Variance Risk Premiums

Peter Carr
Bloomberg LP and Courant Institute, New York University

Liuren Wu
Zicklin School of Business, Baruch College

We propose a direct and robust method for quantifying the variance risk premium on
financial assets. We show that the risk-neutral expected value of return variance, also
known as the variance swap rate, is well approximated by the value of a particular portfolio
of options. We propose to use the difference between the realized variance and this synthetic
variance swap rate to quantify the variance risk premium. Using a large options data set,
we synthesize variance swap rates and investigate the historical behavior of variance risk
premiums on five stock indexes and 35 individual stocks. (JEL G10, G12, G13)

It has been well documented that return variance is stochastic. When investing
in a security, an investor faces at least two sources of uncertainty, namely the
uncertainty about the return as captured by the return variance, and the uncer-
tainty about the return variance itself. It is important to know how investors deal
with the uncertainty in return variance to effectively manage risk and allocate
assets, to accurately price and hedge derivative securities, and to understand
the behavior of financial asset prices in general.
We develop a direct and robust method for quantifying the return variance
risk premium on an asset using the market prices of options written on this
asset. Our method uses the notion of a variance swap, which is an over-the-
counter contract that pays the difference between a standard estimate of the
realized variance and the fixed variance swap rate. Since variance swaps cost
zero to enter, the variance swap rate represents the risk-neutral expected value
of the realized variance. We show that the variance swap rate can be synthesized
accurately by a particular linear combination of option prices. We propose to

We thank Yacine Aı̈t-Sahalia (the editor), an anonymous referee, and Turan Bali, David Bates, Menachem
Brenner, Mikhail Chernov, Robert Engle, Stephen Figlewski, Rajna Gibson, Alfredo Ibanez, George Jiang, Dilip
Madan, Steven Posner, Anders Bjerre Trolle, Benjamin Wurzburger, Jin Zhang, Chu Zhang, and seminar partici-
pants at the 5th Conference on Financial Risks at Verona (Italy), the 2005 American Finance Association meetings,
Baruch College, Hong Kong University of Science and Technology, New York University, and University of
Zurich for comments. We also thank David Hait and OptionMetrics for providing the option data, Alex Mayus
for clarifying trading practices, and Rui Yao for help on data processing. We assume full responsibility for any
remaining errors. Liuren Wu gratefully acknowledges the support by a grant from The City University of New
York PSC-CUNY Research Award Program. Send correspondence to Liuren Wu, Department of Economics
and Finance, Zicklin School of Business, Baruch College, CUNY, One Bernard Baruch Way, Box B10-225,
New York, NY 10010, telephone: (646) 312-3509, fax: (646) 312-3451. E-mail: Liuren Wu@[Link].

C The Author 2008. Published by Oxford University Press on behalf of The Society for Financial Studies.
All rights reserved. For Permissions, please e-mail: [Link]@[Link]
doi:10.1093/rfs/hhn038 Advance Access publication April 10, 2008
The Review of Financial Studies / v 22 n 3 2009

use the difference between the ex post realized variance and this synthetic
variance swap rate to quantify the variance risk premium.
Using a large options data set, we synthesize variance swap rates on five
stock indexes and 35 individual stocks over a seven-year period. We compare
the synthetic variance swap rates to realized variance, and study the historical
behaviors of variance risk premium on different assets.
We find that the average variance risk premiums are strongly negative for
the S&P 500 and 100 indexes and for the Dow Jones Industrial Average.
The estimates on individual stocks show large cross-sectional variation. We
conjecture that there exists a common stochastic variance risk factor in the
stock market that asks for a highly negative risk premium. When we use the
variance on the S&P 500 index as a proxy for this common variance risk factor
and estimate a variance beta for each stock by regressing the stock’s return
variance on the index variance, we find that the variance risk premiums are
more negative for stocks with higher variance beta. The negative sign on the
variance risk premium indicates that variance buyers are willing to accept a
negative average excess return to hedge away upward movements in stock
market volatility. In other words, investors regard increases in market volatility
as unfavorable shocks to the investment opportunity.
Return variance varies stochastically either due to its correlation with the
stock price or return (e.g., the constant elasticity of variance model of Cox
(1996) and the local volatility model of Dupire (1994); and Derman and Kani
(1994)), or due to its independent variation as a separate source of risk (e.g., the
stochastic volatility models of Heston (1993); and Hull and White (1987)), or
both. Accordingly, variance risk premiums can come from either its correlation
with the return risk and return risk premium, or a separate premium on the
independent variance variation, or both. We investigate whether the classic
capital asset pricing model can explain the negative variance risk premiums.
We find that the negative correlation between stock index returns and the return
variance generates a strongly negative beta, but this negative beta only explains
a small portion of the negative variance risk premiums. Other risk factors
identified by the recent literature, such as size, book-to-market, and momentum,
cannot explain the strongly negative variance risk premiums, either. Therefore,
we conclude that the majority of the market variance risk premium is generated
by an independent variance risk factor.
We also analyze the dynamics of the variance risk premiums by formulat-
ing expectation hypothesis regressions. Under the null hypothesis of constant
variance risk premiums, a regression of the realized variance on the variance
swap rate generates a slope estimate of 1. However, the slope estimates from
our regressions are significantly lower than one for the S&P indexes, the Dow
Jones Industrial Average, and also for many of the individual stocks, suggesting
that the market variance risk premiums are time-varying and correlated with the
variance swap rate. Nevertheless, when we regress the log-realized variance on
the log variance swap rate, the slope estimates are much closer to 1, suggesting

1312
Variance Risk Premiums

that although the log variance risk premiums are strongly negative, they are not
strongly correlated with the logarithm of the variance swap rate.
We check the robustness of our results from several aspects. First, we use
numerical analysis to gauge the magnitude of approximation errors in synthe-
sizing the variance swap rates due to jumps and discretization. We find that
under commonly used models and model parameters, the approximation er-
rors from the two sources are small. Second, we measure the impacts of the
options bid-ask spreads on the variance risk premium estimates, and find that
the variance risk premiums on S&P and Dow indexes remain strongly nega-
tive, regardless of whether we synthesize the variance swap rates using bid,
mid, or ask option prices. Third, we evaluate the error-in-variable issue in our
expectation hypothesis regressions. We find that measurement errors in the
synthetic variance swap rate do bias our slope estimates toward zero, but that
our general conclusions remain valid after correcting for the biases: the market
variance risk premiums are time-varying and correlated with the variance swap
rate when defined in dollar terms, but become closer to an independent series
when defined in log returns. Finally, we divide our data into two subsample
periods, with one corresponding broadly to a bullish market and the other to a
bearish market. We find that the variance risk premiums on stock indexes are
significantly negative under both bullish and bearish market conditions.
In related works, Bakshi and Kapadia (2003a,b) consider the profit and loss
arising from delta-hedging a long position in a call option. They argue that
this profit and loss is approximately neutral to the directional movement of the
underlying asset return, but is sensitive to the movement in the return volatility.
Thus, by analyzing the profit and loss from these delta-hedged positions, they
can infer useful qualitative properties for the variance risk premiums without
referring to a specific model. Our approach maintains and enhances the robust-
ness of their model-free approach, as we provide a quantitative measure of the
variance risk premiums. As a result, we can analyze not only the sign, but also
the quantitative properties of the premiums.
Bates (1996, 2000, 2003); Pan (2002); Jones (2003); and Eraker (2004)
analyze variance risk premiums in conjunction with return risk premiums by
estimating various parametric option pricing models with either Bayesian meth-
ods or efficient methods of moments. Most recently, Aı̈t-Sahalia and Kimmel
(2007) propose a maximum likelihood method for estimating stochastic volatil-
ity dynamics and volatility risk premiums based on closed-form approximations
(developed in Aı̈t-Sahalia (2002, 2008)) to the true likelihood function of the
joint observations on the underlying asset and option prices. Wu (2005) pro-
poses to estimate the variance dynamics and variance risk premiums without
specifying the return dynamics using realized variance estimators from high-
frequency return data and variance swap rates synthesized from option prices.
Bollerslev, Gibson, and Zhou (2004) construct a risk aversion index using real-
ized variance estimators and the VIX, which approximates the 30-day variance
swap rate on the S&P 500 index (Carr and Wu 2006).

1313
 The Review of Financial Studies / v 22 n 3 2009

Ang et al. (2006) form stock portfolios ranked by their sensitivity to volatility
risk and analyze the difference among these different stock portfolios. From the
analysis, they infer the impact of volatility risk on the expected stock return.
Coval and Shumway (2001) study how returns on option investment vary with
strike choices and whether the classic capital asset pricing theory can explain
the returns. Bondarenko (2004) links the market price of variance risk to hedge
fund behavior.
Our empirical analysis of the variance risk premiums is based on our theoret-
ical work on synthesizing a variance swap using European options and futures
contracts. Carr and Madan (1998); Demeterfi et al. (1999); and Britten-Jones
and Neuberger (2000) use the same replicating strategy, but under the assump-
tion of continuity in the underlying asset price. Jiang and Tian (2005) extend the
result to a jump-diffusion stochastic volatility model. Our derivation is under
the most general setting possible. We also quantify the approximation error
induced by jumps. Most importantly, we exploit the theoretical developments
in synthesizing variance swaps for variance risk premium analysis.
The remainder of the paper is organized as follows. Section 1 lays out the
theoretical foundation on how we synthesize the variance swap from vanilla
options and how we infer the variance risk premiums based on the difference
between the synthetic variance swap rate and the realized return variance.
Section 2 describes the data and the methodologies used to synthesize the
variance swap rates and to calculate the realized variance and variance risk
premiums. Section 3 investigates the historical behavior of the variance risk
premiums. Section 4 performs robustness analysis. Section 5 concludes.

1. Variance Swap Rates and Variance Risk Premiums

A return variance swap has zero net market value at entry. At maturity, the
payoff to the long side of the swap is equal to the difference between the
realized variance over the life of the contract and a constant called the variance
swap rate,


RVt,T − SWt,T L , (1)

where RVt,T denotes the realized annualized return variance between time t
and T , SWt,T denotes the fixed variance swap rate that is determined at time t
and paid at time T , and L denotes the notional dollar amount that converts the
variance difference into a dollar payoff. No arbitrage dictates that the variance
swap rate equals the risk-neutral expected value of the realized variance


SWt,T = EQ t RVt,T , (2)

where EQ t [·] denotes the time-t conditional expectation operator under some
risk-neutral measure Q.

1314
Variance Risk Premiums

1.1 Synthesizing variance swap rates from options

We use St to denote the time-t spot price of an asset, and Ft its time-t futures
price of maturity T > t. We assume that the futures contract is marked to market
continuously. No arbitrage dictates that there exists a risk-neutral probability
measure Q defined on a probability space (
, F, Q) such that the futures price
Ft solves the following stochastic differential equation:

 
d Ft = Ft− σt− dWt + Ft− e x − 1 [µ(d x, dt) − νt (x)d xdt] , (3)
R0

where Wt is a Q-standard Brownian motion, R0 denotes the real line excluding

zero, Ft− denotes the futures price just prior to any jump at time t, and the
random counting measure µ(d x, dt) realizes to a nonzero value for a given x
if and only if the futures price jumps from Ft− to Ft = Ft− e x at time t. The
process νt (x) compensates the jump process, so that the last term in Equation (3)
is the increment of a Q-pure jump martingale. Therefore, Equation (3) models
the futures price change as the summation of the increments of two orthogonal
martingales: a purely continuous martingale and a purely discontinuous (jump)
martingale. This decomposition is generic for any continuous-time martingales
(Jacod and Shiryaev, 1987). To avoid notational
 complexity, we assume that
the jump process exhibits finite variation: R0 (|x| ∧ 1)νt (x)d x < ∞.
The time subscripts on σt− and νt (x) indicate that both are stochastic and
predictable with respect to the filtration Ft . We further restrict σt− and νt (x) so
that the futures price Ft is always positive. Finally, we assume deterministic
interest rates so that the futures price and the forward price are identical. So
long as futures contracts trade, we need no assumptions on dividends.
Under the specification in Equation (3), the quadratic variation on the futures
return from time t to T is
 T  T 
Vt,T = σs2− ds + x 2 µ(d x, ds). (4)
t t R0

The annualized return variance is RVt,T = T 1−t Vt,T . We show that this return
quadratic variation can be replicated up to a higher-order error term by a static
position in a portfolio of options of the same horizon T and a dynamic position
in futures. As the risk-neutral expected value of futures trading is zero, the
risk-neutral expected value of the quadratic variation can be approximated by
the value of the options in the static portfolio.

Proposition 1. Under no arbitrage, the time-t risk-neutral expected value

of the return quadratic variation of an asset over horizon [t, T ] defined in
Equation (4) can be approximated by the continuum of European out-of-
the-money option prices across all strikes K > 0 and at the same maturity

1315
The Review of Financial Studies / v 22 n 3 2009

date T
 ∞

 2 t (K , T )
EQ
t RVt,T = d K + ε, (5)
T −t 0 Bt (T )K 2
where Bt (T ) denotes the time-t price of a bond paying one dollar at T , t (K , T )
denotes the time-t value of an out-of-the-money option with strike price K > 0
and maturity T ≥ t (a call option when K > Ft and a put option when K ≤ Ft ),
and ε denotes the approximation error, which is zero when the futures price
process is purely continuous. When the futures price can jump, the approxima-
tion error ε is of order O(( dFt−
Ft 3
) ) and is determined by the compensator of the
discontinuous component
   
−2 Q T x2
ε= Et ex − 1 − x − νs (x)d xds. (6)
T −t t R0 2

Refer to the Appendix for the derivation. Equation (5) serves our theoretical
basis for inferring variance swap rates from vanilla options.
1.2 Quantifying variance risk premiums using variance swap rates and
realized variance
Using P to denote the statistical probability measure, we link the variance swap
rate to the realized variance through the following valuation equation:


EPt Mt,T RVt,T

SWt,T = P

 = EPt m t,T RVt,T , (7)
Et Mt,T

where Mt,T denotes a pricing kernel and m t,T = Mt,T /EPt [Mt,T ]. No arbitrage
guarantees the existence of at least one such pricing kernel that prices all traded
assets (Duffie, 1992).
Equation (7) can be decomposed into the following two terms:



SWt,T = EPt m t,T RVt,T = EPt RVt,T + CovPt (m t,T , RVt,T ). (8)

The first term EPt [RVt,T ] represents the time-series conditional mean of the
realized variance. The second term captures the conditional covariance between
the normalized pricing kernel and the realized variance. The negative of this
covariance defines the return variance risk premium. Thus, a direct estimate
of the average variance risk premium is the sample average of the difference
between the variance swap rate and the realized variance, R Pt,T ≡ RVt,T −
SWt,T . This difference also measures the terminal profit and loss from a long
variance swap contract and holding it to maturity.
Dividing both sides of Equation (8) by SWt,T , we can represent the decom-
position in excess returns,
   
RVt,T RVt,T RVt,T
1 = EPt m t,T = EPt + CovPt m t,T , . (9)
SWt,T SWt,T SWt,T

1316
 Variance Risk Premiums

If we regard SWt,T as the forward cost of a variance swap investment,

(RVt,T /SWt,T − 1) captures the excess return from the investment. The sample
average of the excess return represents an estimate of the negative of the covari-
ance term in Equation (9), hence the risk premium. To make the distribution
closer to normality, we represent the excess return in continuously compounded
form and label it as the log variance risk premium, L R Pt ≡ ln(RVt,T /SWt,T ).

2. Data and Methodologies

The options data are from OptionMetrics, which provides historical prices of
options based on closing quotes at the Chicago Board of Options Exchange.
Our data sample starts in January 1996 and ends in February 2003. From
the data set, we filter out market prices of options on five stock indexes and
35 individual stocks. The list of securities is selected mainly based on quote
availability. Specifically, we compute the number of valid option quotes on each
security in the data sample, and select the securities with the highest number of
valid option quotes. In computing the number of valid quotes, we only retain
options that have time-to-maturities within one year, and have strictly positive
bid quotes and strictly positive bid-ask spreads. Options on some securities
are very actively quoted, but only during a short period of our data sample.
In selecting our samples, we further require that the number of active days be
greater than 900 for stock indexes and 600 for individual stocks. We apply
the following criterion to determine the number of active days: (i) the nearest
available maturity must be within 90 days; (ii) the actual stock-price level must
be greater than one dollar; and (iii) the number of strikes is at least three at each
of the two nearest maturities. We compute the synthetic variance swap rates
only on the active days defined above.
Table 1 lists the five stock indexes and 35 individual stocks in our sample. For
each security, the table lists the company name, the starting and ending dates,
the number of active days (N ), and the average number of strikes (N K ) at the
chosen maturities. The index options on the S&P 500 index, the Dow Jones
Industrial Index, and the NASDAQ-100 Index are European options on the
spot indexes. Options on the S&P 100 index and the other 35 individual stocks
and the QQQ (the NASDAQ-100 tracking stock) are all American options
on the underlying spot. The data set includes closing bid and ask quotes for
each option contract and the Black-Scholes implied volatilities based on the
mid quote. For the European options, implied volatilities are directly inferred
from the Black-Scholes option pricing formula. For the American options,
OptionMetrics employs a binomial tree approach that takes account of the
early exercise premium. The data set also includes the interest rate curve
and the projected dividend yield. Our analysis directly employs the implied
volatilities provided by OptionMetrics.
We choose a 30-day horizon for the synthetic variance swap rates. At each
date for each stock, we choose the two nearest maturities, except when the

1317
The Review of Financial Studies / v 22 n 3 2009

Table 1
List of stocks and stock indexes in the sample
No. Ticker Starting date Ending date N NK Name

1 SPX 04-Jan-1996 28-Feb-2003 1779 26 S&P 500 index

2 OEX 04-Jan-1996 28-Feb-2003 1780 27 S&P 100 index
3 DJX 06-Oct-1997 28-Feb-2003 1333 12 Dow Jones Industrial Average
4 NDX 04-Jan-1996 28-Feb-2003 1722 19 NASDAQ 100 Stock Index
5 QQQ 10-Mar-1999 28-Feb-2003 978 22 NASDAQ-100 Index Tracking Stock
6 MSFT 04-Jan-1996 28-Feb-2003 1766 9 Microsoft Corp
7 INTC 04-Jan-1996 28-Feb-2003 1653 8 Intel Corp
8 IBM 04-Jan-1996 28-Feb-2003 1768 9 International Business Machines Corp
9 AMER 04-Jan-1996 28-Feb-2003 1648 9 Nanobac Pharmaceuticals Inc
10 DELL 04-Jan-1996 28-Feb-2003 1650 7 Dell Inc
11 CSCO 04-Jan-1996 28-Feb-2003 1554 7 Cisco Systems Inc
12 GE 04-Jan-1996 28-Feb-2003 1458 6 General Electric Co
13 CPQ 04-Jan-1996 03-May-2002 1272 6 Compaq Computer Corp
14 YHOO 09-Sep-1997 28-Feb-2003 1176 14 Yahoo! Inc
15 SUNW 04-Jan-1996 28-Feb-2003 1395 8 Sun Microsystems Inc
16 MU 04-Jan-1996 28-Feb-2003 1720 8 Micron Technology Inc
17 MO 04-Jan-1996 28-Feb-2003 1474 5 Altria Group Inc
18 AMZN 19-Nov-1997 28-Feb-2003 1078 12 [Link] Inc
19 ORCL 04-Jan-1996 28-Feb-2003 1104 6 Oracle Corp
20 LU 19-Apr-1996 28-Feb-2003 981 7 Lucent Technologies Inc
21 TRV 04-Jan-1996 28-Feb-2003 1279 5 Thousand Trails Inc
22 WCOM 04-Jan-1996 21-Jun-2002 1104 6 WorldCom Inc
23 TYC 05-Jan-1996 28-Feb-2003 979 6 Tyco International Ltd
24 AMAT 04-Jan-1996 28-Feb-2003 1671 8 Applied Materials Inc
25 QCOM 04-Jan-1996 28-Feb-2003 1613 8 Qualcomm Inc
26 TXN 04-Jan-1996 28-Feb-2003 1610 7 Texas Instruments Inc
27 PFE 04-Jan-1996 28-Feb-2003 1420 6 Pfizer Inc
28 MOT 04-Jan-1996 28-Feb-2003 1223 6 Motorola Inc
29 EMC 04-Jan-1996 28-Feb-2003 1188 7 EMC Corp
30 HWP 04-Jan-1996 28-Feb-2003 1395 6 Hewlett-Packward Co
31 AMGN 04-Jan-1996 28-Feb-2003 1478 6 Amgen Inc
32 BRCM 28-Oct-1998 28-Feb-2003 1003 12 Broadcom Corp
33 MER 04-Jan-1996 28-Feb-2003 1542 6 Merill Lynch & Co Inc
34 NOK 04-Jan-1996 28-Feb-2003 1176 6 Nokia OYJ
35 CHL 04-Jan-1996 28-Feb-2003 1422 5 China Mobile Hong Kong Ltd
36 UNPH 16-Sep-1996 28-Feb-2003 745 12 JDS Uniphase Corp
37 EBAY 01-Feb-1999 28-Feb-2003 1000 12 eBay Inc
38 JNPR 07-Oct-1999 28-Feb-2003 627 15 Juniper Networks Inc
39 CIEN 14-May-1997 28-Feb-2003 998 9 Ciena Corp
40 BRCD 30-Nov-1999 28-Feb-2003 693 10 Brocade Communications Systems Inc

Entries list the ticker, the starting date, the ending date, the sample length (N ), the average number of available
strikes per maturity (N K ), and the full name for each of the five stock indexes and 35 individual stocks used in
the study.

shortest maturity is within eight days. Then we switch to the next two maturities
to avoid potential microstructure effects of very short-dated options.
At each maturity, we first linearly interpolate implied volatilities at different
moneyness levels, defined as k ≡ ln(K /F), to obtain a fine grid of implied
volatilities. For moneyness below the lowest available moneyness level in the
market, we use the implied volatility at the lowest strike price. For k above the
highest available moneyness, we use the implied volatility at the highest strike.
Using this interpolation and extrapolation procedure, we generate a fine grid
of 2,000 implied volatility points with a strike range of ±8 standard deviations
from the current spot price. The standard deviation is approximated by the

1318
 Variance Risk Premiums

average implied volatility. Given the fine grid of implied volatilities, we com-
pute the out-of-the-money option prices using the Black-Scholes formula and
replicate the variance swap rate according to a discretization of Equation (5).
At each date t, we interpolate the synthetic variance swap rates at the two
maturities to obtain the variance swap rate at a fixed 30-day horizon. The
interpolation is linear in total variance
 
1 SWt,T1 (T1 − t)(T2 − T ) + SWt,T2 (T2 − t)(T − T1 )
SWt,T = , (10)
T −t T2 − T1
where T1 and T2 denote the two maturity dates, and T denotes the interpolated
maturity date such that T − t is 30 days. We have experimented with different
interpolation schemes, but found that our main conclusions are not materially
affected by the particular choice of the interpolation method.
Corresponding to each 30-day variance swap rate, we also compute the
annualized 30-day realized variance,
30 2
365 Ft+i,t+30 − Ft+i−1,t+30
RVt,t+30 = , (11)
30 i=1
Ft+i−1,t+30

where Ft,T denotes the time-t forward price with expiry date at time T (in days).
A small difference exists between the return variance defined in Equation (11)
and the quadratic variation in Equation (4) due to the difference between
daily monitoring and continuous monitoring. Since the stock prices in the
OptionMetrics data set are not adjusted for stock splits, we manually adjust the
stock splits for each stock in calculating the realized variance. We have also
downloaded stock prices from Bloomberg to check for robustness. Term sheets
of variance swaps vary on how the realized variance is calculated: whether to
use log returns or percentage returns, whether to demean the returns or not, and
whether to use an actual/365 or a business days/252 day-count convention. For
robustness, we have experimented with alternative realized variance measures.
These variations do not alter our conclusions.
Table 2 reports the summary statistics of the annualized realized variance
(RV) and the synthetic variance swap rate (SW). The sample averages of the
variance swap rates are higher than the average realized variance for all the five
stock indexes and most of the individual stocks. The realized variance series
are persistent given the overlapping nature of the estimates. The variance swap
rates are also highly persistent, reflecting the persistence of the return variance
process. Both variance swap rates and the realized variance show positive
skewness and positive excess kurtosis for most stocks and indexes.

3. Historical Behavior of Variance Risk Premiums

To analyze the historical behavior of variance risk premiums, we first establish
the existence, sign, and average magnitude of the variance risk premiums.

1319
The Review of Financial Studies / v 22 n 3 2009

Table 2
Summary statistics for the realized variance and the synthetic variance swap rate
Ticker Panel A: Realized variance, RV×100 Panel B: Variance swap rate, SW×100
Mean Std. dev. Auto Skew Kurt Mean Std. dev. Auto Skew Kurt

SPX 4.07 3.43 0.98 2.23 8.76 6.81 3.87 0.88 2.45 13.95
OEX 4.53 3.82 0.98 2.13 8.26 6.90 3.65 0.96 1.77 6.94
DJX 4.39 3.66 0.98 2.15 7.55 6.98 3.60 0.93 2.18 9.59
NDX 16.69 15.62 0.98 2.31 9.61 19.12 11.86 0.98 0.96 3.32
QQQ 22.61 16.53 0.98 1.60 5.32 26.54 11.29 0.94 0.82 3.58
MSFT 16.59 13.44 0.98 2.33 9.10 19.79 11.80 0.76 4.80 63.40
INTC 27.67 23.25 0.98 2.25 8.49 25.17 14.06 0.94 1.80 7.97
IBM 15.15 11.39 0.97 1.76 6.52 16.83 8.56 0.89 2.19 11.64
AMER 41.13 25.49 0.97 1.12 4.07 44.64 21.04 0.90 1.25 5.37
DELL 32.90 21.95 0.97 1.60 5.48 37.33 17.29 0.93 2.24 12.97
CSCO 31.10 28.53 0.98 2.09 7.83 33.40 20.96 0.94 2.01 7.92
GE 11.91 8.83 0.98 1.84 6.54 14.15 7.93 0.90 1.38 5.53
CPQ 31.01 21.46 0.96 1.66 5.94 32.14 16.48 0.81 2.06 10.81
YHOO 72.25 43.52 0.97 0.91 3.29 73.92 43.08 0.78 2.78 18.63
SUNW 36.71 27.14 0.98 1.83 6.61 37.63 23.13 0.84 3.22 25.06
MU 56.68 31.16 0.97 1.43 5.37 59.43 25.53 0.87 1.75 12.10
MO 13.63 11.04 0.96 1.99 8.24 15.44 9.79 0.84 2.41 16.68
AMZN 88.62 48.69 0.96 0.63 2.99 103.81 55.77 0.95 1.80 7.60
ORCL 43.72 33.83 0.97 1.70 5.84 48.11 43.62 0.94 4.40 29.09
LU 31.43 30.86 0.97 3.22 17.31 31.25 29.02 0.68 6.22 79.86
TRV 19.36 17.40 0.95 2.46 9.24 19.02 11.44 0.93 2.83 15.27
WCOM 26.84 24.12 0.97 2.76 13.82 27.81 20.81 0.92 2.22 8.94
TYC 32.61 38.22 0.98 2.32 8.71 40.74 53.24 0.90 4.46 30.54
AMAT 43.89 26.91 0.97 1.89 7.37 45.78 20.18 0.91 1.30 4.98
QCOM 46.98 32.35 0.98 1.31 4.26 48.70 23.65 0.93 1.41 5.45
TXN 37.24 25.65 0.98 1.90 7.24 35.40 18.21 0.94 1.27 4.65
PFE 12.65 7.96 0.97 1.69 7.06 14.19 5.88 0.88 0.97 4.69
MOT 29.28 25.44 0.96 2.04 7.23 27.09 18.70 0.84 1.97 9.56
EMC 41.93 37.67 0.98 2.62 10.39 38.05 22.29 0.91 1.71 6.10
HWP 25.19 17.24 0.96 1.34 4.34 24.94 13.52 0.91 1.58 6.18
AMGN 23.78 17.93 0.98 1.82 6.75 25.66 15.64 0.95 1.42 4.71
BRCM 91.22 57.64 0.98 1.68 6.06 90.56 45.18 0.95 1.68 6.56
MER 23.26 15.12 0.97 1.77 7.12 24.05 11.51 0.93 1.61 8.81
NOK 33.67 20.99 0.96 0.97 3.32 32.15 15.99 0.83 1.34 7.90
CHL 19.23 17.65 0.98 3.12 16.78 20.10 14.30 0.96 2.62 13.02
UNPH 83.74 59.74 0.97 1.47 5.40 77.16 42.70 0.95 1.02 3.44
EBAY 69.16 59.28 0.98 1.29 4.18 73.10 42.97 0.96 1.00 3.81
JNPR 104.72 56.79 0.97 1.02 3.30 114.50 49.41 0.93 0.99 3.54
CIEN 96.26 65.37 0.97 1.26 4.39 91.04 47.70 0.92 1.05 5.51
BRCD 110.11 65.55 0.97 1.01 3.45 100.75 40.78 0.92 0.38 2.51

Entries report summary statistics for the annualized realized variance RV and the synthetic variance swap rate
SW. Columns under Mean, Std. dev., Auto, Skew, Kurt report the sample average, standard deviation, daily
autocorrelation, skewness, and excess kurtosis, respectively.

Then, we investigate whether the risk premiums can be explained by classic

risk factors. Finally, we analyze the dynamic properties of the premiums using
expectation hypothesis regressions.

3.1 Do investors price variance risk?

If investors price variance risk, the sample averages of the realized variance
will differ from the average variance swap rates. Table 3 reports the summary
statistics of the difference between the realized variance and the variance swap
rate, R P = (RVt,T − SWt,T ) × 100, in Panel A and the log difference L R P =

1320
Variance Risk Premiums

Table 3
Summary statistics of variance risk premiums
Ticker Panel A: (RV − SW ) × 100 Panel B: ln (RV /SW ) IR
Mean Std. dev. Auto Skew Kurt t Mean Std. dev. Auto Skew Kurt t

SPX −2.74 3.63 −0.04 −1.44 17.86 −8.39 −0.66 0.57 0.05 0.18 3.23 −11.83 0.98
OEX −2.36 3.57 −0.07 0.21 6.69 −7.02 −0.58 0.56 0.06 0.36 2.90 −10.34 0.85
DJX −2.58 3.86 −0.05 −0.15 8.28 −6.37 −0.61 0.58 0.07 0.63 3.31 −9.07 0.87
NDX −2.43 10.24 0.05 1.49 9.42 −2.54 −0.28 0.47 0.11 0.40 3.41 −6.49 0.55
QQQ −3.93 12.55 0.12 0.83 4.68 −2.62 −0.29 0.48 0.21 0.16 2.88 −4.91 0.55
MSFT −3.20 12.31 −0.10 −1.91 57.68 −3.32 −0.30 0.52 −0.08 0.08 3.48 −6.62 0.55
INTC 2.49 19.07 −0.13 1.91 8.96 1.34 −0.02 0.51 −0.19 0.42 3.31 −0.44 0.04
IBM −1.68 10.24 −0.04 0.67 9.38 −1.80 −0.24 0.60 0.02 0.01 2.94 −4.35 0.36
AMER −3.51 23.76 −0.08 0.63 4.62 −2.05 −0.17 0.57 −0.06 0.03 3.06 −3.79 0.33
DELL −4.43 21.35 0.14 0.49 7.05 −2.15 −0.23 0.55 0.18 0.22 3.07 −4.17 0.36
CSCO −2.30 20.31 −0.11 1.51 10.80 −1.42 −0.27 0.83 0.05 −6.14 70.41 −4.06 0.36
GE −2.24 7.63 −0.14 0.59 7.68 −3.52 −0.25 0.49 −0.04 0.28 3.32 −5.60 0.51
CPQ −1.14 20.66 −0.02 0.17 7.66 −0.62 −0.13 0.59 −0.07 0.14 3.27 −2.54 0.25
YHOO −1.67 43.58 −0.21 −1.17 16.82 −0.42 −0.09 0.56 −0.16 0.07 2.97 −1.63 0.17
SUNW −0.92 20.32 0.03 −1.97 40.32 −0.53 −0.11 0.48 −0.01 −0.08 3.32 −2.54 0.24
MU −2.75 29.38 −0.12 0.68 6.71 −1.12 −0.10 0.47 −0.12 0.22 3.25 −2.70 0.23
MO −1.81 11.76 0.10 0.52 9.99 −1.66 −0.24 0.69 0.05 0.21 3.29 −3.79 0.34
AMZN −15.19 59.66 −0.06 −0.30 5.14 −2.14 −0.22 0.59 −0.14 0.10 2.75 −3.33 0.35
ORCL −4.39 46.26 0.04 −4.55 33.85 −0.82 −0.14 0.66 0.04 −1.59 8.93 −1.92 0.20
LU 0.18 29.35 0.06 −3.43 79.81 0.08 −0.08 0.54 0.18 0.05 3.53 −1.55 0.17
TRV 0.35 16.04 0.20 2.38 10.88 0.22 −0.12 0.64 0.14 0.93 5.03 −1.85 0.18
WCOM −0.97 21.22 −0.11 1.39 11.29 −0.38 −0.13 0.63 −0.07 −0.10 2.84 −1.78 0.19
TYC −8.13 48.26 0.06 −2.16 26.08 −1.55 −0.34 0.74 −0.13 0.90 4.13 −4.00 0.45
AMAT −1.89 24.20 −0.08 1.08 7.08 −0.93 −0.11 0.48 −0.14 0.20 3.21 −2.72 0.23
QCOM −1.73 28.20 −0.11 0.82 5.18 −0.69 −0.15 0.59 −0.06 −0.14 3.77 −2.77 0.24
TXN 1.84 19.93 −0.07 1.22 7.33 1.00 −0.02 0.47 −0.11 0.16 2.97 −0.59 0.05
PFE −1.54 8.08 −0.06 1.44 7.92 −1.92 −0.21 0.60 0.03 −0.14 4.17 −3.35 0.31
MOT 2.19 20.13 −0.12 0.09 10.91 1.23 −0.02 0.57 −0.08 −0.29 3.43 −0.31 0.03
EMC 3.87 28.91 0.35 2.25 11.62 1.21 −0.02 0.53 0.05 0.55 3.64 −0.47 0.05
HWP 0.25 14.56 −0.01 0.71 7.28 0.19 −0.08 0.54 −0.01 0.27 3.72 −1.67 0.16
AMGN −1.88 15.11 −0.07 0.71 5.94 −1.26 −0.16 0.55 −0.02 0.10 2.76 −3.01 0.27
BRCM 0.66 48.32 0.07 0.64 4.89 0.12 −0.05 0.48 0.17 0.20 2.41 −0.99 0.11
MER −0.79 13.11 0.05 1.04 5.90 −0.68 −0.11 0.50 −0.13 0.26 2.99 −2.50 0.22
NOK 1.51 18.73 0.07 −0.11 7.86 0.79 −0.03 0.55 0.08 0.07 3.31 −0.54 0.05
CHL −0.87 15.10 −0.11 1.91 14.02 −0.60 −0.15 0.54 −0.14 0.36 3.61 −2.84 0.26
UNPH 6.59 49.42 −0.16 0.79 4.88 1.12 −0.01 0.58 −0.04 −0.32 3.00 −0.25 0.03
EBAY −3.95 45.88 −0.05 1.44 6.54 −0.69 −0.27 0.58 0.17 0.35 2.97 −3.48 0.38
JNPR −9.79 54.78 −0.15 −0.21 3.74 −1.27 −0.14 0.52 −0.12 −0.54 3.20 −2.07 0.29
CIEN 5.22 62.53 −0.07 0.80 6.64 0.68 −0.03 0.61 −0.07 0.47 4.18 −0.33 0.04
BRCD 9.36 59.99 0.10 0.59 3.59 1.08 0.00 0.56 0.08 −0.25 2.63 0.02 −0.00

Entries report summary statistics of variance risk premiums, defined as the difference between the realized
variance and the variance swap rate in Panel A and as the log difference in Panel B. Columns under Mean,
Std. dev., Auto, Skew, Kurt report the sample average, standard deviation, average nonoverlapping 30-day
autocorrelation, skewness, and excess kurtosis, respectively. Columns under t report the t-statistics of the mean
risk premiums, which are adjusted for serial dependence according to the Newey-West method with a lag of
30 days. The last column of the table under “IR” reports the annualized Sharpe ratio of shorting the 30-day
variance swap contracts, computed as the mean of − ln(RV /SW ) divided by its Newey-West standard deviation
(with 30 lags), and then annualized by 365/30.

ln(RVt,T /SWt,T ) in Panel B. The variance risk premiums R P show large

kurtosis and sometimes large skewness. The skewness and kurtosis are much
smaller for the log variance risk premiums LRP.
The sample averages of the variance risk premiums and log variance risk
premiums are negative for all the five stock indexes and most of the individual
stocks. Table 3 also reports the t-statistics on the significance of the mean

1321
 The Review of Financial Studies / v 22 n 3 2009

risk premiums, adjusted for serial dependence according to Newey and West
(1987) with 30 lags. The largest t-statistics come from the S&P 500 and
S&P 100 indexes and the Dow Jones Industrial Average, which are strongly
significant for both variance risk premiums and log variance risk premiums.
The t-statistics for the NASDAQ-100 Index and its tracking stock are lower,
but remain strongly significant.1
The two definitions of variance risk premiums in Table 3 represent two ways
of computing returns for variance swap investments. The mean estimates in
Panel A, (RV − SW ) × 100, represent the average dollar profit and loss for
each $100 notational investment in the variance swap contract. Thus, if we long
a 30-day variance swap contract with a notional of $100 on the S&P 500 index
and hold the contract to maturity, during our sample period the average return
per $100 notional investment is −$2.74.
Alternatively, if we regard the variance swap rate as the forward cost of the
variance swap contract, the log variance risk premium ln(RV /SW ) in Panel B
of Table 3 can be thought of as the continuously compounded excess return to
going long the variance swap contract and holding it to maturity. Based on this
calculation, the average excess return is −66% for long 30-day variance swap
contracts on the S&P 500 index. The different magnitudes in the two panels
mainly come from different scaling. Panel A regards the $100 notional as the
initial investment, whereas Panel B uses the forward cost (i.e., the variance
swap rate) as the initial investment. For the S&P 500 index, a $100 notional
corresponds to an average forward cost of $6.81 (Table 2). For the same dollar
profit and loss, the smaller base number generates larger return estimates in
Panel B.
Despite the different representations, it is clear that investors are willing
to accept a significantly negative average return to long variance swaps on
the S&P and Dow indexes. Accordingly, shorting variance swap contracts on
the indexes generates significantly positive average excess returns during the
sample period. To gauge the profitability of such a trading strategy, we estimate
the annualized Sharpe ratio on shorting the 30-day variance swap contracts,
and report them in the last column of Table 3 under “IR.” The Sharpe ratio is
computed as the sample mean of the log excess return − ln(RV /SW ) divided
by its standard deviation and multiplied by 365/30 for annualization. The
standard deviation is adjusted for serial dependence according to Newey and
West (1987) with 30 lags. The Sharpe ratio estimates are 0.98, 0.85, and 0.87 for
shorting variance swaps on the S&P 500, the S&P 100, and the Dow indexes,
which are significantly higher than an average stock portfolio investment.
Nevertheless, it is important to point out that the Sharpe ratios are computed
using synthetic variance swap rates. The actual profitability depends on several
practical factors, such as the actual availability of variance swap quotes, their

1
The variance risk premiums on the NASDAQ-100 Index and its tracking stock QQQ also show some differences
due to, among other things, their different sample periods

1322
Variance Risk Premiums

bid-ask spreads, and their similarity to our synthetic values. Furthermore, given
the nonlinear payoff structure, caution should be applied when interpreting
Sharpe ratios on derivative trading strategies (e.g., Goetzmann et al., 2002).
The average variance risk premiums and log variance risk premiums on in-
dividual stocks show large cross-sectional variation. The standard deviation
estimates on the variance risk premiums (RP) of the individual stocks are all
larger than those on the S&P and Dow indexes. As a result, out of the 35
individual stocks, only seven generate variance risk premiums that are signifi-
cantly negative at the 95% confidence level. By contrast, the standard deviation
estimates on the log variance risk premiums (LRP) are much more uniform
across all stock indexes and individual stocks. For 23 out of 35 individual
stocks, the mean log variance risk premiums are significantly negative at the
95% confidence level.
The cross-sectional variation of the variance risk premiums possibly suggests
that the market does not price all return variance risk in each single stock, but
only prices a systematic variance risk component in the stock market portfolio.
Based on this hypothesis, the average variance risk premium on each stock is not
proportional to the total variation of the return variance, but to the covariation
of the return variance with the market portfolio return variance. To test this
hypothesis, we use the realized variance on the S&P 500 index return as the
market portfolio variance, and estimate the “variance beta” for each stock as

βVj = Cov(ln RV j , ln RVSPX )/Var(ln RVSPX ), (12)

where the variance and covariance are measured using the common sample
of the two realized variance series. We estimate the variance beta using log
variance for better distributional behaviors.
Given the variance beta estimates, our hypothesis suggests that the average
variance risk premiums are more negative on stocks with higher variance beta.
Regressing the average log variance risk premiums on the variance beta across
the 40 stocks and stock indexes generates the following estimates:

LRP j = 0.0061 − 0.3283 βVj +e, R 2 = 18.4%,

(13)
(0.09) (−2.96)

with t-statistics reported in the parentheses below the estimates. Consistent

with our hypothesis, the slope estimate is negative and statistically significant.
Therefore, we identify a systematic variance risk factor that the market prices
heavily. The negative sign on the market variance risk premiums suggests that
investors are willing to pay a premium to hedge away upward movements in the
return variance of the stock market. In other words, investors regard increases
in market volatility as unfavorable shocks to the investment opportunity and
demand a high premium for bearing such shocks.
Table 3 also reports the average nonoverlapping 30-day autocorrelation
for the variance risk premiums. The autocorrelation estimates are low,

1323
 The Review of Financial Studies / v 22 n 3 2009

averaging −0.023 for RP and −0.006 for LRP. Therefore, although return
variance is strongly predictable, investors have priced this predictability into
options, so that excess returns on synthetic variance swap investments are no
longer strongly predictable.

3.2 Can classic risk factors explain the variance risk premiums?
Return variance can vary either by itself as in the stochastic volatility models
of Heston (1993); and Hull and White (1987), or it can vary as a function of
the stock price as in the constant elasticity of variance model of Cox (1996)
and the local volatility model of Dupire (1994); and Derman and Kani (1994).
In the first case, the independent variance variation represents an additional
source of risk (in addition to the return risk), which can ask for a risk premium
in addition to the premium on the return risk. In the latter case, the variance
risk premium is induced purely by the underlying return risk and return risk
premium.
The classic capital asset pricing model (CAPM) argues that the expected
excess return on an asset is only proportional to the beta of the asset on the
market portfolio. Under this model, variance risk premium cannot come from
an independent source of risk, but can only come from the variance swap’s
correlation with the market portfolio. Qualitatively, the negative excess return
on the variance swap contract on the stock indexes is consistent with the
well-documented negative correlation between index returns and index return
variance. The question is whether this negative correlation can fully account
for the negative variance risk premiums.
To answer this question, we estimate the following CAPM regressions:

ln RVt,T /SWt,T = α + β j E Rt,T

m
+e (14)

for the five stock indexes and 35 individual stocks, where ERm denotes the
excess return on the market portfolio, for which we consider two proxies. First,
we use the S&P 500 index to proxy for the market portfolio and compute the
m
excess return as E Rt,T = ln STm /Ft,T
m
. Our second proxy is the value-weighted
return on all NYSE, AMEX, and NASDAQ stocks (from CRSP) minus the
one-month Treasury bill rate (from Ibbotson Associates). This excess return
is publicly available at Kenneth French’s online data library.2 The data are
monthly. The sample period that matches our options data is from January
1996 to December 2002. We estimate the regressions using the generalized
methods of moments (GMM), with the weighting matrix computed according
to Newey and West (1987) with 30 lags for the overlapping daily series and six
lags for the nonoverlapping monthly series.
Table 4 reports the estimates (and t-statistics in parentheses). The results from
the two market portfolio proxies are similar. The beta estimates are strongly

2
The address is: [Link] [Link].

1324
Variance Risk Premiums

negative for all the stock indexes and most of the individual stocks. The es-
timates are the most negative for the S&P and Dow indexes. Nevertheless,
the intercept α estimates remain strongly negative, especially for the S&P and
Dow indexes, indicating that the negative beta cannot fully account for the neg-
ative variance risk premiums. Indeed, the estimates for α are not much smaller
than the mean variance risk premiums reported in Table 3. The results call for
additional risk factors.
Kraus and Litzenberger (1976) propose a three-moment CAPM, in which
the excess return on a security is proportional not only to the excess return on
the market portfolio, but also to the squared deviation of the market portfolio
return from its expected value
 2
ln RVt,T /SWt,T = α + β E Rt,T
m
+ γ Rm − R̄ m + e, (15)
m
where R m denotes the market portfolio return and R denotes its expected
value. We use E R m to proxy R m in constructing the squared deviation factor.
The loading coefficient estimates for γ are mostly insignificant and the α esti-
mates are close to what we have obtained from the regression in Equation (14).
Hence, (R m − R̄ m )2 is not the factor that we are looking for in explaining the
variance risk premiums. To save space, we do not report the estimation results
but they are available upon request.
Fama and French (1993) identify two additional risk factors in the stock mar-
ket that are related to the firm size (SMB) and book-to-market value (HML),
respectively. We investigate whether these additional common risk factors ex-
plain the variance risk premiums. We estimate the following relations:

ln RVt,T /SWt,T = α + β E Rt,T

m
+ s S M Bt,T + h H M L t,T + e. (16)

Data on all three risk factors are available on Kenneth French’s online data
library. We refer the interested readers to Fama and French (1993) for details
on the definition and construction of these common risk factors. The sample
period that overlaps with our options data is monthly from January 1996 to
December 2002. Again, ERm denotes the excess return to the market portfolio.
Furthermore, both SMB and HML are in terms of excess returns on zero-cost
portfolios. Therefore, the intercept α represents the expected excess return on
an investment that is neutral to all three risk factors. Table 5 reports the GMM
parameter estimates and t-statistics. The intercept estimates for the indexes
remain strongly negative, the magnitudes only slightly smaller than the average
variance risk premiums in Table 3. Thus, the Fama-French risk factors can only
explain a small portion of the variance risk premiums.
In the regression, both the market ERm and the size SMB factors generate
significantly negative loadings, indicating that the return variance is not only
negatively correlated with the market portfolio return, but also with the SMB
factor. Hence, going long the variance swap contract also serves as an insurance

1325
The Review of Financial Studies / v 22 n 3 2009

Table 4
Explaining variance risk premiums with CAPM beta
Proxy Panel A: S&P 500 index Panel B: Valued-weighted market portfolio
α β R2 α β R2

SPX −0.646 (−13.554) −4.510 (−5.644) 0.173 −0.641 (−10.236) −5.508 (−4.751) 0.245
OEX −0.562 (−11.630) −4.473 (−5.653) 0.175 −0.571 (−9.112) −5.536 (−4.936) 0.247
DJX −0.613 (−10.820) −4.681 (−5.165) 0.206 −0.617 (−9.362) −4.668 (−3.828) 0.205
NDX −0.273 (−6.689) −2.450 (−3.742) 0.078 −0.237 (−5.998) −3.617 (−3.225) 0.182
QQQ −0.301 (−5.257) −1.157 (−1.707) 0.018 −0.320 (−5.330) −2.964 (−1.702) 0.117
MSFT −0.287 (−6.722) −2.168 (−4.134) 0.048 −0.339 (−5.981) −2.495 (−2.870) 0.068
INTC −0.011 (−0.252) −2.211 (−2.706) 0.051 −0.045 (−0.912) −3.793 (−3.021) 0.150
IBM −0.235 (−4.302) −2.181 (−2.737) 0.037 −0.269 (−4.448) −2.185 (−1.759) 0.041
AMER −0.164 (−3.595) −2.133 (−3.161) 0.038 −0.242 (−5.077) −1.695 (−1.604) 0.026
DELL −0.214 (−4.145) −2.723 (−3.582) 0.068 −0.271 (−3.645) −3.527 (−3.673) 0.130
CSCO −0.269 (−3.583) −0.966 (−0.603) 0.004 −0.272 (−3.948) −1.998 (−0.888) 0.043
GE −0.240 (−5.855) −2.648 (−3.868) 0.087 −0.287 (−5.461) −1.648 (−1.488) 0.046
CPQ −0.107 (−1.959) −2.426 (−2.336) 0.039 −0.028 (−0.558) −3.292 (−2.646) 0.099
YHOO −0.089 (−1.630) −0.582 (−0.800) 0.003 −0.171 (−2.540) 0.752 ( 0.791) 0.006
SUNW −0.095 (−2.250) −2.295 (−3.378) 0.054 −0.111 (−1.977) −3.978 (−3.162) 0.178
MU −0.096 (−2.576) −1.215 (−2.055) 0.018 −0.105 (−2.364) −2.452 (−3.664) 0.093
MO −0.246 (−3.818) 0.375 (0.400) 0.001 −0.253 (−3.980) 0.539 (0.473) 0.002
AMZN −0.217 (−3.323) 0.075 (0.072) 0.000 −0.164 (−1.722) 0.486 (0.464) 0.002
ORCL −0.125 (−1.805) −2.311 (−2.706) 0.032 −0.179 (−2.488) −3.606 (−2.408) 0.099
LU −0.065 (−1.269) −1.463 (−1.740) 0.018 −0.047 (−0.877) −2.713 (−2.757) 0.099
TRV −0.122 (−1.885) −2.075 (−2.584) 0.034 −0.167 (−1.647) −1.025 (−0.626) 0.010
WCOM −0.099 (−1.432) −3.588 (−3.550) 0.078 −0.066 (−0.715) −4.256 (−2.821) 0.140
TYC −0.349 (−4.025) −1.726 (−1.591) 0.017 −0.405 (−3.885) 0.778 (0.260) 0.002
AMAT −0.104 (−2.641) −0.972 (−1.700) 0.011 −0.119 (−2.482) −2.839 (−3.515) 0.110
QCOM −0.149 (−2.747) −1.062 (−1.358) 0.009 −0.164 (−2.629) −2.573 (−2.129) 0.059
TXN −0.022 (−0.546) −0.603 (−1.150) 0.005 −0.086 (−1.596) −0.783 (−1.046) 0.009
PFE −0.206 (−3.307) −2.016 (−1.862) 0.035 −0.230 (−3.541) −1.907 (−1.213) 0.035
MOT −0.003 (−0.055) −1.955 (−1.956) 0.030 −0.024 (−0.329) −3.479 (−1.926) 0.099
EMC −0.007 (−0.150) −2.865 (−2.967) 0.079 −0.069 (−0.845) −4.090 (−3.950) 0.167
HWP −0.070 (−1.444) −1.669 (−2.037) 0.024 −0.048 (−0.802) −1.996 (−1.650) 0.052
AMGN −0.159 (−2.979) −0.947 (−1.028) 0.009 −0.128 (−1.478) −0.164 (−0.116) 0.000
BRCM −0.046 (−0.867) 0.861 (1.140) 0.010 −0.082 (−0.950) −0.824 (−0.423) 0.006
MER −0.106 (−2.419) −1.184 (−1.601) 0.016 −0.145 (−2.354) −1.294 (−1.147) 0.021
NOK −0.031 (−0.537) −1.708 (−1.959) 0.028 −0.037 (−0.563) −2.132 (−1.765) 0.074
CHL −0.144 (−2.752) −1.662 (−1.840) 0.029 −0.105 (−2.405) −1.968 (−2.173) 0.050
UNPH −0.014 (−0.236) −1.676 (−1.440) 0.023 −0.057 (−1.276) −3.015 (−1.209) 0.074
EBAY −0.266 (−3.401) 0.214 (0.219) 0.000 −0.293 (−2.267) −0.145 (−0.122) 0.000
JNPR −0.147 (−2.151) −0.484 (−0.572) 0.003 −0.184 (−2.494) −2.091 (−1.082) 0.042
CIEN −0.020 (−0.258) −2.320 (−1.738) 0.039 −0.062 (−0.653) −4.433 (−1.929) 0.148
BRCD 0.004 (0.054) 0.182 (0.190) 0.000 0.006 (0.060) −2.790 (−2.301) 0.104

Entries report the GMM estimates (and t-statistics in parentheses) of the following relation,

ln RV t,T /SWt,T = α + β j E Rt,T

m
+ e,

where ERm denotes the excess return on the market portfolio, which is proxyed by the return on the S&P 500
index forward in Panel A and the excess return on the CRSP valued-weighted stock portfolio in Panel B. The
t-statistics are computed according to Newey and West (1987) with 30 lags for the overlapping daily series in
Panel A and six lags for the nonoverlapping monthly series in Panel B. Columns under “R 2 ” report the unadjusted
R-squared.

against the SMB factor going up. The loading estimates on the HML factor are
mostly insignificant.
Fama and French (1993) also consider two bond market factors, related to
the bond maturity (TERM) and default (DEF) risks. Furthermore, Jegadeesh
and Titman (1993) identify a momentum phenomenon, that past winners often

1326
Variance Risk Premiums

Table 5
Explaining variance risk premiums with Fama-French risk factors
Ticker α ERm SMB HML R2

SPX −0.633 (−9.070) −5.205 (−3.853) −2.858 (−2.098) −0.195 (−0.227) 0.287
OEX −0.561 (−8.294) −5.268 (−4.030) −3.292 (−2.472) −0.443 (−0.477) 0.300
DJX −0.604 (−8.251) −4.601 (−3.246) −3.603 (−3.074) −1.346 (−1.743) 0.275
NDX −0.235 (−5.943) −2.851 (−2.614) −1.958 (−2.216) 1.391 (1.780) 0.269
QQQ −0.304 (−5.213) −2.204 (−1.461) −1.694 (−1.548) 1.543 (2.470) 0.229
MSFT −0.321 (−6.617) −2.530 (−2.758) −4.960 (−4.839) −1.855 (−2.164) 0.225
INTC −0.037 (−0.886) −3.875 (−3.119) −2.961 (−3.026) −1.177 (−1.530) 0.205
IBM −0.259 (−4.963) −2.027 (−1.567) −3.063 (−1.944) −0.660 (−0.428) 0.089
AMER −0.229 (−5.607) −1.692 (−1.225) −3.084 (−2.154) −0.964 (−0.930) 0.080
DELL −0.268 (−3.664) −2.987 (−3.459) −2.968 (−2.293) 0.420 (0.344) 0.203
CSCO −0.282 (−4.458) −1.084 (−0.479) 1.461 (1.092) 2.130 (2.473) 0.081
GE −0.267 (−6.495) −1.537 (−1.281) −2.834 (−2.871) −0.818 (−0.973) 0.132
CPQ −0.030 (−0.653) −2.907 (−2.005) 0.790 (0.661) 0.910 (0.906) 0.106
YHOO −0.170 (−2.554) 0.104 (0.087) 0.669 (0.432) −0.876 (−0.783) 0.024
SUNW −0.117 (−2.220) −3.066 (−2.066) −1.507 (−1.466) 1.119 (1.188) 0.227
MU −0.101 (−2.400) −2.650 (−3.792) −0.646 (−0.811) −0.745 (−1.166) 0.101
MO −0.255 (−4.129) 0.709 (0.529) −0.527 (−0.286) 0.556 (0.394) 0.005
AMZN −0.173 (−2.025) −0.110 (−0.069) −1.775 (−0.746) −1.828 (−1.182) 0.035
ORCL −0.174 (−2.355) −3.814 (−3.008) −0.110 (−0.068) −0.334 (−0.334) 0.100
LU −0.046 (−0.859) −3.377 (−2.586) −0.918 (−0.546) −1.333 (−1.287) 0.125
TRV −0.129 (−1.581) −0.496 (−0.283) −5.906 (−4.769) −1.194 (−1.066) 0.209
WCOM −0.073 (−0.835) −4.910 (−2.768) −2.263 (−1.419) −1.756 (−1.660) 0.167
TYC −0.367 (−3.334) 0.253 (0.107) −4.234 (−1.611) −2.668 (−1.875) 0.078
AMAT −0.100 (−2.769) −2.617 (−2.867) −4.062 (−4.045) −1.103 (−1.667) 0.247
QCOM −0.164 (−2.657) −1.657 (−1.330) −3.863 (−3.097) 0.749 (0.716) 0.136
TXN −0.073 (−1.382) −1.007 (−1.600) −4.353 (−4.575) −1.997 (−2.073) 0.163
PFE −0.208 (−3.746) −1.504 (−0.834) −3.795 (−2.047) −1.163 (−0.582) 0.107
MOT −0.029 (−0.354) −2.860 (−1.650) 0.706 (0.452) 1.002 (0.869) 0.109
EMC −0.071 (−0.994) −2.482 (−2.356) −1.771 (−2.035) 2.003 (2.848) 0.270
HWP −0.044 (−0.751) −2.456 (−1.744) −0.918 (−0.816) −1.138 (−1.026) 0.067
AMGN −0.122 (−1.555) −0.421 (−0.276) −1.367 (−1.030) −1.184 (−1.352) 0.019
BRCM −0.049 (−0.633) −0.313 (−0.148) −3.039 (−2.407) −0.432 (−0.389) 0.079
MER −0.138 (−2.573) −0.899 (−0.737) −2.030 (−1.382) 0.313 (0.346) 0.063
NOK −0.023 (−0.339) −2.001 (−1.745) −2.276 (−2.111) −0.626 (−0.800) 0.127
CHL −0.102 (−2.498) −1.830 (−1.927) −2.866 (−2.234) −0.565 (−0.568) 0.112
UNPH −0.044 (−0.604) −1.688 (−0.757) −1.547 (−0.868) 1.382 (0.999) 0.152
EBAY −0.263 (−2.003) 0.446 (0.412) −2.835 (−2.125) 0.336 (0.296) 0.073
JNPR −0.105 (−1.506) −3.192 (−2.421) −3.368 (−2.146) −3.227 (−3.041) 0.158
CIEN −0.040 (−0.452) −5.340 (−2.403) −2.960 (−2.350) −2.574 (−1.565) 0.202
BRCD 0.060 (0.510) −2.025 (−1.699) −3.496 (−2.457) −0.723 (−0.637) 0.223

Entries report the GMM estimates (and t-statistics in parentheses) of the following relation,

ln RV t,T /SWt,T = α + β E Rt,T

m
+ s S M Bt,T + h H M L t,T + e,

where the regressors are the three stock market risk factors defined by Fama and French (1993): the excess return
on the market portfolio (ERm ), the size factor (SMB), and the book-to-market factor (HML). Data are monthly
from January 1996 to December 2002. The t-statistics are computed according to Newey and West (1987) with
six lags. Columns under “R 2 ” report the unadjusted R-squared of the regression.

continue to outperform past losers. We construct the TERM and DEF factors
using Treasury and corporate yield data from the Federal Reserve Statisti-
cal Release. Kenneth French’s data library also provides a momentum factor
(UMD) similar to that from Carhart (1997). However, single-factor regressions
on these three factors show that none of these factors have a significant loading
on the variance risk premiums. Therefore, they cannot explain the variance risk
premiums, either.

1327
The Review of Financial Studies / v 22 n 3 2009

The bottom line here is that classic risk factors cannot fully account for the
negative variance risk premiums on the stock indexes. Either there exists a large
inefficiency in the market for index variance or else the majority of the variance
risk is generated by an independent risk factor that the market prices heavily.
Investors are willing to receive a negative excess return to hedge against market
volatility going up, not only because market volatility movement is negatively
correlated with stock market portfolio return, but also because investors regard
market volatility hikes by themselves as unfavorable shocks and demand high
compensation for bearing such shocks.
There are several potential reasons for the negative variance risk premiums.
Take the market portfolio of stocks as an example, which the market holds
in aggregate. With the same expected return, the increase in return variance
implies a decline in performance in terms of the Sharpe ratio. Hence, one way
to guarantee a minimum performance is to buy options to hedge against return
variance increases. Furthermore, going long the variance swap contract is an
effective strategy to hedge against risks associated with the random arrival of
discontinuous price movements. Finally, considerations on meeting value-at-
risk requirements and preventing shortfalls and draw-downs also make long
variance swap an attractive strategy that investors are willing to take even with
a negative expected excess return.

3.3 Are variance risk premiums constant or time-varying?

To understand the dynamic behavior of variance risk premiums, we run the
expectation hypothesis regression

RVt,T = a + b SWt,T + e. (17)

Under the null hypothesis of zero variance risk premiums in dollar terms:
CovPt (m t,T , RVt,T ) = 0 as defined in Equation (8), we have a = 0 and b = 1.
In particular, the slope estimate deviating from zero would suggest that the
variance risk premiums are time-varying and correlated with the variance swap
rate.
Table 6 reports the GMM estimates of Equation (17) and the t-statistics under
the null hypotheses of a = 0 and b = 1 in Panel A. All the slope estimates are
positive, but many are lower than 1. The t-statistics show that over half of
the stock indexes and individual stocks generate regression slopes that are
significantly lower than the null value of 1.
Since the variance and variance swap rates show positive skewness (Table 2),
we also run the expectation hypothesis regression in log terms and report the
results in Panel B of Table 6,

ln RVt,T = a + b ln SWt,T + e. (18)

Under the null hypothesis of zero variance risk premiums in return terms:
CovPt (m t,T , SW
RVt,T
t,T
) = 0 as defined in Equation (9), the slope estimate b should

1328
Variance Risk Premiums

Table 6
Expectation hypothesis regressions on constant variance risk premiums
Ticker Panel A: RVt,T = a + bSWt,T + e Panel B: ln RVt,T = a + b ln SWt,T + e
a b R2 a b R2

SPX 0.010 (1.416) 0.455 (−4.596) 0.262 −0.891 (−2.593) 0.919 (−0.684) 0.378
OEX 0.006 (0.981) 0.568 (−3.933) 0.294 −0.600 (−1.797) 0.992 (−0.065) 0.408
DJX 0.013 (1.524) 0.443 (−4.046) 0.190 −1.210 (−2.859) 0.781 (−1.467) 0.253
NDX −0.023 (−1.329) 0.995 (−0.042) 0.571 −0.170 (−1.233) 1.060 (0.876) 0.672
QQQ −0.027 (−0.887) 0.953 (−0.326) 0.424 −0.281 (−1.466) 1.007 (0.060) 0.445
MSFT 0.046 (1.804) 0.605 (−2.726) 0.282 −0.465 (−2.677) 0.903 (−1.040) 0.395
INTC 0.038 (1.067) 0.948 (−0.302) 0.328 −0.263 (−1.724) 0.839 (−1.922) 0.404
IBM 0.039 (2.164) 0.670 (−2.707) 0.253 −0.594 (−2.975) 0.814 (−1.881) 0.264
AMER 0.145 (4.138) 0.596 (−5.697) 0.242 −0.408 (−5.318) 0.743 (−3.014) 0.271
DELL 0.126 (3.260) 0.543 (−3.954) 0.183 −0.583 (−4.082) 0.668 (−2.934) 0.202
CSCO −0.009 (−0.295) 0.957 (−0.419) 0.494 −0.117 (−0.863) 1.127 (0.983) 0.343
GE 0.026 (2.557) 0.657 (−4.075) 0.348 −0.660 (−4.239) 0.803 (−3.013) 0.455
CPQ 0.129 (2.922) 0.562 (−2.858) 0.187 −0.540 (−3.257) 0.673 (−2.455) 0.229
YHOO 0.354 (4.335) 0.499 (−4.679) 0.244 −0.204 (−3.569) 0.730 (−2.823) 0.299
SUNW 0.065 (1.295) 0.802 (−1.293) 0.468 −0.233 (−2.021) 0.892 (−1.134) 0.445
MU 0.221 (4.816) 0.582 (−5.174) 0.228 −0.309 (−4.773) 0.656 (−4.078) 0.255
MO 0.072 (5.406) 0.415 (−7.421) 0.135 −0.971 (−4.791) 0.641 (−3.795) 0.222
AMZN 0.565 (5.120) 0.309 (−8.470) 0.125 −0.246 (−4.035) 0.652 (−2.559) 0.238
ORCL 0.323 (3.758) 0.238 (−3.502) 0.095 −0.513 (−2.013) 0.600 (−1.847) 0.255
LU 0.141 (2.435) 0.554 (−2.232) 0.271 −0.330 (−2.348) 0.817 (−2.128) 0.454
TRV 0.066 (2.557) 0.673 (−2.921) 0.196 −0.722 (−3.327) 0.665 (−2.911) 0.220
WCOM 0.087 (3.201) 0.652 (−2.520) 0.316 −0.500 (−2.721) 0.751 (−2.305) 0.358
TYC 0.185 (3.862) 0.347 (−17.988) 0.233 −0.535 (−4.928) 0.851 (−2.009) 0.462
AMAT 0.132 (3.524) 0.670 (−3.464) 0.252 −0.377 (−4.362) 0.690 (−3.847) 0.275
QCOM 0.117 (2.474) 0.724 (−2.833) 0.281 −0.361 (−3.516) 0.747 (−2.439) 0.249
TXN 0.056 (1.489) 0.893 (−0.813) 0.402 −0.211 (−1.983) 0.839 (−1.971) 0.435
PFE 0.059 (4.270) 0.473 (−6.469) 0.122 −1.052 (−6.084) 0.587 (−4.779) 0.160
MOT 0.063 (2.393) 0.846 (−1.133) 0.387 −0.325 (−2.022) 0.794 (−2.348) 0.427
EMC 0.006 (0.099) 1.087 (0.461) 0.414 −0.203 (−1.450) 0.838 (−1.532) 0.399
HWP 0.069 (2.836) 0.733 (−2.849) 0.330 −0.378 (−2.659) 0.804 (−2.159) 0.349
AMGN 0.061 (2.617) 0.691 (−2.832) 0.363 −0.451 (−3.461) 0.809 (−2.395) 0.405
BRCM 0.240 (2.010) 0.742 (−1.692) 0.338 −0.112 (−1.802) 0.705 (−2.868) 0.322
MER 0.061 (2.844) 0.714 (−2.766) 0.295 −0.479 (−3.292) 0.758 (−2.726) 0.346
NOK 0.119 (2.872) 0.676 (−2.395) 0.265 −0.366 (−2.913) 0.734 (−2.776) 0.323
CHL 0.051 (2.525) 0.704 (−2.373) 0.326 −0.347 (−2.277) 0.889 (−1.377) 0.491
UNPH 0.213 (2.056) 0.809 (−1.080) 0.334 −0.092 (−1.050) 0.809 (−1.702) 0.372
EBAY 0.047 (0.506) 0.882 (−0.882) 0.408 −0.166 (−2.100) 1.207 (2.481) 0.637
JNPR 0.422 (2.837) 0.546 (−2.997) 0.226 −0.121 (−1.728) 0.553 (−3.302) 0.185
CIEN 0.435 (3.038) 0.580 (−2.784) 0.179 −0.103 (−1.362) 0.675 (−2.582) 0.305
BRCD 0.386 (2.561) 0.710 (−1.702) 0.195 −0.026 (−0.318) 0.666 (−2.341) 0.231

Entries report the GMM estimates (and t-statistics in parentheses) of the following relations,

Panel A: RVt,T = a + b SWt,T + e,

Panel B: ln RVt,T = a + b ln SWt,T + e.

The t-statistics are calculated according to Newey and West (1987) with 30 lags, under the null hypothesis of
a = 0, b = 1. Columns under “R 2 ” report the unadjusted R-squared.

be zero and the intercept estimate should be lower than zero due to the convexity
term induced by the variance of the log variance risk premiums. The estimation
results in Panel B of Table 6 show that for all the stock indexes and many of the
individual stocks, the slope estimates are no longer significantly different from
one at the 95% confidence level. The difference between the slope estimates
of the two regressions indicates that the risk premiums defined in log returns

1329
 The Review of Financial Studies / v 22 n 3 2009

is closer to a constant or independent series than the risk premiums defined in

dollar terms.

4. Robustness Analysis
Our results on variance risk premiums rely on the accuracy of the variance swap
rates that we synthesize from the options market. We perform several exercises
to verify the robustness of our results. First, we gauge the approximation error
of the synthetic variance swap rate due to price jumps and discretization.
Second, we analyze the impact of options bid-ask spreads on our results.
Third, we evaluate the impacts of error-in-variable problems on our expectation
hypothesis regressions where the synthetic variance swap rate is used as a
regressor. Finally, we analyze whether the variance risk premiums behavior
varies significantly over different subsample periods.

4.1 Replication errors due to price jumps and discretization

The replication of the payoff to a variance swap in Equation (5) has an in-
stantaneous error of order O(( dFt−Ft 3
) ). We refer to this error as jump error as
it vanishes under continuous path monitoring if there are no jumps. Further-
more, Equation (5) asks for a continuum of option prices at all strikes. We use
a simple interpolation/extrapolation scheme to generate 2,000 option prices
over ±8 standard deviations from the available option quotes. We then sum
the 2,000 option prices to replace the integration in Equation (5). The scheme
introduces a second source of error due to the interpolation/extrapolation and
the discrezation of the integral. We refer to this error as the discretization error.
To gauge the magnitude of these two sources of errors, we numerically illus-
trate three standard option pricing models: (i) the Black-Scholes model (BS),
(ii) the Merton (1976) jump-diffusion model (MJD), and (iii) a combination of
the MJD model with Heston (1993) stochastic volatility (MJDSV), as in Bates
(1996); and Bakshi, Cao, and Chen (1997). The risk-neutral dynamics of the
underlying futures price process under these three models are
BS: d Ft /Ft = σdWt ,
MJD: d Ft /Ft− = σdWt + d J (λ) − λgdt, (19)
√
MJDSV: d Ft /Ft− = vt dWt + d J (λ) − λgdt,
where Wt denotes a standard Brownian motion and J (λ) denotes a compound
Poisson jump process with constant intensity λ. Conditional on a jump oc-
curring, the MJD model assumes that the size of the jump in log price is
normally distributed with mean µ J and variance σ2J , with the mean percentage
price change induced by a jump given by g = eµ J + 2 σ J − 1. In the MJDSV
1 2

model, the diffusion variance rate vt is stochastic and follows a mean-reverting

square-root process
√
dvt = κ (θ − vt ) dt + σv vt d Z t , (20)

1330
Variance Risk Premiums

where Z t is another standard Brownian motion, correlated with Wt by

E[d Z t dWt ] = ρdt.
The MJDSV model nests the MJD model, which in turn nests the BS model.
We regard the progression from BS to MJD and then from MJD to MJDSV
as one of increasing complexity. All three models are analytically tractable,
allowing us to numerically calculate risk-neutral expected values of variance.
The difference in the BS model between the synthetic variance swap rate and
the constant variance rate are purely due to the discretization. The increase
in the error due to the use of the MJD model instead of the BS model allows us
to numerically gauge the magnitude of the jump error in the presence of discrete
strikes. The change in approximation error from MJD to MJDSV allows us to
numerically gauge the impact of stochastic volatility in the presence of discrete
strikes and jumps.
For the numerical analysis, we normalize the current futures price to $100
and assume a constant risk-free rate at r = 5.6%. We set σ = 0.37 in the BS
model and 0.35 in the MJD model. The other parameters are set to λ = 0.4,
µ J = −0.09, σ J = 0.18, κ = 1.04, θ = 0.35, σv = 0.9, and ρ = −0.7. These
parameters reflect approximately those estimated from the S&P 500 index
option prices (e.g., Bakshi, Cao, and Chen, 1997).
In parallel to our empirical study, we fix the option maturity to 30 days. We
assume that only five option quotes are available at this maturity at strikes of
$80, $90, $100, $110, and $120. Since all the stock indexes and individual
stocks in our data sample average no less than five strikes at each chosen
maturity, the choice of five options for the numerical analysis is reasonable
and conservative. First, we compute the prices of the five options under each
model and convert them into implied volatilities. Second, we employ the same
interpolation/extrapolation method as in our empirical study to obtain a fine
grid of 2,000 implied volatilities. Third, we convert the fine grid of implied
volatilities into out-of-the-money option prices and approximate the integral in
Equation (5) with a summation. From this procedure, we compute the synthetic
variance swap rate, SW t,T , where the hat stresses the approximations involved.
The difference between this approximate synthetic variance swap rate SW and
the analytically computed variance swap rate EQ t [RVt,T ] represents the total
approximation error.
Under the BS model, the annualized return variance rate is constant at σ2 .
Under the MJD model, this variance rate is constant at σ2 + λ(µ2J + σ2J ). In
both cases, the variance swap rate equals the constant variance. Under MJDSV,
the variance rate is stochastic, and the variance swap rate depends on the current
level of the instantaneous variance rate vt ,

  2 
EQ
t RVt,T = σt + λ µ J + σ J ,
2 2
(21)

1331
The Review of Financial Studies / v 22 n 3 2009

where σt2 is given by


1 T
1 − e−κ(T −t)
σt2 ≡ EQ
t vs ds = θ + (vt − θ) . (22)
T −t t κ(T − t)
Our replicating strategy for the variance swap contract is exact when the
underlying dynamics are purely continuous, but has a higher-order approxima-
tion error in the presence of jumps. Thus, under the BS model, the theoretical
approximation error is zero: ε = 0. Under the two jump-diffusion models, MJD
and MJDSV, the compound Poisson jump component has the following com-
pensator:
(x−µ J )2
1 −
ν(x) = λ  e 2σ2J
, (23)
2πσ2J

from which we can compute the jump-induced error ε according to Equation (6),
1 2 
ε = −2λ eµ J + 2 σ J − 1 − µ J −
1 2
µ J + σ2J . (24)
2

Table 7 reports the analytical variance swap rate (EQ t [RV ]), the synthetic
variance swap rate ( SW ), the total approximation error (EQ t [RV ] − SW ), and
the jump-induced error (ε) under each model. Under the BS model, the jump
error (ε) is zero. Furthermore, since the implied volatility is constant and equal
to σ at all strikes, there is no interpolation or extrapolation error. The only
potential error comes from the discretization of the integration. Table 7 shows
that this error is practically zero.
Under the MJD model, the jump error (ε) is 0.0011, which is merely 0.79%
of the variance level at 0.1387. The total error is 0.0021, only 1.51% of the
variance level.
Under the MJDSV model, we consider different instantaneous variance lev-
els, represented by its log difference from the mean, ln(vt /θ). As the variance
level vt varies, the jump error is fixed at 0.0011 because the jump arrival rate
does not change. Table 7 indicates that the total approximation error increases
with the volatility level. The largest absolute error is 0.0221 when the variance
swap rate reaches as high as 2.3782. The error is less than 1% of the variance
level. Therefore, even under stochastic volatility and when the volatility level is
very high, the interpolation and extrapolation across the five implied volatility
quotes do not add much additional approximation error. The numerical exercise
shows that our simple interpolation and extrapolation method works well.

4.2 Bid-ask spreads

We synthesize variance swap rates by interpolating implied volatilities com-
puted from the mid quotes of the option prices. The mid-quote may not reflect
the fair price if the bid and ask quotes are not symmetric around the fair price.

1332
Variance Risk Premiums

Table 7
Numerical illustration of the approximation error for variance swap rates
ln vt /θ EQ [RV ] SW Total error (EQ [RV ] − SW ) Jump error (ε)

The Black-Scholes model

0.0 0.1369 0.1369 0.0000 0.0000

The Merton jump-diffusion model

0.0 0.1387 0.1366 0.0021 0.0011

The MJD–stochastic volatility model

−3.0 0.0272 0.0273 −0.0001 0.0011

−2.5 0.0310 0.0313 −0.0003 0.0011
−2.0 0.0372 0.0376 −0.0004 0.0011
−1.5 0.0475 0.0477 −0.0001 0.0011
−1.0 0.0645 0.0637 0.0008 0.0011
−0.5 0.0925 0.0905 0.0020 0.0011
0.0 0.1387 0.1356 0.0031 0.0011
0.5 0.2148 0.2107 0.0041 0.0011
1.0 0.3403 0.3353 0.0051 0.0011
1.5 0.5472 0.5410 0.0062 0.0011
2.0 0.8884 0.8799 0.0085 0.0011
2.5 1.4509 1.4377 0.0132 0.0011
3.0 2.3782 2.3561 0.0221 0.0011

Entries report the analytical 30-day variance swap rate (EQ [RV ]), the synthetic approximation of the
variance swap rate ( SW ) based on interpolation and extrapolation over five implied volatility quotes,
the total approximation error (Total error = EQ [RV ] − SW ), and the error induced by jumps in the
underlying asset price (ε) under each model. For the MJD–stochastic volatility model, the first column
denotes the log difference between the current instantaneous variance level vt and its long-run mean θ.
For ease of comparison, we represent all swap rates and errors in volatility percentage points.

To gauge how much our conclusions are affected by the mid-quote choice, we
reconstruct the synthetic variance swap rates using bid and ask option prices,
respectively. When direct quotes are not available, we can regard these as
synthetic bid and ask swap rate quotes, respectively. For European options,
we directly convert the bid and ask option prices into bid and ask implied
volatilities, and perform the interpolation and extrapolation on each side. For
American options, we first convert the OptionMetrics implied volatility into a
mid-quote European option price. Then we superimpose the bid-ask spread of
the American option quotes on this mid price to generate bid and ask European
option prices, from which we compute the bid and ask implied volatilities,
respectively.
Table 8 reports the sample averages of the synthetic bids and asks of the
variance swap rates, as well as the variance risk premiums defined in both dollar
terms and log returns. For the risk premiums, we also report in parentheses the
serial-dependence-adjusted t-statistics on the significance of the mean value.
The bid-ask spreads for the synthetic variance swap rates range from 0.82 to
16.75. Converting the variance swap rates into volatility percentage points per-
industry quoting convention, we obtain the average bid-ask spreads ranging

1333
The Review of Financial Studies / v 22 n 3 2009

Table 8
Mean synthetic variance swap rates and variance risk premiums from bid and ask option prices
Ticker Panel A: SW×100 Panel B: (RV − SW ) × 100 Panel C: ln(RV /SW )
Ask Bid Ask Bid Ask Bid

SPX 7.52 6.41 −3.45 (−9.81) −2.34 (−7.44) −0.76 (−13.48) −0.60 (−10.81)
OEX 7.44 6.62 −2.90 (−8.32) −2.08 (−6.26) −0.65 (−11.73) −0.54 (−9.53)
DJX 7.90 6.33 −3.51 (−8.15) −1.94 (−4.94) −0.73 (−10.93) −0.51 (−7.60)
NDX 20.76 18.09 −4.06 (−4.29) −1.40 (−1.43) −0.36 (−8.29) −0.23 (−5.30)
QQQ 29.48 24.68 −6.88 (−4.51) −2.07 (−1.39) −0.40 (−6.51) −0.22 (−3.74)
MSFT 22.34 18.12 −5.75 (−5.59) −1.53 (−1.62) −0.41 (−9.18) −0.21 (−4.61)
INTC 27.69 23.60 −0.03 (−0.01) 4.07 ( 2.15) −0.11 (−2.37) 0.04 ( 0.90)
IBM 18.23 16.06 −3.08 (−3.24) −0.92 (−0.99) −0.32 (−5.73) −0.20 (−3.50)
AMER 48.68 42.26 −7.55 (−4.37) −1.13 (−0.65) −0.26 (−5.73) −0.12 (−2.49)
DELL 41.51 34.57 −8.61 (−4.17) −1.67 (−0.81) −0.33 (−6.14) −0.15 (−2.69)
CSCO 37.03 31.15 −5.93 (−3.96) −0.05 (−0.03) −0.37 (−5.54) −0.21 (−3.00)
GE 16.12 12.80 −4.21 (−5.97) −0.89 (−1.46) −0.37 (−8.16) −0.14 (−3.35)
CPQ 37.18 28.60 −6.17 (−3.26) 2.40 ( 1.34) −0.27 (−5.05) −0.02 (−0.31)
YHOO 79.74 70.77 −7.49 (−1.82) 1.48 ( 0.38) −0.16 (−2.94) −0.05 (−0.83)
SUNW 43.28 33.69 −6.58 (−3.72) 3.02 ( 1.71) −0.25 (−5.58) 0.01 ( 0.12)
MU 64.64 56.35 −7.96 (−3.21) 0.33 ( 0.13) −0.18 (−4.93) −0.05 (−1.24)
MO 17.84 13.72 −4.21 (−3.55) −0.09 (−0.09) −0.38 (−5.94) −0.12 (−1.94)
AMZN 112.47 98.81 −23.84 (−3.08) −10.18 (−1.53) −0.30 (−4.39) −0.17 (−2.66)
ORCL 53.46 44.59 −9.73 (−1.82) −0.86 (−0.16) −0.25 (−3.58) −0.05 (−0.65)
LU 34.63 29.07 −3.20 (−1.52) 2.36 ( 1.07) −0.17 (−3.45) −0.01 (−0.23)
TRV 21.38 17.41 −2.02 (−1.26) 1.95 ( 1.20) −0.24 (−3.70) −0.03 (−0.40)
WCOM 32.49 24.44 −5.65 (−2.04) 2.40 ( 0.98) −0.29 (−3.85) 0.02 ( 0.31)
TYC 45.10 37.94 −12.49 (−2.18) −5.33 (−1.07) −0.45 (−5.22) −0.26 (−3.10)
AMAT 51.27 42.13 −7.38 (−3.59) 1.76 ( 0.86) −0.23 (−5.74) −0.01 (−0.36)
QCOM 53.38 45.76 −6.40 (−2.54) 1.22 ( 0.49) −0.25 (−4.49) −0.08 (−1.47)
TXN 39.23 32.91 −1.99 (−1.07) 4.32 ( 2.35) −0.13 (−3.08) 0.05 ( 1.31)
PFE 16.14 12.82 −3.48 (−4.21) −0.17 (−0.21) −0.34 (−5.34) −0.11 (−1.68)
MOT 30.43 24.80 −1.16 (−0.64) 4.48 ( 2.49) −0.13 (−2.46) 0.08 ( 1.50)
EMC 42.83 34.83 −0.90 (−0.29) 7.10 ( 2.19) −0.15 (−2.81) 0.08 ( 1.48)
HWP 27.60 23.20 −2.41 (−1.80) 1.99 ( 1.57) −0.19 (−3.74) −0.00 (−0.03)
AMGN 28.33 23.95 −4.55 (−3.02) −0.17 (−0.11) −0.27 (−5.10) −0.08 (−1.42)
BRCM 97.50 86.72 −6.28 (−1.11) 4.50 ( 0.81) −0.12 (−2.29) −0.01 (−0.20)
MER 26.18 22.75 −2.92 (−2.50) 0.52 ( 0.44) −0.20 (−4.48) −0.05 (−1.10)
NOK 35.57 29.92 −1.90 (−0.97) 3.74 ( 1.95) −0.13 (−2.18) 0.04 ( 0.67)
CHL 22.59 18.38 −3.36 (−2.18) 0.85 ( 0.59) −0.27 (−5.02) −0.05 (−1.05)
UNPH 82.99 74.07 0.75 ( 0.13) 9.68 ( 1.61) −0.09 (−1.44) 0.03 ( 0.44)
EBAY 78.23 70.37 −9.07 (−1.59) −1.22 (−0.21) −0.34 (−4.37) −0.23 (−2.95)
JNPR 122.76 110.14 −18.05 (−2.24) −5.43 (−0.73) −0.21 (−2.98) −0.10 (−1.53)
CIEN 99.05 86.25 −2.79 (−0.34) 10.01 ( 1.33) −0.11 (−1.38) 0.03 ( 0.38)
BRCD 110.90 94.16 −0.80 (−0.09) 15.95 ( 1.80) −0.09 (−1.20) 0.07 ( 0.88)

Entries compare sample mean of synthetic variance swap rates, synthetic volatility swap rate, and variance risk
premiums when the synthesis is based on ask and bid option prices, respectively. We also report in parentheses
the t-statistics on the significance of the mean variance risk premiums, which are adjusted for serial dependence
according to the Newey-West method with a lag of 30 days.

from 1.55 to 8.28 volatility percentage points. Overall, the spreads are larger
for individual stocks than for stock indexes.
It is important to point out that currently there exists an active over-the-
counter market for variance swap contracts on stock indexes. Although it is
difficult to retrieve long histories, current quotes from several broker dealers
are readily available from common financial data sources. The bid-ask spreads
on these variance swap rate quotes are normally within one volatility percentage
point. Our synthetic variance swap bid-ask spreads for the five stock indexes

1334
Variance Risk Premiums

are from 1.55 to 4.62 volatility points, much wider than the actual spreads from
the over-the-counter market.
Nevertheless, even with the exaggerated bid-ask spreads, our main conclu-
sions on the variance risk premiums remain valid whether we measure the
premiums using the synthetic bid swap rates or ask swap rates. Using the syn-
thetic ask rates makes the variance risk premiums even more negative. Using
the bid swap rates lowers the absolute magnitude of the negative risk premiums.
However, even when we use the synthetic bid swap rates to compute the vari-
ance risk premiums, the premiums remain significantly negative for the S&P
and Dow indexes, whether the premiums are measured in dollar terms or log
returns.

4.3 Errors in variables

Since the synthetic variance swap rates are measured with error, the error-
in-variable issue arises when they are used as regressors in the expectation
hypothesis regressions in Equations (17)–(18). Thus, the fact that the slope
estimate for Equation (17) is significantly below the null hypothesis of one for
the S&P and Dow indexes could be either due to time-varying risk premium, as
we have conjectured, or simply due to the bias induced by the error-in-variable
problem.
To gauge the size of the bias caused by the error-in-variable problem in
Equation (17), we propose the following expanded formulation:

RVt,T = a + bSWt,T + e,
(25)
SW t,T = SWt,T + η,

where SW denotes the true swap rate and SW denotes the synthetic swap rate,
which is regarded as a noisy estimator of the true swap rate, with η capturing
the measurement error. Furthermore, we specify an auxiliary AR(1) dynamic
for the true 30-day variance swap rate

SWt+1 = θ(1 − φ) + φSWt+1 + εt+1 . (26)

If we assume independent normal distributions on the error terms (e, η, ε)

with variance (σe2 , ση2 , σε2 ), respectively, we can use the maximum likelihood
method jointly with Kalman filter to estimate the parameters of the system. In
this estimation, we regard Equation (26) as the state-propagation equation and
Equation (25) as the measurement equation. Given initial parameter guesses,
we use Kalman filter to obtain the forecasted mean values and variances on the
measurement series. Then, we construct the likelihood based on the forecast-
ing errors, which are normally distributed under our assumption. The model
parameters (a, b, θ, φ, σe2 , ση2 , σε2 ) are estimated by maximizing the likelihood
value. Using this method, we learn not only the bias-corrected expectation hy-
pothesis coefficients (a, b), but also the variances of the measurement errors
and the true swap rates.

1335
The Review of Financial Studies / v 22 n 3 2009

Table 9
Maximum likelihood estimates of the expectation hypothesis regression on the S&P 500 index
Ticker SPX OEX DJX

a −0.144 (0.162) −0.151 (0.160) −0.110 (0.184)

b 0.618 (0.012) 0.698 (0.013) 0.569 (0.013)
θ 3.719 (1.383) 3.720 (1.422) 3.748 (0.921)
φ 0.988 (0.002) 0.990 (0.003) 0.995 (0.003)
σe2 6.782 (0.152) 6.803 (0.136) 7.083 (0.168)
ση2 1.989 (0.031) 1.851 (0.105) 1.122 (0.049)
σε2 0.549 (0.028) 0.422 (0.044) 0.536 (0.036)

Entries report the maximum likelihood estimates of the parameters (and standard errors in parentheses) of the
following system of equations:

RVt,T = a + bSWt,T + e,
SW t,T = SWt,T + η,
SWt+1 = θ(1 − φ) + φSWt+1 + εt+1 ,

where the error terms are independently normal with zero mean and variances (σe2 , ση2 , σε2 ). The realized variance
and swap rates are both scaled up by 100 for the estimation.

We perform the likelihood estimation on the three S&P and Dow indexes that
have generated regression slope coefficients significantly lower than 1. Table 9
reports the maximum likelihood estimates and standard errors of the parameters.
Take the S&P 500 index as an example. The slope estimate is 0.618, larger than
the least squares estimate of 0.455 reported in Table 6. The difference between
the two estimates captures the bias induced by the measurement errors in the
synthetic variance swap rates. Nevertheless, after correcting for this bias, the
slope coefficient on the S&P 500 index remains significantly lower than the
null value of 1. The results are similar for the other two indexes. Therefore,
our earlier conclusion remains valid after controlling for the error-in-variable
issue. Especially for the S&P and Dow indexes, the expectation hypothesis
regression slope estimate is significantly below the null value of 1, suggesting
that the variance risk premium in dollar terms is time-varying and correlated
with the variance swap rate.

4.4 Subsample analysis

The stock market had been largely bullish since the beginning of our sample
in 1996 until the burst of the NASDAQ bubble in March 2000, after which the
stock market has been going down till the end of our sample in 2003. As a
concrete example, the S&P 500 index started at around 600 in January 1996,
and climbed over 1,500 points before it started to fall after March 2000. By the
end of our sample in February 2003, the S&P 500 index retreated to around
800. Thus, we can largely divide our whole sample into two subsample periods,
a bullish period from 1996 to March 2000, and a bearish period after March
2000.
To study whether the variance risk premiums behavior varies in bullish
versus bearish market conditions, we divide our sample into two subsamples,

1336
Variance Risk Premiums

Table 10
Summary statistics of variance risk premiums from different subsamples
Ticker RV×100 SW×100 (RV − SW ) × 100 ln(RV /SW )
S1 S2 S1 S2 S1 S2 S1 S2

SPX 3.23 5.31 6.12 7.83 −2.89 (−8.99) −2.52 (−7.83) −0.76 (−14.26) −0.52 (−9.91)
OEX 3.51 6.06 5.84 8.48 −2.32 (−8.02) −2.42 (−6.28) −0.65 (−12.55) −0.47 (−8.30)
DJX 3.84 4.88 6.91 7.03 −3.07 (−6.96) −2.15 (−6.19) −0.72 (−10.49) −0.51 (−8.62)
NDX 9.84 26.51 12.19 29.05 −2.35 (−4.78) −2.55 (−1.90) −0.32 (−8.15) −0.23 (−4.86)
QQQ 15.15 25.36 18.74 29.42 −3.59 (−5.45) −4.06 (−2.42) −0.31 (−8.93) −0.28 (−4.37)
MSFT 12.95 22.12 16.37 24.96 −3.43 (−6.26) −2.84 (−2.21) −0.31 (−8.37) −0.27 (−5.24)
INTC 18.03 42.50 18.06 36.13 −0.03 (−0.04) 6.37 ( 2.32) −0.05 (−1.52) 0.03 ( 0.53)
IBM 13.06 18.26 14.55 20.21 −1.49 (−1.85) −1.95 (−1.78) −0.23 (−4.47) −0.27 (−4.27)
AMER 44.84 34.67 48.24 38.35 −3.41 (−1.94) −3.69 (−2.24) −0.16 (−3.67) −0.20 (−4.02)
DELL 30.24 36.85 37.71 36.75 −7.47 (−5.10) 0.10 ( 0.04) −0.27 (−6.61) −0.17 (−2.40)
CSCO 18.78 55.54 24.37 51.31 −5.59 (−5.97) 4.23 ( 1.94) −0.40 (−5.67) −0.02 (−0.54)
GE 8.63 15.99 10.33 18.91 −1.70 (−3.43) −2.92 (−3.80) −0.24 (−5.53) −0.26 (−5.71)
CPQ 25.87 48.62 28.96 43.06 −3.09 (−2.32) 5.56 ( 2.08) −0.17 (−3.56) −0.01 (−0.13)
YHOO 67.28 77.79 65.33 83.51 1.95 ( 0.48) −5.72 (−1.54) −0.06 (−0.91) −0.12 (−2.82)
SUNW 26.75 64.20 30.03 58.59 −3.29 (−2.57) 5.61 ( 2.38) −0.17 (−3.74) 0.03 ( 0.92)
MU 49.94 67.42 50.57 73.56 −0.63 (−0.33) −6.14 (−2.04) −0.07 (−1.96) −0.15 (−3.85)
MO 13.03 14.53 13.91 17.74 −0.88 (−0.77) −3.21 (−3.26) −0.16 (−2.55) −0.37 (−5.93)
AMZN 99.63 77.32 93.34 114.57 6.30 ( 1.38) −37.24 (−5.15) 0.00 ( 0.02) −0.44 (−7.69)
ORCL 32.94 61.56 41.31 59.37 −8.37 (−1.34) 2.19 ( 0.70) −0.17 (−2.15) −0.07 (−1.50)
LU 21.70 61.12 23.49 54.93 −1.79 (−1.50) 6.19 ( 1.80) −0.12 (−2.59) 0.06 ( 1.23)
TRV 19.25 19.47 18.76 19.26 0.49 ( 0.44) 0.21 ( 0.11) −0.05 (−0.99) −0.19 (−2.50)
WCOM 18.68 47.27 19.67 48.19 −0.99 (−0.73) −0.91 (−0.21) −0.15 (−2.22) −0.08 (−0.97)
TYC 24.83 37.35 25.30 50.17 −0.47 (−0.22) −12.81 (−2.03) −0.21 (−2.68) −0.42 (−4.88)
AMAT 37.21 54.36 37.99 57.98 −0.78 (−0.51) −3.62 (−1.45) −0.08 (−2.16) −0.15 (−3.65)
QCOM 38.79 57.82 40.01 60.22 −1.22 (−0.54) −2.40 (−0.86) −0.15 (−2.55) −0.16 (−3.09)
TXN 27.67 51.51 25.81 49.70 1.86 ( 1.43) 1.80 ( 0.75) −0.00 (−0.06) −0.06 (−1.44)
PFE 13.11 12.07 14.16 14.23 −1.05 (−1.66) −2.16 (−2.24) −0.12 (−2.96) −0.33 (−4.08)
MOT 20.00 49.25 18.78 44.97 1.21 ( 1.34) 4.28 ( 1.55) 0.01 ( 0.29) −0.08 (−1.19)
EMC 27.37 66.85 28.58 54.28 −1.20 (−0.83) 12.57 ( 2.74) −0.08 (−1.82) 0.07 ( 1.15)
HWP 19.88 36.02 19.05 36.95 0.83 ( 0.62) −0.94 (−0.78) −0.06 (−1.17) −0.12 (−3.01)
AMGN 21.81 26.00 19.52 32.60 2.29 ( 2.07) −6.60 (−4.01) −0.04 (−0.81) −0.30 (−5.66)
BRCM 73.73 99.72 64.07 103.44 9.67 ( 3.55) −3.72 (−0.60) 0.09 ( 2.43) −0.12 (−2.22)
MER 23.24 23.30 21.65 27.01 1.59 ( 1.28) −3.71 (−4.15) −0.03 (−0.64) −0.21 (−5.61)
NOK 25.34 43.30 23.27 42.43 2.08 ( 1.07) 0.86 ( 0.46) −0.02 (−0.24) −0.05 (−1.17)
CHL 15.39 25.06 14.85 28.09 0.54 ( 0.50) −3.03 (−1.62) −0.10 (−1.92) −0.22 (−4.45)
UNPH 58.47 112.73 56.11 101.30 2.37 ( 0.53) 11.43 ( 1.61) −0.04 (−0.65) 0.01 ( 0.23)
EBAY 94.39 58.85 94.19 64.49 0.21 ( 0.03) −5.64 (−1.28) −0.05 (−0.70) −0.36 (−4.89)
JNPR 82.78 109.59 90.77 119.78 −7.99 (−1.25) −10.19 (−1.32) −0.14 (−1.83) −0.14 (−2.20)
CIEN 80.18 115.23 71.03 114.64 9.15 ( 1.27) 0.59 ( 0.07) 0.03 ( 0.35) −0.09 (−1.28)
BRCD 86.29 112.83 68.60 104.42 17.69 ( 3.46) 8.41 ( 0.97) 0.09 (1.61) −0.01 (−0.11)

Entries report the sample averages of the annualized realized variance, the synthetic variance swap rates, and
the (log) variance risk premiums during two subsamples. The first subsample (S1) is from 4 January 1996 to
23 March 2000. The second subsample (S2) is from 24 March 2000 to 28 February 2003. We also report in
parentheses the t-statistics on the significance of the mean variance risk premiums, which are adjusted for serial
dependence according to the Newey-West method with a lag of 30 days.

with 24 March 2000 as the dividing point. The first subsample includes dates
before 23 March 2000. The second subsample includes 24 March 2000 and
after. Table 10 reports the summary statistics of the realized variance, variance
swap rates, and the variance risk premiums under the two subsample periods.
On average, both the realized variance and the variance swap rates are higher
during the bearish period (the second subsample) than during the bullish period.
Nevertheless, the variance risk premiums are strongly negative under both
market conditions for the S&P and Dow indexes.

1337
 The Review of Financial Studies / v 22 n 3 2009

5. Conclusion
In this paper, we propose a direct and robust method to quantify the variance
risk premiums on financial assets underlying options. Our method uses the
notion of a variance swap, which is an over-the-counter contract that pays
the difference between the realized variance and the fixed swap rate. Since
the variance swap rate represents the risk-neutral expected value of the realized
variance, we propose to use the difference between the realized variance and
the variance swap rate to quantify the variance risk premium. We show that
the variance swap rate can be well approximated by the value of a particular
portfolio of options. Using a large options data set, we synthesize variance
swap rates and analyze variance risk premiums on five stock indexes and 35
individual stocks.
We find that the variance risk premiums are strongly negative for the S&P and
Dow indexes. Further analysis shows that there exists a systematic variance risk
factor in the stock market that asks for a highly negative risk premium. When
we investigate whether the classic asset pricing model can explain the negative
variance risk premiums, we find that the well-documented negative correlation
between index returns and volatility generates a strongly negative beta, but this
negative beta can only explain a small portion of the negative variance risk
premiums. The Fama-French factors cannot account for the strongly negative
variance risk premiums, either. Therefore, we conclude that either there is a
large inefficiency in the market for index variance or else the majority of the
variance risk is generated by an independent risk factor that the market prices
heavily. The negative sign on the variance risk premiums indicates that investors
regard market volatility going up as an unfavorable shock, and are willing to
pay a large premium to hedge against market volatility going up.
To analyze the dynamic properties of the variance risk premiums, we formu-
late expectation hypothesis regressions. When we regress the realized variance
on the variance swap rate, we obtain slope estimates that are significantly lower
than 1, the null value under the hypothesis of constant or independent variance
risk premiums. The slope estimates become closer to 1 when the regression is
on the logarithm of variance. These regression results indicate that although
the log variance risk premiums are strongly negative, they are not strongly
correlated with the logarithm of the variance swap rate.
The simple, direct, and robust method that we propose to measure variance
risk premium opens fertile ground for future research. Given the evidence on
stochastic variance and strongly negative variance risk premiums, it is important
to understand the pricing kernel behavior as a function of both the market
portfolio return and return variance. Recent studies (e.g., Jackwerth, 2000; and
Engle and Rosenberg, 2002) have found some puzzling behaviors on the pricing
kernel projected on the equity index return alone. Accurately estimating the
pricing kernel as a joint function of the index return and return variance can
prove fruitful not only for understanding the variance risk premiums behavior,

1338
Variance Risk Premiums

but also for resolving the puzzling behaviors observed on the pricing kernels
projected on the index return alone.
The empirical analysis in this paper focuses on the variance swap rate and
variance risk premium over a fixed 30-day horizon. As over-the-counter vari-
ance swap rate quotes are becoming increasingly available at many different
maturities, an important line for future research is to design and estimate
stochastic return variance models that can capture the term structure of vari-
ance swap rates and variance risk premiums (e.g., Egloff, Leippold, and Wu,
2006).

Appendix: Synthesizing Variance Swap Contracts

Let f (F) be a twice differentiable function of F. By Itô’s lemma for semi-martingales,
 T  T
1
f (FT ) = f (Ft ) + f  (Fs− )d Fs + f  (Fs− )σs2− ds
t 2 t
 T 
+ [ f (Fs− e x ) − f (Fs− ) − f  (Fs− )Fs− (e x − 1)]µ(d x, ds). (A1)
t R0

Applying Equation (A1) to the function f (F) = ln F, we have

 T  T  T 
1 1
ln(FT ) = ln(Ft ) + d Fs − σs−
2
ds + [x − e x + 1]µ(d x, ds). (A2)
t Fs− 2 t t R0
T
Adding and subtracting 2[ FFTt − 1] + t x 2 µ(d x, ds) and rearranging, we obtain the following
representation:
 T  T    T  
FT FT 1 1
Vt,T ≡ σs−
2
ds + x 2 µ(d x, ds) = 2 − 1 − ln +2 − d Fs
t t Ft Ft t Fs− Ft
 T   
x2
−2 ex − 1 − x − µ(d x, ds). (A3)
t R0 2

A Taylor expansion with remainder of ln FT about the point Ft implies

 Ft  ∞
1 1 1
ln FT = ln Ft + (FT − Ft ) − (K − FT )+ d K − (FT − K )+ d K . (A4)
Ft 0 K2 Ft K2

Combining Equations (A3) and (A4) and noting that FT = ST , we have

 Ft  ∞ 
1 + 1 +
Vt,T = 2 (K − S T ) d K + (S T − K ) d K
0 K2 Ft K
2
 T  
1 1
+2 − d Fs
t Fs− Ft
 T   
x2
−2 ex − 1 − x − µ(d x, ds). (A5)
t R0 2

Thus, we can replicate the return quadratic variation up to time T by the sum of (i) the payoff
from a static position in 2dK K2 European options on the underlying spot at strike K and maturity T
(first line), (ii) the payoff from a dynamic trading strategy holding 2Bs (T )[ F1s− − F1t ] futures at
time s (second line), and (iii) a higher-order error term induced by the discontinuity in the futures

1339
The Review of Financial Studies / v 22 n 3 2009

price dynamics (third line). The options are all out-of-the-money forward, i.e., call options when
Ft > K and put options when K ≤ Ft .
Taking expectations under measure Q on both sides, we obtain the risk-neutral expected value
of the quadratic variation on the left-hand side. We also obtain the forward value of the startup cost
of the replicating strategy and the replication error on the right-hand side,
 ∞    

 2t (K , T ) T x2
EQ
t Vt,T = d K − 2EQ
t ex − 1 − x − νs (x)d xds. (A6)
0 Bt (T )K 2 t R0 2

By the martingale property, the expected value of the gains from dynamic futures trading is zero
under the risk-neutral measure. Dividing by (T − t) on both sides, we obtain the result on the
annualized return quadratic variation.

References
Aı̈t-Sahalia, Y. 2002. Maximum-Likelihood Estimation of Discretely-Sampled Diffusions: A Closed-Form Ap-
proximation Approach. Econometrica 70(1):223–62.

Aı̈t-Sahalia, Y. 2008. Closed-Form Likelihood Expansions for Multivariate Diffusions. Annals of Statistics,
36(2):906–37.

Aı̈t-Sahalia, Y., and R. Kimmel. 2007. Maximum Likelihood Estimation of Stochastic Volatility Models. Journal
of Financial Economics 83(2):413–52.

Ang, A., R. J. Hodrick, Y. Xing, and X. Zhang. 2006. The Cross-Section of Volatility and Expected Returns.
Journal of Finance 61(1):259–99.

Bakshi, G., C. Cao, and Z. Chen. 1997. Empirical Performance of Alternative Option Pricing Models. Journal
of Finance 52(5):2003–49.

Bakshi, G., and N. Kapadia. 2003a. Delta-Hedged Gains and the Negative Market Volatility Risk Premium.
Review of Financial Studies 16(2): 527–66.

Bakshi, G., and N. Kapadia. 2003b. Volatility Risk Premium Embedded in Individual Equity Options: Some New
Insights. Journal of Derivatives 11(1):45–54.

Bates, D. 1996. Jumps and Stochastic Volatility: Exchange Rate Processes Implicit in Deutsche Mark Options.
Review of Financial Studies 9(1):69–107.

Bates, D. 2000. Post-’87 Crash Fears in the S&P 500 Futures Option Market. Journal of Econometrics 94(1–
2):181–238.

Bates, D. S. 2003. Empirical Option Pricing: A Retrospection. Journal of Econometrics 116(1–2):387–404.

Bollerslev, T., M. Gibson, and H. Zhou. 2004. Dynamic Estimation of Volatility Risk Premia and Investor Risk
Aversion from Option-Implied and Realized Volatilities. Working Paper, Duke University and Federal Reserve
Board.

Bondarenko, O. 2004. Market Price of Variance Risk and Performance of Hedge Funds. Working Paper, University
of Illinois at Chicago.

Britten-Jones, M., and A. Neuberger. 2000. Option Prices, Implied Price Processes, and Stochastic Volatility.
Journal of Finance 55(2):839–66.

Carhart, M. M. 1997. On Persistence in Mutual Fund Performance. Journal of Finance 52(1):57–82.

Carr, P., and D. Madan. 1998. Towards a Theory of Volatility Trading, in R. Jarrow (ed.), Risk Book on Volatility.
New York: Risk, pp. 417–27.

Carr, P., and L. Wu. 2006. A Tale of Two Indices. Journal of Derivatives 13(3):13–29.

Coval, J. D., and T. Shumway. 2001. Expected Option Returns. Journal of Finance 56(3):983–1009.

1340
Variance Risk Premiums

Cox, J. C. 1996. The Constant Elasticity of Variance Option Pricing Model. Journal of Portfolio Management
23(1):15–17.

Demeterfi, K., E. Derman, M. Kamal, and J. Zou. 1999. A Guide to Volatility and Variance Swaps. Journal of
Derivatives 6(4):9–32.

Derman, E., and I. Kani. 1994. Riding on a Smile. Risk 7(2):32–9.

Duffie, D. 1992. Dynamic Asset Pricing Theory, 2nd ed. Princeton, NJ: Princeton University Press.

Dupire, B. 1994. Pricing with a Smile. Risk 7(1):18–20.

Egloff, D., M. Leippold, and L. Wu. 2006. Variance Risk Dynamics, Variance Risk Premia, and Optimal Variance
Swap Investments. Working Paper, Zurich Cantonalbank, Imperial College, and Baruch College.

Engle, R. F., and J. V. Rosenberg. 2002. Empirical Pricing Kernels. Journal of Financial Economics 64(3):341–72.

Eraker, B. 2004. Do Stock Prices and Volatility Jump? Reconciling Evidence from Spot and Option Prices.
Journal of Finance 59(3):1367–1404.

Fama, E. F., and K. R. French. 1993. Common Risk Factors in the Returns on Stocks and Bonds. Journal of
Financial Economics 33(1):3–56.

Goetzmann, W. N., J. E. Ingersoll Jr., M. I. Spiegel, and I. Welch. 2002. Sharpening Sharpe Ratios. ICF Working
Paper 02-08, Yale.

Heston, S. 1993. Closed-Form Solution for Options with Stochastic Volatility, with Application to Bond and
Currency Options. Review of Financial Studies 6(2):327–43.

Hull, J., and A. White. 1987. The Pricing of Options on Assets with Stochastic Volatilities. Journal of Finance
42:281–300.

Jackwerth, J. C. 2000. Recovering Risk Aversion from Option Prices and Realized Returns. Review of Financial
Studies 13(2):433–51.

Jacod, J., and A. N. Shiryaev. 1987. Limit Theorems for Stochastic Processes. Berlin: Springer.

Jegadeesh, N., and S. Titman. 1993. Returns to Buying Winners and Selling Losers: Implications for Stock
Market Efficiency. Journal of Finance 48(1):65–91.

Jiang, G. J., and Y. S. Tian. 2005. The Model-Free Implied Volatility and Its Information Content. Review of
Financial Studies 18(4):1305–42.

Jones, C. S. 2003. The Dynamics of Stochastic Volatility: Evidence From Underlying and Options Markets.
Journal of Econometrics 116(1–2):181–224.

Kraus, A., and R. H. Litzenberger. 1976. Skewness Preference and the Valuation of Risk Assets. Journal of
Finance 31(4):1085–1100.

Merton, R. C. 1976. Option Pricing When Underlying Stock Returns Are Discontinuous. Journal of Financial
Economics 3(1):125–44.

Newey, W. K., and K. D. West. 1987. A Simple, Positive Semi-Definite, Heteroskedasticity and Autocorrelation
Consistent Covariance Matrix. Econometrica 55(3):703–8.

Pan, J. 2002. The Jump-Risk Premia Implicit in Options: Evidence from an Integrated Time-Series Study. Journal
of Financial Economics 63(1):3–50.

Wu, L. 2005. Variance Dynamics: Joint Evidence from Options and High-Frequency Returns. Working Paper,
Baruch College.

1341