Vous êtes sur la page 1sur 12

Stochastic Local Volatility

Grigore Tataru, Travis Fisher


Quantitative Development Group, Bloomberg
gtataru@bloomberg.net, tfisher6@bloomberg.net
Version 1 – Feb 5, 2010

1. Introduction

Barrier prices and path dependent options in general are not determined by the vanilla market
quotes, they also depend on the dynamics of the market. Matching this extra market dimension is
a modeling requirement to obtain good barrier prices. Classical models such as local volatility
(LV) and stochastic volatility (SV) are in fact calibrated to and completely determined by the
vanilla market and, as such, they offer little or no extra flexibility in matching the market
dynamics. The Stochastic Local Volatility (SLV) model we present here combines features of an
LV model with features of an SV model to give rise to dynamics that are unachievable by the
simpler models.

During the past few years, practitioners have settled on a consensus of using some variation of a
mixed stochastic/local volatility model as the standard for pricing barrier options in FX markets.
Details vary widely from implementation to implementation. The most naive approach is simply
to calibrate a local volatility model and, completely independently, calibrate a stochastic
volatility model. Then the prices for exotic options are estimated as being some 𝑋% of the SV
model price and 100 − 𝑋 % of the LV model price. There is an obvious flaw to this approach
in that the only dynamics consistent with the model are completely unnatural — randomly
choose at time zero to follow the SV model with probability 𝑋 or the LV model with probability
1 − 𝑋. Another approach extends the LV model to include several volatility states, with some
dynamics to jump between these states. For more sophisticated approaches which model
realistic dynamics, the challenge is how to perform a good calibration of the model in a
reasonable amount of time. Some practitioners resort to carefully chosen functional forms for
the local volatility surface to allow semi-analytic pricing of vanilla options. There are problems
that arise from that approach: the functional form required (typically quadratic) may imply that
the exchange rate can reach zero or infinity. Calibration speed is still an issue, as is the stability
of the best fit.

In contrast, we allow a non-parametric form for the local volatility component of our SLV
model: the ―leverage surface‖. We calibrate it quickly and stably using a fixed point approach
centered around the solution of the forward Kolmogorov PDE for the transition density. This
gives us near complete freedom of choice of stochastic parameters while still matching the entire
volatility surface of vanilla options.

In this paper we will describe in detail the SLV model as implemented in the Bloomberg OVML
function. We will give motivation for the choices we have made, particularly regarding the
calibration freedom of the model. We will also describe the mathematical approach we follow.
We will start in the next section with the description of our stochastic local volatility model, and

1
then continue with extensive details regarding the model calibration. Finally, we will discuss
pricing and make comparisons with real-world data.

The first reference that we are aware of describing a SLV model is a paper by Jex, Henderson
and Wang [3]. Many other authors have contributed to this literature in the intervening years.
Our approach is similar to Lipton [4] and Ren, Madan, Qian [5] who describe a calibration
method based on the forward Kolmogorov PDE. The fixed point approach for the Kolmogorov
PDE that we use here is based on a method implemented at Bear Stearns in 2007 by the first
author, along with Christian Gilles and Graham Wells and with strong market insight coming
from Andrew Dexter. We offer thanks to Apollo Hogan, Ashish Midha, Peter Carr, Mark
Rubery, Leif Andersen, Peter Lee, and Oleg Kovrizhkin for helpful advice.

2. Model Details

In the SLV model, the foreign exchange rate dynamics are governed by the stochastic differential
equations:
𝑑𝑆𝑡
= 𝑟𝑑 − 𝑟𝑓 𝑑𝑡 + 𝐿 𝑆𝑡 , 𝑡 ∙ 𝑉𝑡 𝑑𝑊𝑡1 ,
𝑆𝑡
(1)
𝑑𝑉𝑡 = 𝜅 𝛳 − 𝑉𝑡 𝑑𝑡 + 𝜉𝑉𝑡 𝑑𝑊𝑡2 ,

𝑑𝑊𝑡1 ⋅ 𝑑𝑊𝑡2 = 𝜌 𝑑𝑡.

Here St represents the spot exchange rate process (given as units of domestic currency for a unit
of foreign currency), 𝑉𝑡 is the stochastic volatility process, and 𝑊𝑡1 and 𝑊𝑡2 are independent
standard Brownian motions. We use deterministic rate curves for 𝑟𝑑 (the domestic interest rate)
and 𝑟𝑓 (the foreign interest rate).
𝐿 𝑆, 𝑡 represents the local volatility component of the model, the leverage surface. We take a
non-parametric form for 𝐿 𝑆, 𝑡 , calibrated numerically to match the implied volatility surface.
The other parameters have the same meaning as in purely stochastic models: 𝜅 = speed of mean
reversion, 𝛳 = long-term mean level of 𝑉𝑡 , 𝜉 = volatility of the volatility process 𝑉𝑡 , 𝜌 =
correlation of the Brownian increments. We allow a term structure of these parameters with
piecewise constant values.

One should note that by setting 𝜉 = 0 the SLV model degenerates to a local volatility model. At
the other extreme, setting 𝐿 𝑆, 𝑡 ≡ 1 gives a purely stochastic vol model. The mixing of the LV
and SV features in the general SLV is mostly controlled by the volatility of volatility 𝜉, followed
in importance by the correlation ρ. To make it more intuitive, we simplify this mixing to a one-
parameter family:
𝜉 = 𝜆 ⋅ 𝜉𝑚𝑎𝑥
0% ≤ 𝜆 ≤ 100%,
ρ = λ ⋅ ρmax
where 𝜉𝑚𝑎𝑥 and ρmax correspond to what we will define later as the set of maximally stochastic
parameters. Again, we allow a term structure of these parameters, including the mixing fraction
λ, with piecewise constant values.

2
The dependence of barrier prices on the mixing of stochastic and local components is shown in
Figure 1. Different mixing fractions of the SLV model give different prices, all consistent with
the same implied volatility surface. The graph shows spot scenario value curves for a double-no-
touch option. As discussed later, our calibration includes matching of the spot/smile dynamics to
choose between these different possibilities.

35.00

30.00

25.00 100% Stochastic

20.00
50% Stochastic

15.00
0% Stochastic
(Local Vol)
10.00
Black-Scholes
5.00

0.00
1.58 1.63 1.68

Figure 1

We have chosen a lognormal process for the volatility process, as opposed to the familiar square-
root process found, for example, in the Heston model. We get more realistic behavior for the
paths of the volatility process and for the dynamics of the volatility surface. Also, the volatility
process 𝑉𝑡 does not reach 0, whereas it is possible in some cases for the Heston model. One
advantage of the Heston model, due to the square-root variance process, allows for analytic
formulas for vanillas. This is lost in the more general case of stochastic local volatility models,
though approximate formulas can be obtained for very particular choices of the leverage
function. Since we are using a general leverage surface and numerical solutions of the PDEs, our
method is not resting on the existence of analytical formulas, and as such we have made a choice
for the volatility process that we believe is more in line with the market behavior.

Models with a lognormal volatility process can have infinite moments and thus give infinite
prices for some instruments. This is not an issue for the model described here, mostly due to the
choice of a non-parametric leverage surface and its construction. Any tendency of the stochastic
volatility to produce high volatility in the wings will be annihilated by a leverage surface that
decreases fast in the wings.

3
3. Calibration Overview
Given a choice of stochastic parameters, one can calibrate the leverage surface 𝐿 𝑆𝑡 , 𝑡 to match
the vanilla market. A different set of stochastic parameters, along with the recalibrated 𝐿 𝑆𝑡 , 𝑡 ,
will again match the vanilla market, but will correspond to a different dynamics.

The calibration process has two main steps: find stochastic parameters to match the dynamics
(not to be recalibrated often) and then calibrate the leverage surface 𝐿 𝑆𝑡 , 𝑡 to match the vanillas
(recalibrated more often to the continuously changing market quotes for vanillas).

Both steps are essential for a successful model, but the second one is more important: not
matching the vanillas gives a bigger mispricing error of barrier options than a similar error in
stochastic parameters (the vanillas are the ones that every market participant agrees on, without
the same being true about exotics, or at least not to the same degree). The second step is also
more technically complex and is the one that corrects any misspecification of the stochastic
parameters.

The first calibration step—choosing stochastic parameters—is as much of an art as a science for
our situation. We will achieve an equally good fit to the current vanilla market regardless of this
choice, so there is no information in the current volatility surface that forces a particular set of
stochastic parameters. In many cases market makers who use versions of the SLV model may
hand-tune one or more parameters for a given market, or even for different instruments trading
on the same market. Also market participants may calibrate their models to barrier option prices
available to them. Neither of these approaches is available for Bloomberg’s default calibration.
Instead we have focused on matching particular statistics relating the historical dynamics of
volatility skew and spot. Details of this step will be presented later in the document.

4. Leverage Surface Calibration


Given a set of stochastic parameters, interest rate curves, and an implied volatility surface, we
calibrate a non-parametric leverage surface which matches the entire volatility surface. This
calibration is carried out by solving the forward Kolmogorov partial differential equation for the
SLV model while simultaneously solving for the unknown leverage surface in order to match the
conditional forward spot densities implied by the vanilla volatility surface.

The computation is organized into two main steps. First match the market with a local volatility
model. Second, for the given stochastic parameters, find a leverage function such that the SLV
has the same marginal distributions as the local volatility model.

Local Volatility

We start by calibrating a local volatility model:


𝑑𝑆𝑡
= 𝑟𝑑 − 𝑟𝑓 𝑑𝑡 + 𝜎 𝑆𝑡 , 𝑡 𝑑𝑊𝑡 .
𝑆𝑡
Using the Dupire formula to strip the implied volatility surface we get the local volatility 𝜎 𝑆, 𝑡 .
The implied volatility surface is interpolated from available data; in the usual case this is the
Bloomberg generic volatility data which is itself a filtered, smoothed, and averaged version of
the data obtained directly from contributors.

4
Leverage Surface

The marginal distribution of the SLV model is (see Gyöngy [1]) the same as the distribution of a
local volatility model with 𝜎 𝑆, 𝑡 given by:

2 𝜎(𝑆, 𝑡)2 = 𝐸 𝐿(𝑆𝑡 , 𝑡)2 𝑉𝑡2 𝑆𝑡 = 𝑆 = 𝐿(𝑆, 𝑡)2 𝐸 𝑉𝑡2 𝑆𝑡 = 𝑆 .

We want 𝜎(𝑆, 𝑡) to be the local volatility calibrated above which will guarantee that our SLV is
calibrated to the market. The difficulty is to find 𝐿(𝑆, 𝑡) such that the relation (2) holds. The
difficulty is increased by the fact that the conditional expectation above, 𝐸 𝑉𝑡2 𝑆𝑡 = 𝑆 , itself
depends on 𝐿(𝑆, 𝑡) in a nontrivial way.

To see this, consider the forward Kolmogorov PDE for the transition probability density
𝑝 𝑡, 𝑆, 𝑉 of (1) from the initial state 𝑡0 , 𝑆0 , 𝑉0 to 𝑡, 𝑆, 𝑉 :

𝜕𝑝 1 𝜕 2 2 2 2
1 2 𝜕2 2
𝜕2
3 – 𝑆 𝐿 𝑆, 𝑡 𝑉 𝑝 − 𝜉 𝑉 𝑝 − 𝜌 𝜉 𝑆𝐿 𝑆, 𝑡 𝑉 2 𝑝
𝜕𝑡 2 𝜕𝑆 2 2 𝜕𝑉 2 𝜕𝑆𝜕𝑉
𝜕 𝜕
+ 𝑟𝑑 − 𝑟𝑓 𝑆 𝑝 + 𝜅 𝛳 − 𝑉 𝑝 = 0,
𝜕𝑆 𝜕𝑉
𝑝 𝑡0 , 𝑆, 𝑉 = 𝛿 𝑆 − 𝑆0 𝛿 𝑉 − 𝑉0 ,
where 𝛿 𝑥 is the Dirac distribution centered at 𝑥 = 0.

We calculate the conditional expectation of the instantaneous variance given future spot and time
as an integral against the density:
∞ 2
0
𝑉 𝑝 𝑡, 𝑆, 𝑉 𝑑𝑉
2
𝐸 𝑉𝑡 𝑆𝑡 = 𝑆 = ∞
0
𝑝 𝑡, 𝑆, 𝑉 𝑑𝑉
and can rewrite (2) as:
∞ 2
𝑉 𝑝 𝑡, 𝑆, 𝑉 𝑑𝑉
2 2 0
4 𝜎(𝑆, 𝑡) = 𝐿(𝑆, 𝑡) ∞ .
0
𝑝 𝑡, 𝑆, 𝑉 𝑑𝑉
Both 𝐿(𝑆, 𝑡) and 𝑝 𝑡, 𝑆, 𝑉 are unknown at this stage, while all the other terms in the equation (3)
and (4) are known, including the stochastic parameters and the local volatility. Note that
knowing 𝑝 gives 𝐿 by (4), while knowing 𝐿 gives 𝑝 by solving the PDE (3). We will set up an
iterative argument in the next section to find 𝑝 𝑡, 𝑆, 𝑉 , and hence 𝐿(𝑆, 𝑡). This will give an SLV
model calibrated to the vanilla market for the given set of stochastic parameters.

Fixed point reformulation


Finding 𝑝 𝑡, 𝑆, 𝑉 to satisfy (3) and (4) amounts to solving the nonlinear PDE:

𝜕𝑝 1 𝜕2 0
𝑝 𝑡, 𝑆, 𝑣 𝑑𝑣 1 𝜕2
5 − 𝑆 2 𝑉 2 𝜎 𝑆, 𝑡 2
∞ 2 𝑝 − 𝜉2 2 𝑉2𝑝
𝜕𝑡 2 𝜕𝑆 2 𝑣 𝑝 𝑡, 𝑆, 𝑣 𝑑𝑣 2 𝜕𝑉
0
∞ 1
2
𝜕2 0
𝑝 𝑡, 𝑆, 𝑣 𝑑𝑣 𝜕 𝜕
− 𝜌𝜉 𝑆𝑉 2 𝜎 𝑆, 𝑡 ∞ 2 𝑝 + 𝑟𝑑 − 𝑟𝑓 𝑆 𝑝 + 𝜅 𝛳 − 𝑉 𝑝 = 0.
𝜕𝑆𝜕𝑉 𝑣 𝑝 𝑡, 𝑆, 𝑣 𝑑𝑣 𝜕𝑆 𝜕𝑉
0

5
This can be reformulated as a fixed point problem 𝑞 → 𝑃 𝑞 = 𝑝 where the operator 𝑃 ∙ is the
solution of the linear PDE:

𝜕𝑝 1 𝜕2 0
𝑞 𝑡, 𝑆, 𝑣 𝑑𝑣 1 𝜕2
6 − 𝑆 2 𝑉 2 𝜎 𝑆, 𝑡 2
∞ 2 𝑝 − 𝜉2 2 𝑉2𝑝
𝜕𝑡 2 𝜕𝑆 2 𝑣 𝑞 𝑡, 𝑆, 𝑣 𝑑𝑣 2 𝜕𝑉
0
∞ 1
2
𝜕2 0
𝑞 𝑡, 𝑆, 𝑣 𝑑𝑣 𝜕 𝜕
− 𝜌𝜉 𝑆𝑉 2 𝜎 𝑆, 𝑡 ∞ 2 𝑝 + 𝑟𝑑 − 𝑟𝑓 𝑆 𝑝 + 𝜅 𝛳 − 𝑉 𝑝 = 0.
𝜕𝑆𝜕𝑉 𝑣 𝑞 𝑡, 𝑆, 𝑣 𝑑𝑣 𝜕𝑆 𝜕𝑉
0

One can construct a recursive sequence 𝑝(𝑛) → 𝑃 𝑝(𝑛 ) = 𝑝(𝑛+1) by solving the above PDE
repeatedly. Corresponding to each 𝑝(𝑛) 𝑡, 𝑆, 𝑉 we have a leverage function 𝐿(𝑛) (𝑆, 𝑡) defined by
(4) and the procedure above can be understood alternatively as a fixed point problem for 𝐿 𝑆, 𝑡 .
We will solve the previous PDE numerically. Convergence of the sequences 𝑝(𝑛) and 𝐿(𝑛) , at
least numerically, will give us the calibrated leverage surface.

Note that we expect 𝑝𝐿 𝑡, 𝑆 = 0 𝑝 𝑡, 𝑆, 𝑣 𝑑𝑣 to be the transition density of the local volatility
model, assumed to be known, for example by solving the forward Kolmogorov PDE for the LV
model:
𝜕𝑝𝐿 1 𝜕2 2 𝜕
7 − 2
𝑆 𝜎 𝑆, 𝑡 2 𝑝𝐿 + 𝑟𝑑 − 𝑟𝑓 𝑆 𝑝𝐿 = 0.
𝜕𝑡 2

𝜕𝑆 ∞
𝜕𝑆
We could have replaced 0 𝑝 𝑑𝑣 and 0 𝑞 𝑑𝑣 with 𝑝𝐿 in the (5) and (6) to end up with a slightly
different construction for 𝑝 𝑛 and 𝐿(𝑛 ) . As a sanity check, one would like to know that

𝑝𝐿 𝑡, 𝑆 = 0 𝑝 𝑡, 𝑆, 𝑣 𝑑𝑣 is indeed true for 𝑝 obtained by solving (5) (or the modified version
of (5)). This follows by simple integration of (5) in the 𝑉 direction and the uniqueness of
solutions for equation (7).

Finite difference scheme

To solve (6) we use a finite difference method, a variation of the Douglas scheme as described in
in ’t Hout and Foulon [2]. This is an operator splitting scheme of ADI type with a predictor-
corrector treatment of the mixed derivative. To solve the one-dimensional problems in the 𝑆
and 𝑉 direction we use fully implicit time-stepping, while for the mixed partial derivative in 𝑆
and 𝑉 representing the correlation term we use fully explicit time-stepping.

5. Risk-Reversal Dynamics in the SLV model


In a typical local volatility model, the risk-reversal will increase when spot increases and
decrease when spot decreases. Schematically, this can be seen as a result of the curvature of the
local volatility surface as illustrated in Figure 2. Suppose the local volatility surface is a function
of spot as given by the curve shown. For a lower level S1 of the spot, the 25 delta call and put
strikes are C1 and P1, respectively. For a slightly higher level S2 of the spot, the 25 delta call and
put strikes are C2 and P2, respectively. The implied volatility for a vanilla option is related to a
weighted average of the local volatilities between the spot and strike levels. Thus the implied
volatility for strike C2 will be higher than the implied volatility for strike C1, and the implied
volatility for strike P2 will be lower than the implied volatility for strike P1. Since the risk-

6
reversal is given by the difference of the implied volatility for the call minus the implied
volatility for the put, the risk-reversal is increased by increasing spot.

C2
C1
P1
P2

Figure 2

In the stochastic volatility model, the correlation between spot and risk-reversal is neutralized.
Our SLV model, with the leverage surface set equal to one, becomes the following stochastic
volatility model:
𝑑𝑆𝑡
= 𝑟𝑑 − 𝑟𝑓 𝑑𝑡 + 𝑉𝑡 𝑑𝑊𝑡1
𝑆𝑡
𝑑𝑉𝑡 = 𝜅 𝛳 − 𝑉𝑡 𝑑𝑡 + 𝜉𝑉𝑡 𝑑𝑊𝑡2

Note that this model has a clear spot scaling property. Rescaling the spot process results in
rescaled but otherwise unchanged dynamics. The volatility smile expressed in terms of deltas
shares the same spot scaling property and is unchanged by rescaling spot. So for the pure
stochastic volatility model, the risk-reversal is unchanged by changing the initial spot level.

For our SLV model, we never use a completely pure stochastic volatility model. We always
have a leverage surface calibrated to meet the residual prices of the vanilla options from the full
volatility surface. Still the overall relationship is the same: on the local volatility side of the
spectrum the SLV model has a strong positive relationship between the risk-reversal and the spot
level. On the maximally stochastic end of the spectrum, the SLV model has risk-reversal levels
nearly unaffected by changes of the spot level. For mixing fractions in between, there are
intermediate relationships between the risk-reversal and spot dynamics.

Figure 3 shows normalized risk-reversal vs. spot for the calibrated SLV model, together with
local volatility and maximally stochastic volatility extremes of the SLV model. This uses the
USDJPY volatility surface as of 1 Feb 2010. Note the expected trend for the local volatility
model compared to the near-constant trend for the maximally stochastic model. The calibrated
SLV model gives an intermediate result.

7
5.00%
Calibrated SLV Local Vol Stochastic Vol
0.00%

-5.00%
RR Vol as Percent of ATM Vol

-10.00%

-15.00%

-20.00%

-25.00%

-30.00%

-35.00%
80 82 84 86 88 90 92 94 96 98 100
USDJPY Spot

Figure 3

6. Risk-Reversal Dynamics and One-Touch Options


The dynamics of the risk-reversal in relation to the spot are also closely related to barrier option
pricing. For example consider a down one-touch option paying in domestic currency. In the
Black-Scholes model an approximate hedge for it would be a European digital put option struck
at the barrier level paying double the touch option’s notional. If the barrier is hit, the digital
option is at-the-money and so has approximately (ignoring rates, etc.) 50% chances of paying. A
50% chance at double the payoff for the digital option vs. 100% chance at the payoff for the
touch option at the barrier hit would give these the same value. In a model with a skew,
however, this hedge will fail because the digital put price depends on the implied volatility skew
at its strike: the higher the skew, the higher the value of the digital option. Thus the necessary
notional amount of the hedge is no longer double the one-touch notional, but is instead related to
the at-the-money implied volatility skew that will be in effect when the spot has moved to the
barrier level. If the skew is higher the digital price is higher and less notional is required,
whereas if skew is lower the value of the digital is lower and more notional is required. In
conclusion, if the one-touched is hedged as above, its value on trade date depends on the notional
of the hedging digital, which depends on the implied volatility skew when the spot reaches the
barrier.

8
7. Historical Trends of Risk-Reversal Dynamics
Particularly for skewed currency pairs, there is often a very clear relationship between the spot
exchange rate and the volatility skew. This relationship is immediately clear in a scatter-plot of
risk-reversal level vs. spot. A market like USDJPY may stay in the same regime for many
months, revisiting the same spot and risk-reversal levels. Unfortunately for modeling, the market
may also shift to a different regime, re-normalizing around a different base range of spot and a
different base level of risk-reversal. This re-normalizing is beyond the range of market dynamics
the SLV model can capture. Still we are interested in historical calibration, and for that purpose
we want to answer the question: to the extent that the risk-reversal dynamics are predictable,
what is the best prediction of the relationship of spot to risk-reversal?

We have developed an approach which helps to isolate the predictable trends from the
unpredictable regime changes. To do this we study deviations of the risk-reversal from its
exponentially weighted moving average, in comparison to deviations of the exchange rate from
its exponentially weighted moving average. Figure 4 illustrates the advantage of this approach.
On the left, we look at the normalized risk reversal (one month RR vol as a percent of ATM vol)
compared to exchange rate. We split the data into three consecutive time periods. In each time
period, the normalized risk-reversal shows a strong trend with spot. But for the entire time
period, the market shifts overwhelm the trends. On the right, we look at the deviation from
exponentially weighted moving average, using a half-life of 12 business days. Here the trend is
clear.

Figure 4

9
8. Historical Stochastic Parameter Calibration—Implementation Details
As discussed above, the key degree of freedom of the SLV model is the ability to interpolate
between a local volatility model and a stochastic volatility model. In terms of the dynamics of
the volatility smile, this degree of freedom is reflected in different relationships of how the risk-
reversals change when spot moves. Stylistically, for a pure local volatility model with a smile in
the volatility surface, the risk-reversal increases as spot increases and decreases as spot
decreases. For a purely stochastic volatility model, the risk-reversal is largely unchanged by
moves of the spot. This stylistic view is made a little cleaner even if we normalize the market
25% delta risk-reversal to a ―normalized risk-reversal‖ by dividing by the at-the-money
volatility. The separation of the risk-reversal strikes is approximately proportional to the overall
level of volatility, so the normalization gives a value that more cleanly reflects the degree of
skewness of the volatility smile.

We calibrate stochastic parameters in a two-phase process. The first phase is to find a set of
―maximally stochastic‖ parameters, where the residual shape of the leverage surface is nearly flat
in the main region of interest. The second phase is to find a term structure of mixing fractions to
match the desired relationship of normalized risk-reversal move to spot move.

The calibration of maximally stochastic parameters proceeds in a bootstrap fashion. For each
maturity in the term-structure of parameters, a Levenberg-Marquardt nonlinear least squares
optimization is used. The mean reversion rate 𝜅 is fixed to 1.0. The initial 𝑉0 and long-term
mean level 𝛳 of 𝑉𝑡 are both set to 1.0 as well, reflecting a scaling of the leverage surface
comparable to the local volatility surface. The parameters to be calibrated are 𝜉 and 𝜌. For each
test set of parameters a leverage surface calibration is performed, and the resulting leverage
surface is measured for flatness. The objective function consists of three components. For
calibrating the 𝜉 and ρ in effect from time 𝑡𝑖 to 𝑡𝑖+1 , we take a vega-weighted average of 20
leverage surface points distributed between the 10 delta put and the 10 delta call strikes. The
objective function penalizes deviation of leverage surface points from the average level. The
second component is forward-looking regularization that penalizes deviation of the leverage
surface points from the average deviation at the succeeding maturity 𝑡𝑖+2 . The third component
is a backward-looking regularization that penalizes deviation of 𝜉𝑖 from 𝜉𝑖−1 and deviation of 𝜌𝑖
from 𝜌𝑖−1 .

After the maximal stochastic parameter calibration, a further regularization is done to reduce
intraday variation. Instead of the calibrated parameters being used directly, an exponentially
weighted moving average is used with the new calibration weighted at 20%.

The mixing fraction calibration computes a trend of how normalized risk-reversal moves are
related to moves of the forward exchange rate. It performs a Levenberg-Marquardt nonlinear
least squares optimization to find a term structure of mixing fractions to match the model to the
historical relationship in both the short and long term. Daily historical data of the risk-reversal,
at-the-money, and forward exchange rate are used. The computed trend measures deviation of
normalized risk-reversal from its exponentially weighted moving average. This is compared to
deviation of forward exchange rate from its exponentially weighed moving average. A linear
trend of these quantities is estimated, again using exponentially decreasing weights. The weights
used correspond to a halflife in days of 50*t, where t is the tenor as a year fraction, truncated to a

10
minimum of 12 business days for short tenors and a maximum of 35 business days for long
tenors.

Currently the term structure of mixing fraction is calibrated to be controlled by two points—the
historical trends computed at the one month and one year tenors. The term structure of
parameters varies linearly in time for intermediate parameter points.

For each parameter choice, the leverage surface is calibrated corresponding to those parameters.
A series of option scenario pricings are performed to find the model normalized risk-reversal
values at a future time and across a range of spot levels. For each tenor t, the scenario horizon
time is set to (1/2)t and option maturity is set to (3/2)t. A set of strikes are chosen in the range
between the currently 1% delta put and 1% delta call range. For each strike, a scenario pricing is
carried out giving prices at a series of spot levels chosen in the range between the currently 25%
delta put and 25% delta call. The implied volatility is computed for each strike and spot pair,
giving an implied volatility smile for the strikes involved. For each spot, the 25% delta risk-
reversal and at-the-money volatilities are interpolated from this scenario volatility smile. A
linear trend is estimated between the scenario normalized risk-reversals and the scenario spot
levels.

The optimization penalizes differences between the scenario normalized risk-reversal trend and
the computed historical trend. The objective function also includes regularization penalties for
deviation of the mixing fraction parameter from 50%.

After the mixing fraction calibration, a further regularization is done to reduce intraday variation.
Instead of the calibrated mixing fraction being used directly, an exponentially weighted moving
average is used with the new calibration weighted at 20%.

9. Pricing
Assume that we have a calibrated model (1). Pricing is done by solving backwards the following
PDE for the price 𝐶 𝑡, 𝑆, 𝑉 of a claim:
𝜕𝐶 1 2 2 2
𝜕2𝐶 2
𝜕2𝐶 1 2 2 𝜕2𝐶 𝜕𝐶
+ 𝑆 𝐿 𝑆, 𝑡 𝑉 2
+ 𝜌 𝜉𝑆𝐿 𝑆, 𝑡 𝑉 + 𝜉 𝑉 2
+ 𝑟𝑑 − 𝑟𝑓 𝑆
𝜕𝑡 2 𝜕𝑆 𝜕𝑆𝜕𝑉 2 𝜕𝑉 𝜕𝑆
𝜕𝐶
+ 𝜅 𝛳−𝑉 = 𝑟𝑑 𝐶,
𝜕𝑉
with the appropriate boundary conditions and final value for the payoff. We solve this PDE
numerically using a variation of the Douglas finite difference scheme, just as we use for the
leverage surface calibration PDE.

10. Calibration Speed and Accuracy

As a representative example, we calibrated the SLV model to market data from February 1,
2010. The calibrated term structure of stochastic parameters is as follows. Calibration of the set
of maximal stochastic parameters followed by the historical estimation of mixing fraction takes
about one minute. The parameters not shown are mean reversion 𝜅 (equal to 100% across the
entire term structure) and the initial value 𝑉0 and the long-term mean level 𝛳 of 𝑉𝑡 (both equal to
1.0).

11
Vol-of-Vol Rho Mixing Fraction
0-2W 128.71% -0.1455 62.39%
2W-1M 79.60% -0.2063 62.39%
1M-2M 57.97% -0.2843 62.37%
2M-3M 52.35% -0.2787 61.90%
3M-6M 61.39% -0.254 61.42%
6M-1Y 64.99% -0.2626 59.96%
1Y-2Y 64.05% -0.2587 57.05%

Using these parameters, we calibrate to the implied volatility surface. Our non-parametric
calibration matches the entire volatility surface to the very distant wings. Here we show the
results of testing the calibration on the key tenors and market strikes. Calibration time is about 1
second, using a single iteration of the fixed point solution. Option pricing times vary by
maturity, with an average pricing time in the table below taking about one half second. The mean
absolute volatility error in this example is 5.6 basis points while the root mean square error is 6.7
basis points. The maximum error of 17 basis points occurs in the high-volatility corner at the 2Y
maturity. In all cases the calibrated prices are within the bid/ask spread of the market data.
Table entries in the following table show the market level of volatility, followed in square
brackets by the error in implied volatility achieved when re-pricing each option using the SLV
pricing PDE.

10 Put 25 Put ATM 10 Call 25 Call


1W 14.14 [0.07] 13.36 [-0.08] 12.88 [-0.13] 12.89 [-0.05] 13.31 [0.14]
2W 13.95 [0.04] 13.08 [-0.05] 12.43 [-0.08] 12.31 [-0.02] 12.58 [0.10]
3W 14.35 [0.02] 13.37 [-0.04] 12.60 [-0.06] 12.43 [-0.01] 12.63 [0.08]
1M 14.42 [0.02] 13.39 [-0.03] 12.56 [-0.04] 12.31 [0.00] 12.45 [0.07]
2M 15.26 [-0.02] 13.94 [-0.07] 12.92 [-0.07] 12.54 [-0.02] 12.67 [0.06]
3M 16.06 [-0.04] 14.42 [-0.07] 13.20 [-0.07] 12.66 [-0.02] 12.77 [0.05]
4M 16.55 [-0.04] 14.74 [-0.06] 13.43 [-0.06] 12.81 [-0.02] 12.89 [0.04]
6M 17.17 [-0.04] 15.14 [-0.05] 13.70 [-0.05] 12.98 [-0.01] 13.05 [0.04]
9M 17.91 [-0.05] 15.57 [-0.06] 13.97 [-0.06] 13.13 [-0.02] 13.18 [0.03]
1Y 18.46 [-0.08] 15.85 [-0.08] 14.11 [-0.06] 13.17 [-0.02] 13.18 [0.02]
18M 19.06 [-0.12] 16.23 [-0.11] 14.35 [-0.09] 13.28 [-0.04] 13.29 [0.01]
2Y 19.36 [-0.17] 16.41 [-0.15] 14.49 [-0.11] 13.26 [-0.05] 13.24[0.00]

References

1. Gyöngy, I., Mimicking the One-Dimensional Marginal Distributions of Processes Having


an Ito Differential, Probability Theory and Related Fields 71 (1986)
2. in ’t Hout, K.J., Foulon, S., ADI finite difference schmes for option pricing in the Heston
model with correlation, Working paper (2007)
3. Jex, M., Henderson, R., Wang, D., Pricing Exotics Under the Smile, J.P.Morgan (1999)
4. Lipton, A., The vol smile problem, Risk (2002)
5. Ren, Y., Madan, D., Qian, M.Q., Calibrating and pricing with embedded local volatility
models, Risk (2007)

12

Vous aimerez peut-être aussi