – model values are very close to
observed equity values, and convergence was quick. In cases where the
Solver has approached the solution but stopped before errors were close
to zero, try running the Solver again. In cases where the Solver procedure
stops because it considers a value of zero for the asset volatility, add the
constraint B10≥0.000001 in the Solver window. In some cases, playing
around with the objective function might also help.
To compute the default probability, we again need the drift rate of asset
returns. We could, for example, obtain it in a fashion similar to the previ-
ous section. Apply the calculations from Table 2.5 to a series of dates in
the past, obtain a series of asset values, and use the CAPM as in Table 2.2.
For simplicity, we do not spell out the calculations but rather use the drift
rate obtained in the previous section, which was 4.5%. The default proba-
bility can then be determined as in Table 2.3, which gives 0.38%.

Comparing different approaches

The following table summarizes the key results that we obtained with the
two different approaches:

The iterative procedure and the procedure based on solving a system of

two equations yield asset values that are relatively close (the asset value
from the 2-equation approach is 1.9% lower than the one from the itera-
tive approach). The asset volatilities, however, differ dramatically, which
is also the main reason why the default probabilities differ in the way
they do.

This may seem odd because we used the same one-year history of equity
prices in both approaches. However, we used them in different ways. In
the 2-equation approach, we estimated the equity volatility from those
prices. This is a good way of estimating a volatility if we believe it to be
constant across time. But equity is a call option in the Merton world, with
risk varying if the asset-to-equity ratio At/Et varies (see Equation 2.12).
Equivalently, we could also say that equity risk varies with leverage, be-
cause leverage can be measured through (At – Et)/At = 1 – Et/At.

During the time period of the analysis, Enron's asset-to-equity ratio

changed dramatically. Using the figures from Table 2.1, it increased from
1.52 in August 2000 to 2.96 in August 2001. Leverage increased from 34%
to 66%. The equity volatility measured with past values thus mixes obser-
vations from a low volatility regime with the ones from a high volatility
regime.
By contrast, in the iterative approach, we model changes in leverage.
Recall that we had collected the history of liabilities, which then entered
the Black-Scholes equations. We do rely on the assumption that the asset
volatility is constant across time, but this is an assumption that is also im-
plicitly included in the 2-equation approach (equity volatility is constant
if both leverage and asset volatility are constant). For data characterized
by large changes in leverage, one can therefore make a case for prefer-
ring the iterative approach.

We can also compare our estimates to the ones from a commercial imple-
mentation of the Merton model: the EDF™ measure by Moody's KMV
(MKMV, see Box 2.2 for a brief description). One element that MKMV adds
to the simple Merton approach is calibration of the model outcome to de-
fault data. For various reasons (e.g., non-normal asset returns) Merton de-
fault probabilities can underestimate the actual default probabilities.
MKMV thus uses a historical default database to map model default prob-
abilities into estimates of actual default probabilities. Partly for this rea-
son, EDFs are usually larger than the default probabilities we get from
implementations like the ones we followed here.

Box 2.2 The EDF™ measure by Moody's KMV

A commercial implementation of the Merton model is the EDFTM mea-

sure by Moody's KMV (MKMV). Important modeling aspects are the
following:

MKMV uses a modified Black-Scholes valuation model that allows for

different types of liabilities.
In the model, default is triggered if the asset value falls below the sum
of short-term debt plus a fraction of long-term debt. This rule is de-
rived from an analysis of historical defaults.
The distance-to-default that comes out of the model is transformed
into default probabilities by calibrating it to historical default rates.

For August 2001, the EDF for Enron was roughly 2%,4 which is larger than
the result we get from the 2-equation approach, but smaller than the re-
sult from the iterative approach. The latter is somewhat atypical. One rea-
son could be that the critical value that triggers default in the EDF model
is not total liabilities, but short-term debt plus a fraction of long-term lia-
bilities. Although this adjustment increases the quality of EDFs on aver-
age, it may have led to an underestimation for Enron. Due to off-balance
sheet transactions, financial statements understated the Enron's liabili-
ties. By using the total liabilities, we may have unwittingly corrected this
bias.

To sum up, the case that we have examined may be somewhat atypical in
the sense that a simple implementation of the Merton model yields rela-
tively high default probabilities, which also seem to be close to the true
default probability. In many cases and especially for short maturities, a
simple Merton approach will produce default probabilities that are very
low, such as 0.0000001%, even though we have good reason to believe
that they should be much higher. We would then be hesitant to use the
model result as an estimate of the actual default probability. Empirical
studies, however, show that the results can nonetheless be very useful for
ordering firms according to their default risk (see Vassalou and Xing,
2004).

IMPLEMENTING THE MERTON MODEL WITH A T-

YEAR HORIZON

So far, we have implemented the Merton model by setting debt maturity

to one year – an arbitrary, but convenient assumption. Typically, the aver-
age maturity of a firm's debt is larger than one year. So can we hope to
get better results by aligning the maturity in the model with the actual
debt maturities? The answer is not immediately obvious. If the only thing
that we change is the horizon (e.g., change cell B6 of Table 2.5 from 1 to
5), we would have failed to model the fact that the firm makes payments
before maturity – like regular interest on bonds and loans, or dividends.
It may be safe to ignore such interim payments over a horizon of one
year. A one-year bond with annual coupon payments is in fact a zero-
coupon bond, and firms usually do not pay out large dividends shortly be-
fore default. However, for a horizon of several years interim payments
should enter our valuation formula in a consistent way.

In the following, we implement such an approach. It maintains the set-up

of the Merton model in the sense that there is only one date at which lia-
bilities are due, however, we take interim payments into account. The key
steps are as follows:
 1. Assume that the firm has issued only one coupon bond with maturity
equal to the average maturity of liabilities.
2. Accrue interest and dividend payments to the maturity assumed in
step 1; i.e., hypothetically shift their payment dates into the future.
3. Since accrued dividends and interest are assumed to be due at matu-
rity, even though they are actually paid before, treat them as liabilities
that have higher priority than the principal of the bond.

We start with step 1. In the balance sheet of a firm, liabilities are split up
into current liabilities (maturity less than one year) and long-term liabili-
ties (maturity larger than one year). Additional information on maturity
can be obtained from the annual report, an examination of outstanding
bonds or other sources. Usually, however, this information is not suffi-
cient for precisely determining the average maturity. Here, we follow a
simple rule that requires little information: assuming that current (long-
term) liabilities have an average maturity of 0.5 (10) years, the average
maturity obtains as (L is total liabilities, CL is current liabilities)

[0.5 × CL + 10 × (L – CL)]/L

With the balance sheet from Enron's quarterly report for June 2001, this
leads to a maturity of T – t = 5.53 years.

Having fixed T, we can proceed to step 2 and compute the value of ac-
crued dividends and interest payments at T. We assume that dividends
are paid annually, and that they grow at an annual rate of g. With the div-
idend just paid (D0), the end value of the dividend stream, which we de-
note by D, then obtains as

From Enron's annual report, the dividend for 2000 was D0 = 368m, up
3.66% on the 1999 dividend. This motivates our assumption of g = 3%.
Note that we accrue dividends at the risk-free rate r, which we take to be
the yield of five-year treasuries. Using the risk-free rate seems ad hoc, be-
cause dividends are risky, but it has some justification because dividends
will be treated senior to debt, and so it is probably a better choice than
the yield on Enron's debt.
Interest payments are treated in a similar fashion. Assuming that they are
due annually, and that the coupon rate is c, the end value of interest pay-
ments (denoted by I) is

We could infer the coupon rate by examining the coupons on Enron

bonds outstanding at t. Here, we just assume a value c = 4%.

Now we can move to step 3. As in the first two sections, the analysis will
rest on the option pricing formula that returns equity value as a function
of the asset value, liabilities and asset volatility. To understand how eq-
uity should be valued, we examine the pay-off to equity holders at matu-
rity T. Assuming that accrued dividends D have priority over the princi-
pal L, and that accrued interest I and accrued dividends have equal prior-
ity, we can distinguish three regimes (AT is the asset value at maturity):

AT < D + I : Firm is in default, and asset value is not sufficient to cover

claims from dividends and interest. The equity holders receive their
share D/(D + I)AT.
L + D + I > AT > D + I : Asset value suffices to cover claims from divi-
dends and interest, but the firm is in default because the principal L is
not fully covered. Equity holders receive only accrued dividends D.
AT > L + D + I : Asset value suffices to cover all claims. Equity holders
receive AT – L – I. Note that this includes the dividend claims D.
Explicitly stated equity holders receive D + (AT – L – I – D) = AT – L – I.

Table 2.6 Pay-off structure if accrued dividends and interest have priority
over other liabilities L
The pay-off structure is shown in Table 2.6 for example values for D, I
and L.

Carefully inspecting the pay-off structure, it is an exercise in financial en-

gineering to replicate the pay-off to equity with a portfolio of call options
and direct investments in the underlying assets. Specifically, equity is
equivalent to

a share of D/(D + I) in the assets, plus

a share of D/(D + I) in a short call on assets with strike D + I, plus a call on

assets with strike L + D + I

= equity value.

Here is a graphical depiction:

We can then use the standard Black-Scholes option pricing formula to

model today's value of equity. We obtain
with

As before, we can derive a second equation relating equity volatility to as-

set volatility

and determine the unknowns At and σ by solving (2.16) and (2.19).

In Table 2.7, the approach is applied to Enron. Dividends and interest are
accrued in E1:G12 using Equations (2.14) and (2.15), respectively.

The starting value for the asset value is equity value plus book value of li-
abilities; the starting value for the asset correlation is equity correlation
times Et/At. Cells B19:B24 contain the formulae (2.16) to (2.19). We then
use the Solver to minimize the squared percentage errors between the
observed values (for equity value and volatility) and their model
counterparts.

We also determine the default probability (cell B29). Assuming the drift
rate to be 4.5% as in the previous section, we get a default probability of
31.37%. Note that this is a default probability over a horizon of 5.5 years.
Within our framework, it is not obvious how to convert it to an annual
default probability, because the model does not allow interim defaults. To
get some indication, we can derive an annual default probability under
the assumption that default probabilities are constant across time. This
leads to

Prob(default p.a.) = 1 – (1 – 0.3137)1/5.5 = 6.58%

When comparing this figure to the previous results, note that there are
several effects at work. The assumptions about dividends and interest are
not the only difference between the multi-year approach and the one-
year approach. The sensitivity of the default probability to a given asset
drift and a given asset volatility also changes with the horizon. This is evi-
dent from the results. The asset volatility in Table 2.7 is closer to the one
we received from the one-year, 2-equation approach. The default proba-
bility, on the other hand, is closer to the one from the iterative approach.

Table 2.7 Calibrating the multi-period model to equity value and equity
volatility

CREDIT SPREADS

With the estimates from Table 2.7, we can also determine the yield on
Enron's liabilities. In our model, the firm has just one bond that pays L + I
at maturity (if the firm is not in default). The current value of the bond Bt
is the payment in T discounted at the yield y. We also know that it is equal
to the asset value minus the equity value. Therefore, we have
Solving for the yield y we get

Here, we have inserted the results from Table 2.7. The spread s (i.e., the
difference between the corporate bond yield and the risk-free rate) is

In accordance with the relatively large default probability, we get a

spread that is typical of relatively risky debt, which appears to be
sensible.

In empirical studies, however, spreads produced by Merton models are

often found to be lower than observed market spreads. One reason could
be that the Merton model tends to underestimate default risk (see the dis-
cussion in the previous section). In addition, market spreads compensate
investors for the illiquidity and tax disadvantages of corporate bonds, i.e.,
factors that are separate from default risk.

CREDITGRADES

In this section we implement a widely used alternative structural model-

ing approach called CreditGrades (CG, hereafter). In this model default oc-
curs if the asset value of the firm drops below a random barrier. The ran-
domness of the barrier is the main difference to the models presented so
far. It captures the fact that we usually do not know the level of a firm's li-
abilities until the firm defaults. (Balance sheet information is only avail-
able at a quarterly frequency, and it could be distorted or manipulated.)
In addition, the CG model assumes that default can occur at any time,
whereas default in the classical Merton model occurs only at maturity.
The CG approach is illustrated in Figure 2.3, which can be compared to
Figure 2.1.
Figure 2.3 Illustration of the CreditGrades model

We follow the CG documentation and express the relevant firm variables

on a per-share basis. We denote today's asset value per-share by A. As
above, asset values are assumed to follow a lognormal distribution. The
random default barrier is given as

default barrier = Λ D

where D is today's debt-per-share. Λ represents the uncertainty in recov-

ery values and is a random variable, which is assumed to follow a lognor-
mal distribution with mean and standard deviation λ, independent of
the asset value process.5 As in the classical Merton model, we need to esti-
mate today's asset value, its volatility and its drift rate. In CG, they are de-
termined as follows.

The drift rate is set to zero. This is justified if firms tend to maintain a
constant leverage over time. Assets may rise at some rate but if debt rises
at the same rate, the distance to the default barrier remains constant and
we can describe the situation by assuming a zero drift for both assets and
debt.

For the initial asset value, the CG model suggests the approximation A = S
+ D where S is the firm's stock price. This choice is based on an inspec-
tion of boundary cases; details can be found in Finger et al. (2002, p. 10).
The asset volatility is set to
where σE is equity volatility. This expression is an approximation of the
theoretical relationship between asset volatility and equity volatility that
we also used in previous sections (see Equations (2.12 and 2.13)).

Moving on, one can derive an approximation for the probability that the
firm survives until time t, as seen from today (t = 0):

with

As shown by Kiesel and Veraart (2008), it is also possible to derive an ex-

act solution which would be straightforward to implement with the tools
presented in this book.6

Next, we describe how the parameters of Equation (2.23) are specified in

the CG model. The number of shares used for computing per-share fig-
ures is the number of common shares plus the number of preferred
shares. The number of common shares can be determined by dividing the
company's market capitalization by the current stock price; the number
of preferred shares is calculated by dividing the book value of preferred
shares by the price of common stock from the reporting date of the book
value. Furthermore, the number of preferred shares is capped at half the
number of common shares.

The debt-per-share measure D is calculated by dividing liabilities by the

number of shares. For industrial firms this is done by calculating the fi-
nancial debt as the sum of short-term borrowings, long-term borrowings
and one half times the sum of other short-term liabilities and other long-
term liabilities. The latter two are weighted by ½ to correct for their in-
clusion of non-financial liabilities such as deferred taxes or provisions.
‘Accounts payable’ is not included in the financial debt calculation.
Liabilities of subsidiaries are fully included in a consolidated balance
sheet even though the parent may own less than 100% of the subsidiary.
To adjust for this effect, the final debt measure is found by reducing the
financial debt by min{0.5 × financial debt, minority interest}.

For financial firms and companies with large financial subsidiaries the
debt calculation is adjusted by either excluding the financial subsidiary
or changing the whole financial debt calculation procedure. For banks
some authors suggest eliminating the short-term borrowings. Another
way to obtain an estimate for the debt-per-share ratio is to compare the
leverage ratios of the company with its peers.

The recovery rate and its volatility λ can be based on recovery statistics
published by rating agencies, e.g., see Chapter 7, Table 7.6. CG proposes
= 0.5 and λ = 0.3. Equity volatility can be estimated as described above
(see Table 2.4). Alternatively, option implied volatilities can be used.

Several modifications are possible. When leverage data is difficult to ob-

tain or if market conditions change quickly, a promising extension can be
to use additional information in order to back out an implied leverage.7
Recall that the survival probability of Equation (2.23) depends on the
debt-per-share. The basic idea is to back out an implied debt-per-share
from market observed credit default swap (CDS) spreads and insert it into
Equation (2.23) in place of balance sheet information.8

We move on to implement the CG model in Excel. All calculations can be

done in the spreadsheet, with VBA merely adding convenience. We start
with a simple example and later extend the methodology to a case study
of Lehman.

In Table 2.8 we estimate the CreditGrades model for John Wiley & Sons as
of December 09. The parameters for the global recovery Λ and the volatil-
ity of the default barrier are set to 50% (cell B4) and 30% (cell B5) re-
spectively. The time horizon can be specified in cell B6.

From the SEC filings we obtain the balance sheet information needed to
calculate the financial debt: the short-term borrowings (cell B10), long-
term borrowings (B11) and the other short-term (B12) and long-term
(B13) liabilities.
Both the preferred equity and minority interest (cells B14/B15) are zero
for Wiley. The financial debt is then given in cell B17 as the sum of the
former, with a weighting of ½ on other liabilities:

=B10+B11+0.5*(B12+B13).

Table 2.8 Calibrating the CreditGrades model

The final debt variable is found by subtracting the minority interest

capped at ½ of the financial debt. Cell B18 therefore reads:

=B17-MIN(0.5*B17,B15).

Before estimating the debt-per-share variable D, we need to estimate the

number of shares. We divide the market capitalization (cell E10) by the
stock price (E11) in cell E12. Since there are no preferred shares outstand-
ing the total number of shares in E14 is the sum of common shares and
zero. Debt-per-share in cell B19 is simply the debt divided by the number
of shares: =B18/E14.
We specify the volatility measure in cell E18 as the option implied volatil-
ity of 33%.9

In cell B24 we calculate the asset value according to the CG approxima-

tion introduced above. When calculating the parameter d according to
Equation (2.23) we can refer to cell B24. Thus d is defined as

=(B24*EXP(B5^2))/(B4*B19)

The final auxiliary parameter α is determined in cell B26 according to

Equation (2.23):

=(((E18*E11)/(B24))^2*B6 + B5^2)^0.5

Now we estimate the probability of survival in cell B28 according to

Equation (2.23). The survival probability over the next year is 99.997%.
This implies a one-year default probability of 0.003%.

In the following case study we will show the calculation of a time series of
CreditGrade survival probabilities for one company, Lehman Brothers. As
mentioned before, the adjustments necessary when dealing with finan-
cial companies are not standard. The estimation of the debt-per-share
variable should be adjusted in order not to overweight short-term bor-
rowings. Furthermore, the volatility of the barrier is usually set to a lower
level, e.g., 10% as proposed by Veraart (2004). Here we show that these
simple adjustments are sufficient to obtain a timely credit assessment.10
When dealing with time series it is convenient to comprise the calculation
into one function; here CG_PS(S, sigma_E, D, Lambda, sigma_B, t) will di-
rectly estimate the survival probability according to Equation (2.23):11

Function CG_PS(S, sigma_E, D, Lambda, sigma_B, t)

‘S= stock price, sigma_E=equity vola, D=Debt-per-share
‘Lambda=global recovery, s_B=vola of barrier, t=time
Dim dl, alpha
d1 = (S + Lambda * D) * Exp(sigma_B ^ 2) / (Lambda * D)
alpha = (((sigma_E * S) / (S + Lambda * D)) ^ 2 * t + sigma_B ^ 2) ^ 0.5
CG_PS = Application.NormSDist(-(alpha / 2) + (Log(d1) / alpha)) - _
d1 * Application.NormSDist(-(alpha / 2) - (Log(d1) / alpha))
End Function
Now let us look into the balance sheet of Lehman. We calculate the debt-
per-share ratios from each quarterly report between Q4 2003 and Q2
2008 (Lehman defaulted in September 2008) and report the numbers in
Table 2.9.12

Using end-of-quarter stock prices and historical volatilities over 90 days,

we estimate the one-year probabilities of default for Lehman setting λ =
10% and = 50% (see Figure 2.4).

S&P rated Lehman Brothers at A+ from October 05 to June 08 when it

downgraded the company to an A flat rating. According to S&P's yearly
default studies, these rating categories translate into a yearly historical
default rate of 0.05–0.07%; the CreditGrades implied PD is lower during
Q4 2003–Q4 2006,13 and in Q1 2007 the model matches the 0.05%. In 2007,
the decline of the stock price paired with the increase of the volatility re-
sults in sharper PD estimates implying S&P ratings between BBB and
BB–.14 In 2008, CreditGrades implies PDs of 17–27% per year, indicating a
CCC/C level and thus pre-default. However, in the second quarter of 2008
one might have believed in a model error, rather than an early warning.
Let us look at the CDS levels at that time together with the CreditGrades
PDs in Figure 2.5.

Table 2.9 Case study on Lehman Brothers

Figure 2.4 Share price and historical volatility for Lehman Brothers

Figure 2.5 Comparing PD from the CreditGrades model with different

specifications and the CDS spread for Lehman Brothers

At the end of the second quarter 2008, when Lehman published its last
quarterly report, the Lehman CDS trades around 230 basis points.
Assuming that the CDS spread correctly prices the default risk, we obtain
a market-implied debt-per-share ratio of $62.15 Using this implied debt-
per-share ratio lowers the PD implied by CreditGrades compared to the
use of balance sheet information, but it is still large enough to create seri-
ous concerns about the solvency of Lehman. The subsequent increase of
both PD measures is driven by the increasing stock volatility and the de-
crease of the stock price, predictors that are particularly useful in such
crisis scenarios. Concluding, we find that CreditGrades is able to provide
quite timely credit signals, even in difficult cases.

APPENDIX
Here we collect some equations of the CreditGrades model mentioned in
the main text. The formulae can be used to estimate the implied asset
volatility and implied barrier. Furthermore we give the CDS spread for-
mula for estimating the implied debt-per-share from observed CDS
spreads. To start, the price of an European put option under the CG
framework is given as16

where X is the strike, T the maturity of the option, B the random default
barrier, Φ(x, y) the integral over the normal distribution density from x to
y and Φ(x) the corresponding integral over [–∞, x]. I is given as

where

which depend on the distance to default measures

With two at-the-money put prices P1,2 we can use Equation (2.A1) to solve
for the implied asset volatility and the implied barrier. However here it is
easier to solve
Converting the survival probability into a credit spread we specify R as
the recovery rate of the underlying credit17 and obtain18

where the function G is given as

In VBA we define the corresponding functions as

Function CG_G(x, d1, r, sigma)

Dim z
z = (0.25 + (2 * r / sigma ^ 2)) ^ (0.5)
CG_G = d1 ^ (z + 0.5) * Application.NormSDist(-Log(d1) / _
(sigma * x ^ 0.5) - z * sigma * x ^ 0.5) + d1 ^ (-z + 0.5) _
* Application.NormSDist(-Log(d1) / (sigma * x ^ 0.5) + _
z * sigma * x ^ 0.5)
End Function

and the main function

Function CG_CDS(S, sigma_E, d, Lambda, sigma_B, t, Rec, r)

Dim dl, sigma, xi, H
d1 = (S + Lambda * d) * Exp(sigma_B ^ 2) / (Lambda * d)
sigma = (sigma_E * S) / (S + Lambda * d)
xi = sigma_B ^ 2 / sigma ^ 2
H = Exp(r * xi) * (CG_G(t + xi, d1, r, sigma) - CG_G(xi, d1, r, sigma))

CG_CDS = (r * (1 - Rec) * (1 - CG_PS(S, sigma_S, d, Lambda, sigma_B,

0#) _ + H)) / CG_PS(S, sigma_E, d, Lambda, sigma_B, 0#) - _
CG_PS(S, sigma_E, d, Lambda, sigma_B, t) * Exp(-r * t) - H)
End Function

Using observed CDS spreads we can now use Solver in combination with
the CG_CDS function, e.g., to find the implied debt-per-share ratio or any
other parameter of interest.
 NOTES AND LITERATURE

Assumptions

The Merton model, like any model, simplifies the reality to make things
tractable. Important assumptions in Merton (1974) are: no transactions
cost; no bankruptcy cost; no taxes; unrestricted borrowing and lending at
the risk-free interest rate; no short selling restrictions; no uncertainty
about liabilities; lognormally distributed assets. Many extensions to
Merton (1974) have been proposed and tested, and the design and practi-
cal application of structural models is still high on the agenda in credit
risk research.

Literature

The seminal paper is Merton, R.C., 1974, On the pricing of corporate debt.
The risk structure of interest rates, Journal of Finance 29, 449–470.

The iterative method is used, for example, in Vassalou, M. and Xing, Y.,
2004, Default risk in equity returns, Journal of Finance 59, 831–868. Our
multi-year analysis follows Delianedis, G. and Geske, R., 2001, The compo-
nents of corporate credit spreads. Default, recovery, tax, jumps, liquidity,
and market factors, Working paper, UCLA. The approach behind Moody's
KMV EDFs is described in Kealhofer, S., 2003, Quantifying credit risk I:
Default prediction, Financial Analysts Journal 59(1), 30–44.

The CreditGrades model is extensively described in Finger, C., et al., 2002,

CreditGrades. Technical document, Riskmetrics.com (accessed August 9,
2010). The extensions with respect to implied volatility and implied lever-
age are based on Hull, J., Nelken, I. and White, A., 2005, Merton's model,
credit risk and volatility skews, Journal of Credit Risk 1, 3–28 and
Stamicar, R. and Finger, C., 2005, Incorporating equity derivatives into the
CreditGrades model, Riskmetrics.com (accessed August 9, 2010). An exact
solution to the CreditGrades survival probability is derived by Kiesel, R.
and Veraart, L., 2008, A note on the survival probability in CreditGrades,
Journal of Credit Risk 4(2), 65–74. The adjustments for financials are dis-
cussed by Veraart, L., 2004, Asset-based estimates for default probabilities
for commercial banks, mimeo, University of Ulm.
1 A variable X whose logarithm is normal with mean E(ln X) and variance
σ2 has expectation E[X] = exp(E(ln X) + σ2/2). Denoting the expected
change of ln X by E(ln X) = μ – σ2/2 rather than by μ has the effect that the
expected change of X is E[X] = exp(μ) and thus depends only on the cho-
sen drift parameter, and not on the variance σ2.

2 Data can, for example, be obtained from www.econstats.com.

3
Prices should be adjusted for stock splits etc.

4 See EDF Case Study: Enron,

http://www.moodyskmv.com/research/files/Enron.pdf (accessed August 18,
2010).

5 In order to comply with our previous notation we changed some sym-

bols from the CreditGrades technical document. In the original CG model
our Λ is called L and they use At to denote our αt.

6
The exact solution involves the bivariate normal distribution, which can
be evaluated with a function used in Chapter 6.

7 See Hull, Nelken and White (2005) for a discussion of implied leverage
in the Merton model.

8
See Chapter 10 of this book for an extensive discussion of CDS and im-
plied probabilities. See Equation (2.A2) in the Appendix of this chapter
for a closed form solution of a CDS spread under the CG model. Another
approach to market implied leverage is the use of two at-the-money op-
tion prices to estimate both the implied asset volatility and implied lever-
age using the option pricing formula (2.A1).
9 Cells E16–E17 present both the 30 days historical volatility and the op-
tion implied volatility. In cell E18 we enter our assumed volatility.

10 In the Appendix to this chapter, we further demonstrate how the debt-

per-share can be implied by CDS data.

11 In the VBA code we cannot distinguish between the capital lambda Λ

and the small λ, thus we denote the latter by sigma_B.

12 Note that Lehman undertook two stock splits (2:1), one in September
2000, the other in June 2006. The share prices we use adjust for these
splits.

13
Increasing the global recovery parameter, e.g., to 60%, as examined for
financials by Veraart (2004), improves the match to the rating implied PDs
during that period.

14
In early 2007 the historical volatility climbed from around 30% to 50%.
With option implied volatility this pattern is even more pronounced. The
stock price started dipping two quarters later.

15 See the Appendix to this chapter.

16
See Finger et al. (2002).

17 Note the difference between R and Λ. The former refers to the expected
recovery of a specific part of the firm's liabilities, e.g., a bond, whereas
the expected value of the random variable Λ gives the expected average
recovery over all assets of the firm.

18 For more details on CDS spreads and the procedure underlying the de-
viation of this equation, see Chapter 10.