Electronic copy available at: http://ssrn.

com/abstract=899273
A Behavioral Foundation of Reward-Risk Portfolio
Selection and the Asset Allocation Puzzle

Enrico De Giorgi

University of Lugano
Thorsten Hens

University of Zurich
Janos Mayer

University of Zurich
First Draft: June 10, 2005
This Version: June 10, 2008

Financial support from the BSI Gamma Foundation, the Foundation for Research and Development of
the University of Lugano, the University Research Priority Programme Finance and Financial Markets
of the University of Zurich, and the National Center of Competence in Research Financial Valuation and
Risk Management (NCCR-FINRISK) is gratefully acknowledged. We are grateful to Thierry Post, Hersh
Shefrin, Meir Statman and Siegfried Trautmann, as well as to seminar participants at the Universities of
Zurich, Lugano and St. Gallen and to participants at the 2006 Zurich Wealth Forum, at the 2006 NHH
Workshop in Bergen, at the BSI Gamma Foundation Conference on Behavioral Finance held in Frankfurt,
and at the EFA 2006 Conference for their useful comments and suggestions.

Swiss Finance Institute, University of Lugano, via Bu 13, CH-6900 Lugano, Switzerland, Tel. +41 +58
6664677, Fax. +41 +58 6664734 and NCCR-FINRISK, email: [email protected] (corresponding
author).

Swiss Finance Institute Professor at the Swiss Banking Institute, University of Zurich, Plattenstrasse
32, CH-8032 Z urich, Switzerland, Tel. +41 +44 6343706, Fax. +41 +44 6374970, and Norwegian School of
Economics and Business Administration, Helleveien 30, N-5045, Bergen, Norway, email: [email protected].

Institute for Operations Research, University of Zurich, Moussonstrasse 15, CH-8044 Z urich, Tel. +41
+44 6343774, Fax. +41 +44 6344920, email: [email protected].
1
Electronic copy available at: http://ssrn.com/abstract=899273
A Behavioral Foundation of Reward-Risk Portfolio Selection and
the Asset Allocation Puzzle
Abstract
In this paper we suggest a behavioral foundation for the reward-risk approach to
portfolio selection based on prospect theory. We identify sucient conditions for two-
fund separation in reward-risk models in general, and for the behavioral reward-risk
model in particular. It is shown that a prospect theory investor with piecewise-power
function satises two-fund separation if the reference point is the risk-free rate, while
two-fund separation fails if the reference point is higher than the risk-free rate. We
derive a multiple-account version of the behavioral reward-risk model and we perform
an empirical analysis on U.S. data to show that this model explains the asset allocation
puzzle.
Keywords: Portfolio selection, asset allocation puzzle, prospect theory, coherent risk
measures, mental accounting.
JEL Classication: G11, D81.
2
Introduction
The modern portfolio theory of Markowitz (1952) is a rich source of intuition and also the
basis for many practical decisions. Markowitzs seminal idea was to evaluate portfolios by
using two opposing criteria: reward and risk.
1
From a reward-risk perspective investors may
dier with respect to the degree they are willing to trade o reward against risk but all
investors will choose from the set of ecient portfolios: those portfolios that for any given
level of risk achieve maximal reward. Moreover, under certain conditions, the reward-risk
model of portfolio selection leads to two-fund separation, i.e., all investors hold a combination
of the same portfolio of risky assets combined with the risk-free asset. Two-fund separation
greatly simplies the advice one should give to a heterogenous set of investors since the
proportion of risky assets in the optimal portfolio is then independent of investors risk
aversion. Moreover, two-fund separation implies a simple asset pricing structure in which a
single risk factor based on the optimal portfolio of risky assets explains the rewards investors
get in a nancial market equilibrium.
The purpose of this paper is threefold. First, we derive the conditions for two-fund
separation in a general reward-risk model. Second, we suggest giving the reward-risk model
a behavioral microeconomic foundation based on prospect theory (Kahneman and Tversky
1979). Finally, we combine the behavioral risk-reward model with the behavioral portfolio
theory of Shefrin and Statman (2000) to provide a behavioral solution to a well known puzzle
arising in the mean-variance model. Even though practitioners adhere to the reward-risk
3
methodology, their advices do not seem to follow the two-fund separation property.
2
This
so-called asset allocation puzzle was rst observed by Canner, Mankiw, and Weil (1997) who
found that in practitioners advice the more risk-averse the client, the larger the bonds-to-
stocks ratio. Moreover, Wang (2003) identies an asset allocation sub-puzzle, i.e., in nancial
advisors recommendations the proportion of large cap stocks relative to total holding of
stocks increases with increasing risk aversion.
In the general reward-risk model we consider, investors preferences are increasing func-
tions of a reward measure and decreasing functions of a risk measure. This justies the
terminology reward and risk, i.e., investors utility increases when reward is higher
while it decreases when risk is higher. We show that two-fund separation holds if reward
and risk measures can be transformed by means of strictly increasing functions into pos-
itively homogeneous, translation invariant or translation equivariant functionals. Positive
homogeneity says that reward or risk of a positive number of units of a given asset cor-
respond to the same number of units multiplied by reward or risk of a single position in
the asset. Translation equivariance says that adding a risk-free position to a portfolio in-
creases or decreases reward or risk by the same amount as the risk-free addition. Finally,
translation invariance says that reward or risk are unchanged when adding a risk-free po-
sition to the portfolio. Translation equivariance is often assumed when the risk measure is
interpreted as risk capital requirement: the capital requirement diminishes when risk-free
positions are added to the portfolio. By contrast translation invariance often holds for risk
measures used in reward-risk models since the addition of risk-free positions is captured
4
by the reward measure, while the risk measure remains unchanged. Several reward and risk
measures introduced in the literature satisfy positive homogeneity and translation invariance
or equivariance. Grootveld and Hallerbach (1999) and Brogan and Stidham (2008) show for
example that two-fund separation holds for the mean-lower partial moment (LPM) portfolio
model of Harlow and Rao (1989) if the target value corresponds to the risk-free asset, or
to the expected portfolio return. For those target values, LPM can be transformed into a
positively homogeneous, translation invariant risk measure, so this is a special case of our
results. Our ndings also imply that two-fund separation holds if investors possess mean-risk
preferences, where risk is evaluated according to coherent risk measures (Artzner, Delbaen,
Eber, and Heath 1999), Choquet risk measures (De Giorgi 2005), or general deviation mea-
sures (Rockafellar, Uryasev, and Zabarankin 2006). Since these latter measures are dened
based only on few principles of rationality, it follows that two-fund separation is a common
property to many rational mean-risk models, including even those which are consistent with
second order stochastic dominance as the mean-risk model introduced by De Giorgi (2005).
This result is surprising because strong conditions on investors utility functions are needed in
order to obtain two-fund separation within expected utility theory (Cass and Stiglitz 1970).
It also follows that deviations from two-fund separation are dicult to explain within many
(rational) reward-risk models and not only within the mean-variance model, conrming the
diculties encountered by Canner, Mankiw, and Weil (1997) in nding a rational model of
preferences that explains the asset allocation puzzle. This motivates the development of a
behavioral reward-risk model.
5
The behavioral reward-risk model that we propose in this paper is based on the trade-
o between gains and losses that is incorporated in the prospect theory of Kahneman and
Tversky (1979), or any reference-dependent choice theory.
3
In this paper we focus on the
prospect theory. The behavioral reward-risk model could also be developed using cumulative
prospect theory (Tversky and Kahneman 1992) and our results will apply to this latter model
as well. Also note that common applications of cumulative prospect theory in nance ignores
the probability weighting function and in this case prospect theory is identical to cumulative
prospect theory.
4
For given asset payos, we dene reward as the prospect theory value
function applied over gains, while risk is the negative of the prospect theory value function
applied over losses and normalized by the index of loss aversion. The normalization for losses
implies that loss aversion describes the investors tradeo between gains and losses, while
it doesnt impact how investors measure losses. Gains and losses are dened with respect
to a subjective reference point, which describes investors target returns or aspiration levels.
Consequently, in our framework, the risk measure describes the risk that assets payos are
below the reference point. This also relates to the bounded rationality model of Simon (1955)
and to the satiscing measures studied by Brown and Sim (2008).
The reward and risk measures that we dened based on prospect theory obviously depend
on the parametrization of prospect theory by means of the utility index, as well as on the
choice of the reference point which denes what is perceived as a gain and what is perceived
as a loss. As a straightforward application of our result for the general reward-risk model, we
show that positively homogeneous utility indexes and positively homogeneous and translation
6
equivariant reference points imply the separation property for optimal portfolio allocations.
Examples of reward and risk measures leading to this result are those dened by means of
the piecewise-power value function suggested by Tversky and Kahneman (1992) and having
the risk-free return as reference point, which are common specications of prospect theory
in behavioral nance.
In general, optimal solutions to the behavioral reward-risk model do not satisfy the two-
fund separation property. For example, if reward and risk measures are dened according to
a piecewise-power value function and the reference point is higher than the risk-free return
and identical for all investors, then two-fund separation is violated. Nevertheless, when
the reference point is identical for all investors, deviations from two-fund separation in the
behavioral reward-risk model are not systematic, e.g., the bonds-to-stocks ratio and the ratio
between large cap stocks and the total holding of stocks are not monotonic as a function of
investors loss tolerance. However, this is due to the assumption that all investors possess
the same reference point, which is not realistic, e.g., we expect investors with higher loss
tolerance to also possess higher reference points. Consequently, we extend the behavioral
reward-risk model in order to allow investors possessing dierent or even multiple (if multiple
investment goals exist) reference points.
We combine the behavioral reward-risk model with the multiple-account version of the
behavioral portfolio theory of Shefrin and Statman (2000). According to this theory investors
possess dierent mental accounts which correspond to dierent aspiration levels, investments
goals, or, in our framework, reference points.
5
Low aspiration accounts refer to need for
7
security, while high aspiration accounts refers to hope for richness. In the behavioral reward-
risk model, risk increases with the reference point, since the value and the probability of losses
obviously increase when the reference point is higher. Therefore, investors with low degrees
of loss tolerance mainly invest on accounts with low reference points, while investors with
higher degrees of loss tolerance put a higher proportion of their wealth into accounts with
high reference points. For each account, investors determine the minimum risk portfolio.
Indeed, given their reference point for the corresponding mental account, investors goal is
to minimize the risk of being below the reference point, while higher reward refers to mental
accounts with higher reference points or aspiration levels. Finally, investors allocate their
wealth between the dierent accounts in order to maximize their total reward, given the
loss constraint implied by their loss tolerance. This step is a simple linear program since
investors treat mental accounts separately, e.g., aggregate the dierent accounts ignoring
co-movements between the payos of accounts specic portfolios.
We perform an empirical analysis on US data. We specify the behavioral reward-risk
model using the piecewise-exponential value function, as suggested by De Giorgi, Hens, and
Levy (2003), K obberling and Wakker (2005) and De Giorgi, Hens, and Post (2005). We
dene investors account using deterministic reference points, which corresponds to various
target returns. We compute optimal portfolios as a function of investors loss tolerance
according to the behavioral reward-risk model with multiple accounts. We show that the
bonds-to-stocks ratio is lower in high aspiration accounts relative to low aspiration accounts.
Similarly, the ratio between large cap stocks and the total holding of stocks is higher for low-
8
to-medium aspiration accounts relative to very high aspiration accounts. Since investors with
lower loss tolerance mainly invest in low aspiration accounts, their portfolios also present a
lower bonds-to-stocks ratio and a lower ratio between large cap stocks and total holding of
stocks. These ndings are consistent with the nancial advisors recommendations reported
by Canner, Mankiw, and Weil (1997) and Wang (2003).
I Related literature
This paper contributes to the literature on portfolio selection in a static reward-risk model.
Since the seminal work of Markowitz (1952) several authors have suggested to replace the
variance with other measures of risk. Already Markowitz (1959) realized that variance might
not be an appropriate measure of risk because it considers both positive and negative devi-
ations from the mean as risky outcomes, while positive deviations are obviously welcome by
investors. Markowitz (1959) suggested to replace variance with lower semi-variance, which
only measures negative deviations from the mean. Other authors followed by suggesting gen-
eral lower partial moments (Bawa and Lindenberg 1977, Harlow and Rao 1989), value-at-risk
(Jorion 1997), coherent risk measures (Artzner, Delbaen, Eber, and Heath 1999), convex risk
measures (F ollmer and Schied 2002, Frittelli and Rosazza Gianin 2004), spectral risk mea-
sures (Acerbi 2002), Choquet risk measures (De Giorgi 2005), or general deviation measures
(Rockafellar, Uryasev, and Zabarankin 2006). By contrast, only few authors suggested alter-
native measures of reward. De Giorgi (2005) proposes an axiomatization of reward measures
9
based on second order stochastic dominance, but shows that only the expected value satises
his axioms.
While some of these extensions of Markowitz (1952) have been justied from the reward-
risk perspective, i.e., by suggesting well-behaved measures of reward and risk to replace
mean and variance, respectively, they have however no microeconomic foundation based
on expected utility theory, because they usually violate second-order stochastic dominance.
6
Moreover, for most of these cases it is not known whether the central property of the reward-
risk model of Markowitz (1952), the two-fund separation, still holds. Generalizations of
mutual-fund separation have mainly been derived for the microeconomic model of portfolio
selection, generalizing the case of quadratic utility to expected utility with hyperbolic ab-
solute risk aversion. The seminal paper in this literature is Cass and Stiglitz (1970). For a
recent monograph including the expected utility approach to two-fund separation see Gollier
(2001).
This paper also contributes to the literature on the asset allocation puzzle. In their
seminal work, Canner, Mankiw, and Weil (1997) relax the key assumptions leading to the
mean-variance two-fund separation theorem, which are: (i) existence of a riskless asset, (ii)
mean-variance objective functions, (iii) investors use historical distributions, (iv) assets can
be freely traded (i.e., there are no short-sale or borrowing constraints), (v) investors operate
over a one-period planning horizon, (vi) there is no background risk like human capital.
However, they conclude that deviating from these assumptions does not provide satisfactory
explanations of the recommended portfolio allocations, in particular of the relationship be-
10
tween the bonds-to-stocks ratio and risk aversion. Indeed, the authors state that it is hard
to explain recommended portfolio allocations with a rational model.
Several authors have suggested solutions to the asset allocation puzzle. We discuss here
some of them and motivate why we think that our behavioral solution adds to this litera-
ture. Some authors are skeptical about the conclusion of Canner, Mankiw, and Weil (1997)
concerning the inconsistency of nancial advisors with respect to the modern portfolio the-
ory. Elton and Gruber (2000), for example, claim that the bonds-to-stocks ratio test is not
sucient to assert that nancial advisors do no follow the modern portfolio theory. In par-
ticular they point out that violations of two-fund separation might result from constraints
that advisors are obliged to satisfy, like short-sale constraints, and the bonds-to-stocks ratio
can be increasing or decreasing depending on the set of historical data or forecasted expected
returns used to derive optimal strategies. They conclude that in order to test deviations of
advisors recommendations from the modern portfolio theory of Markowitz (1952) one should
also take into account the input data used by nancial advisors to derive the recommend
allocations. Indeed, Siebenmorgen and Weber (2003) asked German advisors to provide
both recommended allocations for dierent investors and input data used to compute them.
They found that recommended allocations were dicult to explain within the mean-variance
model of Markowitz (1952) even when using the input data provided by the advisors.
Shalit and Yitzhaki (2003) relax the assumption about mean-variance preferences and test
the eciency of nancial advisors portfolio allocations with respect to second-order stochas-
tic dominance (SSD). They show that these allocations are not inecient, i.e., not dominated
11
with respect to SSD by any alternative allocation. Therefore, even if the recommended allo-
cations are not mean-variance ecient, they are optimal for at least one risk-averse expected
utility maximizer, i.e., there is no alternative allocation that is preferred by all risk-averse
investors. Our concern about this approach is that the set of portfolio allocations that are
not dominated by others with respect to second order stochastic dominance is large. Thus,
several portfolios could be justied using second order stochastic dominance, while the fact
that the bonds-to-stocks ratio of recommended allocations show a specic shape as function
of risk tolerance remains unexplained. Note that also in the mean-risk model introduced by
De Giorgi (2002) investors select portfolios that are not dominated with respect to second
order stochastic dominance, but two-fund separation is satised.
Other authors argue that a static portfolio model is not able to capture important aspects
of the portfolio decision process and suggest relaxing assumption (v) above, advocating inter-
temporal hedging activities as a solution of the asset allocation puzzle. Brennan and Xia
(2000, 2002), Campbell and Viceira (2001, 2002), and Bajeux-Besnainou, Jordan, and Portait
(2001) consider a dynamic model with stochastic interest rate where bonds can be used to
hedge against the interest rate risk and show that the bonds-to-stocks ratio increases with
risk aversion due to the hedging component of investors optimal portfolios. Mougeot (2003)
introduces a dynamic model with stochastic interest rate and estimation risk, i.e., uncertainty
about markets excess return, and shows that hedging components for estimation risk and
interest rate risk can rationalize the asset allocation puzzle.
While dynamic models of portfolio selection represent a theoretical framework for solving
12
the asset allocation puzzle that is very appealing to economists, Lioui (2007) shows that the
results obtained are mainly driven by the assumption that bonds perfectly hedge the interest
rate risk and the market price of risk is constant, and these assumptions lack empirical
support. He shows that in a more realistic dynamic model for portfolio selection the asset
allocation puzzle might even be more puzzling. Moreover, Wang (2003) points out that inter-
temporal hedging might help explaining the bonds-to-stocks ratio, but it cannot explain why
the proportion of large cap stocks relative to the total holding of stocks increases with risk
aversion in nancial advisors recommendations.
Gomes and Michaelides (2004) provide a human capital explanation of the puzzle: in-
vestors face a stochastic uninsurable labor income and more risk averse households invest a
smaller percentage of their assets in stocks since they prefer labor income substitutes such
as long-term bonds. Again, introducing human capital does not help understanding the
sub-puzzle identied by Wang (2003).
This paper suggests a solution to the asset allocation puzzle by addressing assumption
(ii) above for mean-variance two-fund separation, i.e., that investors possess mean-variance
objective functions. Indeed, we provide a behavioral explanation of the asset allocation
puzzle and the sub-puzzle assuming that investors possess reward-risk preferences founded
in the prospect theory of Kahneman and Tversky (1979) and, additionally, have dierent
mental accounts which correspond to dierent aspiration levels, or reference points. We are
not the rst suggesting a behavioral explanation to the asset allocation puzzle. Siebenmorgen
and Weber (2003) use a static mean-variance portfolio selection model where investors are
13
assumed to calculate portfolios variance without taking correlations into account (pure risk),
combined with a naive diversication criterion.
7
Wang (2003) also analyzes investors choices
in a static mean-variance portfolio model with pure risk (correlations are ignored), but
instead of using naive diversication, investors are assumed to be averse to extreme losses.
In the model of Wang (2003), aversion to extreme losses is given by a worst-case threshold
which is assumed to depend on investors tolerance to volatility. However, the author does
not give any functional form to the relationship between volatility aversion and loss aversion.
Our solution to the asset allocation puzzle diers from these papers by the fact that
we depart from the mean-variance setup of Markowitz (1952) and suggest a behavioral
reward-risk model that is founded in the prospect theory, using the reference point which
characterizes prospect theory preferences to dene investors mental accounts (Thaler 1985,
1999). Extensive experimental evidence supports prospect theory as a descriptive model of
decision-making under risk, showing that people have dierent attitudes to risk when facing
gains or losses, and that people are loss averse, i.e., dislike symmetric payos around their
reference points. By contrast, mean-variance preferences treat gains and losses in the same
way. Prospect theory also represents the natural framework to study portfolio selection
with multiple mental accounts, which reect dierent investment goals or reference points.
Finally, aversion to extreme losses can be easily incorporated into prospect theory.
8
Our
approach is also related to the recent behavioral literature applying prospect theory and loss
aversion to explain low participation in equity markets, under-diversication, the disposition
eect and the equity premium puzzle: see Benartzi and Thaler (1995), Barberis, Huang, and
14
Santos (2001), Gomes (2005), Berkelaar, Kouwenberg, and Post (2004), Barberis, Huang,
and Thaler (2006), Barberis and Huang (2008a, 2007), Jin and Zhou (2007), Dimmock and
Kouwenberg (2007). We add to this important literature by showing that prospect theory
also helps explaining the asset allocation puzzle and the sub-puzzle.
The remainder of the paper is organized as follows. In Section II we introduce the general
reward-risk model for portfolio selection and we derive sucient conditions for reward and
risk measures in order to obtain the two-fund separation property for optimal allocations.
In Section III we dene the behavioral reward-risk model which is based on prospect theory.
Section IV presents an empirical analysis of the asset allocation puzzle on US data. Section V
concludes our proposal. All proofs are given in the Appendix.
II General reward-risk model and two-fund separation
We consider a two-period nance economy. Let = {1, . . . , S}, S < , denote the states
of nature in the second period. F = 2

is the power algebra on , i.e., the set of all possible

events arising from . Uncertainty is modeled by the probability space (, F, P), where the
probability measure P on satises p
s
= P
_
{s}

> 0 for all s = 1, . . . , S, i.e., every state

of the world has strictly positive probability to occur. The space of real-valued measurable
functions on (, F, P) is denoted by G.
There are K + 1 assets with payos A
k
and prices q
k
, k = 0, . . . , K. R
k
=
A
k
q
k
is the
gross return of asset k and R = (R
0
, . . . , R
K
)

is the vector of gross returns. We denote

15
by R
1
= (R
1
, . . . , R
K
)

the vector of risky gross returns. Asset 0 is the risk-free asset with
payo A
0
= 1 and price q
0
=
1
1+r
, where r = R
0
1 is the risk-free rate of return. We
assume that short-sale is allowed and the set of attainable portfolio payos corresponds to
X = {

K
k=0

k
A
k
|(
0
, . . . ,
K
)

R
K+1
}.
9
Let : G R and : G R be real-valued functions on X. We call the reward
measure and the risk measure. We state the following assumption:
Assumption 1 ((, )-preferences). The investors objective function can be represented as
U(X) = V ((X), (X)), for all X X
where V : R
2
R is continuously dierentiable and satises for all and
(i)
V

(, ) > 0,
(ii)
V

(, ) < 0,
(iii) V (, ) is quasi-concave.
Assumption 1 simply states that investors evaluate assets payos only through reward
and risk measures and , respectively. The trade-o between reward and risk is specied
by the function V , which must be increasing in the reward measure and decreasing in the risk
measure, i.e., the value of assets payos increases when reward increases and risk decreases,
in the spirit of the reward-risk analysis. Mean-variance preferences are a well-known example
of preferences satisfying Assumption 1. In this case, (X) = E
_
X

, (X) = Var(X) and, e.g.,

16
V (, ) = for some > 0. Therefore, Assumption 1 generalizes the mean-variance
model of Markowitz (1952) to any measure of reward and risk, respectively.
The investor maximizes her objective function under her budget constraint, i.e., the
portfolio choice problem is
max
XX
U(X), q(X) w
0
, (1)
where w
0
> 0 is the investors initial wealth and q : X R is the linear pricing functional
on X, i.e., for X X, X =

K
k=0

k
A
k
for R
K+1
, we have q(X) =

K
k=0

k
q
k
.
10
An alternative representation of the portfolio optimizing problem (1) is to maximize the
reward given a maximal level of risk. The precise relation of these two perspectives is given
by the following observations. Given Assumption 1, for any solution X

to (1) there exists

a parameter such that X

solves
max
XX
(X) s.t. (X) , q(X) w
0
. (2)
This can be easily seen by putting = (X

), where X

is a solution to the optimization

Problem (1). Hence the investors reward-risk goal function V determines the risk tolerance
. On the other hand, we need stronger conditions on , , and V such that any optimal
solution to (2) also solves (1). We will discuss some advantages of representation (2) in
Section III, where we introduce a (, )-representation of prospect theory. We assume that
the budget restriction in (2) holds with equality, i.e., q(X) = w
0
for any optimal allocation.
This assumption is satised if the reward measure increases when units of the risk-free asset
are added to the portfolio.
17
Let
k
=

k
q
k
w
0
be the proportion of initial wealth invested in asset k for k = 0, . . . , K and
let = (
0
, . . . ,
K
)

. Then, the (, )-optimization problem (2) can equivalently be written

as follows:
max
R
K+1
(

Rw
0
) s.t. (

Rw
0
) ,

e = 1, (3)
where e = (1, . . . , 1)

R
K+1
and

Rw
0
is the portfolios payo for the investment and
initial wealth w
0
.
The following denition introduces some basic properties of reward and risk measures.
Denition 1. Let : G R be a real-valued function on G. We say that is:
(i) translation invariant
11
if
(X +a) = (X)
for all a R and X G,
(ii) translation equivariant if
(X +a) = (X) +a
for all a R and X G, or
(X +a) = (X) a
for all a R and X G. In the rst case, we say that is positive translation
equivariant, while in the second case it is negative translation equivariant.
When reward or risk measures are translation invariant, then adding a risk-free position
to the portfolio doesnt change risk or reward. By contrast, when reward or risk measures are
18
translation equivariant, then adding a risk-free position increases (positive translation equiv-
ariant) or decreases (negative translation equivariant) reward or risk by the same amount as
the risk-free addition. Negative translation equivariance is often assumed for risk measures
that are interpreted as risk capital requirement: adding a risk-free position to the current
portfolio reduces the capital requirement. Positive translation equivariance and translation
invariance are satised by many reward and risk measures, respectively, used in reward-risk
models for portfolio selection.
If investors possess mean-variance preferences, i.e., (X) = E
_
X

and (X) = Var(X) in

Assumption 1, any solution X

to the portfolio selection problem (2) satises the two-fund

separation property (Tobin 1958), i.e.,
X

=
0
w
0
R
0
+ (1
0
) w
0
X
M
,
where X
M
X. Written in terms of portfolio shares

(i.e., X

= R

w
0
), if no redundant
assets exist, then the two-fund separation property implies:

=
0
e
0
+ (1
0
)
M
,
where
M
satises X
M
= R

M
, e

M
= 1, and e
0
= (1, 0, . . . , 0)

R
K+1
. Consequently,
mean-variance investors optimally allocate their wealth between the risk free asset and a
single portfolio
M
, which is also called the tangency portfolio (for a detailed description,
see De Giorgi 2002). The proportion of wealth invested in the tangency portfolio depends on
the investors risk aversion, which is measured by means of her target risk (see Equation
(2)). By contrast, the tangency portfolio does not depend on investors preferences, but only
19
on investors beliefs concerning expected returns, variances, and correlations between asset
returns.
The mean-variance portfolio choice model is a special case of a general model with reward-
risk preferences. However, as we will show in this section, two-fund separation does not
require mean-variance preferences but also holds for the general reward-risk model under
few assumptions on reward and risk measures. We now investigate these assumptions and
we give sucient conditions on reward and risk measures in order to derive the separation
property of optimal portfolio choices.
Proposition 1. Let : G R and : G R be a reward and risk measure on G,
respectively. Suppose that strictly increasing transformations T

and T

exist, such that

= T

and = T

are positively homogeneous of degree 1 and satisfy one of the

following two properties:
(i) translation invariance,
(ii) translation equivariance.
Then, there exists a portfolio allocation

p
, such that for any (, )-optimal portfolio alloca-
tion

, there exists
0
with

=
0
e
0
+ (1
0
)

p
,
where e
0
= (1, 0

R
K+1
.
Proposition 1 states that two-fund separation holds in reward-risk models when the ef-
cient frontier is a straight line in the reward-risk diagram, after transforming reward and
20
risk measures by means of strictly increasing transformations. Since increasing transforma-
tions change the risk-reward trade-o but not the ranking of asset payos, the transformed
reward-risk model delivers the same ecient frontier as the original reward-risk model.
For many applications the reward and the risk measures are themselves homogeneous so
that it makes sense to emphasize this case. Two-fund separation can then easily be derived
from Proposition 1.
Corollary 1.
Let : G R and : G R be reward and risk measures on G, respectively. Suppose that
one of the following two properties holds for both and :
(i) translation invariance and (positive) homogeneity of degree > 0,
(ii) Translation equivariance and (positive) homogeneity of degree 1.
12
Then two-fund separation holds.
Note that the expected value (the mean) is a linear measure, thus it is positively ho-
mogeneous of degree 1 and translation equivariant. Variance is positively homogeneous of
degree 2 and translation invariant, since in fact Var(X) =
2
Var(X) for all R and
Var(X + a) = Var(X) for all a R. Consequently, according to Corollary 1 two-fund sep-
aration holds for mean-variance optimal allocations, as discussed at the beginning of this
section.
The same argument applies to the lower partial moment with the risk-free return as target
21
value. In fact, the lower partial moment with target is dened by:
LPM(X; , ) =
S

s=1
p
s
[max(0, (X) X(s))]

(4)
for some 1. In general might depend on portfolio X, thus it is a function on the space
of asset payos. If the target value is the risk-free return, then (X) = q(X) R
0
where q(X)
is the price at time 0 for asset X.
13
Consequently,
LPM(X +a; , ) =
S

s=1
p
s
[max(0, q(X +a)R
0
X(s) a)]

=
S

s=1
p
s
[max(0, q(X)R
0
+a X(s) a)]

= LPM(X; , )
for all a R. Moreover,
LPM(X; , ) =
S

s=1
p
s
[max(0, q(X)R
0
X(s))]

s=1
p
s
[max(0, q(X)R
0
X(s))]

LPM(X; , )
for all > 0. Thus, the lower partial moment with (X) = R
0
q(X) is translation invariant
and positively homogeneous of degree . The same conclusion applies to any target value
(X) which is positive translation equivariant and positively homogeneous of degree one as,
for example, (X) = E
_
X

or (X) = V aR

(X), i.e., the -quantile of X. V aR

(X)
corresponds to the value-at-risk of X with condence level , which is dened as negative of
the -quantile. These results on two funds separation for LPM generalizes the observation
22
made by Grootveld and Hallerbach (1999) and Brogan and Stidham (2008). Note that for
(X) = V aR

(X) and = 1 we have

LPM(X; 1, V aR

(X)) =
S

s=1
p
s
[max(0, V aR

(X) X(s))]
= E
_
X 1
{XV aR

(X)}

V aR

(X) P
_
X V aR

(X)

= (ES

(X) V aR

(X))
where ES

(X) =
1

_
E
_
X 1
{XV aR

(X)}

+V aR

(X)
_
P
_
X V aR

(X)

__
is the
expected shortfall with condence level (Acerbi and Tasche 2002). The relationship be-
tween LPM, V aR and ES has already been derived in Jarrow and Zhao (2006), who con-
sider the special case P
_
X V aR

(X)

= . Expected shortfall is a coherent measure

of risk. Coherent risk measures (Artzner, Delbaen, Eber, and Heath 1999), which became
very popular in the mathematical nance literature and also among practitioners, are by
denition positively homogeneous of degree one and translation equivariant. Similarly, Cho-
quet measures (De Giorgi 2005) and general deviation measures (Rockafellar, Uryasev, and
Zabarankin 2006) are positively homogeneous of degree one and translation invariant. There-
fore, on the basis of Proposition 1, two-fund separation also holds if investors have mean-risk
preferences where risk is measured by means of any coherent, Choquet or general deviation
measure. These measures are considered well-behaved measures of risk since they are dened
based on only few principles or axioms which correspond to minimal rationality requirements
on the way people should measure risk. Therefore, two-fund separation is a common prop-
erty of rational reward-risk models, including even those which are consistent with second
23
order stochastic dominance like the mean-risk models introduced by De Giorgi (2005). At
a rst glance, this result might appear surprising because strong assumptions on investors
utilities are needed to obtain two-fund separation under expected utility theory preferences
(Cass and Stiglitz 1970). However, the normative framework suggested in the literature
to dene well-behaved measures of reward and risk diers from the one of expected utility
theory, and rational reward-risk models arise from expected utility theory only under strong
assumptions on investors utility functions or on the distribution of assets payos. The
observation that two-fund separation holds for rational reward-risk models also conrms the
diculties encountered by Canner, Mankiw, and Weil (1997) in nding a rational model of
preferences which explain the asset allocation puzzle. This also motivates our approach, i.e.,
providing to the reward-risk perspective a behavioral foundation.
III The behavioral reward-risk model
The prospect theory by Kahneman and Tversky (1979) and its extension, the cumulative
prospect theory by Tversky and Kahneman (1992), are one of the cornerstones of behavioral
economics. The kinked and convex-concave value function and the probability weighting
function capture well-known deviations from expected utility theory. There exists growing
consensus that prospect theory is superior to expected utility theory in terms of descriptive
validity.
Prospect theory suggests modeling investors preferences as follows:
14
24
Assumption 2 (Prospect Theory Preferences). Investors evaluate portfolio payos accord-
ing to the value function:
V (X) =
S

s=1
v(X(s) RP(X))
s
(5)
for all X X where
(i) v is a two-time dierentiable function on R \ {0}, strictly increasing on R, strictly
concave on (0, ) and strictly convex on (, 0), with v(0) = 0,
(ii)
s
= w(p
s
) and w is a dierentiable, non-decreasing function from [0, 1] onto [0, 1]
with w(p) = p for p = 0 and p = 1 and with w(p) > p (w(p) < p) for p small (large),
(iii) RP : X R is a subjective reference point. If RP(X) = RP for all X, then the
reference point is xed.
The prospect theory assumes that investors code payos in terms of gains and losses,
i.e., payos above and below a subjective reference point, respectively. Moreover, it also
assumes dierent risk attitudes with respect to gains than with respect to losses (reection
principle). These assumptions suggest a natural way to dene a reward-risk trade-o that
describes investors preferences. We rewrite the value function (5) by separating portfolio
outcomes which are above the reference point and portfolio outcomes which are below the
reference point and we obtain:
V (X) =
S

s=1
v (max(0, X(s) RP(X)))
s
+
S

s=1
v (min(0, X(s) RP(X)))
s
(6)
25
The trade-o between gains and losses is determined by investors loss aversion. Following
Benartzi and Thaler (1995), Kobberling and Wakker (2005), and Schmidt and Zank (2008),
we dene the index of loss aversion by:
=
lim
x0
v

(x)
lim
x0
v

(x)
1
and we rewrite equation (6) as follows:
V (X) =
S

s=1
v (max(0, X(s) RP(X)))
s

S

s=1
_
1

_
v (min(0, X(s) RP(X)))
s
.
We now dene reward and risk measures by
PT
+
(X) =
S

s=1
v (max(0, X(s) RP(X)))
s
, (7)
PT

(X) =
1

s=1
v (min(0, X(s) RP(X)))
s
(8)
respectively. PT
+
measures positive deviations of assets returns with respect to the reference
point. By contrast, the reward measure ignores assets payos which are below the reference
point, i.e., losses. Rather than aecting portfolio reward, we believe that losses should impact
portfolios risk. In fact, losses are determined by means of the risk measure PT

. Finally,
the index of loss aversion measures the investors tradeo between gains and losses, since
PT

has been normalized by and therefore does not account for loss aversion. In our
framework, the risk measure gives the risk of being below the reference point, i.e., of failing
the target return. The idea that investors evaluate assets payo based on given aspiration
levels and consider risk as the possibility of failing the aspiration level goes back to the
26
model of bounded rationality introduced by Simon (1955) and has recently been formalized
by Brown and Sim (2008), who introduce so-called satiscing measures.
The reward measure PT
+
is non-negative, monotone, but in general not positively ho-
mogeneous, translation equivariant, or invariant. The risk measure PT

is non-negative,
anti-monotone, but in general not positively homogeneous, translation equivariant, or in-
variant.
15
Moreover, the risk measure PT

is generally not convex. Convexity is a common

assumption for risk measures since it guarantees that investors benet from diversication,
i.e., assign lower risk to diversied portfolios than to the single positions (see F ollmer and
Schied 2002). This reects investors preferences for diversied portfolios, as described by a
concave utility index. However, Kahneman and Tversky (1979) found that people are risk
seeking when facing losses, and this implies that undiversied portfolios might be preferred
to diversied portfolios.
Obviously,
V (X) = PT
+
(X) PT

(X)
and therefore prospect theory investors possess (PT
+
, PT

)-preferences (see Assumption 1).

The portfolio choice problem (2) becomes:
max
R
K+1
PT
+
(

Rw
0
) s.t. PT

Rw
0
) pt

e = 1. (9)
The parameter pt

represents investors risk aversion and is closely related to her index of

loss aversion . In the sequel, we will consider pt

as the parameter reecting investors loss

aversion. We call Problem (9) the behavioral reward-risk model.
27
As discussed in Section II, under some conditions on PT

and PT
+
, optimization prob-
lem (9) is equivalent to the maximization of the investors value function (5) under her
budget constraint, provided that solutions to this latter problem exist. However, if we
specify prospect theory using the piecewise-power value function suggested by Tversky and
Kahneman (1992),
v(x) =
_

_
x

+
x > 0
(x)

x 0
, (10)
and we take the reference point equal to the risk-free return RP(X) = R
0
q(X), investors
would optimally leverage innitely any risky portfolio with strictly positive prospect theory
value (see De Giorgi, Hens, and Levy 2003).
16
The existence of a risky portfolio with positive
prospect theory value can be easily checked. For example, for a yearly horizon, and assuming
the usual parameters for the piecewise-power function, i.e.,
+
=

= 0.88 and = 2.25,

the prospect theory value of the market portfolio of Fama and French (1993) is strictly
positive, if using the empirical distribution estimated on yearly returns from 1927 to 2007
as a proxy for the returns distribution. The reward-risk representation of prospect theory
imposes an additional assumption on investors preferences if compared to the value function
representation. The reward-risk representation additionally assumes that investors loss tol-
erance is bounded. An innite leverage might be optimal from a value function perspective
when losses are compensated by gains. However, in the case of an innite leverage both
gains and losses are unbounded, i.e., investors should be willing and able to suer extreme
28
losses in order to innitely leverage a portfolio. Yet, usually, investors dislikes extreme losses.
Recently, Bosch-Dom`enech and Silvestre (2006) conducted experiments involving lotteries
with large stakes and real payos. Their results suggest that decision makers are risk seeking
with respect to small losses, but become risk averse when large losses are involved. Kahne-
man and Tversky (1979) conducted experiments involving only small stakes or hypothetical
payos, so risk aversion for large losses are not in contrast with their ndings. Risk aversion
for large losses also implies that investors do not take an innite leverage. Similarly, the
reward-risk representation of prospect theory implies that investors do not take an innite
leverage since their loss tolerance is bounded.
The conditions on reward and risk measures given in Proposition 1 in order to obtain
two-fund separation can be reformulated for the utility index and the reference point dening
prospect theory measures PT
+
and PT

. The following corollary sets sucient conditions

on the utility index v and the reference point RP such that (PT
+
, PT

)-optimal allocations
satisfy two-fund separation.
Corollary 2. Let PT
+
and PT

be dened as in equations (7) and (8), respectively. Suppose

that v is positively homogeneous of degree + on R
+
, and positively homogeneous of degree
on R

. Moreover, assume that RP is positive translation equivariant and positively

homogeneous of degree 1, then two-fund separation holds for all optimal solutions to the
portfolio choice problem (9).
The assumption that the reference point is positive translation equivariant implies that
29
adding a sure payo to the current portfolio will aect the reference point by the same
amount. In other words, investors reward and risk measures only capture gains and losses
deriving from the risky part of their portfolios, while certain payos are already incorporated
in the reference point, because investors can achieve them in all states of nature.
17
Corollary 2 immediately implies the following result for Tversky and Kahneman (1992)
specication of the prospect theory value function.
Corollary 3. Let v be dened as in equation (10) and suppose that RP is positive translation
equivariant and positively homogeneous of degree 1. Then two-fund separation holds for all
optimal solutions to the portfolio choice problem (9).
Examples of reference points which are positively homogeneous of degree 1 and posi-
tive translation equivariant are the risk-free return RP(X) = R
0
q(X), the expected value
RP(X) = E
_
X

, and any -quantile RP(X) = F

1
X
() for some (0, 1), where F
X
is the
cumulative distribution function of X and F
1
X
its generalized inverse.
If the utility index v is dened as in equation (10) and there is no probability weighting
(w(p) = p for p [0, 1], see Assumption 2), then
PT

(X) = LPM(X; , RP).

Moreover, for = 1 and any target value RP, the ratio
PT
+
(X)
PT

(X)
corresponds to the omega ratio introduced by Shadwick and Keating (2002) for the valuation
of hedge fund portfolios.
30
IV Empirical application
This section derives the optimal (PT
+
, PT

)-portfolio allocations between cash, bonds,

small-mid caps and large caps using yearly historical returns from 1927 to 2007. We show
that the prospect theory reward-risk model with mental accounting (Thaler 1985, 1999)
oers an explanation to the asset allocation puzzle and the sub-puzzle.
A Data
We use yearly nominal returns of the CRSP market portfolio (market), of a bond index
(bonds), and of the US one-month Treasury Bill (cash). In addition to this, we also
include yearly nominal returns of two Fama and French US common stock portfolios formed
on market capitalization of equity: small-mid caps (small-mid) and large caps (large).
18
The equity data are obtained from Kenneth Frenchs online data library; the bond index
corresponds to the US intermediate-term government bond index maintained by Ibbotson
Associates; T-Bill data are also from Ibbotson Associates. The sample covers the period from
January 1927 to December 2007, with a total of 81 yearly observations. Table I gives the
summary statistics of our data. In our empirical analysis, similarly to Canner, Mankiw, and
Weil (1997), we assume that cash is risk-free with a yearly gross return R
0
= 1 +r = 1.0378
corresponding to the mean gross return of the T-Bill.
[Table 1 about here.]
31
B Prospect theory and two-fund separation
We compute optimal (PT
+
, PT

)-portfolios by solving Problem (9) using historical dis-

tributions as proxies for future returns, i.e., we assume that future returns are uniformly
distributed on the set {r
t
: t = 1927, . . . , 2007}, where r
t
is the vector of observed returns for
year t. We impose short-sale constraints. This is consistent with the observation that nan-
cial advisors recommendations reported by Canner, Mankiw, and Weil (1997) and Wang
(2003) only presents long positions on all asset classes. In general, the behavioral reward-risk
model is non-convex and non-dierentiable and thus we cannot apply Lagrange methods.
We refer to De Giorgi, Hens, and Mayer (2007) for a detailed discussion of the numerical
methodologies used to solve the (PT
+
, PT

)-optimization.
We initially assume that all investors possess the same reference point. For this case, in
order to compare with the results of Section III (in particular Corollary 3), we specify reward
and risk measures PT
+
and PT

, respectively, using the piecewise-power value function of

Equation (10). The reference point is assumed to correspond to a xed rate of return, i.e.,
RP(X) = (1 +) q(X) for some R
+
. If = r (i.e., 1 + = R
0
), Corollary 3 implies that
two-fund separation holds for the reward-risk model. Indeed, in this case, the reference point
is positive homogenous of degree one and translation equivariant. If > r, the reference
point RP(X) = (1 + ) q(X) is positively homogeneous of degree one, but not translation
equivariant.
19
We compute optimal portfolio allocations between cash, bonds and the market
portfolio. Figure 1 reports the bonds-to-stocks ratio and the proportion invested into cash
32
as function of the stock holding, which reects the investors risk tolerance.
[Figure 1 about here.]
The bonds-to-stocks ratio for = r is constant as long as short-sale constraints are not
binding, which is consistent with two-fund separation. However, if > r two-fund separation
is violated and the bonds-to-stocks ratio rst increases and then decreases, i.e., deviations
from two-fund separation are not systematic as described by Canner, Mankiw, and Weil
(1997) and the asset allocation puzzle remains unexplained. We also observe that the bonds-
to-stocks ratio is much higher when the reference point is = 4% compared to the case where
it corresponds to the higher target return 5%. We will analyze below how optimal portfolio
allocations depend on the reference point, but Figure 1 already suggests that investors with
higher reference points and, consequently, higher loss tolerance, will have lower bonds-to-
stocks ratios. We also computed optimal allocations between cash, bonds, small-mid caps and
large caps and we found similar results for the bonds-to-stocks ratio. However, when reward
and risk are dened using a piecewise-power function, we found that optimal portfolios
do not contain large cap stocks. This conrms the ndings of De Giorgi, Hens, and Post
(2005), who show that while cumulative prospect theory with piecewise-power value function
rationalizes the equity premium puzzle (Benartzi and Thaler 1995), the same does not apply
to the size premium puzzle, or the historically favorable risk-return trade-o of small cap
stocks relative to large cap stocks (rst documented by Banz 1981). De Giorgi, Hens, and
Post (2005) solve the size premium puzzle by replacing the piecewise-power function with a
33
piecewise-exponential function.
20
Therefore, in the application of the behavioral reward-risk
model with multiple mental accounts that we present below, we will specify PT
+
and PT

using a piecewise-exponential function.

The assumption that investors possess the same reference point is not realistic. Indeed,
we expect that investors with higher tolerance to losses will also have higher aspiration lev-
els. Moreover, not only investors possess dierent reference points, but they also possess
multiple reference points, depending on their investment goals. Therefore, we combine the
behavioral reward-risk model with the multiple-account version of the behavioral portfolio
theory of Shefrin and Statman (2000). According to this theory investors possess dierent
mental accounts, which correspond to dierent aspiration levels or investment goals. In the
prospect theory reward-risk model this is captured by assuming that investors possess multi-
ple reference points, which characterize their aspiration levels and the corresponding mental
accounts. Low aspiration accounts reect need for security, while high aspiration accounts
reects hope to achieve richness. Risk, as measured by PT

, obviously increases with the

reference point. Consequently, investors with low loss tolerance only invest into low aspira-
tion accounts. For each account, optimal (PT
+
, PT

)-portfolios are derived, where reward

and risk measures are dened using the account-specic reference point. Finally, wealth is
allocated to the dierent accounts according to investors loss tolerance. In performing this
step, investors treat accounts specic portfolios as separate entities ignoring dependencies
between portfolios payos.
We compute optimal (PT
+
, PT

)-portfolios for each account by solving Problem (9), as

34
for the case discussed above where a unique reference point exist. As previously discussed,
we now specify reward and risk measures PT
+
and PT

, respectively, using a piecewise-

exponential value function:
v(x) =
_

_
1 exp(x) x 0
(1 exp(x)) x < 0
. (11)
Again, the reference point is assumed to correspond to a xed rate of return, i.e., RP(X) =
(1 + ) q(X) for some R
+
, which corresponds to the accounts aspiration level. We
take reference points with values ranging from = 3% (mean ination rate over the sample
period) to = 15.5% (approximatively the mean return of small-mid cap stocks). Table II
reports the minimum risk portfolios as function of the reference point, while Figure 2 displays
their risk-reward tradeo.
[Table 2 about here.]
[Figure 2 about here.]
The minimum risk portfolio minimizes PT

for the corresponding reference point. We

compute minimum risk portfolios since investors loss tolerance only determine how wealth
is allocated between the dierent accounts, while given a specic reference point all investors
minimize the risk of missing the reference target return. As we expected, minimum risk
portfolios strongly depend on the choice of the reference point. As the reference point
increases, allocation to cash is reduced, while allocation to bonds and stocks increases and
35
the ratio between bonds and stocks decreases. Finally, for high values of the reference point,
minimum risk portfolios only contain stocks, which become small-mid cap stocks for extreme
values of the reference point.
Investors portfolios are obtained by allocating wealth between the dierent accounts.
Let pt

k
be the risk of the minimum risk portfolio for account k with corresponding reference
point
k
for k = 1, . . . , K, where
1
<
2
< <
K
. Obviously we have pt

1
pt

2

pt

K
, since PT

increases with the reference point. Let

pt
+
k
=
k
+ pt
+
k
be the reward of
the minimum risk portfolio for account k, plus the corresponding target return. Investors
allocate wealth between the dierent accounts in order to maximize their total reward given
the risk constraint implied by their loss tolerance. Moreover, the dierent accounts are
considered separately, i.e., the covariance between accounts specic portfolios are ignored.
Consequently, the investors (global) portfolio choice problem can be written as a simple
linear program:
max
K

k=1

k

pt
+
k
, s.t.
K

k=1

k
pt

k
pt

,
K

k=1

k
= 1,
k
0 for k = 1, . . . , K. (12)
Table II, i.e., the mental accounts are dened for reference points RP(X) = (1 + ) q(X),
where takes values 3.0%, 4.2%, . . . , 15.5% as given in the rst column of Table II. We
solve Problem (12) for pt

[pt

1
, pt

K
], which represent investors tolerance to losses. The
range [pt

1
, pt

K
] corresponds to the values of PT

that can be reached when short-selling

are not allowed, so it makes sense to restrict investors tolerance to it. We obtain that
investors with low tolerance to losses only invest into the accounts with reference points 3%
36
and 7.8%, while investors with higher loss tolerance invests into the accounts with reference
points 6.6%, 7.8% and 15.5%. Therefore, investors with higher loss tolerance also possess
high aspiration accounts, while investors with lower loss tolerance mainly invest into low-
aspiration accounts. Figure 3 shows the bonds-to-stocks ratio as function of investors loss
tolerance pt

. The bonds-to-stocks ratio is decreasing due to the fact that investors with
higher loss tolerance put a higher proportion of their wealth into high aspiration accounts.
Similarly, Figure 4 shows the ratio between the proportion invested in large cap stocks and
the total holding of stocks. This ratio is constant for low degrees of loss tolerance, while
it decreases as function of pt

when the degree of loss tolerance is higher. These results

are consistent with the portfolio recommendations reported by Canner, Mankiw, and Weil
(1997) and Wang (2003).
[Figure 3 about here.]
[Figure 4 about here.]
V Conclusion
We analyzed general reward-risk models and showed that only a few properties for reward
and risk measures imply the two-fund separation property for optimal allocations. If reward
and risk measures can be transformed by means of increasing transformations into positively
homogeneous and translation invariant or equivariant functionals, then reward-risk optimal
37
allocations satisfy the two-fund separation property. Since positive homogeneity and trans-
lation invariance or equivariance are often used as properties to dene well-behaved (i.e.,
rational) measures of risk (see, for example, coherent risk measures), our result on two-fund
separation conrms the diculties encountered by Canner, Mankiw, and Weil (1997) in
nding a rational model of preferences explaining the asset allocation puzzle.
We then presented a behavioral reward-risk model for portfolio selection that is founded
in a descriptive model of preferences, i.e., the prospect theory of Kahneman and Tversky
(1979). In this model, the reward measure only captures portfolio payos above the subjec-
tive reference point (gains), while the risk measure only captures portfolio payos below the
reference point (losses). Finally, loss aversion determines the investors trade-o between
gains and losses. This model is the behavioral counterpart of the mean-variance model of
Markowitz (1952).
The two-fund separation property holds for the behavioral reward-risk model if re-
ward and risk measures are specied using a piecewise-power value function (Tversky and
Kahneman 1992) and a positively homogeneous (of degree one) and positive translation
equivariant reference point. Examples of positively homogeneous of (degree one) and trans-
lation equivariant reference points are the expected value and the risk-free rate of return.
The piecewise-power value function and the risk-free return as reference point are common
specication of prospect theory in behavioral nance. However, in general, the reward-risk
model violates two-fund separation, but deviations from two-fund separation are not sys-
tematic when the reference point is identical for all investors.
38
We then combined the behavioral reward-risk model with the multiple accounts version
of the behavioral portfolio theory of Shefrin and Statman (2000), assuming that investors
might possess dierent and even multiple reference points. Indeed, we expect investors with
higher loss tolerance to also possess higher reference points. In an empirical analysis on
US data, we showed that the behavioral reward-risk model with multiple accounts explains
both the asset allocation puzzle documented by Canner, Mankiw, and Weil (1997) and the
sub-puzzle identied by Wang (2003).
Further research on the behavioral reward-risk model could address asset pricing implica-
tions. More specically, the derivation of a behavioral asset pricing model that can explain
well-documented violations of the mean-variance CAPM; see De Giorgi and Post (2008).
39
A Proofs
A Proof of Proposition 1
We prove the following theorem:
Proposition 1
Let : G R and : G R be a reward and risk measure on G, respectively. Suppose
that strictly increasing transformations T

and T

exist, such that = T

and = T

are positively homogeneous of degree 1 and satisfy one the following two properties:
(i) translation invariance,
(ii) translation equivariance.
Then, there exists a portfolio allocation

p
, such that for any (, )-optimal portfolio allo-
cation

, there exists
0
with

=
0
e
0
+ (1
0
)

p
,
where e
0
= (1, 0

R
K+1
.
Proof. The ( , ) portfolio optimization problem is
max
R
K+1
(

Rw
0
) s.t. (

Rw
0
) ,

e = 1. (13)
We rst show that the (, )-portfolio optimization problem (3) with target risk is equiv-
alent to the ( , )-optimization problem (13) with target risk = T

().
40
Let

be a solution to the (, )-optimization (3) and suppose that

exists such that

_

Rw
0
_
T

() =
and

_

Rw
0
_
>
_

Rw
0
_
.
Then, since T

and T

are strictly increasing,

Rw
0
_

and

Rw
0
_
>
_

Rw
0
_
.
This contradicts the optimality of

. The same argument applies to T
1

and T
1

, since the
inverses are also strictly increasing. Thus the equivalence of (3) and (13) follows.
Now, we show mutual separation for the ( , )-portfolio choice. Suppose rst that and
are positively homogeneous of degree 1 and positive translation equivariant. Let us take
a portfolio

p
that solves the optimization problem (13) for the target risk
p
with w
0
= 1.
The portfolio gross return is denoted by

R
p
=

p
R. Let us now take another target risk
and initial wealth w
0
> 0, and nd
0
such that =
0
w
0
R
0
+ (1
0
) w
0

p
, i.e.,

0
=
w
0

p
w
0
R
0
w
0

p
.
Without loss of generality we assume 1
0
= (w
0
R
0
)/(w
0
R
0
w
0

p
) > 0.
21
Consider
41
the linear combination

=
0
e
0
+ (1
0
)

p
.
Then

_

Rw
0
_
=
0
w
0
R
0
+ (1
0
) w
0
(

R
p
)
0
w
0
R
0
+ (1
0
) w
0

p
=
and

_

Rw
0
_
=
0
w
0
R
0
+ (1
0
) w
0
(

R
p
)
is maximal. In fact, if a portfolio

exists that further increases the reward for the target
risk , then one can nd a linear combination of this portfolio and the risk-free asset that
further increases the reward for the target risk
p
, a contradiction to the denition of

p
.
Indeed, let be the risk of the portfolio

and let

0
=
0
/(1
0
). Then

p
=

0
e
0
+ (1

0
)

.
satises

_

p
R
_
=

0
R
0
+ (1

0
) (

R) =

0
1
0
R
0
+
1
w
0
(1
0
)
(

Rw
0
)


0
1
0
R
0
+
1
w
0
(1
0
)
=

0
R
0
w
0
w
0
(1
0
)
=
p
and

_

p
R
_
=

0
R
0
+ (1

0
)
_

R
_
>

0
R
0
+ (1

0
)
_

R
_
=

0
R
0
+ (1

0
)
_

0
R
0
+ (1
0
) (

R
p
)
_
=
_

R
p
_
42
i.e.,

p
is preferred to

p
by the investor with target risk
p
and w
0
= 1, a contradiction to
the optimality of

p
.
It follows that given an optimal solution

p
for a given target risk
p
, then any optimal
solution for any target risk corresponds to a linear combination of

p
and the risk-free
asset. This proves two-fund separation for ( , ). Because of equivalence of (3) and (13)
from the rst step of this proof, the statement of the proposition follows directly. The same
argument applies to the case of translation invariant and negative equivariant reward or risk
measures.
B Proof of Corollary 1
Let

be the degree of homogeneity of and , respectively. Take the strictly increasing

transformations T

: R R, x x
1

and T

: R R, x x
1

. Then the transformed

measures = T

and = T

are positively homogeneous of degree 1 and translation

invariant or equivariant. Thus, by Proposition 1, two-fund separation holds for (, ).
C Proof of Corollary 2
Let > 0 then
PT
+
( X) =
S

s=1
u(max(0, X(s) RP(X)))
s
=
S

s=1
u(max(0, X(s) RP(X)))
s
=
S

s=1
u( max(0, X(s) RP(X)))
s
=
+
S

s=1
u(max(0, X(s) RP(X)))
s
=
+
PT
+
(X).
43
Let a R, then
PT
+
(X +a) =
S

s=1
u(max(0, X(s) +a RP(X +a)))
s
=
S

s=1
u(max(0, X(s) +a RP(X) a))
s
=
S

s=1
u(max(0, X(s) RP(X)))
s
= PT
+
(X).
Therefore PT
+
is translation invariant and positively homogeneous of degree +. Similarly,
one can show that PT

is translation invariant and positively homogeneous of degree .

Consequently, according to Corollary 1, two-fund separation holds for any optimal portfolio
choice.
44
References
Acerbi, C. (2002): Spectral Measures of Risk: A Coherent Representation of Subjective
Risk Aversion, Journal of Banking and Finance, 26(7), 15051518.
Acerbi, C., and D. Tasche (2002): Expected Shortfall: A Natural Coherent Alternative
to Value at Risk, Economic Notes by Banca Monte dei Paschi di Siena SpA, 31(2), 379
388.
Artzner, P., F. Delbaen, J.-M. Eber, and D. Heath (1999): Coherent Measure of
Risk, Mathematical Finance, 9(3), 203228.
Bajeux-Besnainou, I., J. Jordan, and R. Portait (2001): An Asset Allocation
Puzzle: Comment, American Economic Review, 91(4), 11701179.
Banz, R. (1981): The Relationship Between Return and Market Value of Common Stocks,
Journal of Financial Economics, 9(1), 318.
Barberis, N., and M. Huang (2007): Preferences with Frames: A New Utility Speci-
cation that Allows for Framing of Risks, Working paper, Yale University.
(2008a): The Loss Aversion / Narrow Framing Appraoch to the Equity Premium
Puzzle, in Handbook of Investments: Equity Premium, ed. by R. Mehra, chap. 6, pp.
199228. North Holland, Amsterdam.
45
Barberis, N., and M. Huang (2008b): Stocks as Lotteries: The Implications of Proba-
bility Weighting for Security Prices, American Economic Review, forthcoming.
Barberis, N., M. Huang, and T. Santos (2001): Prospect Theory and Asset Prices,
Quarterly Journal of Economics, 116(1), 153.
Barberis, N., M. Huang, and R. H. Thaler (2006): Individual Preferences, Mone-
tary Gambles, and Stock Market Participation: A Case for Narrow Framing, American
Economic Review, 96(4), 10691090.
Basili, M., R. Ren
`
o, and C. Zappia (2008): Asset Prices and Multiple Reference
Points, Journal of Financial Decision Making, 4(1), Forthcoming.
Bawa, V., and E. Lindenberg (1977): Capital Market Equilibrium in a Mean-Lower
Partial Moment Framework, Journal of Financial Economics, 5, 189200.
Bell, D. E. (1982): Regret in Decision Making under Uncertainty, Operations Research,
30(5), 961981.
(1983): Risk Premiums for Decision Regret, Management Science, 29(10), 1156
1166.
(1985): Disappointment in Decision Making under Uncertainty, Operations Re-
search, 33(1), 127.
46
Benartzi, S., and R. Thaler (1995): Myopic Loss Aversion and the Equity Premium
Puzzle, Quarterly Journal of Economics, 110(1), 7592.
Berkelaar, A. B., R. Kouwenberg, and T. Post (2004): Optimal Portfolio Choice
Under Loss Aversion, The Review of Economics and Statistics, 86(4), 973987.
Bosch-Dom
`
enech, A., and J. Silvestre (2006): Reections on Gains and Losses: A
2 2 7 Experiment, Journal of Risk and Uncertainty, 33(3), 217235.
Bossaerts, P., M. Hsu, and K. Preuschoff (2008): The Neurobiological Founda-
tions of Valuation in Human Decision Making under Uncertainty, Working Paper, EPFL
Lausanne.
Brennan, M., and Y. Xia (2000): Stochastic Interest Rates and the Bond-Stock Mix,
European Finance Review, 4(2), 197210.
(2002): Dynamic Asset Allocation under Ination, Journal of Finance, 57(3),
12011238.
Brogan, A., and S. Stidham (2008): Non-Separation in the Mean-LPM Portfolio Op-
timization Problem, European Journal of Operational Research, 184(2), 701710.
Brown, D. B., and M. Sim (2008): Satiscing Measures for Analysis of Risky Positions,
Working Paper, Duke University.
47
Campbell, J., and L. Viceira (2001): Who Should Buy Long-Term Bonds?, American
Economic Review, 91(1), 99127.
Canner, N., G. Mankiw, and D. Weil (1997): An Asset Allocation Puzzle, American
Economic Review, 87(1), 181191.
Cass, D., and J. Stiglitz (1970): The Structure of Investor Preferences and Asset
Returns, and Separability in Portfolio Allocation: A Contribution to the Pure Theory of
Mutual Funds, Journal of Economic Theory, 2(2), 122160.
Cillo, A., and P. Delqui

e (2006): Disappointment Witout Prior Expectations: A

Unifying Perspective on Decision under Risk, Journal of Risk and Uncertainty, 33(3),
197215.
De Giorgi, E. (2002): A Note on Portfolio Selection Under Various Risk
Measures, NCCR-FINRISK Working paper, no. 9, Available at SSRN:
http://ssrn.com/abstract=762104.
(2005): Reward-Risk Portfolio Selection and Stochastic Dominance, Journal of
Banking and Finance, 29(4), 895926.
De Giorgi, E., T. Hens, and H. Levy (2003): Existence of CAPM Equilibria with
Prospect Theory Preferences, NCCR-FINRISK Working Paper, no. 85, Available at
SSRN: http://ssrn.com/abstract=420184.
48
De Giorgi, E., T. Hens, and J. Mayer (2007): Computational Aspects of Prospect
Theory with Asset Pricing Applications, Computational Economics, 29(3-4), 267281.
De Giorgi, E., T. Hens, and T. Post (2005): Myopic Loss Aversion and the Size and
Value Premium Puzzles, Mimeo, University of Zurich.
De Giorgi, E., and T. Post (2008): Second Order Stochastic Dominance, Reward-
Risk Portfolio Selection and the CAPM, Journal of Financial and Quantitative Analysis,
Forthcoming.
DeMiguel, V., L. Garlappi, and R. Uppal (2007): Optimal versus Naive Diversi-
cation: How Inecient Is the 1/N Portfolio Strategy?, Review of Financial Studies,
forthcoming.
Dimmock, S. G., and R. Kouwenberg (2007): Loss-Aversion and Household Portfolio
Choice, Working Paper.
Elton, E., and M. Gruber (2000): The Rationality of Asset Allocation Recommenda-
tions, Journal of Financial and Quantitative Analysis, 35(1), 2741.
Fama, E., and K. French (1993): Common Risk-factors in the Returns on Stocks and
Bonds, Journal of Financial Economics, 33(1), 356.
F

ollmer, H., and A. Schied (2002): Convex Measures of Risk and Trading Constraints,
Finance and Stochastics, 6(4), 429447.
49
Frittelli, M., and E. Rosazza Gianin (2004): Dynamic Convex Risk Measures, in
Risk Measures for the 21st Century, ed. by G. Szego. J. Wiley.
Gollier, C. (2001): The Economics of Risk and Time. MIT Press, Cambridge, Mas-
sachusetts.
Gomes, F. (2005): Portfolio Choice and Trading Volume with Loss-Averse Investors,
Journal of Business, 78(2), 675706.
Gomes, F., and A. Michaelides (2004): A Human Capital Explanation for an Asset
Allocation Puzzle, Financial Market Group (FMG) Discussion Paper.
Grootveld, H., and W. Hallerbach (1999): Variance vs Downside Risk: Is There
Really That Much Dierence?, European Journal of Operational Research, 114(2), 304
319.
Gul, F. (1991): A Theory of Disappointment Aversion, Econometrica, 59(3), 667686.
Harlow, W., and R. Rao (1989): Asset Pricing in a Generalized Mean-Lower Par-
tial Moment Framework: Theory and Evidence, Journal of Financial and Quantitative
Analysis, 25, 285311.
Jalal, A., E. Jondeau, and M. Rockinger (2007): Crash Aversion, Lottery-like
Behavior and Asset Prices, Working Paper, University of Lausanne.
50
Jarrow, R., and F. Zhao (2006): Downside Loss Aversion and Portfolio Management,
Management Science, 52(4), 558566.
Jin, H., and X. Y. Zhou (2007): Behavioral Portfolio Selection in Continuous Time,
Mathematical Finance, forthcoming.
Jorion, P. (1997): Value at Risk. Irwin, Burr Ridge, IL.
Kahneman, D., and A. Tversky (1979): Prospect Theory: An Analysis of Decisions
Under Risk, Econometrica, 47(2), 263291.
K

obberling, V., and P. Wakker (2005): An Index of Loss Aversion, Journal of

Economic Theory, 122(1), 119131.
Lioui, A. (2007): The Asset Allocation Puzzle Is Still A Puzzle, Journal of Economic
Dynamics & Control, 31, 11851216.
Loomes, G., and R. Sugden (1982): Regret Theory: An Alternative Theory of Rational
Choice Under Uncertainty, The Economic Journal, 92(368), 805824.
(1986): Disappointment and Dynamic Consistency in Choice under Uncertainty,
Review of Economic Studies, 53(2), 271282.
Markowitz, H. (1952): Portfolio Selection, Journal of Finance, 7(1), 7791.
(1959): Portfolio Selection. John Wiley & Sons, New York.
51
Mougeot, N. (2003): Interest Rate, Learning and the Asset Allocation Puzzle, EFA 2003
Annual Conference Paper, No. 89, Available at SSRN: http://ssrn.com/abstract=419740.
Ogryczak, W., and A. Ruszczynski (1999): From Stochastic Dominance to Mean-Risk
Models: Semideviations as Risk Measures, European Journal of Operational Research,
116, 3350.
Rockafellar, R. T., S. Uryasev, and M. Zabarankin (2006): Generalized Devia-
tions in Risk Analysis, Finance and Stochastics, 10(1), 5174.
Schmeidler, D. (1986): Integral Representation Without Additivity, Proceedings of the
American Mathematical Society, 97(2), 255261.
Schmidt, U., and H. Zank (2008): Risk Aversion in Cumulative Prospect Theory,
Management Science, 58(1), 208216.
Shadwick, W., and C. Keating (2002): A Universal Performance Measure, The Jour-
nal of Performance Measurement, 6(3), 5984.
Shalit, H., and S. Yitzhaki (2003): An Asset Allocation Puzzle: Comment, American
Economic Review, 93(3), 10021008.
Shefrin, H., and M. Statman (2000): Behavioral Portfolio Theory, The Journal of
Financial and Quantitative Analysis, 35(2), 127151.
52
Siebenmorgen, N., and M. Weber (2003): A Behavioural Model for Asset Allocation,
Financial Markets and Portfolio Management, 17(1), 1542.
Simon, H. A. (1955): A Behavioral Model of Rational Choice, Quarterly Journal of
Economics, 69(1), 99118.
Sugden, R. (2003): Reference-Dependent Subjective Expected Utility, Journal of Eco-
nomic Theory, 111, 172191.
Thaler, R. (1985): Mental Accounting and Consumer Choice, Marketing Science, 4,
199214.
(1999): Mental Accounting Matters, Journal of Behavioral Decision Making,
12(3), 183206.
Tobin, J. (1958): Liquidity Preference as Behaviour Towards Risk, Review of Economic
Studies, 25(2), 6586.
Tversky, A., and D. Kahneman (1992): Advances in Prospect Theory: Cumulative
Representation of Uncertainty, Journal of Risk and Uncertainty, 5, 297323.
Wang, H. (2003): An Asset Allocation Sub-Puzzle, Working Paper, Yale University.
53
Notes
1
Recent neurobiological research also shows that the computational model for choices
under risk which emerges from the study of brain activities is one of a tradeo between
reward and risk (Bossaerts, Hsu, and Preuscho 2008).
2
The reward-risk methodology is the main paradigm used by nancial advisors to com-
municate with clients.
3
See, for example, the disappointment theory (Bell 1985, Loomes and Sugden 1986, Gul
1991, Cillo and Delquie 2006), the regret theory (Loomes and Sugden 1982, Bell 1982, Bell
1983), and the reference-dependent subjective theory with stochastic reference point (Sugden
2003).
4
Few papers consider the probability weighting function in cumulative prospect theory;
see De Giorgi, Hens, and Levy (2003), Barberis and Huang (2008b), and Jin and Zhou
(2007).
5
Mental accounting refers to the set of cognitive operations used by individuals [...] to
organize, evaluate, and keep track of nancial activities (Thaler 1999). Here we consider the
component of mental accounting which relates to the way investors assign dierent activities,
or investment goals, to dierent (mental) accounts.
6
Few authors analyze the consistency of the reward-risk model with the expected utility
54
theory. Ogryczak and Ruszczynski (1999) introduce mean-dispersion risk measures and show
that under some conditions these are consistent with second order stochastic dominance.
De Giorgi (2005) axiomatizes a general reward-risk model that is consistent with second-
order stochastic dominance and shows that such a model can be linked to the Choquet
expected utility theory of Schmeidler (1986).
7
Siebenmorgen and Weber (2003) assume that investors solve a tradeo between pure
risk and naive diversication, where naive diversication is measured as the distance from
the 1/n portfolio strategy (see DeMiguel, Garlappi, and Uppal 2007).
8
See Jalal, Jondeau, and Rockinger (2007) who propose a model with both loss aversion
and crash aversion. Basili, Ren`o, and Zappia (2008) consider multiple reference points
to capture dierent degrees of loss aversion with respect to moderate and extreme losses,
respectively.
9
Imposing short-sale constraints help explaining violations of mean-variance two-fund
separation also within the mean-variance model. However, Canner, Mankiw, and Weil (1997)
point out that advisors recommendations do not satisfy the two-fund property even if short-
sale constraints are not binding, i.e., advisors recommend a strictly positive allocation to
any of the asset classes considered.
10
For convenience, we write the portfolio choice problem (1) over the set of budget feasible
portfolio payos X X instead of the set of assets holdings = (
0
, . . . ,
K
)

. However,
55
under the assumption that there are no redundant assets, for each X X there exists a
unique vector R
K+1
such that X =

K
k=0

k
A
k
.
11
Referring to the mathematical properties of a functional, Artzner, Delbaen, Eber, and
Heath (1999) use the terminology translation invariant for the property that we call neg-
ative translation equivariant. In order to give an intuitive meaning to the two translation
properties considered in this paper, we associate the word invariant to the property of a
functional of being invariant in terms of its value with respect to a deterministic addition
to its argument, while we associate the word equivariant to the property of varying in the
same way, in terms of its value, as its argument.
12
Suppose that is translation equivariant and homogeneous of degree > 0. Then
((X +a)) = ||

(X +a) = ||

((X) +a) = ||

(X) + ||

a.
On the other hand, we also have:
((X +a)) = (X +a) = (X) +a = ||

(X) +a.
Since this is true for all R (or > 0) and a R, we must have = 1.
13
The risk-free return depends on portfolio initial value, which is its price at time 0.
Therefore, the risk-free rate of return is in general a moving target. But, if we assume that
investors optimally invest their total wealth w
0
at time 0, then any optimal portfolio X

satises q(X

) = w
0
and thus (X

) = q(X

) R
0
= w
0
R
0
which is xed and independent
from the portfolio choice.
56
14
We dene the behavioral reward-risk model based on prospect theory, but we could
similarly apply cumulative prospect theory. The results we obtain in this section on two-fund
separation also hold for cumulative prospect theory, since positive homogeneity, translation
invariance or equivariance are unaected by the weighting of the cumulative distribution
functions instead of probabilities. Note also that typical applications of cumulative prospect
theory in nance ignore probability weighting and in this case prospect theory and cumulative
prospect theory are identical.
15
Let : G R be a real-valued function on G. We say that is:
(i) monotone if
(X) (Y )
for all X, Y G with X Y ,
(ii) anti-monotone if
(X) (Y )
for all X, Y G with X Y .
16
In a dynamic setting, the innite leverage problem has been discussed by Jin and Zhou
(2007), who show that the portfolio optimization problem with cumulative prospect theory
preference can be ill-posed. In their setting, the portfolio selection problem is said to be
ill-posed when the supremum of the objective function over the budget set is innite.
57
17
Kahneman and Tversky (1979) call this mental operation the segregation of certain
payo and include it into the editing phase of prospect theory.
18
The returns of the small-mid caps portfolio are obtained from the Lo30 and the Med40
portfolios of Fama and French. Based on the number of rms n
Lo30
t
and n
Med40
t
contained
in these portfolios and the average market capitalization amc
Lo30
t
and amc
Med40
t
, we derive
at each time t the weights
Lo30
t
= n
Lo30
t
amc
Lo30
t
/mc
tot
t
and
Med40
t
= n
Med40
t
amc
Med40
t
/mc
tot
t
where mc
tot
t
= n
Lo30
t
mc
Lo30
t
+ n
Med40
t
amc
Med40
t
. The returns of the small-mid caps portfolio
are then obtained as r
sm
t
=
Lo30
t
r
Lo30
t
+
Med40
t
r
Med40
t
. The large caps portfolio corresponds
to the Hi30 portfolio of Fama and French.
19
The pricing functional q is a linear function and satises q(a) = a/R
0
for all a R.
Consequently, for RP(X) = (1 +) q(X) we have
RP(X) = (1 +) q(X) = (1 +) q(X) = RP(X)
and
RP(X +a) = (1 +) q(X +a) = (1 +) (q(X) +a/R
0
) = RP(X) +a (1 +)/R
0
.
It follows that RP(X) = (1+) q(X) is positively homogeneous of degree one and translation
equivariance if and only if 1 +/R
0
= 1, i.e., = r.
20
This value function has been suggested for modeling prospect theory and cumulative
prospect theory by De Giorgi, Hens, and Levy (2003) and Kobberling and Wakker (2005).
58
De Giorgi, Hens, and Levy (2003) show that nancial market equilibria do not exist when in-
vestors possess heterogeneous cumulative prospect theory preferences with piecewise-power
value function and returns are Gaussian distributed, while existence of equilibria can be
established when cumulative prospect theory is modeled using a piecewise-exponential func-
tion. Kobberling and Wakker (2005) show that a piecewise-power function for prospect
theory violates loss aversion and also suggest using the piecewise-exponential function.
21
The assumption 1
0
> 0 holds in general because investors target risk
p
and are
greater than the risk free return. We also assume that a risky portfolio exists, that provides
greater reward than the riskfree asset.
59
(a)
Assets T-Bill Bond Index Market Small-Mid Large
Mean 3.78 5.34 12.01 15.63 11.74
Std deviation 3.11 5.27 20.11 27.63 19.37
Skewness 0.95 1.16 -0.36 0.27 -0.37
Kurtosis 0.96 2.14 -0.02 0.96 0.00
Maximum 14.72 26.02 57.50 111.45 53.10
Minimum -0.04 -5.17 -44.35 -46.82 -43.21
(b)
T-Bill Bond Index Market Small-Mid Large
T-Bill 1.00 0.49 -0.03 -0.10 -0.02
Bond Index 0.49 1.00 0.03 -0.03 0.06
Market -0.03 0.03 1.00 0.93 0.99
Small-Mid -0.10 -0.03 0.93 1.00 0.89
Table I: The Data. Panel (a) reports the summary statistics of yearly nominal returns (in
%) for the Treasury Bill, the bond index, the CRSP market portfolio and the two Fama and
French US stock portfolios formed on market capitalization (small-mid and large). There are
81 observations ranging from 1927 to 2007. The bond index and the Treasury Bill data are
from Ibbotson Associates, the stock portfolios and the CRSP market portfolio are from the
Kenneth French data library (http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/).
Panel (b) reports the correlation between any pair of assets.
60
(%) Cash Bonds Small-Mid Large
3.0 95 5 0 0
4.2 90 10 0 0
5.4 70 25 5 0
6.6 40 45 10 5
7.8 0 70 15 15
9.1 0 60 20 20
10.3 0 55 25 20
11.5 0 35 40 25
12.7 0 20 45 35
13.9 0 0 55 45
15.1 0 0 50 50
15.5 0 0 100 0
Table II: Minimum Risk Portfolios. The table reports minimum risk portfolios for the
prospect theory reward-risk model for dierent reference points RP(X) = (1+) q(X). The
numbers are in percentage and rounded to nearest 5% for portfolio holdings. The reward
and risk measures PT
+
and PT

, respectively, are specied using a piecewise-exponential

function v(x) = 1 exp(x) for x 0 and v(x) = (1 exp(x)) for x < 0, = 10.
61
Figure 1: Bonds-to-Stocks Ratio with Single Account. The gure shows the bonds-
to-stocks ratio (upper plot) and proportion invested into cash (lower plot) for the behavioral
reward-risk model, as function of stock holding. The reward and risk measure are specied
using a piecewise-power value function with = 1 and reference points RP(X) = (1+) q(X)
for = r, 4%, 5%.
62
Figure 2: Reward-Risk Tradeos. The gure displays reward and risk for the min-
imum risk portfolios for the prospect theory reward-risk model with dierent reference
points (which are indicated above the points). The prospect theory reward-risk model is
specied using a piecewise-exponential value function v(x) = 1 exp(x) for x 0
and v(x) = (1 exp(x)) for x < 0, = 10, and deterministic reference points
RP(X) = (1 + ) q(X). In order to compare reward and risk for portfolios computed
assuming dierent reference points, reward is given as +PT
+
, while risk is given as PT

.
63
Figure 3: Bonds-to-Stocks Ratio with Multiple Accounts. The gure shows the
bonds-to-stocks ratio for optimal portfolios as function of the degree of loss tolerance pt

.
Investors mental accounts correspond to the reference points presented in Table II. The
behavioral reward-risk model is specied using a piecewise-exponential value function v(x) =
1 exp(x) for x 0 and v(x) = v(x) for x < 0, = 10.
64
Figure 4: Large Cap Stocks to Total Stocks Ratios with Multiple Accounts. The
gure shows the ratio between the proportion of large cap stocks and the total holding of
stocks as function of the degree of loss tolerance pt

. Investors mental accounts correspond

to the reference points presented in Table II. The behavioral reward-risk model is specied
using a piecewise-exponential value function v(x) = 1 exp(x) for x 0 and v(x) =
(1 exp(x)) for x < 0, = 10.
65