Pricing and Hedging Volatility Derivatives

Mark Broadie
a
Ashish Jain
b
January 10, 2008
Abstract
This paper studies the pricing and hedging of variance swaps and other volatility derivatives,
including volatility swaps and variance options, in the Heston stochastic volatility model. Pricing
and hedging results are derived using partial dierential equation techniques. We formulate an
optimization problem to determine the number of options required to best hedge a variance
swap. We propose a method to dynamically hedge volatility derivatives using variance swaps
and a nite number of European call and put options.
a
Mark Broadie is the Carson Family Professor of Business at Columbia University in New
York, NY. e-mail: [email protected].
b
Ashish Jain is a senior associate at Lehman Brothers in New York, NY. e-mail:
[email protected]

This paper was presented at the Fall 2006 INFORMS annual conference, the London Business School, the
2005 Winter Simulation Conference and at the Columbia Business School. This work was partly supported by
NSF grant DMS-0410234.
1
1 Introduction
Volatility derivatives are securities whose payo depends on the realized variance of an underly-
ing asset or an index return. Realized variance is the variance of the underlying assets return
over the life of the volatility derivative. A variance swap has a payo which is a linear function
of the realized variance, a volatility swap has a payo which is a concave function of the realized
variance and a variance call options payo is a convex function of the realized variance. We
provide denitions of various volatility derivatives in Section 2.
In this paper we propose a methodology for hedging volatility swaps and variance options using
variance swaps. Since the price of both variance swaps and volatility swaps depend on the real-
ized variance of the underlying asset, there must be a relationship between their prices to avoid
arbitrage. Since variance swaps can be priced and hedged using actively traded European call
and put options, by exploiting the no-arbitrage relationship between volatility derivatives and
variance swaps we can price and hedge volatility derivatives.
The volatility of asset prices is an indispensable input in both pricing options and in risk man-
agement. Through the introduction of volatility derivatives, volatility is now, in eect, a tradable
market instrument. Previously traders would use a delta-hedged option position as a means to
trade volatility. However, this does not provide a pure volatility exposure since the return also
depends on the underlying stock price. Variance and volatility swaps provide pure exposure to
volatility and have become quite popular in the market. Three dierent groups of traders have
emerged: directional traders, spread traders and volatility hedgers. Directional traders specu-
late on the future level of volatility, while spread traders bet on the dierence between realized
and implied volatility. In contrast, a volatility hedger typically covers short volatility positions.
For example, life insurance companies now oer many products with guaranteed benets (e.g.,
variable annuities or with-prots funds) and these expose them to short volatility positions that
may be oset by using variance swaps. Variance and volatility swaps capture the volatility of the
underlying asset over a specied time period and are eective hedging instruments for volatility
exposure. Based on the demand from volatility traders, the market in volatility and variance
swaps has developed rapidly over the last few years and is expected to grow more in the future.
Estimating total trading volume is problematic, as with any OTC market, but recent estimates
for daily trading volume on indices are in the region of $3035 million notional. Hence the pricing
and hedging of these derivatives have become an important research problem in academia and
industry.
1
Building on the work of Neuberger (1994), Demeter, Derman, Kamal and Zou (1999) examined
properties of variance and volatility swaps. They showed that variance swaps can be replicated
by a static position in European call and put options of all strikes and a dynamic trading strategy
in the underlying asset. Brockhaus and Long (2000) provided an analytical approximation for
the pricing of volatility swaps. Javaheri, Wilmott and Haug (2002) discussed the valuation of
volatility swaps in the GARCH(1,1) stochastic volatility model. They used a partial dierential
equation approach to determine the rst two moments of the realized variance and then used
a convexity approximation approximation to price the volatility swaps. Lipton (2000) priced
volatility swaps using a PDE approach. Little and Pant (2001) developed a nite dierence
method for the valuation of variance swaps in the case of discrete sampling in an extended
Black-Scholes framework. Detemple and Osakwe (2000) priced European and American options
on the terminal value of volatility when volatility follows a diusion process. Carr, Geman,
Madan and Yor (2005) priced options on realized variance by directly modeling the quadratic
variation of underlying process using a Levy process. Carr and Lee (2005) priced arbitrary pay-
os of realized variance provided a zero correlation assumption between stock price process and
variance process. Broadie and Jain (2007) show that the convexity correction approximation
doesnt provide a good estimate of fair volatility strikes in the Heston stochastic volatility and
the Merton jump-diusion models.
In this paper we price variance and volatility swaps when the variance process is a continuous
diusion given by the Heston stochastic volatility model. We compute fair volatility strikes and
price variance options by deriving a partial dierential equation that must be satised by volatil-
ity derivatives. We compute the risk management parameters (greeks) of volatility derivatives
by solving a series of partial dierential equations. Independently, Sepp (2006) priced options
on realized variance in the Heston stochastic volatility model by solving a partial dierential
equation. We present a numerical method to determine the number of options required to hedge
a variance swap. We propose a method to dynamically hedge volatility derivatives using variance
swaps and a nite number of European call and put options.
The rest of the paper is organized as follows. We begin by briey introducing volatility derivatives
in Section 2. In Section 3 we present the pricing of volatility swaps and the variance options using
a partial dierential equation approach in the Heston stochastic volatility model. In Section 4
we present the computation of greeks of volatility derivatives in the Heston stochastic volatility
model. In Section 5 we present an optimization approach to hedge variance swaps using a nite
2
number of options. We also present a dynamic approach to hedge volatility swaps using variance
swaps. Concluding remarks are given in Section 6.
2 Volatility Derivatives
Volatility and variance swaps are forward contracts in which one counterparty agrees to pay the
other a notional amount, N, times the dierence between a xed level and a realized level of
volatility and variance, respectively. The xed level is called the variance strike for variance
swaps and the volatility strike for volatility swaps. Realized variance is determined by the vari-
ance of the assets return over the life of the swap.
The variance swap payo is dened as
(V
d
(0, n, T) K) N
where V
d
(0, n, T) is the realized variance of stock return (dened below) over the life of the
contract, [0, T], where n is the number of sampling dates, the subscript d is used to empha-
size that the variance is computed discretely (i.e., with a nite number of sampling dates, n),
K is the variance strike, and N is the notional amount of the swap in dollars. The holder of
a variance swap at expiration receives N dollars for every unit by which the stocks realized
variance V
d
(0, n, T) exceeds the variance strike K. The variance strike is quoted in units of
volatility squared, e.g., (20%)
2
. For example, suppose an investor takes a long position in a
variance swap with strike (20%)
2
= 0.04 and a notional of one million dollars. If, over the life
of the contract, the realized variance is (25%)
2
= 0.0625, the investor would make a prot of
(0.0625 0.04) 1,000,000 = $22,500.
The volatility swap payo is dened as
(
_
V
d
(0, n, T) K) N
where
_
V
d
(0, n, T) is the realized stock volatility (quoted in annual terms as dened below) over
the life of the contract, where n is the number of sampling dates, K is the volatility strike, and
N is the notional amount of the swap in dollars. The volatility strike, K, is typically quoted
in units of percent, e.g., 20%. An investor who is long a volatility swap with strike 20% and a
notional of one millon dollars would make a prot of (0.25 0.2) 1,000,000 = $50,000 in the
3
previous example.
The procedure for calculating realized volatility and variance is specied in the derivative contract
and includes details about the source and observation frequency of the price of the underlying
asset, the annualization factor to be used in moving to an annualized volatility and the method
of calculating the variance. Let 0 = t
0
< t
1
< ... < t
n
= T be a partition of the time interval
[0, T] into n equal segments of length t, i.e., t
i
= iT/n for each i = 0, 1, ..., n. Most traded
contracts dene realized variance to be
V
d
(0, n, T) =
AF
n 1
n1

i=0
_
ln
_
S
i+1
S
i
__
2
(1)
for a swap covering n return observations. Here S
i
is the price of the asset at the i
th
observation
time t
i
and AF is the annualization factor, e.g., 252 (= n/T) if the maturity of the swap, T, is
one year with daily sampling. This denition of realized variance diers from the usual sample
variance because the sample average is not subtracted from each observation. Since the sample
average is approximately zero, the realized variance is close to the sample variance.
We call V
d
(0, n, T) the discretely sampled realized variance and V
c
(0, T) the continuously sampled
realized variance. The oating leg of variance swap, or discrete realized variance, in the limit
approaches the continuously sampled realized variance, that is,
V
c
(0, T) lim
n
V
d
(0, n, T) (2)
In this paper we price volatility derivatives assuming sampling is done continuously. Broadie
and Jain (2007) compute fair variance strikes and fair volatility strikes when realized variance is
computed discretely.
A European variance call option gives the holder the right to receive a payo V
c
(0, T) in exchange
for paying the strike K at the maturity of variance call option, i.e., its payo is
C
T
= max(V
c
(0, T) K, 0) N (3)
Similarly the payo of the variance put option is:
P
T
= max(K V
c
(0, T), 0) N (4)
where N is the notional amount in dollars. Unlike European equity options, the payo of vari-
ance options depends on realized variance V
c
(0, T) which is not a traded instrument in the market.
4
We assume the risk-neutral dynamics of the underlying asset S
t
follows the Heston (1993) stochas-
tic volatility (hereafter SV) model:
dS
t
= rS
t
dt +

v
t
S
t
(dW
1
t
+
_
1
2
dW
2
t
) (5)
dv
t
= ( v
t
)dt +
v

v
t
dW
1
t
(6)
Equation (5) gives the dynamics of the stock price: S
t
denotes the stock price at time t,

v
t
is
the volatility at time t and r is the riskless interest rate. Equation (6) species the evolution
of the variance as a mean-reverting process: is the long-run mean variance, represents the
speed of mean reversion, and
v
is a parameter which determines the volatility of the variance
process. The processes W
1
t
and W
2
t
are two independent standard Brownian motions under the
risk-neutral measure Q, and represents the instantaneous correlation between the return and
volatility processes. The initial value of the stock price is denoted by S
0
and the variance process
by v
0
. The variance process v
t
, unlike S
t
, is unobservable, so it needs to be estimated from data
(e.g., option prices or a time series of S
t
).
In the SV model, continuous realized variance is given by
V
c
(0, T) =
1
T
_
T
0
v
s
ds (7)
The fair variance strike, K

var
, is dened as the value which makes the contracts net present
value equal to zero, i.e., it is the solution of
E
Q
0
_
e
rT
(V
c
(0, T) K

var
)
_
= 0 (8)
where the superscript Q indicates the risk-neutral measure and the subscript 0 denotes expecta-
tion at time t = 0. In the SV model, the fair variance strike is given by
K

var
= E[V
c
(0, T)] = E
_
1
T
_
T
0
v
s
ds
_
= +
v
0

T
(1 e
T
) (9)
where the last equality follows, e.g., from Broadie and Jain (2007). The fair volatility strike is
dened as the value which makes the contract net present value equal to zero, i.e., it solves the
equation
E
0
_
e
rT
(
_
V
c
(0, T) K

vol
)
_
= 0 (10)
Hence, the fair volatility strike can be expressed as
K

vol
= E
_

1
T
_
T
0
v
t
dt
_
= E[
_
V
c
(0, T)] (11)
5
Using Jensens inequality
1
we can obtain an upper bound on the fair volatility strike:
K

vol
= E
0
[
_
V
c
(0, T)]
_
E
0
[V
c
(0, T)] =
_
K

var
(12)
Hence, the fair volatility strike is bounded above by the square root of the fair variance strike.
The dierence in the square root of the fair variance strike and the fair volatility strike is called
the convexity correction. Some authors have obtained an approximation of this convexity cor-
rection using Taylors expansion, but Broadie and Jain (2007) show that it is not necessarily
accurate in the SV model. We compute fair volatility strikes by deriving a partial dieren-
tial equation which exploits a no-arbitrage relationship between variance and volatility swaps.
Gatheral (2006) provides a numerical integration approach for computing fair volatility strikes
in the SV model.
3 Pricing Volatility Derivatives
Over the past two decades, the volatility of an underlying stock or an index has developed as
an asset class in its own right. Variance swaps are very liquid instruments which can be used to
trade volatility and they can be regarded as underlying assets in order to price other volatility
sensitive instruments, including volatility swaps, variance options, VIX futures, etc. Using a
no-arbitrage argument, we derive a partial dierential equation to price volatility derivatives,
compute the fair volatility strike and price variance call and put options.
3.1 Pricing Volatility Swaps
Dene X
T
t
to be the price process of the oating leg of a variance swap:
X
T
t
= E
Q
t
_
1
T
_
T
0
v
s
ds
_
This security price X
T
t
depends on the variance, v
s
, of the underlying asset from time t = 0 until
maturity T. It has a payo at maturity, T, which is same as the oating leg of a continuous
variance swap. At time 0 it represents the fair variance strike:
K

var
= X
T
0
(13)
From equation (9) we know the value of this security at time 0 and we can derive the stochastic
dierential equation satised by the security X
T
t
:
dX
T
t
=
1 e
k(Tt)
kT

v
v
t
dW
1
t
(14)
6
This price process has zero drift since it is a forward price process. The process X
T
t
is driven by
the same Brownian motion W
1
t
as the variance process in the SV model. The volatility of the
price process, X
T
t
, goes to 0 as t approaches T.
Next we dene the price process of a security Y
T
t
which represents the oating leg of a volatility
swap:
Y
T
t
= E
Q
t
_

1
T
_
T
0
v
s
ds
_
This security has a payo at time T which depends on the variance process from time t = 0 until
maturity. At time T it represents the payo of the oating leg of the volatility swap. At time 0
it gives the fair volatility strike:
K

vol
= Y
T
0
These securities are similar to interest rate derivatives. The price of a zero coupon bond trading
in the market depends on the interest rate process from time 0 until the maturity of bond. An
interest rate is not a tradable market instrument, so for hedging any interest rate product we
use some other interest rate derivatives which are traded in the market. Similarly, the security
Y
T
t
depends on the variance process, v
s
, which is not a traded instrument in the market. Since
the security X
T
t
also depends on the variance process, there must be a relationship between the
price processes of Y
T
t
and X
T
t
to avoid arbitrage in the market. Using that relationship we can
hedge volatility derivatives using variance swaps.
Next we dene a state variable I
t
to measure the accumulated variance so far:
I
t
=
_
t
0
v
s
ds
This state variable is a known quantity at time t and satises the dierential equation:
dI
t
= v
t
dt
The forward price process, Y
T
t
, can be expressed as
Y
T
t
= E
t
_

1
T
_
I
t
+
_
T
t
v
s
ds
__
= F(t, v
t
, I
t
)
and is a function of time, the stochastic variance v
t
and a deterministic quantity I
t
. Applying
It os lemma to F() we get
dF =
F
t
dt +
F
v
dv +
F
I
dI +
1
2

2
F
v
2
dv
2
7
which can be simplied using equation (6) to
dF =
_
F
t
+
F
v
( v
t
) +
F
I
v
t
+
1
2

2
F
v
2
v
t

2
v
_
dt +
F
v

v

v
t
dW
1
t
(15)
Since F is a forward price process, its drift under the risk-neutral measure must be zero. Hence,
F
t
+
F
v
( v
t
) +
F
I
v
t
+
1
2

2
F
v
2
v
t

2
v
= 0 (16)
Thus, the forward price process satises the partial dierential equation (16) in the SV model.
We solve the partial dierential equation (16) in the region: 0 t T, I
min
I I
max
, v
min

v v
max
with the boundary condition
Y
T
T
= F(T, v
T
, I
T
) =
_
I
T
T
(17)
At other boundaries (I and V ) we set the second order variation of the price process to zero. In
particular, we use the boundary conditions:

2
F
I
2

(I=Imax,I
min
)
= 0

2
F
v
2

(v=vmax,v
min
)
= 0 (18)
Thus by solving the equation (16) with boundary conditions (17) and (18) we can compute
the fair volatility strike. By solving this partial dierential equation we get the price at all
times until maturity. The variance swap forward price process X
T
t
satises the same dierential
equation (16). The boundary condition in the case of a variance swap will be dierent at maturity
and is given by
X
T
T
= G(T, v
T
, I
T
) =
I
T
T
(19)
The analytical formula for the variance strike given by equation (9) solves the partial dierential
equation (16) with boundary conditions (18) and (19).
Exhibit 1 about here
Next we present numerical results to illustrate the computation of fair variance and fair volatility
strikes. We use model parameters similar to those estimated in Due, Pan and Singleton (2000),
which were found by minimizing mean-squared dierences between model and market S&P 500
8
option prices on November 2, 1993. We adjust the parameters slightly so that the fair continuous
variance strike is same in the two models. We assume a risk free rate of 3.19%. Exhibit 1 gives
these parameters. Exhibit 2 shows the fair variance strike and fair volatility strike of a one-
year maturity swap computed by solving the partial dierential equation (16) with appropriate
boundary conditions. We solve the PDE (16) on a three-dimensional grid with 400 points each in
the V - and I-directions and 2000 intervals in the t-direction. We also compute the fair variance
and fair volatility strikes using Monte Carlo simulation and a numerical integration approach
given in Broadie and Jain (2007). The theoretical value of the fair variance strike is computed
using equation (9). All results are computed under the dynamics of the SV model. We report
the square-root of fair variance strike,
_
K

var
, in the results. The fair variance strike for the
parameters in Exhibit 1 is (13.261%)
2
= 0.017585. The results from the PDE approach in this
section match the values obtained by other methods.
Exhibit 2 about here
Exhibit 3 illustrates the dependence of fair variance and fair volatility strikes on initial variance.
One advantage of the PDE method over simulation is that we get fair variance and fair volatil-
ity strikes for all values of initial variance and accumulated variance. Also, this approach gives
prices at all times until maturity. The left graph in Exhibit 3 presents the fair variance strike
(plotted as the square root of fair variance strike,
_
K

var
) and the fair volatility strike versus
initial volatility

v
0
. Equation (9) shows that the fair variance strike is a linear function of the
initial variance. The fair volatility strike is a not a linear function of the initial variance since its
payo is not a linear function of realized variance. Also, as given by the inequality (12), the fair
volatility strike is less than the fair variance strike.
Exhibit 3 about here
The convexity value is the dierence between the square root of fair variance strike and the fair
volatility strike. The right graph in Exhibit 3 plots the convexity value with initial volatility.
This illustrates that the convexity value is a decreasing function of initial volatility,

v
0
.
9
3.2 Pricing Variance Options
The price of a variance call option is given by:
C
t
= E
Q
t
[e
r(Tt)
max(V
c
(0, T) K, 0)] N (20)
We derive a partial dierential equation to price a variance call option using a no-arbitrage ar-
gument similar to that in the previous section. The payo of a variance call can be replicated by
continuous trading in a variance swap, and the replicating portfolio gives the price of variance
call option.
We form a portfolio of one variance call option and units of variance swaps. At time 0 the
portfolio value is

0
= (X
T
0
K

var
) +C
0
(21)
This portfolio value is the same as the variance call option value since there is no cost to buy
one unit of a variance swap at the inception of the contract. The variance call price process, C
T
t
,
can be represented as
C
T
t
= G(t, v
t
, I
t
)
Applying It os lemma to G() we get
dG =
G
t
dt +
G
v
dv +
G
I
dI +
1
2

2
G
v
2
dv
2
(22)
which can be simplied using equation (6) to
dG =
_
G
t
+
G
v
( v
t
) +
G
I
v
t
+
1
2

2
G
v
2
v
t

2
v
_
dt +
G
v

v

v
t
dW
1
t
(23)
From equation (21), the change in portfolio value in a small time dt is
d
t
= dF +dG (24)
Substituting equations (16), (15) and (23) in (24) and simplifying we obtain
d
t
=
_
F
v

v

v
t
dW
1
t
_
+
_
G
t
+
G
v
( v
t
) +
G
I
v
t
+
1
2

2
G
v
2
v
t

2
v
_
dt +
G
v

v

v
t
dW
1
t
(25)
If we choose =
G
v
/
F
v
then the stochastic component in the portfolio vanishes and equa-
tion (25) simplies to
d
t
=
_
G
t
+
G
v
( v
t
) +
G
I
v
t
+
1
2

2
G
v
2
v
t

2
v
_
dt (26)
10
Since the portfolio
t
is riskless, it should earn the risk free rate of return, and so
_
G
t
+
G
v
( v
t
) +
G
I
v
t
+
1
2

2
G
v
2
v
t

2
v
_
dt = rGdt (27)
which can be rewritten as
G
t
+
G
v
( v
t
) +
G
I
v
t
+
1
2

2
G
v
2
v
t

2
v
rG = 0 (28)
We solve the partial dierential equation (28) in the region: 0 t T, I
min
I I
max
, v
min

v v
max
with the boundary conditions (18).
We compute the price of variance call and variance put options of maturity one year for dierent
strikes. The at-the-money strike is K = (13.261%)
2
= 0.017585 from Exhibit 2. The other strikes
are given in Exhibit 4. We use the SV parameters in Exhibit 1. We solve the partial dierential
equation on a three-dimensional grid with 400 points each in the V - and I-directions and 2000
intervals in the t-direction. We assume a notional N = $1000 in our calculations. Option prices
are given in Exhibit 4. When the call and put options are both at-the-money, their prices are
the same due to the put-call parity relationship:
C
t
P
t
= X
T
t
Ke
r(Tt)
(29)
Exhibit 4 about here
4 Risk Management Parameters of Volatility Derivatives
In this section we compute greeks of variance and volatility swaps using partial dierential
equations and discuss properties of the greeks. These greeks are required for hedging volatility
derivatives. Delta units of volatility derivatives are used to dynamically hedge volatility swaps
with variance swaps as explained in Section 5.2. Other greeks are useful to understand the
sensitivity of the price of volatility derivatives to various parameters of the SV model.
11
4.1 Delta of Volatility Derivatives
We dene the delta of variance and volatility swaps as the rst-order variation in the fair strike
with respect to the variance, v
t
. Thus, the delta of a variance swap is
v
t

X
T
t
v
t
(30)
and the delta of a volatility swap is dened similarly. We compute the delta of a variance swap
analytically using equation (9) to get
v
t
=
X
T
t
v
t
=
1 e
(Tt)

(31)
Exhibit 5 about here
The delta of variance swap is constant and positive since the payo of the variance swap is a
linearly increasing function of realized variance. The delta of the variance swap approaches zero
as time to maturity decreases, since at maturity the payo of the variance swap is independent
of the initial variance. We compute the delta of the volatility swap numerically using rst-order
nite dierences. The left plot in Exhibit 5 shows the delta of variance and volatility swaps
versus initial volatility.
To make variance and volatility swap deltas comparable, note that:

_
X
T
0
v
0
=
1
2
_
X
T
0
X
T
0
v
0
=
1
2
_
K

var
X
T
0
v
0
(32)
and in Exhibit 5 we plot
_
X
T
0
/v
0
as the variance swap delta and Y
T
t
/v
t
as the volatility
swap delta. We computed volatility swap deltas numerically using rst-order nite dierence
and we used (31) and (32) to compute
_
X
T
0
/v
0
. This procedure is used for all greeks in the
following subsections to make the sensitivities comparable.
Using the parameters from Exhibit 1, the delta of fair variance strike,
_
X
T
0
/v
0
, is 60.6%. We
approximate the change in the fair variance strike as follows. A change in initial volatility from
10.1% to 11% implies a change in initial variance from 0.010201 to 0.0121 or v
0
= 0.001899.
The change in the fair variance strike is X
T
0
2
_
K

var
v
0
0.606 0.000303. This
12
implies the fair variance strike changes from (13.261%)
2
= 0.017586 to 0.017889 = (13.375%)
2
.
The actual value of fair variance strike at an initial volatility of 11% is (13.375%)
2
. As shown in
Exhibit 5, the delta of the fair volatility strike is a positive and decreasing function of variance
and volatility. Since the volatility swap payo is a concave function of realized variance, its delta
decreases with initial variance and volatility. The right plot in Exhibit 5 shows the dierence in
the deltas of variance and volatility swaps versus initial volatility.
Next we dene the sensitivities of strikes with respect to the parameters of the model.
4.2 Volatility Derivatives:
We dene as rst-order variation in fair strikes with respect to the mean reversion speed, .
For variance swaps it is dened as:

X
T
t

(33)
Using equation (9) we get
=
X
T
t

= (v
t
)
_
(T t)e
(Tt)
kT

1 e
(Tt)

2
T
_
(34)
Observe that approaches zero as time to maturity decreases, since at maturity the realized
variance is xed so all the sensitivities must approach zero. We compute the of the volatility
swap by dierentiating the partial dierential equation (16) with respect to the parameter :

t
+

v
( v) +
F
v
( v) +
1
2

2

v
2
v
2
v
= 0 (35)
We solve this partial dierential equation in the same domain 0 t T, I
min
I I
max
, V
min

V V
max
with the boundary conditions:
|
(t=T)
= 0 (36)

2

I
2

(I=Imax,I
min
)
= 0

2

v
2

(v=vmax,v
min
)
= 0 (37)
Exhibit 6 about here
13
The left plot in Exhibit 6 shows the sensitivity of fair strikes to the mean reversion speed as a
function of initial volatility. The variance strike sensitivity plotted in Exhibit 6 is:

_
X
T
0

=
1
2
_
X
T
0
X
T
0

=
1
2
_
K

var
X
T
0

(38)
We plot = Y
T
0
/ for the volatility swap, which we compute by solving the PDE (35).
The fair variance strike sensitivity,
_
X
T
0
/, to mean reversion speed, , is approximately
0.081% = 0.00081 at an initial volatility of

v
0
= 10.10%. We compute the approximate
change in the fair variance strike if the mean reversion speed, , changes from its initial level
6.21 to 7.21 at an initial volatility, 10.1%, as follows. The change in the fair variance strike is
X
T
0
2
_
K

var
0.00081 0.000215. This implies the fair variance strike changes
from (13.261%)
2
= 0.017586 to 0.017801 = (13.342%)
2
. The actual value of fair variance strike
at = 7.21 is 0.017781 = (13.334%)
2
. The graphs show that the sensitivity changes sign from
positive to negative as initial variance increases and the sign change occurs at the long-run mean
variance . When initial variance is lower than the long-run mean variance, , increasing the
mean reversion speed will result in an increase in the variance level, the realized variance will be
higher and hence a positive . The right plot in Exhibit 6 shows the dierence in the sensitivity
of the fair variance strike and the fair volatility strike to the mean reversion speed versus initial
volatility.
4.3 Volatility Derivatives:

We dene

as the rst-order variation in the fair strike with respect to the long-run mean
variance, . For variance swaps it is dened as:

X
T
t

(39)
Using equation (9) we compute the

of the variance swap and get

=
X
T
t

=
T t
T

1 e
(Tt)
T
(40)
The fair variance strike sensitivity to the long-run mean variance is constant and positive since
the realized variance increases as the long-run mean variance increases. We compute the

of
the volatility swap by dierentiating the partial dierential equation (16) with respect to the
parameter :

t
+

v
( v) +
F
v
+
1
2

v
2
v
2
v
= 0 (41)
14
We solve this partial dierential equation in the same domain 0 t T, I
min
I I
max
, V
min

V V
max
with the boundary conditions:

|
(t=T)
= 0 (42)

I
2

(I=Imax,I
min
)
= 0

2

v
2

(v=vmax,v
min
)
= 0 (43)
Exhibit 7 about here
The left plot in Exhibit 7 shows the sensitivity of the fair variance and volatility strikes to the
long-run mean variance as a function of initial volatility. As before, we plot the following quantity
for variance strike sensitivity:

_
X
T
0

=
1
2
_
X
T
0
X
T
0

=
1
2
_
K

var
X
T
0

(44)
For the fair volatility strike sensitivity we plot

= Y
T
0
/ which we compute by solving the
PDE (41). Exhibit 7 shows that the fair variance strike sensitivity,
_
X
T
0
/, is approximately
316% = 3.16 at an initial volatility of 10.1%. We compute the approximate change in the fair vari-
ance strike if long-run mean variance, , changes from 0.019 to 0.021 at an initial volatility 10.1%
as follows. The change in the fair variance strike is X
T
0
2
_
K

var
3.16 0.001676. This
implies the fair variance strike changes from (13.261%)
2
= 0.017585 to 0.019265 = (13.879%)
2
.
The actual value of fair variance strike at = 0.021 is 0.019262 = (13.879%)
2
. The variance
swap strike sensitivity to theta is constant at all variance levels. For volatility swaps

is positive,
which implies that the higher the long-run variance, the higher the fair volatility strike. The

of
the volatility swap is a decreasing function of initial variance. The right plot in Exhibit 7 shows
the dierence in the sensitivity of the fair variance and volatility strikes to the long-run mean
variance as a function of the initial volatility.
15
4.4 Volatility Derivatives:
We dene as the rst-order variation in fair strikes with respect to the volatility of variance
parameter,
v
. For variance swaps it is dened as:

v

X
T
t

v
(45)
Using equation (9) we nd that the fair variance strike is independent of the volatility of variance.
We compute the
v
of the volatility swap by dierentiating the partial dierential equation (16)
with respect to the parameter
v
:

v
t
+

v
v
( v) +
1
2

2

v
v
2
v
2
v
+

2
F
v
2
v
v
= 0 (46)
We solve this partial dierential equation in the domain 0 t T, I
min
I I
max
, V
min
V
V
max
with the boundary conditions:

v
|
(t=T)
= 0 (47)

2

v
I
2

(I=Imax,I
min
)
= 0

2

v
v
2

(v=vmax,v
min
)
= 0 (48)
Exhibit 8 about here
Exhibit 8 shows the sensitivity of fair strikes to the volatility of variance parameter as a function
of initial volatility. Consistent with equation (9), the fair variance strike is independent of the
volatility of variance. The fair volatility strike has a negative dependence on the volatility of
variance, i.e., an increase in the volatility of variance parameter will lead to a decrease in the
fair volatility strike. Since the fair volatility strike is a concave function of realized variance, the
fair volatility strike decreases with the increase in the volatility of variance parameter,
v
. For
convex payo functions, e.g, variance call and put options, the sensitivity with respect to the
volatility of variance is positive.
Thus all the greeks can be computed by either solving the pricing partial dierential equation (16)
and using nite dierence approximations (for delta) or by solving the other related partial
dierential equations with appropriate boundary conditions.
16
5 Hedging Volatility Derivatives
In this section we present an approach to hedge volatility derivatives using variance swaps. Other
authors (Demeter et al. 1999) have shown that variance swaps can be replicated using an innite
number of European call and put options. We formulate an optimization problem to nd the best
portfolio of European call and put options to closely replicate a variance swap for a given nite
number of options. We also analyze how replication error decreases as we increase the number
of European call and put options in the replicating portfolio. Then we present an approach to
dynamically hedge volatility swaps using variance swaps and a nite number of European call
and put options.
5.1 Replicating Variance Swaps
In this section we formulate an optimization problem for replicating a variance swap using a
static portfolio consisting of a nite number of European call and put options. Applying It os
lemma to the stock price diusion (5, 6) we can express realized variance as
V
c
(0, T) =
1
T
_
T
0
v
t
dt =
2
T
__
T
0
dS
t
S
t
ln
S
T
S
0
_
(49)
This result holds in both the Heston stochastic volatility model and the Black-Scholes model.
From (49) the realized variance can be replicated by shorting a log contract and dynamically
holding 1/S
t
shares of stock until the maturity of the contract. Next we review how to replicate
a European style payo, in particular a log contract payo (Neuberger 1994), statically using
call and put options. Let f be a twice continuously dierentiable function which represents the
payo of a European style path-independent derivative security. It can be expressed as (Breeden
and Litzenberger 1978)
f(S
T
) = f(x) +f

(x)(S
T
x) +
_

x
f

(K)(S
T
K)
+
dK +
_
x
0
f

(K)(K S
T
)
+
dK (50)
Thus, the payo function f can be replicated (Carr, Ellis and Gupta 1998) by holding positions
in a zero coupon bond with face value f(x), a forward contract with strike x, and call and put
options of all strikes using equation (50). The time zero value of the claim can be expressed in
terms of the European call C
0
(K) and put P
0
(K) prices of maturity T:
V
0
= E
Q
0
[e
rT
f(S
T
)]
= e
rT
f(x) +f

(x)[C
0
(x) P
0
(x)] +
_

x
f

(K)C
0
(K)dK +
_
x
0
f

(K)P
0
(K)dK
(51)
17
Now, let f(S
T
) = ln(S
T
/S
0
) and x = S
0
, and substitute in equation (50) to get:
ln
_
S
T
S
0
_
=
S
T
S
0
S
0

_

S
0
1
K
2
(S
T
K)
+
dK
_
S
0
0
1
K
2
(K S
T
)
+
dK (52)
Substituting this in equation (49) gives
V
c
(0, T) =
2
T
_ _
T
0
dS
t
S
t

S
T
S
0
S
0
+
_

S
0
1
K
2
C
T
(K)dK +
_
S
0
0
1
K
2
P
T
(K)dK
_
(53)
Thus, the oating leg of the variance swap can be replicated (Demeter et al. 1999) by a portfolio
having a short position in a forward contract struck at S
0
, a long position in 1/K
2
put options
of strike K, for all strikes from 0 to S
0
, a long position in 1/K
2
call options for all strikes from
S
0
to and payos from a dynamic trading strategy which instantaneously holds 1/S
t
shares
of stock worth $1 in the portfolio. In particular, equation (53) shows that continuously realized
variance can be replicated in both the Black-Scholes and the SV model.
Thus, to replicate the variance swap we need a short position in a log contract, and this can be
replicated using call and put options of all strikes (52). In practice we can only form a portfolio
using a nite number of options with a limited set of strikes. We analyze how well we can repli-
cate the log contract (and variance swaps) with a nite number of options.
Suppose we want to replicate the log contract with n
p
put options and n
c
call options of various
strikes and common maturity T. We dene the portfolio of a log contract and a forward contract
as portfolio B. Its payo at maturity T when the stock price is S
T
is given by
V
B
(S
T
) =
S
T
S
0
S
0
ln
_
S
T
S
0
_
(54)
Let w
c
i
represent the number of call options having strike K
c
i
and w
p
i
the number of put options
having strike K
p
i
in portfolio A. The payo of portfolio A at maturity T when the stock price is
S
T
is given by
V
A
(S
T
) =
np

i=1
w
p
i
(K
p
i
S
T
)
+
+
nc

i=1
w
c
i
(S
T
K
c
i
)
+
(55)
where (K
p
i
S
T
)
+
is the payo of the put option and (S
T
K
c
i
)
+
is the payo of the call option.
If we include options of all strikes in portfolio A, then portfolio A exactly replicates portfolio B
from equation (52). The quantities of options in portfolio A are unknown and we compute these
values using optimization so that the payo of portfolios A and B match as closely as possible
18
for xed number (n
c
and n
p
) of call and put options
2
.
To compute the number of options in portfolio A to most closely replicate the payo of portfolio B
we solve the optimization problem:
(P1) min
w
p
,w
c
n

j=1
_
V
A
(S
j
T
) V
B
(S
j
T
)
_
2
s.t.
np

i=1
w
p
i
P
0
(S
0
, K
p
i
) +
nc

i=1
w
c
i
C
0
(S
0
, K
c
i
) = P
B
(S
0
) (56)
In the problem (P1), the decision variables are vectors w
p
, w
c
of sizes n
p
and n
c
, respectively,
which represent the quantities of call and put options in the portfolio. The value V
B
(S
j
T
) is the
payo of the portfolio of log contract and forward contract when the terminal stock price is S
j
T
.
The value V
A
(S
j
T
) is the payo of the portfolio of call and put options when the terminal stock
price is S
j
T
. The value P
B
(S
0
) represents the initial value of the portfolio of the log contract and
forward contract. The value P
0
(S
0
, K
p
i
) represents the initial value of the put option with strike
K
p
i
and C
0
(S
0
, K
c
i
) represents the initial value of the call option with strike K
c
i
. The objective
function in (P1) minimizes the sum of squared dierences in two portfolio payos at the maturity
T over n scenarios. The constraint enforces the initial values of both portfolios to be equal. Thus
the portfolio optimization problem (P1) attempts to make the payos of the two portfolios as
close as possible given the constraint that initial value of the two portfolios must be equal.
To illustrate, we use the Black-Scholes and SV parameters in Exhibit 1, and set the matu-
rity to be one year. We choose the strikes of the call and put options to be equally dis-
tributed in a three standard deviation range denes as follows. For S
0
= $100 we choose
the put strikes to be equally distributed between S
0
e
(rT
1
2

2
T3

T)
= $68 and $100. Here,
we have chosen =
_
K

var
= 13.26%. Thus, for n
p
put options in the (P1), the put strikes
are K
p
i
= 68 + (i 1)(100 68)/(n
p
1), i = 1, . . . , n
p
. Similarly we choose call strikes
to be equally distributed between $100 and S
0
e
(rT
1
2

2
T+3

T)
= $152. The call strikes are
K
c
i
= 100 + (i 1)(152 100)/(n
c
1), i = 1, . . . , n
c
. We choose the n scenarios within
a four standard deviation range. In particular, we take n = 200 scenarios of stock prices,
S
j
T
= 60 + (j 1)(173 60)/(n 1), j = 1, . . . , n. The strike range is narrower than the range
used in the optimization scenarios to reect the practical reality that not all strikes are actively
traded in the market.
19
To compare the performance of the replicating portfolio of call and put options we compute three
types of error:
e
1
:
E

V
A
(S
T
) V
B
(S
T
)

P
B
(S
0
)
(57)
e
2
:

E
_
V
A
(S
T
) V
B
(S
T
)
_
2
P
B
(S
0
)
(58)
e

:
max

V
A
(S
T
) V
B
(S
T
)

P
B
(S
0
)
(59)
where P
B
(S
0
) represents the value of portfolio B, given in (54), at t = 0 when the stock price is
S
0
. The expectation is under the real-world probability measure.
The objective function in (P1) uses equally weighted scenarios, while the error measures e
1
and
e
2
weight the scenarios by their real-world probabilities, so that extreme outcomes have less eect
on the results. The error measure e

will be determined by the single scenario with the most

extreme outcome. We use equal weighting for scenarios in the objective function so that the
portfolio will perform reasonably well under all three error measures. By solving one optimiza-
tion problem instead of three, we reduce the number of results presented. If the user is interested
in one particular error measure, that measure should be substituted in (P1).
We solve (P1) by forming the Lagrangian and solving the resulting system of linear equations.
The solution of optimization problem (P1) gives quantities of call and put options to replicate a
log contract for a given number of puts and calls in the replicating portfolio. Using this portfolio
of call and put options we analyze the dynamic replication of a variance swap using a nite num-
ber of options. We compute these error measures by simulating scenarios in the Black-Scholes
and SV real-world probability measures. The value of the drift in the real-world probability
measures from Exhibit 1 is = 7%.
Exhibit 9 about here
20
Exhibit 10 about here
Equation (53) gives the formula to replicate continuous realized variance using a portfolio of call
and put options and continuous rebalancing of 1/S
t
shares of stock. There are two types of errors
in replicating a variance swap with a nite number of options. The rst type of error comes from
replicating a log contract by a nite number of options. The second type of error comes from
the discrete rebalancing of 1/S
t
shares of stock worth $1 in the portfolio. In practice, realized
variance is computed for a discrete sampling frequency, e.g, daily or weekly. In order to isolate
the eects of the sampling frequency, we present results for the replication of a discrete variance
swap for two sampling frequencies: a high sampling frequency of 16 times per day and a daily
sampling frequency. In the rst case, there are n = 4096 sampling observations during the life
of a T = 1 year variance swap, i.e., 16 samples per day assuming 256 trading days in a year.
The same time interval (t = 1/4096 years) is used to computed the payos from the dynamic
trading strategy. In this case the majority of the replication error is due to a nite number of
options. These results are given in Exhibits 9 and 10.
In Exhibits 9 and 10 a static portfolio of options determined from the solution of (P1) is used to
replicate the log contract. When computing the variance swap replication errors, the portfolio A
payo is the discrete realized variance payo at maturity and the portfolio B payo is composed
of the payos from the options portfolio, a short forward contract and dynamic rebalancing with
the stock (53). To normalize the results, we set P
B
(S
0
) to K

var
(equation (9)) in the three
error measures. To compute the error measures we simulated 10,000 stock price paths under
the real-world measures in the Black-Scholes and SV models. The results show all three error
measures decrease as we increase the number of options in the replicating portfolio.
3
With 16
options (8 puts and 8 calls in the option portfolio) the mean absolute replication error for the
Black-Scholes model is about 1.9% of the initial value of the portfolio and 2.0% for the SV model.
Exhibit 10 shows replication errors and number of options on log scale with dierent number
of options in the Black-Scholes and SV models. These results show error measures e
1
and e
2
converge quadratically to zero as the number of options increases.
Exhibit 11 about here
Next we analyze the eect of a daily sampling interval. We assume that the sampling interval in
computing the realized variance is the same as the rebalancing interval in the dynamic trading
21
strategy. Exhibit 11 shows the daily sampling results using a nite number of options. The
error measures in Exhibits 9 and 11 are very similar, indicating that daily rebalancing is as eec-
tive as more frequent rebalancing, since the error is dominated by using a nite number of options.
5.2 Hedging Volatility Derivatives in the SV Model
In this section we present an approach to dynamically hedge volatility swaps using variance
swaps in an SV model. Suppose we take a long position in one unit of volatility swap at t = 0
of maturity T with fair volatility strike, K

vol
. The volatility swap is initially costless. At time t,
the value of volatility swap contract is
P
t
= E
t
(e
r(Tt)
(
_
V
c
(0, T) K

vol
)) = e
r(Tt)
(Y
T
t
K

vol
) (60)
We assume the notional amount of swap to be $1. To hedge a long position in volatility swap at
time t, we construct a portfolio having one unit of volatility swap and
t
units of variance swaps.
Thus the portfolio value at time t equals

t
= E
t
_
e
r(Tt)
_

t
(V
c
(0, T) K

var
)
_
+ (
_
V
c
(0, T) K

vol
)
_
= e
r(Tt)
_

t
(X
T
t
K

var
) + (Y
T
t
K

vol
)
_
(61)
The change in this portfolio in a small amount of time dt is given by
d
t
= r
t
dt +e
r(Tt)
_

t
dX
T
t
+dY
T
t
_
which can be written using equation (15) as
d
t
= r
t
dt +e
r(Tt)
_

t
__
X
T
t
t
+
X
T
t
v
( v
t
) +
X
T
t
I
v
t
+
1
2

2
X
T
t
v
2
v
2
t

2
v
_
dt +
X
T
t
v

v

v
t
dW
1
t
_
+
_
Y
T
t
t
+
Y
T
t
v
( v
t
) +
Y
T
t
I
v
t
+
1
2

2
Y
T
t
v
2
v
2
t

2
v
_
dt +
Y
T
t
v

v

v
t
dW
1
t
_
Since the processes X
T
t
and Y
T
t
satisfy the pricing partial dierential equation (16), the dt terms
in the previous equation vanish. Hence the change in the portfolio value can be rewritten as:
d
t
= r
t
dt +e
r(Tt)
_

t
X
T
t
v

v

v
t
dW
1
t
+
Y
T
t
v

v

v
t
dW
1
t
_
(62)
We dene
t
as the volatility swap hedge ratio:

t
=
Y
T
t
v
X
T
t
v
(63)
22
If we choose
t
as in equation (63), the stochastic component of portfolio vanishes and the port-
folio value is hedged. Thus for hedging a volatility swap we can take a short position in
t
units
of the variance swap and the portfolio value is hedged dynamically.
Next, we present numerical results for the volatility swap hedging performance. We compute the
prot and loss of two dierent hedging strategies and compare with no hedging. The two hedge
portfolios are: a portfolio containing one unit of volatility swap and
t
(from equation (63))
units of a variance swap and a portfolio containing one unit of volatility swap and a portfolio
of European call and put options which replicates
t
units of variance swap. We compute the
portfolio of call and put options which replicates a variance swap as described in Section 5.1.
No hedging: We price the variance and volatility swap of maturity one year using the partial
dierential equation described in Section 3.1. We also compute the deltas at time zero of the
variance and volatility swaps using (30). Together these give the hedge ratio at time zero. We
generate 4,800 scenarios of the stock price and variance level at t = 1/252 years. The variance
and volatility swaps are initially costless. The prot and loss of a long position in an unhedged
volatility swap at t = 1/252 years is equal to the price of volatility swap contract:
P
t
= E
t
(e
r(Tt)
(
_
V
c
(0, T) K

vol
)) = e
r(Tt)
(Y
T
t
K

vol
) (64)
Hedging with variance swaps: We form a portfolio containing one unit of a volatility swap
and
0
units of variance swaps. The value of this portfolio is zero at t = 0:

0
=
0
(X
T
0
K

var
) + (Y
T
0
K

vol
)
where
0
=
Y
T
0
v
/
X
T
0
v
is the hedge ratio. The prot and loss of this hedged portfolio at t = 1/252
years is equal to the value of this portfolio:

t
= E
t
_
e
r(Tt)
_

0
(V
c
(0, T) K

var
)
_
+ (
_
V
c
(0, T) K

vol
)
_
= e
r(Tt)
_

0
(X
T
t
K

var
) + (Y
T
t
K

vol
)
_
(65)
Exhibit 12 about here
Hedging with options: We form a portfolio containing one unit of volatility swap and
t
units
of a portfolio of call and put options which replicates a variance swap. We replicate a variance
23
swap using portfolio of call and put options as described in Section 5.1. In these results we are
replicating a continuous variance swap with a portfolio of put and call options and a forward
contract and payo from a dynamic trading strategy which holds 1/S
t
shares of stock (see equa-
tion (53)). For simplicity, we have assumed that rebalancing is done continuously in computing
payos from the dynamic trading strategy. In this hedging exercise we present results using eight
(four calls and four puts) options and 32 (16 calls and 16 puts) options. We compute the prot
and loss of this portfolio at time t = 1/252 years in both cases: hedging with eight options and
hedging with 32 options.
Exhibit 12 shows the performance of hedging volatility swaps with variance swaps and a nite
number of options. We compute the error measures e
1
(57), e
2
(58) and e

(59) of prot and

loss using 4,800 scenarios of stock prices and variance levels at t = 1/252 years. In these results
the error measures are normalized by the variance strike, K

var
, dened in (9). We use a batching
method with twelve batches to compute the standard error estimates in error measures. The
hedging errors are for hedging over an interval of 1/252 year compared to the Exhibit 9 where the
hedging interval is one year. From Exhibit 12 we can see that the absolute value of the volatility
swap prot and loss is about 5.29% of the variance strike, K

var
, over a single day. Hedging a
volatility swap with a variance swap reduces this to 0.03%, which is quite signicant. Hence,
a volatility swap can be eectively hedged dynamically using variance swaps. The results also
show that hedging with eight options reduces the absolute value of the prot and loss to 0.15%
and with 32 options to 0.059%. The error in hedging volatility swaps with options decreases as
we increase the number of options. Dynamic hedging of volatility swaps with variance swaps and
options is seen to be quite eective.
6 Conclusion
In this paper we presented a partial dierential equation approach to price volatility derivatives in
Hestons SV model. The pricing of volatility derivatives (including volatility swaps and variance
options) is complicated since the underlying variable, realized variance, is not a market traded
instrument. We exploited a no-arbitrage relationship between variance swaps and other volatility
derivatives to derive a partial dierential equation to price volatility derivatives. We also derived
PDEs for the greeks in order to hedge volatility derivatives. We presented an optimization model
for the practical hedging of variance swaps using a nite number of options. We also presented
24
an approach to hedge volatility derivatives using variance swaps, and showed the hedge to be
very eective.
25
References
Breeden, D. and Litzenberger, R. (1978), Prices of state contingent claims implicit in option
prices, Journal of Business 51(6), 621651.
Broadie, M. and Jain, A. (2007), Discrete and continuously sampled volatility and variance
swaps. Working paper, Columbia Business School.
Brockhaus, O. and Long, D. (2000), Volatility swaps made simple, Risk 19(1), 9295.
Carr, P., Ellis, K. and Gupta, V. (1998), Static hedging of exotic options, Journal of Finance
53(3), 11651190.
Carr, P., Geman, H., Madan, D. and Yor, M. (2005), Pricing options on realized variance,
Finance and Stochastics 9(4), 453475.
Carr, P. and Lee, R. (2005), Robust replication of volatility derivatives. Working paper, Courant
Institute and University of Chicago, http://math.nyu.edu/research/carrp/research.html.
Demeter, K., Derman, E., Kamal, M. and Zou, J. (1999), A guide to volatility and variance
swaps, Journal of Derivatives 4, 932.
Detemple, J. B. and Osakwe, C. (2000), The valuation of volatility options, European Finance
Review 4(1), 2150.
Due, D., Pan, J. and Singleton, K. (2000), Transform analysis and asset pricing for ane jump
diusions, Econometrica 68, 13431376.
Gatheral, J. (2006), The volatility surface: A practitioners guide, John Wiley & Sons, Inc. .
Heston, S. (1993), A closed form solution of options with stochastic volatility with applications
to bond and currency options, The Review of Financial Studies 6(2), 327343.
Javaheri, A., Wilmott, P. and Haug, E. (2002), Garch and volatility swaps. Working Paper,
http://www.wilmott.com.
Lipton, A. (2000), Mathematical methods for foreign exchange: A nancial engineers approach,
World Scientic .
Little, T. and Pant, V. (2001), A nite dierence method for the valuation of variance swaps,
Journal of Computational Finance 5(1), 81103.
26
Neuberger, A. J. (1994), The log contract, Journal of Portfolio Management 20(2), 7480.
Sepp, A. (2006), Pricing options on realized volatility in heston model with volatility jumps.
Working paper, Bear Stearns Inc. and Institute of Mathematical Statistics, University of
Tartu.
Notes
1
For the concave square root function Jensens inequality is:
E(

x)
_
E(x)
2
We have chosen portfolio A to hold both call and put options. We can also choose this portfolio to consist of
call options or put options only as we can replace the put options by call options and stock using put-call parity.
3
The error e

does not decrease as fast because it is sensitive to a single simulation scenario which may lie
outside the range of scenarios used in the optimization.
27
Exhibit 1
Black-Scholes and SV model parameters used in pricing and hedging
Parameters BS SV
correlation n/a -0.7
long-run mean variance n/a 0.019
speed of mean reversion n/a 6.21
volatility of variance
v
n/a 0.31
initial volatility

v
0
13.261% 10.10%
risk free rate r 3.19% 3.19%
real-world growth rate 7.0% 7.0%
28
Exhibit 2
Comparison of fair variance and fair volatility strikes in the SV model using dierent
numerical methods
K

var
(%
2
) K

vol
(%)
Simulation price 13.259 13.094
(standard error) (0.057) (0.002)
PDE 13.261 13.096
Analytical 13.261
Numerical integration 13.096
The rst column shows the fair variance strike com-
puted using the PDE method, simulation and ana-
lytical value in the SV model. The second column
shows the fair volatility strikes computed with the
same methods.
29
Exhibit 3
0 5 10 15 20 25 30
12
13
14
15
16
17
18
Initial volatility (%)
F
a
i
r

V
a
r
i
a
n
c
e

a
n
d

V
o
l
a
t
i
l
i
t
y

S
t
r
i
k
e

(
%
)
fair variance strike
fair volatility strike
0 10 20 30
0.35
0.4
0.45
0.5
0.55
Initial volatility (%)
C
o
n
v
e
x
i
t
y

V
a
l
u
e

(
%
)
The left plot shows the square root of the fair variance strike and fair volatility strike versus initial volatility. The
right plot shows the convexity value (66) versus initial volatility.
30
Exhibit 4
Prices of variance call and put options in the SV model
Strike K (%)
2
Call ($) Put ($)
10.272 7.127 0.314
11.095 5.575 0.465
12.581 3.101 1.398
13.261 2.220 2.220
14.527 1.068 4.475
15.691 0.481 7.295
These prices are for one-year maturity op-
tions corresponding to the SV model pa-
rameters given in Exhibit 1.
31
Exhibit 5
0 5 10 15 20 25 30
50
52
54
56
58
60
62
64
Initial volatility (%)
Variance swap
Volatility swap

v
t
(
%
)
0 5 10 15 20 25 30
4
2
0
2
4
6
8
10
12
Initial volatility (%)
D
i
f
f
e
r
e
n
c
e

i
n

d
e
l
t
a

(
%
)
The left plot shows the sensitivity of fair variance strikes (32) and fair volatility strikes to initial variance as a
function of initial volatility. The right plot shows the dierence in the deltas of fair variance and volatility strikes.
32
Exhibit 6
0 10 20 30
3
2.5
2
1.5
1
0.5
0
0.5
1
Initial volatilty (%)
Variance swap
Volatility swap

(
%
)
0 10 20 30
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
Initial volatility (%)
D
i
f
f
e
r
e
n
c
e

(
%
)
The left plot shows the sensitivity of fair variance strikes (38) and fair volatility strikes (35) to the mean reversion
speed as a function of initial volatility. The right plot shows the dierence between the two sensitivities.
33
Exhibit 7
0 5 10 15 20 25 30
240
260
280
300
320
340
360
Initial volatility (%)
Variance swap
Volatility swap

(
%
)
0 10 20 30
40
20
0
20
40
60
80
Initial volatility (%)
D
i
f
f
e
r
e
n
c
e

(
%
)
The left plot shows the sensitivity

of fair variance strikes (44) and fair volatility strikes (41) to long-run mean
variance (39) as a function of initial volatility. The right plot shows the dierence between the two sensitivities.
34
Exhibit 8
0 5 10 15 20 25 30
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0.2
Initial volatility (%)
Variance swap
Volatility swap

v
(
%
)
0 10 20 30
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
Initial volatility (%)
D
i
f
f
e
r
e
n
c
e

(
%
)
The left plot shows the sensitivity
v
of fair variance strikes and fair volatility strikes to the volatility of variance

v
(45) as a function of initial volatility. The right plot shows the dierence between the two sensitivities.
35
Exhibit 9
Error in the dynamic replication of a variance swap with a nite number of options
Number of Black-Scholes SV
Options Error e
1
Error e
2
Error e

Error e
1
Error e
2
Error e

4 0.381 0.483 1.574 0.392 0.494 5.024

8 0.091 0.115 0.280 0.093 0.128 4.248
16 0.019 0.024 0.120 0.020 0.050 3.411
32 0.004 0.006 0.077 0.005 0.037 2.986
The rst column shows the number of options used in replicating a continuous (almost)
variance swap of maturity one year. The second column shows the error measure dened
in (57) (divided by the variance strike to normalize the error) in the Black-Scholes model
and third column shows the error measure dened in (58). The fourth column shows the
error measure dened in (59). The fth, sixth and seventh column shows the respective
error measures in the SV model. These error measures are computed for a time interval
of one year for a variance swap in which realized variance is computed using (n = 4096)
sampling observations and rebalancing is done every dt = 1/4096 year in computing payos
from dynamic trading strategy.
36
Exhibit 10
5 10 20 40
0.005
0.01
0.05
0.1
Replication error in the BS model
Number of options
R
e
p
l
i
c
a
t
i
o
n

e
r
r
o
r
Error e1
Error e2
5 10 20 40
0.005
0.01
0.05
0.1
Replication error in the SV model
Number of options
R
e
p
l
i
c
a
t
i
o
n

e
r
r
o
r
Error e1
Error e2
Replication error in a variance swap as a function of the number of options. The gures show error measures
dened in (57) and (58) in replicating a variance swap with a nite number of options in the Black-Scholes and
SV models. These error measures are computed for a time interval of one year for a variance swap in which
realized variance is computed using (n = 4096) sampling observations and rebalancing is done every dt = 1/4096
year in computing payos from dynamic trading strategy. The gures are plotted on a log-log scale.
37
Exhibit 11
Error in the dynamic replication of a discrete variance swap with a nite number of
options
Number of Black-Scholes SV
Options Error e
1
Error e
2
Error e

Error e
1
Error e
2
Error e

4 0.381 0.485 1.573 0.393 0.496 5.032

8 0.090 0.114 0.280 0.096 0.127 4.254
16 0.019 0.024 0.124 0.019 0.050 3.421
32 0.004 0.005 0.046 0.005 0.036 2.986
The rst column shows the number of options used in replicating a discrete variance swap of
maturity one year with daily sampling. The rebalancing interval in computing payos from
the dynamic trading in underlying stock is daily as well. The second column shows the error
measure dened in (57) (divided by the variance strike to normalize the error measures) in the
Black-Scholes model and third column shows the error measure dened in (58). The fourth
column shows the error measure dened in (59). The fth, sixth and seventh column shows
the respective error measures in the SV model.
38
Exhibit 12
Hedging a volatility swap using variance swaps and a nite number of options
Error e
1
Error e
2
Error e

Avg. (SE) (%) Avg. (SE) (%) Avg. (SE) (%)

Volswap Unhedge 5.293 (0.066) 6.646 (1.044) 21.811 (0.745)
Varswap Hedge 0.029 (0.001) 0.042 (0.001) 0.289 (0.021)
Options Hedge (8) 0.146 (0.002) 0.184 (0.031) 0.617 (0.018)
Options Hedge (32) 0.057 (0.001) 0.076 (0.014) 0.342 (0.018)
The second, third and fourth column show the error measures e
1
(57), e
2
(58) and
e

(59) of prot and loss in dierent hedging strategies for hedging volatility swap
over an interval of 1/252 years. The rst row shows the mean and standard error
of prot and loss of unhedged volatility swap. The second row shows the mean and
standard error of prot and loss of a hedged volatility swap using variance swaps.
The third row shows the mean and standard error of prot and loss of a hedged
volatility swap with eight options. The fourth row shows the mean and standard
error of prot and loss of hedged volatility swap with 32 options.
39