This action might not be possible to undo. Are you sure you want to continue?
Volatility Model with Regime Switching
Robert J. Elliott
∗
Tak Kuen Siu
†
Leunglung Chan
‡
January 16, 2006
Abstract
We develop a model for pricing volatility derivatives, such as variance swaps and
volatility swaps under a continuoustime Markovmodulated version of the stochastic
∗
The Corresponding Author: RBC Financial Group Professor of Finance, Haskayne School of
Business, University of Calgary, Calgary, Alberta, Canada, T2N 1N4; Email: relliott@ucalgary.ca;
Fax: 4037708104; Tel: 4032205540
†
Lecturer, Department of Actuarial Mathematics and Statistics, HeriotWatt University, Edin
burgh, United Kingdom; Email: T.K.Siu@ma.hw.ac.uk
‡
Ph.D. candidate, Department of Mathematics and Statistics, University of Calgary, Calgary,
Alberta, Canada; Email: lchan@math.ucalgary.ca
1
volatility (SV) model developed by Heston (1993). In particular, we suppose the
parameters of our version of Heston’s SV model depend on the states of a continuous
time observable Markov chain process, which can be interpreted as the states of an
observable macroeconomic factor. The market we consider is incomplete in general,
and hence, there is more than one equivalent martingale pricing measure. We adopt
the regime switching Esscher transform used by Elliott et al. (2005) to determine
a martingale pricing measure for the valuation of variance and volatility swaps in
this incomplete market. We consider both the probabilistic approach and the partial
diﬀerential equation, (PDE), approach for the valuation of volatility derivatives.
Key words: Regime Switching Esscher Transform; Markovmodulated Heston’s
SV model; Observable Markov Chain Process; Volatility Swaps; Variance Swaps;
Regime Switching OUprocess
2
§1. Introduction and Summary
Volatility is one of the major features used to describe and measure the ﬂuctuations
of asset prices. It is popular as a measure of risk and uncertainty. It plays a signiﬁcant
role in three pillars of modern ﬁnancial analysis: risk management, option valuation
and asset allocation. There are diﬀerent measures of volatility, including realised
volatility, implied volatility and modelbased volatility. Realised volatility is the
standard deviation of the historical ﬁnancial returns; implied volatility is the volatility
inferred from the market option price data based on an assumed option pricing model,
such as the BlackScholes model; modelbased volatility includes both parametric and
nonparametric speciﬁcations of the volatility dynamics, such as the ARCH models
introduced by Engle (1982), the GARCH models introduced by Bollerslev (1986) and
Taylor (1986), independently, and their variants. These models are introduced to
provide a better speciﬁcation, measurement and forecasting of volatilities of various
ﬁnancial assets. Recent major ﬁnancial issues, such as the collapse of LTCM, the
Asian ﬁnancial crisis and the problems of Barings and Orange Country, reveal that
the global ﬁnancial markets have become more volatile. Due to large and frequent
shifts in the volatilities of various assets in the recent past, there has been a growing
and practical need to develop some models with related ﬁnancial instruments to
hedge volatility risk. On the other hand, market speculators may be interested in
3
guessing the direction that the volatility may take in the future. These practical needs
facilitate the growth of the market for derivative products related to volatility. The
initial suggestions for such products included options and futures on the volatility
index. Brenner and Galai (1989, 1993) proposed the idea of developing a volatility
index. The Chicago Board Options Exchange (CBOE) introduced a volatility index
in 1993, called VIX, which was based on implied volatilities from options on the S&P
500 index. The VIX Index reﬂects the market expectations of nearterm volatility
inferred from the S&P 500 stock index option prices. For a detailed discussion about
various volatility derivative products, see Brenner et al. (2001).
Variance swaps and volatility swaps are popular volatility derivative products
and they have been actively traded in overthecounter markets since the collapse
of LTCM in late 1998. In particular, the variance swaps and the volatility swaps
on stock indices, currencies and commodities are quoted and traded actively. These
products are popular among market practitioners as a hedge for volatility risk. The
variance swap is a forward contract in which the long position pays a ﬁxed amount
K
var
/$1 nominal value at the maturity date and receives the ﬂoating amount (σ
2
)
R
/$1
nominal value, where K
var
is the strike price and (σ
2
)
R
is the realized variance.
The volatility swap is the same as the variance swap, except that the realized
variance (σ
2
)
R
is replaced by the realized volatility (σ)
R
.
4
Recently, there has been a considerable interest in pricing and hedging variance
swaps and volatility swaps. Gr¨ unbuchler and Longstaﬀ (1996) developed pricing
models for options on variance based on Heston’s stochastic volatility model. Carr
and Madan (1998) showed how derivatives on volatility can be valued. Demeterﬁ
et al. (1999) discussed the properties and valuation of volatility swaps. Heston
and Nandi (2000) discussed the valuation of options and swaps on variance and
volatility in the context of discretetime GARCH models. Brockhaus and Long (2000)
discussed the pricing issues of volatility swaps based on Heston’s stochastic volatility
model. Javaheri et al. (2002) developed a partial diﬀerential equation approach for
pricing volatility swaps under a continuoustime version of the GARCH(1, 1) model.
Brenner et al. (2001) considered the use of the SteinStein model to value volatility
swaps. Matytsin (2000) considered the valuation of volatility swaps and options on
variance for a stochastic volatility model with jumpdiﬀusion dynamics. Howison et
al. (2004) also considered the valuation of volatility swaps for a stochastic volatility
model driven by a jumpdiﬀusion. They used an asymptotic approximation for the
solution of the partial diﬀerential equation for pricing. Carr et al. (2005) valued
options on variance which were based on quadratic variation. Swishchuk (2004) used
an alternative probabilistic approach to value variance and volatility swaps under
the Heston (1993) stochastic volatility model. Elliott and Swishchuk (2004) valued
5
variance swaps in the context of a fractional BlackScholes market. Swishchuk (2005)
developed a model for valuing variance swaps under a stochastic volatility model with
delay.
In this paper, we develop a model for pricing volatility derivatives, such as vari
ance swaps and volatility swaps under a continuoustime Markovmodulated version
of Heston’s stochastic volatility (SV) model, (Heston (1993)). In particular, the pa
rameters of a continuoustime version of Heston’s SV model depend on the states of
a continuoustime observable Markov chain process, which can be interpreted as the
states of an observable macroeconomic factor, such as an observable economic indica
tor for business cycles and the sovereign ratings for the region. The market described
by the Markovmodulated model is incomplete in general, and hence, there is more
than one equivalent martingale pricing measure. We adopt the regime switching Es
scher transform introduced by Elliott et al. (2005) to determine a martingale pricing
measure for the valuation of variance and volatility swaps. We consider both the
probabilistic approach and the PDE approach for the valuation of volatility deriva
tives. We assume that the Markov chain process is observable in both the proba
bilistic approach and the PDE approach. We shall document economic consequences
for the prices of the variance swaps and volatility swaps of the incorporation of the
regimeswitching in the volatility dynamics by conducting a Monte Carlo experiment.
6
The next section describes the model. Section three demonstrates the use of the
probabilistic approach for pricing volatility and variance swaps under the continuous
time Markovmodulated version of Heston’s stochastic volatility model. Section four
considers the use of the PDE approach for the valuation. In Section ﬁve, we shall con
duct the Monte Carlo experiment for comparing the prices implied by the stochastic
volatility models with and without switching regimes. We shall document the eco
nomic consequences for the prices of switching regimes. The ﬁnal section suggests
some possible topics for further investigation.
§2. The Model
In this section, we adopt the probabilistic approach for the valuation of volatility
derivatives under a continuoustime Markovmodulated stochastic volatility model.
Our model can be considered the regimeswitching augmentation of the
model by Swishchuk (2004) for pricing and hedging volatility swaps. The
Markovmodulated version of Heston’s stochastic volatility model can describe the
consequences for the asset price and volatility dynamics of the transitions of the
states of an observable macroeconomic factor, which aﬀects the asset prices and
volatility dynamics, such as an observable economic indicator of business cycles, or
the sovereign ratings of the region by some international rating agencies, such as
7
the Standard & Poors, Moodys and Fitch, etc. In particular, the parameters in the
asset price dynamics and the stochastic volatility dynamics depend on the state of
an economic indicator, which is described by the observable Markov chain.
First, consider a continuoustime ﬁnancial model with two primary securities,
namely a riskfree bond B and a risky stock S. Fix a complete probability space
(Ω, F, P) with P being the realworld probability measure. Let T be the time index
set [0, ∞). On (Ω, F, P) we consider a continuoustime ﬁnitestate observable Markov
chain X := {X
t
}
t∈T
with state space S which might be the set {s
1
, s
2
, . . . , s
N
}, where
s
i
∈ R
N
, for i = 1, 2, . . . , N. The states of the Markov chain process X describe the
states of an observable economic indicator. Without loss of generality, the state space
of the chain can be identiﬁed with the set {e
1
, e
2
, . . . , e
N
} of unit vectors in R
N
.
Write Π(t) for the generator, or Qmatrix, [π
ij
(t)]
i,j=1,2,...,N
of X. Following Elliott
et al. (1994), the following semimartingale representation theorem for the process
X can be obtained:
X
t
= X
0
+
_
t
0
Π(s)X
s
ds +M
t
. (2.1)
Here {M
t
}
t∈T
is an R
N
valued martingale increment process with respect to the
natural ﬁltration generated by X.
Let W
1
:= {W
1
t
}
t∈T
and W
2
:= {W
2
t
}
t∈T
denote two correlated standard
8
Brownian Motions on (Ω, F, P) with respect to the Paugmentation of the ﬁltra
tion F
W
:= {F
W
t
}
t∈T
, where W
t
:= (W
1
, W
2
). We assume that
Cov(dW
1
t
, dW
2
t
) = ρdt , (2.2)
where ρ ∈ (−1, 1) is the instantaneous correlation coeﬃcient between W
1
and W
2
.
We further suppose that the processes X is independent with (W
1
, W
2
). Pan
(2002) documented from her empirical studies on stock indices that the
correlation coeﬃcient between the diﬀusive shocks to the volatility level
and the level of the underlying price is signiﬁcantly negative. The model
considered here can incorporate this negative correlation coeﬃcient.
Let {r(t, X
t
)}
t∈T
be the instantaneous market interest rate of the bond B, which
depends on the state of the economic indicator described by X; that is,
r(t, X
t
) =< r, X
t
> , t ∈ T , (2.3)
where r := (r
1
, r
2
, . . . , r
N
) with r
i
> 0 for each i = 1, 2, . . . , N and < ·, · > denotes
the inner product in R
N
.
To simplify the notation, write r
t
for r(t, X
t
). Then, the dynamics of the price
9
process {B
t
}
t∈T
for the bond B are described by:
dB
t
= r
t
B
t
dt ,
B
0
= 1 . (2.4)
Suppose the expected appreciation rate {µ
t
}
t∈T
of the risky stock S depends on the
state of the economic indicator X and is described by:
µ
t
:= µ(t, X
t
) =< µ, X
t
> , (2.5)
where µ := (µ
1
, µ
2
, . . . , µ
N
) with µ
i
∈ R, for each i = 1, 2, . . . , N.
Let {α
t
}
t∈T
denote the longterm volatility level. We assume that α
t
depends on
the state of the economic indicator X and is given by:
α
t
:= α(t, X
t
) =< α, X
t
> , (2.6)
where α := (α
1
, α
2
, . . . , α
N
) with α
i
> 0, for each i = 1, 2, . . . , N.
Let β and γ denote the speed of mean reversion and the volatility of volatility,
respectively. (In general, we could consider a more general case where both the
speed of mean reversion β and the volatility of volatility γ depend on the states of
the economic indicator X.) However, in order to make our model more analytically
tractable, we suppose that both are constant. Suppose that the dynamics of the price
10
process {S
t
}
t∈T
and the shortterm volatility process σ := {σ
t
}
t∈T
of the risky stock
are governed by the following equations:
dS
t
= µ
t
S
t
dt +σ
t
S
t
dW
1
t
,
dσ
2
t
= β(α
2
t
−σ
2
t
)dt +γσ
t
dW
2
t
, (2.7)
where Cov(dW
1
t
, dW
2
t
) = ρdt.
Note that the variance process σ
2
t
follows a Cox, Ingersoll and Ross (1985) process.
Let ˜ ρ :=
√
1 −ρ; Write W := {W
t
}
t∈T
for a standard Brownian motion, which
is independent of W
1
and X. Then, we can write the dynamics of {S
t
}
t∈T
and σ := {σ
t
}
t∈T
as:
dS
t
= µ
t
S
t
dt +σ
t
S
t
dW
1
t
,
dσ
2
t
= β(α
2
t
−σ
2
t
)dt +ργσ
t
dW
1
t
+ ˜ ργσ
t
dW
t
. (2.8)
Let Y
t
denote the logarithmic return ln(S
t
/S
0
) over the interval [0, t]. Then,
Y
t
=
_
t
0
_
µ
u
−
1
2
σ
2
u
_
du +
_
t
0
σ
u
dW
1
u
. (2.9)
In our model, there are three sources of randomness: X, W
1
and W
2
. Let F
X
:=
{F
X
t
}
t∈T
, F
W
1
:= {F
W
1
t
}
t∈T
and F
W
2
:= {F
W
2
t
}
t∈T
be the Paugmentation of the
natural ﬁltrations generated by {X
t
}
t∈T
, {W
1
t
}
t∈T
and {W
2
t
}
t∈T
, respectively. Let
11
F
S
:= {F
S
t
}
t∈T
denote the Paugmentation of the natural ﬁltration generated by
{S
t
}
t∈T
.
Our model is a regime switching version of Heston’s stochastic volatility model
and is in general incomplete. There are, therefore, inﬁnitely many equivalent martin
gale pricing measures. In the sequel, we shall adopt the regimeswitching Esscher
transform developed in Elliott et al. (2005) to determine an equivalent martin
gale pricing measure for the volatility and variance swaps. The regimeswitching
Esscher transform provides market practitioners with a convenient and
ﬂexible way to determine an equivalent martingale measure in the incom
plete market. The choice of the equivalent martingale measure can be
justiﬁed by the regimeswitching minimal entrophy equivalent martingale
(MEMM) measure (See Elliott et al. (2005)). One drawback of using the
Esscher transform is that the pricing rule by the Esscher transform is not
linear, which is considered by ﬁnancial economists as a desirable property
for a pricing rule. There are other possible approaches for determining an
equivalent martingale measure in an incomplete market, for instance, the
minimum variance hedging in Duﬃe and Richardson (1991) and Schweizer
(1992). In the minimumvariance hedging, the instrinsic value process
corresponding to a given contingent claim is used as the optimal tracking
12
process. The instrinsic value process is deﬁned by the minimal martingale
measure introduced in F¨ollmer and Schweizer (1991) and Schweizer (1991)
and corresponds to the riskneutral approach for valuing the claim. The
minimumvariance hedging can provide a pertinent solution to address
the pricing and hedging of a contingent claim while the Esscher transform
mainly deals with the valuation of the claim. The hedging strategies are
optimal in sense of minimizing the expected quadratic costs. The Esscher
transform can provide a convenient and intuitive way to price a claim.
Let G
t
denote the right continuous completion of the σalgebra F
X
t
∨F
W
1
t
∨F
W
2
t
,
for each t ∈ T . Let Θ
t
:= Θ(t, X
t
, σ
t
) denote a regime switching Esscher process,
which is written as follows:
Θ
t
= Θ(t, X
t
, σ
t
) =< Θ(t, σ
t
), X
t
> , (2.10)
where Θ(t, σ
t
) := (Θ(t, σ
t
, e
1
), Θ(t, σ
t
, e
2
), . . . , Θ(t, σ
t
, e
N
)) and Θ(t, σ
t
, e
i
) is F
W
2
t
measurable, for each i = 1, 2, . . . , N. So, Θ(t, X
t
, σ
t
) is an Ndimensional F
X
t
∨F
W
2
t

measurable random vector.
Following Elliott et al. (2005), for such a process Θ we deﬁne the regime switching
Esscher transform Q
Θ
∼ P on G
t
with respect to a family of parameters {Θ
u
}
u∈[0,t]
13
by:
dQ
Θ
dP
¸
¸
¸
G
t
=
exp
_
_
t
0
Θ
u
dY
u
_
E
P
_
exp
_
_
t
0
Θ
u
dY
u
_¸
¸
¸F
X
t
∨ F
W
2
t
_ , t ∈ T . (2.11)
Note that
_
t
0
Θ
u
dY
u
F
X
t
∨F
W
2
t
∼ N(
_
t
0
Θ
u
(µ
u
−
1
2
σ
2
u
)du,
_
t
0
Θ
2
u
σ
2
u
du), that is a normal
distribution with mean
_
t
0
Θ
u
(µ
u
−
1
2
σ
2
u
)du and variance
_
t
0
Θ
2
u
σ
2
u
du, under P. Then
given F
X
t
∨ F
W
2
t
, the RadonNikodym derivative of the regime switching Esscher
transform is given by:
dQ
Θ
dP
¸
¸
¸
G
t
= exp
_
_
t
0
Θ
u
σ
u
dW
1
u
−
1
2
_
t
0
Θ
2
u
σ
2
u
du
_
. (2.12)
The central tenet of the fundamental theorem of asset pricing, which is also called
the fundamental theorem of ﬁnance in Ross (2005), established the equivalence be
tween the absence of arbitrage opportunities and the existence of a positive linear
pricing operator (or positive state space prices). It was ﬁrst established by Ross
(1973) in a ﬁnite state space setting. Cox and Ross (1976) provided the ﬁrst state
ment of riskneutral pricing. Ross (1978) and Harrsion and Kreps (1979) then ex
tended the fundamental theorem in a general probability space and characterized the
riskneutral pricing as the expectation of the discounted asset’s payoﬀ with respect
to an equivalent martingale measure. Harrison and Kreps (1979), and Harrison and
Pliska (1981, 1983) established the equivalence between the absence of arbitrage and
the existence of an equivalent martingale measure under which all discounted price
14
processes are martingale. The fundamental theorem was then extended by several
authors, including Dybvig and Ross (1987), Back and Pliska (1991), Schachermayer
(1992) and Delbaen and Schachermayer (1994). In our setting, let
˜
Θ := {
˜
Θ
t
}
t∈T
denote a process for the riskneutral regime switching Esscher parameters. Due to
the presence of the uncertainty generated the processes X and W
2
, the martingale
condition is characterised by considering an enlarged ﬁltration and requiring:
S
0
= E
Q
˜
Θ
_
exp
_
−
_
t
0
r
s
ds
_
S
t
¸
¸
¸F
X
t
∨ F
W
2
t
_
, for any t ∈ T . (2.13)
One can interpret this condition as one when information about the Markov chain
and the stochastic volatility process are known to the market’s agent in advance.
Give these arguments which are similar to those in Elliott et al. (2005), it can be
shown that the martingale condition (2.11) implies that
˜
Θ
t
:=
˜
Θ(t, X
t
, σ
t
) should be
given by
˜
Θ
t
=
r(t, X
t
) −µ(t, X
t
)
σ
2
t
= −
λ(t, X
t
, σ
t
)
σ
t
, t ∈ T , (2.14)
where λ
t
:= λ(t, X
t
, σ
t
) ∈ G
t
is the market price of risk at time t.
Then
˜
Θ
t
=<
˜
Θ(t, σ
t
), X
t
>, where
˜
Θ(t, σ
t
) = (
r
1
−µ
1
σ
t
,
r
2
−µ
2
σ
t
, . . . ,
r
N
−µ
N
σ
t
). This is an
Ndimensional F
W
2
t
measurable random vector.
15
Using (2.12), the RadonNikodym derivative of Q
˜
Θ
with respect to P is given by:
dQ
˜
Θ
dP
¸
¸
¸
G
t
= exp
_
_
t
0
_
r
u
−µ
u
σ
u
_
dW
1
u
−
1
2
_
t
0
_
r
u
−µ
u
σ
u
_
2
du
_
. (2.15)
By Girsanov’s theorem,
˜
W
1
t
= W
1
t
+
_
t
0
(
r
s
−µ
s
σ
s
)ds is a standard Brownian motion with
respect to {G
t
}
t∈T
under Q
˜
Θ
. Since W and X are independent of W
1
, W is a
standard Brownian motion under Q
˜
Θ
and X remains unchanged under the
change of the probability measure from P to Q
˜
Θ
. Let ˜ α
2
t
:= α
2
t
−ργ(r
t
−µ
t
).
Then, the dynamics of S and σ under Q
˜
Θ
are
dS
t
= r
t
S
t
dt +σ
t
S
t
d
˜
W
1
t
,
dσ
2
t
= β(˜ α
2
t
−σ
2
t
)dt +ργσ
t
d
˜
W
1
t
+ ˜ ργσ
t
dW
t
. (2.16)
Let
˜
W
2
t
:= ρ
˜
W
1
t
+ ˜ ρW
t
. Then, the dynamics of σ can be written as:
dσ
2
t
= β(˜ α
2
t
−σ
2
t
)dt +γσ
t
d
˜
W
2
t
. (2.17)
When there is no regime switching, (i.e. the Markov chain X is degenerate), the
riskneutral dynamics under Q
˜
Θ
reduce to the riskneutral dynamics in Heston (1993).
§3. The Probabilistic Approach
In this section, we assume the dynamics of the longterm volatility level {α
t
}
t∈T
switch
over time according to one of the regimes determined by the state of an observable
Markov chain. We may interpret the states of the observable Markov chain as those of
16
some observable economic indicator. First, we shall consider the valuation of variance
swaps, which is more simple than the valuation of volatility swaps. Then, we shall
discuss the hedging of variance swaps and volatility swaps.
§3.1. Valuation
A variance swap is a forward contract on annualized variance, which is the square
of the realized annual volatility. Let σ
2
R
(S) denote the realized annual stock variance
over the life of the contract. Then,
σ
2
R
(S) :=
1
T
_
T
0
σ
2
u
du . (3.1)
In practice, variance swaps are written on the realized variance evaluated
based on daily closing prices with the integral in (3.1) replaced by a dis
crete sum. Hence, variance swaps with payoﬀs depending on the realized
variance deﬁned in (3.1) are only approximations to those of the actual
contracts. See Javaheri et al. (2002) for discussions on this point.
Let K
v
and N denote the delivery price for variance and the notional amount
of the swap in dollars per annualized volatility point squared. Then, the payoﬀ of
the variance swap at expiration time T is given by N(σ
2
R
(S) − K
v
). Intuitively, the
buyer of the variance swap will receive N dollars for each point by which the realized
annual variance σ
2
R
(S) has exceeded the variance delivery price K
v
. We can adopt the
17
riskneutral regime switching Esscher transform Q
˜
Θ
for the valuation of the variance
swap. In fact, the value of the variance swap can be evaluated as the expectation of
its discounted payoﬀ with respect to the measure Q
˜
Θ
, which is exactly the same as
the value of a forward contract on future realized variance with strike price K
v
.
As in Elliott, Sick and Stein (2003), we initially consider the evaluation of the
conditional value, or price, of a derivative given the information about the sample
path of the Markov chain process from time 0 to time T, say F
X
T
. In particular, given
F
X
T
, the conditional price of the variance swap P(X) is given by:
P(X) = E
Q
˜
Θ
[e
−
R
T
0
r
u
du
N(σ
2
R
(S) −K
v
)F
X
T
]
= e
−
R
T
0
r
u
du
NE
Q
˜
Θ
(σ
2
R
(S)F
X
T
) −e
−
R
T
0
r
u
du
NK
v
. (3.2)
Hence, the valuation of the variance swap given F
X
T
can be reduced to the problem
of calculating the mean value of the underlying variance E
Q
˜
Θ
(σ
2
R
(S)F
X
T
).
Note that under Q
˜
Θ
, the volatility dynamics can be written as:
σ
2
t
= σ
2
0
+
_
t
0
β(˜ α
2
s
−σ
2
s
)ds +
_
t
0
γσ
s
d
˜
W
2
s
. (3.3)
Given F
X
T
, ˜ α
2
t
is a known function of time t. Hence,
E
Q
˜
Θ
(σ
2
t
F
X
T
) = σ
2
0
+
_
t
0
β[ ˜ α
2
s
−E
Q
˜
Θ
(σ
2
s
F
X
T
)]ds . (3.4)
18
This implies that
dE
Q
˜
Θ
(σ
2
t
F
X
T
)
dt
= β[ ˜ α
2
t
−E
Q
˜
Θ
(σ
2
t
F
X
T
)] . (3.5)
Solving (3.5) gives:
E
Q
˜
Θ
(σ
2
t
F
X
T
) = σ
2
0
e
−βt
+β
_
t
0
˜ α
2
s
e
−β(t−s)
ds . (3.6)
By Itˆo’s lemma,
σ
4
t
= σ
4
0
+
_
t
0
[2βσ
2
s
(˜ α
2
s
−σ
2
s
) +γ
2
σ
2
s
]ds +
_
t
0
2γσ
3
s
d
˜
W
2
s
. (3.7)
Hence,
dE
Q
˜
Θ
(σ
4
t
F
X
T
)
dt
= (2β˜ α
2
t
+γ
2
)E
Q
˜
Θ
(σ
2
t
F
X
T
) −2βE
Q
˜
Θ
(σ
4
t
F
X
T
) . (3.8)
Solving (3.8) gives:
E
Q
˜
Θ
(σ
4
t
F
X
T
) = σ
4
0
e
−2βt
+
_
t
0
e
−2β(t−s)
(2β˜ α
2
s
+γ
2
)
_
σ
2
0
e
−βs
+β
_
s
0
˜ α
2
u
e
−β(s−u)
du
_
ds . (3.9)
Hence,
V ar
Q
˜
Θ
(σ
2
t
F
X
T
)
= E
Q
˜
Θ
(σ
4
t
F
X
T
) −E
2
Q
˜
Θ
(σ
2
t
F
X
T
)
=
σ
2
0
γ
2
β
(e
−βt
−e
−2βt
) +
_
t
0
_
e
−2β(t−s)
_
2β
2
˜ α
2
s
+βγ
2
_
_
s
0
˜ α
2
u
e
−β(s−u)
du
_
ds
−β
2
_
_
t
0
˜ α
2
s
e
−β(t−s)
ds
_
2
. (3.10)
19
The results for E
Q
˜
Θ
(σ
2
t
F
X
T
) and V ar
Q
˜
Θ
(σ
2
t
F
X
T
) are consistent with those in Shreve
(2004).
Write V := σ
2
R
(S) to simplify the notation. Then, E
Q
˜
Θ
(V F
X
T
) can be calculated
as follows:
E
Q
˜
Θ
(V F
X
T
) =
1
T
_
T
0
_
σ
2
0
e
−βt
+β
_
t
0
˜ α
2
s
e
−β(t−s)
ds
_
dt
=
σ
2
0
βT
(1 −e
−βT
) +
β
T
_
T
0
_
_
t
0
˜ α
2
s
e
−β(t−s)
ds
_
dt , (3.11)
We evaluate V ar
Q
˜
Θ
(V F
X
T
) as follows:
V ar
Q
˜
Θ
(V F
X
T
)
= Cov
Q
˜
Θ
(V, V F
X
T
)
= 1/T
2
_
T
0
_
T
0
Cov
Q
˜
Θ
(σ
2
t
, σ
2
s
F
X
T
)dtds (3.12)
We ﬁrst derive an expression for Cov
Q
˜
Θ
(σ
2
t
, σ
2
s
F
X
T
). Without loss of generality, we
suppose that t > s. Then, we deﬁne η
t,s
as follows:
η
t,s
:= σ
2
t
−E
Q
˜
Θ
(σ
2
t
F
W
2
s
∨ F
X
T
)
= σ
2
t
−σ
2
s
e
−β(t−s)
−β
_
t
s
˜ α
2
u
e
−β(t−u)
du . (3.13)
Then, it is immediate that
E
Q
˜
Θ
(η
t,s
F
W
2
s
∨ F
X
T
) = 0 , (3.14)
20
E
Q
˜
Θ
(η
t,s
σ
2
s
F
X
T
) = 0 , (3.15)
so
Cov
Q
˜
Θ
(η
t,s
, σ
2
s
F
X
T
) = 0 . (3.16)
Hence,
Cov
Q
˜
Θ
(σ
2
t
, σ
2
s
F
X
T
)
= Cov
Q
˜
Θ
(η
t,s
+σ
2
s
e
−β(t−s)
+β
_
t
s
˜ α
2
u
e
−β(t−u)
, σ
2
s
F
X
T
)
= e
−β(t−s)
V ar
Q
˜
Θ
(σ
2
s
F
X
T
)
= e
−β(t−s)
_
σ
2
0
γ
2
β
(e
−βs
−e
−2βs
) +
_
s
0
_
e
−2β(s−u)
_
2β
2
˜ α
2
u
+βγ
2
_
_
u
0
˜ α
2
z
e
−β(u−z)
dz
_
du −β
2
_
_
s
0
˜ α
2
u
e
−β(s−u)
du
_
2
_
. (3.17)
Therefore,
V ar
Q
˜
Θ
(V F
X
T
)
= 1/T
2
_
T
0
_
T
0
e
−β(t−s)
_
σ
2
0
γ
2
β
(e
−βs
−e
−2βs
) +
_
s
0
_
e
−2β(s−u)
_
2β
2
˜ α
2
u
+βγ
2
_
_
u
0
˜ α
2
z
e
−β(u−z)
dz
_
du −β
2
_
_
s
0
˜ α
2
u
e
−β(s−u)
du
_
2
_
dtds
=
σ
2
0
γ
2
β
2
T
2
_
(1 −e
−βT
)T +
1
4β
(1 −e
−2βT
)
2
_
+ 1/T
2
_
T
0
_
T
0
e
−β(t−s)
_
_
s
0
_
e
−2β(s−u)
_
2β
2
˜ α
2
u
+βγ
2
_
_
u
0
α
2
z
e
−β(u−z)
dz
_
du −β
2
_
_
s
0
˜ α
2
u
e
−β(s−u)
du
_
2
_
dtds . (3.18)
For each i = 1, 2, . . . , N, let ˜ α
i
:= α
i
−ργ(r
i
−µ
i
); Write ˜ α := (˜ α
1
, ˜ α
2
, . . . , ˜ α
N
). Given
21
F
X
T
, the conditional price of the variance swap is given by:
P(X) = e
−
R
T
0
r
u
du
N
_
σ
2
0
βT
(1 −e
−βT
) +
β
T
_
T
0
_
_
t
0
< ˜ α
2
, X
t
> e
−β(t−s)
ds
_
dt −K
v
_
. (3.19)
Consider now the valuation of the volatility swap given F
X
T
. A stock volatility
swap is a forward contract on the annualized volatility. Let K
s
denote the annualized
volatility delivery price and N is the notational amount of the swap in dollar per
annualized volatility point. Then, the payoﬀ function of the volatility swap is given
by N(σ
R
(S) − K
s
), where σ
R
(S) :=
_
1
T
_
T
0
σ
2
u
du. In other words, the payoﬀ of the
volatility swap is equal to the payoﬀ of the variance swap when σ
2
R
(S) is replaced by
σ
R
(S) and K
v
is replaced by K
s
. Given F
X
T
, the conditional price of the volatility
swap is given by:
P
s
(X) = E
Q
˜
Θ
[e
−
R
T
0
r
u
du
N(σ
R
(S) −K
s
)F
X
T
]
= e
−
R
T
0
r
u
du
NE
Q
˜
Θ
(σ
R
(S)F
X
T
) −e
−
R
T
0
r
u
du
NK
s
= e
−
R
T
0
r
u
du
NE
Q
˜
Θ
(
√
V F
X
T
) −e
−
R
T
0
r
u
du
NK
s
. (3.20)
For the valuation of the volatility swap, we need to evaluate E
Q
˜
Θ
(
√
V F
X
T
). We
adopt the approximation for E
Q
˜
Θ
(
√
V F
X
T
) introduced by Brockhaus and Long (2000),
based on the secondorder Taylor expansion for the function
√
V . This approxima
tion method has also been adopted in Javaheri et al. (2002) and Swishchuk (2004).
22
It gives
E
Q
˜
Θ
(
√
V F
X
T
) ≈
_
E
Q
˜
Θ
(V F
X
T
) −
V ar
Q
˜
Θ
(V F
X
T
)
8[E
Q
˜
Θ
(V F
X
T
)]
3/2
, (3.21)
where the term
V ar
Q
˜
Θ
(V F
X
T
)
8[E
Q
˜
Θ
(V F
X
T
)]
3/2
is the convexity adjustment.
Hence, given F
X
T
, the conditional price of the volatility swap can be approximated
as:
P
s
(X) ≈ e
−
R
T
0
r
u
du
N
__
E
Q
˜
Θ
(V F
X
T
) −
V ar
Q
˜
Θ
(V F
X
T
)
8[E
Q
˜
Θ
(V F
X
T
)]
3/2
−K
s
_
. (3.22)
23
§3.2. Hedging
Hedging variance swaps and volatility swaps is a challenging but practically im
portant task. Javaheri et al. (2002) contended that hedging these products is diﬃcult
in practice. Hence, they considered the pricing of a variance swap and a volatility
swap by the expectations of the discounted payoﬀs under the realworld probability
measure. This is an example of actuarialbased valuation method. In this section, we
shall discuss the hedging of variance swaps and volatility swaps. Diﬀerent methods on
hedging variance swaps, have been proposed in the literature. These methods include
the simple delta hedging, the deltagamma hedging, hedging using option portfolios,
hedging using a log contract and the vega hedging, etc. For a comprehensive overview
of various hedging strategies, see Demeterﬁ et al. (1999), Howison et al. (2004) and
Windcliﬀ et al. (2003). Simple delta hedging does not work well since it is an implicit
linear approximation, which cannot incorporate the eﬀects of large price movements
and the realized variance or volatility will increase substantially when the underlying
share prices move either up or down dramatically. This is the case even one considers
a Geometric Brownian Motion for the price dynamics of the underlying share. Simple
delta hedging is even more diﬃcult to apply when one considers a stochastic volatility
model and a regimeswitching stochastic volatility model, which is even more com
plicated. Hedging using option portfolios and hedging using a log contract works
24
well when one considers an asset price model with nonstochastic volatility. The vega
hedging provides market practitioners with a convenient way to hedge variance swaps
and volatility swaps under a stochastic volatility model, in particular, the Heston SV
model (see Howison et al. (2004) and Carr (2005)). Howison et al. (2004) considered
the use of Vega to hedge volatililty derivatives and derived a general formula for the
Vega of a volatility derivative. Here, following Howison et al. (2004), we adopt the
Vega hedging for a variance swap and a volatility swap since it can provide a conve
nient way to hedge these products under the regimeswitching Heston’s SV model. In
the case of the volatility swap, we shall derive an approximate formula for the Vega
of the contract based on the approximate price of the contract.
First, we consider the hedging of the variance swap. Let I
t
:=
_
t
0
σ
2
u
du. Write
F
˜
W
2
t
for the Paugmentation of the σalgebra generated by the values of
˜
W
2
up to
and including time t. Note that given F
X
T
, F
˜
W
2
t
is equivalent to F
W
2
t
. Then, using
the results in Section 3.1, the price of the variance swap P(t, X) at time t is given
by:
P(t, X) =
1
T
e
−
R
T
t
r
u
du
N
_
I
t
+E
Q
˜
Θ
_
_
T
t
σ
2
u
du
¸
¸
¸F
X
T
∨ F
W
2
t
_
−TK
v
_
=
1
T
e
−
R
T
t
r
u
du
N
_
I
t
+
σ
2
t
β
(e
−βt
−e
−βT
) +β
_
T
t
_
_
s
t
< ˜ α
2
, X
u
>
e
−β(s−u)
du
_
ds −TK
v
_
. (3.23)
25
Let ˆ σ
2
t
:= γσ
2
t
, which represents the instantaneous volatility of the variance process
at time t. Then, the Vega of the variance swap is given by:
∂P(t, X)
∂ˆ σ
=
2Nˆ σ
t
Tβγ
2
(e
−βt
−e
−βT
)
=
2Nσ
t
Tβ
(e
−βt
−e
−βT
) , (3.24)
which can be evaluated given the current value of the volatility level σ
t
.
We shall consider the hedging of the volatility swap. First, we deﬁne R
1
(t, σ
2
t
),
R
2
(t, σ
2
t
) and R
3
(t, σ
2
t
) as follows:
R
1
(t, σ
2
t
) = I
t
+
σ
2
t
β
(e
−βt
−e
−βT
) +β
_
T
t
_
_
s
t
< ˜ α
2
, X
u
> e
−β(s−u)
du
_
ds ,
R
2
(t, σ
2
t
) =
σ
2
t
γ
2
T
2
β
3
[−e
−β(T−t)
2
−
1
β
e
−2β(T−t)
] +
1
T
2
_
T
t
_
T
t
_
e
−β(h−s)
_
s
0
_
e
−2β(s−u)
_
2β
2
˜ α
2
u
+βγ
2
_
_
u
0
˜ α
2
z
e
−β(u−z)
dz
_
du −β
2
_
_
s
0
˜ α
2
u
e
−β(s−u)
du
_
2
_
dhds ,
and
R
3
(t, σ
2
t
) =
σ
2
t
γ
2
β
2
T
2
_
(1 −e
−βt
)t +
1
4β
(1 −e
−2βt
)
2
_
+ 1/T
2
_
t
0
_
t
0
_
e
−β(u−s)
_
s
0
_
e
−2β(s−u)
_
2β
2
˜ α
2
u
+βγ
2
_
_
u
0
α
2
z
e
−β(u−z)
dz
_
du −β
2
_
_
s
0
˜ α
2
u
e
−β(s−u)
du
_
2
_
duds .
26
From the results in Section 3.1, the price of the volatility swap P
s
(t, X) at time t can
be approximated as:
P
s
(t, X) ≈ e
−
R
T
t
r
u
du
N
__
E
Q
˜
Θ
(V F
X
T
∨ F
W
2
t
) −
V ar
Q
˜
Θ
(V F
X
T
∨ F
W
2
t
)
8[E
Q
˜
Θ
(V F
X
T
∨ F
W
2
t
)]
3/2
−K
s
_
= e
−
R
T
t
r
u
du
N
_
_
R
1
(t, σ
2
t
) −
R
3
(t, σ
2
t
) +R
2
(t, σ
2
t
)
8R
1
(t, σ
2
t
)
3/2
−K
s
_
. (3.25)
Write R
i
for R
i
(t, σ
2
t
) (i = 1, 2, 3). Then, the Vega of the volatility swap is approxi
mated as:
∂P
s
(t, X)
∂ˆ σ
= e
−
R
T
t
r
u
du
N
_
σ
t
β
R
1
−1/2
(e
−βt
−e
−βT
) +
3σ
t
8β
(R
3
+R
2
)R
1
−5/2
(e
−βt
+e
−βT
) +
σ
t
γ
2
R
1
4T
2
β
3
_
e
−β(T−t)
2
+
1
β
e
−2β(T−t)
__
, (3.26)
which can be evaluated given the current value of the volatility level σ
t
.
§4. The P.D.E. Approach
In this section, we adopt a partial diﬀerential equation (P.D.E.) approach for
evaluating the expectations of the discounted values of V and V
2
, which are useful
for computing the prices of the variance swaps and volatility swaps. The P.D.E.
approach has been adopted by Javaheri, et al. (2002) for the valuation and hedging of
volatility swaps within the framework of a GARCH(1, 1) stochastic volatility model.
Here, we provide a regime switching modiﬁcation of the problem and derive regime
switching P.D.E.s and the corresponding systems of coupled P.D.E.s satisﬁed by the
27
expectations of the discounted values of V and V
2
. We adopt a regime switching
version of the FeymanKac formula to obtain the regime switching P.D.Es. The
derivation of the regime switching version of the FeymanKac formula follows from
the martingale approach and Itˆo’s diﬀerentiation rule in Buﬃngton and Elliott (2002).
First, let V
t
:= σ
2
R,t
(S) :=
1
T
_
t
0
σ
2
u
du. Then, given F
W
2
t
∨ F
X
T
, the price of the
variance swap is given by:
P(X, t) = e
−
R
T
t
r
u
du
NE
Q
˜
Θ
(σ
2
R,T
(S)F
W
2
t
∨ F
X
T
) −e
−
R
T
t
r
u
du
NK
v
, (4.1)
and the price of the volatility swap is given by:
P
s
(X, t)
= e
−
R
T
t
r
u
du
NE
Q
˜
Θ
(σ
R,T
(S)F
W
2
t
∨ F
X
T
) −e
−
R
T
t
r
u
du
NK
s
. (4.2)
Now, suppose σ
2
t
= Σ, X
t
= X and V
t
= V are given at time t. Then, the price of
the variance swap is given by:
P(X, Σ, V, t) = E
Q
˜
Θ
(P(X, t)σ
2
t
= Σ, V
t
= V, X
t
= X) , (4.3)
and the price of the volatility swap is given by:
P
s
(X, Σ, V, t) = E
Q
˜
Θ
(P
s
(X, t)σ
2
t
= Σ, V
t
= V, X
t
= X) . (4.4)
Buﬃngton and Elliott (2002a, b) adopted a similar method to determine the price of
a standard European call option.
28
By the double expectation formula,
P(X, Σ, V, t)
= NE
Q
˜
Θ
_
e
−
R
T
t
r
u
du
σ
2
R,T
(S) −e
−
R
T
t
r
u
du
K
v
¸
¸
¸σ
t
= Σ, V
t
= V, X
t
= X
_
, (4.5)
and
P
s
(X, Σ, V, t)
= NE
Q
˜
Θ
_
e
−
R
T
t
r
u
du
σ
R,T
(S) −e
−
R
T
t
r
u
du
K
s
¸
¸
¸σ
2
t
= Σ, V
t
= V, X
t
= X
_
. (4.6)
Hence, for the evaluation of P(X, Σ, V, t) and P
s
(X, Σ, V, t), we need to compute
1. E
Q
˜
Θ
_
e
−
R
T
t
r
u
du
¸
¸
¸σ
2
t
= Σ, V
t
= V, X
t
= X
_
2. E
Q
˜
Θ
_
e
−
R
T
t
r
u
du
σ
2
R,T
(S)
¸
¸
¸σ
2
t
= Σ, V
t
= V, X
t
= X
_
3. E
Q
˜
Θ
_
e
−
R
T
t
r
u
du
σ
R,T
(S)
¸
¸
¸σ
2
t
= Σ, V
t
= V, X
t
= X
_
The ﬁrst expectation is equal to the value of a zerocoupon bond at time t given that
σ
2
t
= Σ, V
t
= V, X
t
= X, which pays one unit of account at maturity time T. It is
given by:
E
Q
˜
Θ
_
e
−
R
T
t
r
u
du
¸
¸
¸σ
2
t
= Σ, V
t
= V, X
t
= X
_
= E
Q
˜
Θ
_
e
−
R
T
t
r
u
du
¸
¸
¸X
t
= X
_
:= B(t, T, X) . (4.7)
29
Let B := diagr − Π
∗
, where diagr is the matrix with the vector r on its diagonal.
Write φ
t
(r) := exp[−B(T − t)]I. Then, by Elliott and Kopp (2004), B(t, T, X) is
given by:
B(t, T, X) =< φ
t
(r), X >=< exp[−B(T −t)], X > I . (4.8)
The third expectation can be approximated by the using the formula in Section
2. More speciﬁcally,
E
Q
˜
Θ
_
e
−
R
T
t
r
u
du
σ
R,T
(S)
¸
¸
¸σ
2
t
= Σ, V
t
= V, X
t
= X
_
≈
_
E
Q
˜
Θ
_
e
−2
R
T
t
r
u
du
σ
2
R,T
(S)
¸
¸
¸σ
2
t
= Σ, V
t
= V, X
t
= X
_
−
V ar
Q
˜
Θ
_
e
−2
R
T
t
r
u
du
σ
2
R,T
(S)
¸
¸
¸σ
2
t
= Σ, V
t
= V, X
t
= X
_
8
_
E
Q
˜
Θ
_
e
−2
R
T
t
r
u
du
σ
2
R,T
(S)
¸
¸
¸σ
2
t
= Σ, V
t
= V, X
t
= X
__ . (4.9)
Hence, in order to provide an approximation to the third expectation, we need to
evaluate
1. E
Q
˜
Θ
(e
−4
R
T
t
r
u
du
σ
4
R,T
(S)σ
2
t
= Σ, V
t
= V, X
t
= X)
2. E
Q
˜
Θ
(e
−2
R
T
t
r
u
du
σ
2
R,T
(S)σ
2
t
= Σ, V
t
= V, X
t
= X).
Suppose H
t
:= σ{W
2
u
, X
u
u ∈ [0, t]}. Since V
t
is a path integral of σ
2
t
and σ
2
t
is a Markov process given knowledge of X, (V
t
, σ
2
t
) is a twodimensional Markov
30
process given the knowledge of X. Since X is also a Markov process, (X
t
, σ
2
t
, V
t
) is a
threedimensional Markov process with respect to the information set H
t
. Hence,
M
1
(X, Σ, V, t) := E
Q
˜
Θ
(e
−
R
T
t
r
u
du
σ
2
R,T
(S)σ
2
t
= Σ, V
t
= V, X
t
= X)
= E
Q
˜
Θ
(e
−
R
T
t
r
u
du
σ
2
R,T
(S)H
t
) , (4.10)
M
2
(X, Σ, V, t) := E
Q
˜
Θ
(e
−2
R
T
t
r
u
du
σ
2
R,T
(S)σ
2
t
= Σ, V
t
= V, X
t
= X)
= E
Q
˜
Θ
(e
−2
R
T
t
r
u
du
σ
2
R,T
(S)H
t
) , (4.11)
and
M
3
(X, Σ, V, t) := E
Q
˜
Θ
(e
−4
R
T
t
r
u
du
σ
4
R,T
(S)σ
2
t
= Σ, V
t
= V, X
t
= X)
= E
Q
˜
Θ
(e
−4
R
T
t
r
u
du
σ
4
R,T
(S)H
t
) , (4.12)
Now, write
˜
M
1
(X, Σ, V, t) := e
−
R
t
0
r
u
du
M
1
(X, Σ, V, t)
= E
Q
˜
Θ
(e
−
R
T
0
r
u
du
σ
2
R,T
(S)H
t
) , (4.13)
˜
M
2
(X, Σ, V, t) := e
−2
R
t
0
r
u
du
M
2
(X, Σ, V, t)
= E
Q
˜
Θ
(e
−2
R
T
0
r
u
du
σ
2
R,T
(S)H
t
) , (4.14)
and
˜
M
3
(X, Σ, V, t) := e
−4
R
t
0
r
u
du
M
3
(X, Σ, V, t)
31
= E
Q
˜
Θ
(e
−4
R
T
0
r
u
du
σ
4
R,T
(S)H
t
) . (4.15)
Then, it can be shown that
˜
M
1
,
˜
M
2
and
˜
M
3
are H
t
martingales under Q
˜
Θ
.
In the sequel, we shall derive the P.D.E. for
˜
M
1
,
˜
M
2
and
˜
M
3
. For each i = 1, 2, 3,
let
˜
M
i
(Σ, V, t) denote the Ndimensional vector (
˜
M
i
(e
1
, Σ, V, t), . . . ,
˜
M
i
(e
N
, Σ, V, t)).
Then,
˜
M
i
(X, Σ, V, t) =<
˜
M
i
(Σ, V, t), X
t
> . (4.16)
Then, by applying Itˆo’s diﬀerentiation rule to
˜
M
i
(X, Σ, V, t),
˜
M
i
(X, Σ, V, t) =
˜
M
i
(X, Σ, V, 0) +
_
t
0
_
∂
˜
M
i
∂u
+β(˜ α
2
u
−σ
2
u
)
∂
˜
M
i
∂Σ
+
1
2
γ
2
σ
2
u
∂
2
˜
M
i
∂Σ
2
+σ
2
u
∂
˜
M
i
∂V
_
du +
_
t
0
∂
˜
M
i
∂Σ
γσ
u
d
˜
W
2
u
+
_
t
0
<
˜
M
i
, dX
u
> , (4.17)
and
dX
t
= Π(t)X
t
dt +dM
t
. (4.18)
Due to the fact that
˜
M
1
,
˜
M
2
and
˜
M
3
are H
t
martingale under Q
˜
Θ
, all terms with
bounded variation in the above Itˆo’s intergral representation for
˜
M
i
(i = 1, 2, 3) must
be identical to zero. Hence, for i = 1, 2, 3,
˜
M
i
satisﬁes the following P.D.E.:
∂
˜
M
i
∂t
+β(˜ α
2
t
−σ
2
t
)
∂
˜
M
i
∂Σ
+
1
2
γ
2
σ
2
t
∂
2
˜
M
i
∂Σ
2
+σ
2
t
∂
˜
M
i
∂V
+ <
˜
M
i
, ΠX >= 0 . (4.19)
32
Now, let M
i
(Σ, V, t) denote the Ndimensional vector (M
i
(e
1
, Σ, V, t), . . . , M
i
(e
N
, Σ, V, t)),
where i = 1, 2, 3. Then,
M
i
(X, Σ, V, t) =< M
i
(Σ, V, t), X
t
> . (4.20)
M
1
satisﬁes the following P.D.E.:
_
exp
_
−
_
t
0
r
u
du
___
−r
t
M
1
+
∂M
1
∂t
+β(˜ α
2
t
−σ
2
t
)
∂M
1
∂Σ
+
1
2
γ
2
σ
2
t
∂
2
M
1
∂Σ
2
+σ
2
t
∂M
1
∂V
+ < M
1
, ΠX >
_
= 0 , (4.21)
with terminal condition M
1
(X, Σ, V, T) = V
T
.
M
2
satisﬁes the following P.D.E.:
_
exp
_
−2
_
t
0
r
u
du
___
−2r
t
M
2
+
∂M
2
∂t
+β(˜ α
2
t
−σ
2
t
)
∂M
2
∂Σ
+
1
2
γ
2
σ
2
t
∂
2
M
2
∂Σ
2
+σ
2
t
∂M
2
∂V
+ < M
2
, ΠX >
_
= 0 , (4.22)
with terminal condition M
2
(X, Σ, V, T) = V
T
.
M
3
satisﬁes the following P.D.E.:
_
exp
_
−4
_
t
0
r
u
du
___
−4r
t
M
3
+
∂M
3
∂t
+β(˜ α
2
t
−σ
2
t
)
∂M
3
∂Σ
+
1
2
γ
2
σ
2
t
∂
2
M
3
∂Σ
2
+σ
2
t
∂M
3
∂V
+ < M
3
, ΠX >
_
= 0 , (4.23)
with terminal condition M
3
(X, Σ, V, T) = V
2
T
.
33
Note that with X = e
j
(j = 1, 2, . . . , N),
r
t
= < r, X
t
>= r
j
,
˜ α
t
= < ˜ α, X
t
>= ˜ α
j
. (4.24)
Let M
ij
:= M
i
(e
j
, Σ, V, T), where i = 1, 2, 3 and j = 1, 2, . . . , N. Then, M
i
=
(M
i1
, M
i2
, . . . M
iN
). Hence, M
1
satisﬁes the following system of N coupled P.D.E.s:
−r
j
M
1j
+
∂M
1j
∂t
+β(˜ α
2
j
−σ
2
t
)
∂M
1j
∂Σ
+
1
2
γ
2
σ
2
t
∂
2
M
1j
∂Σ
2
+σ
2
t
∂M
1j
∂V
+ < M
1
, Πe
j
>= 0 , (4.25)
with terminal condition M
1
(e
j
, Σ, V, T) = V
T
.
M
2
satisﬁes the following system of N coupled P.D.E.s:
−2r
j
M
2j
+
∂M
2j
∂t
+β(˜ α
2
j
−σ
2
t
)
∂M
2j
∂Σ
+
1
2
γ
2
σ
2
t
∂
2
M
2j
∂Σ
2
+σ
2
t
∂M
2j
∂V
+ < M
2
, Πe
j
>= 0 , (4.26)
with terminal condition M
2
(e
j
, Σ, V, T) = V
T
.
M
3
satisﬁes the following system of N coupled P.D.E.s:
−4r
j
M
3j
+
∂M
3j
∂t
+β(˜ α
2
j
−σ
2
t
)
∂M
3j
∂Σ
+
1
2
γ
2
σ
2
t
∂
2
M
3j
∂Σ
2
+σ
2
t
∂M
3j
∂V
+ < M
3
, Πe
j
>= 0 , (4.27)
with terminal condition M
3
(e
j
, Σ, V, T) = V
2
T
.
34
Once M
1
, M
2
and M
3
are solved from the above systems of N coupled P.D.E.s,
we can use them to approximate the prices of the variance swap and the volatility
swap.
§5. Monte Carlo Experiment
In this section, we shall perform a Monte Carlo Experiment for the prices of the
variance swaps and the volatility swaps implied by the regimeswitching Heston’s
stochastic volatility model. We shall document economic consequences for the prices
of the variance swaps and the volatility swaps of a regimeswitching in Heston’s
stochastic volatility model by comparing the prices with those obtained from Heston’s
SV model without regimeswitching. We shall compute the prices of the variance
swaps and the volatility swaps with various delivery prices under both the regime
switching Heston’s SV model and Heston’s SV model without switching regimes by
Monte Carlo simulation. For illustration, we suppose that the number of regimes N =
2 throughout this section. The ﬁrst and second regimes, namely X
t
= 1 and X
t
= 2,
can be interpreted as the “Good” and “Bad” economic states, respectively. We also
assume that Heston’s SV model without switching regimes coincides with the ﬁrst
regime of the regimeswitching Heston’s SV model. In this case, we can investigate
economic consequences for the prices of the variance swaps and the volatility swaps
35
when we allow the possibility that the dynamics of Heston’s SV model switches over
time to the one corresponding to the “Bad” economic states. We generate 10,000
simulation runs for computing each price. All computations were done by C++
codes with GSL functions.
We shall assume some specimen values for the parameters of regimeswitching
Heston’s SV model and the one without switching regimes. When the economy is
good (bad), the interest rate is high (low). Let r
1
and r
2
denote the annual interest
rates for the “Good” state and the “Bad” state, respectively. Then, we suppose that
r
1
= 5% and r
2
= 2%. The appreciation rate of the underlying risky asset is high
(low) when the economy is good (bad). In each case, the appreciation rate should
be higher than the corresponding interest rate. Hence, we suppose that the annual
appreciation rate µ
1
= 7% for the “Good” state and the annual appreciation rate
µ
2
= 5% for the “Bad” state. When the economy is good (bad), the underlying risky
asset is less (more) volatile. Hence, we suppose that α
1
= 0.12 and α
2
= 0.24. The
speed of mean reversion β = 0.2 and the volatility of volatility parameter γ = 0.08.
We also assume that the correlation coeﬃcient ρ is negative and is equal to −0.5.
The transition probabilities of the Markov chain are π
11
= 0.5, π
12
= 0.5, π
21
= 0.5,
π
22
= 0.5. The notational amount of the variance swap or the volatility swap is £1
million. We suppose that the current economic state X
0
= 1 and that the current
36
volatility level V
0
= 0.12. The delivery prices of the variance swap and the volatility
swap range from 80% to 125% of the current levels of the variance and the standard
deviation of the underlying risky asset, respectively. The timetoexpiry of both the
variance swap and the volatility swap is 1 year. Since the regimeswitching Heston
stochastic volatility model is a continuoustime model, we need to discretize it when
we compute the prices of the variance swaps and volatility swaps by Monte Carlo
simulation. We suppose that the number of steps for the discretization is 20. Table
1 displays the prices of the variance swaps for various delivery prices implied by
the regimeswitching Heston stochastic volatility model and its nonregimeswitching
counterpart.
Table 1: Prices of Variance Swaps with and without Switching Regimes
Delivery Prices Prices with Switching Regimes Prices without Switching Regimes
in % in £ million in £ million
80 1.14556 0.847634
85 1.1431 0.845326
90 1.14165 0.843224
95 1.13823 0.840911
100 1.13607 0.838839
105 1.13167 0.836526
110 1.12934 0.834374
115 1.12657 0.832061
120 1.12196 0.829986
125 1.12082 0.82767
37
Table 2 displays the prices of the volatility swaps for various delivery prices im
plied by the regimeswitching Heston stochastic volatility model and its nonregime
switching counterpart.
Table 2: Prices of Volatility Swaps with and without Switching Regimes
Delivery Prices Prices with Switching Regimes Prices without Switching Regimes
in % in £ million in £ million
80 0.632453 0.467974
85 0.624144 0.461601
90 0.616364 0.455246
95 0.607529 0.448786
100 0.599301 0.442436
105 0.589936 0.436079
110 0.581685 0.429701
115 0.573195 0.423343
120 0.563777 0.416992
125 0.55599 0.410641
From Table 1 and Table 2, we see that the prices of the variance swaps and the
volatility swaps implied by the regimeswitching Heston stochastic volatility model
are signiﬁcantly higher than the corresponding prices of the variance swaps and the
volatility swaps, respectively, implied by the standard Heston stochastic volatility
without switching regimes. This reveals that a higher risk premium is required to
compensate for the risk from the structural change of the volatility dynamics to the
one with higher longterm volatility level due to the possible transitions of the states
38
of the economy to the “Bad” state. This illustrates the economic signiﬁcance of
incorporating the switching regimes in the volatility dynamics for pricing variance
swaps and volatility swaps.
§6. Further Research
For further investigation, it is interesting to explore and develop some criteria to de
termine the number of states of the Markov chain in our framework which will incor
porate important features of the volatility dynamics for diﬀerent types of underlying
ﬁnancial instruments, such as commodities, currencies and ﬁxed income securities. It
is also interesting to explore the applications of our model to price various volatility
derivative products, such as options on volatilities and VIX futures, which are a listed
contract on the Chicago Board Options Exchange. It is also of practical interest to
investigate the calibration and estimation techniques of our model to volatility index
options. Empirical studies comparing the performance of models on volatility swaps
are interesting topics to be investigated further.
Acknowledgment
We would like to thank the referee for many valuable comments and suggestions.
39
References
1. Back, K. and Pliska, S. R. (1991) On the fundamental theorem of asset pricing
with an inﬁnite state space, Journal of Mathematical Economics, 20, pp. 118.
2. Bollerslev, T. (1986) Generalized autoregressive conditional heteroskedasticity,
Journal of Econometrics, 31, pp. 307327.
3. Brenner, M., and Galai, D. (1989) New ﬁnancial instruments for hedging changes
in volatility, Financial Analyst Journal, July/August, pp. 6165.
4. Brenner, M., and Galai, D. (1993) Hedging volatility in foreign currencies,
Journal of Derivatives, 1, pp. 5359.
5. Brenner, M., Ou, E., and Zhang, J. (2001) Hedging volatility risk, Working
paper, Stern School of Business, New York University, United States.
6. Brockhaus, O., and Long, D. (2000) Volatility swaps made simple, Risk, 2(1),
pp. 9295.
7. Buﬃngton, J. and Elliott, R. J. (2002a) Regime switching and European op
tions, In Stochastic Theory and Control, Proceedings of a Workshop, Lawrence,
K.S., October, Springer Verlag, pp. 7381.
40
8. Buﬃngton, J. and Elliott, R. J. (2002b) American options with regime switch
ing, International Journal of Theoretical and Applied Finance, 5, pp. 497514.
9. Carr, P. (2005) Delta hedging with stochastic volatility, Working paper, Quan
titative Financial Research, Bloomberg LP, New York.
10. Carr, P., and Madan, D. (1998) Towards a theory of volatility trading, In the
book: Volatility, Risk book publications.
11. Carr, P., Geman H., Madan, D., and Yor, M. (2005) Pricing options on realized
variance, Finance and Stochastics, 9(4), pp. 453475.
12. Cox, J., Ingersoll, J. and Ross, S. (1985) A theory of the term structure of
interest rates, Econometrica, 53, pp. 385407.
13. Cox, J. and Ross, S. (1976) The valuation of options for alternative stochastic
processes, Journal of Financial Economics, 3, pp. 145166.
14. Delbaen, F. and Schachermayer, W. (1994) A general version of fundamental
theorem of asset pricing, Mathematische Annalen, 300, pp. 463520.
15. Demeterﬁ, K., Derman, E., Kamal, M., and Zou, J. (1999) A guide to volatility
and variance swaps, The Journal of Derivatives, 6(4), pp. 932.
41
16. Duﬃe, D. and Richardson, H. R. (1991) Meanvariance hedging in continuous
time, The Annals of Applied Probability, 1, pp. 115.
17. Dybvig, Philip H., and Ross, S. (1987) Arbitrage, In Eatwell J., Milgate M.,
and Newman P. eds. the New Palsgrave: A dictionary of Economics , 1, pp.
100106.
18. Elliott, R. J., Aggoun, L. and Moore, J. B. (1994) Hidden Markov Models:
Estimation and Control. SpringerVerlag: BerlinHeidelbergNew York.
19. Elliott, R. J., Sick, G. A. and Stein, M. (2003) Modelling electricity price risk,
Working paper, Haskayne School of Business, University of Calgary.
20. Elliott, R. J., Chan, L. and Siu, T. K. (2005) Option pricing and Esscher
transform under regime switching, Annals of Finance, 1(4), pp. 423432.
21. Elliott, R. J. and Kopp, E. P. (2004) Mathematics of Financial Markets. Springer
Verlag: BerlinHeidelbergNew York.
22. Elliott, R. J., and Swishchuk, A. (2004) Pricing options and variance swaps
in Brownian and fractional Brownian market, Working paper, University of
Calgary.
42
23. Engle, R. (1982) Autoregressive conditional heteroskedasticity with estimates
of the variance of U.K. inﬂation, Econometrica, 50, pp. 9871008.
24. F¨ollmer, H. and Schweizer, M. (1991) Hedging of contingent claims under
incomplete information, In: M.H.A. Davis and R. J. Elliott (eds.), Applied
Stochastic Analysis, Stochastics Monographs, Vol.5, Gordon and Breach, Lon
don/New York, pp. 389414.
25. Gr¨ unbuchler, A., and Longstaﬀ, F. (1996) Valuing futures and options on
volatility, Journal of Banking and Finance, 20, pp. 9851001.
26. Harrison, J. M. and Kreps, D. M. (1979) Martingales and arbitrage in multi
period securities markets, Journal of Economic Theory, 20, pp. 381408.
27. Harrison, J. M. and Pliska, S. R. (1981) Martingales and stochastic integrals in
the theory of continuous trading, Stochastic Processes and Their Applications,
11, pp. 215280.
28. Harrison, J. M. and Pliska, S. R. (1983) A stochastic calculus model of contin
uous trading: complete markets, Stochastic Processes and Their Applications,
15, pp. 313316.
29. Heston, S. L. (1993) A closedform solution for options with stochastic with
43
applications to bond and currency options, Review of Financial Studies, 6(2),
pp. 237343.
30. Heston, S. L., and Nandi, S. (2000) Derivatives on volatility: some simple
solution based on observables, Working paper, Federal Reserve Bank of Atlanta.
31. Howison, S., Rafailidis A., and Rasmussen, H. (2004) On the pricings and
hedging of volatility derivatives, Applied Mathematical Finance, 11(4), pp. 317
346.
32. Javaheri, A., Wilmott, P., and Haug, E. G. (2002) “GARCH and volatility
swaps, Wilmott Magazine, January, pp. 117.
33. Matytsin, A. (2000) Modeling volatility and volatility derivatives, Working pa
per, Columbia University.
34. Pan, J. (2002) The jumprisk premia implicit in options: evidence from an
integrated timeseries study, Journal of Financial Economics, 63, pp. 350.
35. Ross, S. A. (1973) Return, risk and arbitrage, Wharton Discussion Paper pub
lished in Risk and Return in Finance, edited by I. Friend and J. Bicksler, pp.
189217. Cambridge: Ballinger, 1976.
44
36. Ross, S. A. (1978) “A simple approach to the valuation of risky streams, Journal
of Business, 3, pp. 453476.
37. Ross, S. A. (2005) “Neoclassical Finance”, Princeton and Oxford: Princeton
University Press.
38. Rubinstein, M. (1976) The valuation of uncertain income streams and the pric
ing of options, Bell Journal of Economics and Management Science, 7, pp.
407425.
39. Schachermayer, W. (1992) “A Hilbert space proof of the fundamental theorem
of asset pricing in ﬁnite discrete time, Insurance: Mathematics and Economics,
11, pp. 249257.
40. Schweizer, M. (1991) Option hedging for semimartingales, Stochastic Processes
and Their Application, 37, pp. 339363.
41. Schweizer, M. (1992) Meanvariance hedging for general claims, The Annals of
Applied Probability, 2, pp. 171179.
42. Shreve, S. E. (2004) Stochastic Calculus for Finance, New York: Springer.
43. Swishchuk, A. (2004) Modeling of variance and volatility swaps for ﬁnancial
markets with stochastic volatilities, Wilmott magazine, 2, pp. 6472.
45
44. Swishchuk, A. (2005) Modeling and pricing of variance swaps for stochastic
volatilities with delay, Wilmott magazine, Forthcoming.
45. Taylor, S. J. (1986) Modelling Financial Time Series. New York: John Wiley
& Sons.
46. Windcliﬀ, H., Forsyth, P. A. and Vetzal, K. R. (2003) Pricing methods and
hedging strategies for volatility derivatives, Working paper, University of Wa
terloo.
46