Estimating the Intertemporal Risk Return Relation

Using Option Implied Expected Returns*

Guanglian Hu Hamish Malloch

University of Sydney

June 20, 2024

Abstract

We provide new empirical evidence on the intertemporal risk-return relation in the

aggregate stock market. Our approach estimates the expected excess market returns
from index options, avoiding reliance on noisy realized returns or specifying condi-
tioning variables. We find a positive and statistically significant relationship between
the conditional mean and variance of stock returns with this evidence being consis-
tent across different specifications of the conditional variance. The positive risk-return
relation is time varying, robust to controlling for the omitted variable bias, holds for
different components of the expected return, and persists to longer horizons. Our find-
ings suggest that the measurement of the conditional mean is critical in identifying the
risk return relation.

JEL Classification: G12

Keywords: risk-return trade-off; mean-variance relation; option implied expected re-
turn

* We thank Dirk Baur, Jiali Gao, Philip Gharghori, Idan Hodor, Frank Liu, Stijn Masschelein, Peter
Pham, Raymond Da Silva Rosa, Joshua Shemesh, David Solomon, Jin Yu, and seminar participants at the
University of Sydney, University of Western Australia, Monash, and IAAE 2024 (Xiamen).

1
1 Introduction

Understanding the fundamental risk-return relationship — the way in which the conditional
mean and variance of stock market returns are related to one another — is a key issue for
financial economists. Economic intuition would suggest a positive mean-variance relation as
investors must be compensated with higher expected returns during risky/uncertain times.
Standard theories (e.g., Merton, 1973; Sharpe, 1964; Lintner, 1965) indeed predict that
the expected excess market return is positively related to its conditional variance.1 Despite
strong theoretical support, empirical evidence on this risk-return relationship is surprisingly
mixed, with the literature reporting conflicting findings on its sign and statistical significance.
In particular, the results appear to be highly sensitive to the measure of the conditional
variance used in the empirical analysis.
We argue the measurement of the conditional mean is important in testing the risk return
trade-off. Existing studies often consider future realized returns as a proxy for expected
returns in empirical tests. While realized returns are unbiased, it is well known that they
are notoriously noisy and can be a poor measure of the expected returns especially in the
presence of time varying risk premium (Lundblad, 2007; Elton, 1999). Another popular
method for obtaining estimates of expected returns involves projecting realized returns onto
predetermined conditioning variables and then taking the fitted values from the projection.
However, the choice of the conditioning variables is somewhat arbitrary and can significantly
influence the estimated risk return relation (Harvey, 2001). The primary innovation of our
paper is that we estimate the intertemporal risk-return relation using conditional expected
1
Some studies argue that a negative relationship can obtain in some theoretical setups (e.g., Abel, 1988;
Backus and Gregory, 1993; Whitelaw, 2000). Lettau and Ludvigson (2010) show that those models have
counterfactual implications for other aspects of the data. A negative mean variance relation may also arise
in models with information driven volatility (Ai, Han, and Xu, 2022), biased volatility beliefs (Lochstoer
and Muir, 2022), no-dividend stocks (Atmaz and Basak, 2022), or variance components relating to fear and
euphoria (Guo, Lin, and Pai, 2024).

2
excess market returns directly computed from index option prices. Our approach is attractive
because it does not rely on using noisy realized returns or specifying unknown conditioning
variables, and option implied expected returns are genuinely forward looking.
We focus on the risk-return trade-off for the S&P 500 index over the sample period from
1996 to 2021. We first estimate the risk return relation using realized returns. Confirming
existing studies, when using realized returns in the empirical tests, we find the sign and
statistical significance of the relationship between risk and return is sensitive to the measure
of the conditional variance and sample period. There is a weak or even negative risk return
trade-off. In contrast, empirical evidence is drastically different when we consider option
implied expected returns. We find a positive and statistically significant relationship between
the first two moments of stock returns. The empirical results are remarkably consistent across
different implementations of the conditional variance, with estimates of slope coefficient
ranging from 0.528 to 2.201 over the full sample. The mean-variance relationship is also
economically significant. We find that one standard deviation increase in the conditional
variance is associated with an increase of 0.29% to 0.37% per month in the expected return,
depending on the measure of the conditional variance. Overall our analysis suggests that the
measurement of the conditional mean is critical in detecting a positive risk-return trade-off.
Yu and Yuan (2011) demonstrate that investor sentiment has a significant impact on the
risk-return trade-off. They find that stock returns are positively related to variance in low
sentiment periods but unrelated to variance in high sentiment periods. Consistent with their
results, we find that the risk-return trade-off is indeed stronger in the low sentiment regime.
In contrast to their findings, which are based on realized returns, we find the relation between
option implied expected return and variance remains significant in high sentiment periods.
This suggests that while the impact of sentiment is important, it does not completely override
the fundamental risk-return trade-off.

3
 We also conduct several specification analyses. A series of papers (Guo and Whitelaw,
2006; Scruggs, 1998; Campbell, 1987) argue that the standard empirical tests of how the
expected return is related to its conditional variance is subject to the omitted variable
bias because the expected return depends on not only the conditional variance but also the
covariance between stock returns and state variables that drive investment opportunities. To
address the potential omitted variable bias, we control for several macroeconomic variables
that are commonly used in the literature to track changes in the investment opportunity
set. We also control for lags of the conditional mean and conditional variance because prior
studies find that including the lagged mean and lagged variance in the risk return regression
can have a significant impact on the sign of the estimated risk return relation (Lettau and
Ludvigson, 2010; Brandt and Kang, 2004). We find that the strong positive relation between
the conditional mean and conditional variance of stock returns remains in the presence of
all the controls. Lastly, we show that the squared variance is highly significant and helps
explain the movements in the conditional mean, suggesting that there is non-linearity in the
risk return relation, consistent with the finding in Adrian, Crump, and Vogt (2019).
We decompose the option-implied expected return into three components related to the
risk-neutral volatility, skewness, and kurtosis of the index. Our analysis reveals that the risk
neutral second moment makes the largest contribution to the expected return, but higher
moments become increasingly important at longer horizons and in recent years. We also find
a positive and statistically significant risk-return relation for each component of the expected
return.
Last, we study the term structure of the risk-return trade-off. When using realized
returns, we find that while the mean-variance relation is weak at short horizons, it becomes
significantly positive at longer horizons. Our results are consistent with Bandi and Perron
(2008) and Harrison and Zhang (1999) who also document a significantly positive risk-return

4
trade-off at longer horizons using different sample period and methodology. A positive risk
return trade-off is better revealed at longer horizons because variation in expected return
is easy to detect at longer horizons. In contrast, short term realized returns are noisy in
that most of their variation is unrelated to variation in the expected return. When using
option implied expected returns, we find a significantly positive risk-return relation across
all horizons. The risk return relation also appears stable and does not exhibit significant
variation as a function of return horizon.
Related Literature
This study is most closely related to two strands of literature. First, there is a long standing
literature on estimating the risk-return trade-off for the aggregate stock market with mixed
empirical evidence. See, among others, Merton (1980), Scruggs (1998), Ghysels, Santa-
Clara, and Valkanov (2005), Guo and Whitelaw (2006), Lundblad (2007), Ludvigson and
Ng (2007), Campbell (1987), Turner, Startz, and Nelson (1989), Campbell and Hentschel
(1992), Glosten, Jagannathan, and Runkle (1993), Brandt and Kang (2004), French, Schwert,
and Stambaugh (1987), Yang (2024), and Pastor, Sinha, and Swaminathan (2008). This
literature has largely emphasized on the importance of estimating the conditional variance
with few studies focusing on the measurement of the conditional mean. A notable exception
is Pastor, Sinha, and Swaminathan (2008) which use the implied cost of capital as a proxy for
the conditional expected return and find that it is positively related to variance. We also find
a positive mean-variance relationship, but we use option-based expected returns instead of
implied cost of capital. Overall our analysis complements Pastor, Sinha, and Swaminathan
(2008) and suggests that the measurement of the conditional mean may be of first order
importance in testing the risk-return trade-off.
Second, we build on the recent literature that demonstrates expected stock returns can be
inferred from options. In a highly influential article, Martin (2017) demonstrates that, under

5
the negative correlation condition, the expected market return is related to its risk neutral
variance and therefore can be calculated from index option prices.2 Chabi-Yo and Loudis
(2020) and Bakshi, Crosby, Gao, and Zhou (2019) propose alternative methods to extract
the expected return from options, relaxing some of the assumptions in Martin (2017).3 This
body of literature has convincingly demonstrated that options offer new opportunities for
understanding the variation in the market risk premium. Our paper adopts the approach
in Chabi-Yo and Loudis (2020) which does not rely on any distributional assumptions on
stock returns. Our contribution is to bridge these two previously unconnected literatures by
applying option implied expected returns in testing the relationship between the conditional
mean and conditional variance of stock market returns.
Our work is also related to an emerging strand of literature that uses option-implied
expected return lower bounds as a direct measure of the forward looking expected return.
For instance, using this approach, Cieslak, Morse, and Vissing-Jorgensen (2019) find that the
Fed lowers the equity premium by promising to provide stimulus, hence reducing downside
risk, Sautner, Van Lent, Vilkov, and Zhang (2023) show that firm-level climate change
exposure earns a significant risk premium while Berkman and Malloch (2023) show that
around 40% of the change in equity market values at the onset of the Covid-19 pandemic
can be attributed to changes in short-term discount rates.
More generally, our paper contributes to the extensive literature that aims to understand
the time series properties of aggregate stock market returns. Some studies focus on the
2
The negative correlation condition requires that Covt (rt,t+h , Mt,t+h rt,t+h ) < 0 where Mt,t+h is the
stochastic discount factor over the period t to t + h and rt,t+h is the return over the same period.
3
Related, several studies (Martin and Wagner, 2019; Kadan and Tang, 2020; Chabi-Yo, Dim, and Vilkov,
2023) extend the methodology to estimate the expected returns on individual stocks with some auxiliary
assumptions. Ross (2015) proposes a recovery theorem to extract the physical moments of returns from the
risk neutral measure. Back, Crotty, and Kazempour (2022) conduct a comprehensive study of the empirical
performance of several option-based expected return bounds, finding that the bounds for the market works
better than those for individual stocks.

6
first moment, finding that the equity risk premium appears too large relative to classical
models (e.g., Mehra and Prescott, 1985) and that stock returns seem predictable, especially
at the longer horizons, which in turn implies that the equity risk premium varies over time
(e.g., Campbell and Shiller, 1988; Fama and French, 1988). Existing studies also suggest
that stock return variance, which is the second moment, also varies significantly over time
(e.g., Schwert, 1989). Different from those studies, our focus is the co-movement between
the first two moments. We offer new insights on the nature of the relation between the
conditional mean and conditional variance of stock returns by constructing a novel measure
of conditional expected return from index options.
The paper proceeds as follows. Section 2 contains our main analysis. Section 3 includes
additional discussion and extensions, and Section 4 concludes.

2 Main Analysis

This section contains our main analysis. Section 2.1 provides relevant theoretical background
and an overview of our analysis. Section 2.2 considers the measurement of the conditional
variance. Section 2.3 discusses the existing approaches to estimating the conditional mean.
Section 2.4 introduces the approach we take to construct the expected excess market returns
from index options. Section 2.5 presents the empirical results for estimating the risk return
trade-off.

2.1 Overview

The literature on testing the risk-return trade-off for the aggregate stock market is often
motivated by a stylized version of the Intertemporal Capital Asset Pricing Model (ICAPM)
of Merton (1973), which says the expected excess market return Et (rt+1 ) is positively related

7
to its conditional variance:
Et (rt+1 ) = λ + γV art (rt+1 ) (1)

where γ is the relative risk aversion coefficient and should be positive for a risk-averse agent,
λ should be zero in the absence of market imperfections such as taxes or transaction costs.4
To test equation (1), it is common to adopt the following specification for the empirical
regression:
µ̂t = c + γ vˆt + et+1 (2)

where µ̂t is a measure of conditional mean and vˆt is a measure of conditional variance at
time t. Neither the expected return or its conditional variance is directly observed and must
be estimated from the data. In this paper, we estimate the risk-return relation for the S&P
500 index by estimating the regression in (2) using various measures of conditional expected
return and variance. Our sample period is from 1996 to 2021 for which we have option
data available to compute the expected excess returns for the S&P 500 index. Following the
literature, our baseline analysis focuses on the risk-return trade-off at the monthly frequency.
We also test the risk-return relation for other horizons and report those results in Section
3.4.
Alternatively, one can infer the relationship between the conditional mean and condi-
tional variance by examining the contemporaneous relationship between realized returns
and innovations in variance:
rt+1 = ρ0 + ρ1 ∆v + et+1 (3)

where ∆v = v̂t+1 − vˆt . This indirect test, dating back to French, Schwert, and Stambaugh
4
This equation can also be motivated by the conditional versions of the Sharpe-Lintner CAPM. Merton
(1973) demonstrates that the ICAPM collapses to the Sharpe-Lintner CAPM when the investment oppor-
tunity is static or the agent has a log utility. A positive mean-variance relation also obtains in more recent
dynamic equilibrium models (e.g., Campbell and Cochrane, 1999; Bansal and Yaron, 2004; Cochrane, 2017).

8
(1987), argues that a positive relationship between the conditional mean and conditional
variance (i.e., γ > 0) would imply a negative contemporaneous relationship between realized
returns and variance innovation, namely ρ1 < 0. This is because following an increase in the
conditional variance, the expected return increases which leads to an immediate drop in the
stock price and hence low realized contemporaneous return.5 We report the results for the
indirect tests in Section 3.5.

2.2 Measuring the Conditional Variance

We consider five commonly used measures of the conditional variance in our empirical tests.
First of all, we use lagged realized variance, which is sometimes referred to as the rolling
window estimator. Using lagged variance as a proxy for expected future variance is popular
as the second moment of returns is usually quite persistent. It should be noted though that
lagged realized variance is a biased estimate of expected variance unless variance follows a
random walk. Specifically, following existing studies (Ghysels, Santa-Clara, and Valkanov,
2005; French, Schwert, and Stambaugh, 1987), we calculate realized variance in month t
(RVt ) as the sum of the squared daily excess returns within the month and use it as a proxy
for the expected variance for month t + 1:

22
X
RVt = ri2 (4)
i=1

where ri is the daily excess return of the S&P 500 index on trading day i in month t. We
obtain return data on the S&P 500 index from CRSP.
Second, we compute the conditional expected value of future realized variance, denoted
5
A negative ρ1 could also be consistent with the leverage effect (Black, 1976; Christie, 1982). That is,
following a drop in the stock price, financial leverage increases and therefore the stock becomes riskier.

9
as EV , using the Heterogeneous Autoregressive Model (HAR) of Corsi (2009). Despite its
simplicity, the HAR model demonstrates superior performance and has become a standard
benchmark in the volatility forecasting literature. Following Corsi (2009), we project current
realized variance onto past realized variances computed over the previous day, week, and
month, and take the fitted values as our proxy for expected variance:

W D
RVt = δ0 + δ1 RVt−1 + δ2 RVt−1 + δ3 RVt−1 + ϵt (5)

W D
where RVt and RVt−1 are realized variances in month t and t − 1, RVt−1 and RVt−1 represent
realized variances over the past 5 trading days and 1 trading day respectively, measured on
the last trading day of month t − 1. Realized variances are calculated based on equation
(4).6
We also consider two measures of GARCH-tpye variances. In particular, we consider a
simple GARCH(1,1) model and an EGARCH (1,1) model that extends the GARCH(1,1)
by allowing for the conditional variance to respond asymmetrically to positive and negative
return innovations. We obtain estimates of conditional variance by fitting the two models
to monthly excess returns of the S&P 500 index and we refer to those two measures as GV
and EGV .7
Lastly, we estimate the conditional variance using the mixed data sampling approach
(MIDAS) of Ghysels, Santa-Clara, and Valkanov (2005) and refer to those variance estimates
6
Our implementation is slightly different from Corsi (2009) in that we compute realized variances from
daily returns instead of intraday returns, following most studies in the literature on testing the risk return
trade-off. Bali and Peng (2006) estimate the intertemporal relation between risk and return using high-
frequency data and find a positive relationship at the daily level.
7
When estimating GARCH and EGARCH variances, we do not embed a linear relation between return
and conditional variance in the mean equation to ensure the comparison with other measures of variances is
meaningful.

10
as M V . The MIDAS estimator of the conditional variance of monthly return is given by:

250
X
2
M Vt = ωd rt−d (6)
d=0

where ωd is the weight assigned to the squared return of day t − d. M V is somewhat similar
to the realized variance, but uses longer sample and assigns different weights to past returns.
Following Ghysels, Santa-Clara, and Valkanov (2005), we consider the following parametric
scheme for the weights:
exp (κ1 d + κ2 d2 )
ωd (κ1 , κ2 ) = P250 . (7)
exp (κ i + κ i2)
i=0 1 2

We implement the MIDAS estimator using the parameter values for κ1 and κ2 in Ghysels,
Santa-Clara, and Valkanov (2005).8 Alternatively, we follow their methodology to estimate
κ1 and κ2 using our sample and find similar results.
Table 1 reports summary statistics for monthly S&P 500 returns and five monthly con-
ditional variance measures. All quantities are reported on a monthly basis. Over our sample
period 1996 to 2021, the average monthly excess return of the S&P 500 index is about 59
basis points with a standard deviation of 0.0439, implying an annualized Sharpe ratio of
0.465. The mean of monthly realized variance (RV ), expected variance (EV ), GARCH
variance (GV ), EGARCH variance (EGV ), and MIDAS variance (M V ) is 0.309%, 0.309%,
0.203%, 0.195%, and 0.324%, corresponding to an annualized volatility of 19.25%, 19.26%,
15.61%, 15.32%, and 19.71%. Index returns have a negative skewness while return variances
exhibit positive skewness. These results are consistent with the existing empirical evidence
on the properties of the first two moments of index returns (e.g., Eraker, 2004; Andersen,
Bollerslev, Diebold, and Ebens, 2001).
8
Specifically we set κ1 = −0.005141 and κ2 = −0.0001058, which are estimated based on the longest
sample in their paper.

11
2.3 Measuring the Conditional Mean

This section discusses the current approaches to estimating the conditional mean of stock
returns. There are a variety of methods for estimating this quantity. The most commonly
used approach is to replace µ̂t in equation (2) with future realized returns rt+1 and estimate
the following regression:
rt+1 = c + γ vˆt + et+1 (8)

where vˆt is again a measure of conditional variance at time t. It is natural to consider realized
returns because ex ante expected return is in general difficult to measure and realized returns
are unbiased estimates of expected returns. However, it is well known that realized returns
are notoriously noisy and may be poor measures of expected returns (Elton, 1999). For
example, Lundblad (2007) convincingly shows that when using realized returns as a proxy
for expected returns, even 100 years of data still constitute a small sample for identifying
the risk-return trade-off. The measurement issue is further compounded in the presence of a
time varying risk premium. When the expected return is time varying, a low realized return
would actually suggest that the expected return is rising rather than falling. Indeed Section
2.5 shows that the use of realized returns is at least partially responsible for the conflicting
results in the literature.
It is also popular to estimate expected returns via projection. See, among others, Camp-
bell (1987), Whitelaw (1994), and Lettau and Ludvigson (2010). This usually involves pro-
jecting future realized returns onto a set of predetermined conditioning variables and taking
the fitted values as a measure of expected returns. While theory tells us the conditional mean
is related to the conditional variance, it does not tell us how these conditional expectations
are formed. In practice, the selection of the conditioning variables is somewhat ad-hoc and
the inference is highly sensitive to the choice of the conditioning variables used in empirical

12
tests (see Harvey, 2001).9
Alternatively, one can obtain estimates of expected returns from survey data. For exam-
ple, the Philadelphia Fed’s Livingston Survey and Duke CFO Survey both provide survey
based measures of expected returns. While survey data has its own limitations, existing stud-
ies show they contain rich economic information about investor expectations. See, among
others, Ben-David, Graham, and Harvey (2013), Greenwood and Shleifer (2014), and Gandhi,
Gormsen, and Lazarus (2023). Nagel and Xu (2023) use survey data to study the so called
‘subjective risk-return tradeoff’, finding that subjective risk premium is increasing in sub-
jective risk perceptions.10

2.4 Option Implied Expected Returns

The primary innovation of our paper is to estimate the risk return relationship by computing
the expected return directly from option prices, without relying on noisy realized returns,
specifying conditioning variables or survey data. We now describe the approach, which
follows Chabi-Yo and Loudis (2020), for extracting expected excess returns (or market risk
premiums) from available option prices. We refer to those option implied expected market
returns as µ̂ot .
Assume a market that is free of arbitrage and hence admits a strictly positive stochastic
discount factor (SDF). We denote this SDF over the period t to t + h as Mt+h . We also
denote expectation, variance and covariance under the physical (risk neutral) measure with
Mt+h
a superscript P (Q). By using the fact that EtP (Mt+h )
defines the Radon-Nikodym derivative
9
Another issue with projections is that they sometimes generate negative estimates for the expected stock
returns.
10
Jo, Lin, and You (2022) survey a large number of U.S respondents and find a negative subjective risk
return trade off for stocks. Related, Liu, Su, Wang, and Yu (2024) study the impact of return extrapolation
on the risk return trade off.

13
that allows one to traverse the risk neutral and physical probability measures, Chabi-Yo
and Loudis (2020) show that the expected excess return under the physical measure can be
linked to risk neutral moments via

 
EtP (Mt+h )
EtP (rt+h ) = CovQ
t rt+h , . (9)
Mt+h

Assuming a single period economy populated by a representative agent endowed with a

rational utility function over terminal wealth, Chabi-Yo and Loudis (2020) demonstrate that
the difference between physical (P) and risk neutral (Q) moments may be expressed via an
infinite series. Defining

MPn = EtP ((rt+h )n ); MQ Q n

n = Et ((rt+h ) ),

these authors show that

P∞ Q Q Q
k=1 θk (Mn+k − Mn Mk )
MPn − MQ
n = (10)
1+ ∞
P Q
k=1 θk Mk

where θk are linked to investors risk preferences associated with the corresponding moment.
Chabi-Yo and Loudis (2020) consider two versions of their bound through the choice of θk .
Their unrestricted bound uses historical data to estimate the values of θk whereas their
restricted bound uses economic arguments to replace θk with functions of the risk-free rate.
We elect to use the restricted bound as it is entirely forward looking, does not suffer from
estimation error and produces outcomes that are almost identical to the unrestricted lower
bound.
Implementing the conditions of the restricted lower bound, setting n = 1 and truncating

14
the infinite sum in (10) to four terms provides the expression

1 1 1
1+rf,t
MQ
2 −
(1+rf,t )2
MQ Q
3 + (1+rf,t )3 M4
EtP (rt+h ) ≥ 1 1
. (11)
1− (1+rf,t )2
MQ Q
2 + (1+rf,t )3 M3

Computing equation (11) provides our estimate of the option implied expected excess re-
turn. The lower bound in equation (11) incorporates rich information embedded in higher
moments and Chabi-Yo and Loudis (2020) show that it provides a better measure of con-
ditional expected return than some of the existing bounds. Values for MQ
k (k = 2, 3, 4)

are computed from observed option prices via discretized versions of the spanning relations
found in Carr and Madan (2001) and Bakshi and Kapadia (2003).11 Computing our option
implied expected return lower bounds requires the index level, option prices across a variety
of strikes and maturities and risk-free rates. We obtain all these inputs over the period
January 1996 to December 2021 from the IvyDB Optionmetrics database. To minimize the
impact of stale prices, we filter our data in several standard ways.12
While we use the lower bound as a proxy for the expected return, it should be noted
that our empirical tests are still valid even if the bound is not tight. In particular, when
the slackness of the bound (true conditional expected return minus the bound) is constant
or averages out, the estimate of γ in equation (2) is unaffected. On the other hand, if
the slackness of the bound is time varying and proportional to the level of the bound, our
estimates of γ represent a lower bound for the risk aversion coefficient.
Panel A of Table 2 reports summary statistics of option implied expected returns. At the
11
The relevant formulas can be found in Appendix A.
12
First, we remove options where the bid or ask price violates arbitrage bounds (call prices are less than
the index price, and put prices are less than the strike multiplied by the risk-free bond price). Next, we
remove options where the bid price is zero and where the bid-ask spread is negative. Finally, we remove
options with zero open interest. To achieve a set of standardized maturities of {1, 2, 3, 6, 12}-months, we
linearly interpolate between the two nearest available maturities.

15
1-month horizon, the option implied expected excess return is on average 40 basis points,
which is economically plausible and quite close to the unconditional average realized returns
reported in Table 1. Consistent with Martin (2017) and Chabi-Yo and Loudis (2020), Panel
A shows that the term structure of expected stock returns is also upward sloping in our
sample, with the mean return increasing monotonically from 40 bps per month for the 1-
month horizon to 49 bps per month for the 1-year horizon.
Figure 1 compares the time series of expected 1-month excess returns of the S&P 500
index with that of realized returns. Panel A shows that option implied expected return
is always positive and varies significantly over time. In particular, it tends to increase in
periods when the marginal utility is arguably high, which is consistent with the notion that
investors require a higher compensation to hold stocks in bad times. In comparison, monthly
realized returns in panel B appear to be a noisy measure of expected returns in that they take
negative values frequently. The expected return is also more persistent than realized return.
The AR(1) coefficients for expected and realized returns are 0.74 and 0.03, respectively.
Figure 2 plots option implied expected excess returns for other horizons. As can be
seen from the plot, expected stock returns follow similar patterns and are highly correlated
across horizons. Another interesting result is that there is a significant variation in the
term structure of expected stock returns. During normal times, expected returns are larger
at longer horizons and therefore the term structure is upward sloping. During times of
market turbulence, however, the term structure can become downward sloping with short
term expected returns higher than their long horizon counterparts.
To investigate the validity of option implied expected returns, we estimate a series of
Mincer-Zarnowitz regressions (Mincer and Zarnowitz, 1969) in which we use option implied
expected excess returns to forecast subsequent realized excess returns over the relevant hori-

16
zons, indexed by h ranging from one month to one year:

rt+h = β0 + β1 µ̂ot+h + ϵ (12)

where, with a slight abuse of notation, rt+h denotes the realized return from (end of) t to
(end of) t + h and µ̂ot+h is the option implied expected return for the same period, estimated
at t using option prices. For an unbiased measure of the expected return, the intercept
should be zero and slope coefficient should be one.
Panel B of Table 2 reports the results for the Mincer-Zarnowitz regressions. Point esti-
mates of the intercept are close to zero and point estimates of β1 are close to 1, suggesting
that option implied returns are valid measures of ex ante expected returns. We also conduct
a joint test of whether the estimated intercept and slope are significantly different from zero
and one. The null hypothesis of β0 = 0 and β1 = 1 cannot be rejected with large p-values in
most cases.

2.5 Estimating the Intertemporal Risk Return Relation

We report two sets of results for estimating the intertemporal risk-return relation to highlight
the important role of measuring the conditional mean. First, we use realized returns as a
proxy for expected returns and replicate the conflicting findings on the mean-variance relation
as reported in existing studies. We then present new evidence on the risk-return relation
using option implied expected returns.

17
2.5.1 Using Realized Returns

We first estimate the following monthly time series regression:

rt+1 = c + γ vˆt + et+1 (13)

where rt+1 is the realized excess return of the index in month t + 1 and vˆt is the conditional
variance at the end of month t. We consider five measures of conditional variance as discussed
in Section 2.2.
Table 3 reports the results for testing the risk return relation using realized returns.
Consistent with existing findings, Panel A shows that the estimated risk return relation
is sensitive to the choice of the conditional variance and there is a weak or even negative
risk-return trade-off in our sample period. When using RV , GV , EGV or M V , the slope
coefficient is estimated to be positive whereas the sign flips for EV .13 Moreover, none of
them is statistically significant. Throughout the paper, t-statistics are computed based on
Newey and West (1987) standard errors unless otherwise stated.14 Panels B and C report the
results for two sub-samples of equal length: 1996 to 2008 and 2009 to 2021. We find the risk
return relation is also sensitive to the sample period used in empirical tests. Specifically, the
first half of the sample is associated with a negative risk return trade-off and the relationship
is statistically significant when using RV and M V , with a Newey-West t-statistic of −2.37
and −2.80. In contrast, we find a positive risk return trade-off in the second half of the
sample and the results are statistically significant in three out of five conditional variance
13
Another issue with estimating the regression in equation (13), which is somewhat less emphasized in the
literature, is the Stambaugh bias (Stambaugh, 1999). The estimated γ is biased upwards because there is a
strong negative contemporaneous correlation between realized returns and changes in conditional variance.
The bias also depends on the persistence of the conditional variance measure.
14
The lag selection is based on the sample length recommended in Newey and West (1994). We also vary
the number of lags used in computing standard errors and find the results overall similar.

18
measures, using t = 1.96 as the cutoff value. It is also worth noting that the R2 of the
regressions tends to be quite small. This is not entirely surprising as much of the variation
in realized return is unrelated to the variation in expected return at the monthly frequency.
In summary, we confirm the existing findings that when using realized returns as a
proxy for expected returns, tests of the risk-return relation are sensitive to the empirical
implementation, especially the way the conditional variance is modeled and the sample
period that is selected. The estimated risk return trade-off is often found to be weak or even
negative.

2.5.2 Using Option Implied Expected Return

Now we re-estimate the same regressions, but using option implied expected excess returns
instead of realized returns:
µ̂ot = c + γv̂t + et+1 (14)

where µ̂ot is the expected excess return of the index we construct from options and v̂t refers to
the five measures of conditional variance as before. An additional advantage of our approach
is that we can compute expected returns for different horizons, which allows us to assess the
risk-return trade-off for different horizons. We focus on the 1-month horizon in this section
and report the findings for other horizons in Section 3.4.
Table 4 reports the results for the regressions in (14). When using option implied ex-
pected excess returns, the evidence on the risk-return relation is drastically different from
the findings based on realized returns in Table 3. In particular, across all conditional vari-
ance specifications, we find a positive and statistically significant relationship between the
conditional mean and conditional variance of the S&P 500 index returns. Panel A shows
that over the full sample, the estimate of the risk aversion coefficient γ is 0.528 for RV with

19
a t-statistic of 6.13, 0.707 for EV (t-stat = 6.48), 2.201 for GV (t-stat = 7.48), 2.192 for
EGV (t-stat = 10.61), and 0.712 for M V (t-stat = 10.12). The mean variance relationship
is also economically significant. One standard deviation increase in RV , EV , GV , EGV ,
and M V is associated with an increase of 0.319%, 0.292%, 0.308%, 0.340%, and 0.370%
per month in the expected return. Another interesting result is that R2 s of regressions are
much higher, ranging from 56.82% to 76.80%. This suggests that a substantial portion of
the variation in expected return is indeed related to variation in variance, which is consistent
with the theoretical frameworks discussed in Section 2.1. This is in sharp contrast to realized
returns whose variation is weakly related to volatility. Panels B and C show that the positive
risk-return relation also holds in two subsamples (1996-2008 and 2009-2021).
Overall our analysis suggests that the measurement of the conditional mean is of first-
order importance in detecting the risk-return trade-off. We show that the mixed findings
prior studies report on the risk return relation are partly due to the use of realized returns
as a proxy for expected returns. We demonstrate that using option implied expected excess
returns allows for a more precise estimate of the risk-return trade-off. When using option
implied expected returns, we find a positive and statistically significant risk-return trade-off
and the evidence is remarkably consistent across all specifications of the conditional variance.

3 Extension and Discussion

This section contains additional discussion and extensions to further shed light on the na-
ture of the comovement between the first two conditional moments of stock market returns.
Section 3.1 investigates whether the risk return trade-off is time varying. Section 3.2 con-
ducts specification analysis by adding additional variables to explain the variation in the
conditional mean. Section 3.3 performs a decomposition of option implied expected return

20
into different components and examines the risk return relation for each component. Section
3.4 studies the term structure of risk-return trade-off. Section 3.5 reports the results for the
indirect test discussed in Section 2.1.

3.1 Is the Risk Return Trade-Off Time Varying?

This section investigates whether the risk return relation is time varying. Yu and Yuan (2011)
demonstrate that investor sentiment has a significant impact on the risk return trade-off.
They find that stock return is positively related to variance in low sentiment periods but
unrelated to variance in high sentiment periods. They argue that sentiment traders tend to
be native and inexperienced and therefore the increasing presence and trading of sentiment
traders undermine the risk-return trade-off.
Panels A and B of Table 5 report the results for testing the risk-return relation using
option implied returns (i.e., regression in (14)) in the low and high sentiment periods based
on the sentiment index of Baker and Wurgler (2006).15 Baker and Wurgler (2006) construct
the investor sentiment index based on the first principal component of five (standardized)
sentiment proxies where each of the proxies has first been orthogonalized with respect to a
set of six macroeconomic indicators.16 Consistent with the finding in Yu and Yuan (2011),
the risk return trade-off is indeed stronger in the low sentiment period, judging by slope
coefficients, statistical significance, and R2 . In contrast to their findings, which are based
on realized returns, we find the risk-return relation remains significant in the high sentiment
period. This suggests that while the impact of sentiment is important, it does not entirely
override the fundamental risk-return trade-off.17
15
High (low) sentiment periods are defined as months with positive (negative) Baker and Wurgler sentiment
index values. There are 154 high sentiment months and 158 low sentiment months in our sample.
16
The five sentiment proxies are closed-end fund discount (cefd), IPO volume (nipo), first-day returns on
IPO (ripo), equity share in new issues (s), and value-weighted dividend premium (pdnd).
17
We also consider different regimes characterized by return and volatility (e.g., low return v.s high return,

21
 We also estimate the following regression with a dummy variable to formally investigate
whether the impact of sentiment is statistically significant:

µ̂ot = c1 + γ1 RVt + c2 Dt + γ2 Dt RVt + et+1 (15)

where Dt is a dummy variable that takes the value of 1 if month t is characterized as a high
sentiment month, and we simply use realized variance as the proxy for conditional variance.
γ2 measures the impact of sentiment on the risk return relation. Over our sample, γ2 is
estimated to be negative (−0.233) and statistically significant, suggesting that high investor
sentiment indeed undermines the risk-return trade off. Moreover, by construction, γ1 gives
the coefficient in the low sentiment periods as reported in Panel A and γ1 + γ2 would recover
the coefficient in high sentiment periods as reported in Panel B.

3.2 Specification Analysis

The baseline regression in (14) is likely to be mis-specified for a number of reasons. First, the
simple result that the market risk premium depends on the conditional variance alone does
not hold in more realistic theoretical set-ups. For example, in the ICAPM world, when the
investment opportunity set is time varying, the agent would desire to hedge unfavorable shifts
in the investment opportunities and therefore the expected return depends on not only the
conditional variance but also the covariance between stock returns and state variables that
drive investment opportunities. Guo and Whitelaw (2006) and Scruggs (1998) find that the
omission of those hedging components can have a significant impact on the estimated risk-
return relation. To address the omitted variable bias, we follow existing studies and control
low volatility v.s high volatility). In those cases, we continue to find a significant and positive mean variance
relation across all regimes, but there is no strong pattern emerging. The relative strength of the risk return
trade-off somewhat depends on the measure of the conditional variance.

22
for several macroeconomic variables that are likely to track the changes in the investment
opportunities:
µot = c + γRVt + ηϕt + et+1 (16)

where ϕt is stock price-earnings (PE) ratio, default spread or unemployment.18 For brevity,
we present our findings using the realized variance. The results are similar with other
measures of the conditional variance.
Whitelaw (1994) documents that the conditional mean and variance exhibit some inter-
esting lead and lag effects, and the estimated risk-return relation appears to be sensitive
to whether one controls for lags of conditional mean and conditional variance (Lettau and
Ludvigson, 2010). Given these findings, we also adopt the following regression specification
by adding lags of conditional mean and conditional variance:

µot = c + γRVt + η0 µot−1 + η1 RVt−1 + et+1 . (17)

Lastly, we assess if there is non-linearity in the risk-return relation by adding the squared
term into our baseline regression:

µot = c + γRVt + ηRVt2 + et+1 . (18)

If the expected excess return is a linear function of its conditional variance, we would expect
that η equals zero.
Table 6 reports the results. Columns (1) to (4) demonstrate that the positive mean-
variance relationship remains highly significant in the presence of macroeconomic variables,
18
We obtain data on price-earnings ratio from Robert Shiller’s website and data on unemployment and
default spread from the Federal Reserve Bank of St. Louis website. The default spread is measured as the
difference in yield between AAA and BAA bonds.

23
regardless of whether we control for them individually or collectively. Column (5) shows
that the positive relationship between the conditional mean and variance is also robust
to controlling for their lagged values. The last column of Table 6 reports the results for
the regression in (18). While γ remains positive and significant, we find the coefficient on
the squared variance is highly significant. Moreover, the increase in the R2 from adding
the squared variance is much larger than that from adding other control variables. These
findings suggest that there is some non-linearity in the risk return relation.

3.3 A Decomposition Analysis

As shown in Section 2.4, the expected excess market return is related to risk neutral variance,
skewness, and kurtosis of the index. In this section, we decompose the expected excess market
returns into three components relating to the three risk neutral moments:

µ̂ot+h = µ̂vt+h + µ̂st+h + µ̂kt+h (19)

where µ̂ot+h is the option implied expected excess return, µ̂vt+h , µ̂st+h , and µ̂kt+h denote the
part of the expected return that is driven by risk neutral volatility, skewness, and kurtosis,
respectively, and we then investigate the risk return relation for each return component.
We compute these values following the approach of Chabi-Yo and Loudis (2020). First, to
calculate µ̂vt+h , we set M3 and M4 equal to zero in equation (11). Next, we find µ̂vt+h + µ̂st+h
by setting only M4 equal to 0 in (11) allowing the direct computation of µ̂st+h . Finally, the
difference between the full computation of (11) and µ̂vt+h + µ̂st+h provides µ̂kt+h . We note that
each of the components in (19) is positive under the assumptions of Chabi-Yo and Loudis
(2020) and is almost always positive empirically, allowing us to examine the proportional

24
contribution made by each risk neutral moment.19
Figure 3 plots the proportion of expected excess market return that each risk neutral
moment contributes over time. Consistent with the findings in Chabi-Yo and Loudis (2020),
Panel A shows that the 1-month expected return is dominated by risk neutral variance. It
appears, however, higher-order moments become increasingly significant over time. Panels B
to D present the corresponding plots for 3-month, 6-month, and 12-month expected returns.
While risk neutral variance remains the biggest contributor, higher moments are increasingly
important at longer horizons and in recent periods. For example, the unconditional average
contribution from risk neutral variance, skewness and kurtosis for 12-month expected returns
are 0.738, 0.152, and 0.110, respectively.
We then test the risk return relation for each component of the expected return via the
following regressions:
µ̂it = c + γv̂t + et+1 i = v, s, k, (20)

where µ̂it denotes the part of the one month expected return that is related to each risk neutral
moment. Table 7 reports the results. There is a positive and statistically significant risk
return relation for each component of the expected return. For example, the slope coefficient
on realized variance is 0.427, 0.062, and 0.039 for µ̂vt , µ̂st , and µ̂kt , all of which are statistically
significant. Notice that, by construction, those estimates add up to the original coefficient
reported in Table 4 (0.528). In terms of t-statistics and R2 , the estimates for different return
components are of similar magnitude. In summary, we find that the positive risk return
relation not only obtains for the part of the expected return relating to risk neutral variance
but also the parts of the return relating to risk neutral higher-order moments.
19
While both µ̂vt+h and µ̂kt+h are positive by construction, we require that M3 < 0 to ensure µ̂st+h > 0. We
find this to be true in approximately 99% of our sample.

25
3.4 The Term Structure of Risk-Return Trade-Off

In the main analysis, we follow the literature and focus on the risk return trade-off at the
monthly horizon. However, our approach allows us to measure the expected market risk
premium over different horizons and therefore we are able to test how the mean-variance
trade-off varies as a function of the return horizon. In this section, we provide new insights
about the term structure of the risk return trade-off using option implied expected excess
returns. For comparison, we also report the results for using realized returns. Moreover, we
simply use past realized variance as a proxy for conditional variance to highlight our focus
on the measurement of the conditional mean. This also facilitates the comparison of our
results with existing findings.
We first follow Bandi and Perron (2008) and estimate the following regressions at different
horizons, indexed by h ranging from one month to one year:

rt+h = c + γRVt−h + et+h (21)

where rt+h and RVt−h are future realized returns and past realized variances over different
horizons. The regression is estimated using overlapping observations for return horizons
longer than one month. Panel A of Table 8 shows that when using realized returns, we find
the mean-variance trade-off is weak at short horizons, whereas the relationship is positive
and statistically significant at the 6-month and 12-month horizons with a Newey-West t-
statistic of 2.78 and 2.53 respectively.20 The risk-return relationship is better revealed at
longer horizons, presumably because variation in expected return is easier to detect at longer
horizons. For example, a large literature shows that stock returns are predictable by a number
20
We also consider the Hansen and Hodrick (1980) standard error to account for the use of overlapping
observations and find similar results.

26
of financial variables at longer horizons (e.g., Campbell, 1987; Fama and French, 1988). Our
results, which are based on a more recent sample and linear regressions, are consistent with
existing studies. Bandi and Perron (2008) show that the relation between past realized
variance and future realized excess market returns is statistically mild at short horizons,
but increases with the horizon and is much stronger in the long run. Harrison and Zhang
(1999) also find the risk-return relation is much stronger at long horizons by estimating
the conditional mean and variances over different horizons using semi-nonparametric density
estimation and Monte Carlo integration. Bonomo, Garcia, Meddahi, and Tédongap (2015)
propose a theoretical model incorporating ambiguity to explain the short and long end of
the risk-return trade-off.
Panel B of Table 8 reports the corresponding results for using option implied expected
returns in estimating the term structure of the risk return relation:

µot+h = c + γRVt−h + et+h (22)

where µot+h and RVt−h are option implied expected returns and past realized variances over
relevant horizons, respectively. When using option implied expected returns, we find a
positive and statistically significant risk-return relation for all horizons and the relation seems
quite stable across horizons. The slope coefficient monotonically rises from 0.528 for the 1-
month horizon to 0.672 for the 12-month horizon, but the magnitude of the increase appears
quite small. The pattern in the R2 of the regressions is opposite: The R2 monotonically
decreases from 67.78% for the 1-month horizon to 54.25% for the 12-month horizon, but
again the magnitude of this drop is modest.

27
3.5 Indirect Tests

In this section, we infer the ex ante relation between the first two conditional moments of
stock returns by examining the ex-post relationship between realized returns and changes in
the conditional variance. In particular, we estimate the following regression:

rt+1 = ρ0 + ρ1 ∆v + et+1 (23)

where ∆v = v̂t+1 − vˆt . If there is a positive relation between the conditional mean and
conditional variance of stock market returns, we would expect to see a negative contempora-
neous relationship between realized returns and variance innovation; that is, ρ1 < 0, because
following an increase in the conditional variance, the expected return increases accordingly,
resulting in an immediate drop in the stock price and hence low realized contemporaneous
return.
Table 9 reports the results for the above regressions. We find realized returns are neg-
atively related to changes in the conditional variance. The negative relation is highly sta-
tistically significant and holds for all measures of the conditional variance. These findings
corroborate our main results using expected returns and provide indirect evidence that the
ex ante relationship between the conditional mean and conditional variance is positive.

4 Conclusion

This paper shows that the measurement of the conditional mean is important in estimating
the risk-return trade-off for the aggregate stock market. The use of realized returns, which are
a poor measure of expected returns, is partially responsible for the conflicting findings on the
relationship between risk and return in existing studies. We present novel empirical evidence

28
on this relationship using expected excess returns extracted from index options. We find a
significantly positive relation between the first two conditional moments of index returns
and the evidence is remarkably consistent across different specifications of the conditional
variance. Further analysis suggests that the mean variance relation appears to be time
varying, nonlinear, holds for different components of the expected return, is robust to control
for other variables, and persists to longer horizons.
Our study can be extended in several ways. First, it may prove interesting to replicate
our findings for the U.S. market using international data. Second, we have so far focused
on the risk-return trade-off for the equity risk premium, and extensions to analyzing higher-
order moment risk premia (e.g., variance and skewness risk premium) would be useful. We
plan to address these in future research.

29
Appendix A: Computing Expected Return Lower Bound

from Option Prices

We present here the formulas used to compute the terms Mk (k = 2, 3, 4) in equation (11).
These are reproduced from Appendix B in Chabi-Yo and Loudis (2020). Define rt+h =
St+h
St
− Rf,t+h where Rf,t+h is the gross risk-free rate from t to t + h and apply the spanning
formula of Carr and Madan (2001) to (rt+h )k . This yields
 k−2 
R Ft+h  K
− rf,t (K − St+h )+ dK
 
k(k − 1) 
0 St

(rt+h )k = k−2
St2

 + ∞ K
R

Ft+h St
− r f,t (St+h − K)+ dK 


where Ft+h is the forward price from t to t + h. Taking expectations gives

k
MQ Q
k = Et ((rt+h ) )
  k−2 
 Ft+h K − rf,t
R
P utt+h (K)dK
 
k(k − 1)Rf,t+h 0 St

= k−2
St2
 
 + ∞ K
R

Ft+h St
− rf,t Callt+h (K)dK 


where P utt+h (K) and Callt+h (K) are the prices of put and call options respectively at t with
maturity t + h and strike K. In applications, we replace the integrals with discrete sums
as is standard in the literature (see Martin, 2017; Martin and Wagner, 2019; Chabi-Yo and
Loudis, 2020, among others).

30
References

Abel, Andrew B., 1988, Stock prices under time-varying dividend risk: An exact solution in
an infinite-horizon general equilibrium model, Journal of Monetary Economics 22, 375–
393.

Adrian, Tobias, Richard K. Crump, and Erik Vogt, 2019, Nonlinearity and flight-to-safety
in the risk-return trade-off for stocks and bonds, Journal of Finance 74, 1931–1973.

Ai, Hengjie, Leyla Jianyu Han, and Lai Xu, 2022, Information-driven volatility, Working
Papper, University of Wisconsin Madison.

Andersen, Torben G., Tim Bollerslev, Francis X. Diebold, and Heiko Ebens, 2001, The
distribution of realized stock return volatility, Journal of Financial Economics 61, 43–76.

Atmaz, Adem, and Suleyman Basak, 2022, Stock Market and No-Dividend Stocks, The
Journal of Finance 77, 545–599.

Back, Kerry, Kevin Crotty, and Seyed Mohammad Kazempour, 2022, Validity, tightness, and
forecasting power of risk premium bounds, Journal of Financial Economics 144, 732–760.

Backus, David K., and Allan W. Gregory, 1993, Theoretical relations between risk premiums
and conditional variances, Journal of Business and Economic Statistics 11, 177–185.

Baker, Malcolm, and Jeffrey Wurgler, 2006, Investor sentiment and the cross-section of stock
returns, Journal of Finance 61, 1645–1680.

Bakshi, Gurdip, John Crosby, Xiaohui Gao, and Wei Zhou, 2019, A new formula for the
expected excess return of the market, Fox School of Business Research Paper.

31
Bakshi, Gurdip, and Nikunj Kapadia, 2003, Delta-hedged gains and the negative market
volatility risk premium, Review of Financial Studies 16, 527–566.

Bali, Turan G., and Lin Peng, 2006, Is there a risk-return trade-off? Evidence from high-
frequency data, Journal of Applied Econometrics 21, 1169–1198.

Bandi, Federico M., and Benoıˆt Perron, 2008, Long-run risk-return trade-offs, Journal of
Econometrics 143, 349–374.

Bansal, Ravi, and Amir Yaron, 2004, Risks for the long-run: A potential resolution of asset
pricing puzzles, Journal of Finance 59, 1481–1509.

Ben-David, Itzhak, John R. Graham, and Campbell R. Harvey, 2013, Managerial miscali-
bration, Quarterly Journal of Economics 128, 1547–1584.

Berkman, Henk, and Hamish Malloch, 2023, Stock valuation during the COVID-19 pan-
demic: An explanation using option-based discount rates, Journal of Banking & Finance
147, 106386 Special Issue: The Impact of Global Pandemic on Financial Markets and
Institutions.

Black, Fischer, 1976, Studies of stock price volatility changes, Proceedings of the 1976 Meet-
ings of the American Statistical Association, Business and Economic Statistics Section
pp. 177–181.

Bonomo, Marco, René Garcia, Nour Meddahi, and Roméo Tédongap, 2015, The long and
the short of the risk-return trade-off, Journal of Econometrics 187, 580–592.

Brandt, Michael W., and Qiang Kang, 2004, On the relationship between the conditional
mean and volatility of stock returns: A latent VAR approach, Journal of Financial Eco-
nomics 72, 217–257.

32
Campbell, John Y., 1987, Stock returns and the term structure, Journal of Financial Eco-
nomics 18, 373–399.

Campbell, John Y., and John H. Cochrane, 1999, By force of habit: A consumption-based
explanation of aggregate stock market behavior, Journal of Political Economy 107, 205–
251.

Campbell, John Y., and Ludger Hentschel, 1992, No news is good news: An asymmetric
model of changing volatility in stock returns, Journal of Financial Economics 31, 281–
318.

Campbell, John Y., and Robert J. Shiller, 1988, The dividend-price ratio and expectations
of future dividends and discount factors, Review of Financial Studies 1, 195–228.

Carr, Peter, and Dilip Madan, 2001, Optimal positioning in derivative securities, Quantita-
tive Finance 1, 19–37.

Chabi-Yo, Fousseni, Chukwuma Dim, and Grigory Vilkov, 2023, Generalized bounds on the
conditional expected excess return on individual stocks, Management Science 69, 922–939.

Chabi-Yo, Fousseni, and Johnathan Loudis, 2020, The conditional expected market return,
Journal of Financial Economics.

Christie, Andrew A., 1982, The stochastic behavior of common stock variances: Value,
leverage and interest rate effects, Journal of Financial Economics 10, 407–432.

Cieslak, Anna, Adair Morse, and Annette Vissing-Jorgensen, 2019, Stock returns over the
FOMC cycle, Journal of Finance 74, 2201–2248.

Cochrane, John H., 2017, Macro-finance, Review of Finance 21, 945–985.

33
Corsi, Fulvio, 2009, A simple approximate long-memory model of realized volatility, Journal
of Financial Econometrics 7, 174–196.

Elton, Edwin J., 1999, Presidential address: expected return, realized return, and asset
pricing tests, Journal of Finance 54, 1199–1220.

Eraker, Bjørn, 2004, Do stock prices and volatility jump? Reconciling evidence from spot
and option prices, Journal of Finance 59, 1367–1404.

Fama, Eugene F, and Kenneth R French, 1988, Dividend yields and expected stock returns,
Journal of Financial Economics 22, 3–25.

French, Kenneth R., G. William Schwert, and Robert F. Stambaugh, 1987, Expected stock
returns and volatility, Journal of Financial Economics 19, 3–29.

Gandhi, Mihir, Niels Joachim Gormsen, and Eben Lazarus, 2023, Forward Return Expecta-
tions, Working paper, Working Paper, University of Chicago.

Ghysels, Eric, Pedro Santa-Clara, and Rossen Valkanov, 2005, There is a risk-return trade-off
after all, Journal of Financial Economics 76, 509–548.

Glosten, L.R., R. Jagannathan, and D. Runkle, 1993, On the relation between the expected
value and the volatility of the nominal excess return on stocks, Journal of Finance 48,
1779–1801.

Greenwood, Robin, and Andrei Shleifer, 2014, Expectations of returns and expected returns,
Review of Financial Studies 27, 714–746.

Guo, Hui, Qian Lin, and Yu-Jou Pai, 2024, A tale of fear and euphoria in the stock market,
Journal of Financial and Quantitative Analysis, Forthcoming.

34
Guo, Hui, and Robert F. Whitelaw, 2006, Uncovering the risk-return relation in the stock
market, Journal of Finance 61, 1433–1463.

Hansen, Lars Peter, and Robert J. Hodrick, 1980, Forward exchange rates as optimal pre-
dictors of future spot rates: An econometric analysis, Journal of Political Economy 88,
829–853.

Harrison, Paul, and Harold H. Zhang, 1999, An investigation of the risk and return relation
at long horizons, Review of Economics and Statistics 81, 399–408.

Harvey, Campbell R., 2001, The specification of conditional expectations, Journal of Em-
pirical Finance 8, 573–637.

Jo, Chanik, Chen Lin, and Yang You, 2022, Subjective risk-return trade-off, Working Papper,
Chinese University of Hong Kong.

Kadan, Ohad, and Xiaoxiao Tang, 2020, A bound on expected stock returns, Review of
Financial Studies 33, 1565–1617.

Lettau, Martin, and Sydney C. Ludvigson, 2010, Measuring and modeling variation in the
risk-return trade-off, Handbook of Financial Econometrics: Tools and Techniques pp. 617–
690.

Lintner, John, 1965, Security prices, risk and maximal gains from diversification, Journal of
Finance 20, 587–615.

Liu, Qi, Zhiwei Su, Huijun Wang, and Jianfeng Yu, 2024, Extrapolation and risk-return
trade-offs, PBCSF-NIFR Research Paper Forthcoming.

Lochstoer, Lars, and Tyler Muir, 2022, Volatility expectations and returns, Journal of Fi-
nance 77, 1055–1060.

35
Ludvigson, Sydney C., and Serena Ng, 2007, The empirical risk-return relation: A factor
analysis approach, Journal of Financial Economics 83, 171–222.

Lundblad, Christian, 2007, The risk return tradeoff in the long run: 1836-2003, Journal of
Financial Economics 85, 123–150.

Martin, Ian, 2017, What is the expected return on the market?, Quarterly Journal of Eco-
nomics 132, 367–433.

Martin, Ian W. R., and Christian Wagner, 2019, What is the Expected Return on a Stock?,
Journal of Finance 74, 1887–1929.

Mehra, Rajnish, and Edward Prescott, 1985, The equity premium puzzle, Journal of Mon-
etary Economics 15, 145–161.

Merton, Robert C., 1973, An intertemporal capital asset pricing model, Econometrica 41,
867–887.

Merton, Robert C., 1980, On estimating the expected return on the market: An exploratory
investigation, Journal of Financial Economics 8, 323–361.

Mincer, Jacob A., and Victor Zarnowitz, 1969, The evaluation of economic forecasts, in
(Economic forecasts and expectations: Analysis of forecasting behavior and performance,
).

Nagel, Stefan, and Zhengyang Xu, 2023, Dynamics of subjective risk premia, Journal of
Financial Economics 150, 103713.

Newey, Whitney, and Kenneth West, 1987, A simple, positive semi-definite, heteroskedastic-
ity and autocorrelation consistent covariance matrix, Econometrica 55, 703–708.

36
Newey, Whitney K., and Kenneth D. West, 1994, Automatic lag selection in covariance
matrix estimation, Review of Economic Studies 61, 631–653.

Pastor, Lubos, Meenakshi Sinha, and Bhaskaran Swaminathan, 2008, Estimating the in-
tertemporal risk-return tradeoff using the implied cost of capital, Journal of Finance 63,
2859–2897.

Ross, Steve, 2015, The recovery theorem, Journal of Finance 70, 615–648.

Sautner, Zacharias, Laurence Van Lent, Grigory Vilkov, and Ruishen Zhang, 2023, Pricing
climate change exposure, Management Science.

Schwert, William G., 1989, Why does stock market volatility change over time?, Journal of
Finance 44, 1115–1153.

Scruggs, John T., 1998, Resolving the puzzling intertemporal relation between the market
risk premium and conditional market variance: A two-factor approach, The Journal of
Finance 53, 575–603.

Sharpe, W., 1964, Capital asset prices: A theory of market equilibrium under conditions of
risk, Journal of Finance 19, 425–444.

Stambaugh, Robert F., 1999, Predictive regressions, Journal of Financial Economics 54,
375–421.

Turner, Christopher M., Richard Startz, and Charles R. Nelson, 1989, A Markov model of
heteroskedasticity, risk, and learning in the stock market, Journal of Financial Economics
25, 3–22.

Whitelaw, Robert F., 1994, Time variations and covariations in the expectation and volatility
of stock market returns, Journal of Finance 49, 515–541.

37
Whitelaw, Robert F., 2000, Stock market risk and return: an equilibrium approach, Review
of Financial Studies 13, 521–547.

Yang, Aoxiang, 2024, Understanding negative risk-return trade-offs, Working Paper, Peking
University.

Yu, Jianfeng, and Yu Yuan, 2011, Investor sentiment and the mean-variance relation, Journal
of Financial Economics 100, 367–381.

38
 Table 1: Summary Statistics of S&P 500 Returns and Variances

MEAN×102 STD×102 SKEW KURT

S&P 0.587 4.385 -0.617 1.151
RV 0.309 0.605 7.299 66.583
EV 0.309 0.414 6.108 46.176
GV 0.203 0.140 1.892 4.374
EGV 0.195 0.155 2.952 11.712
MV 0.324 0.438 4.048 18.753

Notes: This table reports mean (MEAN), standard deviation (STD), skewness (SKEW), and kurtosis
(KURT) of monthly excess returns and variances of the S&P 500 index. We consider five variance mea-
sures including realized variance (RV ), the expected variance estimated from the HAR model (EV ), two
conditional variances estimated from GARCH(1,1) and EGARCH(1,1) models (GV and EGV ), and the
MIDAS variance estimator (M V ). Sample period: January 1996 to December 2021.

39
 Table 2: Option Implied Expected Excess Market Returns

Panel A: Summary Statistics

MEAN×102 STD×102 SKEW KURT

30 0.400 0.387 3.505 16.573
60 0.428 0.371 3.247 14.711
90 0.440 0.348 3.041 13.245
180 0.473 0.314 2.585 10.393
360 0.492 0.288 2.562 9.931

Panel B: Mincer-Zarnowitz Regressions

β0 β1 p-value R2
30 0.003 0.645 0.697 0.32%
60 0.001 1.193 0.696 1.96%
90 0.001 1.155 0.650 2.47%
180 -0.003 1.710 0.070 8.16%
360 -0.000 1.194 0.557 5.74%

Notes: Panel A reports summary statistics of option implied expected excess returns for the S&P 500 index
over 30-day, 60-day, 90-day, 180-day, and 360-day horizons. Panel B reports the results for the Mincer-
Zarnowitz predictive regressions in which we use option implied expected returns to forecast future realized
returns over relevant horizons. Sample period: January 1996 to December 2021.

40
 Table 3: Estimating the Risk-Return Relation: Using Realized Returns

Panel A: 1996 to 2021

RV EV GV EGV MV
c 0.005 0.008 -0.000 0.003 0.005
(1.69) (2.17) (-0.05) (0.68) (1.31)
γ 0.083 -0.815 2.889 1.138 0.258
(0.08) (-0.60) (0.97) (0.37) (0.20)
R2 0.01% 0.59% 0.85% 0.16% 0.07%

Panel B: 1996 to 2008

RV EV GV EGV MV
c 0.004 0.005 0.003 0.003 0.006
(1.26) (1.04) (0.37) (0.46) (1.64)
γ -1.338 -1.454 -1.396 -1.469 -1.798
(-2.37) (-1.23) (-0.36) (-0.44) (-2.80)
R2 3.04% 2.79% 0.18% 0.32% 2.69%

Panel C: 2009 to 2021

RV EV GV EGV MV
c 0.007 0.005 -0.002 -0.002 0.005
(2.58) (1.28) (-0.45) (-0.31) (1.27)
γ 1.505 2.398 6.813 7.620 1.953
(4.08) (1.56) (3.28) (2.44) (1.88)
R2 4.94% 2.22% 5.52% 5.52% 4.64%

Notes: This table reports the results for the following regression:

rt+1 = c + γ vˆt + et+1

where rt+1 is the realized excess return of the index in month t + 1 and vˆt refers to the five measures of the
conditional variance discussed in Table 1. Panel A reports the results for the full sample: January 1996 to
December 2021 and Panels B and C report on two sub-samples: 1996 to 2008 and 2009 to 2021. Newey-West
t-statistics that adjust for autocorrelation and heteroskedasticity are reported in parentheses.

41
 Table 4: Estimating the Risk-Return Relation: Using Option Implied Returns

Panel A: 1996 to 2021

RV EV GV EGV MV
c 0.002 0.002 -0.000 -0.000 0.002
(8.05) (5.47) (-0.97) (-0.85) (7.84)
γ 0.528 0.707 2.201 2.192 0.712
(6.13) (6.48) (7.48) (10.61) (10.12)
R2 67.78% 56.82% 63.23% 76.80% 64.66%

Panel B: 1996 to 2008

RV EV GV EGV MV
c 0.002 0.002 -0.001 -0.000 0.002
(10.37) (6.64) (-1.39) (-0.64) (4.36)
γ 0.615 0.587 2.550 2.091 0.816
(14.80) (6.46) (6.03) (7.36) (6.90)
R2 81.28% 57.46% 74.22% 81.95% 70.14%

Panel C: 2009 to 2021

RV EV GV EGV MV
c 0.003 0.001 -0.000 -0.000 0.002
(7.38) (1.98) (-0.09) (-1.40) (7.63)
γ 0.447 1.197 1.885 2.452 0.627
(4.48) (9.72) (8.21) (10.07) (8.39)
R2 55.09% 70.23% 53.39% 72.47% 60.66%

Notes: This table reports the results for the following regression:

µ̂ot = c + γ vˆt + et+1

where µ̂ot is the option implied expected excess return of the index for month t + 1, computed using options
at the end of month t, and vˆt refers to the five measures of the conditional variance discussed in Table 1.
Panel A reports the results for the full sample: January 1996 to December 2021 and Panels B and C report
on two sub-samples: 1996 to 2008 and 2009 to 2021. Newey-West t-statistics that adjust for autocorrelation
and heteroskedasticity are reported in parentheses.

42
 Table 5: Time Varying Risk Return Relation: The Role of Sentiment

Panel A: Low Sentiment

RV EV GV EGV MV
c 0.002 0.002 -0.001 -0.000 0.001
(9.54) (5.05) (-1.52) (-1.33) (5.47)
γ 0.641 0.671 2.369 2.319 0.751
(11.55) (5.28) (6.38) (11.39) (8.22)
R2 84.06% 59.78% 75.91% 85.89% 81.14%

Panel B: High Sentiment

RV EV GV EGV MV
c 0.003 0.002 0.000 0.000 0.002
(8.18) (2.61) (0.51) (0.08) (5.78)
γ 0.408 0.800 1.921 1.993 0.636
(5.71) (3.53) (5.76) (7.22) (4.04)
R2 50.41% 52.87% 44.75% 63.34% 39.99%

Panel C: A Joint Regression

c1 γ1 c2 γ2 R2
RV 0.002 0.641 0.001 -0.233 70.93%
(8.49) (12.86) (1.99) (-2.76)

Notes: Panels A and B report the results for the regression:

µ̂ot = c + γ vˆt + et+1

for low and high investor sentiment periods using the sentiment index of Baker and Wurgler (2006). vˆt refers
to the five measures of the conditional variance discussed in Table 1. Panel C reports the results for

µ̂ot = c1 + γ1 RVt + c2 Dt + γ2 Dt RVt + et+1 .

Newey-West t-statistics that adjust for autocorrelation and heteroskedasticity are reported in parentheses.
Sample period: January 1996 to December 2021.

43
 Table 6: Assessing The Omitted Variable Bias

(1) (2) (3) (4) (5) (6)

c 0.003 0.001 0.002 -0.003 0.001 0.001
(2.14) (0.91) (7.94) (-2.24) (6.40) (8.82)
γ 0.527 0.460 0.529 0.452 0.380 0.947
(6.10) (6.07) (6.08) (6.29) (5.08) (13.16)
PE -0.000 0.000
(-0.15) (2.90)
DS 0.002 0.003
(2.79) (3.97)
UNE 0.002 0.014
(0.24) (2.07)
RV lag -0.038
(-0.82)
µo lag 0.414
(5.16)
RV 2 -8.372
(-6.97)
R2 67.69% 71.19% 67.68% 72.36% 75.79% 76.91%

Notes: This table assesses the significance of the mean-variance relationship controlling for other variables.
For control variables, we consider stock price-earnings ratio (P E), default spread (DS), unemployment
rate(U N E), lagged values of the conditional mean and variance (µo lag and RV lag), as well as the squared
variance (RV 2 ). Newey-West t-statistics that adjust for autocorrelation and heteroskedasticity are reported
in parentheses. Sample period: January 1996 to December 2021.

44
 Table 7: Decomposition Analysis

Panel A: µ̂vt
RV EV GV EGV MV
c 0.002 0.002 -0.000 -0.000 0.002
(8.48) (6.19) (-0.58) (-0.22) (8.65)
γ 0.427 0.577 1.840 1.827 0.589
(5.50) (6.64) (7.95) (12.05) (9.71)
R2 65.12% 55.61% 64.78% 78.30% 64.72%

Panel B: µ̂st
RV EV GV EGV MV
c 0.000 0.000 -0.000 -0.000 0.000
(6.02) (2.35) (-1.84) (-1.90) (3.72)
γ 0.062 0.080 0.224 0.226 0.075
(16.03) (5.02) (5.93) (6.69) (8.82)
R2 68.04% 52.88% 47.83% 59.55% 52.71%

Panel C: µ̂kt
RV EV GV EGV MV
c 0.000 -0.000 -0.000 -0.000 -0.000
(1.19) (-0.67) (-2.66) (-2.91) (-1.00)
γ 0.039 0.050 0.138 0.139 0.049
(6.91) (5.93) (4.48) (5.07) (7.98)
R2 75.16% 57.65% 50.59% 63.03% 61.88%

Notes: We decompose option implied expected return (µ̂ot ) into three components that are related to risk
neutral variance (µ̂vt ), risk neutral skewness (µ̂st ), and risk neutral kurtosis (µ̂kt ) of the index and report the
results for the regressions of each return component on the five conditional variance measures. Newey-West
t-statistics that adjust for autocorrelation and heteroskedasticity are reported in parentheses. Sample period:
January 1996 to December 2021.

45
 Table 8: The Term Structure of the Risk Return Relation

Panel A: Using Realized Returns

Horizon 1m 2m 3m 6m 12m
c 0.005 0.005 0.005 0.002 0.002
(1.69) (1.93) (1.88) (0.98) (0.70)
γ 0.083 0.077 0.172 1.017 1.111
(0.08) (0.09) (0.20) (2.78) (2.53)
R2 0.01% 0.02% 0.11% 4.68% 5.99%

Panel B: Using Option Implied Returns

Horizon 1m 2m 3m 6m 12m
c 0.002 0.003 0.003 0.003 0.003
(8.05) (8.51) (8.81) (9.70) (7.79)
γ 0.528 0.564 0.574 0.614 0.672
(6.13) (6.57) (7.07) (10.27) (6.15)
R2 67.78% 66.19% 64.17% 61.09% 54.25%

Notes: Panel A reports the results for the following regression:

rt+h = c + γRVt−h + et+h

where rt+h and RVt−h are future realized returns and past realized variances over different horizons indexed
by h. Panel B reports the corresponding results for using option implied expected excess returns:

µot+h = c + γRVt−h + et+h .

Newey-West t-statistics that adjust for autocorrelation and heteroskedasticity are reported in parentheses.
Sample period: January 1996 to December 2021.

46
 Table 9: Estimating the Risk-Return Relation: Indirect Test

RV EV GV EGV MV
ρ0 × 102 0.581 0.579 0.579 0.578 0.582
(2.45) (2.36) (2.81) (3.00) (2.82)
ρ1 -2.85 -4.28 -28.73 -29.31 -9.39
(-5.16) (-3.56) (-6.60) (-7.89) (-5.26)
R2 14.78% 16.66% 17.02% 56.40% 16.32%

Notes: This table reports the results for the following regression:

rt+1 = ρ0 + ρ1 ∆v + et+1

where rt+1 is the realized excess return of the index in month t + 1 and ∆v is the change in the conditional
variance from month t to month t+1: ∆v = v̂t+1 − vˆt . Newey-West t-statistics that adjust for autocorrelation
and heteroskedasticity are reported in parentheses. Sample period: January 1996 to December 2021.

47
 Figure 1: Expected and Realized Returns of the S&P 500

Panel A: Option Implied Expected 1-Month Excess Returns

0.04

0.035

0.03

0.025

0.02

0.015

0.01

0.005

0
1997 2000 2002 2005 2007 2010 2012 2015 2017 2020

Panel B: Monthly Realized Excess Returns

0.2

0.15

0.1

0.05

-0.05

-0.1

-0.15

-0.2
1997 2000 2002 2005 2007 2010 2012 2015 2017 2020

Notes: Panel A plots option implied monthly expected excess returns for the S&P 500 index. Panel B plots
monthly realized excess returns for the index. Sample period: January 1996 to December 2021.

48
 Figure 2: Option Implied Expected Returns: Other Horizons

0.04
60-day
90-day
0.035 180-day
360-day

0.03

0.025

0.02

49
0.015

0.01

0.005

0
1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021

Notes: This figure plots option implied expected excess returns for the S&P 500 index over 60-day, 90-day, 180-day, and 360-day
horizons. Sample period: January 1996 to December 2021.
 Figure 3: Decomposition of Expected Excess Market Returns

Panel A: 1-Month Returns Panel B: 3-Month Returns

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 Variance 0.2 Variance

Skewness Skewness
0.1 Kurtosis 0.1 Kurtosis

0 0
1997 2000 2002 2005 2007 2010 2012 2015 2017 2020 1997 2000 2002 2005 2007 2010 2012 2015 2017 2020

Panel C: 6-Month Returns Panel D: 12-Month Returns

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
Variance Variance
Skewness Skewness
0.1 0.1
Kurtosis Kurtosis
0 0
1997 2000 2002 2005 2007 2010 2012 2015 2017 2020 1997 2000 2002 2005 2007 2010 2012 2015 2017 2020

Notes: This figure plots the proportion of expected excess market return that each risk neutral moment
contributes over time. Panel A for 1-month expected return, Panel B for 3-month expected return, Panel C
for 6-month expected return and Panel D for 12-month expected return. Sample period: January 1996 to
December 2021.

50