Electronic copy available at: http://ssrn.

com/abstract=2234438
What does the volatility risk premium say about liquidity
provision and demand for hedging tail risk?

Jianqing Fan
jqfan@princeton.edu
Michael B. Imerman

mbi212@lehigh.edu
Wei Dai
weidai@princeton.edu
March 2013
JEL Classication: C01, C58, G12

Jianqing Fan is the Frederick L. Moore Professor of Finance and Statistics, Department of
Operation Research & Financial Engineering, Princeton University, Princeton, NJ 08544; Michael B.
Imerman is an Assistant Professor, Perella Department of Finance, Lehigh University, Bethlehem, PA
18015; and Wei Dai is a Ph.D. Student, Department of Operations Research & Financial Engineering,
Princeton University, Princeton, NJ 08544.

We would like to thank participants at the 20

th
Annual Conference on Pacic Basin Finance,
Economics, Accounting, and Management at Rutgers University and the Second Measuring Risk
Conference at Princeton University. We are especially grateful for comments from Yacine At-
Sahalia and Peter Carr. Also, feedback from the Derivatives Team at Morgan Stanley provided
useful insight into the events of Fall 2008.

Contact Author Email: mbi212@lehigh.edu; Phone: (610) 758-6380.

Electronic copy available at: http://ssrn.com/abstract=2234438
What does the volatility risk premium say about liquidity
provision and demand for hedging tail risk?
ABSTRACT
This paper examines the volatility risk premium, dened as the dierence
between expected future volatility under the risk-neutral measure and the ex-
pectation under the physical measure. This risk premium represents the price
of volatility risk in nancial markets. It is standard to use the VIX volatility
index to proxy the expectation under the risk-neutral measure. Estimation
under the physical measure is less straightforward. Using ultra-high-frequency
transaction data on SPDR, the S&P500 ETF, we implement a novel approach
for estimating integrated volatility on the frequency domain which allows us
to isolate microstructure noise from the true volatility. Once we compute the
volatility risk premium, we perform a comprehensive econometric analysis to
help identify its determinants. We nd that analyzing the sign and magni-
tude components of the volatility risk premium provides greater insight into
the underlying economic drivers, most notably supply and demand forces in
the market for hedging tail risk as well as the role of intermediaries in this mar-
ket. Our results are consistent with previous studies and are able to reconcile
dierent interpretations of the volatility risk premium in the literature.
1 Introduction
The past decade has seen unprecedented swings in volatility. Consequently market
participants are no longer just concerned with the level of volatility, but also the
risk surrounding future levels of volatility. The desire to price and hedge volatility
risk has lead to new advances in nancial econometrics and risk management. One
way to measure and study how volatility risk is priced in the market is to analyze
the volatility risk premium. The volatility risk premium is dened as the amount
by which some estimate of market-implied volatility exceeds (or falls short of) some
measure of realized volatility. A more technical denition is the dierence between
expected future volatility under the physical measure and the risk-neutral measure.
In general, the volatility risk premium, as dened above, is thought to be negative
which implies that the expected volatility under the pricing measure is greater than
under the physical measure. The negative volatility risk premium indicates that
investors are willing to pay a premium to hedge volatility risk; it further signies
that investors are risk averse. Along these lines, the volatility risk premium contains
a lot of information about the markets, such as expected future volatility, hedging
demand, and liquidity provision. However, we argue that much of this information is
hidden in noise if one looks at the overall changes in the volatility risk premium. To
get around this, we decompose the volatility risk premium into direction (sign) and
size (magnitude) and are able to identify the principal determinants over time. We
construct two testable hypotheses about the information content of these respective
components.
Our contributions are twofold. First, we use ultra-high-frequency data and a
unique methodology to estimate integrated volatility on the frequency domain. This
technique, which has not been widely used in the nancial econometrics literature,
creates an ecient way to eliminate bias due to microstructure noise even when the
noise may be autocorrelated. Then, we use the de-biased estimate of integrated
volatility to study the determinants of the volatility risk premium over time using
traditional risk factors along with other nancial market variables. We nd that,
1
when decomposing the volatility risk premium, the factors are able to explain almost
half of the variation in the magnitude component of the volatility risk premium over
the entire sample period. It is interesting to note that the Fama-French risk fac-
tors have no explanatory power, the TED-spread has very little (or no) explanatory
power, while credit spreads and demand for hedging tail risk are the most signicant
variables. These results have some fascinating interpretations within the framework
that we construct based on recent papers in the academic literature as well as input
from industry professionals.
The remainder of this paper is structured as follows. The next section, Section 2,
reviews the volatility risk premium, establishes the conceptual framework, and pro-
poses our hypotheses. Then, in Section 3, we discuss estimating integrated volatility
using high-frequency data. Section 4 contains our empirical analysis including data
and econometric specications. In Section 5 we present our results with discussion
about the economic interpretations and implications. Section 6 concludes. We have
two Appendices: Appendix A covers the technical details on the Fourier transform
method that we use to address microstructure noise in our estimation of integrated
volatility with the ultra-high-frequency data; Appendix B presents simulations that
demonstrate the benet of using ultra-high-frequency data in estimating integrated
volatility and the extent to which our method performs better than alternatives.
2 Volatility Risk Premium
In this section we give a precise mathematical denition for the volatility risk premium
and discuss the economic relevance of the volatility risk premium in the nancial
markets. Our work builds on a growing body of literature that has studied the
volatility risk premium and variance risk premium, all of which will be put into
context in this section. We should note that although we present all of our results in
terms of the volatility risk premium, everything is robust with respect to the variance
risk premium as well. The reason we chose to use the former is because the results
are easier and more natural to interpret in terms of vol units (as opposed to units
2
squared with the latter). There are instances where using the variance risk premium
may be preferred as we discuss below.
The volatility risk premium is the amount by which some estimate of market-
implied volatility exceeds (or falls short of) some measure of realized volatility. Math-
ematically this is the dierence between expected future volatility under the physical
measure and the risk-neutral measure:
vrp
t
=

E
P
t

T
t

2
u
du

E
Q
t

T
t

2
u
du

. (1)
To the extent that the average realized volatility (RV ) over the time t to T t
represents the expectation under the physical measure and implied volatility (IV ) at
time t looking T t into the future represents the expectation under the pricing, or
risk-neutral measure, (since it is implied by market prices) then we can say Equation
(1) is equivalent to the less formal equation:
vrp
t
= RV
t,Tt
IV
t,Tt
. (2)
Note that we use the lowercase vrp
t
to denote the volatility risk premium at time t.
If the square roots are not taken in Equation (1), then it would represent the variance
risk premium which we would denote as V RP
t
. Evidence that the two may be used
interchangeably is found throughout the literature. A recent example is Drechsler
and Yaron (2011) where volatility is the object of interest, but the quantity used in
the analysis is the variance risk premium. They dene the variance premium as
the dierence between VIX squared and the conditional expectation of the realized
variance. Conceptually, this denition follows the idea that, for a nancial instrument,
the risk premium is the dierence between the price of the contract and the expected
payo of the contract. When dealing with the variance risk premium, the price is
the VIX squared (which represents the expected squared volatility under the pricing
measure) and the payo is the realized variance (with expectation taken under the
physical measure). The reader may note that this is the reverse of our denition
3
given in Equations (1) and (2). The decision of how to sign the vrp (or V RP) is a
matter of personal preference and perspective. We follow Carr and Wu (2009) who
take the perspective that the negative sign reects investors willing to pay to hedge
their volatility risk. The idea of hedging downside tail risk will become a central
theme in explaining our empirical results. However, we further assert that the sign
of the vrp plays a secondary role to the magnitude of the vrp when trying to nd the
underlying economic determinants over time.
Whether to use the volatility or variance risk premium is also a matter of prefer-
ence that may reect the modelers specic needs. Some industry professionals and
nancial engineers prefer to work in terms of the V RP because of a more funda-
mental link to pricing and structuring variance swaps. Demeter, Derman, Kamal,
and Zou (1999) show that variance swaps can be valued using no arbitrage principals
by creating a replicating portfolio of options with dierent strike prices. Therefore,
when working with variance swaps it is more logical to quantify the market price
of volatility risk using the variance risk premium. This is true in the Carr and Wu
(2009) paper where the authors create a synthetic variance swap in computing the
variance risk premium in their empirical analysis.
To empirically compute the vrp we must choose some proxy for the expectations
given in Equation (1). We use the daily opening value of the VIX volatility index
as the proxy for the expected future volatility of the market under the risk-neutral
measure. VIX, which will be discussed more in Section 4.2, is widely accepted as the
appropriate proxy for this quantity in the literature and in practice. There is less
consensus on how to proxy the expected future volatility under the physical measure.
We use the average ex-post realized volatility, which is similar to the approach taken
by others (see, e.g., Bollerslev, Tauchen, and Zhou (2009), Carr and Wu (2009)). This
is detailed in Section 4.2 as well.
Recent studies have been able to establish some interesting empirical properties of
the volatility/variance risk premium. Carr and Wu (2009) note that traditional risk
factors have very little explanatory power for the variance risk premium (we are able
to conrm this in our empirical analysis of the volatility risk premium). They suggest
4
that there is an independent risk factor that is responsible for driving the principally
negative variance risk premium. Furthermore, they nd evidence that the V RP is
time-varying. Bollerslev, Tauchen, and Zhou (2009) study the predictability of the
variance risk premium on stock market returns from 1990 to 2005. They nd that
there is a strong, statistically signicant positive relationship between the V RP and
quarterly future stock returns. They note that the predictive power is better than
other nancial and macroeconomic factors that are typically used in stock market
return forecasting. Bollerslev, Gibson, and Zhou (2011) examine the volatility risk
premium and its relation to several macro-nancial state variables. They nd that
the vrp exhibits signicant temporal dependencies related to the macro-nance state
variables and is also able to help predict stock market returns. Zhou (2011) stud-
ies the predictability of the variance risk premium across nancial markets through
equity returns, bond returns, and credit spreads. He observes that the VRPs pre-
dictability maximizes typically in the one to four month horizon, and the short-run
risk premium dynamics can be interpreted within a general equilibrium model which
prices stochastic economic uncertainty. The calibrated model can help explain the
equity premium puzzle and the credit spread puzzle in the short-run. However, it
remains a challenge to incorporate long-run predictability patterns of consumption
growth and asset returns found in literature.
Several studies have examined the role that jumps play. Using high-frequency
index futures data, Wu (2011) computes maximum likelihood estimators of the in-
stantaneous realized return variance. His analysis shows that index return variance
jumps frequently with time-varying intensity. Both the jump arrival rate and the ab-
solute value of the negative variance risk premium are proportional to variance rate
level. This last nding, with respect to the absolute value, will be very relevant in
developing our hypotheses below and in interpreting the results of our analysis.
Todorov (2010) analyzes the variance risk premium under a semi-parametric stochas-
tic volatility model with the inclusion of price jumps. The model parameters are
estimated by GMM with high-frequency data on the ve-minute return of S&P 500
index futures contract from 1990 to 2002. The results provide empirical evidence that
5
investors are willing to pay for protection against jumps, especially when preceded
by recent jumps, which supports the hypothesis that risk aversion is time-varying.
The main takeaways are that the volatility risk premium appears to be a priced
risk factor in the capital markets (both equity as well as credit) and investors are
willing to pay a premium to hedge their downside risk, especially when uncertainty
is high. However, our knowledge is still very limited about the determinants of the
volatility risk premium and we do not have sound empirical evidence documenting
what exactly the volatility risk premium says about the mechanics of the market for
pricing and hedging risk.
Bollerslev, Gibson, and Zhou (2011) refer to their estimate of the vrp as a risk
aversion index. It seems that many practitioners agree with this interpretation of
the volatility risk premium, and we are able to nd some evidence that supports
this point. This leads to something of a paradox: suppose we choose to dene the
volatility risk premium such that it is typically negative, thereby indicating that
market participants are willing to pay to hedge their volatility risk. Then, when the
vrp gets more negative it indicates that investors are becoming more risk averse. But
then how do we explain the occurrence of a large positive spike in the volatility risk
premium; investors becoming risk loving? As we will show, the data indicates several
instances over our ve year sample period where the vrp turns positive, most notably
during the Financial Crisis. Surely, investors did not become risk loving during the
Financial Crisis.
Our prior is that the positive spikes in the volatility risk premium reect liquid-
ity conditions in the nancial markets. Consequently, several recent papers in the
nancial economics literature have linked the vrp to liquidity, intermediation, and
hedging demand. These papers provide the conceptual underpinning that we use to
construct stylized facts and testable hypotheses about the economic meaning of the
volatility risk premium. First, the volatility risk premium represents option market
makers willingness to absorb inventories and provide liquidity (G arleanu, Pedersen,
and Poteshman (2009), Nagel (2012)). Also, investors are net buyers of index options
(G arleanu, Pedersen, and Poteshman (2009)). To the extent that investors use index
6
put options to hedge their downside tail risk, then we should be able to use option
market data to draw inferences about investors demand for hedging downside tail
risk and intermediaries willingness to meet this demand (i.e. provide liquidity). The
volatility risk premium can, therefore, naturally be interpreted as the compensation
that option market makers receive for this intermediation and liquidity provision to
meet hedging demand. Adrian and Shin (2010) nd evidence of this interpretation in
the expansion and contraction of nancial intermediaries balance sheets.
Even within this conceptual framework of intermediation and liquidity provision,
the existence of a positive vrp is still a bit puzzling. Does this mean that periods of
positive vrp indicate that sellers of volatility have to pay hedgers in order to meet their
demand? The answer is: not quite. By some accounts, traders view the sign of the
volatility risk premium as indicative of nothing more than the markets expectation
of future levels of volatility. It is then the magnitude of the volatility risk premium
that represents the actual price of volatility risk. If this price is negative, then it just
implies that the market expects realized volatility levels to decrease (RV < IV ); if
the price is positive, then it implies that the market expects realized volatility levels
to increase (RV > IV ). The magnitude of the volatility captures the extent to which
market makers are willing to absorb inventory, provide liquidity, and meet hedging
demand. When demand for hedging downside tail risk increases, market makers will
take the short side (sell volatility) but must be compensated appropriately. The price
of volatility increases and implied volatility rises relative to realized levels. When
demand for hedging downside tail risk decreases, there will be a sello of volatility
and market makers will take the other side, but only at a substantial discount. Implied
volatility falls relative to realized levels. Therefore, the magnitude captures the extent
to which market makers must be compensated to provide liquidity to the options
markets, either as a premium or discount if intermediaries are selling volatility to meet
hedging demand or buying it back in response to a reduction in hedging demand.
There is another plausible interpretation of the sign of the volatility risk premium.
Bakshi and Kapadia (2003) associate the sign (negative) of the volatility risk premium
with gains from delta-hedging. In fact, we will argue that this interpretation of the
7
sign of the vrp is consistent with and complementary to the interpretation that the
magnitude captures information about intermediation and liquidity provision in the
options markets. We formalize these hypotheses below.
In our empirical analysis, we nd that isolating the direction or sign from the
magnitude of the vrp provides a new degree of insight into how the market prices
volatility risk and the role that intermediaries play. We are able to more precisely test
specic hypotheses about the determinants and economic meaning of the volatility
risk premium with its decomposition into direction (sign) and magnitude (absolute
value).
Hypotheses:
H1A: The sign of the volatility risk premium reects the markets expectation
about the about future changes in volatility.
1
H1B: The sign of the volatility risk premium reects the delta-hedged gains or
losses for option market makers.
H2A: The magnitude of the volatility risk premium reects investors demand for
hedging tail risk.
H2B: The magnitude of the volatility risk premium reects the willingness of
option market makers to absorb inventory and provide liquidity.
H1A and H1B are the hypotheses regarding the direction (sign) of the vrp. H2
are our hypotheses regarding the magnitude (absolute value) of the vrp; H2A can be
thought of as demand-side eects and H2B can be thought of as supply-side eects.
We should note that the hypotheses are not mutually exclusive, but rather they may
be complementary (i.e. supply and demand forces working with or against each other
to determine the magnitude of the vrp at a given time).
In order to econometrically test the volatility risk premium, we must come up
with an accurate and clean measure of the actual volatility in the market. With the
1
We thank Yacine At-Sahalia for pointing out that H1A is similar to the Expectation Hypothesis
in At-Sahalia, Karaman, and Mancini (2012). Their analysis helps to understand the empirical
results, and potential shortcomings, when testing H1A.
8
growth of high-frequency nancial data and the application of continuous time nance
to the analysis of such data, the tools for estimating the integrated volatility of a price
process have become plentiful. In the next section, we give some background on the
mathematics and statistical properties of various estimators before presenting our
methodology for computing the volatility risk premium using ultra-high-frequency
data.
3 Estimating Integrated Volatility with High-Frequency
Data
Typically, we will assume that the latent true (log)price X
t
follows an Ito process
dX
t
=
t
dt +
t
dW
t
(3)
where W
t
is a standard Brownian Motion and
t
and
t
are time-varying drift and
volatility, respectively, that may or may not follow stochastic processes themselves.
However, what we observe is the the transaction price, or its logarithm, Y
t
at times
t
i
[0, T], which are related to X
t
according to
Y
t
i
= X
t
i
+
t
i
. (4)
Note that the
t
i
in Equation (4) represent microstructure noise.
We would like to use observable price data to estimate the volatility, or in
Equation (3). This exercise is of clear interest for derivative pricing, risk management,
and in the construction and testing of trading strategies. In this section we review
some of the existing methodologies for estimating volatility, with emphasis on recent
advances in the use of high-frequency-data. It is important to note the dierent
restrictions that are placed on the structure of and (in Equations (3) and (4),
respectively) as quite often these are the subtleties that set one method apart from
another.
9
First, suppose we observe regularly spaced Y
t
i
, where t
i
t
i1
= . Then, let
us dene n =
T

; i.e. n is the number of sampled data points. If

t
is modeled
parametrically as constant , and the noise distribution is assumed to be Gaussian
with mean 0 and variance a
2
, then the log-likelihood function of Y
i
= Y
t
i
Y
t
i1
is
l(
2
, a
2
) =
1
2
log det()
n
2
log(2)
1
2
Y

1
Y, (5)
where is the covariance matrix of Y and
1
can be calculated explicitly.
Choosing to maximize Equation (5) gives the Maximum Likelihood Estimator,
or MLE, of volatility. It can be shown that the MLE is consistent for both the volatil-
ity component and the noise component at rates O
p
(n
1
4
) and O
p
(n
1
2
), respectively.
Moreover, misspecication of the marginal distribution of does not have any adverse
consequences. (At-Sahalia, Mykland, and Zhang (2005))
The assumption that volatility is constant is probably not very reasonable. There
is considerable evidence of time-varying volatility which means that we have to come
up with a way to estimate the instantaneous volatility process
t
either parametrically
or nonparametrically (see, e.g., Andersen, Bollerslev, and Diebold (2004)). Quite
often we are interested in estimating the integrated volatility over a period of time.
This is done by making use of the quadratic variation, X, X
T
, of the stochastic
process described by Equation (3). The quadratic variation is
X, X
T
=

T
0

2
t
dt. (6)
We want to estimate this quantity using the observable price data. A nave estimator
would be the Realized Volatility (RV) estimator
[Y, Y ]
T
=
n

i=1
(Y
t
i+1
Y
t
i
)
2
, (7)
which is a consistent estimator in a noise-free model. However, since the observable
price process given by Equation (4) is contaminated by the microstucture noise this
RV estimator is both biased and inconsistent.
10
Statistical theory indicates that we should be able to improve the accuracy and
precision of our estimate by increasing the rate at which we sample the data; hence
the value of ultra-high-frequency data.
2
This would be the case if we could observe
X
t
directly; but microstructure noise introduces an added dimension of complexity
to the problem. In fact, assuming iid noise, the bias of the RV estimator is 2nE[
2
].
This tells us that as we increase the frequency of the price data, the eect from noise
become more overwhelming. Indeed, much of the literature on estimating integrated
volatility with high frequency data attempts to address this bias that arises from
microstructure noise.
One way to address the problem of noise when sampling at too high of a frequency
is to which is to sample sparsely and use the corresponding RV estimator. This prac-
tice, known as the subsampling approach, was rst introduced by Zhou (1996). How-
ever, even when sampling sparsely at the optimally-determined frequency, the fact
that large portions of data are discarded violates important statistical principles. Fur-
thermore, Zhang, Mykland, and At-Sahalia (2005) argue that sampling over longer
horizons merely reduces the impact of microstructure, rather than quantifying and
correcting its eect for volatility estimation.
The pre-averaging approach of Jacod, Li, Mykland, Podolskij, and Vetter (2009)
uses all or most of the data, but averages over a moving window. The averages are
then used to compute the realized volatility, which then have to be adjusted by an
additive term to eliminate bias. The result is a consistent estimator of integrated
volatility in the presence of microstructure noise. In many ways, one can think of
the pre-averaging approach to be a compromise between subsampling and using a full
data sample. Additionally, pre-averaging is an eective method of data cleaning.
One of the solutions to incorporate the full data sample is Two Scales Realized
Volatility (TSRV) as proposed by Zhang, Mykland, and At-Sahalia (2005). The
TSRV estimator is based on subsampling, averaging, and bias-correction. The n
observations from the original sample is partitioned into K subsamples of longer
2
See Appendix B where we use simulations to illustrate the benet of using ultra-high-frequency
data.
11
time intervals. Note that here it is not required to sample regularly, and asymptotic
analysis is under max
i
t
i
0 and the average size n =
n
K
as n . Dene
the averaged subsampling estimator as
[Y, Y ]
(avg)
T
=
1
K
K

k=1
[Y, Y ]
(sparse,k)
T
. (8)
This provides an improvement in the order of asymptotic bias and variance as com-
pared to [Y, Y ]
(all)
T
from Equation (7). The bias can be estimated as

E[
2
] =
1
2n
[Y, Y ]
(all)
T
and the de-biased estimator becomes:

X, X
(TSRV )
T
= [Y, Y ]
(avg)
T

n
n
[Y, Y ]
(all)
T
. (9)
The estimation of

E[
2
] is justied by its asymptotic normality (conditional on X)
and proved through the martingale central limit theorem.
3
Similar asymptotic results
can be derived for the noise term

X, X
(tsrv)
T
[X, X]
(avg)
T
and the discretization term
[X, X]
(avg)
T
X, X
T
. In particular, their convergence is conditional on X with order
O
p
(

n
K
) and stable in law with order O
p
( n
1/2
) Choosing the optimal sampling step
K = cn
2/3
yields
n
1/6
(

X, X
(tsrv)
T
X, X
T
)
L
(8c
2
(E
2
)
2
+ c
2
T)
1/2
N(0, 1) (10)
where
2
is a limiting random variable that comes from the discretization eect (equal
to
4
3

T
0

4
t
dt when t
i
= ), and the convergence is stable in law. The TSRV esti-
mator is shown to outperform the standard RV estimator empirically in the study by
At-Sahalia and Mancini (2008).
A closely related estimator is Multiple Scale Realized Volatility (MSRV), which
is a generalization of TSRV. The MSRV estimator is proposed and derived in Zhang
3
For technical details of the proof, please see the Appendix of Zhang, Mykland, and At-Sahalia
(2005).
12
(2006). It has the form of

X, X
(msrv)
T
=
M

i=1
a
i
[Y, Y ]
(K
i
)
T
+ 2

E[
2
]. (11)
The convergence rate was shown to be n
1/4
with suitably selected weights a

i
s. The
TSRV and MSRV estimators are studied further in Fan and Wang (2007) and Fan
and Wang (2008).
Returning to the parametric approach, if we mistakenly assume that volatility
is constant and noise normally distributed with variance a
2
, the quasi-log likeli-
hood function remains the same as in Equation (5). The statistical properties of
the quasi-estimator QMLE(
2
, a
2
) are derived in Xiu (2010). It is shown that the
QMLE(
2
, a
2
) is still consistent under the assumption that the volatility process is
a positive and locally bounded Ito semimartingale, and that the noise is iid and in-
dependent of X. Furthermore, assume the volatility process has mean 0, variance a
2
,
and a nite fourth moment.
Finally, another interesting approach involves working on the frequency domain
rather than the time domain. As such, it relies on the Fourier transform (see Olhede,
Sykulski, and Pavliotis (2009)). The procedure computes a consistent and unbiased
estimator of integrated volatility at ultra-high-frequencies under very general and un-
restrictive specications of the microstructure noise process. This is the methodology
that we employ in estimating integrated volatility for our study. With the frequency
domain estimation method, integrated volatility is estimated through the variance
of the Fourier transform of the increment process. Under the rationale that the
high-frequency coecients are more heavily contaminated by the noise, the de-bias
procedure is done locally at each frequency. The unknown parameters involved in the
de-bias are estimated through MLE using a Whittle likelihood function. Technical
details regarding this methodology can be found in our Appendix A.
The frequency domain method for estimating integrated volatility has several de-
sirable features, both in terms of the statistical properties and the practicality in
applying to real nancial data. Perhaps most important from a practical nance
13
point-of-view, working in the frequency domain allows for more robust characteriza-
tion of the microstructure noise process. For the most part, the methods discussed
above all assume that the noise process,
t
, is iid. However, in practice i.e. in deal-
ing with real nancial data this is an unreasonable assumption.
4
Autocorrelated
microstructure noise may be a more reasonable assumption as large disturbances this
second may be highly correlated with large disturbances last second if there is a lot
of noise in the market. This may give the impression that the market is more turbu-
lent or volatile, when in fact the persistent volatility in our observed time series, Y ,
is coming from microstructure noise. Thus, we need a clean way to strip away the
true volatility of the price process in the presence of microstructure noise at ultra-
high-frequencies. This is why we use the frequency domain estimation method in
computing integrated volatility. In the next section we discuss our data in detail: the
collection, cleaning, and processing; as well as the construction of our volatility risk
premium time series and the other nancial market and economic variables that are
used in our econometric analysis of the volatility risk premium.
4 Empirical Analysis
4.1 Data
The data used for the empirical analysis came from several dierent sources. The
empirical work was conducted on three levels: rst we computed integrated volatil-
ity using ultra-high-frequency data on SPY transaction prices; then the integrated
volatility was used together with daily data on VIX to construct our time series of
the volatility risk premium; and nally we used a battery of econometric tests to
analyze the volatility risk premium in terms of relevant nancial market factors. At
each step of the process value was added: rst in our use of the novel Fourier-based
4
While most previous approaches assume iid microstructure noise, recent work by At-Sahalia,
Mykland, and Zhang (2011) addresses the complicated issue of estimating volatility from ultra-high-
frequency data with dependent microstructure noise. Our methodology is similar in that it also
permits estimation of integrated volatility for more general classes of microstructure noise, but in a
parametric framework.
14
estimation of integrated volatility on the frequency domain; then in our construction
of the daily time series of the volatility risk premium; and nally in the analysis of the
statistical and economic signicance of the nancial market factors. We should note
here that the our results in terms of computing the integrated volatility are consistent
with other popular methods (i.e. TSRV of Zhang, Mykland, and At-Sahalia (2005)),
and simply illustrates the added benet of estimating on the frequency domain when
dealing with ultra-high-frequency data in the presence of microstructure noise. Our
time series of the volatility risk premium is also qualitatively similar to other available
times series.
5
Transaction price data for trading in the SPDR ETF (ticker SPY) was obtained
from the TAQ database within WRDS. The sample period we looked at goes from
July 2006 to June 2011. Over these ve years there were a total of 523,814,632
trades. For our integrated volatility estimation method to work best, we need as
many observations as possible.
6
Trade volume decreases considerably as we go further
back in time which is why we stop at 2006. The rst year of data (2006-2007) has
approximately one-quarter the number trades as the nal year of data (2010-2011).
Additionally, this sample period contains about the same number of observations
Pre-Crisis, Crisis, and Post-Crisis.
For data cleaning and processing purposes, we ltered the data based on the
Correction Indicator (CORR) and Sale Condition. We kept only transactions
where CORR=00; these represent regular trades that were not cancelled or corrected.
This resulted in only 0.003% of the data being removed from the sample, leaving
us with 523,796,850 trades remaining. We also eliminated any Special Condition
Trades which introduced suspicious and irregular patterns in the transaction price
5
See, for instance, the data collected and managed by Hao Zhou on his website.
6
We illustrate this principle with simulations in Appendix B. The simulations show that, under
our method, sampling at higher frequencies allows for the most precise estimation of integrated
volatility. Our method performs better than nave subsampling rules that are typically used in
high-frequency studies, and as noted earlier, has the added benet that the microstructure noise
can be autocorrelated and so we need not restrict ourselves to the case where microstructure noise
is independent over time.
15
sequences (i.e. large jumps that were immediately reversed). This resulted in 1.8% of
the data being removed from the sample leaving us with 514,270,624 trades remaining.
Since multiple trades can occur in any given second, we next introduced an aggre-
gation step in the data processing. This would allow us to have a second-by-second
time series of SPY prices. We tried two methods for aggregation: median and size-
weighted average price and did not nd signicant aberrations. Finally, we had to
include an expansion step to account for seconds where no trades were executed. To
address these instances we used piecewise constant interpolation; i.e. if there was no
trade at second t then we lled it with the last executed price t 1. This resulted in
29,461,859 second-by-second data points covering 1,259 trading days. This was the
data that was used to compute our daily time series of monthly realized volatility (on
a rolling 21 trading day basis) via the frequency domain estimation methodology.
The daily opening level of the VIX volatility index was downloaded from the
CBOE database and serves as our time series of the expected integrated volatility
under the risk-neutral, or pricing, measure.
The explanatory variables that we use in our regressions also come from multiple
sources. First, we have the traditional risk factors from the Fama-French Three
Factor Model (Fama and French (1993)); the data for the Fama-French factors are
available from Kenneth Frenchs website.
7
We also include the credit spread, also
known as the default risk premium, which is the dierence in yield on Baa-rated
and Aaa-rated corporate debt. The yields on corporate debt, by Moodys rating, are
available from the FRED database. Use of the credit spread as a risk factor in asset
pricing studies goes back to Chen, Roll, and Ross (1986) and has the interpretation
as a measure of investor risk aversion. It has subsequently been used in volatility
risk premium studies such as Zhou (2011). We also include the TED-spread capture
liquidity eects in the nancial markets and a measure of distress in the nancial
system. The TED-spread is the dierence between 3-month Eurodollar rates and 3-
month Treasury rates, both of which are also available through the FRED database.
The interpretation of the TED-spread follows from the logic that as uncertainty in the
7
http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data library.html.
16
nancial system heightens, nancial institutions charge more to each other for short-
term borrowing this is reected in Eurodollar rates; at the same time they require
better collateral, which drives up demand for Treasury Bills and pushes down their
rates (Brunnermeier (2009)). As a nal explanatory variable we include the daily open
interest for put options on SPY, which is obtained from the OptionMetrics database.
This serves as our proxy for investors demand for hedging tail risk. The logic for this
is as follows. Garleanu, Pedersen, and Poteshman (2009) nd that investors are net
buyers of index options. Put options on SPY allow investors to hedge their downside
tail risk against market crashes.
The explanatory variables are summarized in Table 1. Descriptive statistics for
all variables are given in Table 2, which will be referred to throughout the analysis
and discussion.
4.2 Time Series of the Volatility Risk Premium
In this section we discuss how we constructed our time series of the volatility risk pre-
mium over the sample period, vrp
t

2011.06
t=2006.07
, and provide some preliminary analysis.
The vrp time series is constructed from two primary sources: rst, there is the daily
value of the VIX volatility index obtained from the CBOE dataset within WRDS.
The VIX is essentially a measure of implied volatility on near-term and next-term put
and call options on the S&P500 Index and is computed using a model-free approach.
8
The VIX is our proxy for the expected volatility over the next month under the Q-
measure; this is commonly accepted in the literature (see Carr and Wu (2009), Wu
(2011)). We use the Open value (rather than the Close) so as to be consistent with
our realized volatility estimate in terms of the 21-trading-day period for which we are
looking at on any given day. The time series of the VIX Open value, V IX
t

2011.06
t=2006.07
is the rst component of our vrp time series.
8
For details on the the methodology used in constructing the VIX volatility index please see
CBOE (2009). A similar methodology is employed in Jiang and Tian (2005) where the information
content of model-free implied volatility is studied. The CBOE volatility index is studied in Carr and
Wu (2006) and Jiang and Tian (2007).
17
The second component is computed using the ultra-high-frequency data on the
SPDR S&P500 ETF (ticker SPY). Using the second-by-second cleaned transaction
price data on SPY, we apply the frequency domain methodology, described in Section
3 and detailed in Appendix A, to estimate the integrated volatility for each trading
day. This gives us a daily time series of the realized volatility, but what we need
is some proxy of for the expected value of this realized volatility under the physical
measure for the next 21 trading days. Similar to Carr and Wu (2009) who use the ex-
post realized volatility on S&P500 futures, we use the average of the ex-post forward
realized volatilities on SPY for the 21-trading-day period coinciding with the same
interval covered by todays VIX level. Dene this as

RV
t

2011.06
t=2006.07
=
1
21
21

i=1
RV
i
where
RV
i
represents the realized volatility for day i computed using the frequency domain
methodology. Note that on day i, V IX
i
is supposed to cover the same 21-trading-day
period.
The last step is to subtract the rst time series, V IX
t

2011.06
t=2006.07
, from the second
time series,

RV
t

2011.06
t=2006.07
which gives
vrp
t
= RV
t
V IX
t
. (12)
A time series of our computed vrp over the sample period is plotted in Figure
1. Looking at the Figure, two things stand out immediately: rst, the risk premium
is negative throughout most of the sample period; second, there are a few pockets
where the vrp goes positive most notably in the third quarter of 2008. That large
positive spike which extended for a period of more than two months seemed to be
anomalous to what most of the literature says. We were having trouble explaining this
positive spike, both statistically and economically. To the extent that the negative
vrp represents investors being risk averse, does a positive vrp mean that investors
went from risk averse to risk loving during this time? That certainly does not seem
right, since that period includes the failure of Lehman Brothers and the plunging of
the global economy into the worst nancial crisis in history. So perhaps it means that
investors who typically pay to hedge volatility risk were no longer willing to, but rather
required compensation (i.e. be paid) to enter into any volatility related transaction?
18
A similar story was told about the negative yields on T-Bills during the Fall of 2008,
but it doesnt seem to t with what is going on in our data. So perhaps it is a
ctitious by-product of the frequency domain estimation methodology? Fortunately,
that was not the case either as we were able to conrm positive vrp including the
third quarter of 2008 using other methods for computing realized volatility, including
the TSRV estimator of Zhang, Mykland, and At-Sahalia (2005).
We actually examine this phenomenon extensively using various econometric tech-
niques.
4.3 Econometric Tests
4.3.1 Preliminary Regressions
As a rst step, we would like to examine how the volatility risk premium is related
to VIX during the Financial Crisis. We run the simple linear regression
vrp
t
=
0
+
1
V IX
t
+
t
. (13)
The results are given in Table 3. VIX is highly signicant and negative. Since a
normal volatility risk premium is negative, this implies that as VIX increases the
market price for volatility risk goes up. However, this does not tell us anything about
the underlying economics of what is happening. So, next, we add the traditional risk
factors from the Fama-French Three Factor Model (Fama and French (1993)).
9
The
Fama-French regression is
vrp
t
=
0
+
1
(r
m
r
f
)
t
+
2
SMB
t
+
2
HML
t
+
t
. (14)
The results for the Fama-French regression can be found in Table 4. Note that
although the market risk premium is highly signicant (HML is also signicant at the
9
Data for the Fama-French factors are available from Kenneth Frenchs website:
http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data library.html.
19
< 10% level), the R-squared is very small (less than 0.03) indicating that traditional
risk factors have very little explanatory power for the volatility risk premium. This
is consistent with previous ndings in other studies.
Then we introduce additional risk factors that may have theoretical links to the
volatility risk premium as indicated by the literature (e.g. G arleanu, Pedersen, and
Poteshman (2009), Nagel (2012), etc.). We want to be able to capture the demand
for hedging tail risk as well as liquidity and stability of the nancial system. Our
proxy for the demand for hedging tail risk is the open interest on S&P500 index
options (SPX). The TED spread is viewed by many as a proxy for liquidity risk and
a measure of distress in the nancial sector (see Brunnermeier (2009)). The credit
spread, or the default risk premium, captures elements of both potential theoretical
determinants. This next regression is specied as
vrp
t
=
0
+
1
POI
t
+
2
(r
m
r
f
)
t
+
3
SMB
t
+
4
HML
t
+
6
CS
t
+
7
TED
t
+
t
. (15)
The results for this regression, with Newey-West corrected standard errors, for
the whole sample period are reported in Table 5. Here we see that the inclusion
of the additional factors credit spread, TED spread, and put option open interest
improve the explanatory power substantially as the R-squared is over 40%. The
market risk premium remains signicant at the < 1% level; the credit spread and
put option open interest are also signicant at the < 1% level indicating that the
default risk premium, a widely-used proxy for risk aversion, in addition to our proxy
for hedging demand, are important factors in explaining the volatility risk premium
over time. Note that the TED spread, our proxy for liquidity risk, is not signicant.
Because of the large positive spike in the volatility risk premium that occurs right
in the middle of the nancial crisis, it is worth looking at the results for the regression
in Equation (15) for the Crisis subperiod. These results are reported in Table 6,
also with Newey-West corrected standard errors. We can see that the explanatory
power is even better during the crisis, with the R-squared at almost 50%. While
the the market risk premium and the credit spread are still signicant at the 0.1%
20
level, put option open interest is no longer signicant. Also, the TED spread appears
as minimally signicant (at the 10% level), indicating that perhaps there is a minor
liquidity eect that can help explain the positive spike in the volatility risk premium.
However, note that the HML factor is also signicant at the 10% level and there is no
clear way to interpret this within the economic framework that we have established.
The next step is to separate the volatility risk premium into the magnitude and di-
rection components to test our hypotheses H1 and H2 presented in Section 2. Options
traders believe that the sign of the volatility risk premium has very little information
about economic fundamentals or risk factors, but rather just reects the markets
expectations about future movements in volatility (hypothesis H1A). Traders believe
that the actual information content is rather reected in the magnitude of the volatil-
ity risk premium. The idea is that when the volatility risk premium is negative then
there are one or more fundamental risk factors that will push it further into negative
territory, and these same risk factors explain large positive spikes in the volatility
risk premium when it is positive. Therefore, the direction or sign of the vrp i.e.
whether it is positive or negative has no indication of traders willingness to make
a market or of the price they will charge to provide liquidity. An options dealer will
always make a market for the right price; this price, however, should be reective
of the dealers risk tolerance. We posit that this price for supplying liquidity to the
options market is the magnitude of the volatility risk premium, as an absolute value
(H2). It is possible that the sign of the volatility risk premium does reect the gains
or losses on dealers delta hedge as per Bakshi and Kapadia (2003) rather that the
practitioner view about expectations about volatility (alternative hypothesis H1B),
but regardless, disentangling the eects still requires that we decompose the time
series of vrp into direction and magnitude to empirically test. We start with the
magnitude regressions, since they are more straightforward that tests on the sign of
the volatility risk premium.
21
4.3.2 Magnitude Regressions
Our next econometric specication is to use the explanatory variables from Equation
(15), but now regressing the magnitude component of the volatility risk premium on
them to see what additional insights might be obtained within the context of our
hypothesis H2. The magnitude regression is specied as
[vrp
t
[ =
0
+
1
POI
t
+
2
(r
m
r
f
)
t
+
3
SMB
t
+
4
HML
t
+
6
CS
t
+
7
TED
t
+
t
. (16)
The results of the magnitude regression, with robust Newey-West standard er-
rors, for the whole sample period are shown in Table 7. Comparing Table 7 with
Table 5 i.e. the results of the magnitude regression specied by Equation (16)
with the results of the original vrp regression specied by Equation (15) reveals a
few interesting features. First of all, by regressing the absolute value of vrp on the
explanatory variables, the R-squared goes from just under 41% to over 49%. While
both of these statistics are very impressive in nancial econometrics (without any
lagged y-variables as regressors), the latter has the nice interpretation that over our
entire sample period, we nd that the factors are able to explain almost half of the
variation in the magnitude of the volatility risk premium. Furthermore, since the
most signicant variables are still the credit spread and the put option open interest,
we may conclude that after controlling for the traditional risk factors credit spreads
and put option open interest can be attributed the explanatory power; note that the
market risk premium signicance was slightly reduced.
We the performed the magnitude regression in Equation (16) for three subperiods
Pre-Crisis, Crisis, and Post-Crisis to see if we can identify any patterns
that might coincide with the dramatic changes in nancial markets as a result of
the Financial Crisis of 2007-2009. It is very hard to assign start and end points
to nancial crises, since they are not as clearly dened as business cycles. To best
address this, we use the ocial NBER recession dates which puts our Crisis period
from December 2007 to June 2009. The results for the magnitude regressions for the
22
three subperiods are reported in Tables 8, 9, and 10, for Pre-Crisis, Crisis, and
Post-Crisis, respectively.
An initial comparison of the results in Tables 8, 9, and 10 with those in Table 7
reveals that two preliminary observations. First, the variables seem to do a better job
explaining the volatility risk premium over the entire sample period rather than any
of the subperiods, as indicated by the R-squared being lower in any of the subperiod
regressions when compared to the whole sample period (although, the R-squared of
45.53% for the Crisis subperiod is still quite impressive). Second, there is not one
single factor that appears as consistently signicant over all three subperiods. In the
Pre-Crisis subperiod the TED spread is the sole signicant explanatory variable (at
the < 1% level). In the Crisis subperiod, the credit spread is highly signicant (at
the < 1% level) and the market risk premium has minimal signicance (10% level). In
the Post-Crisis subperiod, credit spread is still highly signicant (< 1% level) and
put option open interest, our proxy for demand for hedging tail risk, is signicant at
the < 5% level. It is possible that there was a structural change in the economic forces
underlying the volatility risk premium, but actually when examined together the
results support an interesting story that supports our theoretical framework that the
volatility risk premium (specically the magnitude component) represents the price
that options dealers require to provide liquidity and investors pay to hedge their tail
risk. Rather, the trend seems to indicate that the increase in high-frequency data will
allow for more precise and accurate measurement of the volatility risk premium and
that going forward the results will provide even stronger support for this framework.
The magnitude regression for the Pre-Crisis subperiod (Table 8) has an R-
squared of 39.64% indicating that the factors are able to explain almost 40% of the
variability in the magnitude component of the volatility risk premium. The only factor
that is statistically signicant is the TED spread. While this appears to support a
liquidity-driven explanation for the volatility risk premium, there are two reasons
why we hesitate to draw such a conclusion. First, the estimated coecient on TED
is 5.1614 which says that a 100bps increase in the TED spread should result in a
516bps increase in the price of volatility risk. Such an extreme causal eect does not
23
seem to be supported by the economics. Supposing that the TED spread is indeed a
good proxy for funding liquidity risk, this implies that when liquidity becomes worse
and funding is more expensive, options market makers are going to increase the price
at which they are willing to provide liquidity to the market for hedging tail risk by
a factor of ve. Conversations with derivative traders conrm that this is not what
happens in the market not before, during, or after the Crisis. The traders did believe
that the vrp would be correlated with the TED spread but that the relationship is
not causality based. To this end, our second reason for not giving too much weight
to the Pre-Crisis results is that this is the only magnitude regression in which the
TED spread is signicant. Therefore, we believe that it is most likely reective of
spurious correlation.
During the Crisis subperiod, the R-squared of the magnitude regression implies
that the factors are able to explain more than 45% of the variation in the magnitude
component of the volatility risk premium. The most signicant explanatory variable
is the credit spread, which not only provides validation for the risk aversion inter-
pretation of the volatility risk premium, but also supports the hypothesis that the
magnitude of the volatility risk premium reects the willingness of market makers to
absorb inventory and take risk onto their balance sheet. This is actually a pivotal
point in our analysis and below (in Section 5) we will argue why this particular result
represents an important nding within the context of our proposed framework.
Next, the critical reader might wonder why we still feel this magnitude regression
is more informative than the original vrp regression for the Crisis subperiod, when
it is not overwhelmingly clear that the magnitude component captures more infor-
mation or better reects the underlying economics. After all, the R-squared for the
magnitude regression is 45.53% compared to 49.29%. First we should note that a
lower R-squared should actually be expected in the magnitude regression, since the
variability of variability of the absolute value of vrp is less than the variability of the
actual vrp, especially during the Crisis subperiod where there is a large swing from
from positive to negative in late 2008. Since the total variability of the regressand is
the denominatator of the R-squared, then it should be lower for the magnitude regres-
24
sion. In that case, the the explanatory power R-squared of 45.53% can be considered
as good as 49.29% when comparing the two regressions. Also, both regressions have
the credit spread as the most signicant variable and the p-value associated with
the magnitude regression is smaller than that on the original vrp regression for the
Crisis subperiod. The coecient of CS in the magnitude regression is positive,
but negative in the original regression. However, the dierence in signs is perfectly
explainable. Since vrp can be negative the coecient estimate of -8.6987 indicates
that in normal times when the negative vrp indicates what investors are willing to
pay to hedge their volatility risk, a 100bps widening of the credit spread results in an
increase of approximately 870bps in the cost of hedging such risk (i.e. more negative).
This is consistent with, although more extreme than, the coecient estimate of 4.922
in the magnitude regression. Since, under hypothesis H2, the it is the size in absolute
value terms of the vrp that represents this price, the interpretation is that a 100bps
widening of the credit spread results in an increase of approximately 492bps in this
price. While this is almost half of the marginal eect of credit spreads on the vrp in
the original regression, it is still consistent with the intuition; an intuition for which
we are able to justify within the context of liquidity provision and intermediation in
the options markets (see Section 5). However, this intuition falls apart when we con-
sider the portion of the Crisis subperiod where vrp is positive (e.g. December 2007
through January 2008, and August 2008 through October 2008). The coecient esti-
mate of -8.6987 basically says that a 100bps widening of the credit spread results in a
decrease in the positive vrp of almost 870bps. This interpretation is not economically
justied, especially within the context of the framework we propose. In summary,
the fact that the explanatory power of the factors in the Crisis subperiod is almost
as strong combined with the better interpretation of the coecient estimates leads
us to believe the magnitude regressions still do a better job overall at illustrating the
underlying economics of the volatility risk premium. It is clear that for the entire
sample period the magnitude regression performs better than the regression of the
vrp time series, and it better supports our hypotheses H2, with the largest R-squared
and the highest signicance for some of the most relevant variables of interest.
25
Lastly, there is the Post-Crisis subperiod. While the R-squared is relatively low
(25.85%), it is still fairly good in a nancial econometrics study without any lagged
y-variables as regressors. One explanation for the R-squared being lower is that there
is less variability in the vrp during this subperiod. Note that both the credit spread
and the put option open interest are highly signicant. The economic interpretation
of the coecients is consistent with the ndings from the whole sample period and
supports our framework as well as H2. These are discussed in much greater detail
below in Section 5, where we provide a comprehensive synthesis of the ndings within
the context of the economic framework and hypotheses that we established. Next we
examine the direction or sign of the volatility risk premium.
4.3.3 Sign Tests
While a negative volatility risk premium, as dened by us in Equation (1), is jus-
tied economically and for the most part supported by the data, the large positive
spike in vrp in the middle of the Financial Crisis as well as several other positive
spikes throughout the entire sample period provides a paradox. To help reconcile
this paradox within a accepted economic framework, we decompose the volatility risk
premium into direction (or sign) and magnitude (absolute value). The latter was
tested econometrically in the last subsection. It is much less straightforward to test
the hypotheses on the sign of the vrp. Recall, H1A echos the view of some derivative
traders that the direction, or sign, of volatility risk premium reects the markets ex-
pectation of future changes in volatility. When the vrp is negative, then the average
or expected realized volatility in the equity market is less than the implied volatility
extracted from option prices. Volatility is priced higher in the forward-looking op-
tions market indicating that market participants expect realized volatility to increase
in the future. When the vrp is positive, then the average or expected realized volatil-
ity in the equity market is greater than the implied volatility extracted from option
prices. Volatility is priced lower in the forward-looking options market indicating
that market participants expect realized volatility to decrease in the future. This is
similar to the Expectation Hypothesis in At-Sahalia, Karaman, and Mancini (2012).
26
We seek to test Hypothesis H1A using a modied version of the regression pro-
posed in At-Sahalia, Karaman, and Mancini (2012) to test the Expectation Hy-
pothesis. We modify the specication for two reasons. First, if we were to use the
exact specication with our data, it would be a regression of the form: RV
t,t+21
=
+
1
V IX
t
+
2
vrp
t
+
t
. However, this would essentially be a perfect t or very close
to it, since we dene the vrp time series as the dierence between the ex-post average
realized volatility from day t over the next 21 trading days and the VIX index opening
level at time t (see Equation (12)). The second reason is that this specication does
not allow us to evaluate the information that the sign component of the volatility risk
premium contains about the future actual realized volatility. For these reasons, we
modify the regression to be specied as:
RV
t,t+21
= +
1
V IX
t
+
2
abs(vrp
t
) +
3
sgn(vrp
t
) +
t
(17)
where abs(vrp
t
) denotes the absolute value of the volatility risk premium at time
t and sgn(vrp
t
) denotes the sign of the volatility risk premium at time t, that is
sgn(vrp
t
) = 1 if vrp
t
< 0 and sgn(vrp
t
) = +1 if vrp
t
> 0.
We can then interpret the results both in terms of the theoretical predictions
from At-Sahalia, Karaman, and Mancini (2012) as well as evaluate whether or not
the direction of the volatility risk premium contains information about the future
levels of realized volatility, i.e. test H1A.
The results of the regression in Equation (17) are given in Table 11. The the-
oretical predictions of At-Sahalia, Karaman, and Mancini (2012) are as follows. If
the expected value of the vrp is zero and it is uncorrelated with the VIX level, then
= 0 and
1
= 1 which conrms their Expectation Hypothesis. An alternative is
that the vrp is stochastic but with a constant nonzero mean and still uncorrelated
with VIX, then ,= 0 but its estimate represents the expected value of the vrp; under
the Expectation Hypothesis, it should still be that
1
= 1. However, they are able
to prove, within a theoretical model, that
1
< 1 indicating that implied volatilities
tend to overpredict realized volatility and thus must be discounted. Looking at the
27
results in Table 11, we see that the intercept estimate is highly signicant (at the
< 1% level) which implies that the vrp is not zero; additionally, the estimated coe-
cient of VIX is also highly signicant (at the < 1% level) and less than one providing
evidence of overpredicting. Note that the estimated coecient of the absolute value
of the vrp is insignicant which helps validate our search for an economic meaning of
the magnitude component of the volatility risk premium separate from the direction
or sign. Under our hypotheses H2, tested in the previous subsection, we propose that
the magnitude of the vrp at time t represents a price for volatility risk and we make
no claim that it has predictive power for future levels of volatility. However, under
hypothesis H1A, we propose that the direction of the vrp, or whether it is positive
or negative, does contain information about future realized volatility. It turns out
that sgn(vrp) is signicant at the < 5% level, and so it would seem that there is
some evidence in favor of H1A. The coecient estimate is 7.48 which does have an
interesting implication when looking at the entire set of regression results. Suppose
that the vrp is negative. Then, from the regression results in Table 11, we can fore-
cast the future realized volatility of the S&P500 over the next 21 trading days will be
86% of the opening level of the VIX index less an additional 1.612 points. So, if the
opening VIX is say 20 (about the sample average in our data), then the forecast of the
realized volatility on the S&P500 would be 15.59%. Note that the negative volatil-
ity risk premium implies there should be some additional discounting of the implied
volatility given by VIX. Next, suppose the volatility risk premium is positive. Again,
from the regression results in Table 11, we can forecast the future realized volatility
of the S&P500over the next 21 trading days as 86% of the opening level of the VIX
index plus an additional 13.35 points. Assuming the opening VIX is 20 again, now
the forecast of the realized volatility on the S&P500 would be 30.55%. Note that
under the rare circumstances when the volatility risk premium is positive, then the
expected future realized volatility should be some fraction of the current level of the
VIX index plus a premium to account for an expected increase in volatility.
These results should be met with some degree of skepticism. At-Sahalia, Kara-
man, and Mancini (2012) show that as the forecasting horizon increases, the parame-
28
ter estimates of the Expectation Hypothesis regression become biased and inecient.
So, while our results do perhaps provide some evidence in in favor of hypothesis H1A,
the statistical inferences may not be sound. Furthermore, it does not say much about
the economic meaning of a positive or negative volatility risk premium. Alterna-
tively, Bakshi and Kapadia (2003) provide an explanation that is more consistent
with our motivating theme about market making and intermediation in the options
market. They show, both theoretically and with empirical evidence on index options,
that a negative volatility risk premium is representative of the underperformance of
a delta-neutral portfolio, where the trader sells calls and purchases units of the the
underlying as a hedge or sells puts and shorts units of the the underlying as a hedge.
Within the theoretical framework and empirical setup that we have established, we
are interested in market makers who provide liquidity to investors that wish to hedge
downside tail risk with put options on the market (S&P500). Therefore, the market
maker must short the underlying (e.g. SPDRs) in order to be delta-neutral. Con-
sequently, the market maker will have a gain on the delta-hedge when the S&P500
is down and a loss on the delta-hedge when the S&P500 is up. Table 12 shows the
returns on the S&P500 index when the volatility risk premium is positive (Panel A)
and negative (Panel B). We can see that for every period that the vrp is positive,
the S&P500 has negative returns. This is consistent that the less frequent instances
when the vrp is positive, traders making a market in S&P500 index put options are
making money on their delta-neutral hedge. More often, than not, when the vrp is
negative, traders making a market in S&P500 index put options are losing money on
their delta-neutral hedge as evidenced by the majority of periods that show positive
returns on the S&P500. This suggests strong evidence in favor of H1B.
5 Discussion and Interpretation of Results
In this section we provide a more in-depth analysis of the empirical results, specically
focusing on the results of the magnitude regressions and the sign tests. We seek to
combine all of the results into a comprehensive synthesis with the goal of interpreting
29
them within the context of our conceptual framework and addressing the hypotheses
developed in Section 2. We argue that the magnitude regression provides added
value and insight beyond just regressing vrp on the explanatory variables. This can
be justied from a statistical standpoint because allowing the volatility risk premium
to change sign introduces additional noise that cannot be explained by the data.
Furthermore, it is very dicult to assign economic meaning to the volatility risk
premium becoming increasingly positive, when it is supposed to represent the price
that is paid to hedge volatility risk. By viewing the sign of the volatility risk premium
as a binary event (negative or positive) that contains information distinct from how
large the volatility risk premium is, we are able to test specic hypotheses about the
motivating economic forces behind it. Then, we can separately examine the size of
the volatility risk premium in absolute value terms and interpret it as the price that
investors will pay to hedge volatility risk and market makers are willing to accept
to provide liquidity to that market; that is, understand it in terms of supply and
demand in a competitive market that prices risk.
Let us start with the hypotheses in H1 which attempt to explain the sign of the
volatility risk premium. The volatility risk premium is typically negative, but occa-
sionally does go positive. Practitioners like to think of the sign of the volatility risk
premium as conveying information about the markets expectations of future real-
ized volatility. A negative volatility risk premium indicates that volatility implied by
option prices (i.e. expectation under the risk-neutral, or pricing, measure) is higher
than the actual volatility (computed under the physical measure). If the forward-
looking options market is pricing volatility higher than the underlying market, then
it can be an indicator that volatility will rise over the period in question. Uncer-
tainty about future levels of risk tend to result in be pessimistic, which explains why
the volatility risk premium is typically negative. A positive volatility risk premium
suggests the reverse: implied volatility is below actual levels of volatility, implying
that the forward-looking options market is pricing down volatility risk thereby indi-
cating expectations that volatility will fall over the period in question. This is the
essence of hypothesis H1A which bears resemblance to the Expectation Hypothesis
30
in At-Sahalia, Karaman, and Mancini (2012). It turns out that it is very dicult
to empirically test whether or not the sign of the volatility risk premium is able to
predict future realized volatility. We ran a modied version of the regression pro-
posed in At-Sahalia, Karaman, and Mancini (2012), and we did nd that the sign
component of the vrp does appear to contain information about the future level of
realized volatility. However, in light of the potential statistical problems and the lack
of a strong economic interpretation of the results, we cannot accept H1A (nor can we
necessarily reject it).
The Bakshi and Kapadia (2003) explanation for the negative volatility risk pre-
mium reects delta-hedged gains (or losses). This provides the theoretical basis for
our hypothesis H1B. Rather than perform rigorous statistical tests of this hypothesis,
we used our sample data to carry out an experiment to test the economics behind
H1B. We looked at the returns on the S&P500 over the periods when vrp < 0 versus
the returns when vrp > 0. The results are shown in Table 12. Panel A shows that a
trader who was making a market in S&P500 index put option and delta-hedged by
shorting the SPDR ETF would have made money on the hedge during times when
the vrp was positive, and would have lost money on the hedge in all but one of the
periods where the vrp was negative. On average the results do seem to conrm H1B.
Wu (2011) shows that the absolute value of the variance risk premium is propor-
tional to the level of volatility in the market. Therefore, the variance risk premium
(and consequently the volatility risk premium) is either very negative or very positive
when volatility levels are high. This provides added justication for examining the
absolute value of the volatility risk premium to make inferences about the price of
volatility risk. We seek to do this within the context of demand for hedging tail risk
and liquidity provision to the volatility market (hypotheses H2A and H2B).
Looking at the magnitude regression results for the entire sample period, we see
that the independent variables are able to explain almost half of the variability in the
size of the volatility risk premium (R-squared of 48.99% in Table 7). The interpreta-
tion of the magnitude or size of the volatility risk premium, in absolute terms, within
our framework is the price that investors are willing to pay to hedge their downside
31
tail risk and the price that market makers are willing to accept as compensation for
providing liquidity to meet this demand. When market makers are selling volatility,
they want to hedge their exposure to the underlying, usually through delta-hedging.
When they are providing liquidity to investors that wish to hedge downside tail risk
with put options on the market (S&P500), the market maker must short the underly-
ing (e.g. SPDRs) in order to be delta-neutral. Therefore, the market maker will have
a gain on the delta-hedge when the S&P500 is down and a loss on the delta-hedge
when the S&P500 is up. Bakshi and Kapadia (2003) were the rst to hypothesize
and provide empirical evidence for this interpretation of the sign of the volatility risk
premium. Indeed we are able to nd evidence of this in the SPDR ETF data, as
discussed above. There is an observed pattern that when the S&P500 is down, the
VIX tends to be up; this is a phenomenon referred to as the leverage eect (see
Black (1976), Christie (1982)). However, this relationship does not seem to hold when
using the realized volatility; a phenomenon referred to as the leverage eect puzzle
(At-Sahalia, Fan, and Li (2013)). The fact that the relationship between the market
and delta-hedged gains/losses cannot be extended to understand the protability of
volatility positions is yet further justication for why we should separate the sign
of the vrp from its magnitude. Market makers are going to meet demand that in-
vestors have for hedging downside tail risk regardless of whether implied volatilities
are greater than or less than realized volatilities. When demand is higher, we expect
to see market makers widen spreads thereby increasing the eective price of volatility
risk the magnitude of the volatility risk premium. This is exactly what we see in
the data, where put option open interest is our proxy for hedging downside tail risk.
For the entire sample period, the estimated coecient of put option open interest
is 7.1225 10
7
and is statistically signicant at the < 1% level. The interpretation
and economic signicance of this result is that, over the entire sample period, a 1
million unit increase in hedging demand results in an increase in the volatility risk
premium of approximately 71bps. This is what can be considered demand-side eects,
holding supply constant. For the entire sample period, the data does support H2A,
both statistically and economically.
32
The other highly signicant variable in the magnitude regression for the entire
sample period is the credit spread with an estimated coecient of 4.1455. The inter-
pretation is that, for the entire sample period, holding the other factors constant, a
100bps widening of the credit spread can be expected to result in approximately a
415bps increase in the price of volatility risk. Credit spreads widening can be indica-
tive of many things, including increased risk aversion (Chen, Roll, and Ross (1986)),
which has been linked to the volatility risk premium in previous studies. However,
there is also a supply-side interpretation that is consistent with H2B and has very
strong economic signicance.
The supply-side eect is, holding demand for hedging tail risk constant, if mar-
ket makers are less willing to take on additional risk then as a result the price for
volatility risk will increase. There is evidence of this through the credit spread vari-
able. In addition to being used in the literature as a proxy for risk aversion, more
recently some papers have suggested that the credit spread represents global risk
appetite (Bekaert, Hoerova, and Scheicher (2009), Bekaert, Harvey, Lundblad, and
Siegel (2011)). Professional traders view the magnitude of the volatility risk pre-
mium as a reection of the risk tolerance of market makers, along the lines of this
risk appetite interpretation. We further believe that credit spreads, or the dier-
ence between yields on speculative and investment grade debt, captures a related
supply-side eect that is present for the entire sample period and seems to dominate
during the Crisis subperiod. When a nancial institution is concerned with the risk
on its balance sheet, one way to reduce the overall risk exposure is to de-leverage.
De-leveraging can be achieved through the right-hand-side of the balance sheet by
altering the capital structure: buying back debt, issuing equity, or both. However,
there is evidence that large nancial institutions have a preference to de-leverage
through the left-hand-side of the balance sheet: i.e. reducing its holding of risky as-
sets (Adrian and Shin (2010)). Bai and Collin-Dufresne (2011) show that the credit
spread actually picks up this eect quite well, particularly during the 2007- 2009 Fi-
nancial Crisis. They present evidence that the large nancial institutions classied as
primary dealers in the credit markets sold o their holdings of risky corporate debt
33
which would have exerted downward pressure on speculative grade bond prices and
increases in yields relative to investment grade debt. During the crisis, de-leveraging
no doubt played a role in the extreme widening of credit spreads in late 2008 into
early 2009. This can explain the dominant impact of the credit spread variable in
explaining the changes in the volatility risk premium during the Crisis subperiod in
our results, since there is a high degree of overlap in the set of institutions that serve
as primary dealers in the credit markets and those that are market makers in index
options.
10
During the Crisis subperiod, recall that the most statistically signicant
explanatory variable was the credit spread (see Table 9). The economic signicance
of the coecient estimate is that for a 100bps widening of the credit spread we would
expect a 492bps increase in the price of volatility risk. As can be seen from Figure
2, credit spreads increased dramatically at the end of 2008. While this can certainly
be viewed as an increase in risk aversion or decrease in global risk appetite, it is
also reective of the massive de-leveraging that occurred after the failure of Lehman
Brothers. As large nancial institutions reduced their holdings of speculative grade
debt they were also reluctant to take on additional risk in other markets. It is rea-
sonable to conclude that during this time dealers in index options increased the price
at which they were willing to make a market for hedging downside tail risk, thus
increasing the magnitude of the volatility risk premium during this time. Because
of the inherent leverage in option positions, it is not surprising to see a multiplier
eect to the order of 4 to 5 times that of what occurs in the credit markets. This
supply-side eect was clearly the driving force behind the increase in the magnitude
of the volatility risk premium, not only because it is the only statistically signicant
explanatory variable during the Crisis subperiod, but also because there were no
sharp increases in demand for put options on the S&P500 index as can be seen in
Figure 3 where open interest uctuated between 5,000,000 and 10,000,000 when the
range for the entire sample period is 2,000,000 to nearly 16,000,000; this can also be
seen in Table 2 by comparing the standard deviation for the put open interest variable
10
Compare the Federal Reserve Bank of New Yorks list of primary deal-
ers at http://www.newyorkfed.org/markets/pridealers current.html with the mem-
bers of the Options Clearing Corporation that are dealers of index options at
http://www.optionsclearing.com/membership/member-information/.
34
across subperiod. While it may seem curious that during the most uncertain point in
the nancial crisis investors were not aggressively trying to hedge their downside tail
risk, but there is an intuitive explanation for this: the so-called Fed put. After the
failure of Lehman Brothers and the subsequent bailout of the nancial industry, it
became clear that the federal government would provide a backstop either implicitly
or explicitly. Therefore, there was little need for investors to pay the high price to
hedge their downside tail risk. After the nancial crisis, however, demand for hedging
downside tail risk returned (see Figure 3) and, indeed, put option open interest was
highly signicant in the Post-Crisis subperiod (see Table 10). This is in addition
to the credit spread which still represents the supply-side eect.
Overall, we nd that there is evidence, albeit weak, in favor of H1A. We nd
more economic signicance supporting H1B. There is also very strong statistical and
economic relationships found in the data to support H2A and H2B.
6 Conclusion
Using ultra-high-frequency transaction level data on the SPDR ETF, we apply a
novel method to de-bias microstructure noise to estimate the integrated volatility
of the S&P500 on the frequency domain. We then use this estimate along with
the VIX volatility index to construct a time series of the volatility risk premium
which quanties the dierence between expected volatility under the physical and
risk-neutral measures. This is widely understood to represent the price of volatility
risk in the market and many papers have proposed various economic interpretations;
however, little work has been done to empirically support these interpretations in
a unied way. We use our time series of the volatility risk premium to conduct a
detailed analysis of its determinants. We nd that by decomposing the volatility risk
premium into sign and magnitude, we are able to better understand the determinants
within a unifying framework of intermediation in the market for hedging tail risk.
Relying on previous studies to guide our research methodology, we are able to syn-
thesize a conceptual framework for what the sign and the magnitude of the volatility
35
risk premium says about the market for hedging tail risk. Based on various litera-
tures, we propose a set of hypotheses about the meaning of the sign (H1) and the
magnitude (H2) of the volatility risk premium. Specically, H1A links the sign of the
volatility risk premium to the markets expectation about future levels of volatility.
This is a view held by practitioners and while we are able to nd some evidence in
favor of this hypothesis in the data, statistical issues raise doubt on the validity of
the inferences and there is no clear economic interpretation within the conceptual
framework we established. H1B links the sign of the volatility risk premium to the
gains and losses on traders delta-hedged positions when making a market for index
options. Bakshi and Kapadia (2003) were the rst to propose this interpretation,
and we are able to nd evidence supporting it in the market for index put options
which complements our hypotheses in H2. H2A proposes that the magnitude of the
volatility risk premium reects investors demand for hedging tail risk. Using S&P500
index put option open interest as our proxy for the demand for hedging downside tail
risk, we nd very strong evidence both statistically and economically in favor for
this interpretation. However, this demand-side interpretation is not the only mar-
ket force at work. Some recent papers (G arleanu, Pedersen, and Poteshman (2009),
Nagel (2012)) helped to motivate the nal hypothesis, H2B, which takes a supply-side
perspective. H2B proposes that the magnitude of the volatility risk premium reects
the willingness of option dealers to absorb inventory and provide liquidity to this mar-
ket. There is also strong evidence both statistically and economically supporting
this hypothesis, especially during the Crisis subperiod. The credit spread helps to
explain much of these supply-side forces in the market for volatility risk and hedging
downside tail risk, which is consistent with our conceptual framework.
When examined together the results depict an interesting story that supports our
theoretical framework which proposes that the volatility risk premium (specically
the magnitude component) represents the price that options dealers require to provide
liquidity and investors pay to hedge their tail risk. This holds for the entire sample
period, as well as the Crisis and Post-Crisis subperiods. The trend seems to
indicate that the increase in high-frequency data will allow for more precise and
36
accurate measurement of the volatility risk premium and that going forward the
results will provide even stronger support for this framework based on intermediation
in the market for volatility and volatility risk.
37
Appendix
A The Fourier Transform Method
A.1 Fourier Domain Representation
To simplify the notion, the drift will not be considered, as it only accounts for a term
of order log(t)

t. First dene the discrete Fourier transform of the increment

process U
t
j
= U
t
j+1
U
t
j
of a sample from a generic time series U
t
j
, j = 1, , N,
J
(U)
k
=

1
N
N

j=1
U
t
j
e
2it
j
f
k
, f
k
=
k
T
, U = X, Y, . (18)
The rst and second order structures of J
(X)
k

k
are as follows:
EJ
(X)
k
= 0,

(X)
k
1
,k
2
CovJ
(X)
k
1
, J
(X)
k
2

(19)
In particular,

(X)
k,k
=
1
N

T
0
E
2
s
ds (20)
A.1.1 De-bias
Now comes the de-bias procedure. By calculating the rst and second order structures
of J
(X)
k

k
, an oracle shrinkage estimator (in which
(X)
k,k
is unknown) would be
L
k
=

(X)
k,k

(X)
k,k
+ a
2
[2 sin(f
k
t)[
2
)
, (21)
38
which yields an estimator

X, X
(L
k
)
T
=

N1
k=0
L
k
[J
(Y )
k
[
2
. To estimate L
k
, the Whittle
log-likelihood is proposed, such that

L
k
=

2
X

2
X
+ a
2
[2 sin(f
k
t)[
2
, (22)
where (
2
X
, a
2
) comes from maximizing the modied Whittle log-likelihood
l(
2
X
, a
2
) =
N/21

k=1
log(
2
X
+a
2
[2 sin(f
k
t)[
2
)
N/21

k=1
[J
Y
k
[
2

2
X
+ a
2
[2 sin(f
k
t)[
2
, (23)
which gives a nal estimator

X, X
(

L
k
)
T
=
N1

k=0

L
k
[J
(Y )
k
[
2
. (24)
A.1.2 Properties
It was proved that (rst is asymptotically unbiasedness, and second is consistency)
E

X, X
(

L
k
)
T
= E

T
0

2
t
dt + O(t
1/4
), (25)

X, X
(

L
k
)
T
=

T
0

2
t
dt + O
p
(t
1/4
). (26)
39
B Simulations
We simulate data using a Heston (1993) model (following the simulation in Olhede,
Sykulski, and Pavliotis (2009)), and compare the performance of Fourier method
and naive subsampling at dierent sampling frequency. The Heston (1993) model is
specied as:
dX
t
= (
t
/2)dt +
t
dB
t
,
d
t
= (
t
)dt +
1/2
t
dW
t
,
where
t
=
2
t
. The parameters are set as follows: = 0.05, = 5, = 0.04, = 0.5,
and the correlation between the two Brownian motions B
t
and W
t
is = 0.5.
11
The
initial values are X
0
= 0 and
0
= 0.04. We take T as one day, and simulate data
with
t
= 0.1s, which yields a sample path of length N = 234, 000 in one trading
day. We rst calculate the underlying true integrated volatility by a Riemann sum
approximation of the integral, i.e.:
T
N

N
i=1

2
i
=

T
0

2
t
dt. Then we add iid noise
A(0,
2

) to get the observed data Y

i
= X
i
+
i
, where we set

= 5 10
4
.
We estimate the integrated volatility using two methods, the Fourier method and
the naive subsampling, which yields < X, X >
Fourier
T
and < X, X >
subsampling
T
. We
calculate the RMSE (root-mean-square error) of the estimates to the truth over 200
simulated sample paths. To further illustrate the eect of high frequency data, we
evaluate two methods from
t
= 1s up to
t
= 600s. Below is a gure showing the
RMSE of the Fourier method and the naive subsampling against decreasing sampling
frequencies.
The takeaway of this gure is two folds. First, the Fourier method can eectively
lter the microstructure noise, and works better than naive sampling method (we
didnt implement other more sophisticated methods for comparison, as the simulation
is not to illustrate Fourier method is superior, but rather to justify the use of high
frequency data). Second, if we are able to lter the microstructure noise, higher
11
These are the same as those used in the Olhede, Sykulski, and Pavliotis (2009) simulations.
40
frequency gives us a better estimate as we are able to utilize more data (hence more
information).
41
References
Adrian, T., and H. S. Shin, 2010, Liquidity and leverage, Journal of Financial
Intermediation, 19(3), 418 437.
At-Sahalia, Y., J. Fan, and Y. Li, 2013, The leverage eect puzzle: Disentangling
sources of bias at high frequency, Journal of Financial Economics, forthcoming.
At-Sahalia, Y., M. Karaman, and L. Mancini, 2012, The Term Structure of Vari-
ance Swaps, Risk Premia and the Expectation Hypothesis, Available at SSRN:
http://ssrn.com/abstract=2136820.
At-Sahalia, Y., and L. Mancini, 2008, Out of sample forecasts of quadratic varia-
tion, Journal of Econometrics, 147(1), 17 33.
At-Sahalia, Y., P. A. Mykland, and L. Zhang, 2005, How Often to Sample a
Continuous-Time Process in the Presence of Market Microstructure Noise, Re-
view of Financial Studies, 18(2), 351416.
, 2011, Ultra high frequency volatility estimation with dependent microstruc-
ture noise, Journal of Econometrics, 160(1), 160 175.
Andersen, T. G., T. Bollerslev, and F. X. Diebold, 2004, Parametric and Non-
parametric Volatility Measurement, in in L.P. Hansen and Y. Ait-Sahalia (eds.),
Handbook of Financial Econometrics. North-Holland.
Bai, J., and P. Collin-Dufresne, 2011, The Determinants of the CDS-
Bond Basis During the Financial Crisis of 2007-2009, Available at SSRN:
http://ssrn.com/abstract=2024531.
Bakshi, G., and N. Kapadia, 2003, Delta-Hedged Gains and the Negative Market
Volatility Risk Premium, Review of Financial Studies, 16(2), 527566.
Bekaert, G., C. R. Harvey, C. T. Lundblad, and S. Siegel, 2011, What Segments
Equity Markets?, Review of Financial Studies, 24(12), 38413890.
42
Bekaert, G., M. Hoerova, and M. Scheicher, 2009, What Do Asset Prices Have to Say
about Risk Appetite and Uncertainty?, ECB Working Paper No. 1037., Available
at SSRN: http://ssrn.com/abstract=1358848.
Black, F., 1976, Studies of stock price volatility changes, in Proceedings of the 1976
Meetings of the American Statistical Association, Business and Economics Section,
pp. 177181, Chicago.
Bollerslev, T., M. Gibson, and H. Zhou, 2011, Dynamic estimation of volatility risk
premia and investor risk aversion from option-implied and realized volatilities,
Journal of Econometrics, 160(1), 235 245.
Bollerslev, T., G. Tauchen, and H. Zhou, 2009, Expected Stock Returns and Variance
Risk Premia, Review of Financial Studies, 22(11), 44634492.
Brunnermeier, M. K., 2009, Deciphering the Liquidity and Credit Crunch 2007-
2008, Journal of Economic Perspectives, 23(1), pp. 77100.
Carr, P., and L. Wu, 2006, A Tale of Two Indices, The Journal of Derivatives,
13(3), 1329.
, 2009, Variance Risk Premiums, Review of Financial Studies, 22(3), 1311
1341.
CBOE, 2009, The CBOE Volatility Index - VIX, White paper, Chicago Board
Options Exchange, Chicago, IL.
Chen, N.-F., R. Roll, and S. A. Ross, 1986, Economic Forces and the Stock Market,
The Journal of Business, 59(3), pp. 383403.
Christie, A. A., 1982, The stochastic behavior of common stock variances: Value,
leverage and interest rate eects, Journal of Financial Economics, 10(4), 407
432.
Demeter, K., E. Derman, M. Kamal, and J. Zou, 1999, A guide to volatility and
variance swaps, The Journal of Derivatives, 6(4), 932.
43
Drechsler, I., and A. Yaron, 2011, Whats Vol Got to Do with It, Review of Finan-
cial Studies, 24(1), 145.
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.
Fan, J., and Y. Wang, 2007, Multi-Scale Jump and Volatility Analysis for High-
Frequency Financial Data, Journal of the American Statistical Association,
102(480), 13491362.
, 2008, Spot volatility estimation for high-frequency data, Statistics and its
Interface, 1(2), 279288.
G arleanu, N., L. H. Pedersen, and A. M. Poteshman, 2009, Demand-Based Option
Pricing, Review of Financial Studies, 22(10), 42594299.
Heston, S. L., 1993, A closed-form solution for options with stochastic volatility
with applications to bond and currency options, Review of Financial Studies,
6(2), 327343.
Jacod, J., Y. Li, P. A. Mykland, M. Podolskij, and M. Vetter, 2009, Microstructure
noise in the continuous case: The pre-averaging approach, Stochastic Processes
and their Applications, 119(7), 2249 2276.
Jiang, G. J., and Y. S. Tian, 2005, The Model-Free Implied Volatility and Its Infor-
mation Content, Review of Financial Studies, 18(4), 13051342.
Jiang, G. J., and Y. S. Tian, 2007, Extracting Model-Free Volatility from Option
Prices, The Journal of Derivatives, 14(3), 3560.
Nagel, S., 2012, Evaporating Liquidity, Review of Financial Studies, 25(7), 2005
2039.
Olhede, S., A. Sykulski, and G. Pavliotis, 2009, Frequency Domain Estimation of
Integrated Volatility for Ito Processes in the Presence of Market-Microstructure
Noise, SIAM - Multiscale Modeling and Simulation, 8(2), 393427.
44
Todorov, V., 2010, Variance Risk-Premium Dynamics: The Role of Jumps, Review
of Financial Studies, 23(1), 345383.
Wu, L., 2011, Variance dynamics: Joint evidence from options and high-frequency
returns, Journal of Econometrics, 160(1), 280 287.
Xiu, D., 2010, Quasi-maximum likelihood estimation of volatility with high fre-
quency data, Journal of Econometrics, 159(1), 235 250.
Zhang, L., 2006, Ecient Estimation of Stochastic Volatility Using Noisy Observa-
tions: A Multi-Scale Approach, Bernoulli, 12(6), pp. 10191043.
Zhang, L., P. A. Mykland, and Y. At-Sahalia, 2005, A Tale of Two Time Scales:
Determining Integrated Volatility with Noisy High-Frequency Data, Journal of
the American Statistical Association, 100(472), pp. 13941411.
Zhou, B., 1996, High-Frequency Data and Volatility in Foreign-Exchange Rates,
Journal of Business Economic Statistics, 14(1), pp. 4552.
Zhou, H., 2011, Variance Risk Premia, Asset Predictability Puzzles, and Macroeco-
nomic Uncertainty, Available at SSRN: http://ssrn.com/abstract=1783624.
45
2007 2008 2009 2010 2011

3
0

2
0

1
0
0
1
0
2
0
3
0
V
o
l
a
t
i
l
i
t
y

R
i
s
k

P
r
e
m
i
u
m
Figure 1: Time Series of the Volatility Risk Premium
46
Figure 2:
Top: Time series of the absolute value of the volatility risk premium or the magnitude
component of the vrp
Bottom: Time series of the credit spread, dened as the yield on Baa-rated and Aaa-
rated corporate debt.
Note: The vertical lines separate the three subperiods.
47
"
#$%%%$%%%
&$%%%$%%%
'$%%%$%%%
($%%%$%%%
)%$%%%$%%%
)#$%%%$%%%
)&$%%%$%%%
)'$%%%$%%%
)($%%%$%%%
*
+
,
)
+
%
'

-
+
,
%
+
%
'

)
)
+
,
%
+
%
'

)
+
,
)
+
%
*

,
+
,
)
+
%
*

.
+
,
)
+
%
*

*
+
,
)
+
%
*

-
+
,
%
+
%
*

)
)
+
,
%
+
%
*

)
+
,
)
+
%
(

,
+
,
)
+
%
(

.
+
,
)
+
%
(

*
+
,
)
+
%
(

-
+
,
%
+
%
(

)
)
+
,
%
+
%
(

)
+
,
)
+
%
-

,
+
,
)
+
%
-

.
+
,
)
+
%
-

*
+
,
)
+
%
-

-
+
,
%
+
%
-

)
)
+
,
%
+
%
-

)
+
,
)
+
)
%

,
+
,
)
+
)
%

.
+
,
)
+
)
%

*
+
,
)
+
)
%

-
+
,
%
+
)
%

)
)
+
,
%
+
)
%

)
+
,
)
+
)
)

,
+
,
)
+
)
)

.
+
,
)
+
)
)

!"# %&'( )(#'*'+#
Figure 3: Open interest for put options on the S&P500 index over the sample period.
Put open interest is our proxy for investor demand for hedging downside tail risk.
48
0 100 200 300 400 500 600
0
.
0
0
0
0
.
0
0
2
0
.
0
0
4
0
.
0
0
6
0
.
0
0
8
delta t
R
M
S
E
Fourier
Subsampling
Figure 4: RMSE of Two Methods
49
Variable Name Description
r
m
r
f
(Mkt) Fama-French market risk factor; market return minus risk-free rate
SMB Fama-French size factor
HML Fama-French value factor
Credit Spread (CS) Dierence in yield on Baa-rated and Aaa-rated corporate debt
TED Spread (TED) Dierence between 3-month Eurodollar rate and 3-month Treasury rate
Put Open Interest (POI) Daily open interest for put options on SPY
Table 1: Explanatory variables
50
W
h
o
l
e
-
S
a
m
p
l
e
P
r
e
-
C
r
i
s
i
s
C
r
i
s
i
s
P
o
s
t
-
C
r
i
s
i
s
(
N
=
1
2
3
8
)
(
N
=
3
3
8
)
(
N
=
3
7
5
)
(
N
=
5
2
5
)
M
e
a
n
M
e
d
S
t
d
M
e
a
n
M
e
d
S
t
d
M
e
a
n
M
e
d
S
t
d
M
e
a
n
M
e
d
S
t
d
v
r
p
-
6
.
5
0
-
7
.
5
5
7
.
5
8
-
2
.
1
8
-
1
.
9
6
5
.
1
3
-
6
.
9
6
-
7
.
6
9
1
0
.
8
0
-
8
.
9
5
-
9
.
1
7
4
.
1
4
a
b
s
(
v
r
p
)
8
.
2
3
8
.
1
7
5
.
6
5
3
.
9
9
2
.
5
5
3
.
8
9
1
0
.
5
8
9
.
0
9
7
.
2
8
9
.
1
7
9
.
2
8
3
.
3
5
V
I
X
2
4
.
2
0
2
1
.
8
8
1
1
.
7
0
1
5
.
5
6
1
3
.
5
4
5
.
1
6
3
4
.
5
8
2
8
.
1
9
1
4
.
6
8
2
2
.
3
4
2
1
.
8
4
4
.
8
6
M
k
t
0
.
0
2
0
.
1
0
1
.
6
0
0
.
0
4
0
.
1
1
0
.
8
6
-
0
.
0
9
-
0
.
0
1
2
.
4
5
0
.
0
9
0
.
1
4
1
.
1
0
S
M
B
0
.
0
1
6
0
.
0
2
0
.
6
4
-
0
.
0
2
-
0
.
0
4
0
.
4
2
-
0
.
0
2
-
0
.
0
3
0
.
8
8
0
.
0
3
0
.
0
5
0
.
5
4
H
M
L
-
0
.
0
0
7
-
0
.
0
1
0
.
7
3
-
0
.
0
3
-
0
.
0
3
0
.
2
6
-
0
.
0
0
4
0
.
0
1
1
.
1
3
0
.
0
0
7
-
0
.
0
1
0
.
5
5
C
r
e
d
i
t
S
p
r
e
a
d
1
.
3
5
1
.
0
9
0
.
6
8
0
.
9
0
0
.
9
0
0
.
0
5
2
.
0
7
1
.
5
6
0
.
8
0
1
.
1
2
1
.
0
7
0
.
2
5
T
E
D
S
p
r
e
a
d
0
.
8
7
0
.
4
4
0
.
8
8
0
.
6
1
0
.
4
0
0
.
5
1
1
.
7
8
1
.
3
8
1
.
0
3
0
.
3
8
0
.
3
3
0
.
1
7
P
u
t
O
p
e
n
I
n
t
e
r
e
s
t
(

1
0
6
)
7
.
5
8
7
.
0
4
3
.
1
5
4
.
1
3
3
.
8
6
1
.
5
7
6
.
5
9
6
.
5
0
0
.
9
2
1
0
.
5
1
1
0
.
6
5
2
.
0
5
T
a
b
l
e
2
:
D
e
s
c
r
i
p
t
i
v
e
s
t
a
t
i
s
t
i
c
s
f
o
r
a
l
l
v
a
r
i
a
b
l
e
s
a
c
r
o
s
s
a
l
l
p
e
r
i
o
d
s
.
51
Variable Estimate Std Error t-value Pr(> [t[) Signicance
(Intercept) 1.081349 2.17043 0.4982 0.618418
VIX.open -0.313211 0.11200 -2.7966 0.005245 ***
Adjusted R-squared 0.2333
F-statistic 377.4 on 1 and 1236 DF
p-value < 2.2e 16
Table 3: Regression of vrp on contemporaneous VIX. Signicance levels are < 1%, < 5%, and < 10% for
***, **, and *, respectively.
Variable Estimate Std Error t-value Pr(> [t[) Signicance
(Intercept) -6.468910 1.306755 -4.9504 8.434e-07 ***
Mkt -0.993864 0.297025 -3.3461 0.0008443 ***
SMB -0.043943 0.362837 -0.1211 0.9036251
HML 1.064161 0.622588 1.7093 0.0876554 *
Adjusted R-squared 0.02938
F-statistic 13.48 on 3 and 1234 DF
p-value 1.166e 08
Table 4: Regression of vrp on Fama-French risk factors. Signicance levels are < 1%, < 5%, and < 10% for
***, **, and *, respectively.
52
Variable Estimate Std Error t-value Pr(> [t[) Signicance
(Intercept) 6.8384e+00 2.2206e+00 3.0796 0.0021187 ***
Mkt -8.0864e-01 1.6562e-01 -4.8824 1.186e-06 ***
SMB 2.1647e-01 2.5942e-01 0.8344 0.4041895
HML 7.0313e-01 4.4234e-01 1.5896 0.1121893
Credit Spread -7.1259e+00 2.0875e+00 -3.4136 0.0006619 ***
TED Spread 3.0345e+00 2.8259e+00 1.0738 0.2831139
Put Open Interest -8.3635e-07 1.6677e-07 -5.0149 6.083e-07 ***
Adjusted R-squared 0.4088
F-statistic 143.6 on 6 and 1231 DF
p-value < 2.2e 16
Table 5: Regression of vrp on explanatory variables: Whole Sample Period. Signicance levels are < 1%,
< 5%, and < 10% for ***, **, and *, respectively.
Variable Estimate Std Error t-value Pr(> [t[) Signicance
(Intercept) 8.7621e+00 7.8219e+00 1.1202 0.26336
Mkt -9.2040e-01 1.8273e-01 -5.0370 7.424e-07 ***
SMB -3.4024e-01 3.8535e-01 -0.8830 0.37784
HML 7.9593e-01 4.7605e-01 1.6719 0.09539 *
Credit Spread -8.6987e+00 2.0488e+00 -4.2457 2.762e-05 ***
TED Spread 4.3272e+00 2.3333e+00 1.8545 0.06446 *
Put Open Interest -8.3427e-07 1.3475e-06 -0.6191 0.53621
Adjusted R-squared 0.4929
F-statistic 61.58 on 6 and 368 DF
p-value < 2.2e 16
Table 6: Regression of vrp on explanatory variables: Crisis Subperiod. Signicance levels are < 1%, < 5%,
and < 10% for ***, **, and *, respectively.
53
Variable Estimate Std Error t-value Pr(> [t[) Signicance
(Intercept) -3.7271e+00 1.5454e+00 -2.4118 0.0160208 **
Mkt 2.4442e-01 1.1868e-01 2.0595 0.0396607 **
SMB -9.4849e-02 1.8744e-01 -0.5060 0.6129336
HML 1.3585e-01 1.9148e-01 0.7095 0.4781550
Credit Spread 4.1455e+00 1.0831e+00 3.8276 0.0001359 ***
TED Spread 1.1003e+00 1.5480e+00 0.7107 0.4773794
Put Open Interest 7.1225e-07 1.2624e-07 5.6420 2.083e-08 ***
Adjusted R-squared 0.4899
F-statistic 199 on 6 and 1231 DF
p-value < 2.2e 16
Table 7: Regression of [vrp[ on explanatory variables: Whole Sample Period. Signicance levels are < 1%,
< 5%, and < 10% for ***, **, and *, respectively.
Variable Estimate Std Error t-value Pr(> [t[) Signicance
(Intercept) 1.2395e+00 1.1019e+01 0.1125 0.9105
Mkt -1.7846e-03 1.7880e-01 -0.0100 0.9920
SMB 4.3882e-01 5.4235e-01 0.8091 0.4190
HML 6.5857e-01 8.6073e-01 0.7651 0.4447
Credit Spread -1.3464e-02 1.0063e+01 -0.0013 0.9989
TED Spread 5.1614e+00 8.4138e-01 6.1345 2.437e-09 ***
Put Open Interest -9.3716e-08 6.0789e-07 -0.1542 0.8776
Adjusted R-squared 0.3964
F-statistic 37.89 on 6 and 331 DF
p-value < 2.2e 16
Table 8: Regression of [vrp[ on explanatory variables: Pre-Crisis Subperiod. Signicance levels are < 1%,
< 5%, and < 10% for ***, **, and *, respectively.
54
Variable Estimate Std Error t-value Pr(> [t[) Signicance
(Intercept) -6.8637e+00 5.5810e+00 -1.2298 0.21954
Mkt 2.7594e-01 1.5493e-01 1.7811 0.07572 *
SMB 3.1343e-02 3.3766e-01 0.0928 0.92609
HML 1.2792e-01 3.0538e-01 0.4189 0.67555
Credit Spread 4.9220e+00 1.0948e+00 4.4959 9.294e-06 ***
TED Spread 1.4905e+00 2.0325e+00 0.7333 0.46382
Put Open Interest 7.0008e-07 9.0752e-07 0.7714 0.44095
Adjusted R-squared 0.4553
F-statistic 53.1 on 6 and 368 DF
p-value < 2.2e 16
Table 9: Regression of [vrp[ on explanatory variables: Crisis Subperiod. Signicance levels are < 1%, < 5%,
and < 10% for ***, **, and *, respectively.
Variable Estimate Std Error t-value Pr(> [t[) Signicance
(Intercept) -6.4221e+00 3.7675e+00 -1.7046 0.08886 *
Mkt 5.2905e-02 2.1812e-01 0.2425 0.80845
SMB -2.4493e-01 2.3927e-01 -1.0237 0.30647
HML 2.4199e-01 3.7528e-01 0.6448 0.51932
Credit Spread 1.0837e+01 2.7372e+00 3.9593 8.567e-05 ***
TED Spread -4.9425e+00 3.2354e+00 -1.5276 0.12722
Put Open Interest 5.1633e-07 2.0308e-07 2.5425 0.01130 **
Adjusted R-squared 0.2585
F-statistic 31.45 on 6 and 518 DF
p-value < 2.2e 16
Table 10: Regression of [vrp[ on explanatory variables: Post-Crisis Subperiod. Signicance levels are < 1%,
< 5%, and < 10% for ***, **, and *, respectively.
55
Variable Estimate Std Error t-value Pr(> [t[) Signicance
(Intercept) 5.869269 1.427331 4.1121 4.181e-05 ***
VIX 0.860373 0.068876 12.4916 < 2.2e 16 ***
abs(vrp) -0.391898 0.402971 -0.9725 0.33098
sgn(vrp) 7.481117 3.005251 2.4893 0.01293 **
Adjusted R-squared 0.8198
F-statistic 1877 on 3 and 1234 DF
p-value < 2.2e 16
Table 11: Regression to test the Expectation Hypothesis and the forecasting power of the direction (sign)
of the vrp on future realized volatility. Signicance levels are < 1%, < 5%, and < 10% for ***, **, and *,
respectively.
56
Positive vrp Dates Trading Days Avg(vrp) S&P500 Return
2006/11/15 - 2006/11/27 8 0.400 -0.0105
2007/01/11 - 2007/02/27 32 8.033 -0.0174
2007/07/06 - 2007/07/31 18 2.258 -0.0491
2007/12/21 - 2008/01/16 17 1.267 -0.0749
2008/08/15 - 2008/10/15 43 14.615 -0.301
2010/04/08 - 2010/05/04 19 4.521 -0.0108
Average S&P500 daily return when vrp is positive: -0.00477
(a) Panel A: S&P500 Returns when vrp is positive
Negative vrp Dates Trading Days Avg(vrp) S&P500 Return
2006/07/31 - 2006/11/14 76 -1.677 0.0913
2006/11/28 - 2007/01/10 29 -1.098 0.0203
2007/02/28 - 2007/07/05 89 -2.313 0.0843
2007/08/01 - 2007/12/24 102 -7.420 0.0209
2008/01/17 - 2008/08/14 146 -5.785 -0.0302
2008/10/16 - 2010/04/07 370 -11.843 0.249
2010/05/05 - 2011/06/29 291 -9.3773 0.121
Average S&P500 daily return when vrp is negative: 0.000777
(b) Panel B: S&P500 Returns when vrp is negative
Table 12: Test of Delta-Hedged Gain/Loss Hypothesis
57