Professional Documents
Culture Documents
Returns
Abstract
We study equilibrium asset prices in a model where investors are loss averse, paying
particular attention to what they are loss averse about. We consider two possibilities,
which correspond to dierent assumptions about how people do mental accounting, or
about how they evaluate their investment performance. In one case, investors track
their performance stock by stock, and are loss averse over individual stock
uctuations.
In the other case, they measure their performance at the portfolio level, and are loss
averse only over portfolio
uctuations.
We nd that loss aversion over individual stock
uctuations is a helpful ingredient
for explaining a wide range of empirical facts, both in the time series and in the
cross-section. In simulated data, individual stock returns have a high mean, excess
volatility, and are slightly predictable in the time series; moreover, there are large
\value" and \size" premia in the cross-section. The case where investors are loss averse
over portfolio
uctuations, although normatively more appealing, is less successful
in explaining the facts: individual returns are insuÆciently volatile and excessively
correlated, while the premia for value and size largely disappear.
We thank John Campbell, David Hirshleifer, Andrei Shleifer, Jeremy Stein, Richard Thaler, and seminar
participants at Harvard University and the Kellogg Conference on Market Frictions and Behavioral Finance
for helpful comments on an earlier draft. Further comments welcome at nick.barberis@gsb.uchicago.edu or
mhuang@leland.stanford.edu.
1
1 Introduction
Over the past two decades, researchers analyzing the structure of individual stock returns
have uncovered a wide range of phenomena, both in the time series and the cross-section. In
the time series, individual stock returns have a high mean, excess volatility, and are slightly
predictable using lagged variables. In the cross-section, there is a substantial \value" pre-
mium, in that stocks with low ratios of price to fundamentals have higher average returns;
and also a \size" premium, in that stocks with lower market capitalizations have higher
average returns. These ndings have attracted a good deal of attention from nance the-
1
orists; it has proved something of a challenge, though, to explain both the time series and
cross-sectional eects in the context of an equilibrium model where investors maximize a
clearly specied utility function. 2
In this paper, we study equilibrium asset prices in models that are based on insights
from experimental work that psychologists have done in their eort to understand how
people evaluate risky gambles. Our aim is to investigate to what extent these insights can
be useful to nancial economists trying to explain the behavior of individual stock returns.
The two ideas we focus on are loss aversion and narrow framing.
Loss aversion is a feature of Kahneman and Tversky's (1979) descriptive model of decision
making under risk, prospect theory, which uses experimental evidence to argue that people
get utility from changes in wealth, rather than from absolute levels. The specic nding
known as loss aversion is that people are more sensitive to losses than to gains. Since our
framework is explicitly intertemporal, we also make use of more recent evidence on dynamic
aspects of loss aversion. This evidence suggests that the degree of loss aversion depends on
prior gains and losses: a loss that comes after prior gains is less painful than usual, because
it is cushioned by those earlier gains. On the other hand, a loss that comes after other losses
is more painful than usual: after being burnt by the rst loss, people become more sensitive
to additional setbacks.
In order to study the implications of loss aversion, we need to make an assumption on
what investors are loss averse about. Are they loss averse only over changes in their total
wealth ? Or do they take a narrower view, and also experience loss aversion over changes in the
value of one part of their total wealth, namely their stock portfolio ? Or do they take an even
narrower perspective, and feel loss aversion over
uctuations in the value of individual stocks
1 The size premium was originally noted by Banz (1981), and the value premium by Rosenberg, Reid and
Lanstein (1985); Fama and French (1992) provide more recent evidence. Vuolteenaho (1999) documents the
excess volatility and time series predictability of rm level stock returns.
2 Fama and French (1996) demonstrate that a three-factor asset pricing model can explain the value and
size premia, but their model has not yet been shown to result from an explicit economic equilibrium with
utility maximizing agents.
2
that they own? Put dierently, how do investors do their \mental accounting"? Mental
accounting, a term coined by Thaler (1980), refers to the system people use to summarize
and think about their nancial transactions. In our context, it is the system people use to
make their investment decisions ex-ante and to evaluate their investment outcomes ex-post.
Numerous experimental studies by psychologists suggest that an important feature of
mental accounting is narrow framing, the idea that people take too narrow a perspective
when making decisions. Kahneman and Lovallo (1993) summarize the evidence as suggesting
that \people tend to consider decision problems one at a time, often isolating the current
problem from other choices that may be pending, as well as from future opportunities to
make similar decisions."
In our analysis, we study the equilibrium behavior of individual stock returns when
investors are subject to narrow framing in their mental accounting. We investigate two
kinds of narrow framing, one narrower than the other. By comparing equilibrium returns
under each kind of framing, we can isolate the eect of mental accounting on prices.
In the rst economy we consider, investors care about consumption, but over and above
that, they are loss averse over individual stock
uctuations. In other words, they evalu-
ate their investment performance stock by stock, and get utility from each stock's gain or
loss; moreover, how painful a loss on a particular stock is, depends on that stock's prior
performance. We refer to this as \individual stock accounting", a severe form of narrow
framing.
In the second economy, investors care about consumption, but over and above that, they
are loss averse over
uctuations in the value of their overall portfolio of stocks. In other
words, they evaluate their investment performance at the portfolio level, and get utility from
gains and losses in portfolio value; moreover, how painful a drop in portfolio value is, depends
on the portfolio's prior performance. We call this \portfolio accounting", a form of narrow
framing, although not as severe as individual stock accounting.
In our rst set of results, we show that individual stock accounting can be a helpful
ingredient for understanding a wide range of empirical phenomena. In equilibrium, under
this form of mental accounting, individual stock returns have a high mean, are more volatile
than their underlying cash
ows and are slightly predictable in the time series. In the cross-
section, there is a large \value" premium: stocks with low price-dividend ratios have much
higher average returns than stocks with high price-dividend ratios; and a substantial \size"
premium: stocks with lower prices have higher average returns. At the same time, the model
matches empirical features of aggregate stock returns: in equilibrium, they have a high mean,
excess volatility and are moderately predictable in the time series.
Second, we nd that the investor's system of mental accounting aects asset prices in a
signicant way. As we broaden the investor's decision frame from individual stock accounting
3
to portfolio accounting, the equilibrium behavior of individual stock returns changes consid-
erably: their mean value falls, they become much less volatile, and much more correlated
with each other. Moreover, the value and size premia in the cross-section disappear.
To understand where our results come from, consider rst the case of individual stock
accounting. Many of the eects here derive from a single source, namely a discount rate
for individual stocks that changes as a function of the stock's past performance. If a stock
has had good recent performance, the investor gets utility from this gain, and becomes less
concerned about future losses on the stock because any losses will be cushioned by the prior
gains. In eect, the investor thinks of the stock as less risky and discounts its future cash
ows
at a lower rate. Conversely, if one of his stocks performs dismally, he nds this painful and
becomes more sensitive to the possibility of further losses on the stock: in eect, he views
the stock as riskier and raises its discount rate.
This changing discount rate has many implications. It gives individual stocks some time
series predictability: a lower discount rate pushes up the price-dividend ratio and leads
to lower subsequent returns, which means that the lagged price-dividend ratio can predict
returns. It makes stock returns more volatile than underlying cash
ows: a high cash
ow
pushes the stock price up, but this prior gain also lowers the discount rate on the stock,
pushing the stock price still higher. It also generates value and size premia in the cross-
section : stocks with high
D ratios (growth stocks) or high P values (large stocks) have often
P
done well in the past, accumulating prior gains for the investor who then views them as less
risky and requires a lower average return. Stocks with low DP ratios (value stocks) or low
P values (small stocks) have often had dismal prior performance, burning the investor, who
now thinks of them as riskier, and charges a higher risk premium. Finally, since the investor
is loss averse over individual stock
uctuations, he hates the frequent losses that individual
stocks often produce, and charges a high average return as compensation.
The reason the results are dierent under portfolio accounting is that in this case, changes
in discount rates on stocks are driven by
uctuations in the value of the overall portfolio:
when the portfolio does well, the investor is less concerned about losses on any of the stocks
that he holds, since the prior portfolio gain will cushion any such losses. Eectively, he
views all stocks as less risky. Discount rates on all stocks therefore go down simultaneously.
Conversely, discount rates on all stocks go up after a prior portfolio loss.
This discount rate behavior is the key to many of the portfolio accounting results. Stock
returns are less volatile here than under individual stock accounting. In the latter case, stocks
are highly volatile because good cash
ow news is always accompanied by a lower discount
rate, kicking the price up even more. Under portfolio accounting, good cash
ow news on a
particular stock will only coincide with a lower discount rate on the stock if the portfolio
as a whole does well. There is no guarantee of this, and so volatility is not amplied by as
4
much. Note also that since shocks to discount rates are perfectly correlated across stocks,
individual stock returns are highly correlated with one another. Moreover, the value and
size premia largely disappear since a stock's past performance no longer aects its discount
rate, which is now determined at the portfolio level. Finally, while there is a substantial
equity premium, it is not as large as under individual stock accounting. The investor is loss
averse over portfolio level
uctuations, which are sizeable but not as severe as the swings on
individual stocks. The compensation for risk is therefore more moderate.
While individual stock accounting can be helpful in understanding the data, we emphasize
that it is only a potential ingredient in an equilibrium model, not a complete description of
the facts. For one, we show that it underpredicts the correlation of stocks with each other,
and argue that a model which combines elements of both individual stock accounting and
portfolio accounting is likely to be superior to a model which uses individual stock accounting
alone.
The fact that we study equilibrium returns under both individual stock accounting and
portfolio accounting can also be useful for making additional predictions for future testing.
If individual stock accounting is relatively more prevalent among individual investors as
opposed to institutional investors, we would expect to see stocks held primarily by individuals
{ small stocks, for example { exhibit more of the features associated with individual stock
accounting. Other predictions arise, if, over time, investors change the way they do their
mental accounting. For example, the increased availability of mutual funds since the early
1980's may have caused a shift away from individual stock accounting and towards portfolio
accounting, since funds automatically prevent investors from worrying about individual stock
uctuations. Our analysis predicts that stocks that were once held directly but are now held
indirectly through mutual funds should exhibit specic changes in pricing behavior. Among
other predictions, such a shift should lead to higher price-to-fundamentals ratios and lower
value and size premia.
Loss aversion has already been applied with some success to understanding the aggregate
stock market. Benartzi and Thaler (1995) analyze the static portfolio problem of a loss
averse investor trying to allocate his wealth between T-Bills and the stock market. They
nd that the investor is reluctant to allocate much to stocks, even if the expected return
on the market is set equal to its high historical value. Motivated by this nding, Barberis,
Huang, and Santos (1999) introduce loss aversion into a dynamic equilibrium model and nd
that it captures a number of aggregate market phenomena. They do not address the time
series or cross-sectional behavior of individual stocks. Moreover, since they consider only
one risky asset, they cannot investigate the impact of dierent forms of mental accounting,
which is our main focus in this paper.
In Section 2, we propose two dierent specications for investor preferences: in one case,
5
the investor is loss averse over
uctuations in the value of individual stocks in his portfolio.
In the other, he is loss averse only over
uctuations in overall portfolio value. Section 3
derives conditions that govern equilibrium prices in economies with investors of each type.
In Section 4, we use simulations to analyze equilibrium stock returns for the two kinds of
mental accounting. Section 5 discusses the results further and Section 6 concludes.
6
decides that for both stocks, the share value a year from now will be distributed as
(
$150, with probability 1
2
$70, with probability ; 1
2
2 12 w(50) + 21 w( 30) = 10;
which is no longer as attractive.
Which form of mental accounting is a better description of individual behavior? From
a normative perspective, it makes sense to use as broad a form of accounting as possible.
This point of view is adopted by traditional asset pricing models which typically specify
utility only over total wealth or over consumption, and not over individual stock
uctuations.
Experimental work by psychologists, however, suggests that people often think about decision
problems too narrowly, segregating them from other relevant choices, a phenomenon known
as narrow framing.
In some of these experimental studies, subjects are oered a gamble which is unattractive
when viewed in isolation, but attractive when viewed as part of a broader portfolio of gambles
the subject is already taking. These studies typically nd that the gamble is rejected,
suggesting that it is being evaluated in isolation, rather than as part of the portfolio. Put
dierently, these studies suggest that subjects use too narrow a frame. Redelmeier and
Tversky (1992) and Rabin and Thaler (2000) discuss examples of this kind.
Other studies oer subjects the same gamble, but presented in two dierent ways: in
aggregate form, as a single gamble, or disaggregated, as a combination of several dierent
3 This calculation says: with probability 1 , both stocks will gain $50, for a total gain of $100; with
4
probability 21 , one stock will gain $50, the other will lose $30, for a total gain of $20; and with probability
1
4 , both
4
stocks will lose $30, for a total loss of $60.
This calculation says: for each stock, there is an equal chance of a gain of $50 and a loss of $30.
7
subgambles. Of course, how a gamble is presented should not matter. However, subjects
typically nd the gamble much less attractive in disaggregated form, suggesting that the way
information is presented leads them to evaluate the subgambles separately, making the overall
gamble unattractive, much as in our numerical example above. This is particularly true in
situations where it is not trivial to compute the probability distribution of the aggregate
gamble from information about each subgamble. Thaler, Tversky, Kahneman and Schwartz
(1997), Gneezy and Potters (1997), Benartzi and Thaler (1999), and Redelmeier and Tversky
(1992) present examples of such experiments.
In summary, when people have disaggregated information about the gambles they face,
they are often drawn into adopting a narrow frame when evaluating risk. There is good
reason to think that this idea may be relevant for investing in stocks. An investor building
a portfolio of stocks may devote a good deal of time to analyzing each stock he considers,
leading to a well thought out subjective probability distribution for the individual stock.
However, computing the implied portfolio distribution is a much harder task, requiring as
it does detailed knowledge of the correlation between each pair of stocks. As a result, the
investor may simply content himself with using individual stock accounting to evaluate his
risks. The fact that brokerage houses send investors detailed information about each of their
stock holdings in addition to information about their overall portfolio position may only
serve to reinforce this narrow framing. 5
In what follows, we consider two dierent kinds of narrow framing. First, we suppose
that investors care about consumption, but that over and above that, they experience loss
aversion over individual stock
uctuations. In other words, they evaluate their investment
performance stock by stock, and get utility from each stock's gain or loss. We refer to this
as \individual stock accounting". 6
In our second piece of analysis, we suppose that investors care about consumption, but
that over and above that, they are loss averse over
uctuations in the value of their over-
all portfolio of stocks. In other words, they evaluate their investment performance at the
portfolio level, and get utility from gains and losses in portfolio value. We call this \port-
folio accounting". While this is a broader form of mental accounting than individual stock
accounting, it still represents narrow framing: the investor is segregating his stock portfolio
5 Read, Loewenstein, and Rabin (1999) survey recent evidence on narrow framing. They also discuss
various explanations of why people frame decisions the way they do.
6 A skeptic could argue that an investor who does individual stock accounting will be reluctant to take on
blatantly attractive opportunities, such as exploiting a relative mispricing between two stocks by going long
one and short the other: even if he is sure to make $5 on the long position and to lose only $3 on the short,
he will code this as 5 2(3), which does not look attractive. However, since the correlation between returns
on the two legs of the trade is known to be -1, it is easy for investors to evaluate the strategy as a single
entity, even if they are subject to individual stock accounting in other circumstances where the correlation
is not so obvious. He should therefore be happy to take on such an opportunity.
8
from his other forms of wealth such as human capital, and is focusing on its
uctuations
separately.
We now show how these two forms of mental accounting can be incorporated into a
traditional asset pricing framework, starting with individual stock accounting in Section 2.1
and then moving to portfolio accounting in Section 2.2. In both cases, there are two kinds of
assets: a riskfree asset in zero net supply, paying a gross interest rate of Rf;t between time
t and t + 1; and n risky assets { \stocks" { each with a total supply of one unit. The gross
return on stock i between time t and t + 1 is Ri;t . +1
When the investor is loss averse over individual stock
uctuations, he chooses consumption
Ct and an allocation Si;t to stock i to maximize
1 " #
X Ct1
X
n
E t
1
+ b Ct
0
t+1
v (Xi;t+1 ; Si;t ; zi;t ) : (1)
t=0 i=1
The rst term in this preference specication, utility over consumption Ct, is a standard
feature of asset pricing models. Although the framework does not require it, we specialize
to power utility, the benchmark case studied in the literature. The parameter is the time
discount factor, and
> 0 controls the curvature of utility over consumption. 7
The second term models the idea that the investor is loss averse over changes in the value
of individual stocks that he owns. The variable Xi;t measures the gain or loss on stock i
+1
between time t and time t + 1, a positive value indicating a gain and a negative value, a loss.
The utility the investor receives from this gain or loss is given by the function v. We think
of v as capturing utility which is \over and above" any consumption-related utility.
The b C t
coeÆcient on the loss aversion terms is a scaling factor which ensures that
0
risk premia in the economy remain stationary even as aggregate wealth increases over time.
It involves per capita consumption C t which is exogeneous to the investor, and so does not
aect the intuition of the model. The constant b controls the importance of the loss aversion
0
terms in the investor's preferences; setting b = 0 reduces our framework to the much studied
0
9
a model of the aggregate stock market, and we borrow their specication. The dependence
on prior outcomes is the reason that v contains arguments beyond just the size of the gain or
loss Xi;t : it also depends on Si;t, the time t value of the investor's holdings of stock i, and
+1
zi;t , a variable that measures the investor's prior gains and losses on stock i as a fraction of
Si;t . By including Si;t and zi;t as arguments in v , the investor's prior returns on a stock can
aect the way subsequent losses on that stock are experienced.
In the remainder of this section, we make the denitions of Xi;t , zi;t and v precise. BHS
+1
provide more supporting detail: we limit ourselves to describing the essential structure. The
gain or loss on stock i between time t and t + 1 is measured as
Xi;t = Si;t Ri;t
+1 Si;t Rf;t :
+1 (2)
In words, the gain is the value of stock i at time t + 1 minus its value at time t multiplied by
the riskfree rate. Multiplying by the riskfree rate models the idea that investors may only
view a particular stock return as a gain if it exceeds the riskfree rate. The unit of time is
a year, so that gains and losses are measured annually. While the investor may check his
holdings much more often than that, even several times a day, we assume that it is only once
a year, perhaps at tax time, that he confronts his past performance in a serious way.
The variable zi;t tracks prior gains and losses on stock i. It is the ratio of another variable,
Zi;t , to Si;t , so that zi;t = ZS . BHS call Zi;t the \historical benchmark level" for stock i, to
i;t
be thought of as the investor's memory of an earlier price level at which the stock used to
i;t
trade. When Si;t > Zi;t, or zi;t < 1, the stock price today is higher than what the investor
remembers it to be, making him feel as though he has accumulated prior gains on the stock,
to the tune of Si;t Zi;t. When Si;t < Zi;t, or zi;t > 1, the current stock price is lower than
where it used to be, so that the investor feels that he has had past losses, again of Si;t Zi;t .
The point of introducing zi;t is to allow v to capture experimental evidence suggesting
that the pain of a loss depends on prior outcomes. This is done by dening v in the following
way.
When zi;t = 1, (
Xi;t Xi;t 0
v (Xi;t ; Si;t; 1) = for +1
; +1
(3)
+1
X X
i;t+1 <0 i;t+1
with > 1. For zi;t < 1,
(
Si;t Ri;t+1 Si;t Rf;t Ri;t zi;t Rf;t
v (Xi;t+1 ; Si;t ; zi;t ) = for +1
;
Si;t (zi;t Rf;t Rf;t ) + Si;t (Ri;t+1 zi;t Rf;t ) Ri;t < zi;t Rf;t
+1
(4)
and for zi;t > 1,
(
Xi;t+1 Xi;t+1 0
v (Xi;t+1 ; Si;t ; zi;t ) = for ; (5)
(zi;t )Xi;t+1 Xi;t+1 < 0
10
with
(zi;t ) = + k(zi;t 1) > : (6)
It is easiest to understand these equations graphically. Figure 1 shows the form of v: the
solid line for zi;t = 1, the dash-dot line for zi;t < 1, and the dashed line for zi;t > 1. When
zi;t = 1, the case where the investor has neither prior gains nor prior losses on stock i, v is
a piecewise linear function with a slope of one in the positive domain and a slope of in
the negative domain. This gives it a kink at the origin where the gain equals zero. Since
gains are rewarded at a rate of 1, and losses penalized at a rate of > 1, this is a simple
representation of loss aversion.
When zi;t < 1, the investor has accumulated prior gains on stock i. The form of
v (Xi;t ; Si;t ; zi;t ) is the same as for v (Xi;t ; Si;t ; 1) except that the kink is no longer at
+1 +1
the origin but a little to the left; how far to the left depends on the size of the prior gain.
This formulation captures the idea that prior gains may cushion subsequent losses. In par-
ticular, the graph shows that a small loss on stock i is penalized at the gentle rate of 1,
rather than : since this loss is cushioned by the prior gain, it is less painful. If the loss is
so large that it depletes the investor's entire reserve of prior gains, it is once again penalized
at the more severe rate of > 1.
The remaining case is zi;t > 1, where stock i has been losing money. The form of
v (Xi;t ; Si;t ; zi;t ) in this case has a kink at the origin just like v (Xi;t ; Si;t ; 1), but diers
+1 +1
from v(Xi;t ; Si;t; 1) in that losses are penalized more heavily than , capturing the idea
+1
that losses that come after other losses are more painful than usual. How much higher than
the penalty is, is determined by (6), and in particular by the constant k. Note that the
larger the prior loss, as measured by zi;t, the more painful any subsequent losses will be.
To complete the model description, we need an equation for the dynamics of zi;t. Based
on BHS, we use !
Ri
zi;t = zi;t + (1 ) (1); (7)
+1
R i;t+1
where Ri is a xed parameter. Note that if the return on stock i is particularly good, so
that Ri;t > Ri , the state variable zi;t = ZS falls in value. This means that the benchmark
+1
i;t
level Zi;t rises less than the stock price Si;t, increasing the investor's reserve of prior gains. In
i;t
other words, equation (7) captures the idea that a particularly good return should increase
the amount of prior gains the investor feels he has accumulated on the stock. It also says
that a particularly poor return depletes the investor's prior gains: if Ri;t < Ri, then zi;t
+1
goes up, showing that Zi;t falls less than Si;t, decreasing Si;t Zi;t.
Implicit in equation (7) is an assumption that the evolution of zi;t is unaected by the
actions the investor takes, such as buying or selling shares of the stock. In most cases, this
is very reasonable: if the investor sells some shares for consumption purposes, it is plausible
11
that any prior gains on the stock are reduced in proportion to the amount sold, in other
words, that zi;t remains constant. More extreme transactions, such as selling one's entire
holdings of the stock, might plausibly aect the way zi;t evolves. In assuming that they do
not, we are making a strong assumption, but one that is very useful in keeping our analysis
tractable.
Ri is not a free parameter, but is determined endogeneously by imposing the requirement
that in equilibrium, the median value of zi;t be equal to one. The idea behind this is that
half the time, the investor feels as though he has prior gains, and the rest of the time he
feels as though he has prior losses. It turns out that Ri is typically of similar magnitude to
the average stock return.
The second form of narrow framing we consider is portfolio accounting, where investors
track their investment performance at the portfolio level, and so are loss averse only over the
overall value of their portfolio. In particular, they choose consumption Ct and an allocation
Si;t to stock i to maximize
1 " #
X Ct1
E t
1
+ b Ct 0
t+1
v (Xt+1 ; St ; zt ) : (8)
t=0
When zt = 1, v is dened as
(
Xt+1 Xt+1 0
v (Xt+1 ; St ; 1) = for ; (10)
Xt+1 Xt+1 < 0
with > 1. For zt < 1,
(
St Rt+1 St Rf;t Rt zt Rf;t
v (Xt+1 ; St ; zt ) = for +1
; (11)
St (zt Rf;t Rf;t ) + St (Rt+1 zt Rf;t ) Rt < zt Rf;t
+1
12
with
(zt ) = + k(zt 1) > : (13)
In summary, the functional form of v is identical to what it was in the individual stock
accounting case. The only dierence is that in (1), the investor experiences loss aversion v
over changes in the value of each stock that he owns, while in (8), he is loss averse only over
overall portfolio
uctuations.
Finally, the dynamics of zt are given by
!
R
zt+1 = zt + (1 ) (1): (14)
Rt+1
3 Equilibrium Prices
We now derive conditions that govern equilibrium prices in two dierent economies. The rst
economy is populated by investors who are loss averse only over individual stock
uctuations,
and have preferences like those in (1)-(7). Investors in the second economy are loss averse
about overall portfolio
uctuations, and have preferences as in (8)-(14). In both cases, there
are a continuum of investors, with a total \mass" of one.
In each economy, we want to compute the price of a stock, say stock i, which we model
as a claim to a stream of perishable output given by the dividend sequence fDi;tg, where
!
log DDi;t+1 = gi + ii;t+1 ; (15)
i;t
We do not impose the Lucas (1978) restriction that aggregate consumption equal the
aggregate dividend. The advantage of this is that it allows the volatility of consumption
growth and of dividend growth to be very dierent in our model, as they are in the data.
Given that aggregate consumption diers from the aggregate dividend, we ll the gap by
13
assuming that each agent also receives a stream of nonnancial income fYt g { labor income,
say. We assume
Pn
that fYtg and fDi;tgi ;:::;n form a joint Markov process whose distribution
=1
gives C t i Di;t + Yt the distribution in (16). For simplicity, we assume that agents are
=1
not loss averse over labor income
uctuations, although this can be relaxed without aecting
the main features of our results.
Consider rst an economy where investors have the preferences in (1)-(7). Our assumptions
so far allow us to construct a Markov equilibrium in which the riskfree rate is constant
and the state variable zi;t determines the distribution of returns on stock i. Specically, we
assume that the price-dividend ratio of stock i is a function of the state variable zi;t ,
fi;t Pi;t =Di;t = fi (zi;t ); (18)
and then show that there is indeed an equilibrium satisfying this assumption. Note that
under this one-factor assumption, the distribution of the stock return Ri;t is determined +1
Intuitively, the value of stock i can change because of news about dividends "i;t , or +1
because its price-dividend ratio fi changes. Changes in this ratio are driven by changes in
zi;t , which tracks the past performance of the stock. Recent gains (losses) on the stock make
the investor perceive the stock as less (more) risky, raising (lowering) its price-dividend ratio.
In equilibrium, and under rational expectations about stock returns and aggregate con-
sumption levels, the agents in our economy must nd it optimal to hold the market supply
of zero units of the riskfree asset and one unit of each stock at all times, and to consume
their labor income and the dividend on each stock every period. The proposition below
8
14
and fi(:), the price-dividend ratio of stock i is given by:
" #
1 = e Et 1 +ff(iz(zi;t) ) e
!
gi
gc + 21
2 c2 (1 !ci
2) +1 ( i ci c )i;t+1
i i;t
" !#
+b Et vb 1 + f i (zi;t ) g +1
e ; zi;t ;
+1 i+ i i;t
(21)
0
fi (zi;t )
where for zi;t < 1,
(
Ri;t+1 Rf;t Ri;t zi;t Rf;t
vb(Ri;t+1 ; zi;t ) = for +1
; (22)
(zi;tRf;t Rf;t ) + (Ri;t+1 zi;t Rf;t ) Ri;t < zi;t Rf;t
+1
We now compute the price of stock i, the claim to the dividend stream fDi;tg, in a second
economy where investors are loss averse only about changes in overall portfolio value, so that
they have the preferences described in (8)-(14).
In the case of portfolio accounting, we need to price the portfolio before we can price
any individual stocks. The portfolio is a claim to the aggregate dividend, which follows the
process
log DDt = gp + p"t ; +1
+1 (24)
t
with "t i.i.d. N (0; 1) and
+1
(
corr(t ; "t ) =
0
!cp
0 for tt = t
6 t0
= (25)
(
corr("i;t; "t) = !ip
0 for tt = t
6= t0 : (26)
Note that the dividend processes for stocks 1 through n in (15) will not in general \add
up" to the aggregate dividend process in (24). Without additional structure, we cannot
think of the n stocks as a complete list of all stocks in the portfolio: we therefore imagine
15
that there are some other securities in the economy whose dividends are distributed in such
a way that the total dividend indeed adds up to the aggregate dividend in (24). For the
purpose of choosing parameters, it helps to have a setup where the n stocks on their own do
\add up", and we present this special case in Section 4.
Our assumptions allow us to construct a Markov equilibrium in which the riskfree rate
is constant and the portfolio-level state variable zt determines the distribution of returns on
all stocks. Specically, we assume that stock i's price-dividend ratio is a function of zt ,
fi;t Pi;t =Di;t = fi (zt ); (27)
and then show that there is indeed an equilibrium satisfying this assumption. Given this
one-factor assumption, the distribution of the stock return Ri;t is determined by zt and +1
Pi;t
+1
As in the rst economy, the price of stock i can change because of dividend news "i;t or +1
because its price-dividend ratio fi changes. The key dierence with the previous case is that
changes in this ratio are not driven by a stock-level state variable zi;t but by the portfolio-level
state variable zt , which tracks prior gains and losses on the overall portfolio. Recent gains
(losses) on the portfolio make the investor perceive the entire portfolio as less (more) risky,
raising (lowering) the price-dividend ratio of every stock in the portfolio simultaneously.
In equilibrium, and under rational expectations about stock returns and aggregate con-
sumption levels, the agents in our economy must nd it optimal to hold the market supply
of zero units of the riskfree asset and one unit of each stock at all times and to consume
their labor income and the aggregate dividend stream every period. The proposition below
characterizes the equilibrium.
Proposition 2. For the preferences in (8)-(14), there exists an equilibrium in which the
riskfree rate is constant at 2 2
Rf = e
g
= ; 1 c
(29) c 2
i t
" !#
+b Et ve 1 + f i (zt ) g ! +1 21 2
+1
e i+ !2
i
; Rt ; zt ;
ip t + (1
(30) )
fi (zt )
i ip
0 +1
16
where for zt < 1,
(
Ri;t+1 Rf;t Rt zt Rf;t
ve(Ri;t+1 ; Rt+1 ; zt ) = for +1
; (31)
(zt Rf;t Rf;t ) + (Ri;t+1 zt Rf;t ) Rt < zt Rf;t
+1
Rt =
1 + f (zt ) eg " +1 +1 p+ p t
(33)
f (zt )
+1
and (34), and then plug this into condition (30) governing the individual stock return.
Table 1 summarizes our choice of parameters. We divide the table into two panels, to
separate the two types of parameters: those that determine the distributions of consumption
and dividend growth; and those that determine investor preferences.
17
For the mean gc and standard deviation c of log consumption growth, we follow Ceccheti,
Lam, and Mark (1990) who obtain gc = 1:84% and c = 3:79% from a time series of annual
data from 1889 to 1985.
In principle, specifying parameters for individual and aggregate dividend growth is a
daunting task. Equations (15) and (24) show that we need gi, i , !ci, and !ip for each
stock, !ij for each pair of stocks, and gp, p , and !cp for the aggregate portfolio, a total of
n2
2
+ + 3 parameters. Fortunately, it turns out that with two simplifying assumptions, we
7
2
n
can specify the dividend processes with just four parameters, and yet still convey most of
the essential economics. Essentially, we assume that all stocks are identically distributed, so
that all stock-specic parameters are the same across stocks. In Section 4.5, we relax this
restriction.
For now, we assume that the mean and standard deviation of log dividend growth is the
same for all stocks,
gi = g; i = ; 8i. (35)
Second, we assume a simple factor structure for individual stock dividend growth innovations,
q
"i;t+1 = !p "t+1 + "bi;t+1 1 !p2 : (36)
In words, the cash
ow shock to stock i has one component due to the aggregate dividend
innovation "t introduced in (24), and one idiosyncratic component, "bi;t N (0; 1). The
+1 +1
relative importance of the two components is controlled by a new parameter !p. The idiosyn-
cratic component is orthogonal to consumption growth shocks, aggregate dividend growth
shocks, and the idiosyncratic shocks on other stocks:
corr("bi;t; t) = corr("bi;t; "t) = corr("bi;t; "bj;t) = 0; 8i; j: (37)
This immediately implies
!ci = !p!cp; 8i, (38)
!ij = !p2; 8i; j; (39)
!ip = !p; 8i: (40)
Another attractive feature of this simple factor structure is that in the limit, as we add
more and more stocks, their total dividend is also i.i.d lognormal:
Pn
i=1 Di;t+1 1 2 !p2 )+!p "t+1
lim
n!1
P n
Di;t
! eg + 2 (1 : (41)
i=1
18
(41) with (24), we obtain
1
gp = g + 2 (1 !p2) (42)
2
p = !p: (43)
Equations (35), (38)-(40), and (42)-(43) show that the entire structure of dividend growth
can be determined from gp, p, , and !cp alone. We choose these four quantities as the
basis variables rather than any other four because they can be estimated relatively straight-
forwardly. First, we estimate the mean and standard deviation of aggregate dividend growth
using NYSE data from 1925-1995 from CRSP which gives gp = 0:015 and p = 0:12. We set
, the volatility of individual stock dividend growth equal to the median dividend growth
volatility for stocks in the Compustat database, which gives = 0:25. Finally, the cor-
9
relation between shocks to consumption growth and dividend growth, !cp, we take from
Campbell (1999) who estimates it in the neighbourhood of 0.15. Table 1 shows what these
values imply for the remaining parameters governing the dividend processes. 10
We choose preference parameters to match the behavior of actual asset returns as closely
as possible for the case of individual stock accounting; we can then judge to what extent
they are reasonable. These parameter values are summarized in the lower panel of Table 1.
We choose the curvature
of utility over consumption and the time discount factor so as
to produce a sensibly low value for the riskfree rate. Given the values of gc and c, equation
(20) shows that
= 1:0 and = 0:98 bring the riskfree interest rate close to Rf 1 = 3:86%.
The value of determines how keenly losses are felt relative to gains in the case where the
investor has no prior gains or losses. We take = 2:25, which is also the value Tversky and
Kahneman (1992) estimate by oering subjects isolated gambles in experimental settings.
The parameter k determines how much more painful losses are when they come on the
heels of other losses. We choose k = 3. To interpret this, suppose that the state variable zi;t
is initially equal to 1, and that stock i then experiences a sharp fall of 10%. From equation
(7) with 1, this means that zi;t goes up by about 0.1, to 1.1. From (6), the pain of any
additional losses will now be penalized at 2:25+3(0:1) = 2:55, a slightly more severe penalty.
The parameter arises in the denition of the state variable dynamics. It controls the
persistence of the state variable zi;t , which in turn controls the autocorrelation of price-
dividend ratios. We nd that = 0:9 brings this autocorrelation close to its empirical
value.
9 More precisely, we take all stocks in the annual Compustat database for which at least 11 consecutive
years of dividend data are recorded, compute dividend growth volatility for each, and record the median.
10 Note that while the mean log dividend growth gi is negative, the mean simple dividend growth equals
exp( 0:0091 + 0:25
2
2 ) 1 = 2:24%, a positive number.
19
Finally, b determines the relative importance of the loss aversion term in the investor's
0
4.2 Methodology
For the case of individual stock accounting, we start by using an iterative technique to solve
equation (21) for the price dividend ratio fi(:) of an individual stock. The only diÆculty is
that the state variable zi;t is endogenous: it tracks prior gains and losses, which depend on
past returns, themselves endogenous. To deal with this, we use the following procedure. We
guess a solution to (21), fi say. We then construct a function zi;t = hi (zi;t ; "i;t ) that
(0)
+1
(0)
+1
solves
Ri;t =
1 + fi(zi;t ) eg +1 +1
(44)
i+ i i;t
f (z )
+1
i i;t
and !
R
zi;t+1 = zi;t + (1 ) (1) (45)
Ri;t+1
simultaneously for this particular fi = fi . Given hi , we get a new candidate solution fi
(0) (0) (1)
This hi gives us a new candidate fi = fi from (46). We continue this process until we
(1) (2)
Once we obtain the nal price-dividend ratio, we use simulated data to see how returns
behave in equilibrium. We simulate dividend shocks f"i;tg for n = 100 stocks and for 10,000
time periods, subject to the specication in (15) and the parameters in Table 1. We then
apply the price-dividend function fi(:) to this dividend data to see what realized returns look
like. More precisely, we use the zi;t = hi (zi;t; "i;t ) function described above to generate
+1 +1
the series of zi;t implied by the dividend shocks and then set the return on stock i between
time t and t + 1 equal to
Ri;t =
1 + fi(zi;t ) eg " +1 . +1
(47)
i+ i i;t
f (z )
+1
i i;t
This gives n time series of individual stock returns. We can then compute moments {
volatility, say { for each stock, and then average these moments across dierent stocks.
20
This provides a sense of how the \typical" stock behaves, and we report the results of such
calculations later in the section.
We can also use our n time series of individual stock returns to compute an equal-weighted
average
Rp;t =
1Xn
Ri;t (48)
n+1
i=1
+1
For the case of portfolio accounting, we start out by using iteration in equation (34) to
compute the aggregate price dividend ratio. As before, we iterate between guesses f = f j ( )
Rt =
1 + f (zt ) eg " +1+1 p+ p t
(49)
f (zt )
+1
and !
R
zt = z t + (1 ) (1) (50)
+1
R t+1
the resulting functions f (:) and zt = h(zt ; "t ) and iterate in equation (30) over guesses
+1 +1
fi j (:) for stock i's price-dividend ratio, converging eventually to the solution fi (:).
( ) 12
Simulation then illustrates the behavior of individual stock returns. We again generate
dividend shocks f"i;tg for n = 100 stocks and also for the portfolio, f"tg, for 10,000 time
periods. The function zt = h(zt ; "t ) generates the time series for the aggregate state
+1 +1
variable zt implied by these f"t g. The time series of returns for stock i is then given by
Ri;t =
1 + fi(zt ) eg " +1 ,
+1 i+ i i;t
(51)
+1
fi (zt )
while the aggregate return is measured by
=1 R .
X n
R p;t+1
n i=1 i;t+1 (52)
Table 2 summarizes the properties of individual and aggregate stock returns in simulated
data from two economies: one in which investors use individual stock accounting, and an-
other in which they use portfolio accounting. The nal column presents empirical values for
11 The aggregate return Rp;t+1 diers from the aggregate return Rt+1 described earlier only in that it is
equal-weighted rather than value-weighted.
12 The way f (:) enters equation (30) is through the portfolio return Rt+1 in (31) and (32). Note that Rt+1
depends on f (:) as shown in (33).
21
comparison. The top panel in the table reports time series properties of individual stocks;
the middle panel describes the time series properties of aggregate returns; nally, the bot-
tom panel summarizes the cross-sectional patterns in average returns. Both simulated and
empirical values are expressed in annual terms. In this section, as well as in Sections 4.4
13
and 4.5, we lay out the results and explain the intuition behind them. Section 5 discusses
the broader implications of our ndings.
Our rst set of results suggest that loss aversion over individual stock
uctuations oers
a way of understanding a wide range of empirical phenomena, both in the time series and
the cross-section. We look rst at the time series.
Under individual stock accounting, the typical individual stock has a high average excess
return, E (R) Rf = 8:3%. It is volatile, and in particular, \excessively" volatile, in the
14
sense of Shiller (1981) and Leroy and Porter (1981): its standard deviation, R = 43:3%; is
higher than the standard deviation of underlying dividend growth, = 25%. Its returns are
also slightly predictable: a regression of 4 year cumulative log returns on the lagged dividend
yield,
log(Ri;t Ri;t Ri;t Ri;t ) = + D
+1 +2 +3 +4
P
i;t
+ vi;t ; +4
i;t
produces an R of 4.8%.
2
The last variable in the top panel, !R ; measures the average contemporaneous correlation
of individual stocks; we calculate it to be 0.25 in a way that we explain shortly.
The next panel shows that the aggregate stock market also has a high excess average
return, E (RM ) Rf = 8:3%, and is excessively volatile, with a standard deviation of M =
21:5% that exceeds the 12% volatility of aggregate dividend growth. Aggregate stock returns
are also slightly predictable: a regression of 4 year cumulative log aggregate returns on the
lagged aggregate dividend yield gives an R of 7:5%. 2 15
The market volatility M helps us measure the average correlation !R between stocks
13 The empirical estimates of the size and value premia come from Fama and French (1992); the measure of
\value" is price scaled by book value. The De Bondt-Thaler premium, described later in this section, is taken
from Chopra, Lakonishok, Ritter (1992), who update De Bondt and Thaler's original results. The aggregate
equity premium and volatility are based on data used in Fama and French (2000), while the measure of
aggregate predictability comes from Fama and French (1988). The typical volatility of an individual stock
comes from Elton and Gruber (1995), and the measure of time series predictability from Vuolteenaho (1999).
Finally, we impute the typical correlation between individual stocks from the aggregate and individual stock
volatility estimates, in a manner to be described shortly, and also impute the typical mean of an individual
stock from the aggregate mean.
14 As described in Section 4.2., the results for individual stocks come from computing the relevant moment
for each stock in the simulated sample and then averaging across stocks.
15 Note that the table uses the notation R2 (M ) to distinguish this from the R2 obtained earlier in the
individual stock return regression.
22
2
reported earlier: we compute it as . This calculation is exact if all stocks have the
M
same standard deviation and correlation with one another, as they do in our simple economy,
R
since
!
R ;t + : : : + Rn;t R 2
1
= lim V ar(
2
n!1
1 +1
) = lim
+1
+ (1 ) ! = ! .
n!1
2
R R
2
R R
M
n n n
The bottom panel in the table describes the cross-sectional features of individual stock
returns. Our simulated data reproduces the predictive power of scaled-price variables { in
our case, the price-dividend ratio { for the cross-section of average returns, a phenomenon
sometimes known as the \value" premium. Each year, we sort stocks into deciles based on
this ratio, and measure the returns of the top and bottom decile portfolios over the next
year. The time series mean of the dierence in returns between the two deciles is a very
substantial 21:2%.
Our data also produces a \size" eect, in that stocks with smaller market capitalizations
have higher average returns. Each year, we sort stocks into deciles based on their total value,
P , and measure the returns of the top and bottom decile portfolios over the next year. The
time series mean of the dierence in returns between the two deciles is 6.3%.
Our simulated data also replicates a well known study of De Bondt and Thaler (1985)
which nds that long term prior losing stocks on average outperform long term prior winning
stocks. At any point in time, we sort stocks into deciles based on their 3 year prior return,
and measure the average annual returns of the best and worst prior performing deciles over
the next three years. The time series mean of the dierence in average returns between the
two deciles over all non-overlapping periods in our simulated data is 13:3%.
Many of the eects we obtain under individual stock accounting derive from a single
source, namely a discount rate for individual stocks that changes as a function of the stock's
past performance. If a stock has had good recent performance, the investor gets utility from
this gain and becomes less concerned about future losses on the stock because any losses
will be cushioned by the prior gains. In eect, the investor thinks of the stock as less risky
and discounts its future cash
ows at a lower rate. Conversely, if one of his stocks performs
dismally, he nds this painful and becomes more sensitive to the possibility of further losses
on the stock: in eect he views the stock as riskier and raises its discount rate.
This changing discount rate has many implications. It gives individual stocks some time
series predictability: a lower discount rate pushes up the price-dividend ratio and leads
to lower subsequent returns, which means that the lagged price-dividend ratio can predict
returns. It makes stock returns more volatile than underlying cash
ows: a high cash
ow
pushes the stock price up, but this prior gain also lowers the discount rate on the stock,
pushing the stock price still higher. It also generates value and size premia in the cross-
section : stocks with high
D ratios (growth stocks) or high P values (large stocks) have often
P
23
done well in the past, accumulating prior gains for the investor who then views them as less
risky and requires a lower average return. Stocks with low DP ratios (value stocks) or low
P values (small stocks) have often had dismal prior performance, burning the investor, who
now thinks of them as riskier, and charges a higher risk premium.
The high equity premia we obtain under individual stock accounting derive from a dier-
ent source: since the investor is loss averse over individual stock
uctuations, he hates the
frequent losses that individual stocks often produce, and charges a high average return as
compensation. Other papers, such as Benartzi and Thaler (1995) and Barberis, Huang, and
Santos (1999) have also suggested loss aversion as way of understanding a high equity pre-
mium. The eect we obtain here, though, is one level stronger than in those earlier papers,
since the investor is now loss averse over individual stock
uctuations rather than over the
less dramatic
uctuations in the diversied aggregate market.
Our next set of results show that the investor's system of mental accounting matters a great
deal for the behavior of asset prices. As we broaden the investor's frame from individual
stock accounting to portfolio accounting, individual stocks exhibit quite dierent features in
equilibrium.
Table 2 shows that under portfolio accounting, the average excess return on a typical
individual stock is 2.9% { not insubstantial, but rather lower than under individual stock
accounting. At 30.4%, individual stock volatility is also lower than under individual stock
accounting; in particular, excess volatility of returns over dividend growth is much smaller.
The time series predictability of individual stock returns is also minimal.
The average excess return on the aggregate market is 2.9%. Interestingly, aggregate
returns are roughly as volatile here as they were under individual stock accounting. Since
individual stocks are much less volatile here than under individual stock accounting, this
must mean that stocks are more highly correlated than before, and indeed, !R = 0:41.
Finally, aggregate stock returns are slightly predictable.
We now turn to the cross-section. One disadvantage of our assumption that all stock-
specic parameters { dividend growth mean and standard deviation, and correlations with
the overall portfolio and with consumption { be the same for all stocks is that there is no
cross-sectional dispersion in price-dividend ratios in the case of portfolio accounting. This
assumption will be relaxed in Section 4.5. For now, the lack of dispersion means that we
cannot check for a value premium in the simulated data.
We can, however, still look to see if there is a size premium, or if there is a De Bondt-
Thaler premium to prior losers. As the table shows, these eects are no longer present under
24
portfolio accounting.
The reason the results are dierent under this form of accounting is that now, changes
in discount rates on stocks are driven by
uctuations in the value of the overall portfolio:
when the portfolio does well, the investor is less concerned about losses on any of the stocks
that he holds, since the prior portfolio gain will cushion any such losses. Eectively, he
views all stocks as less risky. Discount rates on all stocks therefore go down simultaneously.
Conversely, discount rates all go up simultaneously after a prior portfolio loss.
This discount rate behavior is the key to understanding a number of our results. Stock
returns are less volatile here than under individual stock accounting because changes in
discount rates are now less correlated with cash
ow news about any one stock. Under
individual stock accounting, stocks are highly volatile because good cash
ow news is always
accompanied by a lower discount rate, kicking the price up even more. Under portfolio
accounting, good cash
ow news on a particular stock will only coincide with a lower discount
rate on the stock if the portfolio as a whole does well. There is no guarantee of this, and
so volatility is not amplied by as much. Note also that since shocks to discount rates are
perfectly correlated across stocks, individual stock returns are highly correlated with one
another. Moreover, a stock with long term poor prior performance will not necessarily have
a higher expected return than a long term prior winner, in the way that De Bondt and
Thaler found. The discount rate on both stocks will be determined by portfolio-level events,
and hence will be the same.
Finally, note that while there is a substantial equity premium, it is not as large as under
individual stock accounting. The investor is loss averse over portfolio level
uctuations,
which are sizeable but not as severe as the swings on individual stocks. The compensation
for risk is therefore more moderate.
Our analysis so far has assumed that mean dividend growth rates gi and dividend growth
volatilities i are equal across stocks. In particular, Table 1 shows that we have assumed
gi = 0:0091 and i = 0:25, 8i. We now relax this restriction. This should allow for a more
realistic comparison of the cross-sectional features of the simulated data with those of actual
data.
As before, we simulate dividend data for 100 dierent stocks over 10; 000 time periods.
This time, however, we draw gi and i for each stock from a distribution, independently
across dierent stocks:
gi N ( 0:0091; g )2
i N (0:25; S )
2
25
where g = 0:01 and S = 0:05. In other words, the average gi across dierent stocks is the
same as before, -0.0091, but we allow for some dispersion. Similarly, the average i is the
same as before, 0.25, but we now allow for dispersion. We use the dispersion in dividend
growth volatilities estimated from Compustat data as a guide to choosing S ; g is much
harder to estimate, but we nd that an g of 0:01 leads to a realistic cross-sectional dispersion
in price-dividend ratios.
Table 3 repeats the cross-sectional calculations shown in the bottom panel of Table 2 for
our new simulated data. We look rst at the results for individual stock accounting.
Under individual stock accounting, the value premium, size premium and the De Bondt-
Thaler premium to prior losers remain as strong as in Table 2. One mechanism generating
the value and size premia here is changing discount rates on individual stocks, exactly as
described earlier. Now that we have allowed for cross-sectional dispersion in the volatility of
dividend growth, there is an additional mechanism contributing to the value and size premia.
A rm with more volatile cash
ow growth has more volatile returns, which scares the loss
averse investor into charging a higher risk premium. Since the investor is applying a high
discount rate to future cash
ows, the rm will have a low DP ratio and a low P value, thus
generating a cross-sectional link between DP ratios (or P values) and average returns.
It is not easy to isolate the eect of this additional mechanism by comparing the \In-
dividual Accounting" column in Table 3 to the \Individual Accounting" column in Table
2. The reason is that we have introduced dispersion in both the mean and volatility of
dividend growth, and it is not possible to separate out the eect of these two changes. Table
4 addresses this. The results here are based on simulated data in which we allow dispersion
in means, gi, but not in volatilities, i. As we go from the \Individual Accounting" column
in Table 4 to the corresponding column in Table 3, we see that the size and value premia
increase in magnitude, a result that we can now cleanly attribute to dispersion in the i .
It is interesting to note that for individual stock accounting, the size and value premia in
Tables 3 and 4 are lower than the premia in Table 2. The reason for this is the following. In
the Table 2 simulations, a stock only has a low DP ratio if investors decide to assign it a high
discount rate. Portfolios of low DP stocks therefore earn high returns on average. In the more
realistic simulations we do in Tables 3 and 4, a low DP ratio can be a sign of a high discount
rate, but it can also mean that the stock has a low gi, in other words, that investors expect
low cash
ow growth on the stock. Portfolios of low DP stocks still have high returns, but the
eect is diluted since the portfolio now includes stocks with low expected growth rates and
only average discount rates.
Tables 3 and 4 also presents results for the portfolio accounting case. There is now
nontrivial dispersion in price-dividend ratios, which means that we can examine whether
price-dividend ratios have any predictive power in the cross-section; they do not. In other
26
words, there is no value premium. The attempt to generate a size premium or to replicate
De Bondt and Thaler's ndings is also a failure, as it was in Table 2.
5 Further Discussion
We draw a number of conclusions from the results in Tables 2, 3, and 4. First, loss aversion
over individual stock
uctuations appears to be a helpful device for understanding many
features of asset returns, at least some of which are not easily generated in traditional
equilibrium models: a high mean, volatility, and modest predictability for individual and
aggregate stock returns in the time series, as well as substantial \value" and \size" premia
in the cross-section.
We emphasize two aspects of these results. First, it is interesting that a single mechanism
{ discount rates that change depending on the stock's past performance { generates several
dierent kinds of eects in our simulated data. Put dierently, we have shown that empirical
phenomena which are often thought of as being distinct { time series predictability and the
cross-sectional value premium, for example { may in fact be closely related. Of course, this
is a specic feature of our model, and one that can in principle be tested.
Another noteworthy aspect of our results is the fact that we are able to produce a sizeable
value premium. The empirical value of this premium is over 10% per year on average, which
initially seems a dauntingly high number to explain in the context of an equilibrium model
with rationally maximizing agents. Our analysis shows that mild changes in loss aversion
over individual stocks { or more specically, a slight decrease (increase) in the discount rate
of a stock after a prior gain (loss) { are more than suÆcient to generate a value premium
similar to that in the data. At the same time, these mild changes in loss aversion generate
only mild time series predictability, which also matches the data.
While our results suggest that individual stock accounting can be a useful device for
understanding features of asset prices, we emphasize that it is at most a potential ingredi-
ent in an equilibrium model, and cannot provide a complete description of the facts. For
example, it predicts that the correlation between returns on dierent stocks is the same as
the correlation between their cash
ows. This can be seen in Table 2, which reports the typ-
ical correlation between stocks to be !R = 0:25, almost identical to the correlation between
cash
ow shocks, !ij = 0:23; listed in Table 1. The reason for this is that under individual
stock accounting, shocks to discount rates are only as correlated as the cash
ows shocks.
However, Vuolteenaho (1999) nds that shocks to expected returns on dierent stocks are
actually much more correlated than cash
ow shocks, which immediately implies that stocks
are more correlated with each other than are their underlying cash
ows. A model which com-
bines individual stock accounting and portfolio accounting may therefore be more fruitful
27
than one with individual stock accounting alone: portfolio accounting introduces a common
component in discount rate variation across stocks and therefore makes stocks move together
more than their cash
ows do.
Although individual stock accounting on its own is not a complete description of the
facts, a model without any individual stock accounting at all is much less successful. Table 2
shows that if we are not prepared to allow for individual stock accounting, then some other
additional structure has to be imposed if the empirical facts are to be understood: under
portfolio accounting alone, individual stock returns lack volatility and are too highly corre-
lated, and there is no value premium, no size premium, and no De Bondt-Thaler premium
to prior losers in the cross-section.
Although we make this last point in the context of a specic model based on loss aver-
sion, it may apply more generally to any model that uses a broad form of mental accounting,
including most standard consumption-based models. For example, in consumption-based
habit formation models, discount rates on individual stocks depend on an aggregate state
variable, just as in our portfolio accounting model: when consumption approaches habit, the
investor becomes more risk averse and hence raises the discount rates on all stocks simul-
taneously, pushing them all down together. Following the same intuition as in Section 4.4.,
habit formation models may share the drawbacks of our portfolio accounting model: without
additional structure, individual stock returns may be insuÆciently volatile and excessively
correlated and there may be no value premium nor size premium in the cross-section.
The fact that we study equilibrium returns under both individual stock accounting and port-
folio accounting can also be useful for making additional predictions for future testing. If
individual stock accounting is relatively more prevalent among individual investors as op-
posed to institutional investors, we would expect to see stocks held primarily by individuals
{ small stocks, for example { exhibit more of the features associated with individual stock
accounting. In particular, by comparing the \Individual Accounting" and \Portfolio Ac-
counting" columns in Table 2, the specic prediction is that small stocks should have higher
mean returns and more excess volatility than large stocks, be more predictable in the time
series and less correlated with each other, and that the value and De Bondt-Thaler premia
be stronger among small stocks.
Other predictions arise, if, over time, investors change the way they do their mental
accounting. For example, the increased availability of mutual funds since the early 1980's
may have caused a shift away from individual stock accounting and towards portfolio ac-
counting, since funds automatically prevent investors from worrying about individual stock
28
uctuations. Our analysis predicts that stocks that were once held directly but are now
increasingly held indirectly through mutual funds should exhibit specic changes in pricing
behavior. Among other predictions, such a shift should lead such stocks to have higher
price-to-fundamentals ratios and lower value and size premia.
In Section 4.1., we introduced a one-factor cash
ow structure for stocks so as to simplify the
calibration process as much as possible. If in reality, cash
ows did indeed have a one-factor
structure, it would be relatively straightforward for investors who do not use narrow framing
to take advantage of investors who do, thus attentuating their eects. For example, they
could buy a portfolio of stocks with price-dividend ratios in the lowest decile and short a
portfolio of stocks with price-dividend ratios in the highest decile. This strategy earns the
sizeable value premium documented in Tables 2 through 4, and if implemented with a very
large number of stocks, becomes almost riskless: idiosyncratic risk gets washed away within
the portfolio while the long-short position nets out the market factor. Table 5 provides some
numbers: the column marked \one-factor" reports the standard deviation of the strategy just
described for various values of n, the total number of stocks. As n increases, the standard
deviation falls, dropping to as low as 11:5% for an n of 1000. Coupling this standard deviation
with the kind of value premia reported in Table 3 produces very attractive Sharpe ratios.
The problem with such a strategy in reality is that stock returns are driven not by one,
but by many dierent factors, making it much harder to reduce the strategy's risk, even
with many stocks. Industry shocks are one example of such additional factors. To illustrate,
suppose that each of our n stocks falls into one of a small number of industry sectors, labelled
s = 1; : : : ; S . Suppose that the cash
ow shock to stock i has the following structure:
q
"i;t = !p"t + !s"s;t + "bi;t 1 !p !s :
+1 +1 +1 +1
2 2
In addition to the market-wide shock "t and the idiosyncratic shock "bi;t , there is sector-
+1 +1
level cash
ow shock "s;t , where s is the sector to which stock i belongs. All shocks are
+1
distributed as N (0; 1), and their relative importance is determined by !p and !s. If we
assume for simplicity that these shocks are orthogonal to one another, the correlation !ij
between the cash
ow shocks on stock i and j will satisfy:
!ij = !p , for i 6= j and i; j in dierent sectors
2
!ij = 1, for i = j .
We can now simulate long time series of dividend data on n stocks for this correlation
structure, and then compute the resulting stock prices in a world with individual stock
29
accounting. The column titled \multi-factor" in Table 5 reports the standard deviation of
the same long-short strategy described at the beginning of this section. Note that once
16
we recognize the additional factors that aects stock returns, the strategy becomes much
more risky: going from 100 to 1000 stocks only reduces risk by about a fth, in contrast
to the almost 50% reduction in risk in the \one-factor" case. Intuitively, many value stocks
now belong to the same sector and hence comove: there may have been bad news at the
sector level, pushing all stocks in the sector down and leading investors to view them all as
more risky, thus giving them low price-dividend ratios and making them all value stocks.
Similarly, many growth stocks now belong to the same sector, and therefore also comove. A
long value, short growth strategy does not net out these industry factors, leaving it far from
riskless.
There may be other reasons why exploiting investors who engage in narrow framing may
be harder in reality than it appears in theory. First, it is likely to take many years of data
before arbitrageurs can be statistically condent of the existence of a value premium, and
even more time before they can convince themselves that it is not simply compensation for
a lurking risk factor. Put dierently, the eects of narrow framing may persist for a very
long time before they can be detected and exploited. 17
Second, the long-short strategy we have focused on involves buying stocks with very poor
prior performance, and shorting stocks with very good prior performance. For institutional
investors, such a strategy is hard to justify to sponsors: stocks with dismal prior performance
are not considered \prudent" investments, and are therefore avoided as much as possible,
while stocks with good recent performance are thought of as prudent and institutions would
rather buy them than short them.
Although we have argued that exploiting investors who use individual stock accounting is
likely to be a diÆcult enterprise, we accept that it will be tempting enough for arbitrageurs
to engage in at least to a partial degree. In other words, arbitrageurs will to some extent
buy beaten-down stocks from individuals who frame narrowly and hence nd those stocks
more risky. Moreover, they will, to some extent, sell high-
ying stocks to those same narrow
framers who nd these stocks less risky. Thinking of the narrow framers as individual
investors and the arbitrageurs as institutions, we would then expect to see individuals selling
to institutions in market troughs and institutions selling to individuals at market peaks.
There is in fact evidence of such an eect: Cohen (1999) examines the long-term buying
and selling patterns of individuals and institutions, and nds exactly the pattern we predict,
namely individuals selling to institutions in market troughs and vice-versa at market peaks.
16 In this particular example, we take !p = 0:48 as before, and !s = 0:5.
17 After all, the value premium in our simulated data, while certainly strong, has been computed using
10,000 years of data. Arbitrageurs looking for exploitable anomalies have much less data at their disposal.
30
It is interesting to compare our prediction (and Cohen's results) with the ndings of
Odean (1997), who shows that individual investors are reluctant to sell stocks with short-
term prior losses, preferring to sell prior short-term winner stocks. The dierence between
18
this nding and our prediction may lie in the horizon over which investors come to terms
with an investment loss. If a stock has experienced a short-term loss, the investor may not
count this as a loss, prefering to hold on the stock in the hope of breaking even down the line.
However, if the stock experiences a sustained, long-term drop, the investor may eventually
accept that the investment has been a failure; he will accept the loss, view the stock as
riskier, and then be happy to sell it.
6 Conclusion
In this paper, we incorporate loss aversion into an equilibrium asset pricing model, pay-
ing particular attention to the issue of what people are loss averse about. We consider
two possibilities, which correspond to dierent assumptions about how investors do mental
accounting, or about how they track their investment performance. First, we analyze indi-
vidual stock accounting, where people evaluate their performance stock by stock and feel
loss aversion over individual stock
uctuations. We also consider portfolio accounting, where
investors track their overall portfolio performance and feel loss aversion only over portfolio
uctuations.
We nd that a model with individual stock accounting generates a rich set of eects,
closely matching well-known phenomena in actual data. Individual stock returns have a
high mean, excess volatility, and are moderately predictable in the time series; there are also
substantial \value" and \size" premia in the cross-section.
A model with portfolio accounting alone is rather less successful at matching observed
empirical facts. Individual stock returns are insuÆciently volatile, and too highly correlated.
Moreover, there is no value premium nor size premium to speak of in the cross section.
We note that while individual stock accounting may be a useful ingredient in equilibrium
models, it cannot in and of itself be a complete description of the facts. We argue that
a model which combines elements of both individual and portfolio accounting may oer a
better hope of understanding the data.
18 This is often known as the disposition eect, and is also discussed by Shefrin and Statman (1985).
31
7 References
Banz, R. (1981), \The Relation between Return and the Market Value of Common Stocks",
Journal of Financial Economics 9, 3-18.
Barberis, N., M. Huang, and T. Santos (1999), \Prospect Theory and Asset Prices", forth-
coming, Quarterly Journal of Economics.
Benartzi, S., and R. Thaler (1995), \Myopic Loss Aversion and the Equity Premium Puzzle",
Quarterly Journal of Economics 110, 73-92.
Benartzi, S. and R. Thaler (1999), \Risk Aversion or Myopia? Choices in Repeated Gambles
and Retirement Investments", Management Science 45, 364-381.
Campbell, J.Y., \Asset Price, Consumption and the Business Cycle," forthcoming in Hand-
book of Macroeconomics, John Taylor and Michael Woodford eds. (Amsterdam: North
Holland, 1999) .
Campbell, J.Y., and R. Shiller (1988), \Stock Prices, Earnings, and Expected Dividends",
Journal of Finance 43, 661-676.
Ceccheti, S., P. Lam, and N. Mark (1990), \Mean Reversion in Equilibrium Asset Prices,"
American Economic Review 80, 398-418.
Elton, E., and M. Gruber (1995), Modern Portfolio Theory and Investment Analysis (Wiley,
New York).
Fama, E., and K. French (1989), \Dividend Yields and Expected Stock Returns", Journal
of Financial Economics 22, 3-27.
Fama, E., and K. French (1992), \The Cross-section of Expected Stock Returns", Journal
of Finance 47, 427-465.
Fama, E., and K. French (1996), \Multifactor Explanations of Asset Pricing Anomalies",
Journal of Finance 51, 55-84.
Fama, E., and K. French (2000), \The Equity Premium", Working Paper, University of
Chicago.
32
Gneezy, U., and J. Potters (1997), \An Experiment on Risk Taking and Evaluation Periods",
Quarterly Journal of Economics 112, 631-645.
Kahneman, D., and D. Lovallo (1993), \Timid Choices and Bold Forecasts: A Cognitive
Perspective on Risk Taking", Management Science 39, 17-31.
Kahneman, D., and A. Tversky (1979), \Prospect Theory: An Analysis of Decision Under
Risk," Econometrica 47, 263-291.
Leroy, S., and R. Porter (1981), \The Present Value Relation: Tests Based on Variance
Bounds", Econometrica 49, 555-577.
Lucas, R. (1978), \Asset Prices in an Exchange Economy," Econometrica 46, 1419-1446.
Odean, T. (1998), \Are Investors Reluctant to Realize their Losses?", Journal of Finance
53, 1775-1798.
Rabin, M., and R. Thaler (2000), \Anomalies: Risk Aversion", Working Paper, University
of Chicago.
Read, D., G. Loewenstein, and M. Rabin (1999), \Choice Bracketing", Journal of Risk and
Uncertainty 19, 171-197.
Redelmeier, D., and A. Tversky (1991), \On the Framing of Multiple Prospects", Psycho-
logical Science 3, 191-193.
Rosenberg, B., K. Reid, and R. Lanstein (1985), \Persuasive Evidence of Market IneÆ-
ciency", Journal of Portfolio Management 11, 9-17.
Shefrin, H., and M. Statman (1985), \The Disposition to Sell Winners too Early and Ride
Losers too Long: Theory and Evidence", Journal of Finance 40, 777-790.
Shiller, R. (1981), \Do Stock Prices Move too Much to be Justied by Subsequent Changes
in Dividends?", American Economic Review 71, 421-436.
Thaler, R. (1980), \Toward a Positive Theory of Consumer Choice", Journal of Economic
Behavior and Organization 1, 39-60.
Thaler, R., and E. Johnson (1990), \Gambling with the House Money and Trying to Break
Even: The Eects of Prior Outcomes on Risky Choice," Management Science 36, 643-660.
Thaler, R., A. Tversky, D. Kahneman, and A. Schwarz (1997), \The Eect of Myopia and
Loss Aversion on Risk Taking: An Experimental Test", Quarterly Journal of Economics
112, 647-661.
Tversky, A., and D. Kahneman (1981), \The Framing of Decisions and the Psychology of
Choice," Science 211, 453-458.
33
Tversky, A., and D. Kahneman (1992), \Advances in Prospect Theory: Cumulative Repre-
sentation of Uncertainty," Journal of Risk and Uncertainty 5, 297-323.
Vuolteenaho, T. (1999), \What Drives Firm Level Stock Returns?", Working Paper, Uni-
versity of Chicago.
34
8 Appendix: Proof of Propositions
Proof of Proposition 1: In the case of individual stock accounting, we conjecture that in
equilibrium, the riskfree gross interest rate is constant at Rf given by (20) and each stock's
returns have a one-factor Markov structure given by (7), (18), and (19), with fi() satisfying
(21) for all zi;t. We then show that under rational expectations, each agent indeed consumes
their labor income and the dividend of each stock and holds the total supply of assets at
each time t.
Each agent's optimization problem is
1 " #
X Ct1
X
n
max E
fC ;fS gg
t
1
+ b Ct 0
t+1
v (Xi;t+1 ; Si;t ; zi;t ) ; (53)
t i;t
t=0 i=1
2 !
3 " #
Et 4Ri;t+1 C t+1 5 = Et 1 +ff(iz(zi;t) ) eg +1 i+ i i;t+1
e
(gc +c t+1 )
Ct i i;t
" #
= e gi
gc
Et E[e
c t+1
ji;t ]e i i;t+1 1 + fi(zi;t ) +1
+1
fi (zi;t )
" #
= e gi
gc + 21
2 c2 (1 !ci
2)
Et 1 +ff(iz(zi;t) +1 ) e
! +1 :
( i ci c ) i;t
i i;t
35
Applying this, we nd that the strategy of consuming C t and holding the market supply of
all securities indeed satises Euler equation (56) for all i.
Proof of Proposition 2: In the case of portfolio accounting, we conjecture that in equi-
librium, the riskfree gross interest rate is constant at Rf given by (29), each stock's returns
have a one-factor Markov structure given by (14), (27), and (28), with fi() satisfying (30) for
all zi;t, and aggregate stock returns are given by (33) with f () satisfying (34) for all zt . We
then show that, under rational expectations, each agent indeed consumes all the dividend of
each stock and holds the total market supply of assets at each time t.
Each agent's optimization problem is
1 " #
X Ct1
max E
fC ;fS gg
t
1
+ b Ct 0
t+1
v (Xt+1 ; St ; zt ) ; (57)
t i;t
t=0
The choice of the riskfree rate in (29) allows us to satisfy Euler equation (59), given
(16). A similar procedure to that used in the proof of Proposition 1 for individual stocks
can be used to prove that the strategy of consuming C t and holding the market supply of
all securities indeed satises Euler equation (61) under the conjectured Markov structure of
market returns.
36
For individual stock i, assume that its log dividend growth process is joint-normally
distributed with the log aggregate dividend and consumption growth processes according to
(15)-(17) and (24)-(26). Then, we have
h i 1 2 2 )+i !ip t+1
E eg i+ i i;t+1
jt+1 = eg i+ 2 (1 !ip
; (62)
and
h i
E e
(gc +c t+1 ) gi +i i;t+1
e jt+1 (63)
gc + 21
2 c2 (1 2 )+ 1 i2 (1 2)
= eg i !cp 2 !ip
c i (!ci !cp !ip )
e( !
i ip
c !cp )t+1
: (64)
Applying these two equations, and doing iterated expectations by observing that zt is +1
determined by zt and t given the conjectured Markov structure of stock returns, we can
+1
show that the Euler equation (60) for stock i reduces to (30) and is indeed satised by the
strategy of consuming C t and holding the market supply of all stocks.
37
Table 1: The table shows the parameter values used in the simulations.
The top panel contains parameters that govern the dividend and aggregate
consumption processes; the bottom panel lists the preference parameters.
Parameter
gc 1.84%
c 3.79%
gi -0.91%, 8i
i 25.0%, 8i
!ci 0.072, 8i
!ij 0.23, 8i; j
gp 1.5%
p 12.0%
!cp 0.15
!ip 0.48, 8i
1.0
0.98
2.25
k 3
0.9
b0 0.7
38
Table 2: The table reports the moments of simulated asset returns under
individual stock accounting and under portfolio accounting, with empirical
values alongside for comparison.
39
Table 3: The table reports the cross-sectional features of simulated returns,
allowing for dispersion in both the mean and volatility of rm level dividend
growth rates.
Individual Portfolio
Accounting Accounting
Cross-section
Value Premium 19.1% 0.5%
Size Premium 5.9% 0.9%
De Bondt/Thaler Premium 14.4% 0.7%
Individual Portfolio
Accounting Accounting
Cross-section
Value Premium 15.1% 0.2%
Size Premium 2.8% 0.2%
De Bondt/Thaler Premium 12.6% 0.4%
40
Table 5: The table reports the standard deviation, in simulated data, of a
strategy which each year, sorts stocks by price-dividend ratio and buys the
bottom decile (value stocks) and shorts the top decile (growth stocks). The
calculation is done for the case when rm level cash
ows follow a one-factor
structure and for the case where they have a multi-factor structure which
includes industry shocks. There are n stocks in total, and S = 5 industries.
41
zt<1
0.5 zt=1
z >1
t
0.4
0.3
0.2
0.1
Utility
−0.1
−0.2
−0.3
−0.4
−0.5
−0.25 −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2 0.25
Gain/loss
Figure 1. Utility of gains and losses when the investor has prior gains (dash-dot
line), prior losses (dashed line), or neither prior gains nor prior losses (solid line).
42