A Non-Structural Investigation of VIX Risk Neutral Density

Andrea Barletta, Paolo Santucci de Magistris and Francesco

Violante

CREATES Research Paper 2017-15

Department of Economics and Business Economics Email: [email protected]

Aarhus University Tel: +45 8716 5515
Fuglesangs Allé 4
DK-8210 Aarhus V
Denmark
 A Non-Structural Investigation of VIX Risk Neutral
Density ∗
Andrea Barletta† Paolo Santucci de Magistris ‡ Francesco Violante §
March 31, 2017

Abstract
We propose a non-structural pricing method to derive the risk-neutral density (RND)
implied by options on the CBOE Volatility Index (VIX). The methodology is based on or-
thogonal polynomial expansions around a kernel density and yields the RND of the under-
lying asset without the need for a parametric specification. The classic family of Laguerre
expansions is extended to include the GIG and the generalized Weibull kernels, thus relax-
ing the conditions required on the tail decay rate of the RND to ensure convergence. We
show that the proposed methodology yields an accurate approximation of the RND in a
large variety of cases, also when the no-arbitrage and efficient option prices are contami-
nated by measurement errors. Our empirical investigation, based on a panel of traded VIX
options, reveals some stylized facts on the RND of VIX. We find that a common stochastic
factor drives the dynamic behavior of the risk neutral moments, the probabilities of volatil-
ity tail-events are priced in the options as jumps under the risk-neutral measure, and the
variance swap term structure depends on two factors, one accounting for the slope and one
for the mean-reverting behavior of the VIX.

Keywords: VIX options, orthogonal expansions, risk-neutral moments, volatility jumps,

variance swaps

JEL Classification: C01, C02, C58, G12, G13 .

∗ We thank Fulvio Corsi, Friedrich Hubalek, Elisa Nicolato and Viktor Todorov for useful comments and sug-

gestions. We also thank the participants of CREATES seminar series, the CFE 2015 conference in London, the Em-
pirical Finance workshop in Paris, the VCMF 2016 conference in Vienna and the HAX 2017 conference in Aarhus.
The research leading to these results received funding from the European Union Seventh Framework Programme
(FP7/2007-2013) under grant agreement No. 289032 (HPCFinance). Paolo Santucci de Magistris and Francesco
Violante acknowledge the research support of CREATES, funded by the Danish National Research Foundation
(DNRF78). Francesco Violante also acknowledges the research support of the Danish Council for Independent
Research, Social Sciences, Individual Research grant and Sapere Aude Research Talent grant.
† Department of Economics and Business Economics, Aarhus University, Denmark. [email protected]
‡ CREATES, Department of Economics and Business Economics, Aarhus University, Denmark. psan-
[email protected]
§ CREATES, Department of Economics and Business Economics, Aarhus University, Denmark. fvi-
[email protected]

1
1 Introduction
The Volatility Index (VIX) was introduced in 1993 by the Chicago Board Options Exchange
(CBOE) to measure the market expected volatility. In its first formulation, the VIX was de-
fined as an average of S&P 100 call and put implied volatilities. In response to the growing
interest in volatility trading, in 2004 the CBOE introduced the VIX futures alongside a revised
formulation of the VIX which was based on the replication of variance swap contracts written
on the broader S&P 500 (SPX) index. Specifically, in its current formulation, see CBOE (2015),
the VIX is computed as the present value of a portfolio of SPX call and put options constructed as
a static replication of a 30-days variance swap. In 2006, options written on the VIX also started
trading. Since then, several authors have studied the pricing of VIX options. The main strand
of literature addresses VIX derivative pricing under stochastic volatility models, mostly within
the affine class. This branch is pioneered by Zhang and Zhu (2006) and Zhu and Zhang (2007),
who derive dynamics for the VIX starting from a square-root model for the spot variance. The
works of Sepp (2008a,b) extend this approach by introducing jumps in the spot variance within
the affine jump-diffusion (AJD) framework of Duffie et al. (2000). The recent paper by Bardgett
et al. (2014) further generalizes the framework of Sepp (2008a,b) by allowing for a stochastic
long-run mean of variance. Non-affine pure-diffusion extensions of the square-root model for
the spot variance are in Gatheral (2008) and Bayer et al. (2013). Finally, modeling frameworks
based on infinite-dimensional specifications of the variance swap term-structure are proposed
by Buehler (2006), Bergomi (2008), and Cont and Kokholm (2013).
A common feature of these contributions is that the risk-neutral density (RND) is assumed
to be fully described by stochastic dynamic equations of the state-variables, which are functions
of the underlying model parameters. Unfortunately, fully parametric specifications of the dy-
namics of price and volatility come at the cost of an intrinsic risk of misspecification, see for
example Cont (2006) for a discussion. The problem of correct model specification in VIX option
pricing is particularly troublesome since the linkage between VIX and SPX is not fully explicit,
and they both depend on the variance, which is an unobservable quantity. Even when modeling
is solely addressed to the marginal density of the VIX, a comparative analysis of the performance
of simple stochastic volatility models in pricing of VIX options tends to confirm the potential
issue of misspecification. For example, Christoffersen et al. (2010) and Wang and Daigler (2011)
find some evidence in favor of models that assume log-normal dynamics for the instantaneous
variance, although none of these models achieve small pricing errors over the entire range of
strike prices. Moreover, the econometric analysis carried out by Mencia and Sentana (2013) re-
veals that the risk of model misspecification in structural pricing of VIX options is particularly
high during financial crises. This reflects the general disagreement in the literature on the ”na-
ture” and the roughness of the instantaneous volatility. In particular, although the instantaneous
volatility is most commonly modeled as a jump-diffusion process, Todorov and Tauchen (2011)
find that it is best described by a pure jump process, with clear consequences on the VIX index

2
and its related options. Reducing the model risk concerned with VIX option pricing is possible
but often comes at the cost of analytical tractability and availability of closed-form solutions.
Thus, modeling frameworks conceived for capturing stylized facts of the VIX are rarely suited
for estimation purposes.
In this view, non-structural methods for estimating the RND directly from VIX options rep-
resent a viable alternative to stochastic modeling. The term ”non-structural” is referred, in gen-
eral, to any option pricing method that does not rely on the postulation of a specific parametric
expression for the RND. This entails considerable reduction of the risk associated with mis-
specification. The idea that vanilla option prices can be linked to the RND through an explicit
non-structural relation was pioneered by Breeden and Litzenberger (1978). In the context of
VIX option pricing, a non-structural technique has recently been employed by Song and Xiu
(2016) with the purpose of estimating the volatility pricing kernel, which is the ratio between
the physical and the risk-neutral density of the VIX. More precisely, as in many other works
addressing non-structural option pricing, the authors propose to retrieve the RND by inferring
the second-order derivatives of option prices with respect to strikes directly from the market.
The non-structural approach proposed in this paper does not directly consider the second-order
derivatives of option prices, but it recovers the RND by means of an orthogonal polynomial ex-
pansion around a kernel density, see for instance Szegö (1939). Classic examples of orthogonal
expansions are the Hermite, which are obtained when the kernel is a Gaussian density, and the
Laguerre, which are obtained when the kernel is an exponential density. The key feature of or-
thogonal expansions is that they yield an explicit functional form of the RND without the need
of specifying stochastic dynamics of the state-variables. Instead, this method imposes mild in-
tegrability conditions on the form of the RND, proving to be particularly robust to misspecifica-
tion. There is extensive literature on the use of orthogonal expansions in financial applications.
Seminal examples are Jarrow and Rudd (1982), Corrado and Su (1996b), Madan and Milne (1994),
Coutant et al. (2001), and Jondeau and Rockinger (2001), while more recent contributions are
˜
Rompolis and Tzavalis (2008), Zhang et al. (2011), Nı́guez and Perote (2012), and Xiu (2014). In
all these cases, the expansions are provided in terms of Hermite polynomials. Our methodology
can be thought of as an alternative of Hermite expansions to the case of kernels with positive
support. Indeed, adapting the expansion kernel to the data (for instance by choosing a kernel
with support on the positive axis only) may be a better alternative to the inverse approach of
adapting the data to the kernel (for instance by log-change or standardization). In particular, we
extend the Laguerre expansions, used recently by Filipovic et al. (2013) and Mencia and Sentana
(2016), by introducing a family of kernels that encompasses well known distributions such as
the exponential, the Gamma, the Weibull and the GIG, among others. We show that the intro-
duction of the extended Laguerre (eLaguerre) kernels increases the adaptability of the approach
by reducing the number of restrictions to be imposed on the form of the RND.
We contribute to the VIX option pricing literature on several aspects. First, we provide
general convergence conditions of orthogonal expansions to the true RND. These conditions

3
relate to the rate of tail decay of the expansion kernel. We show that the log-normal density, due
to its slow tail decay rate, does not represent a suitable candidate for the expansion kernel as this
would generally lead to inaccurate approximations. Instead, our extended Laguerre kernels are
better suited to approximate the RND associated with the VIX options, due to the very flexible
decay rate on both tails. Indeed, owing to the irregular nature of the instantaneous volatility,
see Todorov and Tauchen (2011) and Todorov et al. (2014), the tails of the RND of the VIX are
expected to display features that can only be captured if a flexible choice of the kernel density
is adopted. Second, in the spirit of Aı̈t-Sahalia and Lo (1998), Jondeau and Rockinger (2001) and
Aı̈t-Sahalia and Duarte (2003), we propose a robust methodology to estimate the parameters of
the polynomial expansion by a minimum distance criterion based on the observed option prices.
We prove that the proposed methodology yields a very accurate approximation of the RND also
when the no-arbitrage and efficient option prices are contaminated by measurement errors.
Although this paper focuses on VIX options, our methodology is outlined in full generality, and
hence it can be applied to any option to recover the implied RND.
The analysis of the RND is carried out on a panel of VIX options collected at monthly fre-
quency for the period January 2010 - April 2016. The results highlight the reliability of our
methodology to recover RND up to negligible rounding errors and to mimic relevant quantities
embedded in the VIX options, such as the volatility of volatility index, i.e. VVIX. The time se-
ries of the first four risk-neutral moments of VIX display some interesting clues about volatility
expectations. First, we discover the presence of a positive correlation between mean and vari-
ance of VIX (reversed leverage effect) as well as a common factor structure across moments and
times-to-maturity. Second, by fitting a multiplicative error model (MEM) on the time series of
the VIX risk-neutral moments, we find strong empirical evidence in favour of a non-negligible
volatility jump probability under Q that is priced in the options. Third, the variance swap term
structure is empirically studied. The variance swap term-structure implicit in the VIX second
moments is coherent with the one directly computed from SPX options, proving that the two
markets are consistent with each other. This result entails that the second moment of VIX can
be traded through a combination of long and short positions on SPX options. Finally, we show
that it is crucial to account for the mean-reverting behavior of the realized variance to describe
the term-structure with a constant slope term.
The paper is organized as follows. Section 2 recaps the VIX formula and defines its RND,
while Section 3 summarizes some properties of orthogonal polynomial expansions. Section 4
discusses the estimation procedure based on principal components regression of the expansion
coefficients, under additional consistency constraints. Section 5 addresses whether and how
the estimated RND is affected by option prices contaminated by measurement errors associated
with no-arbitrage violations. Finally, Section 6 presents the empirical applications with real
data. Appendix A contains the proofs of the theorems related to the orthogonal polynomials,
while Appendix B contains some supplementary material with details on the fitting procedure,
the numerical experiments and the robustness to no-arbitrage violations.

4
2 The VIX index
Introduced in 1993 by the Chicago Board Options Exchange (CBOE), the VIX is a risk-neutral
forward measure of market volatility calculated as the fair value of a 30-days variance swap
rate. Futures and options on the VIX are the first derivatives on volatility to be traded on a
regulated security exchange, see Carr and Wu (2006) and Carr and Lee (2009) for a detailed
historical review of the VIX. At the core of the VIX construction, which is detailed in CBOE
(2015), lies the model-free methodology of Carr and Madan (1998), Britten-Jones and Neuberger
(2000), and Jiang and Tian (2005), to price a variance swap via static replication. According to
this methodology, under standard risk-neutrality and no-arbitrage assumptions, the s-variance
swap rate at time t, VSt,s , can be computed as
" !#
2 Q SPXt+s
VSt,s = − Et log , (1)
s xF

where Q denotes the risk-neutral pricing measure and x F is the SPX forward price observed at
time t for maturity t + s. The right-hand term in (1) can be replicated by the following portfolio
of SPX calls and puts
Z xF Z ∞
2e −r ·s
" !# !
2 SPXt+s 1 SPX 1 SPX
− Et log = P (x )dx + C (x )dx , (2)
s xF s 0 x2 xF x
2

where r is the risk-free rate and C SPX (x ) and P SPX (x ) denote prices of SPX call and put options
observed at time t, expressed as functions of the strike x. The square of VIX is defined by fixing
s to 30 days and by discretizing the infinite strip of out-of-the-money options in (2) over the
finite set of available strikes

VIXt2 2e −r ·s * X ∆xi SPX X ∆xi 1 xF

" #2
= P (xi ) + C (xi ) / +
SPX + −1 , (3)
s xi2 xi2 s x0
.
100 x ≤x x ≥x
, i 0 i 0 -

where ∆xi is half the distance between xi+1 and xi−1 , x 0 is the greatest available strike below x F .
In practice, when options with maturity exactly equal to s are not available, the VIX is calculated
by interpolating the values of (3) calculated by using SPX options with the largest available
maturity below 30 days (near term) and the smallest available maturity above 30 days (next
term). We should also remark that, in principle, formula (1) is not exact when the trajectories of
the SPX are subject to jumps. However, as pointed out by Carr and Wu (2009), the contribution
of price jumps in the computation of (1) can be neglected in practice.
The VIX index typically exhibits high negative correlation to the stock market, thus proving
an effective tool to hedge (or leverage) volatility separately from directional price moves. As a
matter of fact, many investors deem the VIX to be a leading indicator of market sentiment - the
index is often referred to as the market fear gauge. This is reflected by high trading volumes of
VIX options, which stand at approximately 37% of the average daily volume of SPX options, as

5
noticed by Mencia and Sentana (2013). Under risk-neutrality and no-arbitrage assumptions, for
fixed observation time t, time-to-maturity τ , and strike K, the price of a VIX call option CK (t, τ )
and the price of a VIX put option PK (t, τ ) are given by
Z ∞ Z ∞
+
CK (t, τ ) = e −rτ
f Q (t, τ ; x )(x − K ) dx, PK (t, τ ) = e −rτ
f Q (t, τ ; x )(K − x ) +dx , (4)
0 0

where f Q (t, τ ; ·) is the conditional RND of VIXt+τ , given all the past information up to time t.
When not misleading, we will omit the dependence of the RND on t, τ and simply denote it by
f Q . The RND of the VIX is the object of interest in this paper. In the next section, we discuss a
robust methodology to retrieve f Q directly from the VIX option prices.

3 Orthogonal polynomials
The methodology that we adopt builds upon the following expansion of the RND

n
f Q(n) (x )
X
fQ ≈ := ϕ (x ) 1 +
* ek (x ) + , n ≥ 1, (5)
, k=1 -

where ϕ is a chosen probability density function (kernel) and ek are corrective factors with
e 0 = 1. The kernel function ϕ can be seen as the 0-order term in (5)

f Q(0) (x ) := ϕ (x ) ,

and it can be interpreted as an initial proxy for the RND. In this work, we consider expansions
where the corrective factors e 1 , . . . , en take the following form

ϕ
ek (x ) = ck hk (x ),

ϕ
where, for every k = 1, . . . , n, ck is a real constant and hk is a polynomial of degree k in the
ϕ ϕ
state variable x. The polynomials h 1 , . . . , hn only depend on the kernel ϕ, and therefore the
coefficients c 1 , . . . , cn embed all the information on the RND.
The consistency of this approach relies on two results of functional analysis discussed in
ϕ ϕ
Sections 3.1 and 3.2. The first result ensures that if h 1 , . . . , hn are orthogonal to each other, in a
sense that will be clarified below, then by construction the resulting approximate density f (n)
enjoys unitary mass and other desirable properties regarding its moments. The second result
reveals that the relation expressed in (5) is non-structural, meaning that no specific form of
the RND needs to be postulated in order to ensure that (5) is an admissible expansion, i.e., that
f Q(n) → f Q as n → ∞. The results in Sections 3.1 and 3.2 are obtained under the maintained
assumption that ϕ is a probability density function with support D ⊆ R and possessing finite

6
polynomial moments, that is
Z
|x |k ϕ (x )dx < +∞, ∀k ∈ N.
D

3.1 Properties of orthogonal polynomials

ϕ
The elements of the family (hk )k∈N are said to be orthogonal polynomials with respect to ϕ if, for
all k ∈ N and all j ∈ N such that j , k,
 ϕ Z
ϕ ϕ
deg hk = k and hk (x )h j (x )ϕ (x )dx = 0. (6)
D

The existence of a family (hk )k∈N of orthogonal polynomials with respect to ϕ can be shown con-
structively, e.g., by applying the Gram-Schmidt orthogonalization to the basis 1, x, . . . , x n , . . ..
ϕ ϕ
For fixed n ∈ N, h 1 , . . . , hn are uniquely determined, up to a sign, if they also obey to the fol-
lowing normality condition Z
ϕ
hk (x ) 2ϕ (x )dx = 1 . (7)
D
ϕ
Henceforth, for a given kernel ϕ possessing finite moments, we denote by (hk )k∈N the unique
(up to a sign) family of related orthogonal polynomials satisfying condition (7) for every k ∈ N.
Furthermore, for a given n ∈ N, we denote by W := (wi,j ) the (n + 1) × (n + 1) lower triangular
ϕ ϕ
matrix containing the coefficients of h 0 , . . . , hn . Specifically, for i = 1, . . . , n we have

ϕ
hi (x ) = wi,0 + wi,1x + . . . + wi,i x i , (8)

and wi,j = 0 for j > i.

ϕ ϕ
For the numerical computation of the basis h 1 , . . . , hn , the following recurrence relation can
be used as a more efficient alternative to the Gram-Schmidt procedure. Define

ϕ ϕ x − Mϕ
h 0 (x ) = 1, h 1 (x ) = √ ,
Vϕ

where M ϕ and V ϕ are the mean and the variance of ϕ, respectively. The remaining terms,
ϕ ϕ
h 2 , . . . , hn , can be computed as

ϕ 1 f ϕ ϕ
g
hk (x ) = (x − ak )hk−1 (x ) − bk hk−2 (x ) , (9)
Ck

where Z Z
ϕ ϕ ϕ
ak = xhk−1 (x ) 2ϕ (x ) 2dx , bk = xhk−1 (x )hk−2 (x )ϕ (x ) 2dx ,
D D
Z ! 12
ϕ ϕ
f g2
Ck = (x − ak )hk−1 (x ) − bk hk−2 (x ) ϕ (x ) 2dx .
D

7
 An important property of orthogonal polynomials is that they fulfill a mass conservation
principle that extends to all moments. More specifically, for fixed p ≥ 0
Z Z
(p)
x p
f Q(n) (x )dx = x p f Q (x )dx, ∀n ≥ p. (10)
D D

A major consequence of (10) is that, irrespective of the order n, the approximated RND obtained
by (5) always integrates to 1, irrespective of the order n, i.e.,
Z
f Q(n) (x )dx = 1 ∀n ∈ N. (11)
D

Furthermore, in view of (10), we can interpret c 1 , . . . , cp as corrective factors of the first p mo-
ments of ϕ to match the corresponding moments of f Q . In general, f Q(n) is not guaranteed to
be a positive function over its support, even under the assumption that ϕ is a positive function.
This property is recovered when the n-th coefficient cn fulfills some constraints. More precisely,
f Q(n) ≥ 0 if and only if
cninf ≤ cn ≤ cn ,
sup
(12)

where
f Q(n−1) (x ) f Q(n−1) (x )
cninf = − , cn = .
sup
sup ϕ
inf ϕ
x
ϕ
: hn (x )>0 hn (x ) ϕ
x : hn (x )<0 hn (x )
The result above follows after noticing that

ϕ
f Q(n) = f Q(n−1) + cnhn .

3.2 Admissibility of orthogonal expansions

This paragraph is devoted to discussing the admissibility of expansion (5) for the RND. The fol-
lowing theorem represents a prerequisite for setting up a mathematically well-posed procedure
to estimate the expansion coefficients c 1 , . . . , cn based on observed option prices.

Theorem 3.1. Assume that supp( f Q ) ⊆ D ⊆ R+ and that ϕ −1 f Q2 is integrable over its support.
1
Moreover, assume that limx→+∞ ϕ (x )e ςx 2 = 0 for some ς > 0 and that pϕ is bounded for some
polynomial p. Then:

(a) there exists a sequence (ck )k∈N such that, for a proper subsequence of indexes1 (kn )n∈N ,

kn
ϕ
X
f Q (x ) = lim ϕ (x ) .1 +
* ck hk (x ) +/ for a.e. x ∈ D ;
n→+∞
, k=1 -

1 There is no need for a subsequence if the RND fulfills additional regularity assumptions, e.g., smoothness.

8
 (b) the following holds in the limit
Z +∞ Z +∞
lim Π(x ) f Q(n) (x )dx = Π(x ) f Q (x )dx , (13)
n→+∞ 0 0

for any function Π such that Πϕ 2 ∈ L2 (D).

1

Proof. See Appendix A.1. 

Point (a) in Theorem 3.1 provides sufficient conditions to ensure that the RND admits the
representation (5) or, in other words, that f Q(n) converges to f Q for some c 1 , . . . , cn , . . .. Point
(b) ensures that we can move the limit within the integral pricing formulas in (4) and obtain
convergent expansions for prices of call and put options. Looking at the hypotheses of Theorem
3.1, we note that they are essentially sharp conditions, meaning that they cannot be further
relaxed and, in fact, they sharpen the assumption of Filipovic et al. (2013), which allows for
Gamma-type decay in the kernel. Indeed, if ϕ −1 f Q2 is not integrable, then the expansion (5) is not
well-defined since it may diverge as n → ∞. This means that the kernel should not decay faster
1 −γ
than the RND. On the other hand, if limx→+∞ ϕ (x )e ςx 2 > 0 for some γ , ς > 0, then Theorem
3.1 is still well-defined but will not necessarily converge to f Q , as pointed out by Theorem A.3-
(ii), reported in the Appendix A.1. This suggests that the expansion (5) may not be sufficiently
√
informative on the RND when the kernel decays too fast, i.e., faster than e − x . However, in
view of Theorem 3.1, the kernel ϕ mostly serves as an initializing state of the expansion (5).
Therefore, as long as the choice of ϕ complies with the hypotheses underlying Theorem 3.1, its
impact on the form of f Q(n) will be of marginal importance, provided that n is sufficiently large.
The validity of this statement is supported by numerical illustrations provided in Section B.2 of
the Supplementary material. Finally, the following remark highlights the existence of a linear
mapping between the coefficients (ck )k∈N and the moments of f Q .

Remark 3.2. If the hypotheses underlying Theorem 3.1 are satisfied, then one can show that for
every k ∈ N

Z k Z +∞
ϕ
x i f Q (x )dx
X
ck = hk (x ) f Q (x )dx = wk,i (14)
supp(ϕ) i=0 −∞

ϕ
where wk,i is the i-th coefficient of hk .

3.3 The extended Laguerre and the log-Hermite expansions

In this sub-section, we study the properties of a class of kernel densities that can form the basis
for the polynomial expansions used to retrieve the RND of VIX. Based on restrictions imposed
by Theorem 3.1, we propose the following family of kernels with support on D = [0, +∞[,

ϕ (x ) ∝ x α−1e − ( βx ) 1D (x ),
p +ξ x −1
α, β, ξ , p ∈ Θ, (15)

9
where
Θ = α, β, ξ , p ∈ R | β > 0, 0 < p ≤ 1, (α > 0, ξ = 0) ∨ (α ∈ R, ξ > 0) .


The specification (15) embeds a number of notable sub-cases such as the Gamma (for p = 1, ξ =
0), the generalized inverse Gaussian (GIG, for p = 1), and the generalized Weibull (GW, for
ξ = 0) kernel. Therefore, the orthogonal expansions arising from (15) extend the classic La-
guerre expansions that are associated with a Gamma kernel. For this reason, we refer to these
expansions as extended Laguerre. This family is flexible enough to capture different tail behav-
iors and it proves effective in reproducing the peculiar tail behavior of the RND implied by VIX
options, as later shown in Section B.2. Indeed, the tail behavior of the GIG and GW kernels char-
acterizes their ability to meet the condition ϕ − 2 f Q ∈ L2 (D), required by Theorem 3.1. Looking at
1

the left tail, the GW kernel clearly has the slowest decay, and thus it is the most flexible in terms
1
of behavior around the origin. However, Theorem 3.1 also requires that limx→+∞ ϕ (x )e ςx 2 = 0
for some ς > 0. This condition is always met by the GIG kernel, and it is met by the GW kernel
when p is restricted to fall between [ 12 , 1].
As alternative to the extended Laguerre kernel family, one could consider the log-normal
(LN) kernel, that is
ϕ (x ) ∝ e − 2σ 2 (log(x )−µ) , µ ∈ R, σ > 0.
1 1 2
(16)
x
Expanding the underlying RND based on the LN kernel is conceptually similar to applying a
Hermite expansion to the logarithm of the underlying. Furthermore, there is documented em-
pirical evidence that the volatility, a quantity comparable to the VIX, is roughly log-normally
distributed, see e.g. Christoffersen et al. (2010), Wang and Daigler (2011) and Bayer et al. (2013).
This makes the LN kernel an interesting competitor for either the GIG or the GW kernel. Con-
cerning the condition ϕ − 2 f Q ∈ L2 (D), the LN kernel ensures the greatest flexibility in terms of
1

1
requirements on the right tail of the RND. However, the condition limx→+∞ ϕ (x )e ςx 2 = 0 for
some ς > 0 is never met by the LN kernel. Thus, the LN kernel does not guarantee that the
RND is fully recovered by the expansion (15). Therefore, using the LN kernel inherently entails
further restrictions on the form of f Q , and it may reduce the ”non-structurality” of the approach.

4 Retrieving the RND from the option prices

In this section, we outline a procedure to estimate the coefficients c 1 , . . . , cn of the expansion
(5) by minimizing the distance between the RND implied and the observed option prices. The
key feature of the proposed procedure is to estimate the option pricing formulas in (4) based
on the observed market prices by choosing a sufficiently large n in the expansion (5), where
the coefficients c 1 , . . . , cn are the unknown terms that convey the information about f Q . The
consistency of this procedure relies on Theorem 3.1.
For fixed K ≥ 0, n ∈ N, and c = [c 1 , . . . , cn ]> ∈ Rn , the prices associated with the expansion

10
of order n are defined as
Z +∞ n
(ϕ)
CK(n) (c)
X
:= ϕ (x ) 1 +
* ck hk + (x − K ) dx ,
K , k=1 -
Z K n
(ϕ)
PK(n) (c) :=
X
ϕ (x ) *1 + ck hk + (K − x ) dx .
0 , k=1 -

The expressions above can be rewritten in the following compact form

CK(n) (c) = A0(K ) + A(K )c , PK(n) (c) = B 0(K ) + B (K )c (17)

where Z +∞ Z K
A0(K ) = ϕ (x ) (x − K ) dx, B 0 (K ) = ϕ (x ) (K − x ) dx ,
K 0

and A(K ) and B (K ) are 1 × n vectors, whose i-th element is given by

i Z +∞ i Z K
Ai(K ) x j+1 − Kx j ϕ (x )dx , Bi(K ) Kx j − x j+1 ϕ (x )dx .
X   X  
= wi,j = wi,j (18)
j=0 K j=0 0

Therefore, for chosen ϕ and n, one may estimate c 1 , . . . , cn by collecting a cross-section of

undiscounted market prices, CKObs m
(t, τ ) and PKObs
m
(t, τ ), for m = 1, . . . , M, and by finding the
solution ĉ = [ĉ 1 , . . . , ĉn ]> of the following optimization problem

ĉ = argmin Q (t, τ ; c) , (19)

c∈Rn

where Q (t, τ ; ·) defines an objective function to be minimized. Given that the expressions in
(17) are linear in the coefficients c, natural choice for Q (t, τ ; ·) is the criterion function of the
minimum least squares problem for the following linear model

Y = X0 + Xc + ε, (20)

where the 2M × n matrix X is

 A1(K1 ) . . . An(K1 ) 
.. .. ..
 
 . . . 
A1(K M ) (K M )
. . . An
 
X =   ,
B 1(K1 ) . . . Bn(K1 ) 
.. ..

... 
. .
 
B 1(K M ) . . . Bn(K M )
 
 

Y = [CKObs
1
(t, τ ), . . . , CKObs
M
(t, τ ), PKObs
1
(t, τ ), . . . , PKObs
M
(t, τ )]0, and X0 = [AK0 1 , . . . , AK0 M , B K0 1 , . . . , B K0 M ]0.
The 2M × 1 vector ε represents the error term, whose properties are discussed more in detail in

11
Section 5. In this way, the objective function takes the following quadratic form

Q (t, τ ; c) = (Y∗ − Xc) 0 (Y∗ − Xc), (21)

where Y∗ = Y − X0 is a 2M × 1 vector and X0 is the option price vector generated by the kernel.
Unfortunately, the columns of X are functions of the first non-standardized n moments. As
a consequence, they display an increasing degree of multicollinearity as the expansion order n
grows. Employing the standard LS minimization to solve (21) is therefore not suitable when n
is large. This is a well-known problem in the literature on orthogonal polynomial expansions.
For example, Jarrow and Rudd (1982) and Corrado and Su (1996a,b) consider expansions only
up to the fourth order, i.e., n = 4, and they calibrate the standardized skewness and kurtosis to
the options on the SPX. Similarly, Jondeau and Rockinger (2001) estimate the RND of the Franc-
Mark exchange rate by matching only the first four moments, which implies once again that n
does not exceed 4. In this regard, it is important to stress that the RND of VIX is expected to be
characterized by a long right tail, meaning that the moments higher than the fourth may provide
significant information on the shape of the RND. We solve the problem of multicollinearity by
orthogonalization of the regressors in (20). The latter, accomplished by means of PCA, also
allows to achieve a dimensionality reduction of the problem in (19) without discarding a priory
potentially relevant information. The Supplementary material, in Section B.1, provides further
details on the implementation of the PCA analysis in this context. The vector of coefficients that
minimizes the quadratic objective quadratic function under the PCA constraints is denoted by
c̃. Given the vector c̃, the estimated RND function f˜Q(n) is determined as

n
(n) ϕ
X
f Q (x ) = ϕ (x; θ ) 1 +
˜ * c̃k hk (x ) + . (22)
, k=1 -

Note that the kernel ϕ in (22) is now expressed as function of an additional term, to highlight
its dependence on a set of parameters θ ∈ Θ. For example, θ = [α, β, ξ ]0 for the GIG kernel,
θ = [α, β, p]0 for the GW kernel, and θ = [µ, σ 2 ] for the LN kernel. The choice of θ does not
play a crucial role in this context as long as ϕ (·; θ ) fulfills the assumptions of Theorem 3.1. In
practice, we set θ to the value that minimizes the residuals variance for the expansion of order
0. The minimization is performed under the restriction of zero-mean residuals, which implies
absence of systematic pricing errors. A number of numerical examples, reported in Section
B.2 of the Supplementary material, display the robustness and the flexibility of the proposed
methodology.

5 No-arbitrage violations
Market prices are typically subject to a number of frictions that are typically functions of the
market liquidity. Obviously, possible arbitrage opportunities can hardly be exploited due to the

12
presence of transaction costs in the form of bid and ask spread. However, from a purely mathe-
matical perspective, the fact that the mid-quote is adopted to approximate the latent arbitrage-
free option price can be seen as a violation of the no-arbitrage assumption. This means that
even in absence of discretization/truncation errors it is not possible to achieve exact matching
of all the observed option prices by minimizing (19), since the option prices obtained through a
RND are free of (static) arbitrages by construction. Therefore, in defining the linear model (20)
on which we build the estimation procedure based on orthogonal polynomials, we can assume
that the error term ε subsumes all the uncertainty associated with the fact that the polynomial
expansion is truncated to a finite n, that the number of available strikes M is finite and that
the market prices may be subject to no-arbitrage violations. In other words, the error term in
(20) can be split as ε = δ + ϵ, where δ is a vector of non-stochastic terms coming from the
fact that both n and M are finite, while ϵ is a random term that approximates all the deviations
from the latent arbitrage-free option prices resulting from the trading activity. In particular, we
assume that ϵ is a vector of random variables with zero mean and that the vector Y = Y − ϵ
is arbitrage-free. An example of specific distributional form for ϵ is given in Section B.6 in the
Supplementary material.
In the following, an infill asymptotic analysis is carried out to show that, as the observed
prices are sampled increasingly over a fixed interval of strikes and as n → ∞, then δ → 0 and the
only remaining error term is the noise associated with the no-arbitrage violations. Assuming
that the observed option strikes fall in a fixed finite interval I = [K 1 , KM ], we define the ”infill
version” of (21) as
Z 
CKObs (t, τ ) − CK(n) (c) + PKObs (t, τ ) − PK(n) (c) dK .
(n) 1 2  2
Q (t, τ ; c = [c 1 , . . . , cn ] ) :=
>
KM − K1 I
(23)
Note that there are other ways to define an ”infill counterpart” of Q consistently. The definition
that we adopt, justified by mathematical convenience, builds on the fact that the integral in (23)
is the limit of M1 Q, under continuity assumptions for the integrand. Since multiplying Q by any
constant does not affect the solution ĉ of (19), then (23) can be interpreted as a valid infill version
of (21). The observed prices C Obs (t, τ ), P Obs (t, τ ) appearing in (19) are assumed to take the form

CKObs (t, τ ) = CK (t, τ ) + ϵKC , PKObs (t, τ ) = PK (t, τ ) + ϵKP ,

where C· (t, τ ), C· (t, τ ) ∈ C 2 (I ) are defined as in (4), while ϵ C = (ϵKC )K ∈I and ϵ P = (ϵKP )K ∈I are
zero mean processes on a probability space (Ωϵ , F ϵ , P ϵ ), belonging to L2 (Ωϵ × I ). Under these
assumptions, the infill target function Q (n) is well-defined and has finite expected value for every
t, τ , n.

Proposition 5.1 (Infill asymptotics). Assume that the hypotheses of Theorem 3.1 are satisfied and

13
denote by c ? (n) the minimum of Q (n) (t, τ ; ·), for every n ∈ N. The following inequality holds
"Z    2 #
(n) ? 1
ϵK + ϵKP dK .
C 2
f g
lim E Q (t, τ ; c (n)) ≤ E (24)
n→+∞ KM − K1 I

The inequality in (24) becomes an equality under the following additional hypotheses:

(i) There exists n̄ ∈ N such that, for all n ≥ n̄, c ? (n) is obtained by constraining (19) to the space
of coefficients c 1 , . . . , cn such that

kn
ϕ
X
ϕ .1 +
* ck hk +/ ≥ 0 on D.
, k=1 -

(ii) Let C·? (t, τ ) ∈ C 2 (I ) and P·? (t, τ ) ∈ C 2 (I ) be arbitrage-free call and put curves, respectively.
Then, almost surely
Z  Z  
Obs ? Obs ?
ϵKC + ϵKP dK .
2  2 2  2
CK (t, τ ) − CK (t, τ ) + PK (t, τ ) − PK (t, τ ) dK ≥
I I

Proof. See Appendix A.2. 

The inequality (24) defines an upper-bound on the expected value of the target function,
Q (n) (t, τ ; c ? (n)).
In particular, the expected value of the target function evaluated in c ? (n) is
lower than the variance of the non-arbitrage residuals. Under the additional assumptions (i)-
(ii), Proposition 5.1 states that our estimation method provides the arbitrage-free prices closest
to the observed ones. In particular, assumption (i) requires that the estimation always returns a
probability density function, while assumption (ii) can be interpreted as a uniqueness require-
ment on the target RND. This establishes an interesting linkage with the work of Aı̈t-Sahalia and
Lo (1998). Furthermore, under no-arbitrage (i.e. ϵ C = 0 and ϵ P = 0), Proposition 5.1 ensures that
the sum of the squared residuals goes to zero as n → ∞, so that the estimated and the observed
prices coincide.
Summing up, Proposition 5.1 provides conditions ensuring that the estimation procedure
based on orthogonal polynomials is robust to the presence of measurement errors in the option
prices. This is a remarkable feature that is often lacked by non-structural techniques based on
the straightforward computation of second-order derivatives, whose estimation is typically very
sensitive to data inconsistencies. Below, we derive a theoretical lower bound for the estimation
residuals that is inferred directly from the put-call parity violations affecting the observed option
prices. In Section B.6 in the Supplementary material we show that this lower bound can be
consistently used as a proxy for (24). Moreover, we show that the validity of Proposition 5.1 is
empirically confirmed by a number of numerical tests.

14
5.1 An observable lower bound for the estimation residuals
Following Proposition 5.1, it seems natural to set a tolerance on the variability of the residuals,
which are defined as ε˜ = Y∗ − Xc̃. This tolerance should quantify the presence of arbitrage
opportunities. Given a fixed threshold ∆Q > 0, we define as admissible any RND implying
option prices whose distance from the observed ones is below ∆Q . Consistently, we say that a
density f˜Q(n) is admissible if

M M
(n) 2
PKm − P̃K(n)m ≤ ∆Q ,
1 X  Obs  1 X  Obs 2
CKm − C̃Km +
M m=1 M m=1

where C̃K(n)m and P̃K(n)m are call and put prices associated with f˜Q(n) . Since the option data are always
affected by some noise, in view of Proposition 5.1 the existence of admissible RNDs is not guar-
anteed when ∆Q is chosen to be too small. A lower bound for the set of all possible values of ∆Q
can be inferred from put-call parity violations. Given f˜Q(n) , denote by MeanQ the risk-neutral
futures price computed as Z +∞
MeanQ = x f˜Q(n) (x )dx ,
0
and by ∆pcp the variance of the put-call parity violations

M
1 X  Obs
CKm (t, τ ) − PKObs Obs 2

∆pcp = (t, τ ) + Km − Mean , (25)
M m=1 m

1 PM  Obs
where MeanObs = Obs (t, τ ) + K . It follows that

C (t, τ ) − P m
M m=1 Km Km

M
1 X f Obs (n) (n)
CKm (t, τ ) − PKObs Obs 2
 g
∆ pcp
= (t, τ ) − C Km + P Km + F Q
− F
M m=1 m

M M 
CKm (t, τ ) − C̃K(n)m + (n) 2
1 X  Obs 2 1 X
PKObs Obs 2
  
≤ (t, τ ) − P̃ Km + Mean Q
− Mean ,
M m=1 M m=1 m

which yields the following inequality

∆Q ≥ ∆pcp − MeanQ − MeanObs .

 2
(26)

From (26) we obtain a lower bound for the tolerance level that must be allowed on the estimation
residual. Moreover, it proves that admissible solutions of (19) with tolerance level lower than
Obs 2
 
∆pcp − MeanQ 1 − Mean do not exist. Therefore, (26) suggests that setting the tolerance level
as ∆Q = ∆pcp is a convenient choice since ∆pcp is an observable quantity and is expected to be
only slightly greater than the lower bound.

15
6 Empirical Analysis
In this section we estimate the RND of the VIX on a panel of option prices observed in the period
from January 2010 to April 2016, sampled at monthly frequencies, and with time-to-maturity,
τ , ranging from 1 to 5 months.2 The data is obtained from the OptionMetrics database. For each
month in the sample, we collect option prices observed on the first Tuesday following the third
Friday of the month. This ensures that observations will not overlap in the base monthly fre-
quency, meaning that 1-month options observed at time t always expire prior to the subsequent
observation occurring at time t + 1. We operate minimum pre-filtering of the data. More specif-
ically, we exclude all OTM puts (calls) with mid-quote below 0.025$ together with the ITM calls
(puts) with the same strikes. These contracts turn out to be highly illiquid (if traded at all) and
therefore they are likely subject to mispricing. With this filtering criterion, we have an average
total number of 56 available contracts for each date and time-to-maturity. Under normal mar-
ket conditions, the strike values that are taken into consideration typically fall between 10$ and
45$, with this range remaining quite stable over time and maturities due to the mean-reverting
behavior of the VIX. However, the interval of available strikes enlarges during turmoil periods,
with maximum values reaching peaks of 100$. For further details see Table 7 in Section B.7
of the Supplementary material. In the following, we focus on a specific date and maturity in
the sample to illustrate the effectiveness of the orthogonal polynomials in correcting ϕ and to
confirm the robustness of the methodology to the choice of the kernel.

6.1 November 16, 2011

On November 16, 2011, the cross-section of VIX options expiring on December 21, 2011 (τ =
1) quoted by the CBOE consists of 64 contracts. The choice of this date is not coincidental.
Indeed, the end of the year 2011 is characterized by high levels of market volatility, registered
in connection with the European sovereign debt crisis and the US sovereign debt downgrading.
As a consequence, on November 16, 2011, the VIX index reached a value of 33.51%. At the
same time, the trading of VIX options spanned strike values in the range between 15$ and 90$.
After the pre-filtering, we end up with 52 contracts (26 calls and 26 puts) with strikes ranging
between 21$ and 80$. Hence, on the chosen date, trading of deep out-of-the-money options is
sufficiently high to ensure informative market prices in a wide range of strikes and, in particular,
suggests uncommonly long right-tail in the RND. To highlight the robustness of our results to
the choice of the kernel, we consider both the GIG and the GW kernels. The parameters of
the two kernels are chosen to minimize the variance of Y ∗ . As discussed in Section 4, we can
increase the expansion order n to high values, thanks to the orthogonalization of regressors
and the dimension reduction thereby operated by means of PCA. Therefore, we set the order
to a relatively high value n = 18. We find that 5 and 6 principal components explain the 99%

2 Weekly VIX options and options with time-to-maturity τ = 6 are also traded, but unfortunately they are not

available for all dates in the sample.

16
of the total variance of X for the GW and the GIG kernel, respectively. The estimated RNDs
obtained using expansions of order n = 18 are reported in Figure 1 (solid lines), together with the
corresponding kernels (dashed lines). From the visual inspection of Figure 1, it emerges that the

Expansion (GIG) Expansion (GW) GIG kernel GW kernel

0.07 -1
10

0.06

-2
10
0.05

0.04
-3
10
0.03

15% - 18%

75% - 71%

95% - 96%

98% - 98%

99% - 99%
0.02

1% - 2%
-4
10

0.01

0 -5
10
10 20 30 40 50 60 70 80 90 100 20 30 40 50 60 70 80 90 100

Figure 1: Estimated RNDs. The left panel depicts the graphs of the two kernels considered in this section
together with the corresponding estimated RNDs obtained for n = 18. The right panel reports the same
contents of the left panel in semi-log scale, to highlight tail features that are difficult to observe in linear
scale. Each couple of percentage values denotes, from top to bottom, the average mass levels of the two
kernels and the two corresponding estimated RNDs, related to the quantiles identified by the dashed
vertical lines.

RND of VIX should have short left tails. The densities and the associated kernels display relevant
differences on the right tails. The differences mostly occur in a part of the domain above the
98% quantile, thus signaling the relevance of a proper characterization of the right tail in pricing
OTM calls and ITM puts of VIX. The importance of the correction provided by the orthogonal
polynomials is better understood by comparing the implied volatility curves generated by the
two kernels and their corresponding expansions, reported in Figure 2. The implied volatilities
generated by both kernels are considerably different from those generated by market prices,
while the expansions are able to produce implied volatilities that closely replicate the observed
ones. All these observations point in the same direction: neither of the two baseline kernels is
able to reproduce the tail-features of the RND of VIX. In particular, they both display positive
excess mass between the 75% and 95% quantiles while, on the other hand, they display negative
excess mass in the area covered by the last 5 percentiles. Coherently with Theorem 3.1, the
expansions obtained using the two different kernels converge to the same density, as the two
functions diverge from each other only starting from roughly the 99.5% quantile.

6.2 Stylized facts on VIX risk-neutral moments

The analysis carried out for November 11, 2016 is replicated for all dates and maturities in the
sample. From Figure 3, it emerges that the orthogonal expansion (blue line) outperforms the
GW kernel (red line) and, in all cases, it generates a root mean square error that lies around the

17
 1.2 1.2

1.1 1.1
Implied Volatility

Implied Volatility
1 1

0.9 0.9

0.8 Market 0.8 Market

Expansion order 18 (GIG) Expansion order 18 (GIG)
0.7 Expansion order 18 (GW) 0.7 Expansion order 18 (GW)
GIG kernel GIG kernel
0.6 GW kernel 0.6 GW kernel

30 40 50 60 70 80 30 40 50 60 70 80
Strike Strike
(a) Call Imp. Vol. (b) Put Imp. Vol.

Figure 2: Black and Scholes implied volatility curves obtained from market prices, GIG kernel, GW
kernel, and the resulting approximated RNDs with expansion order equal to n = 18.

√
market implied threshold ∆pcp .3 This signals the goodness of fit for all the estimated RNDs in

0.12
"pcp 0.25 "pcp
0.1 Kernel Kernel
Expansion Expansion
0.2
0.08
0.15
0.06

0.1
0.04

0.02 0.05

0 0
10

11

12

13

14

15

16

10

11

12

13

14

15

16
20

20

20

20

20

20

20

20

20

20

20

20

20

20

(a) τ = 1 (b) τ = 5

Figure 3: Root mean square error between market VIX option prices and approximate prices implied
by both the GW kernel and its related expansion. Option prices have been collected based on monthly
observations taken from January 2010√to April 2016 with τ = 1 and τ = 5 months. The grey area is the
market implied threshold, defined as ∆pcp , see equation (25).

the sample and the accuracy gain achieved by correcting the GW kernel through the orthogonal
expansion. To further assess the consistency of the estimated RND with market data, we can
look at the VVIX, i.e., the volatility-of-volatility index inferred from VIX options through the
same algorithm used for the VIX itself. The VVIX can be linked to the RND of the VIX through

3 Similar results are obtained with other maturities, τ = 2, 3, 4 and are reported in the Supplementary material

together with a table reporting the summary statistics associated with the fit.

18
the following formula, which is analogous to (1)

+∞
x +
2 Z
VVIXt,τ 2
=− log * f Q (t, τ ; x )dx , (27)
100 τ 0
Q
, Meant,τ -

where Meant,τQ
:= EtQ [VIXt+τ ] denotes the forward price of VIX at time t for maturity t + τ .
Figure 4, which reports the observed time series of the VVIX and that computed by (27), confirms
that the estimated RND implies generally consistent estimates of the VVIX. Discrepancies are
small in size and can be attributed to discretization errors affecting the CBOE formula for the
computation of the VVIX.

150 VVIX, observed VVIX, observed

VVIX, implied by RND 70 VVIX, implied by RND
140

130
65
120

110
%

100 60

90

80
55
70

60
2011 2012 2013 2014 2015 2016 2011 2012 2013 2014 2015 2016
Year Year

(a) τ = 1 (b) τ = 5

Figure 4: VVIX time series. Comparison between observed values of the VVIX and those obtained by
formula (27), for τ = 1 and τ = 5 months.

By direct inversion of the linear relation in (14), it is possible to retrieve the first k risk-neutral
moments of the VIX for each value of t and τ . An empirical analysis of the risk-neutral moments
of the VIX reveals a set of stylized facts related to the market expectation of volatility (and its
Q Q
powers). Figure 5 reports the time series of mean (Meant,τ ) and variance (Vart,τ ), for all times-
to-maturity. From Panels (a) and (b), we notice that the first two expected risk-neutral moments
exhibit large variations associated with the escalation of the European sovereign debt crisis in
the second half of 2011. From 2013 to the end of the sample, both moments display lower levels
and more stable patterns reflecting agents’ confidence in more stable market conditions. For
what concerns the first moment, the spread between maturities has remained largely uniform
along the sample, i.e., an average spread between τ = 5 and τ = 1 of 3.094 with a standard
deviation of 2.135. Additionally, from Panel (c) we can see that changes of sign in the slope of
the term structure are sporadic and tend to occur concurrently with extreme market conditions,
in analogy with the well-known case of inverted yield curve. The periods with negative slope
of the mean term structure are associated with the first Greek debt crisis (April-May 2010), the
escalation of the European sovereign debt crisis (October-November 2011), the second Greek
debt crisis (July 2015) and the slowdown of Chinese production (February 2016), which all reflect

19
expectations of increasing market volatility in the short term.

35

250

30
200

25
150

20 100

50
15

2011 2012 2013 2014 2015 2016 2011 2012 2013 2014 2015 2016

(a) Mean (b) Variance

6 160

140
4
120

2 100

80

0
60

40
-2
20

2011 2012 2013 2014 2015 2016 2011 2012 2013 2014 2015 2016

(c) Slope of MeanQ (d) Slope of VarQ

Figure 5: Time series of the VIX risk-neutral mean and variance. Panel a) and b) report the time series of
Meant,τQ Q
) and Vart,τ for different times-to-maturity, τ = 1 (blue), τ = 2 (red), τ = 3 (yellow), τ = 4 (purple)
and τ = 5 (green). Panels c) and d) also report the time series of the slope of MeanQ and VarQ . The slope
Q
is defined as Meant,1 Q
− Meant,1 Q
and Vart,1 Q
− Vart,1 ∀t = 1, . . . ,T for mean and variance, respectively.

Q
A generally positive slope in the term structure is also observed for Vart,τ , whose dynam-
Q
ics closely resemble those followed by Meant,τ , see Panel (b). The slope of the variance term
structure, measured by the spread between values of Vart,τQ
observed for τ = 5 and τ = 1, peaks
during periods of market turmoil up to 177.022, against an average value of 73.751 over the
whole period. This behavior reflects agents’ lack of confidence about the level of market stabil-
ity in the long term. Only in one case, during the second Greek debt crisis in July 2015, does
the slope not increase to abnormal levels in response to the market turmoil, although it remains
positive.
Q Q
We further investigate the link between Meant,τ and Vart,τ to measure the leverage effect.
The leverage effect generally relates to the negative correlation between the change of an asset
price and its volatility. We perform regressions of ∆Meant,τQ
on a constant and ∆Vart,τ
Q
for each
maturity τ . The coefficients associated with the change in Vart,τ are all positive and increasing
Q

20
in τ , ranging from 0.241 (τ = 1) to 0.315 (τ = 5), and the regressions R-squared are around 75%.
Such evidence suggests the presence of a leverage effect with positive (reversed) sign. As the
uncertainty about future levels of market volatility increases, it is arguable that this has direct
implications on the expected level of market volatility itself, which increases proportionally.
By means of the normalized versions of the third and fourth moments reported in Figure
Q Q
6, namely the skewness (Skt,τ ), and kurtosis (Kurtt,τ ), we can study how the shape of the RND
Q Q
changes over time. Both Skt,τ and Kurtt,τ appear highly volatile and share similar dynamics,
as also suggested by their sample correlations reaching levels close to 97%. Contrary to what
Q Q
we have observed for mean and variance, Skt,τ and Kurtt,τ plummet during periods of market
turmoil. Thus, the third and the fourth moment increase at a slower rate with respect to the
variance. As the conditional mean and variance of VIX increase, moving away from the zero
lower bound, the distribution tends to become more symmetric and thinner-tailed. Notably,

5.5
50

5 45

40
4.5
35
4
30

3.5 25

20
3
15
2.5 10

2 5

2011 2012 2013 2014 2015 2016 2011 2012 2013 2014 2015 2016

(a) Skewness (b) Kurtosis

Figure 6: Time series of the VIX risk-neutral skewness and kurtosis. Panel a) and b) report the time
Q
series of Skt,τ Q
) and Kurtt,τ for different times-to-maturity, τ = 1 (blue), τ = 2 (red), τ = 3 (yellow), τ = 4
(purple) and τ = 5 (green).

Q Q
both Skt,τ and Kurtt,τ exhibit a term structure that is systematically downward sloped. Sample
averages decrease from 3.43 (τ = 1) to 2.78 (τ = 5) for the Skt,τ Q
and from 22.90 (τ = 1) to
15.55 (τ = 5) for Kurtt,τ . As the prediction horizon increases, the distribution implied by the
Q

VIX options reflects higher expectations and volatility for the level of VIX, together with a more
symmetric and platykurtic distribution.
The similarity in the dynamic behavior of the first four moments becomes striking after
simple rescaling. As an example, Figure 7 plots the standardized time series of the first four
moments for τ = 1 and τ = 5, with lines tracking one another. The sample correlations are
well above 90% for both τ = 1 and τ = 5, supporting the existence of a strong link between
(standardized) moments. To further explore the behavior of the implied moments both over the
time dimension and the term-structure, we test for the existence of a common factor structure
across maturities and moment orders. This is conceptually similar to the analysis carried out
by Andersen et al. (2015b) on the SPX implied volatility surfaces. The PCA, carried out on the

21
 4 4

3 3

2 2

1
1

0
0

-1
-1

2011 2012 2013 2014 2015 2016 2011 2012 2013 2014 2015 2016

(a) τ = 1 (b) τ = 5

Figure 7: Standardized moments for short (τ = 1) and long τ = 5 time-to-maturity. The standardization
is performed as Z i = MS i(M
−M̄ i
i)
for i = 1, . . . , 4 where M̄ and S (Mi ) are the time-series sample mean and
standard deviation of the i-th moment.

cross-section of τ , shows that the first principal component of Meant,τ Q

explains 98% of its total
Q
covariance. Similarly, the first principal component of Vart,τ explains almost 96% and, for the
third and fourth moments, the first principal component explains 89% and 82% of the related
cross sectional (across maturities) variability.
Q
When regressing Meant,τ on a constant and the first principal component, we find regression
intercepts that increase with the maturity, ranging from 19.26 (τ = 1) to 22.17 (τ = 5), but
slopes that are fairly invariant. This suggests that VIX options prices incorporate premia for
the uncertainty about future market conditions as the horizon increases, but also that these
premia are constant over time, and thus deterministic in nature. Similar results are obtained
when repeating the same exercise for the variance, although in this case the slope coefficients,
increasing with τ from 12.1 to 21.8, reflect a (less than proportional) increase in the variability of
Q
Vart,τ with the time-to-maturity. The PCA for the third and fourth moments shows qualitatively
the same results, up to higher orders of magnitude. We also test whether and to what extent the
high degree of correlation between the four moments, as illustrated in Figure 7, is due to the
presence of a common factor. We find that the first principal component explains nearly 93% of
the total covariation between the 20 variables. This result provides strong evidence of a main
driver for the RND of VIX over time.

6.3 VIX jumps under Q

The peculiar proportional structure of the moments of the VIX under risk neutrality suggests
that there might exist a multiplicative error model, MEM, driving the dynamics of future states
of the VIX over time under the Q-measure. The MEM model is defined as

X t = µ t ηt , t = 1, . . . ,T , (28)

22
where, for a fixed maturity τ , Xt = VIXt+τ , µt is the expectation of VIXt+τ computed at the time
t, ηt is an i.i.d. stochastic term, independent of µt , with positive support and with mean equal to
1, and T is the sample size. MEM type dynamics for the VIX index are also studied in Mencia and
Sentana (2016), who adopt a Laguerre expansion for the innovation term ηt under the physical
measure P. They also analyze the process under Q, using the VIX futures and assuming an
exponentially affine stochastic discount factor to model the risk premium. In this way, they
retrieve information on the risk-neutral higher moments of VIX by a structural assumption to
link P and Q. We instead perform a direct analysis of the distributional assumptions on the VIX
under Q, exploiting higher risk-neutral moments obtained from the VIX options. The idea of
using transformations of option prices to back out estimates of a parametric model is in line
with the methodology adopted in Pastorello et al. (2000) and Pan (2002), among others.
By setting ηt ∼ i.i.d.Γ(1, ν ) (mean-shape form) in (28), which corresponds to a zero-order
Laguerre expansion we obtain
1 
EtQ [VIXt+τ ] = µt , EtQ [VIXt+τ
2
] = µt2 +1 ,
   ν6
3 2 3 
Et [VIXt+τ ] = µt 2 + + 1 ,
Q 3
EtQ [VIXt+τ
4
] = µt4 +
11 6
+ + .
ν ν ν3 ν2 ν
1

Under the MEM-Gamma the proportionality between second, third and fourth moment is con-
trolled by a single parameter, ν . The MEM-Gamma can be augmented with jumps, thus obtaining
the MEM-J of Caporin et al. (2017), defined as

X t = µ t Z t ηt , t = 1, . . . ,T , (29)

where Zt is the VIX jump component, and the innovation ηt ∼ i.i.d.Γ (1, ν ). The jump term, Zt
is defined as
dλ

 Nt = 0
Zt =  (30)
 PNt Yj,t Nt > 0

 j=1
  −1
where Nt is a Poisson random variable with intensity λ > 0 and dλ = e −λ + λ is a scalar
positive function of λ, denoting the baseline value of Zt in absence of jumps. When Nt > 0, the
process Zt is a compound Poisson process. Following Caporin et al. (2017), we assume that the
jump sizes are driven by a Gamma distribution Γ (dλ , ς ) (in mean-shape form), from which it
follows that the innovation term ξt = Zt ηt is driven by a countably infinite mixture of Gamma

23
 MEM-Gamma MEM-J
ν̂ χ 2 (2) ν̂ λ̂ χ 2 (1)
τ =1 26.38a 0.000 55.79a 0.351a 0.053
τ =2 12.87a 0.000 47.33a 0.513a 0.200
τ =3 12.65a 0.000 41.04a 0.591a 0.161
τ =4 9.433a 0.000 26.65a 0.585a 0.117
τ =5 7.411a 0.000 19.66a 0.563a 0.054

Table 1: GMM estimates of the parameters of the MEM-Gamma and the MEM-J. The estimation is per-
formed by matching the parametric expressions of EtQ [VIXt2+τ ], EtQ [VIXt3+τ ] and EtQ [VIXt4+τ ] outlined
above with their empirical counterpart obtained from the RNDs retrieved from the option prices. The
first moment is exactly matched, i.e. Meant,τ Q
:= EtQ [VIXt +τ ] = µ t , by construction, and it is used as a
driver for the dynamics of the higher order moments. The subscripts a, b and c stand for significance at
1%, 5% and 10%, respectively. χ 2 (2) and χ 2 (1) are the p-values of the J-test for over-identifying restric-
tions whose distribution is χ 2 (m − r ), where m = 3 is the number of moment conditions adopted in the
estimation and r is the number of free parameters.

and Kappa distributions with closed-form moments given by

EtQ [VIXt,τ ] = µ t ,
2 λ
"  #  
Et [VIXt,τ ] = µ t
Q 2
+ e + λ + λ d λ2 1 + ν −1 ,
−λ 2
ς

3  −λ * λ
2d 3 d λ3 
Et [VIXt,τ ] = µ t e
Q 3
+ 3 + d λ + (λ + 3λ + λ)d λ ς Cν,ς + 3(λ + λ)d λ ςCν,ς + 2λd λCν,ς  ,
3+ 3 2 3 2 2 3 3
 , ν 2 ν - 
 d4 d4 d4 
EtQ [VIXt,τ
4
] = µ t4 e −λ *6 λ3 + 11 λ2 + 6 λ + d λ4 + + (λ4 + 6λ3 + 7λ2 + λ)d λ4 ς 3 D ν,ς + 6(λ3 + 3λ2 + λ)d λ4 ς 2 D ν,ς 
 , ν ν ν - 
 
+ µ t4 11(λ2 + λ)d λ4 ςD ν,ς + 6λd λ4 D ν,ς ,
(31)

where
ν 2 + 3ν + 2 ν 3 + 6ν 2 + 11ν + 6
Cν,ς = , Dν,ς = .
ν 2ς 2 ν 3ς 3
The MEM-J generates a mixture of distributions conceptually similar to the mixture generated
by the Laguerre expansion used in Mencia and Sentana (2016), but it has a further interpretation
for the tail-events as generated by jumps.
For every τ = 1, . . . , 5, based on the parametric expressions (31), we estimate the MEM-
Gamma and the MEM-J models by GMM exploiting the time-series of the first four risk-neutral
moments of the VIX. The GMM estimates of the parameters are reported in Table 1. The es-
timates of the parameter ν obtained under the Gamma distribution are always significant and
they decrease with maturity. Since ν is the inverse of the variance of ηt in the Gamma specifica-
tion, this evidence correctly signals the increase in the uncertainty around future values of the
VIX at longer horizons. However, the MEM with Gamma distributed innovations does a poor
job matching the structure of the higher moments of the VIX under the Q-measure. Indeed, the

24
 MEM-J MEM-Gamma MEM-J MEM-Gamma

0.16 0.16

0.14 0.14

0.12 0.12

0.1 0.1

0.08 0.08

0.06 0.06

0.04 0.04

0.02 0.02

0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60

Figure 8: Steady-state risk neutral densities of VIXτ implied by the MEM-Gamma model and by the
MEM-J model of Caporin et al. (2017) for τ = 1 (upper panel) and τ = 5 (lower panel).

J-test for over-identifying restrictions, which is distributed as a χ 2 with 2 degrees of freedom,

strongly rejects the null hypothesis that the higher moments are those implied by a Gamma
distribution. Instead, adding flexibility to the higher moments of VIX by adding the jump term,
as in the MEM-J specification, provides a good description of the linkages between the VIX risk-
neutral moments, since the p-value of the J-test is larger than 5% for every τ = 1, . . . , 5. Looking
at the parameter estimates, we note that ν̂ decreases with maturity for both MEM specifications,
thus signaling an increase of the variance at longer horizons. In the MEM-J, the mixing parame-
ter λ, which governs the expected number of jumps in each month, is rather stable across τ and
close to 0.5. This means an average of 1 jump every second month. Notably, the parameter λ is
always significant, thus indicating that tail events, namely those generated by jumps, are priced
in the VIX options. A graphical illustration of contribution of jumps on the probability of tail
events is provided in Figure 8. The figure reports the model-implied RND of the VIX under Q in
”steady state” that is obtained by setting µt to its long-run value represented by its sample mean,
µ̄ = T1 Tt=1 Meant,τ
Q
P
. In steady state, the option-implied probability of observing VIX larger than
31% is negligible under the MEM-Gamma for τ = 1, while it is approximately 4.5% if jumps are
included in the model. Analogous evidence emerges for τ = 5.

6.4 Variance swap term-structure

In this paragraph, we analyze the term structure of the (annualized percentage) realized variance
(RV) in terms of variance swaps (VS). As discussed in Section 2, the square of VIX can be seen
as the swap rate of a VS maturing in one month, that is VSt,t+1 = EtQ [RVt,t+1 ] = VIXt2 , and we
assume zero interest rate, which is indeed negligible in the sample under investigation. This
relationship can be generalized to link 1-month forward VS prices to the second moment of the
VIX, as shown below
EtQ [RVt+τ ,t+τ +1 ] = EtQ [VIXt+τ
2
]. (32)

25
Equation (32) clearly generalizes the definition of VIX, which is recovered for τ = 0. By aggre-
gating the terms on the RHS of (32) over τ one can obtain variance swap prices for maturities
longer than one month, that is
n n
1X Q 1X Q
VSt,t+n = EtQ [RVt,t+n ] = Et [RVt+τ −1,t+τ ] = Et [VIXt+τ
2
−1 ]. (33)
n τ =1 n τ =1

The relation in (33) holds under the inherent assumption that the joint market of VIX and RV is
priced consistently under the unique risk neutral measure Q, so that
f Q g
EtQ [VIXt+τ
2
] = EtQ Et+τ [RVt+τ ,t+τ +1 ] = EtQ [RVt+τ ,t+τ +1 ] .

In principle, by generalizing the VIX formula (3), one can always replicate VSt,τ through a port-
folio of SPX options expiring in n months. This provides a tool to assess whether the SPX and
the VIX markets are consistent with each other. In Figure 9 we report the time series of VSt,t+τ
for τ = 3 months (left panel) and τ = 4 months (right panel). Each plot displays the VS com-
puted from SPX options by extending formula (3) (red line) and the VS implied by the RND of
VIX through (33) (blue line). Figure 9 suggests that there are no profitable arbitrage opportu-
nities based on trading VS across the SPX and the VIX markets. In particular, this means that
the second risk-neutral moment of the VIX can be regarded as an equity derivative, as it can be
replicated by combining short and long positions on SPX options expiring in τ + n − 1 and τ + n
months, respectively.

35 35
Implied by term-structure of VIX2 Implied by term-structure of VIX2
Implied by SPX options Implied by SPX options

30 30

25 25

20 20

15 15

2011 2012 2013 2014 2015 2016 2011 2012 2013 2014 2015 2016

(a) τ = 3 (b) τ = 4

Figure 9: Time series of 3-months and 4-months VS. The figures display VS implied by SPX market (red
line) and the VS implied by the VIX market (blue line).

Based on the time series of EtQ [VIXt+τ

2
], we characterize the driving factors of the VS term-
structure. As starting point we consider the simple model below

yt,τ = e κ (τ +1) xt (34)

26
 τ =0 τ =1 τ =2 τ =3 τ =4 τ =5
κ̂ 0.3130 0.2795 0.2447 0.2110 0.1835 0.1627
S.E.(κ) 0.0676 0.0342 0.0247 0.0194 0.0162 0.0139
t-test 4.6303a 8.1650a 9.9131a 10.8702a 11.3064a 11.7156a

Table 2: Estimate of the time-to-maturity compounding factor. The standard errors are computed with
the Newey-West robust estimator. a, b and c stand for significance at 1%, 5%, and 10%, respectively. For
τ = 0, EtQ [VIXt2+τ ] is VIXt2 .

where yt,τ = EtQ [RVt+τ ,t+τ +1 ], xt = RVt−1,t . This model represents a spot-future parity, where κ
is a time-to-maturity compounding factor, which determines the slope of the term-structure of
VS.
For each given τ , the parameter κ can be estimated by a simple regression in logs as

ζt,τ = κ (τ + 1) + ut,τ , t = 1, . . . T (35)

where ζt,τ = log(EtQ [VIXt+τ

2
]) − log(RVt−1,t ), with observations on RVt−1,t constructed by cu-
mulating daily RV over the 30 days period between t − 1 and t. The daily RV is constructed
from 5 minute returns on SPX. The estimation results are reported in Table 2. The value of κ
estimated for τ = 0 is twice as large as the value of κ estimated for τ = 5, thus signaling a
rather steep downward sloped term structure of VS. The decreasing behavior of the estimates of
k signals that model (34) does not provide an accurate description of the term structure of VS,
as it prescribes a constant positive drift in RV under Q.
We therefore suggest a generalization of (34) by including a term accounting for the mean-
reversion of RV, that is
yt,τ = e κ (τ +1) xt · mrt (36)

where mrt is a multiplicative term that adjusts for the mean-reversion. We assume that mrt is a
function of the ratio between RVt,t−1 and VIXt , and it determines the level of the term-structure
of VS relative to the current level of the underlying. The econometric specification becomes

ζt,τ = κ (τ + 1) + αξt + ut , t = 1, . . . T (37)

where ξt = log(RVt−1,t ) − log(VIXt2 ) can be interpreted as the residuals from the long-run equi-
librium between RV and VIX2 . This term is mean-reverting due to the cointegration relationship
between the two quantities, see Bollerslev et al. (2013) among others. Table 3 shows that the
correction based on the mean-reversion makes the parameter κ much more stable across τ ,
and approximately equal to 13.8%. We can interpret this number as the monthly growth rate,
adjusted by mean-reversion, of the VS. Similarly, Johnson (2012) finds that the term structure
implicit in VIX futures cannot be well described by an unique slope factor, but another state-
dependent term is needed to provide a good approximation of the surface.

27
 τ =1 τ =2 τ =3 τ =4 τ =5
κ̂ 0.1388 0.1547 0.1461 0.1324 0.1204
S.E.(κ) 0.0105 0.0118 0.0107 0.0095 0.0086
t-test(κ) 13.2102 13.0934 13.6426 13.9938 14.0555

Table 3: Estimate of the time-to-maturity compounding factor. The standard errors are computed with
the Newey-West robust estimator. a, b and c stand for significance at 1%, 5% and 10%, respectively.

7 Conclusion and directions for future research

In this paper, we proposed a methodology based on a finite orthogonal expansion to infer the
RND underlying the VIX option prices. The method generalizes the Laguerre expansions and
suits cases where the density is supported over the positive real axis. The approach is non-
structural since it does not require restrictive parametric assumptions on the underlying asset
dynamics, reducing the number of restrictions to be imposed on the form of the RND. Therefore,
we drastically reduce the intrinsic risk of misspecification entailed in parametric models. While
in our applications we addressed the RND underlying VIX options, the same technique can
be applied to different classes of financial derivatives sharing the same characteristics, e.g., to
interest rates and inflation. Our empirical study on VIX options highlights the usefulness of this
technique to directly study several features of the VIX through its RND and associated moments.
The empirical analysis suggests that the proposed methodology may be particularly use-
ful when the study of risk-premia embedded in option prices is of interest, see Bollerslev and
Todorov (2011), Andersen et al. (2015b) and Schneider (2015), or to compute a VaR on volatility
(VolaR, see Caporin et al., 2017), adjusted for risk aversion as in Aı̈t-Sahalia and Lo (2000). A
multivariate functional dynamic model for the RND would be a further natural extension of this
work, see Grith et al. (2013). This would provide an alternative to the parametric methodology
of Andersen et al. (2015a) to carry out inference on the underlying processes based on panels of
options. Along these lines, the study of the risk premia embedded in VIX options could be car-
ried out by studying the shape and the variability over time of the pricing kernel in a bivariate
setting that also includes the SPX. This would further extend the work of Song and Xiu (2016) to
potentially assess the risk aversion parameter of investors using the VIX for hedging purposes
as opposed to those taking speculative positions on it.

References
Abramowitz, M. and Stegun, I. A. (1964). Handbook of mathematical functions with formulas,
graphs, and mathematical tables. Dover, New York.
Aı̈t-Sahalia, Y. and Duarte, J. (2003). Nonparametric option pricing under shape restrictions.
Journal of Econometrics, 116(1-2):9–47.
Aı̈t-Sahalia, Y. and Lo, A. W. (1998). Nonparametric estimation of state-price densities implicit
in financial asset prices. The Journal of Finance, 53(2):499–547.

28
Aı̈t-Sahalia, Y. and Lo, A. W. (2000). Nonparametric risk management and implied risk aversion.
Journal of Econometrics, 94(1fi?!2):9 – 51.
Andersen, T. G., Fusari, N., and Todorov, V. (2015a). Parametric inference and dynamic state
recovery from option panels. Econometrica, 83(3):1081–1145.
Andersen, T. G., Fusari, N., and Todorov, V. (2015b). The risk premia embedded in index options.
Journal of Financial Economics, 117(3):558 – 584.
Bardgett, C., Gourier, E., and Leippold, M. (2014). Inferring volatility dynamics and risk premia
from the S&P 500 and VIX markets. Technical report, Swiss Finance Institute Research Paper
No. 13-40.
Bayer, C., Gatheral, J., and Karlsmark, M. (2013). Fast Ninomiya–Victoir calibration of the
double-mean-reverting model. Quantitative Finance, 13(11):1813–1829.
Bergomi, L. (2008). Smile dynamics III. Risk, October:90–96.
Bollerslev, T., Osterrieder, D., Sizova, N., and Tauchen, G. (2013). Risk and return: Long-run
relations, fractional cointegration, and return predictability. Journal of Financial Economics,
108(2):409 – 424.
Bollerslev, T. and Todorov, V. (2011). Tails, fears and risk premia. The Journal of Finance,
66(6):2165–2211.
Breeden, D. T. and Litzenberger, R. H. (1978). Prices of state-contingent claims implicit in option
prices. The Journal of Business, 51(4):621–51.
Brigo, D. and Mercurio, F. (2002). Lognormal-mixture dynamics and calibration to market
volatility smiles. International Journal of Theoretical and Applied Finance, 5(4):427–446.
Britten-Jones, M. and Neuberger, A. (2000). Option prices, implied price processes, and stochas-
tic volatility. The Journal of Finance, 55(2):839–866.
Buehler, H. (2006). Consistent variance curve models. Finance and Stochastics, 10(2):178–203.
Caporin, M., Rossi, E., and Santucci de Magistris, P. (2017). Chasing volatility - a persistent mul-
tiplicative error model with jumps. Technical report, Forthcoming on the Journal of Econo-
metrics.
Carr, P. and Lee, R. (2007). Realized volatility and variance: Options via swaps. Risk, 20(5):76–83.
Carr, P. and Lee, R. (2009). Volatility derivatives. Annual Review of Financial Economics, pages
14.1–14.21.
Carr, P. and Madan, D. (1998). Towards a theory of volatility trading. In Volatility: New estimation
techniques for pricing derivatives, ed. R. Jarrow, chap. 29, pages 417–427. Risk Publications.
Carr, P. and Wu, L. (2006). A tale of two indices. The Journal of Derivatives, 13(3):13–29.
Carr, P. and Wu, L. (2009). Variance risk premiums. Review of Financial Studies, 22(3):1311–1341.
CBOE (2015). The CBOE volatility index VIX. White Paper. Available at .
Christoffersen, P., Jacobs, K., and Mimouni, K. (2010). Volatility dynamics for the S&P500: ev-
idence from realized volatility, daily returns, and option prices. Review of Financial Studies,
23(8):3141–3189.
Cont, R. (2006). Model uncertainty and its impact on the pricing of derivative instruments.
Mathematical Finance, 16(3):519–547.
Cont, R. and Kokholm, T. (2013). A consistent pricing model for index options and volatility
derivatives. Mathematical Finance, 23(2):248–274.

29
Corrado, C. J. and Su, T. (1996a). Skewness and kurtosis in S&P 500 index returns implied by
option prices. Journal of Financial Research, 19(2):175–192.
Corrado, C. J. and Su, T. (1996b). S&P 500 index option tests of Jarrow and Rudd’s approximate
option valuation formula. Journal of Futures Markets, 16(6):611–629.
Coutant, S., Jondeau, E., and Rockinger, M. (2001). Reading PIBOR futures options smiles: The
1997 snap election. Journal of Banking & Finance, 25(11):1957–1987.
Duffie, D., Pan, J., and Singleton, K. (2000). Transform analysis and asset pricing for affine
jump-diffusions. Econometrica, 68(6):1343–1376.
Filipovic, D., Mayerhofer, E., and Schneider, P. (2013). Density approximations for multivariate
affine jump-diffusion processes. Journal of Econometrics, 176(2):93 – 111.
Gatheral, J. (2008). Consistent modeling of SPX and VIX options. In Presentation at the Fifth
World Congress of the Bachelier Finance.
Grith, M., Hardle, W., and Park, J. (2013). Shape invariant modeling of pricing kernels and risk
aversion. Journal of Financial Econometrics, 11(2):370.
Hazewinkel, M. (1988). Encyclopaedia of mathematics: C An updated and annotated translation
of the Soviet ’Mathematical Encyclopaedia’. Springer.
Hewitt, E. (1954). Remark on orthonormal sets in L2(a, b). The American Mathematical Monthly,
61(4):249–250.
Huskaj, B. and Nossman, M. (2013). A term structure model for VIX futures. Journal of Futures
Markets, 33(5):421–442.
Jarrow, R. and Rudd, A. (1982). Approximate option valuation for arbitrary stochastic processes.
Journal of Financial Economics, 10(3):347–369.
Jiang, G. J. and Tian, Y. S. (2005). The model-free implied volatility and its information content.
Review of Financial Studies, 18(4):1305–1342.
Johnson, T. L. (2012). Equity risk premia and the VIX term structure. Technical report, Forth-
coming on the Journal of Financial and Quantitative Analysis.
Jondeau, E. and Rockinger, M. (2001). Gram-Charlier densities. Journal of Economic Dynamics
and Control, 25(10):1457–1483.
Lee, R. and Wang, D. (2009). Displaced lognormal volatility skews: analysis and applications to
stochastic volatility simulations. Annals of Finance, 8(2):159–181.
Madan, D. B. and Milne, F. (1994). Contingent claims valued and hedged by pricing and investing
in a basis. Mathematical Finance, 4(3):223–245.
Mencia, J. and Sentana, E. (2013). Valuation of VIX derivatives. Journal of Financial Economics,
108(2):367 – 391.
Mencia, J. and Sentana, E. (2016). Volatility-related exchange traded assets: an econometric in-
vestigation. Technical report, Forthcoming on the Journal of Business & Economic Statistics.
˜
Nı́guez, T.-M. and Perote, J. (2012). Forecasting heavy-tailed densities with positive Edgeworth
and Gram-Charlier expansions. Oxford Bulletin of Economics and Statistics, 74(4):600–627.
Pan, J. (2002). The jump-risk premia implicit in options: evidence from an integrated time-series
study. Journal of Financial Economics, 63(1):3 – 50.
Pastorello, S., Renault, E., and Touzi, N. (2000). Statistical inference for random-variance option
pricing. Journal of Business & Economic Statistics, 18(3):358–367.

30
Rompolis, L. S. and Tzavalis, E. (2008). Recovering risk neutral densities from option prices: a
new approach. Journal of Financial and Quantitative Analysis, 43(04):1037–1053.
Rudin, W. (1987). Real and complex analysis. Tata McGraw-Hill Education.
Schneider, P. (2015). Generalized risk premia. Journal of Financial Economics, 116(3):487 – 504.
Sepp, A. (2008a). Pricing options on realized variance in the Heston model with jumps in returns
and volatility. Journal of Computational Finance, 11(4):33–70.
Sepp, A. (2008b). VIX options pricing in a jump-diffusion model. Risk, pages 84–89.
Shohat, J. (1942). Note on closure for orthogonal polynomials. Bulletin of the American Mathe-
matical Society, 48(6):488–490.
Song, Z. and Xiu, D. (2016). A tale of two option markets: pricing kernels and volatility risk.
Journal of Econometrics, 90:176–196.
Szegö, G. (1939). Orthogonal polynomials. American Mathematical Society.
Todorov, V. and Tauchen, G. (2011). Volatility jumps. Journal of Business & Economic Statistics,
29(3):356–371.
Todorov, V., Tauchen, G., and Grynkiv, I. (2014). Volatility activity: Specification and estimation.
Journal of Econometrics, 178, Part 1:180 – 193.
Wang, Z. and Daigler, R. T. (2011). The performance of VIX option pricing models: Empirical
evidence beyond simulation. Journal of Futures Markets, 31(3):251–281.
Xiu, D. (2014). Hermite polynomial based expansion of European option prices. Journal of
Econometrics, 179(2):158–177.
Zhang, J. E. and Zhu, Y. (2006). VIX futures. Journal of Futures Markets, 26(6):521–531.
Zhang, L., Mykland, P. A., and Aı̈t-Sahalia, Y. (2011). Edgeworth expansions for realized volatil-
ity and related estimators. Journal of Econometrics, 160(1):190–203.
Zhu, Y. and Zhang, J. E. (2007). Variance term structure and VIX futures pricing. International
Journal of Theoretical and Applied Finance, 10(01):111–127.

31
A Appendix

A.1 Proof of Theorem 3.1

Some preliminary are stated before the main proof. First, we recall a standard result of functional
analysis. For a proof the reader may refer, e.g., to Rudin (1987)-Theorem 4.14.

Lemma A.1. Assume that ϕ − 2 f Q ∈ L2 (D) and supp( f Q ) ⊆ D. Consider the Hilbert space
1

(Hϕ , h· , ·i) defined by

Z
( 1 ) 1
Hϕ = ψ , ϕ − 2 ψ ∈ L2 (D) , ψ 1 ,ψ 2 := ψ 1 (x )ψ 2 (x ) dx , ∀ψ 1 ,ψ 2 ∈ Hϕ ,


D ϕ (x )

and the subspace

 ( ϕ )
Hϕ∗ = Cl span ϕhk , k ∈ N ⊆ Hϕ .

Then, there exists a sequence (ck )k∈N such that function

kn
ϕ
f Q(∞)
X
:= lim ϕ .1 +
* ck hk +/ in Hϕ
n→+∞
, k=1 -

solves the minimum distance problem

E1
f Q(∞) = argmin ψ − f Q ,ψ − f Q 2 .
D
(38)
ψ ∈Hϕ∗

In particular, if Hϕ∗ = Hϕ we have f Q(∞) = f Q almost everywhere.

Definition A.2 (Closed polynomial set in Hϕ ). The kernel ϕ is said to generate closed polyno-
mial sets if
Cl span x k , k ∈ N = Lϕ2 (x )dx (D).
 ( )
(39)
ϕ
In this case, we say that either (x k )k∈N or (hk )k∈N is closed with respect to ϕ.

The following result provides necessary and sufficient conditions to determine whether ϕ
generates closed polynomial sets. The results, whose proof is deferred to the end of the section,
extends the classic result of closure of Laguerre polynomials.
ϕ
Theorem A.3 (Conditions to the closure of (hk )k∈N ). Let ϕ be a positive integrable function and
D = [0, +∞[.
1
(i) If limx→+∞ ϕ (x )e ςx 2 = 0 for some ς > 0 and there exists a polynomial p such that pϕ is
bounded, then ϕ generates closed polynomial sets.

32
 1 −γ
(ii) If limx→+∞ ϕ (x )e ςx 2 > 0 for some γ , ς > 0, then ϕ does not generate closed polynomial
sets.

We can now prove Theorem 3.1. In the following, Hϕ and Hϕ∗ refer to the Hilbert spaces
defined in Lemma A.1.
ϕ
If (hk )k∈N is closed with respect to ϕ, then Hϕ = Hϕ∗ and from Lemma A.1 it follows that
f Q∗ = f Q whenever ϕ − 2 f Q ∈ L2 (D). The first implication can be readily shown by noticing that
1

ϕ
f Q ∈ Hϕ implies ϕ −1 f Q ∈ Lϕ2 (x )dx (D). Then, the closure of (hk )k∈N implies that ϕ −1 f Q can be
approximated by a certain polynomial series a 0 + a 1x + a 2x 2 + . . . in Lϕ2 (x )dx (D) or equivalently
ϕ ϕ ϕ
that f Q can be approximated by a certain series c 0ϕh 0 + c 1ϕh 1 + c 2ϕh 2 . . . in Hϕ .
Then, (a) follows immediately from Lemma A.1 by noticing that the assumptions on ϕ imply
Hϕ = Hϕ∗ , in view of Theorem A.3-(i). To prove (13), we observe that for every ϕ ∈ Hϕ∗ and
every n ∈ N

Z +∞ (n)
Z +∞
(∞)
Z +∞
Π(x )  f Q(n) (x ) − f Q(∞) (x )  dx

 Π(x ) f Q (x )dx − Π(x ) f Q (x )dx  ≤
 0Z 0  0
E1
ϕ 2 (x )Π(x ) ϕ − 2 (x ) f Q (x ) − ϕ − 2 (x ) f Q (x )  dx ≤ b · f Q(n) − f Q(∞) , f Q(n) − f Q(∞) 2 ,
(n) (∞)
1 1 1 D
=  
D

D 1 E1
where b = ϕ 2 Π, ϕ 2 Π 2 is finite by hypothesis. Then, (13) follows from Lemma A.1.
1

Proof of Theorem A.3

The following lemma is needed to prove one part of the theorem.

Lemma A.4. Suppose that ϕ ∗ generates closed polynomial sets and ϕ = h · ϕ ∗ , where h is bounded
and positive a.e. on D. Then ϕ generates closed polynomial sets.

Proof. By the Riesz-Fischer characterization it suffices to prove that if there exists f ∈ Lϕ2 (x )dx (D)
such that Z
f (x )x k ϕ (x )dx = 0 ∀k ∈ N,
D

then it must hold that f (x ) = 0 a.e. on D. Define д(x ) = h(x ) f (x ), then

Z Z
д (x )ϕ (x )dx ≤ max h(x ) ·
2 ∗
f 2 (x )ϕ (x )dx < +∞
D x ∈D D

which proves д ∈ Lϕ2 ∗ (x )dx (D). Furthermore

Z Z
k ∗
д(x )x ϕ (x )dx = f (x )x k ϕ (x )dx = 0
D D

for every k ∈ N, which implies in view of hypothesis that д(x ) = 0 a.e. on D and therefore
f (x ) = 0 a.e. on D due to positivity assumptions on h(x ). 

33
 To prove statement (i) we start by recalling a classic result due to Hewitt (1954) showing that
every bounded function ψ supported on the entire real line and such that

lim ψ (x )e ς |x | = 0 (40)
|x |→+∞

generates closed polynomial sets. Based on this result, statement (i) can be proven under the
additional hypothesis that ϕ is bounded. Indeed, under this assumption, the function ψ (x ) =
|x |ϕ (|x | 2 ) is bounded on R and satisfies (40), and therefore it generates closed polynomial sets.
1

Statement (i) is then a straightforward consequence of the main theorem reported in Shohat
(1942). To prove statement (i) with no additional requirements on ϕ, we remark that by hypoth-
esis there exist a polynomial p and ς ∗ > 0 such that the function ϕ ∗ defined by

∗ √x
ϕ ∗ (x ) := p(x )e ς ϕ (x )

is bounded on D. Since ϕ ∗ clearly preserves the same integrability and asymptotic properties
of ϕ, then it generates closed polynomial sets. Now, consider f such that f 2ϕ is integrable and
Z
f (x )x k ϕ (x )dx = 0, ∀k ∈ N.
D

Moreover, define д as
∗ √x
д(x ) = e −ς f (x ), x ∈ D.

We have
∗√
Z Z
д(x ) ϕ (x )dx ≤ sup p(x )e −ς x 
2 ∗
f 2 (x )ϕ (x )dx < +∞.
D x ∈D D

On the other hand, for every k ∈ N

Z Z
k ∗
д(x )x ϕ (x )dx = f (x )x k p(x )ϕ (x )dx = 0,
D D

which proves д(x ) = 0 and therefore f (x ) = 0 a.e. on D. Then, statement (i) is proved.
The proof of statement (ii) is based on a known counterexample in the theory of orthogonal
polynomials (cf. entry ”Closed system of elements” in Hazewinkel (1988)), showing that every
function ψ of the form
2m
ψ (x ) = e −|x | , x ∈ R, m ∈ N ,
2m+1

does not generate closed polynomial sets. By combining this counterexample with the results
of Shohat (1942), we can prove that the function ψ supported on [0, +∞[ and defined by
m
ψ (x ) = e −x , x ≥ 0, m ∈ N.
2m+1

34
does not generate closed polynomial sets. By a change of variable and through the Riesz-Fischer
characterization, one can extend the latter result to the case where ψ is supported on [x 0 , +∞[
and is of the form
m
ψ (x ) = e −ς (x−x 0 ) , x ≥ x 0 , m ∈ N.
2m+1

for some ς > 0 and x 0 ≥ 0. To prove statement (ii), then, we proceed by contradiction
and suppose that there exists an integrable function ϕ, supported on [0, +∞[ and such that
1 −γ
limx→+∞ ϕ (x )e ςx 2 > 0 for some γ , ς > 0, which generates closed polynomial sets. To this aim,
we observe that by the hypothesis made on the right-tail of ϕ, there exists x 0 ≥ 0 such that
ϕ (x ) > 0 for all x ≥ x 0 . The closure property of polynomial sets with respect to ϕ holds in
particular when the support is restricted, by truncation, to [x 0 , +∞[. Furthermore, the function
h defined by
m
h(x ) = e −ς (x−x 0 ) ϕ (x ) −1 ,
2m+1

is bounded on [x 0 , +∞[, for m sufficiently large. Then, as a consequence of Lemma A.4, the
m
function e −ς (x−x 0 ) 2m+1 generates closed polynomial sets on [x 0 , +∞[, which is a contradiction.
The proof is thereby concluded.


A.2 Proof of Proposition 5.1

For the notational simplicity, throughout the proof we omit the dependence on t, τ of Q (n) , C Obs ,
P Obs , C, and P. We also set r = 0. Moreover, we denote by (ck )k∈N and f Q(n) the quantities defined
in Theorem 3.1-(a). For every n ∈ N we have
"Z  2 #
(n) ? Obs (n) ? Obs (n) ?
f g 2 
(KM − K 1 ) · E Q (c (n)) = E CK − CK (c (n)) + PK − PK (c (n)) dK
I
"Z  2 #
C (n) P (n)
2 
≤E CK + ϵK − CK (c) + PK + ϵK − PK (c) dK
I
"Z   #
(n) C C (n)
2  2 
=E CK − CK (c) + ϵK + 2ϵK CK − CK (c) dK
I
"Z   #
(n) P P (n)
2  2 
+E PK − PK (c) + ϵK + 2ϵK PK − PK (c) dK
I
Z  "Z    2 #
(n) (n)
ϵK + ϵKP dK .
C 2
2  2
= CK − CK (c) + PK − PK (c) dK + E
I I

35
Then, by (4), we get
Z 
CK − CK(n) (c) + PK − PK(n) (c) dK
2  2
I
Z "Z  # 2 "Z #2
(n) +
f Q(n) (x ) +
  
= f Q (x ) − f Q (x ) (K − x ) dx + f Q (x ) − (x − K ) dx dK
I D D
Z "Z #2 Z
≤  f (x ) − f (n) (x )  |K − x |dx dK ≤ b ·  f (x ) − f (n) (x ) 2 1 dx ,
I D
 Q Q  D
 Q Q  ϕ (x )

+∞
where b = (K − x ) 2ϕ (x )dxdK is finite. Then, in view of Theorem 3.1
R R
I 0
Z 
CK − CK(n) (c) + PK − PK(n) (c) dK = 0 ,
2  2
lim
n→+∞ I

which proves (24). The proof is concluded by noticing that, under the additional hypotheses (i)
and (ii) "Z  
1  2 #
(n) ?
ϵK + ϵKP dK ,
C 2
f g
E Q (c (n)) ≥ E ∀n ≥ n̄ .
KM − K1 I

36
B Supplementary material

B.1 A Robust Technique

B.1.1 Orthogonal regressors

We propose to solve the problem of multicollinearity outlined in Section 4 by means of a prin-

cipal component analysis (PCA), which allows to estimate the coefficients of an expansion of
any arbitrarily large order n. The PCA analysis is implemented as follows: first, to avoid scale
effects, we standardize each column of X, as
1 P2M
Xi − j=1 Xji
Zi = q 2 , i = 1, . . . , n
2M
 (41)
1 P2M 1 P2M
2M−1 j=1 Xji − 2M j=1 Xji

where Xi and Xji denote the i-th column and the j, i-th element of X, respectively. Then, we
determine the 2M × n matrix of principal components as V = PZ and the n × n orthonormal
matrix of weights P from the spectral decomposition ZP = PΛ. Lastly, we extract the sub-matrix
Ṽ = V·,1:s of the first s principal components, associated with a given threshold on the explained
total variance (e.g. 99%), to be used as regressor. For example, when n > 10, the first 4-5 principal
components typically explain at least the 99% of the total variance.
Once we have obtained V, we estimate the coefficients γ̂ = (γ̂ 1 , . . . , γ̂s ) of the following
regression,
Y∗ = Ṽγ + u, (42)

where γ represents the loading on the first s principal components. The estimated coefficients
c̃ are finally retrieved by reverting the orthogonalization as follows

s s  0
2M − 1 2M − 1
c̃ = (Oγ̂ ) ◦  P2M
 , . . . , P2M  , (43)
 j=1 (X j1 − 1 P2M
2M X
j=1 j1 ) 2
j=1 (X jn − 1 P2M
2M X )
j=1 jn 
2 

where O is the n × s matrix obtained from the first s columns of P and ◦ denotes the Hadamard
product.

B.1.2 Regression through the origin

In regression (42), there is no intercept and the columns of Ṽ have zero-mean by construction,
while Y∗ has zero mean if and only if the sample mean of Y and X0 coincide. To enforce that
E (Y∗ ) = E (Y − X0 ) = 0, the initial estimation of the kernel parameters, θ , must be constrained
such that E (X0 ) = E (Y). This represents a very mild constraint but it has several practical
advantages. First, it ensures that the approximation of order 0 does not produce systematic
mispricing, since the observed market prices are centered around the estimated price curve

37
generated by the kernel. Second, the residuals of (42) have zero mean for any order n ≥ 1 by
construction.
Since the principal components are constructed from the standardized regressors Z, when
remapping the solution of (42) onto the√ solution of (20), a constant term equal to ni=1 di Ri
P
1 P2M
j=1 X ji
appears, where R = Oγ̂ and di = P2M
q 2M −1
, i = 1, . . . , n. Therefore, in order to
j=1 (X ji − 2M j=1 X ji )
1 P2M 2

guarantee that the relation in equation (20) holds for any n ≥ 1, the following constrained
optimization is performed

[γ 1 , . . . , γs ] = argmin Q̃ (t,T ; γ 1 , . . . , γs ), (44)

γ 1 ,...,γs
n
X
s.t. di Ri = 0
i=1

where Q̃ (t,T ; γ 1 , . . . , γs ) = (Y − Ṽγ ) 0 (Y − Ṽγ ). This restriction also guarantees that c̃ is such that
there is no systematic pricing error or, in other words, that ε˜ = Y − X0 − Xc̃ are centered around
zero for any n.

B.1.3 Positivity and unitary mass

Since the estimation of c̃ 1 , . . . , c̃n is performed on a finite set of option prices, the estimated RND
f˜ (n) could display significant negative mass even for large values of n. Therefore, we add an
Q
extra implicit constraint to the optimization problem (44). In particular, the optimal parameters
γ 1 , . . . , γs are found by solving the following constrained minimum distance problem

[γ 1 , . . . , γs ] = argmin Q̃ (t,T ; γ 1 , . . . , γs |θ ), (45)

γ 1 ,...,γs
n
X
s.t. di Ri = 0
i=1
Z ∞  n 
ϕ +
X
s.t. 1−∆ pos
< ϕ (x; θ ) *1 + ck (γ )hk  dx < 1 + ∆pos
0 k=1
 , -

where ∆pos = is the tolerance on the unity mass constraint, e.g. ∆pos = 0.000001, while the
coefficients c 1 (γ ), . . . , cn (γ ) are functions of the parameters γ 1 , . . . , γs , determined as in (43).

B.1.4 Kernel displacement

Consistency conditions ensured by Theorem 3.1 are rather flexible with respect to the support
of the kernel. In principle, it is sufficient that the support of the RND is contained in the support
of the kernel. However, if the support of ϕ is too large with respect to the support of f Q , then
the expansion (5) is ”forced” to converge to zero for all points that are outside the support of

38
f Q . This has clear disadvantages from an empirical perspective, since the kernel, which is the
starting point of the optimization in (45), associated with c 1 , . . . , cn = 0, does not satisfy the
constraint of unit mass in supp( f Q ). Even if in general we may assume that supp( f Q ) ⊆ R+ ,
when the left tail of the true RND is particularly short around a point K min > 0, this implies that
that nearly the whole probability mass is concentrated away from the origin. Since the put price
curve of VIX contract normally becomes quickly linear as the strike price approaches the deep
OTM region, the RND is expected to display a strong negative skewness associated with a very
short left tail. Hence, when f Q displays such a behavior on the left tail, it may be convenient to
choose the kernel ϕ so that the following condition is satisfied
Z Kmin
ϕ (x, θ )dx = 0. (46)
0

A simple way to guarantee (46) is to displace the domain of a kernel by Kmin . The idea of using
displaced densities is not new to finance, see e.g. Brigo and Mercurio (2002), and has attracted
particular interest in the context of volatility derivatives, see e.g. Carr and Lee (2007) and Lee and
Wang (2009). The kernel displacement is done by considering a set K ∗ of shifted strikes defined
as K ∗ = [K 1 − Kmin , . . . , KM − Kmin ], and defining the matrix of regressors X with respect to
K ∗ . Once the optimal c̃ are obtained as solution of the of the problem (45) based on K ∗ , then the
estimated RND is determined as follows
n
ϕ
f˜Q(n) (x ) = ϕ (x − Kmin ; θ ) *1 +
X
c̃k hk (x − Kmin ) + . (47)
, k=1 -
Kmin ˜ (n)
f Q (x )dx = 0. The choice of Kmin is based on the analysis of the
R
which guarantees that 0
convexity of deeply OTM put prices, see the discussion in Section 6.

B.2 Numerical illustrations

In this section, we test the accuracy of the proposed approach by means of two numerical exam-
ples under no-arbitrage. The purpose here is to show that the orthogonal polynomials are able
to approximate RNDs, belonging to different families, with a high degree of accuracy. Therefore,
we perform the estimation on option prices generated by structural models for which the RND
is known in closed-form. The option prices that are thereby considered are arbitrage-free by
construction. In this section we illustrate the practical relevance of the asymptotic conditions
on the RND required in Theorem 3.1 to ensure a convergent estimation. To obtain the target
RND and the related option prices, we consider two simple but popular models. In the first case,
the VIX is determined as a function of the instantaneous variance process of the Heston model,
as explained in Zhang and Zhu (2006). In the second case, the RND of the log-VIX is assumed to
be normal inverse Gaussian (NIG), an approach that is adopted in Huskaj and Nossman (2013).

39
In both cases, the estimation of c 1 , . . . , cn is performed according to the methodology outlined in
Section 4 on a set of M = 42 option prices relative to strikes in the interval [K 1 , KM ] = [10, 55].
The option prices to be matched are generated through direct integration of the RND implied
by the two models. The expansions order is set to 20, which is sufficiently high to ensure that
the fitting cannot be further improved by adding more terms to the expansion. Furthermore,
choosing a high order for the expansion illustrates the convergence and stability properties of
the approach. Notably, the true risk-neutral densities associated with the two models display
different decay rates on the tails, thus offering an interesting evaluation on how violating the
conditions of Lemma A.1 may possibly generate divergent expansions. Moreover, this numerical
exercise provides valuable information on the robustness of the estimates to the initial choice
of the expansion kernel. In particular, although the asymptotic properties of ϕ have a tangible
effect on the accuracy of the approximation, the choice of its parameters has only a marginal
impact, provided that it guarantees that convergence and closure conditions are respected (see
Theorem 3.1).

B.3 Heston model

Under Heston dynamics, the undiscounted SPX price (St )t ≥0 and its variance (vt )t ≥0 are gener-
ated according to the following SDE

1 √
d log St = − vt dt + vt dWt ,
2
√
dvt = k (v̄ − vt )dt + η vt dWt∗ ,

where dWt and dWt∗ are correlated Brownian motions with constant correlation ρ. The param-
eters k and v̄ govern the speed of mean reversion and long-run value of vt respectively, while
η is the volatility-of-volatility parameter. Following the approach of Zhang and Zhu (2006), un-
der the Heston model the square of the VIX at time T can be expressed as the following linear
function of vT

1 1 − e kτ 30
VIXT = 100 · (a 1 · vT + a 2 ) 2 , a1 = , a 2 = v̄ (1 − a 1 ), τ = .
kτ 365

Moreover, the density of vT given vt = z has the following closed-from expression

k v̄
− 12 − 2ks √
(vT | vt = z) ∼ д, д(s) = C 1s η 2 e η 2 (1−e −kT ) I 2k v̄ −1 (C 2 s),
η2

√
where C 1 = 2k
η 2 (1−e −k (T −t ) )
and C 2 = 2C 1 e −k (T −t ) z do not depend on the state variable s and Iν
denotes the modified Bessel function of first kind of order ν . Hence, the RND of VIXt is also

40
known in closed-form
x2 − b
!
2
f Q (x ) = x · д
a a
and vanilla options prices can be generated through the integral formulas in (4). The support
√
of f Q is [ a 2 , +∞[ and, by the asymptotic properties of Iν (see for example Abramowitz and
Stegun, 1964), it can be shown that f Q (x ) ∼ x α e −β
∗ ∗ x 2 +γ ∗ x
as x → +∞, where the leading term is
−β ∗x 2 √
clearly e . Moreover, whenever the support of ϕ strictly contains [ a 2 , +∞[, the left-tail de-
cay of f Q does not influence the integrability of f Q2ϕ −1 . Therefore, the condition ϕ 2 f ∈ L2 (D) of
1

Theorem 3.1 is met for any choice of the kernel among the families of GIG, GW and LN densities.
Figure 10 portrays the true RND of the Heston model and the related orthogonal polynomial
expansions based on different choices of the kernel. The approximated densities reported in
Figure 10 highlight the ability of the expansions based on the GIG and the GW kernels to well
recover the original density f Q . On the contrary, the LN kernel fails in approximating f Q , al-
though several corrective terms are considered in the expansion and the convergence condition
f Qϕ −1/2 ∈ L2 (D) is satisfied. The expansion based on the LN kernel proves particularly ineffec-
tual on both tails of f Q . This is a practical consequence of the fact that the LN density does not
generate closed polynomial sets (see Theorem A.3), and may serve as an interpretive example
of the importance of the hypotheses required by Theorem 3.1.

True GIG GW LN
0.05

0.045

0.04

0.035 10 -2

0.03

0.025

0.02 10 -3
0.015

0.01
15%

75%

95%

98%
99%
1%

0.005
10 -4
0
10 15 20 25 30 35 40 45 50 55 10 15 20 25 30 35 40 45 50 55

Figure 10: Probability density functions in standard scale (left) and semi-logarithmic scale (right). Com-
parison between the true density of VIX implied by the Heston model and the estimated RNDs of order 20.
The parameters for the Heston model are: k = 1.71, v̄ = 0.097, η = 0.577, v (0) = v̄ and T −t = 30/365. The
dashed vertical lines on the right panel identify several relevant probability levels and the corresponding
quantiles.

41
B.4 NIG distribution
We assume that the log-VIX at maturity T follows a NIG distribution, that is
 p 
K 1 α δ 2 + (s − µ) 2
log (VIXT ) ∼ д, д(s) = C · e β (s−µ) ,
δ + (s − µ)
p
2 2

αδ δγ
where C = π e is the normalization constant and Kν denotes the modified Bessel function
of the second kind (cf. Abramowitz and Stegun (1964)). Therefore, by the change of variable
s = log(x ) we obtain the RND of the VIX

1
f Q (x ) = · д (log(x )) .
x

The asymptotic properties of Kν determine polynomial decay of f Q both on the right and the left
tail. It follows that none of the kernels considered here meets the condition f Qϕ − 2 ∈ L2 (R+ ).
1

Figure 11 reports the true RND implied by the log-NIG density, and related expansions based
on different choices of the kernel. As expected, in all cases the main convergence issues involve
the tails. In particular, the expansion based on the GIG kernel is defective on both tails, which
is consistent with the fact that a GIG kernel decays more rapidly than the true RND, at both
sides. Due to the polynomial decay of the GW kernel on left tail, which accommodates the slow
decay of the true RND, the GW-based expansion proves inexact only on the right tail. Finally,
the LN kernel provides again the weakest performance, but it is worth noticing that here the
approximation is more accurate than in the previous test. This is a consequence of the fact that
the LN is nested within the log-NIG family, and therefore here f Q is intuitively ”closer” to a
log-normal than in the previous case.

B.5 Robustness to kernel specification

We now test the robustness of our estimation to the initialization of the kernel parameters, θ .
So far, the parameters of the kernels were optimally determined by minimizing the residuals
variance for the expansion of order 0. However, it is interesting to empirically assess how the
initial choice of the parameters θ ∈ Θ affects the accuracy of (5). To answer this question,
we perturb the parameters of optimally calibrated kernels, so that the moments and the option
prices implied by the kernels heavily mismatch those generated by the true f Q . In Table 4
we report the first four moments of the GIG and the GW kernels, where mean and variance
are drastically perturbed as compared to the values implied by the true density of the Heston
model. The last two columns of the table highlight the capability of the polynomial expansions
to yield precise fitting of the moments of the RND, even when the kernel largely deviates from
the true density. It is inherently assumed, however, that the assumptions of Theorem 3.1 are

42
 True GIG GW LN
0.08 10 -1

0.07

0.06
10 -2
0.05

0.04

10 -3
0.03

0.02

15%

75%

95%
98%
99%
1%
0.01 10 -4

0
10 20 30 40 50 60 70 20 30 40 50 60 70

Figure 11: Probability density functions in standard scale (left) and semi-logarithmic scale (right). Com-
parison between the true density of VIX implied by the NIG density and the estimated RNDs of order
20. The parameters for the NIG density are chosen as follows: α = 14.36, β = 9.8, µ = 2.97, γ = 0.38.
The dashed vertical lines on the right panel identify relevant probability levels and the corresponding
quantiles.

True GIG kernel GW kernel GIG (order 20) GW (order 20)

Mean 30.13 27.65 35.44 30.14 30.17
Variance 65.36 50.79 165.78 65.27 65.81
Skewness 50.26 21.07 56.86 50.16 50.40
Kurtosis 2.86 0.80 6.07 2.82 2.87

Table 4: The table reports mean, variance, standardized skewness and kurtosis of the true density of
the Heston model, of the calibrated kernel densities (GIG kernel and GW kernel) and of their related
expansions of order larger than 20.

always satisfied. Figure 12 portrays the true density implied by the Heston model, the perturbed
kernels, and the RNDs obtained by estimating the coefficients of the corresponding orthogonal
expansions.

43
 True Expansion (GIG) Expansion (GW) GIG kernel GW kernel
0.06

0.05

10 -2
0.04

0.03

10 -3
0.02

0.01

15%

75%

95%

98%
99%
1%
10 -4
0
10 15 20 25 30 35 40 45 50 55 10 15 20 25 30 35 40 45 50 55

Figure 12: Probability density functions in standard scale (left) and semi-logarithmic scale (right). Com-
parison between the true density implied by the Heston model, the ”mismatching” kernels, and the related
estimated expansions. The dashed vertical lines locate some relevant mass levels and the corresponding
quantiles.

The non-calibrated kernels clearly mismatch the true RND and totally deviate from each
other, but almost perfect approximations of the RND are attained in both cases through expan-
sions of order 20. Thus, the estimation based on the orthogonal expansions proves to be very
robust with respect to the initialization of ϕ.

25 Market 25 Market
Expansion order 20(GIG) Expansion order 20(GIG)
20 Expansion order 20(GIG) 20 Expansion order 20(GIG)
GIG kernel GIG kernel
GW kernel GW kernel
Price

Price

15 15

10 10

5 5

10 20 30 40 50 10 20 30 40 50
Strike Strike

(a) Call options (b) Put options

Figure 13: Call and put option prices implied by the Heston model, the mismatching GIG and GW
kernels, and related expansions of order 20.

Figure 13, depicting the option prices generated by the densities reported in Figure 12, con-
firms that the accuracy of the estimated RND is affected by the choice of the kernel only to a
minor extent. Indeed, the kernel has almost no impact on the estimation, provided that the con-
ditions of convergence have been guaranteed and that the expansion order can be set sufficiently
large - which is our case.

44
B.6 Robustness to no-arbitrage violations
We assess here the practical validity of Proposition 5.1 by means of Monte Carlo simulations.
To this end, we evaluate how adding a random noise to a discrete set of arbitrage-free prices
obtained from a known RND affects the estimates of RND obtained by solving the problem in
(19). Consistently with notations of Section 4, we assume the following form for the vector Y of
observed prices
Y = Y + ϵ,

where Y are arbitrage free option prices and ϵ is a vector of random shocks embedding all the
violations from the no-arbitrage assumption. Specifically, we assume that the vector of observed
call and put prices are given by

C = C + ϵC , P = P + ϵP

where C = [CK1 (t, τ ), . . . , CK M (t, τ )]0, P = [PK1 (t, τ ), . . . , PK M (t, τ )]0, ϵ C and ϵ P are indepen-
dent vectors of independent centered Gaussian variables with non-constant variance

σC,i
2
= Var[ϵiC ], σP,i
2
= Var[ϵiP ], i = 1, . . . , M.

Choosing a non-constant variance is owed to the fact that the magnitude of no-arbitrage viola-
tions must be consistent with the magnitude of option prices, which are monotonic quantities.
Therefore, σC2 = [σC,1
2 , . . . , σ 2 ] and σ 2 = [σ 2 , . . . , σ 2 ] are assumed to be an increasing and
C,M P P,1 P,M
a decreasing vector, respectively. To identify σC2 and σP2 we further assume that the arbitrage
error ϵ F induced on the vector F of future prices implied by the put-call parity consists of i.i.d.
components, that is
F = C − P + K = F + ϵF ,

where
F = Ci − Pi + Ki ∀i = 1, . . . , M

is the unique arbitrage-free future price. Hence ϵ F = ϵ C − ϵ P and E[ϵ F ] = 0. Assuming σF2 :=
Var[ϵ F ] < ∞, identification of Var[ϵ C ] and Var[ϵ P ] can therefore be achieved by

σC,i
2
Ci
σC,i
2
+ σP,i
2
= σF2 , = , i = 1, . . . , M,
σP,i
2 Pi

which gives

σF2
σP,i
2
= , σC,i
2
= σF2 − σP,i
2
, i = 1, . . . , M. (48)
Ci2
1+ Pi2

45
 Note that observable quantity ∆pcp defined in (25) is a sample counterpart of σF2 . Therefore,
2 + σ 2 by construction, up to switching the integration order in (24), ∆pcp can
since σF2 = σC,i P,i
consistently approximate the right hand of (24). We therefore carry out Monte Carlo simulations
with the purpose of investigating the robustness of the orthogonal polynomial expansion to the
no-arbitrage violations and the usefulness of the threshold ∆pcp to provide an indication for
the lower bound on the variance of residuals. Each Monte Carlo simulation consists of a set of
perturbed option prices Y over a fixed number M = 25 of strikes. The vector of arbitrage-free
call and put prices, Y, is generated only once, by direct integration of the VIX-RND implied by
the Heston model, with parameters k = 1.71, v̄ = 0.097, η = 0.577, v (0) = v̄ and τ = 30/365.
The arbitrage components ϵ C and ϵ P for each Monte Carlo simulation are obtained as

 ϵ C   σ 
  =  C  ◦ R,
P
 ϵ   σP 

where ◦ denotes the Hadamard product, σC , σP are determined as in (48), and R is a 2M ×1 vector
of i.i.d. standard Gaussian realizations, symmetrically truncated to ensure Y ≥ 0. We repeat the
procedure based on either the GIG or the GW kernel, and for σF = 0.01, 0.03, 0.05.4 The results
of these Monte Carlo simulations are summarized in Table (5). The so called divergence rate,
√
which is associated to the cases in which the RSME exceeds the threshold 2 ∆pcp , is intended
to approximate the frequency of violations of the conditions of Proposition 5.1. On the other
√
hand, the second column of Table 5 endorses the validity of (24), since the RSME is below ∆pcp
in a large percentage of cases. Furthermore, by looking at the Monte Carlo average of the RMSE,
it emerges that the variance of the error associated to the expansion of order 10 decreases with
σF and it is of the same order of σF in most cases. Differently, the two kernels on average are not
associated to a residual variance that is comparable to ∆pcp and the RMSE remains very high
also when σF = 0.01. The third column of Table 5 reports the filtering rate as a measure of how
often, among the convergent cases, the noise produced on data does not affect the estimated
RND. As expected, the filtering rate increases as the level of noise, namely σF , decreases. This is
consistent with Proposition 5.1 since the hypotheses 5.1.i)-ii) are expected to be less restrictive
as σF decreases. These additional hypotheses require that the estimated RND is constrained
to be positive - which is the case here - and that the observed prices do not embed multiple
arbitrage-free curves. Intuitively, under these hypotheses, the estimated RND is not affected
by the arbitrage noise existing in the observed prices. The figures reported in Table 5 provide
a solid confirmation of this intuition, since the percentage of cases where the noise does not
affect the estimated RND grows as σF2 decreases, which in turn implies reducing the uncertainty
on the RND. Consistently, the L2 distance between the estimated and the true RND decreases as

√
4 Typicalvalues of ∆pcp determined on real data fall in the interval [0.01, 0.05], which roughly correspond to
an uncertainty between 1 and 5 cents of dollar on the futures prices implied by the put-call parity.

46
σF2 decreases, as shown in the fifth column. Finally, the last column of Table 5 confirms that the
estimation based on expansions of order 10 outperforms the related kernel in all cases.

Order 10 Kernel
Div. rate Fitt. rate Filt. rate RMSE L2 RMSE L2

GIG 14.4 % 71.4 % 100 % 0.0123 0.0069 0.0969 0.0218

σF = 0.01
GW 15.6 % 71.3 % 100 % 0.0125 0.0070 0.0962 0.0217
GIG 5.6 % 74 % 82.57 % 0.0362 0.0135 0.0988 0.0217
σF = 0.03
GW 4.7 % 76.3 % 80.47 % 0.0356 0.0131 0.0980 0.0217
GIG 6.8 % 81.5 % 56.81 % 0.0713 0.0186 0.1024 0.0218
σF = 0.05
GW 4.7 % 83.2 % 57.81 % 0.0536 0.0174 0.1017 0.0217

Table 5: The table summarizes the results of N = 1000 Monte Carlo tests described above, correspond-
ing to different kernels and different values of σ F2 . The first column reports the divergence rate of the
estimation, determined
√ as the percentage of tests such that the residual root-mean squared error (RMSE)
is greater than 2 ∆ . The second column reports the rate of optimal fitting according to Equation (26),
pcp

that is the percentage of tests yielding a RMSE lower or equal to σ F . The third column reports the per-
centage of tests for which the arbitrage component is successfully ”filtered”. The arbitrage component
is considered to be filtered when the RND estimated on the perturbed data and the RND estimated on
arbitrage-free data achieve the same level of accuracy, in terms of magnitude (∼ 10−3 ) of their distance
from the true RND, measured as L2 norm (L2). Only convergent tests are considered in this computation.
The last four columns report the Monte Carlo average of RMSE and L2 relative to the expansion of order
10 and kernel (order 0).

B.7 Details on the Empirical Analysis

GIG GW
Par. Estimate Par. Estimate
α -0.899 α 2.467
β 0.090 β 0.874
ξ 33.99 p 0.605

Table 6: The table reports the kernel parameters for the GIG and GW kernels respectively, estimated
using the procedure detailed in Section B.1.2.

47
 τ = 1 month τ = 2 months τ = 3 months τ = 4 months τ = 5 months

M Min Max M Min Max M Min Max M Min Max M Min Max
20 Jan 2010 21 17 60 22 15 75 20 18 75 19 15 70 20 15 75
24 Feb 2010 19 17 50 23 17 75 18 18 70 21 10 75 20 15 80
24 Mar 2010 18 15 42.5 21 15 55 21 15 80 20 15 80 20 15 80
21 Apr 2010 20 15 47.5 21 15 60 22 15 70 24 15 80 23 15 75
26 May 2010 25 19 100 28 10 90 28 15 100 26 15 90 31 15 100
23 Jun 2010 21 20 80 23 18 90 28 15 100 30 16 100 20 15 80
21 Jul 2010 21 17 75 23 17 85 31 15 100 19 20 80 21 10 80
25 Aug 2010 19 22.5 75 27 18 95 22 10 80 21 10 80 21 10 80
22 Sep 2010 22 17 65 27 15 80 22 20 80 20 20 75 22 15 80
20 Oct 2010 23 15 55 23 18 75 22 18 80 24 10 80 21 15 80
24 Nov 2010 21 15 47.5 24 15 70 23 15 75 22 15 80 22 15 80
22 Dec 2010 21 15 50 26 15 75 23 15 75 21 15 75 20 15 75
26 Jan 2011 18 14 40 23 13 55 27 13 75 27 13 75 26 13 70
23 Feb 2011 22 15 60 25 15 75 25 14 70 26 14 75 28 13 80
23 Mar 2011 20 15 47.5 25 15 70 25 15 75 27 14 80 26 13 70
20 Apr 2011 22 13 47.5 27 14 75 27 14 75 27 14 80 25 15 75
25 May 2011 19 14 42.5 25 13 60 25 13 65 25 15 75 24 15 70
22 Jun 2011 21 15 50 24 14 65 24 15 70 24 15 70 26 15 80
20 Jul 2011 22 14 55 24 15 70 25 15 75 25 15 75 28 15 80
24 Aug 2011 22 21 90 26 19 100 25 19 95 29 16 90 31 17 95
21 Sep 2011 24 22.5 95 26 20 95 30 16 95 31 18 100 33 16 100
26 Oct 2011 23 19 75 28 17 90 32 17 100 33 16 100 34 15 100
23 Nov 2011 24 21 90 30 19 100 34 15 100 35 10 100 34 15 100
21 Dec 2011 26 17 70 29 16 80 31 15 85 33 15 95 31 16 90
25 Jan 2012 25 15 55 27 16 70 32 16 95 31 15 85 32 15 90
22 Feb 2012 26 15 60 30 16 85 29 15 75 31 15 85 33 15 95
21 Mar 2012 28 12 55 31 12 70 31 12 70 34 10 75 35 11 85
25 Apr 2012 24 13 45 30 12 65 33 12 80 32 12 75 34 10 95
23 May 2012 27 16 70 31 14 80 33 14 90 31 15 85 30 15 80
20 Jun 2012 31 14 80 36 13 100 30 15 80 29 15 75 30 15 80
25 Jul 2012 28 14 65 31 15 85 30 15 80 32 15 90 30 15 80
22 Aug 2012 31 13 75 31 13 75 31 13 75 32 13 80 31 13 75
26 Sep 2012 26 13 50 32 11 70 30 12 65 31 12 70 32 12 75
24 Oct 2012 27 13 55 31 12 70 31 12 70 31 12 70 32 13 80
21 Nov 2012 26 13 50 33 12 80 31 12 70 30 13 70 30 13 70
26 Dec 2012 25 14 50 32 12 75 29 13 65 31 12 70 32 12 75
23 Jan 2013 20 12 32.5 28 11 50 34 11 80 31 10 60 31 10 60
20 Feb 2013 22 12 37.5 30 11 60 33 10 70 35 10 80 30 11 60
20 Mar 2013 25 11 42.5 30 11 60 35 9 75 30 11 60 34 11 80
24 Apr 2013 23 11 37.5 31 11 65 33 11 75 36 9 80 31 11 65
22 May 2013 25 12 45 31 11 65 32 11 70 31 10 60 31 11 65
26 Jun 2013 22 13 40 28 13 60 33 11 75 31 12 70 30 12 65
24 Jul 2013 20 12 32.5 28 11 50 29 11 55 31 11 65 28 12 55
21 Aug 2013 23 12 40 30 11 60 29 11 55 32 10 65 30 11 60
25 Sep 2013 19 12 30 25 12 45 29 12 60 31 11 65 30 11 60
23 Oct 2013 22 12 37.5 28 12 55 29 11 55 31 11 65 27 12 50
20 Nov 2013 22 12 37.5 27 12 50 31 11 65 31 11 65 29 11 55
24 Dec 2013 21 13 37.5 27 12 50 28 11 50 31 11 65 30 11 60
22 Jan 2014 25 12 45 29 11 55 30 11 60 30 11 60 30 11 60
26 Feb 2014 21 12 35 27 12 50 31 11 65 30 12 65 31 11 65
26 Mar 2014 20 13 35 24 13 45 30 12 65 30 11 60 30 11 60
23 Apr 2014 19 12 30 24 12 42.5 29 12 60 32 11 70 29 11 55
21 May 2014 19 12 30 23 12 40 30 11 60 30 11 60 33 11 75
25 Jun 2014 23 10.5 28 24 10 37.5 29 10 50 30 11 60 30 10 55
23 Jul 2014 23 10.5 28 30 10 40 29 10 50 31 10 60 31 10 60
20 Aug 2014 26 10.5 32.5 29 10 50 31 10 60 31 10 60 33 10 70
24 Sep 2014 24 11.5 32.5 30 11 45 30 10 55 33 10 70 32 11 70
22 Oct 2014 31 12.5 60 31 11 65 30 12 65 32 11 70 31 11 65
26 Nov 2014 25 11.5 35 32 12 60 31 11 65 31 11 65 30 12 65
24 Dec 2014 28 12 45 34 11.5 65 30 11 60 30 12 65 32 11 70
21 Jan 2015 28 13.5 55 30 12 65 29 12 60 30 12 65 31 12 70
25 Feb 2015 25 12.5 40 31 12.5 60 31 12 70 31 12 70 33 11 75
25 Mar 2015 23 13 37.5 31 12 55 29 12 60 30 12 65 32 11 70
22 Apr 2015 26 12 40 31 12 55 30 12 65 32 11 70 31 12 70
20 May 2015 25 12 37.5 31 11 65 31 12 70 31 11 65 32 11 70
24 Jun 2015 27 11.5 40 32 11.5 55 31 11 65 31 12 70 31 12 70
22 Jul 2015 26 11.5 37.5 29 11 55 32 11 70 30 12 65 31 12 70
26 Aug 2015 31 14 75 35 12.5 80 31 12 70 30 12 65 32 11 70
22 Sep 2015 29 14 65 32 13 70 30 12 65 31 12 70 31 12 70
20 Oct 2015 30 12.5 55 29 12 60 31 11 65 32 11 70 31 12 70
24 Nov 2015 25 12 37.5 32 12.5 65 30 12 65 31 12 70 31 12 70
22 Dec 2015 29 13 55 33 12 65 31 11 65 33 12 80 31 12 70
19 Jan 2016 30 13.5 65 29 14 70 32 13 80 31 12 70 32 11 70
23 Feb 2016 25 14.5 50 29 14.5 70 28 14 65 30 13 70 32 13 80
22 Mar 2016 28 13 50 32 12.5 65 29 12 60 31 12 70 30 12 65
19 Apr 2016 27 12 42.5 29 12 60 30 12 65 31 12 70 31 12 70

Table 7: Summary of the panel of VIX options.

48
 √ τ = 1 month √ τ = 2 months √ τ = 3 months √ τ = 4 months √ τ = 5 months
∆pcp Kernel Expans. ∆pcp Kernel Expans. ∆pcp Kernel Expans. ∆pcp Kernel Expans. ∆pcp Kernel Expans.
20 Jan 2010 0.0886 0.0621 0.0538 0.0556 0.103 0.0362 0.0664 0.0799 0.0355 0.0807 0.116 0.0438 0.0654 0.13 0.037
24 Feb 2010 0.0744 0.075 0.0431 0.032 0.0943 0.0239 0.0458 0.103 0.0333 0.0532 0.106 0.0514 0.07 0.128 0.0364
24 Mar 2010 0.0554 0.0616 0.0322 0.0416 0.0782 0.0232 0.0344 0.0689 0.022 0.0518 0.103 0.0297 0.072 0.0877 0.0402
21 Apr 2010 0.0296 0.0948 0.0235 0.0326 0.0775 0.0206 0.0337 0.0935 0.0239 0.0416 0.0948 0.03 0.0717 0.106 0.0486
26 May 2010 0.0761 0.0968 0.0562 0.0798 0.253 0.0603 0.0776 0.17 0.0459 0.0869 0.151 0.0511 0.0767 0.172 0.0578
23 Jun 2010 0.0321 0.123 0.0267 0.0459 0.121 0.0393 0.0465 0.177 0.0379 0.0608 0.0992 0.0419 0.0668 0.11 0.0433
21 Jul 2010 0.0305 0.096 0.0205 0.0403 0.0812 0.0257 0.0713 0.125 0.0465 0.0685 0.0906 0.0383 0.0708 0.128 0.0376
25 Aug 2010 0.0328 0.0733 0.0226 0.0463 0.075 0.0301 0.0469 0.129 0.0272 0.085 0.143 0.0448 0.0749 0.133 0.0414
22 Sep 2010 0.0443 0.0572 0.0305 0.0369 0.063 0.0239 0.0389 0.0935 0.0287 0.0367 0.0783 0.0209 0.0493 0.145 0.031
20 Oct 2010 0.0471 0.0982 0.0327 0.0706 0.118 0.0503 0.0405 0.124 0.0248 0.0813 0.142 0.042 0.087 0.13 0.0474
24 Nov 2010 0.0383 0.0713 0.0248 0.0367 0.0738 0.0281 0.0757 0.145 0.0482 0.0555 0.172 0.0427 0.0698 0.122 0.0436
22 Dec 2010 0.0402 0.0496 0.0278 0.0558 0.123 0.0336 0.0333 0.109 0.0308 0.0585 0.105 0.0346 0.0347 0.107 0.0254
26 Jan 2011 0.0224 0.0681 0.0169 0.023 0.0706 0.0176 0.0351 0.0994 0.0227 0.0334 0.0812 0.0228 0.0383 0.0708 0.0229
23 Feb 2011 0.028 0.0691 0.023 0.0519 0.0971 0.0327 0.0491 0.0757 0.0352 0.0547 0.0828 0.037 0.0536 0.113 0.0402
23 Mar 2011 0.0329 0.0506 0.0276 0.0303 0.0935 0.0262 0.0349 0.0805 0.0246 0.0339 0.097 0.021 0.0442 0.118 0.0357
20 Apr 2011 0.0295 0.0952 0.019 0.03 0.0852 0.0227 0.039 0.114 0.0359 0.0526 0.107 0.0534 0.0406 0.117 0.0351
25 May 2011 0.0282 0.0761 0.0187 0.0267 0.102 0.0217 0.0227 0.0927 0.0202 0.0322 0.0871 0.0231 0.0458 0.115 0.0386
22 Jun 2011 0.0236 0.079 0.0209 0.027 0.077 0.0215 0.0296 0.0789 0.0257 0.0374 0.127 0.036 0.0312 0.106 0.0224
20 Jul 2011 0.023 0.0906 0.0193 0.0278 0.0873 0.024 0.0299 0.143 0.0226 0.0331 0.0917 0.0211 0.0323 0.0851 0.0382
24 Aug 2011 0.0371 0.0754 0.0352 0.0445 0.121 0.0362 0.0385 0.143 0.0336 0.0411 0.183 0.0417 0.0779 0.137 0.0577
21 Sep 2011 0.0358 0.0903 0.0274 0.0273 0.0985 0.022 0.0385 0.187 0.0307 0.0583 0.118 0.059 0.0795 0.15 0.0432
26 Oct 2011 0.0238 0.0843 0.0172 0.0284 0.0808 0.0252 0.0272 0.0714 0.0203 0.0333 0.0999 0.0324 0.0641 0.136 0.0359
23 Nov 2011 0.026 0.0692 0.0245 0.0423 0.0925 0.033 0.0398 0.115 0.0274 0.0571 0.261 0.0329 0.0738 0.275 0.0421
21 Dec 2011 0.0304 0.112 0.0235 0.0343 0.0889 0.0289 0.0442 0.153 0.0286 0.0462 0.156 0.0265 0.0437 0.103 0.0437
25 Jan 2012 0.0326 0.0831 0.0221 0.0337 0.0569 0.0284 0.0438 0.0543 0.032 0.0447 0.0853 0.0337 0.058 0.0974 0.0352
22 Feb 2012 0.028 0.0762 0.0222 0.0305 0.0684 0.0291 0.0289 0.175 0.0286 0.0553 0.161 0.0322 0.0837 0.167 0.0447
21 Mar 2012 0.0302 0.0704 0.0191 0.0334 0.0546 0.0218 0.04 0.0614 0.0319 0.0442 0.157 0.0305 0.0468 0.143 0.0302
25 Apr 2012 0.02 0.0771 0.0164 0.0423 0.107 0.0268 0.0394 0.0654 0.0232 0.0437 0.074 0.0271 0.0599 0.144 0.0322
23 May 2012 0.042 0.0611 0.025 0.0461 0.0676 0.0295 0.0475 0.061 0.0314 0.035 0.153 0.0261 0.0594 0.106 0.0362
20 Jun 2012 0.028 0.0742 0.0198 0.0426 0.0661 0.0271 0.0308 0.0687 0.0236 0.0317 0.0823 0.031 0.0418 0.0797 0.0363
25 Jul 2012 0.0276 0.066 0.0205 0.0386 0.083 0.0254 0.0453 0.0804 0.0293 0.0536 0.159 0.0305 0.0645 0.0837 0.0363
22 Aug 2012 0.029 0.07 0.0198 0.042 0.0744 0.0281 0.0303 0.0598 0.0232 0.0439 0.0523 0.0334 0.0415 0.0548 0.0266
26 Sep 2012 0.0227 0.0757 0.019 0.0279 0.118 0.0215 0.0263 0.0737 0.0197 0.0294 0.0545 0.0205 0.0394 0.0672 0.0244
24 Oct 2012 0.0318 0.075 0.0297 0.034 0.0794 0.0278 0.0331 0.0857 0.0259 0.0375 0.0662 0.0338 0.0532 0.0702 0.0342
21 Nov 2012 0.0281 0.0771 0.0191 0.0417 0.0331 0.0255 0.0302 0.0704 0.0296 0.0404 0.0721 0.0267 0.04 0.0734 0.0264
26 Dec 2012 0.0322 0.0525 0.0203 0.0291 0.0958 0.024 0.0408 0.0572 0.0292 0.0561 0.0945 0.0324 0.0593 0.0617 0.0323

Table 8: Root mean square error (RMSE): errors between the observed VIX option prices and the approximate
prices implied by GW kernel and related expansion.

49
 √ τ = 1 month √ τ = 2 months √ τ = 3 months √ τ = 4 months √ τ = 5 months
∆pcp Kernel Expans. ∆pcp Kernel Expans. ∆pcp Kernel Expans. ∆pcp Kernel Expans. ∆pcp Kernel Expans.
23 Jan 2013 0.0311 0.0566 0.0201 0.0235 0.053 0.0192 0.0495 0.0672 0.0303 0.0485 0.106 0.0313 0.0434 0.113 0.0302
20 Feb 2013 0.0281 0.0605 0.0214 0.036 0.0576 0.0233 0.0385 0.0862 0.0284 0.0597 0.108 0.0376 0.0523 0.0642 0.0316
20 Mar 2013 0.0186 0.0624 0.0168 0.0402 0.0674 0.0309 0.0364 0.123 0.0311 0.0487 0.0676 0.0348 0.0641 0.0561 0.0398
24 Apr 2013 0.0203 0.0556 0.0162 0.0293 0.0621 0.0238 0.045 0.0602 0.0347 0.0539 0.117 0.0384 0.0523 0.0551 0.0337
22 May 2013 0.034 0.0641 0.0219 0.0295 0.0651 0.0213 0.0325 0.0483 0.0264 0.038 0.105 0.033 0.0369 0.0805 0.028
26 Jun 2013 0.0254 0.0758 0.0209 0.0235 0.0591 0.0228 0.037 0.102 0.0304 0.0366 0.0417 0.0289 0.0332 0.0438 0.0272
24 Jul 2013 0.0224 0.0554 0.0171 0.0271 0.0612 0.0199 0.0334 0.0544 0.0227 0.0442 0.0363 0.0252 0.0445 0.0473 0.0278
21 Aug 2013 0.0361 0.0629 0.0317 0.049 0.0986 0.0371 0.0249 0.0913 0.0235 0.0573 0.103 0.0408 0.0531 0.0609 0.0371
25 Sep 2013 0.0257 0.0335 0.0175 0.0246 0.0588 0.0203 0.0447 0.0646 0.0367 0.0489 0.0588 0.0373 0.045 0.045 0.0359
23 Oct 2013 0.0261 0.0721 0.0222 0.0286 0.0712 0.0214 0.0305 0.051 0.0254 0.0439 0.0461 0.0265 0.0373 0.0432 0.0317
20 Nov 2013 0.0289 0.0608 0.0205 0.0272 0.0663 0.0195 0.0347 0.0918 0.0208 0.0353 0.0405 0.0244 0.0388 0.0406 0.0273
24 Dec 2013 0.0277 0.0372 0.0272 0.0305 0.0884 0.029 0.0655 0.0897 0.0624 0.0386 0.0668 0.0299 0.037 0.0561 0.0277
22 Jan 2014 0.0256 0.0509 0.0212 0.0236 0.128 0.0233 0.0369 0.0933 0.0341 0.041 0.0546 0.0273 0.05 0.0677 0.0363
26 Feb 2014 0.0269 0.0565 0.0187 0.0189 0.0841 0.0179 0.0253 0.108 0.0224 0.0303 0.0815 0.0249 0.0306 0.0703 0.03
26 Mar 2014 0.0253 0.0545 0.016 0.027 0.04 0.0263 0.0249 0.056 0.022 0.0363 0.107 0.0367 0.0354 0.0681 0.0338
23 Apr 2014 0.0328 0.0514 0.0228 0.0242 0.0195 0.0182 0.0362 0.0652 0.0337 0.0467 0.0985 0.0421 0.0405 0.0633 0.036
21 May 2014 0.0316 0.0379 0.02 0.0255 0.0652 0.0208 0.0237 0.037 0.0193 0.0306 0.0635 0.0267 0.0421 0.0576 0.0353
25 Jun 2014 0.0215 0.0369 0.0147 0.021 0.0492 0.0194 0.0304 0.0303 0.0196 0.0241 0.0631 0.0184 0.0341 0.0614 0.0312
23 Jul 2014 0.0199 0.0503 0.017 0.0234 0.091 0.02 0.0207 0.0628 0.0174 0.0337 0.0869 0.027 0.0326 0.0858 0.0319
20 Aug 2014 0.0187 0.0343 0.0148 0.0233 0.056 0.0179 0.0209 0.0623 0.0213 0.0425 0.106 0.039 0.0415 0.0609 0.0327
24 Sep 2014 0.0239 0.0713 0.0198 0.0237 0.0823 0.018 0.0309 0.0622 0.0257 0.0392 0.131 0.0361 0.0393 0.0687 0.0379
22 Oct 2014 0.0223 0.071 0.0185 0.0259 0.102 0.0237 0.0314 0.0834 0.0307 0.0284 0.0793 0.0222 0.0262 0.0752 0.0252
26 Nov 2014 0.0332 0.0618 0.0221 0.0261 0.129 0.0187 0.0221 0.127 0.0221 0.0295 0.0838 0.0298 0.027 0.0613 0.0271
24 Dec 2014 0.0303 0.0745 0.0198 0.0337 0.129 0.0213 0.0371 0.104 0.0353 0.0347 0.0778 0.0311 0.0413 0.063 0.028
21 Jan 2015 0.0319 0.0564 0.0214 0.0264 0.122 0.0265 0.0179 0.103 0.018 0.0279 0.0522 0.0224 0.0395 0.0704 0.038
25 Feb 2015 0.0265 0.0684 0.0172 0.0208 0.066 0.0209 0.0149 0.0722 0.0141 0.0261 0.0616 0.0251 0.0368 0.139 0.0289
25 Mar 2015 0.0285 0.0629 0.0177 0.0179 0.0928 0.0171 0.0263 0.0521 0.0262 0.0296 0.0693 0.0296 0.0299 0.0673 0.0283
22 Apr 2015 0.0248 0.0859 0.0207 0.0222 0.0833 0.0204 0.0233 0.0584 0.0204 0.0232 0.11 0.0236 0.0276 0.0741 0.0242
20 May 2015 0.0208 0.0767 0.0195 0.0287 0.106 0.0208 0.0275 0.0756 0.0239 0.0224 0.102 0.0184 0.033 0.139 0.0307
24 Jun 2015 0.0281 0.089 0.0207 0.0342 0.0696 0.0308 0.0343 0.124 0.032 0.0348 0.0744 0.0274 0.0507 0.0778 0.0395
22 Jul 2015 0.0219 0.0908 0.02 0.0255 0.0421 0.0219 0.031 0.0983 0.0279 0.0411 0.0894 0.0384 0.0418 0.0635 0.0368
26 Aug 2015 0.0307 0.0442 0.0235 0.046 0.0786 0.044 0.0565 0.0702 0.0418 0.0425 0.0974 0.041 0.0593 0.0715 0.0379
22 Sep 2015 0.0291 0.0969 0.0257 0.0286 0.0768 0.0236 0.0314 0.102 0.0309 0.0481 0.115 0.0489 0.0491 0.0959 0.0407
20 Oct 2015 0.0223 0.0883 0.0183 0.023 0.0408 0.021 0.0206 0.11 0.0208 0.0364 0.116 0.0264 0.0445 0.0888 0.0479
24 Nov 2015 0.03 0.0684 0.0232 0.0335 0.044 0.0261 0.0422 0.0701 0.0418 0.0396 0.0875 0.0379 0.0495 0.0809 0.0472
22 Dec 2015 0.0291 0.073 0.0202 0.027 0.114 0.0272 0.0369 0.101 0.0249 0.0571 0.0638 0.0427 0.0632 0.0746 0.0485
19 Jan 2016 0.0312 0.0699 0.0214 0.0249 0.059 0.0219 0.0448 0.0816 0.0438 0.0453 0.105 0.0455 0.0377 0.133 0.0401
23 Feb 2016 0.025 0.0518 0.0201 0.0282 0.0698 0.0215 0.036 0.0683 0.0338 0.0338 0.0708 0.03 0.0532 0.0699 0.0525
22 Mar 2016 0.0249 0.0775 0.0171 0.0324 0.0768 0.0254 0.0312 0.0772 0.0214 0.0472 0.0661 0.0436 0.0506 0.0771 0.0477
19 Apr 2016 0.0249 0.0623 0.0191 0.0233 0.098 0.022 0.0282 0.0524 0.0256 0.0362 0.0525 0.0306 0.042 0.0609 0.0343

Table 9: Root mean square error (RMSE): errors between the observed VIX option prices and the approximate
prices implied by GW kernel and related expansion.

50
Research Papers
2016

2016-31: Carlos Vladimir Rodríguez-Caballero: Panel Data with Cross-Sectional

Dependence Characterized by a Multi-Level Factor Structure
2016-32: Lasse Bork, Stig V. Møller and Thomas Q. Pedersen: A New Index of Housing
Sentiment
2016-33: Joachim Lebovits and Mark Podolskij: Estimation of the global regularity of a
multifractional Brownian motion
2017-01: Nektarios Aslanidis, Charlotte Christiansen and Andrea Cipollini: Predicting
Bond Betas using Macro-Finance Variables
2017-02: Giuseppe Cavaliere, Morten Ørregaard Nielsen and Robert Taylor: Quasi-
Maximum Likelihood Estimation and Bootstrap Inference in Fractional Time
Series Models with Heteroskedasticity of Unknown Form
2017-03: Peter Exterkate and Oskar Knapik: A regime-switching stochastic volatility
model for forecasting electricity prices
2017-04: Timo Teräsvirta: Sir Clive Granger’s contributions to nonlinear time series
and econometrics
2017-05: Matthew T. Holt and Timo Teräsvirta: Global Hemispheric Temperatures and
Co–Shifting: A Vector Shifting–Mean Autoregressive Analysis
2017-06: Tobias Basse, Robinson Kruse and Christoph Wegener: The Walking Debt
Crisis
2017-07: Oskar Knapik: Modeling and forecasting electricity price jumps in the Nord
Pool power market
2017-08: Malene Kallestrup-Lamb and Carsten P.T. Rosenskjold: Insight into the
Female Longevity Puzzle: Using Register Data to Analyse Mortality and Cause
of Death Behaviour Across Socio-economic Groups
2017-09: Thomas Quistgaard Pedersen and Erik Christian Montes Schütte: Testing for
Explosive Bubbles in the Presence of Autocorrelated Innovations
2017-10: Jeroen V.K. Rombouts, Lars Stentoft and Francesco Violante: Dynamics of
Variance Risk Premia, Investors' Sentiment and Return Predictability
2017-11: Søren Johansen and Morten Nyboe Tabor: Cointegration between trends and
their estimators in state space models and CVAR models
2017-12: Lukasz Gatarek and Søren Johansen: The role of cointegration for optimal
hedging with heteroscedastic error term
2017-13: Niels S. Grønborg, Asger Lunde, Allan Timmermann and Russ Wermers:
Picking Funds with Confidence
2017-14: Martin M. Andreasen and Anders Kronborg: The Extended Perturbation
Method: New Insights on the New Keynesian Model
2017-15: Andrea Barletta, Paolo Santucci de Magistris and Francesco Violante: A Non-
Structural Investigation of VIX Risk Neutral Density