9、Cumulative Prospect Theory, Option Returns, And the Variance Premium（2019RFS）

Cumulative Prospect Theory, Option
Returns, and the Variance Premium

Lieven Baele
Tilburg University
Joost Driessen
Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

Tilburg University
Sebastian Ebert
Frankfurt School of Finance and Management and Tilburg University
Juan M. Londono
International Finance at the Federal Reserve Board of Governors
Oliver G. Spalt
Tilburg University
We develop a tractable equilibrium asset pricing model with cumulative prospect theory
(CPT) preferences. Using GMM on a sample of U.S. equity index option returns, we show
that by introducing a single common probability weighting parameter for both tails of the
return distribution, the CPT model can simultaneously generate the otherwise puzzlingly
low returns on both out-of-the-money put and out-of-the-money call options as well as the
high observed variance premium. In a dynamic setting, probability weighting and time-
varying equity return volatility combine to match the observed time-series pattern of the
variance premium. (JEL C15, G11, G13)
Received May 30, 2017; editorial decision August 10, 2018 by Editor Andrew Karolyi.
A central empirical success of prospect theory (Kahneman and Tversky 1979;

Tversky and Kahneman 1992) is its ability to explain major puzzles in financial
economics. For example, Benartzi and Thaler (1995) show that prospect theory
can help us understand the equity premium puzzle, and Barberis, Huang,
We thank editors Robin Greenwood and Andrew Karolyi; two anonymous referees; Nicholas Barberis, Ing-Haw
Cheng (discussant), Shimon Kogan (discussant), Lei Sun (discussant), and David Veredas (discussant); and
seminar participants at Tilburg University, the 2012 EFA meeting in Copenhagen, the 2015 FIRS conference in
Reykjavik, the 2016 Belgian Financial Research Forum in Brussels, the 2016 Research in Behavioral Finance
Conference in Amsterdam, and the 2017 AEA meeting in Chicago for useful comments and suggestions.
Send correspondence to Joost Driessen, Tilburg University, PO Box 90153, 5000 LE, Tilburg, the Netherlands;
telephone: +31-13-4662324. E-mail: j.j.a.g.driessen@uvt.nl.
© The Author(s) 2018. Published by Oxford University Press on behalf of The Society for Financial Studies.
All rights reserved. For permissions, please e-mail: journals.permissions@oup.com.
doi:10.1093/rfs/hhy127 Advance Access publication December 7, 2018
[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3667 3667–3723

The Review of Financial Studies / v 32 n 9 2019
and Santos (2001) demonstrate how a dynamic prospect theory framework

can simultaneously generate a high equity premium, predictability of equity
returns, and excess volatility. In these and other settings, prospect theory offers
a rigorous and testable alternative framework to more traditional asset pricing
theories (see, e.g., Barberis 2013 for a survey).
In this paper, we show that prospect theory can help us understand three
important puzzles in the recent asset pricing literature simultaneously: the low
returns on out-of-the-money (OTM) equity index put options, the low returns
on OTM equity index call options, and the large variance premium. OTM put

returns, which are the first puzzle we address, are too low to be compatible
with standard asset pricing models. This well-known fact is the subject of the
sizable literature we discuss below. OTM calls, which are the second puzzle we
address, also have low returns. This call option puzzle has received less attention
in the prior literature, and existing models proposed to fitting put returns are
usually not equipped to also fit call returns. The third puzzle we address is
the variance premium, defined as the difference between the option-implied
and the expected realized variance of stock returns. In the data, the variance
premium is time varying and strongly positive. This is a major puzzle because
the standard consumption-based model with constant relative-risk aversion
(CRRA) preferences cannot generate a nonzero variance premium, regardless
of the risk-aversion level and even when consumption variance varies over time
(see, e.g., Drechsler and Yaron 2011). Quantitatively, the variance premium is
a first-order phenomenon for investors.1
In this paper we show that cumulative prospect theory (CPT) can jointly
explain the three puzzles above, by introducing a single common probability
weighting parameter for both tails of the equity return distribution. A key
insight we use is that the variance premium can be written as the expected
return on a portfolio of OTM calls and OTM puts with different strike prices.
As a consequence, fitting the returns of the underlying options accurately is a
sufficient condition for explaining the variance premium. Our central result is
that a calibrated equilibrium model with CPT investors can generate option
returns close to those observed in the data. Therefore, we show that CPT
can explain the variance premium by solving the more general problem of
accurately fitting OTM call and put option returns.
The central intuition, formalized in a theoretical CPT model below, is as
follows. Start with the standard pricing equation, like in Cochrane (2005),
applied to investors with CRRA preferences:

E [Ri ]−R f = −R f cov mCRRA ,Ri , (1)
1 For example, Coval and Shumway (2001), Driessen and Maenhout (2007), and Eraker (2013) find that the Sharpe
ratio for volatility-selling strategies, such as shorting straddles, which effectively bet on the variance premium, is
at least twice the Sharpe ratio of the underlying equity index (see also Bakshi and Kapadia 2003; Jiang and Tian
2005; Bakshi and Madan 2006; Carr and Wu 2009; Bollerslev, Tauchen, and Zhou 2009; Bollerslev, Gibson, and
Zhou 2011, among others). If the equity premium is a puzzle, then, by this metric, the variance premium is an
even bigger puzzle.
3668

Cumulative Prospect Theory, Option Returns, and the Variance Premium
which says that the expected return in excess of the risk-free rate of an asset is
a function of the covariance between the asset’s return and the CRRA pricing
kernel mCRRA . The CRRA pricing kernel mCRRA increases in the marginal utility
of the investor. Equation (1) thus implies that assets that pay off in states where
the investor’s marginal utility is low have high expected returns, because the
covariance term is negative.
Even though the CRRA pricing equation is widely used, it does not explain
observed option returns. Figure 1 illustrates the disconnect between the CRRA
model and the data by plotting the observed average option returns for different

strike prices (black squares) against option returns implied by a standard CRRA-
model. Panel A shows that the CRRA model is a complete failure when pricing
calls. First, because call options pay off when investors’ marginal utility is
low, which makes the covariance term negative, Equation (1) predicts that calls
should have positive expected excess returns. In the data, however, calls have
negative excess returns across most strike prices. Second, panel A of Figure
1 shows that, in the data, returns from investing in call options decrease for
higher strike prices, whereas Equation (1) implies that OTM call returns should
increase for higher strike prices, because the covariance between payoffs and
marginal utility becomes more negative for higher strike prices. The CRRA
model thus fails on two fronts: it predicts the wrong sign on call option returns
and, on top, it predicts the wrong sign on the relation between call option returns
and moneyness.
Panel B of Figure 1 presents analogous results for put returns. Qualitatively,
the standard CRRA model does better on put returns than it does on calls. In
particular, the model is consistent with the observed negative returns on puts,
and it correctly predicts that OTM puts have the lowest returns. Quantitatively,
however, the fit of the model is poor. For reasonable parametrizations, standard
CRRA models predict dramatically higher returns for puts than those observed
in the data. In Figure 1, panel B, this is reflected by the vertical difference
between the data (black squares) and CRRA-implied values (blue x’s).
We show formally below that when investors have CPT preferences in our
model, the equilibrium pricing equation becomes

E [Ri ]−R f = −R f cov mCRRA ,Ri −R f cov mCPT ,Ri , (2)
which says that the expected return in excess of the risk-free rate of an asset is
equal to the return implied by the CRRA model, minus an additional covariance
term. This additional covariance term depends on mCPT , which we label the
“CPT pricing kernel.” Equation (2) highlights an advantage of our CPT model
in terms of theory building: it nests the standard CRRA model as a special case,
and thus allows us to directly compare the standard model with our proposed
extension.
Figure 1 shows the fit of our CPT model. The CPT model-implied option
returns (red crosses) line up remarkably well with the actual data for both calls
3669

A Average call returns
-20
%
-40

Observed
CPT-implied
-60 CRRA-implied
0.92 0.94 0.96 0.98 1 1.02 1.04 1.06 1.08

Moneyness
B Average put returns
-20
%
-40
-60
0.92 0.94 0.96 0.98 1 1.02 1.04 1.06 1.08

Moneyness
Figure 1
Average option returns and model fit
This figure compares the average observed 30-day return computed from all options in our sample with the
returns predicted by our benchmark CPT representative agent model (specification (2) in Table 2) as well as by a
CRRA model—a restricted CPT model with c = 1 and b0 = 0 (specification (1) in Table 2). The sample comprises
daily S&P 500 index option returns from 1996 to March 2016 for various strike prices. Of the 26 options in the
sample, 13 are calls, with deltas ranging from 0.2 to 0.8 (from OTM to ITM). The other 13 options are puts, with
deltas ranging from −0.8 to −0.2 (from ITM to OTM). The option returns are plotted by moneyness (K/S0 ) for
a given delta, separately for calls (panel A) and puts (panel B). We use a discrete approximation of the lognormal
distribution for the S&P 500 index, taking a grid of 200 potential outcomes for the 1-month-ahead equity return
distribution, ranging from −50% to 150%.
(panel A) and puts (panel B), which sharply contrasts with the poor fit of the
CRRA model. Because fitting returns of puts and calls is a sufficient condition
for explaining the variance premium, Figure 1 also implies that the CPT model
does a good job fitting the variance premium.
3670

To understand why the CPT model is successful in fitting options returns,

one should look more closely at the CPT pricing kernel mCPT . Panel A of
Figure 2 plots it as a function of equity index returns. The crucial feature of
the CPT pricing kernel is that it has a U-shape, which implies that the price
of an asset is high, and its expected return low, if it pays off in extreme states
of the world, regardless of whether those states are good or bad. The CPT
model therefore predicts simultaneously low returns for both OTM call options
and OTM put options. For comparison, panel A of Figure 2 also plots mCRRA ,
which is monotonically decreasing, and can therefore not explain low OTM

call returns. It is also too flat to fit OTM put returns.
The key driver of mCPT is the probability weighting feature of CPT.
Probability weighting is a nonlinear transformation of objective probabilities
whose main implication for our setting is that the tails of a distribution
are overweighted when evaluating its attractiveness. Intuitively, probability
weighting is a modeling device that captures demand for lottery tickets and
insurance through the same underlying mechanism. An extensive literature
in decision science documents that probability weighting is a pervasive trait
of human decision making (see, e.g., Epper and Fehr-Duda 2012). Barberis
(2013, p. 191) surveys a growing literature in finance and economics on
probability weighting, and writes that “in risk-related fields of economics,
such as finance, insurance, and gambling, more empirical support exists
for probability weighting than for loss aversion, an arguably better-known
component of prospect theory.”
Probability weighting is relevant in the context of option pricing because
OTM calls and OTM puts, like lottery tickets and insurance, pay off in unlikely
states of the world, which makes them attractive to CPT investors. Probability
weighting thus explains the U-shape of the CPT pricing kernel in Figure 2,
panel A. One-to-one mapping occurs between the CPT pricing kernel and
state-dependent relative-risk aversion, which we plot in Panel B of Figure 2.
The figure shows that, in contrast to a standard CRRA model, a CPT model
implies that relative-risk aversion varies across states. CPT investors in our
model put a high decision weight on bad outcomes, which increases the implied
risk aversion and thus leads to expensive put options. At the same time, CPT
investors overweigh good outcomes, which decreases effective risk aversion
and thus leads to expensive call options. Both features combine to generate the
CPT model’s remarkably good fit documented in Figure 1.
Our paper proceeds as follows. We first introduce an equilibrium asset
pricing model with CPT preferences in Section 1. The model builds on the
work of Barberis, Huang, and Santos (2001), who consider a representative
agent model in which the agent’s preferences are the sum of a CRRA utility
function and a gain-loss prospect theory utility function. Because our focus is
on pricing options, we do not explicitly model consumption and dividends but
directly depart from a given distribution of stock market returns. In contrast to
3671

A Pricing Kernel
2.5
CPT
2 CRRA
1.5
m

0.5
0
0.92 0.94 0.96 0.98 1 1.02 1.04 1.06 1.08
Market return (R E )
B Implied Risk Aversion

20
CPT
CRRA
10
0
*
-10
-20
0.92 0.94 0.96 0.98 1 1.02 1.04 1.06 1.08
Market return (R E )
Figure 2
The pricing kernel and implied risk aversion
Panel A shows the CRRA and CPT components of the pricing kernel implied by our CPT setting (see Equation
(10)) as a function of the market return, R E . Panel B shows the state-dependent risk aversion, (RiE ), implied by
∂ln(m(RiE )) E
the pricing kernel, m = mCPT +mCRRA . The implied risk aversion is calculated as = − Ri , which is
∂RiE
consistent with the definition of local risk aversion in Ait-Sahalia and Lo (2000). In both panels, the reference
level, WRef , is calculated using a risk-free rate of 0.19%; the level of risk aversion, γ , is fixed at 1; the parameter
driving probability weighting, c1 = c2 = c, is fixed at 0.65; and the scale, b0 , is fixed at 0.95, which is the scale
that yields a 50% contribution of CPT to total utility for this set of parameters. Note that these parameter values
match those in the benchmark GMM estimation (see Table 2).
Barberis, Huang, and Santos (2001), we take into account the probability
weighting feature of prospect theory and analyze its ability to fit option returns
and the variance premium puzzle. Our model is not designed to address
3672

the equity premium puzzle.2 We show that the resultant model yields the
equilibrium pricing Equation (2) above, which allows us to derive the expected
return on the stock market index and on a range of options on that index. We
can then derive the model-implied variance premium from these equilibrium
returns.
A key advantage of the model is that it is parsimonious and very tractable. It
allows us to prove that the model-implied portfolio weights are unique, finite,
and wealth independent. This is relevant because exploding portfolio weights
are a well-known challenge for existing CPT portfolio choice models (see,

e.g., Ang, Bekaert, and Liu 2005; Barberis and Huang 2008; Bernard and
Ghossoub 2010). In a multiperiod extension of the model, we prove that if,
and only if, the CRRA component of preferences is given by log-utility, the
CPT investor behaves like a one-period myopic investor, even when returns
are not independently and identically distributed (iid). This result is interesting
because it provides a set of conditions under which insights obtained for a one-
period setting extend to a multiperiod setting. Under these conditions, we can
study a dynamic setting with non-iid returns using a conditional one-period
equilibrium model.
In Section 2, we bring the model to the data. Based on S&P 500 equity returns
and S&P option prices from 1996 to 2016, we construct monthly call and put
option returns for 13 different strike levels. Using the generalized method of
moments (GMM), we show that the model yields a very good fit for the cross-
section of expected option returns, and, therefore, for the variance premium.
All our tests show that probability weighting is the central determinant of the
model’s fit, and reject the null hypothesis of no probability weighting. In our
benchmark setup, the estimate of the Tversky-Kahneman probability weighting
parameter is 0.65, and thus very similar to values found for subjects in lab
experiments, and to values that are commonly used in calibrations of prospect
theory models in the finance and economics literatures.
When we allow probability weighting to be different for gains and losses, we
find that distorting probabilities in both tails is important. Distorting only gains
yields an acceptable fit for calls, but the model then fails to explain puts well.
Distorting only losses yields the opposite result. As we explain in greater detail
below, the ability to distort also the right tail of the distribution distinguishes
our CPT model from well-known alternative approaches to explain the variance
premium in the literature that focus on the left tail (e.g., rare disasters, long-run
risks). Our results suggest that models which distort either marginal utility,
probabilities, or both—but only on the downside—are likely to fail to price
call options well, even though the model may fit the variance premium. Our
2 We leave fully integrating the intertemporal consumption-based model of Barberis, Huang, and Santos (2001)
with our model that features probability weighting to future research. Because dynamic probability weighting
models imply, in general, that investors behave in a time-inconsistent fashion, such models pose significant
modeling challenges that are beyond the scope of this paper.
3673

approach of allowing for probability distortions for both gains and losses
is consistent with findings in the psychology and experimental economics
literatures which suggest probability weighting for both gains and losses is
exhibited by a majority of decision makers (e.g., Tversky and Kahneman 1992;
Abdellaoui 2000; Bruhin, Fehr-Duda, and Epper 2010).
In Section 3, we show that the model can also generate meaningful time
variation in the variance premium. We focus on three potential drivers: time-
varying equity return volatility, time-varying probability weighting, and time-
varying loss aversion. The main finding is that, with time-varying volatility,

the CPT model captures the dynamics of the variance premium well, even
when probability weighting is kept fixed at the benchmark value. Hence,
we do not need time variation in investor preferences to explain the time
variation in the variance premium. Moreover, allowing for such time variation
in preferences hardly improves the fit. The intuition for the effect of volatility is
straightforward: the power of probability weighting comes from overweighting
the tails of the wealth distribution. Higher return volatility translates into longer
tails, which, in turn, are then overweighted more. As a result, the variance
premium is positively correlated with the volatility level, in line with empirical
observations.
Our paper contributes to the literature that studies the usefulness of prospect
theory in finance and to a growing number of recent papers exploring the
usefulness of probability weighting in asset pricing contexts (see Barberis 2013
for a recent survey). In particular, probability weighting has been used to explain
a number of otherwise puzzling facts in the data. For example, Polkovnichenko
(2005) argues that CPT can explain household underdiversification. Driessen
and Maenhout (2007) empirically examine portfolio choice with CPT investors
and show that probability weighting can rationalize why investors hold puts
and straddles in an optimal portfolio. Barberis and Huang (2008) provide a
theoretical framework to explain why idiosyncratic skewness of an asset can
be priced in equilibrium. De Giorgi and Legg (2012) show that probability
weighting can provide a solution to the stock market nonparticipation puzzle
and the equity premium puzzle. Barberis, Mukherjee, and Wang (2016) show
that prospect theory, and in particular the probability weighting feature, is useful
for understanding the cross-section of equity returns in the United States and
also in a majority of 46 other national stock markets. A central implication of
CPT’s probability weighting feature is a preference for skewness and lower
returns of skewed assets, a prediction which has been found across many asset
classes, including stocks and options (e.g., Kumar 2009; Boyer, Mitton, and
Vorkink 2010; Conrad, Dittmar, and Ghysels 2012; Boyer and Vorkink 2014;
Eraker and Ready 2015). Applications of the idea that overweighting the tails
of a distribution, like in prospect theory, affects prices and returns include
Mitton and Vorkink (2007), who explain the diversification discount, Green
and Hwang (2012), who explain the returns to initial public offerings (IPOs),
and Schneider and Spalt (2016), who explain distortions in capital budgeting.
3674

Ilmanen (2012) provides a survey of financial market facts consistent with a

preference for skewness as induced by probability weighting. Related to our
paper, Polkovnichenko and Zhao (2013) and Kliger and Levy (2009) use index
option prices to provide evidence consistent with inverse S-shaped probability
weighting functions. Outside finance, probability weighting has been used to
explain the attractiveness of casino gambling (Barberis 2011), and the favorite
long-shot bias on race tracks (Snowberg and Wolfers 2010). To the best of our
knowledge, our paper is the first to develop an equilibrium prospect theory
model to jointly explain the variance premium as well as the returns of OTM

put and OTM call options.
We also contribute to the recent literature that studies potential explanations
for the variance premium puzzle. Existing attempts to explain the variance
premium are numerous and include rare disaster models (e.g., Gabaix 2012),
models based on long-run risks (e.g., Bollerslev, Tauchen, and Zhou 2009;
Drechsler and Yaron 2011; Londono 2014; Shaliastovich 2015), models based
on habit formation (in combination with rare jumps, like in Du 2011, or
with bad environment-good environment (BEGE) dynamics, like in Bekaert
and Engstrom 2016), as well as other approaches (e.g., Drechsler 2013;
Schreindorfer 2014; Jin 2015). Several notable differences exist between these
approaches and our paper. First, our model with CPT preferences generates
a substantial variance premium even when asset returns are iid lognormal.
As such, rare disaster models, such as the one in Gabaix (2012), present the
polar opposite to our approach. These models maintain the CRRA nature
of preferences, but change the return generating process. Second, our paper
analyzes both the variance premium as well as option returns, while most
existing papers analyze the variance premium only. This is relevant because,
as we show below, it is possible to fit the variance premium perfectly while
fitting OTM option returns poorly. Third, existing work, either via the assumed
preferences, or via the assumed data generating process, focuses predominantly
on the left tail of outcome distributions. By contrast, our CPT model, via
the probability weighting feature, speaks to both tails of the distribution, and
can thus jointly explain returns for put and call options. Fourth, an important
difference to our model is that the above models are necessarily dynamic,
whereas the CPT model can generate a variance premium even in a one-
shot game with known probabilities. Therefore, it is possible that some of
the variance premium is driven by a fundamentally different mechanism than
suggested by the existing literature.
Our paper also relates to empirical studies on option returns. Many studies
have documented low returns on OTM put options (see, among many others,
Bondarenko 2003, 2014; Jones 2006; Santa-Clara and Saretto 2009). Fewer
studies document low OTM call option returns, including Bakshi, Madan,
and Panayotov (2010), Constantinides, Jackwerth, and Savov (2013), and
Chaudhuri and Schroder (2015).
3675

Finally, our paper is related to studies on the shape of the pricing

kernel. Seminal studies by Ait-Sahalia and Lo (2000), Jackwerth (2000),
and Rosenberg and Engle (2002) show that pricing kernels estimated from
index option data are not monotonically decreasing, as predicted by standard
economic theory, but instead often follow a U-shaped pattern (see, e.g.,
Cuesdeanu and Jackwerth 2018 for a recent survey of that literature). Several
potential explanations have been offered for this “pricing kernel puzzle,”
including heterogeneous beliefs (e.g., Ziegler 2007; Bakshi, Madan, and
Panayotov 2010), missing state variables and state-dependent utility (e.g.,

Chabi-Yo, Garcia, and Renault 2008; Chabi-Yo 2012; Christoffersen, Heston,
and Jacobs 2013; Song and Xiu 2016), or ambiguity aversion (e.g., Gollier
2011). We contribute to this literature by showing that our CPT model can
provide a microfoundation for U-shaped pricing kernels.
1. The Model
1.1 Preferences
Following Barberis, Huang, and Santos (2001) (BHS), the representative
agent’s total utility is given by a weighted sum of expected utility over terminal
wealth WT and CPT utility over the gain or loss XT (Tversky and Kahneman
1992). The final gain or loss is defined as XT = WT −WRef , where WRef denotes
the reference point, which determines whether terminal wealth levels are
perceived as gains (WT ≥ WRef ) or losses (WT < WRef ). Formally, the investor
seeks to maximize
EU [WT ]+b0 CP T [XT ], (3)
where b0 ≥ 0 is a scaling term that governs the relative importance of the
expected utility part, EU [WT ], and the CPT part, CP T [XT ]. Taking a
combination of “traditional” expected utility preferences over terminal wealth
and a “behavioral” prospect theory term over gains and losses—a hybrid model
of traditional and behavioral preferences—is crucial to the results in this paper.
We provide details about why we believe these preferences are economically
and psychologically more attractive than a pure CPT model at the end of this
subsection.
For the expected utility part, we assume that EU [WT ] = E[U (WT )] satisfies
CRRA, that is, the utility function U is given by
1−γ
WT
γ = 1
U (WT ) = 1−γ for , (4)
lnWT γ =1
where γ is the risk aversion coefficient. To define the CPT part, first, let v denote
the value function, which is defined over gains and losses XT . We assume that
v takes the piecewise linear form

XT XT ≥ 0
v(XT ) = for , (5)
λXT XT < 0
3676

where λ ≥ 1 is called the loss aversion parameter. For example, if λ = 2, then

a $1 loss resonates twice as much as a $1 gain. Second, probabilities may be
processed nonlinearly. We define probability weighting functions w + and w− ,
for gains and losses, respectively, by
p c1 p c2
w− (p) = +
1/c1 , w (p) = 1/c2 , (6)
p c1 +(1−p)c1 p c2 +(1−p)c2
where c1 , c2 ∈ [0.28;1] in Equation (6) control the curvature of each weighting

function.3
The weighting functions are inverse S shaped, which means that probabilities
close to zero are overweighted (w(p) > p), while probabilities close to 1
are underweighted (w(p) < p). The CPT of Tversky and Kahneman (1992),
however, distorts cumulative and decumulative probabilities rather than
marginal probabilities p to obtain decision weights for each state. Intuitively,
using decision weights instead of distorted marginal probabilities leads to an
overweighting of unlikely and extreme states, that is, the tails of the distribution.
Formally, assume that there are N states of the world at time T , each occurring
with objective probability p i , and that each state is associated with a specific
final wealth level WT ,i . All states are then ordered (“ranked”, see Quiggin
1982) from worst to best, and related to the investor’s reference point WRef :
W1 ≤ ··· ≤ Wk−1 ≤ WRef ≤ Wk ≤ ... ≤ WN . Then the decision weight πi of state
i is given by
πi = w − (p1 +···+pi )−w − (p1 +···+pi−1 ) for 2 ≤ i < k

πi = w + (pi +···+pN )−w + (pi+1 +···+pN ) for k ≤ i ≤ N −1,
where π1 = w − (p1 ) and πN = w + (pN ). The CPT value of the gain or loss XT is
then computed as

N
CP T (XT ) = πi v(XT ,i ). (7)
i=1
When c1 = c2 = 1, the weighting functions in (6) reduce to w + (p) = w − (p) = p,

so that overall utility is the sum of terminal wealth and gain-loss utility, as is
the case in BHS. For b0 = 0, our preference specification nests pure expected
utility, which is a particularly attractive feature of our setup that allows us to
directly gauge the relative impact of the CPT part of the total utility function.
Following BHS, we make the assumption that gain-loss utility scales with
−γ
wealth according to b0 = b̂W0 for some b̂ ≥ 0. Moreover, following the prior
literature (e.g., Barberis and Huang 2008; Barberis and Xiong 2009, 2012), we
3 The lower bound is a technical condition that ensures that decision weights are positive (Rieger and Wang (2006);
Ingersoll (2008)).
3677

assume that WRef = W0 R f ; that is, the investor experiences positive gain-loss
utility if and only if her investment yields more than the risk-free return.
Our baseline preference specification in Equation (3) is a combination of
rational (consumption or terminal wealth) and behavioral (reference-dependent
or gain-loss) elements. The use of such a hybrid model is motivated by BHS,
as well as by Kőszegi and Rabin (2006, 2007), who argue strongly in favor of
a hybrid setup. Two reasons make such a setup attractive. First, in contrast to
a pure CPT model, a hybrid model can capture the fact that investors may also
care about absolute wealth levels. As Kőszegi and Rabin (2006, p. 1138) put it,

“while preferences are reference dependent, gains and losses are clearly not all
that people care about. The sensation of gain or avoided loss from having more
money does significantly affect our utility—but so does the absolute pleasure
of consumption we purchase with the money. Therefore, in contrast to prior
formulations based on a “value function” defined solely over gains and losses,
our approach [of using a hybrid model] makes explicit the way preferences also
depend on absolute levels.”
Second, the hybrid model can be given a psychological foundation in the
dual-system framework of decision making (e.g., Kahneman 2011). In that
framework, System 1 represents intuitive decision making which is prone to
behavioral biases. System 2 represents “rational,” effortful, thinking. The key
idea is that for any given decision System 1 proposes a, potentially suboptimal,
path of action. System 2 may subsequently overrule System 1 and lead to
decisions more in line with normative prescriptions—in our case, the normative
prescriptions of expected utility theory (EUT). The hybrid model can be
interpreted as a stylized way to capture the relative importance of Systems 1
and 2 for investors’ decisions. If System 2 is perfectly successful in correcting
System 1, then we are left with only the expected utility part. The prospect
theory value function can be viewed as a deviation from this rational benchmark
introduced by System 1. An attractive feature of our model is that we can bring
it to the data in order to quantify the relative impact of the expected utility
benchmark and the CPT component.
The hybrid preference model also has specific advantages for modeling
investment behavior. Bernard and Ghossoub (2010) and Barberis and Huang
(2008) find that if the CPT value of the market return is positive, a representative
investor with pure CPT preferences would like to invest an infinite amount.4
Similarly, Ebert and Strack (2015) show that a CPT agent would invest
arbitrarily large amounts into gambles with negative expectation. Our results
below show that the expected utility theory part of our hybrid model eliminates
the implausible infinite leverage result, while retaining the economically
interesting predictions of CPT.
4 Barberis and Huang (2008) show that an equilibrium can exist in a CPT setting with multiple investors. Allowing
for multiple investors, instead of a representative investor, would substantially complicate our analysis, which is
why we leave it for future research.
3678

Finally, note that probability distortion is effective only in the CPT part of
preferences, but not in the EUT part. This modeling choice allows us again to
interpret the CPT part as a deviation from the rational EUT benchmark, and
it allows us to quantify the economic importance of that deviation from EUT
rationality for explaining option returns and the variance premium. Applying
probability weighting only to the CPT part of the utility function is in line with
earlier work of De Giorgi and Legg (2012). While we believe setting up our
model in this way is the most natural baseline in light of the existing literature,
alternative ways to setting up the model are conceivable, and we will explore

various alternatives in our analysis below.
1.2 The investor’s problem

We first consider a simple market with (1) a risk-free asset with a constant
and exogenously given gross return R f and (2) an equity index that pays, at
time T , xiE in state i, which occurs with probability pi . The representative
investor’s problem is to determine the share of wealth to be invested in the
risky asset, α E , such that total utility in Equation (3) is maximized. Given the
investor’s
f initial wealth
W0 , her terminal wealth in state i is given by WT ,i =
R +α E (RiE −R f ) W0 and, given our assumptions in the previous section, the
portfolio gain or loss experienced is XT ,i = α E (RiE −R f )W0 , where RiE = xiE /s0E
with s0E the time-0 equity price. The CPT value of the portfolio can be written
as
CPT (XT ) = α E W0 CPT (R E −R f ). (8)
Taking the derivative of equation (8) shows that, in a pure representative
agent CPT model, without an expected utility part like in Equation (3), the
optimal portfolio weight in equity will be (minus) infinity if a total investment
of one dollar into the equity index yields positive (negative) CPT value.
We now show the hybrid model in Equation (3) does not make the unrealistic
prediction of infinite equity investment. To see this, take the derivative of
Equation (3), which yields the first-order condition (FOC) of the investor’s
problem:
−γ
0= (RiE −R f ) pi R f +α E RiE −R f + b̂πi (1+(λ−1)1RE <Rf ) , (9)
i
i
where the indicator function 1RE <Rf is equal to one if the investment in state
i yields a loss and equal to zero otherwise. The FOC shows that the hybrid
model inherits two attractive properties from the standard CRRA model. First,
the second-order condition (the derivative of the right-hand side of the FOC with
respect to the portfolio weight, set equal to zero) is identical to the one obtained
in a pure CRRA model. Therefore, the second-order condition is satisfied, which
implies a unique and finite optimal portfolio weight. Second, like in the model
of BHS, our choice of b0 ensures that the optimal portfolio weight is wealth
independent. We summarize our results in the following proposition.
3679

Proposition 1. The investor’s optimal portfolio weight is unique, finite, and

wealth independent.
Our baseline model above is static. On the one hand, this simplicity is a feature
of our modeling strategy, because it emphasizes that we do not need to resort
to a dynamic argument for generating a variance premium. Instead, our model
shows that probability weighting can generate a variance premium even in a
static setting, which is different from most explanations in the existing literature.

On the other hand, introducing dynamics would be attractive, because we could
align our model closer with standard intertemporal asset pricing models, and
closer with the dynamic loss aversion arguments in BHS. Unfortunately, it is
well-known that introducing probability weighting into a dynamic model is
difficult (e.g., Barberis 2011; Ingersoll and Jin 2013). A central problem is that
using probability weighting in a dynamic model introduces the problem of time
inconsistency. As explained by Machina (1989), time inconsistency essentially
arises whenever preferences do not satisfy the independence axiom. Because the
independence axiom does not hold for cumulative prospect theory preferences,
an intertemporal model with CPT preferences will in general feature a time-
inconsistency problem. Fully incorporating dynamics is therefore beyond the
scope of our current paper and a topic we leave for future research. However, we
can show that the static setting is not as restrictive as it may appear. Specifically,
the next proposition (proven in Appendix A) presents one set of conditions
under which the optimal portfolio chosen by an investor in the static model is
identical to the optimal portfolio chosen in the same period by an investor in a
multiperiod version of the model.
Proposition 2. Consider a multiperiod version of the model above where

the investor has CRRA terminal wealth utility, can trade each period, and, in
addition, experiences prospect theory utility over portfolio gains and losses each
period. The investor’s optimal portfolio weights are, in each period, independent
of her investment horizon (i.e., myopic) if, and only if, the investor has log utility
in the CRRA part of the utility function.
We think of the preferences underlying Proposition 2 as follows. While the

investor consumes only in the terminal period, she pays attention to the returns
generated by her investment each period. When the investment gains, she feels
good, and when the investment loses, she feels bad. Hence, we think of the
period-by-period gain-loss utility as capturing the psychological pleasure or
pain of seeing the portfolio performing better or worse than expected, which is
directly motivated by the treatment of gain-loss utility in BHS.5
5 Note that log utility of terminal wealth can be rewritten as the sum of log utility of the per-period wealth returns.
The investor thus evaluates per-period returns in both components of the total utility function.
3680

Proposition 2 is special, because it represents one case in which a dynamic

probability weighting model yields myopic, and therefore time-consistent,
choice. Intuitively, the reason is that a one-period investor can, by definition,
not be time inconsistent. Then, if a multiperiod investor is myopic in the sense
that she behaves like a one-period investor, her portfolio choice is necessarily
time consistent. In general, however, dynamic probability weighting models
will yield time-inconsistent investor behavior, and results like Proposition 2
will not obtain.
Note that the proposition does not assume that returns are independently

and/or identically (iid) distributed. If returns are iid, then we can show that
the investor chooses the same weight each period. This result is related to
the seminal results of Merton (1969) and Samuelson (1969), who find that
portfolio choice is myopic for investors if they either have log utility, or if
they have CRRA utility when returns are iid.6 Proposition 2 shows that a
dynamic extension of our CPT model yields myopia for log utility, but not
for the combination of CRRA utility and iid returns.
We emphasize that log utility is a channel that connects a one-period setting
with its multiperiod counterpart for the specific multiperiod model underlying
Proposition 2. That this also would be the case for alternative multiperiod
models, such as ones that feature intertemporal consumption, is not guaranteed.
In the next subsection, we derive the pricing kernel for our static baseline
model. It follows from the above discussion that this pricing kernel is identical
to the kernel in the multiperiod version of our model under the assumptions of
Proposition 2.
1.3 Equilibrium pricing kernel and equity risk premium

In equilibrium, the market must clear, and the representative investor must
hold the entire supply of equity, which implies α E = 1. The FOC in Equation
(9) becomes

E −γ πi
0= pi (Ri −R ) Ri
E f
+ b̂ 1+(λ−1)1RE <Rf
i
pi i

⇐⇒ 0 = E (R E −R f ) mCRRA +mCPT , (10)
where mCRRA and mCPT refer to the random variables that, in state i, yield
−γ
outcome RiE and b̂ pπii 1+(λ−1)1RE <Rf , respectively. Given a risk-free
i
rate, probabilities pi , and preference parameters of the investor, we can use
Equation (10) to determine the equilibrium expected return on the equity index,
E[R E ].
6 Our results are related to, but different from those of Merton (1969) and Samuelson (1969). One way in which
their results are different is that they model consumption, whereas we do not.
3681

We now introduce derivatives on the equity index. Each derivative d on

the equity index pays xid in state i of the world at time T . Derivatives are
in zero net supply. Hence, in equilibrium, the representative investor will not
hold derivatives in her portfolio. We assume that the investor ranks states i
according to equity returns.7 Then the random variable m := mCRRA +mCPT in
Equation (10) is a pricing kernel that can be used to price any derivative on the
equity index, such as a put or call option. Hence, applying (10) to the return on
a derivative R d and solving for the expected excess return, we obtain:

Proposition 3. The equilibrium expected return of a derivative d in the above
model is given by

d f − cov(mCRRA ,R d )+cov(mCPT ,R d )
E[R ]−R = . (11)
E[mCRRA +mCPT ]
If the pricing kernel also prices the risk-free asset, we have Rf = 1/E[mCRRA +
mCPT ] and Equation (11) can be rewritten into Equation (2), which we have
already discussed in the Introduction.8
To gain some intuition on the workings of the model, panel A of Figure 2
illustrates the two components of the pricing kernel. If b̂ = 0, then mCPT = 0,
−γ
and the second covariance term vanishes. Because mCRRA = R E describes
the investor’s marginal utility from equity returns, mCRRA is large when equity
payout is low. Hence, by the return Equation (11), the expected returns of
derivatives whose payoffs correlate with the market (cov(mCRRA ,R d ) > 0) are
lower. As such, calls should have high expected returns while puts should have
low returns.
When b̂ > 0, derivative returns are influenced by a second covariance term
cov(mCPT ,R d ). The prospect theory part of the pricing kernel, mCPT , is larger
when pπii is large and/or when the equity payoff is experienced as a loss (i.e.,
when the equity return is smaller than R f ). To see when pπii is large, note that
πi
pi
is approximately equal to the derivative of the weighting function used in
the computation of πi . Because the probability weighting functions have an
inverse S shape, the derivatives of these functions are large for both small and
large probabilities. The pricing kernel is thus elevated in states of extreme
7 This is a technical assumption to ensure the existence of a representative agent equilibrium. Without this
assumption, the ranking of states i may be different for different portfolios, such that the investor might
endogenously influence the ranking via her portfolio choice. Such “endogenous rankings” would complicate
the analysis substantially, as shown by Ingersoll (2016). Using the terminology from Ingersoll (2016), the above
assumption implies that we are analyzing an “order-constrained” optimization problem, where the ordering is
derived from the equity index. Note that, in equilibrium, where the investor holds only equity and no derivatives,
the rankings of equity states and wealth states coincide.
8 In the empirical analysis, we do not price the risk-free asset. Instead, we take the (observed) risk-free rate as
given and directly focus on excess returns like in Equation (11). Note that, to match any given risk-free rate level,
our model can be generalized by scaling the total utility function (and hence the pricing kernel mCRRA +mCPT )
with an additional free parameter. This would not affect Equation (11) or change any of our results.
3682

equity payouts, both large and small. With respect to losses, the factor (1+(λ−
1)1RE <Rf ) increases the CPT contribution to marginal utility by the constant
i
loss aversion parameter λ if and only if state i is a loss state, thus leading to a
discontinuity in the pricing kernel.
To get an idea of the quantitative effect of adding CPT preferences to a CRRA
framework, panel B of Figure 2 shows the state-dependent risk aversion implied
by the pricing kernel given our benchmark parameter estimates (described in
Section 2). To calculate state-dependent relative-risk aversion, we follow Ait-
Sahalia and Lo (2000) and Chabi-Yo, Garcia, and Renault (2008), who show that

CRRA CPT
it equals − ∂ln(m ∂RE+m ) RiE . The results in Figure 2 show that state-dependent
i
risk aversion is increasing in losses and decreasing in gains. Relative to standard
CRRA utility, this implies that a CPT investor is substantially more risk averse
for losses, and substantially more risk loving for gains. The magnitudes are
notable. For example, risk aversion increases to around 18 for losses greater
than 8%. Moreover, for gains greater than 1%, effective risk aversion becomes
negative, and the investor becomes risk loving. The latter effect is pronounced
enough to make the combined pricing kernel in panel A U shaped. Our results
are thus related to an existing literature on U-shaped pricing kernels discussed in
the literature section. The large magnitudes for implied risk aversion illustrate
the economic significance of these option pricing puzzles, and the ability of the
CPT model to generate such pronounced deviations from the CRRA benchmark
for the tails of the distribution is a key strength of the model.
In sum, the key feature of the novel pricing kernel derived above is that
expected returns of assets that pay off in extremely bad and/or extremely good
states are low, while, under standard CRRA, expected returns are always lower
for bad states than for good states.
1.4 The model-implied variance premium

In this subsection, we define the variance premium and show how the variance
premium can be expressed in terms of expected option returns.
As is common in the literature, we define the variance premium as the
difference between the risk-neutral and the actual expected variance of
the equity market return (see, e.g., Bollerslev, Tauchen, and Zhou 2009).
Specifically, define the actual probability measure P as the measure under
which the actual equity payoffs xiE are generated, and Q as the risk-neutral
measure under which equity and derivatives are priced. The variance premium
is then defined by
V P ≡ V Q −V P
2 E 2
= qi ln RiE −E Q ln R E − pi ln Ri −E P ln R E ,
i i
(12)
3683

where pi and qi denote the actual and risk-neutral probability of state i,

respectively. If investors were risk-neutral, P and Q would be identical and
the variance premium would be equal to zero. Hence, the variance premium
depends on investor preferences about variation in equity returns.
In our model, the risk-neutral probability for state i is given by
CRRA
mi +mCPT
qi = pi P CRRA i
. (13)
E m +mCPT

Using the expressions for mCRRA i +mCPT
i from Equation (10), and with an
assumption on the actual distribution of returns, we can calculate the model-
implied variance premium. Equation (13) shows how the difference between the
actual and the risk-neutral probability is determined by the preferences of the
investor, and illustrates how our setting differs from more traditional models.
In addition to the standard component in traditional CRRA-based asset pricing
models given by mCRRA
i , the risk-neutral probability qi depends on the CPT part
of the pricing kernel, mCPT
i .
The following proposition, which we prove in Appendix B, shows that the
variance premium and expected option returns are closely related.
Proposition 4. Let R c (K) and R p (K) denote the return to maturity on a call
and put option with strike K, respectively. If there are no arbitrage opportunities,
we have
E
s0 ∞

VP ≈ v p (K)E P R p (K)−R f dK + v c (K)E P R c (K)−R f dK, (14)
0 s0E
where s0E is the time-0 price of equity and v c (K) and v p (K) are scalars whose
expressions are given in Appendix B.
Proposition 4 shows that the variance premium is a weighted average of

the expected returns on OTM put and call options across strikes. Equation
(14) integrates over put options for strikes K < s0E and over call options for
strikes K > s0E . In Appendix B, we show that the weights v c (K) and v p (K) are
effectively always negative. Hence, one can “earn” the variance premium by
selling put and call options with various strikes. The resultant strategy forms
a portfolio of short positions in so-called “strangles,” where each strangle is a
combination of an OTM put and an OTM call option. The variance premium,
V Q −V P , thus represents the expected profit on this portfolio, which effectively
sells both insurance against bad states of the world and sells a lottery-like
payoff in the good states of the world. Hence, the variance premium reflects
information about the compensation received in extreme states of the underlying
asset (the equity market in case of index options), as these very negative and
3684

positive outcomes determine the value of OTM puts and calls, respectively.
Equation (14) thus shows that the pricing of both tails of the return distribution
affects the variance premium.
Note that Equation (14) is “model-free” in the sense that it neither requires
assumptions on preferences nor on the equity return distribution. The only
assumption is the absence of arbitrage opportunities. Hence, the equation holds
both in the data, as well as in the model. We can thus write the difference between
the model-implied variance premium, V PCP T , and the observed variance
premium, V P , as follows

E
s0
p p

V PCP T − V P≈ v p (K) ECP
P
T R (K) − Ê R (K) dK (15)
0
∞
c c
+ v c (K) ECP T R (K) − Ê R (K) dK,
P
s0E
P
where ECP T [R] denotes the expected return implied by the model and Ê[R]
the sample average return for any return R.9 Thus, if the model-implied option
returns match the option returns in the data, then the variance premium is the
same for both the model and the actual data. Pricing the cross-section of OTM
options is therefore a sufficient condition for fitting the variance premium.10
The equation also shows that a model can match the variance premium while, at
the same time, generating completely counterfactual option prices. Thus, fitting
the variance premium by fitting options, as we do in this paper, is a stricter test
than fitting only the variance premium.
Finally, as an important point of reference for our subsequent analysis,
note that the variance premium is zero for the CRRA case if the market
payoff, x E (and hence the terminal wealth of the investor, WT ) have a
lognormal distribution, and if the representative agent has pure CRRA
preferences (b0 = 0) (see, e.g., the results in Samuelson and Merton 1969;
Rubinstein 1976).11
In Appendix C we further decompose the variance premium into upside and
downside variance premiums by conditioning on positive and negative market
9 In small samples, Equation (15) is subject to the sampling error in the estimates of the average returns and the
variance premium.
10 This reasoning comes with two minor caveats. First, Equation (14) is only an approximation. We show in Appendix
B that the approximation error is equal to the difference between central and noncentral second moments. For our
application, this term is numerically small, as we focus on 1-month returns—it is equal to 4.4 in our benchmark
model, while empirically the variance premium is equal to 146.7 (both numbers are in percentage-squared).
Second, empirically, we only have options for a finite number of strike prices.
11 In a setting with discrete trading and lognormal returns, the results in Brennan (1979) imply that CRRA
preferences are the only preferences that generate a zero variance risk premium. Drechsler and Yaron (2011)
show that CRRA preferences generate a zero variance premium also in a multiperiod setting with long-run risk.
3685

returns, respectively. We show that both the upside and downside variance
premium contribute positively to the total variance premium, and that our model
generates a decent fit of these components.
2. Empirical Results
This section explores the ability of the CPT model described in Section 1 to
fit the data. We first show that the model yields a good fit for observed option
returns for a set of plausible input parameter values, which implies that the

model also yields a good fit for the variance premium under those parameters.
We then conduct a comparative statics exercise that conveys the intuition for
most of the results that follow. Finally, we use a GMM approach to match the
CPT model to the data.
2.1 Data
Our main dataset consists of daily closing midquotes of S&P 500 index options
for various deltas and maturities obtained from the OptionMetrics database from
January 1996 to March 2016. OptionMetrics creates a “surface” of interpolated
option prices for fixed levels of the Black-Scholes delta and for fixed maturities
(30 calendar days, 60 days, etc.) for both puts and calls. We focus on 30-day
options because they are most liquid and the ones most frequently used in the
literature. For each day and option delta, and for both put and call options,
we construct the return on buying an option and holding it to maturity, that is,
for 30 calendar days. We thus create overlapping returns. This yields a panel
containing the returns on 26 options across deltas. Of these 26 options, 13 are
calls with deltas ranging from 0.2 to 0.8 (from OTM to in-the-money (ITM)),
and 13 are puts with deltas ranging from −0.8 to −0.2 (from ITM to OTM).
Table 1 presents summary statistics for option returns. The observed values
reported in Figure 1 (black squares) are identical to the numbers in Table
1 except that deltas are replaced by the average moneyness of the options
(K/S0 ) for a given delta. The observed option returns are in line with the
documented stylized facts in the literature (e.g., Coval and Shumway 2001;
Jones 2006; Driessen and Maenhout 2007; Bakshi, Madan, and Panayotov
2010; Constantinides, Jackwerth, and Savov 2013).12
We focus on option returns, instead of option prices, for two reasons. First,
Proposition 4 shows that the variance premium is directly related to expected
option returns, and that matching option returns is a sufficient condition for
matching the variance premium. Second, option returns are directly informative
12 Bakshi, Madan, and Panayotov (2010), Constantinides, Jackwerth, and Savov (2013), and Chaudhuri and
Schroder (2015) also find call returns to be low compared to the underlying index return, although in our
sample call returns seem to be somewhat lower than in previous work. This is due to (1) differences in sample
periods and (2) the fact that, to make more efficient use of data, we use overlapping option returns. Given that
the return distribution of OTM options is skewed, using overlapping returns helps to reduce estimation error
(Broadie, Chernov, and Johannes 2009).
3686

[13:59 5/8/2019 RFS-OP-REVF180130.tex]

Table 1
Summary statistics for put and call option returns
Call option returns
Option delta 0.200 0.250 0.300 0.350 0.400 0.450 0.500 0.550 0.600 0.650 0.700 0.750 0.800
Average return −0.244 −0.179 −0.133 −0.096 −0.067 −0.044 −0.026 −0.012 −0.001 0.006 0.011 0.015 0.019
t-statistic average −2.718 −2.144 −1.701 −1.317 −0.972 −0.691 −0.441 −0.208 −0.019 0.126 0.254 0.383 0.551
Return standard deviation 2.014 1.803 1.632 1.489 1.364 1.254 1.154 1.061 0.976 0.896 0.809 0.728 0.643
Put option returns

Option delta −0.800 −0.750 −0.700 −0.650 −0.600 −0.550 −0.500 −0.450 −0.400 −0.350 −0.300 −0.250 −0.200
Average return −0.143 −0.163 −0.182 −0.203 −0.224 −0.246 −0.271 −0.298 −0.331 −0.373 −0.423 −0.489 −0.572
t-statistic average −3.017 −3.131 −3.263 −3.415 −3.569 −3.741 −3.930 −4.139 −4.410 −4.787 −5.224 −5.864 −6.742
Return standard deviation 0.863 0.946 1.020 1.090 1.159 1.228 1.298 1.369 1.441 1.517 1.596 1.674 1.744
This table shows summary statistics of returns on S&P 500 index options with 30 calendar days to maturity for puts and calls with various strike levels as measured by their Black-
Scholes delta. The data are collected from OptionMetrics, which provides daily interpolated option prices for various deltas and maturities. The sample period runs from January
1996 to March 2016 at a daily frequency and with overlapping 30-day returns. The t-statistic for the average return is calculated using Newey-West standard errors (using 30 lags).
Page: 3687
3687
3667–3723
about the preference parameters. This can be seen easily from Equation (11),
where the expected excess option return depends on the covariance of the
option payoff and the pricing kernel (which, in turn, depends on the preference
parameters). In contrast, option prices depend on this covariance term as well as
on the expected option payoff. Given that the expected option payoff depends
strongly on the assumed equity distribution, if we fitted to option prices rather
than returns, the results would be very sensitive to the assumption on this
underlying return distribution. As shown in Section 2.4.3, where we fit to
option returns, our estimates hardly depend on the assumed distribution for

the underlying asset return.
In addition to option returns, we construct the corresponding S&P 500 index
returns for the 30-day holding period at a daily frequency. This gives an average
return of 0.75% per 30 calendar days and a standard deviation of 4.84% per
month. The average 1-month Treasury-bill rate over the sample period is 0.19%,
which gives an in-sample equity premium of 0.56% per month.
The variance premium in the data is calculated as the difference between the
risk-neutral expected variance, V Q , and the actual expected variance of equity
index returns, V P . We follow the existing literature and use the square of the
VIX as our measure of risk-neutral expected variance (e.g., Bollerslev, Tauchen,
and Zhou 2009; Carr and Wu 2009).13 To calculate the actual expected variance,
V P , we regress realized variance, calculated using daily returns over the last 22
trading days, on the 1-month-lagged realized variance and the square of the VIX
level at the beginning of the month over which the realized variance is measured.
The expected realized variance is then calculated as the forecast implied by
this regression model (see Londono 2014). In our sample, the annualized
average of the squared VIX equals 510.1%2 , and the annualized average actual
expected variance equals 363.4%2 . This gives an average variance premium
of 146.7%2 , which is similar to values found in the related literature (e.g,
Bollerslev, Tauchen, and Zhou 2009, Londono 2014).
2.2 Benchmark parametrization

Before we estimate the model parameters with GMM, we assess the
implications of our model for a benchmark parametrization. We assume that
E
the monthly
E equity return, R , is lognormally distributed with parameter
μE = E R , which is to be determined in equilibrium, and exogenous volatility
σE . Because our model is set up for a finite number of states, we use a
discrete approximation of the lognormal distribution, taking a grid of 200
potential outcomes for the 1-month-ahead equity return ranging from −50% to
13 Formally, this risk-neutral variance is given by the integrated variance of continuous-time returns over a given
time period. We work in discrete time. If the continuous-time returns are uncorrelated over time, this integrated
variance equals the variance of the return over a discrete time period.
3688

150%.14 We assume a one-period constant risk-free rate of R f = 0.19% and an

unconditional volatility of equity returns of σE = 4.84%, which are, respectively,
the observed unconditional mean of the risk-free rate approximated by the U.S.
1-month Treasury-bill rate and the unconditional monthly volatility of the S&P
500 index returns over our sample period.
We parametrize preferences using standard values in the literature.
Specifically, following Tversky and Kahneman (1992), we use a loss aversion
parameter of λ = 2.25 and, to parametrize the probability weighting function in
Equation (6), we use c1 = c2 = 0.65. We use a benchmark coefficient of relative-

risk aversion of γ = 1, which implies log utility for the CRRA component of the
investor’s total utility. Because risk preferences in our model are determined
by both the CRRA component and the CPT component, assuming γ = 1 is less
restrictive than it would be in a pure CRRA model.
Finally, we need to make an assumption on the parameter b0 , which governs
the relative importance of the CRRA and gain-loss utility components. In the
absence of clear guidance from the existing literature, we assume b0 = 0.95,
which, in combination with the other parameters in the benchmark specification,
makes CPT’s contribution to total utility equal to 50%.15
Based on these inputs, we can calculate the model-implied variance premium

as follows. First, we solve Equation (10) for the return on equity E R E .
Second, once E R E is fixed, we use the pricing kernel given by Equation (10)
to price the cross-section of options. Finally, we use Proposition 4 to compute
the variance premium from the option prices calculated in the previous step.16
Finally, for the benchmark parameters in this section, we obtain an annualized
value for the variance premium of 162.4%2 . This is close to the actual value
of 146.7%2 in the data. The CPT model can thus generate a variance premium
similar to empirically observed levels for a plausible set of input parameters.
2.3 Comparative statics

Because a pure CRRA model yields a zero variance premium, predicts the
wrong sign on call option returns, and predicts much higher returns for OTM put
returns, the good fit of the CPT model documented in the previous section must
be due to the CPT component in the investor’s total utility. In this section, we
use comparative statics to investigate which CPT parameters drive the model’s
ability to generate a variance premium and fit the returns of options, especially
those out of the money.
We start with the scale parameter b0 , which governs the weight of the CPT
component in the investor’s utility. Panel A of Figure 3 shows model-implied
14 Our main results for the ability of our model with CPT to match the observed variance premium remains robust
for grids ranging from 50 to 500 outcomes and for equity returns for narrower (-25% to 125%) and wider (−75%
to 175%) intervals.
15 The relative contribution of CPT to total is calculated as b0 CP T [XT ]
EU [WT ]+b0 CP T [XT ] (see Equation (3)).
16 Equivalently, one could derive the VP by computing the risk-neutral probabilities and then using Equation (12).
3689

variance premiums for values of b0 ranging from zero, in which case the CPT
contribution to the investor’s utility is also zero, to infinity, in which case
the CPT contribution approaches 100%. The solid line in panel A shows that
varying b0 induces economically significant variation in the variance premium.
As b0 approaches zero, the variance premium approaches zero. The variance
premium increases monotonically with b0 and reaches a level above 300 for
b0 → ∞, which is about twice the level of the variance premium observed in
the data.
The dashed line in panel A of Figure 3 offers first insights into why CPT can

explain the variance premium. The dashed line is constructed identically to the
black line, with the difference being that probability weighting is set to zero
(c1 = c2 = 1), such that the decision weights used by the investor coincide with the
actual probabilities. Once we switch probability weighting off, the model loses
its ability to generate a substantial variance premium. This finding indicates
that it is not loss aversion which induces the variance premium (because loss
aversion is still set at its benchmark value) but probability weighting.
Panel C of Figure 3 explores the importance of probability weighting in
greater detail by varying probability weighting from 0.4 (strong weighting) to
1 (no weighting). We again find that varying the degree of probability weighting
induces substantial changes in the model-implied VP, with values ranging from
about 0%2 (for the case of no probability weighting) to 400%2 (for the case
of strong weighting). Probability weighting thus is the central driver of the
variance premium in our model.
The mechanism behind the effect of probability weighting on the variance
premium is straightforward. In the presence of probability weighting, the state
prices of extreme outcomes increase as the investor attaches a higher decision
weight to these states. Intuitively, investors with probability weighting find both
insurance and lottery tickets attractive, and are thus willing to pay higher prices
for OTM put and call options. As shown in panels B and D of Figure 3, all
else equal, increasing the CPT contribution or increasing probability weighting
decreases option returns. Equivalently, buying strangles is very attractive for
a CPT investor, because strangles provide both lottery and insurance. By
Proposition 4, higher prices for strangles, which are the result of stronger
probability weighting, imply a positive variance premium.
Contrasting the strong effect of probability weighting, panel E shows that
the variance premium varies very little with the degree of loss aversion. For
the benchmark set of parameters, we find a weakly positive relation between
the loss aversion parameter λ and the variance premium. Panel F shows that
the effect of loss aversion on the return of OTM calls goes in the opposite
direction as that for OTM puts. More loss aversion increases the equilibrium
expected equity return. Given the positive exposure of calls to equity returns,
this also increases the expected call return. Puts have negative exposure to
equity, and the expected put returns thus decrease. The overall effect of loss
aversion on the variance premium will thus depend on whether the put or call
3690

A CPT contribution (scale) B CPT contribution (scale)
400 benchmark OTM put

c=1 0 OTM call
%2
%
200
-50
0
-100
0 20 40 60 80 100 0 20 40 60 80 100

% %
C c (prob. distortion) D c (prob. distortion)
400
0
%2
%
200
-50
0
-100
0.5 0.6 0.7 0.8 0.9 1 0.5 0.6 0.7 0.8 0.9 1
c c
E O (loss aversion) F O (loss aversion)
400
0
%2
200
-50
0
-100
1 2 3 4 1 2 3 4
O O
Figure 3
CPT-implied variance premium and OTM option returns for alternative preference parameter values
This figure shows the annualized CPT-implied variance premium (left) and 30-day returns for out-of-the-money
(OTM) puts and calls (right) as a function of the scale b0 (panels A and B), which determines the contribution
of CPT utility to total utility, probability weighting, c (panels C and D), and loss aversion, λ (panels E and F).
The representative agent’s CPT preferences are defined in Section 1.1. The variance premium is defined as the
difference between the risk-neutral and the expected realized variance of equity returns, as explained in Section
1.4. We report the CPT-implied returns for puts and calls with 0.95 and 1.05 moneyness (K/S ), respectively. In
all panels, we set c = 0.65, b0 = 0.95, and λ = 2.25, apart from the parameter that is varied.
effect dominates. Our results in panel E show that, qualitatively, the put effect
dominates (because the variance premium depends negatively on expected put
and call returns). Quantitatively, however, loss aversion has only a minor effect
on the variance premium.
3691

2.4 GMM estimation

We now provide more formal evidence on the ability of our CPT model to match
the data. Specifically,
we use GMM to estimate optimal preference parameters
θ = γ ,λ,b0 ,c1 ,c2 , under different sets of parameter restrictions.
2.4.1 GMM specification. We estimate optimal preference parameters θ

using the following set of moment conditions:
⎡ ⎤
μE (θ)−T −1 T1 RtE

⎢ V P (θ )−T −1 T V Pt ⎥

g(θ) = ⎢ ⎥
⎣ μp (θ )−T −1 T R p ⎦ ,
1
1 t
μc (θ )−T −1 T1 Rtc
where μE (.), μp (.), and μc (.) are the model-implied equilibrium expected
returns for equity, calls, and puts, respectively, and V P (·) is the model-implied
variance premium, all matched to their empirical counterparts in our sample.
We include 13 calls and 13 puts with different strike prices, so that μp (.), and
μc (.) are both 13×1 vectors. We thus have 28 moment conditions in total. We
find the
θ that minimizes the GMM goal function V F = g(θ) Wg(θ ), where W is
the weighting matrix for the moment conditions. We use a diagonal weighting
matrix which assigns an equal weight of one-fourth to the equity, call, put,
and variance premium moments. We calculate Newey-West corrected standard
errors for the estimated parameters to account for potential autocorrelation
and heteroscedasticity in the residuals due to overlapping returns as we
construct holding-to-maturity returns each day. Thus, we can do inference on
the estimated parameters by exploiting that

θ ∼ N (θ,V /T ),
where V = [G W G]−1 G W SW G[G W G]−1 ,S is the standard Newey-West-
corrected covariance matrix, and G = δg/δθ . The matrix G captures the
sensitivity of the moment conditions to the parameters, as analyzed in
Section 2.3.
2.4.2 GMM baseline results. Table 2 presents our baseline GMM estimates.
As a point of reference, specification (1) sets b0 = 0 and γ = 1, which means that
our model collapses to the pure CRRA case with log utility. Not surprisingly,
the statistics shown in Table 2 reflect the inability of the CRRA model to fit the
option data. The average put option return estimated in this case is −5.86%,
while the first column shows that the return in the data is −30.14%. The CRRA
model also fits calls poorly. While the model predicts a positive 6.50% average
call return, the return in the data is −5.77. The blue x’s in Figure 1, which
we have already discussed in the introduction, show the model-implied returns
from specification (1) for each of our 26 options.
The variance premium is zero for the CRRA specification (1), as it should
be. Finally, the CRRA model slightly undershoots the equity premium, but note
3692

Table 2
GMM estimates in the unconditional setting
(1) (2)
CRRA Benchmark
Observed log-utility CPT
γ (risk aversion) 1 1
b0 (scale) 0 0.95
λ (loss aversion) 1.39
[0.30]
c (probability weighting) 0.65
[0.00]

CPT contribution (%) 0 50.00
Variance premium (%2 ) 146.72 0 146.33
Equity premium (%) 0.56 0.23 0.53
Average call return (%) −5.77 6.50 −9.23
Average put return (%) −30.14 −5.86 −31.71
Average OTM call return (%) −11.27 7.79 −16.94
Average OTM put return (%) −41.46 −7.31 −43.01
VF 5,942.41 3.47
RA without CPT 1.00 2.27
This table reports GMM estimates for the degree of loss aversion, λ, and probability weighting, c (c1 = c2 ), for
two alternative preference specifications. In all specifications, the reference level, WRef , is calculated using a
risk-free rate of 0.19%, and the level of risk aversion, γ , is fixed at 1. For the benchmark CPT specification
(specification (2)), the scale parameter, b0 , is fixed at 0.95, which implies a 50% contribution of CPT to total
utility. p-values of tests for the respective hypotheses that λ = 1 and c = 1 are reported in brackets. The p-values
are calculated using Newey-West standard errors. For each specification, we report CPT’s relative contribution
b CP T [X ]
to total utility, which is calculated as EU [W0 ]+b CPTT [X ] . We also compare the observed variance and equity
T 0 T
risk premiums, the average model-implied call and put returns, and the average model-implied OTM call and
put returns with those implied by each alternative specification. The overall fit is summarized by the level of the
value function (VF) evaluated at the optimum. Finally, we report the level of risk aversion (RA) that would be
needed in the no-CPT case (b0 = 0) to obtain the model-implied equity premium. These parameters are estimated
by applying GMM to returns on the S&P 500 and on S&P 500 call and put options with different strikes and 30
days to maturity and to the variance premium. The weighting matrix gives an equal weight of one-fourth to the
equity, call, put, and variance premium moments. The sample period runs from January 1996 to March 2016.
that we do not model consumption, so fitting the equity premium is less of an

achievement than it would be otherwise. The real test in our setting is pricing
the cross-section of option returns, and the CRRA model fails this test.
Because the variance premium is guaranteed to be zero in a CRRA model
regardless of the coefficient of risk aversion, using alternative values of γ would
not meaningfully alter any conclusions. We will thus restrict γ to be equal to
one in the remainder of the paper. This assumption has two advantages. First,
by virtue of Proposition 2, the results we derive for a one-period CPT setting
remain valid for a multiperiod version of the model in which the investor is
allowed to adjust portfolio weights each period. Second, because γ captures
the curvature of the CRRA part of the investor’s utility, and because λ alone
captures the concavity of the gain-loss function in the CPT part, identifying γ
and λ separately is empirically challenging. By restricting γ = 1, we get better
estimates of loss aversion λ.
Specification (2) estimates the CPT model. We start by restricting probability
weighting to be the same for losses and gains (c1 = c2 = c), a restriction we relax
3693

below. The CPT model matches the variance premium almost perfectly with
a value of 146.33%2 , relative to 146.72%2 in the data. At the same time, our
model also fits the underlying option returns well. For example, the average
call return is −9.23% in the model, compared to −5.77% in the data, and the
average put return in the model is −31.71%, compared to −30.14% in the data.
The red +’s in Figure 1 show the model-implied returns from specification (2)
for each of our 26 options, and confirm that the model is successful at pricing
options across the whole spectrum of strike prices. Finally, the value of the
GMM objective function, VF, which can be thought of as a weighted average

of the fitting errors obtained for each moment condition, shows that the CPT
model is orders of magnitude more accurate than the CRRA model (V F = 3.47
vs. V F = 5,942.41).
Looking at the parameter estimates we obtain for our benchmark
specification (2), the key result is that probability weighting, c, is very precisely
estimated at 0.65. This number is interesting for two reasons. First, the p-value
implies that the estimated 0.65 is significantly different from the benchmark
case of no probability weighting (c = 1). We can thus formally reject the null
hypothesis of no probability weighting. Second, the degree of probability
weighting, 0.65, is very close to the value estimated by Tversky and Kahneman
(1992), as well as values for probability weighting used in the existing finance
literature (e.g., Barberis and Huang 2008 use 0.65).
When estimating specification (2), we fixed the scale parameter b0 . The
reason is that it is empirically hard to jointly identify b0 and c with
sufficient precision, even though the two parameters are conceptually distinct.
Intuitively, the problem arises because “too little” probability weighting can
be compensated for by giving a higher weight to the CPT part in the utility
function. Conversely, even extreme levels of probability weighting do not
matter much if the CPT component has only little weight. In Appendix D,
we document this identification issue more formally using Gentzkow and
Shapiro (2015) sensitivities. We caution that, because we have fixed the scale
parameter, our benchmark estimate for c is not a fully independent estimate of
probability weighting. But, we nevertheless argue that these estimates constitute
strong evidence for probability weighting, and that it can help us understand
observed option prices. To the extent that a CPT contribution of 50% is
considered a reasonable number, the estimate of 0.65 suggests that levels of
probability weighting measured from the lab are externally valid for actual
option investors. Conversely, to the extent that the lab estimates of probability
weighting are considered valid starting points for examining field data, the
results in specification (2) show that assuming a 50% CPT contribution would
make the model consistent with the data.
2.4.3 GMM extensions and robustness. In Table 3, we further analyze the

impact of the CPT contribution to total utility, governed by the parameter b0 ,
on the model’s fit. In specifications (2) to (4), we use three different values for
3694

Table 3
GMM estimates in the unconditional setting: Alternative scales
(1) (2) (3) (4) (5)
Observed Benchmark Low b0 High b0 Estimated b0
γ (risk aversion) 1 1 1 1 1
b0 (scale) 0.95 0.30 5 100 0.27
– – – – [0.23]
λ (loss aversion) 1.39 1.60 1.31 1.28 1.63
[0.30] [0.37] [0.18] [0.14] [0.37]
c (probability weighting) 0.65 0.49 0.75 0.78 0.48
[0.00] [0.00] [0.00] [0.00] [0.03]

CPT contribution (%) 50.00 21.30 84.40 99.09 19.22
Implied variance premium (%2 ) 146.72 146.33 146.67 146.15 146.10 146.71
Implied equity premium (%) 0.56 0.53 0.50 0.54 0.54 0.49
Average call return (%) −5.77 −9.23 −7.62 −9.95 −10.15 −7.39
Average put return (%) −30.14 −31.71 −30.08 −32.31 −32.44 −29.83
Average OTM call return (%) −11.27 −16.94 −14.07 −18.35 −18.75 −13.66
Average OTM put return (%) −41.46 −43.01 −41.77 −43.24 −43.24 −41.56
VF 3.47 0.73 6.19 7.11 0.69
RA without CPT 2.27 2.14 2.32 2.34 2.13
Specifications (1) to (4) report GMM estimates for the degree of loss aversion, λ, and probability weighting,
c = c1 = c2 , for alternative values of b0 ; that is, the scale parameter controlling the contribution of the CPT
component to total utility. Specification (5) reports the estimates for a specification in which the scale is also
estimated. In all specifications, the level of risk aversion, γ , is fixed at 1, and the reference level, WRef , is
calculated using a risk-free rate of 0.19%. We report the CPT contribution, the model-implied variance and
equity risk premiums, the average model-implied call and put returns, the average model-implied OTM call and
put returns, the value function (VF) evaluated at the optimum, and the level of risk aversion (RA) that would
be needed in the no-CPT case to obtain the model-implied equity premium. These parameters are estimated by
applying GMM to returns on the S&P 500 and on S&P 500 call and put options with different strikes and 30
days to maturity and to the variance premium. The GMM weighting matrix gives an equal weight of one-fourth
to the equity, call, put, and variance premium moments. The sample period runs from January 1996 to March
2016. p-values from tests of the hypotheses that λ = 1, b0 = 0, and c = 1, respectively, are reported in brackets.
These p-values are calculated using Newey-West standard errors.
b0 : a low value of 0.3, which implies a CPT contribution of about 21.3%; a

value of 5, which implies a CPT contribution of 84.4%; and a very high value
of 100, which implies a CPT contribution of 99%.
Several insights are to be had. First, the loss aversion estimate is largely
insensitive to changes in the CPT contribution. Combined with the fact that we
assume γ = 1, this implies that overall risk aversion, and therefore the model-
implied equity premium, remains remarkably constant.
Second, the probability weighting estimate varies across specifications in
a way that is expected. If the contribution of CPT to the utility function
is high, less probability weighting is needed to fit options well. Likewise,
for low CPT contributions, a substantial degree of probability weighting is
needed to fit the data. Remarkably, even as the CPT contribution becomes very
high, the estimated probability weighting parameter does not approach the no-
probability-weighting value of 1, but, rather, approaches a value of about 0.8.
Estimated probability weighting parameters in Table 3 range between 0.49 and
0.78, which is not unreasonable given estimates in the related literature on
decision making under risk (e.g., Camerer and Ho 1994 estimate a value of
0.56; Wu and Gonzalez 1996 estimate 0.71). Echoing our earlier argument,
3695

the estimates in Table 3 imply that, for arguably plausible levels of the
CPT contribution (21% to 99%), the CPT model yields probability weighting
estimates that are in the same region as estimates found in lab studies.
Third, and importantly, across all specifications, we can statistically reject the
hypothesis of no probability weighting. Hence, while the value we pick for b0
has an influence on the probability weighting parameter, it does not change the
central conclusion that probability weighting is the central ingredient to fitting
the data. Finally, specifications (2) to (4) show that the CPT model yields an
excellent fit of the variance premium and the underlying option returns in all

cases, with only minor variations among the different models. In sum, from
specifications (2) to (4), we conclude that our central findings are robust for
reasonable values of b0 .
Specification (5) in Table 3 jointly estimates b0 ,λ, and c. Based on the
point estimates of these parameters, our previous conclusions remain largely
unchanged. The GMM goal function indicates that the model fits the data now
even better, with only small remaining pricing errors for both option returns
and the variance premium. The model yields an estimated value for b0 of 0.27,
which implies a CPT contribution of about 19%. The associated estimate for
the probability weighting parameter is c = 0.48, which is slightly lower than
the value estimated in the lab study of Camerer and Ho (1994). While these
results are reassuring, one caveat is—as highlighted before—that it is very hard
for the model to cleanly identify b0 and c simultaneously with high degrees
of confidence. Consistent with the evidence from the Gentzkow and Shapiro
(2015) sensitivities reported in Appendix D, the estimates for b0 and c are at
best only marginally significant when we jointly estimate these parameters in
specification (5).
In Appendix E we show that our results are not specific to assuming
lognormal equity index returns. Using either a normal distribution for equity
returns instead of a lognormal distribution, or a skewed student-t distribution,
which allows for skewness and fat-tailed returns, yields largely similar results.
In unreported results we further show that our results (1) are robust to using
an alternative number of gridpoints to approximate the lognormal distribution
and that (2) results are robust to reasonable variations in the reference point.
Finally, we reestimate the model for two alternative preference specifications.
First, we replace Equation (3) by a rank-dependent utility specification (Quiggin
1982). That specification uses only CRRA utility, no gain-loss utility, and
applies weighting function w+ to the probabilities in the CRRA part. Second, we
otherwise keep specification (3), but now apply probability weighting to both,
the CRRA and gain-loss parts. The results for both models are presented in
specifications (2) and (3) in Table F.1 in Appendix F. Despite the fact that these
models are conceptually different, both provide a good fit for option returns
and the variance premium, similar to our benchmark model. This robustness
across specifications shows that probability weighting is a powerful driver of
3696

OTM option returns, and that specific assumptions about the curvature of the
utility or gain-loss functions are relatively less important.
2.4.4 Probability weighting on the downside and on the upside: c1 = c2 . We

now relax the assumption that c1 = c2 , thereby allowing for differential levels
of probability weighting for losses (the “downside”) and gains (the “upside”).
These results are shown in Table 4, specification (2). We find that this model
fits the data only marginally better, implying that c1 = c2 is not a particularly
restrictive simplification. In terms of probability weighting estimates, the

results show that the model-implied weighting on the downside becomes more
pronounced, while the weighting on the upside becomes less pronounced
than in the benchmark case c1 = c2 = 0.65. However, quantitatively, at c1 = 0.62
and c2 = 0.69, the optimal values are rather close to the benchmark, and we
cannot formally reject that the values are the same (p = 0.36). Importantly, both
parameters remain significantly different from the no-probability-weighting
case at any conventional significance level.
Specifications (3) and (4) of Table 4 present results for weighting
probabilities only on the downside or only on the upside, respectively. Both
models perform well in terms of fitting the variance premium. However, the
results also show that this good fit of the variance premium masks problems
of the model when fitting the cross-section of option returns, especially those
out of the money. By switching probability weighting off entirely, we are,
effectively, pricing options as we would be under a standard expected utility
model. Hence, the restricted CPT model that only weights probabilities on
the downside (upside) inherits the CRRA model’s poor fit for calls (puts).
Figure 4 documents this result for the individual options in our data. Distorting
only on the downside does reasonably well for puts (the x’s in Panel B). For
calls, however, the relation between model-implied call returns is convex and
upward-sloping for deep OTM calls, which is opposite to the concave and
downward-sloping relation observed in the data (black squares in panel A). A
similar pattern emerges for weighting probabilities only on the upside. These
empirical results argue in favor of a model which distorts both tails of the
distribution, which is in line with extensive laboratory evidence documenting
that most decision makers distort probabilities for both losses and gains (e.g.,
Tversky and Kahneman 1992, Abdellaoui 2000, Bruhin, Fehr-Duda, and Epper
2010).
The findings reported in this section further underscore our earlier point that a
model can produce a good fit for the VP, even if the individual option returns are
not well matched, as long as the fitting errors for expected option returns cancel
out across all options. Comparing models based on their ability to fit the VP
alone may not be a very informative test; looking at the cross-section of options
is stricter, and therefore, all else equal, preferable. The GMM goal function
estimates support this view, as they are substantially higher for specifications
(3) and (4), relative to specification (2).
3697

Table 4
GMM estimates in the unconditional setting: Asymmetric probability weighting
(1) (2) (3) (4)

Benchmark Asymmetric
CPT CPT
γ (risk aversion) 1 1.00 1.00 1.00
b0 (scale) 0.95 0.95 0.95 0.95
λ (loss aversion) 1.39 1.32 1.00 2.12
[0.30] [0.38] [0.40] [0.01]
c1 (weighting losses) 0.65 0.62 0.43 1.00

[0.00] [0.00] [0.00] –
c2 (weighting gains) 0.69 1.00 0.42
[0.00] – [0.00]
CPT contribution (%) 50.00 49.22 44.30 54.24
Implied variance premium (%2 ) 146.72 146.33 146.36 146.77 146.57
Implied equity premium (%) 0.56 0.53 0.50 0.46 0.66
Average call return (%) −5.77 −9.23 −9.16 −4.05 −5.86
Average put return (%) −30.14 −31.71 −31.58 −30.15 −27.49
Average OTM call return (%) −11.27 −16.94 −16.05 −4.72 −18.33
Average OTM put return (%) −41.46 −43.01 −43.74 −46.30 −29.85
VF 3.47 3.08 26.94 44.02
RA without CPT 2.27 2.17 1.98 2.84
This table reports GMM estimates for the degree of loss aversion, λ, and probability weighting for gains and losses,
c1 and c2 , respectively, for different specifications of CPT’s probability weighting feature. In all specifications,
the reference level, WRef , is calculated using a risk-free rate of 0.19%, the level of risk aversion, γ , is fixed at
1, and the scale, b0 , is fixed at 0.95, which is the scale that yields a 50% contribution of CPT to total utility in
the benchmark specification (specification (1)). p-values of the hypotheses that λ = 1, and c = 1, respectively, are
reported in brackets. The p-values are calculated using Newey-West standard errors. For each specification, we
also report CPT’s contribution to total utility, the model-implied variance and equity risk premiums, the average
model-implied call and put returns, the average model-implied OTM call and put returns, the value function (VF)
evaluated at the optimum, and the level of risk aversion (RA) that would be needed in the no-CPT case to obtain
the model-implied equity premium. These parameters are estimated by applying GMM to returns on the S&P
500 index and on S&P 500 call and put options with different strikes and 30 days to maturity and to the equity
and variance premium. The GMM weighting matrix gives an equal weight of one-fourth to the equity, call, put,
and variance premium moments. The sample period runs from January 1996 to March 2016.
The results in Table 4 provide further insight into the workings of the model.
What is notable is the fact that both specifications (3) and (4) feature substantial
degrees of probability weighing. Because the VP can be thought of as a portfolio
of strangles, and because strangles generate a positive return when there is
demand for both lottery and insurance, the absence of a preference for lottery-
like payoffs (induced by weighting on the upside) requires a strong desire
for insurance to still induce high enough strangle returns, which implies a
substantial degree of probability weighting in the model. In the absence of a
desire for insurance, an analogous pattern emerges.
Combined, the results in this section show that distorting probabilities in
both tails of the equity return distribution is important. Distorting only one
tail yields substantial pricing errors for individual options. The ability to price
both puts and calls using only one underlying driver, probability weighting, is
a particular advantage of the CPT model.
3698

A Average Call Returns
0
Percent
-20
-40 Observed

Asymmetric distortion
Left-tail distortion
-60 Right-tail distortion
0.92 0.94 0.96 0.98 1 1.02 1.04 1.06 1.08

Moneyness
B Average Put Returns
0
Percent
-20
-40
-60
0.92 0.94 0.96 0.98 1 1.02 1.04 1.06 1.08

Moneyness
Figure 4
Average option returns and model fit for CPT models with asymmetric probability weighting
This figure compares the average observed 30-day return for all options in our sample with the ones predicted by a
CPT representative agent model with asymmetric probability distortion (c1 = c2 ), a model in which only the left-
tail probabilities are subject to probability distortion (c2 = 1), and a model in which only the right-tail probabilities
are subject to probability weighting (c1 = 1); see specifications (2), (3), and (4) in Table 4, respectively. The sample
comprises daily S&P 500 index option returns from 1996 to March 2016 for various strike prices. Of the 26
options in the sample, 13 are calls, with deltas ranging from 0.2 to 0.8 (from OTM to ITM). The other 13 options
are puts, with deltas from −0.8 to −0.2 (from ITM to OTM). Results are presented by moneyness (K/S0 ) for a
given delta, separately for calls (panel A) and puts (panel B).
3. Time-Varying Variance Premium and Option Fit

In the previous section, we showed that our CPT model can match the
unconditional level of the VP as well as the cross-section of option returns when
allowing for probability weighting. We now investigate whether the CPT model
can also capture time variation in option returns and the variance premium. This
3699

1500
Benchmark
Low vol (V=0.02)
High vol (V=0.1)
1000
%2
500

0
0.5 0.6 0.7 0.8 0.9 1
c
Figure 5
CPT-implied variance premium for alternative volatility levels
This figure shows the annualized CPT-implied variance premium as a function of the probability weighting
parameter, c, for three different values of monthly equity volatility: 2% (low), 4.84% (benchmark), and 10%
(high). We set b0 = 0.95 and λ = 2.25.
is important for two reasons: first, the VP fluctuates widely, from about 40%2
in November 2015 to more than 600%2 in January 2009. Second, fitting the
unconditional level of the VP well does not imply that the model does also a
good job fitting each period’s VP.
Our analysis focuses on time variation in equity market volatility, loss
aversion, and the degree of probability weighting. To guide the intuition for
our subsequent results, we start by exploring the impact of the level of equity
volatility on the variance premium in a comparative statics exercise.
3.1 Comparative statics: The impact of volatility

In the unconditional analysis of Subsection 2.4, we set volatility equal to the
average historical 1-month volatility of the S&P 500 of 4.84% per month. In
the data, however, volatility varies substantially over time.17 Because volatility
widens the tails of the investor’s wealth distribution, both expected option
returns and the variance premium strongly increase with volatility. This is
illustrated in Figure 5, which shows model-implied VPs as a function of
probability distortion for a low (2%), average (4.84%), and high (10%) level
of monthly volatility.
The figure shows that the model-implied VP increases in volatility as long as
there is some probability weighting. Probability weighting matters more for the
VP when volatility is higher. This pattern is intuitive, as higher volatility leads
17 Over our sample, realized monthly volatility ranges from a low of 1.18% to a high of 23.33%. Its interquartile
range is 3.13%–5.57%.
3700

to more extreme outcomes, which are then overweighted by CPT investors. The
figure also shows that high volatility alone is not sufficient to generate a VP:
only when coupled with probability weighting does higher volatility lead to a
higher VP. In sum, Figure 5 suggests that volatility and probability weighting
interact in generating a VP.
3.2 Time-varying CPT setting

We now explore whether allowing for time variation in volatility and probability

weighting, enables the CPT model to also match the large variation in the
variance premium and option returns that we observe over our sample period.
We start by calculating the conditional monthly equity market variance, ht =
Vt [rt+1 ], as the integrated variance based on a simple GARCH(1,1) model
estimated on daily S&P 500 index returns rt+1 over the 1996 to March 2016
period. The estimated GARCH(1,1) specification is
ht+1 = 1.85e −06+0.89ht +0.09(rt −μ)2 , (16)
where rt represents daily S&P 500 index returns. All estimated parameters are
statistically significant at the 1% confidence level. Consistent with previous
evidence, we find the daily conditional variance process to be highly persistent
and stationary. To calculate the conditional variance for the next month, we
take the sum of the model-implied predictions of the variance for all days over
the next month (“integrated variance”).18
The existing literature on probability weighting offers little to no guidance
on how to properly specify time variation in probability weighting. We consider
two specifications. The first one makes the end-of-month distortion parameter
ct a function of a dummy Ic,t that equals 1 when the return on the S&P 500 over
month t belongs to the lowest or highest quartile of returns over our sample
period, and zero otherwise:
ct = ĉ +κc Ic,t , (17)
where ĉ is equal to the unconditional estimate of 0.65 from Table 4. Intuitively,

for κc < 0, this specification captures that investors focus more on extreme
returns, on either side of the distribution, when they have recently experienced
extreme returns, which may be due to future extreme returns becoming more
salient in periods with extreme returns (e.g., Bordalo, Gennaioli, and Shleifer
2012). We also estimate a more flexible specification that makes left (c1 ) and
right (c2 ) tail distortion a function of dummies that are one when the return on
18 Although we believe that the integrated variance from a GARCH process is a good approximation of the actual
variance, in unreported results (available on request), we show that the results in this section are robust to
alternative conditional variance methods, such as an exponentially weighted moving average or, even more
simply, last month’s realized variance.
3701

up
dn
the S&P 500 belongs either to the top (Ic,t ) or bottom (Ic,t ) quartile of returns
over our sample period:
up
c1,t = cˆ1 +κcup I +κcdn
1 c,t
I dn ,
1 c,t
up
c2,t = cˆ2 +κcup I +κcdn
2 c,t
I dn .
2 c,t
(18)
Here, cˆ1 and cˆ2 are the unconditional estimates of 0.62 and 0.69, respectively,
from Table 4. Finally, closely following BHS, we also allow for time variation
in loss aversion through a slope parameter κλ on a state variable zt capturing
prior gains and losses (see Appendix G for details on the construction).19

3.3 Calibration method and results
To calibrate the parameters in the time-varying setting, we extend the moments
in the GMM procedure in Section 2.4 to a time-varying setting as follows. At
each point in time t, we compare the model-implied equity and VPs to empirical
estimates of the time-t conditional equity and VPs. The conditional VP at any
time t is estimated as the difference between the square of the VIX at time t
and the time-t expectation of the realized variance.20
To obtain an estimate for the conditional equity premium, we perform a
predictive regression of monthly S&P 500 returns on the price-dividend ratio
and the VP.21 We use the fitted values of this regression to obtain the empirically
observed conditional equity premium at each point in time. We use the same
method for modeling the conditional expectations of the option returns. To
reduce the computational burden implied by the time-varying setting, we use
monthly instead of daily data, and consider only 10 deltas for the options—5 for
puts and 5 for calls—instead of 26. We obtain a panel of errors by comparing,
at each point in time, the model-implied with the empirically observed values
for the equity premium, the VP, and the option returns. The conditions used to
calibrate the parameters are thus given by
⎡ E
⎤
μE,t (κc ,κλ )−Et (Rt+1 )
⎢ V Pt (κc ,κλ )−Et (V Pt+1 ) ⎥
gt (κc ,κλ ) = ⎢
⎣ μp,t (κc ,κλ )−Et (Rt+1
⎥,
) ⎦
p
c
μc,t (κc ,κλ )−Et (Rt+1 )
where μp,t (κc ,κλ ) and μc,t (κc ,κλ ) are both 5×1 vectors with expected put
and call returns, respectively. All parameters in the model are fixed at their
19 We have also estimated an alternative specification which models time variation in probability weighting
parameters as a function of the same variable zt that we use to model time-varying loss aversion. We find
results which are in line with the ones presented above. Also, we find similar results when we use dummies that
equal 1 when S&P 500 returns are in the top and/or bottom decile, rather than quartile.
20 Because we compare the conditional moments implied by the model with conditional moments obtained from
regressing equity and option returns on predicting variables, and we do not take unconditional expectations of
these conditional moments, our estimation method no longer fits into the GMM framework. The approach in this
section is thus one of calibration. The goal is to see whether our CPT model can capture the time variation in
risk premiums of equity and options, as implied by a data-driven regression model.
21 The work of Bollerslev, Tauchen, and Zhou (2009) provide evidence that the VP has predictive power for equity
returns.
3702

Table 5
Parameter calibration in the time-varying setting
A. Parameter estimates for time-varying specifications
Parameter (1) (2) (3) (4)
κλ 2.42
κc −0.03
up
κc1 −0.02
dn
κc1 0.06
up
κc2 0.06
dn
κc2 −0.05

Avg. value function 116.13 114.53 114.06 114.56
B. Return predictability
Model-implied VP
Coefficient Observed VP (1) (2) (3) (4)
βV P 11.22 5.19 5.36 5.04 5.17
[2.87] [1.97] [2.03] [1.83] [1.84]
βV P +βVGF
P
C 62.12 1.34 1.48 1.17 1.23
[4.21] [0.29] [0.30] [0.24] [0.25]
Panel A shows the estimated parameters and model fit for alternative specifications of the time-varying CPT
setting. For every specification, we report the value function, divided by the number of months (Avg. VF). The
value function is calculated as a weighted average of the squared residuals for the variance and equity premiums
as well as for the 10 options considered in the time-varying setting. In specification (1), we keep all parameters
at their benchmark values, but allow for time-varying volatility. In specification (2), we vary loss aversion over
time according to the specification in Appendix G. In specifications (3) and (4), probability weighting varies over
time according to Equations (17) and (18), respectively. In panel B, to assess the ability of the model-implied
variance premium to predict equity returns, for each specification, we report the coefficients associated with the
variance premium in the following regression:

rt+1 = α0 +αGFC IGFC + βVP +βGFC IGFC VPt +βPD DYt + t+1 ,
where IGF C is a dummy variable that takes the value of 1 for each month between October 2008 and March
2009. For brevity, we only report the two components of the coefficient on V Pt . t-statistics for these coefficients
are reported in brackets.
unconditional GMM estimates except for κc and κλ , which we calibrate by

minimizing the sum of squared errors in the above moment conditions. For
the probability weighting specification in Equation (18), we replace κc by κc1
and κc2 . We sum across assets and over time, using the same weighting of the
equity premium, the VP, and put and call returns, as we did in Section 2.4.
To avoid numerical problems, we use a grid of values for the different κ’s to
identify the values that minimize the weighted sum of the squared errors for
each specification.
Table 5, panel A, presents results for alternative specifications of the time-
varying CPT setting. For each specification, we report parameter estimates as
well as the level of the value function, divided by the number of months. Our
point of departure is a restricted model in which only the equity return volatility
is time varying, while the CPT parameters are fixed at their unconditional
estimates. That is, we set κc = κλ = 0. This model isolates the impact of time
variation in volatility on the fit of the model.
3703

Figure 6 compares the model-implied VP and equity premium of this model

with their observed counterparts. Panel A shows that this simplest version of
the conditional CPT model fits the level and dynamics of the variance premium
remarkably well, although it overestimates the variance premium in late 2008,
around the collapse of Lehman Brothers, because the equity return volatility is
unusually elevated during this period. While not our primary focus, the model
also provides a reasonable fit of the equity premium (panel B), although it misses
some of its larger movements. In sum, when we keep all preference parameters
constant over time, a CPT model with time-varying equity volatility generates a

remarkably good fit of the time-series variation in the VP, despite its simplicity.
Turning to the setting with time-varying loss aversion (specification (2) in
Table 5, panel A), we find that our estimate of κλ is positive. This estimate
supports one of the key assumptions in BHS, namely that loss aversion increases
after losses. The estimate for κλ generates a range in loss aversion, from about
1.28 to 1.88. However, this variation has only a small effect on the model fit, as
can be seen from the small improvement in the value function. This is intuitive,
because our previous results have shown that loss aversion does not have a large
impact on the VP.
Turning to the specifications with time-varying probability weighting,
specification (3) shows a negative value for κc , which, as argued above, is
consistent with the intuition that future extreme returns become more salient
for investors in periods with extreme returns. The results from the more flexible
specification (4) imply that distortion in the left tail decreases following losses
but increases following gains, and vice versa for the right tail of the distribution.
Interestingly, these results are less consistent with a model in which recent
extreme losses induce decision makers to focus more on losses, and recent
extreme gains induce decision makers to focus more on gains. They are instead
more consistent with a “gambling for resurrection” interpretation, in which
experiencing periods of particularly bad outcomes increases the willingness of
investors to take a long shot. On the upside they are consistent with investors
who are concerned about hedging their recent gains. While our results in Table
5, panel A, are by no means a comprehensive examination of how to best model
probability weighting in a time-varying fashion, we nevertheless view them as
informative for future research on how to model time variation, especially
because we currently know little about how probability weighting varies over
time.
Overall, we conclude that making probability weighting depend on recent
extreme returns improves the fit of the model, in the sense of a lower value
function, but that the effects are quantitatively small. In addition, the estimated
parameters in specifications (3) and (4) do not lead to any drastic swings in
the implied levels of probability weighting. The results in specification (3)
imply that the distortion parameter ct varies within the narrow range of 0.65
to 0.68. In specification (4), the ranges are somewhat larger but still relatively
narrow in economic terms: c1,t varies between 0.60 and 0.68 and c2,t varies
3704

A Variance risk premium

2000
Observed
CPT-implied
1500
%
1000

500
0
Mar-96 Sep-98 Mar-01 Sep-03 Mar-06 Sep-08 Mar-11 Sep-13 Mar-16
B Equity premium
6
4
%
-1
Mar-96 Sep-98 Mar-01 Sep-03 Mar-06 Sep-08 Mar-11 Sep-13 Mar-16
Figure 6
Observed versus CPT-implied equity and variance premiums
This figure compares the observed equity and variance risk premiums with those implied by a CPT setting in
which the model’s CPT parameters are constant and the stock return volatility is time varying (specification (1)
in Table 5). The variance premium, in panel A, is calculated as the difference between the square of the VIX and
the expected realized variance (see Section 2.4). The expected realized variance is calculated as an in-sample
forecast of the stock return variance using the lagged stock return variance and the square of the VIX. In panel
B, the observed equity premium is calculated as an in-sample forecast of equity returns from two predictors, the
price-dividend ratio and the variance risk premium.
between 0.65 and 0.75. Given that the models in specifications (3) and (4)
are computationally intensive, and given that specifying the functional form
of time-varying probability weights provides additional degrees of freedom to
the model, it is not clear that the small improvements in the value functions
warrant making the probability weighting parameter time varying. In our view,
3705

a key insight from this section is that incorporating exogenous time-varying

volatility allows our simple model to generate a remarkably good empirical fit
to the data even if we restrict ourselves to a specification in which probability
weighting parameters are fixed and not time varying.
3.4 Model-implied variance premium and return predictability

Finally, we investigate whether our model-implied VP has predictive power for
future stock returns. To that end, we run the following regression:

rt+1 = (α0 +αGFC IGFC )+(βVP +βGFC IGFC )VPt +βPD DYt + t+1 , (19)
with rt+1 next month’s market return, V Pt the current model-implied VP, and
DYt the observed dividend yield. Because both the observed and model-implied
VPs are extremely high and volatile during the global financial crisis (GFC),
we also include an interaction term between the VP and a dummy for the GFC,
that is, IGF C = 1 from October 2008 to March 2009, and zero otherwise.22
Table 5, panel B, presents results. To have a meaningful benchmark,
the left-most column reports coefficients (with corresponding t-statistics in
parentheses) from a regression of 1-month- ahead returns on the observed VP
and the dividend yield. Consistent with the previous literature, we find a positive
significant relation between the VP and future stock returns; an increase in VP
is followed by an increase in stock returns. The second column (specification
(1)) reports estimates from a specification that includes the VP implied by our
benchmark model with constant loss aversion, constant probability distortion,
and using our estimates of conditional market volatility. We find model-implied
VPs to be significantly positively related to future returns outside the GFC.
Inside the GFC, the point estimate on the model-implied VP is still positive,
but statistically insignificant.
We find very similar predictive patterns when model-implied variance pre-
miums are generated using models featuring either time-varying loss aversion
(specification (2)) or time-varying distortion parameters (specifications (3) and
(4)). Empirically, the reduced predictive power around the GFC for model-
implied VPs may be because (1) we are estimating the coefficient from 6 data
points only and (2) the observed relation between the VP and subsequent returns
during the crisis may be influenced by a host of factors (such as policy responses
to the crisis), which are crisis specific, and therefore less likely to be captured
by a preference-based model.
Overall, we conclude that our model does a good job in qualitatively
generating the observed positive relation between the VP and subsequent
22 We calculate the observed VP as the difference between the square of the VIX and expected return variance on
the S&P 500 for the next month, following the existing literature. During crisis periods, though, the VIX estimate
may be affected by the poor liquidity of the options it is calculated from. Similarly, expected return variance,
which is based on a predictive regression relating next month’s variance to the current month’s variance and VIX,
is likely estimated with very large confidence bounds during the GFC. These same measurement issues in turn
also affect our estimate of the model-implied VP.
3706

returns. Quantitatively, future returns in our model are not as sensitive to the
VP as they are in the data, especially during the crisis months.
4. Conclusion
This paper investigates the potential of prospect theory to explain the puzzlingly
low average returns on both out-of-the-money put and out-of-the-money call
options on the S&P 500 index, and the level and the dynamics of the variance
premium. We build on the representative investor model of Barberis, Huang,

and Santos (2001), in which preferences are the sum of a CRRA utility function
and a prospect theory value function over gains and losses, by incorporating
probability weighting, an integral part of cumulative prospect theory (Tversky
and Kahneman 1992).
A central finding of our paper is that, for plausible parametrizations, a CPT
model with probability weighting can match observed put and call returns well,
and generate a variance premium similar to that observed in the data. In our
benchmark specification, when we estimate the preference parameters from our
theoretical model using GMM on S&P 500 equity index and option returns from
1996 to 2016, we obtain an estimate for the probability weighting parameter
of 0.65, which is remarkably close to standard values found in experimental
studies and used in the finance and economics literature. The CPT model thus
provides a unifying explanation for two well-known option pricing puzzles (the
low returns on OTM puts and the low returns on OTM calls), as well as for
the variance premium puzzle, by incorporating only one additional modeling
ingredient: probability weighting.
Appendices
A Proof of Proposition 2
Without loss of generality, we focus on one risky asset (equity).23 The investor’s problem is now
given by
−1
T
max V= max E[U (WT )]+ bt CPT (Xt+1 ),
α0E ,...,αTE−1 α0E ,...,αTE−1 t=0
where, as before, we follow Barberis, Huang, and Santos (2001) and let the weight on the prospect
theory utilities be given by a multiple of current period’s marginal utility of wealth, that is, by
−γ
bt = b̂Wt . The investor’s budget constraint is
Wt+1 = Wt (αtE rt+1 +R f ), t = 0,...,T −1,

E −R f denotes the excess return from period t to period t +1. The portfolio gain
where rt+1 := Rt+1
or loss each period is thus given by
Xt+1 = Wt+1 −Wt R f = Wt αtE rt+1 , t = 0,...,T −1.
23 Not introducing d = 1,...,D derivatives (with portfolio shares α d for derivative d in period t ) simplifies the
t
notation in the following derivation significantly and does not cause any loss of economic insight.
3707

Let us first consider the log-utility case (γ = 1), whose proof is simple and intuitive. In that case,
the investor seeks to maximize
−1
T
−1 E
V = E[log((W0 Tt=0 (αt rt+1 +R f ))]+ bt CPT (Wt αtE rt+1 )
t=0
−1
T −1
T
= E[logW0 + log(αtE rt+1 +R f )]+ b̂Wt−1 CPT (Wt αtE rt+1 ).
t=0 t=0
Because the CPT functional is homogeneous of degree 1, it follows that CPT (Wt αtE r t+1 ) =

Wt αtE CPT (r t+1 ). For this reason, the decision variables αtE are no longer nested within the CPT
functional. It follows that
−1
T −1
T
V = logW0 + E[log(αtE rt+1 +R f )]+ αtE b̂CPT (rt+1 )
t=0 t=0
−1
T
= logW0 + E[log(αtE rt+1 +R f )]+αtE b̂CPT (rt+1 ) .
t=0
The objective functional is thus additively separable in the decision variables so that the FOC for
each αtE is independent of the αsE , s = t and structurally identical. In particular, the weights αtE
do not depend on the period t itself or on the number of periods remaining—the myopia result.
When returns are iid, we further have that α0E = ... = αTE−1, so that the investor does not rebalance
her portfolio.
Now consider the case of general CRRA utility, which we prove by dynamic programming. Let
τ denote the number of periods that remain from current time t so that T = t +τ . The value of the
agent’s problem at time t is given by
1−γ t+τ −1
Wt+τ
V (τ,Wt ) = max Et + bs CPT (rs+1 αsE Ws )
{α E }t+τ −1
s s=t
1−γ s=t
−1

1−γ t+τ
Wt+τ
= max Et + b̂αsE Ws1−γ CPT (rs+1 ) ,
{αsE }t+τ −1 1−γ
s=t s=t
where, similarly to the log case, we exploit the fact that the CPT preference functional is
homogeneous of degree 1. The problem has become one of maximizing terminal wealth expected
utility plus a sum of flow utilities, which are linear in the decision variable. In particular, letting
1−γ
ct+1 := b̂CPT (rt+1 ), the time-t value of the next period’s flow utility is given by πt+1 := ct+1 αtE Wt .
By the dynamic programming principle, the investor’s problem at time t can be written as
⎡ ⎤
⎢ ⎥
⎢ 1−γ t+τ −1 ⎥
⎢ ⎥
⎢ Wt+τ 1−γ ⎥
V (τ,Wt ) = max Et ⎢ max Et+1 + cs+1 αsE Ws1−γ + αtE ct+1 Wt ⎥,
αtE ⎢{α E }t+τ −1 1−γ ⎥
⎢ s s=t+1 s=t+1 ⎥
⎣ πt+1 (αtE ,Wt ,ct+1 )⎦
V (τ −1,Wt (αtE rt+1 +R f ))
1−γ
t+τ +c W 1−γ
subject to the terminal condition V (0,Wt+τ ) = 1−γ t+τ αt+τ −1 Wt+τ −1 . This means that τ periods
before the model ends, the value of the problem is determined by the problem’s continuation
value next period, V (τ −1,Wt (αtE rt+1 +R f )), and next period’s flow utility, πt+1 (αtE ,Wt ,ct+1 ).
The notation indicates that the continuation value is that of τ −1 periods before the end, where
3708

wealth from the previous period has changed to Wt (αtE rt+1 +R f ). This continuation value includes
the τ −2 expected flow utilities on from two periods in the future. Note that the value of the next
period’s flow utility (i.e., πt+1 ) is deterministic, because it is a function of the current time t portfolio
weight and current period wealth. Continuing,
⎡ t+τ −1
(Wt (αsE rs+1 +R f ))1−γ
V (τ,Wt ) = max Et ⎣ max Et+1 s=t
αtE {αsE }t+τ −1 1−γ
s=t+1

⎤

t+τ −1
+αtE ct+1 Wt ⎦
1−γ
+ cs+1 αsE Ws1−γ
s=t+1
⎡
(Wt (αtE rr+1 +R f ))1−γ
= max Et ⎣αtE ct+1 Wt
1−γ
+ ·
αtE 1−γ
⎡ "1−γ ⎤⎤
!
t+τ −1
t+τ −1 1−γ
cs+1 αsE Ws (1−γ ) ⎦⎦
max Et+1 ⎣ (αsE rs+1 +R f ) + 1−γ
t+τ −1
{αsE }s=t+1 Wt+1
s=t+1 s=t+1
⎡
(αtE rt+1 +R f )1−γ
= max Wt Et ⎣αtE ct+1 +
1−γ
·
αtE 1−γ
⎡ "1−γ ⎤⎤
!
t+τ −1
t+τ −1 1−γ
cs+1 αsE Ws (1−γ ) ⎦⎦
max Et+1 ⎣ (αsE rs+1 +R f ) + 1−γ
.
t+τ −1
{αsE }s=t+1 Wt+1
s=t+1 s=t+1
With some standard manipulations, one can show that

t+τ −1 1−γ
t+τ −1 !
s
cs+1 αsE Ws (1−γ )
1−γ
= (1−γ ) cs+1 αsE (αs̃E rs̃+1 +R f )1−γ ,
s=t+1 Wt+1 s=t+1 s̃=t+2
t+1
where we use the convention that E f 1−γ
s̃=t+2 (αs̃ r s̃+1 +R ) ≡ 1. Therefore, we obtain

1−γ (αtE rt+1 +R f )1−γ
V (τ,Wt ) = max Wt Et αtE ct+1 + ·
αtE 1−γ
t+τ −1 −1
⎤
!
t+τ !
s
max Et+1 (αsE rs+1 +R f )1−γ +(1−γ ) cs+1 αsE (αs̃E rs̃+1 +R f )1−γ ⎦,
{αsE }t+τ −1
s=t+1 s=t+1 s=t+1 s̃=t+2
3709

and see that the problem is homogeneous in current wealth so that, without loss of generality, we
let Wt ≡ 1. Thus,
⎡
⎢
⎢
⎢ E (α E rt+1 +R f )1−γ
V (τ,Wt ) = max Et ⎢
⎢ αt ct+1 + t · (A1)
αtE ⎢ 1−γ

⎣ πt+1
=:ut+1
t+τ −1 ⎥
⎥

! −1 !
E f 1−γ
t+τ
E
s
E
⎥
f 1−γ ⎥
max Et+1 (αs rs+1 +R ) +(1−γ ) cs+1 αs (αs̃ rs̃+1 +R ) ⎥.
{αsE }t+τ −1 ⎥
⎦
s=t+1 s=t+1 s=t+1 s̃=t+2

=:ψt+1 (τ −1)
Note that the portfolio weight αtE

depends on the investment horizon τ through the next-period
“horizon-effect” term ψt+1 (τ −1), which we now study in more detail. First, by applying some of
the previous transformations in the opposite direction, observe that
t+τ −1
!
ψt (τ ) = max Et+1 (αsE rs+1 +R f )1−γ
{αsE }t+τ −1
s=t s=t
1−γ
−1
=1 =1

t+τ !
s
1−γ
+(1−γ ) cs+1 αsE (αs̃E rs̃+1 +R f )1−γ · Wt
s=t s̃=t+1
−1

ψt (τ ) Wt+τ
1−γ t+τ
⇐⇒ = max Et+1 + cs αsE Ws1−γ
1−γ {α E }t+τ −1 1−γ s=t
s s=t
1−γ t+τ −1
ψt (τ ) Wt+τ
⇐⇒ = max Et+1 + cs αsE Ws1−γ ≡ V (τ,Wt ).
1−γ {α E }t+τ −1 1−γ s=t
s s=t
ψt (τ )
Therefore, we can substitute V (τ,Wt ) in Equation (A1) with to obtain the recursion
1−γ

ψt (τ ) = max Et αtE ct+1 (1−γ )+(αtE rt+1 +R f )1−γ ·ψt+1 (τ −1) .
αtE
We see that when γ = 1, the optimal portfolio weight αtE depends on the investor’s horizon τ, that is,
is nonmyopic. We now show that even when returns are independent (and/or identically distributed),
investment is nonmyopic. When the per-period returns are independent, because ψt+1 (τ −1) is free
of rt+1 , the maximizing αtE satisfies

ct+1 (1−γ )+Et ψt+1 (τ −1) ·Et (αtE rt+1 +R f )−γ rt+1 = 0.
We see now that if, and only if, γ = 1, we can divide by Et [ψt+1 (τ −1)] so that each αtE satisfies

Et (αtE rt+1 +R f )−γ rt+1 = 0,
which says that αtE is horizon independent.24 For all other CRRA utilities, we have shown that the
αtE are time and horizon dependent.
24 Note, however, that the proof for general γ does not nest the proof for γ = 1, because it involves expressions with
1−γ in the denominator.
3710

B Proof of Proposition 4
Without loss of generality, we work in a setting with an infinite number of outcomes for the equity
payoff x E > 0 at time T . As shown by, for example, Bakshi, Kapadia, and Madan (2003) and Carr
and Madan (2001), any continuous and twice-differentiable payoff function G(x E ) can be written
as
∞
G(x E ) = G(x̄)+G (x̄)(x E − x̄)+ G (K)max(x E −K,0)dK
x̄
x̄

+ G (K)max(K −x E ,0)dK, (B.1)
0
for any value of x̄. This equation shows that any payoff function can be replicated by an equity
position and a position in an infinite number of call and put options with different strike prices.
If there are no arbitrage opportunities, a risk-neutral measure exists and the risk-neutral expected
value of the payoff function G can be written as a function of call and put prices
∞ x̄

E Q G(x E ) = G(x̄)+G (x̄)(R f s0E − x̄)+ G (K)R f C(K)dK + G (K)R f P (K)dK, (B.2)
x̄ 0

where we use the risk-neutral pricing equations for equity, s0E = 1f E Q x E , call options
R
with strike K, C(K) = 1f E Q max(x E −K,0) , and put options with strike K, P (K) =
1
R
Q max(K −x E ,0) . Similar to Equation (B.2) for the risk-neutral expectation, we can
Rf E
compute the expected value of the payoff function G under the physical measure P as

E P G(x E ) = G(x̄)+G (x̄)(E[x E ]− x̄) (B.3)
∞ x̄

+ G (K)E max(x E −K,0) dK + G (K)E max(K −x E ,0) dK.
x̄ 0
# # $$2
E
Now choose G(x E ) = ln xE , and by subtracting (B.3) from (B.2) it directly follows that
s0
⎡ ""2 ⎤ ⎡ ""2 ⎤ "
xiE xiE E
E Q⎣
ln ⎦ −E ⎣ ln
P ⎦ = −G (x̄)s0E E[x ] −R f (B.4)
s0E s0E E
s0
∞ # $
E[max(x E −K,0)]
− G (K)C(K) −R f dK
C(K)
x̄
x̄ # $
E[max(K −x E ,0)]
− G (K)P (K) −R f dK,
P (K)
0
# $ # # $$
x 1 2 x
with G (x) = 2ln and G (x) = 1−ln . We consider the approximation at x̄ = s0E ,
s0E x x2 s0E
so that G (x̄) = 0. Defining

R p (K) = E max(K −x E ,0) /P (K)
3711

and

R c (K) = E max(x E −K,0) /C(K),
and defining weights
v c (K) = −G (K)C(K)
and
v p (K) = −G (K)P (K),
then yields the right-hand side of Equation (14).
To see that the right-hand side of Equation (14) is indeed an approximation to the variance

premium, note that adding the term
2 2
E P ln R E −E Q ln R E
to both sides of Equation (B.4) above yields the variance premium exactly, as can be seen by
comparing the left-hand side of the resulting equation with the definition of the variance premium
in Equation (12). The term arises because of the difference between central and noncentral second
moments. For our application, the term is numerically small as we focus on 1-month returns: it is
equal to 1.3 in our benchmark model, while empirically the variance premium is equal to 146.7.
Equation (14) omits the term, which is why it becomes an approximation to the variance premium.
Finally, note that due to our choice of x̄ = s0E , the put strikes in Equation (14) satisfy K ≤ s0E
while the call strikes satisfy K ≥ s0E . In other words, Equation (14) only refers to returns of put
and call options that are OTM. Moreover, note that G (K) ≤ 0 if and only if K ≤ s0E e1 ≈ 2.71s0E .
It thus follows that the weights v c (K) and v P (K) are negative for all put options as well as for
call options with strikes below 2.71s0E ; that is, the weights are negative for all empirically relevant
OTM options we consider.
C Upside and Downside Variance Premiums

In this appendix, we discuss how the variance premium can be decomposed as the sum of an upside
and a downside variance premium. In addition, we use option prices and realized returns to estimate
these upside and downside variance premiums and to assess how well the CPT model fits these
premiums.
We start with an extension of Proposition 4, where we formally decompose the variance premium
into upside and downside components.
Proposition 5. Let R c (K) and R p (K) denote the return on a call and put option with strike K,
respectively. Define equity index returns as R E = x E /s0E . If there are no arbitrage opportunities,
we can define and calculate downside and upside variance premiums as follows:
E
s0
2 2
V Pd ≡ E Q
ln R E 1R E <1 −E P ln R E 1R E <1 = v p (K)E P R p (K)−R f dK, (C.1)
0
∞
2 2
V Pu ≡ E Q ln R E 1R E >1 −E P ln R E 1R E >1 = v c (K)E P R c (K)−R f dK, (C.2)
s0E
where 1R E >1 is the indicator function which equals 1 if the equity return is above 1, and v c (K)
and v p (K) are scalars whose expressions are given in Appendix B. The total variance premium in
Proposition 4 can then be decomposed as
V P ≈ V Pu +V Pd . (C.3)
3712

Proof. The proof of Proposition 5 follows the same steps used in Appendix B, where we add
indicator functions for the final stock price being below or above the current price. For the sake of
completeness, we go through these steps here. Multiplying Equation (B.1) by the indicator function
1x<x̄ ,
G(x E )1x E <x̄ = G(x̄)1x E <x̄ +G (x̄)(x E − x̄)1x E <x̄ + (C.4)
∞ x̄
1x E <x̄ G (K)max(x −K,0)dK +1x E <x̄ G (K)max(K −x E ,0)dK,
E
x̄ 0
(C.5)

which simplifies to
x̄
G(x E )1x E <x̄ = G(x̄)1x E <x̄ +G (x̄)(x E − x̄)1x E <x̄ + G (K)max(K −x E ,0)dK. (C.6)
0
If there are no arbitrage opportunities, a risk-neutral measure exists and the risk-neutral expected
value of the payoff function G can be written as a function of call and put prices, as follows:
x̄

E Q G(x E )1x E <x̄ = G(x̄)E Q 1x E <x̄ +G (x̄)E Q (x E − x̄)1x E <x̄ + G (K)R f P (K)dK,
0
(C.7)
where we use the risk-neutral pricing equations for put options with strike K, P (K) =
# # $$2
1

Q max(K −x E ,0) . Now choose G(x E ) = ln x E
f E E and x̄ = s0E , so that G(x̄) = 0 and
R s0
G (x̄) = 0. We then obtain that
⎡ ""2 ⎤ E
s0
xE
E Q ⎣ ln E 1x E <s E ⎦ = G (K)R f P (K)dK, (C.8)
s0 0
0
which implies that the risk-neutral downside second moment is linked to OTM put option prices
only.
In a similar way, we can compute the expected value of the payoff function G under the physical
measure P as (filling in the choices for G and x̄)
⎡ ""2 ⎤ E
s0
xE
E ⎣ ln E
P
1x E <s E ⎦ = G (K)E P max(K −x E ,0) dK. (C.9)
s0 0
0
Because R p (K) = max(K −x E ,0)/P (K), for v p (K) = −G (K)P (K), we obtain
⎡ ""2 ⎤ ⎡ ""2 ⎤
x E x E
V Pd = E Q ⎣ ln E 1x E <s E ⎦ −E P ⎣ ln E 1x E <s E ⎦
s0 0 s0 0
E
s0

= v p (K)E P R p (K)−R f dK. (C.10)
0
3713

Analogously, for the upside variance premium we obtain

⎡ ""2 ⎤ ⎡ ""2 ⎤
Q⎣ xE ⎦ P⎣ xE
V Pu = E ln E 1x E >s E −E ln E 1x E >s E ⎦
s0 0 s0 0
∞

= v c (K)E P R c (K)−R f dK. (C.11)
s0E
Given Proposition 4, Proposition 5 then directly follows.

Not that, in Proposition 5, we defined the upside and downside variance to be the noncentral
second moment of the stock return, conditional upon the stock price at maturity being above or
below the current stock price. Then the downside and upside variance premiums add up to become
the total variance premiums, as in Proposition 4. In summary, we can decompose the variance
premium into upside and downside components, and both components can directly be calculated
from the expected returns and prices of puts and calls.
How does our decomposition into upside and downside variance compare to that in the literature?
While Held, Kapraun, Omachel, and Thimme (2018) do not provide results for option returns, they
calculate risk-neutral noncentral second moments in the same way as we do. Our approach is
also similar to Andersen and Bondarenko (2009), who define upside and downside variance in a
continuous-time setting by conditioning on whether the stock price is above or below a barrier at
each point in time. However, in some other studies, the conditioning set for the second moment
under Q differs from the conditioning set used under P . This blurs the comparison of the P and Q
moments. This is an issue for the semivariance moments in Held, Kapraun, Omachel, and Thimme
(2018). Also, Kilic and Shaliastovich (2017) use the same upside and downside measures under
the risk-neutral measure as we do, but under the actual measure P they calculate the realized
variance using intraday returns, where for the downside variance they sum all (squared) negative
intraday returns, and for the upside variance they sum all squared positive intraday returns. Hence,
their P -moment conditions on a different set of events compared to the Q-moment, for which
the condition is whether the stock price at maturity is above or below the current price. Feunou,
Jahan-Parvar, and Okou (2017) calculate the P -moments in a similar way as Kilic and Shaliastovich
(2017), while they use the results of Andersen and Bondarenko (2009) to calculate the Q-moments.
Hence, in their case, there is also a difference in the conditioning set when comparing Q-moments
and P -moments. These differences make it hard to compare our results to those in these articles.
We now discuss the empirical results for the upside and downside variance premiums. We start by
calculating empirical estimates of V Pu and V Pd . We do this by using Equations (C.10) and (C.11)
above. For each day in our sample, we use observed option prices to calculate the weights v c (K)
and v p (K), calculate realized option returns across strikes and thus compute sample counterparts
of Equations (C.10) and (C.11).25 We then average these across all days in the sample to obtain the
full-sample V Pu and V Pd . We obtain a value of 118.5 for V Pd and 27.5 for V Pu .26 At first sight,
the estimates for V Pu and V Pd may lead one to conclude that downside risk contributes much more
to the total variance premium. However, to place the estimates for V Pu and V Pd into perspective,
it is useful to think of a benchmark model with CRRA utility and lognormal returns. As discussed
25 For strike levels beyond the lowest and highest strike in our sample, we assume option returns are equal to the
return on the option with the closest strike, and use Black-Scholes option prices to calculate the weights.
26 These add up to 146.0, which is very close to the estimated variance premium of 146.7 that we use in the paper.
As explained in Section 2.1, this latter estimate of the variance premium is calculated as the difference between
the squared VIX and the expected realized variance. Our Proposition 4 shows how the variance premium can
be calculated using either the difference in variances or using option returns, and we thus find numerically very
similar variance premiums using both methods.
3714

in Section 1.4, in that case the total variance premium is equal to zero. However, it is easy to see
that, in this CRRA case, V Pu < 0 and V Pd > 0. This is because, as shown in Figure 1, in such a
CRRA model, OTM calls have positive expected returns and OTM puts have negative expected
returns, and from Equations (C.10) and (C.11), it then directly follows that V Pu < 0 and V Pd > 0
(because the weights are negative). For example, for the pure CRRA model used in Table 2, we find
V Pu = −17.7 and V Pd = 16.4.27 Hence, the positive upside variance premium that we observe in
the data contradicts substantially with the negative V Pu implied by the standard CRRA-lognormal
model.
Our empirical results for the upside and downside variance premiums for the S&P 500 index are
similar to those of Andersen and Bondarenko (2009) and Held, Kapraun, Omachel, and Thimme

(2018) (for the noncentral moments in their Table 2). They also find that, both on the upside
and downside, the risk-neutral second moments are higher than the actual counterparts. Kilic and
Shaliastovich (2017) and Feunou, Jahan-Parvar, and Okou (2017) document results different from
ours, and, as discussed above, comparing our results to theirs is difficult to do.
Finally, we assess how well the CPT model fits the upside and downside variance premiums.
Table E.2 presents the results. We see that, for our benchmark model with lognormal returns, the
upside variance premium implied by the model is somewhat larger than in the data, while the
downside premium is lower than in the data. Because in Figure 1 we show that the CPT model
fits expected option returns very well for both puts and calls, it must be that the model-implied
weights in Equations (C.10) and (C.11) are different from the observed weights. We confirm this in
unreported results. The model-implied weights depend on option price levels. Hence, even though
the CPT model fits option risk premiums very well, the fit for option price levels is somewhat less
good. This is a direct consequence of our assumption that returns follow a lognormal distribution.
In particular, because this distribution misses the negative skewness observed in realized S&P 500
index returns, the model underestimates put option prices and overestimates call option prices,
which directly affects the weights in Equations (C.10) and (C.11). However, Table E.2 also shows
that, once we use a skewed-t distribution for the underlying returns, this issue is resolved. In that
case, the model fits the upside and downside variance premiums very well. As discussed in Section
2.4.3, using the skewed-t distribution does not change any of the other conclusions in our paper.
D Gentzkow and Shapiro (2015) Sensitivities

To assess how well the different parameters in our CPT model are identified, we calculate the
Gentzkow and Shapiro (2015) sensitivities, which capture the sensitivity of the parameter estimates
to the moment conditions. Formally, these sensitivities are defined as the coefficients of a regression
of the estimator θ̂ on the moment conditions ĝ. In a GMM setting, Gentzkow and Shapiro (2015)
show that the sensitivities Λ are given by
Λ = −(G W G)−1 G W, (D.1)
where G = δg/δθ and W is the GMM weighting matrix. Intuitively, the sensitivities measure how
a change in a given moment condition affects the estimate of a parameter and thus capture how
informative each moment condition is about a parameter. In addition, if two parameters have very
similar sensitivities, it will be hard to separately identify these parameters from the given set of
moment conditions. Gentzkow and Shapiro (2015) propose to standardize the sensitivities by the
asymptotic variance of the estimators
V ar(gˆj )
Λ̄ij = Λij . (D.2)
V ar(θˆi )
In Figure D.1, we plot these standardized sensitivities for the case where we estimate b0 , λ,
and c (panel A), and for the benchmark case where we fix b0 and estimate only λ and c (panel
2 2
27 These add up to −1.3, and adding E P ln R E −E Q ln R E gives a total variance premium equal to
zero.
3715

B). In panel A, we see that the sensitivities of b0 and λ are very similar. As discussed in Section
2.4.2, this problem arises because “too little” probability weighting can be compensated for by
giving a higher weight to the CPT part in the utility function. Conversely, even extreme levels of
probability weighting do not matter much if the CPT component enters the investor’s total utility
function with only little weight. Once we fix b0 , panel B shows that we can separately identify λ
and c as their sensitivities differ substantially across moments. In particular, the variance premium
affects the estimator for c much more than the estimator for λ, in line with the comparative statics
analysis in Section 2.3.
E Alternative Equity Return Distributions

In this section, we show that our baseline results are robust to using a normal distribution, or a
skewed fat-tailed distribution, instead of a standard lognormal distribution, for equity index returns.
We first introduce the standardized skewed-t distribution (see, for example, Lambert, Laurent, and
Veredas (2012) and Bauwens and Laurent (2005)). Then we report the Maximum Likelihood
estimates for the key parameters of these distributions using our sample of S&P 500 index returns.
We then present results for our baseline GMM estimation using these alternative equity return
distributions.
The return on the S&P 500, rt , follows the following process:
rt = μ+εt (E.1)
εt = σ ςt , (E.2)
where the random variable ςt is SKST (0,1,ξ,v) distributed; that is, it follows a standardized
skewed-t distribution with parameters v > 2 (the number of degrees of freedom) and ξ > 0
(a parameter related to skewness). The density of this function is given by
⎧
⎪
⎪ 2
sg ξ (sςt +m)|v
⎨ ξ + ξ1 if ςt < −m/s
f (ςt |ξ,v) = (E.3)
⎪
⎪ 2
sg ξ (sς +m)/ξ |v if ς t −m/s,
⎩ ξ+ 1 t
ξ
where g(·|v) is a symmetric (zero mean and unit variance) Student-t density with v degrees of
freedom defined as
−(v+1)/2
v−1 x2
g(x|v) = √ 2 v 1+ , (E.4)
π (v −2) 2 v −2
and where (·) denotes the gamma function.

)
The constants m = m(ξ,v) and s = s 2 (ξ,v) are,
respectively,
the mean and standard deviation
of the nonstandardized skewed-t density, SKST m,s 2 ,ξ,v , and are defined as follows:
√ # $
v−1 v −2 1
m(ξ,v) = √2 v ξ− (E.5)
π 2 ξ
# $
1
s 2 (ξ,v) = ξ 2 + 2 −1 −m2 . (E.6)
ξ
The parameters ξ and v are related to the distribution’s skewness and kurtosis, respectively. Because
ξ 2 can be shown to be equal to the ratio of the probability masses above and below the mode, the
distribution has zero skewness when ξ = 1, negative skewness when ξ < 1, and positive skewness
when ξ > 1. The fatness of tails (kurtosis) decreases with the degrees of freedom v, and converges
to a skewed normal distribution as v → ∞. When v → ∞ and ξ = 1, this distribution collapses to a
standard normal distribution.
3716

A Scale b0 estimated
0.3
0.2
0.1

-0.1
b
-0.2 O
c
-0.3
EP VP put(0.95) put(1.04) call(0.95) call(1.05)
B Benchmark (scale b0 fixed)

0.2
-0.2
-0.4
-0.6
-0.8 O
c
-1
EP VP put(0.95) put(1.04) call(0.95) call(1.05)
Figure D.1
Sensitivity of CPT parameter estimates to sample moments
This figure shows the standardized sensitivities of the CPT parameter estimates to the following sample moments:
the equity premium, the variance premium, and the put and call option returns. To facilitate the interpretation
of sensitivities, we orthogonalized option returns with respect to the equity and the variance premium. Panel A
shows the sensitivities for the setting in which the scale is also estimated (specification (5) in Table 3), and panel
B shows the sensitivities for the benchmark CPT setting (specification (1) in Table 3). Standardized sensitivities
are calculated following Gentzkow and Shapiro (2015) as the sensitivity of the expected value of each parameter
to a 1-standard-deviation change in the realization of each moment condition.
Table E.1 reports estimates for θ = (μ,σ,ξ,v), using our daily sample of overlapping 30-day log
returns on the S&P 500 over the period running from 1996 to 2010. ξ is well below 1, implying
negative skewness. The ratio of the probability masses above and below the mode is equal to
ξ 2 = 0.59, implying that the degree of negative skewness is economically important. The estimated
degrees of freedom parameter, v, equals 4.58, indicating that the return distribution is not only
left-skewed but also fat-tailed.
Table E.2 reports results for our GMM estimation using these parameter values in the skewed-t
distribution, and using the normal distribution in specifications (2) and (1), respectively.
3717

Table E.1
Parameter estimates of the skewed-t distribution
μ σ ξ v
Estimate 0.0046 0.0533 0.7692 4.5776
This table reports the estimated parameters for the Skewed-t distribution described in Equations (E.1) to (E.6).
Table E.2
GMM estimates in the unconditional setting: Alternative equity return distributions

(1) (2) (3)
Observed Benchmark Normal Skewed-t
γ (risk aversion) 1 1 1
b0 (scale) 0.95 0.95 0.95
ξ 0.75
υ 4.58
λ (loss aversion) 1.39 1.34 1.26
[0.30] [0.35] [0.37]
c (probability weighting) 0.65 0.66 0.77
[0.00] [0.00] [0.00]
CPT contribution (%) 50.00 49.58 50.30
Implied variance premium (%2 ) 146.72 146.33 146.41 147.03
Implied downside VP (%2 ) 118.49 91.85 97.72 117.91
Implied upside VP (%2 ) 27.53 54.48 48.70 29.12
Implied equity premium (%) 0.56 0.53 0.53 0.51
Average put return (%) −5.77 −9.23 −8.87 −5.98
Average call return (%) −30.14 −31.71 −31.42 −28.19
Average OTM call return (%) −11.27 −16.94 −16.31 −12.43
Average OTM put return (%) −41.46 −43.01 −42.46 −37.61
VF 3.47 2.75 3.87
RA without CPT 2.27 2.30 2.19
This table reports GMM estimates for the degree of loss aversion, λ, and probability weighting, c (c1 = c2 ),
for equity return distributions that are different from the benchmark lognormal distribution. In particular, we
compare the benchmark lognormal distribution with a normal distribution (specification (2)) and with a skewed-t
distribution (specification (3)). For the skewed-t distribution, the parameters driving the skewness and kurtosis—
ξ and υ , respectively—are calibrated to match the observed monthly equity returns (see main text). In all
specifications, the level of risk aversion, γ , is fixed at 1, the scale, b0 , is fixed at 0.95, which is the scale that
yields a 50% contribution of CPT to total utility in the benchmark specification (specification (1)), and the
reference level, WRef , is assumed equal to the risk-free rate (0.19%). For each specification, we report the CPT
contribution to total utility, the model-implied variance and equity risk premiums, the average model-implied call
and put returns, the average model-implied OTM call and put returns, the value function (VF) evaluated at the
optimum, and the level of risk aversion (RA) that would be needed in the no-CPT case to obtain the model-implied
equity premium. These parameters are estimated by applying GMM to returns on the S&P 500 and on S&P 500
call and put options with different strikes and 30 days to maturity and to the equity and variance premium. The
weighting matrix gives an equal weight of one-fourth to the equity, call, put, and variance premium moments.
The sample period runs from January 1996 to March 2016. p-values of test of the hypotheses that, respectively,
λ = 1 and c = 1, are reported in brackets. These p-values are calculated using Newey-West standard errors.
F Results for a Model with Distortion in Both CRRA and CPT Components
Table F.1 reports GMM estimates for the degree of loss aversion, λ, and probability weighting, c,
when the CRRA and CPT components of total utility are distorted. For comparison, we also show
the results for the benchmark specification (specification (1)), in which only the CPT component of
the utility function is distorted using the probability weighting parameter cCPT . In all specifications,
the reference level, WRef , is assumed equal to the risk-free rate (0.19%), the level of risk aversion,
γ , is fixed at 1. In specifications (1) and (2), the scale, b0 , is fixed at 0.95, which is the scale that
yields a 50% contribution of CPT to total utility in the benchmark specification (specification (1)).
3718

Table F.1
GMM estimates in the unconditional setting: Distortion of the CRRA and CPT components
(1) (2) (3)
Distorted CPT
Benchmark and CRRA Quiggin (82)
b0 (scale) 0.95 0.95 0
γ (risk aversion) 1.00 1.00 1.00
λ (loss aversion) 1.39 1.40
[0.30] [0.23]
cCRRA 0.78 0.78
[0.00] [0.00]
cCPT 0.65 0.78

[0.00] [0.00]
CPT contribution (%) 50.00 53.11 0.00
Implied Variance Premium (%2 ) 146.72 146.33 146.11 146.18
Implied downside VP (%2 ) 118.49 91.85 90.09 83.59
Implied upside VP (%2 ) 27.53 54.48 56.02 62.58
Implied equity premium (%) 0.56 0.53 0.54 0.29
Average call return (%) −5.77 −9.23 −10.13 −14.38
Average put return (%) −30.14 −31.71 −32.43 −27.76
Average OTM call return (%) −11.27 −16.94 −18.45 −22.31
Average OTM put return (%) −41.46 −43.01 −43.52 −39.16
VF 3.47 6.59 16.18
RA without CPT 2.27 1.67 1.00
In specification (3), the scale, b0 , is fixed at 0, a setting which is effectively the rank-dependent
utility model from Quiggin (1982). p-values of the hypotheses that λ = 1 and c = 1, respectively,
are reported in brackets. The p-values are calculated using Newey-West standard errors. For each
specification, we also report the CPT contribution to total utility, the model-implied variance and
equity risk premiums, the average model-implied call and put returns, the average model-implied
OTM call and put returns, the value function (VF) evaluated at the optimum, and the level of
risk aversion (RA) that would be needed in the no-CPT case to obtain the model-implied equity
premium. These parameters are estimated by applying GMM to returns on the S&P 500 index and
on S&P 500 call and put options with different strikes and 30 days to maturity, and to the equity
and variance premium. The GMM weighting matrix gives an equal weight of one-fourth to the
equity, call, put, and variance premium moments. The sample period runs from January 1996 to
March 2016.
G Details on the Modeling of Time-Varying Loss Aversion

In this appendix, we provide details on the modeling of time-varying loss aversion, as referred to
in Section 3. We closely follow Barberis, Huang, and Santos (2001), who posit that loss aversion
varies with the recent performance of the representative agent’s portfolio. Specifically, time-varying
loss aversion λt depends on a state variable zt as follows:
λt = λ̂+κλ (zt − z¯t ),
where λ̂ is the baseline loss aversion parameter, z¯t is the average of zt , and κλ is the sensitivity of
loss aversion to recently realized gains and losses as measured by zt . Following BHS, we assume
that the representative agent compares the current price of the risky asset with a benchmark level,
where the benchmark level of the price responds sluggishly to changes in the value of equity. This
sluggishness is motivated in BHS (p. 14) as follows: “when the stock price moves up by a lot,
the benchmark level also moves up, but by less. Conversely, if the stock price falls sharply, the
3719

benchmark level does not adjust downwards by as much.” Formally,
RE
zt+1 = η(zt E
)+(1−η),
Rt+1
where R E is a fixed parameter calibrated to guarantee that half of the time the agent has prior gains
(RtE > R E ) and the rest of the time she has prior losses (RtE < R E ). The parameter η measures
the degree of sluggishness, which can be interpreted as the agent’s memory horizon. We assume
η = 0.5, which implies that the half life of the representative agent’s memory is 1 month.

References
Abdellaoui, M. 2000. Parameter-free elicitation of utility and probability weighting functions. Management
Science 46:1497–512.
Ait-Sahalia, Y., and A. Lo. 2000. Nonparametric risk management and implied risk aversion. Journal of
Econometrics 94:9–51.
Andersen, T. G., and O. Bondarenko. 2009. Dissecting the market pricing of return volatility. Working Paper.
Ang, A., G. Bekaert, and J. Liu. 2005. Why stocks may disappoint. Journal of Financial Economics 76:471–508.
Bakshi, G., and N. Kapadia. 2003. Delta-hedged gains and the negative market volatility risk premium. Review
of Financial Studies 16:527–66.
Bakshi, G., N. Kapadia, and D. Madan. 2003. Stock return characteristics, skew laws, and the differential pricing
of individual equity options. Review of Financial Studies 16:101–43.
Bakshi, G., and D. Madan. 2006. A theory of volatility spreads. Management Science 52:1945–56.
Bakshi, G., D. Madan, and G. Panayotov. 2010. Returns of claims on the upside and the viability of U-shaped
pricing kernels. Journal of Financial Economics 97:130–54.
Barberis, N. 2011. A model of casino gambling. Management Science 58:35–51.
Barberis. 2013. Thirty years of prospect theory in economics: A review and assessment. Journal of Economic
Perspectives 27:173–96.
Barberis, N., and M. Huang. 2008. Stocks as lotteries: The implications of probability weighting for security
prices. American Economic Review 98:2066–100.
Barberis, N., M. Huang, and T. Santos. 2001. Prospect theory and asset prices. Quarterly Journal of Economics
116:1–53.
Barberis, N., A. Mukherjee, and B. Wang. 2016. Prospect theory and stock returns: An empirical test. The Review
Barberis, N., and W. Xiong. 2009. What drives the disposition effect? An analysis of a long-standing preference-
based explanation. Journal of Finance 64:751–84.
——-. 2012. Realization utility. Journal of Financial Economics 104:251–71.
Bauwens, L., and S. Laurent. 2005. A new class of multivariate skew densities, with applications to GARCH
models. Journal of Business and Economic Statistics 3:346–54.
Bekaert, G., and E. Engstrom. 2016. Asset return dynamics under bad environment-good environment
fundamentals. Journal of Political Economy 125:713–60.
Benartzi, S., and R. Thaler. 1995. Myopic loss aversion and the equity premium puzzle. Quarterly Journal of
Economics 110:75–92.
Bernard, C., and M. Ghossoub. 2010. Static portfolio choice under cumulative prospect theory. Mathematics and
Financial Economics 2:277–306.
3720

Bollerslev, T., M. Gibson, and H. Zhou. 2011. Dynamic estimation of volatility risk premiums and investor risk
aversion from option-implied and realized volatilities. Journal of Econometrics 160:235–45.
Bollerslev, T., G. Tauchen, and H. Zhou. 2009. Expected stock returns and variance risk premiums. Review of
Financial Studies 22:4463–92.
Bondarenko, O. 2003. Statistical arbitrage and securities prices. Review of Financial Studies 16:875–919.
———. 2014. Why are puts so expensive? Quarterly Journal of Finance 4:1–50.
Bordalo, P., N. Gennaioli, and A. Shleifer. 2012. Salience theory of choice under risk. Quarterly Journal of

Boyer, B., T. Mitton, and K. Vorkink. 2010. Expected idiosyncratic skewness. Review of Financial Studies
23:169–202.
Boyer, B. H., and K. Vorkink. 2014. Stock options as lotteries. Journal of Finance 69:1485–527.
Brennan, M. J. 1979. The pricing of contingent claims in discrete time models. Journal of Finance 34:53–68.
Broadie, M., M. Chernov, and M. Johannes. 2009. Understanding index option returns. Review of Financial
Studies 22:4493–529.
Bruhin, A., H. Fehr-Duda, and T. Epper. 2010. Risk and rationality: Uncovering heterogeneity in probability
distortion. Econometrica 78:1375–412.
Camerer, C. F., and T. Ho. 1994. Violations of the betweenness axiom and nonlinearity in probability. Journal
of Risk and Uncertainty 8:167–96.
Carr, P., and D. Madan. 2001. Optimal positioning in derivative securities. Quantitative Finance 1:19–37.
Carr, P., and L. Wu. 2009. Variance risk premiums. Review of Financial Studies 22:1311–41.
Chabi-Yo, F. 2012. Pricing kernels with stochastic skewness and volatility risk. Management Science 58:624–40.
Chabi-Yo, F., R. Garcia, and E. Renault. 2008. State dependence can explain the riskaversion puzzle. Review of
Chaudhuri, R., and M. Schroder. 2015. Monotonicity of the stochastic discount factor and expected option returns.
Review of Financial Studies 28:1462–505.
Christoffersen, P., S. Heston, and K. Jacobs. 2013. Capturing Option Anomalies with a Variance-Dependent
Pricing Kernel, Review of Financial Studies 26:1962–2006.
Cochrane, J. 2005. Asset pricing. Princeton: Princeton University Press.
Conrad, J., R. F. Dittmar, and E. Ghysels. 2012. Ex ante skewness and expected stock returns. Journal of Finance
68:85–124.
Constantinides, G. M., J. C. Jackwerth, and A. Savov. 2013. The puzzle of index option returns. Review of Asset
Pricing Studies 3:229–57.
Coval, J., and T. Shumway. 2001. Expected option returns. Journal of Finance 56:983–1009.
Cuesdeanu, H., and J. C. Jackwerth. 2018. The pricing kernel puzzle: Survey and outlook. Annals of Finance
14:289–329.
De Giorgi, E., and S. Legg. 2012. Dynamic portfolio choice and asset pricing with narrow framing and probability
weighting. Journal of Economic Dynamics and Control 36:951–72.
Drechsler, I. 2013. Time-varying fear, and asset prices. Journal of Finance 68:1843–89.
Drechsler, I., and A. Yaron. 2011. What’s vol got to do with it? Review of Financial Studies 24:1–45.
Driessen, J., and P. Maenhout. 2007. An empirical portfolio perspective on option pricing anomalies. Review of
Finance 11:561–603.
3721

Du, D. 2011. General equilibrium pricing of options with habit formation and event risks. Journal of Financial
Ebert, S., and P. Strack. 2015. Until the bitter end: On prospect theory in a dynamic context. American Economic
Review 105:1618–33.
Epper, T., and H. Fehr-Duda. 2012. Probability and risk: Foundations and economic implications of probability-
dependent risk preferences. Annual Review of Economics 4:567–93.
Eraker, B. 2013. The performance of model based option trading strategies. Review of Derivatives Research
16:1–23.
Eraker, B., and M. Ready. 2015. Do investors overpay for stocks with lottery-like payoffs? An examination of

the returns of OTC stocks. Journal of Financial Economics 115:486–504.
Feunou, B., M. R. Jahan-Parvar, and C. Okou. 2017. Downside variance risk premium. Journal of Financial
Gabaix, X. 2012. Variable rare disasters: An exactly solved framework for ten puzzles in macro-finance. Quarterly
Journal of Economics 127:645–700.
Gentzkow, M., and J. M. Shapiro. 2015. Measuring the sensitivity of parameter estimates to sample statistics.
Working Paper.
Gollier, C. 2011. Portfolio choices and asset prices: The comparative statics of ambiguity aversion. Review of
Economic Studies 78:1329–44.
Green, C., and B.-H. Hwang. 2012. Initial public offerings as lotteries: Skewness preference and first-day returns.
Management Science 58:432–44.
Held, M., J. Kapraun, M. Omachel, and J. Thimme. 2018. Up and downside variance risk premia in global equity
markets. Working Paper.
Ilmanen, A. 2012. Do financial markets reward buying or selling insurance and lottery tickets? Financial Analysts
Journal 68:26–36.
Ingersoll, J. E. 2008. Non-monotonicity of the Tversky-Kahneman probability-weighting function: A cautionary

note. European Financial Management 14:385–90.
——–. 2016. Cumulative prospect theory, aggregation, and pricing. Critical Finance Review 5:305–50.
Ingersoll, J. E., and L. J. Jin. 2013. Realization utility with reference-dependent preferences. Review of Financial
Studies 26:723–67.
Jackwerth, J. 2000. Recovering risk aversion from option prices and realized returns. Review of Financial Studies
13:433–67.
Jiang, G., and Y. Tian. 2005. Model-free implied volatility and its information content. Review of Financial
Studies 18:1305–42.
Jin, L. J. 2015. A speculative asset pricing model of financial instability. Working Paper.
Jones, C. 2006. A nonlinear factor analysis of S&P 500 index option returns. Journal of Finance 61:2325–63.
Kahneman, D. 2011. Thinking, fast and slow. New York: Farrar, Straus and Giroux.
Kahneman, D., and A. Tversky. 1979. Prospect theory: An analysis of decision under risk. Econometrica 47:
263–91.
Kilic, M., and I. Shaliastovich. 2017. Good and Bad Variance Premia and Expected Returns. Management Science.
Advance Access published January 16, 2018.
Kliger, D., and O. Levy. 2009. Theories of Choice under Risk: Insights from Financial Markets. Journal of
Economic Behavior and Organization 71:330–46.
3722

Kőszegi, B., and M. Rabin. 2006. A Model of Reference-Dependent Preferences. Quarterly Journal of Economics
121:1133–66.
———. 2007. Reference-Dependent Risk Attitudes. American Economic Review 97:1047–73.
Kumar, A. 2009. Who Gambles in the Stock Market? Journal of Finance 64:1889–933.
Lambert, P., S. Laurent, and D. Veredas. 2012. Testing conditional asymmetry. A residual-based approach.
Journal of Economic Dynamics and Control 36:1229–47.
Londono, J. 2014. The Variance Risk Premium around the World. Working Paper, Federal Reserve Board.
Machina, M. J. 1989. Dynamic Consistency and Non-Expected Utility Models of Choice under Uncertainty.

Journal of Economic Literature 27:1622–68.
Merton, R. C. 1969. Lifetime Portfolio Selection under Uncertainty: The Continuous-Time Case. Review of
Economics and Statistics 51:247–57.
Mitton, T., and K. Vorkink. 2007. Equilibrium Underdiversification and the Preference for Skewness. Review of
Polkovnichenko, V. 2005. Household Portfolio Diversification: A Case for Rank-Dependent Preferences. Review
Polkovnichenko, V., and F. Zhao. 2013. Probability Weighting Functions Implied in Option Prices. Journal of
Financial Economics 107:580–609.
Quiggin, J. 1982. A theory of anticipated utility. Journal of Economic Behavior and Organization 3:323–43.
Rieger, M. O., and M. Wang. 2006. Cumulative prospect theory and the St. Petersburg paradox. Economic Theory
28:665–79.
Rosenberg, J., and R. F. Engle. 2002. Empirical pricing kernels. Journal of Finanical Economics 64:341–72.
Rubinstein, M. 1976. The valuation of uncertain income streams and the pricing of options. Bell Journal of
Samuelson, P. A. 1969. Lifetime Portfolio Selection By Dynamic Stochastic Programming. Review of Economics
and Statistics 51:239–46.
Samuelson, P. A., and R. C. Merton. 1969. A Complete Model of Warrant Pricing that Maximizes Utility.
Industrial Management Review 10:17–46.
Santa-Clara, P., and A. Saretto. 2009. Option strategies: Good deals and margin calls. Journal of Financial
Markets 12:391–417.
Schneider, C., and O. Spalt. 2016. Conglomerate Investment, Skewness, and the CEO Long-Shot Bias. Journal
of Finance 71:635–71.
Schreindorfer, D. 2014. Tails, fears, and equilibrium option prices. Working Paper.
Shaliastovich, I. 2015. Learning, confidence, and option prices. Journal of Econometrics 187:18–42.
Snowberg, E., and J. Wolfers. 2010. Explaining the favorite-long shot bias: Is it risk-love or misperceptions?
Journal of Political Economy 118:723–46.
Song, Z., and D. Xiu. 2016. A tale of two option markets: Pricing kernels and volatility risk. Journal of
Tversky, A., and D. Kahneman. 1992. Advances in prospect theory, cumulative representation of uncertainty.
Journal of Risk and Uncertainty 5:297–323.
Wu, G., and R. Gonzalez. 1996. Curvature of the probability weighting function. Management Science 42:1676–
90.
Ziegler, A. 2007. Why does implied risk aversion smile? Review of Financial Studies 20:859–904.
3723

9、Cumulative Prospect Theory, Option Returns, And the Variance Premium（2019RFS）

Uploaded by

Copyright:

Available Formats

You might also like

9、Cumulative Prospect Theory, Option Returns, And the Variance Premium（2019RFS）

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

9、Cumulative Prospect Theory, Option Returns, And the Variance Premium（2019RFS）

Uploaded by

Copyright:

Available Formats

Cumulative Prospect Theory, Option

Returns, and the Variance Premium

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

A central empirical success of prospect theory (Kahneman and Tversky 1979;

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3667 3667–3723

and Santos (2001) demonstrate how a dynamic prospect theory framework

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3668 3667–3723

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3669 3667–3723

A Average call returns

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

0.92 0.94 0.96 0.98 1 1.02 1.04 1.06 1.08

B Average put returns

0.92 0.94 0.96 0.98 1 1.02 1.04 1.06 1.08

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3670 3667–3723

To understand why the CPT model is successful in fitting options returns,

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3671 3667–3723

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

B Implied Risk Aversion

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3672 3667–3723

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3673 3667–3723

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3674 3667–3723

Ilmanen (2012) provides a survey of financial market facts consistent with a

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3675 3667–3723

Finally, our paper is related to studies on the shape of the pricing

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3676 3667–3723

where λ ≥ 1 is called the loss aversion parameter. For example, if λ = 2, then

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

πi = w − (p1 +···+pi )−w − (p1 +···+pi−1 ) for 2 ≤ i < k

When c1 = c2 = 1, the weighting functions in (6) reduce to w + (p) = w − (p) = p,

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3677 3667–3723

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3678 3667–3723

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

1.2 The investor’s problem

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3679 3667–3723

Proposition 1. The investor’s optimal portfolio weight is unique, finite, and

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

Proposition 2. Consider a multiperiod version of the model above where

We think of the preferences underlying Proposition 2 as follows. While the

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3680 3667–3723

Proposition 2 is special, because it represents one case in which a dynamic

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

1.3 Equilibrium pricing kernel and equity risk premium

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3681 3667–3723

We now introduce derivatives on the equity index. Each derivative d on

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3682 3667–3723

Downloaded from https://academic.oup.com/rfs/article/32/9/3667/5232703 by Zhejiang University user on 14 December 2020

1.4 The model-implied variance premium

[13:59 5/8/2019 RFS-OP-REVF180130.tex] Page: 3683 3667–3723

where pi and qi denote the actual and risk-neutral probability of state i,