Investors Facing Risk€¦ · Investors Facing Risk: Splitting Money Between Risky and Risk-Free...
Transcript of Investors Facing Risk€¦ · Investors Facing Risk: Splitting Money Between Risky and Risk-Free...
Investors Facing Risk:
Splitting Money Between Risky and Risk-Free Assets
Erick W. Rengifo∗ Emanuela Trifan†
October 6, 2006
Abstract
This paper studies the impact of loss aversion on decisions regarding the alloca-
tion of wealth between risky and risk-free assets. We use a Value-at-Risk portfolio
model with endogenous desired risk levels that are individually determined in an
extended prospect theory framework. This framework allows for the distinction be-
tween gains and losses with respect to a subjective reference point as in the original
prospect theory, but also for the influence of past performance on the current per-
ception of the risky portfolio value. We show how the portfolio revision frequency
impacts investor decisions and analyze the role of past gains and losses in the cur-
rent wealth allocation. The perceived portfolio value exhibits distinct evolutions in
two frequency segments delimitated by what we consider to be the optimal revision
horizon of one year. Our empirical results suggest that previous research relying on
fixed confidence portfolio risk levels underestimates the loss aversion of individual
investors.
Keywords : prospect theory, Value-at-Risk, loss aversion, portfolio revision
JEL Classification: C32, C35, G10
∗Fordham University, New York, Department of Economics. 441 East Fordham Road, Dealy Hall,
Office E513, Bronx, NY 10458, USA, phone: +1(718) 817 4061, fax: +1(718) 817 3518, e-mail: rengi-
[email protected]†Darmstadt University of Technology, Institute of Economics, Department of Applied Economics and
Econometrics, 15 Marktplatz, 64283 Darmstadt, Germany, phone: +49(0) 6151 162636, fax: +49(0) 6151
165652, e-mail: [email protected]
1
1 Introduction
Optimal portfolio allocation models represent important tools in helping investors to de-
cide upon how to split their wealth among assets. The goal of such models is to find what
is called the optimal allocation, i.e. the one that maximizes expected portfolio returns at
a given risk level. One of the most well known and broadly used portfolio optimization
settings is the mean-variance model, which suggests the variance of portfolio returns as
risk measure. Recent research employs other ways of quantifying risk, such as the so called
Value-at-Risk (VaR), defined as the highest expected loss from financial investments over
a specified time horizon and subject to a certain confidence level.
Acting on the VaR-concept, Campbell, Huisman, and Koedijk (2001) develop a model
for maximizing expected returns subject to both a budget and a VaR constraint, where
the latter requires the maximum expected loss to meet an exogenously specified VaR-limit
(the so called desired VaR, shortly VaR*). One important feature of this model is that,
as in the classical mean-variance framework, the two fund separation theorem applies
namely, neither the investors’ initial wealth nor the desired VaR*, affect the maximiza-
tion procedure. Thus, investors interested in allocating wealth among risky assets can
first determine the risky portfolio composition, and then decide upon the extra amount of
money to be borrowed or lent (i.e. invested in risk-free assets), according to the individual
degree of risk aversion measured by the selected VaR*. In practice, the former decision
is often made by professional portfolio managers in charge of the construction of an op-
timal risky portfolio for their clients. These clients, usually non-professional investors,
concentrate sooner on the second decision step, by choosing the amount of money to be
invested in the risky portfolio as a whole, and implicitly by fixing the level of the risk-free
investment.1
In this context, the focus of our paper is on the decisions of non-professional investors.
The first and most important of these decisions refers to the formulation of VaR*. Its
value decides then on the optimal portfolio composition, hence on the sum of money to
be invested in risk-free assets. As mentioned above, finding the optimal risky portfolio
represents the task of professional portfolio managers and was extensively studied in
previous research on portfolio optimization. However, the resulting values of the total
risky vs. the risk-free investment as percentages of total wealth directly concerns non-
professional investors, becoming thus an object of study in the present work.
Our model builds on the portfolio optimization setting with exogenous desired VaR*
1In other words, non-professional investors consider the risky portfolio as exogenously given (fixed by
the manager). They are exclusively concerned with determining the final position in risky vs. risk-less
assets (i.e. how much money to put in the risky portfolio as a whole, while the rest is allocated to risk-free
assets), according to their own level of risk aversion.
2
in Campbell, Huisman, and Koedijk (2001) that we extend by explicitly accounting for the
formation of the individual VaR*-levels. They rely on the subjective perception of non-
professional investors over the risky portfolio performance and over utility in general. In
other words, we analyze how non-professional investors set their subjective VaR* and how
this (now endogenous) VaR* impacts on the wealth allocation between risk-free assets and
the risky portfolio. Our paper also incorporates the idea that individual risk perception
is affected by the previous evolution of financial wealth, so that facing past gains (losses)
induces a more (less) aggressive behavior, hence an increase (decrease) in risky portfolio
holdings. Thus, we present evidence for how different investment decisions of individual
investors can be interpreted as a consequence of different financial performance histories,
how these decisions change subject to the individual degree of loss aversion, and how our
results conform with previously documented findings.
Our paper comes in line with previous research aiming at studying the influence of
behavioral aspects on financial decisions. Benartzi and Thaler (1995) develop a plausible
explanation for the equity premium puzzle that relies on the interaction between loss
aversion and frequent portfolio evaluations, denoted as “myopic loss aversion”. Their
findings support the idea that, when investors review the performance of their portfolios
yearly, the resulting empirical equity premium is consistent with the loss aversion values
estimated in the standard prospect theory framework. Furthermore, Barberis, Huang, and
Santos (2001) apply the main concepts of the prospect theory to asset pricing, showing
that investors derive utility not only from consumption but also from variations in the
perceived value of financial investments. Moreover, investors not only distinguish between
gains and losses with respect to a subjective reference and warily avoid losses, as stated
in the original Kahneman and Tversky’s prospect theory, but also their perception over
losses appears to be affected by previous portfolio performance, i.e. by gains and losses
accumulated from past trades and referred as “cushions”.
We assume that, at the beginning of the trade, non-professional investors already
hold well diversified portfolios such as a market index (i.e. the empirical part of our
paper considers the SP500 index as proxy for the risky portfolio). Thus, the problem
they actually face reduces to the allocation of wealth between this risky portfolio (as a
whole) and the risk-free investment alternative. This separation of the risky and risk-
free investments complies with the concept of “mental accounting”, as first introduced
in Thaler (1980). According to Thaler (1992), people manifest the tendency to frame
(i.e. code and evaluate) outcomes in several non-fungible mental categories or accounts
(such as accounts for current income, current wealth or future income) with different
consumption propensities. This mental categorization decides upon the perceived utility
of those outcomes.
3
In our model, investors find the optimal solution to their decision problem by maxi-
mizing subjectively perceived utility. This utility is assumed to be derived merely from
changes in financial wealth.2 In line with the prospect theory (introduced in Tversky and
Kahneman (1979) and extended in Kahneman and Tversky (1992)), the perceived value
of risky investments is denoted as the prospective value. It relies on the subjective value
generated by one unit of risky project that is captured by the so-called value function.
Yet, in line with Barberis, Huang, and Santos (2001), we reconsider the original prospect
theory definition of the value function in order to account for the possible influence of past
performance. The idea that past gains and losses may change the current risk aversion,
hence financial decisions, is supported by an empirically observed phenomenon denoted
as the “house money effect” and documented in Thaler and Johnson (1990). Accordingly,
subjects who made money in past gambles appear to behave less risk aversely in subse-
quent bets. In other words, past gains make future losses less painful, while prior losses
may increase the risk aversion. A neurobiological explanation of this human reaction is
provided by the “somatic marker hypothesis” in Damasio (1994). Accordingly, preexist-
ing somatic (i.e. bodily) states can influence new ones by inducing modifications in the
level of activation (threshold) of the new state. As suggested in Bechara and Damasio
(2005), prior somatic states (in our case generated by past series of gains or losses) can
reinforce (impede) the perception of new ones (here, currently expected gains and losses)
by congruous (incongruous) valence (i.e. positivity/negativity). Also, even when prior
performance induces only weak somatic states (known as background states), it appears
to exert an impact on risk aversion. For instance, negative background states diminish
the risk aversion in face of sure losses (for that the fear of experience one more loss after a
series of past losses is higher and makes investors more risk loving in the hope of recovering
those losses), while positive background states enhance risk aversion in face of sure gains
(i.e. once several gains are experienced, investor predisposition to gambling diminishes).
In our framework, the desired Value-a-Risk (VaR*) is endogenously defined as the max-
imum expected loss perceived by individual investors and depends on past performance,
loss aversion, current value of the risky investment and the expected return premium.
We first compute the VaR* and then derive the desired level of investment in the risky
portfolio relative to the risk-free allocation. This allows us to draw a conclusion about the
investor risk aversion and to provide a comparison with the exogenous VaR* framework
in Campbell, Huisman, and Koedijk (2001).
We design two ways of assessing the prospective utility (which reduces in our setting to
2In other words, investors are interested only in the (perceived) value of their financial investments
(and not in other determinants of utility, such as consumption). This can rely on the fact that investors
narrow frame, i.e. put excessive emphasis on, the importance of financial investments and the utility they
generate. According to Barberis, Huang, and Thaler (2003), this is a common situation in practice.
4
the expected value of the risky investment). One definition relies on the original prospect
theory, and another one answers what we call a “worst case scenario”, where investors
are assumed to be concerned with the maximum sustainable (and not with the expected)
loss in the risky investment. Moreover, we study how investor decisions change according
to different market conditions, as captured by the prospect theory-part of the utility
function, and how different ways of representing loss aversion can influence utility. For
instance, we expect that risk averse investors reduce their risky holdings, shifting their
positions to more secure investment alternatives. Also, we derive equivalent significance
levels for VaR* at each trading time and compare them to the corresponding significance
levels used in the original model of Campbell, Huisman, and Koedijk (2001). In addition,
we introduce the notions of actual loss aversion and of global first order risk aversion that
we consider better suited than the simple coefficient of loss aversion to measuring the real
investor attitude to losses.
Finally, we study how investment decisions change under different portfolio revision
horizons (one day, one month, one quarter, four months, one year, two years, and up to
eight years), in other words how the revision frequency exerts influence on the risk percep-
tion and wealth allocation. In this context, we estimate the evolution of the prospective
value, as well as of the actual loss aversion and of the global first order risk aversion, as
functions of the revision frequency and address the problem of optimal revision horizons.
The theoretical findings in the first part of our paper are implemented and amended
in the subsequent empirical part. In so doing, we use real market data, such as the SP500
index as proxy for the risky portfolio and the US 10-year bond accounting for the risk-free
investment alternative.
Our empirical findings lead to several interesting conclusions. First, the risky holdings
of non-professional investors performing annual portfolio revisions represent on average
42% of the investors’ wealth and decrease to a value very close to zero for the revision
horizon of one day. Second, even when the coefficient of loss aversion remains constant
over trading dates, financial wealth fluctuations determined by the success of previous
decisions play a key role in the current portfolio allocation, Thus, when investors are
unable to accumulate positive significant cushions (calculated as the difference between
the risky return and the reference point of the value function represented here by the
risk-free return), most of their wealth is directed to the risk-free investment. Third, the
creation of positive and significant cushions is inversely related to the revision frequency.
As this frequency increases, the ability to accumulate profits decreases and a lower wealth
portion will be invested in risky assets. Fourth, as in Benartzi and Thaler (1995), our
results support the idea that one year appears to be the most plausible revision frequency
to be used by non-professional loss averse investors in practice. Further estimations show
5
that the evolution of the perceived portfolio utility (i.e. of the prospective value) for
different revision frequencies can be decomposed into two distinct intervals, namely one
for high frequency revisions (less than one year), and a second for low frequency revisions
(more than one year). The prospective value on the first interval can be analytically
represented as a quadratic convex function of the revision horizon, while on the second
segment, its evolution is rather upward sloping and of higher degree (yet, a simple line
represents a satisfactory approximation). A similar segmentation can be observed also
for the actual loss aversion and the global first order risk aversion. Fifth, the VaR*-levels
assessed from real data point out that, in practice, the loss aversion of real non-professional
investors may be higher than the values obtained for common confidence levels such as
90%, 95% and 99% in previous theoretical papers. Finally, the equivalent average loss
aversion coefficient computed for fixed confidence levels of 99% and 90% lie far below
the widely documented and empirically supported value of 2.25. Again, this implies that
previous research considering these confidence levels underestimates the aversion to losses
manifest in practice. We also note the inverse relationship between the revision frequency
and the associated loss aversion coefficient. For the maximal considered frequency of one
day, the coefficient of loss aversion becomes one, which implies that investors treat gains
and losses in the same way.
The remainder of the paper is organized as follows. Section 2 presents the main
theoretical considerations. We start by a brief review of the optimal portfolio selection
model with exogenous VaR* as in Campbell, Huisman, and Koedijk (2001), on which we
build our own theoretical structure. Section 2.2 takes on the reformulation of the value
functions in Barberis, Huang, and Santos (2001), out of which we derive what we define
as actual loss aversion. In Section 2.3 we introduce the notions of VaR* and of global first
order risk aversion and propose different ways of quantifying the endogenous VaR*. The
subsequent Section 2.4 frames distinct ways of assessing the value of the risky portfolio
as perceived by individual investors, while Section 2.5 treats the influence of a variable
revision frequency on the prospective value, the actual loss aversion and the global risk
aversion. Section 3 illustrates the empirical implementation of our theoretical model.
Section 3.1 discusses the impact of variable revision frequencies and of the cushion on the
evolution of wealth percentages invested in the risky portfolio. In Section 3.2, we analyze
the evolution of the prospective utility in time, as well as in the revision frequency domain.
Finally, Section 3.3 restates our model in terms of previous research with exogenous
VaR*, in that equivalent significance levels of portfolio risk and of loss aversion coefficients
that result from the average VaR* computed from our data and according to our model
equations are inferred. Section 4 concludes and provides an outlook for future research.
Graphics and further results are included in the Appendix.
6
2 Theoretical model
This section contains the main theoretical considerations of our work. We take up pre-
senting the model of portfolio selection with VaR as the risk measure and an exogenous
desired risk aversion (VaR*) of Campbell, Huisman, and Koedijk (2001). This model has
motivated us in extending the analysis for the case with endogenous VaR*. Subsequently,
we formulate our own model referring the individual perception of risky projects, detailing
the construction of the endogenous measure of risk aversion VaR* and its implications for
individual investor decisions. More precisely, we show how the investor desired VaR* can
be formulated and how it flows into the prospective value of the risky investment that
investors aim at maximizing. Also, we detail the definition of the real investor attitude to
losses, by introducing the notions of actual loss aversion and of global first order risk aver-
sion. Moreover, we analyze how the prospective value and these two further risk attitude
measures vary subject to different portfolio revision frequencies.
2.1 Optimal portfolio selection with exogenous VaR*
Let us first refresh the portfolio selection model with exogenous VaR* introduced in
Campbell, Huisman, and Koedijk (2001). Accordingly, financial assets are allocated by
maximizing the expected return subject to the common budget constraint, as well as to an
additional risk constraint, where risk is measured by the so-called Value-at-Risk (VaR).
The optimal portfolio is derived such that the maximum expected loss does not exceed
the VaR* indicated by non-professional investors, for a chosen investment horizon and at
a given confidence level. Additionally, investors can borrow or lend money at the fixed
market interest rate.
We denote by Wt the investor wealth at time t, by Bt the amount of money to borrow
(Bt > 0) or to lend (Bt < 0) at the fixed risk-free gross return rate Rf , and by VaR*
the individually desired VaR (specified later in this section). Let the risky portfolio
consist in i = 1 . . . n financial assets with single time t prices pi,t and define Ωt ≡ [wt ∈Rn :
∑ni=1 wi,t = 1] the set of portfolio weights at time t, such that xi,t = wi,t(Wt +
Bt)/pi,t represents the number of shares of the asset i contained in the portfolio at time
t. Obviously, the portfolio gross return at next trade (Rt+1) depends on the portfolio
composition at the current date wt. With the budget constraint:
Wt + Bt =n∑
i=1
xi,tpi,t = x′tpt, (2.1)
the value of the portfolio at t + 1 results in:
Wt+1(wt) = (Wt + Bt)Rt+1(wt)−BtRf . (2.2)
7
The VaR is defined as the maximum expected loss over a given investment horizon
and for a given confidence level 1-α, thus:
Pt[Wt+1(wt) ≤ Wt − VaR∗] ≤ 1− α, (2.3)
where Pt is the conditional probability on the available information at time t and VaR∗
is the the investor desired VaR level. Equation (2.3) represents the risk constraint that
(professional) investors have to take into account in addition to the budget constraint
(2.1), when searching for optimal portfolio weights.
The portfolio optimization problem can be now expressed in terms of the maximization
of expected portfolio returns Et[Wt+1(wt)], subject to both the budget restriction and the
VaR-constraint:
w∗t ≡ arg max
wt
(Wt + Bt)Et[Rt+1(wt)]−BtRf, s.t. (2.1) and (2.3). (2.4)
Here, Et[Rt+1(wt)] represents the expected return of the portfolio given the information
at time t.
The optimization problem may be rewritten in an unconstrained way, by replacing
(2.1) in (2.2) and taking expectations:
Et[Wt+1(wt)] = x′tpt(Et[Rt+1(wt)]−Rf ) + WtRf . (2.5)
Equation (2.5) points out that risk-averse investors are going to put a fraction of their
wealth in risky assets if the expected risky portfolio return is higher than the risk-free
rate Et[Rt+1(wt)] ≥ Rf .
Substituting (2.5) (before taking expectation) in (2.3) gives:
P [x′tpt(Rt+1(wt)−Rf ) + WtRf ≤ Wt − VaR∗] ≤ 1− α,
so that
P
[Rt+1(wt) ≤ Rf − VaR∗ + Wt(Rf − 1)
x′tpt
]≤ 1− α, (2.6)
defines the quantile qt(wt, α) of the distribution of portfolio returns for a given confidence
level 1-α or probability of occurrence (α).
Thus, the budget constraint can be restated as:
x′tpt =VaR∗ + Wt(Rf − 1)
Rf − qt(wt, α). (2.7)
Finally, substituting (2.7) in (2.5) and dividing by the initial wealth Wt, we obtain a
new expression to be maximized:
Et[Wt+1(wt)]
Wt
=VaR∗ + Wt(Rf − 1)
WtRf −Wtqt(wt, α)(Et[Rt+1(wt)]−Rf ) + Rf . (2.8)
8
Given that at moment t of maximization, Wt is known and Rf is fixed, the optimal
portfolio composition can be derived as:
w∗t ≡ arg max
wt
Et[Rt+1(wt)]−Rf
WtRf −Wtqt(wt, α). (2.9)
Equation (2.9) shows that, similarly to the traditional mean-variance framework, the
two fund separation theorem applies, i.e. neither the (non-professional) investor’s initial
wealth nor the desired VaR* affect the maximization procedure. In other words, investors
can first allocate wealth inside of the risky portfolio (i.e. among different risky assets)
and second fix the extra amount money to be borrowed or lent (i.e. invested in risk-free
assets). The latter reflects by how much the portfolio VaR varies according to the investor
degree of risk aversion, which is measured by the selected (desired) VaR* level. Replacing
(2.1) in (2.7), we further derive:
Bt =VaR∗ + VaR
Rf − qt(w∗t , α)
(2.10a)
VaR = Wt[qt(w∗t , α)− 1]. (2.10b)
2.2 The value function
Coming from the main ideas of the Campbell, Huisman, and Koedijk (2001)’s setting,
our model goes a step further by asking how individual investors set their desired level
of risk aversion VaR*. We elaborate on the construction of an endogenous VaR* and its
implications for the wealth allocation between risky and the risk-free assets.
Investors’ desires depend on their perception upon the value of financial investments.
Prospect theory suggests how individual perceptions of financial performance can be for-
malized by means of the so-called value function. According to Tversky and Kahneman
(1979) and Kahneman and Tversky (1992), to human minds, the actual carriers of value
are not absolute outcomes of a project, but deviations from an individual reference point,
where the deviations above (below) this reference are labelled as gains (losses). Thus,
the value function is kinked at the reference point and exhibits distinct evolution in the
domains of gains and losses, i.e. it is steeper for losses (a property known as loss aversion).
Also, it shows diminishing sensitivity in both domains (namely, it is concave for gains but
convex for losses).
As noted in Barberis, Huang, and Santos (2001), the view of the original prospect
theory over individual perceptions of risky investments can be enriched by accounting for
the potential impact of past performance (i.e. in addition to the mental distinction be-
tween gains and losses). Accordingly, the value function additionally reflects the influence
of a so-called cushion, that is defined as the difference between the current value of the
risky investment St and a benchmark level from the past Zt (e.g. the purchasing price
9
of the stock). When this difference is positive, investors made money from past risky
investment, otherwise they accumulated losses.
Our approach relies on the formulation of the value function proposed in Barberis,
Huang, and Santos (2001). In their Equations (15) and (16), the reference point changes
with the past performance (from ztRft for zt ≤ 1 to Rft for zt > 1, where zt = Zt/St). We
restate these definitions, in order to obtain identical reference points and similar courses in
the loss domain for both considered cases with positive and negative cushions, like in the
original prospect theory formulation, where gains are defined as the difference between the
value function argument (here Rt+1) and the reference point. Thus, we fix the reference
value in both cases (with prior gains zt ≤ 1 and prior losses zt > 1) to Rft and rearrange
the terms in Equations (15) and (16) in Barberis, Huang, and Santos (2001), obtaining:
v =
St(Rt+1 −Rft) , for Rt+1 ≥ Rft
λSt(Rt+1 −Rft) + (λ− 1)(St − Zt)Rft , for Rt+1 < Rft
, for zt ≤ 1(⇔ Zt ≤ St),
(2.11)
and
v =
St(Rt+1 −Rft) , for Rt+1 ≥ Rft
λSt(Rt+1 −Rft) + k(Zt − St)(Rt+1 −Rft) , for Rt+1 < Rft
, for zt > 1(⇔ Zt > St).
(2.12)
Here, λ is the coefficient of loss aversion, while the parameter k > 0 captures the
influence of previous losses on the perception of current ones (i.e. the larger the previous
loss is, the more painful the next losses become). We observe that, while the gain branches
of both value functions are invariable to past performance zt, the loss branches contain
a first term that resembles the original prospect theory, i.e. λSt(Rt+1 − Rft), but also a
second one revealing the impact of the cushion St−Zt. Moreover, the time t + 1-value of
the risky investment is derived as:
St+1 = (Wt + Bt)Rt+1. (2.13)
Of note is also that the simultaneous impact of the loss aversion coefficient λ and of
past portfolio performance k changes the actual investor aversion to losses. This can be
easily deduced by merging Equations (2.11) and (2.12) to:
v =
Stxt+1 , for xt+1 ≥ 0
[λSt − (1− πt)k(St − Zt)]xt+1 + πt(λ− 1)Rft(St − Zt) , for xt+1 < 0,
where πt = Pt(zt ≤ 1) is the probability of experiencing past gains and xt+1 = Rt+1−Rft
the equity return premium. According to the original prospect theory, the loss aversion
turns out to be risk aversion of first order in the loss domain, hence identical to the
10
derivative of the loss branch of the value function. Let us define the risk aversion of first
order as:
Λ(xt+1) =∂v
∂xt+1
=
St , for xt+1 ≥ 0
λSt − (1− πt)k(St − Zt) , for xt+1 < 0.(2.14)
Clearly, the actual loss aversion is given by the loss-branch of the above Function
(2.14), i.e. λSt− (1−πt)k(St−Zt), which is more than the simple loss aversion coefficient
λ from the original prospect theory. In other words, the aggregation of positive past
results (i.e. a positive cushion St − Zt ≥ 0) lowers the actual investor aversion to losses,
for that they are more confident in being able to cover prospective losses by past gains.
Henceforth, in line with the original prospect theory formulation, we mostly refer as “loss
aversion” the loss aversion coefficient λ (and not to the actual loss aversion derived above).
2.3 The endogenous VaR*
Our first goal is to formulate the maximum loss expected by individual investors, i.e. the
individual desired VaR*. This value will subsequently enter the optimization problem
when deciding between borrowing or lending.
To this end, we start from the literal definition of VaR*, concentrating on the notions of
“maximum”, “loss”, and “individual”. First, VaR* quantifies losses. However, according
to the prospect theory, what actually counts for individual investors is not the absolute
magnitude of a loss, but rather the subjectively perceived one, as captured by the value
function. Hence, VaR* should rely on the subjective value (or utility) of losses expressed in
the loss branches of the value functions (2.11) and (2.12). It depends on individual investor
characteristics (originated in the subjective view over gains and losses) and can vary over
time. Second, VaR* should represent a (subjective) expectation, because the next period
returns Rt+1 on which the evaluation of risky investments relies, are still unknown at the
decision time t. Third, we are looking for a maximal value, such that in calculating VaR*,
investors must ascribe a maximal occurrence probability Pt(Et[Rt+1] < Rft) = 1 to the
losses in the value function.
Therefore, we propose that VaR* accounts for the maximum expectation of sustainable
losses as resulting from individual valuations of the risky investment. However, we consider
investors to be sophisticated enough in order to consider that not only the mean, but also
the variation of prospective losses should be considered in order to accurately ascertain
the maximum acceptable loss level. Thus, in a second approximation, we enlarge the
VaR*-definition by adjusting for loss variance.
Henceforth, we consider that value functions are weighted by the pure probabilities
of occurrence (and not by non-linear probability functions as stated in the cumulative
11
prospect theory). According to Equations (2.11) and (2.12), we then derive:
Et[loss-utilityt+1] = πt[λSt(Et[Rt+1]−Rft) + (λ− 1)(St − Zt)Rft]
+ (1− πt)[λSt(Et[Rt+1]−Rft) + k(Zt − St)(Et[Rt+1]−Rft)]
= λSt(Et[Rt+1]−Rft) + [πt(λ− 1)Rft − (1− πt)k(Et[Rt+1]−Rft)](St − Zt)
(2.15a)
V art[loss-utilityt+1] = Et[loss-utility2t+1]− E2
t [loss-utilityt+1]
= πt[λSt(Et[Rt+1]−Rft) + (λ− 1)(St − Zt)Rft]2
+ (1− πt)[λSt(Et[Rt+1]−Rft) + k(Zt − St)(Et[Rt+1]−Rft)]2
− E2t [loss-utilityt+1]
= [λSt(Et[Rt+1]−Rft)]2
+ [πt(λ− 1)2R2ft − (1− πt)k
2(Et[Rt+1]−Rft)2](St − Zt)
2
+ 2[πt(λ− 1)Rft − (1− πt)k(Et[Rt+1]−Rft)]λSt(Et[Rt+1]−Rft)(St − Zt)
− [λSt(Et[Rt+1]−Rft)]2
− [πt(λ− 1)Rft − (1− πt)k(Et[Rt+1]−Rft)]2(St − Zt)
2
− 2[πt(λ− 1)Rft − (1− πt)k(Et[Rt+1]−Rft)]λSt(Et[Rt+1]−Rft)(St − Zt)
= πt(1− πt)[(λ− 1)Rft + k(Et[Rt+1]−Rft)]2(St − Zt)
2. (2.15b)
Note that, while the first term of the expected losses (2.15a) is similar to the loss
formulation in the original prospect theory, the remaining terms point out the influence
of the cushion accumulated over past trades. In contrast, the variance of losses (2.15b)
is exclusively dictated by the cushion-part, as individually perceived by investors, and
depends on the probability of having made gains or losses in the past, on the variance of
expected returns with respect to the reference risk-free rate, and on the squared cushion.
As mentioned above, in a first approximation, we stick to the literal definition of VaR*
as an expectation and design VaR* as the maximum expected loss:
VaR∗1t+1 = Et[loss-utilityt+1]
= λSt(Et[Rt+1]−Rft) + [πt(λ− 1)Rft − (1− πt)k(Et[Rt+1]−Rft)](St − Zt).
(2.16)
However, investors may consider loss-variance as an equally important parameter for
determining the maximal sustainable loss. Then, assuming that VaR* follows a cer-
tain distribution (i.e. normal or Student-t)3 with the value ϕ, we introduce the second
3Although VaR is a very popular measure of risk, it has been criticized because it does not satisfy
one of the four properties for coherent risk measure, namely subadditivity (see Artzner, Delbaen, Eber,
and Heath (1999), Rockafellar and Uryasev (2000) and Szego (2002)). However, according to Embrechts,
McNeil, and Straumann (1999), VaR becomes subadditive and can be considered as a coherent risk
12
(variance-adjusted) VaR* definition:
VaR∗t+1 = Et[loss-utilityt+1]− ϕ
√V art[loss-utilityt+1], (2.17)
which, according to Equations (2.15a) and (2.15b), results in:
VaR∗t+1 = λSt(Et[Rt+1]−Rft)
+ [(πt − ϕ√
πt(1− πt))(λ− 1)Rft − (1− πt + ϕ√
πt(1− πt))k(Et[Rt+1]−Rft)](St − Zt).
(2.18)
Again, expression (2.18) encompasses the twofold loss effect stemming from the loss
aversion coefficient of the original prospect theory and from the cushion introduced in
Barberis, Huang, and Santos (2001).
It is worth noting that, for sure gains (i.e. when πt = Pt(zt ≤ 1) = 1), both VaR*
expressions (2.16) and (2.18) reach a common upper bound:
VaR∗1,upt+1 = VaR∗up
t = λSt(Et[Rt+1]−Rft) + (λ− 1)Rft(St − Zt), (2.19)
while for sure losses (i.e. when πt = Pt(zt ≤ 1) = 0), the lowest value of:
VaR∗1,lot+1 = VaR∗lo
t = λSt(Et[Rt+1]−Rft)− k(Et[Rt+1]−Rft)(St − Zt) (2.20)
is attained.
The definition of VaR* serves to determining the optimal level of borrowing or lending
(Bt) from Equation (2.10a). When VaR* lies “to the right” of the portfolio VaR (i.e. it
is higher in absolute value than VaR), Bt is negative, hence investors become more risk
averse and save money. On the contrary, for a VaR* lower than VaR in absolute value,
investors augment their risky investment by borrowing extra money. Thus, an empirical
analysis of the evolution of Bt (as conducted in Section 3) can shed some light on the
investor risk behavior.
Also, one interesting topic to invetigate consists in estimating the equivalent loss aver-
sion parameter λ∗t that can be obtained for a VaR∗t+1 = VaR∗ that is fixed for commonly
used significance levels such as 1, 5 or 10%. The derivation can be readily done from the
definition (2.18) of VaR*:
λ∗t+1 =VaR∗ + [(πt − v
√πt(1− πt))Rft + (1− πt + v
√πt(1− πt))k(Et[Rt+1]−Rft)](St − Zt)
St(Et[Rt+1]−Rft) + (πt − v√
πt(1− πt))Rft(St − Zt).
(2.21)
Moreover, since λ∗t+1 depends on the fixed (thus exogenous) VaR∗, there should exist
no further causal relationship between past and future losses, such that we can set k = 0.
Accordingly, Equation (2.21) becomes:
λ∗t+1 =VaR∗ + [(πt − v
√πt(1− πt))Rft](St − Zt)
St(Et[Rt+1]−Rft) + (πt − v√
πt(1− πt))Rft(St − Zt). (2.22)
measure, if used in the case of elliptic joint distributions, such as normal and Student-t with finite
variance.
13
2.4 The prospective value of the risky investment
The estimation of the maximum acceptable individual loss level represents only the first
step of our analysis. As shown in Section 2.1, one of its consequences with direct impact
on non-professional investors resides in the determination of the optimal borrowing level.
This results as a byproduct of the optimization inside the risky portfolio that is undertaken
by the professional manager. For the non-professional client, it amounts to the optimal
choice in terms of wealth percentages invested, between risky and risk-less assets.
When investors decide about the optimal sum of money to be put in the risky portfolio
(equivalently in risk-free assets), they might not exclusively think in terms of VaR*,
but sooner aim at maximizing the utility generated by their financial investments. This
utility is encompassed in the prospect theory by the so called prospective value of the
risky investment Vt+1.4 Denoting the expected equity return premium by Et[xt+1] =
Et[Rt+1] − Rft and the probability of a positive premium by ωt = Pt(Et[Rt+1] ≥ Rft) =
Pt(Et[xt+1] ≥ 0), the prospective value of the risky investment can be formulated as:
Vt+1 = [ωt+(1−ωt)λ]StEt[xt+1]+(1−ωt)πt(λ−1)Rft−(1−πt)kEt[xt+1](St−Zt). (2.23)
Furthermore, we resolve to analyze the evolution of the prospective value for different
portfolio evaluation frequencies, on the grounds that revising portfolio performance at
different time intervals, implies drawing back on distinct return values, hence on different
return premia. This implicitly changes the values of several model parameters such as St,
Zt, πt, or ωt affecting the prospective value (2.23), a topic that detailed in Section 2.5.
Yet, in practice, risk averse investors may rely on a slightly different method for eval-
uating expected values of risky prospects. For instance, they may continue to consider
gains as unsure events and account for them as “wishes” (i.e. expectations). However,
losses would be assessed at their maximal impact, so to speak in a “worst case scenario”.
This being the case, gains flow into the definition of the prospective value as expected
gains, exactly as in Equation (2.23), while losses take the form of VaR*. In other words,
investors are sufficiently wary in order to count on the possibility of experiencing a max-
imal loss, hence to put an upper bound (in absolute value) on expected losses. It is this
upper bound that now generates utility (value) to the individual investor, and not the
expected loss. These considerations entail an alternative definition V ∗t+1 of the prospective
4Remember that our investors are not concerned with consumption and derive utility merely from the
financial wealth fluctuations.
14
value:
V ∗t+1 = ωtStEt[xt+1] + VaR∗
t+1
= (ωt + λSt)Et[xt+1]
+ [(πt − ϕ√
πt(1− πt))(λ− 1)Rft − (1− πt + ϕ√
πt(1− πt))kEt[xt+1]](St − Zt),
(2.24)
where the latter expression was derived according to Equation (2.18). In Section 3.2.2,
we investigate the evolution and implications of both prospective value definitions stated
here.
Before closing this section, we introduce a new definition that provides a basis to the
empirical analysis in Section 3.2.1. Similarly to the actual loss aversion defined in the
above Equation (2.14), a corresponding notion of global first order risk aversion can be
assessed from the prospective value as:
Λ(xt+1) =∂V
∂xt+1
= [ωt + (1− ωt)λ]St − (1− ωt)(1− πt)k(St − Zt). (2.25)
2.5 The impact of the portfolio revision frequency
As shown in the previous sections, the expected portfolio returns Et[Rt+1] (hence the
expected return premium Et[xt+1]) play a major role in the formulation of the value
function and consequently of almost all other variables of interest in our model (such as
VaR*, the prospective value, the optimal borrowing level, and also future cushions, gain
probabilities, etc.). Therefore, it is essential to notice that the value of returns directly
depends on the time horizon τ over which they are computed, i.e. or on the portfolio
revision frequency 1/τ . We hypothesize that different revision frequencies impact on
investor risk behavior leading to different investment decisions. The main reason for this
resides in the dependence of the computed performance of the risky portfolio on expected
returns, which further gives rise to the dependency of the investor attitude towards the
risky deposit and of the money invested in it on the portfolio revision frequency. The
higher the frequency of evaluation of risky investments is, the less likely is that risky
returns lie above risk-less ones, thus the more pronounced the investor disappointment
concerning the risky portfolio performance. Thus, given that, according to the prospect
theory, registered losses are perceived as more painful than gains of similar size, risky
investments become even less attractive.
The idea that the joint effect of narrow framing (myopia) over financial decisions and
reluctance to losses can dramatically impact risk perception and hence the subjective de-
sirability of risky investments, comes in line with the so-called myopic loss aversion. This
notion was was firstly introduced in Benartzi and Thaler (1995), and subsequently tested
in an experimental context in Thaler, Tversky, Kahneman, and Schwartz (1997), Gneezy
15
and Potters (1997), and Gneezy, Kapteyn, and Potters (2003). According to Barberis and
Huang (2004a), p. 4, myopia refers strictly to annual evaluations of gains and losses, hence
the term of narrow framing is better suited in describing the underlying phenomenon. In
a financial context, narrow framing illustrates the isolate evaluation of stock market risk
(i.e. unrelated to overall wealth risk). As underlined in Barberis and Huang (2004b), this
isolate evaluation determines an underestimation of the stock desirability, even though,
viewed in a wide utility-risk frame, they represent a good diversification modality.
According to Barberis and Huang (2004b), narrow framing can be interpreted as a
consequence of regret of not having taken another decision (non-consumption utility ex-
planation). Another explanation relies on the (higher) accessibility of (financial) infor-
mation that justifies its over-important role in final decisions. As referred in Kahneman
(2003), the easily accessible information is very appealing for the intuitive (for that spon-
taneous, effortless) way in which people use to make decisions. Our work draws upon
the latter motivation, namely accessibility. We consider it as better suited to financial
decisions, for that nowadays, investors are exposed to a tremendously high quantity of
financial information and need to make decisions in a fast changing financial environment.
Consequently, they tend to perform more frequent checks of their investments.
The empirical part of our paper (Section 3) analyzes closely the impact of various
revision horizons (ranging from one day to eight years) on the risk-free investment and
on the prospective value, where the focus lies on high revision frequencies (hence the ones
that are more plausible in practice), such as one day, one week, one month or one quarter.
In Section 3.2.1, we also plot and empirically assess the analytical form of the actual loss
aversion from Equation (2.14) and of the global risk aversion of first order from Equation
(2.25). However, in order to better understand how the revision frequency impacts the
prospective value and the investor attitude to risk, further explanations are necessary and
the rest of this section is dedicated to detailing this problem.
We start by noting that the first variable affected by the revision horizon τ is the
gross return value Rt(τ) = log(Pt/Pt−τ ) that accounts for the price variation over the
time interval τ . Therefore, the expected return premium Et[xt+1(τ)] = Et[Rt+1(τ)]−Rft
depends on the revision frequency.5 For instance, if prices are highly volatile in the short
run but do not change very much in mean in the long-run, a higher τ should generate
higher returns. However, even though there are more parameters (such as St, Zt, πt, etc.)
that are computed from Rt(τ), being thus affected by τ , in our (empirical) analysis, we
assume all λ, k, and Rft as fixed (i.e. independent of τ). Therefore, the changes of the
prospective value Vt+1 documented in Section 3.2.1 are a consequence of a chain impact
5For simplicity reasons, we henceforth drop most of the time-indices at places where we discuss the
dependence of the variables calculated at (the fixed) time t on τ .
16
whose very first seed is the revision horizon, but that does not imply the loss aversion
coefficient λ.6 Obviously, this chain reaction (hence its source, τ) also affects the actual
loss aversion and the first order risk aversion of first, but does not change the simple
coefficient of loss aversion:
Λloss(τ) = λSτ − (1− πτ )k(Sτ − Zτ ) (2.26a)
Λ(τ) = [ωτ + (1− ωτ )λ]Sτ − (1− ωτ )(1− πτ )k(Sτ − Zτ ). (2.26b)
In addition, we address a further theoretical issue which is closely related to the impact
of the portfolio revision frequency discussed above. Given that this frequency appears to
affect the investor risk perception, thus the level of risky investments, could the reverse
causality hold as well? In other words, for a certain loss aversion value (at date t), is there
a revision frequency that is optimal in terms of maximization of the prospective value?
In order to answer this question, we analyze the direct impact of Rt(τ) on the utility
maximization problem of individual investors. To this end, the c.p. dependence of the
prospective value V (x) from Equation (2.23)7 on x(τ) at time t is taken into account.
In other words, we study the direct dependence of utility on returns, but discard the
indirect effects generated by other model parameters influenced by returns.8 Under this
assumptions, the prospective value V (τ) at time t is linear in the return premium x,
with the proportionality coefficient δ = [ω + (1− ω)λ]S − (1− ω)(1− π)k(S − Z).9 The
prospective value is therefore maximized at each t for one of the extreme values of x(τ)
(the highest or the lowest, depending on the sign of δ) and, assuming there exists an
invertible computable form for the function x(τ), then τopt = x−1(τ). Naturally, if we
can find an invertible function V (τ), then the optimal revision horizon τ can be directly
computed at each trade t from the maximization problem of the prospective value. In fact,
this last simplified method is adopted in the empirical Section 3.2.1, where we suggest an
optimal value for the portfolio revision horizon τ based on practical considerations and
our data set.
6This chain reaction takes place in successive steps: (1) τ → Et[xt+1] =: xτ , (2) xτ → St =: Sτ , (3)
S1, S2, ...St → St − Zt =: Sτ − Zτ .7Or the corresponding V ∗(x) from Equation (2.24).8In essence, this can be considered a plausible assumption. The choice of an optimal current τ takes
place at the fixed time t where the model parameters indirectly affected by τ (i.e. St, Zt, πt), depend
on past values of x. The only exception is ωt(τ) that depends on τ through Et[xt+1(τ)], but assuming
that investors also assess ωt on the basis of past experience (e.g. as the frequency of past positive return
premia), we can confine ourselves to analyze the isolate role of Et[xt+1(τ)] in the prospective value
function.9The corresponding proportionality coefficient for V ∗(τ) yields δ∗ = (ω + λ)S − [1 − π +
ϕ√
π(1− π)]k(S − Z).
17
3 Empirical results
This chapter presents empirical findings complying with the theoretical results derived in
Section 2.
We consider that non-professional investors perceive risky investments according to
the value functions in Equations (2.11) and (2.12), and calculate the maximum loss level
according to Equation (2.18). The empirical analysis is based on daily data for the SP500
and the 10-year bond nominal returns (considered as the risky and the risk-free investment
alternative, respectively), ranging from 01/02/1962 to 03/09/2006 (11,005 observations).10
From this data set, we construct weekly, monthly, three-monthly, quarterly, yearly and
further lower frequency returns (ranging from two to eight years with an one-year in-
crement). We divide our sample in two equal parts and use the first one to estimate the
empirical mean and the standard deviation of the portfolio returns. The second half of the
data allows us to run the model on the basis of Sections (2.1) and (2.2) and to derive the
desired VaR*, as well as the wealth proportion invested in the risky portfolio (considered
here to be the index SP500). The remaining money is assumed to be automatically put
in the risk-free 10-year bond. Investors are assumed to start trading with an even initial
wealth allocation between the risky portfolio and the bond.
3.1 The evolution of the risky investment
3.1.1 The impact of the revision frequency
According to Benartzi and Thaler (1995), loss averse investors who evaluate the perfor-
mance of their portfolios once a year and employ a linear value function with standard
prospect theory parameter values, give rise to a market evolution that can explain the
equity premium observed in practice. We are interested in how varying the frequency
of portfolio evaluation can change investor decisions, hence the market evolution in our
setting.
We start by computing the portfolio VaR in Equation (2.10b) for either normally or
Student-t distributed portfolio gross returns and for a significance level of 5%. Then,
taking λ = 2.25 and k = 3 as in Barberis, Huang, and Santos (2001), as well as πt
identical to the empirical frequency of the cases where zt ≤ 1, we derive VaR∗t+1 according
to Equation (2.18). In this context, we consider different alternatives for the estimation of
expected portfolio returns, namely as the unconditional mean returns until the last date
before decision time, a zero mean process, or an AR(1) process. The value Zt of past
portfolio performance that impacts the valuation of current losses is taken to be identical
10Descriptive statistics can be found in Tables 8 and 9 of the Appendix.
18
to the last period risky asset holding Zt = St−1. The derived VaR∗t+1 is then plugged into
Equation (2.10a) in order to determine the optimal level Bt of borrowing or lending.
Table 1 presents the average percentages of wealth invested in the risky portfolio
St/Wt, where St is derived according to Equation (2.13), for different portfolio evaluation
horizons τ , normally and Student-t distributed portfolio returns Rt, and expected returns
Et[Rt+1] computed as the unconditional mean of past returns.11
Table 1: Percentage investments in SP500 for myopic cushions
Revision frequencyExpected returns
Normal Student-t
1 year 34.46 42.91
4 months 15.79 22.32
3 months 13.79 16.10
1 month 7.25 8.53
1 week 3.66 3.72
1 day 1.79 1.91
Accordingly, when investors are loss averse and use the VaR∗t+1 from Equation (2.18) as
measure of the maximal acceptable risk, higher portfolio revision frequencies entail lower
investments in the risky portfolio. This result is consistent with previous findings, such
as Benartzi and Thaler (1995) and Barberis, Huang, and Santos (2001). For normally
distributed returns, the risk aversion of investors appears to be higher compared to the
case with Student-t distribution.
3.1.2 The impact of the cushion
In a next step, we are interested in the interdependence among risky portfolio returns,
cushions and wealth percentages invested in the risky portfolio. In order to analyze this
issue, we fix the revision frequency at one year and plot annual returns on the SP500,
the evolution of the cushion St − Zt generated by series of past gains or losses, and the
resulting yearly percentage of wealth invested in the risky portfolio. As mentioned above,
the past performance benchmark is derived as the risky investment value in the previous
year Zt = St−1. The sample covers the 22 years of analysis (from 1983 to 2005).
At a first inspection, Figure 1 points to a positive correlation of the three variables
(SP500 yearly returns, yearly cushions, and yearly percentage investments in the risky
11Similar results are obtained when expected returns are derived as the zero mean or the AR(1) process.
See Tables 10 and 11 in the Appendix.
19
portfolio). The proportion of wealth invested the risky portfolio reaches its maximum
of 64.4% at the time when the yearly cushion attains its highest value, which is mainly
generated by the previous bull market observable in the SP500 returns. The importance of
the cushion for investor decisions can be traced back on Equation (2.18) that emphasizes
the twofold structure of the individual VaR*. The first term on the left-hand side of
this equation accounts for the expectation of future portfolio returns weighted by the loss
aversion parameter λ, while the second term is responsible for the influence of previous
performance (as encompassed by the cushion St − Zt). We denote them as the PT-term
and the cushion-term, respectively. Accordingly, positive expectations with respect to the
future evolution of the risky portfolio coupled with a positive cushion (i.e. past gains)
should render investors less loss averse. Consequently, given that VaR* directly enters Bt
and hence St, the wealth proportion invested in the risky portfolio St/Wt increases, as
illustrated in Figure 1. This effect is reversed when both return expectations and cushions
become negative.
Moreover, it is interesting to observe that small changes in the cushion at the beginning
of the effective trade period12 entails high variations in the portfolio allocation. This first
investor reaction turns strongly against investing money in risky assets, but the increase
in cushion makes it smooth over time, so that it ends by following fairly close the cushion
evolution. This result is again in line with the concept of loss aversion, i.e. the lower the
cushion of wealth accumulated in past trades is, the more loss averse investors become,
because they dispose of less back-up for later contingent losses. This lowers the wealth
fraction invested in risky assets.13
Further, we ask which is the impact of different ways of assessing the cushion on
investor decisions. In line with Barberis, Huang, and Santos (2001), we assume that
investors account for cumulative cushions (instead of the yearly cushions considered up to
this point), where the cumulation of past performance is carried out since the beginning
of the estimation sample. Table 2 presents the resulting wealth percentages invested in
the risky portfolio for various revision frequencies and different return distributions.
Interestingly, the yearly results with cumulative cushions under the normal distribution
exactly match the so called TIAA-CREF typical allocation mentioned in Benartzi and
Thaler (1995). This time, investors appear to be less risk averse for normally than for
Student-t distributed returns. Moreover, when investors make very frequent evaluations
(e.g. daily), they become extremely loss averse and end up by putting all money in the
12Remember that the effective trade (i.e. the observations that effectively underly the estimation
procedure) begins in the second half of our sample.13Gneezy and Potters (1997) test for the influence of experienced gains and losses on risk behavior, but
find no significant effect. However, as noted on p. 641, their experimental framework deviates from real
market settings, as considered in our model.
20
0 5 10 15 20 25−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
0.3
(a) Yearly SP500 returns.
0 5 10 15 20 25−5000
−4000
−3000
−2000
−1000
0
1000
2000
3000
4000
5000
(b) Yearly cushions.
0 5 10 15 20 250.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
(c) Yearly percentage investments in SP500.
Figure 1: Evolution of yearly SP500 returns, yearly cushions, and yearly percentages
invested in the risky portfolio.
21
Table 2: Percentage investments in SP500 for cumulative cushions
Revision frequencyExpected returns
Normal Student-t
1 year 49.93 41.53
4 months 14.55 14.52
3 months 5.68 5.89
1 month 1.87 1.32
1 week 0.50 0.10
1 day 0.00 0.00
risk-free asset.
Comparing Tables 1 and 2, we find out that, for a revision frequency of one year,
investors who rather cumulate than myopically treat past gains and losses to cushions, are
less risk averse. However, this situation reverses as the frequency of evaluation increases
(in fact, this phenomenon already occurs at a four month revision horizon). Also, the
speed at which the risk aversion increases is much higher for cumulated cushions. Thus,
investors relying only on short-term cushions continue to allocate modest but positive sums
to the risky portfolio, even when they check performance very often (e.g. every day). In
contrast, the net risky investment quickly converges to zero when cushions are based on all
previous trades. In general, these results support the idea that loss averse investors who
narrow frame financial projects and perform high frequency revisions become extremely
risk averse.
At this point, a further interesting empirical question arises: how long does it take
for an investor performing frequent evaluations to quit the risky market? In order to
answer this question, let us further assume that investors start with an initial investment
in risky assets of 50% of the total wealth. Figure 2 points out the dramatic effect of a high
revision frequency, i.e. when portfolio performance is checked every single day, investors
get out of the risky market in only few days. This behavior can be also explained in the
context of Equation (2.18), according to which, highly volatile SP500-returns and very low
cumulative cushions (as generated by the daily change in position) result in an enhanced
acceptable risk level VaR*. This captures the picture of an extremely risk averse investor.
3.1.3 An analysis with unadjusted VaR*
Finally, we analyze the investor behavior for a VaR* that exclusively accounts for maxi-
mum expected losses as in Equation (2.16). The results are quite similar to the previous
ones obtained for the variance adjusted VaR* from Equation (2.18). The explanation
22
0 1000 2000 3000 4000 5000 6000−0.25
−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
(a) Daily SP500 returns.
0 1000 2000 3000 4000 5000 6000−60
−40
−20
0
20
40
60
(b) Daily cushions.
0 1000 2000 3000 4000 5000 60000
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
(c) Daily percentage investments in SP500.
Figure 2: Evolution of daily SP500 returns, daily cushions, and daily percentages invested
in the risky portfolio.
23
becomes apparent in Figure 3, that illustrates the evolution of the probability of accu-
mulating prior gains πt = Pt(zt ≤ 1) for yearly and daily revision horizons, respectively.
Clearly, πt lies very close to one (especially for daily revisions) which renders the term√πt(1− πt) close to zero, hence makes the difference between the formulas (2.16) and
(2.18) disappear. Thus, both ways of estimating the VaR* introduced in Section 2.3
lead to similar results, more precisely close to the upper bound from Equation (2.19). For
daily revisions, the probabilities πt are almost one during the whole trading interval, which
yields zt = Zt/St ≤ 1 but also St − Zt ≈ 0. In other words, for this revision frequency,
the influence of the cushion St − Zt in Equations (2.16) or (2.18) becomes insignificant,
implying that the value of the VaR* depends solely on the first term in the definition
(i.e. what we denoted as the PT-term) λStEt[xt+1]. This underlines again the important
impact of cushions on investment decisions. When cushions are almost constant in time,
the PT-term (that is determined by the loss aversion coefficient λ and the risk premium
Et[xt+1] = Et[Rt+1] − Rft) dominates the maximum acceptable risk level VaR*. Thus,
loss aversion coupled with high market volatilities makes investors to renounce their risky
portfolios.
24
0 5 10 15 20 250.925
0.93
0.935
0.94
0.945
0.95
0.955
0.96
0.965
(a) Yearly revisions.
0 1000 2000 3000 4000 5000 60000.9997
0.9997
0.9998
0.9998
0.9999
0.9999
(b) Daily revisions.
Figure 3: Evolution of the probability of prior gains πt = Pt(zt ≤ 1).
25
3.2 The evolution of the prospective value
3.2.1 The impact of the revision frequency
According to the results in Section 3.1, the measured performance of the risky portfolio
varies with the revision horizon τ .
In order to closer analyze the impact and to determine an optimal value of τ at
each decision time t, we first recall the observation made in Section 2.5 that τ exerts
direct influence on the expected returns, thus on the expected return premium Et[xt+1] =
Et[Rt+1] − Rft. Therefore, the revision time affects the prospective value of the risky
investment from Equation (2.23).14 Here we distinguish between two terms with relevant
contribution to the formation of V (Et[xt+1]), namely the first term on the right hand side
of the Equation (2.23) that stands for the prospective value as considered in the original
prospect theory (that we denote as the PT-effect), and the second one (called the cushion-
effect) generated by the cushions of past gains or losses suggested in Barberis, Huang, and
Santos (2001).
Figure 4 illustrates the evolution and the contribution of these two effects to the final
prospective value in Equation (2.23), for revision frequencies of one day and one year,
respectively. The cushion is again myopically assessed, i.e. Zt = St−1.
At a first inspection of the upper panel in Figure 4, we find that the prospective value
V (Et[xt+1]) relies mainly on the PT-effect, that appears to be negative for the first years
of the total trading period (for the first 500 days approximately). This figure provides
also an additional explanation of why daily revisions result in lower risky investments.
The daily volatility does not allow to investors to accumulate significant cushions from
trading risky assets, hence the cushion-effect cannot overcome the PT-effect. Indeed, the
upper panel points out that, for portfolio checks performed more often than once a year,
the cushion is most of the time positive, but very low. However, the opposite occurs for
yearly revisions (lower panel), where the cushion-effect has a much stronger impact on
the prospective utility.
In the subsequent Figure 5, we plot again the prospective value and its two components
(the PT- and the cushion-effect) now as a function of the revision horizon τ , which ranges
from one day, one week, one month, three months, four months, one year, up to eight
years.15
Let us have a closer look at the curve in Figure 5. First, we note a visible kink of V (τ)
at τ = 1 year (corresponding to the fifth point on the horizontal axis16) that delimitates
14The effects of the alternative prospective value definition (2.24) are discussed in Section 3.2.2.15A revision frequency of eight years implies that investors can only make three portfolio checks during
our estimating sample. Therefore, a further increase of the revision time becomes senseless.16Each point on the abscissa stays for one of the considered revusion frequencies from one day to eight
26
0 1000 2000 3000 4000 5000 6000−0.25
−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
0.2
0.25
V(t)
Cushion−Effect
PT−Effect
(a) Daily revisions.
0 5 10 15 20 25−1500
−1000
−500
0
500
1000
1500
2000
2500
V(t)
Cushion−Effect
PT−Effect
(b) Yearly revisions.
Figure 4: Prospective value evolution for daily and yearly revisions.
27
0 2 4 6 8 10 12 14−2000
0
2000
4000
6000
8000
10000
12000
Cushion−Effect
V(t)
PT−Effect
Figure 5: Prospective value evolution for different revision frequencies.
two segments of different evolution. This reinforces the idea that, in practice, one year
represents indeed a “critical” evaluation frequency. As documented in Benartzi and Thaler
(1995), a decade ago, investors used to perform yearly portfolios checks. Nowadays, due to
the high amount of information available at almost no cost and to the enhanced dynamic of
the market events, we claim that investors manifest the tendency to reconsider the problem
of splitting their money between risky and risk-free assets much more often. Yet, one year
remains an important anchor in the investor minds, given that, on one hand, various
events (such as release of annual activity reports) take place with this frequency and,
on the other hand, non-professional investors may not be sufficiently impatient (perhaps
because they do not dispose of enough time and financial resources) to perform more often
portfolio checks.
The evident segmentation of the prospective value for revision frequencies lower and
higher than one year motivates us to have a closer look at the two separate frequency
segments, as encompassed in Figure 6. Our goal is to find an analytical form that underlies
this evolution and that would allow us to determine the optimal revision frequency.
Optically, for τ < 1 year, the prospective value function V (τ) appears to exhibit a
parabolic (convex) form, while for τ ≥ 1 year, it rather evolves in periodical waves of small
amplitude following an ascending trend. In effect, fitting an analytical model to the data
yields a parabola for τ < 1 year and a polynomial of fifth degree for τ ≥ 1 year, whose
curvature coefficient estimates are given in Table 3. Of note is that already a linear model
of the type a1x + a2, where a1 = 1, 292 (with 95%-CI (1, 135; 1, 450)), and a2 = −1, 089
(with 95%-CI (−1, 884;−293.3)), fits the data in the right segment reasonably well (R-
square: 0.9853, Adjusted R-square: 0.9829, RMSE: 417.2) with respect to the best fit
obtained for the left revision horizon segment.
years.
28
1 1.5 2 2.5 3 3.5 4 4.5 5−10
0
10
20
30
40
50
60
70
80
90
V(t)
Cushion−Effect
PT−Effect
(a) Daily, weekly, quarterly and four-monthly revisions.
1 2 3 4 5 6 7 8−2000
0
2000
4000
6000
8000
10000
12000
Cushion−Effect
V(t)
PT−Effect
(b) Revisions from one to eight years.
Figure 6: Prospective value evolution on the two relevant revision frequency segments.
29
Table 3: Estimated prospective value evolution as a function of the portfolio revision
frequency for λ = 2.25, k = 3.
Fitted model Coefficient 95%-CI Goodness of fit
τ < 1 year year
a1x2 + a2x + a3
a1 = 7.152 (1.709; 12.6) R-square: 0.9909
a2 = −22.46 (−55.75; 10.83) Adjusted R-square: 0.9819
a3 = 15.47 (−28.21; 59.15) RMSE: 4.734
τ ≥ 1 year
b1 = 4.076 (−6.174; 14.33)
b2 = −84.45 (−315.7; 146.8)
b1x5 + b2x
4 + b3x3 b3 = 630.9 (−1, 305; 2, 566) R-square: 0.9993
+b4x2 + b5x + b6 b4 = −1, 994 (−9, 365; 5, 377) Adjusted R-square: 0.9975
b5 = 3, 518 (−8, 885; 15, 920) RMSE: 159.1
b6 = −1, 343 (−8, 397; 5, 711)
30
In sum, table 3 suggests the following analytical form of the prospective value as a
function of the portfolio revision horizon:
V (τ) =
a1τ2 + a2τ + a3, for τ ≤ 1 year
b1τ5 + b2τ
4 + b3τ3 + b4τ
2 + b5τ + b6, for τ > 1 year,(3.1)
which is plotted in the subsequent Figure 7. We note that there is a high jump (kink)
in the prospective value at what we consider to be the reference frequency of τ = 1 year,
namely V (1 year−) = 0.16 ¿ 731.53 = V (1 year+)). For the linear model fitting the right
segment of the curve, we obtain V (1 year+) = 203 À V (1 year−).
1 1.5 2 2.5 3 3.5 4 4.5 5
0
10
20
30
40
50
60
70
80
(a) Daily, weekly, quarterly and four-monthly revisions.
1 2 3 4 5 6 7 8
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
(b) Revisions from one to eight years.
Figure 7: Curve fitting for the prospective value on the two relevant revision frequency
segments.
Interestingly, the two-segment evolution illustrated in Figure 5 and formally assessed
in Equation (3.1) offers an original perspective over the prospective value in the revision
frequency domain. It relies on the visual similarity to the value function in terms of gains
and losses, as formulated in the original prospect theory. More precisely, the prospective
31
value of the risky portfolio in time units appears to mirror the value function in money
units. For better understanding this claim, recall first that, in the revision horizon do-
main, the individually perceived value of prospective risky investments exhibits a twofold
evolution, namely it unfolds convexly for revision horizons lower than one year, while its
course is (almost) linear for revision times higher than one year. In other words, the
prospective value drops much faster for low revision horizons (in the left time segment)
than for high ones (in the right time segment). Thus, in the left revision horizon segment,
individual investors appear to upgrade much faster risky prospective values for revision
horizons approaching the kink-value of one year, while in the right segment there is almost
no significant difference in the valuation change for different frequencies. However, the
slope of the prospective value in revision time domain appears to be lower to the left than
to the right of the reference horizon of one year (i.e. in Equation (3.1), a2 for τ < 1 year
is lower in absolute value than b5 for τ > 1 year), as opposed to the value function where
loss aversion becomes manifest in the left domain (λ > 1 for losses). However, low revision
horizons are equivalent to high revision frequencies, so we conclude on a steeper slope for
low frequencies in revision time domain, similar to the value function evolution for neg-
ative monetary units. We denote this phenomenon as loss aversion in revision frequency
domain. Also, the two frequency segments are separated by the critical revision frequency
of one year. According to the practical considerations (concerning real market conditions,
information overflow and impatience of real investors, as presented below), this critical
frequency can be considered as an anchor (reference point) that provides mental support
to individual investors who resolve to invest in risky portfolios.
Section 5.3 in the Appendix presents the results of various robustness checks performed
for further values of the loss aversion coefficient λ and of the cushion sensitivity parameter
k. Figures 11, 12, and 13 confirm the important contribution of the cushion-effect to the
prospective value, as well as the segmentation for revision horizons lower and higher than
one year. Interestingly, loss loving investors (i.e. λ < 1) appear to perceive the risky
prospect as having a negative and concavely decreasing value in the revision horizon
domain. Formally, this draws back on Equation (2.23), where λ < 1 and k > 0 entail a
negative cushion-effect. Given that this effect dictates the course of the entire prospective
value and that it increases in absolute value for higher revision horizons, the prospective
value decreases in τ . For loss neutral investors (i.e. λ = 1), the cushion effect becomes
nil, hence the PT-effect controls the evolution of the prospective value. However, note
that magnitude of the PT-effect is similar to the case with λ = 2.25. Again, Equation
(2.23) and the probability of experiencing past gains (πt) which lies in our data set very
close to one (see Figure 3) provide explanations of this fact. Finally, the prospective value
obtained for λ = 3 exhibits similar evolution to the original case with λ = 2.25, with
32
a higher peak at the revision horizon of eight years and a more pronounced sinusoidal
component. In sum, the attitude to losses captured by the loss aversion coefficient λ
appears to decide upon which of the cushion- or PT-effects has the major contribution to
the final prospective value, thus affects the current perception of the risky investment at
different revision frequencies.
In addition, an analogous curve fitting procedure performed for these all three further
cases reinforces these conclusions. For each λ ∈ 0.5; 1; 3, a parabola continues to best
describe the prospective value evolution in the high revision frequency segment. In con-
trast, for the low frequency segment, a parabola attains the best fit for loss loving investors
λ = 0.5, but more complex dependencies emerge for higher loss aversion, namely a sixth
order polynomial for loss neutral investors λ = 1 and a fourth order one for extremely
loss averse investors λ = 3. Overall, the kink at the reference revision horizon of one year
is again evident and the slope of the fitted curves remains smaller in the low-frequency
domain underlining the same phenomenon that we called the loss aversion in revision
frequency domain.
Tables 14 and 15 show that the parameter k exerts no significant influence on the
evolution of the prospective value. Confirming the qualitative results in Figures 11, 12,
and 13, the prospective value is negative and decreases in the revision horizon for λ ≤ 1,
but becomes positive and growing for loss averse investors λ > 1.
Furthermore, as discussed in Section 2.5, the portfolio revision frequency reflects on
the actual loss aversion defined in Equation (2.14) as well as on the global first order risk
aversion from Equation (2.25). Subsequently, we conduct an empirical analysis similar
to the above investigations with respect to the evolution of the prospective value. First,
we observe that both actual loss aversion and global first order risk aversion exhibit
resembling patterns, being derived from two interrelated functions, namely the value
function and the prospective value (that results from the value function), respectively.
Also, the obtained values for both measures of the real investor risk attitude are of the same
order of magnitude as the prospective value in the revision frequency domain. As depicted
in Figure 8, there is an apparent similar segmentation around the revision frequency of
one year (corresponding to the fifth point on the horizontal axis).
33
0 2 4 6 8 10 12 140
0.5
1
1.5
2
2.5
3
3.5
4
4.5x 10
4
Actual loss aversion
Global loss aversion
(a) All frequency revisions.
1 1.5 2 2.5 3 3.5 4 4.5 50
1000
2000
3000
4000
5000
6000
7000
8000
9000
Actual loss aversion
Global loss aversion
(b) Daily, weekly, quarterly and four-monthly revisions.
1 2 3 4 5 6 7 81.5
2
2.5
3
3.5
4
4.5x 10
4
Actual loss aversion
Global loss aversion
(c) Revisions from one to eight years.
Figure 8: Evolution of the actual loss aversion and global first order risk aversion on the
two relevant revision frequency segments.
34
In mathematical terms, the evolution on the left (right) revision horizon segment is one
order of complexity higher than the prospective value, namely it corresponds to a third
(sixth) order polynomial. The curvature parameters estimates are presented in Table
4 (and Table 16 in the Appendix), while and the corresponding theoretical courses are
illustrated in Figure 9.
Table 4: Estimated evolution of the actual loss aversion as a function of the portfolio
revision frequency for λ = 2.25, k = 3.
Fitted model Coefficient 95%-CI Goodness of fit
τ < 1 year
a1 = −122.3 (−694; 449.5)
a1x3 + a2x
2 a2 = 1, 337 (−3, 841; 6, 515) R-square: 0.9992
+a3x + a4 a3 = −2, 415 (−16, 370; 11, 540) Adjusted R-square: 0.9968
a4 = 1, 959 (−8, 714; 12, 630) RMSE: 170.8
τ ≥ 1 year
b1 = 49.17 (−141.1; 239.5)
b2 = −1, 343 (−6, 487; 3, 801)
b1x6 + b2x
5 + b3x4 b3 = 14, 150 (−40, 470; 68, 760) R-square: 0.9956
+b4x3 + b5x
2 + b6x + b7 b4 = −71, 930 (−360, 000; 216, 200) Adjusted R-square: 0.969
b5 = 180, 000 (−603, 600; 963, 700) RMSE: 1,327
b6 = −196, 600 (−1, 215, 000; 822, 100)
b6 = 96, 460 (−381, 500; 574, 500)
35
1 1.5 2 2.5 3 3.5 4 4.5 5
1000
2000
3000
4000
5000
6000
7000
8000
actual loss aversion
global risk aversion
(a) Daily, weekly, quarterly, four-monthly, and yearly revisions.
1 2 3 4 5 6 7 8
2
2.5
3
3.5
4
4.5
x 104
actual loss aversion
global risk aversion
(b) Revisions from two to eight years.
Figure 9: Curve fitting for the actual loss aversion and the global first order risk aversion
on the two relevant revision frequency segments.
36
Finally, we return to the question concerning the optimal revision frequency. It appears
natural to assume that investors concerned exclusively with financial investments (and not
with other sources of utility such as consumption) attempt to maximize the prospective
value of their risky portfolios. In fact, smart investors could look for an optimal revision
frequency, i.e. that maximizes the prospective value at a given decision time t. According
to the functional forms fitted to our data set for the prospective value (3.1), as well as for
the actual loss aversion and global first order risk aversion (from Tables 4 and 16), these
variables describing the investor risk behavior are increasing in τ in each of the two relevant
segments. Thus, the optimal τ lies at the upper end of the corresponding segment (i.e.
is identical to the maximal τ of the segment).17 Consequently, in each revision frequency
segment, (loss averse) investors should check the performance of their risky investment
as seldom as possible in order to maximize the corresponding prospective value of this
investment.18 Again, the highest revision frequency of one day entails a minimal expected
value of the risky portfolio, pushing investors to step out of the risky market and to
allocate all their money to risk-free assets. Yet, we know that practical reasons (such as
the huge amount of financial data available almost at no costs to every individual investor
and the high interest raised by financial events in general) entail an increase of the revision
frequency below the limit of one year. Thus, investor perceptions lie sooner in the left
revision horizon domain and the crossover to the second segment is improbable. Therefore,
accounting for the today’s financial market evolution and importance, we consider τ = 1
year as the most reasonable revision time that would increase the perceived returns of
risky investments.
3.2.2 An analysis under the “worst case scenario”
For the “worst case scenario” described in Section 2.4, investors may use a slightly different
definition of the prospective value, namely according to Equation (2.24). Figure 10 plots
the evolution of both Vt+1 in line with the original prospect theory and the new V ∗t+1,
that appear to be almost identical. Thus, the discussion on the evolution in the revision
frequency domain in Section 3.2.1 is also valid the “worst case scenario”. We conclude
that the hypothesis that prudent investors perceive risks in this “worst case” appears to
be acceptably realistic in the context of variable revision frequencies, because it generates
17The situation is inverted for λ ≤ 1, as presented in Section 5.3 of the Appendix. However, the same
arguments hold if we consider that loss loving and loss neutral investors attempt to maximize the absolute
prospective value of the risky investment.18In fact, the optimality of the revision frequency should be sooner understood from the viewpoint of
portfolio managers, whose interest is to attract more clients willing to invest money in risky assets. Rec-
ommending to these clients to undertake performance checks in the “optimal” frequency should maximize
the budget at manager’s disposal.
37
results that are similar those stemming from considerations of the original prospect theory.
0 5 10 15 20 25−1000
−500
0
500
1000
1500
2000
2500
3000
V(t)*
V(t)
Figure 10: Prospective values in the original prospect theory-formulation vs. the “worst
case scenario” for yearly revisions.
3.3 A comparison to the portfolio optimization framework
3.3.1 VaR*-equivalent significance levels
One further question of interest arises from the use of the VaR∗ as a measure of risk in
the portfolio optimization model in Section 2.1. Statistically, VaR∗ represents the lower
quantile of portfolio returns at a given (i.e. fixed) significance level α (or confidence
level 1 − α), where usually α ∈ [0.01, 0.1]. The individually optimal VaR∗t+1 (that is in
fact previously derived by investors on the basis of subjective considerations according
to Equation (2.18)), is compared to the portfolio VaR in Equation (2.10b), in order to
determine how investor wealth is going to be split between the risky portfolio and the
risk-free bond (where the sum to be invested in risk-free assets is formalized in Equation
(2.10a)). We denote by α∗t the significance level that corresponds to the VaR∗t+1 com-
puted in our model. Thus, if the portfolio VaR at time t corresponds to an α < α∗t (or
equivalently, to a confidence level 1 − α > 1 − α∗t ), then the sign of Equation (2.10a) is
negative. In words, too much risk would arise by putting the entire wealth in the risky
portfolio, so that, in order to accommodate the desired (lower) risk level, a percentage of
the investor wealth should be lent, i.e. invested in the risk-free asset (Bt < 0). On the
contrary, if α > α∗t , then the portfolio risk meets the individual risk requirements (being
lower than the subjective risk threshold) and investors borrow extra money (Bt > 0) in
order to increase their SP500-holdings.
In this section, we determine the significance level corresponding to the value of VaR∗t+1
as derived from Equation (2.18), using either normal or the Student-t return distributions
38
and considering again the initial setting with short-term cushions Zt = St−1. Table 5
presents equivalent significance levels averaged over time and denoted as α∗.
Table 5: Equivalent significance levels of the estimated VaR∗t+1.
Revision frequencyα∗
Normal Student-t
1 year 34.94 43.95
4 months 40.49 46.50
3 months 43.18 46.34
1 month 45.23 46.67
1 week 47.69 48.01
1 day 48.90 49.06
Table 5 emphasizes an interesting empirical fact. As stated above, classical portfolio
selection models based on VaR assume that the significance level chosen by investors is
low (i.e. α ∈ [0.01, 0.1]). By contrast, our results (i.e. the much higher values α∗ > 0.3
obtained for any revision frequency higher than one year) show that this assumption does
not match real market data and that investors might be in practice substantially more loss
averse than considered in theory. The lowest α∗ in Table 5 is obtained for the standard
revision frequency of one year but lies still far above commonly assumed significance levels
(between 1−10%). Also, when investors revise their portfolios more frequently (e.g. every
day), α∗ enhances tremendously (up to approximatively 50%) pointing out an excessive
risk aversion.
3.3.2 VaR*-equivalent loss aversion levels
In the same context of equivalency, we now address the impact of an exogenous VaR* as
originally employed in Campbell, Huisman, and Koedijk (2001), on the values of the loss
aversion coefficient λ∗t+1, computed according to Equation (2.22) in our model. To this
end, we go back on the commonly used significance levels of 10% and 1% and estimate
an equivalent exogenous VaR* as derived from Equation (2.10b). Then, we compute the
corresponding λ∗t+1 from Equation (2.22).
Tables 6 and 7 present the wealth percentages invested in the risky portfolio subject
to the assumed portfolio returns distribution (i.e. either normal or Student-t) as resulting
from our model, as well as the generated average equivalent loss aversion coefficient λ∗ for
the two significance levels mentioned above (10% and 1%, respectively). Remember that
the portfolio VaR in Equation (2.10b) is estimated using a 5% significance level that is
going to be considered as the benchmark for the values in these tables.
39
Table 6: Percentage investments in SP500 and the average λ∗ for yearly cushions, expected
returns = unconditional mean and α = 0.10.
Revision frequency
Expected returns
Normal Student-t
Wealth % λ∗ Wealth % λ∗
1 year 77.21 1.43 72.80 1.29
4 months 77.66 1.20 73.07 1.21
3 months 77.73 1.07 73.13 1.08
1 month 77.90 1.08 73.25 1.09
1 week 77.94 0.86 73.28 0.86
1 day 77.94 0.93 73.27 0.93
Table 7: Percentage investments in SP500 and the average λ∗ for yearly cushions, expected
returns = unconditional mean and α = 0.01.
Revision frequency
Expected returns
Normal Student-t
Wealth % λ∗ Wealth % λ∗
1 year 136.50 1.32 160.49 1.40
4 months 139.76 1.18 164.80 1.26
3 months 140.14 1.15 165.29 1.23
1 month 140.91 1.09 166.30 1.14
1 week 141.25 1.05 166.75 1.07
1 day 141.35 1.02 166.89 1.03
40
Tables 6 and 7 show that the equivalent recommendations of our model for 10% (1%)
significance lie well below (above) the benchmark VaR at 5%. This points out a higher
(lower) risk aversion as resulting in our endogenous VaR*-framework (after restating it
in terms of the exogenous VaR model) relative to the portfolio risk measured by VaR.
However, these percentages are still much higher than those in Table 1, where VaR* is
treated as endogenous. This difference is also observed in the lower values of the equivalent
loss aversion coefficient λ∗ implied by significance levels commonly assumed in exogenous
settings.
Moreover, according to the value function from Equations (2.11) and (2.12), for λ = 1
(and recalling that k = 0), gains and losses exhibit identical values in the investor percep-
tion. An equivalent extreme situation (with λ ≈ 1) can be observed in Tables 6 and 7 for
high revision frequencies (over one month). In the rest of the cases, even for low frequency
revisions (e.g. one year), λ∗ is higher than one (indicating slight loss aversion) but lower
than 2.25 (the empirical level estimated in the original prospect theory and largely used in
previous empirical research19). This reinforces our earlier claim that assuming low signif-
icance levels (as it is the common case in previous portfolio optimization research) entails
an underestimation of the real investor loss aversion captured by the specific coefficient
λ.
4 Conclusions
In this paper we investigate the risk behavior of non-professional investors faced with the
problems of fixing a maximal acceptable level of financial losses and of splitting money
between risk-free assets and a risky portfolio. We assume that these investors are loss
averse, narrow frame financial investments and perceive future portfolio returns subject
to past performance.
We extend the portfolio allocation model developed in Campbell, Huisman, and Koedijk
(2001) in order to incorporate the effect of a desired VaR*, that is now subjectively assessed
by individual loss averse investors. Thus, the first task of non-professional investors con-
sists in fixing the VaR*-level that is subsequently communicated to professional portfolio
managers in charge of finding the optimal portfolio composition. The portfolio optimiza-
tion procedure delivers also the optimal sum of money to be invested in risk-free assets,
which represents another important decision variable for the non-professional investor.
In modelling the investor perception over the risky investment that yields the subjec-
tive VaR*, we rely on the notion of myopic loss aversion introduced in Benartzi and Thaler
(1995) and employ the extended subjective valuation of prospective risky investments pro-
19Such as Barberis, Huang, and Santos (2001), Benartzi and Thaler (1995).
41
posed in Barberis, Huang, and Santos (2001). We integrate these behavioral explanations
in the portfolio decision framework mentioned above, enriching the two models with orig-
inal findings that stem both from theoretical consideration and empirical results obtained
on the basis of real market data (such as SP500 and US 10-year bond price series).
Considering that investors are concerned merely with financial investments, we theo-
retically model their perception over the utility of risky assets and define the maximum
individually sustainable level of financial losses VaR*. This level serves to deciding upon
the optimal amount of money to be invested in the risky portfolio. Also, we assess the
utility of risky prospect captured by the prospective value and suggest two ways of quan-
tifying actual risk attitudes. Moreover, we investigate the influence of different revision
frequencies on the prospective value and the actual risk attitude and derive an optimal
performance check horizon under consideration of practical constraints.
The theoretical results are supported and extended by our empirical findings which, in
sum, show that non-professional investors allocate the main part of their wealth to risk-free
assets. A smaller sum is put into the risky portfolio for increasing frequencies of revising
its performance. Also, financial wealth fluctuations determined by the success of previous
decisions exert a significant impact on the current portfolio allocation, making investors
without substantial gain cushions to firmly refuse holding risky assets. One year appears
to be a critical revision frequency, optimal from the viewpoint of maximizing risky holdings
and commonly used in practice. This revision frequency splits individual perceptions over
risky investments (captured by the prospective value) and over market risk in general
(captured by the actual loss aversion and the global first order risk aversion) in two
qualitatively different segments with distinct evolutions. Estimates of the analytical form
of these functions point out a more pronounced slope for revision frequencies higher than
one year, which we denote as loss aversion in the revision frequency domain. Moreover, the
computation of equivalent loss aversion values for confidence levels commonly assumed
in previous research suggests an underestimation of the loss aversion coefficient of real
non-professional investors.
42
5 Appendix
5.1 Descriptive statistics
Table 8: Log-difference of the SP500 index for quarterly and yearly portfolio revisions.
SP500Revision frequency
Quarterly Yearly
Mean 0.017 0.073
Median 0.017 0.070
Std.Dev. 0.006 0.026
Kurtosis 0.623 0.974
Skewness 0.951 1.042
Max. 0.036 0.142
Min. -0.009 0.037
Obs. 175 43
Table 9: Log-difference of the 10-year bond for quarterly and yearly portfolio revisions.
10-yearRevision frequency
Quarterly Yearly
Mean 0.017 0.066
Median 0.018 0.071
Std.Dev. 0.079 0.137
Kurtosis 2.661 -0.659
Skewness -0.671 -0.205
Max. 0.290 0.345
Min. -0.302 -0.207
Obs. 175 43
43
5.2 Percentage invested in SP500 using different expected re-
turns
Table 10: Percentage investment in SP500 for myopic cushions and expected returns =
zero mean.
Revision frequencyExpected returns
Normal Student-t
1 year 29.63 24.42
4 months 15.55 12.85
3 months 12.43 10.28
1 month 7.45 6.14
1 week 3.70 3.04
1 day 1.79 1.46
Table 11: Percentage investment in SP500 for myopic cushions and expected returns =
AR(1)0.
Revision frequencyExpected returns
Normal Student-t
1 year 31.55 25.31
4 months 16.22 13.29
3 months 14.01 11.36
1 month 7.27 6.02
1 week 0.00 0.00
1 day 0.00 0.00
5.3 The prospective value evolution as a function of the portfolio
revision frequency for different parameter values
44
0 2 4 6 8 10 12 14−4000
−3500
−3000
−2500
−2000
−1500
−1000
−500
0
500
PT−Effect
Cushion−Effect
V(t)
(a) All revision frequencies.
1 1.5 2 2.5 3 3.5 4 4.5 5−25
−20
−15
−10
−5
0
PT−Effect
Cushion Effect
V(t)
(b) Daily, weekly, quarterly, and four-monthly revisions.
1 2 3 4 5 6 7 8−4000
−3500
−3000
−2500
−2000
−1500
−1000
−500
0
500
PT−Effect
Cushion−Effect
V(t)
(c) Revisions from one to eight years.
Figure 11: Prospective value evolution for different revision frequencies, λ = 0.5, and
k = 3.
45
0 2 4 6 8 10 12 14−1600
−1400
−1200
−1000
−800
−600
−400
−200
0
200
Cushion−Effect
PT−Effect = V(t)
(a) All revision frequencies.
1 1.5 2 2.5 3 3.5 4 4.5 5−0.9
−0.8
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0
Cushion−Effect
PT−Effect = V(t)
(b) Daily, weekly, quarterly, and four-monthly revisions.
1 2 3 4 5 6 7 8−1600
−1400
−1200
−1000
−800
−600
−400
−200
0
200
Cushion−Effect
PT−Effect = V(t)
(c) Revisions from one to eight years.
Figure 12: Prospective value evolution for different revision frequencies, λ = 1, and k = 3.
46
0 2 4 6 8 10 12 14−5000
0
5000
10000
15000
20000
Cushion−Effect
V(t)
PT−Effect
(a) All revision frequencies.
1 1.5 2 2.5 3 3.5 4 4.5 5−20
0
20
40
60
80
100
120
140
160
V(t)
Cushion−Effect
PT−Effect
(b) Daily, weekly, quarterly, and four-monthly revisions.
1 2 3 4 5 6 7 8−5000
0
5000
10000
15000
20000
Cushion−Effect
V(t)
PT−Effect
(c) Revisions from one to eight years.
Figure 13: Prospective value evolution for different revision frequencies, λ = 3, and k = 3.
47
Table 12: Estimated prospective value evolution as a function of the portfolio revision
frequency for τ < 1 year.
Fitted model Coefficient 95%-CI Goodness of fit
λ = 0.5, k = 3
a1x2 + a2x + a3
a1 = −1.957 (−3.972; 0.05797) R-square: 0.9861
a2 = 5.561 (−6.763; 17.88) Adjusted R-square: 0.9722
a3 = −3.471 (−19.64; 12.7) RMSE: 1.752
λ = 1, k = 3
a1x2 + a2x + a3
a1 = −0.06707 (−0.1443; 0.01011) R-square: 0.9847
a2 = 0.1749 (−0.2971; 0.6469) Adjusted R-square: 0.9695
a3 = −0.1023 (−0.7217; 0.517) RMSE: 0.06712
λ = 3, k = 3
a1x2 + a2x + a3
a1 = 14.69 (5.439; 23.94) R-square: 0.9932
a2 = −48.4 (−105; 8.154) Adjusted R-square: 0.9863
a3 = 34.82 (−39.4; 109) RMSE: 8.042
48
Table 13: Estimated prospective value evolution as a function of the portfolio revision
frequency for τ ≥ 1 year.
Fitted model Coefficient 95%-CI Goodness of fit
λ = 0.5, k = 3
b1x2 + b2x + b3
b1 = −54.5 (−86.1;−22.89) R-square: 0.9888
b2 = −15.41 (−306.8; 276) Adjusted R-square: 0.9844
b3 = −171.2 (−742.8; 400.3) RMSE: 159.4
λ = 1, k = 3
b1 = 2.875 (−7.904; 13.66)
b2 = −81.35 (−372.8; 210.1)
b1x6 + b2x
5 + b3x4 b3 = 893.2 (−2201; 3987) R-square: 0.9972
+b4x3 + b5x
2 + b6x b4 = −4808 (−21, 130; 11, 510) Adjusted R-square: 0.9801
+b7 b5 = 13, 150 (−31, 250; 57, 540) RMSE: 75.19
b6 = −16, 940 (−74, 650; 40, 770)
b7 = 7781 (−19, 300; 34, 860)
λ = 3, k = 3
b1 = 60.27 (−72.43, 193)
b1x4 + b2x
3 + b3x2 b2 = −1071 (−3470; 1329) R-square: 0.8996
+b4x + b5 b3 = 6536 (−8232; 21, 300) Adjusted R-square: 0.7658
b4 = −14, 570 (−49, 980; 20, 840) RMSE: 1774
b5 = 11, 000 (−15, 440; 37, 510)
49
1 1.5 2 2.5 3 3.5 4 4.5 5
−25
−20
−15
−10
−5
0
(a) Daily, weekly, quarterly, four-monthly, and yearly revisions.
1 2 3 4 5 6 7 8−4000
−3500
−3000
−2500
−2000
−1500
−1000
−500
0
(b) Revisions from two to eight years.
Figure 14: Curve fitting for the prospective value on the two relevant revision frequency
segments for λ = 0.5, k = 3.
50
1 1.5 2 2.5 3 3.5 4 4.5 5
−0.9
−0.8
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0
(a) Daily, weekly, quarterly, four-monthly, and yearly revisions.
1 2 3 4 5 6 7 8
−1600
−1400
−1200
−1000
−800
−600
−400
−200
0
200
(b) Revisions from two to eight years.
Figure 15: Curve fitting for the prospective value on the two relevant revision frequency
segments for λ = 1, k = 3.
51
1 1.5 2 2.5 3 3.5 4 4.5 5
0
20
40
60
80
100
120
140
160
(a) Daily, weekly, quarterly, four-monthly, and yearly revisions.
1 2 3 4 5 6 7 8
0
2000
4000
6000
8000
10000
12000
(b) Revisions from two to eight years.
Figure 16: Curve fitting for the prospective value on the two relevant revision frequency
segments for λ = 3, k = 3.
52
Table 14: Prospective value evolution for different revision frequencies (lower than one
year) and different parameter values.
Revision frequency
1 day 1 week 1 month 3 months 4 months
λ = 0.5
k = 3 -0.03 -0.35 -2.90 -14.38 -23.93
k = 10 -0.03 -0.35 -2.90 -14.38 -23.93
k = 20 -0.03 -0.35 -2.90 -14.38 -23.93
λ = 1
k = 3 0.00 -0.02 -0.13 -0.55 -0.88
k = 10 0.00 -0.02 -0.13 -0.55 -0.88
k = 20 0.00 -0.02 -0.13 -0.55 -0.88
λ = 2.25
k = 3 0.08 0.85 7.91 44.45 80.55
k = 10 0.08 0.85 7.91 44.45 80.55
k = 20 0.08 0.85 7.91 44.45 80.55
λ = 3
k = 3 0.13 1.42 13.69 82.36 158.27
k = 10 0.13 1.42 13.69 82.36 158.27
k = 20 0.13 1.42 13.69 82.36 158.27
53
Table 15: Prospective value evolution for different revision frequencies (higher than one
year) and different parameter values.
Revision frequency
1 year 2 years 3 years 4 years 5 years 6 years 7 years 8 years
λ = 0.5
k = 3 -151.83 -534.59 -741.86 -1054.10 -1740.59 -1957.99 -3065.79 -3794.99
k = 10 -151.83 -534.59 -741.86 -1054.10 -1740.59 -1957.99 -3065.79 -3794.99
k = 20 -151.83 -534.59 -741.86 -1054.10 -1740.59 -1957.99 -3065.79 -3794.99
λ = 1
k = 3 -7.65 -99.63 104.55 -185.69 -373.27 26.59 -245.66 -1553.26
k = 10 -7.65 -99.63 104.55 -185.69 -373.27 26.59 -245.66 -1553.26
k = 20 -7.65 -99.63 104.55 -185.69 -373.27 26.59 -245.66 -1553.26
λ = 2.25
k = 3 736.51 1510.73 2540.90 3618.29 5345.62 6437.69 7744.96 9872.30
k = 10 736.51 1510.73 2540.90 3618.29 5345.62 6437.69 7744.96 9872.30
k = 20 736.51 1510.73 2540.90 3618.29 5345.62 6437.69 7744.96 9872.30
λ = 3
k = 3 2063.66 -65.86 3512.46 2208.50 7055.31 5088.58 6904.74 11504.94
k = 10 2063.66 -44.52 3786.25 2271.02 7055.31 5297.73 7043.24 11504.94
k = 20 2063.66 89.26 3590.36 2360.35 7055.31 5596.52 7241.11 11504.94
54
Table 16: Estimated evolution of the first order global risk aversion as a function of the
portfolio revision frequency for λ = 2.25, k = 3.
Fitted model Coefficient 95%-CI Goodness of fit
τ < 1 year
a1 = −110.8 (−613; 391.4)
a1x3 + a2x
2 a2 = 1, 207 (−3, 342; 5, 755) R-square: 0.9992
+a3x + a4 a3 = −2, 175 (−14, 430; 10, 080) Adjusted R-square: 0.997
a4 = 1, 766 (−7, 610; 11, 140) RMSE: 150
τ ≥ 1 year
b1 = 27.87 (−107.2; 163)
b2 = −759.2 (−4, 412; 2, 893)
b1x6 + b2x
5 + b3x4 b3 = 7, 876 (−30, 910; 46, 660) R-square: 0.9974
+b4x3 + b5x
2 + b6x + b7 b4 = −38, 690 (−243, 300; 165, 900) Adjusted R-square: 0.9815
b5 = 90, 249 (−466, 200; 646, 700) RMSE: 942.5
b6 = −83, 320 (−806, 700; 640, 100)
b6 = 43, 670 (−295, 700; 383, 100)
55
References
Artzner, P., F. Delbaen, J. Eber, and D. Heath (1999): “Coherent Measures of
Risk,” Mathematical Finance, 9, 203–228.
Barberis, N., and M. Huang (2004a): “The Loss Aversion/Narrow Framing Approach
to Stock Market Pricing and Participation Puzzles,” Working paper, University of
Chicago, Stanford University.
(2004b): “Preferences with Frames: A New Utility Specification that Allows for
the Framing of Risks,” Working paper, University of Chicago, Stanford University.
Barberis, N., M. Huang, and T. Santos (2001): “Prospect Theory and Asset
Prices,” Quarterly Journal of Economics, 116, 1–53.
Barberis, N., M. Huang, and R. Thaler (2003): “Individual Preferences, Monetary
Gambles and the Equity Premium,” Working paper, University of Chicago, Stanford
University.
Bechara, A., and A. Damasio (2005): “The somatic marker hypothesis: A neural
theory of economic decision,” Games and Economic Behavior, 52, 336–372.
Benartzi, S., and R. H. Thaler (1995): “Myopic Loss Aversion and the Equity
Premium Puzzle,” Quarterly Journal of Economics, 110, 73–92.
Campbell, R., R. Huisman, and K. Koedijk (2001): “Optimal Portfolio Selection
in a Value-at-Risk Framework,” Journal of Banking and Finance, 25, 1789–1804.
Damasio, A. (1994): Descartes error: Emotion, reason and the human brain. G.P. Put-
nams Sons, New York.
Embrechts, P., A. McNeil, and D. Straumann (1999): “Correlation and
Dependance in Risk Management: Properties and Pitfalls,” Available from
¡www.defaultrisk.com¿.
Gneezy, U., A. Kapteyn, and J. Potters (2003): “Evaluation Periods and Asset
Price in a Market Experiment,” Journal of Finance, 58, 821–837.
Gneezy, U., and J. Potters (1997): “An Experiment on Risk Taking and Evaluation
Periods,” Quarterly Journal of Economics, 102, 631–645.
Kahneman, D. (2003): “Maps fo Bounded Rationality: Psychology of Behavioral Eco-
nomics,” The American Economic Review, pp. 1449–1475.
Kahneman, D., and A. Tversky (1992): “Advances in prospect theory: cumulative
representation of uncertainty,” in Choices, values and frames, ed. by D. Kahneman,
and A. Tversky, pp. 44–66. Cambridge University Press, Cambridge.
56
Rockafellar, R., and S. Uryasev (2000): “Optimization of Conditional Value-At-
Risk,” The Journal of Risk, 2, 21–41.
Szego, G. (2002): “Measures of Risk,” Journal of Banking and Finance, 26, 1253–1272.
Thaler, R. (1980): “Towards a positive theory of consumer choice,” Journal of Eco-
nomic Behavior and Organization, 1, 39–60.
(1992): “Savings, Fungibility, and Mental Accounts,” in Choice over time, ed.
by G. Lowenstein, and J. Elster, pp. 177–209. Russell Sage Foundation, New York.
Thaler, R., and E. Johnson (1990): “Gambling with the house money and trying to
break even: The effects of prior outcomes on risky choice,” Management Science,
36, 643–660.
Thaler, R. H., A. Tversky, D. Kahneman, and A. Schwartz (1997): “The Effect
of Myopia and Loss Aversion on Risk Taking,” Quarterly Journal of Economics, 102,
647–661.
Tversky, A., and D. Kahneman (1979): “Prospect theory: An analysis of decision
under risk,” Econometrica, 47, 263–291.
57