Tails, Fears, and Risk Premia: Tim Bollerslev and Viktor Todorov

THE JOURNAL OF FINANCE VOL. LXVI, NO.
6 DECEMBER 2011
Tails, Fears, and Risk Premia

TIM BOLLERSLEV and VIKTOR TODOROV
ABSTRACT
We show that the compensation for rare events accounts for a large fraction of the
average equity and variance risk premia. Exploiting the special structure of the jump
tails and the pricing thereof, we identify and estimate a new Investor Fears index. The
index reveals large time-varying compensation for fears of disasters. Our empirical
investigations involve new extreme value theory approximations and high-frequency
intraday data for estimating the expected jump tails under the statistical probability
measure, and short maturity out-of-the-money options and new model-free implied
variation measures for estimating the corresponding risk-neutral expectations.
So what are policymakers to do? First and foremost, reduce uncertainty.

Do so by removing tail risks, and the perception of tail risks.
Olivier Blanchard, chief economist, IMF, The Economist, January 31,
2009.
LACK OF INVESTOR CONFIDENCE is frequently singled out as one of the main
culprits behind the massive losses in market values in the advent of the Fall
2008 financial crises. At the time, the portrayal in the financial news media of
different doomsday scenarios in which the stock market would have declined
even further were also quite commonplace. Motivated by these observations,
we provide a new theoretical framework for better understanding the way in
Bollerslev is with Duke University and Todorov is with Kellogg School of Management, Northwestern University. We thank Yacine At-Sahalia, Torben G. Andersen, Snehal Banerjee, Robert
Barro, Geert Bekaert, Jules van Binsbergen, Oleg Bondarenko, Mikhail Chernov, Rama Cont,
Dobrislav Dobrev, Darrell Duffie (Acting Editor), Eric Engstrom, Jeremy Graveline, Friedrich
Hubalek, Jean Jacod, Claudia Kluppelberg, Kai Li, Sydney Ludvigson, Lasse Pedersen, Mark
Podolskij, Petra Posadel, Walter Schachermayer, George Tauchen, and Amir Yaron, our discussants
Alexandra Dias, Xavier Gabaix, and Yuhang Xing, two anonymous referees, as well as seminar
participants at EDHEC, the Federal Reserve Board, UIC, the June 2009 CREATES-Stevanovic
Center conference in Skagen, the October 2009 SoFiE conference on Liquidity, Credit Risk and
Extreme Events in Chicago, the January 2010 conference on the Interplay between Financial and
Insurance Mathematics, Statistics and Econometrics at Wolfgang Pauli Institute, Vienna University, the 2010 Warwick Business School Conference on Derivatives, Volatility and Correlation, the
June 2010 FERM symposium in Taipei, the 2010 WFA meetings, and the 2010 NBER Fall Asset
Pricing meeting for their comments and helpful discussions related to the paper. We also thank
OptionMetrics for providing us with the options data, Lai Xu for help with extracting and organizing the data, and Nicola Fusari for help with the option pricing codes used in the Monte Carlo
simulations. The research was partly funded by NSF grant SES-0957330 to the NBER. Bollerslev
also acknowledges the support of CREATES, funded by the Danish National Research Foundation.
2165
2166
The Journal of FinanceR
which the market prices and perceives jump tail risks. Our estimates rely on
the use of actual high-frequency intraday data and short maturity out-of-themoney options. Our empirical results based on data for the S&P 500 index
spanning the period from 1990 to mid-2007 show that the market generally
incorporates the possible occurrence of rare disasters in the way it prices risky
payoffs, and the fear of such events accounts for a surprisingly large fraction
of the historically observed equity and variance risk premia. Extending our
results through the end of 2008, we document an even larger role of investor
fears during the recent financial market crises. As such, our findings implicitly
support the conjecture quoted above in that removing the perception of jump
tail risk was indeed crucial in restoring asset values at the time.
More precisely, by exploiting the special structure of the jump tails and the
pricing thereof, we identify and estimate a new Investors Fears index.1 Our
identification of investors fears is based on the distinctly different economic
roles played by the compensation for the pathwise variation in asset prices versus the compensation for the possible occurrence of large rare jump events.
These two separate sources of variation have traditionally been treated as
latent constituents of the market risk premia. Our findings suggest that compensation for the former, which is naturally associated with temporal variation
in the investment opportunity set, is rather modest, while the compensation
attributable to the latter, and the fear of rare events, is both time varying and
often quite large.2
Our empirical estimates rely on a series of new statistical procedures for
backing out the objective expectations of jump tail events and the markets pricing thereof. In particular, building on new extreme value theory (EVT) developed in Bollerslev and Todorov (2011), we use the relatively frequent medium
sized jumps in high-frequency intraday returns for reliably estimating the
expected interdaily tail events under the statistical probability measure,
thereby explicitly avoiding peso type problems in inferring the actual jump
tails.
Similarly, our estimates for the market risk-neutral expectations are based
on actual short maturity out-of-the-money options and the model-free variation
measures originally used by Carr and Wu (2003) and further developed here to
recover the complete risk-neutral tail jump density. Intuitively, while the continuous price variation and the possibility of jumps both affect short maturity
options, their relative importance varies across different strikes, which allows
us to separate the valuation of the two risks in a model-free manner.
1 Our Investor Fears index is conceptually different from standard measures of tail risks that
only depend on the actual probabilities for the occurrences of tail events but not the pricing thereof;
see, for example, the discussion in Artzner et al. (1999). Further, most traditional tail risk measures
apply to any portfolio, while the Investor Fears index pertains explicitly to the pricing of jump tail
risk in the aggregate market portfolio.
2 The relatively minor role played by temporal variation in the investment opportunity set for
explaining the risk premia suggested by our results is also consistent with the evidence in Chacko
and Viceira (2005), who find only modest gains from standard intertemporal hedging in the context
of dynamic consumption and portfolio choice.
2167
The general idea behind the paper is related to an earlier literature that
seeks to explain the observed differences in the time series behavior of options
prices and the prices of the underlying asset through the pricing of jump risk;
see, for example, Andersen, Benzoni, and Lund (2002), Bates (1996, 2000),
Broadie, Chernov, and Johannes (2009a), Eraker (2004), and Pan (2002). The
paper is also related to the work of At-Sahalia, Wang, and Yared (2001), who
compare nonparametric estimates of the risk-neutral state price densities from
time series of returns to the densities estimated directly from options prices. All
of these earlier studies rely on specific, typically affine, parametric stochastic
volatility jump diffusion models, or an assumption of no unspanned stochastic
volatility. Our approach is distinctly different in that we rely on flexible nonparametric procedures that are able to accommodate rather complex dynamic
tail dependencies as well as time-varying stochastic volatility. Our focus on the
aggregate market risk premia and the notion of investor fears implicit in the
wedge between the estimated objective and risk-neutral jump tails, along with
the richer data sources used in the estimation, also set the paper apart from
this earlier literature.
Our empirical results suggest that on average close to 5% of the equity
premium (in absolute terms) may be attributed to the compensation for rare
disaster events. Related option-based estimates for the part of the equity premium due to compensation for jump risk have previously been reported in the
literature by Broadie, Chernov, and Johannes (2009a), Eraker (2004), Pan
(2002), and Santa-Clara and Yan (2010), among others. However, as noted
above, these earlier studies are based on fairly tightly parameterized jumpdiffusion models and restrictive assumptions about the shapes and dependencies in the tails.3 Counter to these assumptions, our new nonparametric
approach reveals a series of intriguing dynamic dependencies in the tail probabilities, including time-varying jump intensities and a tendency for fatter tails
in periods of high overall volatility. At the same time, our results also suggest
that the risk premium for tail events cannot solely be explained by the level of
the volatility.
In parallel to the equity premium, defined as the difference between the
statistical and risk-neutral expectation of the aggregate market returns, the
variance risk premium is naturally defined as the difference between the statistical and risk-neutral expectation of the corresponding forward variation.4
Taking our analysis one step further, we show that on average more than half
of the historically observed variance risk premium is directly attributable to
disaster risk, or differences in the jump tails of the risk-neutral and objective
3 Anderson, Hansen, and Sargent (2003) and Broadie, Chernov, and Johannes (2009) both forcefully argue from very different perspectives that misspecified parametric models can result in
highly misleading estimates for the risk premia.
4 Several recent studies seek to explain the existence of a variance risk premium, and how the
premium correlates with the underlying returns, within the context of various equilibrium-based
pricing models; see, for example, Bakshi and Kapadia (2003), Bakshi and Madan (2006), Bollerslev,
Tauchen, and Zhou (2009), Carr and Wu (2009), Drechsler and Yaron (2011), Eraker (2008), and
Todorov (2010).
2168
distributions. Moreover, while the left and right jump tails in the statistical
distribution are close to symmetric, the contribution to the overall risk-neutral
variation coming from the left tail associated with dramatic market declines
is several times larger than that attributable to the right tail. Therefore, even
though the Chicago Board Options Exchange (CBoE) VIX volatility index formally contains compensation for different risks, that is, time-varying volatilities and jump intensities as well as fears for jump tail events, our separation
shows that a nontrivial portion of the index may be attributed to the latter
component and notions of investor fear.5
At a more general level, our results hold the promise of better understanding
the economics behind the historically large average equity and variance risk
premia observed in the data. Most previous equilibrium-based explanations
put forth in the literature effectively treat the two risk premia in isolation,
relying on different mechanisms for mapping the dynamics of the underlying
macroeconomic fundamentals into asset prices. Instead, our analysis shows
that high-frequency aggregate market and derivative prices together contain
sufficient information to jointly identify both premia in an essentially modelfree manner, and that the fear of rear events plays an important role in determining both. As such, any satisfactory equilibrium-based asset pricing model
must be able to generate large and time-varying compensations for the possible
occurrence of rare disasters.6
The plan for the rest of the paper is as follows. We begin in the next section
with a discussion of the basic setup and assumptions, along with the relevant
formal definitions of the equity and variance risk premia. Section II outlines
the procedures that we use for separating jumps and continuous price variation under the risk-neutral distribution. Section III details the methods that
we use for the comparable separation under the statistical measure. Section IV
discusses the data and our initial empirical results. Our main empirical findings related to the tail risks and the risk premia are given in Section V.
Section VI reports on various robustness checks related to the dependencies
in the tail measures and the accuracy of the asymptotic approximations underlying our results. Section VII concludes. Most of the technical proofs are
deferred to the Appendix, as are some of the details concerning our handling of the options and intraday data, as well as the econometric modeling
procedures.
5 As such, this does lend some credence to the common use of the term investors fear gauge
as a moniker for the VIX volatility index, albeit a rather imperfect proxy at that; see, for example,
Whaley (2009). Instead, the new Investor Fears index developed here cleanly isolates the fear
component.
6 The idea that rare disasters, or tail events, may help explain the equity premium and other
empirical puzzles in asset pricing finance dates back at least to Rietz (1988). While the original
work by Rietz (1988) is based on a peso type explanation and probabilities of severe events that
exceed those materialized in sample, Barro (2006) argues that, when calibrated to international
data, actual disaster events do appear frequent enough to meaningfully impact the size of the
equity premium. Further building on these ideas, Gabaix (2010) and Wachter (2010) have shown
that explicitly incorporating time-varying risks of rare disasters in otherwise standard equilibriumbased asset pricing models may help explain the apparent excess volatility of aggregate market
returns.
2169
I. Risk Premia and Jumps

The continuous-time no-arbitrage framework that underlies our empirical
investigations is very general and essentially model-free. It includes all parametric models previously analyzed and estimated in the literature as special
cases. We begin with a discussion of the basic setup and assumptions.
A. Setup and Assumptions
To lay out the notation, let Ft denote the futures price for the aggregate
market portfolio expiring at some undetermined future date, and let the corresponding logarithmic price be denoted by ft log (Ft ). The absence of arbitrage
implies that the futures price should be a semimartingale.7 Restricting this
just slightly by assuming an Ito semimartingale along with jumps of finite
variation, the dynamics of the futures price may be expressed as8

dFt
= t dt + t dWt +
dx),
(1)
(e x 1) (dt,
Ft
R
where t and t denote locally bounded but otherwise unspecified drift and instantaneous volatility processes, respectively, and Wt is a standard Brownian
motion. The last term accounts for any jumps, or discontinuities, in the
price process through the so-called compensated jump measure (dt,
dx)
(dt, dx) tP (dx) dt, where (dt, dx) is a simple counting measure for the jumps
intensity) of the jumps for
and tP (dx) dt denotes the compensator (stochastic

tP (dx) being predictable and such that R (x 2 1)tP (dx) is locally integrable.9
This is a very general specification, and, aside from the implicit weak integrability conditions required for the existence of some of the expressions discussed
below, we do not make any further assumptions about the properties of the
jumps or the form of the stochastic volatility process.10
To better understand the impact and pricing of the two separate components,
it is instructive to consider the total variation associated with the market price.
Specifically, let QV [t,T] denote the quadratic variation of the log-price process
over the [t, T] time interval,
T
T
QV[t,T ]
s2 ds +
x 2 (ds, dx).
(2)
t
7 For the connection between the semimartingale assumption for the price and the notion of no
arbitrage, see, for example, Delbaen and Schachermayer (1994).
8 The assumption of finite variation for the jumps is inconsequential and for expositional purposes only. The big jumps are always of finite activity, that is, only a finite number over any finite
time interval.
x
9 Intuitively, this renders the demeaned sum of the jumps
dx) a (local)
R (e 1)(dt,
martingale.
10 The specification in (1) does exclude semimartingale processes whose characteristics are
not absolutely continuous with respect to Lebesgue measure, for example, certain time-changed
Brownian motions where the time change is a discontinuous process.
2170
The first term on the right-hand side represents the variation attributable to
the continuous-time stochastic volatility process s , that is, the variation due
to small price moves. The second term measures the variation due to jumps,
that is, large price moves. These risks are fundamentally different and present
different challenges from a risk management perspective. While diffusive risks
can be hedged (managed) by continuously rebalancing a portfolio exposed
to those risks, the locally unpredictable nature of jumps means that such a
strategy will not work. Consequently, these two types of risks may also demand
different compensation, as manifest in the form of different contributions to the
aggregate equity and variance risk premia.
B. Equity and Variance Risk Premia
The equity risk premium represents the compensation directly associated
with the uncertainty about the future price level. The variance risk premium
refers to the compensation for the risk associated with temporal changes in the
variation of the price level.11
In particular, let Q denote the risk-neutral distribution associated with the
general dynamics in equation (1). The equity and variance risk premia are then
formally defined by12

FT Ft
FT Ft
1
EQ
,
(3)
EPt
ERPt
t
T t
Ft
Ft
and
VRPt

1 P
Et (QV[t,T ] ) EQ
t (QV[t,T ] ) ,
T t
(4)
which correspond to the expected excess return on the market and the expected payoff on a (long) variance swap on the market portfolio, respectively.
Empirically, of course, the sample equity risk premium is generally positive and
historically large, while the sample variance risk premium as defined above is
on average negative and equally large and puzzling.13
11 Investor desire to hedge against intertemporal shifts in riskiness has led to the recent advent
of many new financial instruments with their payoffs directly tied to various notions of realized
price variation. Especially prominent among these are variance swap, or forward, contracts on
future realized variances. Exotic so-called gap options explicitly designed to hedge against large
price moves over short daily time intervals have also recently been introduced on the OTC market.
12 From here on we will suppress the dependence on the horizon T for notational convenience.
Also, the discrete-time equity premium defined here should not be confused with the instantaneous
premium, or drift, defined in equation (5) below. The latter serves a particular convenient role in
explicitly separating the different types of risks, as manifest in equation (12).
13 The equity premium puzzle and the failure of standard consumption-based asset pricing
models to explain the sample ERPt is arguably one of the most studied issues in academic finance
over the past two decades; see, for example, the discussion in Campbell and Cochrane (1999)
and Bansal and Yaron (2004) and the many references therein. Representative agent equilibrium models involving time-separable utility also rule out priced volatility risk and corresponding
hedging demands, and in turn imply that VRPt = 0. Recent rational equilibrium-based pricing
2171
To disentangle the distributional features that account for these welldocumented differences in the P and Q expectations, let tQ (dx) dt denote the
compensator, or intensity, for the jumps under Q (recall our absolute continuity assumption for the P jump compensator and the fact that the probability
measures P and Q are equivalent). Also, let t denote the drift that turns Wt
t
into a Brownian motion under Q, that is, WtQ = Wt + 0 s ds.14 The drift term
for the futures price in the P distribution in equation (1) must therefore satisfy

t = t t + (e x 1) tP (dx) (e x 1) tQ (dx).
(5)
R
This general expression directly manifests the compensation that is required

by investors to accept the different types of equity pricerelated risks.
Using this result, the equity risk premium is naturally decomposed as
ERPt = ERPct + ERPdt,
where
ERPct
and
ERPdt
1
EP
T t t

T
t

R
1
EP
T t t

T
t
(6)

Fs
s s ds ,
Ft
Fs x
(e 1) sP (dx) ds
Ft

R
(7)

Fs x
Q
(e 1) s (dx) ds ,
Ft
(8)
with the two terms representing the unique contribution to the premium
coming from diffusive and jump risk, respectively.
Similarly, the variance risk premium may be expressed as
VRPt = VRPct + VRPdt,
where
VRPct
and
VRPdt

T
T
1
Q
P
2
2
=
s ds Et
s ds
,
Et
T t
t
t
(9)
(10)

T
T
1
Q
P
2 P
2 Q
=
x s (dx) ds Et
x s (dx) ds
, (11)
Et
T t
R
R
t
t
models designed to help explain the positive variance risk premium by incorporating jumps and/or
time-varying volatility-of-volatility in the consumption growth process of the representative agent
include Bollerslev, Tauchen, and Zhou (2009), Drechsler and Yaron (2011), and Eraker (2008).
14 Standard asset pricing theory implies that all futures (and forward) contracts must be
martingales under the risk-neutral Q measure; see, for example, Duffie (2001). This result is subject to a boundedness condition on the interest rate process, but such assumption can be further
relaxed; see Pozdnyakov and Steele (2004).
2172
representing the diffusive and jump risk components, respectively. As

equations (10) and (11) make clear, the presence of jumps may result in a
nonzero premium even with i.i.d. returns.15
The jump risk directly links the equity and variance risk premia. By contrast,
the price of diffusive price risk does not enter the variance risk premium.
Previous attempts in the literature to separately identify and estimate the
risk premia due to jumps have been based on explicit and somewhat restrictive
parametric models that result in different and sometimes conflicting numerical
findings; see, for example, the discussion in chapter 15 of Singleton (2006).
Meanwhile, as equations (6) and (9) make clear, the compensation for jumps
generally affects the equity and variance risk premia in distinctly different,
and statistically readily identifiable, ways.
In particular, using actual high-frequency data and short maturity options,
it is possible to nonparametrically estimate16

T
1
P
x
P
ERPt (k) =
Et
(e 1) s (dx) ds
T t
t
|x|>k

T
EQ
t
|x|>k
(e x 1) sQ (dx) ds
(12)
and

T
T
1
Q
P
2 P
2 Q
VRPt (k) =
x s (dx) ds Et
x s (dx) ds
,
Et
T t
|x|>k
|x|>k
t
t
(13)
where k > 0 refers to the threshold that separates large from small jumps.17
Comparing these estimates with the corresponding estimates for ERPt with
VRPt , we may therefore gauge the importance of fears and rare events in
explaining the equity and variance risk premia from directly observable financial data. We begin our discussion of the relevant empirical procedures with
the way in which we quantify the risk-neutral Q measures.
15 It is possible to further separate VRP c into two components: diffusive and jump-type changes
t
in the stochastic volatility. We purposely do not explore this, however, as our main focus is on
the separation of the fear of disasters from the premia attached to changes in the investment
opportunity set, and the type of changes underlying the evolution in t has no effect on that
separation. Further decomposing VRP ct would also necessitate stronger parametric assumptions
than those employed here.
16 The integrands in equations (8) and (12) obviously differ by F /F . Alternatively, we could have
s t
defined the equity risk premium in terms of logarithmic returns, in which case Fs /Ft would not
appear in equation (8). In that case, however, the basic definition would instead involve convexity
adjustment terms.
17 The choice of k and other practical implementation issues are discussed further in the
Appendix below.
2173
II. Risk-Neutral Measures

There is long history of using option prices to infer market expectations
and corresponding option implied volatilities. Traditionally, these estimates
have been based on specific assumptions about the underlying P distribution
and the pricing of systematic risk(s). However, as emphasized in the literature
more recently, the risk-neutral expectation of the total quadratic variation of
the price process in equation (2), normalized by the horizon T t,

T
T
1
1
Q
Q
Q
2
2
QVt
s ds +
x (ds, dx) ,
(14)
Et
Et
T t
T t
t
t
R
may be estimated in a completely model-free manner by a portfolio of options
with different strikes; see, for example, Bakshi, Kapadia, and Madan (2003),
Britten-Jones and Neuberger (2000), and Carr and Wu (2009).18 Meanwhile,
as the expression in equation (14) makes clear, and in line with the discussion
above, the expected quadratic variation under Q is naturally decomposed into
the expected variation associated with diffusive, or small price moves, and
discontinuous, or large price moves. Building on these insights, we show below
how option prices may be used in estimating these two separate parts of the
risk-neutral distribution in an essentially model-free fashion.19
A. Jumps and Tails
Our estimates for the jump risk perceived by investors are based on closeto-maturity deep out-of-the-money options. The use of deep out-of-the-money
options to more effectively isolate jump risk has previously been applied in
the estimation of specific parametric affine jump-diffusion stochastic volatility
models by Bakshi, Cao, and Chen (1997), Bates (1996, 2000), Pan (2002), and
Eraker (2004) among others, while Pan and Liu (2003) and Pan, Liu, and
Longstaff (2003) consider similar fully parametric approaches in the context of
hedging portfolio jump risks.20 In contrast to these studies, our estimates for
the jump risk premia are entirely nonparametric.
18 This also mirrors the way in which the CBOE now calculates the VIX volatility index for
the S&P 500. The approximation used by the CBOE formally differs from the expression in
T
x2
x
equation (14) by T1t EQ
t ( t
R 2(e 1 x 2 )(ds, dx)); for further discussion of the approximation errors involved in this calculation see the white paper on the CBOE website, as well as
Carr and Wu (2009) and Jiang and Tian (2005, 2007).
19 Previous attempts at nonparametrically estimating the risk-neutral distribution implied by
options prices include At-Sahalia and Lo (1998, 2000), Bakshi, Kapadia, and Madan (2003), and
Rosenberg and Engle (2002), among others. In particular, At-Sahalia, Wang, and Yared (2001) and
seek to quantify the amount of additional jump risk needed to make the P tails comparable to the
options-implied Q tails, by calibrating the jump frequency to match the estimated skewness (or
kurtosis) of the state price density. To the best of our knowledge, however, none of these studies
have sought to estimate the portion of the two risk premia that are directly attributable to the
jump tails and the fears thereof.
20 As previously noted, the recent exotic so-called gap options, or gap risk swaps, are explicitly
designed to hedge against gap events, or jumps, in the form of a one-day large price move in the
2174
In particular, building on the ideas in Carr and Wu (2003), who use the
different rates of decay of the diffusive and jump risks for options with various
levels of moneyness as a way to test for jumps, we rely on short-maturity
options to actually identify the compensation for jump risk. Intuitively, shortmaturity out-of-the-money options remain worthless unless a rare event, or a
big jump, occurs before expiration. As such, this allows us to infer the relevant
risk-neutral expectations through appropriately scaled option prices.
Specifically, let Ct (K) and Pt (K) denote the price of a call and a put option,
respectively, with a strike of K recorded at time t expiring at some future date,
T. It follows that for T t and K > Ft limst Fs , where limst denotes the limit
from the left,

T

+ Q
Q
r(t,T ]
x
Ct (K)
Et
1{Fs <K} Fs e K s (dx) ds,
(15)
e
R
while for K < F t ,

e
r(t,T ]
Pt (K)
EQ
t
1{Fs >K} K
+
Fs e x sQ (dx)

ds,
(16)
where x+ max{0, x}. Importantly, the expressions on the right-hand sides of

(15) and (16) involve only the jump measure. Over short time intervals, changes
in the price due to the continuous component are invariably small relative to the
possible impact of large jumps, and the diffusive part may simply be ignored.
Moreover, for T t there will be at most one large jump. The right-hand sides
of equations (15) and (16) represent the conditional expected values of the two
different option payoffs under these approximations. The payoffs equal zero if
no large jump occurred and otherwise are equal to the difference between the
price after the jump and the strike price of the option. The proof of Proposition
1 below given in Subsection A of the Appendix contains a more formal analysis
of the errors involved in the approximations and their orders of magnitude.
Going one step further, the expressions in equations (15) and (16) may be used
in the construction of the following model-free risk-neutral jump tail measures:
T
1
er(t,T ] Ct (K)
Q
(e x ek)+ EQ
,
(17)
RTtQ (k)
t (s (dx)) ds
T t t
(T t)Ft
R
and
LTtQ (k)
T t
T
t
Q

er(t,T ] Pt (K)
(ek e x )+ EQ
,
t s (dx) ds
(T t)Ft
R
(18)
where k ln( FKt ) denotes the log-moneyness. Unlike the variance replicating
portfolio used in the construction of the VIX index and the approximation to
underlying; see Cont and Tankov (2009) and Tankov (2009) for further discussion on the pricing
of these contracts using different parametric models as well as their hedging using standard
exchanged-traded out-of-the-money options.
2175
QVtQ in equation (14), which holds true for any value of T, the approximations
for RTtQ (k) and LTtQ (k) in equations (17) and (18) depend on T t.
In practice, of course, we invariably work with close-to-maturity options
and a finite horizon T t, so the effect of the diffusive volatility might be
nontrivial. In Section VI, we report representative results from an extensive Monte Carlo simulation study designed to investigate the finite sample
accuracy of the approximations across different models, option strikes K, and
time-to-maturity T t. Taken as a whole, these results suggest that the error
involved in approximating LTtQ through equation (18) is in fact quite trivial for options and empirical settings designed to mimic those of the actual
data. The estimation error for RTtQ in equation (17) tends to be somewhat
larger, albeit still fairly small in an absolute sense. Intuitively, investors primarily fear large negative price moves, making negative jumps in the riskneutral distribution more pronounced and easier to separate from the diffusive
component.
Although the RTtQ (k) and LTtQ (k) measures are related to the tail jump density, they do not directly correspond to the risk-neutral expectations required
for the calculation of the ERPt (k) and VRPt (k) jump premia. In order to get
at these and the risk-neutral tail density, we rely on an additional Extreme
Value Theory (EVT) type approximation.21 This approximation allows us to
easily extrapolate the tail behavior from the estimates of RTtQ (k) and LTtQ (k)
for a range of different values of k. The only additional assumption required
for these calculations concerns the format of the risk-neutral jump density.
Specifically, we assume that

(19)
tQ (dx) = t+ 1{x>0} + t 1{x<0} Q (x) dx,
where the unspecified
t stochastic processes allow for temporal variation in
the jump arrivals and Q (x) denotes any valid Levy density. This is a very
weak assumption on the general form of the tails, merely restricting the time
variation to be the same across different jump sizes. This assumption is trivially
satisfied by all of the previously estimated parametric jump-diffusion models
2
referred to above, most of which restrict
t 1 or t t .
The following proposition, the proof of which is deferred to Subsection A of the
T
x
Appendix, formalizes the results underlying our estimation of EQ
t ( t
|x|>k(e

T
Q
Q
Q
1)sQ (dx) ds) and Et ( t |x|>k x 2 sQ (dx) ds) from RTt (k) and LTt (k). Intuitively,
by tracking the slope of the latter for increasingly deeper out-of-the-money
options, it is possible to infer the tail behavior of tQ (dx).22
PROPOSITION 1: Assume that Ft and tQ (dx) satisfy equations (1) and (19),
Q
respectively. Denote + (x) = ex 1 and Q+ (x) = (ln(x+1))
for x > 0, and
x+1
Q
Q
Q
Q ( ln(x))
x
for x < 0. Define (x) = x (u) du, and
(x) = e and (x) =
x
21
For a general discussion of EVT, see Embrechts, Kluppelberg, and Mikosch (2001).
Our actual estimation is slightly more general than the result of Proposition 1 in the sense
that it does not restrict a scaling parameter; the details are in Appendix, Subsection A.
22
2176
pQ
assume that Q
L (x) for pQ
> 0, where L (x) satisfy L (tx)/L (x) = 1
(x) = x
+ O( (x)) as x for each t > 0, and (x) > 0, (x) 0 as x , and (x)
are nonincreasing. Then, for arbitrary t > 0, T t0, K 1 , and K 2 , with
K 2 /K 1 > 1 constant,

RTtQ ( K1 /Ft )
K2 Ft
P +
,
(20)
log
pQ 1 log
K1 Ft
RTtQ ( K2 /Ft )
while for K 1 0 and K 2 0 with K 1 /K 2 > 1 constant,

log
LTtQ ( K1 /Ft )
LTtQ ( K2 /Ft )
pQ

+ 1 log
K1
K2

.
(21)
The risk premia and the ERPt (k) and VRPt (k) measures, of course, also depend on the jump tails in the statistical distribution. We next discuss the highfrequency realized variation measures and reduced-form forecasting models
that we rely on in estimating the tails under P.
III. Objective Measures
Our estimates of the relevant expectations under the objective, or statistical, probability measure P are based on high-frequency intraday data and
corresponding realized variation measures for the diffusive and discontinuous
components of the total variation in equation (2). In addition, we rely on newly
developed EVT-based approximations and nonparametric reduced form modeling procedures for translating the ex-post measurements into forward-looking
P counterparts to the QVtQ , RTtQ , and LTtQ expectations defined above. The
foundations for the EVT approximations are formally based on the general theoretical results in Bollerslev and Todorov (2011), and we provide details on the
calculations involved in these approximations in the Appendix. We continue
the discussion here with a description of our high-frequency-based realized
variation measures.
A. Realized Jumps and Volatility
For ease of notation, we normalize the unit time interval to represent a day,
where, by convention, the day starts with the close of trading on the previous
day. The resulting daily [t, t + 1] time interval is naturally broken into the
[t, t + t ] overnight time period and the [t + t , t + 1] active part of the trading
day. The opening time of the market t + t is obviously fixed in calendar
time. Conceptually, however, it is useful to treat { t }t=1,2,... as a latent discretetime stochastic process that implicitly defines the proportion of the total day
t variation due to the overnight price change. With n equally spaced highfrequency price observations, this leaves us with a total of n 1 intraday
2177
t
increments for day t + 1, each spanning a time interval of length
n,t 1
,
n
n,t
23
say
i f ft+t +i
n,t ft+t +(i1)
n,t for i = 1, . . . , n 1.
The realized variation for day t, denoted RVt , is simply defined as the
summation of the squared high-frequency increments over the active part of
the trading day. The realized variation concept was first introduced into the
economics and finance literatures by Andersen and Bollerslev (1998),
Andersen et al. (2003), and Barndorff-Nielsen and Shephard (2001). The key
insight stems from the basic result that, for increasingly finer-sampled increments, or n , RVt consistently estimates the total ex-post variation of the
price process. Formally,
RVt

n1
n,t 2 P
i f
t+1
t+t
i=1

s2 ds +
t+1

R
t+t
x 2 (ds, dx),
(22)
where the right-hand side equals the quadratic variation of the price over the
[t + t , t + 1] time interval, as defined by the corresponding increment to the
total variation in equation (2).
As discussed above, the quadratic variation obviously consists of two distinct
components: the variation due to the continuous sample price path and the
variation coming from jumps. It is possible to separately estimate these two
components by decomposing the summation of the squared increments into
separate summations of the small and large price changes, respectively. In
particular, it follows that, for the continuous variation,

n1
n,t 2
P
n,t
i f 1{|
i f |
n,t }
CVt
i=1
t+1
t+t
s2 ds,
(23)
while, for the jump variation,

n1
n,t 2
P
i f 1{|
in,t f |>
n,t }
JVt
i=1
t+1
t+t

R
x 2 (ds, dx),
(24)
where > 0 and (0, 0.5) denote positive constants.24 These measures
for the continuous and jump variation were first implemented in the context
of high-frequency financial data by Mancini (2001). They are based on the
intuitive idea that, over short time intervals, the continuous component of the
price will be normally distributed with variance proportional to the length of
the time interval. Thus, if the high-frequency price increment exceeds a certain
threshold, it must be associated with a jump in the price.
23 Even though it is convenient to treat the sampling times (t + , t + +
, . . . , t + 1) as
t
t
n,t
stochastic, they are obviously fixed in calendar time. As discussed further in the Appendix, all we
need from a theoretical perspective is for the value of t to be (conditionally) independent of the ft
logarithmic price process.
24 The actual choice of and , together with our normalization procedure designed to account
for the strong intraday pattern in the volatility, are discussed in Subsection B of the Appendix.
2178
Building on this same idea, the JVt jump component may be further split
into the variation coming from positive, or right tail, jumps,

n1
n,t 2
P

i f 1{
in,t f >
n,t }
RJVt
i=1
t+1
t+t
x 2 1{x>0} (ds, dx),
(25)
and the variation coming from negative, or left tail, jumps,

n1
n,t 2
P

n,t

i f 1{
i f <
n,t }
LJVt
i=1
t+1
t+t

R
x 2 1{x<0} (ds, dx).
(26)
It follows, of course, by definition that RVt = CVt + RJVt + LJVt .

The different ex-post variation measures discussed above all pertain to the
actual realized price process. By contrast, the option-implied diffusive and
jump tail measures discussed in Section II provide ex-ante estimates of the
risk-neutral expectations of the relevant sources of risks. We next discuss our
translation of the actual realized risks into comparable measures of objective,
or statistical, ex-ante expectations of these risks.
B. Expected Jumps and Tails

The most challenging problem in quantifying the statistical expectations
in the equity and variance risk premia in equations (6) and (9), respectively,
concerns the P equivalent to the previously discussed RTtQ (k) and LTtQ (k) riskneutral tail measures, say RTtP (k) and LTtP (k), respectively. These measures
pertain to the truly large jumps. However, these types of tail events are invariably rare, and possibly even nonexistent, over a limited calendar time span.
This is akin to a peso type problem, and as such straightforward estimation of
the tail expectations based on sample averages generally will not be reliable.
Instead, we rely on EVT and the more frequent medium-sized jumps implicit
in the high-frequency intraday data to meaningfully extrapolate the behavior
of the large jumps, and more precisely estimate the extreme parts of the jump
tails.
Compared to conventional fully parametric procedures hitherto used in the
literature for estimating the expected impact of jumps, our approach is much
more flexible and has the distinct advantage of not relying on a tight connection
between the jumps of all different sizes. Indeed, as discussed further below, the
statistical properties of the small and large jumps differ quite dramatically, as
indirectly evidenced by the markedly different estimates for the tails obtained
across competing parametric specifications that use all of the jumps for the tail
estimation.
A completely general nonparametric approach allowing jumps of different
sizes to exhibit arbitrary temporal variation is obviously impossible. Instead,
2179
we merely assume that the temporal variation in the jump intensity is a linear
function of the stochastic volatility, as in25

tP (dx) = 0 1{x<0} + 0+ 1{x>0} + (1 1{x<0} + 1+ 1{x>0} )t2 P (x) dx,
(27)
where P (x) is a valid Levy density, and

0 and 1 are nonnegative free
parameters. This semiparametric structure does admittedly restrict the form
of the temporal dependencies somewhat. Importantly, however, the assumption
in equation (27) is only needed to pin down the dynamics of the risk premia.
All of the unconditional results reported in the paper, that is, the level of the
risk premia, continue to hold under the analogous weaker assumption for the
Q jump tails given in equation (19).26 The assumption in equation (27) still
provides a very flexible specification, and we provide direct support for its empirical validity in Section VI below. Virtually all parametric models hitherto
estimated in the literature are based on simplified versions of this tP (dx) measure, typically restricting
1 = 0. By contrast, the more general specification in
equation (27) explicitly allows for time-varying jump intensities as a function
of the latent volatility process 2t . This includes so-called self-affecting jumps
2
in which
1 > 0, and t depends on past realized jumps, as suggested by the
data. Of course, the specification also allows for the relative magnitudes of
the constant and time-varying intensities to differ across positive and negative
jumps.
Relying on (27) to describe the temporal dependencies in the jump intensity, the equivalents to the RTtQ (k) and LTtQ (k) measures under the statistical
distribution may be expressed as

T
1+ P
+
P
2
E
RTt (k) = 0 +
s ds
(e x ek)+ P (x) dx,
(28)
T t t
t
R
and

LTtP (k)
1 P
E
+
T t t

t

s2 ds
(ek e x )+ P (x) dx.
(29)
To estimate RTtP (k) and LTtP (k), and in turn QVtP , we therefore only need to
quantify the integrals with respect to the Levy density P (x) and the conditional
T
expectation of the integrated volatility EPt ( t s2 ds).
Similar to our estimation of the Q jump tails, our estimates of the integrals
with respect to P (x) are again based on the medium-sized jumps coupled with
EVT for meaningfully extrapolating the behavior of the tails. The separation
of the jump intensity into the constant and time-varying part, that is,
0 and
,
respectively,
is
effectively
achieved
through
the
correlation
between
the
1
25 This parallels the significantly weaker assumption for the risk-neutral jump intensity in
equation (19). Extensions to allow for nonlinear functions of the volatility, or other directly observable state variables, are straightforward.
26 Also, given the differences in orders of magnitude of the P and Q tails reported in the empirical
part, the practical import of assumption (27) for the actual empirical results is rather limited.
2180
occurrence of jumps and past volatility, as measured directly by the realized

continuous variation CVt in equation (23). A more detailed description of the
relevant calculations is again given in the Appendix.
Our estimates for the expected integrated volatility, and the time-varying
parts of RTtP , LTtP , and QVtP , are based on reduced-form time series modeling
procedures, as originally advocated by Andersen et al. (2003). This provides a
particular convenient and parsimonious approach for capturing the dynamic
dependencies in the integrated volatility process and corresponding forecasts.
From a statistical perspective, this is also a much easier problem than that of
directly estimating the tails, as we have plenty of reliable data and previous
empirical insights to go by.27
In particular, it is well established that integrated volatility is a stationary,
but highly persistent, process; see, for example, Bollerslev et al. (2009) and
the reduced form model estimates reported therein. It has also previously been
documented that jumps in the price tend to be accompanied by jumps in the
spot volatility t , and that negative and positive jumps impact volatility differently, that is, a leverage effect through jumps. To accommodate all of these
nonlinear dependencies in the construction of our estimates for the expected
integrated volatility, we formulate a vector autoregression (VAR) of order 22,
or 1 month, for the four-dimensional vector consisting of CVt , RJVt , LJVt , and
the squared overnight increments ( ft ft+t )2 . To keep the model tractable,
we follow the HAR-RV modeling approach of Corsi (2009) and Andersen,
Bollerslev, and Diebold (2007) and restrict all but the daily, weekly, and monthly
lags, that is lags 1, 5, and 22, respectively, to be equal to zero. The forecasts
T
for EPt ( t s2 ds) is then simply obtained by recursively iterating the VAR to the
relevant horizon. Further details concerning the model estimates and forecast
construction are given in Subsection C of the Appendix.
Before turning to a discussion of the actual empirical results related to the
P and Q tail measures and risk premia, we briefly discuss the data underlying
our estimates.
IV. Data and Preliminary Estimates
A. High-Frequency Data and Statistical Measures
Our high-frequency data for the S&P 500 futures contract come from Tick
Data Inc. and span the period from January 1990 to December 2008. The prices
are recorded at 5-minute intervals, with the first price for the day at 8:35 (CST)
and the last price at 15:15. All in all, this leaves us with a total of 81 intraday
price observations for each of the 4,751 trading days in the sample.
The resulting realized variation measures all exhibit the well-known volatility clustering effects, with a sharp increase in their levels toward the end of
27 The theoretical results in Andersen, Bollerslev, and Meddahi (2004) and Sizova (2009) also
indicate that the loss in forecast efficiency from reduced-form procedures is typically small relative
to fully efficient parametric procedures, and that the reduced-form procedures have the added
advantage of a built-in robustness to model misspecification.
2181
the sample and the Fall 2008 financial market crises. The realized variation
measures in turn form the basis for our nonparametric estimates of the statistical expectations of the total quadratic variation and the left and right tail
measures, QVtP , LTtP , and RTtP . We do not discuss the details of the estimation results here, instead referring interested readers to the Appendix and the
actual estimates reported therein. Nonetheless, it is instructive to highlight a
few of the key empirical findings.
Directly in line with the discussion in the previous section, the stochastic
volatility process is highly persistent. In addition, negative realized jumps
significantly impact future volatility. In contrast to the strong own dynamic
dependencies in the volatility, the effect of negative jumps is relatively short
lived, however. Our results also indicate that positive jumps do not affect future
volatility. These findings are consistent with the empirical results in BarndorffNielsen, Kinnebrock, and Shephard (2010) and Todorov and Tauchen (2011).
They also support a two-factor stochastic volatility model, as in Todorov (2010),
in which one of the factors is highly persistent while the other factor is short
lived and mainly triggered by (negative) jumps.
Our estimation results also point to the existence of nontrivial temporal
variation in the tail jump intensities, as manifested by the highly statistically
significant estimates for
1 . Interestingly, the right and left jump tails behave
quite similar under the statistical distribution with almost identical tail decays. Taken as a whole, however, our empirical results suggest much more
complicated dependencies than those allowed by the parametric models hitherto estimated in the literature, further underscoring the advantages of our
flexible nonparametric approach for characterizing the behavior of the P jump
tails.
B. Options Data and Risk-Neutral Measures
Our options data come from OptionMetrics. The data consist of closing bid
and ask quotes for all S&P 500 options traded on the CBOE. The data span
the period from January 1996 to the end of December 2008, for a total of
3,237 trading days. To avoid obvious inconsistencies and problems with the
price quotes, we apply similar filters to those used by Carr and Wu (2003,
2009) for cleaning the data.
The estimates for the two tail measures RTtQ (k) and LTtQ (k) are based on
the closest-to-maturity options for day t with at least 8 calendar days to expiration.28 More specifically, we first calculate BlackScholes implied volatilities
for all out-of-the-money options with moneyness k using linear interpolation of
options that bracket the targeted strike.29 From these implied volatilities we
28 The options always expire on a Saturday with the settlement at the opening of the preceding
Friday. The median time-to-expiration of the options in our sample is 14 working days.
29 These prices are taken directly from OptionMetrics and are based on the corresponding mid
quotes. In the absence of any deeper out-of-the-money option than the desired moneyness k, we
use the implied volatility for the deepest out-of-the-money option that is available as a proxy.
2182
Figure 1. Implied measures. The estimates are based on S&P 500 options spanning the period
from January 1996 to December 2008, for a total of 3,237 trading days. The log-moneyness of the
options used for the left and right tail measures is fixed at k = 0.9 and k = 1.1, respectively.
then compute an out-of-the money option price with moneyness k, and in turn
the RTtQ (k) and LTtQ (k) measures as described in Section II above.
Our calculation of the implied total variance measure QVtQ follows Carr and
Wu (2009) and is similar to the approach used by the CBOE in its calculation
of the VIX index. There are two differences, however, between our construction
of QVtQ and the calculation of the VIX index, as described in the white paper
available on the CBOE website. First, we rely on trading, or business, time,
while the VIX index is constructed using a calendar-time convention. Second,
for compatibility with the tail measures RTtQ (k) and LTtQ (k), we use the same
closest-to-maturity options in the calculation of QVtQ , whereas the VIX is based
on linear interpolation of options that bracket a 1-month maturity.
The three resulting risk-neutral variation measures are plotted in Figure 1.
All of the measures are reported in annualized percentage form. Guided by
the results from the Monte Carlo simulation study designed to investigate the
accuracy of the approximations in equations (17) and (18) and discussed in
more detail in Section VI below, here and throughout the rest of the analysis,
the moneyness for the left and right tails is fixed at 0.9 and 1.1, respectively. As
2183
Table I
Mean Jump Intensities

The jump intensities are reported in annualized units. The jump sizes are percentage changes in
the price level. The estimates under Q are based on S&P 500 options data from January 1996
through June 2007, while the estimates under P rely on high-frequency 5-minute S&P futures
prices from January 1990 through June 2007. Standard errors for the estimates are reported in
parentheses.
Jump Size
>7.5%
>10%
>20%
<7.5%
<10%
<20%
Under Q
Under P
0.5551(0.0443)
0.2026(0.0206)
0.0069(0.0014)
0.9888(0.0525)
0.5640(0.0346)
0.0862(0.0084)
0.0098(0.0142)
0.0050(0.0083)
0.0010(0.0022)
0.0036(0.0048)
0.0017(0.0026)
0.0002(0.0005)
is immediately evident from Figure 1, the magnitude of the left tail measure
substantially exceeds that of the right tail. This asymmetry in the tails under
Q contrasts sharply with the previously discussed approximately symmetric
tail behavior under P. This result, of course, is also consistent with the wellknown fact that out-of-the-money puts tend to be systematically overpriced
when assessed by standard pricing models; see, for example, the discussion in
Bates (2009), Bondarenko (2003), Broadie, Chernov, and Johannes (2009b), and
Foresi and Wu (2005). Moreover, the three risk-neutral variation measures all
clearly vary over time, reaching unprecedented levels in Fall 2008. We study
these dynamic patterns more closely below.
V. Tails, Fears, and Risk Premia
We begin our discussion of the equity and variance risk premia and the
relationship between the P and Q measures in Table I with a direct comparison
of the average intensity for large jumps under the two different measures. In an
effort to focus on normal times and prevent our estimates of the risk premia
and investor fears to be unduly influenced by the recent financial market crises,
we explicitly exclude the July 2007 through December 2008 part of the sample
from this estimation.
As is immediately evident from Table I, the estimated frequencies for the
occurrences of large jumps are orders of magnitude smaller than the implied
jump intensities under the risk-neutral measure.30 Since we explicitly rely on
the frequently occurring medium-sized jumps and EVT for meaningfully uncovering the P jump tails, this marked difference cannot simply be attributed to
a standard peso type problem and a nonrepresentative sample. The table also
reveals a strong sense of investor fears, as manifested by the much larger intensities for the left jump tails under Q and the willingness to protect against
30
This is consistent with the aforementioned earlier empirical evidence reported in At-Sahalia,
Wang, and Yared (2001); see their Section 3.5.
2184
negative jumps. For instance, the estimated risk-neutral jump intensity for
negative jumps less than 20% is more than 12 times larger than that for
positive jumps in excess of 20%, and the difference is highly statistically significant. The differences in the left and right P jump tails, on the other hand,
are not nearly as dramatic and also statistically insignificant.
Although the option-implied risk-neutral intensities for positive jumps are
much smaller than the intensities for negative jumps, they are still much
higher than the estimated intensities for positive jumps under the statistical measure. This contrasts with previous estimates from parametric models,
which typically imply trivial and even negative premia for positive jumps.
Most of these findings, however, are based on tightly parameterized models
that constrain the overall jump intensity, that is, the intensity for jumps of
any size, to be the same under the risk-neutral and statistical measures.
Thus, if negative jumps are priced, in the sense that their risk-neutral intensity exceeds their statistical counterpart, this automatically implies that
positive jumps will carry no or even a negative risk premium. Meanwhile,
our nonparametric approach, which does not restrict the shape and relation
between the P and Q jumps, suggests that the large positive jumps do in
fact carry a premia, albeit of a much smaller magnitude than the negative
jumps.31
Turning next to the actual risk premia, Figure 2 plots the components of the
equity and variance risk premia that are due to compensation for rare events,
that is, ERP(k) and VRP(k), respectively. As mentioned before, we explicitly
exclude the last 18 months of the sample in the estimation of the volatility
forecasting model and different jump tail measures, so that the premia in
the shaded parts of the figure, coinciding with the advent of the recent financial market crises, are based on the in-sample estimates with data through
June 2007 only. Also, for ease of interpretation, both of the premia are reported
at a monthly frequency based on 22-day moving averages of the corresponding
daily estimates.32
There are obvious similarities in the general dynamic dependencies in the
two jump risk premia, with many of the peaks in ERP(k) coinciding with the
troughs in VRP(k). There are also important differences, however, in the way
31 As discussed further below, the temporal variation in the jump intensities means that positive jumps may indeed carry a risk premium. This is reminiscent of the U-shaped pattern in
the projection of the pricing kernel on the space of market returns previously documented in
At-Sahalia and Lo (2000) and Rosenberg and Engle (2002), among others.
32 The time-to-expiration of the options used in the calculation of the risk premia changes systematically over the month, inducing monthly periodicity in the day-by-day estimates. To illustrate,
consider the popular jump-diffusion model of Duffie, Pan, and Singleton (2000). For simplicity, suppose there is only a single continuous volatility factor. Let P and P denote the mean-reversion
parameter and the mean of 2t under P, with the corresponding parameters under Q denoted by
Q and Q , respectively. The unconditional mean of the risk premium may then be expressed as
Q
K0 + K1 ( P Q )(1 e (T t) )( Q (T t))1 , where the two constants K 0 and K 1 are determined
by the risk premia parameters. This expected premium obviously depends on the horizon and
the relative import of the continuous stochastic volatility component, in turn inducing monthly
periodicity in the daily measures.
2185
Figure 2. Equity and variance risk premia due to large jumps. The estimates for the risk
premia are based on 5-minute S&P 500 futures prices and options. The values for the premia in
the shaded area, corresponding to July 2007 through December 2008, are based on the parameter
estimates for the P and Q measures obtained using data through June 2007. The log-moneyness
of the options used for the left and right tails is fixed at k = 0.9 and k = 1.1, respectively.
in which the compensation for the tail events manifest in the two premia. We
begin with a more detailed discussion of the equity jump risk premium depicted
in the top panel.
A. Equity Jump Risk Premium

Most of the peaks in the equity jump risk premium are readily associated
with specific economic events. In particular, the October 1997 mini-crash and
the August to September 1998 turmoil associated with the Russian default
and the long-term capital management (LTCM) debacle both resulted in sharp,
but relatively short lived, increases in the required compensation for tail risks.
A similar albeit much smaller increase is observed in connection with the
September 11, 2001 attacks. Interestingly, the magnitude of the equity jump
risk premium in October 2008 is about the same as the premium observed in
August to September 1998.
To better understand where this compensation for tail risk is coming from, it
is instructive to decompose the total jump risk premium into the parts associ
ated with negative and positive jumps, say ERP+
t (k) and ERPt (k), respectively.
2186
Figure 3. Decomposition of equity jump premia. The estimates for the equity risk premia
components are based on 5-minute S&P 500 futures prices and options. The values for the premia
in the shaded area, corresponding to July 2007 through December 2008, are based on the parameter
estimates for the P and Q measures obtained using data through June 2007. The log-moneyness
of the options used for the left and right tails is fixed at k = 0.9 and k = 1.1, respectively.
This decomposition, depicted in Figure 3, shows that, although the behavior

of the two tails is clearly related, the contributions to the overall risk premium are far from symmetric. Most noticeably, the October 1997 mini-crash
seems to be associated with an increased fear among investors of additional
sharp market declines and a corresponding peak in ERP
t (k), while there is
(k).
Similarly,
the
August
to
September
1998 market
hardly any effect on ERP+
t
+
(k)
than
it
did
on
ERP
turmoil had a much bigger impact on ERP
t
t (k). Conversely, the dramatic stock market declines observed in July and September
2002 in connection with the burst of the dot-com bubble resulted in peaks and
troughs in both ERP+

t (k) and ERPt (k), with less of a discernable impact on the
total jump risk premium ERPt (k). Likewise, the March 2003 start of the Second
Gulf War is hardly visible in ERPt (k), while it clearly influenced both ERP+
t (k)
and ERP
t (k). Most dramatic, however, the sharp increase in ERPt (k) associated with the Fall 2008 financial crises masks even larger offsetting changes
in ERP+
t (k) and ERPt (k) that dwarf the separate jump risk premia observed
during the rest of the sample. As such, this directly underscores the notion that
investors simply did not know what the right price was at the time.
Looking at the typical sample values, again excluding the July 2007 to
December 2008 part of the sample, the median of the estimated equity risk
2187
premia due to rare events equals 5.2%.33 This is quite high, and, compared to
the prototypical estimate of 8% for the equity risk premia in postwar U.S. data
(see, for example, Cochrane (2005)), our results imply that fears of rare events
account for roughly two-thirds of the total expected excess return. These numbers contrast with the estimation results reported in Broadie, Chernov, and
Johannes (2009a), which imply a mean ERPt (k) of 1.85%, and the estimates in
Eraker (2004), which imply a premium for rare events of only 0.65%.34
At a more general level, our findings suggest that excluding fears and tail
events, the magnitude of the equity risk premium is quite compatible with
the implications from standard consumption-based asset pricing models with
reasonable levels of risk aversion. Of course, this raises the question of why
investors price tail risk so high. Recent pricing models that pay special attention to tail risks include Pan, Liu, and Wang (2005) and Bates (2008); the work
by Bansal and Shaliastovich (2009) based on the concept of confidence risk; and
Drechsler (2011), who emphasizes the role of time-varying model uncertainty
in amplifying the impact of jumps.
Further corroborating these ideas, we show next that fears of rare events
account for an even larger fraction of the historically large and difficult to
explain variance risk premium.
B. Variance Jump Risk Premium
The general dynamic dependencies in the jump variance risk premium depicted in the second panel in Figure 2 fairly closely mirror those of the jump
equity premium in the first panel. Similarly, decomposing VRPt (k) into the
parts associated with negative and positive jumps, say VRP+

t (k) and VRPt (k),
respectively, reveals a rather close coherence between the two tail measures.35
With the notable exception of the aforementioned October 1997 mini-crash
and the August to September 1998 Russian default and LTCM debacle, most
of the spikes coincide between the two time series plotted in Figure 4 . The
+
magnitude of VRP
t (k), of course, typically exceeds that of VRPt (k) by a fac36
tor of two to four. Do these findings of large tail risk premia imply special
treatment, or compensation, for rare events? If large (or jump-related) risks
were treated the same as small (or diffusive) risks, what would the magnitudes
of the tail risk premia be?
33
Including the more recent period, the median increases to 5.6%.

Both of these numbers are based on a specific stochastic volatility model with correlated
jumps, referred to in the literature as a stochastic volatility with correlated jumps (SVCJ) type
model, in which the parametric form of the jumps effectively implies an exponential decay.
35 This result is not simply an artifact of the assumption for the P jump intensity in equation (27).
The P tail moments are orders of magnitude smaller than the moments under Q, and as such the
dynamic dependencies seen in the figure are largely driven by the dependencies in the Q jump
intensity estimated under the less restrictive assumption in (19).
36 This is also in line with the recent empirical findings in Andersen and Bondarenko (2010)
related to their so-called up and down implied variance measures.
34
2188
Figure 4. Decomposition of variance jump premia. The estimates for the variance risk premia components are based on 5-minute S&P 500 futures prices and options. The values for the
premia in the shaded area, corresponding to July 2007 through December 2008, are based on
the parameter estimates for the P and Q measures obtained using data through June 2007. The
log-moneyness of the options used for the left and right tails is fixed at k = 0.9 and k = 1.1,
respectively.
To begin answering these questions, it is instructive to further decompose the

variance jump tail premia. In particular, from the formal definition of VRP+
t (k),
it follows that
VRP+
t (k)

T
T
1
Q
P
2
P
2
P
=
x dss (dx) Et
x dss (dx)
Et
T t
t
t
x>k
x>k

T
T
1
Q
2
P
2
Q
+
Et
x dss (dx)
x dss (dx) , (30)
T t
t
t
x>k
x>k
with a similar decomposition available for VRP

t (k).
The first term on the right-hand side in equation (30) reflects the compensation for time-varying jump intensity risk. It mirrors the compensation
demanded by investors for the continuous part of the quadratic variation,
that is, VRP ct (k) in equation (10). As such, it is naturally associated with investors willingness to hedge against changes in the investment opportunity set.
2189
Specifically, for the jump intensity process in equation (27), the relevant difference between the P and Q expectations takes the form
EPt
=

x
1+
x>k
dssP (dx)

2 P
EPt
x (x) dx
x>k
EQ
t

t
t
T
x 2 dssP (dx)
x>k
s2 ds
EQ
t

s2 ds
(31)
For Levy type jumps, that is, jumps with constant jump intensity, this part of
the variance jump risk premium will be identically equal to zero.
By contrast, the second term on the right-hand side in equation (30) involves
the wedge between the risk-neutral and objective jump intensities. But this
difference is evaluated under the same probability measure, and as such it is
effectively purged of the premia due to temporal variation in the jump intensities. If jumps were treated the same as diffusive risks, this component would
be zero, since the predictable quadratic variation of the continuous price does
not change when changing the measure from P to Q.
In more technical terms, the quadratic variation of the log-price process may
be split into its predictable component and a residual martingale component
solely due to jumps,
QV[t,T ] =
t

s2 ds +

R
fs , fs [t,T ] +
t
x 2 dssP (dx) +

R

t

R
x 2 (ds,
dx)
x 2 (ds,
dx),
where fs , fs [t,T] denotes the predictable quadratic variation; see, for example,
Jacod and Shiryaev (2003). The pricing of the first term, fs , fs [t,T] , is naturally
associated with changes in the investment opportunity set, while the pricing
of the martingale part reflects any special treatment of jump risk.
The high-frequency-based estimation results discussed in the previous section indicate significant temporal variation in the tail jump intensities. Part
of the tail risk premia therefore comes from the first term in equation (30).
However, our estimation results, see Table A.I in the Appendix, also suggest that the left and right tails behave quite similarly under P, so that in
particular37

+
2 P
1
x (x) dx 1
x 2 P (x) dx.
x>k
x<k
37 Note that this result is not by construction as assumption (27) allows for asymmetric tails
since the left and right tails are allowed to have different loadings on the constant and time-varying
parts of the jump intensity.
2190
Figure 5. Investor fears. The estimates for the FI(k) Investor Fears index are based on 5-minute
S&P 500 futures prices and options. The values in the shaded area, corresponding to July 2007
through December 2008, are based on the parameter estimates for the P and Q measures obtained
using data through June 2007. The log-moneyness of the options used for the left and right tails is
fixed at k = 0.9 and k = 1.1, respectively.
Combining this approximation with (31) and the comparable expression for the
left tail measure, it follows that the difference
+
FIt (k) = VRP
t (k) VRPt (k)
(32)
will be largely void of the risk premia due to temporal variation in the jump
intensities. Consequently, FIt (k) may be interpreted as a direct measure of
investor fears, or crash-o-phobia.38
The corresponding plot in Figure 5 shows that the sharpest increase in investor fears over the sample did indeed occur during the recent financial crises.
However, the Russian default and long-term capital management collapse in
August to September 1998 resulted in a spike in the Investor Fears index of almost two-thirds of its recent peak. Slightly less dramatic increases are manifest
in connection with the October 1997 mini-crash; September 11, 2001; and the
summer of 2002; and the burst of the dot-com bubble. The figure also reveals
systematically low investor fears from 2003 through mid-2007, corresponding
to the general run-up in market values.
The average sample values reported in Table II further corroborate the idea
that much of the variance risk premium comes from crash-o-phobia, or special
38 Rubinstein (1994) first attributed the smirk-like pattern in post October 1987 BlackScholes
implied volatilities for the aggregate market portfolio when plotted against the degree of moneyness
as evidence of crash-o-phobia; see also Foresi and Wu (2005) for more recent related international
empirical evidence.
2191
Table II
Variance Risk Decomposition

The average sample estimates are reported in annualized form, with standard errors in parentheses. The estimates under Q are based on S&P 500 options data from January 1996 through June
2007, while the estimates under P rely on 5-minute high-frequency S&P 500 futures data from
January 1990 through June 2007.
Under Q

E( T1t
E( T1t
t T x>ln 1.1
t
x<ln 0.9
E(QVtQ )
Q
x 2 EQ
t s (dx) ds)
0.0032(0.0004)
Q
x 2 EQ
t s (dx) ds)
0.0180(0.0015)
0.0472(0.0046)
Under P

E( T1t
E( T1t
t
t
E(QVtP )
x>ln 1.1
x<ln 0.9
x 2 EPt sP (dx) ds)
0.000196(0.000483)
x 2 EPt sP (dx) ds)
0.000062(0.000115)
0.0269(0.0031)
compensation for tail risks. Excluding the post-July 2007 crises period, the
average sample variance risk premium equals E(VRPt ) = 0.020. Of that total
premium, E(VRP+
t (k))/E(VRPt ) = 14.8% is due to compensation for right tail
risks, while E(VRP
t (k))/E(VRPt ) = 88.4% comes from left tail risks. Looking at
the difference, our results therefore suggest that close to three-quarters of the
variance risk premium may be attributed to investor fears as opposed to more
standard rational-based pricing arguments.39
Comparing these numbers to the implications from the aforementioned
parametric model estimates reported in the literature, the study by Broadie,
Chernov, and Johannes (2009a) implies that E(VRPt (k))/E(VRPt ) = 24.4%,
while the results in Eraker (2004) suggest that E(VRPt (k))/E(VRPt ) = 20.0%.40
These numbers obviously differ quite dramatically from the results obtained
with our nonparametric approach, which effectively suggests that all or most of
the variance risk premium may be attributed to the compensation for jump tail
39 Note that these proportions are maturity specific, and based on a median maturity of 14 days.
To check the sensitivity of our results with respect to the maturity and statistical uncertainty
associated with our nonparametric inference procedure, we redid the estimation using monthly
data comprised of options on the fifth business day of each month. The resulting monthly estimates
for E(VRP+
t (k))/E(VRPt ) = 8.0% and E(VRPt (k))/E(VRPt ) = 60.8% still imply that the fear component accounts for more than half of the variance risk premium. Thus, even though the relative
importance of VRPt (k) drops somewhat, the general findings remain intact. We note that these
differences are also consistent with the biases observed in the Monte Carlo simulations reported
in Section VI below.
40 These numbers are calculated on the basis of the specific parameter estimates for the SVCJ
models reported in the two studies, with the horizon fixed at 14 days, corresponding to the median
time-to-expiration of the options used in our analysis, along with k = 10%.
2192
risk. Importantly, however, existing parametric studies generally constrain the

wedge between the risk-neutral and statistical jump distributions to a negative
shift in the mean jump size;41 see, for example, the discussion in Singleton
(2006). The results reported here indicate that the difference in the jump behavior under the two probability measures is much more intricate than that,
and that time-varying jump intensities play an especially important role for
properly understanding the risk premia and the inherent fear component.
We next check the robustness of our findings with respect to the general
assumptions concerning the P and Q dynamic tail dependencies, and the specific
asymptotic approximations underlying our jump tail estimates.
VI. Robustness Checks
Our approach for estimating the tail risk premia and the Investor Fears
index is extremely flexible, with the only parametric assumptions relating to
the very general form of the time-varying jump intensities for the P and Q
jump tails in equations (27) and (19), respectively. We begin our discussion
below with a series of formal rank tests for the empirical validity of these
assumptions that underlie our estimation of the RTt and LTt tail measures. We
then present summary results from a detailed Monte Carlo simulation study
designed to investigate the accuracy of our Q jump tail estimation. We also
briefly comment on some additional robustness checks related to the use of
lower frequency daily data and shorter-lived options.
A. Time Variation in P Jump Tails
The relevant assumption in equation (27) restricts the intensities for the
positive and negative jumps to be affine functions, possibly with different
coefficients, of the latent instantaneous volatility 2t . This in turn implies
t+1
that the left and right realized jump variances, t+t x>0 x 2 (ds, dx) and
t+1
2
t+t x<0 x (ds, dx), depend upon the past only through the temporal variation in the integrated stochastic volatility. Consequently, the unconditional
(auto)covariances equal

t
t
t+1
t+1
x 2 (ds, dx),
s2 ds = K1 cov
s2 ds,
s2 ds ,
cov
t+t
t1+t
x>0
and

t+1
t+t
x (ds, dx),

= K1 cov
x>0
t+1

x (ds, dx)
t+t
x>0
t1+t

s2 ds,
t1+t
cov
41
t+t
x (ds, dx) ,
2
t1+t
x>0
Broadie, Chernov, and Johannes (2009) also consider changing the variance of the jumps.
2193
Table III
Rank Tests for Affine P Jump Intensities

The table reports the results of the Gill and Lewbel (1992) rank test, as detailed in Appendix
Subsection D. The realized variation measures are based on 5-minute high-frequency S&P
500 futures data from January 1990 through December 2008. The test has a limiting 2d distribution under the null hypothesis that the autocovariance matrix has a reduced rank of 1. The
calculation of the asymptotic variance for the autocovariance matrices are based on a Parzen kernel
with a lag length of 70.
Autocovariance Matrix
Test Stat.
p-value
3.7937
1.2616
0.2590
3.5299
2
1
1
1
0.1500
0.2613
0.6108
0.0603
(CVt , RJVt , LJVt )

(CVt , RJVt )
(CVt , LJVt )
(RJVt , LJVt )
where K1 = (1+ )2

Zt

x>0
t+1
t+t
x 2 P (x) dx. Now, defining the vector

s2 ds,
t+1

x (ds, dx),
t+1

x (ds, dx)
t+t
x>0
t+t
x<0
it follows that the autocovariance matrix for Zt should be of reduced rank.42

We state this testable implication in the following proposition.
2: If 2t is a covariance stationary and ergodic process, and
P
ROPOSITION
2 P
R x (x) dx < , the assumption in equation (27) implies that rank
(cov(Zt , Zt1 )) = 1.
In practice, of course, the vector Zt is not directly observable. However, as discussed in Section III above, it may be accurately estimated by (CVt , RJVt , LJVt ).
Even so, this vector will invariably include some estimation error, which will
render the resulting autocovariance matrix of full rank even if cov(Zt , Zt1 ) is
not. Hence, in order to formally test whether the deviations from reduced rank
in the actually observed sample autocovariance matrices for (CVt , RJVt , LJVt ),
and subsets thereof, are statistically significant, we rely on the test developed
by Gill and Lewbel (1992); further details concerning the implementation are
given in Subsection D of the Appendix.43
None of the resulting test statistics reported in Table III reject the reduced
rank hypothesis. As such, the results do not provide any evidence against the
affine restriction in equation (27) related to the actual jump intensities.44
42 The possibility of jumps in 2 means that the covariance matrix of Z will not necessarily be
t
t
of reduced rank.
43 This test is relatively easy to implement, and it overcomes some of the computational difficulties associated with other tests for rank deficiency based on the smallest eigenvalues of the matrix.
The Monte Carlo simulation evidence discussed in Gill and Lewbel (1992) also suggests that the
test exhibits desirable size and power properties.
44 The test statistics reported in the table are based on the realized variation measures over
the full sample. However, very similar test statistics are obtained for the shorter January 1990
2194
B. Time Variation in Q Jump Tails

In contrast to the affine dependencies for the P jump intensities underlying
our estimates of the statistical tail measures, our estimates for the risk-neutral
tail measures rely on the less restrictive condition in equation (19). This assumption merely requires the temporal variation to be the same across different
jump sizes, and as such is trivially satisfied by all parametric models hitherto
estimated in the literature. This level of generality for the Q jump measures
may be particularly important to properly accommodate the seemingly complex
dynamic dependencies in investor fears.
To more formally investigate this hypothesis, suppose that the Q jump intensities satisfied the analogous affine dependencies assumed for the P jump
measure,

tQ (dx) = 0+Q 1{x>0} + 0Q 1{x<0} + 1+Q 1{x>0} + 1Q 1{x<0} t2 Q (x) dx, (33)
where Q (x) denotes a valid Levy density, and 0Q and 1Q are nonnegative
constants.45 Similar to the definition of Zt in Proposition 2 above, define the vector of risk-neutral measures Yt (RTtQ (k), LTtQ (k), QVtQ ) . If (33) holds, then
RTtQ (k), LTtQ (k), and QVtQ should all be linear functions of the risk-neutral exT 2
pected integrated volatility EQ
t ( t s ds). Consequently, the covariance matrix,
and equivalently the lag-one autocovariance matrix, for Yt should be of reduced
rank one. If (33) does not hold, QVtQ may still depend linearly on the integrated
T 2
volatility EQ
t ( t s ds). However, in this situation additional time-varying factors will typically affect Yt , so that rank(var(Yt )) rank(cov(Yt , Yt1 )) > 1. The
following testable proposition summarizes this implication.
3: If Yt is a covariance stationary and ergodic process, and
PROPOSITION
x Q
e
(x)
dx
< for some > 0, the assumption in equation (33) implies
|x|>
that rank(cov(Yt , Yt1 )) = 1.
The results from the rank tests for the Q tail measures are reported in
Table IV. As seen from the table, all of the tests clearly reject the affine structure for the risk-neutral jump intensities.46 Of course, this contrasts sharply
with the comparable test results for the P jump intensities in Table III, which
generally support the affine structure.
Taken as a whole, the robustness tests in Tables III and IV pertaining to the
temporal variation in the jump intensities are consistent with the key minimal
through June 2007 sample employed in the tail estimation. We also implemented the same set of
1/2
1/2
tests with all of the variables replaced by their square roots, that is, (CV t , RJVt 1/2 , LJV t ). All of
these tests strongly rejected the reduced rank hypothesis, highlighting the power of the procedure.
45 This assumption, together with equation (27), implies that the temporal variation in all risk
premia would be spanned by the level of the stochastic volatility.
46 The tests reported in the table are based on the full sample through the end of 2008. As
before, very similar results obtain when excluding the July 2007 through December 2008 part of
the sample.
2195
Table IV
Rank Tests for Affine Q Jump Intensities

The table reports the results of the Gill and Lewbel (1992) rank test, as detailed in Appendix
Subsection D. The variation measures are based on S&P 500 options data from January 1996
through December 2008. The test has a limiting 2d distribution under the null hypothesis that the
autocovariance matrix has a reduced rank of 1. The calculation of the asymptotic variance for the
autocovariance matrices are based on a Parzen kernel with a lag length of 70.
Autocovariance Matrix
Test Stat.
p-value
65.8861
29.3789
31.7486
12.4262
2
1
1
1
0.0000
0.0000
0.0000
0.0000
(QVtQ , RTtQ , LTtQ )

(QVtQ , RTtQ )
(QVtQ , LTtQ )
(RTtQ , LTtQ )
assumptions underlying our nonparametric inference procedures. Importantly,

the test results clearly point to the existence of much more complex dynamic
dependencies in the risk-neutral jump tails than hitherto entertained in the
literature, and in the form of dependencies that cannot simply be spanned by
the level of the stochastic volatility. As such, this also calls into question the
validity of parametric procedures that restrict the form of the tails to be the
same across the P and Q distributions when trying to understand notions of
investor fears, or crash-o-phobia, embedded in market prices.
C. Estimation of Q Tail Measures
To gauge the magnitude of the approximation errors in Proposition 1 and
the accuracy of the estimates for the LTtQ and RTtQ tail measures, we simulate
artificial option prices from four different parametric models. Model A refers to
the classical Merton model with constant volatility and normally distributed
compound Poisson jumps. Following Carr and Wu (2003), we fix the continuous
volatility at 14%, the jump intensity at two, the mean log jump size at 5%,
and the volatility of the log jump size at 13%, all in annualized units.47
The remaining three models, Models B, C, and D, all assume that the instantaneous volatility follows a square-root diffusion process,

dt2 = i i t2 dt + t dWt , i = p, q,
(34)
where = 0.1, and the rest of the parameters under the P and Q probability
measures are fixed at p = 0.132 , p = 5.04, q = 0.142 , and q = 4.3457, respectively. Following Bates (2000), Model B further assumes normally distributed
47
The only difference from Carr and Wu (2003) is the reduction in the mean log-jump size from
10% to 5% in order to more closely resemble the sample averages in our data.
2196
Table V
Estimation of Risk-Neutral Tail Measures

The table reports summary statistics for the estimation of the LTt and RTt measures based on
out-of-the-money options computed with simulated data from Models A-D as described in the text.
Time-to-maturity T t is fixed at 15 days, and the log-moneyness k for the left and right tails are
set at 0.9 and 1.1, respectively. The length of the simulated series is 5, 000 days. MAD denotes the
median absolute deviation.
LTtQ
Model A
Model B
Model C
Model D
RTtQ
True
Median
MAD
True
Median
MAD
0.0484
0.0418
0.0407
0.0380
0.0496
0.0437
0.0406
0.0408
0.0072
0.0056
0.0070
0.0201
0.0173
0.0112
0.0101
0.0238
0.0205
0.0128
0.0139
0.0039
0.0021
0.0032
jumps with jump intensity proportional to the instantaneous variance,

t (dx) =
e
i t2
0.5(xi )2 /i2
2 i2
, i = p, q,
(35)
with the P and Q parameters set at p = 2/ p , p = 0, p = 0.0561, q = 2/ q ,

q = 0.05, and q = 0.13, respectively. These values for the Q jump parameters
ensure that, conditional on the mean level of the volatility, the jumps have the
same distribution and intensity as in Model A.
Finally, Models C and D postulate a generalized tempered stable process for
the jump compensator, as in, for example, Bates (2009) and Carr et al. (2003),
t (dx) =

t2 Mi |x|
c e
1{x<0} + c+ eNi x 1{x>0} , i = p, q,
+1
|x|
(36)
where for Model C the parameters are fixed at Mp = Np = 13.0543, Mq =

3, Nq = 8.1313, = 0.0, and c = c+ = 0.3600/ p , while for Model D we set
Mp = Np = 19.5026, Mq = 3, Nq = 9.1990, = 0.5, and c = c+ = 0.2053/ p . The
Q jump parameters are chosen so that the expected values of LTtQ and RTtQ
implied by tQ (dx) are in line with their corresponding sample means from our
data. The P jump parameters imply that the mean value of the daily P jump
variation equals 0.25 in daily percentage units, which is generally consistent
with recent nonparametric empirical evidence reported in the literature.
Turning to the results, Table V reports the median and median absolute
deviation (MAD) estimates for the two tail measures obtained across 5,000
simulated days for each of the four models for options with T t set to 15 days
to maturity and k equal to 0.9 and 1.1 for the left and right tails, respectively.48
These characteristics correspond directly to the average maturity time and
moneyness for the options used in the actual estimation. As is evident from
48
The results for the Merton Model A are exact, so the MADs are identically equal to zero.
2197
the table, the left skew in the risk-neutral jump distributions renders the left
tail measures two to three and a half times larger than the corresponding
right tail measures and generally also easier to estimate. The slight upward
bias observed in both tails is directly attributable to the continuous variation
invariably clouding the identification of the true jump tails. All in all, however,
the results clearly underscore the accuracy of the theoretical approximations
underlying our estimation of the jump tails.49
D. Daily Data and Shorter Options
To check the robustness of our main empirical findings with respect to the use
of intraday data for the extraction of the P jump tails, we redid the estimation
of the tail parameters using only daily data. Restricting the analysis to a
coarser daily frequency invariably handicaps the detection of jumps and the
corresponding jump tail estimation. Nonetheless, the key features pertaining to
the behavior of the jump tails remain intact, albeit not as precisely estimated.50
In a similar vein, to corroborate the simulation results discussed above and
check the robustness of our empirical results with respect to the options used
in extracting the Q jump tails, we re-estimated the tails with only monthly
data. Specifically, by restricting our sample to the 5th business day of each
month, it is possible to reduce the median time-to-maturity of the options from
14 to only 7 business days. Nonetheless, the resulting monthly Investor Fears
index looks almost identical to the plot discussed in Section IV above. Further
details concerning these and other robustness checks are again available in the
Internet Appendix.
VII. Conclusion
The recent market turmoil has spurred renewed interest in the study and
economic impact of rare disaster type events. Most of these studies resort to
peso type explanations for the occurrence of rare events along with specific
parametric modeling assumptions for carrying out statistical inference. In contrast, the new methodologies developed here are based on flexible nonparametric procedures, which rely on high-frequency data and EVT for effectively
measuring the actual expected tail events together with a rich set of options
for uncovering the pricing of tail events.
Allowing the data to speak for themselves, our empirical findings suggest
that much of the historically large equity and variance risk premia may be
49 Much more detailed simulation results and further discussion for a range of different maturity
times and moneyness are available in the supplementary Internet Appendix (available online in the
Supplements and Datasets section at http://www.afajof.org/supplements.asp). Not surprisingly,
these results reveal that the quality of the tail estimates generally deteriorates the closer to the
money and the longer lived are the options used in the estimation.
50 By the basic statistical principle that more data are always better, see, for example, the
pertinent discussion in Zhang, Mykland, and At-Sahalia (2005); the high-frequency data, if used
correctly, will always result in more accurate estimates.
2198
ascribed to compensation for jump tail risk. These types of risks are poorly
described by traditional mean-variance type frameworks. Meanwhile, excluding the part of the risk premia directly attributable to tail events, the average
equity and variance premia observed in the data are in line with the general
implications from standard equilibrium-based asset pricing models with reasonable preference parameters. Our decomposition of the variance risk premia
further suggests that investors fears, or crash-o-phobia, play a crucial role in
explaining the overall large magnitude of the tail risk premia. At the same time,
the perception of jump tail risks appears to vary quite dramatically over time,
and changes therein may help explain some of the recent sharp movements in
market valuations.
Appendix
A. Estimation of Jump Tails
We start with some general remarks that apply for both the P and the Q
distributions, and then subsequently specialize the discussion to the statistical
and risk-neutral jump tail estimation. For ease of notation, we suppress the
dependence on P and Q in the Levy densities P (x) and Q (x) in this initial
discussion.
To begin, define + (x) = ex 1 for x R+ and (x) = ex for x R , where
+
(x) maps R+ into R+ and (x) maps R into (1, +). Both mappings are
for y (0, ) and (y) = (yln y) .
one-to-one. Further, denote + (y) = (ln(y+1))
y+1
Finally, denote the tail jump measures

(x) =
(u) du,
(A1)
x
with x > 0 for +

(x) and x > 1 for (x).
Our estimation under both P and Q is nonparametric and based on the assumption that the tail Levy densities
belong to the domain of attraction
of an extreme value distribution; see, for example, the general discussion in
Embrechts, Kluppelberg, and Mikosch (2001), along with the recent empirical analysis related to extreme market moves in Bakshi, Madan, and Panayotov (2009). This is a very wide class of jump measures, and it includes most
parametric specifications previously used in the literature.
The key implication of the assumption (see, for example, Smith (1987)), is
that
(u + x) appr
G(u; , ),
(x)
where
G(u; , ) =
u > 0, x > 0,
(A2)
1 (1 + u/ )1/ , = 0, > 0
1 eu/ ,
= 0, > 0
(A3)
2199
with G(u; , ) the cdf of a generalized Pareto distribution. In what follows, we

assume that the approximation in (A2) is exact when x is large enough; see the
formal proofs in Bollerslev and Todorov (2011) for further details concerning
the approximation error and the exact conditions under which the error is
indeed asymptotically negligible.
Before we turn to our discussion of the estimation of the parameters for the
generalized Pareto distribution approximation to the tail Levy densities, we
present several integrals with respect to the Levy density (x) that appear in
ERPt (k), VRPt (k), RTtP (k), LTtP (k), RTtQ (k), and LTtQ (k). In particular, assuming
that the approximation in (A2) holds, it follows that, for k > 0 and some tr+ <
ek 1,

k
(e x 1)(x) dx = (ek 1) +
(e
1)
+
(e x ek)+ (x) dx,
x>k

R
k +
(e e ) (x) dx =
x
1)+ + (x) dx
(x e +
k
(0,)
+
+
(tr )
+
+
(tr )
ek 1

(e x 1 x)(x) dx =
x>k
x>ek 1
1 1
+
1 +

=
ek 1
+
(x)
+
(x) dx
+
+
(tr )
dx

11/ +
+ k
+
1 + + (e 1 tr )
,
(x ln(x + 1))+ (x) dx

11/ +
+ k
+
1 + + (e 1 tr )
=
1)(e 1 k) +
+
+
+
1 1
(tr + )( + )1/ +
1 + (tr + 1) 1
;
,
F
;
1
+
;
2
1
+
+ +
+ k
(ek + )1/
+
e
+

ln(x + 1) +
k
x 2 (x) dx = k2 +
(x) dx
(e 1) + 2
k
x+1
x>k
x>e 1
k
+
(e
+
+
(tr )
+
1 +
+
k
+
= k2 +
(e 1) + 2 (tr ) K1 ,
and
+
K1 = ek/ +
+
+ 1/
+ +
(tr + 1) 1
+
3 F2 1 , 1 , 1 ; 1 + 1 , 1 + 1 ; +
+ + +
+
+
+
k
e +
1 1
1
+ k2 F1
+ , + , 1 + + ;
+
(tr +
+
+ 1)1
,
+
ek +
2200
where 3 F 2 ( , , ; , ; ) and 2 F 1 ( , , ; ) are the hypergeometric

functions. The above

(x)dx
<
for some > 0,
integrals are all well defined provided that x> x +
that is, + < 1. Similar calculations for k < 0 and 1 < tr < ek yield,

k
+

k
e e x (x) dx,
(e
)
(e x 1) (x) dx = ek 1
x<k

R

(x)
1 +
k
e
(x) dx =
dx
k
x
x2
(0,)
e

k1+1/ 1/

e
= tr
+ 1
1 tr
1 1
1
,
2 F1
1 + ; ; 2 + ;
ek
k
+
e e x (x) dx =

1
1 + ln x (x) dx
(e 1 x)(x) dx =
k
x
x<k
x>e

k 1/
k
k
= (e )(e 1 k) + tr
e
K2 ,
1 1
1 tr 1
K2 = 2 F1
,
,
1
+
;

k
e
tr
1 1
ek
1
,

2 F1 1 + ; ; 2 + ;
+1

k
e

x<k
k
x 2 (x) dx = k2
(e ) + 2
and
K3 = ek/
1/

x>ek
ln(x)
(x) dx = k2
(e ) + 2 (tr ) K3 ,
x

1 1 1
1
1 tr 1
3 F2
,
,
;
1
+
,
1
+
;
ek
tr 1
1 1
1
.
k2 F1
,
,
1
+
;

k
e
2201
A.1. Estimation of P Jump Tails

The estimation builds on the general theoretical results in Bollerslev and
Todorov (2011). In particular, let

, , 0
(tr ), 1 (tr ), , E(t )
denote the parameter vector that we seek to estimate. Our estimation is then
based on the scores associated with the log-likelihood function of the generalized Pareto distribution,

1
1
u
1
u
1+
1 (u) = + 2 1 +
,
( )
2 (u) =

1

u
1
u
u
1
+
1
+
ln
1
+
.
( )2
The specific estimation equations that we use are

N n1

1

n,t
i
j p tr 1{ (
n,t
= 0,
j p)>tr }
N
i = 1, 2,
t=1 j=1
N n1
1
1{ (
n,t
(1 ) (tr ) (tr )CVt = 0,
0
1
j p)>tr }
N
t=1 j=1
N
n1
1{ (
n,t
(1 ) (tr ) (tr )CVt CVt1 = 0,
0
1
j p)>tr }
N
t=2
j=1
and
(1 )
N
N
2
1

1
pt+t pt
RVt = 0.
N
N
t=1
t=1
The only remaining issue relates to the choice of tr . On the one hand, high
thresholds are preferred for the asymptotic extreme value approximation in
(A2) to work the best. On the other hand, we also need sufficiently many jumps
above the thresholds to make the estimation reliable. Faced with this tradeoff,
we set the values for tr so that they correspond to jumps in the log-prices of
0.6%. In the sample, we have 229 such jumps, 127 of them negative and 102 of
them positive. This choice for tr exceeds the maximum threshold value for t,i
2202
Table A.I
Estimates for P Tail Parameters

The estimates are based on 5-minute high-frequency S&P 500 futures data from January 1990
through June 2007. The truncation levels are fixed at tr = e0.006 1.006 and tr+ = e0.006 1
0.006 for the left and right tails, respectively.
Left Tail
Parameter
Right Tail
Estimate
0.3865
0.1709
0.0084
0.0333
100P
0 P
(tr )
)
1 P
(tr
St. Error
Parameter
Estimate
St. Error
0.1050
0.0222
0.0030
0.0050
P+
0.4361
0.1740
0.0002
0.0361
0.1558
0.0272
0.0021
0.0041
100P+
+
0+ P+
(tr )
+)
1+ P+
(tr
used in the construction of CVt and JVt (see equation (A10) below), so that the
continuous component in the price will have negligible effect in our estimation.
The actual results from the estimation are reported in Table A.I. As discussed
in the main text, the P tails are clearly time varying, but otherwise appear
close to symmetric. Not reported in the table, our estimate for E(t ) equal
to 0.2076 (0.0448) implies that on average roughly 20% of the total variation is
due to changes in the price during nontrading hours.
A.2. Estimation of Q Jump Tails
We start our discussion of the Q jump tail estimation with a formal proof of
Proposition 1.
Proof of Proposition 1: Using the Meyer-Ito formula (Theorem 68 in Protter
(2004)) for the payoff function of out-of-the-money calls, the definition of local
time (and the behavior of the transitional distribution of a Levy process over a
small time interval), together with the fact that futures prices are martingales
under the risk-neutral measure, we can write for arbitrary K > F t and T t

T

Q
+
Q
r(t,T ]
x
x +
1{Fs <K} Fs e K + 1{Fs >K} K Fs e
s (dx) ds,
e
Ct (K)
Et
t
(A4)
f (x)
where f (x) g(x) is to be interpreted as limx0 g(x)
= 1. A similar result
holds for close-to-maturity out-of-the-money puts. This result was first derived
in Carr and Wu (2003) in the proof of their Theorem 1.
Next, note that since Ft is a martingale under Q, using Doobs inequality for
K > F t , that is, for out-of-the-money call options, it follows that

(A5)
Qt sup |Fu Ft | K Ft C (T t),
u[t,T ]
where the constant C is adapted to the information set at time t, and is otherwise proportional to the risk-neutral (conditional to time t) expectation of the
2203
quadratic variation of the Q-martingale process {Ft }t . With this, we may write

e
r(t,T ]
Ct (K) =
EQ
t

=

R
1{Fs <K} (Fs e x K)+ sQ (dx) ds + O((T t)2 )
EQ
t 1{Ft <K}

R
(Fs e x K)+ sQ (dx) ds + O((T t)2 ).
(A6)
To continue further, note that for every x R,

|(Fs e x K)+ (Ft e x K)+ | e x |Fs Ft |.
(A7)
Now, making use of the Burkholder-Davis-Gundy inequality, it follows that

EQ
t
sup |Fu Ft | C T t,
(A8)
u[t,T ]
where the constant C satisfies the same restrictions spelled out immediately
above. Combining these results, we have
er(t,T ] Ct (K) = (T t)Ft RTtQ (k) + O((T t)3/2 ),
(A9)
where RTtQ (k) is formally defined in equation (17).

Using the decomposition of the risk-neutral Levy density in (19), it follows
further that
x
k1 +
RTtQ (k1 )
R (e e ) Q (x) dx

,
=
x
k2 +
RTtQ (k2 )
R (e e ) Q (x) dx
where k1 = FKt1 and k2 = FKt2 . Finally, applying the above formulas with tr+ =
K 1 /F t 1, together with the result in Smith (1987) concerning the residual
function + (x), concludes the proof for the right tail. The proof for the puts and
the left tail proceeds in a completely analogous fashion.

In the proof immediately above, we used a somewhat stronger version of the
approximation in (A2), and the fact that the true value of is /x. A similar
restriction underlies the popular Hill estimator of the tail index. However, this
stronger version of (A2) is not scale invariant. Therefore, while Proposition 1
conveys the main idea, in our actual estimation we only make use of (A2).
In particular, recall that our estimation of the jump tail probabilities under
Q only requires the weak separability condition for the time dependence and
jump sizes implicit in the Levy density in equation (19). The parameter vector
that we seek to estimate is

Q
(tr ) .
Q , Q , Q
2204
Table A.II
Estimates for Q Tail Parameters

The estimates are based on S&P 500 options data from January 1996 through June 2007. The
log-moneyness of the options used for the left and right tails is fixed at k = 0.9 and k = 1.1,
respectively.
Left Tail
Parameter
Q
Q
Q
1
( 10.075
)
Q
Estimate
0.2581
0.0497
0.9888
Right Tail
St. Error
Parameter
Estimate
St. Error
0.02816
0.0021
0.0525
Q+
Q+
+ Q+
Q (0.075)
0.0793
0.0238
0.5551
0.0147
0.0010
0.0443
From the calculations above, we have that for 0 < tr+ ek 1,

11/Q+

+
+

Q
Q
+ Q+
E RTtQ (k) = Q
(tr + )
,
1 + + (ek 1 tr + )
1 Q+
Q
while for 1 < tr ek ,

Q
E LTtQ (k) = Q
(tr )
Q + 1
k1+1/Q
e
1/Q
tr 1
Q
1 1
1
,
1
+
2 F1
;
;
2
+
;
Q Q
Q
Q
ek
Q
T +
Q T
+
1
where Q
E( T1t EQ
t ( t s ds)) and Q E( T t Et ( t s ds)).
Replacing the unconditional risk-neutral expectations of the jump tail measures with the sample analogues of RTtQ (k) and LTtQ (k) for three different
levels of moneyness for each of the tails results in an exactly identified
GMM estimator. The levels of moneyness that we use in the estimation are
K
= 0.9250, 0.9125, and 0.9000 for the left tail, and FKt = 1.0750, 1.0875, and
Ft
1.1000 for the right tail. These levels of moneyness are sufficiently deep in
the tails to guarantee that the effect of the diffusive price component in measuring LTtQ (k) and RTtQ (k) is minimal, and that the extreme value distribution
provides a good approximation to the jump tail probabilities. Also, the truncation levels are fixed at tr = (1 0.075)1 and tr+ = 0.075 for the left and
right tails, respectively. This corresponds directly to the moneyness for the
closest-to-the-money options used in the estimation.
The results from the estimation are reported in Table A.II. Not only is the left
tail under the Q distribution at a higher level, it also decays at a slower rate.
Although not directly comparable, the standard errors for the Q tail parameters are much smaller than the ones for the P tail parameters reported in Table
2205
A.I above. This is quite intuitive as our estimation for the risk-neutral distribution is based on expectations for the large jumps, while the estimates for the
statistical distribution rely on actual jump realizations in the high-frequency
data for inferring the relevant expectations, thus involving an extra layer of
estimation error uncertainty.
B. Estimation of Realized Measures
This section of the Appendix provides further details concerning the estimation of CVt , RJVt , and LJVt in equations (23) to (26) and the relevant truncation
levels. Theoretically, any > 0 and (0, 0.5) will work. In the actual estimation, however, we fix = 0.49. Our choice of is a bit more involved and
takes into account the fact that volatility changes across days and also has a
strong diurnal pattern over the trading day.
To account for the latter feature, we estimate nonparametrically a time-ofday factor TODi , i = 1,2, . . . , n,
T ODi
N
( f(t1+t )+i
n,t f(t1+t )+(i1)
n,t )2 1{| f(t1+t )+i
n,t f(t1+t )+(i1)
n,t |
0.49
n }
= NOIi
t=1
M
N
( f(t1+t )+i
n,t f(t1+t )+(i1)
n,t )2
,
(A10)
t=1 i=1
where
n1
N
NOIi =
1{| f(t1+t )+i

n,t f(t1+t )+(i1)
n,t |
0.49
n }
t=1 i=1
N
1{| f(t1+t )+i

n,t f(t1+t )+(i1)
n,t |
0.49
n }
t=1
=3
2
"
#
n1
N
#1
$
| f(t1+t )+i
n,t f(t1+t )+(i1)
n,t || f(t1+t )+(i1)
n,t f(t1+t )+(i2)
n,t |,
N
t=1 i=2
and
n 1n . The truncation level is based on the average volatility in the
sample, as measured by the so-called bipower variation measure defined in
Barndorff-Nielsen and Shephard (2004). Consequently, (1 E(t ))E(t2 ).
The NOIi factor puts the numerator and denominator in equation (A10) in
the same units. This factor is approximately equal to n 1. This calculation
of TODi implicitly assumes that t and 2t are both stationary and ergodic
2206
Figure A.1. Time-of-day factor. The estimates are based on 5-minute high-frequency S&P 500
futures data from January 1990 through June 2007.
processes. The resulting TODi estimates are plotted in Figure A.1. The figure
shows the previously well-documented U-shape in the average volatility over
NYSE trading hours, with separate end effects for the futures before and after
the cash market opens; see, for example, the related discussion in Andersen
and Bollerslev (1997) pertaining to an earlier sample period.
To account for the time-varying volatility across days, we use the estimated
continuous volatility for the previous day (for the first day in the sample we
use ). Putting all of the pieces together, our time-varying threshold may be
succinctly expressed as t,i = 3 CVt1 T ODi
0.49
n .
C. Estimation of Expected Integrated Volatility
T
Our estimation of Et ( t s2 ds) is based on the following restricted 22-order
VAR,
Xt = A0 + A1 Xt1 + A5
i=1
Xti
%
%
22
5 + A22
Xti 22 + t ,
(A11)
i=1
for the four-dimensional vector

Xt CVt RJVt LJVt ( pt+t pt )2 .
We approximate the multiperiod expected integrated volatility from the sum
of the relevant forecasts for the first element in Xt , which we in turn approximate by easy-to-calculate recursive linear projections using the Kalman filter.
Similar reduced-form forecasting procedures for realized variation measures
2207
Table A.III
Estimation Results for the Restricted VAR(22) Model in (A.11)

The realized variation measures underlying the estimates are based on 5-minute high-frequency
S&P 500 futures data from January 1990 through June 2007. Standard errors are reported in
parentheses.
A0
0.0368 (0.0190)
0.0010 (0.0216)
0.0430 (0.0188)
0.0153 (0.0456)
A1
0.6956 (0.0316)
0.0449 (0.0360)
0.0028 (0.0313)
0.0831 (0.0758)
0.0728 (0.0324)
0.0348 (0.0369)
0.0175 (0.0321)
0.0250 (0.0778)
0.2022 (0.0380)
0.1145 (0.0432)
0.0367 (0.0376)
0.1788 (0.0910)
0.0618 (0.0146)
0.0299 (0.0167)
0.0211 (0.0145)
0.0613 (0.0351)
A5
0.0672 (0.0462)
0.0734 (0.0525)
0.0001 (0.0457)
0.2325 (0.1107)
0.0325 (0.0841)
0.0515 (0.0957)
0.0018 (0.0833)
0.0495 (0.2016)
0.0974 (0.0947)
0.0018 (0.1077)
0.0063 (0.0938)
0.1477 (0.2270)
0.0709 (0.0341)
0.0532 (0.0388)
0.0512 (0.0338)
0.2938 (0.0818)
A22
0.1669 (0.0464)
0.0399 (0.0528)
0.0308 (0.0460)
0.0097 (0.1113)
0.2100 (0.1786)
0.2202 (0.2031)
0.0964 (0.1769)
0.0355 (0.4280)
0.0962 (0.1857)
0.3187 (0.2113)
0.1724 (0.1840)
0.6235 (0.4452)
0.0285 (0.0606)
0.0353 (0.0689)
0.0386 (0.0600)
0.1462 (0.1452)
have been demonstrated to work well empirically by Andersen et al. (2003)

among many others.
The resulting parameter estimates for the VAR are reported in Table A.III.
Not surprisingly, the estimates for the (1, 1) elements in the A1 , A5 , and A22
matrices point to the existence of strong own persistence in the continuous
volatility component. The estimates also reveal significant asymmetries in the
dynamic dependencies between the continuous volatility and both the tails
and the overnight increments, underscoring the importance of including these
variables in the construction of the integrated volatility forecasts.
D. Rank Test
We briefly sketch the basic ideas and calculations required for the rank
tests that form the basis for assessing Propositions 2 and 3; additional details
concerning the implementation of the test are given in Gill and Lewbel (1992).
Let A denote a general p p matrix. The hypothesis of interest may then be
expressed as
H0 : rank(A) r
vs.
H1 : rank(A) > r,
r < p,
where we have an asymptotically normally distributed estimator of A, say
L
A)
Nvec( A
N (0, V ).
Also, let V denote a consistent estimator of V. The rank test is based on LDU
(lower diagonal upper triangular) decomposition with complete pivoting, that
is, the elements of the diagonal matrix are decreasing in absolute value. Toward
2208
this end, write the matrices associated with the LDU decomposition as
A
Q
=L
D
U
,
P
and Q
are permutation matrixes,
where P

1
0
D
L = L11 c 0
, D=
,
21
2
I pr
L
0 D

=
U
11
U
0
12
U
22
U
11 , U
11 , and U
22 are unit lower triangular. In addition, define
and L
1 U
12 U
1
U
'
&
11
22
=
= L
1 L
1 1
.
K
H
22 21 L11 L22 0
1
U
22
Further, let i denote the (p r) (p r) matrix with ones along the ith
diagonal and zeros everywhere else, and define
1
.
=
..
n
and
= ( K
H)(
Q
P)
V ( Q
P
)( K
H
).
W
2 ). The rank test statistic,
Finally, let d 2 = diag( D
1 d 2 ,
= N d 2 W
is then distributed as 2pr under the null hypothesis that rank(A) r.
REFERENCES
At-Sahalia, Yacine, and Andrew Lo, 1998, Nonparametric estimation of state-price densities implicit in financial asset prices, Journal of Finance 53, 499548.
At-Sahalia, Yacine, and Andrew Lo, 2000, Nonparametric risk management and implied risk
aversion, Journal of Econometrics 94, 951.
At-Sahalia, Yacine, Yubo Wang, and Francis Yared, 2001, Do option markets correctly price the
probabilities of movement of the underlying asset?, Journal of Econometrics 102, 67110.
Andersen, Torben, Luca Benzoni, and Jesper Lund, 2002, An empirical investigation of continuoustime equity return models, Journal of Finance 57, 12391284.
Andersen, Torben, and Tim Bollerslev, 1997, Intraday periodicity and volatility persistence in
financial markets, Journal of Empirical Finance 4, 115158.
Andersen, Torben, and Tim Bollerslev, 1998, Answering the skeptics: Yes, standard volatility
models do provide accurate forecasts, International Economic Review 39, 885905.
Andersen, Torben, Tim Bollerslev, and Francis Diebold, 2007, Roughing it up: Disentangling continuous and jump components in measuring, modeling and forecasting asset return volatility,
Review of Economics and Statistics 73, 251277.
2209
Andersen, Torben, Tim Bollerslev, Francis Diebold, and Paul Labys, 2003, Modeling and forecasting
realized volatility, Econometrica 71, 579625.
Andersen, Torben, Tim Bollerslev, and Dobrislav Dobrev, 2007, No-arbitrage semi-martingale
restrictions for continuous-time volatility models subject to leverage effects, jumps and
i.i.d. noise: Theory and testable distributional implications, Journal of Econometrics 138,
125180.
Andersen, Torben, Tim Bollerslev, and Nour Meddahi, 2004, Analytic evaluation of volatility forecasts, International Economic Review 45, 10791110.
Andersen, Torben, and Oleg Bondarenko, 2010, Dissecting the market pricing of volatility, Working
paper, Northwestern University.
Anderson, Evan, Lars Peter Hansen, and Thomas Sargent, 2003, A quartet of semigroups for
model specification, robustness, prices of risk, and model detection, Journal of the European
Economic Association 1, 68123.
Artzner, Philippe, Freddy Delbaen, Jean Marc Eber, and David Heath, 1999, Coherent measures
of risk, Mathematical Finance 9, 203228.
Bakshi, Gurdip, Charles Cao, and Zhiwu Chen, 1997, Empirical performance of alternative option
pricing models, Journal of Finance 52, 20032049.
Bakshi, Gurdip, Nikunj Kapadia, and Dilip Madan, 2003, Stock return characteristics,
skew laws, and the differential pricing of individual equity options, Review of Financial Studies
16, 101143.
Bakshi, Gurdip, and Dilip Madan, 2006, A theory of volatility spreads, Management
Science 52, 19451956.
Bakshi, Gurdip, Dilip Madan, and George Panayotov, 2009, A theory of volatility spreads, Management Science 52, 19451956.
Bansal, Ravi, and Ivan Shaliastovich, 2009, Confidence risk and asset prices, Working paper, Duke
University.
Bansal, Ravi, and Amir Yaron, 2004, Risks for the long run: A potential resolution of asset pricing
puzzles, Journal of Finance 59, 14811509.
Barndorff-Nielsen, Ole, Silja Kinnebrock, and Neil Shephard, 2010, Measuring downside risk
realised semivariance, in T. Bollerslev, J. Russell, and M. Watson, eds. Volatility and Time
Series Econometrics, Essays in Honor of Robert F. Engle (Oxford University Press).
Barndorff-Nielsen, Ole, and Neil Shephard, 2001, Non-Gaussian Ornstein-Uhlenbeck-based models and some of their applications in financial economics, Journal of the Royal Statistical
Society: Series B 63, 167241.
Barndorff-Nielsen, Ole, and Neil Shephard, 2004, Power and bipower variation with stochastic
volatility and jumps, Journal of Financial Econometrics 2, 137.
Barro, Robert, 2006, Rare disasters and asset markets in the twentieth century, Quarterly Journal
of Economics 121, 823866.
Bates, David, 1996, Jumps and stochastic volatility: Exchange rate processes implicit in Deutsche
Mark options, Review of Financial Studies 9, 69107.
Bates, David, 2000, Post-87 crash fears in S&P 500 future options, Journal of Econometrics 94,
181238.
Bates, David, 2008, The market for crash risk, Journal of Economic Dynamics and Control 32,
22912321.
Bates, David, 2008, U. S. stock market crash risk, 19262006, Working paper, University of Iowa.
Bollerslev, Tim, Uta Kretschmer, Christian Pigorsch, and George Tauchen, 2009, A discrete-time
model for daily S&P 500 returns and realized variations: Jumps and leverage effects, Journal
of Econometrics 150, 151166.
Bollerslev, Tim, George Tauchen, and Hao Zhou, 2009, Expected stock returns and variance risk
premia, Review of Financial Studies 22, 44634492.
Bollerslev, Tim, and Viktor Todorov, 2011, Estimation of jump tails, Econometrica, forthcoming.
Bondarenko, Oleg, 2003, Why are put options so expensive?, Working paper, University of Illinois,
Chicago.
2210
Britten-Jones, Mark, and Anthony Neuberger, 2000, Option prices, implied price processes, and
stochastic volatility, Journal of Finance 55, 839866.
Broadie, Mark, Mikhail Chernov, and Michael Johannes, 2009a, Specification and risk premiums:
The information in S&P 500 futures options, Journal of Finance 62, 14531490.
Broadie, Mark, Mikhail Chernov, and Michael Johannes, 2009b, Understanding index options
returns, Review of Financial Studies 22, 44934529.
Campbell, John, and John Cochrane, 1999, By force of habit: A consumption based explanation of
aggregate stock market behavior, Journal of Political Economy 107, 205251.
Carr, Peter, Helyette Geman, Dilip Madan, and Marc Yor, 2003, Stochastic volatility for Levy
processes, Mathematical Finance 13, 345382.
Carr, Peter, and Liuren Wu, 2003, What type of process underlies options? A simple robust test,
Journal of Finance 58, 25812610.
Carr, Peter, and Liuren Wu, 2009, Variance risk premia, Review of Financial Studies 22, 13111341.
Chacko, George, and Luis Viceira, 2005, Dynamic consumption and portfolio choice with stochastic
volatility in incomplete markets, Review of Financial Studies 18, 13691402.
Cochrane, John, 2005, Asset Pricing (Princeton University Press, Princeton, NJ).
Cont, Rama, and Peter Tankov, 2009, Constant proportion portfolio insurance in the presence of
jumps in asset prices, Mathematical Finance 19, 379401.
Corsi, Fulvio, 2009, A simple approximate long-memory model of realized volatility, Journal of
Financial Econometrics 7, 174196.
Delbaen, Freddy, and Walter Schachermayer, 1994, A general version of the fundamental theorem
of asset pricing, Mathematische Annalen 300, 520563.
Drechsler, Itamar, 2011, Uncertainty, time-varying fear, and asset prices, Working paper, New
York University.
Drechsler, Itamar, and Amir Yaron, 2011, Whats Vol got to do with it?, Review of Financial Studies
24, 145.
Duffie, Darrell, 2001, Dynamic Asset Pricing Theory (Princeton University Press, Princeton, NJ).
Duffie, Darrell, Jun Pan, and Ken Singleton, 2000, Transform analysis and asset pricing for affine
jump-diffusions, Econometrica 68, 13431376.
Embrechts, Paul, Claudia Kluppelberg, and Thomas Mikosch, 2001, Modelling Extremal Events
(Springer-Verlag, Berlin).
Eraker, Bjorn, 2004, Do stock prices and volatility jump? Reconciling evidence from spot and option
prices, Journal of Finance 59, 13671403.
Eraker, Bjorn, 2008, The volatility premium, Working paper, University of Wisconsin.
Foresi, Silverio, and Liuren Wu, 2005, Crash-o-phobia: A domestic fear or a worldwide phenonemon?, Journal of Derivatives 9, 821.
Gabaix, Xavier, 2010, Variable rare disasters: An exactly solved framework for ten puzzles in
macro-finance, Working paper, New York University.
Gill, Len, and Arthur Lewbel, 1992, Testing the rank and definiteness of estimated matrices with
applications to factor, state-space and ARMA models, Journal of the American Statistical
Association 87, 766776.
Jacod, Jean, and Albert Shiryaev, 2003, Limit Theorems for Stochastic Processes (Springer-Verlag,
Berlin).
Jiang, George, and Yisong Tian, 2005, The model-free implied volatility and its information content,
Review of Financial Studies 18, 13051342.
Jiang, George, and Yisong Tian, 2007, Extracting model-free volatility from option prices: An
estimation of the VIX index, Journal of Derivatives 14, 126.
Mancini, Cecilia, 2001, Disentangling the jumps of the diffusion in a geometric Brownian motion,
Giornale dellIstituto Italiano degi Attuari 64, 1947.
Pan, Jun, 2002, The jump-risk premia implicit in options: Evidence from an integrated time-series
study, Journal of Financial Economics 63, 350.
Pan, Jun, and Jun Liu, 2003, Dynamic derivative strategies, Journal of Financial Economics 69,
401430.
Pan, Jun, Jun Liu, and Francis Longstaff, 2003, Dynamic asset allocation with event risk, Journal
of Finance 58, 231259.
2211
Pan, Jun, Jun Liu, and Tan Wang, 2005, An equilibrium model of rare-event premia and its
implication for option smirks, Review of Financial Studies 18, 131164.
Pozdnyakov, Vladimir, and J. Michael Steele, 2004, On the martingale framework for futures
prices, Stochastic Processes and their Applications 109, 6977.
Protter, Philip, 2004, Stochastic Integration and Differential Equations (Springer-Verlag, Berlin).
Rietz, Thomas, 1988, The equity risk premium: A solution, Journal of Monetary Economics 22,
117131.
Rosenberg, Joshua, and Robert Engle, 2002, Empirical pricing kernels, Journal of Financial Economics 64, 341372.
Rubinstein, Mark, 1994, Implied binomial trees, Journal of Finance 69, 771817.
Santa-Clara, Pedro, and Shu Yan, 2010, Crashes, volatility and the equity premium: Lessons from
S&P 500 options, Review of Economics and Statistics 92, 435451.
Singleton, Ken, 2006, Empirical Dynamic Asset Pricing (Princeton University Press, Princeton,
NJ).
Sizova, Natalia, 2009, Integrated variance forecasting: Model-based vs. reduced-form, Working
paper, Rice University.
Smith, Richard, 1987, Estimating tails of probability distributions, Annals of Statistics 15,
11741207.
Tankov, Peter, 2009, Pricing and hedging gap risk, Working paper, Ecole Polytechnique.
Todorov, Viktor, 2010, Variance risk premia dynamics: The role of jumps, Review of Financial
Studies 23, 345383.
Todorov, Viktor, and George Tauchen, 2011, Volatility jumps, Journal of Business and Economic
Statistics 29, 356371.
Wachter, Jessica, 2010, Can time-varying risk of rare disasters explain aggregate stock market
volatility?, Working paper, University of Pennsylvania.
Whaley, Robert, 2009, Understanding VIX, Journal of Portfolio Management 35, 98105.
Zhang, Lan, Per Mykland, and Yacine At-Sahalia, 2005, A tale of two time scales: Determining integrated volatility with noisy high-frequency data, Journal of the American Statistical
Association 100, 13941411.

Tails, Fears, and Risk Premia: Tim Bollerslev and Viktor Todorov

Uploaded by

Copyright:

Available Formats

Tails, Fears, and Risk Premia: Tim Bollerslev and Viktor Todorov

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Tails, Fears, and Risk Premia: Tim Bollerslev and Viktor Todorov

Uploaded by

Copyright:

Available Formats

THE JOURNAL OF FINANCE VOL. LXVI, NO.

Tails, Fears, and Risk Premia

So what are policymakers to do? First and foremost, reduce uncertainty.

The Journal of FinanceR

Tails, Fears, and Risk Premia

The Journal of FinanceR

Tails, Fears, and Risk Premia

I. Risk Premia and Jumps

The Journal of FinanceR

Tails, Fears, and Risk Premia

This general expression directly manifests the compensation that is required

The Journal of FinanceR

representing the diffusive and jump risk components, respectively. As

Tails, Fears, and Risk Premia

II. Risk-Neutral Measures

The Journal of FinanceR

while for K < F t ,

where x+ max{0, x}. Importantly, the expressions on the right-hand sides of

Tails, Fears, and Risk Premia

The Journal of FinanceR

while for K 1 0 and K 2 0 with K 1 /K 2 > 1 constant,

Tails, Fears, and Risk Premia

while, for the jump variation,

The Journal of FinanceR

x 2 1{x>0} (ds, dx),

and the variation coming from negative, or left tail, jumps,

x 2 1{x<0} (ds, dx).

It follows, of course, by definition that RVt = CVt + RJVt + LJVt .

B. Expected Jumps and Tails

Tails, Fears, and Risk Premia

where P (x) is a valid Levy density, and

(ek e x )+ P (x) dx.

The Journal of FinanceR

occurrence of jumps and past volatility, as measured directly by the realized

Tails, Fears, and Risk Premia

The Journal of FinanceR

Tails, Fears, and Risk Premia

Mean Jump Intensities

The Journal of FinanceR

Tails, Fears, and Risk Premia

A. Equity Jump Risk Premium

The Journal of FinanceR

This decomposition, depicted in Figure 3, shows that, although the behavior

troughs in both ERP+

Tails, Fears, and Risk Premia

parts associated with negative and positive jumps, say VRP+

Including the more recent period, the median increases to 5.6%.

The Journal of FinanceR

To begin answering these questions, it is instructive to further decompose the

with a similar decomposition available for VRP

Tails, Fears, and Risk Premia

The Journal of FinanceR

Tails, Fears, and Risk Premia

Variance Risk Decomposition

x 2 EPt sP (dx) ds)

x 2 EPt sP (dx) ds)

The Journal of FinanceR

risk. Importantly, however, existing parametric studies generally constrain the

Tails, Fears, and Risk Premia