Sunteți pe pagina 1din 32

Time Varying Expected Returns, Stochastic

Dividend Yields, and Default Probabilities:


Linking the Credit Risk and Equity Literatures
1
George Chacko Peter Hecht Jens Hilscher
Kite Partners Allstate Investments Brandeis University
First draft: September 2002
This version: September 2007
Abstract
Motivated by the equity literature, which has argued that returns are time
varying and that the dividend yield predicts future returns, this paper intro-
duces time varying expected returns into a structural model of rm value.
Our model produces a stochastic dividend yield in equilibrium. It implies
signicant dierences in the reaction of the default probability to dierent
sources of news; a reduction in asset value due to discount rate news has
a smaller impact on the default probability than a reduction due to cash
ow news. In general, model implied default probabilities are less volatile
than those generated from a standard Merton (1974) model. The model is
consistent with the standard credit rating agencies view that focuses on fun-
damental analysis, in addition to market information, when assessing credit
risk. It is particularly relevant in periods of overvaluation and low expected
returns such as the recent technology bubble.
JEL classications: G13, G33, G35
Keywords: Stochastic dividend yield, Default probabilities, Merton model,
Bond pricing
1
Corresponding author: Jens Hilscher, Brandeis University IBS, 215 South Street, Waltham MA,
02453, USA, Tel. +1 781 736 2261, email: hilscher@brandeis.edu. An earlier version of this paper was
circulated under the title Corporate Bond Pricing and Dierent Sources of Asset Return Volatility
as HBS working paper 03-046. We would like to thank Malcolm Baker, John Campbell, Robert C.
Merton, Ken Froot, Jeremy Stein, Peter Tufano, Luis Viceira, and Josh White and seminar participants
at Moodys KMV for helpful discussions.
1 Introduction
What is the best way to measure changes in a particular rms default probability?
Traditionally, credit rating agencies, such as Moodys and S&P, perform fundamental
analysis to determine creditworthiness. This usually entails analyzing a rms account-
ing nancials, calculating nancial ratios, and understanding its business model. Since
fundamental analysis relies heavily on accounting information, it is susceptible to pro-
ducing a stale and potentially backward looking measure of credit risk.
1
A second approach is to assess credit risk by interpreting market and accounting
information in a manner consistent with the structural form Merton (1974) bond pricing
model. In the model, expected returns and the growth rate of rm value are constant.
At maturity, the rm defaults if rm value lies below the face value of debt. This
approach yields a default probability that moves with market based information at
high frequency,
2
while fundamental analysis provides no incremental information about
default.
In this paper we introduce time varying expected returns and dividend yields into a
structural rm value model and consider the implications for default probabilities. Re-
laxing the assumption of constant expected returns, which is common to most corporate
bond pricing models, is motivated by the ndings in the equity literature which argues
that expected returns and dividend yields are time varying and persistent and that
the dividend yield predicts future returns.
3
In this setting rm value and the default
probability move in response to two types of news discount rate and cash ow news.
Distinguishing between these price changes is important because they may have very
dierent impacts on the default probability: If expected returns are time varying and
mean-reverting, a reduction in asset value due to discount rate news will have a smaller
impact on the default probability than a reduction due to cash ow news. Default
probabilities in our model are less volatile, overcoming a major weakness of standard
Merton-type models emphasized by the credit rating agencies. The model is consistent
with the standard credit rating agencies view that focuses on fundamental analysis, in
addition to market information, when assessing credit risk.
We begin our analysis by pointing out the restrictions a constant dividend yield places
on the possible causes of price movements. Either expected returns and dividend growth
rates must be constant, or movements in expected returns must be exactly oset by
dividend growth rate movements, keeping the dividend yield constant. Expected return
news can never independently cause price movements. Therefore, only fundamentals
induced price changes move the probability of default in models with a constant dividend
1
Similarly, Altmans Z-score (1968), Ohlsons O-score (1980), and Zmijewski (1985) use accounting
based nancial ratios to estimate default probabilites.
2
See, for example, Hillegeist, Cram, Keating, and Lunstedt (2003), Vassalou and Xing (2004), Due,
Saita, and Wang (2007), Bharath and Shumway (2006).
3
See, for example, Shiller (1981), Campbell (1991), Campbell and Shiller (1988a, 1988b), Cochrane
(1991, 1992, 1994), Vuolteenaho (2002) and Hecht and Vuolteenaho (2004).
1
yield.
To get a sense of whether or not a stochastic dividend yield is an important feature
of rm asset value processes we consider the empirical relationships found in S&P 500
rms. For these rms the dividend yield is both large and stochastic. We conclude
that relaxing the assumption of constant expected returns in a rm value process may
have a large impact on predicted default probabilities.
We next introduce our model of rm value. We specify a discrete time model with
time varying expected returns and a stochastic dividend yield. Movements in asset value
are due to cash ow or expected return news, where cash ow news is commonly thought
of as being permanent news to fundamentals and expected return news is temporary.
The rm defaults at maturity if rm value lies below the face value of debt.
The stochastic dividend yield in our model is critical to our results and is the main
dierence between our model and the standard approach of structural bond pricing
models. Price changes contain two sources of information: information about future
dividends or cash ows (fundamentals) and/or information about future discount rates
or expected returns (non-fundamentals).
4
We identify fundamental price moves by
movements in price that are driven by shocks to dividends, keeping the dividend yield
constant. On the other hand, we identify non-fundamental price moves by movements
in price that are unaccompanied by movements in dividends, changing the dividend yield
and highlighting the fact that future expected returns (discount rates) have changed.
In order to investigate the dierence between the impact on default probabilities of
the two types of price movements, we calculate default probabilities for various types
of price movements under various sets of parameters. Our key nding is the following:
price movements that occur that do not move the dividend yield, i.e. price movements
due to fundamentals, move default probabilities much more than price movements
that move without a change in the level of the dividend, i.e. a non-fundamental or
discount rate movement.
The intuition is simple, but powerful. Price drops that occur due to increases
in discount rates, i.e. increases of the dividend yield, are partially oset by higher
expected price growth, mitigating the impact that the price movement has on the default
probability. The expected price growth increases because the discount rate is mean-
reverting. The higher discount rate today is expected to decrease over time, producing
higher expected price growth. When the change in discount rates is less persistent, the
impact on the expected price growth is magnied since the discount rate is expected to
move back to its long run mean quickly, dampening the link between price changes and
changes in the default probability. In the extreme case where the discount rate process
exhibits zero persistence, price movements that occur with no movement in the level of
the dividend contain zero information about default.
4
Throughout the paper, dividends, cash ows, and fundamentals will be used interchangeably
as will expected returns, discount rates, and non-fundamentals.
2
Since price movements are no longer a sucient statistic for changes in default proba-
bilities when the model allows for time varying expected returns via a stochastic dividend
yield, our results suggest a role for both fundamental- and market-based measures of
credit risk. The role of fundamental-based measures of credit risk will be most impor-
tant for rms with discount rates (dividend yields) that display low persistence. It will
also be particularly relevant in periods of overvaluation and low expected returns such
as the recent technology bubble.
Our model also has implications for risk neutral default probabilities, i.e. bond
prices. Unlike the constant dividend yield, continuous time pricing models of Black and
Scholes (1973) and Merton (1974), we calculate risk neutral default probabilities and
bond prices in a discrete time model.
5
In standard structural bond pricing models, the
rm defaults if the market value of assets lies below the face value of debt at maturity.
In the Merton model the rm is costlessly liquidated in the case of default. Risky debt
is then equivalent to the combination of a safe bond and a short put option on the rms
assets with a strike price equal to the face value of debt. We calculate bond prices using
our models rm value process while maintaining the assumption of default at maturity
and costless liquidation.
Similar to the true default probability results, we nd that movements in the un-
derlyings asset value can contain a varying degree of information about changes in risk
neutral default probabilities, depending on how much of the underlyings price move-
ment was due to news about cash ows (vs. the discount rate). However, in contrast
to the true default probability results, our results suggest underlying asset price changes
due to expected return news play a more important role when calculating risk neutral
default probabilities since expected return news moves the dividend yield and, thus, the
asset value risk neutral drift. News about cash ows only aects the price level without
changing the dividend yield.
6
Our model brings insights from the equity literature and applies them to credit risk
and risky debt pricing. If expected dividend growth rates are constant, time varying
expected returns must imply a time varying dividend yield. The related equity literature
is large and rich. It argues that expected returns and dividend yields are time varying
and persistent, that the dividend yield predicts future returns, that price movements are
the result of both cash ow (dividend) and expected return news, and that stock price
volatility is dierent across dierent horizons.
7
Another important insight is that the
empirical ndings in the equity literature are internally consistent with standard present
value models, which produce the mechanical relationship between the current dividend
5
Our approach is similar to Rubinstein (1976). Our work is also related to Brennan (1979) and
Stapleton and Subrahmanyam (1984), who use risk neutral valuation when pricing contingent claims in
discrete time.
6
This is true under the assumption that expected dividend growth rates are constant, a maintained
assumption thorughout the paper.
7
See, for example, Shiller (1981), Campbell (1991), Campbell and Shiller (1988a, 1988b), Cochrane
(1991, 1992, 1994), Vuolteenaho (2002) and Hecht and Vuolteenaho (2004).
3
yield and future expected returns and dividend growth rates.
8
There is also a large related credit risk literature. Our work is related to many
studies that have considered the determinants of default probabilities.
9
The rm value
process we specify is related to the large and rich literature on structural bond pricing
models.
10
One thing that all of the these models have in common is the assumption of
constant expected returns and/or constant dividend yield.
11
Our approach to modeling the rm value process is also related to Bekaert and
Grenadier (2001) who price stocks by specifying processes for state variables and the
stochastic discount factor in an a discrete time, ane economy. Our approach is
more reduced form since we specify a price process. However, similar to Bekaert and
Grenadier, we make use of the no arbitrage condition and nd that the dividend yield
will be stochastic in equilibrium if expected returns are time varying.
The rest of paper is divided into ve sections. Since the stochastic dividend yield is
critical to our results, Section 2 explains why time-varying expected returns forces us to
abandon the standard constant dividend yield assumption from a theoretical perspec-
tive. Section 3 motivates our models main assumption from an empirical perspective
by documenting the dividend yield time series behavior for the rms within the S&P
500. Section 4 describes the simple model used to generate the underlying asset price
movements and, thus, estimate true and risk neutral default probabilities. Section 5
discusses the analytical and numerical results, and Section 6 concludes.
2 Implications of a constant dividend yield for ex-
pected returns
Is a Constant Dividend Yield Consistent With Time-Varying Expected Returns and
Constant Expected Dividend Growth?
Before we move to the actual model that will generate prices and default proba-
bilities, it is important to understand why we move away from the typical constant
dividend yield assumption. On the one hand, we would like our economy to exhibit
time-varying expected returns (discount rates) and constant expected dividend growth,
properties that are both reasonable on economic grounds and consistent with many
empirical ndings. On the other hand, we would like our model to be as similar as
8
See for example the loglinear asset pricing framework of Campbell and Shiller 1988a.
9
See Shumway (2001), Chava and Jarrow (2004), Campbell, Hilscher, and Szilagyi.
10
Some examples include Black and Cox (1976), Longsta and Schwartz (1995), Collin-Dufresne and
Goldstein (2001), and Leland (2002).
11
Huang and Huang (2003) introduce a model with a counter-cyclical asset risk premium. However,
they do this in a reduced form manner and do not allow for a stochastic dividend yield. As shown later
in the paper, it is virtually impossible to have a time varying asset premium and constant dividend
yield.
4
possible to the standard Merton model. From a pure statistical standpoint, there is no
problem with simultaneously having time-varying expected returns, constant expected
dividend growth, and a constant dividend yield. However, from an economic stand-
point, is a model with time-varying expected returns and constant expected dividend
growth economically consistent with a constant dividend yield? What types of economic
constraints does a constant dividend yield assumption place on a model? We briey
investigate these questions in the following section.
Using only the denition of return, we rst derive necessary conditions for a constant
dividend yield model. Rewriting the denition of return,
1
t+1
=
1
t+1
+1
t+1
1t
1
t+1
=
1
t+1
1t
_
1
t+1
1
t+1
1t
1t
+
1t
1t
_
If the dividend yield is constant, this reduces to
1
t+1
=
1
t+1
1t
_
1 +
1
1
_
log (1
t+1
) = log
_
1
t+1
1t
_
+ log
_
1 +
1
1
_
(1)
Thus, when the dividend yield is constant, realized returns and realized dividend
growth are perfectly positively correlated. The same is true for log returns and log
dividend growth. For logs, the volatilities are identical, and revisions in future expected
dividend growth rates and revisions in future expected returns are identical too. So, any
news about future expected dividend growth rates are exactly oset by future expected
returns, by construction. Thus, 100% of return variation is explained by the contempo-
raneous dividend growth. It is impossible to even entertain the possibility that a shock
to expected returns moved prices since this shock is always exactly oset by a shock
to future expected dividend growth. News about the future is irrelevant. Only the
current shock to dividends can move prices. All models with a constant dividend yield
assumption place these extreme restrictions on the underlying economic processes.
If we add in the assumption that expected log dividend growth is constant, then it
is impossible to have an economy with time-varying expected returns and a constant
dividend yield. Even without the constant expected dividend growth assumption, as
shown above, it is impossible to identify price movements that are due to movements
in expected returns (discount rates). In eect, all movements in prices must be "fun-
damental" when the constant dividend yield assumption is imposed on an economic
model.
In sum, the dividend yield reects information about future expected dividend growth
and future expected returns
12
. If expected dividend growth is constant, movements in
12
An alternative framework linking the mechanical relationship between dividend yields, expected
returns and dividend growth rates is provided by Campbell and Shiller (1988a). They show, to a rst
5
expected returns will cause the dividend yield to change. Thus, the stochastic dividend
yield assumption used in this model should not be viewed as a standalone exogenous
assumption. It is a result - a result that must follow from a model that wants to
be economically consistent with time-varying expected returns and constant expected
dividend growth. Since the dividend yield will pick up movements in discount rates,
it will be important to watch both price movements and movements in the dividend
yield. Movements in prices that occur without a change in the dividend yield are, in a
sense, "fundamental". In contrast, movements in prices that occur in combination with
a movement in dividend yields will partially reect the movement in expected returns,
a "non-fundamental" source of price variation. In the extreme case where prices move
without any move in the level of the dividend, the price move will be entirely "non-
fundamental". These identications provide the foundation for our numerical results.
3 Dividend yields for S&P 500 rms
In the last section, we showed, from a theoretical perspective, that a model with time
varying expected returns must have a stochastic dividend yield - the key feature of our
new model. In this section, we look at the time series properties of the asset dividend
yield for the aggregate S&P 500 and the asset dividend yield
13
for individual S&P 500
rms actually observed in the data. In addition to empirically justifying the main
feature of our model, we will use the rm level statistics as inputs to the calibration
exercises performed in Section 5. For both the aggregate S&P 500 and the typical S&P
500 rm, we nd that the asset dividend yield displays large variation through time and
is highly autocorrelated.
In order to construct the asset dividend yield series for the S&P 500 rms, we con-
struct payout and asset value series. Payout is measured as the sum of dividends on
common equity, dividends on preferred equity, and interest payments. Asset value is
equal to the sum of common equity value, preferred equity value, and debt value. We
use monthly data from CRSP and annual data from COMPUSTAT to construct these
series. We use book value of preferred equity as well as debt to proxy for their market
values. Book value of debt is equal to long term debt plus debt in current liabilities.
We measure variables at an annual frequency and cover the period from 1980-2002. The
asset dividend yield is dened as the payout over the period divided by the end of period
asset value. Because this data is likely to be more familiar, we also calculate the equity
order approximation, that the log dividend yield equals the discounted innite sum of future expected
returns minus the discounted innite sum of future expected dividend growth rates.
13
Thoughout the paper, unless otherwise stated, the dividend yield refers to the total payout of the
rm (dividends and interest) divided by the value of all securities issued by the rm (equity and debt).
When it is unclear, the terms "equity dividend yield" (dividend divided by equity value) and "asset
payout yield" are used (dividend plus interest divided by equity and debt value).
6
dividend yield series. It is equal to dividend payments paid during the period divided
by the end of period value of common equity. To deal with outliers in the rm level
series, we winsorize the payout yield at the 0.5% level of the entire distribution, i.e. we
replace values below the 0.5th percentile with the 0.5th percentile, and values above the
99.5th percentile with the 99.5th percentile.
We would like to consider data for both the index as well as individual rms. How-
ever, we cannot directly observe the aggregate payout of S&P 500 rms. In addition,
the composition of the S&P 500 changed over the 1980-2002 period. We therefore col-
lect data on rms that were in the S&P in 1990. We then construct aggregate series by
taking cross-sectional averages. We use data from CRSP and/or COMPUSTAT over
the 1980-2002 period. We include rms with scal years ending in December. This
leaves us with a little over 7,000 rm year observations. We construct the aggregate
S&P 500 data series by calculating cross sectional means of this data.
14
Figures 1 and 2 plot the aggregate equity and asset dividend yield, respectively. We
can immediately see that both the equity and asset dividend yield are highly serially
correlated and persistent. Data points at the beginning and end of the series suggest
that both series are mean reverting. Table 1, Panel A provides summary statistics.
The mean asset and equity aggregate dividend yield are large at 4.52% and 3.38%,
respectively. The unconditional asset and equity aggregate dividend yield volatility
are 1.50% and 1.25%, respectively. Table 2, Panel A reports summary statistics from
running AR(1) regressions of the log asset dividend yield. The log asset dividend yield
process is highly persistent with an AR(1) coecient of 0.97. Shocks to the dividend
yield process are volatile at 8.48%.
Figures 3, 4, and 5 graph the rm level asset dividend yield for three S&P 500 rms:
FirstEnergy, Merrill Lynch, and PACCAR. Similar to the aggregate series, the rm
level dividend yield is highly serially correlated and persistent. The entire pooled rm
level summary statistics are reported in Panel B of Tables 1 and 2. When looking at the
rm level, the mean asset and equity dividend yield across all rm years are similar in
magnitude to the aggregate results at 4.72% and 3.27%, respectively. The unconditional
asset and equity aggregate dividend yield volatility across all rm years are 3.05% and
2.57%, respectively, displaying both within rm and across rm variation. Although the
log asset dividend yield process is less persistent at the individual rm level, the median
rm is still highly persistent with an AR(1) coecient of 0.66. Relative to the aggregate
results, the shocks to the median rm level dividend yield process are much more volatile
at 23.61%. This result should not be surprising considering that the aggregate index is
a portfolio and, thus, exhibits diversication-type eects.
In sum, the rm level asset dividend yield for the typical S&P 500 rm is large and
persistent with volatile shocks. In fact, the typical rm level asset dividend yield is
14
Comparing the constructed equity dividend yield series to that calculated using S&P 500 data
directly from CRSP, no material dierences are found.
7
substantially more volatile than its aggregate counterpart. While modeling the asset
dividend yield as an AR(1) is a simplication, it provides a reasonable, parsimonious
approximation.
Clearly, the constant dividend yield assumption is inconsistent with the data in a
rst-order manner. Allowing the dividend yield to be stochastic and persistent provides
a much better t. These empirical features are the foundation of our theoretical model
and calibration exercises.
4 A simple model of rm value
This section provides the processes, distributional assumptions, and parameter restric-
tions that generate the data in our economy. It is important to keep in mind that this
model is not meant to be all encompassing. For example, the model does not incorpo-
rate stochastic interest rates and stochastic volatility. This exclusion does not intend to
say that these properties are unimportant. The model is simply a vehicle to highlight
the stochastic dividend yield (induced by time-varying expected returns) channel and
also help provide a transparent map between dierent types of price movements and
default probabilities.
4.1 Firm value process
In order to generate true default probabilities, risk neutral default probabilities, and
bond prices, we assume the log rmprice growth (j
t+1
= log
_
1
t+1
1t
_
), log dividend yield
(o
t+1
= log
_
1
t+1
1
t+1
_
), and log stochastic discount factor (log SDF, :
t+1
= log (`
t+1
)) are
conditionally multivariate normal with the following distributional parameters:
j
t+1
= j
j,t
+
j,t+1
; \ c:
t
[j
t+1
] = o
2
j
o
t+1
= j
c
+o (o
t
j
c
) +
c,t+1
; \ c:
t
[o
t+1
] = o
2
c
:
t+1
= :
)

1
2
o
2
n
+
n,t+1
; \ c:
t
[:
t+1
] = o
2
n
Co
t
[:
t+1
. j
t+1
] = o
nj,t
; Co
t
[:
t+1
. o
t+1
] = o
nc
Co
t
[o
t+1
. j
t+1
] = o
cj
(2)
The processes for the log price growth and log dividend yield imply a process for the log
dividend growth (d
t+1
= log
_
1
t+1
1t
_
) via an accounting identity.
d
t+1
= j
t+1
+o
t+1
o
t
d
t+1
= j
o
+
o,t+1
; \ c:
t
[d
t+1
] = o
2
o
(3)
All of the volatilities associated with the shocks are constant. The log price growth
has a time-varying mean while the log dividend yield follows a stationary AR(1) with
8
0 < o < 1. Out of all of the specied covariances, only the covariance between the log
SDF and log price growth is time-varying. The risk free rate and expected log dividend
growth are assumed to be constant.
Before moving further, it is important to note that this reduced form model for
the price and dividend yield process can be justied from a more general equilibrium
perspective. Given the dividend, SDF, and market price of dividend risk processes,
the price process is exactly determined by the expected innite sum of stochastically
discounted future dividends. If the log dividend growth, log SDF, and market price
of dividend risk processes are conditionally multivariate normal with the market price
of dividend risk following an AR(1) process, it can be shown that the equilibrium log
dividend yield process follows a conditionally normal AR(1) process to a rst order ap-
proximation.
15
The log dividend yield process inherits the dynamics of the market price
of dividend risk process, consistent with the discussion in Section 2 of this paper. Since
the log dividend growth and log dividend yield processes are conditionally multivariate
normal to a rst order approximation, by denition, the log price process also follows
a conditionally normal process to a rst order approximation. This more primitive
underlying model justies the reduced form model used throughout this paper.
Many of the parameters from our reduced form model are constrained by distribu-
tional assumptions, the no-arbitrage condition associated with the existence of an SDF,
and accounting identities. First, by no-arbitrage, the expected value of the rms return
multiplied by the SDF must be 1.
1 = 1
t
[`
t+1
1
t+1
]
1 = 1
t
_
`
t+1
1
t+1
1t
_
1 +
1
t+1
1
t+1
__
(4)
Rearranging and using our distributional assumptions,
1 = 1
t
[exp (:
t+1
+ j
t+1
)] +1
t
[exp (:
t+1
+ j
t+1
+o
t+1
)]
1 = exp
_
:
)
+j
j,t
+o
nj,t
+
1
2
o
2
j
_ _
1 + exp
_
j
c
+o (o
t
j
c
) +
1
2
o
2
c
+o
nc
+o
cj
__
0 = :
)
+j
j,t
+o
nj,t
+
1
2
o
2
j
+ log
_
1 + exp
_
j
c
+o (o
t
j
c
) +
1
2
o
2
c
+o
nc
+o
cj
__
j
j,t
= :
)
o
nj,t

1
2
o
2
j
log
_
1 + exp
_
j
c
+o (o
t
j
c
) +
1
2
o
2
c
+o
nc
+o
cj
__
(5)
This is one restriction placed on the parameters of the model. Another restric-
tion results from the constant expected log dividend growth assumption. Using the
accounting identity from before relating prices, dividends, and dividend yields,
15
This derivation is available from the authors upon request.
9
j
t+1
= d
t+1
o
t+1
+o
t
j
j,t
= j
o
(j
c
+o (o
t
j
c
)) +o
t
j
j,t
= j
o
+ (1 o) (o
t
j
c
)
(6)
Equating the two expressions for the expected log price growth, we nd that our
model implies that movements in the log dividend yield are directly related to the market
price of price risk, o
nj,t
. Given the discussion in Section 2 of the paper relating
dividend yields and expected returns in addition to the underlying general equilibrium
discussion above, this nding is not too surprising.
o
nj,t
= j
o
(1 o) (o
t
j
c
) +:
)

1
2
o
2
j
log
_
1 + exp
_
j
c
+o (o
t
j
c
) +
1
2
o
2
c
+o
nc
+o
cj
__
(7)
This relationship between o
nj,t
and o
t
is easier to see if we linearize the log dividend
yield around its long run mean.
o
nj,t

= j
o
(1 o) (o
t
j
c
) +:
)

1
2
o
2
j
(c
1
+c
2
(o
t
j
c
))
o
nj,t

= j
o
(1 o) o
t
+ (1 o) j
c
+:
)

1
2
o
2
j
c
1
c
2
o
t
+c
2
j
c
o
nj,t

= (1 o +c
2
) o
t
j
o
+ (1 o +c
2
) j
c
+:
)

1
2
o
2
j
c
1
(8)
where c
1
and c
2
are linearization constants
c
1
= log
_
1 + exp
_
j
c
+
1
2
o
2
c
+o
nc
+o
cj
__
c
2
=
0 exp(j

+
1
2
o
2

+o
m
+o
p)
1+exp(j

+
1
2
o
2

+o
m
+o
p)
(9)
Since 1 o +c
2
0, a higher o
nj,t
(lower risk) corresponds to a lower log dividend
yield. The log dividend yield is perfectly correlated with risk (o
nj,t
) to a rst order
approximation. Thus, to a rst order approximation, since the log dividend yield follows
an AR(1) with persistence o, the process for risk (o
nj,t
) will also follow an AR(1) with
persistence o. By assuming a constant expected log dividend growth, assuming a
process for the log dividend yield is equivalent to assuming a process for risk (o
nj,t
).
Reenforcing the discussion from Section 2 and the general equilibrium discussion above,
time-varying risk has a one-to-one map with the dividend yield and, thus, is economically
inconsistent with a constant dividend yield assumption.
10
4.2 Adjusting the rm value process to the risk neutral mea-
sure
All of the analysis from the previous section relates back to the natural measure. In
order to calculate risk neutral probabilities and bond prices, we will need to specify
the risk neutral processes for the log price growth and log dividend yield. Based on
the conditional multivariate normality assumption from before, it can be shown (e.g.
Brennan (1979)) that the log price growth and log dividend yield are still conditionally
multivariate normal with the same conditional covariance matrix under the risk neutral
measure. However, the new conditional means are equal to the original conditional
means plus a risk correction term, i.e. the covariance between the particular shock
and the log SDF.
1
Q
t
[j
t+1
] = j
j,t
+o
nj,t
1
Q
t
[o
t+1
] = (j
c
+o (o
t
j
c
)) +o
nc
=
_
j
c
+
o
m
10
_
+o
_
o
t

_
j
c
+
o
m
10
__
(10)
Under the risk neutral measure, the log dividend yield still follows a stationary
AR(1), but it now has a new long run mean, j
c
+
o
m
10
. The no-arbitrage condition
derived before for the rm allows us to conveniently substitute out the expression for
the log price growths risk neutral conditional mean.
j
j,t
= :
)
o
nj,t

1
2
o
2
j
log
_
1 + exp
_
j
c
+o (o
t
j
c
) +
1
2
o
2
c
+o
nc
+o
cj
__
j
j,t
+o
nj,t
= :
)

1
2
o
2
j
log
_
1 + exp
_
j
c
+o (o
t
j
c
) +
1
2
o
2
c
+o
nc
+o
cj
__
(11)
Now, all of the relevant processes have been specied under the natural and risk
neutral measure.
4.3 Calculating default probabilities
We now consider a zero coupon bond where default is dened by the normal Merton
(1974) boundary conditions. In the standard Merton setup, the bondholders receive
their promised payment (1) at maturity t + , if the value of the rm is greater than or
equal to 1. Otherwise, the rm liquidates, and bondholders receive the value of the
rm.
Payo to bondholders at maturity t +,: min (1. 1
t+)
)
If 1
t+)
1, bondholders receive 1.
If 1
t+)
< 1, default occurs and bondholders receive 1
t+)
.
(12)
In order to calculate a true probability of default, we need to solve for the distrib-
11
ution of rm value at the bonds maturity date conditional on todays information set.
Working in logs and using the accounting identity for log prices,
log (1
t+)
) = log (1
t
) +

)
I=1
j
t+I
= log (1
t
) +,j
o
+ (1 o)

)1
I=0
(o
t+I
j
c
) +

)
I=1

j,t+I
= log (1
t
) +,j
o
+ (1 o)
_

)1
I=0
o
I
(o
t
j
c
) +

)1
I=1
_
10
jk
10
_

c,t+I
_
+

)
I=1

j,t+I
= log (1
t
) +,j
o
+ (1 o)
_
(o
t
j
c
)
_
10
j
10
_
+

)1
I=1
_
10
jk
10
_

c,t+I
_
+

)
I=1

j,t+I
.
(13)
We can reduce this further:
log (1
t+)
) = log (1
t
) +,j
o
+ (o
t
j
c
)
_
1 o
)
_
+

)1
I=1
_
1 o
)I
_

c,t+I
+

)
I=1

j,t+I
.
(14)
Since the log price at maturity is just the sum of i.i.d. multivariate normal shocks, the
log price at maturity is normal too. Lets calculate the mean and variance.
1
t
[log (1
t+)
)] = log (1
t
) +,j
o
+ (o
t
j
c
)
_
1 o
)
_
\ c:
t
[log (1
t+)
)] = o
2

)1
I=1
_
1 o
)I
_
2
+,o
2
p
+ 2o
cj

)1
I=1
_
1 o
)I
_
(15)
Given the mean and variance of the normally distributed log price at maturity, we can
easily calculate default probabilities in closed form.
111 =
_
log(1)1t[log(1
t+j
)]
p
\ ovt[log(1
t+j
)]
_
(16)
where () represents the standard normal cdf and 111 represents the true default
probability.
We can go through a similar exercise to calculate the risk neutral probability. How-
ever, since the risk neutral conditional mean for the log price (refer to Section 3.2) is a
nonlinear function of a normally distributed random variable (o
t
), the log price at matu-
rity will not be normally distributed. Thus, we will have to simulate the processes under
the risk neutral measure and estimate the risk neutral default probability (1`11) from
the simulations.
Similar to continuous-time, contingent claim analysis, we value the , period, zero
coupon bond (1
t
) by discounting back the expected payo at maturity under the risk
neutral measure by the constant risk free rate.
1
t
= exp (,:
)
) 1
Q
t
[min (1. 1
t+)
)] (17)
12
As with the risk neutral default probability calculation, we can not solve for the bond
price in closed form since the log price at maturity does not have a normal distribution.
The bond price can be obtained via simulation methods too.
5 Economic magnitude of changes in default proba-
bilities
This section analytically measures the reaction of the true and risk neutral probability
of default in response to cash ow versus expected return news. In order to gauge
the economic signicance of the results, we also calibrate our model using the S&P 500
rm level parameters estimated in Section 3. Cash ow news is identied as a price
movement with no change in the dividend yield while expected return news is identied
as a price movement with no change in the level of the dividend, i.e. an opposite move in
the dividend yield. In general, we nd that true default probabilities are more sensitive
to cash ow news relative to expected return news, especially when the dividend yield
process is less persistent. In contrast, we nd that risk neutral default probabilities are
more sensitive to expected return news relative to cash ow news since expected return
news moves the dividend yield and, thus, the risk neutral drift.
5.1 Sensitivity of the default probability to dierent sources of
news
In the last section, we showed that under the natural measure
1
t
[log (1
t+)
)] = log (1
t
) +,j
o
+ (o
t
j
c
)
_
1 o
)
_
\ c:
t
[log (1
t+)
)] = o
2

)1
I=1
_
1 o
)I
_
2
+,o
2
p
+ 2o
cj

)1
I=1
_
1 o
)I
_
(18)
Given the mean and variance of the normally distributed log price at maturity, in the
last section, we also calculated default probabilities in closed form.
111 =
_
log(1)1t[log(1
t+j
)]
p
\ ovt[log(1
t+j
)]
_
(19)
where () represents the standard normal cdf and 111 represents the true default
probability.
Since cash ow news is identied as a price movement with no change in the dividend
yield while expected return news is identied as a price movement with no change in the
level of the dividend, it will be helpful to substitute out log (1
t
) with log (1
t
) o
t
from
the 1
t
[log (1
t+)
)] expression. Thus, we can rewrite the expected log price at maturity
as
13
1
t
[log (1
t+)
)] = log (1
t
) o
t
+,j
o
+ (o
t
j
c
)
_
1 o
)
_
1
t
[log (1
t+)
)] = log (1
t
) + (o
t
) +,j
o
((o
t
) +j
c
)
_
1 o
)
_
(20)
We are interested in measuring the default sensitivity to a cash ow-induced price
change, i.e. movement in log (1
t
), and an expected return-induced price change, i.e.
movement in o
t
. Thus, we need to calculate
0T11
0 log(1t)
and
0T11
0[ct]
. Since we are considering
price increases, we take the partial derivative with respect to the negative log dividend
yield. Holding the level of the dividend constant, a decrease in the dividend yield
corresponds to an increase in price. For the cash ow-induced price move,
0T11
0 log(1t)
= :
_
log(1)1t[log(1
t+j
)]
p
\ ovt[log(1
t+j
)]
_
01t[log(1
t+j
)]
0 log(1t)
= :
_
log(1)1t[log(1
t+j
)]
p
\ ovt[log(1
t+j
)]
_
(21)
where :() represents the standard normal pdf. For the expected return-induced price
move,
0T11
0[ct]
= :
_
log(1)1t[log(1
t+j
)]
p
\ ovt[log(1
t+j
)]
_
01t[log(1
t+j
)]
0[ct]
= :
_
log(1)1t[log(1
t+j
)]
p
\ ovt[log(1
t+j
)]
_
_
1
_
1 o
)
__
= :
_
log(1)1t[log(1
t+j
)]
p
\ ovt[log(1
t+j
)]
_
o
)
(22)
The default probability is more sensitive to a cash ow-induced price movement.
Changes in expected returns are temporary, having a smaller impact on default prob-
abilities. For rms with more persistent expected return processes, i.e. higher o, this
eect is mitigated since the mean reversion is slower. Also, for rms with shorter term
debt, i.e. lower ,, this eect is mitigated since less mean reversion can occur over a
shorter time period.
Unfortunately, we cannot perform a similar closed form analysis for the risk neutral
probability of default. Since the risk neutral conditional mean for the log price is
a nonlinear function of a normally distributed random variable (o
t
), the log price at
maturity will not be normally distributed. However, if we linearize the conditional mean,
the log price at maturity will be normally distributed to a rst order approximation. To
provide some insight and to complement the true default probability results, we perform
the partial derivative analysis with the linearized model. In the next section, we will
conrm our rst order approximation results here with a simulation exercise that draws
from the exact processes and distributions. If we linearize the risk neutral drift of the
log price process around the risk neutral unconditional mean for o
t
, it can be shown that
the log price at maturity is normally distributed with the following mean and variance:
14
1
Q
t
[log (1
t+)
)] = log (1
t
) +,
_
:
)

1
2
o
2
j
c

1
_
c

2
o
_
_
o
t

_
j
c
+
o
m
10
__
_
10
j
10
__
= log (1
t
) o
t
+,
_
:
)

1
2
o
2
j
c

1
_
c

2
o
_
_
o
t

_
j
c
+
o
m
10
__
_
10
j
10
__
= log (1
t
) + (o
t
) +,
_
:
)

1
2
o
2
j
c

1
_
+c

2
o
_
_
(o
t
) +
_
j
c
+
o
m
10
__
_
10
j
10
__
\ c:
Q
t
[log (1
t+)
)] = (c

2
o)
2
o
2
c

)1
I=1
_
10
jk
10
_
2
+,o
2
j
2c

2
oo
cj

)1
I=1
_
10
jk
10
_
(23)
where
j

c
= j
c
+
0o
m
10
c

1
= log (1 +exp (j

c
))
c

2
=
exp(j

)
1+exp(j

)
= exp
_
o
nc
+o
cj
+
1
2
o
2
c
_
(24)
Both c

1
and c

2
are the linearization constants.
Similar to the true default probability analysis, we can easily calculate risk neutral
default probabilities in closed form.
1`11 =
_
log(1)1
Q
t
[log(1
t+j
)]
p
\ ov
Q
t
[log(1
t+j
)]
_
(25)
where 1`11 represents the risk neutral default probability. The sensitivity of the
risk neutral default probability to a cash ow-induced price movement (
01.11
0 log(1t)
) and
expected return-induced price movement (
01.11
0[ct]
) is the following:
01.11
0 log(1t)
= :
_
log(1)1
Q
t
[log(1
t+j
)]
p
\ ov
Q
t
[log(1
t+j
)]
_
01
Q
t
[log(1
t+j
)]
0 log(1t)
= :
_
log(1)1
Q
t
[log(1
t+j
)]
p
\ ov
Q
t
[log(1
t+j
)]
_
(26)
and
01.11
0[ct]
= :
_
log(1)1
Q
t
[log(1
t+j
)]
p
\ ov
Q
t
[log(1
t+j
)]
_
01
Q
t
[log(1
t+j
)]
0 log(1t)
= :
_
log(1)1
Q
t
[log(1
t+j
)]
p
\ ov
Q
t
[log(1
t+j
)]
_
_
1 +c

2
o
_
10
j
10
__
(27)
In contrast to the true default probability results, risk neutral default probabilities,
i.e. bond prices, are more sensitive to expected return-induced price movements since
c

2
o
_
10
j
10
_
0 for 0 < o < 1. When the price goes up due to lower than expected
future returns, the risk neutral price drift increases since the expected future dividend
15
yield is now lower. This eect does not occur for a cash ow-induced price movement.
For rms with more persistent expected return processes, i.e. higher o, this eect is
exacerbated since future risk neutral price drifts are more likely to be aected by the
original expected return shock that moved the dividend yield. Also, for rms with
longer term debt, i.e. higher ,, this eect is greater since the original expected return
shock that moved the dividend yield will have a larger cumulative eect over a longer
time period.
5.2 Predicted changes in default probabilities for S&P 500
rms
In the last section, we analytically derived the true and risk neutral default probability
sensitivities to a cash ow and expected return-induced price movement. Cash ow-
induced price movements are more important for the true default probability while
expected return-induced price movements are more important for the risk neutral default
probability, i.e. bond prices. In this section we calibrate our model to the actual
data, providing a means to judge and measure the economic signicance of our results.
Also, since the analytical results for the risk neutral default probabilities were only
approximate, this section provides additional support for our sensitivity results.
Most of the parameters were estimated using the actual rm level S&P 500 data
described in Section 3 and reported in Tables 1 and 2, Panel B. For all of the numerical
results, we used the following median annualized parameters: time to maturity (,) =
5 years, exp (:
)
) 1 = 2.0%, o
nc
= 0, o
j
= 30%, o
c
= 24%,
o
p
o

o
p
= -0.63, j
c
=
-3.16, j
o
= 0.054, face value of debt (1) = 2,722, o = 0.66, 1
t
= 10,533. For the
holding D/P constant results, i.e. cash ow-induced price movement, we use a o
t
= -3.42, the December 2002 value for the median S&P 500 rm. For the holding D
constant results, i.e. expected return-induced price movement, we use log (1
t
) = 5.84,
the implied number from the December 2002 value of o
t
(= -3.42) and 1
t
(= 10,533)
for the median S&P 500 rm. We set the market price of dividend yield risk (o
nc
) to
zero in order to emphasize that this new risk premium was not important to our main
results. The 30% annual rm level asset price volatility can be implied from the median
S&P 500 rm leverage, a stock price volatility of 40%, a risky bond price volatility of
10%, and a 0.1 correlation between stock and bond prices.
Given the median S&P 500 rm level parameters described above, we calculate the
change in default probability to various instantaneous moves in the underlying asset
price. This is performed for both true and risk neutral default probabilities, cash ow
and expected return-induced price movements, and various dividend yield persistence
parameters (o). The size of the instantaneous price moves can be interpreted as a one
standard deviation in the daily, weekly, monthly, and annual underlying asset price.
In order to approximate the daily, weekly, and monthly underlying asset volatility, we
use the 30% annual number and implicitly assume that asset price changes are i.i.d.,
16
i.e. volatility increases with the square root of time. This exercise produces a 1.89%,
4.16%, and 8.66% for the daily, weekly, and monthly price volatility, respectively.
16
In
addition to the median S&P 500 rm level persistence parameter o = 0.66, we also
show results for lower and higher values: o = 0.00, 0.25, 0.75, 0.95. When calculating
the risk neutral default probability, we simulate 1,000,000 draws from the actual risk
neutral distribution. No rst order approximation is used. Consistent with the previous
sections, cash ow news is identied as a price movement with no change in the dividend
yield while expected return news is identied as a price movement with no change in
the level of the dividend, i.e. an opposite move in the dividend yield.
Table 3 reports the calibration results for the true default probability (TDP). For
each of the dividend yield persistence parameters, the true default probability is more
sensitive to a cash ow-induced price move. The percentage change in the default
probability is always higher for the cash ow-induced price move. This general result
is most dramatic when the persistence parameter is low. In fact, when o = 0.00, an
expected return-induced price movement has no impact on the default probability. Even
with o = 0.25, an expected return-induced price movement still has almost no impact
on the default probability.
The results for o = 0.66, the median rm level S&P 500 value, are most relevant for
gauging economic signicance. For a price drop equivalent to a one standard deviation
daily move, the TDP increases 9.7682% in response to a cash ow shock while only
increasing 1.1797% for an expected return shock. In other words, the TDP sensitivity
to a cash ow-induced price movement is 8.28 times higher. These results become
even more striking when considering a price drop equivalent to a one standard deviation
annual move. The TDP increases 382.6429% in response to a cash ow shock while
only dropping 24.1806% for an expected return shock, making the TDP sensitivity to a
cash ow-induced price movement 15.82 times higher. From the perspective of a bond
rating agency or any institution interested in predicting default, there is a signicant
dierence between a 3.3519% and 0.8624% default probability, which correspond to the
"cash ow" TDP and "expected return" TDP level for the "annual shock" to the price,
respectively. Figure 6 provides a graphical representation of our main result: true
default probabilities are more sensitive to cash ow-induced price movements.
Table 4 reports the calibration results for the risk neutral default probability (RNDP).
In contrast to the results for the TDP, for each of the dividend yield persistence para-
meters (excluding o = 0.00), the risk neutral default probability is more sensitive to a
expected return-induced price move. There is no dierence between the default sen-
sitivities for the o = 0.00 case since any move in the current dividend yield due to an
expected return shock has no impact on the future expected dividend yield. The im-
material dierences that show up in Table 4 for the o = 0.00 case are due to simulation
error. Excluding the o = 0.00 special case, the percentage change in the default prob-
ability is always higher for the expected return-induced price move. However, for the
16
This calculation assumes 252 trading days in a year, 52 weeks in a year, and 12 months in a year.
17
o = 0.25 case, it is interesting to note that the dierences between the "cash ow" and
"expected return" RNDP sensitivities are quite small. In contrast to the TDP results,
the RNDP general result is most dramatic when the persistence parameter is high. For
example, for the o = 0.95 case, a one standard deviation annual price decline increases
the RNDP 181.6663% for the expected return-induced movement and 143.9657% for
the cash ow-induced movement. All of the numerical results for the RNDP reinforce
and complement the approximate rst order analytical results reported in the previous
section.
Similar to the TDP results, the results for o = 0.66, the median rm level S&P 500
value, are most relevant for gauging economic signicance. For a price drop equivalent
to a one standard deviation daily move, the RNDP increases 5.7206% in response to a
expected return shock while only increasing 5.2479% for a cash ow shock. In other
words, the RNDP sensitivity to a expected return-induced price movement is only 1.09
times higher. In absolute terms, these dierences are more pronounced when considering
a price drop equivalent to a one standard deviation annual move. The RNDP increases
161.6567% in response to a expected return shock while only increasing 146.3626% for a
cash ow shock. In relative terms, however, the RNDP sensitivity to a expected return-
induced price movement is only 1.10 times higher. From a bond valuation perspective,
there is still signicant dierence between a 16.9373% and 15.9473% risk neutral default
probability, which correspond to the "expected return" RNDP and "cash ow" RNDP
level for the "annual shock" to the price, respectively. Under the assumption of zero
recovery for simplicity, these RNDP dierences (16.9373% vs. 15.9473%) correspond to
overpaying for the bond by 1.19% if the price move is incorrectly assumed to be cash
ow-induced. Relative to the TDP results, the RNDP results are less dramatic, but
still economically signicant.
6 Conclusion
Most credit risk measures produced by traditional credit agencies were too slow to react
to the multiple, high prole, corporate bankruptcies of the recent past Many see this
as a failure of fundamental-based measures of credit risk.
During the tech boom of the 1990s, asset valuations sky-rocketed, predicting un-
believably low default rates via a standard Merton model. Many see this as a failure of
market-based measures of credit risk.
This paper presents a model of default that explains and provides a solution to
the above-mentioned criticisms, all within a rational, Merton-type setup. Motivated
by both theory and the empirical relationships found in S&P 500 rms, in the model,
expected dividend growth is constant and discount rates (expected returns) are time-
varying, producing a stochastic dividend yield in equilibrium. As a result, asset prices
can move due to news about dividends (cash ows or fundamentals) and/or news
about future discount rates (expected returns or non-fundamentals). Movements in
18
the dividend yield identify changes in the discount rate. Asset prices are no longer a
sucient statistic for fundamentals.
The model predicts that asset price movements that are driven by fundamentals,
i.e. price moves that occur with no change in the dividend yield, have a greater impact
on true default probabilities than asset price movements due to changes in discount
rates, i.e. price moves that occur with no change in the level of the dividend and, thus, a
change in the dividend yield. Intuitively, movements in the discount rate are expected
to mean-revert, causing expected price growth to move in the opposite direction of the
current price change. This osetting move in expected price growth dampens the eect
that the current price change has on the default probability. The osetting expected
price growth is largest when the change in discount rate is less persistent. In the extreme
case where discount rate changes display no persistence, movements in asset prices due
to a discount rate shock contain no information about changes in default probability.
Our stochastic dividend yield model also has implications for risk neutral default
probabilities, i.e. bond prices. Similar to the true default probability results, we nd
that movements in the underlyings asset value can contain a varying degree of informa-
tion about changes in risk neutral default probabilities, depending on how much of the
underlyings price movement was due to news about cash ows (vs. the discount rate).
However, in contrast to the true default probability results, our results suggest under-
lying asset price changes due to expected return news (or "non-fundamentals") play a
more important role when calculating risk neutral default probabilities since expected
return news moves the dividend yield and, thus, the asset value risk neutral drift. News
about cash ows only aects the price level without changing the dividend yield.
The results for both the true and risk neutral default probabilities are more than
just analytically signicant. Using parameters estimated from actual S&P 500 rm
level data, the model produces economically signicant results too. However, the true
default probability results are much more dramatic.
The tech boom of the 1990s was a time where prices and dividend yields moved
in opposite directions, suggesting a discount rate-induced price change and, thus, a
price change that contained little information about default probabilities. The recent,
high prole bankruptcies were preceded by price drops with little movements in cash
ow yields, suggesting a "fundamental"-induced price change and, thus, a price change
that contained substantial information about default probabilities. In sum, since price
changes contain varying degrees of information about default, our model provides a role
for both fundamental-based and market-based measures of credit risk, all within a ratio-
nal, Merton-type model. It also identies those rms where fundamental analysis will
add the most value: rms with less persistent discount rate (dividend yield) processes.
19
References
Altman, Edward I., 1968, Financial ratios, discriminant analysis and the prediction of
corporate bankruptcy, Journal of Finance 23, 589-609.
Bekaert, Geert, and Steven R. Grenadier, 2001, Stock and bond pricing in an ane
economy, Working Paper: Columbia University.
Bharath, Sreedhar and Tyler Shumway, 2004, Forecasting default with the Merton
model, unpublished paper, University of Michigan.
Black, Fischer, and Myron Scholes, 1973, The pricing of options and corporate liabili-
ties, Journal of Political Economy 81, 637654.
Black, Fischer, and John C. Cox, 1976, Valuing corporate securities: Some eects of
bond indenture provisions, Journal of Finance 31, 351367.
Brennan, Michael J., 1979. The pricing of contingent claims in discrete time models,
Journal of Finance 34, 5368.
Campbell, John Y., 1991 A variance decomposition for stock returns, Economic Journal
101, 157179.
Campbell, John Y., Jens Hilscher, and Jan Szilagyi, 2007, In Search of Distress Risk,
Journal of Finance, forthcoming.
Campbell, John Y., and Robert J. Shiller, 1988a, The dividend-price ratio and expec-
tations of future dividends and discount factors, Review of Financial Studies 1,
195228.
Campbell, John Y., and Robert J. Shiller, 1988b, Stock prices, earnings, and expected
dividends, Journal of Finance 43, 661676.
Chava, Sudheer and Robert A. Jarrow, 2004, Bankruptcy prediction with industry
eects, Review of Finance 8, 537569.
Cochrane, John H., 1991, Volatility tests and ecient markets: A review essay, Journal
of Monetary Economics 27, 463485.
Cochrane, John H., 1992, Explaining the variance of price-dividend ratios, Review of
Financial Studies 5, 243-280.
Cochrane, John H., 1994, Permanent and transitory components of GNP and stock
prices, Quarterly Journal of Economics 109, 241-265.
Collin-Dufresne, Pierre, and Robert S. Goldstein, 2001, Do credit spreads reect sta-
tionary leverage ratios? Journal of Finance 56, 19291957.
20
Delianedis, Gordon and Robert Geske, 1999, Credit risk and risk neutral default proba-
bilities: Information about rating migrations and defaults, Working Paper, UCLA.
Due, Darrell, Leandro Saita, and Ke Wang, 2007, Multi-Period Corporate Failure
Prediction with Stochastic Covariates, Journal of Financial Economics 83, 635
665.
Hillegeist, Stephen A., Elizabeth Keating, Donald P. Cram and Kyle G. Lunstedt, 2004,
Assessing the probability of bankruptcy, Review of Accounting Studies 9, 534.
Hecht, Peter, and Tuomo Vuolteenaho, 2004, Explaining returns with cash-ow proxies,
Review of Financial Studies, forthcoming.
Huang, Ming, and Jing-zhi Huang, 2003, How much of the corporate-treasury yield
spread is due to credit risk? Working Paper, Standford University.
Leland, Hayne E., 2002, Predictions of expected default frequencies in structural models
of debt, Working Paper, UC Berkeley.
Longsta, Francis and Eduardo Schwartz, 1995, A simple approach to valuing risky
xed and oating rate debt, Journal of Finance 50, 789-819.
Merton, Robert. C., 1974, On the pricing of corporate debt: The risk structure of
interest rates, Journal of Finance 29, 449469.
Ohlson, James A., 1980, nancial ratios and the probabilitic prediction of bankruptcy,
Journal of Accounting Research 18, 109-131.
Rubinstein, M., 1976, The Valuation of Uncertain Income Streams and the Pricing of
Options, Bell Journal of Economics and Management Science 7, 407-25.
Shumway, Tyler, 2001, Forecasting bankruptcy more accurately: a simple hazard
model, Journal of Business 74, 101-124.
Stapleton, R. C., and M. G. Subrahmanyam, 1984, The valuation of multivariate con-
tingent claims in discrete time models, Journal of Finance 39, 207228.
Vassalou, Maria and Yuhang Xing, 2004, Default Risk in Equity Returns, Journal of
Finance 59, 831-868.
Vasicek, O., 1977, An equilibrium characterizations of the term structure, Journal of
Financial Economics 5, 177188.
Vuolteenaho, Tuomo , 2002, What drives rm-level stock returns?, Journal of Finance
57, 233264.
Zmijewski, Mark E., Methodological issues related to the estimation of nancial distress
prediction models, Journal of Accounting Research 22, 59-82.
21
Figure 1: Aggregate S&P 500 Equity Dividend Yield Over Time

This figure graphs the annual equity dividend yield for the aggregate S&P 500 over the 1980 2002
period. The annual aggregate equity dividend yield is computed by taking the cross-sectional sum
of firm level common dividends divided by the cross-sectional sum of firm level common equity
values. To be consistent with our firm level sample, we only include firms that were in the S&P500
in 1990. The firm level equity data is computed from CRSP's RET and RETX series.

S&P 500 Aggregate Equity Dividend Yield
0.015
0.020
0.025
0.030
0.035
0.040
0.045
0.050
0.055
0.060
1
9
8
0
1
9
8
1
1
9
8
2
1
9
8
3
1
9
8
4
1
9
8
5
1
9
8
6
1
9
8
7
1
9
8
8
1
9
8
9
1
9
9
0
1
9
9
1
1
9
9
2
1
9
9
3
1
9
9
4
1
9
9
5
1
9
9
6
1
9
9
7
1
9
9
8
1
9
9
9
2
0
0
0
2
0
0
1
2
0
0
2
Year
E
q
u
i
t
y

D
P

Figure 2: Aggregate S&P 500 Asset Dividend Yield Over Time

This figure graphs the annual asset dividend yield for the aggregate S&P 500 over the 1980 2002
period. The annual aggregate asset dividend (payout) yield is computed by taking the cross-
sectional sum of firm level payouts divided by the cross-sectional sum of firm level assets. To be
consistent with our firm level sample, we only include firms that were in the S&P500 in 1990. The
firm's asset dividend (payout) yield is defined as dividend +preferred dividend +interest (from
COMPUSTAT) paid during the year divided by the end of year market value of equity +book
preferred equity +book debt (from CRSP and COMPUSTAT), where book debt includes short and
long term maturities.

S&P 500 Aggregate Asset Dividend Yield
0.020
0.030
0.040
0.050
0.060
0.070
1
9
8
0
1
9
8
1
1
9
8
2
1
9
8
3
1
9
8
4
1
9
8
5
1
9
8
6
1
9
8
7
1
9
8
8
1
9
8
9
1
9
9
0
1
9
9
1
1
9
9
2
1
9
9
3
1
9
9
4
1
9
9
5
1
9
9
6
1
9
9
7
1
9
9
8
1
9
9
9
2
0
0
0
2
0
0
1
2
0
0
2
Year
A
s
s
e
t

D
P

Figure 3: FirstEnergy Firm Level Asset Dividend Yield Over Time

This figure graphs the annual asset dividend yield for FirstEnergy, an S&P 500 firm, over the 1980
2002 period. The firm's asset dividend (payout) yield is defined as dividend +preferred dividend
+interest (from COMPUSTAT) paid during the year divided by the end of year market value of
equity +book preferred equity +book debt (from CRSP and COMPUSTAT), where book debt
includes short and long term maturities.

FirstEnergy Asset Dividend Yield
0.030
0.040
0.050
0.060
0.070
0.080
0.090
0.100
0.110
0.120
1
9
8
0
1
9
8
1
1
9
8
2
1
9
8
3
1
9
8
4
1
9
8
5
1
9
8
6
1
9
8
7
1
9
8
8
1
9
8
9
1
9
9
0
1
9
9
1
1
9
9
2
1
9
9
3
1
9
9
4
1
9
9
5
1
9
9
6
1
9
9
7
1
9
9
8
1
9
9
9
2
0
0
0
2
0
0
1
2
0
0
2
Year
A
s
s
e
t

D
P

Figure 4: Merrill Lynch Firm Level Asset Dividend Yield Over Time

This figure graphs the annual asset dividend yield for Merrill Lynch, an S&P 500 firm, over the
1980 2002 period. The firm's asset dividend (payout) yield is defined as dividend +preferred
dividend +interest (from COMPUSTAT) paid during the year divided by the end of year market
value of equity +book preferred equity +book debt (from CRSP and COMPUSTAT), where book
debt includes short and long term maturities.

Merrill Lynch Asset Dividend Yield
0.030
0.040
0.050
0.060
0.070
0.080
0.090
0.100
0.110
0.120
0.130
1
9
8
0
1
9
8
1
1
9
8
2
1
9
8
3
1
9
8
4
1
9
8
5
1
9
8
6
1
9
8
7
1
9
8
8
1
9
8
9
1
9
9
0
1
9
9
1
1
9
9
2
1
9
9
3
1
9
9
4
1
9
9
5
1
9
9
6
1
9
9
7
1
9
9
8
1
9
9
9
2
0
0
0
2
0
0
1
2
0
0
2
Year
A
s
s
e
t

D
P

Figure 5: PACCAR Firm Level Asset Dividend Yield Over Time

This figure graphs the annual asset dividend yield for PACCAR, an S&P 500 firm, over the 1980
2002 period. The firm's asset dividend (payout) yield is defined as dividend +preferred dividend +
interest (from COMPUSTAT) paid during the year divided by the end of year market value of
equity +book preferred equity +book debt (from CRSP and COMPUSTAT), where book debt
includes short and long term maturities.

PACCAR Asset Dividend Yield
0.020
0.030
0.040
0.050
0.060
0.070
0.080
0.090
0.100
1
9
8
0
1
9
8
1
1
9
8
2
1
9
8
3
1
9
8
4
1
9
8
5
1
9
8
6
1
9
8
7
1
9
8
8
1
9
8
9
1
9
9
0
1
9
9
1
1
9
9
2
1
9
9
3
1
9
9
4
1
9
9
5
1
9
9
6
1
9
9
7
1
9
9
8
1
9
9
9
2
0
0
0
2
0
0
1
2
0
0
2
Year
A
s
s
e
t

D
P

Figure 6: True Default Probability vs. Cash Flow and Expected Return-Induced Price
Changes

This figure graphs the true default probability (TDP) as a function of the underlying asset price.
The underlying asset price moves either due to news about cash flows or expected returns. The
TDP is calculated based on the model presented in Section 4, where the parameters of the model are
chosen to match the actual median values found in the firm level S&P 500 data reported in Tables 1
and 2 (discussed in Section 3 of the paper). The base line underlying asset price, denoted by a star
on the graph, is $10,533, the 2002 S&P 500 median value in millions. Four instantaneous price
decrease sizes from the base line are highlighted in the graph: $199.07, $438.17, $912.16, and
$3,159.90. These four perturbations (shocks) can be interpreted as a one standard deviation daily,
weekly, monthly, and annual price movement, respectively. The Cash Flow-Induced line graphs
the TDP under the assumption that the entire price movement is cash flow-induced. The Expected
Return-Induced line graphs the TDP under the assumption that the entire price movement is
expected return-induced. The bond is assumed to be zero coupon and matures in five years.

0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
7
,
3
7
3
7
,
5
7
7
7
,
7
8
1
7
,
9
8
5
8
,
1
8
9
8
,
3
9
3
8
,
5
9
7
8
,
8
0
1
9
,
0
0
5
9
,
2
0
9
9
,
4
1
3
9
,
6
1
7
9
,
8
2
1
1
0
,
0
2
5
1
0
,
2
2
9
1
0
,
4
3
3
1
0
,
6
3
7
1
0
,
8
4
1
1
1
,
0
4
5
1
1
,
2
4
9
1
1
,
4
5
3
1
1
,
6
5
7
1
1
,
8
6
1
1
2
,
0
6
5
1
2
,
2
6
9
1
2
,
4
7
3
1
2
,
6
7
7
1
2
,
8
8
1
1
3
,
0
8
5
1
3
,
2
8
9
1
3
,
4
9
3
Asset Pri ce
T
r
u
e

D
e
f
a
u
l
t

P
r
o
b
a
b
i
l
i
t
y
Cash Flow-Induced Expected Return-Induced
B
a
s
e

L
i
n
e
D
a
i
l
y

S
h
o
c
k
W
e
e
k
l
y

S
h
o
c
k
M
o
n
t
h
l
y

S
h
o
c
k
A
n
n
u
a
l

S
h
o
c
k

Table 1: S&P 500 Firm Level Asset Dividend Yield Summary Statistics

This table reports annual summary statistics for S&P500 firms. EquityDP is the equity dividend
yield, computed from CRSP's monthly RET and RETX series, DP is the firm's payout yield (asset
dividend), where payout is equal to dividend +preferred dividend +interest (COMPUSTAT annual
data items 19, 15) paid during the year and assets are measured as end of year market value of
equity +book preferred equity +book debt (CRSP prc*shrout, COMPUSTAT data items 130,
9+34), where book debt is measured as long term debt +debt in current liabilities. Variables are
measured at the end of each calendar year. logDP is the natural log of DP. LEV is the firms
financial leverage, defined as book debt divided by assets. logD Grow is the log growth rate of the
firm's payout, and P is the value of assets in millions. Panel A reports aggregate level statistics.
The aggregate asset dividend yield is computed by taking the cross-sectional sum of firm level
payouts divided by the cross-sectional sum of firm level assets. The aggregate equity dividend
yield is computed similarly. Panel B reports the pooled firm level statistics. We include
observations from 1980-2002 for firms that were in the S&P500 in 1990. For logDP and P, we also
include statistics for only 2002. We report mean and standard deviation as well as percentiles of the
distribution.

Panel A: Aggregate Level S&P 500, Across Time, 1980 - 2002
Statistics Equity DP DP logDP LEV logD Grow
Mean 0.0338 0.0452 -3.1544 0.3954 0.0698
SD 0.0125 0.0150 0.3519 0.0557 0.0580
Min 0.0156 0.0247 -3.7029 0.2815 -0.0299
p25 0.0230 0.0304 -3.4931 0.3488 0.0272
p50 0.0317 0.0456 -3.0872 0.4152 0.0723
p75 0.0457 0.0561 -2.8804 0.4361 0.1101
Max 0.0581 0.0722 -2.6282 0.4617 0.1961
N 23 23 23 23 22
Panel B: Firm Level S&P 500, Across Firms and Time, 1980 - 2002
2002 ONLY 2002 ONLY
Statistics Equity DP DP logDP LEV logD Grow logDP P
Mean 0.0327 0.0472 -3.3036 0.3144 0.0676 -3.5593 34199
SD 0.0257 0.0305 0.8100 0.2295 0.3701 0.7140 80738
Min 0.0000 0.0009 -7.0069 0.0000 -6.3238 -7.0069 190
p25 0.0153 0.0254 -3.6726 0.1318 -0.0280 -3.9175 3810
p50 0.0278 0.0423 -3.1641 0.2730 0.0541 -3.4193 10533
p75 0.0447 0.0632 -2.7612 0.4545 0.1507 -3.0735 30390
Max 0.1356 0.1608 -1.8276 0.9915 10.5132 -1.8276 917720
N 7160 7160 7160 7093 6677 211 211

Table 2: S&P 500 Firm Level Asset Dividend Yield AR(1) Estimation Summary Statistics

This table reports summary statistics for annual log asset dividend yield AR(1) estimations
performed on S&P500 firms. The firm payout yield, DP, and firm assets, P, are defined as in Table
1. logDP is the natural log of DP. Corr(logP,logDP) is the conditional correlation between
changes in log asset growth and the log payout yield (using our model, we estimate this as the
unconditional correlation between the shock to logDP and log dividend growth minus shock to
logDP). Theta is the AR(1) coefficient from the log payout yield regression; we refer to it as the
persistence parameter. SD(logDP Shock) is the standard deviation of the log payout yield AR(1)
residual. Panel A reports statistics for aggregate level variables. The annual aggregate asset
dividend yield is equal to the cross-sectional sum of firm level payouts divided by the cross-
sectional sum of firm level assets. Panel B reports the firm level statistics. The AR(1) is estimated
over the 1980-2002 period for those firms that were in the S&P500 in 1990. In order to be
included, a firm is required to have logDP data for at least 10 years. We report statistics for those
firms with persistence parameters between 0 and 1 in order to be consistent with our model.

Panel A: Aggregate Level AR(1) Parameters: S&P 500 Firms, 1980 - 2002
Statistics DP SD(DP) SD(logDP) SD(logDP Shock) Corr(logP,logDP) Theta
Mean 0.0452 0.0150 0.3519 0.0848 -0.7536 0.9668
Panel B: Firm Level AR(1) Parameters: S&P 500 Firms, 1980 - 2002
Statistics DP SD(DP) SD(logDP) SD(logDP Shock) Corr(logP,logDP) Theta
Mean 0.0475 0.0194 0.4205 0.2832 -0.5660 0.6309
SD 0.0212 0.0113 0.2332 0.1763 0.2809 0.2211
Min 0.0015 0.0006 0.1023 0.0686 -0.9226 0.0187
p25 0.0323 0.0113 0.2814 0.1791 -0.7645 0.5064
p50 0.0482 0.0174 0.3656 0.2361 -0.6310 0.6649
p75 0.0630 0.0265 0.4911 0.3331 -0.4355 0.7990
Max 0.1053 0.0603 1.5247 1.1861 0.5740 0.9929
N 327 327 327 327 327 327
Table 3: True Default Probability Sensitivities to Cash Flow and Expected Return-Induced Price Movements

This table reports true default probability (TDP) sensitivities to cash flow and expected return-induced underlying price movements based on the
model presented in Section 4, where the parameters of the model are chosen to match the actual median values found in the firm level S&P 500 data
reported in Tables 1 and 2 (discussed in Section 3 of the paper). The base line underlying asset price is $10,533, the 2002 S&P 500 median value in
millions. Four instantaneous price decrease sizes from the base line are considered: $199.07, $438.17, $912.16, and $3,159.90. These four
perturbations (shocks) can be interpreted as a one standard deviation daily, weekly, monthly, and annual price movement, respectively. TDP CF
shock reports the TDP under the assumption that the entire price movement is cash flow-induced. TDP ER shock reports the TDP under the
assumption that the entire price movement is expected return-induced. % increase in TDP over Base Line reports the percentage increase in the TDP
due to the price movement over the Base Line TDP. This is calculated for both the cash flow and expected return cases. Persistence Parameter refers
to the log asset dividend yield AR(1) coefficient used in the calibration. Although the median AR(1) persistence parameter from the sample of S&P
500 firms is 0.66, other values are considered. In all of the calibrations, the bond is assumed to be zero coupon and matures in five years. All
reported results are exact, i.e. no simulations. NA refers to not applicable.

True Default Probability (TDP)
Description Change in TDP TDP % increase in TDP over % increase in TDP over Persistence
Asset Price Asset Price Asset Price CF shock ER shock Base Line - CF shock Base Line - ER shock Parameter
Base Line 10,533.00 $ NA 0.7779% 0.7779% NA NA 0
Daily Shock 10,333.93 $ (199.07) $ 0.8534% 0.7779% 9.6994% 0.0000% 0
Weekly Shock 10,094.83 $ (438.17) $ 0.9546% 0.7779% 22.7191% 0.0000% 0
Monthly Shock 9,620.84 $ (912.16) $ 1.1961% 0.7779% 53.7632% 0.0000% 0
Annual Shock 7,373.10 $ (3,159.90) $ 3.7038% 0.7779% 376.1234% 0.0000% 0
Base Line 10,533.00 $ NA 0.7424% 0.7424% NA NA 0.25
Daily Shock 10,333.93 $ (199.07) $ 0.8153% 0.7425% 9.8259% 0.0092% 0.25
Weekly Shock 10,094.83 $ (438.17) $ 0.9134% 0.7425% 23.0318% 0.0205% 0.25
Monthly Shock 9,620.84 $ (912.16) $ 1.1476% 0.7427% 54.5859% 0.0437% 0.25
Annual Shock 7,373.10 $ (3,159.90) $ 3.6027% 0.7437% 385.2852% 0.1722% 0.25
Base Line 10,533.00 $ NA 0.6945% 0.6945% NA NA 0.66
Daily Shock 10,333.93 $ (199.07) $ 0.7623% 0.7027% 9.7682% 1.1797% 0.66
Weekly Shock 10,094.83 $ (438.17) $ 0.8535% 0.7129% 22.8924% 2.6438% 0.66
Monthly Shock 9,620.84 $ (912.16) $ 1.0712% 0.7341% 54.2376% 5.7103% 0.66
Annual Shock 7,373.10 $ (3,159.90) $ 3.3519% 0.8624% 382.6429% 24.1806% 0.66
Base Line 10,533.00 $ NA 0.6839% 0.6839% NA NA 0.75
Daily Shock 10,333.93 $ (199.07) $ 0.7496% 0.6990% 9.6005% 2.2084% 0.75
Weekly Shock 10,094.83 $ (438.17) $ 0.8377% 0.7180% 22.4810% 4.9765% 0.75
Monthly Shock 9,620.84 $ (912.16) $ 1.0476% 0.7583% 53.1726% 10.8715% 0.75
Annual Shock 7,373.10 $ (3,159.90) $ 3.2273% 1.0194% 371.8680% 49.0500% 0.75
Base Line 10,533.00 $ NA 0.7323% 0.7323% NA NA 0.95
Daily Shock 10,333.93 $ (199.07) $ 0.7950% 0.7804% 8.5503% 6.5617% 0.95
Weekly Shock 10,094.83 $ (438.17) $ 0.8782% 0.8431% 19.9135% 15.1272% 0.95
Monthly Shock 9,620.84 $ (912.16) $ 1.0734% 0.9860% 46.5708% 34.6392% 0.95
Annual Shock 7,373.10 $ (3,159.90) $ 2.9753% 2.2197% 306.2729% 203.1046% 0.95


Table 4: Risk Neutral Default Probability Sensitivities to Cash Flow and Expected Return-Induced Price Movements

This table reports risk neutral default probability (RNDP) sensitivities to cash flow and expected return-induced underlying price movements based on
the model presented in Section 4, where the parameters of the model are chosen to match the actual median values found in the firm level S&P 500
data reported in Tables 1 and 2 (discussed in Section 3 of the paper). The base line underlying asset price is $10,533, the 2002 S&P 500 median value
in millions. Four instantaneous price decrease sizes from the base line are considered: $199.07, $438.17, $912.16, and $3,159.90. These four
perturbations (shocks) can be interpreted as a one standard deviation daily, weekly, monthly, and annual price movement, respectively. RNDP CF
shock reports the RNDP under the assumption that the entire price movement is cash flow-induced. RNDP ER shock reports the RNDP under the
assumption that the entire price movement is expected return-induced. % increase in RNDP over Base Line reports the percentage increase in the
RNDP due to the price movement over the Base Line RNDP. This is calculated for both the cash flow and expected return cases. Persistence
Parameter refers to the log asset dividend yield AR(1) coefficient used in the calibration. Although the median AR(1) persistence parameter from the
sample of S&P 500 firms is 0.66, other values are considered. In all of the calibrations, the bond is assumed to be zero coupon and matures in five
years. All reported results are based on simulation via 1,000,000 draws. NA refers to not applicable.

Risk Neutral Default Probability (RNDP; Based on simulation via 1,000,000 draws)
Description Change in RNDP RNDP % increase in RNDP over % increase in RNDP over Persistence
Asset Price Asset Price Asset Price CF shock ER shock Base Line - CF shock Base Line - ER shock Parameter
Base Line 10,533.00 $ NA 6.3717% 6.3717% NA NA 0
Daily Shock 10,333.93 $ (199.07) $ 6.7332% 6.7250% 5.6735% 5.5448% 0
Weekly Shock 10,094.83 $ (438.17) $ 7.2236% 7.2194% 13.3701% 13.3041% 0
Monthly Shock 9,620.84 $ (912.16) $ 8.2616% 8.2788% 29.6608% 29.9308% 0
Annual Shock 7,373.10 $ (3,159.90) $ 16.0364% 16.0792% 151.6817% 152.3534% 0
Base Line 10,533.00 $ NA 6.4249% 6.4249% NA NA 0.25
Daily Shock 10,333.93 $ (199.07) $ 6.7518% 6.8016% 5.0880% 5.8631% 0.25
Weekly Shock 10,094.83 $ (438.17) $ 7.2662% 7.2667% 13.0944% 13.1021% 0.25
Monthly Shock 9,620.84 $ (912.16) $ 8.2802% 8.3360% 28.8767% 29.7452% 0.25
Annual Shock 7,373.10 $ (3,159.90) $ 16.0222% 16.2463% 149.3766% 152.8646% 0.25
Base Line 10,533.00 $ NA 6.4731% 6.4731% NA NA 0.66
Daily Shock 10,333.93 $ (199.07) $ 6.8128% 6.8434% 5.2479% 5.7206% 0.66
Weekly Shock 10,094.83 $ (438.17) $ 7.3051% 7.3906% 12.8532% 14.1740% 0.66
Monthly Shock 9,620.84 $ (912.16) $ 8.3317% 8.4287% 28.7127% 30.2112% 0.66
Annual Shock 7,373.10 $ (3,159.90) $ 15.9473% 16.9373% 146.3626% 161.6567% 0.66
Base Line 10,533.00 $ NA 6.4477% 6.4477% NA NA 0.75
Daily Shock 10,333.93 $ (199.07) $ 6.8470% 6.8409% 6.1929% 6.0983% 0.75
Weekly Shock 10,094.83 $ (438.17) $ 7.2607% 7.3921% 12.6091% 14.6471% 0.75
Monthly Shock 9,620.84 $ (912.16) $ 8.2894% 8.4994% 28.5637% 31.8206% 0.75
Annual Shock 7,373.10 $ (3,159.90) $ 15.9746% 17.1614% 147.7566% 166.1631% 0.75
Base Line 10,533.00 $ NA 6.4166% 6.4166% NA NA 0.95
Daily Shock 10,333.93 $ (199.07) $ 6.8012% 6.8285% 5.9938% 6.4193% 0.95
Weekly Shock 10,094.83 $ (438.17) $ 7.1892% 7.3721% 12.0406% 14.8911% 0.95
Monthly Shock 9,620.84 $ (912.16) $ 8.1984% 8.5770% 27.7686% 33.6689% 0.95
Annual Shock 7,373.10 $ (3,159.90) $ 15.6543% 18.0734% 143.9657% 181.6663% 0.95

S-ar putea să vă placă și