A General Affine Earnings

Valuation Model∗

Andrew Ang†
Columbia University and NBER

Jun Liu‡
UCLA

First Version: April 21, 1998

This Version: July 1, 2001

JEL Classification Codes: G12, M41.

Keywords: stock valuation, earnings, residual income model,

asset-pricing, affine model, linear information dynamics

∗ Thispaper was originally circulated as the working paper, “A Generalized Earnings Model of Stock
Valuation”. We thank Geert Bekaert, Michael Brennan, Ron Kasznik, Charles Lee, Jing Liu, Stephen
Penman, Stefan Reichelstein, Ken Singleton, Ramu Thiagarajan and seminar participants at Stanford
University and Mellon Capital for comments. We especially thank James Ohlson for encouragement,
mentorship and valuable advice.
† 3022 Broadway, 805 Uris, New York, NY 10027. email:aa610@columbia.edu, ph: (212) 854-9154.
‡ Anderson School C509, UCLA, CA 90095. jliu@anderson.ucla.edu, ph: (310) 825-7132.
 Abstract

We introduce a methodology, with two applications, that incorporates stochastic interest

rates, heteroskedasticity and risk aversion into the residual income model. In the first appli-
cation, goodwill is an affine (constant plus linear term) function where the constant and linear
coefficients are time-varying. Homoskedastic risk gives rise to a constant risk premium, while
heteroskedastic risk gives rise to linear state-dependent risk premiums. In the second applica-
tion, we present a class of models where a non-linear function for the price-to-book ratio can
be derived. We show how interest rates, risk, profitability and growth affect the price-to-book
ratio.
 This paper provides a parametric class of models that shows how a firm’s market value
relates to accounting data under stochastic interest rates, heteroskedasticity and adjustments for
risk aversion. We use the framework of the Residual Income Model (RIM), which expresses
the value of a stock as the firm’s book value plus the expected future discounted value of the
firm’s abnormal (or residual) earnings. Our methodology builds on the framework of Feltham
and Ohlson (1999), who extend the RIM to a no-arbitrage setting to accommodate time-varying
interest rates and risk aversion. Feltham and Ohlson give a partial parametric model of stock
valuation using accounting information in this setting. In a heteroskedastic environment with
stochastic interest rates and risk aversion, we extend this analysis in several ways. First, we
apply this methodology to the case where the dynamics of accounting variables are expressed
in dollar amounts as in Feltham and Ohlson (1995). Second, we derive a solution for the price-
to-book ratio of a firm as a function of stochastic interest rates, accounting rates of return and
growth in book.
Our first result extends the Linear Information Model (LIM) developed in Ohlson (1995)
and Feltham and Ohlson (1995). The LIM presents firm value as a linear function of current
observable accounting information and is derived under constant discount rates. This assump-
tion leads to a standard simplification where a single discount factor can be applied to all future
periods. The discount factor can incorporate an ad hoc adjustment for risk. There are certain
questions, however, which cannot be addressed under this assumption. For instance, is it always
possible to incorporate risk aversion as a spread in a constant discount factor? Can a linear so-
lution be found under the addition of time-varying interest rates, heteroskedasticity and risk
aversion? Feltham and Ohlson (1999) show how to adjust the RIM for risk but do not provide
a complete parametric model to answer these questions directly. They hint, however, that a
model as tractable as the LIM might exist under more general conditions. We propose a class
of models where an extended Feltham-Ohlson linear form can be preserved under stochastic
interest rates and time-varying risk premiums. Under risk neutrality and constant interest rates,
our model reduces to the Feltham-Ohlson LIM.
Our extension to the LIM expresses firm value as an affine combination (constant plus linear

1
form) of abnormal earnings and book value. We show that it is possible to choose a parameter-
ization so that an affine form for goodwill (the difference between price and book value) exists
under stochastic interest rates and risk aversion, provided that the interest rate process is un-
correlated with accounting variables. The coefficients in the affine form are time-varying, and
reflect the dependence on current zero coupon bond prices. Risk aversion potentially affects
both the constant and the linear terms. With risk aversion, homoskedasticity is captured conve-
niently only in the constant term, while risk aversion combined with heteroskedastic risk leads
to state-dependent risk premiums which appear in the linear terms.
Under the LIM, dollar amounts of residual earnings are assumed to follow a stationary pro-
cess and firm value is a linear function of contemporaneous earnings information. Questions
about the rate of earnings growth rather than the dollar amount of earnings, or firm growth,
cannot easily be addressed in this setting. Our second application focuses on the price-to-book
ratio, rather on than the dollar difference in the price and book, as in the LIM. We apply our pric-
ing methodology to the RIM framework of Feltham and Ohlson (1999), with a normalization
by book-value.1 This framework models the price-to-book as a function of stochastic interest
rates, a rate of return measure based on profitability (accounting returns of earnings in excess of
the risk-free rate), and firm growth. Risk aversion and heteroskedasticity of all three variables
are explicitly modeled.
Ratio analysis highlights the effect of the rate of profitability and growth on valuation. We
might anticipate that higher profitability of a firm would lead to higher price-to-book ratios, as
higher profitability increases firm value relative to current book value. However, the effect of
growth in book on the price-to-book is not so clear.2 By directly parameterizing growth in book
we can determine how the price-to-book ratio behaves when parameters underlying the process
for growth in book are changed. Another comparative static of interest is the mean-reversion
of profitability and growth. Standard economic arguments in a competitive environment argue
that these variables are mean-reverting.3 Does higher mean-reversion in profitability or growth
lead to higher or lower price-to-book ratios? Finally, risk-averse agents are affected by the
volatility of interest rates, profitability and growth. Hence, under risk aversion, volatility of

2
these variables affects the price-to-book.
We present a closed-form non-linear solution of the price-to-book ratio. The solution for-
mula allows characteristics of the behavior of the price-to-book to be examined by comparative
statics. As expected, increasing profitability increases the price-to-book ratio. Increasing the
conditional mean of book growth unambiguously increases the price-to-book ratio. We find that
increasing the autocorrelation of profitability or growth in book increases the price-to-book.
That is, ceteris paribus, firms with highly mean-reverting profitability or growth have lower
price-to-book ratios.
Conducting comparative statics with respect to the dynamics of the interest rate yields one
surprising behavior of the price-to-book which seems counter-intuitive. We may expect that in-
creasing the volatility of interest rates would decrease the price-to-book, since nominal interest
rates are positive (which is the case under a Cox-Ingersoll-Ross (1985) term structure model)
and increasing the volatility of interest rates implies higher discount factors in future periods.
However, under conservative accounting and risk neutrality, the price-to-book is an increasing
function of the volatility of all state variables including, surprisingly, the volatility of the inter-
est rate. This counter-intuitive behavior is due to a Jensen’s inequality effect which dominates
under risk neutrality. When risk aversion is introduced, the price-to-book ratio may fall as the
volatility of the interest rate increases.
Our methodology incorporates several stylized facts of interest rates and risk aversion which
bear on accounting valuation. First, interest rates are time-varying, which affects the discount
factors used in future periods. This is a “denominator” effect but interest rates also predict
future abnormal earnings, which is a “numerator” effect.4 Our formal methodology simul-
taneously handles both the denominator and numerator effect of time-varying interest rates.
Moreover, predictability of accounting information by any variable, not just interest rates, can
be accommodated. Second, risk-averse agents in the economy need to be compensated for the
uncertainty in the evolution of financial statement information because accounting information
is a driving factor of prices. Viewed another way, the risk premiums associated with the uncer-
tainties of accounting information are reflected in discount factors, as discussed by Feltham and

3
Ohlson (1999). Our methodology tractably incorporates risk aversion and shows how it affects
the LIM and book-to-market ratio dynamics.
Our methodology tractably and parsimonously incorporates rich dynamics of interest rates
and accounting variables by using “affine” processes (Duffie and Kan (1996)), where both the
conditional mean and conditional volatility take on affine forms (constant plus linear terms).
Ohlson (1995) and Feltham and Ohlson (1995) rely on simple AR(1) or modified AR(1) pro-
cesses to derive the LIM. These are special cases of the affine set-up. The affine processes also
formally encompass Feltham and Ohlson (1999)’s partial model, since they can incorporate
both heteroskedasticity of the driving variables and risk adjustments.
The rest of the paper is organized as follows. In Section 1 we describe the role of a “pricing
kernel” in no-arbitrage valuation. Section 2 presents the affine extension of the LIM and shows
that linearity can survive the introduction of stochastic interest rates and risk aversion. Section 3
applies the methodology to the case of ratio dynamics, and presents a closed-form model of the
price-to-book ratio. Comparative static exercises show how changes in the interest rate process,
profitability, growth and risk aversion affect the price-to-book of the firm. Section 4 concludes.

1 A Parameterization of the No-Arbitrage RIM

This section introduces notation to interpret no-arbitrage valuation. This is accomplished by

specifying a tractable “pricing kernel,” which we parameterize as a log-affine form in Section
1.1. Similar to Feltham and Ohlson (1999)’s analysis, we bring the RIM into this framework in
Section 1.2.

1.1 A Log-Affine Pricing Kernel

The assumption of no-arbitrage, together with some technical conditions (see Harrison and
Kreps (1979)), guarantees the existence of a random process which prices all assets in the econ-
omy. This random process is called a “pricing kernel,” which we denote by πt+1 . The pricing
kernel is unique if markets are complete, meaning that the number of securities is sufficient to

4
insure against all possible sources of risk in the economy. While the assumption of no-arbitrage
guarantees the existence of πt+1 , no-arbitrage gives no information, however, about the true
functional form of πt+1 , only that some πt+1 exists.5
The pricing kernel πt+1 relates the price of a security today with its payoffs in the next
period. For any asset:6

Pt = Et [πt+1 Zt+1 ] , (1)

where Pt is the price of an asset, and Zt+1 are its payoffs at time t + 1. Note that if Zt+1 = 1, a
unit payoff, then Pt is the price of a one period bond. In the case of a stock St , the price of the
stock is related to its dividends δt by:

St = Et [πt+1 (St+1 + δt+1 )] . (2)

Time-varying interest rates and risk aversion are captured by the pricing kernel πt+1 . To
separate out the role of the short rate rt and risk in πt+1 , we introduce another random variable
ξt+1 which we define as:

πt+1
ξt+1 = . (3)
Et (πt+1 )

Recall that Et (πt+1 ) = exp(−rt ) is the price of the one-period risk-free bond at time t. This
enables us to rewrite equation (2) as:

St = Et [exp(−rt )ξt+1 (St+1 + δt+1 )] , (4)

where πt+1 = exp(−rt )ξt+1 . Equation (4) separately decomposes the role of the pricing kernel
into the short rate process rt , and ξt+1 , which allows for explicit adjustments for risk aversion
in the no-arbitrage environment.
The most general parameterization of no-arbitrage is determined by (i) ξt+1 > 0 and (ii)
Et (ξt+1 ) = 1. We now assume a parameterization for ξt+1 . Specifically, we assume that ξt+1 is

5
log-normally distributed:

1
ξt+1 = exp − γ ′ σt σt′ γ + γ ′ σt ǫt+1 .


(5)
2

where ǫt+1 is a K × 1 vector IID N(0,I), γ is a K × 1 vector and σt is a K × K matrix. The

subscript t on σt indicates that σt may be a function of time t information, and hence may
vary through time. The errors ǫt+1 represent all shocks to K driving variables in the economy.
For now, these driving variables remain unspecified, but they can be any variable which affects
prices in the economy. In Section 2 we specify the driving variables to be accounting variables
in levels and in Section 3 we specify the driving variables to be ratios.
The log-normal pricing kernel in equation (5) is a valid pricing kernel. First, it satisfies strict
positivity as exp(·) > 0. Second, by property of the log-normal distribution (see Appendix),
Et [ξt+1 ] = 1.7
We refer to γ as the price of risk which captures the risk aversion of agents. The following
example illustrates the role of γ. Let time t = 0, and suppose, without loss of generality, the
prevailing short rate r0 = 0. There is only one source of risk in the economy, say from earnings,
so K = 1, and ǫ1 ∼ N(0,1). We would like to price a claim which has a payoff σ0 ǫ1 , that is the
security’s payoff is the unanticipated factor shock. The price of this security P0 is given by:

P0 = E0 [ξ1 σ0 ǫ1 ]
 
1 2 2

= E0 exp − γ σ0 + γσ0 ǫ1 σ0 ǫ1
2
= γσ02 . (6)

The last equality can be derived using a lemma in the Appendix. Under the case of risk neutral-
ity, γ = 0 and the price of the security is zero. This is expected, because under risk neutrality
the price of the security is just the expected value of the security’s payoffs, which is zero. Under
risk aversion γ 6= 0, and risk-averse agents must be compensated to take on a risk with a zero
expected value payoff. If γ < 0, the price of the security is negative, which is less than the
risk-neutral price of zero. That is, risk-averse agents must be paid to bear this risk. The greater

6
the degree of risk aversion, the more negative the value of γ, and the more risk averse agents
must be compensated for bearing risk.
Another interpretation of the role of γ is to look at the role risk plays in factor pricing.
Risk is related to the degree of correlation between shocks in the driving variables (factors) and
the negative of the pricing kernel. To make this concrete, again suppose that there is only one
driving variable in the economy, K = 1, time t = 0, and ǫ1 ∼ N(0,1). Using the lemma stated
in the Appendix, the conditional covariance of ǫ1 with −ξ1 at time t = 0 is given by:
 
1 2 2

cov0 (ǫ1 , −ξ1 ) = cov0 ǫ1 , − exp − γ σ0 + γσ0 ǫ1
2
= −γσ02 . (7)

If factor shocks have a correlation of zero, then their prices of risk are zero. If factor shocks
have non-zero correlation with the pricing kernel, then the degree of correlation is a measure
of the degree of risk associated with that variable. Only if γ is non-zero will the covariance of
the factor shock with the signed normalized pricing kernel ξ1 be non-zero. In particular, if γ is
negative, the covariance with −ξ1 is positive, which translates to non-zero risk aversion. As the
degree of risk aversion becomes greater, γ becomes more negative, and the correlation between
the driving factor and −ξ1 becomes more positive.

1.2 The RIM and No-Arbitrage Valuation

We now introduce notation to be used for applying the log-normal parameterization of the pric-
ing kernel to level and ratio analysis. This notation enables the role of ξt+1 to be used in evalu-
ating the expectations of future periods. We start by working with pricing kernel πt+1 notation,
and then we define notation so that expectations of future periods can be taken with respect to
ξt+1 . These transformed expectations are equivalent to the pricing kernel expectations.
The pricing kernel πt+1 relates the price of a stock to its future payoffs. Repeating equation
(2) we have:
St = Et [πt+1 (St+1 + δt+1 )] .

7
Iterating this forward and assuming transversality Et (πτ Sτ ) → 0 as τ → ∞, we get the Divi-
dend Discount Model (DDM):
"∞ #
X
St = Et πt+i δt+i . (8)
i=1

To relate the dividend process back to accounting variables, residual accounting makes the
clean surplus accounting assumption:

yt = yt−1 + xt − δt (9)

where yt is the book value of equity and xt represents net earnings at time t. This says that the
increase in the book value of equity comes from earnings less dividends paid.
Feltham and Ohlson (1999) develop a generalized RIM under time-varying interest rates
and risk aversion. They show that by using clean surplus accounting and the DDM in equation
(8), the equity value of a firm can be written as:8
"∞ #
X
St = yt + Et πt+i xat+i (10)
i=1

where xat are abnormal earnings:

xat = xt − exp(rt−1 ) − 1 yt−1

(11)

where rt−1 is the risk-free short rate from time t − 1 to time t, which can be stochastic. In the
basic RIM with constant short rates, the appropriate capital charge is the (constant) risk-free
rate. In a setting with stochastic short rates, the relevant capital charge component of abnormal
earnings is the riskless one-period interest rate applied to the book value at the start of the
period. Note that risk is embedded in the Feltham and Ohlson (1999) framework by using the
pricing kernel to take the future expectations of equation (10).
We shall now re-write equation (10) so that the expectation is taken with respect to ξt+1 ,

8
called the “risk-neutral measure”. Part of the role of ξt+1 in the pricing kernel is to separate the
effects of the risk-free rate and risk aversion in πt+1 . Repeating equation (4) we have:

St = Et [exp(−rt )ξt+1 (St+1 + δt+1 )] .

This expectation is taken under the “real measure” (the probability density function existing in
the real world). It can be rewritten as:

St = EQ
t [exp(−rt )(St+1 + δt+1 )] . (12)

This expectation is taken under the measure, or probability density function, Q. The measure
Q is called the risk-neutral measure because under Q the price of equity is just the expected
discounted payoffs of the security. This is also called the equivalent martingale measure.9
The pricing kernel πt+1 and ξt+1 are related recursively by:

t
πt Y
= ξj exp(−rj−1 ). (13)
π0 j=0

Using the risk-neutral measure we can equivalently rewrite the value of equity using the RIM
in equation (10), by substituting for the pricing kernel in the infinite sum, as:

∞
" i−1
! #
X Y
St = y t + EQ
t exp(−rt+j ) xat+i (14)
i=1 j=0

This is an equivalent representation of the RIM valuation equation.

The advantage of the parameterization of equation (14) is that it explicitly shows the inter-
Q
action of stochastic short rates (through the stochastic discount factor exp(−rt+j )) and risk
aversion (through the density ξt used to evaluate the expectation). If there is no risk, then γ = 0
so ξt = 1 and the real and risk-neutral measure coincide.
The special case of the RIM presented in Ohlson (1995) shows the simplifications which
arise using constant discount rates in the valuation problem. The following claim shows how

9
the RIM with constant interest rates and a constant ad hoc risk adjustment is a special case of
our general valuation methodology.

Claim 1.1 Assume that

1. short rates are constant, so exp(rt ) = Rf

s
2. the risk premium is constant, covt (rt+1 , −ξt ) = σ̄, where rs is the return on the stock
s
rt+1 ≡ (St+1 + δt+1 )/St

Then the traditional RIM of stock valuation holds:

∞
X
R−i Et xat+i
 
St = y t + (15)
i=1

where St is the value of firm equity, R = Rf + σ̄, yt is book value, and xat are abnormal
earnings:
xat = xt − (R − 1)yt−1

where xt represents firm earnings.

Compared to the traditional RIM, a pricing kernel methodology must be employed once
interest rates are stochastic and risk premiums are time-varying. Under this generalized setting
the valuation problem becomes much more complex. Since the discount factor R is constant
in Claim 1.1, it passes through the expectation operator in equation (14). When short rates are
Q
stochastic, the discount factor exp(−rt+j ) is time-varying and it cannot be passed through
the expectation operator. In Claim 1.1, risk aversion is captured by a constant risk premium
s
R − Rf > 0. If risk premiums are time-varying, covt (rt+1 , −ξt ) is no longer constant and ξt
plays a role in valuation. In a generalized setting of stochastic interest rates and risk aversion,
we must value equity using equation (14) where rt and ξt need to be parameterized.

10
2 An Affine Information Model

2.1 Discrete-Time Affine Processes

We now specify what are the driving variables in the economy and how they evolve over time.
The generalized RIM in equation (14) is able to take into account stochastic short rates and risk
aversion. However, equation (14) does not relate realized financial numbers with firm value,
since the values on the right hand side of equation (14) are forecasts. In this form, equation
(14) gives us a theoretical framework to study risk and return, but not a parametric form to
relate firm value with realized financial accounting reports. This section develops a parametric
no-arbitrage model of firm value when accounting information is given in dollar amounts.
The LIM of Ohlson (1995) and Feltham and Ohlson (1995) is a parametric formulation of
the RIM that expresses the value of a stock as a linear function of current accounting variables
specified in dollars rather than in terms of abstract future expectations. However, it is developed
under constant interest rates and risk-neutrality. Our aim is to extend the LIM to account for
stochastic interest rates and risk aversion. We specifically ask under what assumptions and
parameterizations linearity can survive under more general no-arbitrage conditions. We show it
is not always possible to incorporate risk aversion by adjusting a constant discount factor.
For concreteness, we assume the same driving variables in the economy as Feltham and
Ohlson (1995). Specifically, suppose the driving variables are denoted by Xt , and let:

Xt = (xat oat v1t v2t )′

where xat denoting abnormal earnings, oat operating assets, v1t and v2t “other” information at
time t. We also assume that Xt follows a discrete-time affine process (Duffie and Kan (1996)).
(The term “affine” refers to a constant plus linear term.)

Definition 2.1 A K × 1 vector Xt is said to follow a discrete-time affine process if:

Xt+1 = µ + AXt + σt ǫt+1 , (16)

11
where µ is an K × 1 vector, A and σt are K × K matrices and ǫt ∼ N(0,I). The conditional
mean Et (Xt+1 ) = µ + AXt is affine in Xt , and the conditional covariance is also affine in Xt
and is given by:

σt σt′ = h + H · Xt . (17)

The notation “·” represents a tensor product, and is interpreted as:

K
X
H · Xt ≡ Xtj H (j) (18)
j=1

where Xtj refers to the jth element of Xt . The K × K matrices h and H (j) are symmetric.

Discrete-time affine processes are an attractive parsimonious class of models which can
capture feedback (mean-reversion) and stochastic volatility. They capture feedback through the
companion matrix A in the conditional mean. Stochastic volatility depends on the level of the
variables through the tensor product in the conditional volatility. Hence, the variables in Xt
may be heteroskedastic (H (j) 6= 0 for some j). Homoskedasticity occurs as a special case when
H (j) = 0 ∀j and h 6= 0.
We now give two examples of affine processes. First, if there is no heteroskedasticity in the
conditional covariances (H (j) = 0 ∀j), the process reduces to a Vector Autoregression of first
order (VAR(1)). This is the process used in the Feltham-Ohlson LIM, where Xt follows the
following VAR(1):

Xt+1 = AXt + σǫt+1 (19)

where Xt = (xat oat v1t v2t )′ , and σσ ′ = h, where h is a constant symmetric matrix. In Feltham

12
and Ohlson (1995) the companion matrix A takes the form:
 
ω11 ω12 1 0 
 
 0 ω22 0 1 
A= . (20)
 
 0 0 γ1 0 
 
 
0 0 0 γ2

Note that the LIM extends to other more general forms of A, rather than to just this special
form. This is a special case of the discrete-time affine process, formed by setting µ = 0 and
specifying the covariance have no heteroskedasticity, so H (j) = 0 ∀j.
A second example of an affine process having heteroskedasticity is the Cox, Ingersoll and
Ross (CIR) (1985) model of term structure. If Xt = rt , the univariate short rate, then setting
h = 0 gives a discretized CIR model, where the variance is proportional to the level of the
interest rate:
Xt+1 = µ + AXt + σt ǫt+1

where σt2 = H (1) Xt , and H (1) is a positive scalar. Under a CIR model, the yield curve can
assume a variety of shapes including upward sloping, humped and downward sloping yield
curves. The stochastic movements of all interest rates are inferred from the short rate rt , once
the dynamic of rt in the CIR model is specified. We use the CIR term structure model to
incorporate time-varying interest rates.

2.2 An Affine Information Model (AIM)

Under the LIM, firm value is a linear function of accounting information. We state assumptions
under which linearity, or an affine form, can be maintained under a more general setting of
heteroskedasticity, risk aversion and time-varying interest rates.
For most of this section we assume that spot interest rates are independent of the accounting
variables. The interest rate process can be very general. At the end of the section we comment
on the case where interest rates and accounting variables are correlated.

13
2.2.1 An AIM under Independent Interest Rates

We let the K driving variables Xt = (xat vt′ )′ , with xat denoting abnormal earnings, and vt
a vector representing “other” information at time t. This formulation subsumes Feltham and
Ohlson (1995)’s parameterization, by letting vt = (oat v1t v2t )′ . We use the affine process of
Section 2.1:

Assumption 2.1 Xt is a K × 1 vector with first element abnormal earnings xat and other el-
ements representing other information at time t. Xt follows a discrete-time affine process as
defined by Definition 2.1.

Next we assume that interest rates follow a process that is consistent with no-arbitrage but
independent of accounting information:

Assumption 2.2 The economy is arbitrage-free, and spot interest rates rt follow a process
independent of Xt . The random variable defined in equation (3) ξt+1 is the product of two
(r) (X)
factors ξt+1 and ξt+1 :

(r) (X)
ξt+1 = ξt+1 · ξt+1 , (21)

(r) (X)
where ξt+1 and ξt+1 are independent and

(X) 1
ξt+1 = exp − γ ′ σt σt′ γ + γ ′ σt ǫt+1 ,


(22)
2

(r)
where ǫt+1 are the shocks of the process of Xt . The only requirement for the factor ξt+1 is that
it remains arbitrage-free.
{n}
The process for short rates leads to zero coupon bond prices Λt for period n:
"n−1 #
{n}
Y
Λt = EQ
t exp(−rt−j ) . (23)
j=0

One example of an admissible process for rt is a CIR model uncorrelated with Xt , but any term
structure uncorrelated with Xt is possible. In the case of the CIR model, zero coupon bond

14
prices are exponential affine functions of rt :

{n}
Λt = exp(a(n) + b(n)rt ),

where the a(n) and b(n) coefficients are closed-form (see Cox, Ingersoll and Ross (1985)).
Using the methodology presented in Section 1, an extended version of the Feltham-Ohlson
LIM holds, where we extend the LIM to an affine setting.

Proposition 2.1 Under Assumptions 2.1 and 2.2 the valuation function gt = St − yt can be
expressed as:

gt = αt + βt′ Xt , (24)

where the constant coefficient αt and the linear coefficient βt of the affine form are given by:

∞ ∞
!
(n+k)
X X
αt = Λt e′1 (A + H̄)n (µ + hγ)
n=0 k=1
∞
{n}
X
βt = Λt e′1 (A + H̄)n , (25)
n=1

where e1 is a K × 1 vector with first element 1 and the rest zero, A is the companion matrix of
the process for Xt in equation (16), and H̄ is a K × K matrix defined as:

K
(j)
X
H̄ij = Hik γk ,
k=1

(j)
where H̄ij is the element in the i-th row, j-th column of H̄, Hik is the element in the i-th row,
k-th column of H (j) (the K × K matrix in equation (18)) and γk is the k-th element of γ.

We make several comments on the affine form of valuation in Proposition 2.1. First, the en-
vironment is very general, and Proposition 2.1 shows that an affine form survives with stochas-
tic interest rates, risk aversion and heteroskedasticity in accounting information. However, this
affine form in accounting variables Xt depends crucially on the assumption that the interest

15
rate is orthogonal to accounting information. Given this restriction, spot interest rates may take
on any dynamic consistent with no-arbitrage. Second, when interest rates are not constant,
the affine coefficients αt and βt depend on time t through their dependence on the time t zero
{n}
coupon bond prices Λt . In particular, βt can be interpreted as the valuation of a perpetuity
whose payments are risk-adjusted future residual earnings. The discount factors on the perpe-
{n}
tuity are Λt . Without risk, H̄ = 0 and e′1 An Xt is the future expected abnormal earnings n
periods into the future, assuming abnormal earnings have zero mean as in the LIM. Under risk
aversion, H̄ 6= 0, and the future expected abnormal earnings in period t + n incorporate a risk
adjustment.
Finally, risk aversion contributes to both αt and βt . In the case of homoskedasticity (H (j) =
0 ∀j and h 6= 0), risk aversion gives rise only to state-independent risk premiums in αt . In this
case, H̄ = 0 so the effect of homoskedasticity enters through the action of the hγ term. Under
heteroskedasticity (H (j) 6= 0 for some j), risk aversion enters the βt terms. In this case, the risk
premium is state-dependent, through the non-zero H̄ term.

2.2.2 An AIM Under Constant Interest Rates

To focus on the effect of risk aversion and heteroskedasticity, we analyze the case where interest
{n}
rates are constant (rt = rf ∀t, Rf = exp(rf ) and Λt = Rf−n ). In this case the valuation
formula in Proposition 2.1 can be further simplified because all bond pricies are geometrically
related. We look at several examples, including the original LIM. We also determine when
linearity can be maintained under risk aversion and heteroskedasticity.

{n}
Corollary 2.1 In the case of constant interest rates, Λt = Rf−n , the coefficients are constant:
αt = α and βt = β and are given by:

Rf
α= e′ (Rf − A)−1 (µ + hγ)
Rf − 1 1
β = (Rf − (A + H̄)′ )−1 (A + H̄)′ e1 . (26)

16
Note the constant term α can arise through either a non-zero µ, or as a constant risk premium
through homoskedastic risk (h 6= 0).
Previously, applications of the RIM account for risk by using an ad hoc adjustment to a
constant discount factor. Corollary 2.1 shows that with constant interest rates, it may not be
possible to incorporate the effects of risk aversion in this way. The traditional RIM model in
Claim 1.1 incorporates risk aversion by setting the constant discount rate R to be R = Rf + σ̄,
where the spread σ̄ over the risk-free rate takes risk aversion into account:

∞
X
gtcdf R−i Et xat+i ,
 
= (27)
i=1

where “cdf” denotes constant discount factor. The following lemma shows that under a constant
discount rate R, the valuation function gtcdf is linear in Xt and does not have a constant term:

Lemma 2.1 Under Assumption 2.1 and µ = 0, gtcdf is a linear function of Xt .

Corollary 2.1 shows that the constant term is zero only if µ = 0 and γ = 0, or µ = 0 and h = 0.
If either condition is not satisfied, the value function gt cannot be described as in equation (27)
using a constant discount factor R, so we have the following corollary:

Corollary 2.2 It is not always possible, even under constant interest rates, to have a constant
discount factor.

Note that Feltham and Ohlson (1995) parameterize xat to have zero mean, but if some state
variables in Xt have non-zero mean then α is no longer zero. In this case, only the assumption
of risk neutrality guarantees the absence of a constant risk premium.
The following example shows how the AIM nests the LIM of Ohlson (1995) and Feltham
and Ohlson (1995) as a special case.

17
Example 2.1 The Feltham-Ohlson LIM. In the case of µ = 0, homoskedasticity (H (j) = 0 so
H̄ = 0) and risk neutrality, γ = 0, the AIM reduces to the Feltham-Ohlson LIM, where:

α = 0

β = (Rf − A′ )−1 A′ e1 .

We comment that under homoskedasticity, if µ 6= 0 then α is no longer zero but the linear
coefficient β in the traditional LIM remains unchanged.
In the next example we state an alternative set of conditions under which linearity can be
maintained under risk aversion.

Example 2.2 In the case µ = 0, h = 0 and risk aversion (γ 6= 0), the valuation function gt has
a linear form, where:

α = 0

β = (Rf − (A + H̄)′ )−1 (A + H̄)′ e1 .

Since we assume no homoskedastic risk (h = 0), Xt must exhibit heteroskedasticity (H (j) 6= 0

for some j) to be non-degenerate. In this case, the effect of risk aversion is absorbed into the lin-
ear coefficient β (through the H̄ term). That is, state-dependent risk (through heteroskedasticity
and risk aversion) gives rise to state-dependent risk premiums.

2.2.3 The Case of Correlated Interest Rates

When interest rates are correlated with the accounting variables, they must be included as a
state variable in Xt . That is, we re-define Xt as Xt = (rt xat vt′ )′ . Now, the random variable ξt+1
can no longer be factored into two independent terms, one depending on rt and one depending
only on accounting variables. In this case, an affine solution for gt is no longer possible, but we

18
can still find a functional form to relate gt with contemporaneous Xt . It can be shown that:

∞
X ′
gt = ea(i)+b(i) Xt (c(i) + d(i)′ Xt ),
i=1

where a(i) and c(i) are scalars, and b(i) and d(i) are vectors. The coefficients a(i), b(i), c(i)
and d(i) are constant and can be derived similarly to Proposition 3.1.10 This formula still relates
gt to observed values of Xt but the relation is now non-linear.

3 A Parametric Generalized Earnings Model

While the AIM or LIM presents firm value as a linear function of earnings, it is not convenient
for determining how growth in earnings or growth of the firm affects valuation. This section
uses a measure of firm profitability and growth in book, together with the stochastic short rate,
as driving factors to build a model of the price-to-book ratio. This allows direct inference of
how the rate of profitability and firm growth affect ratio valuation. We introduce the framework
in Section 3.1. Section 3.2 presents an example of a specialized process for the driving variables
to motivate how the model can be used in comparative statics. We derive the model in Section
3.3. Finally, Section 3.3 conducts comparative statics using the price-to-book formula.

3.1 Normalizing the RIM by Book Value

We start by repeating the no-arbitrage RIM in equation (14):

∞
" i−1
! #
X Y
St = y t + EQ
t exp(−rt+j ) xat+i .
i=1 j=0

To deal with accounting ratios, we normalize and divide each side by book value yt :
" ∞ i−1 !  #
St X Y x t+i y t+i−1
= 1 + EQ
t exp(−rt+j ) − (ert+i−1 − 1) (28)
yt i=1 j=0
y t+i−1 yt

We introduce some definitions to convert the accounting variables in levels to ratios or

19
growth rates:

Bt = St /yt

egt = yt /yt−1

rte = xt /yt−1

zt = rte − (exp(rt−1 ) − 1) (29)

In equation (29), Bt is the price-to-book ratio and gt is the growth rate in book value. The
next equation for rte is the accounting return on equity (earnings to book). Higher accounting
returns denote higher profitability. The variable zt is the accounting return in excess of the risk-
free rate, which we define as “the abnormal accounting return”. It is derived using the Feltham
and Ohlson (1999) definition of abnormal earnings, and then normalizing the abnormal earnings
by book value:

xat = xt − (exp(rt−1 ) − 1)yt−1

xat xt yt−1
zt ≡ = − (exp(rt−1 ) − 1) (30)
yt−1 yt−1 yt−1

Under mark-to-market accounting Bt = 1 and zt = 0. Under conservative accounting zt may

be non-zero.
Using the above definitions we can rewrite equation (28) as:
" ∞ i−1
! #
X Y
Bt = 1 + e−rt EQ
t zt+1 + e−(rt+j −gt+j ) zt+i
i=2 j=1
" ∞ i−1 ! #
X Y
−rt Q −(rt+j −gt+j )
= 1 + e Et e zt+i (31)
i=1 j=1

where we assume that the product term is equal to 1 if the index is negative. We assume that
the variables Bt , rt , zt and gt are stationary.11
Equation (31) rewrites the RIM, but expressing the price-to-book ratio as discounted ab-
normal returns. It still remains in the no-arbitrage setting of the RIM. We remark that the

20
normalized setting of equation (31) is still Miller-Modigliani (1961) consistent, so dividend
policy does not influence value because the original RIM in levels is Miller-Modigliani consis-
tent (see Ohlson (1995)). The driving variables behind the price-to-book are accounting returns
of earnings (rather than earnings for price levels), growth in book value (rather than book value
in levels), and accounting abnormal returns of earnings (rather than abnormal earnings). This
moves us from the setting of dollar amounts or price levels to ratios or growth rates.
Although equation (31) shows the price-to-book ratio to be a function of spot rates, abnormal
returns and growth, the numbers on the right hand side of equation (31) are forecasts. We
need a parameterization of the driving variables to relate the price-to-book to contemporaneous
accounting information. We do in this in the following section.

3.2 A Specialized Process for Interest Rates, Profitability and Growth

We specify the driving variables of the economy as the risk-free rate, abnormal returns of earn-
ings and growth in book value. Denote Xt = (rt zt gt )′ and assume that Xt follows a discrete-
time affine process, as defined in equations (16) and (17). To give a concrete example, suppose
Xt is given by:

√
rt+1 = µr + ρr rt + σr rt ǫ1t+1

zt+1 = µz + α1 rt + α2 zt + σz ǫ2t+1

gt+1 = µg + α3 rt + α4 zt + α5 gt + σg ǫ3t+1 (32)

with ǫt = (ǫ1t ǫ2t ǫ3t )′ ∼ N(0,I).

This structure may seem overly restrictive, but it is only intended as a simple example of
how feedback dynamics can be accomplished in the affine system. Assuming the errors are
independent implies that interest rates, abnormal earnings and growth in equity are subject
to independent shocks. We make this assumption so that we can analyze the effect of each
of the variables in Xt separately. The first equation is a discretized square root process (the
workhorse CIR model of term structure asset pricing) of the short rate. Through the variance

21
term, conditional volatility increases proportionally with the level of the interest rate.
The second equation is a Gaussian process which says that abnormal earnings are autocorre-
lated, and that the short rate Granger-causes abnormal returns. As interest rates go up, we expect
abnormal earnings to decrease (see Nissim and Penman (2000)). This is captured by a negative
α1 coefficient. Increasing interest rates decrease the discount factors applying in future periods
and decrease the price-to-book through a “denominator” effect. The predictability of account-
ing returns by interest rates is a “numerator” effect because it decreases cashflows in future
periods. Both the denominator and numerator effect are handled simultaneously by the dynam-
ics of companion matrix A in the discrete-time affine process (equation (16)). The coefficient
α2 captures the mean-reversion of profitability. As profitability becomes more mean-reverting
(or less persistent), α2 decreases.
The last equation parameterizes growth as a Gaussian process. The conditional mean of
growth in equity is a predictable function of past growth in equity, lagged short rates and
abnormal earnings. In particular, we would expect firm growth and profitability to be posi-
tively related; this would be captured by a positive α4 coefficient. The α5 coefficient reflects
mean-reversion or persistence of growth in book. As mean-reversion increases (or persistent
decreases), α5 decreases.
In terms of the notation of Section 1.1, this imposes the following structure on the com-
 
ρ 0 0
 r 
panion matrix A, and the matrices h and H driving the covariances: A = α1 α2 0 ,
 
 
α3 α4 α5
   
0 0 0 σ2 0 0
   r 
h = 0 σz 0 , H =  0 0 0, and H (2) = H (3) = 0.
 2  (1)  
   
2
0 0 σg 0 0 0
We examine several interesting effects on the price-to-book ratio from this parameteriza-
tion. In particular, we separately determine the effect of changing the parameters of the short
rate, abnormal return or growth in book on the price-to-book. Changing the mean reversion
of profitability or growth is accomplished by changing α2 and α5 . Risk-averse agents are af-
fected by changes in volatility, so σr , σz and σg affect valuation under risk aversion. To conduct

22
comparative statics, however, we need to derive an analytical expression for the price-to-book
ratio.

3.3 The Parametric Earnings Model

We now develop a non-linear formula for the price-to-book ratio. The case presented in the
previous section is only an example of a particular affine system, but our derivations presented
here apply to the most general affine model. This formula is like the LIM in that it relates
the price-to-book to observed accounting data rather than to forecasts in equation (31), but
with abnormal returns and growth rates a linear solution is no longer possible. The price-to-
book valuation formula is given in the following proposition, which evaluates the conditional
expectation of the infinite sum in equation (31) as a function only of time t information.

Proposition 3.1 Suppose the variables Xt = (rt zt gt )′ follow a discrete-affine process in Defi-
nition 2.1 and the default-adjusted pricing kernel πt+1 takes the log-linear form in equation (5).
Then the price-to-book ratio can be written only as a function of time t information:
" ∞ i−1 ! #
X Y
Bt = 1 + e−rt EQ
t e−(rt+i −gt+i) zt+i
i=1 j=0
∞
X ′
= 1 + e−rt ea(i)+b(i) Xt (c(i) + d(i)′ Xt ), (33)
i=1

where a(i) and c(i) are scalars, and b(i) and d(i) are 3 × 1 vectors. The constant coefficients

23
a(i), b(i), c(i) and d(i) are given by:

a(i) = a(i − 1) + (−e1 + e3 + b(i − 1))′ (µ + hγ)

1
+ (−e1 + e3 + b(i − 1))′ h(−e1 + e3 + b(i − 1))
2
b(i)′ Xt = (−e1 + e3 + b(i − 1))′ (AXt + H · Xt γ)
1
+ (−e1 + e3 + b(i − 1)′ (H · Xt )(−e1 + e3 + b(i − 1))
2
c(i) = c(i − 1) + d(i − 1)′ (µ + h(γ − e1 + e3 + b(i − 1)))

d(i)′ Xt = d(i − 1)′ (AXt + (H · Xt )(γ − e1 + e3 + b(i − 1))), (34)

where ej denotes a vector of zeros with a 1 in the jth place. The initial conditions are given by:

a(1) = 0

b(1)′ Xt = 0

c(1) = e′2 (µ + hγ)

d(1)′ Xt = e′2 ((H · Xt )γ + AXt ) (35)

Let us interpret the valuation model of Proposition 3.1 by comparing it to the LIM and the
AIM in Section 2. Although the LIM is given in dollar amounts and Proposition 3.1 gives
a model in ratio terms the two approaches are similar. First, the LIM relates how prices are
related to current observable accounting information rather than to forecasts. Proposition 3.1
does the same thing. It assumes a parameterization of accounting ratios and growth rates (Xt =
(rt zt gt )′ ) that enables the infinite sum involving expectations in equation (31) to be a function
only of observable Xt information. Second, the AIM presented in Section 2 incorporates time-
varying interest rates and risk aversion under heteroskedasticity. There, to maintain an affine
form rt must be uncorrelated with accounting variables. Here, rt is correlated with accounting
ratios and included as a state variable. Third, both the AIM and Proposition 3.1 are derived
under no-arbitrage framework. Finally, the LIM is closed-form and gives price as a linear
function of accounting information. Proposition 3.1 is also closed-form, but the solution is

24
non-linear.
We interpret the coefficients a(i), b(i), c(i) and d(i) in the closed-form solution as follows.
As long as transversality is satisfied, a(i) → −∞ when i → ∞ so the exponential tends to
zero, and the individual terms in the sum quickly become small. Practially, this means that the
sum in equation (33) can be evaluated very quickly without many terms.
The a(i) and b(i) coefficients result from the effect of the r − g discounting terms. In
equation (34), the terms quadratic in −e1 +e3 +b(i−1) are Jensen’s inequality terms. The terms
linear in −e1 + e3 + b(i − 1) involving γ result from risk aversion. The Jensen’s inequality terms
are always positive, while the risk premium terms can be negative if γ is negative. Increasing
the volatility of a factor increases the Jensen’s inequality terms, unless the risk premium terms
in the a(i) and b(i) recursions outweigh the effect of the Jensen’s inequality terms. This implies
that in a risk-neutral setting, increasing the volatility of a factor increases, ceteris paribus, the
terms in the exponential. This increases the price-to-book. However, in a risk-averse setting,
that is when γ < 0, the price-to-book can decrease with volatility. We illustrate this in the next
section.
The c(i) and d(i) coefficients value the abnormal return stream. Notice from the initial
conditions in equation (35) that the e2 vector pulls out only the abnormal returns terms. The
terms involving µ and AXt result from the action of the conditional mean of the process of Xt ,
and the terms involving γ are the risk premiums which act on the covariances. In the recursions
for c(i) and d(i) in equation (34) the Jensen term −e1 + e3 + b(i − 1) also enters, along with a
risk premium effect.
Equations (34) and (35) completely determine the response of the price-to-book (equation
(33)) in terms of parameters to the underlying process Xt . However, the action of the price to
the parameters is not immediately transparent due to the recursive nature of the coefficients in
the valuation equation. We now conduct a series of exercises in comparative statics to further
analyze the effects of parameter changes on the price-to-book.

25
3.4 Comparative Statics of the Price-to-Book

In this section we show how varying the parameters of the processes of the short rate, abnormal
returns and growth in book affect the price-to-book ratio of a firm, using the motivating example
in equation (32). In a previous version of this paper (Ang and Liu (1998)), we calibrated this
model to several individual stocks. We use the estimated parameters of Intel, from January 1975
to June 1997, as a basis for illustrating the comparative statics. These parameters are listed in
Table 1. We conduct our comparative statics at the base case of the sample mean for Xt over
the sample.12 In our plots, we show this baseline case as a circle.
In our comparative static exercises we wish to clarify the role of risk aversion. To do this,
only one factor, the growth in equity gt , is priced, and we set the prices of risk for the short
rate rt and abnormal returns zt to zero. The first assumption is close to reality, because term
structure estimations have found insignificant prices of interest rate risk. The action of the price
of risk of zt is very similar to the price of risk of gt , so we concentrate only on the action of one
price of risk. Hence we set γ = (0 0 γ3)′ where γ3 is the price of risk of growth in book.
For reference, we repeat the system Xt = (rt zt gt )′ here:

√
rt+1 = µr + ρr rt + σr rt ǫ1t+1

zt+1 = µz + α1 rt + α2 zt + σz ǫ2t+1

gt+1 = µg + α3 rt + α4 zt + α5 gt + σg ǫ3t+1 . (36)

3.4.1 Effect of the Short Rate on the Price-to-Book

We first examine the effect of the short-rate parameters. As µr increases, short rate levels
increase and future abnormal earnings are discounted back at higher rates. This decreases the
price-to-book. In Figure 1 we see the price-to-book as a function of the persistence ρr , and
the volatility σr . We first discuss the effect of ρr . As interest rates become more persistent,
the price-to-book decreases. Intuitively, increasing the persistence increases the unconditional
mean of the short rate. As cash flows are generally valued back at a higher discount factor, the
price-to-book falls.

26
 In Figure 1 the price-to-book increases when σr increases. At first glance, this would seem
counter-intuitive, for two reasons. First, we would expect agents to dislike volatility, so that
price-to-book would decrease when volatility increases. Second, the interest rate is bounded at
zero in the CIR model, and increasing the volatility of the short rate increases the unconditional
mean of interest rates. This implies that the discount factors which value back future abnormal
returns are higher. However, in our base-line case, agents are risk-neutral with respect to the
interest rate risk. The increase in the price-to-book when σr increases is purely due to a Jensen’s
inequality effect. Only if interest rate risk is priced is it possible to cancel the Jensen’s inequality
effect.

3.4.2 Effect of Profitability on the Price-to-Book

We turn now to the effect of profitability on the price-to-book. As expected, increasing µz

increases the price-to-book as a higher µz implies greater profitability. The price-to-book ratio
also increases as the predictability coefficient α1 increases, if zt is positive. Economically,
this occurs because increasing abnormal returns causes cashflows in future periods to increase,
and hence this increases the price-to-book ratio. Correspondingly, decreasing α1 decreases
abnormal returns. The effect is opposite for negative levels of zt .
The top panel of Figure 2 shows the effect of altering the persistence (α2 ) of abnormal re-
turns. Increasing the persistence of zt , or decreasing the mean-reversion, increases the price-to-
book. The statistical interpretation is as follows. The unconditional mean of abnormal returns
rises as the persistence rises. Higher average abnormal returns then imply higher price-to-book.
Economically, we expect a firm’s profitability to mean-revert within and across industries (see
Fama and French (2000)). Higher mean reversion (or lower persistence of abnormal earnings)
means that high relative earnings in the short term persist for fewer periods. This lower prof-
itability decreases the price-to-book ratio.
The bottom panel of Figure 2 presents the price-to-book as a function of abnormal return
volatility σz . In our parameterization, increasing the volatility of abnormal returns increases
the price-to-book. This is because the price of risk of abnormal returns γ2 is zero, so agents

27
are risk-neutral with respect to profitability. The Jensen’s effect causes the price-to-book to
increase when σz increases. However, if γ2 is negative and agents are risk-averse with respect to
abnormal returns, then it may be possible for the price-to-book to fall as volatility of profitability
increases.

3.4.3 Effect of Growth on the Price-to-Book

Any parameters which increase the growth in book increase the price-to-book ratio. For exam-
ple, increasing µg , α3 or α4 when zt is positive increase the price-to-book because each param-
eter raises growth in book. The intuition is that increasing growth increases the likelihood of
higher cashflows in future periods.
The parameter α5 captures the persistence, or mean-reversion, of firm growth. The top panel
of Figure 3 shows that increasing the persistence of the growth in book increases the price-to-
book. This effect is similar to increasing the persistence of zt (α2 ), since the unconditional
mean of gt rises as the persistence of gt increases. Higher persistence implies that firm growth
mean-reverts to an industry or market average at a faster rate. Hence the firm has fewer periods
to enjoy the benefits of relatively higher growth.
The bottom panel of Figure 3 shows the effect of a decreasing price-to-book with increas-
ing volatility in growth in equity (σg ). This is in line with intuition: we would expect, ce-
teris paribus, normal risk-averse investors to lower their valuations the greater the volatility in
growth. We obtain this result because there is a large non-zero price of risk on growth in equity
γ3 and this causes the price-to-book to decrease as volatility increases. In this case, in addition
to the Jensen’s inequality effect (which increases with σg ), there is also a risk aversion effect
which counteracts the Jensen’s inequality effect.
Finally, Figure 4 shows the effect of risk aversion on the price-to-book. In Figure 4, for risk
aversion levels below γ3 = −15 the price-to-book is very flat, but as investors approach risk
neutrality the price-to-book becomes very large.

28
4 Conclusion

This paper introduces a methodology that incorporates stochastic interest rates, risk aversion
and heteroskedasticity into the Residual Income Model (RIM). We provide two applications
of the methodology. First, in applying the methodology to dollar amounts, we show that the
Ohlson (1995) and Feltham and Ohlson (1995) Linear Information Model generalizes to an
affine (constant plus linear terms) model under time-varying interest rates and risk aversion.
The processes for accounting information may be heteroskedastic. The interest rate process is
very general but is assumed to be uncorrelated with the processes governing the evolution of
accounting information.
Second, in applying the methodology to ratio dynamics, we provide a non-linear closed-
form formula for the price-to-book ratio in terms of stochastic short rates, profitability and firm
growth. In comparative static exercises, increasing the growth in book increases the price-to-
book ratio and increasing the mean-reversion of profitability or growth decreases the price-to-
book. The effect of interest rates on the price-to-book depends on the degree of risk aversion.
Under sufficiently high risk aversion, increasing the mean or volatility of the short rate decreases
the price-to-book ratio.

29
Appendix Proofs

We start by stating the following Lemma which gives the expectation of the product of a normal
with the exponential of a normal, which can be proved by evaluating the expectation. This is
used in some of the proofs below.

A Lemma

Lemma A.1 If Y is distributed as a K-variate normal with Y ∼ N(0, Σ), and γ and δ are
K × 1 constant vectors then

′ 1 ′
E(δ ′ Y eγ Y ) = δ ′ Σγe 2 γ Σγ (A-1)

B Proof of Claim 1.1

Starting from the relation St = Et [πt+1 (St+1 + δt+1 )] we can substitute for πt+1 = Rf−1 ξt+1
using the assumptions about a flat term structure to get:

St = Rf−1 Et ξt+1 (St+1 + δt+1 ) .

(B-1)

s s
Using the definition of the return of the stock rt+1 ≡ (St+1 + δt+1 )/St and covt (rt+1 , −ξt+1 ) =
s
σ̄, we can write Et (rt+1 ) = Rf + σ̄ ≡ R, with R a constant.
In this case:

1
St = Et (St+1 + δt+1 ), (B-2)
R

and iterating this equation forward and assuming transversality we obtain:

∞
X
St = R−i Et (δt+i ). (B-3)
i=1

30
 Then, as in Ohlson (1995), abnormal earnings become xat = xt − (R − 1)yt−1 , and substi-
tuting for dividends in the previous equation yields:
"∞ # ∞
X X
St = yt + Et R−i xat+i = yt + R−i Et (xat+i ) (B-4)
i=1 i=1

because the telescoping sum collapses, assuming R−τ Et (yt+τ ) → 0 as τ → ∞.

C Proof of Proposition 2.1

From equation (14) we can write:

∞
" i−1
! #
X Y
gt = EQ
t exp(−rt+j ) xat+i (C-1)
i=1 j=0

where gt = St − yt . An equivalent statement is:

gt = Et exp(−rt )ξt+1 (gt+1 + xat+1 )

 

exp(rt )gt = Et (gt+1 + xat+1 ) + covt (gt+1 + xat+1 , ξt+1 ), (C-2)

noting that Et (ξt+1 ) = 1.

Next, conjecture an affine solution for gt which has the form:

gt = αt + βt Xt ,

{n}
where αt and βt can depend on zero coupon bond prices Λt , but not the state variables Xt .
We can rewrite equation (C-2) as:

{1}
(Λt )−1 (αt + βt′ Xt ) = Et (αt+1 + (βt+1 + e1 )′ Xt+1 )covt ((βt+1 + e1 )′ σt ǫt+1 , ξt+1 ). (C-3)

{1}
Here note that exp(rt ) = (Λt )−1 . Using Lemma A.1 and evaluating the conditional mean of

31
Xt+1 gives:

{1} {1}
(Λt )−1 αt + (Λt )−1 βt′ Xt = Et (αt+1 ) + (Et (βt+1 ) + e1 )′ µ + (Et (βt+1 ) + e1 )′ AXt

+(Et (βt+1 ) + e1 )′ (h + H · Xt )γ, (C-4)

after substituting σt σt′ = h + H · Xt . Equating the coefficients of Xt we have:

{1}
(Λt )−1 αt = Et (αt+1 ) + (Et (βt+1 ) + e1 )′ (µ + hγ)
{1}
(Λt )−1 βt = (A + H̄)′ (Et (βt+1 ) + e1 ). (C-5)

One can show by substitution that the above equations are solved by setting:

∞ ∞
!
(n+k)
X X
αt = Λt e′1 (A + H̄)n (µ + hγ)
n=0 k=1
∞
{n}
X
βt = Λt e′1 (A + H̄)n . (C-6)
n=1

D Proof of Lemma 2.1

Since R is a constant, xat = xt − (R − 1)yt−1 is a linear function of the state variables Xt . That
is, xat = L′ Xt for some constant vector L. Then, Et (xat+i ) = L′ Ai Xt is a linear function of Xt .
Hence:

∞
X
gt = R−i L′ Ai Xt = L′ A(R − A)−1 Xt , (D-1)
i=1

which is a linear function of Xt , that is, there is no constant term.

32
E Proof of Proposition 3.1

To prove Proposition 3.1 we show that a single term in the infinite sum at horizon T sum takes
the form:
" T −1
! #
Y
EQ
′
t e−(rt+i −gt+i) zt+T = ea(T )+b(T ) X(t) (c(T ) + d(T )′ X(t)) (E-1)
i=1

We show that the coefficients a(T ), b(T ), c(T ), and d(T ) are given by the Ricatti differ-
ence equations in equation (34) with initial conditions in equation (35). Once this is shown,
Proposition 3.1 follows immediately from evaluating each individual term in the infinite sum.
We prove equation (E-1) by induction. Assume validity of equation (E-1) for T . We show
that the equation holds for T + 1. Using iterative expectations we can write:
" T
! # " " T −1
! ##
Y Y
EQ
t e−(rt+i −gt+i) zt+T +1 =EQ
t e−(rt+1 −gt+1) EQ
t+1 e−(rt+1+i −gt+1+i ) zt+T +1
i=1 i=1
h i
Q −(e1 −e3 )′ Xt+1 a(T )+b(T )′ Xt+1 ′
=Et e e (c(T ) + d(T ) Xt+1 ) .

(E-2)

The induction assumption is used in the last equality. We then observe that Xt+1 under Q
satisfies (this is the discrete-time version of Girsanov’s theorem):

Xt+1 = µ + AXt + σt σt′ γ + σt ǫQ

t+1 , (E-3)

33
where ǫQ
t+1 is a (mean-zero) standard normal random variable under Q. Then:

" T
! #
Y
EQ
t e−(rt+i −gt+i) zt+T +1
i=1
′ ′
=ea(T )+(−e1 +e3 +b(T )) (µ+AXt +σt σt γ)
h i
(γ−(e1 −e3 )+b(T ))′ σt ǫQ
× EQ t e t+1 (c(T ) + d(T )′ (µ + AX + σ σ ′ γ) + d(T )′ σ ǫ
t t t
Q
t t+1 )
′ ′ 1 ′ ′
=ea(T )+(−e1 +e3 +b(T )) (σt σt γ+µ+AXt ) e 2 (−e1 +e3 +b(T )) σσt (−e1 +e3 +b(T ))
 
′ ′ ′ ′
× c(T ) + d(T ) (µ + AXt + σt σt γ) + d(T ) σt σt (−e1 + e3 + b(T )) .

(E-4)

The last equality is obtained by employing Lemma A.1. Equating coefficients gives us the
result. To obtain the initial conditions we directly evaluate EQ
t (zt+1 ) using Lemma A.1.

34
Notes
1
Penman (1991) argues that the RIM has implications for ratio analysis, particularly for the
price-to-book and price-to-earnings ratios.

2
Zhang (1998) demonstrates that for a given level of earnings the relationship between book
and equity value is ambiguous.

3
See Fama and French (2000), and Ou and Penman (1989), among others.

4
Nissim and Penman (2000) demonstrate that interest rates have predictive power to forecast
future accounting returns.

5
Feltham and Ohlson (1999) use the notation ξt+1 for the pricing kernel. We use the no-
tation πt+1 as we save ξt+1 to describe a component of πt+1 more in line with standard asset
pricing notation. In certain situations the pricing kernel has a known form, such as the case of
consumption-based asset pricing (Lucas (1978)), or with complete markets (Black and Scholes
(1973)). Both of these functional forms are not valid here because we do not specify a represen-
tative agent who has utility over consumption and because markets are incomplete with respect
to accounting information (market prices are not available for earnings or book values).

6
An equivalent representation of the pricing kernel is:

mt Pt = Et (mt+1 Zt+1 )

where mt+1 satisfies the equation:

mt+1
πt+1 = .
mt

See Harrison and Kreps (1979).

7
Although this parameterization may look restrictive, it is very flexible. The strict positivity
requirement is almost equivalent to specifying ξt+1 = exp(f (ǫt+1 )) for some function f (·). The
sources of risk arise from the shocks of the driving variables in the economy, which are ǫt+1 .

35
For small variability in the driving factors, we can approximate the variable part of f (ǫt+1 ) by a
first-order approximation γ ′ σt ǫt+1 . This leads to the form of the pricing kernel in equation (5).
This form of pricing kernel has been used in many financial applications including Duffie and
Liu (2001), Bakshi and Chen (2001), Bekaert and Grenadier (2001), and others.

8
Strictly speaking Feltham and Ohlson (1999) consider a countable state space and a finite
horizon. This can be generalized to most uncountable state spaces with some technical as-
sumptions (see Harrison and Kreps (1979)), and applied to an infinite horizon sum by assuming
transversality.

9
The process ξt+1 converts the risk-neutral measure to the real measure. By definition the
following relationship holds between the real measure and Q: EQ
 
t Zt+1 ] = Et ξt+1 Zt+1 ] for

any t + 1 measurable random variable Zt+1 . In technical terms, the process ξt+1 is a Radon-
Nikodym derivative of the real measure with respect to the risk-neutral measure Q. See Harrison
and Kreps (1979).

10
We omit this proof as the derivation is very similar to the derivation of Proposition 3.1.

11
Note that gt does not appear in the first term of the infinite sum. The reason is that at time
t + 1, the opportunity costs of the firm are (ert − 1)yt, which are known at time t. At time
t + 2 the opportunity costs of the firm are based on yt+1 = yt egt . This lag in the timing of gt
disappears if the model is formulated in continuous time. See Ang and Liu (1998).

12
The sample mean of Xt is X̄t = (0.0817/4, 0.1366, 0.2773)′ where the interest rate is
annualized, but used as quarterly.

36
References

Ang, A., and J. Liu. (1998). “A Generalized Earnings Model of Stock Valuation,” Stanford
Research Paper 1491.

Bakshi, G., and Z. Chen. (2001). “Stock Valuation in Dynamic Economies,” forthcoming
Journal of Finance

Bekaert, G., and S. Grenadier. (2001). “Stock and Bond Pricing in an Affine Economy,”
working paper, Columbia University.

Black, F., and M. Scholes. (1973). “The Pricing of Options and Corporate Liabilities,” Journal
of Political Economy 81, 637-654.

Cox, J. S., J. Ingersoll, and S. Ross. (1985). “A Theory of the Term Structure of Interest
Rates,” Econometrica 53, 385-408.

Duffie, D., and R. Kan. (1996). “A Yield-Factor Model of Interest Rates,” Mathematical
Finance 6, 4, 379-406.

Duffie, D., and J. Liu. (2001). “Floating-Fixed Credit Spreads,” Financial Analysts Journal 57,
3, 76-87.

Fama, E. F., and K. R. French. (2000). “Forecasting Profitability and Earnings,” Journal of
Business 73, 2, 161-175.

Feltham, G. A., and J. A. Ohlson. (1995). “Valuation and Clean Surplus accounting for
Operating and Financial Activities,” Contemporary Accounting Research 11, 2, 689-731.

Feltham, G. A., and J. A. Ohlson. (1999). “Residual Earnings Valuation with Risk and
Stochastic Interest Rates,” The Accounting Review 74, 2, 165-183.

Harrison, J. M., and D. M. Kreps. (1979). “Martingales and Arbitrage in Multiperiod

Securities Markets,” Journal of Economic Theory 2, 3, 381-408.

37
Lucas, R. (1978). ”Asset Prices in an Exchange Economy”, Econometrica 46, 6, 1429-45.

Miller, M., and F. Modigliani. (1961). “Dividend Policy, Growth and the Valuation of Shares,”
Journal of Business 34, 4, 411-433.

Nissim, D., and S. H. Penman. (2000). “The Empirical Relationship Between Interest Rates
and Accounting Rates of Return,” working paper, Columbia University.

Ohlson, J. A. (1990). “A Synthesis of Security Valuation Theory and the Role of Dividends,
Cash Flows, and Earnings,” Journal of Contemporary Accounting Research 6, 2, 648-676.

Ohlson, J. A. (1995). “Earnings, Book Values and Dividends in Equity Valuation,” Journal of
Contemporary Accounting Research 11, 2, 661-687.

Ou, J. A., and S. H. Penman. (1989). “Accounting Measurement, Price-Earnings Ratio, and the
Information Content of Security Prices,” Journal of Accounting Research 27, 0, 111-44.

Penman, S. H. (1991). “An Evaluation of Accounting Rate-of-Return,” Journal of Accounting,

Auditing and Finance 6, 2, 233-255.

Zhang, G. (1998). “Accounting Information, Capital Investment Decisions, and Equity

Valuation: Theory and Empirical Implications,” working paper, Hong Kong University of
Science and Technology.

38
 Table 1: Parameters for Comparative Statics

Baseline Parameters
µr 0.0054
ρr 0.7548
σr 0.0374
µz 0.0263
α1 -0.8695
α2 0.7720
σz 0.0176
µg 0.0247
α3 1.1829
α4 0.6008
α5 -0.0105
σg 0.0698
γ3 -13.1982

These parameters are the base-line case for the comparative

statics exercises. The equations for the processes are given
by equation (32). We set γ1 = γ2 = 0.

39
 Changing mean−reversion of interest rate

12

10

8
price to book

0.7 0.72 0.74 0.76 0.78 0.8 0.82 0.84 0.86 0.88 0.9
ρ
r

Changing volatility of interest rate

5.1

5
price to book

4.9

4.8

4.7

4.6

0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05

σ
r

Figure 1: Comparative Statics for Short Rate Parameters

40
 Changing α2 of abnormal returns

6.5

5.5
price to book

4.5

3.5

0.7 0.71 0.72 0.73 0.74 0.75 0.76 0.77 0.78 0.79 0.8
α
2

Changing σz of abnormal returns

6.5
price to book

5.5

0.01 0.015 0.02 0.025 0.03 0.035 0.04

σ
z

Figure 2: Comparative Statics for Abnormal Return Parameters

41
 Changing α5 of growth in equity

5.6

5.4

5.2
price to book

4.8

4.6

−0.1 −0.08 −0.06 −0.04 −0.02 0 0.02 0.04 0.06 0.08 0.1
α
5

Changing σg of growth in equity

7
price to book

0.065 0.07 0.075 0.08

σ
g

Figure 3: Comparative Statics for Growth in Equity Parameters

42
 Changing γ3 growth in equity risk

35

30

25
price to book

20

15

10

−20 −19 −18 −17 −16 −15 −14 −13 −12 −11 −10
γ
3

Figure 4: Comparative Statics for Price of Risk of Growth in Equity

43
Figure Legends

Figure 1 plots the price-to-book as a function of persistence of the short rate ρr (top plot), and
as a function of the short-rate volatility σr (bottom plot). The base-line case is shown as a circle.

Figure 2 plots the price-to-book as a function of the persistence of abnormal returns α2 (top
plot), and the volatility of abnormal returns σz (bottom plot). The base-line case is shown as a
circle.

Figure 3 plots the price-to-book as a function of tje persistence of growth in equity α5 (top
plot), and the volatility of growth in equity σg (bottom plot). The base-line case is shown as a
circle.

Figure 4 plots the price-to-book as a function of the price of risk of growth in equity (γ3 ). The
base-line case is shown as a circle.

44