The extended SSVI volatility surface

Sebas Hendriks1,2 , Claude Martini1

1 Zeliade Systems, Paris 2 Delft University of Technology

Abstract
We extend Gatheral and Jacquier SSVI volatility surface parameterisation by making
the correlation maturity-dependent, obtaining necessary and sufficient conditions for
no calendar-spread arbitrage. Parametric families for the correlation are provided for
which those conditions are explicit. This extension of SSVI typically increases the
calibration accuracy for short maturities, and may also be more robust in stressed
market conditions.
Keywords and phrases: volatility smile, volatility surface, no arbitrage.

1 Introduction
Explicit formulas for implied volatility are of utmost practical interest: practitioners are
able to explain the market with few parameters and, most importantly, they do so without
the need to go through intricate mathematical models, complex pricing algorithms and the
additional layer of an algorithm to compute the implied volatility from the price. Implied
volatility surfaces are used directly by risk teams to store historical implied volatility data in
a compact way which enables the design of forward-looking risk measures from the history
of the calibrated parameters, and also by traders as pre-calibration smoothers.
One of the most successful volatility slice (meaning: only a given maturity is looked at)
models is the Stochastic Inspired Volatility model (SVI, [1]) used internally at Merrill Lynch
and publicly disclosed by Jim Gatheral in 2004. Several equivalent parameterisations of SVI
are available which differ by the geometrical interpretation of the parameters. SVI is just
a five-parameters formula, yet despite its simplicity, no necessary and sufficient conditions
are known ensuring the absence of arbitrage. In practice SVI is massively used on Equity
markets.
Numerous applied research teams have attempted to extend SVI to a whole surface
model. This has been achieved in 2012 by Gatheral and Jacquier with the Surface SVI
(SSVI, [2]) model. SSVI is parameterized with the ATM total variance θt , in such a way
that SSVI slices at a given maturity t are SVI slices with only 3 parameters (so a sub-family
of SVI). This restriction allows to get explicit sufficient conditions for absence of arbitrage,
while allowing enough flexibility for calibration, at least on liquid Equity Indexes (most
academic papers on SSVI test calibration on SPX 500). Besides the curve θt , which is
assumed to be read on the market, SSVI has a constant leverage parameter ρ which drives

Electroniccopy
Electronic copy available
available at:
at: https://ssrn.com/abstract=2971502
https://ssrn.com/abstract=2971502
the skew, and a curvature function ϕ(θt ). Parametric families have been proposed for ϕ,
like the power law ηθt−λ .
SSVI is a major achievement: it provides a tractable arbitrage-free parameterisation of
the volaility surface, which calibrates quite well. It has a few shortcomings though. Firstly
the ATM volatility is taken directly as a parameter, which is quite unusual in the modeling
litterature. To our knowledge, SSVI is the only model to do this. For instance Bergomi’s
last generation of volatility models takes the Variance Swap curve as a parameter instead,
since Variance Swap quotes are directly available on the market. Secondly, SSVI slices
depend only on 3 parameters, which is probably too few in some situations and certainly
too few for a very accurate fit. Thirdly, the leverage parameter ρ is constant for the whole
surface (in other words, among their 3 respective parameters, the same ρ will be shared by
2 slices at different maturities).
An attempt to address the second point has been made in [3], where a general non-
parametric shape generalizing the SVI shape at a given maturity has been investigated and
necessary and sufficient conditions for no arbitrage have been obtained.
The goal of this paper is to address the third point and investigate an extension of
SSVI, which we will call eSSVI, where the correlation parameter ρ may also depend on the
maturity through θt to allow an even greater accuracy of the surface calibration.
In section 2, we recall Gatheral and Jacquier results on SSVI and motivate the introduc-
tion of eSSVI. We start then by studying the consistency of two SSVI slices (so, in discrete
time) with different correlation parameters ρ1 and ρ2 and obtain an explicit necessary and
sufficient condition for the absence of calendar-spread arbitrage. Passing formally to the
continuous-time limit, we obtain a set of no (calendar-spread) arbitrage conditions and
prove rigorously that they are necessary and sufficient in section 4. In the last section we
illustrate the calibration of eSSVI on Equity market data.
It should be noted that we make use of the word formula and not of the word model:
volatility surfaces are to some extent model-free in the sense that they don’t prescribe in
general a dynamic for the underlying: they are by essence of a purely static nature. In
principle, a local volatility model can always be associated to a volatility surface, so that
American options and path-dependent claims can be priced in a consistent way if needed.
In practice, some care is needed with the parameters to ensure that the corresponding
local volatility function has no (or only mild) singularities and satisfies the required growth
behaviour so that the solution to the associated SDE is a true martingale.
We would like to thank Antoine Jacquier, Stefano De Marco, Ismail Laachir and Cornelis
Oosterlee for fruitful discussions at Zeliade or at Delft University. All errors are ours.

2 From the SSVI to the eSSVI surface

Gatheral and Jaquier derived SSVI as an extension of the well known SVI slice formula for
a whole volatility surface with fewer parametersat a given maturity. SSVI is specified as a
function of the log-forward-moneyness k := log FKt , with K the the option strike and Ft
2 (0, t).
the forward price and of the at-the-money (ATM) implied total variance θt := tσBS

Electroniccopy
Electronic copy available
available at:
at: https://ssrn.com/abstract=2971502
https://ssrn.com/abstract=2971502
It is assumed that the ATM implied total variance is a function in time of at least class
C 1 on R∗+ , and that: limt→0 θt = 0. For a smooth curvature function: ϕ : R∗+ 7→ R∗+ such
that limt→0 θt ϕ(θt ) exists in R, SSVI consists in the following formula for the total implied
2 (k, t):
variance w(k, θt ) := tσBS

θt  p 
w(k, θt ) = 1 + ρϕ(θt )k + (ϕ(θt )k + ρ)2 + 1 − ρ2 (2.1)
2
where ρ is a constant with |ρ| ≤ 1 representing the correlation between the stock price
and its instantaneous volatility.
The main attractive feature of SSVI is the availability of tractable conditions for no
arbitrage: a necessary and sufficient condition for no calendar-spread arbitrage, and a
sufficient condition for no butterfly arbitrage:

Theorem 2.1 (Theorem 4.1 in [2]). The SSVI surface is free of calendar-spread arbitrage
if and only if:

1. ∂t θt ≥ 0, for all t ≥ 0;
 p 
2. 0 ≤ ∂θ (θϕ(θ)) ≤ ρ12 1 + 1 − ρ2 ϕ(θ), for all θ > 0.

Theorem 2.2 (Theorem 4.2 in [2]). The SSVI surface is free of butterfly arbitrage if for
all θ > 0:

1. θϕ(θ)(1 + |ρ|) < 4;

2. θϕ2 (θ)(1 + |ρ|) ≤ 4.

One of the shortcomings of the SSVI surface is that the correlation ρ is assumed to
remain constant across maturities. We would like instead to allow it to depend on the
maturity, or, equivalently, on θt : this will be the eSSVI, or extended SSVI surface, where
the constant ρ is replaced by a function: ρ : R∗+ 7→ (−1, 1), such that:

θt  p 
w(k, θt ) = 1 + ρ(θt )ϕ(θt )k + (ϕ(θt )k + ρ(θt ))2 + 1 − ρ(θt )2 (2.2)
2
To motivate this extension, it is instructive to go through the following basic experiment:
on a set of liquid vanilla option quotes at a given time, calibrate an SSVI slice on each
available maturity so that the only difference with the classical SSVI is that the parameter
ρ is not constant giving a different correlation value for each slice. The typical time-
dependency pattern for the calibrated ρ, on Equity indices, is as shown in Figure 1.

Electronic copy available at: https://ssrn.com/abstract=2971502

 Figure 1: Typical pattern for a slicewise calibrated ρ.

We can see that the correlation parameter does exhibit some sort of constant behaviour
for longer maturities, however for shorter maturities it appears that letting ρ vary will yield
an enhanced calibration accuracy.
The issue at hand is to obtain no-arbitrage conditions for eSSVI. For butterfly arbitrage,
since the two surfaces have similar structures for fixed maturities, nothing has changed and
we can use the SSVI condition. For calendar-spread arbitrage conditions the story is more
complicated. To determine these conditions we will start by investigating the case of two
(e)SSVI slices with two different correlation parameters attached to two different maturities.

3 A tale of two SSVI slices

In this section we consider two SSVI slices: w(k, θt1 ) and w(k, θt2 ) with maturities t1 < t2 ,
defined by parameters (θti , ρ(θti ), ϕ(θti )). We simplify notation by defining the following:
wi := w(k, θti ), θi := θti , ϕi := ϕ(θi ) and ρi := ρ(θi ). So we have:

θ1
q
w1 = (1 + ρ1 ϕ1 k + ϕ21 k 2 + 2ρ1 ϕ1 k + 1),
2
θ2
q
w2 = (1 + ρ2 ϕ2 k + ϕ22 k 2 + 2ρ2 ϕ2 k + 1).
2
We look for conditions on the parameters ensuring there is no calendar-spread arbitrage,
that is ∀k, w2 ≥ w1 , i.e.:

Electronic copy available at: https://ssrn.com/abstract=2971502

 q q
θ2 (1 + ρ2 ϕ2 k + ϕ22 k 2 + 2ρ2 ϕ2 k + 1) ≥ θ1 (1 + ρ1 ϕ1 k + ϕ21 k 2 + 2ρ1 ϕ1 k + 1). (3.1)

Remark 3.1. The same problem is addressed in [2], for two SVI slices, and leads to the
study of the roots of a quartic polynomial. The corresponding conditions on the parameters
in this case were judged to be too intricate by the authors. Here we consider an easier
problem within the sub-family of SSVI slices.
In particular taking k = 0 and k = ±∞ gives:

θ2 ≥ θ1
 when k = 0,
θ2 ϕ2 (1 + ρ2 ) ≥ θ1 ϕ1 (1 + ρ1 ) when k = ∞, (3.2)

θ2 ϕ2 (1 − ρ2 ) ≥ θ1 ϕ1 (1 − ρ1 ) when k = −∞.


θ2 θ2 ϕ2 1+ρ1 θ2 ϕ2 1−ρ1
Since it is assumed that all ρi ∈ (−1, 1), we get: θ1 ≥ 1, θ1 ϕ1 ≥ 1+ρ2 and θ1 ϕ1 ≥ 1−ρ2 .

θ2 ϕ2
Writing: x := ϕ1 k, θ := θ1 , ϕ := ϕ1 , we get:
p p
θ(1 + ρ2 ϕx + ϕ2 x2 + 2ρ2 ϕx + 1) ≥ (1 + ρ1 x + x2 + 2ρ1 x + 1), (3.3)
and our conditions read:
1 + ρ1 1 − ρ1
θ ≥ 1, θϕ ≥ and θϕ ≥ . (3.4)
1 + ρ2 1 − ρ2
So the question is: assuming the previous necessary conditions hold, is the inequality
(3.3) in force, or do we need additional conditions?

3.1 From a quartic to a quadratic polynomial Q

The inequality (3.3) is rewritten as α + θz2 ≥ z1 where:

α := θ(1 + ρ2 ϕx) − (1 + ρ1 x) = θ − 1 + (θρ2 ϕ − ρ1 )x,

p p
z1 := x2 + 2ρ1 x + 1 and z2 := ϕ2 x2 + 2ρ2 ϕx + 1.

To analyse if the opposite can be true we look at the roots of the equation:

α + θz2 = z1 . (3.5)
Squaring twice, we get the following function P (z):

P := 4α2 θ2 z22 − (z12 − α2 − θ2 z22 )2 = 0, (3.6)

where any root of (3.5) is also a root of P and if at least two of these roots exist, the
slices intersect and calendar-spread arbitrage exists. For any root of P we again have the
same three possibilities:

Electronic copy available at: https://ssrn.com/abstract=2971502

 
2 2 2 2
2αθz2 = −(z1 − α − θ z2 )
 or,
2αθz2 = z12 − α2 − θ2 z22 and α + θz2 = −z1 or, (3.7)

2αθz2 = z12 − α2 − θ2 z22 and α + θz2 = z1 ,


where only the last case corresponds to an intersection of slices. Filling in the values for
α, z1 and z2 the polynomial P becomes:

P (x) = − (ρ1 − θϕρ2 )2 − (θϕ − 1)2 (ρ1 − θϕρ2 )2 − (θϕ + 1)2 x4

+4θ (ρ1 + ϕρ2 )((ρ1 − θϕρ2 )2 − θ2 ϕ2 − 1) + 2θϕ(ϕρ1 + ρ2 ) x3

+4θ(θ − 1)(ρ21 − θϕ2 ρ22 + θϕ2 − 1)x2 .

Here x2 is a factor of P , which can again be dropped as x = 0 is not a root of (3.5).
Hence, a polynomial Q of degree 2 remains, where:

P = x2 Q. (3.8)

3.2 Study of the roots of Q

The last stage is to investigate how the roots of the quadratic polynomial Q depend upon
the parameters. The discriminant of Q reads:

D = 16θ(ρ21 θ2 ϕ2 ρ22 + θ2 ϕ2 1)2 ((ρ1 θϕρ2 )2 − (θ − 1)(θϕ2 − 1)). (3.9)

It is positive and so two real roots exist if and only if:

(ρ1 − θϕρ2 )2 − (θ − 1)(θϕ2 − 1) > 0. (3.10)

If we define: Z(x) = z12 − θz22 − α2 , we can see that non zero roots x1 , x2 of Q(x) also
correspond to roots of (3.5) if Z(x)α < 0, or when α ≥ 0 and θz2 + α > 0.

3.2.1 Roots as a function of ρ1 and ρ2

Clearly, Q only has two real roots on thedomain I of all ρ = (ρ1 , ρ2 ) with ρ1 , ρ2 ∈ (−1, 1)
such that (3.10) holds. We define ρ∗ = ρ : ((ρ1 − θϕρ2 )2 − (θ − 1)(θϕ2 − 1) = 0 . If we
let x be a root of Q, with ρ in this domain, this reads:

2θ (ρ1 + ϕρ2 )((ρ1 − θϕρ2 )2 − θ2 ϕ2 − 1) + 2θϕ(ϕρ1 + ρ2 )

xρ =
((ρ1 − θϕρ2 )2 − (θϕ − 1)2 )((ρ1 − θϕρ2 )2 − (θϕ + 1)2 )
p
2(ρ21 − θ2 ϕ2 ρ22 − θ2 ϕ2 + 1) θ ((ρ1 − θϕρ2 )2 − (θ − 1)(θϕ2 − 1))
± .
((ρ1 − θϕρ2 )2 − (θϕ − 1)2 ) ((ρ1 − θϕρ2 )2 − (θϕ + 1)2 )
Using this value of xρ explicit expressions for α and Z can be obtained, namely α(xρ ) and
Z(xρ ). Furthermore, we also see that, since under (3.4) (ρ1 − θϕρ2 )2 < (θϕ − 1)2 , (θϕ + 1)2
they are continuous functions of ρ, we have the following:

Electronic copy available at: https://ssrn.com/abstract=2971502

Lemma 3.2. (ρ1 , ρ2 ) 7→ α(xρ )Z(xρ ) does not change sign for ρ1 , ρ2 ∈ (−1, 1).

Proof. Assume the opposite, meaning that for some α(xρ )Z(xρ ) = 0 for some ρ. Since:
2αθz2 = ±Z, we see that Z(xρ ) = 0 ⇔ α(xρ ) = 0, so if α(xρ ) = Z(xρ ) = 0:

( (
z12 − θ2 z22 = 0 θ2 (ϕ2 x2 + 2ρ2 ϕx + 1) = x2 + 2ρ1 x + 1
⇒
θ(1 + ρ2 ϕx) = 1 + ρ1 x θ2 (ρ22 ϕ2 x2 + 2ρ2 ϕx + 1) = ρ21 x2 + 2ρ1 x + 1
⇒ (1 − ρ22 )θ2 ϕ2 x2 = (1 − ρ21 )x2 (subtracting the two equations)
1 − ρ21
⇔ θ 2 ϕ2 = .
1 − ρ22
1+ρ1 1−ρ1
Under condition (3.4) this means that: θϕ = 1+ρ2 = 1−ρ2 . Hence: ρ1 = ρ2 = 1, which
contradicts their definition.

Lemma 3.3. For ρ∗ as defined above, Z(xρ∗ ) is strictly negative, and since α(xρ∗ ) ≥ 0 ⇔
0 < ϕ ≤ 1, it holds that: α(xρ∗ )Z(xρ∗ ) < 0 ⇔ 0 < ϕ ≤ 1.

Proof. First off, note that when ϕ = 1 the slices do not cross as we never have two roots
since this requires (ρ1 − θρ2 )2 > (θ − 1)2 , which does not hold. So we can assume ϕ 6= 1.
When ϕ < 1 it is possible that θϕ2 −1 < 0, meaning that (ρ1 −θϕρ2 )2 −(θ −1)(θϕ2 −1) > 0,
thus we always have two roots for any values of ρ1 , ρ2 ∈ (−1, ∗
q 1) and ρ does q not exist. If
we set in this case ρ∗ = (0, 0), we get α(xρ∗ ) = θ − 1, z1 = x2ρ∗ + 1 and z2 = ϕ2 x2ρ∗ + 1.
Hence: α(xρ∗ ) > 0 always holds and:

Z(xρ∗ ) = x2ρ∗ (θϕ2 − 1) − θ(θ − 1) < 0, (3.11)

also always holds. If however for ϕ < 1 it holds that θϕ2 − 1 > 0, then ρ∗ always exists,
and if we now define y 2 := (ρ1 − θϕ2 ) = (θ − 1)(θϕ2 − 1) we can find:

2θ (ρ1 + ϕρ2 )((ρ1 − θϕρ2 )2 − θ2 ϕ2 − 1) + 2θϕ(ϕρ1 + ρ2 )

xρ ∗ =
((ρ1 − θϕρ2 )2 − (θϕ − 1)2 )((ρ1 − θϕρ2 )2 − (θϕ + 1)2 )
2(θϕρ2 − ϕρ2 + y)
= .
(ϕ − 1)(ϕ + 1)

For our function Z(x) = z12 − α2 − θ2 z22 , we can see that:

Z(x) = x2 + 2ρ1 x + 1 − (θ − 1 − (ρ1 − θϕρ2 )x)2 − θ2 (ϕ2 x2 + 2ϕρ2 x + 1)

= x2 −(ρ1 − θϕρ2 )2 − θ2 ϕ2 + 1 + 2θx ((ρ1 − θϕρ2 ) − ρ2 ϕ(θ − 1)) − 2θ(θ − 1).

Now, if we fill xρ∗ into this equation, we get:

Electronic copy available at: https://ssrn.com/abstract=2971502

 2θ(θ − 1)
(2ϕy + ρ2 (2θϕ2 − ϕ2 − 1))2 + (1 − ρ22 )(ϕ − 1)2 (ϕ + 1)2 ,


Z(xρ∗ ) = − 2 2
(ϕ − 1) (ϕ + 1)
(3.12)
which is always negative regardless of the sign of y or ρ2 .

Now, we will look at α(xρ∗ ), which can be written as θ − 1 − xy. If we fill in our value
for xρ∗ we get:

θ−1
α(xρ∗ ) = − (2ϕρ2 y + 2θϕ2 − ϕ2 − 1). (3.13)
(ϕ − 1)(ϕ + 1)
2ϕρ2 y+2θϕ2 −ϕ2 −1
It follows that if θ > 1 and 0 < ϕ < 1, we have α(xρ∗ ) > 0 ⇔ ϕ−1 < 0. The
numerator is always strictly positive, since this is equivalent to:

(θϕ2 − 1) + ϕ2 (θ − 1) > −2ϕρ2 y, (3.14)

where we know the LHS of this inequality is always positive. So either the RHS is
negative, and it is true, or it is positive. In the latter case, we have:

4θ2 ϕ4 − 4θϕ4 − 4θϕ2 + ϕ4 + 2ϕ2 + 1 > 4ϕ2 ρ22 (θ2 ϕ2 − θϕ2 − θ + 1)

⇔ 4ϕ2 (θ − 1)(θϕ2 − 1)(1 − ρ22 ) + (ϕ2 − 1)2 > 0

Since this last inequality always holds, we can conclude that α(xρ∗ ) > 0 ⇔ ϕ < 1.
Combining this with the result for Z(xρ∗ ) allows to conclude.

By combining Lemmas 3.2 and 3.3, we see that if conditions (3.4) hold with ϕ ≤ 1, then
α(x)Z(x) < 0. This would mean that in this case the slices never intersect. Next we look
at what happens when ϕ > 1.
Lemma 3.4. If P has a root xρ , then: (α + θz2 )(xρ ) > 0 ⇔ ϕ > 1.
Proof. We start by showing that the function xρ 7→ (α + θz2 )(xρ ) never changes sign. This
is clear, since if the opposite were true, by continuity there would exist ρ1 , ρ2 such that
(α + θz2 )(xρ1 ,ρ2 ) = 0. However, (3.5) then implies that: z1 (xρ1 ,ρ2 ) = 0, which is impossible
since ρ1 , ρ2 ∈ (−1, 1).

Knowing this, we look at the value of (α + θz2 )(xρ∗ ) to see whether this function is
always positive or negative. We find that:

p !
(2ϕy + ρ2 (2θϕ2 − ϕ2 − 1))2 + (ϕ2 − 1)2 (1 − ρ22 ) − 2ϕρ2 y − 2θϕ2 + ϕ2 + 1
(α + θz2 )(xρ∗ ) = θ
ϕ2 − 1
2ϕρ2 y + 2θϕ2 − ϕ2 − 1
+ ,
ϕ2 − 1

Electronic copy available at: https://ssrn.com/abstract=2971502

 −ϕ −1 2 2
From the proof of Lemma 3.3 it follows that 2ϕρ2 y+2θϕ
ϕ2 −1
> 0 ⇔ ϕ > 1. Moreover
p
(2ϕy + ρ2 (2θϕ − ϕ − 1)) + (ϕ − 1) (1 − ρ2 ) − 2ϕρ2 y − 2θϕ + ϕ2 + 1 = 0, since:
2 2 2 2 2 2 2

q
(2ϕy + ρ2 (2θϕ2 − ϕ2 − 1))2 + (ϕ2 − 1)2 (1 − ρ22 ) = 2ϕρ2 y + 2θϕ2 − ϕ2 − 1
⇔ (2ϕy + ρ2 (2θϕ2 − ϕ2 − 1))2 + (ϕ2 − 1)2 (1 − ρ22 ) = (2ϕρ2 y + 2θϕ2 − ϕ2 − 1)2
⇔ 4ϕ2 (y 2 − θ2 ϕ2 + θϕ2 + θ − 1)(1 − ρ22 ) = 0,

where the second step of squaring both sides is allowed since it is known from the
previous results that both sides are always positive. Since y 2 = θ2 ϕ2 − θϕ2 − θ + 1, we see
that this final equality holds. Combining this gives: (α + θz2 )(xρ∗ ) > 0 ⇔ ϕ > 1 and the
result follows.

As a conclusion, we have investigated all the cases in which the roots of Q are also roots
of (3.5). We can state the following necessary and sufficient conditions:
θ2 ϕ2
Proposition 3.5. Let θ = θ1 and ϕ = ϕ1 .
Then, for the 2 slices w1 and w2 to satisfy
 
w2 ≥ w1 , it is necessary that θ > 1 and θϕ > max 1+ρ 1 1−ρ1
,
1+ρ2 1−ρ2
and sufficient that

ϕ ≤ 1 or (ρ1 − θϕρ2 )2 ≤ (θ − 1)(θϕ2 − 1). (3.15)

This result is already of practical interest: those consistency constraints can be used
to calibrate jointly several maturity slices, and the discrete set of calibrated slices can be
extended continuously to an arbitrage-free volatility surface using for instance the method
described in [2].

4 Going to continuous time

To obtain continuous-time conditions, we informally write θ1 = θ and θ2 − θ1 = dθ, and
compute the 1st order expansions of the previous conditions in dθ. Likewise, we write
ϕ0(θ)dθ
ϕ2 = ϕ1 + ϕ0(θ)dθ so that ϕ2
ϕ1 = 1 + ϕ . Let us set:

1 d(θϕ) ϕ0(θ)
γ := =1+θ . (4.1)
ϕ dθ ϕ

4.1 Necessary conditions

θ2 dθ
We rewrite the necessary condition θ1 ≥ 1 as dt ≥ 0. Regarding the second necessary
condition:
(
1−ρ1
θϕ ≥ 1−ρ2 ,
1+ρ1 . (4.2)
θϕ ≥ 1+ρ2

Electronic copy available at: https://ssrn.com/abstract=2971502

 we set: ρ1 = ρ, and ρ2 = ρ + ρ0(θ)dθ.

dθ ϕ0(θ)
1 + ϕ dθ = 1 + γ dθ


Using θϕ = 1 + θ θ we get:
( 1−ρ
1 + γ dθ
θ ≥ 1−ρ−ρ0(θ)dθ ,
1+ρ (4.3)
1 + γ dθ
θ ≥ 1+ρ+ρ0(θ)dθ ,
or:
(
1 − ρ − ρ0(θ)dθ + (1 − ρ)γ dθ 2
θ + O(dθ ) ≥ 1 − ρ, (4.4)
1 + ρ + ρ0(θ)dθ + (1 + ρ)γ dθ 2
θ + O(dθ ) ≥ 1 + ρ.
This is equivalent to
(
(−θρ0(θ) + (1 − ρ)γ) dθ
θ ≥ 0,
dθ
(4.5)
(θρ0(θ) + (1 + ρ)γ) θ ≥ 0.
Now, given that dθ/θ ≥ 0, setting:

δ := θρ0(θ). (4.6)
we can write:
(
δ ≤ (1 − ρ)γ,
(4.7)
δ ≥ −(1 + ρ)γ.
So the temptative continuous time necessary conditions are:

dθ
≥ 0 and − γ ≤ δ + ργ ≤ γ.
dt

4.2 Sufficient conditions

Besides these necessary conditions, we saw from Proposition 3.5 that the slices did not cross
if either: ϕ ≤ 1 or (ρ1 − θϕρ2 )2 ≤ (θ − 1)(θϕ2 − 1). So in this infinitesimal case we get:

ϕ0(θ) ϕ0(θ)
1+ dθ ≤ 1 ⇔ ≤0
ϕ ϕ
ϕ0(θ)
⇔ 1+θ ≤1
ϕ
⇔ γ ≤ 1.

Setting ρ1 = ρ and ρ2 = ρ + ρ0(θ)dθ we get:

(ρ − (ρ + ρ0(θ)dθ)θϕ)2 ≤ (θ − 1)(θϕ2 − 1)
⇔ (ρ(1 − θϕ) − θϕρ0(θ))2 ≤ (θ − 1)(θϕ2 − 1).

10

Electronic copy available at: https://ssrn.com/abstract=2971502

 ϕ0(θ)
Like in the previous cases, when we substitute θ and ϕ with (1 + dθθ ) and (1 + ϕ dθ)
respectively, we get that: (1 − θϕ) = − dθ dθ 2 dθ
θ γ, (θ − 1) = θ and (θϕ − 1) = θ (2γ − 1). Filling
this into the equation above results in:

   2
dθ dθ 2 dθ
(−ργ − γ + 1 ρ0(θ)dθ) ≤ (2γ − 1)
θ θ θ
 2  2
dθ 2 4 dθ
⇔ (−ργ − θρ0(θ)) + O((dθ) ) ≤ (2γ − 1).
θ θ
By eliminating the higher order term of dθ and by cancelling out (dθ/θ)2 , using δ : θρ0(θ),
we get:

δ 2 + 2ργδ + (ρ2 γ 2 − 2γ + 1) ≤ 0, (4.8)

where this inequality is only true when δ lies between the roots of this polynomial, so
that:
p
−2ργ ± 4ρ2 γ 2 − 4ρ2 γ 2 + 8γ − 4 p
δ1,2 = = −ργ ± 2γ − 1. (4.9)
2
√ √
Here the condition becomes − 2γ − 1 ≤ δ + ργ ≤ 2γ − 1.

4.3 eSSVI continuous calendar-spread conditions

We get the conditions: (δ+ργ)2 ≤ γ 2 and either: (δ+ργ)2 ≤ γ or (δ+ργ)2 ≤ 2γ−1. Observe
that when γ ≤ 1, (δ + ργ)2 ≤ γ 2 implies (δ + ργ)2 ≤ γ so that the second set of conditions
is automatically fulfilled. When γ > 1, the conditions are equivalent to (δ + ργ)2 ≤ 2γ − 1.
The following result is therefore suggested from the previous small dθ expansions:
1
Theorem 4.1. Define γ := ϕ(θ) and δ := θ∂θ (ρ(θ)). The eSSVI surface is free of calendar-
spread arbitrage if and only if ∂t θt > 0 and
(
2 γ2 if 0 ≤ γ ≤ 1,
(δ + ργ) ≤ (4.10)
2γ − 1 if γ > 1.
It remains to prove this result rigorously. This is done in the Appendix, using the same
kind of arguments as the discrete time.

5 Parametric families for ρ(θ)

In this section we obtain a general expression for ρ(θ) in term of ϕ(θ), and use it to build
parametric families for ρ(θ) for the two types of power law families for ϕ(θ) proposed
by Gatheral and Jacquier which will fulfill the no calendar-spread conditions up to some
maximum maturity corresponding to a total ATM variance θmax . Note that this restriction
is not important in practice: typically on Equity markets, liquid vanilla options are available
up to 2 years only.

11

Electronic copy available at: https://ssrn.com/abstract=2971502

5.1 General function ϕ(θ)
In most practical cases it will be found that γ ≤ 1. Therefore, for the calendar-spread
abitrage conditions, looking at the condition |δ + ργ| ≤ γ, which can be written as:

∂θ (θϕ(θ))
|θ∂θ (ρ(θ)) + ρ(θ)∂θ (θϕ(θ))| ≤ , (5.1)
ϕ(θ)
∂(θϕ(θ)ρ(θ)) ∂(ρ(θ)) ∂(θϕ(θ))
= θϕ(θ) + ρ(θ) . (5.2)
∂θ ∂θ ∂θ
Dividing (5.1) by ϕ(θ) our condition becomes:

|∂θ (θϕρ)| ≤ ∂θ (θϕ). (5.3)

Therefore there exists a function u(θ) with values in [−1, 1] such that: ∂θ (θϕ(θ)ρ(θ)) =
u(θ)∂θ (θϕ(θ)). Integrating both sides gives us, using θϕ(θ) → 0 as θ → 0:
Z θ
θϕ(θ)ρ(θ) = u(τ )∂τ (τ ϕ(τ ))dτ. (5.4)
0
If both sides are divided by θϕ(θ) and the following change of variable is performed:
τ = rθ, with r ∈ [0, 1] the following proposition can be found:
Proposition 5.1. Let u any measurable function with values in [−1, 1]. Then the function
ρ defined by Z 1
1
ρ(θ) = u(rθ)∂r (rϕ(rθ))dr (5.5)
ϕ(θ) 0
fulfills the no calendar-spread arbitrage conditions.
From this we see that given a curvature function ϕ(θ) choosing u(θ) such that for all
1
R1
θ > 0: ϕ(θ) 0 u(rθ)∂r (rϕ(rθ))dr ≤ 1, produces an admissible ρ(θ).

From (5.5) and the fact that we can write u(θ) = ∂θ∂(θϕ(θ)ρ(θ))
θ (θϕ(θ))
, we can either choose a
function for ρ(θ) and check whether the resulting function u(θ) lies in [−1, 1], or choose
such a function u(θ) and obtain ρ(θ).

Note that we should check that ρ(θ) lives in (−1, 1). To this end defining: u∗ (θ) =
supr∈[0,1] |u(rθ)| and using ∂θ (θϕ(θ))
ϕ(θ) ≥ 0, we obtain:

Z 1 Z 1
1 1
u(rθ)∂r (rϕ(rθ))dr ≤ |u(rθ)|∂r (rϕ(rθ))dr
ϕ(θ) 0 ϕ(θ) 0
u∗ (θ) 1
Z
≤ ∂r (rϕ(rθ))dr
ϕ(θ) 0
[rϕ(rθ)]10
= u∗ (θ)
ϕ(θ)
= u∗ (θ)

12

Electronic copy available at: https://ssrn.com/abstract=2971502

 From the assumptions on u(θ) we know that u∗ (θ) ≤ 1, and therefore we see that
ρ(θ) ∈ (−1, 1) always holds.

5.2 The first power-law family

Consider the case ϕ(θ) = ηθ−λ so that the condition ∂θ (θϕ(θ))/ϕ(θ) ≤ 1 is equivalent to
0 < λ ≤ 1. Then it is readily checked from the general expression above that the choice
 a
1+a−λ τ
u(τ ) = ρ0 + (ρm − ρ0 ) ,
1−λ θmax
leads to:
 a
θ
ρ(θ) = ρ0 + (ρm − ρ0 ) . (5.6)
θmax
Under the assumption a ≥ 0 it is a monotonous function from (0, ρ0 ) to (θmax , ρm ). The
requirement |u(τ )| ≤ 1 reads:
( (1−λ)(1−ρ )
m
ρm −ρ0 if ρ0 < ρm
a ≤ (1−λ)(1+ρ m)
(5.7)
ρ0 −ρm if ρ0 > ρm

5.3 The second power-law family

Now consider the family ϕ(θ) = ηθ−λ (1 + θ)λ−1 . We get then:

Z 1
1 
−λ λ−1

ρ(θ) = u(rθ)∂ r rη(rθ) (1 + rθ) dr
ηθ−λ (1 + θ)λ−1 0
Z 1
1−λ
= u(rθ)r−λ (1 + rθ)λ−2 dr.
(1 + θ)λ−1 0
The choice
  a
a(1 + τ ) τ
u(τ ) = ρ0 + (ρm − ρ0 ) 1 + ,
1−λ θmax
leads to
 a
θ
ρ(θ) = ρ0 + (ρm − ρ0 ) .
θmax
As above it is a monotonous function from (0, ρ0 ) to (θmax , ρm ) assuming a ≥ 0, and
the requirement |u(τ )| ≤ 1 reads:
( (1−λ)(1−ρ )
m
max )(ρm −ρ0 )
if ρ0 < ρm
a ≤ (1+θ(1−λ)(1+ρm )
(1+θmax )(ρ0 −ρm ) if ρ0 > ρm

13

Electronic copy available at: https://ssrn.com/abstract=2971502

5.3.1 Another example of parametric ρ
We can also go the reverse way to find whether a chosen function of ρ(θ) satisfies the
arbitrage-free conditions. A possible option is to choose the exponential family ρ(θ) =
ρm + (ρ0 − ρm )a−θ . It will be a monotonous function for positive a from ρ(0) = ρ0 to
limθ→∞ ρ(θ) = ρm .
If we work with the first power-law form for ϕ, so: ϕ(θ) = ηθ−λ , this corresponds to the
following choice for u:
 
θ log(a)
u(θ) = ρm + 1 − (ρ0 − ρm )a−θ
1−λ
Solving this shows that u(θ) ∈ [−1, 1] requires, besides the obvious conditions |ρ0 | ≤ 1
and |ρm | ≤ 1:
ρ0 − ρm
|ρm − exp(λ − 2)| ≤ 1
1−λ
This non linear condition on λ can be solved numerically.

5.4 No butterfly arbitrage

Since eSSVI slices are SSVI slices, we have the same sufficient conditions for no butterfly
arbitrage as in SSVI. Yet we must pay attention to the fact that ρ is now a function of θ,
so that we get a family of conditions to fulfill for the parameters driving the function ρ(θ).
Let us illustrate this in the case of the first power-law family, with the parametric family
for ρ designed in section 5.2.
The sufficient condition for no butterfly arbitrage stated in [2] can now be altered to
read for all θ ≤ θmax :
!
4θλ−1 2θλ−1/2
η ≤ min ,p
1 + |ρ(θ)| 1 + |ρ(θ)|
Since ρ is monotonous, the condition above is equivalent to
!
4θλ−1 2θλ−1/2
η ≤ min ,p .
1 + max(|ρm |, |ρ0 |) 1 + max(|ρm |, |ρ0 |)
and using the assumption 0 < λ < 1/2, it is still equivalent to
!
λ−1 λ−1/2
4θmax 2θmax
η ≤ min ,p .
1 + max(|ρm |, |ρ0 |) 1 + max(|ρm |, |ρ0 |)

14

Electronic copy available at: https://ssrn.com/abstract=2971502

6 Calibration results
We now go through the steps required to calibrate the eSSVI model to real market data
and compare the results with a calibrated plain SSVI surface. We will use the power-law
form for ϕ, so: ϕ(θ) =ηθ−λ ,and the parametric shape computed above for the correlation:
a
θ
ρ(θ) = ρ0 − (ρ0 − ρm ) θmax .

6.1 eSSVI parameters versus SSVI ones

In this section, we use DJX vanilla option quotes on the 29/01/2016, obtained from the
freely available delayed option quotes on the CBOE web site.

Figure 2: Comparison of the calibrated functions of ϕ(θt ) and ρ(θt ) for SSVI and eSSVI.

First, the difference in the functions of ϕ and ρ is shown in Figure 2.We have the two
functions for ϕ, which we found were 0.84θt−0.48 for SSVI and 0.98θt−0.42 for eSSVI. The plot
shows that there is hardly any difference here. For ρ, we found the constant value in SSVI of:
ρ = −0.82 and the function found for eSSVI: ρ(θt ) = −0.76 − 0.11 (θt /2.9)0.51 . From (5.2)
it can be seen that these found parameters ensure a surface free of both calendar-spread
and butterfly arbitrage. It can be seen from Figure 2 that this constant value is somewhat
of a weighted average of our function over the given values of θt .

6.2 Calibration accuracy: eSSVI, SSVI and slicewise SVI

Next, from the found parameters we plot the volatilities implied by both the SSVI and
eSSVI model and those implied by the market in Figure 3.

15

Electronic copy available at: https://ssrn.com/abstract=2971502

Figure 3: Market bid and ask and SSVI and eSSVI volatility slices for the first, middle and
last maturity.

Lastly, the average fitting error, for which we took the average 2-norm difference be-
tween model and market price in bps of the spot price per maturity, is displayed in Figure 4
for SSVI, eSSVI and SVI (for each slice). As expected, SVI gives by far the most accurate
fit, however it can also be seen that for short maturities eSSVI outperforms SSVI.

16

Electronic copy available at: https://ssrn.com/abstract=2971502

 Figure 4: Average fitting error per maturity

7 Conclusion
In this paper we have designed an extension of Gatheral and Jacquier SSVI surface where the
leverage parameter ρ is allowed to depend upon the maturity. We have obtained necessary
and sufficient conditions for the absence of calendar-spread arbitrage, an explicit generic
parameterisation for the correlation parameter, and explicit tractable families for the usual
choices of curvature functions ϕ(θ). Our calibration experiments on Equity Indexes suggest
that the gain in fitting quality of eSSVI is small in normal market conditions for long
maturities, but that the fitting quality is practically doubled for short maturities. The
eSSVI surface might also better fit in stressed market conditions.

References
[1] J. Gatheral, The Volatility Surface: A Practitioner’s Guide, Wiley Finance, 2006.

[2] J. Gatheral and A. Jacquier, Arbitrage-free svi volatility surfaces, Quantitative

Finance, 14 (2014), pp. 59–71.

[3] G. Guo, A. Jacquier, C. Martini, and L. Neufcourt, Generalized arbitrage-free

svi volatility surfaces, SIAM Journal on Financial Mathematics, 7 (2016), pp. 619–641.

17

Electronic copy available at: https://ssrn.com/abstract=2971502

A Appendix
In this appendix we provide a direct proof of the no calendar-spread conditions in continous
time (section 4.3). We start by stating the following lemma:

Lemma A.1. For any x ∈ R and fixed ζ, γ such that: ζ 2 > 2γ − 1 and ζ 2 ≤ γ 2 , the
function:
ρ 7→ (xζ + 1)(x2 γ + x(ζ + ρ) + 1), (A.1)
where x2 γ + x(ζ + ρ) + 1 = f (x, ρ)(xζ + 1) with f (x, ρ) a strictly positive function, does not
change sign for any ρ ∈ (−1, 1).

Proof. If the opposite were true, the continuity of the function entails that: (xζ + 1) =
(x2 γ + x(ζ + ρ) + 1) = 0, meaning that:
(
x = −1/ζ,
(A.2)
γ − ζρ = 0.

However, since it is required that |ζ| ≤ γ and ρ ∈ (−1, 1), the second equation of (A.2)
cannot hold, meaning that (A.1) has no zeros and therefore does not change sign.

We turn now to the proof of Theorem 4.1, that we recall for convenience:

Proof. The eSSVI surface is free of calendar-spread arbitrage if and only if: ∂t w(k, θt ) ≥ 0
which is equivalent to the fact that ∂t θt ≥ 0 and:
p p
x2 γ + x(1 + x2 + 2xρ + 1)ζ + xρ + 1 + x2 + 2xρ + 1 ≥ 0, (A.3)
where x := kϕ(θ), and γ and ζ as defined above. When evaluating (A.3) at x → ±∞,
the following condition is found:

|ζ| ≤ γ. (A.4)
Knowing this, it can be asserted that (A.3) holds when the following equation has at
most one real root:
p p
x2 γ + x(1 + x2 + 2xρ + 1)ζ + xρ + 1 + x2 + 2xρ + 1 = 0. (A.5)
In order to evaluate the roots of (A.3), the roots of the following equation are inspected:

2
x2 γ + x(ζ + ρ) + 1 = (xζ + 1)2 (x2 + 2xρ + 1)
⇔ x2 x2 (ζ 2 − γ 2 ) + 2x(ρζ 2 − ζγ + ζ − ργ) + 2ρζ − 2γ + 1 − ρ2 = 0.

As it is clear that x = 0 is not a real root of this equation, this factor can be dropped
and we are left with the roots of the following equation:

18

Electronic copy available at: https://ssrn.com/abstract=2971502

 x2 (ζ 2 − γ 2 ) + 2x(ρζ 2 − ζγ + ζ − ργ) + 2ρζ − 2γ + 1 − ρ2 = 0. (A.6)
There are now two cases to ensure that (A.5) has at most a single real root: either (A.6)
has at most a single root, meaning that its discriminant is non-positive, or the two roots of
(A.6) are not roots of (A.5) as they arose from the squaring process.

For the first case, the discriminant of (A.6) can be computed expicitly:

D = 4(ρζ − γ)2 (ζ 2 − 2γ + 1), (A.7)

which is non-positive if and only if: ζ 2 ≤ 2γ − 1.

If this does not hold, (A.6) has two real roots, x1,2 . However, these roots do not
neccesarily correspond to roots of (A.5). This is the case if and only if (1 + xζ)(x2 γ + x(ζ +
ρ) + 1) < 0. Using Lemma A.1, it is sufficient to show that this holds for ρ = 0, as then the
result follows for all other possible values of ρ. In this case we have the following values for
x1,2 :
p
ζ(1 − γ) ± γ ζ 2 − 2γ + 1
x1,2 = .
γ2 − ζ 2
Now, under the conditions ζ 2 ≤ γ 2 and ζ 2 ≥ 2γ − 1, the following needs to hold:

(1 + x1,2 ζ)(x21,2 γ + x1,2 ζ + 1) ≤ 0. (A.8)

Clearly, from the fact that γ ≥ 0, this requires: 1 + x1,2 ζ ≤ 0 and therefore also:
x21,2 γ + x1,2 ζ + 1 ≥ 0. Firstly, 1 + x1,2 ζ ≤ 0 can be written as:

ζ(1 − γ) + γy γ − ζ 2 + ζy
1+ζ ≤ 0 ⇔ ⇔ γ − ζ 2 + ζy ≤ 0, (A.9)
γ2 − ζ 2 γ2 − ζ 2
p
with y := ± ζ 2 − 2γ + 1. As this needs to hold for both the positive and negative
value of y, this in turn requires γ − ζ 2 < 0, which yields the following:

(γ − ζ 2 )2 ≥ ζ 2 (ζ 2 − 2γ + 1) ⇔ γ 2 ≥ ζ 2 , (A.10)
which indeed holds.

If now also ζ 2 > γ is required, combined with the condition γ 2 ≥ ζ 2 , this implies that
γ > 1. The following can now be written:

γ
x21,2 γ + x1,2 ζ + 1 ≥ 0 ⇔ (ζ 4 + ζ 2 γ 2 − 3ζ 2 γ + ζ 2 − γ 3 + γ 2 + (ζ 3 + ζγ 2 − 2ζγ)y) ≥ 0.
(γ 2 − ζ 2 )2
(A.11)
This now requires:

19

Electronic copy available at: https://ssrn.com/abstract=2971502

 ζ 4 + ζ 2 γ 2 − 3ζ 2 γ + ζ 2 − γ 3 + γ 2 ≥ 0. (A.12)
If this holds, from the fact that (A.11) needs to hold for both values of y, it can be
squared in such a sense that it can be written as:

γ 2 + ζ 2 − 2γ + 1 ≥ 0. (A.13)
which always holds under the stated requirements. However, (A.12) can also be written
as:

ζ 2 (ζ 2 − 2γ + 1) + γ(ζ 2 − γ)(γ − 1) ≥ 0, (A.14)

which also holds under the stated requirements.

In summary, we can conclude that the statement in the theorem holds if besides ∂t θt ≥ 0
it is required that ζ 2 ≤ γ 2 and either: ζ 2 ≤ γ or ζ 2 ≤ 2γ − 1. Observe that when γ ≤ 1,
ζ 2 ≤ γ 2 implies ζ 2 ≤ γ so that the second set of conditions is automatically fulfilled. When
γ > 1, the conditions are equivalent to ζ 2 ≤ 2γ − 1, so that we get eventually:
(
2 γ2 if 0 ≤ γ ≤ 1,
ζ ≤ (A.15)
2γ − 1 if γ > 1.

The result of the theorem then follows. Note that from its definition, ζ can be written
as δ + ργ. From this the result in section 4.3 is proven.

20

Electronic copy available at: https://ssrn.com/abstract=2971502