This action might not be possible to undo. Are you sure you want to continue?

# Electronic copy available at: http://ssrn.

**com/abstract=2013381
**

Working Paper No. 12-10

Fama-Miller Paper Series

A Tale of Two Option Markets: State-Price Densities Implied from

S&P 500 and VIX Option Prices

Zhaogang Song

Federal Reserve Board

Dacheng Xiu

University of Chicago Booth School of Business

This paper also can be downloaded without charge from the

Social Science Research Network Electronic Paper Collection:

http://ssrn.com/abstract=2013381

Electronic copy available at: http://ssrn.com/abstract=2013381

A Tale of Two Option Markets: State-Price Densities Implied from

S&P 500 and VIX Option Prices

∗

Zhaogang Song

†

Federal Reserve Board

Dacheng Xiu

‡

Chicago Booth

This Version: January 2012

Abstract

The S&P 500 and VIX option markets are closely connected as both options depend on the

volatility dynamics. Capturing information in both option prices by nonparametric state-price

densities (SPDs), we look into the dynamics of the index and its volatility, along with interactions

between the two option markets. We ﬁnd that SPDs of the index strongly depend on the current

VIX level, and that such dependence is driven by information implied from VIX options beyond

VIX time series, such as volatility of volatility and volatility skewness. In addition, SPDs of the

VIX document three features of its risk-neutral dynamics, including positive skewness, mean-

reversion, and high persistence. Moreover, the pricing kernel estimates exhibit a U-shape when

the current VIX level is high and a decreasing pattern otherwise, both pre- and post-crisis. This

pattern implies that stochastic volatility may be the missing key state variable for the prefer-

ence of agents, which is responsible for the puzzling U-shape. Finally, we conduct nonparametric

speciﬁcation tests and ﬁnd that the state-of-the-art stochastic volatility models in the literature

cannot capture the S&P 500 and VIX option prices simultaneously. The identiﬁed asymmetric

mean-reversion rate of volatility suggests non-aﬃne speciﬁcation as a necessary extension.

Key Words: Options, Pricing Kernel, State-Price Density, VIX

JEL classiﬁcation: G12,G13

∗

We beneﬁted from discussions with seminar participants at Northwestern University and Princeton University.

This research was funded in part by the Fama-Miller Center for Research in Finance at the University of Chicago

Booth School of Business. The views expressed herein are the authors’ and do not necessarily reﬂect the opinions of

the Board of Governors of the Federal Reserve System.

†

Monetary and Financial Market Analysis Section, Board of Governors of the Federal Reserve System, Mail Stop

165, 20th Street and Constitution Avenue, Washington, DC, 20551. E-mail: Zhaogang.Song@frb.gov.

‡

University of Chicago Booth School of Business, 5807 S. Woodlawn Avenue, Chicago, IL 60637. Email:

dacheng.xiu@chicagobooth.edu.

Electronic copy available at: http://ssrn.com/abstract=2013381

1 Introduction

By reconciling evidence from the S&P 500 index and its options, empirical option pricing literature

has documented stochastic volatility as an indispensable factor of the underlying dynamics of index

returns. Nevertheless, determining the volatility dynamics is not straightforward because, as A¨ıt-

Sahalia et al. (2001) point out, volatility is a nontraded asset, and its risk-neutral behavior cannot

be solely identiﬁed. Therefore, most studies in the literature either rely on strong beliefs in their

favorite parametric models restricted within the aﬃne class, such as Bakshi et al. (1997), Bates

(2000), Pan (2002), Eraker (2004), Broadie et al. (2007), Duan and Yeh (2011), and Amengual and

Xiu (2012), or ignore the unobserved volatility factor when conducting nonparametric analysis of the

index dynamics, as in A¨ıt-Sahalia and Lo (1998), A¨ıt-Sahalia et al. (2001), Buraschi and Jackwerth

(2001), Jackwerth and Rubinstein (1996), and Fan and Mancini (2009).

The lack of observable and tradable volatility has changed substantially since the Chicago Board

of Options Exchange (CBOE) introduced the Volatility Index (VIX) in 1993,

1

and VIX derivatives

such as futures and options in 2004 and 2006, respectively. The VIX provides investors with a

direct measure of the volatility and VIX derivatives oﬀer investors convenient instruments to trade

the volatility of S&P 500 index.

2

The VIX, referred to as the fear gauge, is constantly exposed

in the media spotlight. Also, VIX options achieve huge liquidity, becoming the third most active

contracts at CBOE as of October 2011. Because the VIX is derived from S&P 500 options as the

square root of the risk-neutral expectation of the return’s average variance over the next 30 calendar

days, it uncovers the unobservable volatility in a forward-looking way. In addition, VIX options are

particularly informative about the distribution of the future VIX level. Consequently, two option

markets exist simultaneously that both depend on the volatility dynamics.

Capturing information in both option prices by nonparametric state-price densities (SPDs), we

investigate dynamics of the index and its volatility, as well as the interactions between the two

option markets. These SPDs, also known as the prices of Arrow-Debreu securities, summarize all

the information of the economic equilibrium in an uncertain environment.

3

They not only enable

convenient pricing of vanilla options, but also allow consistent no-arbitrage pricing of exotic and

less-liquid securities such as over-the-counter derivatives. Moreover, distinct from the parametric

1

The VIX, from its inception, was calculated from S&P 500 index options by inverting the Black-Scholes formula.

In 2003, the CBOE amended this approach and adopted a model-free method to calculate the VIX.

2

Before the introduction of VIX derivatives, to trade volatility, investors have to take positions of option portfolios,

such as straddles or strangles.

3

SPDs give for each state x the price of a security paying one dollar if the state falls between x and x + dx. See

A¨ıt-Sahalia and Lo (1998) for detailed discussions.

1

studies in the literature, our nonparametric analysis of SPDs unveils potential misspeciﬁcations of

the parametric forms, and may be advocated as a prerequisite to the construction of parsimonious

models.

Nonparametric studies of SPDs were innovated by A¨ıt-Sahalia and Lo (1998) and Jackwerth

and Rubinstein (1996), disregarding the stochastic volatility factor.

4

Although it may be possible

to estimate SPDs of the S&P 500 index conditional on an ex-post volatility proxy ﬁltered from

historical time series (see Li and Zhao (2009) for such a procedure with interest rate caps), SPDs of the

volatility cannot be determined because options cannot be written on unobservable variables directly.

In contrast, using information in both the S&P 500 and VIX option markets, we recover SPDs of

both the S&P 500 index and VIX and hence indirectly spotlight the unobservable volatility dynamics

without parametric assumptions. Our empirical study of the nonparametric SPDs documents several

important ﬁndings about the dynamics of the index and its volatility and information content of the

two option markets.

First, we ﬁnd that SPDs of the S&P 500 index strongly depend on the current VIX level, not

only in the short run but also for the long term, when VIX options become unavailable. To study the

documented short-run dependence and illustrate how investors employ information in VIX options,

we regress the future average volatility implied from S&P 500 densities on the VIX and moments

of the VIX densities extracted from VIX option prices. We ﬁnd that the short-run dependence is

driven not only by the VIX, but also by information in its SPDs. This ﬁnding implies that the

availability of VIX options indeed deliver incremental information for market participants to pre-

dict future index dynamics and price S&P 500 options accordingly. To understand the documented

long-term dependence, we also implement a simple martingale procedure by which investors interpo-

late implied volatility of today’s long-maturity S&P 500 options using yesterday’s implied volatility

surface. Regression results of implied volatilities of realized option prices on interpolated implied

volatilities show that the martingale interpolation indeed agrees with the market practice and may

oﬀer a potential justiﬁcation for the long-run dependence.

The nonparametric SPDs of A¨ıt-Sahalia and Lo (1998), in the absence of the stochastic volatility

factor, diﬀer considerably from our estimates of SPDs conditional on the VIX when the current

VIX level is either low or high, while only being approximately identical for the average VIX level.

Therefore, with today’s VIX level in the high and low ranges, our nonparametric SPD conditional on

the VIX may provide more-accurate information about the future dynamics of the S&P 500 index.

Second, we document several empirical features of the risk-neutral dynamics of the VIX. In

4

Many follow-up studies have been conducted based on these nonparametric SPDs, such as A¨ıt-Sahalia and Lo

(2000), A¨ıt-Sahalia et al. (2001), Chabi-Yo et al. (2008), and Bakshi et al. (2010)

2

particular, we ﬁnd that SPDs of the VIX are positively skewed across all maturities and current

levels of the VIX. In addition, the risk-neutral dynamics of the VIX exhibit both mean-reversion and

persistence. Although the volatility process under the physical measure is well documented to display

a mean-reverting pattern using historical time series (Mencia and Sentana (2009)), its risk-neutral

behavior is not crystal clear. Our ﬁnding conﬁrms these patterns under the pricing measure without

any parametric restrictions.

Third, we estimate pricing kernels conditional on the VIX, for the periods before and after the

2008 ﬁnancial crisis. We ﬁnd that the estimated pricing kernel exhibits a U-shape conditional on a

high VIX level, while maintaing a decreasing pattern when the current VIX level is low for both pre-

and post-crisis periods. The U-shape of the pricing kernel, described in A¨ıt-Sahalia and Lo (2000),

Jackwerth (2000), and Bakshi et al. (2010) as a puzzle, contradicts standard economic theory, in

which the pricing kernel, equal to the scaled marginal rate of substitution, decreases in market index

return. Our empirical result suggests that the shape of the pricing kernel is signiﬁcantly aﬀected by

the stochastic volatility. The U-shape becomes visible when the market experiences a high volatility.

This ﬁnding echoes the conclusions of equilibrium-based parametric models discussed in Jackwerth

and Brown (2001), Chabi-Yo et al. (2008), Chabi-Yo (2011), and Christoﬀersen et al. (2010a), that

the missing state variables in pricing kernels may result in the U-shape. Without restricting the form

of pricing kernels, our result, however, implies that stochastic volatility is the key but missing state

variable of pricing kernels.

Finally, our SPD estimates enable us to gauge the speciﬁcations of parametric models in pricing

S&P 500 and VIX options simultaneously. Modeling both option prices jointly is important, as they

both rely on the volatility factor: the volatility aﬀects the distribution of the S&P 500 index, while

determining the payoﬀ function of a VIX option. Nevertheless, most studies in the option-pricing

literature treat them separately; they either focus on pricing S&P 500 options (see e.g. Bates (2000),

Pan (2002), Eraker (2004), and Broadie et al. (2007)) or pricing VIX options on top of a standalone

process for VIX (e.g. Whaley (1993), Grunbichler and Longstaﬀ (1996), Detemple and Osakwe

(2000), and Mencia and Sentana (2009)). Although some parametric models have been developed to

capture the evolution of S&P 500 index and VIX, along with joint pricing S&P 500 and VIX options

(Amengual and Xiu (2012)), it remains unclear how well they capture empirical features of these

option prices.

Combining SPDs of both the S&P 500 index and the VIX, we test the speciﬁcation of state-

of-the-art models that resemble those proposed by Pan (2002), Eraker (2004), and Broadie et al.

(2007), but are more ﬂexible in the speciﬁcation of jumps in the return dynamics. Building upon

the diﬀerence between parametric SPD estimates and nonparametric ones, our test statistics reject

3

all these models under consideration. Closer scrutiny of the tests and SPD estimates of the VIX

suggest an asymmetric mean-reversion rate of volatility, i.e., volatility reverts relatively faster when

its current level is low, which further indicates non-aﬃne models may be a promising start point

to jointly pricing the two options. In addition, we also construct test statistics to check whether

the VIX SPD depends on the current information set only through the VIX, for which we ﬁnd no

evidence of rejection. This ﬁnding justiﬁes the VIX to be an eﬀective volatility measure, and helps

simplify our nonparametric analysis.

Our paper is also related to Boes et al. (2007) that tries to estimate the SPDs of returns, with re-

alized volatility constructed from high-frequency data as a proxy for the unobservable spot volatility.

In contrast, our estimation method draws on the observable VIX, and more importantly VIX options,

which are particularly informative about volatility dynamics. Another relevant speciﬁcation analysis

includes Christoﬀersen et al. (2010b), who ﬁnd that non-aﬃne models outperform aﬃne square root

models using realized volatilities, S&P 500 returns, and an extensive panel of option prices. Our

result pinpoints, nonparametrically, the exact inadequacy — constant rate of mean-reversion — of a

variety of aﬃne models. Engulatov et al. (2011), on the other hand, estimate the drift and volatility

functions nonparametrically, and point out similar evidence of asymmetric mean-reversion.

The rest of the paper is organized as follows. In section 2, we discuss the relationship of stochastic

volatility process and VIX, how to obtain SPDs from option prices, and construct our nonparametric

estimators using a multivariate local linear approach. A speciﬁcation test based on the estimated

SPDs is also proposed, and parametric models for pricing S&P 500 options and VIX options con-

sistently are introduced. Section 3 presents our empirical results, including nonparametric SPD

estimates of the S&P 500 index and VIX, analysis of the dependence of S&P 500 index densities on

the current VIX level, estimates of pricing kernels conditional on VIX, and nonparametric speciﬁ-

cation analysis of parametric pricing models. Section 4 concludes. The appendix provides technical

assumptions, proofs, derivations of parametric SPDs, and tests for robustness checks.

2 State-Price Densities: Derivation, Estimation, and Test

2.1 Two Option Markets and State-Price Densities

In this section, we identify the underlying connection between S&P 500 and VIX options and intro-

duce the state-price densities that span the two markets. To ﬁx ideas, we denote the log price of

the S&P 500 index as S

t

, and the VIX as Z

t

. Because the payoﬀs of S&P 500 and VIX options only

depend on their own underlying indices S

T

and Z

T

, the option prices are determined by two SPDs

p(S

T

|F

t

) and p(Z

T

|F

t

) of S

T

and Z

T

, respectively, where F

t

summarizes the information up to time

4

t.

To infer these SPDs from option prices, we denote the instantaneous volatility of S

t

as V

t

, and

write the price of the S&P 500 option with maturity T and strike x as

C(τ, s

t

, v

t

, x) =e

−rτ

E

Q

_

(e

S

T

−x)

+

|S

t

= s

t

, V

t

= v

t

_

=e

−rτ

_

R

(e

s

T

−x)

+

p(s

T

|τ, s

t

, v

t

)ds

T

where p(s

T

|τ, s

t

, v

t

) is the SPD of S

T

, τ = T −t is the time-to-maturity, and r is the constant risk-free

rate between t and T

5

. Building upon the insight of Breeden and Litzenberger (1978), the SPD of

S

T

can be recovered from the second order derivative of C(τ, s

t

, v

t

, x) with respect to x,

p(s

T

|τ, s

t

, v

t

) = e

rτ+s

T

∂

2

C(τ, s

t

, v

t

, x)

∂x

2

¸

¸

¸

x=e

s

T

. (1)

By the same reasoning and assumption, we write the VIX option price H as a function of s

t

, v

t

,

time-to-maturity τ and strike y, i.e., H(τ, s

t

, v

t

, y), and obtain the SPD of Z

T

as

p(z

T

|τ, s

t

, v

t

) = e

rτ

∂

2

H(τ, s

t

, v

t

, y)

∂y

2

¸

¸

¸

y=z

T

. (2)

Note that the derived SPDs of S and Z in (1) and (2) extract all the information in option prices

that can be used to identify the risk neutral measure Q, and the dynamics of S and V .

Nevertheless, these two densities p(s

T

|τ, s

t

, v

t

) and p(z

T

|τ, s

t

, v

t

) are of little use in studying

the two option markets, as the information regarding V

t

is unobservable. Considering the functional

relationship between Z

t

and V

t

,

6

we may rewrite the option prices as a function of z

t

, i.e., C(τ, s

t

, z

t

, x)

and H(τ, s

t

, z

t

, y),

7

and take second order derivatives to obtain

p(s

T

|τ, s

t

, z

t

) =e

rτ+s

T

∂

2

C(τ, s

t

, z

t

, x)

∂x

2

¸

¸

¸

x=e

s

T

p(z

T

|τ, s

t

, z

t

) =e

rτ

∂

2

H(τ, s

t

, z

t

, y)

∂y

2

¸

¸

¸

y=z

T

5

We treat r as a constant here following our empirical setup, which seems valid for the sample period we consider;

see Section 3 for details.

6

The CBOE constructs Zt form a portfolio of options weighted by strikes according to the formula:

(Zt/100)

2

= E

Q

(QVt,T |Ft) =

2e

rτ

τ

F

t

0

P(τ, x)

x

2

dx +

∞

F

t

C(τ, x)

x

2

dx

+

where QVt,T denotes the quadratic variation of the log return process, P(τ, x) and C(τ, x) are put and call options with

time-to-maturity τ and strike x, and Ft = e

rτ+s

t

is the price of futures contracts, see e.g. Britten-Jones and Neuberger

(2000) and Carr and Wu (2009).

7

Strictly speaking, the function C (·) here is a composite function, which is diﬀerent from the previous call option

pricing function. We recycle it to simplify our notations.

5

with an information set that is fully observable. In fact, representing options in terms of S

t

and Z

t

amounts to conditioning on a coarser information set generated by S

t

and Z

t

, which is a subset of

the original information set generated by S

t

and V

t

. Equating the two information sets assumes that

Z

t

, as a function of V

t

and S

t

, is invertible with respect to V

t

.

Such an approximation is justiﬁable for a few reasons. First, VIX is always exposed in the media

spotlight, and used by investors as a measure of market volatility. Second, a simple regression of the

realized volatility (RV) against the lagged VIX and lagged RV indicates that the VIX has far more

explanatory power compared with RV, implying the importance of the VIX in capturing dynamics

of the unobserved volatility (see, e.g., Blair et al. (2001)). Third, although the VIX diﬀers from

unobserved volatility as being forward-looking because it is constructed using S&P 500 options with

maturities up to 30 calendar days, the approximation errors are expected to become smaller and even

negligible as time-to-maturity increases.

8

Finally, for speciﬁcation analysis of parametric models, the

approximation error is zero for most models in the literature as Z

t

is a deterministic and invertible

function of V

t

; see Section 2.5 for details.

In summary, state-price densities p(s

T

|τ, s

t

, z

t

) and p(z

T

|τ, s

t

, z

t

) encapsulate all the price infor-

mation in the two option markets. They complement each other to reveal an intact picture of the

index and volatility dynamics and interactions of the two markets.

2.2 Multivariate Local Linear Estimators

Here we introduce our nonparametric estimation strategies to estimate the SPDs. To ﬁx ideas, we

assume the observed prices

˜

C and

˜

H are contaminated with observation errors, such that

C(τ, s

t

, z

t

, x) = E

_

˜

C

¸

¸

¸˜ τ = τ, S

t

= s

t

, Z

t

= z

t

, X = x

_

H(τ, s

t

, z

t

, y) = E

_

˜

H

¸

¸

¸˜ τ = τ, S

t

= s

t

, Z

t

= z

t

, Y = y

_

.

We then construct nonparametric estimators of C and H, and take derivatives to estimate the SPDs.

Diﬀerent from the multivariate kernel regression approach adopted by A¨ıt-Sahalia and Lo (1998),

we use the multivariate local linear method, which enjoys better boundary performance and reduces

the asymptotic bias without increasing asymptotic variance (Fan and Gijbels (1996)). Moreover,

local linear regression provides a closed-form estimator for both the price function and its ﬁrst order

derivatives, so that estimators for SPDs can be obtained simply by a ﬁrst-order diﬀerentiation with

respect to the strike.

8

Due to this forward-looking feature, we have used S&P 500 option prices with maturities up to 30 calendar days

when changing the information set to that generated by St and Zt. Therefore, caution is needed when employing

p(sT |τ, st, zt) in practice with τ less than 30 calendar days.

6

To estimate the pricing function C as a function of u = (τ, s, z, x)

**, we consider the following
**

minimization problem,

min

α,β

n

i=1

_

C

i

−α −β

(U

i

−u)

_

2

K

h

(U

i

−u)

where U

i

= (τ

i

, S

t

i

, Z

t

i

, X

i

)

and C

i

are the characteristics and price of the i−th option in sample.

K

h

is a kernel function scaled by a bandwidth vector h = (h

τ

, h

s

, h

z

, h

x

):

K

h

(U

i

−u) =

1

h

τ

k

_

τ

i

−τ

h

τ

_

1

h

s

k

_

S

t

i

−s

h

s

_

1

h

z

k

_

Z

t

i

−z

h

z

_

1

h

x

k

_

X

i

−x

h

x

_

(3)

where k (·) is, for example, the density of standard normal distribution. The minimizer has a closed-

form representation:

_

_

´ α

´

β

_

_

(1+4)×1

=

_

Ω

KΩ

_

−1

Ω

KC (4)

where

Ω =

_

¸

¸

¸

_

1

.

.

.

1

(U

1

−u)

.

.

.

(U

n

−u)

_

¸

¸

¸

_

, C =

_

¸

¸

¸

_

C

1

.

.

.

C

n

_

¸

¸

¸

_

, K =

_

¸

¸

¸

_

K

h

(U

1

−u)

.

.

.

K

h

(U

n

−u)

_

¸

¸

¸

_

.

The nonparametric local linear estimator for the option pricing function C(τ, s, z, x) is

´

C(τ, s, z, x) = ´ α = e

1

_

Ω

KΩ

_

−1

Ω

KC,

with e

1

= (1, 0, 0, 0)

and the estimator ´ p(s

**|τ, s, z) for the SPD of S
**

t

is

´ p(s

|τ, s, z) = e

rτ+s

∂

´

β

4

∂x

¸

¸

¸

x=e

s

= e

rτ+s

∂

_

e

4

(Ω

KΩ)

−1

Ω

KC

_

∂x

¸

¸

¸

x=e

s

. (5)

where e

4

= (0, 0, 0, 1)

**. The nonparametric estimator
**

´

H (·) and ´ p(z

|τ, s, z) can be constructed

similarly.

In our empirical implementations, we choose the Gaussian kernel as k (·). Moreover, we set

the bandwidth h

j

(j = τ, s, z, and x) as h

j

= c

j

s

_

U

j

_

n

−1/6

, where s

_

U

j

_

is the unconditional

standard deviation of the regressor U

j

(j = τ, s, z, and x). This bandwidth choice ensures that

the nonparametric pricing function achieves the optimal rate of convergence in the mean-squared

sense among all possible nonparametric estimators for option prices. However, this rate leads to an

asymptotic bias term for the estimator

´

C (·) and ´ p (·), as shown by Fan and Gijbels (1996). Hence,

we follow A¨ıt-Sahalia and Lo (1998) to select c

j

= c

j0

/ ln (n), which results in a slightly slower rate

of convergence, but centers the asymptotic distribution at zero. Similar to Li and Zhao (2009),

7

the constant c

j0

is chosen by minimizing the ﬁnite-sample mean-squared error of the estimator via

simulations.

Suppose the sample sizes of the nonparametric SPD estimates of the S&P 500 index and VIX

are n

C

and n

H

, respectively. Using the equivalent kernels introduced in Fan and Gijbels (1996) and

following the derivation in A¨ıt-Sahalia and Lo (1998), we obtain the asymptotic distributions of these

estimators:

n

1/2

C

h

x

(h

τ

h

s

h

z

h

x

)

1/2

_

´ p(s

|τ, s, z) −p(s

|τ, s, z)

_

(6)

d

−→N

_

0, e

2rτ

__

k

2

(c) dc

_

3

__

_

c

˙

k (c) +k(c)

_

2

dc

_

/

__

k (c) c

2

dc

_

2

s

2

C

(τ, s, z, x) /π

C

(τ, s, z, x)

_

,

n

1/2

H

h

y

_

h

τ

h

s

h

z

h

y

_

1/2

_

´ p(z

|τ, s, z) −p(z

|τ, s, z)

_

d

−→N

_

0, e

2rτ

__

k

2

(c) dc

_

3

__

_

c

˙

k (c) +k(c)

_

2

dc

_

/

__

k (c) c

2

dc

_

2

s

2

H

(τ, s, z, y) /π

H

(τ, s, z, y)

_

,

where s

2

C

(τ, s, z, x) and s

2

H

(τ, s, z, y) are conditional variances for the local linear regressions of C

and H on their state variables respectively, and π

C

(τ, s, z, x) and π

H

(τ, s, z, y) are marginal densities

of these variables. The estimators ´ s

2

C

(·) and ´ s

2

H

(·) for s

2

C

(·) and s

2

H

(·) can be constructed using

similar nonparametric regressions.

2.3 Dimension Reduction

One of the major issues of nonparametric estimation is the curse of dimensionality. The rate of

convergence decreases rapidly as the dimension of state variables increases. In the most general

forms, the pricing functions C (·) and H (·) depend not only on time-to-maturity, strike, VIX, and

the S&P 500 index, but also on interest rates and dividends. Because interest rates and dividends

do not vary much within the period of our empirical studies (see Section 3 for details), we take them

as constants.

Furthermore, following many existing studies such as Li and Zhao (2009), we assume the S&P

500 option price is homogeneous of degree one in the current price level:

C(τ, s, z, x) = e

s

C(τ, 0, z, x/e

s

) = e

s

¯

C(τ, z, m)

where m = x/e

s

represents the moneyness of the S&P 500 option. Consequently, we obtain the

estimate of C(τ, s, z, x) through multiplying the nonparametric estimate of

¯

C (·) by e

s

, and write the

SPD of S

T

as

p(s

T

|τ, s

t

, z

t

) = e

rτ+s

T

−st

∂

2

¯

C(τ, z

t

, m)

∂m

2

¸

¸

¸

m=e

s

T /e

s

t

.

8

As for VIX options, we assume that the information about Z

t

in S

t

is fully incorporated into Z

t

.

In other words, conditional on Z

t

, Z

t

is independent of S

t

, for any t

**> t. This assumption further
**

implies that the SPD of Z

T

, obtained from the VIX option prices, depends on S

t

only through Z

t

,

i.e., p(z

T

|τ, s

t

, z

t

) = p(z

T

|τ, z

t

). Thus, the number of state variables for the SPD of VIX is also

decreased by one. We formally test these two assumptions in Section 3, and the results suggest that

they are not rejected for the sample we consider.

2.4 Nonparametric Estimation of Pricing Kernels

In the equilibrium model of A¨ıt-Sahalia and Lo (2000) with a representative agent, the ratio between

risk-neutral and physical densities, i.e., the pricing kernel, is—up to a scaled factor—the marginal

rate of substitution, which also reﬂects the market’s aggregation of risk preferences implicit in option

prices. In an economy with two state variables, the pricing kernel is supposed to be π(s

T

, z

T

|τ, s

t

, z

t

),

which cannot be identiﬁed nonparametrically. Nevertheless, we may consider an integrated pricing

kernel π(s

T

|τ, s

t

, z

t

) as

π(s

T

|τ, s

t

, z

t

) = E

S

T

(π(s

T

, z

T

|τ, s

t

, z

t

)|s

t

, z

t

) =

p(s

T

|τ, s

t

, z

t

)

¯ p(s

T

|τ, s

t

, z

t

)

where ¯ p(s

T

|τ, s

t

, z

t

) is the conditional physical density of S

T

, p(s

T

|τ, s

t

, z

t

) is the conditional risk

neutral density, and E

S

T

(·|s

t

, z

t

) denotes the projection onto S

T

. Distinct from A¨ıt-Sahalia and Lo

(2000) and Jackwerth (2000), our pricing kernel π(·|·) is conditional on the VIX so that it can reﬂect

the risk preferences with respect to the volatility.

While estimating the risk neutral density p(s

T

|τ, s

t

, z

t

) from option prices, we estimate ¯ p(s

T

|τ, s

t

, z

t

)

using the time series of the S&P 500 index and VIX based on the kernel method. Suppose we have

time series {S

T

i

, S

t

i

, Z

t

i

}

n

i=1

, with T

i

− t

i

= T − t ﬁxed. We construct the estimator of the density

˜ p(s

|τ, s, z) as:

´

¯ p(s

|τ, s, z) =

n

i=1

1

b

s

k

_

S

T

i

−s

b

s

_

1

b

s

k

_

S

t

i

−s

b

s

_

1

b

z

k

_

Z

t

i

−z

b

z

_

n

i=1

1

b

s

k

_

S

t

i

−s

b

s

_

1

b

z

k

_

Z

t

i

−z

b

z

_

with bandwidths given by the following:

b

j

= c

j

σ

j

n

−1/(4+d)

, j = s

, s, z

where σ

j

is the unconditional standard deviation of the data, c

j

is a constant, and d represents the

dimension: d = 3 for those b

j

s that appear in the numerator, and d = 2 for those in the denominator.

9

Consequently, our pricing kernel can be estimated by

´ π(s

|τ, s, z) =

´ p(s

|τ, s, z)

´

¯ p(s

|τ, s, z)

.

As discussed in A¨ıt-Sahalia and Lo (2000), the convergence rate is higher for

´

¯ p(s

|τ, s, z) than for

´ p(s

|τ, s, z), and the asymptotic distribution of ´ π(s

|τ, s, z) is identical to that of ´ p(s

|τ, s, z)/¯ p(s

|τ, s, z),

so that the asymptotic variance can be easily estimated.

2.5 Speciﬁcation Tests Based on State-Price Densities

The estimated SPDs enable us to evaluate the performance of the state-of-the-art parametric models

under the uniﬁed framework of pricing S&P 500 and VIX options. We bring together and test a

variety of stochastic volatility models, including those discussed in Pan (2002), Eraker (2004), and

Broadie et al. (2007), which are shown to be successful in capturing empirical features of S&P 500

options such as volatility smile, smirk, and index option returns. It is important to check whether

they can capture the empirical features of VIX options before employing them in practice, as the two

option markets are highly integrated. However, few paper in the literature have discussed modeling

them jointly. Our SPDs can uncover two distinct aspects— returns and volatility—of the parametric

model dynamics, and hence may shed light on the inadequacy of the model. In particular, comparing

our nonparametric SPD estimates with the parametric ones illustrates whether the rejection is mainly

due to modeling S&P 500 options, VIX options, or both, and which characteristics we should focus

on in order to improve the model speciﬁcation.

The general model we consider is speciﬁed under the risk-neutral measure Q as

dS

t

= (r −d −

1

2

V

t

)dt +

_

V

t

dW

Q

t

+J

Q

S

dN

t

−µλ

t

dt

dV

t

= κ(ξ −V

t

)dt +σ

_

V

t

dB

Q

t

+J

Q

V

dN

t

(7)

where W

Q

t

and B

Q

t

are standard Brownian motions satisfying E(dW

Q

t

dB

Q

t

) = ρdt, J

Q

S

and J

Q

V

are

random jump sizes, dN

t

is a pure-jump process with intensity λ

t

= λ

0

+λ

1

V

t

, and µ = E(e

J

Q

S

−1).

The benchmark model is the Heston stochastic volatility model (SV) without jumps in either

returns or volatility, i.e., λ

0

= λ

1

= 0. To add jumps in returns, we consider two types of speciﬁcation:

SVJ1: J

Q

S

∼ N(µ

S

, σ

S

); or SVJ2: J

Q

S

∼

_

_

_

exp(β

+

) with probability q

− exp(β

−

) with probability 1 −q

where the jump intensity is constant λ

0

, i.e. λ

1

= 0. The last two models incorporate the same

exponentially distributed jumps in volatility with mean β

V

:

J

Q

V

∼ exp(β

V

),

10

and we label them as SVCJ1 and SVCJ2, due to their diﬀerences in jump speciﬁcation for returns.

The intensity in these cases is linear in V

t

. Compared with SVJ1 and SVCJ1, the jump speciﬁcations

of return dynamics in model SVJ2 and SVCJ2 can incorporate heavy-tailed and skewed jumps.

For each model, the theoretical value of the VIX Z

t

can be calculated, as shown by Britten-Jones

and Neuberger (2000) and Carr and Wu (2009):

Z

2

t

= 10

4

·

1

τ

E

Q

t

_

_

t+τ

t

V

s

ds +

Ns≥0, t≤s≤t+τ

(J

Q

S

)

2

_

= aV

t

+b (8)

where a and b are constants determined by the speciﬁc model; see Appendix B for the formulae of

a and b and details of derivation. On top of (8), we derive VIX option prices in addition to S&P

500 option prices.

9

These formulae regarding the VIX and its option prices can greatly simplify

the parametric inference, as the unobservable volatility factor can be replaced by a function of the

VIX.

10

The dynamics of (S

t

, V

t

), combined with (8), imply the dynamics of (S

t

, Z

t

) for which we per-

form our nonparametric speciﬁcation analysis of the model (7). In particular, model (7) speciﬁes

parametric forms of the SPDs of S&P 500 index and VIX, i.e., p(s

T

|τ, s

t

, z

t

;θ) and p(z

T

|τ, z

t

;θ),

respectively (see Appendix B for details of derivation). The idea of our nonparametric analysis is

to compare these parametric SPDs with nonparametric estimates ´ p(s

T

|τ, s

t

, z

t

) and ´ p(z

T

|τ, z

t

) and

check whether the diﬀerences are statistically signiﬁcant.

Recall that we employ two dimension-reduction techniques in obtaining p(s

T

|τ, s

t

, z

t

) and p(z

T

|τ, z

t

),

i.e., homogeneous of degree 1 in S

t

for S&P 500 options and the dependence of VIX densities on S

t

only through Z

t

, both of which are satisﬁed for the parametric model (7). Based on these simpliﬁ-

cations, we propose to test the joint null hypothesis H

0

,

H

0

: Pr {p(m

T

|¯τ, Z

t

; θ

0

) = p(m

T

|¯τ, Z

t

), p(Z

T

|¯τ, Z

t

; θ

0

) = p(Z

T

|¯τ, Z

t

)} = 1

against the alternative hypothesis H

A

,

H

A

: Pr {p(m

T

|¯τ, Z

t

; θ

0

) = p(m

T

|¯τ, Z

t

), p(Z

T

|¯τ, Z

t

; θ

0

) = p(Z

T

|¯τ, Z

t

)} < 1,

where m

T

= e

S

T

/e

St

. If the null hypothesis H

0

is rejected, we test the following two separate

speciﬁcation hypotheses,

H

0,S&P

: Pr {p(m

T

|¯τ, Z

t

; θ

0

) = p(m

T

|¯τ, Z

t

)} = 1,

9

The closed-form pricing formulae have been derived by Amengual and Xiu (2012) for more-general models, see also

Sepp (2008) for derivation under the Heston model.

10

The CBOE VIX contains numerical errors due to jumps and discretization and hence may not equal to the

theoretical VIX in (8) exactly. However, these approximation errors are negligible under commonly used models and

model parameters, as shown by Jiang and Tian (2005) and Carr and Wu (2009).

11

H

0,VIX

: Pr {p(Z

T

|¯τ, Z

t

; θ

0

) = p(Z

T

|¯τ, Z

t

)} = 1,

i.e., whether the model is consistent with S&P 500 option prices and VIX option prices separately.

The test statistics based on the diﬀerences between parametric and nonparametric SPDs are con-

structed and converge to convenient standard normal distributions under the null hypothesis; see

Appendix A for details.

3 Empirical Results

In this section, we estimate nonparametric SPDs implicit in the S&P 500 and VIX options, along

with recovering their interactions. Before delving into the details, we introduce the dataset.

3.1 Data

We collect close prices of S&P 500 Index, VIX, and their options from the OptionMetrics group for

the year 2009. The best bid and oﬀer prices for options are quoted between 3:59 p.m. and 4:00

p.m. EST. Figure 1 plots the joint time series of the S&P 500 index and the VIX, Table 1 provides

their summary statistics, and Table 3 presents summary statistics of the option prices. We follow

the data-cleaning routine commonly used in literature, see, e.g., A¨ıt-Sahalia and Lo (1998). First,

observations with bid or ask prices smaller than 0.05 are eliminated to mitigate the eﬀect of price

discreteness. For each option, we take the midquote as the option price. Due to liquidity concerns,

we eliminate any options with zero open interests or trading volumes, as well as options with time-

to-maturity of less than 10 days. Also, it is clear from Table 2, that in-the-money S&P 500 options

are less liquid than out-of-the-money options. Therefore, we delete in-the-money options, and use

the put-call parity to construct in-the-money call options from out-of-the-money put options. There

is no such pattern for VIX options and hence we only consider VIX call options. The last step is to

eliminate option contracts that violate no-arbitrage conditions. The resulting sample covers a broad

cross section of options, including 128,883 S&P 500 call options, and 14,539 VIX call options.

3.2 Monte Carlo Simulations

We provide simulation evidence for our local linear estimators. The Monte Carlo experiments are

designed to match our empirical studies. First, we select the same option characteristics as those

traded on CBOE in 2009. Second, we ﬁx a sample path generated by the SVCJ2 model. We then

calculate option prices, according to the closed-form formulae given in the Appendix. Finally, we

12

pollute the prices with multiplicative measurement error following log-normal distribution with 1%

standard deviation. Option prices below 0.05 are eliminated to mimic the reality.

Based on the generated sample, we evaluate our nonparametric estimators of index option pricing

functions on the grid of time-to-maturity and current index level, with Z

t

and X ﬁxed at their sample

averages. We also calculate the index densities on the grid of τ and S

T

, with S

t

ﬁxed at the sample

mean, to evaluate our density estimators. The nonparametric estimators of VIX option prices and

densities are evaluated similarly. All of these quantities and their percentage errors are reported in

Figures 2 and 3, averaged over 500 replications. These ﬁgures show that the nonparametric estimates

are within 2% and 5% of their theoretical values for S&P 500 and VIX options, respectively. The

errors for densities are slightly larger, as also shown in both Figures 2 and 3, due to the fact that

derivatives are estimated with slower rates of convergence, the so-called curse of diﬀerentiation.

3.3 State-Price Densities of S&P 500 Index and VIX

3.3.1 Empirical Characteristics of State-Price Densities

Figure 4 presents nonparametric SPD estimates of the S&P 500 index for both low and high levels

of the VIX with time-to-maturity equal to 21, 42, and 84 days. The low and high levels of the

VIX are obtained as the 20% and 80% quantiles, respectively, of the VIX in our sample. The 95%

conﬁdence intervals are obtained through the asymptotic theory given in (6). As expected, Figure 4

shows that the S&P 500 index densities are negatively skewed across all maturities and VIX values.

More importantly, the index densities strongly depend on the VIX Z

t

. Take τ = 21 as an example.

When Z

t

is low, the index density p(s

T

|τ, s

t

, z

t

) has pronounced spikes; as Z

t

rises to the high level,

the density becomes more dispersed and more negatively skewed. This pattern becomes stronger for

longer maturities.

In Figure 5, we benchmark the nonparametric SPD estimates proposed by A¨ıt-Sahalia and Lo

(1998) (AL) that neglect the volatility variable, and compare it with our density estimates conditional

on diﬀerent VIX levels. It is evident that the AL density estimate departs from ours substantially

when Z

t

stays at high or low levels: our estimate is more compact with higher spikes for a low Z

t

but more dispersed with heavier tails for a high Z

t

. In contrast, the two densities are very close

when Z

t

is at the medium level, indicating that the AL density estimator actually approximates our

density estimator for medium VIX levels. Hence, our nonparametric SPD estimator may be more

informative about the future dynamics of the S&P 500 index than the AL density.

Figure 6 provides nonparametric estimates of the VIX SPDs for diﬀerent levels of Z

t

with τ equal

to 21, 42, and 84 days. We ﬁnd three important features of the risk-neutral densities of the VIX.

13

First, the VIX densities are positively skewed across all maturities and current VIX levels. Second,

the risk-neutral dynamics of the VIX exhibit a mean-reverting pattern, which is most apparent when

Z

t

is at the high level. Speciﬁcally, the left tail of the VIX SPD becomes higher as maturity increases,

showing that when Z

t

is high, the probability of Z

T

mean-reverts to the medium level increases with

maturity. This mean-reverting behavior also shows up when Z

t

is at the low level, although less

strongly, as evidenced by the lower peaks at the left end of the VIX densities for longer maturities.

Third, the risk-neutral dynamics of the VIX exhibit high persistence. Conditional on a low Z

t

,

the shape of the VIX density changes slowly as maturity increases. This change implies that the

probability of Z

T

to stay at low levels conditional on a low Z

t

is high. Even conditional on a high Z

t

level when the mean reversion eﬀect is strong, Z

T

can maintain at high levels with large probability

as maturity increases from 21 to 42 days. Moreover, for any ﬁxed maturity, the VIX density is more

compact (dispersed), conditional on a low (high) level of the current VIX.

3.3.2 Information from VIX Options about S&P 500

As shown in the last section, SPDs of the S&P 500 index strongly depend on Z

t

at maturities of

21, 42, and 84 days, for which VIX options are available. Such short-run dependence can be driven

through two channels. First, investors may employ historical time series of the VIX up to time t to

predict the volatility of the future index S

T

. Second, investors may use VIX option prices to obtain

information about the distribution of the future VIX Z

T

and form expectations of the volatility of

S

T

. We now study whether VIX options contain information about market future volatility beyond

VIX time series for the future volatility of index dynamics.

We compute the average volatility over some future 30-calendar-day interval [T, T + 21] by,

V

S

t,T

=

_

Var

t

[S

T+21

] −Var

t

[S

T

]

where Var

t

[S

T

] is the conditional variance of S

T

implied from the density estimate p(s

T

|τ, s

t

, z

t

). We

then regress V

S

t,T

on both Z

t

and characteristics of the VIX density, corresponding to information

in VIX time series and VIX options, respectively. The characteristics of VIX densities we employ

are the conditional standard deviation Std

t

[Z

T

] and conditional skewness Skew

t

[Z

T

] of p(z

T

|τ, z

t

),

which proxy the volatility of volatility and volatility skewness implicit in VIX options. All of the

quantities are annualized and estimated using nonparametric SPDs.

For τ = 21, 42, 63 and 84, we run the following regression

V

S

t,t+τ

= α +β

1

Z

t

+β

2

Std

t

[Z

t+τ

] +β

3

Skew

t

[Z

t+τ

] +ε

t

, t = 1, 2, . . . , 252 (9)

and the results are presented in Table 3. We ﬁnd that the VIX level Z

t

has predictability for the S&P

500 option-implied future volatility at τ = 21 and 84, but is less signiﬁcant with τ = 42 and 63 days.

14

Most importantly, both the volatility of volatility Std

t

[Z

t+τ

] and volatility skewness Skew

t

[Z

t+τ

]

are signiﬁcant in explaining V

S

t,t+τ

, with the former more pronounced for maturities of 63 and 84

days and the latter more evident for all maturities except for 84 days. For example, Panels C and

D show that adding the volatility of volatility Std

t

[Z

t+τ

] to the regression with single explanatory

variable Z

t

increases the adjusted R

2

from less than 1% to over 12%. Moreover, Panels A, B, and C

show that adding volatility skewness Skew

t

[Z

t+τ

] to the regression against Z

t

and Std

t

[Z

t+τ

] leads

to increases in the adjusted R

2

, by 20% for maturities equal to 21 and 63 days and 300% for 42 days.

In summary, we ﬁnd that the dependence of index densities on the VIX is driven not only by

the VIX time series, but also by the volatility of volatility and volatility skewness that can only be

extracted from VIX option prices. That is to say, investors consider both the VIX level Z

t

and the

whole distribution of volatility when forming expectations of future volatility and pricing S&P 500

options. Hence, the availability of VIX options indeed delivers incremental information for market

participants beyond the VIX time series.

3.3.3 In-Sample and Out-of-Sample Forecasts

We benchmark our nonparametric estimator (SX) against two alternative methods discussed in A¨ıt-

Sahalia and Lo (1998): the nonparametric approach without volatility state (AL) for both densities

and option prices, and the martingale approach (MKT) for option prices, which interpolates tomor-

row’s implied volatility using today’s implied volatility surface.

Intuitively, one of the potential advantages of the SX and AL approaches lies in their inclusion of

historical prices of options with similar characteristics into the weighted prediction of option prices.

As opposed to the MKT approach that replies exclusively on the cross section of options at each

date, both SX and AL tend to maintain more-stable pricing structures over time, and hence they are

expected to outperform the MKT approach in out-of-sample forecasting, although they may ﬁt the

cross section of option prices poorly on certain days, whereas perfectly for some other days in-sample.

In terms of the comparison between SX and AL, the SX approach further explores the predictability

from the VIX, hence the SX is expected to outperform the AL in- and out-of-sample for both option

prices and densities.

Panel A of Table 4 reports the forecasting performance of the SX and AL density, whereas Panel

B presents the price prediction comparisons including, additionally, the MKT approach. For each

date t, we adopt a preceding six-month window, within which options are selected to predict the

prices of traded options with strikes equal to 900, 925, 950, and 975 (the four strikes around the

median of the sample) and maturities closest to four months, along with the SPDs, on day t + γ,

for γ = 0 (in-sample), and γ = 7, 14, 21, 28, 35, and 42 days (out-of-sample) progressively. We repeat

15

the procedure for each day t from the second half of the year, and aggregate across t to obtain

the root-mean-squared percentage diﬀerence between the predictions and the values realized on day

t + γ. The realized density on day t + γ is proxied by either the AL or the SX estimator, whereas

the observed market prices are taken to be the realized option prices.

Regarding density forecasts, the AL densities produce a better ﬁt in-sample and over short

horizons (for γ up to 14 days), but only when the realized density is represented by the AL method

(the left part of Panel A). In contrast, the SX estimates are uniformly better over every horizon

in predicting the future density when the realized density is represented by the SX density (the

right part of Panel A), but also when it is proxied by the AL density for γ longer than 21 days.

For market prices, the AL estimates are uniformly outperformed by our estimators at every horizon

except at τ = 14 days. Moreover, the simple martingale interpolation leads to a better ﬁt in-sample

and for less than a month (τ = 0, 7, 14, and 21 days), whereas over longer forecasting horizons, it is

outperformed by the SX densities.

For forecasting option prices, Panel B shows that the MKT approach outperforms the AL ap-

proach uniformly in forecasting prices for the sample period we consider, which is in contrast with

ﬁndings of A¨ıt-Sahalia and Lo (1998), at least for longer horizons. However, this ﬁnding is not

puzzling as the AL estimator behaves as an estimator conditional on the average volatility level and

performs better when the volatility stays around the mean level. In our sample period, the VIX

decreases monotonically and stays at lower levels for the entire second half of the year. Therefore,

the uniform better performance of the MKT over the AL method may be caused by the decreasing

volatility path in 2009, as the assumption of stability across time deteriorates the performance. This

performance actually highlights, indirectly, the importance of including volatility in the prediction

from the fact that the SX outperforms the MKT in the long run, as shown by Panel B.

3.4 Pricing Kernels Conditional on VIX

In addition to SPDs under the risk-neutral measure, we also estimate the physical densities of the

S&P 500 index, and hence obtain estimates of the pricing kernels conditional on VIX by combining

the SPDs under both measures.

To highlight the potential change of average risk preference due to recent ﬁnancial crisis, we

compare the empirical characteristics of the pricing kernels over two diﬀerent periods: the pre-

crisis

11

period from June 1, 2006 to November 30, 2007, and the post-crisis period from June 1, 2009

to November 30, 2010. Moreover, we focus on a 2-month (τ = 42 days) horizon. Figure 7 reports the

11

We follow the NBER Business Cycle Committee and regard the 18-month duration of 2008 ﬁnancial crisis as from

Nov 1, 2007 to May 31, 2009. Then we chose 18-month periods before and after the crisis, respectively.

16

pricing kernel estimates, for both pre-crisis (in the upper panel) and post-crisis (in the lower panel)

at diﬀerent VIX levels. The low and high levels of Z

t

before crisis, are chosen as 11.56 and 20.03,

whereas the low and high VIX levels post-crisis are taken as 20.03 and 35.455, respectively. Note

that the low level of VIX post-crisis is the same as the high level pre-crisis, as the VIX stays in a

high regime after crisis.

We observe that the pre-crisis pricing kernels, along with the post-crisis pricing kernel conditional

on a low level, Z

t

= 20.03, decrease as market index return increases. This decrease is consistent

with standard economic theory that the pricing kernel decreases when wealth rises, as implied by a

decreasing risk aversion. In contrast, the pricing kernel conditional on a high VIX level, Z

t

=35.455,

in the post-crisis period exhibits a pronounced U-shape. Therefore, our empirical result suggests

that the shape of the pricing kernel is inﬂuenced by the stochastic volatility. In particular, the

two pricing kernels conditional on the same VIX level, Z

t

=20.03 (the top right and low left panels),

although separately estimated using non-overlapping data before and after the crisis, exhibit a similar

decreasing pattern.

12

Overall, consistent with Jackwerth and Brown (2001), Chabi-Yo et al. (2008), Chabi-Yo (2011),

and Christoﬀersen et al. (2010a), our nonparametric analysis reveals that the documented U-shape

of pricing kernels could be due to the missing stochastic volatility variable.

3.5 Test Results

Here we report the nonparametric test results of parametric models (7) using S&P 500 and VIX

options. For each parametric model, we estimate the model parameter θ by minimizing the mean

squared percentage pricing errors:

´

θ = min

θ∈Θ

_

n

C

i=1

_

log

_

´

C(τ

i

, s

t

i

, z

t

i

, x

i

)

C(τ

i

, s

t

i

, z

t

i

, x

i

; θ)

__

2

+

n

H

i=1

_

log

_

´

H(τ

i

, z

t

i

, x

i

)

H(τ

i

, z

t

i

, x

i

; θ)

__

2

_

.

It is worth pointing out that our estimation strategy avoids replacing the unobserved volatility process

by some VIX proxy as in A¨ıt-Sahalia and Kimmel (2007), or high frequency volatility measures, e.g.,

Wu (2011). Instead, we derive the exact relationship between the instantaneous volatility and the

VIX, and estimate the model using both VIX time series and VIX option prices, without which

the parameters in the volatility process cannot be identiﬁed. This estimation leads to a secondary

contribution of this paper—a new parametric estimation strategy.

We follow the common practice in the literature to select a subsample of S&P 500 and VIX

options for parametric ﬁtting, in order to accelerate the speed of calculation and ensure the quality

12

The diﬀerence in the supports of these two pricing kernels over the wealth axis is caused by diﬀerent levels of S&P

500 index pre- and post-crisis.

17

of the ﬁtting. For each day in 2009, we ﬁrst choose three shortest, yet, diﬀerent maturities, and for

each maturity, we randomly select three options with diﬀerent moneyness. The resulting subsample

includes a wide cross section of options. Parametric estimation results are shown in Table 5. Recall

that the SVCJ1 and SVCJ2 models are augmented with volatility jumps compared with the SVJ1 and

SVJ2 models. The diﬀerence in ﬁtting results indicates that adding jumps into volatility dynamics,

regardless of jump speciﬁcations, indeed reduces the ﬁtting errors from 60% to 40% and signiﬁcantly

decreases the volatility of volatility (σ). Moreover, the insigniﬁcant λ

0

estimates for both SVCJ1

and SVCJ2 suggest that a state-dependent jump intensity without constant terms seems adequate.

Finally, the estimates for the average jump sizes in S&P 500 returns are negative, as shown by

µ

s

= −0.119 for SVCJ1, and β

−

> β

+

, with the probability of downward jump equal to 1−q = 0.722

for SVCJ2.

Using both parametric and nonparametric estimates, the test statistics discussed in Section 2.5

are computed and reported in Table 6. We ﬁnd that all of the models are rejected with p-values

close to zero, implying that they cannot capture S&P 500 and VIX option prices simultaneously. To

gauge the performance in ﬁtting each type of option prices, Table 6 reports test statistics of separate

hypotheses. The result shows that for SVCJ1 and SVCJ2, the rejection for S&P 500 options is more

serious compared with VIX options, which have smaller values of test statistics. This result is not

surprising though, as adding VIX options in ﬁtting amounts to increasing the weight of volatility

dynamics.

Our SPD estimates can provide more-detailed information about the mis-speciﬁcation of the

model and pave the way for potential improvement. We report in Figures 8 and 9 the nonparametric

and parametric SPDs of S

t

and Z

t

at both high and low levels of the VIX across maturities. It is ob-

vious from the plot that for both index returns and the VIX, parametric estimates of SPDs are closer

to nonparametric ones when the current VIX level is high, whereas they tend to underestimate the

dispersion when volatility is low. This asymmetric pattern suggests an asymmetric mean-reverting

rate of volatility—the speed is high at low levels of volatility and low otherwise—because the aﬃne

models under consideration have a constant mean-reverting rate, κ. Therefore, the parametric mod-

els under consideration are not ﬂexible enough to capture the volatility dynamics implied by S&P

500 and VIX options simultaneously, indicating that non-aﬃne models may be better choices.

13

13

In a related study, Engulatov et al. (2011) nonparametrically estimate the drift function of the volatility process

and document a similar nonlinear pattern.

18

3.6 When VIX Options are Not Available

As shown in Table 1, the longest maturity for VIX option contracts in 2009 is 132 days whereas

S&P 500 options with maturities from 133 days to a year are still very liquid. As a result, investors

who would like to buy S&P 500 options with maturities longer than 132 days have no information

available from VIX options about the volatility of the future index. In this case, do investors predict

the long-term future volatility using current VIX Z

t

and price the S&P 500 options accordingly?

Or do they just price the long-maturity S&P 500 options by taking the future volatility level as a

constant, which may equal the long-run mean? Answering these questions will shed light on how

investors employ historical time series of the VIX up to time t to predict the volatility of the future

S&P 500 index S

T

.

To answer these questions, we direct our nonparametric estimation speciﬁcally to the long-term

SPDs of S&P 500 index with maturities longer than 132 days. The remaining sample size is as large

as 19,876 data points. Figure 10 provides nonparametric SPD estimates with maturities equal to

168, 231, 294, and 357 days for both high and low levels of the VIX. Across the long maturities, we

ﬁnd that the SPDs of the index conditional on high VIX levels diﬀer signiﬁcantly from the SPDs

conditional on low current VIX levels, i.e., the former density is more dispersed than the latter. The

signiﬁcance indicates that investors may predict the future volatility through certain mechanism and

price S&P 500 options accordingly, even when VIX options are not available.

One possible mechanism for such long-run dependence is the simple martingale procedure, which

is shown to work well even in the short run when VIX options are available. Speciﬁcally, investors

can ﬁrst compute implied volatility of yesterday’s S&P 500 option prices, interpolate the implied

volatility of long-term options using today’s corresponding characteristics, and then treat them as

the implied volatility of today’s long-term options. This martingale mechanism may result in the

observed long-run dependence as short-term options may carry information from Z

t

. To study

whether this mechanism works in practice, we run the following regression

IV

τ

t

= α

τ

+β

τ

IIV

τ

t

+ε

τ

t

of the implied volatility IV

τ

t

on the interpolated implied volatility IIV

τ

t

for the maturity τ =

168, 231, 294, and 357 days, corresponding to the documented long-term dependence in Figure 10.

The results in Table 7 show that the β

τ

coeﬃcients are signiﬁcant and close to one across all long

maturities with adjusted R

2

larger than 72%. This result suggests that the martingale procedure

agrees with the market practice and may oﬀer a potential explanation for the long-run dependence

of the future S&P 500 index on current VIX levels.

19

3.7 Robustness Checks

Here we verify, for the sample period, the two dimension-reduction techniques employed in our

nonparametric analysis, including homogeneous of degree 1 in S

t

for S&P 500 options, and the

dependence of VIX option prices on S

t

only through Z

t

.

Suppose

´

C(τ, s

t

, z

t

, x) is the nonparametric estimate of the S&P 500 call option price C(τ, s

t

, z

t

, x)

without the homogeneous of degree 1 assumption, and

´

C(τ, z

t

, m

t

) is the estimate we have been using.

For VIX options, ´ p(z

T

|τ, z

t

) and ´ p(z

T

|τ, s

t

, z

t

) are nonparametric estimates of the SPDs with and

without the assumption. In Figure 11, we plot

´

C(τ, z

t

, m

t

) and ´ p(z

T

|τ, z

t

) on the left panel along with

their percentage errors relative to

´

C(τ, s

t

, z

t

, x) and ´ p(z

T

|τ, s

t

, z

t

) on the right panel. Close scrutiny

of the plots shows that the percentage errors are mostly close to zeros in spite of reaching 20% and

10% around the boundaries where the estimates incur larger biases. Overall, our dimension-reduction

assumptions seem valid for the sample period in the empirical study.

We also propose rigorous statistical testing to check whether these assumptions hold. The null

hypotheses are given below:

H

1,S&P

: Pr {C(τ, S

t

, Z

t

, X) = C(τ, Z

t

, M

t

)} = 1

for S&P 500 options where M

t

= X/e

St

and

H

1,VIX

: Pr {p(Z

T

|τ, Z

t

, S

t

) = p(Z

T

|τ, Z

t

)} = 1

for VIX options. We construct test statistics

´

M

S&P

and

´

M

V IX

based on the diﬀerences between the

two estimators for the same quantity. Under certain regularity conditions, both statistics converge

to standard normal distributions under the null hypotheses; see Appendix C for details. Table 6

reports values of the test statistics and the corresponding p-values. There is no statistical evidence

for rejection with p-values around 0.5.

4 Conclusion

In this paper, we proposed SPD estimates of the index and volatility and document several important

empirical facts relevant to the understanding of the interactions between the two option markets. In

particular, the dependence of SPDs of the index on the VIX suggests pervasive volatility information

in the VIX and its option market. Our results also highlight the missing volatility variable that

may be responsible for the pricing kernel puzzle. In addition, our testing results call the aﬃne

models into question, and suggest that the inadequacy of these models lies in the inability to adjust

mean-reverting speeds for diﬀerent volatility regimes.

20

In fact, our nonparametric pricing method carries over the strategies introduced in A¨ıt-Sahalia

and Lo (1998) to the case with stochastic volatility, by exploring additional information from VIX

and its options. By nature, our method enjoys several beneﬁcial features, such as being model-

free, robust to misspeciﬁcation of models and pricing measures, and computationally eﬃcient. In

contrast, parametric stochastic volatility models face an unfortunate compromise between misspeci-

ﬁcation and computation burden. For example, the convenient aﬃne model with closed-form option

pricing formulae are usually subject to misspeciﬁcation errors. Moreover, even with aﬃne mod-

els, the computation often involves implementation of numerical algorithms, such as inverse Fourier

transforms and partial diﬀerential equations. Our closed-form regression formulae, however, alleviate

the dilemma to a great extent. Of course, eﬀorts are needed to choose suitable bandwidths for our

approach.

21

Appendix

A Speciﬁcation Test

We measure the diﬀerences between the parametric and nonparametric SPDs by the sum of squared

deviations:

M (´ p (·) , p (·)) =E

_

(´ p(m

T

|τ, Z

t

) −p(m

T

|τ, Z

t

, θ

0

))

2

a

C

(τ, Z

t

, m

T

)

+(´ p(Z

T

|τ, Z

t

) −p(Z

T

|τ, Z

t

; θ

0

))

2

a

H

(τ, Z

t

, Z

T

)

_

where a

C

(·) : R

3

→ R

+

and a

H

(·) : R

3

→ R

+

are weighting functions. The idea behind our test is

simple: If the null H

0

holds, then SPDs of both S&P 500 index and VIX should be close to their

nonparametric estimates. Suppose we have a consistent estimator

´

θ for the parameter θ. Then we

can estimate M (´ p (·) , p (·)) by its sample analog

´

M (´ p (·) , p (·)) =

1

n

C

n

C

i=1

_

´ p(m

T

i

|τ

i

, Z

t

i

) −p(m

T

i

|τ

i

, Z

t

i

;

´

θ)

_

2

a

C

(τ

i

, Z

t

i

, m

T

i

)

+

1

n

H

n

H

i=1

_

´ p(Z

T

i

|τ

i

, Z

t

i

) −p(Z

T

i

|τ

i

, Z

t

i

;

´

θ)

_

2

a

H

(τ

i

, Z

t

i

, Z

T

i

) (A.1)

The use of weighting functions a

C

(·) and a

H

(·) are not uncommon in the literature and often used to

remove extreme observations. As noted by A¨ıt-Sahalia et al. (2001), appropriate choices of weighting

functions can reduce the inﬂuences of unreliable estimates and make the tests focus on a particular

empirical question of interest.

Denote the bandwidth for the state variable m

T

as h

m

. Our test statistic for H

0

against H

A

is

an appropriately standardized version of (A.1):

´

M =

_

h

2

m

(h

τ

h

z

h

m

)

1/2

n

C

i=1

_

´ p(m

T

i

|τ

i

, Z

t

i

) −p(m

T

i

|τ

i

, Z

t

i

;

´

θ)

_

2

×a

C

(τ

i

, Z

t

i

, m

T

i

) −

´

C

1

_

/

_

√

2

´

D

1

+

_

h

2

y

_

h

τ

h

z

h

τ

_

1/2

n

H

i=1

_

´ p(Z

T

i

|τ

i

, Z

t

i

) −p(Z

T

i

|τ

i

, Z

t

i

;

´

θ)

_

2

a

H

(τ

i

, Z

t

i

, Z

T

i

) −

´

C

2

_

/

_

√

2

´

D

2

where

´

C

1

=(h

τ

h

z

h

m

)

−1/2

__

k

2

(c) dc

_

2

__

_

c

˙

k (c) +k(c)

_

2

dc

_

/

__

k (c) c

2

dc

_

2

×e

2rτ

_

´ s

2

C

(τ, z, m) a

C

(τ, z, m) dτdzdm

´

C

2

=

_

h

τ

h

z

h

y

_

−1/2

__

k

2

(c) dc

_

2

__

_

c

˙

k (c) +k(c)

_

2

dc

_

/

__

k (c) c

2

dc

_

2

22

×e

2rτ

_

´ s

2

H

(τ, z, y) a

H

(τ, z, y) dτdzdy

´

D

1

=2

_

_ __

_

c

1

˙

k (c

1

) +k(c

1

)

__

(c

1

+c

2

)

˙

k (c

1

+c

2

) +k(c

1

+c

2

)

_

dc

1

_

2

dc

2

_

×

_

_ __

k (c

1

) k (c

1

+c

2

) dc

1

_

2

dc

2

_

2

/

__

k (c) c

2

dc

_

4

×e

4rτ

_

´ s

4

C

(τ, z, m) a

C

(τ, z, m) dτdzdm

´

D

2

=2

_

_ __

_

c

1

˙

k (c

1

) +k(c

1

)

__

(c

1

+c

2

)

˙

k (c

1

+c

2

) +k(c

1

+c

2

)

_

dc

1

_

2

dc

2

_

×

_

_ __

k (c

1

) k (c

1

+c

2

) dc

1

_

2

dc

2

_

2

/

__

k (c) c

2

dc

_

4

×e

4rτ

_

´ s

4

H

(τ, z, y) a

H

(τ, z, y) dτdzdy

where s

C

(τ, z, m) is the square-root of the conditional variance for the local linear regression of C

i

on U

i

and s

H

(τ, z, y) is deﬁned similarly. The bandwidths are chosen to satisfy

ln n

C

n

C

h

4

m

hτ hz

→ 0,

h

2

m

h

τ

h

z

ln n

C

→ 0,

ln n

H

nh

4

y

hτ hz

→ 0, and h

2

y

h

τ

h

z

ln n

H

→ 0 as n

C

, n

H

→ ∞, with n

C

/n

H

converging to

a non-zero constant. Under certain regularity conditions, we have

´

M

d

−→N (0, 1) under H

0

as n

C

, n

H

→∞

When the null hypothesis H

0

is rejected, we follow similar procedures as above to construct test

statistics

´

M

S&P

and

´

M

VIX

to test separate speciﬁcation hypotheses H

0,S&P

and H

0,VIX

, i.e., whether

the model is consistent with S&P 500 option prices and VIX option prices separately.

B Parametric SPDs and Option Pricing

We derive the theoretical value of the VIX Z

t

for Models SVCJ1 and SVCJ2, since the other models

are nested. In fact, we have

Z

2

t

= 10

4

·

1

τ

E

Q

t

_

_

t+τ

t

V

s

ds +

Ns≥0, t≤s≤t+τ

(J

Q

S

)

2

_

= aV

t

+b

where

a =

1 −e

−(κ−β

V

λ

1

)τ

(κ −β

V

λ

1

)τ

(1 +λ

1

χ) · 10

4

b =

_

κξ +β

V

λ

0

κ −β

V

λ

1

_

1 −

1 −e

−(κ−β

V

λ

1

)τ

(κ −β

V

λ

1

)τ

_

(1 +λ

1

χ) +λ

0

χ

_

· 10

4

23

χ =

_

_

_

µ

2

S

+σ

2

S

for Model SVCJ1,

2(qβ

2

+

+ (1 −q)β

2

−

) for Model SVCJ2.

In the following, we derive the conditional densities p(s

T

|τ, s

t

, z

t

;θ) and p(z

T

|τ, s

t

, z

t

;θ). Due to

the fact that Z

t

=

√

aV

t

+b, we have by change of variable that

p(z

T

|τ, s

t

, z

t

; θ) =

2z

T

a

p(v

T

|τ, s

t

, v

t

; θ)

¸

¸

¸

_

v

T

=

z

2

T

−b

a

,vt=

z

2

t

−b

a

_

Therefore, it is suﬃcient to calculate the marginal density of V . Now let Ψ(τ, u, w, s, v) be the time-t

generalized conditional characteristic function of (S

T

, V

T

), that is,

Ψ(τ, (u

s

, u

v

), (w

s

, w

v

), s, v; θ) = E

Q

t

(e

(us+iws)S

T

+(uv+iwv)V

T

).

where τ = T −t, u = (u

s

, u

v

) ∈ R

2

and w = (w

s

, w

v

) ∈ R

2

. The conditional density p(s

T

, v

T

|s

t

, v

t

;θ)

is given by

p(s

T

, v

T

|s

t

, v

t

; θ) =

1

4π

2

_

R

2

e

−iwss

T

−iwvv

T

Ψ(τ, 0, w, s

t

, v

t

; θ)dw

Moreover, the marginal densities p(s

T

|τ, s

t

, v

t

;θ) and p(v

T

|τ, s

t

, v

t

;θ) are:

p(s

T

, τ|s

t

, v

t

; θ) =

1

2π

_

R

e

−iwss

T

Ψ(τ, 0, (w

s

, 0), s

t

, v

t

; θ)dw

s

=

1

π

_

∞

0

Re

_

e

−iwss

T

Ψ(τ, 0, (w

s

, 0), s

t

, v

t

; θ)

_

dw

s

p(v

T

, τ|s

t

, v

t

; θ) =

1

2π

_

R

e

−iwvv

T

Ψ(τ, 0, (0, w

v

), s

t

, v

t

; θ)dw

v

=

1

π

_

∞

0

Re

_

e

−iwvv

T

Ψ(τ, 0, (0, w

v

), s

t

, v

t

; θ)

_

dw

v

As the model is aﬃne, Ψ(τ, u, w, s, v;θ) has a closed form formula. In fact, Ψ(τ, u, w, s, v;θ) satisﬁes

the following PDE:

0 = −

∂Ψ

∂τ

+ (r −d −

1

2

v −µ(λ

0

+λ

1

v))

∂Ψ

∂s

+κ(ξ −v)

∂Ψ

∂v

+

1

2

v

∂

2

Ψ

∂s

2

+

1

2

σ

2

v

∂

2

Ψ

∂v

2

+ρσv

∂

2

Ψ

∂s∂v

+ (λ

0

+λ

1

v)

__

R

2

_

Ψ(τ, u, w, s +η, v +ζ) −Ψ(τ, u, w, s, v)

_

ν

S

(η)ν

V

(ζ)dηdζ

with the initial condition:

Ψ(0, u, w, s, v; θ) = e

(us+iws)s+(uv+iwv)v

Since the model is aﬃne in v, we can obtain a closed-form formula for Ψ:

Ψ(τ, u, w, s, v; θ) = e

(us+iws)s+A(τ;θ)+B(τ;θ)v

where A and B satisfy the system of ordinary diﬀerential equations below:

0 = −

˙

B +

1

2

σ

2

B

2

+ ((u

s

+iw

s

)ρσ −κ)B −(

1

2

+µλ

1

)(u

s

+iw

s

) +

1

2

(u

s

+iw

s

)

2

+λ

1

(l(B) −1)

24

0 = −

˙

A+κξB + (r −d −µλ

0

)(u

s

+iw

s

) +λ

0

(l(B) −1)

where A(0;θ) = 0, B(0;θ) = u

v

+iw

v

, and

l(a) =

__

R

2

e

(us+iws)η+ζa

ν

S

(η)ν

V

(ζ)dηdζ

=

_

_

_

(1 −β

V

a)

−1

e

µs(us+iωs)+

1

2

σ

2

s

(us+iωs)

2

for Model SVCJ1,

(1 −β

V

a)

−1

_

q(1 −(u

s

+iw

s

)β

+

)

−1

+ (1 −q)(1 + (u

s

+iw

s

)β

−

)

−1

_

for Model SVCJ2.

The price of the S&P 500 Call option with strike X and maturity τ is given by:

C(τ, s

t

, v

t

, X; θ) =e

−rτ

E

Q

t

_

(e

S

T

−X)

+

_

=

e

−rτ

π

_

∞

0

Re

_

Ψ((u

s

, 0), (w

s

, 0), τ, s

t

, v

t

; θ)Xe

−(us+iws) log X

(u

s

+iw

s

)(u

s

+iw

s

−1)

_

du

s

for u

s

> 1. Finally, the VIX option price with maturity T and strike price Y is give by:

H(τ, v

t

, Y ; θ) =e

−rτ

E

Q

t

_

(

_

aV

T

+b −Y )

+

_

=e

−rτ

_

a

π

_

∞

0

Re

_Ψ((0, u

v

), (0, w

v

), τ, s

t

, v

t

; θ)e

b

a

(uv+iwv)

erfc

_

Y (

uv+iwv

a

)

1

2

_

2(u

v

+iw

v

)

3

2

_

dw

v

where u

v

> 0, and erfc(x) =

2

√

π

_

∞

x

e

−t

2

dt is the complementary error function. Using the fact that

z

t

=

√

av

t

+b, we can rewrite functions C and H in terms of z

t

, i.e., the VIX, which is observable

from the market.

In order to calculate standard errors for parametric estimation, we need the derivatives of option

prices with respect to θ. This is straightforward for S&P 500 options, since

∂C

∂θ

=

e

−rτ

π

_

∞

0

Re

_

∂Ψ((u

s

, 0), (w

s

, 0), τ, s

t

, v

t

; θ)

∂θ

Xe

−(us+iws) log X

(u

s

+iw

s

)(u

s

+iw

s

−1)

_

du

s

where

∂Ψ(u, w, τ, s

t

, v

t

; θ)

∂θ

=

_

∂A(τ; θ)

∂θ

+

∂B(τ; θ)

∂θ

v

_

Ψ(u, w, τ, s

t

, v

t

; θ)

and

∂A(τ;θ)

∂θ

,

∂B(τ;θ)

∂θ

can be calculated by augmenting the ODE systems with more unknowns. Regard-

ing VIX options, the formula becomes more complicated, since a and b also depend on θ. Therefore,

we have

∂H

∂θ

=e

−rτ

_

1

π

_

∞

0

Re

_

∂Ψ((0, u

v

), (0, w

v

), τ, s

t

, v

t

; θ)

∂θ

√

ae

b

a

(uv+iwv)

erfc

_

Y (

uv+iwv

a

)

1

2

_

2(u

v

+iw

v

)

3

2

_

dw

v

+e

−rτ

_

1

π

_

∞

0

Re

_

Ψ((0, u

v

), (0, w

v

), τ, s

t

, v

t

; θ)

2(u

v

+iw

v

)

3

2

∂

_

√

ae

b

a

(uv+iwv)

erfc

_

Y (

uv+iwv

a

)

1

2

__

∂θ

_

dw

v

and then follow similar calculations as for S&P 500 options.

25

C Tests for Robustness Checks

We measure the diﬀerences between

´

C(τ, S

t

, Z

t

, X) and

´

C(τ, Z

t

, M

t

), and between ´ p(z

T

|τ, z

t

) and

´ p(z

T

|τ, s

t

, z

t

) by the sum of squared deviations. The test statistics are appropriately standardized

versions. For the S&P 500 options, we have

´

M

S&P

=

_

(h

τ

h

z

h

s

h

x

)

1/2

n

C

i=1

_

´

C(τ, S

t

, Z

t

, X) −

´

C(τ, Z

t

, M

t

)

_

2

a

C

(τ

i

, S

t

i

, Z

t

i

, X

i

) −

´

C

S&P

_

/

_

´

D

S&P

where

´

C

S&P

=(h

τ

h

z

h

s

h

x

)

−1/2

__

k

2

(c) dc

_

_

_

´ s

2

C

(τ, z, s, x) + ´ s

2

C

(τ, z, m)

¸

a

C

(τ, z, s, x) dτdzdsdx,

´

D

S&P

=2

_

_ __

k (c

1

) k (c

1

+c

2

) dc

1

_

2

dc

2

_

4

_

_

´ s

2

C

(τ, z, s, x) + ´ s

2

C

(τ, z, m)

¸

2

a

2

C

(τ, z, s, x) dτdzdsdx,

and ´ s

2

C

(τ, z, s, x) and ´ s

2

C

(τ, z, m) are estimators of conditional variances of regressing option prices

C

i

on (τ

i

, S

t

i

, Z

t

i

, X

i

) without and with assuming the homogenous of degree 1.

For the VIX density, we have

´

M

VIX

=

_

h

2

y

_

h

τ

h

z

h

s

h

y

_

1/2

n

H

i=1

[´ p(Z

T

i

|τ

i

, S

t

i

, Z

t

i

) − ´ p(Z

T

i

|τ

i

, Z

t

i

)]

2

a

H

(τ

i

, S

t

i

, Z

t

i

, Z

T

i

) −

´

C

VIX

_

/

_

´

D

VIX

where

´

C

VIX

=

_

h

τ

h

z

h

s

h

y

_

−1/2

__

k

2

(c) dc

_

3

__

_

c

˙

k (c) +k(c)

_

2

dc

_

/

__

c

2

k (c) dc

_

2

×e

2rτ

_

_

´ s

2

H

_

τ, z, s, z

_

+ ´ s

2

H

_

τ, z, z

_¸

a

H

_

τ, z, s, z

_

dτdzdsdz

,

´

D

VIX

=2

_

_ __

_

c

1

˙

k (c

1

) +k(c

1

)

__

(c

1

+c

2

)

˙

k (c

1

+c

2

) +k(c

1

+c

2

)

_

dc

1

_

2

dc

2

_

×

_

_ __

k (c

1

) k (c

1

+c

2

) dc

1

_

2

dc

2

_

3

/

__

c

2

k (c) dc

_

4

×e

4rτ

_

_

´ s

2

H

_

τ, z, s, z

_

+ ´ s

2

H

_

τ, z, z

_¸

2

a

2

H

_

τ, z, s, z

_

dτdzdsdz

,

26

and ´ s

2

H

(τ, z, s, z

) and ´ s

2

H

(τ, z, z

**) are estimators of conditional variances of regressing option prices
**

H

i

on (τ

i

, S

t

i

, Z

t

i

, Y

i

) without and with assuming that H does not depend on S

t

. By similar regularity

conditions and bandwidth choices to those for the speciﬁcation test

´

M, we have

´

M

S&P

d

−→N (0, 1) under H

1,S&P

as n

C

→∞,

´

M

VIX

d

−→N (0, 1) under H

1,VIX

as n

H

→∞.

27

References

A¨ıt-Sahalia, Y. and Kimmel, R. (2007), “Maximum Likelihood Estimation of Stochastic Volatility

Models,” Journal of Financial Economics, 83, 413–452.

A¨ıt-Sahalia, Y. and Lo, A. (1998), “Nonparametric Estimation of State-Price-Densities Implicit in

Financial Asset Prices,” Journal of Finance, 53, 499–547.

— (2000), “Nonparametric Risk Management and Implied Risk Aversion,” Journal of Econometrics,

94, 9–51.

A¨ıt-Sahalia, Y., Wang, Y., and Yared, F. (2001), “Do Option Markets Correctly Price the Probabil-

ities of Movement of the Underlying Asset?” Journal of Econometrics, 102, 67–110.

Amengual, D. and Xiu, D. (2012), “Delving into Risk Premia: Reconciling Evidence from the S&P

500 and VIX Derivatives,” Tech. rep., University of Chicago Booth School of Business.

Bakshi, G., Cao, C., and Chen, Z. (1997), “Empirical Performance of Alternative Option Pricing

Models,” Journal of Finance, 52, 2003–2049.

Bakshi, G., Madan, D., and Panayotov, G. (2010), “Returns of claims on the upside and the viability

of U-shaped pricing kernels,” Journal of Financial Economics, 97, 130–154.

Bates, D. S. (2000), “Post-’87 Crash Fears in the S&P 500 Futures Option Market,” Journal of

Econometrics, 94, 181–238.

Blair, B. J., Poon, S.-H., and Taylor, S. J. (2001), “Forecasting S&P 100 volatility: the incremental

information content of implied volatilities and high-frequency index returns,” Journal of Econo-

metrics, 105, 5–26.

Boes, M., Drost, F., and Werker, B. J. (2007), “Nonparametric risk-neutral return and volatility

distributions,” Tech. rep., Tilburg University.

Breeden, D. and Litzenberger, R. H. (1978), “Prices of State-Contingent Claims Implicit in Option

Prices,” Journal of Business, 51, 621–651.

Britten-Jones, M. and Neuberger, A. (2000), “Option Prices, Implied Price Processes, and Stochastic

Volatility,” Journal of Finance, 55, 839–866.

Broadie, M., Chernov, M., and Johannes, M. S. (2007), “Model Speciﬁcation and Risk Premia:

Evidence from Futures Options,” Journal of Finance, 62.

28

Buraschi, A. and Jackwerth, J. (2001), “The Price of a Smile/; Hedging and Spanning in Option

Markets,” Review of Financial Studies, 14, 495–527.

Carr, P. and Wu, L. (2009), “Variance Risk Premiums,” Review of Financial Studies, 22, 1311–1341.

Chabi-Yo, F. (2011), “Pricing Kernels with Stochastic Skewness and Volatility Risk,” Management

Science.

Chabi-Yo, F., Garcia, R., and Renault, E. (2008), “State dependence can explain the risk aversion

puzzle,” Review of Financial Studies, 21, 973–1011.

Christoﬀersen, P., Heston, S., and Jacobs, K. (2010a), “Option anomalies and the pricing kernel,”

Tech. rep., McGill University.

Christoﬀersen, P., Jacobs, K., and Mimouni, K. (2010b), “Models for S&P 500 dynamics: evidence

from realized volatility, daily returns, and option prices.” Review of Financial Studies, 23, 3141–

3189.

Detemple, J. and Osakwe, C. (2000), “The valuation of volatility options,” European Finance Review,

4, 21–50.

Duan, J. and Yeh, C.-Y. (2011), “Jump and Volatility Risk Premiums Implied by VIX,” Journal of

Economic Dynamics and Control, 34, 2232–2244.

Engulatov, A., Gonzalev, R., and Scaillet, O. (2011), “Recovering Nonlinear Dynamics from Option

Prices,” Tech. rep., University of Geneva.

Eraker, B. (2004), “Do Stock Prices and Volatility Jump? Reconciling Evidence from Spot and

Option Prices,” Journal of Finance, 59.

Fan, J. and Gijbels, I. (1996), Local Polynomial Modelling and Its Applications, London, U.K.:

Chapman & Hall.

Fan, J. and Mancini, L. (2009), “Option pricing with aggregation of physical models and nonpara-

metric statistical learning,” Journal of American Statistical Association, 104, 1351–1372.

Grunbichler, A. and Longstaﬀ, F. (1996), “Valuing futures and options on volatility,” Journal of

Banking and Finance, 20, 985–1001.

Jackwerth, J. (2000), “Recovering risk aversion from option prices and realized returns,” Review of

Financial Studies, 13, 433–451.

29

Jackwerth, J. and Brown, P. (2001), “The pricing kernel puzzle: Reconciling index option data and

economic theory,” Tech. rep., University of Konstanz.

Jackwerth, J. and Rubinstein, M. (1996), “Recovering Probability Distributions from Contempora-

neous Security Prices,” Journal of Finance, 51, 1611–1631.

Jiang, G. J. and Tian, Y. S. (2005), “The Model-Free Implied Volatility and Its Information Content,”

Review of Financial Studies, 18, 1305–1342.

Li, H. and Zhao, F. (2009), “Nonparametric Estimation of State-Price-Densities Implicit in Interest

Rate Cap Prices,” Review of Financial Studies, 22, 4335–4376.

Mencia, J. and Sentana, E. (2009), “Valuation of VIX Derivatives,” Tech. Rep. 0913, CEMFI.

Pan, J. (2002), “The Jump-Risk Premia Implicit in Options: Evidence from an Integrated Time-

Series Study,” Journal of Financial Economics, 63, 3–50.

Sepp, A. (2008), “VIX option pricing in jump-diﬀusion model,” Risk, April, 84–89.

Whaley, R. E. (1993), “Derivatives on Market Volatility: Hedging Tools Long Overdue,” The Journal

of Derivatives, 1, 71–84.

Wu, L. (2011), “Variance Dynamics: Joint Evidence from Options and High-Frequency Returns,”

Journal of Econometrics, 160, 280–287.

30

Table 1: Summary Statistics of the S&P 500 Index and VIX

Mean Std Skew Kurt Median 25% 75%

S&P 500 Index 948.0464 115.5682 -0.1954 2.0342 939.145 855.86 1057.3

S&P 500 Return 0.0007 0.0171 -0.0625 4.9126 0.0018 -0.0081 0.0097

VIX 31.4786 9.0777 0.6825 2.1979 28.57 24.275 39.425

Note: This table reports the summary statistics of the time series of S&P 500 index, return, and VIX from January 2

to December 31, 2009.

31

Table 2: Summary Statistics of S&P 500 and VIX Options

S&P 500 Call S&P 500 Put VIX Call VIX Put

Moneyness OTM ITM OTM ITM OTM ITM OTM ITM

# of Contracts 54,123 21,599 82,658 20,530 11,571 4,083 3,909 4,243

Trading Volume ×10

6

72.2 29.8 135.5 16.9 19.4 1.5 7.0 5.4

Option Price

Min 0.025 6.75 0.025 5.55 0.025 0.325 0.025 0.275

25% 1.00 47.40 1.25 50.40 0.350 5.450 0.175 3.150

50% 6.70 71.15 6.60 83.90 1.000 7.600 0.675 5.950

75% 22.20 111.10 22.75 151.55 2.275 11.350 1.750 11.688

Max 184.60 622.05 197.95 1583.10 7.450 45.10 16.80 73.00

Strike Price

Min 680 300 50 680 20.00 10.00 10.00 20.00

20% 940 800 650 895 35.00 20.00 22.50 27.50

50% 1025 875 775 970 42.50 25.00 27.50 35.00

75% 1125 965 875 1090 55.00 32.50 35.00 45.00

Max 2500 1125 1125 2500 100.00 55.00 55.00 100.00

Time-to-Maturity

Min 3 3 3 3 2 2 2 2

25% 21 17 20 19 26 20 22 17

50% 39 33 38 37 49 40 41 34

75% 80 69 78 86 80 72 67 54

Max 783 780 783 780 132 130 130 131

Implied Volatility

Min 0.136 0.153 0.153 0.105

25% 0.216 0.262 0.319 0.237

50% 0.267 0.317 0.406 0.312

75% 0.336 0.395 0.518 0.377

Max 1.552 2.996 2.103 2.933

Note: This table reports the summary statistics (minimum, quantiles, and maximum) for all traded S&P 500 and VIX

options in 2009, including the option price, strike price, time-to-maturity, implied volatility, and trading volume. In

total, there are 178,910 trading records for S&P 500 options with valid implied volatility and trading volumes, 23,806

records for VIX options with trading volumes.

32

Table 3: Dependence of S&P 500 Index SPD on the VIX SPD

Intercept Z

t

Std

t

[Z

t+τ

] Skew

t

[Z

t+τ

] Adj.R

2

Panel A: Regression Results for τ = 21

48.097 (1.7106) -0.2385 (0.0487) 0.1063

80.409 (27.299) -0.0495 (0.0166) -4.3725 (3.6869) 0.1082

86.670 (27.251) -0.1899 (0.0179) -3.8173 (3.6671) -3.6036 (1.7711) 0.1226

Panel B: Regression Results for τ = 42

39.289 (1.7079) -0.0633 (0.0487) 0.0036

74.546 (27.981) -0.0677 (0.1146) -4.2847 (3.3942) 0.0066

51.251 (30.215) 0.1735 (0.1676) -0.0180 (4.0120) -5.2299 (2.6299) 0.0212

Panel C: Regression Results for τ = 63

-29.282 (1.8518) 0.0202 (0.0528) 0.0008

-85.347 (21.036) 0.1122 (0.0520) 11.402 (2.0853) 0.1270

-64.527 (23.110) 0.0969 (0.0520) 9.9069 (2.1873) -4.3522 (2.0823) 0.1421

Panel D: Regression Results for τ = 84

-24.893 (2.5586) 0.1286 (0.0729) 0.0108

-74.216 (13.050) 0.3859 (0.0720) 8.7851 (1.1396) 0.2416

-69.434 (21.840) 0.3831 (0.0729) 8.4237 (1.7469) -0.8486 (3.1035) 0.2379

Note: This table reports regression results of the future average volatility V

S

t,T

of the S&P 500 index on the square

root of the conditional second moment, conditional skewness, and conditional kurtosis of the VIX densities, with

time-to-maturity equal to 21, 42, 63, and 84 days. The numbers in parentheses are standard errors.

33

Table 4: In-Sample and Out-of-Sample Forecasts

Panel A: Density Forecast Error

Forecast of AL-Density Forecast of SX-Density

γ AL SX AL SX

0 0.0000 1.5582 1.6719 0.0000

7 0.5422 1.4679 1.7012 0.2091

14 0.9451 1.1815 1.7362 0.3322

21 1.2332 1.0107 1.9444 0.4428

28 1.5370 0.7206 2.3313 0.6040

35 1.8695 0.7399 2.5789 0.8338

42 2.0787 0.8326 2.9918 1.1627

Panel B: Market-Price Forecast Error

τ AL SX MKT

0 9.36 8.54 0.00

7 8.79 7.76 4.65

14 6.46 6.97 5.63

21 5.68 4.68 4.62

28 5.79 2.41 3.28

35 4.90 1.74 2.95

42 5.09 1.27 2.11

Note: Panel A reports average forecast errors of the densities produced by the A¨ıt-Sahalia and Lo (1998) estimator

(AL) and our SPD estimator (SX) conditional on the VIX, while Panel B reports those of prices by the AL, SX,

and a martinglae interpolation (MKT) method. The nonparametric option prices and their corresponding SPDs are

estimated with daily data from January 2 to September 30, 2009 and out-of-sample forecasts are generated for various

forecast horizons t on a daily rolling basis from October 1 to December 31, 2009.

34

Table 5: Estimates of Parametric Models

Param SV SVJ1 SVJ2 SVCJ1 SVCJ2

κ 3.672 (0.094) 2.323 (0.086) 3.619 (0.102) 4.362 (0.162) 3.109 (0.159)

ξ 0.145 (0.002) 0.183 (0.004) 0.147 (0.002) 0.028 (0.002) 0.032 (0.002)

σ 1.033 (0.009) 0.921 (0.008) 1.023 (0.010) 0.370 (0.015) 0.410 (0.011)

ρ -0.708 (0.005) -0.953 (0.011) -0.786 (0.011) -0.652 (0.039) -0.611 (0.021)

β

+

0.047 (0.001) 0.002 (0.005)

β

−

0.032 (0.006) 0.138 (0.019)

q 0.128 (0.028) 0.316 (0.100)

β

V

0.067 (0.003) 0.075 (0.004)

µ

s

0.021 (0.013) -0.119 (0.009)

σ

s

0.022 (0.006) 0.154 (0.006)

λ

0

6.768 (2.864) 5.182 (1.371) 0.000 (0.039) 0.001 (0.048)

λ

1

33.276 (2.526) 26.991 (2.096)

RMSE 0.602 0.583 0.585 0.348 0.376

Note: This table reports the parametric estimates for a variety of models. Standard errors are given in parentheses.

Table 6: Nonparametric Speciﬁcation Tests of Parametric Models

Tests SV SVJ1 SVJ2 SVCJ1 SVCJ2

S&P 25.6055 14.5384 11.8035 13.0357 9.8441

(0.0000) (0.0000) (0.0000) (0.0000) (0.0000)

VIX 21.9234 11.0747 15.5492 8.2278 6.9282

(0.0000) (0.0000) (0.0000) (0.0000) (0.0000)

Joint 33.6080 16.1774 19.3413 12.7787 11.8598

(0.0000) (0.0000) (0.0000) (0.0000) (0.0000)

Note: This table reports the nonparametric speciﬁcation tests for a variety of models. P values are given in parentheses.

35

Table 7: Regression Analysis for Long-Term Dependence

Time-to-Maturity 168 231 294 357

Intercept -0.0275 -0.0082 -0.0366 -0.0112

(0.0035) (0.0028) (0.0056) (0.0046)

IIV 1.1165 1.0459 1.1403 1.0476

(0.0108) (0.0090) (0.0173) (0.0142)

Adj.R

2

0.7290 0.7705 0.7967 0.8317

Note: This table reports regression results of implied volatility of observed option prices on the interpolated implied

volatilities for long maturities. Standard errors are given in parentheses.

Table 8: Robustness Tests

S&P 500 VIX

Test Statistics 0.0439 0.0154

p-value 0.5136 0.5025

Note: This table reports the nonparametric tests for the two dimension reduction assumptions: homogeneous of degree

1 for S&P 500 options and dependence of VIX option prices on St only through Zt.

36

Figure 1: Time Series of the S&P 500 Index and the VIX

Jan 09 Mar 09 May 09 Jun 09 Sep 09 Nov 09 Jan 10

600

700

800

900

1000

1100

1200

S

&

P

5

0

0

L

e

v

e

l

Jan 09 Mar 09 May 09 Jun 09 Sep 09 Nov 09 Jan 10

10

20

30

40

50

60

V

I

X

L

e

v

e

l

S&P 500

VIX

Note: This ﬁgure plots the time series of S&P 500 index and VIX from January 2 to December 31, 2009.

37

Figure 2: Nonparametric Option Prices and State-Price Densities of S&P

800

900

1000

1100

1200

1300

1400

20

40

60

80

100

0

50

100

150

200

250

300

350

Current Index Level

Time−to−Maturity

S

&

P

5

0

0

O

p

t

i

o

n

P

r

i

c

e

s

800

900

1000

1100

1200

1300

1400

20

40

60

80

100

0

0.5

1

1.5

2

2.5

3

3.5

x 10

−3

Index Level at Expiration

Time−to−Maturity

S

P

D

o

f

S

&

P

5

0

0

I

n

d

e

x

800

900

1000

1100

1200

1300

1400

20

40

60

80

100

−1

−0.5

0

0.5

1

Current Index Level

Time−to−Maturity

S

&

P

5

0

0

O

p

t

i

o

n

%

E

r

r

o

r

800

900

1000

1100

1200

1300

1400

20

40

60

80

100

−1

−0.5

0

0.5

1

Index Level at Expiration

Time−to−Maturity

S

P

D

o

f

t

h

e

S

&

P

5

0

0

%

E

r

r

o

r

Note: In the upper panel, we plot the nonparametric estimates of S&P 500 option prices and the SPD of the index,

ﬁxing the current volatility level. The percentage estimation errors, averaged by 500 replications, are plotted in the

lower panel. The data are simulated from the model SVCJ2, with empirical parameter estimates given in Section 3.

38

Figure 3: Nonparametric Option Prices and the State-Price Density of the VIX

25

30

35

40

45

20

40

60

80

100

5

10

15

20

25

Current VIX Level

Time−to−Maturity

V

I

X

O

p

t

i

o

n

P

r

i

c

e

s

25

30

35

40

45

20

40

60

80

100

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

VIX Level at Expiration

Time−to−Maturity

S

P

D

o

f

V

I

X

25

30

35

40

45

20

40

60

80

100

−1

−0.5

0

0.5

1

Current VIX Level

Time−to−Maturity

V

I

X

O

p

t

i

o

n

%

E

r

r

o

r

25

30

35

40

45

20

40

60

80

100

−1

−0.5

0

0.5

1

VIX Level at Expiration

Time−to−Maturity

S

P

D

o

f

V

I

X

%

E

r

r

o

r

Note: In the upper panel, we plot the nonparametric estimates of VIX option prices and the SPD of the VIX, ﬁxing

current volatility level at medium. The percentage estimation errors, averaged by 500 replications, are plotted in the

lower panel. The data are simulated from the model SVCJ2, with empirical parameter estimates given in Section 4.

39

Figure 4: Estimates of the State-Price Density of the S&P 500

650 700 750 800 850 900 950 1000 1050 1100 1150

0

1

2

3

4

5

6

7

8

x 10

−3

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

S&P 500 index at Expiration

Time−to−Maturity τ : 21, Current Index Level Zt: 23.49

650 700 750 800 850 900 950 1000 1050 1100 1150

0

1

2

3

4

5

6

7

8

x 10

−3

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

S&P 500 index at Expiration

Time−to−Maturity τ : 21, Current Index Level Zt: 42.455

650 700 750 800 850 900 950 1000 1050 1100 1150

0

1

2

3

4

5

6

x 10

−3

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

S&P 500 index at Expiration

Time−to−Maturity τ : 42, Current Index Level Zt: 23.49

650 700 750 800 850 900 950 1000 1050 1100 1150

0

1

2

3

4

5

6

x 10

−3

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

S&P 500 index at Expiration

Time−to−Maturity τ : 42, Current Index Level Zt: 42.455

650 700 750 800 850 900 950 1000 1050 1100 1150

0

0.5

1

1.5

2

2.5

3

3.5

4

x 10

−3

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

S&P 500 index at Expiration

Time−to−Maturity τ : 84, Current Index Level Zt: 23.49

650 700 750 800 850 900 950 1000 1050 1100 1150

0

0.5

1

1.5

2

2.5

3

3.5

4

x 10

−3

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

S&P 500 index at Expiration

Time−to−Maturity τ : 84, Current Index Level Zt: 42.455

Note: This ﬁgure plots nonparametric SPD estimates (solid lines) of the S&P 500 index at maturities of 21, 42, and

84 days, and for two current VIX levels of 23.490 and 42.455. Dotted lines around each SPD estimate are 95 percent

conﬁdence intervals constructed from the asymptotic distribution theory in (6). The two levels of the current VIX

index are obtained by the 20 and 80 percent quantiles of the VIX time series of 2009 and correspond to low and high

levels of the VIX, respectively.

40

Figure 5: SPDs of S&P 500 Conditional on VIX vs Estimates that Ignore VIX

650 700 750 800 850 900 950 1000 1050 1100 1150

0

1

2

3

4

5

6

x 10

−3

S&P 500 index at Expiration

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

Time−to−Maturity τ : 42, Current Index Level Zt: 23.49

p(S

T

|τ,S

t

)

95% Upper

95% Lower

p(S

T

|τ,S

t

,Z

t

)

95% Upper

95% Lower

650 700 750 800 850 900 950 1000 1050 1100 1150

0

1

2

3

4

5

6

x 10

−3

S&P 500 index at Expiration

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

Time−to−Maturity τ : 42, Current Index Level Zt: 28.905

p(S

T

|τ,S

t

)

95% Upper

95% Lower

p(S

T

|τ,S

t

,Z

t

)

95% Upper

95% Lower

650 700 750 800 850 900 950 1000 1050 1100 1150

0

1

2

3

4

5

6

x 10

−3

S&P 500 index at Expiration

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

Time−to−Maturity τ : 42, Current Index Level Zt: 42.455

p(S

T

|τ,S

t

)

95% Upper

95% Lower

p(S

T

|τ,S

t

,Z

t

)

95% Upper

95% Lower

Note: This ﬁgure overlays nonparametric SPD estimates (solid lines) of the S&P 500 index conditional on the current

VIX level with the nonparametric estimates that ignore the VIX variable (dashed lines). Dotted and dash-dotted lines

are 95 percent conﬁdence intervals constructed from the asymptotic distribution theory in (6). The current values of

the VIX are ﬁxed at 20%, 50% or 80% quantile of the VIX time series in 2009. The time-to-maturity is ﬁxed at 42

days.

41

Figure 6: Estimates of the State-Price Density of the VIX

15 20 25 30 35 40 45 50 55 60

0

0.01

0.02

0.03

0.04

0.05

0.06

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

VIX at Expiration

Time−to−Maturity τ: 21, Current Index Level Zt: 23.49

15 20 25 30 35 40 45 50 55 60

0

0.01

0.02

0.03

0.04

0.05

0.06

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

VIX at Expiration

Time−to−Maturity τ: 21, Current Index Level Zt: 42.455

15 20 25 30 35 40 45 50 55 60

0

0.01

0.02

0.03

0.04

0.05

0.06

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

VIX at Expiration

Time−to−Maturity τ: 42, Current Index Level Zt: 23.49

15 20 25 30 35 40 45 50 55 60

0

0.01

0.02

0.03

0.04

0.05

0.06

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

VIX at Expiration

Time−to−Maturity τ: 42, Current Index Level Zt: 42.455

15 20 25 30 35 40 45 50 55 60

0

0.01

0.02

0.03

0.04

0.05

0.06

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

VIX at Expiration

Time−to−Maturity τ: 84, Current Index Level Zt: 23.49

15 20 25 30 35 40 45 50 55 60

0

0.01

0.02

0.03

0.04

0.05

0.06

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

VIX at Expiration

Time−to−Maturity τ: 84, Current Index Level Zt: 42.455

Note: This ﬁgure plots nonparametric SPD estimates (solid lines) of the VIX at maturities of 21, 42, and 84 days, and

for two current VIX levels at 23.490 and 42.455. Dotted lines around each SPD estimate are 95 percent conﬁdence

intervals constructed from the asymptotic distribution theory in (6). The two levels of the current VIX index are the

20th and 80th percent quantiles of the VIX time series of 2009.

42

Figure 7: Nonparametric Estimates of Pricing Kernels

0.9 0.92 0.94 0.96 0.98 1 1.02 1.04 1.06 1.08 1.1

0

2

4

6

8

10

12

14

16

18

20

S

c

a

l

e

d

M

a

r

g

i

n

a

l

R

a

t

e

o

f

S

u

b

s

t

i

t

u

t

i

o

n

Wealth

06/01/2006−11/30/2007; Zt=11.56

MRS

95% Upper

95% Lower

0.9 0.92 0.94 0.96 0.98 1 1.02 1.04 1.06 1.08 1.1

0

2

4

6

8

10

12

14

16

18

20

S

c

a

l

e

d

M

a

r

g

i

n

a

l

R

a

t

e

o

f

S

u

b

s

t

i

t

u

t

i

o

n

Wealth

06/01/2006−11/30/2007; Zt=20.03

MRS

95% Upper

95% Lower

0.9 0.92 0.94 0.96 0.98 1 1.02 1.04 1.06 1.08 1.1

0

2

4

6

8

10

12

14

16

18

20

S

c

a

l

e

d

M

a

r

g

i

n

a

l

R

a

t

e

o

f

S

u

b

s

t

i

t

u

t

i

o

n

Wealth

06/01/2009−11/30/2010; Zt=20.03

MRS

95% Upper

95% Lower

0.9 0.92 0.94 0.96 0.98 1 1.02 1.04 1.06 1.08 1.1

0

2

4

6

8

10

12

14

16

18

20

S

c

a

l

e

d

M

a

r

g

i

n

a

l

R

a

t

e

o

f

S

u

b

s

t

i

t

u

t

i

o

n

Wealth

06/01/2009−11/30/2010; Zt=35.455

MRS

95% Upper

95% Lower

Note: In the upper panel, we plot nonparametric estimates of the pricing kernels for the 18-month sample period

before the 2008 ﬁnancial crisis. Correspondingly, the post-crisis estimates for the 18-month sample period from June

1, 2009 to December 30, 2010 are reported in the lower panel. Dotted lines around each pricing kernel are 95 percent

conﬁdence intervals. The two levels of Zt, 11.56 and 20.03, represent low and high levels of VIX pre-crisis, whereas

the low and high levels of VIX after the crisis are chosen to be 20.03 and 35.455, respectively. In addition, we choose

time-to-maturity τ to be 42 days when reporting the estimates.

43

Figure 8: Nonparametric Versus Parametric Estimates of the SPDs of the S&P

−0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3

0

1

2

3

4

5

6

7

8

9

10

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

Log−Return

Time−to−Maturity τ : 21, Current Index Level Zt: 23.49

Nonparametric

95% Upper

95% Lower

Parametric: SVCJ1

Parametric: SVCJ2

−0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3

0

1

2

3

4

5

6

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

Log−Return

Time−to−Maturity τ : 21, Current Index Level Zt: 42.455

Nonparametric

95% Upper

95% Lower

Parametric: SVCJ1

Parametric: SVCJ2

−0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3

0

1

2

3

4

5

6

7

8

9

10

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

Log−Return

Time−to−Maturity τ : 42, Current Index Level Zt: 23.49

Nonparametric

95% Upper

95% Lower

Parametric: SVCJ1

Parametric: SVCJ2

−0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3

0

1

2

3

4

5

6

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

Log−Return

Time−to−Maturity τ : 42, Current Index Level Zt: 42.455

Nonparametric

95% Upper

95% Lower

Parametric: SVCJ1

Parametric: SVCJ2

−0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3

0

1

2

3

4

5

6

7

8

9

10

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

Log−Return

Time−to−Maturity τ : 84, Current Index Level Zt: 23.49

Nonparametric

95% Upper

95% Lower

Parametric: SVCJ1

Parametric: SVCJ2

−0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3

0

1

2

3

4

5

6

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

Log−Return

Time−to−Maturity τ : 84, Current Index Level Zt: 42.455

Nonparametric

95% Upper

95% Lower

Parametric: SVCJ1

Parametric: SVCJ2

Note: This ﬁgure overlays nonparametric SPD estimates (solid lines) of the S&P 500 index with the corresponding

parametric SPDs (dashed lines) with time-to-maturity equal to 21, 42, and 84 days, and for two current VIX index levels

of 23.490 and 42.455. Dotted lines are 95 percent conﬁdence intervals constructed from the asymptotic distribution

theory in (6). The parametric SPDs are obtained using formulas in Appendix B. The two levels of the current VIX

are the 20th and 80th percent quantiles of the VIX time series of 2009.

44

Figure 9: Nonparametric Versus Parametric Estimates of the SPDs of VIX

15 20 25 30 35 40 45 50 55 60

−0.01

0

0.01

0.02

0.03

0.04

0.05

0.06

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

VIX at Expiration

Time−to−Maturity τ: 21, Current Index Level Zt: 23.49

Nonparametric

95% Upper

95% Lower

Parametric: SVCJ1

Parametric: SVCJ2

15 20 25 30 35 40 45 50 55 60

−0.01

0

0.01

0.02

0.03

0.04

0.05

0.06

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

VIX at Expiration

Time−to−Maturity τ: 21, Current Index Level Zt: 42.455

Nonparametric

95% Upper

95% Lower

Parametric: SVCJ1

Parametric: SVCJ2

15 20 25 30 35 40 45 50 55 60

−0.01

0

0.01

0.02

0.03

0.04

0.05

0.06

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

VIX at Expiration

Time−to−Maturity τ: 42, Current Index Level Zt: 23.49

Nonparametric

95% Upper

95% Lower

Parametric: SVCJ1

Parametric: SVCJ2

15 20 25 30 35 40 45 50 55 60

−0.01

0

0.01

0.02

0.03

0.04

0.05

0.06

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

VIX at Expiration

Time−to−Maturity τ: 42, Current Index Level Zt: 42.455

Nonparametric

95% Upper

95% Lower

Parametric: SVCJ1

Parametric: SVCJ2

15 20 25 30 35 40 45 50 55 60

−0.01

0

0.01

0.02

0.03

0.04

0.05

0.06

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

VIX at Expiration

Time−to−Maturity τ: 84, Current Index Level Zt: 23.49

Nonparametric

95% Upper

95% Lower

Parametric: SVCJ1

Parametric: SVCJ2

15 20 25 30 35 40 45 50 55 60

−0.01

0

0.01

0.02

0.03

0.04

0.05

0.06

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

VIX at Expiration

Time−to−Maturity τ: 84, Current Index Level Zt: 42.455

Nonparametric

95% Upper

95% Lower

Parametric: SVCJ1

Parametric: SVCJ2

Note: This ﬁgure overlays nonparametric SPD estimates (solid lines) of the VIX with the corresponding parametric

SPDs (dashed lines) with time-to-maturity equal to 21, 42, and 84 days, and for two current VIX index levels of 23.490

and 42.455. Dotted lines are 95 percent conﬁdence intervals constructed from the asymptotic distribution theory in

(6). The parametric SPDs are obtained using formulas in Appendix B. The two levels of the current VIX are the 20th

and 80th percent quantiles of the VIX time series of 2009.

45

Figure 10: Estimates of S&P 500 Index SPDs When VIX Options are Unavailable

650 700 750 800 850 900 950 1000 1050 1100 1150

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3

1.4

1.5

x 10

−3

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

S&P 500 index at Expiration

Time−to−Maturity τ: 168

Zt=23.49

95% Upper

95% Lower

Zt=42.46

95% Upper

95% Lower

650 700 750 800 850 900 950 1000 1050 1100 1150

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3

1.4

1.5

x 10

−3

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

S&P 500 index at Expiration

Time−to−Maturity τ: 231

Zt=23.49

95% Upper

95% Lower

Zt=42.46

95% Upper

95% Lower

650 700 750 800 850 900 950 1000 1050 1100 1150

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3

1.4

1.5

x 10

−3

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

S&P 500 index at Expiration

Time−to−Maturity τ: 294

Zt=23.49

95% Upper

95% Lower

Zt=42.46

95% Upper

95% Lower

650 700 750 800 850 900 950 1000 1050 1100 1150

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3

1.4

1.5

x 10

−3

D

e

n

s

i

t

y

:

P

r

o

b

a

b

i

l

i

t

y

P

e

r

c

e

n

t

S&P 500 index at Expiration

Time−to−Maturity τ: 357

Zt=23.49

95% Upper

95% Lower

Zt=42.46

95% Upper

95% Lower

Note: This ﬁgure plots nonparametric SPD estimates of the S&P 500 index at long maturities of 168, 231, 294, and

357 days, and for two current VIX levels of 23.490 (solid lines) and 42.455 (dashed lines), respectively. Dotted and

dash-dotted lines are 95 percent conﬁdence intervals constructed from the asymptotic distribution theory in (7). The

two levels of the current VIX index are obtained by the 20 and 80 percent quantiles of the VIX time series of 2009 and

correspond to low and high levels of the VIX, respectively.

46

Figure 11: Robustness Checks

700

800

900

1000

1100

1200

0

50

100

150

0

50

100

150

200

250

300

S&P Level St

Time−to−Maturity

S

&

P

5

0

0

O

p

t

i

o

n

P

i

c

e

s

800

900

1000

1100

1200

20

40

60

80

100

120

−1

−0.5

0

0.5

1

S&P Level St

Time−to−Maturity

S

&

P

5

0

0

O

p

t

i

o

n

P

i

c

e

s

%

E

r

r

o

r

10

20

30

40

50

60

0

50

100

150

0.005

0.01

0.015

0.02

0.025

0.03

VIX Level Zt

Time−to−Maturity

S

P

D

o

f

V

I

X

10

20

30

40

50

60

0

50

100

150

−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1

VIX Level Zt

Time−to−Maturity

S

P

D

o

f

V

I

X

%

E

r

r

o

r

Note: In the left panel, we plot nonparametric estimates of the S&P 500 option prices and of the state-price density of

the VIX using the two dimension reduction techniques. The right panel reports the percentage errors relative to the

estimates without using the two dimension reduction techniques.

47

A Tale of Two Option Markets: State-Price Densities Implied from S&P 500 and VIX Option Prices∗

Zhaogang Song† Federal Reserve Board Dacheng Xiu‡ Chicago Booth

This Version: January 2012

Abstract The S&P 500 and VIX option markets are closely connected as both options depend on the volatility dynamics. Capturing information in both option prices by nonparametric state-price densities (SPDs), we look into the dynamics of the index and its volatility, along with interactions between the two option markets. We ﬁnd that SPDs of the index strongly depend on the current VIX level, and that such dependence is driven by information implied from VIX options beyond VIX time series, such as volatility of volatility and volatility skewness. In addition, SPDs of the VIX document three features of its risk-neutral dynamics, including positive skewness, meanreversion, and high persistence. Moreover, the pricing kernel estimates exhibit a U-shape when the current VIX level is high and a decreasing pattern otherwise, both pre- and post-crisis. This pattern implies that stochastic volatility may be the missing key state variable for the preference of agents, which is responsible for the puzzling U-shape. Finally, we conduct nonparametric speciﬁcation tests and ﬁnd that the state-of-the-art stochastic volatility models in the literature cannot capture the S&P 500 and VIX option prices simultaneously. The identiﬁed asymmetric mean-reversion rate of volatility suggests non-aﬃne speciﬁcation as a necessary extension. Key Words: Options, Pricing Kernel, State-Price Density, VIX JEL classiﬁcation: G12,G13

∗

We beneﬁted from discussions with seminar participants at Northwestern University and Princeton University.

This research was funded in part by the Fama-Miller Center for Research in Finance at the University of Chicago Booth School of Business. The views expressed herein are the authors’ and do not necessarily reﬂect the opinions of the Board of Governors of the Federal Reserve System.

†

Monetary and Financial Market Analysis Section, Board of Governors of the Federal Reserve System, Mail Stop

**165, 20th Street and Constitution Avenue, Washington, DC, 20551. E-mail: Zhaogang.Song@frb.gov.
**

‡

University of Chicago Booth School of Business, 5807 S. Woodlawn Avenue, Chicago, IL 60637.

Email:

dacheng.xiu@chicagobooth.edu.

Electronic copy available at: http://ssrn.com/abstract=2013381

1

Introduction

By reconciling evidence from the S&P 500 index and its options, empirical option pricing literature has documented stochastic volatility as an indispensable factor of the underlying dynamics of index returns. Nevertheless, determining the volatility dynamics is not straightforward because, as A¨ ıtSahalia et al. (2001) point out, volatility is a nontraded asset, and its risk-neutral behavior cannot be solely identiﬁed. Therefore, most studies in the literature either rely on strong beliefs in their favorite parametric models restricted within the aﬃne class, such as Bakshi et al. (1997), Bates (2000), Pan (2002), Eraker (2004), Broadie et al. (2007), Duan and Yeh (2011), and Amengual and Xiu (2012), or ignore the unobserved volatility factor when conducting nonparametric analysis of the index dynamics, as in A¨ ıt-Sahalia and Lo (1998), A¨ ıt-Sahalia et al. (2001), Buraschi and Jackwerth (2001), Jackwerth and Rubinstein (1996), and Fan and Mancini (2009). The lack of observable and tradable volatility has changed substantially since the Chicago Board of Options Exchange (CBOE) introduced the Volatility Index (VIX) in 1993,1 and VIX derivatives such as futures and options in 2004 and 2006, respectively. The VIX provides investors with a direct measure of the volatility and VIX derivatives oﬀer investors convenient instruments to trade the volatility of S&P 500 index.2 The VIX, referred to as the fear gauge, is constantly exposed in the media spotlight. Also, VIX options achieve huge liquidity, becoming the third most active contracts at CBOE as of October 2011. Because the VIX is derived from S&P 500 options as the square root of the risk-neutral expectation of the return’s average variance over the next 30 calendar days, it uncovers the unobservable volatility in a forward-looking way. In addition, VIX options are particularly informative about the distribution of the future VIX level. Consequently, two option markets exist simultaneously that both depend on the volatility dynamics. Capturing information in both option prices by nonparametric state-price densities (SPDs), we investigate dynamics of the index and its volatility, as well as the interactions between the two option markets. These SPDs, also known as the prices of Arrow-Debreu securities, summarize all the information of the economic equilibrium in an uncertain environment.3 They not only enable convenient pricing of vanilla options, but also allow consistent no-arbitrage pricing of exotic and less-liquid securities such as over-the-counter derivatives. Moreover, distinct from the parametric

1

The VIX, from its inception, was calculated from S&P 500 index options by inverting the Black-Scholes formula.

In 2003, the CBOE amended this approach and adopted a model-free method to calculate the VIX. 2 Before the introduction of VIX derivatives, to trade volatility, investors have to take positions of option portfolios, such as straddles or strangles. 3 SPDs give for each state x the price of a security paying one dollar if the state falls between x and x + dx. See A¨ ıt-Sahalia and Lo (1998) for detailed discussions.

1

Electronic copy available at: http://ssrn.com/abstract=2013381

studies in the literature, our nonparametric analysis of SPDs unveils potential misspeciﬁcations of the parametric forms, and may be advocated as a prerequisite to the construction of parsimonious models. Nonparametric studies of SPDs were innovated by A¨ ıt-Sahalia and Lo (1998) and Jackwerth and Rubinstein (1996), disregarding the stochastic volatility factor.4 Although it may be possible to estimate SPDs of the S&P 500 index conditional on an ex-post volatility proxy ﬁltered from historical time series (see Li and Zhao (2009) for such a procedure with interest rate caps), SPDs of the volatility cannot be determined because options cannot be written on unobservable variables directly. In contrast, using information in both the S&P 500 and VIX option markets, we recover SPDs of both the S&P 500 index and VIX and hence indirectly spotlight the unobservable volatility dynamics without parametric assumptions. Our empirical study of the nonparametric SPDs documents several important ﬁndings about the dynamics of the index and its volatility and information content of the two option markets. First, we ﬁnd that SPDs of the S&P 500 index strongly depend on the current VIX level, not only in the short run but also for the long term, when VIX options become unavailable. To study the documented short-run dependence and illustrate how investors employ information in VIX options, we regress the future average volatility implied from S&P 500 densities on the VIX and moments of the VIX densities extracted from VIX option prices. We ﬁnd that the short-run dependence is driven not only by the VIX, but also by information in its SPDs. This ﬁnding implies that the availability of VIX options indeed deliver incremental information for market participants to predict future index dynamics and price S&P 500 options accordingly. To understand the documented long-term dependence, we also implement a simple martingale procedure by which investors interpolate implied volatility of today’s long-maturity S&P 500 options using yesterday’s implied volatility surface. Regression results of implied volatilities of realized option prices on interpolated implied volatilities show that the martingale interpolation indeed agrees with the market practice and may oﬀer a potential justiﬁcation for the long-run dependence. The nonparametric SPDs of A¨ ıt-Sahalia and Lo (1998), in the absence of the stochastic volatility factor, diﬀer considerably from our estimates of SPDs conditional on the VIX when the current VIX level is either low or high, while only being approximately identical for the average VIX level. Therefore, with today’s VIX level in the high and low ranges, our nonparametric SPD conditional on the VIX may provide more-accurate information about the future dynamics of the S&P 500 index. Second, we document several empirical features of the risk-neutral dynamics of the VIX. In

4

Many follow-up studies have been conducted based on these nonparametric SPDs, such as A¨ ıt-Sahalia and Lo

(2000), A¨ ıt-Sahalia et al. (2001), Chabi-Yo et al. (2008), and Bakshi et al. (2010)

2

but are more ﬂexible in the speciﬁcation of jumps in the return dynamics. while determining the payoﬀ function of a VIX option. and Broadie et al. and Bakshi et al. we estimate pricing kernels conditional on the VIX. and Mencia and Sentana (2009)). Detemple and Osakwe (2000). as they both rely on the volatility factor: the volatility aﬀects the distribution of the S&P 500 index. we ﬁnd that SPDs of the VIX are positively skewed across all maturities and current levels of the VIX. Chabi-Yo (2011). and Christoﬀersen et al. Combining SPDs of both the S&P 500 index and the VIX. Chabi-Yo et al. The U-shape becomes visible when the market experiences a high volatility.particular. Building upon the diﬀerence between parametric SPD estimates and nonparametric ones. Third. Our ﬁnding conﬁrms these patterns under the pricing measure without any parametric restrictions. Nevertheless. that the missing state variables in pricing kernels may result in the U-shape. contradicts standard economic theory. they either focus on pricing S&P 500 options (see e. in which the pricing kernel. Eraker (2004). and Broadie et al. along with joint pricing S&P 500 and VIX options (Amengual and Xiu (2012)). Modeling both option prices jointly is important. most studies in the option-pricing literature treat them separately. the risk-neutral dynamics of the VIX exhibit both mean-reversion and persistence. We ﬁnd that the estimated pricing kernel exhibits a U-shape conditional on a high VIX level. The U-shape of the pricing kernel. Bates (2000). (2010a). our test statistics reject 3 .g. (2007)) or pricing VIX options on top of a standalone process for VIX (e. its risk-neutral behavior is not crystal clear. we test the speciﬁcation of stateof-the-art models that resemble those proposed by Pan (2002). our result.g. Grunbichler and Longstaﬀ (1996). This ﬁnding echoes the conclusions of equilibrium-based parametric models discussed in Jackwerth and Brown (2001). Without restricting the form of pricing kernels. (2008). Our empirical result suggests that the shape of the pricing kernel is signiﬁcantly aﬀected by the stochastic volatility. Whaley (1993). In addition. decreases in market index return. it remains unclear how well they capture empirical features of these option prices. Jackwerth (2000). for the periods before and after the 2008 ﬁnancial crisis. Pan (2002). while maintaing a decreasing pattern when the current VIX level is low for both preand post-crisis periods. however. Eraker (2004). our SPD estimates enable us to gauge the speciﬁcations of parametric models in pricing S&P 500 and VIX options simultaneously. (2010) as a puzzle. equal to the scaled marginal rate of substitution. described in A¨ ıt-Sahalia and Lo (2000). (2007). Although some parametric models have been developed to capture the evolution of S&P 500 index and VIX. Although the volatility process under the physical measure is well documented to display a mean-reverting pattern using historical time series (Mencia and Sentana (2009)). Finally. implies that stochastic volatility is the key but missing state variable of pricing kernels.

and the VIX as Zt . and an extensive panel of option prices. (2011). estimates of pricing kernels conditional on VIX. which further indicates non-aﬃne models may be a promising start point to jointly pricing the two options. derivations of parametric SPDs. including nonparametric SPD estimates of the S&P 500 index and VIX. we also construct test statistics to check whether the VIX SPD depends on the current information set only through the VIX.. with realized volatility constructed from high-frequency data as a proxy for the unobservable spot volatility. (2010b). we discuss the relationship of stochastic volatility process and VIX. In addition. S&P 500 returns. The rest of the paper is organized as follows. Engulatov et al. and parametric models for pricing S&P 500 options and VIX options consistently are introduced. Section 3 presents our empirical results. and construct our nonparametric estimators using a multivariate local linear approach. In contrast. respectively. and tests for robustness checks.1 State-Price Densities: Derivation. we identify the underlying connection between S&P 500 and VIX options and introduce the state-price densities that span the two markets. and point out similar evidence of asymmetric mean-reversion. and Test Two Option Markets and State-Price Densities In this section. estimate the drift and volatility functions nonparametrically. the option prices are determined by two SPDs p(ST |Ft ) and p(ZT |Ft ) of ST and ZT . on the other hand. Estimation.all these models under consideration. the exact inadequacy — constant rate of mean-reversion — of a variety of aﬃne models. Section 4 concludes. The appendix provides technical assumptions. Our paper is also related to Boes et al.e. and nonparametric speciﬁcation analysis of parametric pricing models. This ﬁnding justiﬁes the VIX to be an eﬀective volatility measure. proofs. for which we ﬁnd no evidence of rejection. In section 2. we denote the log price of the S&P 500 index as St . who ﬁnd that non-aﬃne models outperform aﬃne square root models using realized volatilities. 2 2. (2007) that tries to estimate the SPDs of returns. and helps simplify our nonparametric analysis. Our result pinpoints. and more importantly VIX options. nonparametrically. To ﬁx ideas. which are particularly informative about volatility dynamics. our estimation method draws on the observable VIX. A speciﬁcation test based on the estimated SPDs is also proposed. Closer scrutiny of the tests and SPD estimates of the VIX suggest an asymmetric mean-reversion rate of volatility. where Ft summarizes the information up to time 4 . Because the payoﬀs of S&P 500 and VIX options only depend on their own underlying indices ST and ZT . Another relevant speciﬁcation analysis includes Christoﬀersen et al. volatility reverts relatively faster when its current level is low. analysis of the dependence of S&P 500 index densities on the current VIX level. i. how to obtain SPDs from option prices.

vt .e. we write the VIX option price H as a function of st . st . st . st . the function C (·) here is a composite function. Building upon the insight of Breeden and Litzenberger (1978). Nevertheless. x) and H(τ. i. zt . vt . H(τ. zt . y). x) ∂x2 . x) are put and call options with time-to-maturity τ and strike x. st .T denotes the quadratic variation of the log return process. st . Considering the functional relationship between Zt and Vt . 7 Strictly speaking.T |Ft ) = 2erτ τ Ft 0 P (τ. we denote the instantaneous volatility of St as Vt . the SPD of ST can be recovered from the second order derivative of C(τ. vt . z . s .. P (τ. (1) x=esT By the same reasoning and assumption. vt ) and p(zT |τ.6 we may rewrite the option prices as a function of zt .t. zt ) =erτ +sT ∂ 2 C(τ. vt ) = erτ +sT ∂ 2 C(τ. vt . which is diﬀerent from the previous call option pricing function. and Ft = erτ +st is the price of futures contracts. st .e. Britten-Jones and Neuberger (2000) and Carr and Wu (2009). see Section 3 for details. x) and C(τ. zt . which seems valid for the sample period we consider. x) with respect to x. st . st . x) dx + x2 ∞ Ft C(τ. zt ) =erτ ∂y 2 y=zT 5 We treat r as a constant here following our empirical setup. vt ) are of little use in studying the two option markets. and r is the constant risk-free rate between t and T 5 . st . vt .7 and take second order derivatives to obtain p(sT |τ. x) =e−rτ E Q (eST − x) |St = st . To infer these SPDs from option prices. x) dx + x2 where QVt. st . i. 5 . τ = T −t is the time-to-maturity.g. st . Vt = vt =e−rτ R + (esT − x)+ p(sT |τ. st . st . (2) y=zT Note that the derived SPDs of S and Z in (1) and (2) extract all the information in option prices that can be used to identify the risk neutral measure Q. x) ∂x2 x=esT 2 H(τ. and the dynamics of S and V . time-to-maturity τ and strike y. and obtain the SPD of ZT as p(zT |τ. We recycle it to simplify our notations. as the information regarding Vt is unobservable. see e. C(τ. vt ) = erτ ∂ 2 H(τ. y). and write the price of the S&P 500 option with maturity T and strike x as C(τ. p(sT |τ. vt . vt )dsT where p(sT |τ. y) ∂ t t p(zT |τ. these two densities p(sT |τ. 6 The CBOE constructs Zt form a portfolio of options weighted by strikes according to the formula: (Zt /100)2 = E Q (QVt. st .. vt ) is the SPD of ST . st . y) ∂y 2 . st .

Y = y . zt ) encapsulate all the price information in the two option markets. representing options in terms of St and Zt amounts to conditioning on a coarser information set generated by St and Zt . the approximation error is zero for most models in the literature as Zt is a deterministic and invertible function of Vt . They complement each other to reveal an intact picture of the index and volatility dynamics and interactions of the two markets. Diﬀerent from the multivariate kernel regression approach adopted by A¨ ıt-Sahalia and Lo (1998). a simple regression of the realized volatility (RV) against the lagged VIX and lagged RV indicates that the VIX has far more explanatory power compared with RV. 2. 8 Due to this forward-looking feature. for speciﬁcation analysis of parametric models. zt . e. and take derivatives to estimate the SPDs. which is a subset of the original information set generated by St and Vt . x) = E C τ = τ. Blair et al.2 Multivariate Local Linear Estimators Here we introduce our nonparametric estimation strategies to estimate the SPDs. We then construct nonparametric estimators of C and H. and used by investors as a measure of market volatility. First.8 Finally. zt ) and p(zT |τ. zt . although the VIX diﬀers from unobserved volatility as being forward-looking because it is constructed using S&P 500 options with maturities up to 30 calendar days. VIX is always exposed in the media spotlight. Zt = zt . X = x ˜ ˜ H(τ. which enjoys better boundary performance and reduces the asymptotic bias without increasing asymptotic variance (Fan and Gijbels (1996)). St = st .g. Zt = zt . St = st . as a function of Vt and St . st . Second. Therefore. st . st . 6 . so that estimators for SPDs can be obtained simply by a ﬁrst-order diﬀerentiation with respect to the strike. Equating the two information sets assumes that Zt .5 for details. In summary. we have used S&P 500 option prices with maturities up to 30 calendar days when changing the information set to that generated by St and Zt . In fact. local linear regression provides a closed-form estimator for both the price function and its ﬁrst order derivatives. (2001)). we use the multivariate local linear method. Moreover. Third. the approximation errors are expected to become smaller and even negligible as time-to-maturity increases. To ﬁx ideas. st . see Section 2. zt ) in practice with τ less than 30 calendar days. y) = E H τ = τ. such that ˜ ˜ C(τ. implying the importance of the VIX in capturing dynamics of the unobserved volatility (see. st . state-price densities p(sT |τ. we ˜ ˜ assume the observed prices C and H are contaminated with observation errors. is invertible with respect to Vt ..with an information set that is fully observable. caution is needed when employing p(sT |τ. Such an approximation is justiﬁable for a few reasons.

0. 1) . 0) and the estimator p(s |τ. z) for the SPD of St is p(s |τ. However. This bandwidth choice ensures that the nonparametric pricing function achieves the optimal rate of convergence in the mean-squared sense among all possible nonparametric estimators for option prices. s. α β (1+4)×1 = Ω KΩ −1 Ω KC (4) The nonparametric local linear estimator for the option pricing function C(τ. with e1 = (1. we choose the Gaussian kernel as k (·). hs . 7 . x) . In our empirical implementations. and x) as hj = cj s Uj n−1/6 . for example.. Xi ) and Ci are the characteristics and price of the i−th option in sample. z. . . but centers the asymptotic distribution at zero. Sti . z. we set the bandwidth hj (j = τ. s. min α. this rate leads to an asymptotic bias term for the estimator C (·) and p (·).C = . s. z. Similar to Li and Zhao (2009). s. z. The nonparametric estimator H (·) and p(z |τ. s. (5) where e4 = (0. as shown by Fan and Gijbels (1996). . x) is C(τ. and x). Ω= . . Kh (Un − u) . s. Hence. Cn K= Kh (U1 − u) . Zti . hx ): Kh (Ui − u) = 1 k hτ τi − τ hτ 1 k hs Sti − s hs 1 k hz Zti − z hz 1 k hx Xi − x hx (3) where k (·) is. we follow A¨ ıt-Sahalia and Lo (1998) to select cj = cj0 / ln (n). x) = α = e1 Ω KΩ −1 Ω KC. hz . 1 (Un − u) C1 . 0. which results in a slightly slower rate of convergence. we consider the following minimization problem. z) = e rτ +s ∂ β4 ∂x x=es =e rτ +s ∂ e4 (Ω KΩ)−1 Ω KC ∂x x=es . z. . . . where s Uj is the unconditional standard deviation of the regressor Uj (j = τ. Kh is a kernel function scaled by a bandwidth vector h = (hτ . 0. s. Moreover. 0. s. The minimizer has a closedform representation: where 1 (U1 − u) .To estimate the pricing function C as a function of u = (τ. z) can be constructed similarly. the density of standard normal distribution.β i=1 n 2 Ci − α − β (Ui − u) Kh (Ui − u) where Ui = (τ i . .

x) and s2 (τ. m) ∂m2 m=esT /est . 2. x/es ) = es C(τ. z) − p(s |τ. and write the SPD of ST as p(sT |τ. z. s. s. s. x) /πC (τ. zt ) = erτ +sT −st ¯ ∂ 2 C(τ. y) /πH (τ. s. respectively. s. s. y) are conditional variances for the local linear regressions of C C H and H on their state variables respectively. z. Furthermore. The estimators s2 (·) and s2 (·) for s2 (·) and s2 (·) can be constructed using C H C H similar nonparametric regressions. H where s2 (τ. m) where m = x/es represents the moneyness of the S&P 500 option. s. y) . but also on interest rates and dividends. z) 3 0. x) = es C(τ. z) − p(z |τ. following many existing studies such as Li and Zhao (2009). Suppose the sample sizes of the nonparametric SPD estimates of the S&P 500 index and VIX are nC and nH . we obtain the ¯ estimate of C(τ. Because interest rates and dividends do not vary much within the period of our empirical studies (see Section 3 for details). z. z. C nH hy hτ hs hz hy −→ N d p(z |τ. z.the constant cj0 is chosen by minimizing the ﬁnite-sample mean-squared error of the estimator via simulations. z. Consequently. s. In the most general forms. strike. s. VIX. s. e2rτ k 2 (c) dc ˙ ck (c) + k(c) 2 2 dc / k (c) c2 dc s2 (τ. we assume the S&P 500 option price is homogeneous of degree one in the current price level: ¯ C(τ. s. z.3 Dimension Reduction One of the major issues of nonparametric estimation is the curse of dimensionality. z. x) . and the S&P 500 index. y) are marginal densities of these variables. 0. x) and πH (τ. e2rτ k 2 (c) dc 1/2 ˙ ck (c) + k(c) dc / k (c) c2 dc s2 (τ. Using the equivalent kernels introduced in Fan and Gijbels (1996) and following the derivation in A¨ ıt-Sahalia and Lo (1998). z) −→ N 1/2 d 3 1/2 (6) 2 2 0. x) through multiplying the nonparametric estimate of C (·) by es . The rate of convergence decreases rapidly as the dimension of state variables increases. s. s. z. z. s. zt . z. st . z. and πC (τ. we take them as constants. the pricing functions C (·) and H (·) depend not only on time-to-maturity. 8 . we obtain the asymptotic distributions of these estimators: nC hx (hτ hs hz hx )1/2 p(s |τ.

st .e. the ratio between risk-neutral and physical densities.e. Zt is independent of St . 9 . the pricing kernel is supposed to be π(sT . zt ) = EST (π(sT . This assumption further implies that the SPD of ZT . which also reﬂects the market’s aggregation of risk preferences implicit in option prices. we assume that the information about Zt in St is fully incorporated into Zt . zt ) denotes the projection onto ST . depends on St only through Zt . and EST (·|st . and the results suggest that they are not rejected for the sample we consider.. zT |τ. zt ) where p(sT |τ. zt ) is the conditional physical density of ST . zt ) p(sT |τ. which cannot be identiﬁed nonparametrically. In other words. z) = 1 k bs STi − s bs n i=1 1 k bs Sti − s bs Sti − s bs 1 k bz 1 k bz Zti − z bz Zti − z bz 1 k bs with bandwidths given by the following: bj = cj σj n−1/(4+d) .4 Nonparametric Estimation of Pricing Kernels In the equilibrium model of A¨ ıt-Sahalia and Lo (2000) with a representative agent. Thus. z where σj is the unconditional standard deviation of the data.As for VIX options. and d = 2 for those in the denominator. for any t > t. st . While estimating the risk neutral density p(sT |τ. st . zt ) using the time series of the S&P 500 index and VIX based on the kernel method. we estimate p(sT |τ. the number of state variables for the SPD of VIX is also decreased by one. zt ) as π(sT |τ. zt ) = p(sT |τ. Nevertheless. s. p(zT |τ. conditional on Zt . obtained from the VIX option prices. i. st . with Ti − ti = T − t ﬁxed. Zti }n . j = s . Suppose we have time series {STi . Sti . s. st . zt ). st . s. and d represents the dimension: d = 3 for those bj s that appear in the numerator. is—up to a scaled factor—the marginal rate of substitution. cj is a constant. In an economy with two state variables. i. st . zt )|st .. the pricing kernel. zt ) = p(zT |τ. Distinct from A¨ ıt-Sahalia and Lo (2000) and Jackwerth (2000). zt ) is the conditional risk neutral density. p(sT |τ. We formally test these two assumptions in Section 3. st . zt ) from option prices. st . zT |τ. zt ). st . we may consider an integrated pricing kernel π(sT |τ. our pricing kernel π(·|·) is conditional on the VIX so that it can reﬂect the risk preferences with respect to the volatility. z) as: ˜ n i=1 p(s |τ. 2. We construct the estimator of the density i=1 p(s |τ. st .

In particular. as the two option markets are highly integrated. z) . 2. Q 10 . s. z)/p(s |τ. z). z) = p(s |τ. dNt is a pure-jump process with intensity λt = λ0 + λ1 Vt . However.e. s. and Broadie et al. To add jumps in returns. (2007). Our SPDs can uncover two distinct aspects— returns and volatility—of the parametric model dynamics. smirk. and hence may shed light on the inadequacy of the model.Consequently. Eraker (2004). s. so that the asymptotic variance can be easily estimated. s. λ1 = 0.e. the convergence rate is higher for p(s |τ. z) than for p(s |τ. i. σS ). and which characteristics we should focus on in order to improve the model speciﬁcation. The last two models incorporate the same exponentially distributed jumps in volatility with mean βV : Q JV ∼ exp(βV ). z) p(s |τ. s.5 Speciﬁcation Tests Based on State-Price Densities The estimated SPDs enable us to evaluate the performance of the state-of-the-art parametric models under the uniﬁed framework of pricing S&P 500 and VIX options. The benchmark model is the Heston stochastic volatility model (SV) without jumps in either returns or volatility. s. i. or both. JS and JV are random jump sizes. z) is identical to that of p(s |τ. or SVJ2: JS ∼ − exp(β− ) with probability 1 − q where the jump intensity is constant λ0 . our pricing kernel can be estimated by π(s |τ. comparing our nonparametric SPD estimates with the parametric ones illustrates whether the rejection is mainly due to modeling S&P 500 options. few paper in the literature have discussed modeling them jointly. and index option returns. we consider two types of speciﬁcation: exp(β+ ) with probability q Q Q SVJ1: JS ∼ N (µS . and the asymptotic distribution of π(s |τ. s. As discussed in A¨ ıt-Sahalia and Lo (2000). λ0 = λ1 = 0.. It is important to check whether they can capture the empirical features of VIX options before employing them in practice. The general model we consider is speciﬁed under the risk-neutral measure Q as 1 dSt = (r − d − Vt )dt + 2 dVt = κ(ξ − Vt )dt + σ Q Vt dWtQ + JS dNt − µλt dt Q Q Vt dBt + JV dNt (7) Q Q Q Q where WtQ and Bt are standard Brownian motions satisfying E(dWtQ dBt ) = ρdt. z). We bring together and test a variety of stochastic volatility models. and µ = E(eJS − 1). s. VIX options. including those discussed in Pan (2002). which are shown to be successful in capturing empirical features of S&P 500 options such as volatility smile.

as shown by Britten-Jones and Neuberger (2000) and Carr and Wu (2009): 1 Q 2 Zt = 104 · Et τ t+τ Vs ds + t Ns ≥0. due to their diﬀerences in jump speciﬁcation for returns. If the null hypothesis H0 is rejected. On top of (8).θ). Zt )} = 1 against the alternative hypothesis HA . The intensity in these cases is linear in Vt . we propose to test the joint null hypothesis H0 . respectively (see Appendix B for details of derivation). Zt .e. st . HA : Pr {p(mT |τ . zt . Zt ). p(ZT |τ . H0 : Pr {p(mT |τ . i. see Appendix B for the formulae of a and b and details of derivation. Zt .and we label them as SVCJ1 and SVCJ2. In particular. imply the dynamics of (St . these approximation errors are negligible under commonly used models and model parameters. 9 The closed-form pricing formulae have been derived by Amengual and Xiu (2012) for more-general models. Zt ) for which we perform our nonparametric speciﬁcation analysis of the model (7). see also Sepp (2008) for derivation under the Heston model.9 These formulae regarding the VIX and its option prices can greatly simplify the parametric inference. homogeneous of degree 1 in St for S&P 500 options and the dependence of VIX densities on St only through Zt . Zt . θ 0 ) = p(ZT |τ . t≤s≤t+τ Q (JS )2 = aVt + b (8) where a and b are constants determined by the speciﬁc model.. Zt .e. where mT = eST /eSt . zt ) and check whether the diﬀerences are statistically signiﬁcant. Zt )} < 1. we test the following two separate speciﬁcation hypotheses. However. p(sT |τ. Based on these simpliﬁcations. Vt ). zt ) and p(zT |τ. i. the jump speciﬁcations of return dynamics in model SVJ2 and SVCJ2 can incorporate heavy-tailed and skewed jumps. H0. θ 0 ) = p(mT |τ . as shown by Jiang and Tian (2005) and Carr and Wu (2009). p(ZT |τ . 11 . 10 The CBOE VIX contains numerical errors due to jumps and discretization and hence may not equal to the theoretical VIX in (8) exactly. For each model. both of which are satisﬁed for the parametric model (7). Zt . we derive VIX option prices in addition to S&P 500 option prices. the theoretical value of the VIX Zt can be calculated. Zt ). zt ) and p(zT |τ. zt . zt ). θ 0 ) = p(mT |τ . st . θ 0 ) = p(ZT |τ . Zt )} = 1.10 The dynamics of (St . as the unobservable volatility factor can be replaced by a function of the VIX. combined with (8). θ 0 ) = p(mT |τ . Compared with SVJ1 and SVCJ1. Recall that we employ two dimension-reduction techniques in obtaining p(sT |τ.θ) and p(zT |τ.. st .S&P : Pr {p(mT |τ . The idea of our nonparametric analysis is to compare these parametric SPDs with nonparametric estimates p(sT |τ. model (7) speciﬁes parametric forms of the SPDs of S&P 500 index and VIX.

VIX : Pr {p(ZT |τ . i.g. θ 0 ) = p(ZT |τ . Due to liquidity concerns. Second. and use the put-call parity to construct in-the-money call options from out-of-the-money put options. we 12 . The test statistics based on the diﬀerences between parametric and nonparametric SPDs are constructed and converge to convenient standard normal distributions under the null hypothesis. We follow the data-cleaning routine commonly used in literature. EST. Before delving into the details.1 Data We collect close prices of S&P 500 Index. and Table 3 presents summary statistics of the option prices.m. according to the closed-form formulae given in the Appendix. 3.e. VIX. we take the midquote as the option price. we ﬁx a sample path generated by the SVCJ2 model. For each option. and their options from the OptionMetrics group for the year 2009. as well as options with timeto-maturity of less than 10 days. and 4:00 p. First.H0. it is clear from Table 2.2 Monte Carlo Simulations We provide simulation evidence for our local linear estimators..539 VIX call options. Also. The best bid and oﬀer prices for options are quoted between 3:59 p. The last step is to eliminate option contracts that violate no-arbitrage conditions. 3 Empirical Results In this section. we estimate nonparametric SPDs implicit in the S&P 500 and VIX options. We then calculate option prices. 3. The Monte Carlo experiments are designed to match our empirical studies. Finally. First. we eliminate any options with zero open interests or trading volumes. Figure 1 plots the joint time series of the S&P 500 index and the VIX. observations with bid or ask prices smaller than 0. and 14. Therefore.05 are eliminated to mitigate the eﬀect of price discreteness. whether the model is consistent with S&P 500 option prices and VIX option prices separately. A¨ ıt-Sahalia and Lo (1998). including 128. Zt )} = 1. we introduce the dataset. see. The resulting sample covers a broad cross section of options. There is no such pattern for VIX options and hence we only consider VIX call options. we delete in-the-money options. e. Zt . we select the same option characteristics as those traded on CBOE in 2009. that in-the-money S&P 500 options are less liquid than out-of-the-money options. see Appendix A for details.. Table 1 provides their summary statistics. along with recovering their interactions.m.883 S&P 500 call options.

In Figure 5.pollute the prices with multiplicative measurement error following log-normal distribution with 1% standard deviation. with St ﬁxed at the sample mean. due to the fact that derivatives are estimated with slower rates of convergence. The 95% conﬁdence intervals are obtained through the asymptotic theory given in (6).3. All of these quantities and their percentage errors are reported in Figures 2 and 3. with Zt and X ﬁxed at their sample averages. our nonparametric SPD estimator may be more informative about the future dynamics of the S&P 500 index than the AL density. The low and high levels of the VIX are obtained as the 20% and 80% quantiles. 3. This pattern becomes stronger for longer maturities. The nonparametric estimators of VIX option prices and densities are evaluated similarly. and 84 days. Figure 4 shows that the S&P 500 index densities are negatively skewed across all maturities and VIX values. of the VIX in our sample. In contrast. and compare it with our density estimates conditional on diﬀerent VIX levels. Based on the generated sample. zt ) has pronounced spikes. 42. the index densities strongly depend on the VIX Zt . Figure 6 provides nonparametric estimates of the VIX SPDs for diﬀerent levels of Zt with τ equal to 21. 42. respectively. the density becomes more dispersed and more negatively skewed. Take τ = 21 as an example. Option prices below 0. as also shown in both Figures 2 and 3. and 84 days. the so-called curse of diﬀerentiation. We also calculate the index densities on the grid of τ and ST . 13 . indicating that the AL density estimator actually approximates our density estimator for medium VIX levels. to evaluate our density estimators. the index density p(sT |τ. These ﬁgures show that the nonparametric estimates are within 2% and 5% of their theoretical values for S&P 500 and VIX options. When Zt is low. st . Hence. the two densities are very close when Zt is at the medium level.05 are eliminated to mimic the reality. we evaluate our nonparametric estimators of index option pricing functions on the grid of time-to-maturity and current index level. More importantly. It is evident that the AL density estimate departs from ours substantially when Zt stays at high or low levels: our estimate is more compact with higher spikes for a low Zt but more dispersed with heavier tails for a high Zt .3 3. averaged over 500 replications. respectively. The errors for densities are slightly larger. As expected.1 State-Price Densities of S&P 500 Index and VIX Empirical Characteristics of State-Price Densities Figure 4 presents nonparametric SPD estimates of the S&P 500 index for both low and high levels of the VIX with time-to-maturity equal to 21. as Zt rises to the high level. We ﬁnd three important features of the risk-neutral densities of the VIX. we benchmark the nonparametric SPD estimates proposed by A¨ ıt-Sahalia and Lo (1998) (AL) that neglect the volatility variable.

Third. for which VIX options are available. 14 . We ﬁnd that the VIX level Zt has predictability for the S&P 500 option-implied future volatility at τ = 21 and 84. . as evidenced by the lower peaks at the left end of the VIX densities for longer maturities.T in VIX time series and VIX options. ZT can maintain at high levels with large probability as maturity increases from 21 to 42 days. Second. This change implies that the probability of ZT to stay at low levels conditional on a low Zt is high. For τ = 21. T + 21] by. 63 and 84. zt ). which is most apparent when Zt is at the high level. investors may use VIX option prices to obtain information about the distribution of the future VIX ZT and form expectations of the volatility of ST . . 3. but is less signiﬁcant with τ = 42 and 63 days. 42. investors may employ historical time series of the VIX up to time t to predict the volatility of the future index ST .2 Information from VIX Options about S&P 500 As shown in the last section. st . We compute the average volatility over some future 30-calendar-day interval [T. 252 (9) and the results are presented in Table 3. Moreover. the risk-neutral dynamics of the VIX exhibit high persistence. the probability of ZT mean-reverts to the medium level increases with maturity. the VIX density is more compact (dispersed). corresponding to information t. although less strongly. zt ). 42. Even conditional on a high Zt level when the mean reversion eﬀect is strong. SPDs of the S&P 500 index strongly depend on Zt at maturities of 21. for any ﬁxed maturity. . the VIX densities are positively skewed across all maturities and current VIX levels. showing that when Zt is high. which proxy the volatility of volatility and volatility skewness implicit in VIX options.T Vart [ST +21 ] − Vart [ST ] where Vart [ST ] is the conditional variance of ST implied from the density estimate p(sT |τ.t+τ = α + β1 Zt + β2 Stdt [Zt+τ ] + β3 Skewt [Zt+τ ] + εt . This mean-reverting behavior also shows up when Zt is at the low level. and 84 days. the risk-neutral dynamics of the VIX exhibit a mean-reverting pattern. the left tail of the VIX SPD becomes higher as maturity increases. We then regress VS on both Zt and characteristics of the VIX density. All of the quantities are annualized and estimated using nonparametric SPDs. conditional on a low (high) level of the current VIX. We now study whether VIX options contain information about market future volatility beyond VIX time series for the future volatility of index dynamics. The characteristics of VIX densities we employ are the conditional standard deviation Stdt [ZT ] and conditional skewness Skewt [ZT ] of p(zT |τ.3. Conditional on a low Zt . the shape of the VIX density changes slowly as maturity increases. Speciﬁcally. Such short-run dependence can be driven through two channels. Second. First. VS = t.First. t = 1. 2. respectively. we run the following regression VS t. .

950. For each date t. but also by the volatility of volatility and volatility skewness that can only be extracted from VIX option prices. and the martingale approach (MKT) for option prices.Most importantly.3. and C show that adding volatility skewness Skewt [Zt+τ ] to the regression against Zt and Stdt [Zt+τ ] leads to increases in the adjusted R2 . In summary. We repeat 15 . Hence. That is to say. additionally. hence the SX is expected to outperform the AL in. the availability of VIX options indeed delivers incremental information for market participants beyond the VIX time series. although they may ﬁt the cross section of option prices poorly on certain days. Moreover. 3. Intuitively. Panels C and D show that adding the volatility of volatility Stdt [Zt+τ ] to the regression with single explanatory variable Zt increases the adjusted R2 from less than 1% to over 12%. investors consider both the VIX level Zt and the whole distribution of volatility when forming expectations of future volatility and pricing S&P 500 options. For example. 21. both SX and AL tend to maintain more-stable pricing structures over time. 35. along with the SPDs. whereas perfectly for some other days in-sample. In terms of the comparison between SX and AL. which interpolates tomorrow’s implied volatility using today’s implied volatility surface. the SX approach further explores the predictability from the VIX. with the former more pronounced for maturities of 63 and 84 t.t+τ days and the latter more evident for all maturities except for 84 days. the MKT approach. for γ = 0 (in-sample). 28. we ﬁnd that the dependence of index densities on the VIX is driven not only by the VIX time series. 925. both the volatility of volatility Stdt [Zt+τ ] and volatility skewness Skewt [Zt+τ ] are signiﬁcant in explaining VS . As opposed to the MKT approach that replies exclusively on the cross section of options at each date. and hence they are expected to outperform the MKT approach in out-of-sample forecasting. and γ = 7. 14. Panel A of Table 4 reports the forecasting performance of the SX and AL density. on day t + γ. one of the potential advantages of the SX and AL approaches lies in their inclusion of historical prices of options with similar characteristics into the weighted prediction of option prices. and 42 days (out-of-sample) progressively. we adopt a preceding six-month window. by 20% for maturities equal to 21 and 63 days and 300% for 42 days.and out-of-sample for both option prices and densities. whereas Panel B presents the price prediction comparisons including. B. Panels A.3 In-Sample and Out-of-Sample Forecasts We benchmark our nonparametric estimator (SX) against two alternative methods discussed in A¨ ıtSahalia and Lo (1998): the nonparametric approach without volatility state (AL) for both densities and option prices. and 975 (the four strikes around the median of the sample) and maturities closest to four months. within which options are selected to predict the prices of traded options with strikes equal to 900.

the SX estimates are uniformly better over every horizon in predicting the future density when the realized density is represented by the SX density (the right part of Panel A). it is outperformed by the SX densities. Figure 7 reports the 11 We follow the NBER Business Cycle Committee and regard the 18-month duration of 2008 ﬁnancial crisis as from Nov 1. and the post-crisis period from June 1. This performance actually highlights. Regarding density forecasts. 2010. To highlight the potential change of average risk preference due to recent ﬁnancial crisis. at least for longer horizons. whereas over longer forecasting horizons. we also estimate the physical densities of the S&P 500 index. as the assumption of stability across time deteriorates the performance. the uniform better performance of the MKT over the AL method may be caused by the decreasing volatility path in 2009. but only when the realized density is represented by the AL method (the left part of Panel A). as shown by Panel B. which is in contrast with ﬁndings of A¨ ıt-Sahalia and Lo (1998). and 21 days). For forecasting option prices. and hence obtain estimates of the pricing kernels conditional on VIX by combining the SPDs under both measures.the procedure for each day t from the second half of the year. the VIX decreases monotonically and stays at lower levels for the entire second half of the year.4 Pricing Kernels Conditional on VIX In addition to SPDs under the risk-neutral measure. indirectly. Then we chose 18-month periods before and after the crisis. In contrast. and aggregate across t to obtain the root-mean-squared percentage diﬀerence between the predictions and the values realized on day t + γ. 3. 2009 to November 30. Moreover. 14. the simple martingale interpolation leads to a better ﬁt in-sample and for less than a month (τ = 0. In our sample period. Panel B shows that the MKT approach outperforms the AL approach uniformly in forecasting prices for the sample period we consider. this ﬁnding is not puzzling as the AL estimator behaves as an estimator conditional on the average volatility level and performs better when the volatility stays around the mean level. respectively. we focus on a 2-month (τ = 42 days) horizon. whereas the observed market prices are taken to be the realized option prices. 2007 to May 31. 2009. However. the AL densities produce a better ﬁt in-sample and over short horizons (for γ up to 14 days). we compare the empirical characteristics of the pricing kernels over two diﬀerent periods: the precrisis11 period from June 1. 16 . but also when it is proxied by the AL density for γ longer than 21 days. Therefore. The realized density on day t + γ is proxied by either the AL or the SX estimator. 7. the AL estimates are uniformly outperformed by our estimators at every horizon except at τ = 14 days. 2006 to November 30. the importance of including volatility in the prediction from the fact that the SX outperforms the MKT in the long run. Moreover. For market prices. 2007.

our nonparametric analysis reveals that the documented U-shape of pricing kernels could be due to the missing stochastic volatility variable. Zt =35. decrease as market index return increases. Wu (2011). without which the parameters in the volatility process cannot be identiﬁed. and estimate the model using both VIX time series and VIX option prices. whereas the low and high VIX levels post-crisis are taken as 20. We follow the common practice in the literature to select a subsample of S&P 500 and VIX options for parametric ﬁtting.and post-crisis. Therefore. θ) 2 . Note that the low level of VIX post-crisis is the same as the high level pre-crisis. in the post-crisis period exhibits a pronounced U-shape. 3.455. Zt = 20. our empirical result suggests that the shape of the pricing kernel is inﬂuenced by the stochastic volatility. as the VIX stays in a high regime after crisis. are chosen as 11. θ) 2 nH + i=1 log H(τ i . xi . zti . the pricing kernel conditional on a high VIX level.56 and 20. although separately estimated using non-overlapping data before and after the crisis.12 Overall.03 (the top right and low left panels). sti .g. Chabi-Yo et al. (2010a). This estimation leads to a secondary contribution of this paper—a new parametric estimation strategy. 17 .03. and Christoﬀersen et al. we derive the exact relationship between the instantaneous volatility and the VIX.. we estimate the model parameter θ by minimizing the mean squared percentage pricing errors: nC θ = min θ∈Θ log i=1 C(τ i . as implied by a decreasing risk aversion. Chabi-Yo (2011). xi ) H(τ i . or high frequency volatility measures. This decrease is consistent with standard economic theory that the pricing kernel decreases when wealth rises.03 and 35.pricing kernel estimates. zti . (2008). zti . zti . the two pricing kernels conditional on the same VIX level. xi .5 Test Results Here we report the nonparametric test results of parametric models (7) using S&P 500 and VIX options. along with the post-crisis pricing kernel conditional on a low level. xi ) C(τ i . Instead. In particular. In contrast. consistent with Jackwerth and Brown (2001). for both pre-crisis (in the upper panel) and post-crisis (in the lower panel) at diﬀerent VIX levels. e. Zt =20. For each parametric model.03. We observe that the pre-crisis pricing kernels. exhibit a similar decreasing pattern. It is worth pointing out that our estimation strategy avoids replacing the unobserved volatility process by some VIX proxy as in A¨ ıt-Sahalia and Kimmel (2007). respectively.455. in order to accelerate the speed of calculation and ensure the quality 12 The diﬀerence in the supports of these two pricing kernels over the wealth axis is caused by diﬀerent levels of S&P 500 index pre. sti . The low and high levels of Zt before crisis.

κ.13 13 In a related study. diﬀerent maturities. The result shows that for SVCJ1 and SVCJ2. 18 . and β− > β+ . and for each maturity. This asymmetric pattern suggests an asymmetric mean-reverting rate of volatility—the speed is high at low levels of volatility and low otherwise—because the aﬃne models under consideration have a constant mean-reverting rate. Table 6 reports test statistics of separate hypotheses.of the ﬁtting. Parametric estimation results are shown in Table 5. yet. (2011) nonparametrically estimate the drift function of the volatility process and document a similar nonlinear pattern. the rejection for S&P 500 options is more serious compared with VIX options. We ﬁnd that all of the models are rejected with p-values close to zero. indicating that non-aﬃne models may be better choices. This result is not surprising though. Engulatov et al. the insigniﬁcant λ0 estimates for both SVCJ1 and SVCJ2 suggest that a state-dependent jump intensity without constant terms seems adequate. whereas they tend to underestimate the dispersion when volatility is low. indeed reduces the ﬁtting errors from 60% to 40% and signiﬁcantly decreases the volatility of volatility (σ). For each day in 2009. Our SPD estimates can provide more-detailed information about the mis-speciﬁcation of the model and pave the way for potential improvement. parametric estimates of SPDs are closer to nonparametric ones when the current VIX level is high.722 for SVCJ2. we randomly select three options with diﬀerent moneyness. We report in Figures 8 and 9 the nonparametric and parametric SPDs of St and Zt at both high and low levels of the VIX across maturities. we ﬁrst choose three shortest. the test statistics discussed in Section 2. The diﬀerence in ﬁtting results indicates that adding jumps into volatility dynamics. Therefore. The resulting subsample includes a wide cross section of options. with the probability of downward jump equal to 1 − q = 0. Using both parametric and nonparametric estimates. which have smaller values of test statistics. Moreover. To gauge the performance in ﬁtting each type of option prices.119 for SVCJ1. the estimates for the average jump sizes in S&P 500 returns are negative. as adding VIX options in ﬁtting amounts to increasing the weight of volatility dynamics. the parametric models under consideration are not ﬂexible enough to capture the volatility dynamics implied by S&P 500 and VIX options simultaneously.5 are computed and reported in Table 6. It is obvious from the plot that for both index returns and the VIX. implying that they cannot capture S&P 500 and VIX option prices simultaneously. as shown by µs = −0. regardless of jump speciﬁcations. Finally. Recall that the SVCJ1 and SVCJ2 models are augmented with volatility jumps compared with the SVJ1 and SVJ2 models.

and 357 days. 294. we run the following regression IVtτ = ατ + β τ IIVtτ + ετ t of the implied volatility IVtτ on the interpolated implied volatility IIVtτ for the maturity τ = 168. which may equal the long-run mean? Answering these questions will shed light on how investors employ historical time series of the VIX up to time t to predict the volatility of the future S&P 500 index ST . 231. Across the long maturities. the longest maturity for VIX option contracts in 2009 is 132 days whereas S&P 500 options with maturities from 133 days to a year are still very liquid. The remaining sample size is as large as 19. 231. Figure 10 provides nonparametric SPD estimates with maturities equal to 168. even when VIX options are not available. 294. the former density is more dispersed than the latter. corresponding to the documented long-term dependence in Figure 10.e. The signiﬁcance indicates that investors may predict the future volatility through certain mechanism and price S&P 500 options accordingly.3. One possible mechanism for such long-run dependence is the simple martingale procedure. Speciﬁcally. interpolate the implied volatility of long-term options using today’s corresponding characteristics. which is shown to work well even in the short run when VIX options are available.876 data points. we direct our nonparametric estimation speciﬁcally to the long-term SPDs of S&P 500 index with maturities longer than 132 days. 19 . and then treat them as the implied volatility of today’s long-term options. To answer these questions. investors can ﬁrst compute implied volatility of yesterday’s S&P 500 option prices.. This result suggests that the martingale procedure agrees with the market practice and may oﬀer a potential explanation for the long-run dependence of the future S&P 500 index on current VIX levels. and 357 days for both high and low levels of the VIX. To study whether this mechanism works in practice. As a result. we ﬁnd that the SPDs of the index conditional on high VIX levels diﬀer signiﬁcantly from the SPDs conditional on low current VIX levels. In this case. This martingale mechanism may result in the observed long-run dependence as short-term options may carry information from Zt . do investors predict the long-term future volatility using current VIX Zt and price the S&P 500 options accordingly? Or do they just price the long-maturity S&P 500 options by taking the future volatility level as a constant. investors who would like to buy S&P 500 options with maturities longer than 132 days have no information available from VIX options about the volatility of the future index. i.6 When VIX Options are Not Available As shown in Table 1. The results in Table 7 show that the β τ coeﬃcients are signiﬁcant and close to one across all long maturities with adjusted R2 larger than 72%.

we proposed SPD estimates of the index and volatility and document several important empirical facts relevant to the understanding of the interactions between the two option markets. There is no statistical evidence for rejection with p-values around 0. Overall. 20 . Our results also highlight the missing volatility variable that may be responsible for the pricing kernel puzzle. x) without the homogeneous of degree 1 assumption. Zt )} = 1 for VIX options. and C(τ. We construct test statistics MS&P and MV IX based on the diﬀerences between the two estimators for the same quantity. 4 Conclusion In this paper. st . zt ) are nonparametric estimates of the SPDs with and without the assumption. Under certain regularity conditions. and suggest that the inadequacy of these models lies in the inability to adjust mean-reverting speeds for diﬀerent volatility regimes. our testing results call the aﬃne models into question. the two dimension-reduction techniques employed in our nonparametric analysis. for the sample period. Zt . st . mt ) and p(zT |τ. In particular. including homogeneous of degree 1 in St for S&P 500 options.7 Robustness Checks Here we verify. zt ) on the right panel. In addition. zt . X) = C(τ. p(zT |τ. the dependence of SPDs of the index on the VIX suggests pervasive volatility information in the VIX and its option market. St ) = p(ZT |τ. Table 6 reports values of the test statistics and the corresponding p-values. mt ) is the estimate we have been using. our dimension-reduction assumptions seem valid for the sample period in the empirical study. For VIX options. St . we plot C(τ. Zt . zt ) on the left panel along with their percentage errors relative to C(τ.S&P : Pr {C(τ. Close scrutiny of the plots shows that the percentage errors are mostly close to zeros in spite of reaching 20% and 10% around the boundaries where the estimates incur larger biases. and the dependence of VIX option prices on St only through Zt . We also propose rigorous statistical testing to check whether these assumptions hold. The null hypotheses are given below: H1. zt . both statistics converge to standard normal distributions under the null hypotheses. zt . zt . Mt )} = 1 for S&P 500 options where Mt = X/eSt and H1. Suppose C(τ. zt ) and p(zT |τ.VIX : Pr {p(ZT |τ. st . In Figure 11. see Appendix C for details.3. x) is the nonparametric estimate of the S&P 500 call option price C(τ. Zt . st .5. st . x) and p(zT |τ. zt .

21 . however. eﬀorts are needed to choose suitable bandwidths for our approach. our nonparametric pricing method carries over the strategies introduced in A¨ ıt-Sahalia and Lo (1998) to the case with stochastic volatility. by exploring additional information from VIX and its options. our method enjoys several beneﬁcial features. In contrast. Our closed-form regression formulae. By nature. robust to misspeciﬁcation of models and pricing measures. alleviate the dilemma to a great extent. parametric stochastic volatility models face an unfortunate compromise between misspeciﬁcation and computation burden. For example. such as inverse Fourier transforms and partial diﬀerential equations. Moreover. the computation often involves implementation of numerical algorithms. even with aﬃne models. such as being modelfree.In fact. the convenient aﬃne model with closed-form option pricing formulae are usually subject to misspeciﬁcation errors. and computationally eﬃcient. Of course.

Zti . z.1): nC M = h2 (hτ hz hm )1/2 m + where hy hτ hz hτ 2 i=1 nH 1/2 i=1 p(mTi |τ i . θ) 2 ×aC (τi . Then we can estimate M (p (·) . Zt ) − p(ZT |τ . p (·)) by its sample analog M (p (·) . Zti . p (·)) =E (p(mT |τ . z. Zt . appropriate choices of weighting functions can reduce the inﬂuences of unreliable estimates and make the tests focus on a particular empirical question of interest. Zt . Zti . ZTi ) − C2 / √ 2D1 2 √ 2D2 2 C1 = (hτ hz hm )−1/2 × e2rτ C2 = hτ hz hy k 2 (c) dc ˙ ck (c) + k(c) 2 2 dc / k (c) c2 dc s2 (τ. The idea behind our test is simple: If the null H0 holds. then SPDs of both S&P 500 index and VIX should be close to their nonparametric estimates. θ) i=1 nH 2 aC (τ i . ZT ) where aC (·) : R3 → R+ and aH (·) : R3 → R+ are weighting functions. Zt . mTi ) − C1 / aH (τi . p (·)) = 1 nC nC p(mTi |τ i . Zti . Our test statistic for H0 against HA is an appropriately standardized version of (A. Zti ) − p(mTi |τ i . As noted by A¨ ıt-Sahalia et al. Denote the bandwidth for the state variable mT as hm . θ 0 ))2 aC (τ . θ) i=1 aH (τ i . Zti ) − p(ZTi |τ i .1) The use of weighting functions aC (·) and aH (·) are not uncommon in the literature and often used to remove extreme observations. Zti .Appendix A Speciﬁcation Test We measure the diﬀerences between the parametric and nonparametric SPDs by the sum of squared deviations: M (p (·) . Suppose we have a consistent estimator θ for the parameter θ. mT ) + (p(ZT |τ . mTi ) 2 1 + nH p(ZTi |τ i . m) aC (τ. (2001). m) dτ dzdm C −1/2 2 k 2 (c) dc ˙ ck (c) + k(c) 22 2 2 dc / k (c) c2 dc . Zt . Zti . θ 0 ))2 aH (τ . Zt ) − p(mT |τ . Zti . θ) p(ZTi |τ i . Zti ) − p(ZTi |τ i . Zti . Zti ) − p(mTi |τ i . ZTi ) (A.

VIX . t≤s≤t+τ Q (JS )2 = aVt + b 23 . m) dτ dzdm C 2 ˙ c1 k (c1 ) + k(c1 ) ˙ (c1 + c2 )k (c1 + c2 ) + k(c1 + c2 ) dc1 2 2 4 dc2 × × e4rτ k (c1 ) k (c1 + c2 ) dc1 dc2 / k (c) c2 dc s4 (τ. with nC /nH converging to y a non-zero constant. z. d B Parametric SPDs and Option Pricing We derive the theoretical value of the VIX Zt for Models SVCJ1 and SVCJ2. z. z. we have M −→ N (0.e. we have 1 Q 2 Zt = 104 · Et τ where a= b= 1 − e−(κ−βV λ1 )τ (1 + λ1 χ) · 104 (κ − βV λ1 )τ κξ + βV λ0 1 − e−(κ−βV λ1 )τ 1− (1 + λ1 χ) + λ0 χ · 104 κ − βV λ1 (κ − βV λ1 )τ t+τ Vs ds + t Ns ≥0. y) is deﬁned similarly. m) aC (τ. nH → ∞. z. since the other models are nested. z. y) aH (τ. y) aH (τ.. In fact. The bandwidths are chosen to satisfy h2 hτ hz ln nC → 0. and h2 hτ hz ln nH → 0 as nC . m ln nH nh4 hτ hz y ln nC nC h4 hτ hz m → 0.S&P and H0. m) is the square-root of the conditional variance for the local linear regression of Ci on Ui and sH (τ. z. y) dτ dzdy H 2 ˙ c1 k (c1 ) + k(c1 ) ˙ (c1 + c2 )k (c1 + c2 ) + k(c1 + c2 ) dc1 2 2 4 dc2 × × e4rτ D2 =2 k (c1 ) k (c1 + c2 ) dc1 dc2 / k (c) c2 dc s4 (τ. → 0.× e2rτ D1 =2 s2 (τ. z. i. 1) under H0 as nC . Under certain regularity conditions. whether the model is consistent with S&P 500 option prices and VIX option prices separately. z. nH → ∞ When the null hypothesis H0 is rejected. y) dτ dzdy H where sC (τ. we follow similar procedures as above to construct test statistics MS&P and MVIX to test separate speciﬁcation hypotheses H0.

u.θ)v where A and B satisfy the system of ordinary diﬀerential equations below: 1 1 1 ˙ 0 = −B + σ 2 B 2 + ((us + iws )ρσ − κ)B − ( + µλ1 )(us + iws ) + (us + iws )2 + λ1 (l(B) − 1) 2 2 2 24 with the initial condition: . vt . θ) = e−iws sT Ψ(τ. that is. st . θ) dwv As the model is aﬃne. vt . vt . Ψ(τ. s + η. we can obtain a closed-form formula for Ψ: Ψ(τ. zt . θ)dw R2 Moreover. st . vt . θ) = e(us +iws )s+(uv +iwv )v Since the model is aﬃne in v. θ) dws e−iwv vT Ψ(τ. τ |st . where τ = T − t. w. θ) = e(us +iws )s+A(τ . (ws . w. st . θ)dws R ∞ 0 Re e−iws sT Ψ(τ. it is suﬃcient to calculate the marginal density of V . 0. In fact. 0. v. (ws . wv ). vt . w. st . s. s. (0. θ) = 1 4π 2 e−iws sT −iwv vT Ψ(τ. st . u. for Model SVCJ2. Due to √ the fact that Zt = aVt + b. VT ). s. u. st . uv ) ∈ R2 and w = (ws . θ) = 2π 1 = π p(sT . µ2 + σ 2 S S χ= 2(qβ 2 + (1 − q)β 2 ) + − for Model SVCJ1. v) νS (η)νV (ζ)dηdζ ∂s∂v R2 Ψ(0. wv ) ∈ R2 .θ) satisﬁes the following PDE: 0=− 1 ∂Ψ ∂Ψ 1 ∂ 2 Ψ 1 2 ∂ 2 Ψ ∂Ψ + (r − d − v − µ(λ0 + λ1 v)) + κ(ξ − v) + v 2 + σ v 2 ∂τ 2 ∂s ∂v 2 ∂s 2 ∂v ∂2Ψ + ρσv + (λ0 + λ1 v) Ψ(τ. we derive the conditional densities p(sT |τ. vt .θ) is given by p(sT . s. 0. s. w. θ) a z 2 −b z 2 −b T . u = (us . u. st . vt . Now let Ψ(τ. v. wv ). (0. st . uv ). u. the marginal densities p(sT |τ. zt . 0. st . w. wv ). 0).θ).θ) and p(zT |τ. we have by change of variable that p(zT |τ. st . Ψ(τ. τ |st .θ)+B(τ . v. w.θ) and p(vT |τ. v. θ) = Et (e(us +iws )ST +(uv +iwv )VT ). (us . v) be the time-t generalized conditional characteristic function of (ST . vt . vt . v + ζ) − Ψ(τ. v.θ) are: 1 2π 1 = π 1 p(vT . (ws . vt . vT |st . θ)dwv R ∞ 0 Re e−iwv vT Ψ(τ. u.θ) has a closed form formula. st . 0. The conditional density p(sT . 0). s. Q Ψ(τ.vt = ta a vT = Therefore. u. s. w. vt . w. vT |st . In the following. θ) = 2zT p(vT |τ. zt . vt .

vt . (ws . (ws .θ) = 0. vt . st . the formula becomes more complicated. st . wv ). 0). This is straightforward for S&P 500 options. Using the fact that √ zt = avt + b. θ) = + v Ψ(u. θ) ∂B(τ . w. i. wv ). vt . for Model SVCJ2. Regard- ing VIX options.˙ 0 = −A + κξB + (r − d − µλ0 )(us + iws ) + λ0 (l(B) − 1) where A(0.. θ) =e−rτ Et (eST − X)+ = e−rτ π ∞ Re 0 Ψ((us . 25 . X. the VIX. uv ). θ)Xe−(us +iws ) log X dus (us + iws )(us + iws − 1) for us > 1. θ) =e−rτ Et ( aVT + b − Y )+ =e−rτ a π ∞ Re 0 ∞ Ψ((0. uv ). B(0. vt . 0). and l(a) = R2 e(us +iws )η+ζa νS (η)νV (ζ)dηdζ for Model SVCJ1. θ) ∂θ ∂θ ∂θ and ∂A(τ . we need the derivatives of option prices with respect to θ. θ)e a (uv +iwv ) erfc Y ( uv +iwv ) 2 a 2(uv + iwv ) 2 2 3 b 1 dwv 2 where uv > 0. we have ∂H =e−rτ ∂θ + e−rτ 1 π ∞ √ Re ∂Ψ((0. st . st . vt . wv ). st . θ) ∂ 2(uv + iwv ) 2 3 ae a (uv +iwv ) erfc Y ( uv +iwv ) 2 a ∂θ dwv and then follow similar calculations as for S&P 500 options. τ. w. θ) ∂θ Re 0 ae a (uv +iwv ) erfc Y ( uv +iwv ) 2 a 2(uv + iwv ) 2 √ b 3 b 1 dwv 1 0 1 π ∞ Ψ((0. Finally. (0. vt . (0. 0). τ. (0. vt . st .θ) ∂θ . the VIX option price with maturity T and strike price Y is give by: Q H(τ . τ. which is observable from the market. st . Therefore.e. 0).θ) = uv + iwv . we can rewrite functions C and H in terms of zt . τ. Y . uv ). θ) ∂A(τ . In order to calculate standard errors for parametric estimation. since a and b also depend on θ. since ∂C e−rτ = ∂θ π where ∂Ψ(u.θ) ∂B(τ . ∂θ ∞ Re 0 ∂Ψ((us . τ. θ) Xe−(us +iws ) log X dus ∂θ (us + iws )(us + iws − 1) can be calculated by augmenting the ODE systems with more unknowns. 2 (1 − βV a)−1 eµs (us +iωs )+ 1 σs (us +iωs )2 2 = (1 − βV a)−1 q(1 − (us + iws )β+ )−1 + (1 − q)(1 + (us + iws )β− )−1 The price of the S&P 500 Call option with strike X and maturity τ is given by: Q C(τ. τ. vt . and erfc(x) = √π x e−t dt is the complementary error function. vt . τ. st .

s. st . For the VIX density. s. z. m) are estimators of conditional variances of regressing option prices C C Ci on (τi . s. Zti . we have MVIX = hy hτ hz hs hy where CVIX = hτ hz hs hy × e2rτ DVIX =2 −1/2 3 2 1/2 nH [p(ZTi |τi . z. z. 2 ˙ (c1 + c2 )k (c1 + c2 ) + k(c1 + c2 ) dc1 2 3 4 dc2 × × e4rτ k (c1 ) k (c1 + c2 ) dc1 dc2 / 2 c2 k (c) dc s2 τ. z. zt ) and p(zT |τ. Zt . C and s2 (τ. s. we have MS&P nC = (hτ hz hs hx ) where 1/2 i=1 C(τ. Mt ) 2 aC (τi . St . z. x) and s2 (τ. x) dτ dzdsdx. m) C C dc2 2 a2 (τ. Zti )]2 aH (τi . z + s2 τ. z. C C 2 4 DS&P =2 k (c1 ) k (c1 + c2 ) dc1 s2 (τ. z H H ˙ c1 k (c1 ) + k(c1 ) aH τ. Sti . Sti . z. For the S&P 500 options. X) − C(τ. and between p(zT |τ. H 26 . z. Mt ). Sti . s. Zt . z. Xi ) − CS&P / DS&P CS&P = (hτ hz hs hx )−1/2 k 2 (c) dc s2 (τ. x) + s2 (τ. ZTi ) − CVIX i=1 / DVIX k 2 (c) dc ˙ ck (c) + k(c) 2 2 dc / c2 k (c) dc s2 τ. Zti . z. X) and C(τ. Sti . z dτ dzdsdz . Zt . z H H a2 τ. St . x) + s2 (τ. s. z + s2 τ. Zti ) − p(ZTi |τi . s. x) dτ dzdsdx. z. z dτ dzdsdz . z. s. Xi ) without and with assuming the homogenous of degree 1. Zt . m) aC (τ. Zti . s. z.C Tests for Robustness Checks We measure the diﬀerences between C(τ. The test statistics are appropriately standardized versions. z. zt ) by the sum of squared deviations.

z ) and s2 (τ. MVIX −→ N (0. z ) are estimators of conditional variances of regressing option prices H H Hi on (τi .VIX as nH → ∞. By similar regularity conditions and bandwidth choices to those for the speciﬁcation test M . z. Zti . s. 1) under H1. Sti . Yi ) without and with assuming that H does not depend on St . d d 27 . z.and s2 (τ. we have MS&P −→ N (0. 1) under H1.S&P as nC → ∞.

D. “Nonparametric Estimation of State-Price-Densities Implicit in Financial Asset Prices.” Journal of Econometrics. 2003–2049. S. A. 102. “Returns of claims on the upside and the viability of U-shaped pricing kernels. “Forecasting S&P 100 volatility: the incremental information content of implied volatilities and high-frequency index returns.. Britten-Jones. Y. B. M. Y. 105. “Delving into Risk Premia: Reconciling Evidence from the S&P 500 and VIX Derivatives. Boes. Bakshi. 181–238. 52. Drost. (2010).. Madan. D. Broadie. R. 5–26. and Yared. rep. Z. J. and Neuberger.” Journal of Business. Bates. G. and Johannes. “Maximum Likelihood Estimation of Stochastic Volatility Models. Poon. 67–110.. 621–651.” Journal of Finance.. and Werker. 28 . Cao. J. G. 53. University of Chicago Booth School of Business. 55. 51. M. D. 94. Breeden. and Litzenberger.. Y. S. 839–866. (2012). Amengual. and Xiu.. “Option Prices. C. “Post-’87 Crash Fears in the S&P 500 Futures Option Market. A¨ ıt-Sahalia. 499–547. and Stochastic Volatility. R. G. Tilburg University. (2007). 97.” Journal of Financial Economics. (2000). (2000). A¨ ıt-Sahalia. “Empirical Performance of Alternative Option Pricing Models. B. A. “Nonparametric risk-neutral return and volatility distributions. — (2000). (2007).” Tech. (2001).” Journal of Econometrics..” Journal of Financial Economics. “Do Option Markets Correctly Price the Probabilities of Movement of the Underlying Asset?” Journal of Econometrics.” Tech. (2007)..” Journal of Finance. 83. D. (1978). and Kimmel. Chernov. and Lo. Blair. (1997).-H... and Panayotov. 94. J.... and Taylor. Wang. Bakshi. “Nonparametric Risk Management and Implied Risk Aversion. S. Y. S. M.” Journal of Finance. “Prices of State-Contingent Claims Implicit in Option Prices.. F. 9–51. and Chen. 413–452. rep.” Journal of Econometrics. F. (1998).References A¨ ıt-Sahalia. “Model Speciﬁcation and Risk Premia: Evidence from Futures Options. 130–154. M. 62. D. (2001). H. Implied Price Processes.” Journal of Finance. M.

Jacobs.. “Pricing Kernels with Stochastic Skewness and Volatility Risk.. rep. P. R. and Renault. London. K. (2008). S. J. Gonzalev.. and Mancini. daily returns.” Journal of Economic Dynamics and Control. 21–50. C.: Chapman & Hall.K. (1996). I. O. “State dependence can explain the risk aversion puzzle.. B. and Jackwerth. 985–1001. and Longstaﬀ. E.-Y. (2011).. 20. Christoﬀersen.” Journal of American Statistical Association. Grunbichler. Hedging and Spanning in Option Markets. “Option pricing with aggregation of physical models and nonparametric statistical learning. Local Polynomial Modelling and Its Applications. (1996). University of Geneva. F. “Do Stock Prices and Volatility Jump? Reconciling Evidence from Spot and Option Prices. “The Price of a Smile/. A. Jackwerth. 3141– 3189.” Tech. 14.. (2010a). J.” Review of Financial Studies. and option prices. J..” Review of Financial Studies. and Gijbels. (2011). 23. 22. Chabi-Yo. (2000). J. and Mimouni. A. Chabi-Yo.” Review of Financial Studies. F. 973–1011.” Journal of Finance. P.” Review of Financial Studies. rep. “Models for S&P 500 dynamics: evidence from realized volatility. Fan. Heston. and Yeh. 495–527.” European Finance Review. (2001). L. and Wu. “Jump and Volatility Risk Premiums Implied by VIX. 29 . 433–451.” Review of Financial Studies. R. F.. and Scaillet. P. 1311–1341. McGill University. 104. 34. 4. “Variance Risk Premiums. “The valuation of volatility options. and Osakwe. J. (2009). (2009). 1351–1372. Eraker. and Jacobs.” Management Science. Fan. Detemple. (2004). J. 2232–2244. 59. 21.” Tech. Christoﬀersen. (2010b). (2000). 13. L.Buraschi. “Recovering Nonlinear Dynamics from Option Prices. Garcia. Carr. Engulatov. C. K. Duan. A. “Valuing futures and options on volatility. “Option anomalies and the pricing kernel. (2011).. U. K..” Journal of Banking and Finance. “Recovering risk aversion from option prices and realized returns.

J.” Tech. 1. S. and Rubinstein. P. Rep. (2011). 280–287. “The pricing kernel puzzle: Reconciling index option data and economic theory. L. Wu. “Valuation of VIX Derivatives. 160. and Sentana. “Derivatives on Market Volatility: Hedging Tools Long Overdue. J. and Brown. 1305–1342. 1611–1631. 3–50. (1993). M. Pan. CEMFI. rep. J.Jackwerth. “The Model-Free Implied Volatility and Its Information Content. 84–89.” Journal of Financial Economics. (1996).” Risk. A. April. (2008). Whaley. (2009). F.. “Recovering Probability Distributions from Contemporaneous Security Prices. Sepp. J. 22.” Review of Financial Studies. 0913. J. 71–84. (2009). 18. 30 . 4335–4376. Jiang. and Tian. G.” Tech. 51. “VIX option pricing in jump-diﬀusion model. (2005). E.” The Journal of Derivatives. (2001). R. H. University of Konstanz. Y.” Review of Financial Studies. “Variance Dynamics: Joint Evidence from Options and High-Frequency Returns. “The Jump-Risk Premia Implicit in Options: Evidence from an Integrated TimeSeries Study. 63. E.” Journal of Econometrics. “Nonparametric Estimation of State-Price-Densities Implicit in Interest Rate Cap Prices. Mencia. and Zhao.” Journal of Finance. Li. (2002). Jackwerth.

0625 0.1954 -0.0777 -0.0342 4.0171 9.275 1057.0464 0.0007 31.0097 39.6825 2.3 0.1979 939.425 Note: This table reports the summary statistics of the time series of S&P 500 index.86 -0.145 0. 2009. return.0018 28.Table 1: Summary Statistics of the S&P 500 Index and VIX Mean Std Skew Kurt Median 25% 75% S&P 500 Index S&P 500 Return VIX 948. and VIX from January 2 to December 31.9126 2.57 855.4786 115.0081 24.5682 0. 31 .

75 197. time-to-maturity.950 11.312 0.00 22.175 0. In total.00 42.00 2 17 34 54 131 Note: This table reports the summary statistics (minimum.025 1.20 184.0 0.275 3.5 0.00 20.688 73. implied volatility.150 5.450 7.00 100.00 35.317 0.275 7.95 50 650 775 875 1125 3 20 38 78 783 0.00 6.216 0.00 55.50 55.153 0.262 0.750 16.10 10.60 680 940 1025 1125 2500 3 21 39 80 783 0.50 35.00 100.2 0.10 680 895 970 1090 2500 3 19 37 86 780 0.40 83.75 47. 32 .8 6.267 0.675 1.243 5.350 45.105 0. and maximum) for all traded S&P 500 and VIX options in 2009.4 0.350 1.50 27.70 22.00 2 22 41 67 130 ITM 4. 23.25 6.9 5. strike price.90 151.15 111.530 16.05 300 800 875 965 1125 3 17 33 69 780 0.658 135.025 0.00 25.933 VIX Call OTM 11.55 50.518 2. quantiles.319 0.377 2. there are 178.600 11.996 S&P 500 Put OTM 82.571 19.909 7.083 1. and trading volume.4 0.50 55.00 2 20 40 72 130 VIX Put OTM 3.599 29.50 35.336 1.153 0.00 27.325 5.136 0.40 71.80 10.60 22.5 0.806 records for VIX options with trading volumes.103 ITM 20.10 622.00 32. including the option price.450 20.025 0.000 2.00 20.55 1583.237 0.00 45.123 72.00 2 26 49 80 132 ITM 4.025 1.552 ITM 21.910 trading records for S&P 500 options with valid implied volatility and trading volumes.406 0.Table 2: Summary Statistics of S&P 500 and VIX Options S&P 500 Call Moneyness # of Contracts Trading Volume ×106 Min 25% Option Price 50% 75% Max Min 20% Strike Price 50% 75% Max Min 25% Time-to-Maturity 50% 75% Max Min 25% Implied Volatility 50% 75% Max OTM 54.395 2.

0720) (0.0520) 11.0823) 0.8173 (3.215) -0.7079) (27.2847 -0.1396) (1.6299) 0.840) 0.4237 (1.0180 (3.0487) (0.3522 (2.251 (1.1286 0.981) (30. and 84 days.251) -0.1676) -4.1735 (0.0528) (0.0969 (0.8486 (3. 33 .1082 0.289 74.036) (23.8518) (21. conditional skewness.0729) (0.0495 -0.1226 Panel B: Regression Results for τ = 42 39.110) 0.1899 (0.0202 0.1270 0.409 86.402 9.6036 (1.0036 0.3859 0.0729) 8.0853) (2. 42.Table 3: Dependence of S&P 500 Index SPD on the VIX SPD Intercept Zt Stdt [Zt+τ ] Skewt [Zt+τ ] Adj.2299 (2.6671) -3.097 80.527 (1.893 -74.347 -64.7851 8.0212 Panel C: Regression Results for τ = 63 -29.7106) (27.7711) 0.2416 0.0120) -5. 63.R2 Panel A: Regression Results for τ = 21 48.546 51.0677 0.7469) -0. with time-to-maturity equal to 21.6869) (3.216 -69.670 (1.2379 Note: This table reports regression results of the future average volatility VS of the S&P 500 index on the square t.0520) (0.3831 (0.050) (21.299) (27.0066 0.0166) (0.1122 0. and conditional kurtosis of the VIX densities.2385 -0.3725 -3.1035) 0.0008 0.5586) (13.0633 -0. The numbers in parentheses are standard errors.9069 (2.1421 Panel D: Regression Results for τ = 84 -24.282 -85.3942) (4.1873) -4.434 (2.0487) (0.0179) -4.1146) (0.1063 0.0108 0.T root of the conditional second moment.

8695 2.97 4.8326 AL 1.0787 SX 1.9918 Forecast of SX-Density SX 0.41 1.3313 2.00 4.8338 1.0107 0.28 2.9444 2.11 Note: Panel A reports average forecast errors of the densities produced by the A¨ ıt-Sahalia and Lo (1998) estimator (AL) and our SPD estimator (SX) conditional on the VIX. while Panel B reports those of prices by the AL.1627 Panel B: Market-Price Forecast Error τ 0 7 14 21 28 35 42 AL 9.79 6.76 6.36 8.9451 1.90 5. 34 .0000 0.5422 0.2332 1.54 7. 2009 and out-of-sample forecasts are generated for various forecast horizons t on a daily rolling basis from October 1 to December 31.6719 1.7362 1. The nonparametric option prices and their corresponding SPDs are estimated with daily data from January 2 to September 30. SX.3322 0.7206 0.95 2.4428 0.65 5.5582 1.5370 1.63 4.4679 1. and a martinglae interpolation (MKT) method.27 MKT 0.Table 4: In-Sample and Out-of-Sample Forecasts Panel A: Density Forecast Error Forecast of AL-Density γ 0 7 14 21 28 35 42 AL 0.2091 0.68 2.79 4.5789 2.09 SX 8.1815 1.6040 0.62 3.46 5. 2009.0000 0.68 5.74 1.7012 1.7399 0.

9234 (0.001) (0. P values are given in parentheses.011) SVJ2 3.094) (0.145 1.100) (0.086) (0.585 0.005) SVJ1 2.0747 (0.004) 0.652 (0.047 0.6080 (0.0000) Note: This table reports the nonparametric speciﬁcation tests for a variety of models.019) (0.021 0.708 SV (0.0000) VIX 21.183 0.276 (0.0000) 6. Standard errors are given in parentheses.410 -0.0000) Joint 33.0000) 15.028) 0.067 -0.583 0.5384 (0.8598 (0.991 (0.0000) SVJ2 11.159) (0.003) (0.162) (0.182 (1.033 -0.0000) SVJ1 14.002) (0.002) (0.032 0.6055 (0.119 0.109 0.370 -0.1774 (0.022 6.154 0.004) (0.2278 (0.002) (0.013) (0.7787 (0.768 (0.011) (0.0000) 11.028 0.8441 (0.526) 0.0000) 19.075 (0.005) (0.0000) 8.619 0.006) (2.128 (0.138 0.039) (2.009) (0.010) (0.362 0.021) (0.006) (0.023 -0.009) (0.921 -0.0000) 11.032 0.8035 (0.001 26.348 0.006) (0.376 Note: This table reports the parametric estimates for a variety of models.0357 (0.0000) 12.0000) SVCJ2 9.953 (0.102) (0.371) 3.672 0. 35 .Table 5: Estimates of Parametric Models Param κ ξ σ ρ β+ β− q βV µs σs λ0 λ1 RMSE 0.864) 5.323 0.9282 (0.786 0.000 33.602 0.039) SVCJ2 3.0000) SVCJ1 13.011) (0.316 0.096) SVCJ1 4.5492 (0.002 0.3413 (0. Table 6: Nonparametric Speciﬁcation Tests of Parametric Models Tests S&P SV 25.008) (0.002) (0.147 1.0000) 16.015) (0.611 0.048) (2.

0090) 0. Standard errors are given in parentheses.0108) 0.0112 (0.0028) 1.0439 0.0173) 0.7290 Note: This table reports regression results of implied volatility of observed option prices on the interpolated implied volatilities for long maturities.0035) 231 -0. 36 .0154 0.0046) 1. Table 8: Robustness Tests S&P 500 Test Statistics p-value 0.R2 1.0476 (0.7705 294 -0.0082 (0.Table 7: Regression Analysis for Long-Term Dependence Time-to-Maturity Intercept 168 -0.1403 (0.0459 (0.0142) 0.5136 VIX 0.0056) 1.0366 (0.1165 (0.8317 IIV Adj.5025 Note: This table reports the nonparametric tests for the two dimension reduction assumptions: homogeneous of degree 1 for S&P 500 options and dependence of VIX option prices on St only through Zt .7967 357 -0.0275 (0.

2009.Figure 1: Time Series of the S&P 500 Index and the VIX 1200 S&P 500 VIX 1100 50 60 1000 40 S&P 500 Level 900 30 800 20 700 10 600 Jan 09 Mar 09 May 09 Jun 09 Sep 09 Nov 09 Jan 10 Note: This ﬁgure plots the time series of S&P 500 index and VIX from January 2 to December 31. 37 VIX Level .

with empirical parameter estimates given in Section 3. ﬁxing the current volatility level. are plotted in the lower panel.5 −1 100 80 −1 100 80 1300 60 1100 40 900 1000 20 800 1200 1400 Time−to−Maturity Current Index Level Time−to−Maturity Index Level at Expiration Note: In the upper panel. 38 . averaged by 500 replications. The percentage estimation errors.5 0 100 80 1300 60 1100 40 900 1000 20 800 1200 1400 Time−to−Maturity Current Index Level Time−to−Maturity Index Level at Expiration 1 1 0.5 −0. we plot the nonparametric estimates of S&P 500 option prices and the SPD of the index.Figure 2: Nonparametric Option Prices and State-Price Densities of S&P x 10 350 300 3.5 SPD of the S&P 500 % Error 1400 1300 60 1100 40 900 1000 20 800 1200 S&P 500 Option % Error 0.5 2 1.5 0 0 −0.5 3 −3 S&P 500 Option Prices 250 200 150 100 50 0 100 80 1300 60 1100 40 900 1000 20 800 1200 1400 SPD of S&P 500 Index 2.5 1 0. The data are simulated from the model SVCJ2.

01 10 5 100 80 0 100 80 60 35 40 30 20 25 40 45 Time−to−Maturity Current VIX Level Time−to−Maturity VIX Level at Expiration 1 1 0 SPD of VIX % Error 45 60 35 40 30 20 25 40 VIX Option % Error 0.04 0. with empirical parameter estimates given in Section 4. we plot the nonparametric estimates of VIX option prices and the SPD of the VIX.05 0.5 0 −0. are plotted in the lower panel. averaged by 500 replications.08 0. The data are simulated from the model SVCJ2.03 0. The percentage estimation errors.07 VIX Option Prices 20 0.02 0. ﬁxing current volatility level at medium.5 0. 39 .Figure 3: Nonparametric Option Prices and the State-Price Density of the VIX 25 0.5 −0.06 15 SPD of VIX 45 60 35 40 30 20 25 40 0.5 −1 100 80 −1 100 80 60 35 40 30 20 25 40 45 Time−to−Maturity Current VIX Level Time−to−Maturity VIX Level at Expiration Note: In the upper panel.

42.5 2 2 1. and 84 days.5 0 650 0 650 700 750 800 850 900 950 1000 1050 1100 1150 S&P 500 index at Expiration S&P 500 index at Expiration Note: This ﬁgure plots nonparametric SPD estimates (solid lines) of the S&P 500 index at maturities of 21. Dotted lines around each SPD estimate are 95 percent conﬁdence intervals constructed from the asymptotic distribution theory in (6). Current Index Level Zt: 23.49 Time−to−Maturity τ : 84.5 0.455. Current Index Level Zt: 23. 40 .49 8 x 10 −3 Time−to−Maturity τ : 21. respectively.455 7 7 6 6 Density: Probability Percent 5 Density: Probability Percent 700 750 800 850 900 950 1000 1050 1100 1150 5 4 4 3 3 2 2 1 1 0 650 0 650 700 750 800 850 900 950 1000 1050 1100 1150 S&P 500 index at Expiration 6 x 10 −3 S&P 500 index at Expiration 6 x 10 −3 Time−to−Maturity τ : 42.5 1.490 and 42.5 3 3 Density: Probability Percent 2. Current Index Level Zt: 42.49 Time−to−Maturity τ : 42. and for two current VIX levels of 23.5 Density: Probability Percent 700 750 800 850 900 950 1000 1050 1100 1150 2. Current Index Level Zt: 23. Current Index Level Zt: 42.5 3.5 1 1 0.455 3. Current Index Level Zt: 42. The two levels of the current VIX index are obtained by the 20 and 80 percent quantiles of the VIX time series of 2009 and correspond to low and high levels of the VIX.Figure 4: Estimates of the State-Price Density of the S&P 500 8 x 10 −3 Time−to−Maturity τ : 21.455 5 5 Density: Probability Percent 4 Density: Probability Percent 700 750 800 850 900 950 1000 1050 1100 1150 4 3 3 2 2 1 1 0 650 0 650 700 750 800 850 900 950 1000 1050 1100 1150 S&P 500 index at Expiration 4 x 10 −3 S&P 500 index at Expiration 4 x 10 −3 Time−to−Maturity τ : 84.

41 .Z ) T t t 5 95% Upper 95% Lower Density: Probability Percent 4 3 2 1 0 650 700 750 800 850 900 950 1000 1050 1100 1150 S&P 500 index at Expiration 6 x 10 −3 Time−to−Maturity τ : 42. 50% or 80% quantile of the VIX time series in 2009. Current Index Level Zt: 23.455 p(ST|τ. Current Index Level Zt: 28. The current values of the VIX are ﬁxed at 20%.S . Current Index Level Zt: 42.Zt) 95% Upper 95% Lower 5 Density: Probability Percent 4 3 2 1 0 650 700 750 800 850 900 950 1000 1050 1100 1150 S&P 500 index at Expiration 6 x 10 −3 Time−to−Maturity τ : 42.St) 95% Upper 95% Lower p(ST|τ.905 p(ST|τ.49 p(ST|τ. Dotted and dash-dotted lines are 95 percent conﬁdence intervals constructed from the asymptotic distribution theory in (6).Figure 5: SPDs of S&P 500 Conditional on VIX vs Estimates that Ignore VIX 6 x 10 −3 Time−to−Maturity τ : 42.St.St) 95% Upper 95% Lower p(S |τ.S .St) 95% Upper 95% Lower p(S |τ.Z ) T t t 5 95% Upper 95% Lower Density: Probability Percent 4 3 2 1 0 650 700 750 800 850 900 950 1000 1050 1100 1150 S&P 500 index at Expiration Note: This ﬁgure overlays nonparametric SPD estimates (solid lines) of the S&P 500 index conditional on the current VIX level with the nonparametric estimates that ignore the VIX variable (dashed lines). The time-to-maturity is ﬁxed at 42 days.

Current Index Level Zt: 42. The two levels of the current VIX index are the 20th and 80th percent quantiles of the VIX time series of 2009.455 0.05 0.05 0. 42 . 42.06 0.03 0.49 0.02 0.03 0.05 0. Current Index Level Zt: 42.06 0.04 0. and 84 days. Current Index Level Zt: 23.01 0. Dotted lines around each SPD estimate are 95 percent conﬁdence intervals constructed from the asymptotic distribution theory in (6). Current Index Level Zt: 23.02 0.02 0.04 0.02 0.455 0.03 0.06 VIX at Expiration Time−to−Maturity τ: 84.05 Density: Probability Percent 0.04 Density: Probability Percent 20 25 30 35 40 45 50 55 60 0.03 0.490 and 42.455 0.04 0. and for two current VIX levels at 23.455.06 0.06 Time−to−Maturity τ: 21.03 0.05 Density: Probability Percent 0.01 0.49 0.03 0.04 Density: Probability Percent 20 25 30 35 40 45 50 55 60 0.06 VIX at Expiration Time−to−Maturity τ: 42.05 Density: Probability Percent 0. Current Index Level Zt: 23.01 0 15 0 15 20 25 30 35 40 45 50 55 60 VIX at Expiration Time−to−Maturity τ: 84.Figure 6: Estimates of the State-Price Density of the VIX Time−to−Maturity τ: 21.04 Density: Probability Percent 20 25 30 35 40 45 50 55 60 0.02 0.02 0.01 0 15 0 15 20 25 30 35 40 45 50 55 60 VIX at Expiration Time−to−Maturity τ: 42.01 0.49 0. Current Index Level Zt: 42.01 0 15 0 15 20 25 30 35 40 45 50 55 60 VIX at Expiration VIX at Expiration Note: This ﬁgure plots nonparametric SPD estimates (solid lines) of the VIX at maturities of 21.

9 0. the post-crisis estimates for the 18-month sample period from June 1.1 Wealth Wealth Note: In the upper panel. respectively.1 14 12 12 10 10 8 8 6 6 4 4 2 2 0 0.98 1 1. Dotted lines around each pricing kernel are 95 percent conﬁdence intervals.96 0.94 0.06 1.98 1 1.04 1.02 1.02 1.06 1.92 0. 11.56 20 MRS 95% Upper 95% Lower 20 MRS 95% Upper 95% Lower 06/01/2006−11/30/2007.04 1.02 1.455 18 18 16 16 Scaled Marginal Rate of Substitution 14 Scaled Marginal Rate of Substitution 0.96 0. Correspondingly.92 0. In addition. Zt=35.03 20 MRS 95% Upper 95% Lower 20 MRS 95% Upper 95% Lower Wealth 06/01/2009−11/30/2010.96 0.04 1. Zt=20.92 0. Zt=11.94 0. we plot nonparametric estimates of the pricing kernels for the 18-month sample period before the 2008 ﬁnancial crisis.1 14 12 12 10 10 8 8 6 6 4 4 2 2 0 0.Figure 7: Nonparametric Estimates of Pricing Kernels 06/01/2006−11/30/2007.56 and 20. represent low and high levels of VIX pre-crisis.02 1.08 1.98 1 1.08 1.9 0 0.94 0.06 1.04 1.03 and 35.03.92 0.08 1. 43 . 2010 are reported in the lower panel.94 0.03 18 18 16 16 Scaled Marginal Rate of Substitution 14 Scaled Marginal Rate of Substitution 0.455.9 0 0.08 1. whereas the low and high levels of VIX after the crisis are chosen to be 20. Zt=20.96 0.06 1. we choose time-to-maturity τ to be 42 days when reporting the estimates.9 0.98 1 1. 2009 to December 30.1 Wealth 06/01/2009−11/30/2010. The two levels of Zt .

3 Log−Return Log−Return Note: This ﬁgure overlays nonparametric SPD estimates (solid lines) of the S&P 500 index with the corresponding parametric SPDs (dashed lines) with time-to-maturity equal to 21.3 −0.2 −0.4 0 −0.2 −0.490 and 42.1 0 0.3 −0. Current Index Level Zt: 42. Current Index Level Zt: 42. The two levels of the current VIX are the 20th and 80th percent quantiles of the VIX time series of 2009.49 10 Nonparametric 95% Upper 95% Lower Parametric: SVCJ1 Parametric: SVCJ2 6 Nonparametric 95% Upper 95% Lower Parametric: SVCJ1 Parametric: SVCJ2 Log−Return Time−to−Maturity τ : 42.2 0.1 0.1 0.4 −0.3 4 6 5 3 4 2 3 2 1 1 0 −0.49 10 Nonparametric 95% Upper 95% Lower Parametric: SVCJ1 Parametric: SVCJ2 6 Nonparametric 95% Upper 95% Lower Parametric: SVCJ1 Parametric: SVCJ2 Time−to−Maturity τ : 21.1 0 0. Current Index Level Zt: 23.1 0.49 10 Nonparametric 95% Upper 95% Lower Parametric: SVCJ1 Parametric: SVCJ2 6 Nonparametric 95% Upper 95% Lower Parametric: SVCJ1 Parametric: SVCJ2 Log−Return Time−to−Maturity τ : 84.1 0.2 −0.3 Log−Return Time−to−Maturity τ : 84.2 −0.1 0 0.2 −0.3 Log−Return Time−to−Maturity τ : 42.3 −0.455 9 5 8 Density: Probability Percent 7 Density: Probability Percent −0.2 −0.2 0.3 −0.2 0. The parametric SPDs are obtained using formulas in Appendix B. and for two current VIX index levels of 23.1 0 0.455.4 0 −0.1 0.1 0 0. Current Index Level Zt: 23. Current Index Level Zt: 42.2 0. 42.Figure 8: Nonparametric Versus Parametric Estimates of the SPDs of the S&P Time−to−Maturity τ : 21. Current Index Level Zt: 23.3 −0. 44 .3 −0. Dotted lines are 95 percent conﬁdence intervals constructed from the asymptotic distribution theory in (6).3 4 6 5 3 4 2 3 2 1 1 0 −0.455 9 5 8 Density: Probability Percent 7 Density: Probability Percent −0.2 0. and 84 days.4 0 −0.4 −0.3 4 6 5 3 4 2 3 2 1 1 0 −0.1 0 0.1 0.2 0.455 9 5 8 Density: Probability Percent 7 Density: Probability Percent −0.4 −0.

05 Density: Probability Percent 0.Figure 9: Nonparametric Versus Parametric Estimates of the SPDs of VIX Time−to−Maturity τ: 21. Current Index Level Zt: 23. 45 . Current Index Level Zt: 23.04 Density: Probability Percent 20 25 30 35 40 45 50 55 60 0.01 0. and for two current VIX index levels of 23.04 Density: Probability Percent 20 25 30 35 40 45 50 55 60 0.06 Nonparametric 95% Upper 95% Lower Parametric: SVCJ1 Parametric: SVCJ2 0.01 0 0 −0.02 0.06 VIX at Expiration Time−to−Maturity τ: 42.02 0.49 0.04 0.01 15 −0. Current Index Level Zt: 42.03 0.49 0.01 15 20 25 30 35 40 45 50 55 60 VIX at Expiration Time−to−Maturity τ: 84.455 Nonparametric 95% Upper 95% Lower Parametric: SVCJ1 Parametric: SVCJ2 0.05 0. and 84 days.05 Density: Probability Percent 0.01 15 20 25 30 35 40 45 50 55 60 VIX at Expiration Time−to−Maturity τ: 42.01 15 −0.490 and 42.03 0.455 Nonparametric 95% Upper 95% Lower Parametric: SVCJ1 Parametric: SVCJ2 0. The parametric SPDs are obtained using formulas in Appendix B.03 0.01 0 0 −0.01 15 20 25 30 35 40 45 50 55 60 VIX at Expiration VIX at Expiration Note: This ﬁgure overlays nonparametric SPD estimates (solid lines) of the VIX with the corresponding parametric SPDs (dashed lines) with time-to-maturity equal to 21.01 0 0 −0.06 Nonparametric 95% Upper 95% Lower Parametric: SVCJ1 Parametric: SVCJ2 0.03 0.05 Density: Probability Percent 0.04 0.05 0.06 VIX at Expiration Time−to−Maturity τ: 84. The two levels of the current VIX are the 20th and 80th percent quantiles of the VIX time series of 2009.01 0. 42. Current Index Level Zt: 23.01 15 −0.02 0.01 0. Dotted lines are 95 percent conﬁdence intervals constructed from the asymptotic distribution theory in (6).04 0.03 0.02 0.06 Nonparametric 95% Upper 95% Lower Parametric: SVCJ1 Parametric: SVCJ2 0. Current Index Level Zt: 42.03 0.02 0.455 Nonparametric 95% Upper 95% Lower Parametric: SVCJ1 Parametric: SVCJ2 0.02 0.05 0.49 0.455. Current Index Level Zt: 42.04 Density: Probability Percent 20 25 30 35 40 45 50 55 60 0.06 Time−to−Maturity τ: 21.

2 Density: Probability Percent 700 750 800 850 900 950 1000 1050 1100 1150 1.49 95% Upper 95% Lower Zt=42.6 650 0.3 Density: Probability Percent 1.7 0.3 Density: Probability Percent 1. and 357 days.9 0.46 95% Upper 95% Lower 1.46 95% Upper 95% Lower 1. 294.1 1 1 0.5 x 10 −3 S&P 500 index at Expiration 1.8 0.8 0.2 Density: Probability Percent 700 750 800 850 900 950 1000 1050 1100 1150 1.6 650 0.7 0.7 0.3 1. 231.4 1.49 95% Upper 95% Lower Zt=42.6 650 700 750 800 850 900 950 1000 1050 1100 1150 S&P 500 index at Expiration 1.455 (dashed lines). The two levels of the current VIX index are obtained by the 20 and 80 percent quantiles of the VIX time series of 2009 and correspond to low and high levels of the VIX.5 x 10 −3 Time−to−Maturity τ: 294 Zt=23. and for two current VIX levels of 23.2 1.Figure 10: Estimates of S&P 500 Index SPDs When VIX Options are Unavailable 1.9 0.8 0.46 95% Upper 95% Lower x 10 −3 Time−to−Maturity τ: 231 Zt=23.7 0.2 1.46 95% Upper 95% Lower Time−to−Maturity τ: 357 Zt=23.1 1.490 (solid lines) and 42.4 1.6 650 700 750 800 850 900 950 1000 1050 1100 1150 S&P 500 index at Expiration S&P 500 index at Expiration Note: This ﬁgure plots nonparametric SPD estimates of the S&P 500 index at long maturities of 168.4 1. respectively.5 Zt=23. Dotted and dash-dotted lines are 95 percent conﬁdence intervals constructed from the asymptotic distribution theory in (7).9 0. respectively.49 95% Upper 95% Lower Zt=42.3 1.49 95% Upper 95% Lower Zt=42. 46 .4 1.5 x 10 −3 Time−to−Maturity τ: 168 1.1 1.9 0.8 0.1 1 1 0.

025 0.03 1 0.005 150 100 −1 150 100 50 0 10 20 30 40 50 60 Time−to−Maturity VIX Level Zt Time−to−Maturity VIX Level Zt Note: In the left panel.2 0 −0. we plot nonparametric estimates of the S&P 500 option prices and of the state-price density of the VIX using the two dimension reduction techniques.2 −0.5 0 −0.Figure 11: Robustness Checks 300 250 200 150 100 50 0 150 100 1000 50 800 0 700 900 1200 1100 1 S&P 500 Option Pices % Error S&P 500 Option Pices 0. 47 .6 SPD of VIX % Error 60 50 50 0 20 10 40 30 0.01 0.015 0.02 0.8 SPD of VIX 0.4 0.6 −0.5 −1 120 100 80 60 40 20 800 900 1000 1100 1200 Time−to−Maturity S&P Level St Time−to−Maturity S&P Level St 0. The right panel reports the percentage errors relative to the estimates without using the two dimension reduction techniques.8 0.4 −0.