Quasi-Explicit Calibration of Gatheral’s SVI model

Zeliade White Paper September, 2009

Zeliade Systems

ZWP-0005
Zeliade Systems SAS
56 rue Jean-Jacques Rousseau 75001 Paris France
Phone : +(33) 1 40 26 17 29 Fax : +(33) 1 40 26 17 81
e-mail : [email protected]

T ITLE : Quasi-Explicit Calibration of Gatheral’s SVI model

N UMBER : ZWP-0005
N UMBER OF PAGES : 10
F IRST VERSION : June, 2009
C URRENT VERSION : September, 2009
R EVISION : 1.2.0

This document must not be published, transmitted or reproduced without permission from Zeliade Systems SAS.
Copyright c 2007–2009 Zeliade Systems SAS. All Rights Reserved.
Trademarks: Zeliade Systems is a registered trademark of Zeliade Systems SAS. All other company and product names referenced in this document are used for identification
purposes only and may be trade names or trademarks of their respective owners.
Legal disclaimer: ZELIADE SYSTEMS SAS AND/OR ITS SUPPLIERS MAKE NO REPRESENTATIONS ABOUT THE SUITABILITY OF THE INFORMATION CONTAINED
IN THIS DOCUMENT FOR ANY PURPOSE. THIS DOCUMENT IS PROVIDED ”AS IS”, WITHOUT WARRANTY OF ANY KIND. ZELIADE SYSTEMS SAS AND/OR ITS
SUPPLIERS HEREBY DISCLAIM ALL WARRANTIES AND CONDITIONS OF WITH REGARD TO THIS INFORMATION, INCLUDING ALL IMPLIED WARRANTIES AND
CONDITIONS OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE AND NON INFRINGEMENT. IN NO EVENT SHALL ZELIADE SYSTEMS SAS
BE LIABLE FOR ANY SPECIAL, INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFIT
ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF INFORMATION AVAILABLE IN THIS DOCUMENT.
Contents

1 A simple model and a delicate calibration 1

2 Parameter constraints and limiting cases 2

2.1 Slopes and minimum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2 Arbitrage constraints (b and ρ) . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 Limiting cases σ → 0 and σ → ∞ (almost-affine smiles) . . . . . . . . . . 3

3 Indeed, not more than a linear problem 5

3.1 Dimension reduction: drawing out the linear objective . . . . . . . . . . 5
3.2 Explicit solution of the reduced problem . . . . . . . . . . . . . . . . . . . 6

4 Numerical Results 7

5 Conclusions 9
 Abstract

We present a procedure - based on dimension reduction in parameters space - provid-

ing a quasi-explicit calibration of J. Gatheral’s SVI model for implied variance. The
resulting parameter identification is reliable and stable.
1 A simple model and a delicate calibration

Jim Gatheral’s SVI model [1] describes implied variance with the following parametric
form:  
2
p
v(x) = σBS (x) = a + b ρ(x − m) + (x − m)2 + σ 2 , (1)

where v is the implied variance of market prices of Vanillas at fixed time-to-maturity T ,

x the log-forward moneyness x = log(K/FT ) and a, b, ρ, m, σ are the model parameters.
Parametric models (e.g. SVI or the functional forms obtained by Taylor’s approxima-
tion in CEV or SABR models) are of common use in the treatment of the volatility
surface. Apart from the extrapolation of smile points, they provide a smooth repre-
sentation of the market smile and consequently facilitate the calibration of stochastic
models for the underlying (including the reconstruction of a local volatility surface via
Dupire’s formula, for which interpolation in time must also be taken into account). It
is well known that the SVI parametric form (1) proves to have outstanding calibration
performances to single-maturity slices of the implied smile on many Equity indexes.
Nevertheless, it is also common knowledge that the least-square calibration of (1) is
typically affected by the presence of several local minima. To our experience, even
when SVI parameterization is calibrated to simulated data, i.e. a smile produced by
SVI itself, local minima that are difficult to sort out (least square objective ≈ 10−8 for
√
reasonable volatility values, v ≈ 20% − 40%) are found far away from the global one
(objective = 0).
This unpleasant feature tends to bring some difficulties if one wants to design a pa-
rameter identification strategy for SVI model which is sufficiently robust and stable.
The solution yield by a least square optimizer usually has a strong dependence on the
input starting point. Then, smart initial guesses of parameter sets can be made by look-
ing to the ’geometry’ of the observed smile (asymptotic slopes, minimum value), and
on the other hand the calibration can be restarted from several different initial guesses
and/or using more than one non-linear optimizer. Nevertheless, usual strategies to
find the initial guess are not defect-free and require attention, since the desired smile
features are not available in all the cases and often not for all maturities (e.g. the wings
are not both observed or the smile has no clearly visible minimum). The calibration
reset, though useful, still does not guarantee that one manages to overcome all the lo-
cal minima and may anyhow lead to ambiguous choices of optimal parameters, since
the same smile can be - remarkably well - calibrated with sets of parameters that are
totally different one from the other. The big issue, then, is the stability of calibrated
parameters with respect to time-to-maturity. This is a feature which comes into play in
a significant way when trying to parameterize the whole volatility surface.
This document presents a procedure providing a trustworthy and stable calibration of
SVI parametric form (1), which has the pleasant feature of not being strongly sensitive

c Zeliade Systems SAS

1/10 ZWP-0005– September, 2009
to initial parameter guess. We rely on some simple observations on the symmetries of
the functional form (1) to downsize the minimization problem from dimension 5 (the
number of parameters in (1)) to dimension 2 (namely, m and σ), while the optimization
over the remaining 3 is performed explicitly. Last but not least, the method yields an
optimal parameter set which is automatically consistent with the arbitrage constraint
on the slopes of implied variance. The procedure for the Quasi-Explicit calibration is
presented in section 3, while section 2 discusses the constraints that are introduced in
the parameters space.

2 Parameter constraints and limiting cases

The parameters a, b, ρ, m, σ in general depend on time-to-maturity T . We assume that

b, σ, ρ satisfy
b > 0, σ ≥ 0, ρ ∈ [−1, 1].

Further conditions on b and ρ follow from well known arbitrage conditions (cf section
2.2). In addition, we will discuss some constraints on parameters σ and a which are
related to the well-posedness of the calibration problem (section 2.3). We recall once
again that in the current document we just look to the parameterization of time-slices
of implied variance.

2.1 Slopes and minimum

We review the main interesting properties of the parametric form (1). The left and right
asymptotes are respectively (cf.[1])

vL (x) = a − b(1 − ρ)(x − m),

vR (x) = a + b(1 + ρ)(x − m).

The term adding to a in (1) is always positive and convex w.r.t.x. v has a unique mini-
mum point if ρ2 6= 1, in particular:

1 − ρ2 attained at x∗ = m − √ρσ 2 ;
p
• if ρ2 6= 1, the minimum is a + bσ
1−ρ

• if ρ2 = 1, v is non-increasing for ρ = −1 and non-decreasing for ρ = 1 and

– if σ 6= 0, v is strictly monotone and the minimum is never attained (never-

theless, v → a for very positive or very negative x);
– if σ = 0, v has the shape of a Put or Call payoff of strike m (v is worth a for
x ≥ m if ρ = −1 and for x ≤ m if ρ = 1).

c Zeliade Systems SAS

2/10 ZWP-0005– September, 2009
 0.2

0.16 SVI with ρ ≠ −1 0.19 SVI with ρ = −1

0.18

0.14
0.17

0.16
0.12
* 2 1/2
x = m − ρ σ / (1 − ρ ) 0.15

0.14
0.1
0.13

0.12
2 1/2
min = a + b σ (1 − ρ )
0.08

0.11

0.06 0.1
−0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2

Figure 1: Examples of SVI smile shapes. SVI parameters are a = 0.04, b = 0.4, ρ =
−0.4, m = 0.05, σ = 0.1 (left) and a = 0.04, b = 0.2, ρ = −1, m = 0.1, σ = 0.5 (right).

2.2 Arbitrage constraints (b and ρ)

A necessary condition for the absence of arbitrage is a constraint on the maximal slopes
of total implied variance T v(x). As found in [2], this condition reads

∀ x, ∀ T, |T ∂x v(x)| ≤ 4. (2)

As stated in [1], this translates into the following equivalent condition on b and ρ :

4
b≤ . (3)
(1 + |ρ|)T

2.3 Limiting cases σ → 0 and σ → ∞ (almost-affine smiles)

As observed in section 2.1 in the case ρ2 = 1, letting σ → 0 gives a piecewise affine

parameterisation of variance. In the two regions x < m and x > m, variance reads
respectively
v(x) = a + b(ρ ∓ 1)(x − m). (4)

Smiles which can be excellently fitted with an affine (monotone) parameterisation v(x) =
px + q (we will refer to these as to “almost-affine” smiles) are not uncommon on Equity
indexes, in particular for large maturities. Clearly, the calibration of SVI model to an
almost-affine smile is an ill-posed problem, in the sense that there exists infinitely many
solutions to the minimization problem. Indeed, if we think of a downward smile to fix
ideas, it is sufficient to let σ → 0 and take m to be greater than the largest observed
log-moneyness (to pick the minus sign in (4)) and the matching of the two relevant

c Zeliade Systems SAS

3/10 ZWP-0005– September, 2009
quantities, i.e. smile slope p and intercept q, yields the two equations

b(ρ − 1) = p
a − bm(ρ − 1) = q,

corresponding to infinitely many choices of the parameters a, b, ρ.

The same kind of limiting behaviour is attained in the limit σ → ∞ and a → −∞, in
the precise way specified as follows. A priori, indeed, negative values of a could be
allowed, since the positivity of parameterisation (1) is simply achieved by asking that
the minimum of v (when attained) be non-negative, i.e.
p
a ≥ −bσ 1 − ρ2

(if the minimum is not attained, then ρ2 = 1 and the condition becomes a ≥ 0).
Assume then a < 0 and σ >> 1, so that
 p 
v(x) = a + b ρ(x − m) + σ 2 + (x − m)2
r
(x − m)2
= −|a| + bρ(x − m) + bσ 1 +
σ2
 (x − m)2 
∼ −|a| + bρ(x − m) + bσ 1 +
2σ 2
2
(x − m)
∼|a|=bσ bρ(x − m) + b .
2σ

Hence
lim v(x) = bρ(x − m)
σ→∞,a→−∞,|a|=bσ

for any value of x, and this correspond again to an affine smile whose slope and inter-
cept identify the product bρ and the parameter m, but not b, ρ and m separately.
Smiles tend to flatten with increasing time to maturity, and curved smiles can con-
tinuously deform into almost-affine ones. Since the stability of the calibration is the
features we have in mind, we would like the calibration strategy to avoid falling into
the instable behaviour caused by limiting SVI cases. Hence, we restrict ourselves to the
situation where:
σ ≥ σmin > 0,
(5)
a ≥ 0.
We set the positive lower bound σmin for σ (σmin = 0.005 in our numerical examples)
in the sense that we state that if this threshold is reached, then an unambiguous cali-
bration of SVI is not doable, i.e. any precise choice of model parameters is arbitrary (of
course one can decide, for example, to inherit one of the SVI parameters from the ones
calibrated to the previous time-slice - if any - but this goes back to user choices). In the

c Zeliade Systems SAS

4/10 ZWP-0005– September, 2009
same way, we constrain a to positive values to avoid the σ → ∞, a → −∞ limiting
behaviour. Setting an upper bound for σ would prevent it from assuming too high
values, but this would not avoid the phenomena of “coupling” of low values of a and
high values of σ, where σ sticks to σmax and a becomes very negative.
We just add here that an obvious upper bound on a is:

a ≤ max{vi }, (6)
i

where the vi ’s are the observed variances at the given maturity. Condition (6) simply
follows from a consistent vertical location of the graph of (1): clearly the curve v giving
the optimal fit cannot be systematically greater than the largest observed variance.

3 Indeed, not more than a linear problem

As it stands, the calibration of SVI parametric form cast as a least square problem yields
an optimization problem in dimension 5. We show that, relying on some simple obser-
vations on the properties of the functional form (1), one can reformulate the problem
reducing the main dimension from 5 to 2.

3.1 Dimension reduction: drawing out the linear objective

We focus hereafter on the total variance ṽ = T v rather than on variance. The main
ingredient of the method is the fact that, by means of the change of variables

x−m
y=
σ

the SVI parametrization transforms into

p
ṽ(y) = aT + bσT (ρy + y 2 + 1).

This expression nicely shows how, for fixed values of m and σ, the support of the curve
T v is fully determined by a, ρ and the product bσ. Thus, most important, if we redefine
the parameters as
c = bσT
d = ρbσT
ã = aT,
then ṽ(y) turns out to depend linearly on c, d, a:
p
ṽ(y) = ã + dy + c y 2 + 1.

c Zeliade Systems SAS

5/10 ZWP-0005– September, 2009
 Therefore, for fixed m and σ, we look for the solution of the problem:

(Pm,σ ) min f{yi ,vi } (c, d, ã)

(c,d,ã)∈D

where f{yi ,vi } is the cost function

n q
X
2
f{yi ,vi } (c, d, ã) = f (c, d, ã) = ã + dyi + c yi2 + 1 − ṽi ,
i=1

with ṽi = T vi , and D is the compact and convex domain (a parallelepiped)


 0 ≤ c ≤ 4σ

D= |d| ≤ c and |d| ≤ 4σ − c

0 ≤ ã ≤ maxi {ṽi }


which is obtained from bounds (3) and (5)-(6) on parameters b, ρ and a.

Letting (c∗ , d∗ , ã∗ ) denote the solution of Pm,σ and (a∗ , b∗ , ρ∗ ) the corresponding triplet
for a, b, ρ, then the complete calibration problem is restored as
n
X
(P ) min (vm,σ,a∗ ,b∗ ,ρ∗ (xi ) − vi )2 .
m,σ
i=1

Our goal is therefore to solve Pm,σ in the fastest and most accurate way: once this is
done, the only task left is to look for the solution of the 2-dim problem P .

3.2 Explicit solution of the reduced problem

Pm,σ (the reduced problem) is a convex optimization problem with linear program, and
all the constraints defining the admissible domain D are linear. It is clearly seen, then,
that this problem admits an explicit solution, and becomes extremely easy to deal with.
Since the cost function f is convex, differentiable and its gradient is zero at just one
point (if the target smile contains at least three different points!), only two scenarios
are possible:

• the minimum of f over D is attained at the interior of D, and this is the global
minimum of f ;

• the minimum of f over D is attained on the boundary ∂D.

Then, this yields the simple recipe:

Step 1. find the global minimizer of f , solving the linear system ∇f = 0. If the output
belongs to D, then stop;

c Zeliade Systems SAS

6/10 ZWP-0005– September, 2009
Step 2. if Step 1 yields a global minimum outside D, then look for min∂D f .

The constrained optimization problem addressed in Step 2 can be solved applying a

Lagrange multipliers method for each one of the sides of the domain D. Then, Step 2
involves only the solution of 3 × 3 linear systems (plus a few explicit one-dimensional
minimizations, along the perimeter of the sides).
Once the two steps are accomplished, the solution of the reduced problem (Pm,σ ) is
achieved in explicit form. The calibration of (1) is then carried out solving the 2-
dimensional problem (P ) with some iterative optimizer, whose performances will be
extremely enhanced with respect to the original problem in full dimension.

4 Numerical Results

We display some numerical tests and calibration results showing the performances of
the Quasi-Explicit method.
Table 1 compares the “direct” procedure, i.e. standard least square calibration in di-
mension 5, and the Quasi-Explicit method for a SVI model calibrated to simulated
pP
data, i.e. a smile generated by SVI itself, for fixed T = 1. RMSE is 2
i (v(xi ) − vi ) ,
hence when looking to RMSE values one must take into account that the natural scale
is the one of a variance. For the Standard Least Square calibration, the input value
for a is inferred from the minimum observed variance, and the calibration is restarted
10 times from 10 randomly chosen points (b ∈ [0, 0.5], ρ ∈ [−1, 1], m ∈ [2 min(xi ) <
0, 2 max(xi ) > 0], σ ∈ [0, 1]). We do not go to great effort here in identifying a smart
initial guess for all the parameters since the randomized procedure works quite well
anyway, and our intention is rather to display the performances of the Quasi-Explicit
calibration. We recall that, for the latter, no inputs for a, b and σ are needed; moreover,
we take the initial guesses for m and σ as simple as it might be, i.e. a randomly chosen
point. Standard Least Square optimization is performed with truncated-Newton al-
gorithm, while the optimization over m and σ for the Quasi-Explicit method employs
Nelder-Mead simplex algorithm. As it is seen from Table 1, even if a classical calibra-
tion can work properly, the Quasi-Explicit technique brings the objective to extremely
small values. Moreover, the calibration we have obtained in the case ρ = −0.9 finely
shows how a downward SVI smile can be more than reasonably calibrated with a SVI
smile reaching its minimum (and then pointing upwards) for large values of the log-
moneyness, hence with a set of parameters which is far away from the true one.
Table 2 and Figure 2 display the result of the Quasi-Explicit calibration of SVI model
to the market-implied smile on DAX and EuroStoxx 50 indexes, for two different dates
(20 August and 22 September 2008, respectively). Concerning Table 2, the quality of
the fit is excellent through all maturities, somehow worse just for the very shortest

c Zeliade Systems SAS

7/10 ZWP-0005– September, 2009
 Method Parameters a b ρ m σ RMSE
True value 0.04 0.1 −0.5 0.0 0.1
Standard LS start. pt 0.048 10 r.p. 10 r.p. 10 r.p. 10 r.p.
calibrated 0.041 0.098 −0.51 −4e−4 0.095 7.9e−8
Quasi-Expl. start. pt - - - 1 r.p. 1 r.p.
calibrated 0.040 0.10 −0.50 −8e−7 0.100 5.0e−14
True value 0.1 0.06 −0.9 0.24 0.06
Standard LS start. pt 0.11 10 r.p. 10 r.p. 10 r.p. 10 r.p.
calibrated 0.004 0.11 −0.19 0.73 0.58 8.3e−8
Quasi-Expl. start. pt - - - 1 r.p. 1 r.p.
calibrated 0.10 0.060 −0.90 0.240 0.060 3.4e−17

Table 1: Calibration of SVI model to simulated data (r.p. = random points). The two
calibration strategies, Standard Least Square (dim = 5) and the Quasi-Explicit method
are compared, for T = 1. For Standard Least Square, the starting value for a is inferred
from minimum variance.

one, T = 1 month. Calibrated parameters show a good stability: obviously, a depen-

dence w.r.t. time to maturity is expected - time-dependence is not taken into account
by SVI model - but the important fact here is that the parameters do not show a noisy
behaviour (a smooth time-dependence is particularly seen for a and b). Concerning
the behaviour of ρ: ρ is different than −1 just for the first time-slice, which is the only
non downward-pointing smile (cf. Figure 2, the smile on DAX on 20 August shows
the same feature), and then it sticks to −1, because for all other maturities the smile is
purely decreasing. The Quasi-Explicit method has indeed the tendency to fit decreas-
ing smiles with anti-correlated SVI parameterizations.

T (Yrs) a b ρ m σ RMSE
0.082 0.027 0.234 0.068 0.100 0.028 1.6e−6
0.16 0.030 0.125 −1.0 0.074 0.050 2.8e−7
0.26 0.032 0.094 −1.0 0.093 0.041 2.1e−7
0.33 0.028 0.105 −1.0 0.096 0.072 1.3e−7
0.58 0.026 0.080 −1.0 0.127 0.098 7.1e−8
0.83 0.026 0.066 −1.0 0.153 0.113 1.8e−8
1.33 0.031 0.047 −1.0 0.171 0.065 5.2e−8
1.83 0.037 0.039 −1.0 0.152 0.030 9.1e−10
2.33 0.036 0.036 −1.0 0.200 0.083 1.3e−9
2.82 0.038 0.036 −1.0 0.170 0.139 2.4e−9
3.32 0.034 0.032 −1.0 0.246 0.199 7.2e−10
4.34 0.044 0.028 −1.0 0.188 0.069 2.6e−7

Table 2: Calibration of SVI model to the implied smile on the DAX Index on 20 August
2008. Each maturity is separately calibrated.

c Zeliade Systems SAS

8/10 ZWP-0005– September, 2009
 45
market
SVI fit
T = 0.068
40
Implied Vol (%)

35
T = 0.16

30

25

−0.25 −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2

log−forward moneyness

Figure 2: Calibration of SVI model to the implied smile on the EuroStoxx 50 Index on
22 September 2008, for the two shortest maturities.

5 Conclusions

Given the excellent performances of the Quasi-Explicit method - at least for what con-
cerns calibration on Equity indexes - we claim that this methodology responds prop-
erly to the question of how to obtain an unambiguous identification of a time-slice
of implied variance in terms of a set of SVI parameters. Once this high-quality fit is
achieved, the SVI functional form can serve in many ways. Besides smile point ex-
trapolation, one can recast the calibration of any stochastic model for the underlying
as a calibration to the smooth objective (1). The matching of the geometry (levels,
slopes and curvature) of the two model smiles can lead to explicit mappings of SVI
parameters onto the ones of the chosen model, in the spirit (in another context) of the
calibration methodology of [3]. Of course, this subject includes the issue of extracting a
local volatility surface with Dupire’s formula: since this operation needs interpolation
in time, at this level the time-interpolation mechanism becomes as well a crucial point.

References

[1] J. Gatheral. A parsimonious arbitrage-free implied volatility parameterization

with application to the valuation of volatility derivatives. Global Derivatives & Risk

c Zeliade Systems SAS

9/10 ZWP-0005– September, 2009
 Management, May 2004.

[2] L. C. G. Rogers and M. R. Tehranchi. Can the implied volatility surface move by
parallel shifts? Finance and Stochastics, 2008.

[3] V. Piterbarg. A Stochastic Volatility Forward Libor Model with a Term Structure
of Volatility Smiles. Working Paper at SSRN, Oct 2003.

c Zeliade Systems SAS

10/10 ZWP-0005– September, 2009