Journal of Financial Economics 120 (2016) 1–20

Contents lists available at ScienceDirect

Journal of Financial Economics

journal homepage: www.elsevier.com/locate/finec

Analyzing volatility risk and risk premium in option

contracts: A new theoryR
Peter Carr a, Liuren Wu b,∗
a
Courant Institute, New York University, 251 Mercer Street, New York, NY 10012, USA
b
Baruch College, Zicklin School of Business, One Bernard Baruch Way, New York, NY 10010, USA

a r t i c l e i n f o a b s t r a c t

Article history: We develop a new option pricing framework that tightly integrates with how institutional
Received 29 December 2013 investors manage options positions. The framework starts with the near-term dynamics
Revised 29 June 2015
of the implied volatility surface and derives no-arbitrage constraints on its current shape.
Accepted 6 July 2015
Within this framework, we show that just like option implied volatilities, realized and ex-
Available online 19 January 2016
pected volatilities can also be constructed specific to, and different across, option contracts.
JEL Classification: Applying the new theory to the S&P 500 index time series and options data, we extract
C13 volatility risk and risk premium from the volatility surfaces, and find that the extracted
C51 risk premium significantly predicts future stock returns.
G12
© 2016 Elsevier B.V. All rights reserved.
G13

Keywords:
Implied volatility surface
Option realized volatility
Expected volatility surface
Volatility risk premium
Vega-gamma-vanna-volga
Proportional variance dynamics

1. Introduction

The option pricing literature has made great advances

during the past decade; yet large gaps remain between
theory and practice. First, traditional option pricing models
R
We thank William Schwert (the editor), Peter Christoffersen (the ref- specify the underlying price and variance rate dynamics
eree), Torben Anderson, Hans Buehler, Bruno Dupire, Robert Engle, Travis and derive their implications on option prices; however,
Fisher, Jeremy Graveline, Rachid Lassoued, Alex Levin, Keith Lewis, Dilip
institutional investors manage their volatility views and
Madan, Fabio Mercurio, Attilio Meucci, Alexey Polishchuk, Jason Roth, An-
gel Serrat, Mridul Tandon, Edward Tom, Ilya Ustilovsky, Arun Verma, Scott exchange their quotes not through option prices, but
Weiner, and seminar participants at Baruch College, the Fields Institute, through the option implied volatility computed from the
Northwestern University, Florida State University, Singapore Management Black-Merton-Scholes (BMS) model. This common practice
University, the 2011 Western Finance Association meetings in Santa Fe, does not mean that investors agree with the assumptions
the 2011 China International Conference in Finance in Wuhan, China, and
the 2013 Derivatives Conference at New York University for their com-
made by Black and Scholes (1973) and Merton (1973);
ments and suggestions. Liuren Wu gratefully acknowledges the support rather, they use the BMS model as a transformation to
by a grant (ENHC-44-33) from the City University of New York PSC-CUNY enhance quote stability and to highlight the information
Research Award Program. in the option contract. Second, traditional option pricing
∗
Corresponding author. Tel.: +1 646 312 3509; fax: +1 646 312 3451.
theory requires the full specification of the instantaneous
E-mail address: [email protected] (L. Wu).

http://dx.doi.org/10.1016/j.jfineco.2016.01.004
S0304-405X(16)0 0 0 05-2/© 2016 Elsevier B.V. All rights reserved.
2 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20

variance rate dynamics, not only about its current level, volatility surface may not satisfy no-arbitrage conditions,
but also about its long-run mean; yet in practice, investors may not be extrapolated with stability to regions where
do not observe the instantaneous variance rate, but in- data are sparse or unavailable, and the method does not
stead observe many option implied volatilities across a provide a mechanism to reduce the dimension of the sur-
wide spectrum of strikes and maturities. Furthermore, in- face to a few economically meaningful states. On the other
vestors have much more confidence on how these implied hand, a fully specified parametric model can provide stable
volatilities move in the near term than in the very long and arbitrage-free extrapolation, dimension reduction, and
run. The map between the implied volatility surface and economic interpretation, but it has issues regarding its
the instantaneous variance rate dynamics is not always stability over time, its poor performance when the state
clear or well-determined, forcing modelers to frequently dimension is low, and its numerical complexity and insta-
recalibrate their models to match moving market con- bility when the dimension is high. Our semi-parametric
ditions, with each recalibration generating a new set of theory balances the two by providing a numerically simple
parameters that are supposed to be fixed over time. Such approach to readily interpolate and extrapolate the surface
fudging practices create consistency concerns because the while satisfying dynamic no-arbitrage constraints, and to
option pricing function would differ if one expects these reduce the dimension of the surface to a few economic
parameters to be varying over time. states while leaving the state dynamics unspecified, thus
In this paper, we develop a new option pricing avoiding introducing any fixed model parameters.
framework that tightly integrates with how institutional The new theoretical framework does not replace the
investors manage their option positions, thus closing the role played by fully parametric, equilibrium-based option
gap between theory and practice. Instead of modeling the pricing models; instead, it can provide a bridge between
full dynamics of an unobservable instantaneous variance market observations and the fundamental valuations from
rate and deriving the implication on option prices, the these models. A well-specified parametric option pricing
new framework models the near-term dynamics of the model may not fit the current market observations well,
BMS implied volatility across different strikes and expiries, but its valuation can guide future market implied volatil-
and derives no-arbitrage constraints directly on the shape ity movements. If one believes that option implied volatil-
of the implied volatility surface. Under the assumed im- ities move toward their corresponding fundamental valua-
plied volatility dynamics, the shape of the whole implied tions from a parametric model, the new theory can read-
volatility surface can be cast as the solution to a simple ily embed the fundamental valuations from this model as
quadratic equation. The computational burden is dramat- the near-term targets of the implied volatility movements,
ically reduced compared to the standard option pricing and derives no-arbitrage constraints on the current shape
literature. More importantly, by starting with the whole of the implied volatility surface with the fundamental val-
implied volatility surface instead of a single instantaneous uation as its reference point. To do so, the new theory only
variance rate, the new theory does not need to specify the asks for the numerical valuation results from the paramet-
full dynamics, but just the current levels of the drift and ric model, without needing to know its parametric model
the diffusion processes. The current shape of the implied details.
volatility surface only depends on the current levels of Within the new theoretical framework, we propose a
its drift and diffusion processes, but does not depend on new concept that just like option implied volatilities, both
how these processes will evolve in the future. This “un- realized and expected volatilities can be made specific to,
spanned” nature allows the shape of the current implied and different across, option contracts. We define the op-
volatility surface to be represented as a function of many tion realized volatility (ORV) at each strike and expiry as the
state variables, but with no fixed model parameters. The volatility level at which one achieves zero realized profit if
high dimensionality renders the model flexible enough to one buys the option and performs daily delta-hedge based
fit the observed implied volatility surface well, whereas on the BMS model with this volatility input. Although this
the absence of fixed model parameters dramatically sim- realized volatility can be estimated from the realized secu-
plifies model estimation, alleviates concerns on model rity price sample path, it is defined against a specific op-
stability over time, and allows continuous model recali- tion contract and hence can differ across different strikes
bration to update the state variables without inducing any and expiries of the reference option contract. Since writ-
intertemporal inconsistency. ing the option at this ORV level generates zero profit, the
The fact that the new theory only specifies the near- ex post premium from writing the option at its market
term dynamics of the implied volatility surface while price is directly given by the BMS value difference when
leaving its long-term variation unspecified highlights its evaluated at the option’s implied volatility and its ORV
“semi-parametric” flavor:1 The theory specifies just enough level, respectively. This new option-specific volatility con-
dynamic structure to achieve a fully parametric charac- cept is tightly linked to the common practice of volatil-
terization of the current implied volatility surface, while ity investors, who usually take option positions and per-
saying little about its long-run variation. Traditionally, form dynamic delta hedging to separate the volatility ex-
one can either fit the surface parametrically or nonpara- posure from the directional price movement.2 Taking an
metrically. Nonparametric fitting is easy to do, but with option position with delta hedge exposes the investor to
concerns that the nonparametrically smoothed implied

2
Indeed, most institutional volatility investors and options market
1
We thank the referee for highlighting this feature. makers are required by their institutions to maintain delta neutrality.
 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20 3

variance risk during the life of the option, but the exposure not be larger than the present value of the strike price;
to the different segments of the sample path differs for dif- and put-call parity must hold. Type II bounds are derived
ferent option contracts. The ORV estimate for each option based on no-arbitrage arguments between options of dif-
contract represents a weighted average of the variance risk ferent strikes and maturities, such as the constraints that
over the sample path, with the weighting determined by bull, bear, calendar, and butterfly spreads must be no less
the risk exposure of that contract. than zero. Hodges (1996) shows that by quoting an option
To measure the ex ante volatility risk premium embed- in terms of a positive implied volatility, all Type I bounds
ded in each option contract, we propose to estimate an are automatically guaranteed. This property makes it very
option expected volatility (OEV) at each strike and expiry, attractive for market makers to quote and update implied
defined as the volatility forecast that generates zero ex- volatilities based on options order flows while using an
pected profit if one buys and delta-hedges the option at automated system to update the option prices whenever
this volatility level. The difference between the OEV sur- the underlying security price moves. Unfortunately, quot-
face and the option implied volatility surface defines the ing positive implied volatilities does not exclude Type II
volatility risk premium embedded in the option contracts arbitrages. Our new theory takes advantage of the BMS im-
across different strikes and maturities. Just as the current plied volatility transformation to exclude Type I arbitrages,
shape of the implied volatility surface is constrained by and derives no-arbitrage constraints directly on the current
its near-term risk-neutral dynamics, the current shape of shape of the implied volatility surface based on views on
the OEV surface is analogously constrained by its near- its near-term movements.
term statistical dynamics. These constraints allow us to In related work, Bakshi and Kapadia (2003a, 2003b) ar-
perform dimension reduction and extract meaningful eco- ticulate the idea that one can analyze the volatility risk
nomic states from the two surfaces. premium by investigating the delta-hedged gains from
We apply the new theoretical framework to the S&P each option contract. Our new theoretical framework for-
500 index (SPX) time series and its options. Our data in- malizes their insights via the concept of option-specific ex-
clude nearly 18 years of over-the-counter SPX option im- pected and implied volatilities. To understand the risk pro-
plied volatility quotes from January 1997 to October 2014. file of a portfolio of delta-hedged option positions, Engle
At each date, the quotes are at a fixed grid of five rela- and Figlewski (2015) propose a statistical model for the dy-
tive strikes from 80% to 120% of the spot level and eight namics and correlations of implied volatilities across dif-
fixed time to maturities from 1 month to 5 years. Corre- ferent individual stocks. Also related to our work is the
sponding to these implied volatility quotes, we estimate growing literature on variance risk premium. Carr and Wu
the ORV at the corresponding relative strike and maturity (2009) propose to use the difference between expected fu-
levels based on the realized SPX sample paths, and we also ture realized variance and the variance swap rate to mea-
propose a statistical procedure to estimate the correspond- sure the variance risk premium. A growing list of stud-
ing OEV forecast based on exponential moving averages of ies build upon this variance risk premium measure, from
BMS values of historical ORV estimates. developing theories explaining the large variance risk pre-
Given the unique feature of the new theoretical frame- mium (Baele, Driessen, Londono, and Spalt, 2014; Drech-
work that the volatility surfaces are functions of several sler and Yaron, 2011), modeling the variance swap term
state variables but with no fixed model parameters, we structure and developing variance swap allocation strate-
propose a state-updating procedure based on an extended gies (Egloff, Leippold, and Wu, 2010), documenting vari-
version of the classic Kalman (1960) filter. With this proce- ance risk premium in other markets (Mueller, Vedolin, and
dure, we can fit thousands of volatility surfaces and extract Yen, 2012), to relating the equity variance risk premium
the corresponding state variables in a matter of seconds. to other financial markets (Bollerslev, Tauchen, and Zhou,
By fitting the statistical OEV dynamics to the current 2009; Zhang, Zhou, and Zhu, 2009). Since over-the-counter
OEV surface shape and fitting the risk-neutral implied variance swap rates are not readily available, most of these
volatility dynamics to the current implied volatility surface studies use vanilla options to form a replicating portfo-
shape, we obtain two sets of dynamics estimates that high- lio to approximate the variance swap rate (Carr and Wu,
light how different economic states vary at different histor- 2006; Jiang and Tian, 2005). Our new framework provides
ical sample periods. The differences between the two sets a platform for analyzing volatility risk and volatility risk
of dynamics also highlight how the volatility risk premium premium in each option contract, without resorting to op-
varies over time. We project the volatility risk premium es- tion portfolio formulation.
timates to a return risk premium component based on the Finally, there have been some largely unsuccessful
return-volatility correlation, and find that this projected re- attempts in the literature in directly modeling the implied
turn risk premium can predict future stock returns. volatility dynamics. Examples include Avellaneda and Zhu
In a classic paper, Merton (1973) develops model-free (1998), Ledoit and Santa-Clara (1998), Schonbucher (1999),
bounds on option prices arising from no static arbitrage. Hafner (2004), Fengler (2005), and Daglish, Hull, and Suo
These bounds can be classified into two types. Type I (2007). These models are often called market models of
bounds are derived based on no-arbitrage arguments be- implied volatility. These attempts have completely differ-
tween European options at a fixed strike and maturity ver- ent starting point and ending objectives from our analysis.
sus the underlying security and cash. Examples include: Instead of deriving no-arbitrage constraints on the implied
Call and put prices must not be smaller than their in- volatility surface, these attempts take the observed im-
trinsic value; call prices on a stock must not be larger plied volatility (on a single option, a curve, or over the
than the dividend discounted stock price; put prices must whole surface) as given while specifying the continuous
4 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20

martingale component of the volatility surface. From these standard Brownian motion W under Q such that S solves
two inputs, they try to derive the no-arbitrage restrictions the following stochastic differential equation:
on the risk-neutral drift of the surface. The approach is √
dSt /St = vt dWt , (1)
analogous to the Heath, Jarrow, and Morton (1992) model
on forward interest rates and can in principle be used where vt denotes the time-t instantaneous variance rate.
for pricing derivatives written on the implied volatility We allow v to follow a positive, real-valued stochastic pro-
surface. What this literature fails to recognize is that cess such that there exists a unique solution to (1). How-
the knowledge of the current implied volatility surface ever, in contrast to existing literature, we do not spec-
places constraints on the specification of the continuous ify the risk-neutral dynamics of this process; instead, we
martingale component for its future dynamics. In this will specify the risk-neutral dynamics of the BMS implied
paper, rather than ignoring these constraints, we fully volatility for each vanilla option:
exploit them in building a simple, direct linkage between
dIt (K, T ) = μt dt + ωt dZt , | ln(K/St )| < K, 0 < T − t < T ,
the current shape of the implied volatility surface and its
near-term dynamics. (2)
The remainder of the paper is organized as follows. for some fixed finite levels of absolute moneyness K > 0
Section 2 establishes the new theoretical framework by and time to maturity T > 0. We refer to μt as the drift
specifying near-term implied volatility dynamics and de- process and ωt as the volatility of volatility process (hence-
riving the allowed shapes for the current implied volatility forth “volvol” for short). Both processes can be stochastic
surfaces that exclude dynamic arbitrage. Section 3 defines and they can both depend on deterministic quantities such
the option realized and expected volatility surfaces across as calendar time t, strike price K, and maturity date T. In
strikes and expiries based on the security price sample contrast to μt and ωt , the standard Brownian motion Zt is
paths, and shows how the future statistical dynamics of independent of the strike K and maturity T at all times.
the expected volatility surface determine the current shape The two Brownian shocks on the stock price and the im-
of the surface. Section 4 documents the stylized evidence plied volatilities are allowed to be correlated,
on the option implied and expected volatility surfaces
Et [dWt dZt ] = ρt dt, (3)
for the S&P 500 index. Section 5 proposes a dynamic
estimation procedure for extracting the economic states where ρ t is a stochastic process taking values in the inter-
from the two volatility surfaces. Section 6 discusses the val [−1, 1].
estimation results. Section 7 provides concluding remarks It is worth noting that Eq. (1) assumes a purely contin-
and directions for future research. uous security price dynamics, thus excluding discontinuous
price movements from the security price specification, and
Eq. (2) makes the strong assumption that instantaneously,
2. Implied volatility surface: from near-term dynamics
the implied volatility surface is driven by one Brownian
to current shape
shock. Furthermore, we limit our consideration of the im-
plied volatility surface to a finite range of moneyness K
We consider a market with a riskfree bond, a risky as-
and time to maturity T so that we can specify the dynam-
set, and a continuum of vanilla European options written
ics without worrying about the limiting behaviors of the
on the risky asset.3 For simplicity, we assume zero inter-
implied volatility surface at extreme moneyness and ma-
est rates and zero carrying cost/benefit for the risky asset.
turity.4
In practical implementation, one can readily accommodate
a deterministic term structure of financing rates by model- 2.1. The fundamental PDE governing the implied volatility
ing the forward value of the underlying security and defin- surface
ing moneyness of the option against the forward. The risky
asset can be any types of tradable securities, but we will For concreteness, let the basis option be a call with
refer to it as the stock for concreteness. We assume fric- Ct (K0 , T0 ) denoting its value, and let all other options be
tionless and continuous trading in the riskfree bond, the puts, with Pt (K, T) denoting the corresponding values. Let
stock, and a basis option, and we assume no-arbitrage be- B(S, σ , t; K, T ) : R+ × R+ × [0, T ) → R+ be the BMS model
tween the stock and the bond. As a result, there exists a formula for a European put option:
risk-neutral probability measure Q, equivalent to the sta-  √ 
tistical probability measure P, such that the stock price S ln(K/S ) σ T −t
B(S, σ , t; K, T ) ≡ KN √ +
is a martingale. σ T −t 2
We assume that the stock price S evolves in continuous  √ 
ln(K/S ) σ T −t
time as a strictly positive and continuous semi-martingale. −SN √ − . (4)
By the martingale representation theorem, there exists a σ T −t 2
To reduce notation clutter, we henceforth suppress the no-
tational dependence of B on contract characteristics K and
3
In the US, exchange-traded options on individual stocks are all Amer- T when no confusion shall occur.
ican style. To apply our new theory to American options, a commonly
used shortcut is to extract the BMS implied volatility from the price of an
4
American option based on some tree/lattice method and use the implied See Benaim, Friz, and Lee (2008) for an analysis of the limiting be-
volatility to compute a European option value for the same expiry and haviors of implied volatility at extreme strikes, and Tehranchi (2009)
strike. See Carr and Wu (2010) for a detailed discussion on data process- for an analysis of the implied volatility asymptotic behavior at long
ing on individual stock options. maturities.
 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20 5

By the definition of BMS implied volatility, we can write Not only is the value function B(St , It , t) well known,
both the basis call option and the other put options in so are its various partial derivatives. Plugging these partial
terms of the BMS put formula, derivatives into the PDE in Eq. (6), we can reduce the PDE
into an algebraic equation that links the implied volatil-
Ct (K0 , T0 ) = B(St , It (K0 , T0 ), t ) + St − K0 ,
ity dynamics to the current shape of the implied volatility
Pt (K, T ) = B(St , It (K, T ), t ), (5) surface. This algebraic equation becomes particularly sim-
for all t ≥ 0, K > 0, and T > t. It is well known that the ple if we represent the current implied volatility surface as
function B(S, σ , t) is C2, 2, 1 on R+ × R+ × [0, T ), so Itô’s a function of the relative strike k ≡ ln (K/S) and time to
formula can be used to relate increments of B to the incre- maturity τ ≡ T − t, i.e., It (k, τ ).5
ments of S, I, and t. To shorten the length of the following Proposition 2. The fundamental PDE in (6) can be translated
equations, we let subscripts of B denote partial derivatives into a no dynamic arbitrage constraint on the current shape
and we suppress the arguments of B, which are always (St , of the implied volatility surface It (k, τ ), jointly determined by
It (K, T), t). the current instantaneous variance rate level vt , the current
Requiring the implied volatility for any option at (K, T) levels of the instantaneous drift (μt ) and volatility (ωt ) of
to be positive, It (K, T) > 0, guarantees no static arbitrage the implied volatility dynamics, and the current level of the
between the options at (K, T) and the underlying stock and instantaneous correlation process between return and implied
cash (Hodges, 1996). We further require that no dynamic volatility (ρ t ):
arbitrage be allowed on any option at (K, T) relative to the  
basis option at (K0 , T0 ), the stock, and cash. This require- 1 2 1 ωt √ I2 τ
0= It − μt It τ − vt − ρt vt k + t
ment for no dynamic arbitrage leads to a fundamental par- 2 2 It 2
tial differential equation (PDE) between the functions B(S, 
1 ωt2 2 1 4 2

σ , t) and It (K, T). − k − It τ . (7)
2 It2 4
Proposition 1. Under the stock price dynamics in (1), the im-
plied volatility dynamics in (2), and the correlation specifica- Refer to Appendix A for the proof.
tion in (3), the absence of dynamic arbitrage on an option By specifying particular parametric functional forms for
contract Pt (K, T) relative to the basis option at (K0 , T0 ), the μt and ωt , one can determine the algebraic nature of
stock, and a riskfree bond dictates that the BMS option pric- the manifold (7) in which the implied volatility function
ing function B(S, σ , t) and the implied volatility function It (K, It (k, τ ) resides. Just as integral transforms often convert
T) for this option jointly solve the following fundamental PDE: partial differential equations into algebraic equations, the
use of implied volatility has transformed the second-order
1 √ 1 parabolic PDE in (6) into the simple algebraic relation in
−Bt = μt Bσ + vt S2 BSS + ρt ωt vt St BSσ + ωt2 Bσ σ . (6)
2 t 2 (7). Under our dynamic assumptions for the stock price
Refer to Appendix A for the proof. and the implied volatilities, a necessary condition aris-
In the fundamental PDE in (6), the terms involving par- ing from no dynamic arbitrage is that the current implied
tial derivatives of B are called theta for Bt , vega for Bσ , dol- volatility surface It (k, τ ) resides in the manifold defined by
lar gamma for St2 BSS , dollar vanna for St BSσ , and volga for Eq. (7).
Bσ σ . When μt and ωt are independent of (K, T), Eq. (6) The no-arbitrage constraint embedded in equation in
defines a linear relation between the BMS theta of the op- (7) links the current shape of the implied volatility sur-
tion and its vega, dollar gamma, dollar vanna, and volga. face to the current levels of the drift process μt and the
We christen the class of implied volatility surfaces defined diffusion process ωt for the implied volatility dynamics, as
by the fundamental PDE as the Vega-Gamma-Vanna-Volga well as current levels of the correlation process ρ t and the
(VGVV) model. instantaneous variance rate process vt ; however, the no-
By assuming a zero drift on the implied volatility dy- arbitrage condition places no direct constraints on the ex-
namics, and further assuming that the volvol process ωt act dynamics specifications for these four processes (μt ,
is independent of K and It (K, T0 ), Arslan, Eid, Khoury, and ωt , ρ t , vt ). Thus, the constraint on the current implied
Roth (2009) derive a special case of Eq. (6) that involves volatility surface shape only comes from the near-term dy-
only gamma, vanna, and volga of the option, but not the namics of the implied volatility surface.
option’s vega. They label their model as the GVV model
and use it for calibrating the implied volatility smile at one 2.2. Proportional volatility dynamics
maturity.
It is important to note that the PDE in Eq. (6) is not a Different parameterizations for the drift μt and the dif-
PDE in the traditional sense. Traditionally, a PDE is speci- fusion ωt of the implied volatility dynamics lead to differ-
fied to solve the value function. In our case, the value func- ent functional shapes for the implied volatility surface. As
tion B(St , It , t) is simply the BMS put option formula in (4). an illustrating example, we consider one particularly sim-
Furthermore, the coefficients on traditional PDEs are deter- ple specification, where both the drift and the diffusion are
ministic, but they are stochastic in our PDE. Most impor-
tant, our PDE is not derived to solve the value function, 5
To avoid introducing too many different notations, we use It to denote
but rather to show that the various stochastic quantities both the implied volatility value at time t and the various representations
have to satisfy this particular relation to exclude dynamic of the implied volatility function. The different representations are differ-
arbitrage. entiated by the arguments that follow.
6 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20

proportional to the implied volatility level: At a fixed time to maturity, Eq. (10) describes a hyper-
bola in the variables I2 and k,
dIt (K, T )/It (K, T ) = e−ηt (T −t ) (mt dt + wt dZt ), wt , ηt > 0, 2
It2 (k ) = at + (k − bt )2 + ct , (11)
(8) τ
where mt , wt , and ηt are stochastic processes that do not with
√
depend on K, T, or I(K, T). We constrain wt to be a strictly −2(1 − 2e−ηt τ mt τ − e−ηt τ wt ρt vt τ )
at = ,
positive process with no loss of generality, and we use the e−2ηt τ wt2 τ 2
exponential dampening e−ηt (T −t ) to accommodate the em- √
ρ vt
pirical observation that implied volatilities for long-dated bt = − −η τ ,
e t wt
options tend to move less. √
(1 − ρt2 )vt (1 − 2e−ηt τ mt τ − e−ηt τ wt ρt vt τ )
2
Eq. (8) represents a minimalist structure that captures
ct = + .
the current levels of the drift and diffusion of the implied e−2ηt τ wt2 e −4 ηt τ wt τ
4 2

volatility surface. Full dynamic specifications for stochas- (12)

tic volatilities often accommodate mean reversion in the
drift. Given our near-term focus, we use mt to parsimo- Eq. (11) dictates that the implied variance smile is It2 (k )
niously capture the direction and magnitude of the next convex in k. As k → ±∞, the implied variance smile be-
expected move without delving into the particular drivers haves linearly in k. Furthermore, Eq. (11) shows that the
of the move. implied volatility smile at a fixed maturity is determined
The literature often specifies the dynamics on the vari- by three transformed covariates (at , bt , ct ). The remaining
ance instead of volatility. Under the specification in (8), degrees of freedom determine how the smiles vary across
the implied variance dynamics also possess a proportional different maturities.
structure: Gatheral (2006) proposes a similar, but more general
form for the implied variance smile that involves five free
dIt2 (K, T )/It2 (K, T ) covariates,
 
1 −ηt (T −t ) 2
 
= 2e−ηt (T −t ) mt + e wt dt + wt dZt . (9) It2 (k ) = a + b ρ (k − m ) + ( k − m )2 + σ 2 . (13)
2
It is worth noting that our implied variance diffusion spec- Gatheral labels his specification as the SVI (stochas-
ification deviates from the commonly used affine setting. tic volatility inspired) model. Although Gatheral makes
Eq. (9) dictates that in the limit, the diffusion for the in- some motivational linkages between these coefficients and
stantaneous variance rate vt is also proportional to the stochastic volatility models, no stochastic volatility models
variance rate level. By contrast, the commonly used affine have been proposed that lead exactly to this generalized
variance rate dynamics, e.g., Heston (1993), specifies the SVI form. In particular, it is not clear how the five coeffi-
diffusion for the variance rate as proportional to the square cients should vary across different time to maturities.
root of the variance rate. The traditional literature relies In the limit of τ = 0, the implied variance is quadratic
heavily on the affine setting mainly for pricing tractability. in the relative strike k,
Under our framework, the pricing is extremely tractable for √
It2 (k ) = vt + 2ρt vt wt k + wt2 k2 , (14)
a wide array of dynamics specifications. Our decision to
deviate from the affine setting is motivated by better em- with the volvol process wt dictating the curvature and the
pirical performance. correlation ρ t dictating the asymmetry of the smile.
If we define “at-the-money” as the relative strike equal
Proposition 3. Under the stock price dynamics in (1)–(3), to the conditional mean of the log stock return ln (ST /St )
when the implied variance follows the proportional dynamics under the BMS model, k = − 21 It2 (k, τ )τ , the at-the-money
in (8), no dynamic arbitrage requires that the implied vari- implied variance takes the simple form,
ance surface as a function of relative strike k and time to ma- vt
turity τ , It2 (k, τ ), satisfy the following quadratic equation, At2 (τ ) = , (15)
1 − 2e−ηt τ mt τ
1 −2ηt τ 2 2 4  √  where the at-the-money implied variance term structure
0= e wt τ It + 1 − 2e−ηt τ mt τ − e−ηt τ wt ρt vt τ It2
4 depends on the drift specification (μt ), but is unaffected
 √ 
− vt + 2e−ηt τ wt ρt vt k + e−2ηt τ wt2 k2 . (10) by the choice of the volvol process ωt , nor by the choice
of the correlation process ρ t . The term structure starts at
Refer to Appendix A for the proof. vt at τ = 0. As τ initially increases, the term structure can
An interesting and important feature of Eq. (10) is that be either upward or downward sloping, depending on the
the no-arbitrage constraint depends on the current levels drift of the implied volatility process mt . It is upward slop-
of the five dynamic processes (mt , wt , ηt , vt , ρ t ), but it does ing when the drift is positive (mt < 0) and downward slop-
not depend directly on the exact dynamics of these pro- ing when the drift is negative (mt < 0).
cesses. Thus, the dynamics of the five state variables are
left unspecified. As a result, fitting the relation to observed 2.3. A bridge between market observations and fundamental
implied volatility surfaces only involves extracting the lev- option valuations
els of the five dynamic states, but does not involve the
estimation of any model parameters that govern the state Given the completely different starting points, one
dynamics. naturally wonders whether implied volatility dynamics
 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20 7

specified under the new theoretical framework can be the valuation at different strikes and maturities, with no
mapped tractably to corresponding dynamics for the direct reference to the particular model dynamics specifi-
instantaneous variance rate in the traditional modeling cation. This separation highlights the flexibility of the new
framework, and vice versa. While this endeavor can be an theoretical framework, as it can build upon any fundamen-
interesting direction for future research, making such a tal model by capturing the market reversion to the funda-
mapping is likely to be very difficult, except under certain mental valuation.
special cases (e.g., Carr and Sun, 2007). What makes the
mapping particularly difficult is the fact that under the 3. Option-contract specific realized and expected
new framework, the implied volatility dynamics are not volatilities
fully specified. We specify merely the current drift and
diffusion levels of the implied volatility dynamics, while Corresponding to the option implied volatility surface,
saying nothing about their future movements. This partial we propose the new concept that realized and expected
specification gives us tremendous flexibility in fitting the volatilities can also be defined in reference to a specific op-
implied volatility surface by updating the values of a set tion contract.
of state variables, but without the need to pin down the
dynamics of these state variables.
3.1. Defining realized and expected volatility specific to an
Given the partial specification, the new theory does not
option contract
replace the roles played by fully parametric, equilibrium-
based option pricing models; instead, it can provide a
For each option contract, we define its option realized
bridge between market observations and the fundamental
volatility (ORV) as the volatility input to the BMS model
valuations from these models. A well-specified parametric
such that if one buys the option at this volatility level, with
option pricing model may not fit the current market obser-
the invoice price generated from the BMS pricing formula,
vations well, but its valuation can guide future market im-
and performs daily BMS delta hedge on the option based
plied volatility movements. Specifically, if one believes that
on this volatility level throughout the life of the option
market-observed option implied volatilities tend to move
contract, the terminal profit and loss (PL) is zero. Formally,
toward their corresponding fundamental valuations from
let ORV(K, t, T ) denote the option realized volatility corre-
a parametric model, we can capture this reversion behav-
sponding to the reference option contract initiated at time
ior via the following implied volatility dynamics specifica-
t, struck at K, and expiring at T, and let {t j }Nj=1 denote the
tion,
  sequence of days during the life of the option with t0 = t
dIt2 (K, T ) = κt Mt (K, T ) − It2 (K, T ) dt and tN = T , we can compute the ORV as
+2e−ηt (T −t ) wt It2 (K, T )dZt , (16) ORV(K, t, T ) ≡ x,
where the diffusion component remains the same as (9), s.t. 0 = B(ST , x, T ) − B(St , x, t )
but the size and direction of the drift is dictated by the N   
deviation between the current market implied variance − BS St j−1 , x, t j−1 St j − St j−1 , (18)
level and the model valuation Mt (K, T), which denotes the j=1
value of the BMS implied variance generated from the
where the second line defines the PL from buying the ref-
particular parametric option pricing model. Eq. (16) rep-
erence option contract and performing daily delta hedge,
resents an analogous continuous-time specification to the
with x being the volatility input, B(St j , x, t j ) denotes the
error-correction specification of Engle and Granger (1987).
BMS value of the reference option contract on date tj , and
Through the error-correcting drift specification, Eq. (16)
BS (St j , x, t j ) denotes the corresponding BMS delta of the
drives the implied variance toward the model value Mt (K,
contract. The ORV is the BMS volatility level that makes
T), with κ t controlling the error-correcting speed.
the delta-hedged PL zero.
Given the specification in (9), one can analogously
Although ORV is defined on a particular reference op-
derive no-arbitrage constraints on the observed implied
tion contract, Eq. (18) highlights the fact that the value of
volatility surface by treating Mt (K, T) as a numeric input
ORV does not depend on the market price of that contract,
for each contract.
but rather only on the sample path of the underlying se-
1 −2ηt τ 2 2 4 curity price. Its meaning becomes clearer if we further as-
0= e wt τ It (k, τ )
4 sume that the underlying price dynamics are purely con-
 √ 
+ 1 + κt τ + e−2ηt τ wt2 τ − e−ηt τ wt ρt vt τ It2 (k, τ ) tinuous with stochastic volatility, in which case Carr and
 √  Madan (2002) show that if one buys the options at BMS
− vt + κt Mt (k, τ )τ + 2e−ηt τ wt ρt vt k + e−2ηt τ wt2 k2 . (17)
volatility x and performs continuous delta hedge at this
Eq. (17) essentially contains results from two layers of dy- volatility, the delta-hedged PL can be written as
namic modeling. The first layer is the traditional para- T
1 2
metric option pricing model that generates the benchmark PL = S BSS (Su , x, u )(vu − x2 )du. (19)
t 2 u
implied volatility surface valuation Mt (K, T). The second
layer is our near-term error-correction dynamics assump- Thus, setting this PL to zero to solve for x amounts to com-
tion that dictates how the observed implied volatility sur- puting x2 as a weighted average of the instantaneous vari-
face should vary around the parametric model valuation. ance rate with the weight given by the dollar gamma of
The first layer enters into the second layer only through the option at each point of the sample path. Therefore,
8 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20

based on the same price sample path, one can arrive at dif- 3.2. No-arbitrage constraints on the option expected
ferent variance estimates for different option contracts due volatility surface
to the different dollar gamma weighting. The traditionally
defined realized variance can be regarded as a special case Analogous to the current shape of the implied volatility
of this definition with equal weighting to each day’s real- surface being determined by its future risk-neutral dynam-
ization. Indeed, one can think of the traditionally defined ics, the current shape of the expected volatility surface is
realized variance with equal weighting as an ORV corre- constrained by its future statistical dynamics. In parallel to
sponding to the variance swap contract, which has con- the risk-neutral proportional implied volatility dynamics in
stant dollar gamma. (8), we assume that the option expected volatility is pro-
Given an estimate of the ORV for a particular op- portional to its corresponding option implied volatility and
tion contract, it becomes immediately clear that selling that the two also follow proportional dynamics under the
the contract makes money if its BMS implied volatility is statistical measure P,
higher than the ORV and loses money if its implied volatil-  
dVt (K, T )/Vt (K, T ) = e−ηt (T −t ) mtP dt + wt dZtP , (23)
ity is lower than the ORV. The delta-hedged PL from buy-
ing an option contract can be computed directly as the where we use a different drift process mtP
to capture the
BMS value difference between using the ex post option re- effects of market pricing of the volatility risk dZt . Under
alized volatility and the option’s ex ante implied volatility the continuous price dynamics assumption, both the im-
as inputs, respectively. While the implied volatility It (K, T) plied volatility surface and the expected volatility surface
√
is known at time t, ORV(K, t, T ) is not fully realized until converge to the same instantaneous volatility rate vt as
time T. The ex post realized dollar PL from buying the op- the time to maturity approaches zero. The two surfaces
tion contract and delta-hedging to expiration can be writ- are also governed by the same return-volatility correla-
ten as tion level ρ t . However, when the underlying security price
can jump randomly, the instantaneous variance becomes
PL(t, T ) = B(S, ORV(K, t, T ), t; K, T ) − B(S, It (K, T ), t; K, T ). an expectation of both diffusive movements and random
jumps. When the jump risk is priced, the expectation can
(20)
generate different values under the two measures P and Q.
The expected skewness of the return distribution can also
Since the BMS option pricing formula is monotonic in its
differ under the two measures.6 To partially accommodate
volatility input, the sign of the PL is determined by the
the impacts of random price jumps, we relax the model as-
sign of the difference between the ORV and the implied
sumption and use different variance rates (vtP ) and return-
volatility.
variance correlations (ρtP ) under the statistical measure Pt
Taking expectations on the ex post realized PL, one can
to better match the expected volatility surface behaviors,
obtain the ex ante expected volatility risk premium em-  
bedded in each option contract. To facilitate the ex ante dSt /St = μtP dt + vtP dWt , EtP dWtP dZtP = ρtP dt. (24)
volatility risk premium calculation, we propose a corre-
sponding option expected volatility (OEV) surface, Vt (K, T), Proposition 4. Under the stock price dynamics in (24) and the
defined as the time-t volatility forecast for each option expected volatility dynamics in (23) under the statistical mea-
contract at strike K and expiry T such that the expected PL sure P, the expected variance surface as a function of relative
is zero if one buys this option and delta-hedge to expira- strike k and time to maturity τ , Vt2 (k, τ ), satisfies the follow-
tion at this volatility level: ing quadratic equation,
1 −2ηt τ 2 2 4
  
0= e wt τ Vt + 1 − 2e−ηt τ mtP τ − e−ηt τ wt ρtP vtP τ Vt2
B(S, Vt (K, T ), t; K, T ) = EtP [B (S, ORV(K, t , T ), t ; K, T )], (21) 4
  
− vtP + 2e−ηt τ wt ρtP vtP k + e−2ηt τ wt2 k2 . (25)
where EtP [·] denotes the expectation operator under the
statistical measure P conditional on time-t filtration Ft . Ac-
Refer to Appendix A for the proof.
cording to this definition, if the current implied volatil-
ity level for the option contract is equal to the option’s
3.3. Linking variance risk premium to return risk premium
expected volatility, It (K, T ) = Vt (K, T ), the expected delta-
hedged PL from buying this option would be zero. On the
We assume that the presence of risk premium leads to
other hand, if the implied volatility level differs from the
a difference in the drift process of the implied volatility
expected volatility level, the expected delta-hedged PL, or
dynamics: mt under the risk-neutral measure and mtP un-
the volatility risk premium (VRP) from buying this option
der the statistical measure. The difference between the two
can be computed simply as,
can be regarded as a measure of the instantaneous vari-
ance risk premium. Furthermore, since the return innova-
VRPt (K, T ) = B(S, Vt (K, T ), t; K, T ) − B(S, It (K, T ), t; K, T ). tion dWt and the variance innovation dZt are correlated,
(22) estimates on the variance risk premium have direct impli-
cations on the return risk premium.
Therefore, the difference between this option expected
volatility surface and the implied volatility surface defines 6
See, for example, Polimenis (2006) for an illuminating discussion on
the surface of the volatility risk premium across different how relative risk aversion interacts with the return cumulants under the
strikes and maturities. two measures.
 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20 9

Formally, we can perform a decompose on the stock re- 4.1. The average behavior of volatility surfaces and volatility
turn Brownian stock, risk premiums

dWt = ρt dZt + t ,
1 − ρt2 dW Table 1 reports the sample averages of the 40 implied
t denotes the component of the return risk inde- volatility series in Panel A. At each fixed time to matu-
where dW
rity, the average implied volatility levels are higher for
pendent of the variance risk. With the decomposition, we
low strikes than for high strikes, forming the well-known
assume the following pricing kernel dynamics,
negatively skewed implied volatility smirk pattern that is
√ √ t .
dMt /Mt = −γt vt dZt − ζt vt dW (26) common across all global equity indexes (Foresi and Wu,
2005). At a fixed relative strike level, the average im-
With this pricing kernel specification, the variance risk plied volatility declines with maturity at low strikes but
premium is given by increases with maturity at high strikes. In particular, the
√ at-the-money implied volatilities show an average upward
mt − mtP = −γt wt vt . (27)
sloping term structure.
 instantaneous return risk premium is given by (γt ρt +
The Panel B reports the sample averages of the correspond-
ζt 1 − ρt2 )vt . Without knowing ζ t , we cannot fully iden- ing option expected volatility (OEV) estimates at each rela-
tify the return risk premium, but we can estimate the con- tive strike and maturity. The negative skew along the strike
tribution of the variance risk premium to the return risk dimension also shows up on the OEV surface, but the skew
premium (RRP) as is not as monotone and becomes more of a smile, espe-
cially at short maturities. At a fixed relative strike level,
RRP = γt ρt vt . (28) the OEV term structure is downward sloping at low rela-
tive strikes, but mostly flat at other strikes, forming a con-
When the return-variance correlation is high in absolute
trast with the upward sloping term structures observed on
magnitude, this RRP component becomes the major con-
at-the-money and high-strike implied volatilities. The fact
tributor to the return risk premium.
that the implied volatility mean term structure is more up-
ward sloping than the expected volatility mean term struc-
4. Implied and expected volatility surfaces on S&P 500 ture suggests that on average, mt > mtP and hence the mar-
index options ket price of the variance risk γ t , defined in the pricing ker-
nel specification in (26), is negative.
We use options on the S&P 500 index to perform an The difference between the option expected volatil-
empirical analysis on the new theory. We obtain matrix ity and the implied volatility defines the volatility risk
implied volatility quotes on SPX options from a major premium on each option contract in volatility percentage
bank. The quotes are constructed to match the listed op- points. A positive difference indicates positive expected
tion prices at short maturities and to match the over-the- PL from taking a long position in the option and delta-
counter (OTC) transactions at long maturities. The matrix hedging until expiration, and hence a positive volatility
quotes are on a grid of five fixed relative strikes from 80% risk premium. Panel C of Table 1 reports the average dif-
to 120% of the spot level and eight fixed time to maturi- ference across different strikes and maturities. The aver-
ties from 1 month to 5 years. The data are available from age volatility risk premium is mostly negative except on
January 8, 1997 to October 29, 2014. For our analysis, we high-strike, short-maturity options. The volatility risk pre-
sample the data weekly every Wednesday to avoid week- mium is particularly negative for far out-of-the-money put
day effects. The weekly sampled data include 40 implied options, where the average implied volatility can be over
volatility series over 930 weeks, a total of 37,200 observa- ten volatility points higher than the corresponding average
tions. OEV.
Corresponding to the implied volatility quotes, we ob- To gauge the economic significance of the volatility
tain an extended sample of the SPX daily time series start- risk premium, Panel D reports the annualized informa-
ing from January 8, 1982. At each date t and corresponding tion ratio of a long option strategy: At each date t, we
to each implied volatility quote It (k, τ ), we compute a his- buy the option at (k, τ ) and perform delta-hedge un-
torical option realized volatility, ORV(k, t − τ , t ), using the til expiration. The log expected return from this invest-
SPX time series data from time (t − τ ) to time t. We start ment can be computed as ln B(Vt (k, τ ))/B(It (k, τ )), where
the historical ORV calculation 10 years earlier than the im- B(It (k, τ )) denotes the market cost of buying this option
plied volatility sample from January 8, 1987, which needs and B(Vt (k, τ )) denotes the expected delta-hedged pay-
the time series data going back 5 additional years to Jan- off from the long option position. We define the annu-
uary 8, 1982 for maturities up to 5 years. In estimating the alized information ratio as the ratio of the mean log re-
conditional expectation of the BMS transformation at each turn to the standard deviation of the log return, annualized
time t and for each (k, τ ) reference point, we apply expo- by the square root of the time to maturity of the option.
nential moving average to the BMS value of the historical The information ratio estimates are highly negative for all
ORV estimates, with an exponential decay speed of 0.03 low-strike options and at-the-money options, but they be-
per day. The 10-year additional history is for the exponen- come positive for some short-term high-strike options, as
tial moving average estimates to stabilize. We then invert the OTC implied volatility quotes on these options are on
the moving average of the BMS value to obtain the OEV average lower than the corresponding expected volatility
estimate Vt (k, τ ). estimates.
10 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20

Table 1
Average behavior of option volatilities and volatility risk premiums on the S&P 500 index.
Entries report the sample average of the S&P 500 index option implied volatilities in Panel A, the corresponding
option expected volatilities (OEV) in Panel B, the average expected-implied volatility difference in Panel C, and the
annualized information ratio from long the option contracts in Panel D. The statistics are computed based on 40 over-
the-counter implied volatility quotes series on the S&P 500 index options at a matrix grid of five relative strikes (K/S)
and eight time to maturities (τ , in months). The corresponding OEV series are estimated based on the SPX index
sample path over the same time period. The implied and expected volatility series are sampled weekly from January
8, 1997 to October 29, 2014, 930 weekly observations for each series.

K/S 0.8 0.9 1.0 1.1 1.2 0.8 0.9 1.0 1.1 1.2

Maturity A. Average option implied volatility B. Average option expected volatility

1 34.88 26.81 19.19 15.47 15.31 20.48 20.08 17.56 17.57 18.56
3 30.02 24.79 19.87 16.25 14.92 20.04 19.35 17.59 15.98 16.20
6 27.85 24.00 20.37 17.35 15.55 20.09 19.08 17.72 16.41 15.59
12 26.34 23.55 20.90 18.54 16.74 19.62 18.92 18.20 17.28 16.35
24 25.56 23.54 21.63 19.88 18.37 18.92 18.37 17.96 17.61 17.21
36 25.49 23.84 22.29 20.85 19.56 18.85 18.19 17.75 17.45 17.27
48 25.66 24.24 22.92 21.69 20.56 18.68 17.88 17.43 17.22 17.12
60 25.94 24.69 23.52 22.44 21.43 18.42 17.73 17.08 16.82 16.71

C. Average volatility risk premium D. Annualized information ratio

1 −14.40 −6.73 −1.63 2.10 3.25 −1.47 −2.57 −1.93 2.04 3.04
3 −9.98 −5.44 −2.28 −0.27 1.28 −1.62 −1.88 −1.52 −0.55 0.36
6 −7.75 −4.91 −2.65 −0.93 0.03 −1.22 −1.36 −1.14 −0.66 −0.45
12 −6.72 −4.63 −2.70 −1.26 −0.39 −0.97 −0.90 −0.73 −0.53 −0.31
24 −6.64 −5.18 −3.67 −2.27 −1.16 −0.76 −0.67 −0.54 −0.40 −0.28
36 −6.64 −5.65 −4.54 −3.41 −2.29 −0.59 −0.52 −0.45 −0.36 −0.28
48 −6.98 −6.36 −5.49 −4.47 −3.44 −0.49 −0.46 −0.40 −0.35 −0.29
60 −7.52 −6.97 −6.45 −5.62 −4.72 −0.45 −0.44 −0.40 −0.36 −0.31

4.2. The time-series variation of implied and expected stantaneous variance rate specification in the affine option
volatilities pricing literature, e.g., Heston (1993), which implies that
the variance of the instantaneous volatility changes is in-
Fig. 1 compares the time-series variation of option im- dependent of the volatility level. To investigate how the
plied and expected volatilities. Panels A and C plot the variance of the volatility changes depends on the volatility
time series of the at-the-money implied and expected level, we estimate a constant elasticity of variance (CEV)
volatilities whereas Panels B and D plot the 90% strike mi- specification on the implied volatility time series,
nus 110% strike volatility differences as a measure of skew-  β
1
ness on the return distribution. The three lines in each Vart ( It (k, τ ) ) = C (τ ) It2 (k, τ ) , (29)
panel are for three selected maturities at 1 month (solid t
line), 6 months (dashed line), and 24 months (dash-dotted where the free power coefficient β would be equal to one
line), respectively. under our proportional volatility specification or a log nor-
The time-series variations of at-the-money implied and mal stochastic volatility model (e.g., Hull and White, 1987),
expected volatilities show similar patterns. The volatility but equal to zero under the square-root variance specifica-
series show spikes during the 1998 Asian crises and the tion (e.g. Heston, 1993).
ensuing hedge fund crisis in 1999, during the mild reces- To estimate this power coefficient β , we first estimate
sion in early 20 0 0, and most prominently during the fi- an exponentially weighted variance (EVI) on each implied
nancial market meltdown around 2008. The spike in 2012 volatility series,
corresponds to the European sovereign debt crisis. 
The 90–110% option implied volatility difference is uni- EV It = φ EV It−1 + (1 − φ ) ( It )2 / t , (30)
formly positive over our whole sample period and across where t = 1/52 denotes the weekly sampling frequency,
all option maturities, suggesting that the option implied It denotes the weekly changes on an implied volatility
SPX return distribution is persistently negatively skewed. series, and we set the exponential smoothing coefficient
By contrast, the corresponding option expected volatility φ = 0.97, corresponding to a half life of about half a year.
shows much less negative skewness, and the estimates can Then, we perform the following linear regression to esti-
turn positive from time to time. The large difference in mate the power coefficient,
skewness highlights the market’s extreme aversion to stock
market crashes (Wu, 2006). ln EV It (k, τ ) = ln C (τ ) + β ln It2 (k, τ ) + et . (31)

Table 2 reports the slope estimates for each implied

4.3. Volatility of volatility dependence structure volatility time series. For all 40 series, the slope estimates
are far away from the square-root hypothesis of β = 0, but
Our dynamics specification assumes that the variance of close to our proportional specification of β = 1, suggesting
the volatility changes is proportional to the variance level. that the proportional volatility dynamics enjoys better em-
This specification forms a contrast with the square-root in- pirical support than the square-root specification.
 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20 11

A. At-the-money implied volatility B. 90% – 110% implied volatility skew

70 20

60
15

Option implied volatility skew, %

Option implied volatility, %

50
10

40
5
30

0
20

−5
10

0 −10
97 98 99 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 97 98 99 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15

C. At-the-money expected volatility D. 90% – 110% expected volatility skew

70 20

60
15
Option expected volatility skew, %
Option expected volatility, %

50
10

40
5
30

0
20

−5
10

0 −10
97 98 99 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 97 98 99 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15

Fig. 1. Time-variation in option implied and expected volatilities and volatility skews. Lines plot the time series of option implied (Panels A and B) and
expected volatilities (Panels C and D). Panels A and C are for at-the-money options whereas Panels B and D are for the volatility differences for 90% strike
– 110% strike risk reversals. The three lines in each panel are for three different time to maturities: 1 month (solid lines), 6 months (dashed lines), and the
24-months (dash-dotted lines).

5. Extracting economic states from implied and Among the eight covariates, four (wt , ηt , vt , vtP ) are con-
expected volatility surfaces strained to be strictly positive, two (ρ t and ρtP ) are con-
strained to be between (−1, 1 ). In defining the state vec-
Under the proportional volatility dynamics specifica- tor Xt , we transform these covariates so that they can take
tion, the time-t shape of the option implied volatility sur- values on the whole real line:
face is governed by the time-t values of five covariates 1 + ρ 
Xt ≡ mt , mtP , ln(wt ), ln(ηt ), ln(vt ), ln(vtP ), ln
t
(vt , mt , ρ t , wt , ηt ), and we allow three additional covari- ,
1 − ρt
ates (vtP , mtP , ρtP ) to capture the difference in the shape of  
the expected volatility surface. One particular feature of 1 + ρtP
ln . (32)
the model is that the shapes of the two volatility surfaces 1 − ρtP
only depend on the levels of these state variables, but do
not depend on the particular state dynamics specification. With the transformation, we assume that the state vector
Therefore, the emphasis of the empirical analysis involves propagates as a random walk,
the extraction of the states from the two surfaces, without 
knowing the state dynamics. Based on this unique feature,
Xt = Xt−1 + x t ε , (33)
we cast the model into a state-space form, where we treat where the standardized error vector ε t is normally dis-
the covariates as the hidden states and treat the observed tributed with zero mean and unit variance. We further
option implied and expected volatility estimates as mea- assume that the covariance matrix is a diagonal matrix
surements with errors. with distinct diagonal values so that the states can have
12 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20

Table 2
Constant elasticity of variance dependence of implied volatility dynamics.
For each implied volatility series, we first estimate an exponentially weighted variance series on the weekly implied volatility
changes, and then regress the logarithm of this variance estimator against the logarithm of the implied variance level. The
regression slope captures the power dependence of the implied volatility variance on the implied variance level, β . Entries
report the regression estimates (and standard errors in parentheses) for this power coefficient for each implied volatility series.

Maturity\(K/S) 0.8 0.9 1.0 1.1 1.2

1 1.12 (0.03) 1.11 (0.03) 0.79 (0.02) 1.20 (0.03) 1.55 (0.04)
3 1.25 (0.03) 1.13 (0.03) 0.90 (0.02) 1.03 (0.03) 1.37 (0.04)
6 1.34 (0.03) 1.21 (0.03) 1.02 (0.03) 1.01 (0.03) 1.29 (0.03)
12 1.42 (0.03) 1.28 (0.03) 1.12 (0.03) 1.04 (0.03) 1.17 (0.03)
24 1.49 (0.03) 1.37 (0.03) 1.23 (0.03) 1.14 (0.03) 1.17 (0.03)
36 1.55 (0.03) 1.44 (0.03) 1.32 (0.03) 1.22 (0.03) 1.20 (0.03)
48 1.58 (0.04) 1.48 (0.04) 1.37 (0.04) 1.27 (0.04) 1.21 (0.04)
60 1.58 (0.04) 1.49 (0.04) 1.40 (0.04) 1.30 (0.04) 1.22 (0.04)

different degrees of variation but the movements are auxiliary parameters, and accordingly the optimal state up-
independent of each other. dating speed, by minimizing the sum of squared forecast-
In reality, the eight covariates represent eight different ing errors in a quasi-maximum likelihood setting.
stochastic processes, which can follow much more complex
dynamics than assumed in the state propagation Eq. (33). 6. Pricing performance and state dynamics analysis
However, since their dynamics do not enter the pricing of
the volatility surfaces, we leave them unspecified and use We first examine the pricing performance of the model
the simple random walk assumption to dictate the state on the two volatility surfaces and then analyze the dy-
propagation equation. namic behaviors of the extracted states and their implica-
We define the measurement equations on the logarithm tions.
of the option implied and expected volatility estimates, as-
suming additive, normally distributed errors, 6.1. Pricing performance

yt = h(Xt ) + y et , Panel A of Table 3 reports the average pricing error
h(Xt ) = {ln(I (Xt ; k j , τ j ), ln V (Xt ; k j , τ j )}40
j=1 , (34) on each volatility series. The pricing errors are defined as
the difference between the observed volatility series and
where yt ∈ R80 denotes the logarithm of the 40 implied the model-generated values, in volatility percentage points.
volatility quotes and the 40 corresponding OEV estimates For the implied volatility surface, the most obvious aver-
on date t, and h(Xt ) denotes the logarithm of the model age bias occurs at 1-month maturity, where the observed
value of the implied and expected volatility as a function implied volatilities are on average higher than the corre-
of the states Xt , which can be solved from Eqs. (10) and sponding model values for far out-of-the-money options,
(25) in Propositions 3 and 4, respectively. By defining the but lower for at-the-money options. In essence, the model
measurement equations on the logarithms of the volatili- fails to fully capture the smile shape at short maturities.
ties with additive, normally distributed errors, we guaran- This deficiency comes mainly from the purely continu-
tee the positivity of the volatilities. We assume that the ous price movement assumption. As shown in Carr and
additive pricing errors are independent and identically dis- Wu (2003), continuous and discontinuous price dynamics
tributed (iid) with error variance σI2 for the 40 implied generate very distinct behaviors for short-term out-of-the-
variance quotes and with error variance σV2 for the 40 OEV money options. The data suggest that a jump component is
series. needed to capture the short-term implied volatility smile.
When the state-space model is Gaussian linear, the The average biases at longer maturities are less severe. At
Kalman (1960) filter provides efficient forecasts and up- option maturities 6 months and longer, the model gener-
dates on the mean and covariance of the state and obser- ates more negative skewness along the strike dimension
vations. Our state-propagation equations are constructed than observed from the data. This bias is in part induced
to be Gaussian and linear, but the measurement functions by the model’s difficulty (and over compensation) in fitting
h(Xt ) are not linear in the state vector. We use the un- the negative skewness at short maturities. The average bi-
scented Kalman filter (Wan, 2001) to handle the nonlin- ases on the expected volatility surface show less obvious
earity. structures.
The setup introduces ten auxiliary parameters that de- Panel B reports the explained variation on each series,
fine the covariance matrices of the state-propagation errors defined as one minus the variance ratio of the model’s
and the measurement errors. The relative magnitude of the pricing error to the original volatility series. The measure
state-propagation error variance versus the measurement is analogous to the R2 measure for a regression and cap-
error variance controls the speed with which we update tures the proportion of variation explained by the model.
the states based on new observations. Intuitively, if the On average, the model explains 98% of the variation of
states vary a lot (large x ) and the observations are accu- the implied volatility series, and 80% of the variation
rate (small y ), one would want to update the states faster of the expected volatility series. The lower explanatory
to better match the new observations. We choose these power on the expected volatility series is expected as
 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20 13

Table 3
Model pricing performance on SPX option implied and expected volatilities.
Entries in Panel A report the average pricing error of the model on each volatility series. The pricing error is defined
as the difference between the observed volatility series and their corresponding model values, in volatility percentage
points. Entries in Panel B report the model’s explained variation, defined as one minus the ratio of the pricing error
variance to the variance of the regional volatility series. For each measure, the last row reports the grand average of
the statistic for the 40 implied and 40 expected volatility series, respectively.

Implied volatility surface Expected volatility surface

Maturity\(K/S) 0.8 0.9 1.0 1.1 1.2 0.8 0.9 1.0 1.1 1.2

Panel A: Average pricing error

1 5.80 2.62 −0.97 −1.56 0.44 −1.92 0.83 −0.01 0.23 0.36
3 1.15 0.64 −0.40 −0.98 −0.18 −1.28 0.96 0.76 −0.64 −1.21
6 −0.76 −0.16 −0.12 −0.23 0.06 −0.28 1.37 1.42 0.33 −1.20
12 −1.88 −0.70 −0.04 0.25 0.44 0.12 1.70 2.19 1.49 0.02
24 −2.03 −0.89 −0.15 0.28 0.51 −0.16 0.98 1.50 1.39 0.69
36 −1.46 −0.62 −0.06 0.28 0.45 −0.14 0.44 0.71 0.65 0.34
48 −0.55 −0.02 0.31 0.50 0.57 0.02 0.12 0.19 0.20 0.08
60 0.53 0.79 0.91 0.95 0.91 0.33 0.27 0.01 −0.05 −0.12

Average 0.11 0.32

Panel B: Explained variation

1 0.90 0.95 0.96 0.97 0.90 0.77 0.71 0.73 0.69 0.78
3 0.97 0.99 0.99 0.99 0.99 0.88 0.90 0.91 0.88 0.80
6 0.97 0.99 0.99 0.99 0.99 0.87 0.90 0.88 0.80 0.71
12 0.98 0.99 0.99 0.98 0.98 0.75 0.77 0.73 0.67 0.56
24 0.97 0.99 0.99 0.99 0.98 0.80 0.81 0.78 0.76 0.68
36 0.96 0.99 0.99 0.99 0.98 0.85 0.89 0.88 0.84 0.77
48 0.95 0.98 0.99 0.99 0.99 0.86 0.89 0.91 0.88 0.80
60 0.93 0.96 0.98 0.98 0.97 0.74 0.78 0.78 0.76 0.72

Average 0.98 0.80

these series are noisy estimates of the true expected thus change our valuation. In principle, the optimal updat-
value based on historical movements. Across different ing speed can change with market conditions. For example,
strikes and maturities, the model fits the at-the-money if the observations are becoming more accurate over time
implied volatility better than out-of-the-money implied and/or the market starts to show larger variations, the
volatilities, and fits the moderate-maturity volatility series optimal updating speed should become faster to put more
better than series at very short or very long maturities. weight on the most recent observation. To gauge how sen-
For reasons discussed above, the lowest explanatory sitive the model performance is to these auxiliary param-
power on implied volatilities are at the very short eter estimates, we perform an out-of-sample analysis: we
maturities. only use the first 3 years of data (1997–1999) to perform
In pricing the volatility surfaces, our model only de- the maximum likelihood estimation and use the estimated
pends on the current levels of the state variables, but does parameters to filter the states over the whole sample
not depend on any fixed model parameters. The absence period. We measure the correlation between the two sets
of fixed model parameters greatly simplifies model estima- of states based on the two sets of parameter estimates to
tion and removes potential consistency issues encountered determine how the auxiliary parameter estimates alter the
in model recalibration: a model with recalibrated model values and movements of the extracted states. The correla-
parameters represents essentially a different model and tion estimates are the highest at 99.6% for the two variance
thus generates different pricing and hedging implications rates (vt , vtP ), around 98% for mtP and ρtP , around 93% for mt
from previous calibrations. Such consistency issues do not and wt , and 86% for ρ t . The lowest correlation is between
show up in our model as the pricing relation contains no the two sets of ηt estimates at 80%. The high correlations
fixed parameters. In our state-space approach to extract between the two sets of the extracted states, especially
the state variables, we introduce ten auxiliary parameters for the variance rates, suggest that the filtered states are
to define the state-propagation error variance and the not very sensitive to small variations in the auxiliary
measurement error variance. These variance estimates parameter estimates. The pricing performance is also sim-
control the updating speed of the states based on new ilar under the two sets of auxiliary parameter estimates.
observations, and we use maximum likelihood estimation The average explained variation on the implied volatility
to determine the magnitudes of these parameters and surface is around 98% based on both sets of parameters.
accordingly the optimal updating speed. Altering the state- The explained variation on the expected volatility surface
propagation equation specification and/or the variance es- experiences some deterioration from 80% based on the
timates does not induce consistency issues for the pricing full-sample estimates to 67% based on the 3-year sample
relation, but can nevertheless change the state updates and estimates.
14 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20

70 0.4

60 0.2

Return−volatility correlation
Instantaneous volatility, %

50 0

40 −0.2

30 −0.4

20 −0.6

10 −0.8

0 −1
97 98 99 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 97 98 99 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15

Fig. 2. The time-series variation of instantaneous volatilities. The solid Fig. 4. The time-series variation of the return-volatility correlation. The
√
line represents the time series of the instantaneous volatility ( vt ) ex- solid line represents the time series of the instantaneous correlation be-
tracted from the option implied volatility surface. The dashed line rep- tween the index return and its volatility, extracted from the option im-
resents the corresponding instantaneous volatility extracted from the ex- plied volatility surface. The dashed line represents the corresponding cor-
pected volatility surface. relation extracted from the expected volatility surface.

0.5
line) and the statistical measure (mtP , dashed line). The
0 risk-neutral drift process dictates the term structure shape
of the at-the-money implied volatility, whereas the statis-
Instantaneous volatility drift

−0.5 tical drift process governs the term structure shape of the
expected volatility. The solid line stays above zero most of
−1 the time, except during the 2002 recession and the 2008
financial crisis. The on average positive risk-neutral drift
−1.5 suggests that the at-the-money implied volatility term
structure is upward sloping most of the time. By contrast,
−2 the dashed line stays negative most of the time, suggest-
ing that the expected volatility computed from the histor-
−2.5 ical sample paths has a downward sloping term structure
97 98 99 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15
most of the time. The term structure difference reflects the
Fig. 3. The time-series variation of volatility drift processes. The solid line volatility risk premium. The difference is particularly large
represents the time series of the risk-neutral volatility drift process (mt ), around the two financial crises (1998 and 2008) and dur-
extracted from the option implied volatility surface. The dashed line rep- ing the 2003 recession.
resents the corresponding statistical volatility drift process (mtP ) extracted
from the expected volatility surface.
6.3. Stochastic variation of the return-volatility correlation
and the volatility skew
6.2. The time variation of short and long-term implied and
expected volatilities Fig. 4 plots the time series of the instantaneous correla-
tion between the SPX index return and its volatility, again
Fig. 2 plots the time series of the instantaneous volatil- with the solid line extracted from the implied volatility
√
ity ( vt ), with the solid line extracted from the options surface and the dashed line from the expected volatil-
implied volatility surface and the dashed line from the ex- ity surface. The solid line stays strongly negative over the
pected volatility surface. The time-series variation of the whole sample period, with a maximum of −0.47 and a
two instantaneous volatility series follows closely the time- minimum close to −0.98. These highly negative correlation
series variation of at-the-money implied and expected estimates reflect the persistently negative skew observed
volatilities plotted in Fig. 1. Due to the backward look- from the implied volatility surface. By contrast, the dashed
ing nature of the expected volatility estimation, the in- line varies much closer to zero and can switch signs, sug-
stantaneous volatility series extracted from the expected gesting that the expected volatility surface is not always
volatility surface seems to lag behind the solid line ex- negatively skewed.
tracted from the implied volatility surface. Furthermore, Interestingly, the two financial crises during our sam-
during non-eventful time periods such as the bull mar- ple period (the 1998 Asian crises and the 2008 financial
ket run from 2004 to 2007 and the most recent run since meltdown) are both preceded by a divergence between
2012, the solid line extracted from the implied volatility the two correlation estimates, with the dashed line go-
surface stays above the dashed line extracted from the ex- ing above zero and the solid line reaching its most neg-
pected volatility surface, but the two lines tend to cross in ative level. Before the financial crisis, the options market
the aftermath of a volatility spike. becomes increasingly worried as shown by the extremely
Fig. 3 plots the time series of the instantaneous drift negative implied volatility skew. At the same time, the
processes under both the risk-neutral measure (mt , solid index return dynamics start to show abnormal behaviors
 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20 15

A. Instantaneous volvol coefficient (wt ) B. Maturity decay coefficient (ηt )

0.9 0.6

0.55
0.8
0.5

Maturity decay coefficient

0.7 0.45
Volvol coefficient

0.6 0.4

0.35
0.5
0.3

0.4 0.25

0.2
0.3
0.15
0.2 0.1
97 98 99 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 97 98 99 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15

Fig. 5. The time series of the instantaneous volvol process wt in Panel A and the maturity decay process ηt in Panel B.

as the return volatility starts increasing with rising index Jacobs (2009) on equity options. What is the most interest-
level, whereas at normal times return volatility tends to ing and challenging is to come up with risk and risk pref-
decline with rising index level. These behaviors, combined erence specifications that can accommodate the risk pre-
with the volatility spikes, seem to precede the upcoming mium variations as shown in the different term structure
financial crisis. By contrast, during the mild recession of and skew variations in Figs. 3 and 4 extracted from the two
2003, although the volatility level also spiked up, the op- volatility surfaces.
tion implied volatility skew was not particularly negative,
and the return-volatility correlation extracted from the ex- 6.5. Risk premiums and excess return predictions
pected volatility surface stayed negative. Thus, for future
dynamic model designs, it is important to build different Under the pricing kernel assumption in (26), we can
mechanisms for different types of volatility spikes. identify the market pricing of the variance risk γ t from
the difference between the statistical and risk-neutral drift
6.4. The time-series variation of volatility of volatilities processes (mt and mtP ), as shown in Eq. (27), γt = (mtP −
√
mt )/(wt vt ). The market pricing of the variance risk con-
Fig. 5 plots the extracted time series of the volvol pro- tributes to the instantaneous return risk premium through
cess in Panel A. The volvol estimates tend to be high the return-variance correlation by ρ t as shown in Eq. (28).
when the volatility levels are high. A high volvol coeffi- We label this component of the return risk premium as
cient increases the convexity of the volatility smile along RRP. In this section, we analyze whether this return risk
the strike dimension. premium component has any actual predictive power of
Panel B of Fig. 5 plots the time series of the matu- future excess returns on the SPX index. For comparison, we
rity decay coefficient (ηt ), which lowers the variation for consider two benchmarks. One is the VIX index squared,
long-term implied volatility series. The extracted series are which approximates the 1-month variance swap rate of
stable except during the 2002 recession, when the esti- the S&P 500 index.7 The VIX index is regarded as a fear
mates become much higher. This recession period seems gauge in the industry and has the potential to capture
to be unique in its behaviors, when the short-term volatil- not only the risk level variation, but also risk preference
ity is high, the term structure for both implied and ex- changes over time. The second benchmark is the differ-
pected volatilities are downward sloping, and the short- ence between VIX squared and the 1-month realized vari-
term volatility varies much more than long-term volatil- ance (VIX2 − RV ), which is often labeled as the variance
ities. While the short-term volatility is high during both risk premium (VRP) and has been used to predict future
financial crises and during this recession, the long-dated stock returns by, among others, Bollerslev, Tauchen, and
implied volatilities do not go up as much during the reces- Zhou (2009).
sion, suggesting that investors are much less worried about To obtain an empirical estimate of the return risk pre-
this recession than about the financial crisis. mium, we regress stock index excess returns on each of the
The time-series variations of the different state vari- three predictors,
ables depicted in Figs. 2–5 provide guidance for future ERt, j = ai + bi xti + et,
i
j, (35)
structural model designs. The variation of the instan-
taneous variance rate can be accommodated by most where xti
denotes the time-t value of the ith predictor (VIX,
stochastic volatility models. The return-variance correla- VRP, or RRP) and ERt, j denotes the annualized future index
tion (hence volatility skew) variation can be accommo-
dated by a two-factor volatility structure as in Carr and Wu 7
See Carr and Wu (2006) for a detailed description of this index and
(2007) on currency options and Christoffersen, Heston, and its behaviors. We thank the referee for this suggestion.
16 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20

excess return from time t to j days ahead,8 Table 4

Out-of-sample return forecasting R2 .
j Entries report the out-of-sample return forecasting R2 , defined as one
365
ERt, j = rt+s − Rtf , (36) minus the ratio of sum squared forecasting error of each method to
j that of the historical average benchmark.
s=1

f Horizon 1 3 6 9 12
with rt denoting the daily return at time t, and Rt denotes
the riskfree rate at time t, which we proxy with the US VIX regression 0.001 −0.017 0.008 0.007 0.010
LIBOR rate of the corresponding horizon. The time series VRP regression 0.010 0.040 0.028 0.004 0.002
RRP regression −0.009 −0.015 0.054 0.071 0.086
of the SPX, SPY, VIX, and the LIBOR rates are obtained from
RRP direct 0.017 0.061 0.159 0.223 0.260
Bloomberg. The 30-day realized variance is computed from
the historical return data on the SPX index. In computing
the return risk premium (RRP) according to (28), we use
auxiliary parameters that control the state updates. For the
the variance rate and return-variance correlation extracted
out-of-sample exercise, we use the parameters estimated
from the expected volatility surface.
from the first 3 years of the sample without further up-
We perform an out-of-sample exercise based on the
dating these estimates. As we have shown earlier, the ex-
predictive regression in (35). Starting from January 20 0 0
tracted states are not particularly sensitive to the small
(3 years from the starting date of the data sample), at each
variations in the auxiliary parameters. The extracted states
date t, we estimate each regression using data up to that
are similar whether we re-estimate the model or not. In
point and make predictions for future excess returns from
addition to performing forecasting regressions, given the
that point forward. To reduce the impact of data outliers
structural meaning of RRP being the return risk premium,
on the forecasting results, we follow Campbell and Thomp-
we can also directly use the RRP estimates as the forecast
son (2008) and constrain all annualized excess return fore-
for future excess returns by setting the regression intercept
casts to be within (0, 20%). We compute the out-of-sample
to zero and the slope to one.
forecasting error as the difference between the future re-
Table 4 reports the out-of-sample R2 for the four sets
alized excess return and the forecasted excess return. As
of forecasts. The VIX itself can hardly outperform the his-
in Welch and Goyal (2008), the forecasting performance of
torical average as the R2 estimates are close to zero at
each measure is compared with the historical average of
all forecasting horizons. The variance risk premium regres-
the future excess return up to that point t,
sion can outperform the historical average, with the best
t
1 performance coming at the quarterly forecasting horizon.
ERt, j = ERs− j, j . (37) The performance starts to deteriorate at longer forecasting
t−j
s= j+1
horizons, potentially because the variance risk premium is
The overall out-of-sample forecasting performance of each constructed using only short-term option contracts. The re-
predictor is measured by the sum squared forecasting error turn risk premium regression underperforms the historical
(SSFE) over N out-of-sample observations, average at short forecasting horizons, but outperforms in-
creasingly more as the forecasting horizon increases. The
N  2
i short-horizon underperformance is likely related to the
SSFEi, j,N = E Rt, j − ERt, j , (38)
model’s difficulty in capturing the short-term behavior of
t=1
the implied volatility surface. Its long-horizon outperfor-
i
 denotes the out-of-sample forecast from pre-
where ER t, j
mance, on the other hand, shows the benefit of extracting
dictor i on excess return ERt, j . Using the SSFE on the his- information from the two volatility surfaces.
torical average as the benchmark, More striking is the strong performance of directly ap-
plying the RRP as the excess return forecast. Since the RRP
N  2 represents only part of the return risk premium, without
SSFE0, j,N = E Rt, j − E Rt, j , (39)
accounting for the risk premium on the independent re-
t=1
turn risk, the RRP estimate can be regarded as a conserva-
we measure the relative performance of each predictor tive estimate of the return risk premium. The predictive re-
via an out-of-sample R2 measure as in Rapach and Zhou gression can be used to adjust the scale, but it also brings
(2013), in estimation error, especially out of sample. By discarding
R2i, j = 1 − SSFEi, j,N /SSFE0, j,N . (40) the predictive regression and directly applying the struc-
tural implication of the model, one can avoid the noise in-
A positive R2 estimate indicates that the predictor outper- troduced by predictive regressions and generate much su-
forms the historical average benchmark. perior out-of-sample forecasting performance.9
To compute the return risk premium (RRP) from the As in Rapach and Zhou (2013), Fig. 6 plots the cumula-
VGVV model, we need to estimate the model to obtain the tive squared forecasting error difference,
CFEDi, j,n = SSFE0, j,n − SSFEi, j,n , n = 1, 2, · · · , N. (41)
8
Since the SPX index level does not adjust for dividend payments, log
index level difference does not fully capture the returns from investing
9
in the S&P 500 stocks. To obtain a return series that properly adjusts for Hua and Wu (2015) show that the out-of-sample forecasting perfor-
dividend payments, we use the spider (SPY) adjusted-price time series mance on both interest rates and inflation rates can be significantly im-
instead for the return calculation. The resulting difference from using SPX proved by estimating the predictive relation without running predictive
log index level difference is very small. regressions.
 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20 17

A. VIX regression B. VRP regression

3 4
Cumulative squared forecasting error difference

Cumulative squared forecasting error difference

2.5 3.5

2 3

1.5 2.5

1 2

0.5 1.5

0 1

−0.5 0.5

−1 0

−1.5 −0.5

−2 −1
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15

C. RRP regression D. RRP direct

4 12
Cumulative squared forecasting error difference

Cumulative squared forecasting error difference

3 10

2 8

1 6

0 4

−1 2

−2 0

−3 −2
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15

Fig. 6. Cumulative squared forecasting error difference. The three lines in each panel are the cumulative squared forecasting error difference between each
method and the historical average benchmark, with each line representing one forecasting horizon: 3 months (solid line), 6 months (dashed line), and 12
months (dash-dotted line). The four panels are for four different forecasting methods.

The three lines in each panel denote three selected fore- longer horizons. In particular, the outperformance mainly
casting horizons at 3 months (solid line), 6 months comes from the early 20 0 0 period and during the 2008
(dashed line), and 12 months (dash-dotted line). Panel A financial crisis. The performance shows deterioration dur-
plots the cumulative out-of-sample performance of the VIX ing the 20 03–20 08 stretch and after 2010. However, when
squared regression. The performance is worse than the we discard the regression and directly apply the RRP as
historical average at the 3-month forecasting horizon and the excess return forecast, the cumulative performance in
only becomes slightly better at longer horizons. The dras- Panel D becomes much more uniform over different sam-
tic deterioration in 2009 is caused by the extreme spike ple periods, especially at long forecasting horizons.
in the implied volatility. Panel B shows that the VRP gen- The out-of-sample predictive power of RRP highlights
erates reasonably good out-of-sample forecasting perfor- the information content of the two volatility surfaces
mance at the quarterly forecasting horizon. The forecasting in extracting risk dynamics and risk premiums. Several
performance mainly comes from the early 20 0 0s and the studies, e.g., Bollerslev, Tauchen, and Zhou (2009), Xing,
2008 financial crisis, but otherwise has little power during Zhang, and Zhao (2010), Cremers and Weinbaum (2010),
the long stretches between 2003 and 2008 and after 2010. and Bakshi, Panayotov, and Skoulakis (2011) have found
At longer forecasting horizons, the VRP prediction deterio- equity options to be informative of future stock returns.
rates and can no longer outperform the historical average. More recently in a seminal paper, Ross (2015) shows that
Panel C plots the cumulative performance of the RRP under certain assumptions, one can identify both the
regression. The regression is not stable at short forecasting risk-neutral and statistical dynamics, as well as the pricing
horizons, but generates more consistent performance at kernel that links the two, using only information in the
18 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20

option implied volatility surface alone. Several researchers, in the future does not affect the current volatility shape.
e.g., Borovicka, Hansen, Scheinkman (2014), Hansen and This unique feature allows us to specify a model that has
Scheinkman (2014), Walden (2014), Audrino, Huitema, many state variables but with no fixed model parameters.
and Ludwig (2014), and Qin and Linetsky (2014); 2015) The high dimensional state space allows the model to fit
explore the implications of the underlying assumptions, the observed volatility surface well without extra fudging,
potential extensions, and empirical performance. How to whereas the absence of fixed model parameters drastically
integrate the different perspectives to balance the need simplifies the model estimation process. This feature also
for structural assumptions and data constitutes a deeply makes the model a perfect complement to traditional fully
interesting research direction. parametric option pricing models. In particular, volatility
surface valuations from a chosen parametric model can be
7. Concluding remarks directly used as the starting point by assuming that mar-
ket volatility observations converge to the model valuation
Despite the fact that the BMS model assumptions are via an error-correction specification.
apparently violated, both practitioners and academics have Our new theoretical framework opens ample ground
been accustomed to use the BMS implied volatility surface for future research. First, we use a simple proportional
to represent the information in option contracts. Quoting volatility dynamics for illustration. For future research, one
a positive implied volatility for an option contract directly can explore many different specifications, many of which
excludes arbitrage between this option and the underlying can lead to extremely simple analytical solutions for the
security, adding further attraction to the implied volatility volatility surface. Second, the concept of option-specific re-
quoting convention. Furthermore, delta hedging in practice alized and expected volatility opens a whole new area of
is mostly based on the BMS delta at the implied volatility empirical research on option-specific volatility forecasting.
level. Despite much research, the literature has yet to pro- Third, our current framework assumes diffusion return dy-
pose an alternative delta ratio that outperforms the BMS namics and a one-factor diffusion volatility surface dynam-
delta in all practical situations. ics; future research can investigate how to accommodate
Given this heavy reliance on the BMS implied volatility discontinuous price and volatility movements and multiple
surface, it would be ideal if one can directly model the im- volatility risk factors.
plied volatility dynamics and derive direct implications on
the shape of the implied volatility surface. In this paper, Appendix A. Proofs
we propose a new modeling framework that does just that.
Given a one-factor pure diffusion dynamics on the implied Proof of Proposition 1. First, we form a portfolio between
volatility surface, we transform the dynamic no-arbitrage any put option at (K, T) and the basis call option at (K0 ,
constraint between the underlying stock, a basis option, T0 ) to neutralize the exposure on the volatility risk dZt :
and any other option contract into a simple algebraic con-
straint on the shape of the current implied volatility sur- Bσ (St , It (K, T ), t )ωt (K, T )
face. In particular, under a proportional volatility dynamics − Ntc Bσ (St , It (K0 , T0 ), t )ωt (K0 , T0 ) = 0. (42)
specification, the whole shape of the implied volatility sur- The two-option portfolio with no dZt exposure will in gen-
face becomes the solution to a simple quadratic equation. eral be exposed to dWt . As a result, a three-asset portfolio
As a result, the numerical burden for option pricing and with NtS shares of the underlying stock is determined by
model estimation is dramatically reduced compared to the requiring delta neutrality:
standard option pricing literature.
Corresponding to implied volatility surface, we also BS (St , It (K, T ), t ) − Ntc (1 + BS (St , It (K0 , T0 ), t )) − NtS = 0.
propose a new concept that both realized and statistically (43)
expected volatilities estimated from price sample paths can
Since shares have no vega, this three-asset portfolio retains
vary with the strike and maturity of a reference option
zero exposure to dZt and by construction has zero exposure
contract. The idea is that although taking any option po-
to dWt .
sition with delta hedging exposes the investor to the vari-
By Ito’s lemma, each option in this portfolio has risk-
ance risk during the life of the option, the weighting on
neutral drift given by:
the sample path differs for different option contracts. As
such, one can estimate a realized and expected volatility 1 √ ω2
Bt + μt Bσ + vt S2 BSS + ρt ωt vt St BSσ + t Bσ σ . (44)
corresponding to the risk exposure of each option con- 2 t 2
tract. With this new concept, one can directly measure the No dynamic arbitrage and no rates imply that both option
volatility risk premium embedded in each option contract drifts must vanish, leading to the fundamental partial dif-
as the difference between this contract’s implied and ex- ferential equation (PDE):
pected volatility. Furthermore, the current shape of the ex- 1 √ 1
pected volatility surface is analogously governed by its fu- −Bt = μt Bσ + vt S2 BSS + ρt ωt vt St BSσ + ωt2 Bσ σ . (45)
2 t 2
ture statistical dynamics.
This fundamental PDE applies to any option, as long as we
A unique feature of our modeling framework is that
require that this option allow no dynamic arbitrage relative
by modeling the whole volatility surface, we only need
to the basis option at (K0 , T0 ), the stock, and cash. 
to know the current level of the drift and diffusion of
the volatility surface to determine the current shape of Proof of Proposition 2. The BMS value function B(St , It , t) is
the surface. How the drift and diffusion processes vary well known, so are its various partial derivatives:
 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20 19

σ2 (10). In this case, Eq. (49) only determines the shape of the
Bt = − S2 BSS , Bσ = σ (T − t )S2 BSS , expected volatility Vt (k, τ ), with It being replaced by Vt in
2 √ 
ln (K/S ) σ T −t  the equation. 
SBσ S = √ + T − t S2 BSS ,
σ T −t 2
 
ln (K/S )2 σ 2 (T − t ) References
Bσ σ = − (T − t )S2 BSS (46)
σ 2 (T − t ) 4
Arslan, M., Eid, G., Khoury, J.E., Roth, J., 2009. The gamma vanna volga
cost framework for constructing implied volatility curves. Unpub-
where dollar gamma S2 BSS is the common denominator of lished working paper. Deutsche Bank.
all the partial derivatives. Audrino, F., Huitema, R., Ludwig, M., 2014. An empirical analysis of the
Evaluate these partial derivatives at (S, σ , t ) = ross recovery theorem. Unpublished working paper. University of St.
Gallen and University of Zurich.
(St , It (K, T ), t ), substitute them into the fundamental Avellaneda, M., Zhu, Y., 1998. A risk-neutral stochastic volatility model.
PDE in (6), and divide both sides of the equation by the International Journal of Theoretical and Applied Finance 1, 289–310.
dollar gamma St2 BSS while noting that the dollar gamma Baele, L., Driessen, J., Londono, J.M., Spalt, O.G., 2014. Cumulative prospect
theory and the variance premium. Unpublished working paper.
is strictly positive at T > t, we transform the PDE into an
Tilburg University.
algebraic equation, Bakshi, G., Kapadia, N., 2003. Delta-hedged gains and the negative market
volatility risk premium. Review of Financial Studies 16, 527–566.
1 2 1 Bakshi, G., Kapadia, N., 2003. Volatility risk premium embedded in indi-
0= I (K, T ) − μt It (K, T )(T − t ) − vt
2 t 2 vidual equity options: some new insights. Journal of Derivatives 11,
 √ 
It (K, T ) T − t 
45–54.
√ ln (K/S )
−ρt ωt vt √ + T −t Bakshi, G., Panayotov, G., Skoulakis, G., 2011. Improving the predictabil-
It (K, T ) T − t 2 ity of real economic activity and asset returns with forward variances
  inferred from option portfolios. Journal of Financial Economics 100,
1 2 ln (K/S )2 I2 (K, T )(T − t ) 475–495.
− ω − t (T − t ). (47)
2 t It (K, T )(T − t )
2 4 Benaim, S., Friz, P., Lee, R., 2008. On Black-Scholes implied volatility at ex-
treme strikes. In: Cont, R. (Ed.), In: Frontiers in Quantitative Finance:
Volatility and Credit Risk Modeling. Wiley.
Re-define the implied volatility surface as a function of Black, F., Scholes, M., 1973. The pricing of options and corporate liabilities.
the relative strike k ≡ ln (K/S) and time to maturity τ ≡ T − Journal of Political Economy 81, 637–654.
t, and re-arrange, we obtain the algebraic representation in Bollerslev, T., Tauchen, G., Zhou, H., 2009. Expected stock returns and vari-
ance risk premia. Review of Financial Studies 22, 4463–4492.
(7). 
Borovicka, J., Hansen, L.P., Scheinkman, J.A., 2014. Misspecified recov-
ery. Unpublished working paper. New York University, University of
Proof of Proposition 3. Under the proportional dynamics Chicago, Columbia University.
specification, the drift and volvol of the implied volatility Campbell, J.Y., Thompson, S.B., 2008. Predicting excess stock returns out of
process are sample: can anything beat the historical average? Review of Financial
Studies 21, 1509–1531.
μt = e−ηt (T −t ) mt It (K, t ), ωt = e−ηt (T −t ) wt It (K, T ). (48) Carr, P., Madan, D., 2002. Towards a theory of volatility trading. In: Jar-
row, R. (Ed.), Volatility. Risk Publications, New York, pp. 417–427.
Substitute μt and ωt in (48) into the no dynamic arbitrage Carr, P., Sun, J., 2007. A new approach for option pricing under stochastic
volatility. Review of Derivatives Research 10, 87–150.
restriction (7), fix the relative strike k ≡ ln (K/S) and time
Carr, P., Wu, L., 2003. What type of process underlies options? a simple
to maturity τ ≡ T − t, and re-arrange terms, we can trans- robust test.. Journal of Finance 58, 2581–2610.
form the no-arbitrage algebraic constraint into a quadratic Carr, P., Wu, L., 2006. A tale of two indices. Journal of Derivatives 13, 13–
function of It2 (k, τ ) as shown in (10).  29.
Carr, P., Wu, L., 2007. Stochastic skew in currency options. Journal of Fi-
nancial Economics 86, 213–247.
Proof of Proposition 4. Starting with the statistical dynam-
Carr, P., Wu, L., 2009. Variance risk premiums. Review of Financial Studies
ics in (23) and (24), if we assume that the variance 22, 1311–1341.
risk is not priced, implied volatility would converge to Carr, P., Wu, L., 2010. Stock options and credit default swaps: a joint
the statistical expected volatility, It (k, τ ) = Vt (k, τ ), and framework for valuation and estimation. Journal of Financial Econo-
metrics 8, 409–449.
Eq. (23) would also become the risk-neutral dynam- Christoffersen, P.F., Heston, S.L., Jacobs, K., 2009. The shape and term
ics for the implied volatility. In this hypothetical case, structure of the index option smirk: why multifactor stochastic
Proposition 3 shows that the shape of the implied variance volatility models work so well. Management Science 55, 1914–1932.
Cremers, M., Weinbaum, D., 2010. Deviations from put-call parity and
surface as a function of relative strike and maturity is de- stock return predictability. Journal of Financial and Quantitative Anal-
termined by the following quadratic equation, ysis 45, 335–367.
Daglish, T., Hull, J., Suo, W., 2007. Volatility surfaces: theory, rules of
1 −2ηt τ 2 2 4
0= e wt τ It (k, τ ) thumb, and empirical evidence. Quantitative Finance 7, 507–524.
Drechsler, I., Yaron, A., 2011. What’s vol got to do with it. Review of Fi-
4
   nancial Studies 24, 1–45.
+ 1 + 2e−ηt τ mtP τ − e−ηt τ wt ρtP vtP τ It2 (k, τ ) Egloff, D., Leippold, M., Wu, L., 2010. The term structure of variance swap
rates and optimal variance swap investments. Journal of Financial and
   Quantitative Analysis 45, 1279–1310.
− vtP + 2e−ηt τ wt ρtP vtP k + e−2ηt τ wt2 k2 . (49) Engle, R., Figlewski, S., 2015. Modeling the dynamics of correlations
among implied volatilities. Review of Finance 19, 991–1018.
Furthermore, since It (k, τ ) = Vt (k, τ ) under zero risk pre- Engle, R.F., Granger, C.W., 1987. Cointegration and error correction: repre-
sentation, estimation and testing. Econometrica 55, 251–276.
mium, the same relation in (49) also determines the shape Fengler, M.R., 2005. Semiparametric Modeling of Implied Volatility.
of Vt (k, τ ). Springer-Verlag, Berlin.
With nonzero risk premium, implied variance differs Foresi, S., Wu, L., 2005. Crash-o-phobia: a domestic fear or a worldwide
concern? Journal of Derivatives 13, 8–21.
from the expected realized variance and the shape of the Gatheral, J., 2006. The Volatility Surface: A Practitioner’s Guide. John Wi-
implied volatility is determined by a different relation in ley & Sons, New Jersey.
20 P. Carr, L. Wu / Journal of Financial Economics 120 (2016) 1–20

Hafner, R., 2004. Stochastic Implied Volatility: A Factor-based Model. Polimenis, V., 2006. Skewness correction for asset pricing. Working paper.
Springer-Verlag, Berlin. University of California Riverside.
Hansen, L.P., Scheinkman, J.A., 2014. Stochastic compounding and uncer- Qin, L., Linetsky, V., 2014. Long term risk: a martingale approach. Unpub-
tain valuation. In: Weyl, E.G., Glaeser, E.L., Santos, T., Carr, P. (Eds.), lished working paper. Northwestern University.
In: Après le Déluge: Finance and the Common Good after the Crisis. Qin, L., Linetsky, V., 2015. Positive eigenfunctions of Markovian pricing op-
University of Chicago Press. erators: Hansen-Scheinkman factorization and ross recovery and long-
Heath, D., Jarrow, R., Morton, A., 1992. Bond pricing and the term struc- term pricing. Unpublished working paper. Northwestern University.
ture of interest rates: a new technology for contingent claims valua- Rapach, D.E., Zhou, G., 2013. Forecasting stock returns. In: Elliott, G., Tim-
tion. Econometrica 60, 77–105. mermann, A. (Eds.), Handbook of Economic Forecasting, vol. 2a. Else-
Heston, S.L., 1993. Closed-form solution for options with stochastic vier, North Holland, pp. 328–383.
volatility, with application to bond and currency options. Review of Ross, S., 2015. The recovery theorem. Journal of Finance 70, 615–648.
Financial Studies 6, 327–343. Schonbucher, P.J., 1999. A market model for stochastic implied volatil-
Hodges, H.M., 1996. Arbitrage bounds of the implied volatility strike and ity. Philosophical Transactions: Mathematical, Physical and Engineer-
term structures of European-style options. Journal of Derivatives 3, ing Sciences 357, 2071–2092.
23–32. Tehranchi, M.R., 2009. Asymptotics of implied volatility far from maturity.
Hua, J., Wu, L., 2015. Monetary policy rule as a bridge: predicting inflation Journal of Applied Probability 46, 629–650.
without predictive regressions. Unpublished working paper. Baruch Walden, J., 2014. Recovery with unbounded diffusion processes. Working
College. paper. University of California, Berkeley.
Hull, J., White, A., 1987. The pricing of options on assets with stochastic Wan, E.A., van der Merwe, R., 2001. The unscented Kalman filter. In:
volatilities. Journal of Finance 42, 281–300. Haykin, S. (Ed.), Kalman Filtering and Neural Networks. Wiley & Sons
Jiang, G., Tian, Y., 2005. Model-free implied volatility and its information Publishing, New York, pp. 221–280.
content. Review of Financial Studies 18, 1305–1342. Welch, I., Goyal, A., 2008. A comprehensive look at the empirical perfor-
Kalman, R.E., 1960. A new approach to linear filtering and prediction mance of equity premium prediction. Review of Financial Studies 21,
problems. Transactions of the ASME—Journal of Basic Engineering 82, 1455–1508.
35–45. Wu, L., 2006. Dampened power law: reconciling the tail behavior of fi-
Ledoit, O., Santa-Clara, P., 1998. Relative pricing of options with stochas- nancial asset returns. Journal of Business 79, 1445–1474.
tic volatility. Unpublished working paper. University of California, Los Xing, Y., Zhang, X., Zhao, R., 2010. What does the individual option volatil-
Angeles. ity smirk tell us about future equity returns? Journal of Financial and
Merton, R.C., 1973. Theory of rational option pricing. Bell Journal of Eco- Quantitative Analysis 45, 641–662.
nomics and Management Science 4, 141–183. Zhang, B.Y., Zhou, H., Zhu, H., 2009. Explaining credit default swap
Mueller, P., Vedolin, A., Yen, Y.-M., 2012. Bond variance risk premia. Work- spreads with the equity volatility and jump risks of individual firms.
ing paper. London School of Economics. Review of Financial Studies 22, 5099–5131.