American Options under Stochastic Volatility *

Arun Chockalingam
School of Industrial Engineering, Purdue University, West Lafayette, IN 47907.
arunachalam@purdue.edu

Kumar Muthuraman
McCombs School of Business, University of Texas, Austin, TX 78712.
kumar.muthuraman@mccombs.utexas.edu

The problem of pricing an American option written on an underlying asset with constant price volatility has
been studied extensively in literature. Real-world data, however, demonstrates that volatility is not constant
and stochastic volatility models are used to account for dynamic volatility changes. Option pricing methods
that have been developed in literature for pricing under stochastic volatility focus mostly on European
options. We consider the problem of pricing American options under stochastic volatility which has had
relatively much less attention from literature. First, we develop a transformation procedure to compute the
optimal-exercise policy and option price and provide theoretical guarantees for convergence. Second, using
this computational tool, we explore a variety of questions that seek insights into the dependence of option
prices, exercise policies and implied volatilities on the market price of volatility risk, and correlation between
the asset and stochastic volatility. The speed and accuracy of the procedure is compared against existing
methods as well.

Key words : American option, stochastic volatility, free boundary.

1. Introduction
Options are contracts that give the holder the right to sell (put) or buy (call) an underlying asset
at a pre-determined strike price. A European option allows the holder to exercise the option only
on a pre-determined expiration date, while an American option allows the holder to exercise the
option at any point in time until the expiration date. Option pricing has always played a prominent
role in financial theory as well as real derivative markets. In a celebrated paper, Black and Scholes
(1973) derive a closed-form solution for the price of a European option by characterizing the price
as the expected payoff under a risk-neutral measure. Their model assumes that the underlying
asset price follows a geometric Brownian motion with constant volatility.
Even under the constant volatility assumption, that is, the classical Black-Scholes setting, closed-
form solutions do not exist for American options. Due to the possibility of early exercise, the

* We would like to thank the Associate Editor, Mark Broadie, Haolin Feng, Jose Figueroa-Lopez, Stanley Pliska,
Bruce Schmeiser, Stathis Tompaidis, and two anonymous referees for their comments, suggestions and feedback.

1
 Chockalingam and Muthuraman: Stochastic Volatility
2

American option price depends on the optimal-exercise policy, which can be represented by an
exercise boundary (also known as the free boundary) on the price-time space. The exercise boundary
partitions the price-time space into hold and exercise regions.
Most of option pricing literature consider the constant volatility model. Rubinstein (1994), how-
ever, provides empirical evidence using implied volatilities obtained from index options on the S&P
500 that suggests that the constant volatility assumption does not hold. Using data for the OEX
contract, Broadie et al. (2000) find that dividends alone are not accountable for all aspects of
option pricing and exercise decisions, and suggest that stochastic volatility needs to be included
as well. Furthermore, Scott (1987) and the references therein provide ample evidence of volatility
changing over time. This can also be readily seen from the implied volatilities calculated from
market prices. Implied volatilities are the volatilities that, when used in the Black-Scholes for-
mula, provide European option prices consistent with option prices observed in the market. When
implied volatilities are plotted against strike prices, the plots exhibit a ‘smile’ effect, which refers
to the resulting U-shaped curve, as opposed to a straight line that one would expect if asset prices
had constant volatility. Implied volatilities for in- and out-of-the-money options are observed to
be higher than at-the-money options. Assuming constant volatility therefore leads to considerable
mispricing. Hence, models that allow the volatility of the underlying asset price to be stochastic
are needed to better capture market behavior.
Several models have been proposed to better model the evolution of volatility. One approach
has been to use ARCH models and their variants (Bollerslev et al. (1992)). Using another diffusion
process to model volatility, however, has become a more popular choice. Hull and White (1987),
Scott (1987), Stein and Stein (1991) and Heston (1993) each propose different diffusion processes
to represent the dynamics of asset price volatility.

1.1. Relevant literature

The pricing of American options under constant volatility and European options under stochastic
volatility have both received considerable attention in literature. Numerical methods available to
compute the price and/or exercise boundary for American options under constant volatil-
ity include simulation-based techniques (Broadie and Glasserman (1997), Longstaff and Schwartz
(2001)), binomial lattices (Cox et al. (1979)), partial differential equation (PDE) solution methods
(Brennan and Schwartz (1977), Muthuraman (2008)), front-fixing methods (Wu and Kwok (1997),
Nielsen et al. (2002)), front-tracking methods (Pantazopoulos et al. (1998))) and integral methods
(Kim (1990), Jacka (1991), Carr et al. (1992)). Myneni (1992), Karatzas and Shreve (1998) and
Chockalingam and Muthuraman: Stochastic Volatility
3

Broadie and Detemple (2004) provide an overview of American option pricing under the Black-
Scholes setting.
The pricing of European options under stochastic volatility has been looked at for various
choices of diffusion dynamics to represent the volatility process. Hull and White (1987) use a
lognormal process while Scott (1987) and Stein and Stein (1991) use a mean-reverting Ornstein-
Uhlenbeck (OU) process. Closed-form solutions have been obtained in Heston (1993) when volatility
is modeled as a square-root process. Ball and Roma (1994) also use the same square-root process.
Relatively little attention has been paid to the problem of pricing an American option under
stochastic volatility. Literature on American options under stochastic volatility can be classified
into PDE based and non-PDE based approaches. The PDE methods solve the free-boundary prob-
lem arising from the use of classical dynamic programming arguments and provide the entire price
function and the optimal-exercise policy explicitly. Non-PDE based approaches compute the price
for any given time, asset price and underlying volatility by computing the conditional expectation
under a suitable martingale measure.
A popular approach to solve the related free-boundary PDE problem is to reformulate it as a lin-
ear complementarity problem (LCP). The Projected Successive Over Relaxation (PSOR) method
proposed by Cryer (1971) is widely used to solve these LCPs. Clarke and Parrott (1999) use a
stretching transformation and an adaptive-upwind finite difference approximation to discretize the
LCP, resulting in the need to solve many discrete complementarity problems. A multigrid iter-
ation method is developed to solve these problems. Oosterlee (2003) states that the projected
line Gauss-Seidel smoother used in Clarke and Parrott (1999) is too involved and studies alter-
nate smoothers that can be used in conjunction with the multigrid iteration method, finding that
an alternating line Gauss-Seidel smoother is a better choice, bringing about better convergence.
Oosterlee (2003) also improves upon Clarke and Parrott (1999) using a recombination of iterants.
Ikonen and Toivanen (2004) solve each of the discrete complementarity problems obtained after
time and space discretization using operator splitting methods. The method divides each time step
into two fractional steps, integrates the PDE with an auxiliary variable over the time step, then
updates the solution to satisfy the linear complementarity conditions due to the early-exercise con-
straint. Componentwise splitting methods are utilized in Ikonen and Toivanen (2007a) to solve the
discrete complementarity problems. The componentwise splitting method splits each discretized
LCP into three LCPs, and the use of Strang symmetrization further decomposes the three LCPs
into five LCPs. Each LCP consists of tridiagonal matrices and is solved using the Brennan and
Schwartz (1977) algorithm. Another technique known as the penalty method replaces the unknown
 Chockalingam and Muthuraman: Stochastic Volatility
4

free-boundary with a non-linear penalty term and solves the resulting non-linear fixed boundary
problem. Zvan et al. (1998) use a standard Galerkin finite element method to discretize the aris-
ing PDE, and use penalties to force the discrete problems to satisfy the early-exercise constraint.
The authors highlight the equivalence of the penalty method and the linear complementarity for-
mulation. Ikonen and Toivanen (2007b) compare the five methods for pricing American options
under stochastic volatility, and find that while the error in prices computed by any of these meth-
ods is comparable, the componentwise splitting method is considerably faster. They do, however,
acknowledge that the PSOR method is the easiest to implement while the componentwise splitting
method is the most difficult to implement.
As for non-PDE approaches, a nonparametric approach is utilized in Broadie et al. (2000). Several
simulation methods such as the least squares Monte Carlo approach in Longstaff and Schwartz
(2001), the primal-dual simulation algorithm of Andersen and Broadie (2004) and the stochastic
mesh method of Broadie and Glasserman (2004) that have all been developed for American options
under constant volatility can also be adapted to compute American option prices under stochastic
volatility.

1.2. Contribution and outline

In this paper, we show that the problem of pricing American options under stochastic volatility can,
in fact, be transformed into a sequence of European-type option pricing problems. More specifically,
the free-boundary problem arising in the pricing of American options can be transformed into
a sequence of fixed-boundary problems. By a European-type option, we mean that the exercise
policy is set a priori in the option contract. It is important to note that solving fixed-boundary
problems is not very difficult in three dimensions if one leverages on existing PDE solver packages
or libraries that are available to solve fixed-boundary problems. Software packages such as Comsol
even allow the reuse of contours from one solution as boundaries for another problem - making
the implementation of our method very easy. We also prove the convergence of the sequence of
fixed-boundary problems. Our results show the methodology proposed here is faster than other
available methods, while being as accurate as the PSOR method. Moreover, all it involves is solving
linear PDEs in domains with fixed boundaries.
The approach we present differs from the other papers on American option pricing under stochas-
tic volatility in several ways. First, the model we consider is far more general. The numerical
methods listed above are tailored for the Heston (1993) stochastic volatility model. The methodol-
ogy presented here works for the Hull and White (1987), Scott (1987) and Stein and Stein (1991)
models as well. An assumption that is common to most existing papers is that the market price
Chockalingam and Muthuraman: Stochastic Volatility
5

of volatility risk is zero. As will be discussed in Subsection 2.2, the price of an American option
is, in general, not unique and depends on the market price of volatility risk. Most of these papers
can be extended to handle non-zero volatility risk premiums. However, since they do not consider
non-zero market price of volatility risk, they could not study the effects of volatility risk premium
on the exercise policy and the price. The approach described in this paper readily accommodates
non-zero volatility risk premiums and we study its effects. Several of the existing papers also use
a wrong boundary condition which we correct.
In terms of underlying methodology, the solution technique proposed in this paper is very dif-
ferent from any that is available for pricing American options under a stochastic volatility setting.
Methods, like the penalty method, covert the linear PDE with a free boundary, to a non-linear
PDE with fixed boundary. On the other hand, the transformation we propose retains the linear
PDE but solves it within fixed boundaries in each iteration. Since the complexity of dealing with
a single non-linear PDE is much harder than dealing with a sequence of linear PDEs, it is not
surprising that runtimes are much better in our case. Methods like the componentwise splitting
method introduce new approximations other than the ones due to discretization, while ours does
not. The accuracy depends only on the method used to solve the fixed-boundary problem.
The method proposed in this paper extends the class of methods that are being called moving-
boundary methods. Such methods were developed initially for singular control problems (Muthu-
raman and Kumar (2006) and Kumar and Muthuraman (2004)) and were demonstrated to work
numerically for higher dimensional cases. For American option pricing in the classical setting,
Muthuraman (2008) extends this method and provides theoretical guarantees. This paper extends
the method to optimal stopping problems in a higher setting. Most importantly, due to the chal-
lenges in dealing with higher dimensions theoretically, there had been absolutely no theoretical
guarantees for the method in any problem with more than one space dimensionality. This paper
provides the first set of such guarantees.
The layout of the paper is as follows. Section 2 presents the model formulation. The transforma-
tion procedure is presented in Section 3. A computational illustration of the procedure is provided
in Section 4, together with insights into how stochastic volatility affects option pricing. Speed and
accuracy comparisons are also presented in section 4. We conclude in Section 5. All proofs are
collected in Appendix A and a detailed discussion of the finite difference scheme used to solve the
fixed-boundary problem is presented in Appendix B.
 Chockalingam and Muthuraman: Stochastic Volatility
6

2. Model Formulation
We start this section with a discussion on the use of a second stochastic process to model the
evolution of volatility. We then formulate the free-boundary problem which the American option
price and the optimal-exercise policy jointly solve.

2.1. Stochastic Volatility Models

In stochastic volatility models, asset price (Xt ) evolution is represented by

dXt = µXt dt + σt Xt dWt , (1)

where µ is the constant mean rate of return and Wt is a standard Brownian motion (a Wiener
process). The instantaneous volatility at time t is represented by σt . The volatility σt = f (Yt ),
where Yt is another stochastic process and f (·) is a non-stochastic function. The evolution of Yt is
represented by

dYt = µY (Yt )dt + σ Y (Yt )dẐt , (2)

where µY and σ Y are non-stochastic functions. In Equation (2), Ẑt is another standard Brownian
motion correlated with Wt . We assume a constant correlation ρ ∈ [−1, 1], i.e., dWt dẐt = ρdt. Hence,
Ẑt can be written as a linear combination of Wt and an independent Wiener process Zt such that
√
Ẑt = ρWt + 1 − ρ2 Zt . Here, we have two sources of randomness, namely Wt and Ẑt , but only one
tradable asset, leading to market incompleteness. We refer readers to Björk (2004) and Fouque
et al. (2000) for further discussions on stochastic volatility models and market incompleteness.

µY (y) σ Y (y) f (y) ρ

√ √
Heston (1993) κ(m′ − y) υ y y ρ 6= 0
Stein and Stein (1991) α(m − y) β |y | ρ=0
Scott (1987) α(m − y) β ey ρ=0
√
Hull and White (1987) c1 y c2 y y ρ=0
Table 1 Stochastic Volatility Models

Much of the literature on stochastic volatility focuses on a few specific models (Table 1). Financial
data shows that ρ < 0 (Fouque et al. (2000)). Of the four models, the Heston model is the most
popular and the only one that allows for a non-zero correlation. Also, Dragulescu and Yakovenko
(2002) provides evidence that the Heston model is in agreement with real-world data, leading to
wider adoption of the model by researchers and practitioners. Thus, for expositional ease, we will
restrict much of our attention to the Heston model and provide additional comments relevant to
the other models when necessary. The transformation procedure developed in this paper works for
all models listed in Table 1.
Chockalingam and Muthuraman: Stochastic Volatility
7

2.2. American Option Pricing

Consider an American put option written on an underlying asset with price Xt at time t given by
Equation (1). The instantaneous volatility σt is represented by the Heston model. The put option
with strike price K and maturity time T written on this asset pays max {K − Xt , 0} ≡ (K − Xt )+
at any time t ∈ [0, T ]. The price of an option, p(τ, x, y), is represented as a function of the time to
expiry, τ ≡ T − t, the underlying asset price x and the value, y, of the process Yt . As noted earlier,
the market considered is an incomplete one, as a result of volatility not being a tradable asset.
Assuming that the market selects a unique equivalent martingale measure P ∗(Λ) , prices of deriva-
tives should be priced with respect to this measure to disallow arbitrage giving

p(τ, x, y) = sup E∗(Λ) {e−r(τ −τ̂) (K − XT −τ̂ )+ |Xτ = x, Yτ = y }, (3)

τ̂ ∈(0,τ )

where r > 0 is the risk-free rate of interest. As mentioned earlier, when the volatility of the
asset price is constant, the optimal-exercise policy can be represented by a continuous, non-
increasing exercise boundary (e.g. Karatzas and Shreve (1998)). Analogous to the constant volatil-
ity case, the optimal-exercise policy under stochastic volatility can be represented by a surface
(Fouque et al. (2000)). The exercise surface is a continuous, non-increasing surface x = b(τ, y),
b : R+ × R → R+ . This boundary partitions the state space and dictates the optimal-exercise
policy. The region where it is optimal to hold, known as the continuation region, is defined as
C = (τ, x, y) ∈ R2+ × R | x > b(τ, y) , and the region where it is optimal to exercise, known as the


stopping region, is defined as S = (τ, x, y) ∈ R2+ × R | x ≤ b(τ, y) .



In the continuation region, the price of the American option satisfies the PDE,

1 2 2 ∂2p 2 ∂2p 1 2 2 ∂2p ∂p

σt x 2
+ ρvσt x + v σt 2 + rx
2 ∂x ∂x∂y 2 ∂y ∂x
′ ∂p ∂p
+ (κ(m − y) − vσt Λ) − rp − =0 (4)
∂y ∂τ

for all (τ, x, y) ∈ C .

In Equation (4), Λ denotes the market price of volatility risk. Since volatility cannot be traded, we
have an incomplete market under the stochastic volatility setting. Though we lose the uniqueness
of derivative pricing, for any given Λ, there is a unique price pΛ . An infinitesimal increase in the
volatility risk υσt increases the infinitesimal rate of return on the option by Λ. A discussion on the
market price of volatility risk and its effect on pricing is discussed in Section 4.3. For a detailed
discussion on the interpretation of Λ and its relation to equivalent martingale measures we refer
the reader to Bakshi and Kapadia (2003) and Henderson (2005).
 Chockalingam and Muthuraman: Stochastic Volatility
8

In the exercise region, the price is the payoff

p(τ, x, y) = (K − x)+ (5)

for all (τ, x, y) ∈ S . For notational convenience, we define the differential operator L such that the
LHS of Equation (4) is denoted by Lp.
To solve Equation (4) in the region C , the following boundary conditions are needed;

p(0, x, y) = (K − x)+ , (6)

p(τ, b(τ, y), y) = (K − b(τ, y))+, (7)
∂p
lim = 0, (8)
x→∞ ∂x
∂p
lim = 0, and (9)
y→∞ ∂y
∂p ∂p ∂p
Ap ≡ rx + κm′ − rp − =0 at y = 0. (10)
∂x ∂y ∂τ

Figure 1 illustrates the state space and boundary conditions.

Figure 1 The state space and boundary conditions

Equations (6) and (7) prescribe the value at expiry and at exercise. As the underlying asset price
increases, the probability that the asset price falls below K, before or at expiry, decreases. This
increasingly guarantees a zero pay-off. Equation (8) reflects this behavior of the option. Equation (9)
captures the argument that when volatility is extremely large, a marginal increase in volatility has
little effect on the price.
The boundary condition at y = 0 is directly arrived at by taking y = 0 in Equation (4). Several
papers (including Clarke and Parrott (1999), Oosterlee (2003), Ikonen and Toivanen (2004) and
Chockalingam and Muthuraman: Stochastic Volatility
9

Ikonen and Toivanen (2007a)) that develop innovative numerical methods to price American options
under the Heston model argue that when volatility is zero, since there is no randomness, the pay-
off as well as the price are deterministic, leading to p(τ, x, 0) = (K − x)+ . This is, however, not
the case, since Yt is a mean-reverting process in the Heston model, meaning that if Yt = 0, then
almost surely dYt is positive and hence Yt+ is greater than zero, making the asset price process
non-deterministic. Therefore, the use of p(τ, x, 0) = (K − x)+ , though easier to handle, is incorrect
and needs to be replaced by the PDE in Equation (10).
It is important to note that Equations (4) - (10) can be solved for any sufficiently smooth surface,
i.e., exercise policy, b. The optimal-exercise policy is, however, the only policy for which p is smooth
across the boundary b. This condition, commonly known as the smooth pasting condition, implies
∂p
that p as well as ∂x
are continuous across b and gives rise to (Fouque et al. (2000))

∂p
lim = −1 (11)
x↓b ∂x

for optimality. A solution p(τ, x, y), b(τ, y) to Equations (4) - (11) also has to satisfy the condition

p(τ, x, y) ≥ (K − x)+ ∀(τ, x, y) ∈ S ∪ C . (12)

in order to be the true price function and optimal-exercise policy.

2.3. Other stochastic volatility models

The structure of the pricing problem described by Equations (4) - (11) remains the same for the
other models discussed in Section 2.1. However, the specific PDE and the boundary condition given
in Equation (10) are model specific. This is due to the differences in the evolution of Yt , it’s domain
and the function f (·) that determines the volatility.
Table 2 summarizes the PDE that replaces Equation (4) and the boundary condition that replaces
Equation (10) for each model.

Model PDE
Boundary condition, Boundary
1 2 2 ∂2p ∂2p ∂2 p
Stein and Stein (1991) σ x ∂x2 + ρβσt x ∂x∂y
2 t
+ 12 β 2 ∂y ∂p ∂p ∂p
2 + rx ∂x + (α(m − y) − βΛ) ∂y − rp − ∂τ = 0
∂p
∂y
= 0, y → −∞
1 2 2 ∂2p ∂2p ∂2 p ∂p ∂p ∂p
Scott (1987) σ x ∂x2 + ρβσt x ∂x∂y
2 t
+ 12 β 2 ∂y 2 + rx ∂x + (α(m − y) − βΛ) ∂y − rp − ∂τ = 0
2
1 2∂ p ∂p ∂p ∂p
2
β ∂y 2 + rx ∂x + (αm − βΛ) ∂y − rp − ∂τ = 0, y → −∞
2 2 2
1 2 2∂ p ∂ p 1 2 4∂ p ∂p
´ ∂p ∂p
3
σ x ∂x2 + ρc2 σt x ∂x∂y + 2 c2 σt ∂y 2 + rx ∂x + c1 σt2 − c2 σt2 Λ ∂y
`
Hull and White (1987) 2 t
− rp − ∂τ = 0
+
p = (K − x) , y = 0
Table 2 Boundary conditions
 Chockalingam and Muthuraman: Stochastic Volatility
10

3. Transforming the free-boundary problem

In this section, we first present the transformation procedure that solves the free-boundary problem
described in Section 2.2. We then demonstrate the mechanics of the transformation procedure with
a numerical illustration.
The American option pricing problem is defined by Equations (4) - (11) and is satisfied by the
price function p(τ, x, y) and the optimal-exercise policy b(τ, y). Now consider an arbitrary policy
defined by a surface b0 (τ, y). When such an arbitrary exercise policy is used by an option holder,
the value to the holder will be referred to as the associated value p0 (τ, x, y). Clearly “the price”
is the associated value of using the optimal-exercise boundary, that is, p0 (τ, x, y) = p(τ, x, y) when
b0 = b and p0 (τ, x, y) ≤ p(τ, x, y) for any b0 .
For an arbitrary exercise policy b0 , one can find the associated price p0 by solving
Equations (4) - (10), which can be done using standard PDE techniques, such as finite difference
and finite element schemes. A finite difference scheme for computing the associated values is detailed
in Appendix B. For a given b0 , the uniqueness of the associated value function is established by
the following proposition.

Proposition 3.1. If pn satisfies Equation (4) with the boundary conditions given by Equa-
tions (6) - (10) for a given bn (τ, y), then pn is unique.

All proofs are collected in Appendix A.

Our aim is to construct a transformation procedure that will converge and provide the price
of the option on convergence. Starting from a guess b0 , if we can construct a sequence of policies
b0 , b1 , . . . that are monotonic increasing i.e., bn (τ, y) < bn+1 (τ, y) for all (τ, y) and is bounded above,
then convergence is inevitable. Exercise-policy improvement would also imply that pn (τ, x, y) <
pn+1 (τ, x, y) for all (τ, x, y). We construct and demonstrate such a transformation below.
We require that the initial guess policy b0 be such that

∂p0 

< −1. (13)
∂x (τ,b0 (τ,y)+,y)

Proposition 3.2 guarantees us that such a b0 can be obtained when the chosen b0 lies below the
optimal policy b. This is necessary since we are interested in constructing a transformation proce-
dure that monotonically converges to the optimal b (Theorem 3.1). It has to be noted that though
b is unknown, it is not difficult to choose a b0 < b. A wrong choice would imply that Condition (13)
is violated and a restart with a lower choice b0 would eventually work.

0
Proposition 3.2. If b0 (τ, y) < b(τ, y), then ∂p
∂x 
< −1.
0 (τ,b (τ,y)+,y)
Chockalingam and Muthuraman: Stochastic Volatility
11

Say we begin with an arbitrary guess b0 and solve Equations (4) - (10) and obtain p0 . Now,
on the boundary b0 , p0 = (K − x)+ and immediately above b0 , the derivative w.r.t. x is less than
−1. These together imply that p0 < (K − x)+ in a region immediately above b0 . In this region,
the policy b0 can easily be bettered since an immediate exercise will yield improvement. Hence,
we could choose any b1 in the region where p0 < (K − x)+ and be guaranteed of improving the
exercise-policy. Further guiding our choice of a new exercise-policy is our preference to choose a b1
that is also guaranteed to lie below b (that is b1 (τ, x, y) < b(τ, x, y)) so that we can iterate. Such a
choice is indeed possible and is given by
(  )
n
∂p
bn+1 (τ, y) = sup x |  < −1 (14)
(bn (τ,y),∞) ∂x (τ,x,y)

for all τ and y. The theoretical guarantee that such a bn+1 exists, implies exercise-policy improve-
ment and is still below b is provided by Theorem 3.1.

∂pn 
Theorem 3.1. If pn ∈ C {1,2} solves Equations (4),(6) - (10), with ∂x 
< −1, then
(τ,bn (τ,y)+,y)
bn+1 as defined by Equation (14) exists. Furthermore, the price function p n+1
obtained using bn+1

n+1 
is such that pn+1 > pn and ∂p∂x  < −1.
n+1(τ,b (τ,y)+,y)
∂p
The convergence of the improvement procedure described by Equation (14) to ∂x
= −1 at the
exercise boundary implies the satisfaction of Equation (11), yielding optimality.
The mechanism described above can viewed as decomposing the American option pricing problem
into a sequence of European-style option pricing problems. By European-style, we mean here
that the exercise policy is known a priori, unlike for American options. This reveals that pricing
American options is only as difficult as pricing a sequence of European-style options.

3.1. An illustration
We now illustrate the mechanics of the transformation procedure using a computational example.
Consider the Heston model with parameters used in Clarke and Parrott (1999): strike price K = 10,
time to expiry T = 3 months, risk-free rate of interest r = 10% per annum, mean rate of reversion
κ = 5, long-term mean variance level m′ = 0.16, volatility of the Yt process υ = 0.9, correlation
ρ = 0.1 and market price of volatility risk Λ = 0. We set the initial exercise-policy guess b0 (τ, y) = 1
for all τ and y. The transformation procedure for this set of parameters converges to the optimal
exercise policy shown in Figure 2 after 4 iterations. Figure 2 also shows b0 .
Figure 3 plots the exercise policies in each iteration for the cut taken at y = 1. As can be recalled,
for each exercise policy, the associated value function is first computed. Then, the improved exercise
policy is set as the contour of the associated value function derivative w.r.t. x equalling -1. On
 Chockalingam and Muthuraman: Stochastic Volatility
12

10

Stock price, x
6 b

2 b0

1
0
0.8
0
0.6
1
0.4
2 0.2
3 0
Variance, y
Time to expiry, τ

Figure 2 Initial and final exercise policies

each iteration, the maximum difference between the previous and new associated value functions is
measured for convergence to within a tolerance ǫ. Alternatively, the maximum difference between
the old exercise policy and the new exercise policy can be measured for convergence.

10

6
Stock price, x

4
b3 ≈ b4 ≡ b
b2
3

b1
2

1
b0

0
0 1 2 3
Time to expiry, τ

Figure 3 Exercise policy improvement at y = 1

Figure 4 shows the changes in the first derivative with respect to x through the iterations for
the cut taken at τ = 1.5 months and y = 1. In each figure, the first dotted line represents the
location of the current boundary, and the second represents the location of the largest x such that
Chockalingam and Muthuraman: Stochastic Volatility
13

∂p
∂x
≤ −1. As is evident from Figure 4, the transformation procedure progressively seeks to convert
∂pn
any non-optimal exercise strategies that are exposed by regions with ∂x
< −1. The converged
n
∂p
price function does have ∂x
≥ −1 for all x, satisfying Equation (11).

−0.75 −0.75

−0.8 −0.8

−0.85 −0.85
First derivative w.r.t. x

First derivative w.r.t. x

−0.9 −0.9

−0.95 −0.95

−1 −1

−1.05 −1.05

−1.1 −1.1

0 1 2 3 4 5 6 7 8 4 5 6 7 8 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Stock price, x Stock price, x Stock price, x Stock price, x

b0 b1 b2 b3 ≈ b4 ≡ b
Figure 4 Policy improvement at τ = 1.5 months and y = 1

For further illustrative purposes, also consider the Hull-White model with the following parame-
ters: c1 = 0.3, c2 = 0.6 and ρ = 0. Other parameters are kept the same as in the previously considered
Heston model. Figure 5 plots b0 and the converged b that is obtained using the transformation
procedure. Though the completed optimal-exercise boundaries in Figure 2 (Heston model) and
Figure 5 (Hull-White model) are not directly comparable due to the difference in volatility dynam-
ics and non-comparable parameters, the structural differences at the y = 0 boundary is evident.
Unlike the Heston model, under the Hull-White model, it is strictly optimal to exercise for any x
below K when y = 0. This obviously stems from the fact that the lognormal process used in the
Hull-White model becomes completely deterministic when y = 0, while this is not the case in the
Heston model.

4. Computational Insights and Runtime, Error Comparisons

In this section, we use the transformation procedure described in Section 3 as a computational
tool to study various phenomena. Using an exhaustive set of computational experiments, we seek
insights into the price premium due to stochastic volatility, effects of correlation on option pricing
and exercise policies, effects of market price of volatility risk and the structure of implied volatilities
observed in stochastic volatility models and their dependence on correlation and market price of
volatility risk. Then, we provide runtime/error comparisons against various existing computational
methods discussed earlier in Section 1.
 Chockalingam and Muthuraman: Stochastic Volatility
14

10

Stock price, x
6
b

2
b0

1
0
0 0.8
0.6
1
0.4
2 0.2
3 0
Variance, y
Time to expiry, τ

Figure 5 Optimal exercise policy for the Hull-White model

4.1. Stochastic volatility

Assuming constant volatility, as in the Black-Scholes (BS) model, leads to the mispricing of options.
In this section, we study these price differences, and the effects of introducing stochastic volatility
into the Black-Scholes model.

0.2
0.01
0.05
Difference in option price

0.005
Difference in option price

0.15
Difference in option price

0
0
0.1 −0.05
−0.005
−0.1
−0.01
0.05 0.25
0.25
0.25 −0.015 −0.15
0.2 0.2
0.2 0.15
0 −0.02 0.15 −0.2
20 0.15 20 20
0.1 0.1
15 0.1 15 15
10 10 0.05 10 0.05
0.05
5 5 5
0 0 Time to expiry, τ 0 Time to expiry, τ
0 Time to expiry, τ 0 0
Stock price, x Stock price, x Stock price, x

(a) σ = 0.2 (b) σ = 0.4 (c) σ = 0.6

Figure 6 Price difference (stochastic volatility - constant volatility) at different volatility levels

In Figure 6, we plot the price differences between options priced in a constant volatility setting,
and options priced in a Heston stochastic volatility model setting, for three different volatility
levels. We use the same model parameters considered in Section 3.1, but assume that the volatility
process and the underlying price process are uncorrelated, i.e., ρ = 0. Letting pσ2 (τ, x) denote the
price of a put option with constant volatility σ, we plot p(τ, x, 0.04) − p0.22 (τ, x) in Figure 6(a),
p(τ, x, 0.16) − p0.42 (τ, x) in Figure 6(b), and p(τ, x, 0.36) − p0.62 (τ, x) in Figure 6(c), for various
values of τ and x. The optimal-exercise boundaries for the constant volatility case (dotted line)
and stochastic volatility case (solid line) are plotted on the ceiling/floor of the 3-d plot as well.
Chockalingam and Muthuraman: Stochastic Volatility
15

Figure 6(a) shows that the constant volatility models yields a lower price since it does not account
for the very high likelihood of volatility increasing (σ = 0.2, y = σ 2 = 0.04 and m′ = 0.16) and higher
volatilities yield higher option prices in the classical American and European constant volatility
model. Figure 6(c), however, shows that the constant volatility model yields a higher price than
the stochastic volatility model, as the stochastic volatility model accounts for the likely drop in
volatility. These price differences can be explained by the mean-reverting nature of volatility in the
model. Looking at Figure 6(b), the case when the spot variance is equal to its long-term average, we
find that there is no constant overpricing or underpricing of the option. In this situation, volatility
mean-reversion has negligible effect on the option price. Instead, price differences occur because
of the effect of implied volatilities, since implied volatilities for options at-the-money tend to be
lower than implied volatilities for options in- and out-of-the-money, whereas the constant volatility
model assumes that volatilities are constant everywhere. We also find that the exercise boundaries
for the constant volatility case and stochastic volatility case intersect in Figure 6(b), unlike in
Figures 6(a) and 6(c), where the boundary for the stochastic volatility case is consistently above
or below the constant volatility exercise boundary.

0.5 3.5

0.45
3
2 0.4
Constant
volatility
0.35 2.5
European Option price, p
European Option price, p

European Option price, p

1.5 0.3
2
Stochastic
0.25 volatility Constant
volatility
1.5
1 0.2

Stochastic
volatility 0.15
1 Stochastic
Constant volatility
0.5 volatility 0.1

0.5
0.05

0 0
7.5 8 8.5 9 9.5 10 10.5 11 11.5 12 10.5 11 11.5 12 12.5 13 7 8 9 10 11 12 13 14 15 16
Stock price, x Stock price, x Stock price, x

(a) σ = 0.2 (b) σ = 0.4 (c) σ = 0.6

Figure 7 European option price differences at time T

It would be insightful to study the effects of stochastic volatility on the two components of the
American option price - the associated European option price and the early exercise premium.
We plot European option prices and early exercise premiums, at time T , obtained in the constant
volatility case and the stochastic volatility case in Figures 7 and 8. The effects described in Figure
6 can be readily observed in Figure 7 as well. The European component price reflects that for lower
volatility levels, the mean reverting nature of volatility prices the option higher. However, in Figure
8, for lower volatility levels the premiums are not consistently larger. We find that as the volatility
level increases, early exercise rights for the stochastic volatility setting become relatively more
 Chockalingam and Muthuraman: Stochastic Volatility
16

0.1 0.25

0.09 Constant
volatility 0.2

0.08 0.2
Stochastic
Constant
volatility
volatility
0.07
0.15
Early exercise premium

Early exercise premium

Early exercise premium

0.06 0.15

Stochastic
0.05 volatility
0.1
0.04 0.1
Constant
volatility
0.03
Stochastic
volatility
0.05
0.02 0.05

0.01

0 0 0
9 10 11 12 13 14 15 6 7 8 9 10 11 12 6 7 8 9 10 11 12 13
Stock price, x Stock price, x Stock price, x

(a) σ = 0.2 (b) σ = 0.4 (c) σ = 0.6

Figure 8 Changes in early exercise premium at time T

valuable. This is because early exercise premiums benefit significantly from higher randomness and
as volatility levels increase, randomness in the stochastic volatility setting becomes significantly
higher. One also has to keep in mind that when in the money, the option holder prefers randomness
much less than when out of the money.

0.4 0.45

0.6
0.35 0.4
Deterministic Deterministic
mean−reverting mean−reverting Deterministic
volatility model volatility model mean−reverting
0.35 0.5 volatility
0.3
model

0.3
0.25 0.4
Option price, p

Option price, p
Option price, p

0.25
Stochastic
0.2 model
0.3 Constant
0.2 Stochastic volatility
Stochastic model
0.15 model
0.15
0.2

0.1 Constant
volatility 0.1

0.1
0.05 0.05

0 0 0
10.5 11 11.5 12 12.5 11 11.5 12 12.5 13 11 11.5 12 12.5 13 13.5 14 14.5 15 15.5
Stock price, x Stock price, x Stock price, x

(a) σ = 0.2 (b) σ = 0.4 (c) σ = 0.6

Figure 9 Comparison between deterministic and stochastic mean-reverting volatility at time T

To study solely the effect of stochastic volatility on option pricing, we plot option prices obtained
using a deterministic mean-reverting volatility model, the stochastic volatility model and the con-
stant volatility model at time T in Figure 9, for σ = 0.2 , 0.4 and 0.6. The mean-reverting effect
that was observed in Figure 6 can be observed in Figure 9 as well, as is to be expected. In Fig-
ure 9(b), the plot of option prices for the constant volatility model is not clearly observable, as it is
actually very close to the plots of option prices for the stochastic and deterministic mean-reverting
volatility models.
In all three cases, with respect to the deterministic model, the stochastic nature of volatility
underprices options for lower stock prices and overprices for larger stock prices. For stock prices well
below the option’s strike price, the holder would prefer less randomness in the system to increase
Chockalingam and Muthuraman: Stochastic Volatility
17

the probability of the option remaining, or expiring, in-the-money. A deterministic mean-reverting

volatility model would thus be preferable to a stochastic model. Therefore, option prices for the
deterministic model are higher, reflecting the holder’s preference for the deterministic model. But
for higher stock prices, the holder would prefer more randomness to ensure that the stock price will
fall below the strike price of the option. One can also note that as σ increases, the stock price at
which the two sets of option prices intersect increases as well. This implies that the deterministic
model is preferred over the stochastic model by the option holder for a larger range of stock prices.
Now, in the deterministic model, volatilities tend towards the mean volatility level in a known
fashion. As the volatility level increases, the downward drift of the volatility process increases as
well. Thus, as volatility increases, the deterministic model is capable of bringing about drops in
larger stock prices and increasing the probability of the option expiring in-the-money. As a result,
we find that the threshold over which the stochastic model becomes preferred to the deterministic
model increases as volatility increases.

4.2. Correlation, stochastic volatility and option pricing

10
ρ = −1 ρ = −1
2 ρ=0 ρ=0
ρ=1 ρ=1
9.5

9 τ = 1.5 months
1.5
Option price, p

Stock price, x

8.5

1
8

7.5

0.5

0 6.5
8 8.5 9 9.5 10 10.5 11 11.5 12 0 1 2 3
Stock price, x Time to expiry, τ

(a) Price functions (τ = 1.5 months & σt = 0.4) (b) Exercise policies (σt = 0.4)
Figure 10 Effects of changing correlation ρ

This section considers the effects of correlation between the underlying price process and the
volatility process, otherwise known as the ‘leverage’ effect. The leverage effect is well studied in the
classical setting. In the stochastic volatility setting correlation is often assumed to be zero though
inconsistent with real-world data. Option prices and exercise policies for correlation ρ set to -1, 0
and 1 are shown in Figure 10. Correlation affects the option price differently when the stock price
 Chockalingam and Muthuraman: Stochastic Volatility
18

x is above and below the strike price K. It is interesting to note that the option price is unaffected
by ρ when x = K. As Figure 10(a) illustrates, option prices increase with correlation when x < K.
The opposite is true when x > K.
An increase in correlation can be thought of as a decrease in overall uncertainty in the system.
When out of the money (x > K), it is only natural that the holder prefers more randomness to
increase chances of a positive pay-off, hence explaining option price decreases with increases in
correlation when out of the money. When in the money (x < K), the holder’s preference for less
randomness is indicated by the increase in price with increases in correlation.
Exercise boundaries decrease as correlation increases, as shown by Figure 10(b). This aligns with
the notion that under optimality, higher prices imply later exercise, since at the exercise boundary,
the pay-off is always (K − x)+ .

4.3. Impact of market price of volatility risk on option pricing

Next, lets consider the effect of Λ, the market price of volatility risk. An interpretation for Λ can be
obtained by considering delta-hedged option portfolios that are constructed by buying an option
and hedging it with a fraction of the underlying asset, such that in the Black-Scholes setting, the
rate of return matches the risk free interest rate. The gain on this portfolio is the difference between
the risk-free rate and the earnings from the portfolio. Obviously, in a Black-Scholes setting, when
the delta-hedged portfolio is continuously rebalanced, the delta-hedged gain is always zero. Bakshi
and Kapadia (2003) study and relate the delta-hedged gains to the market price of volatility risk.
They show that when the market price of volatility risk Λ is positive (negative), the expected delta-
hedged gain is positive (negative). It also implies that when volatility is stochastic and volatility
risk not priced, i.e. when Λ = 0, the expected delta-hedge gain is zero. Further, using market data,
they demonstrate empirically that Λ is negative. The finding that Λ is negative is consistent with
the notion that market volatility rises when market return drops.
Figure 11(a) shows that as Λ increases, p decreases. Henderson (2005) shows that for European
put options, as the market price of volatility risk increases, the price of the option decreases. The
question is whether this translates to American option pricing as well. Figure 11 provides numerical
evidence that it does hold true for American put options as well.
An insight into this behavior is revealed by considering the delta-hedged portfolio. When the
market does not price volatility risk, the delta-hedged gain should be zero and the price of the
option matches the price of constructing the portfolio. When the market does price volatility risk,
however, the delta-hedged gain is proportional to the volatility risk premium. Hence, as Λ increases,
the expected gain on this portfolio also increases. In this case, the price of the option would be the
price of constructing the portfolio minus the expected gain.
Chockalingam and Muthuraman: Stochastic Volatility
19

10
2 Λ = −2 Λ = −2
Λ=0 Λ=0
Λ=2 Λ=2
1.8
9.5

1.6

1.4 9
τ = 1.5 months
Option price, p

Stock price, x
1.2

8.5
1

0.8
8
0.6

0.4
7.5

0.2

0 7
8 9 10 11 12 13 14 0 1 2 3
Stock price, x Time to expiry, τ

(a) Price functions (τ = 1.5 months & σt = 0.4) (b) Exercise policies (σt = 0.4)
Figure 11 Effects of changing volatility risk premium Λ

4.4. On implied volatilities

0.45 0.44

Λ = −2
0.42
ρ = −0.8
0.4 0.4
ρ = −0.5
Λ = −1
0.38
ρ=0
Implied volatility
Implied volatility

0.35 0.36
Λ=0

0.34

ρ = 0.5 Λ=1
0.3 0.32

0.3
Λ=2

0.25 ρ = 0.8 0.28

0.26

0.2 0.24
8 8.5 9 9.5 10 10.5 11 11.5 12 8 8.5 9 9.5 10 10.5 11 11.5 12
Strike price, K Strike price, K

(a) Changing ρ (b) Changing Λ

Figure 12 Effects of changing ρ and Λ on implied volatilities

One of the primary arguments favoring stochastic volatility models is that the implied volatilities
computed from observed option prices are not constant and often exhibit a ‘smile’ curve when
plotted against strike prices. We examine the nature of implied volatility curves and their depen-
dence on both correlation and market price of volatility risk in Figure 12. Implied volatilities are
computed for various strike prices from 8 to 12, at an underlying stock price x = 11. We use a
bisection search in conjunction with the binomial-tree method to calculate the implied volatilities.
The smile curves are well-captured in Figure 12(a). They pivot at K = 11, which is the underlying
 Chockalingam and Muthuraman: Stochastic Volatility
20

stock price at which the implied volatilities are computed. When K < x, the implied volatility
decreases as correlation increases. On the other hand, when K > x, a decrease in correlation causes
a decrease in the implied volatility as well. From Section 4.2, we know that for K < x, as ρ increases,
p decreases. This translates directly to decreasing implied volatility. A drop in p therefore implies a
drop in volatility. By a similar reasoning, since p increases as ρ increases for K > x, the monotonic
relationship between option price and volatility leads to the implied volatility increasing.
The smile curves in Figure 12(b) demonstrate that as Λ increases, the implied volatility decreases.
Recall from Section 4.3 that as Λ increases, the option price p decreases. This decrease in option
price leads to a lower implied volatility, because option prices increase monotonically with volatility.

4.5. Runtime and accuracy comparisons

Ikonen and Toivanen (2007b) exhaustively compare the five methods available in literature for
pricing American options under stochastic volatility. These are the PSOR, multi-grid, operator
splitting, penalty and the component-wise splitting (CS) method. Of these, they find that CS
performs fastest, with comparable accuracy to the other methods. From the standpoint of imple-
mentation, the authors do note that the PSOR method is the easiest while CS is the hardest. The
remaining three methods fall in between these two in terms of speed/accuracy and ease of imple-
mentation, highlighting a tradeoff between ease of implementation against the speed/accuracy of
the method. In this section, our objective is to place the method presented in this paper. To this
extent, we compare the speed and accuracy to that of only the CS method and the PSOR method,
since the other methods are known to fall between these two, both in terms of speed/accuracy and
implementation ease. The comparisons were carried out using a C++ implementation with the
GMM++ library on a 2.8 Ghz Intel Xeon Mac Pro with 2 GB RAM and 1.6 Ghz Bus Speed.
It is also important to keep in mind that implementation of the proposed method is quite easy
and straightforward since it solves only a fixed-boundary PDE problem in each iteration and that
the PDE is the classical convection-diffusion PDE with a second-order cross derivative term. Several
off-the-shelf packages can do this readily. We also provide in Appendix B a simple finite difference
scheme that solves the fixed-boundary problem which is used for the results in this section. In
the finite difference scheme, time stepping is done using an implicit Euler method. Note that the
method proposed here, in its current form, works for all the popular stochastic volatility models
and allows for both positive and negative volatility risk premiums.
As earlier, for our method, the first boundary guess is b0 (τ, y) = 1 for all τ and y and we take
the parameter set that has been used in the illustration. The relaxation parameter and stopping
criterion used for the PSOR method are the same as those found in Ikonen and Toivanen (2007b).
Chockalingam and Muthuraman: Stochastic Volatility
21

Tables 3 and 4 list the prices obtained using the respective methods for five initial asset prices and
two volatilities for various grid sizes. For the true values, we use the values listed in Ikonen and
Toivanen (2007b), which the authors obtain from using the CS method in conjunction with a very
fine grid.

X0
Method Grid size 8 9 10 11 12

PSOR (40,16,8) 2.0000 1.0952 0.4966 0.2042 0.0838

(60,32,66) 2.0000 1.1037 0.5142 0.2105 0.0815
(120,64,130) 2.0000 1.1064 0.5182 0.2126 0.0819
(240,128,258) 2.0000 1.1071 0.5193 0.2133 0.0820

Componentwise (40,16,8) 2.0004 1.1003 0.4991 0.2035 0.0828

splitting (60,32,66) 2.0000 1.1043 0.5147 0.2104 0.0813
(120,64,130) 2.0000 1.1066 0.5183 0.2126 0.0819
(240,128,258) 2.0000 1.1073 0.5194 0.2133 0.0820

Transformation (40,16,8) 2.0000 1.0952 0.4966 0.2042 0.0838

procedure (60,32,66) 2.0000 1.1035 0.5142 0.2105 0.0815
(120,64,130) 2.0000 1.1063 0.5181 0.2126 0.0819
(240,128,258) 2.0000 1.1071 0.5193 0.2133 0.0820

True Value 2.0000 1.1076 0.5200 0.2137 0.0820

Table 3 Option prices at y = 0.0625

Figure 13 plots the root mean square errors (RMSE) and run times for the three different methods
for the various grid sizes listed in Tables 3 and 4. The performance of the CS method in relation
to the PSOR method is comparable to the relation found in Ikonen and Toivanen (2007b). As the
figure shows, for the same accuracy the transformation procedure is on average 10 times faster
than the PSOR method and more than twice as fast as the component-wise splitting method.
As the figure shows, the transformation procedure has better accuracy than the CS method,
particularly on coarser grids. This is understandable since the CS method introduces additional
errors when the splitting is performed. At the same time, the speed of our scheme is greater than
that of the PSOR method and the CS method. As mentioned before, the CS method is harder to
implement than the PSOR method, but this leads to the better performance of the CS method. It
is interesting to note, however, that implementing our transformation procedure is no more difficult
than implementing the PSOR method as our transformation procedure is essentially a standard
finite difference implementation, modified to update the boundary during each iteration.
 Chockalingam and Muthuraman: Stochastic Volatility
22

X0
Method Grid size 8 9 10 11 12

PSOR (40,16,8) 2.0691 1.3139 0.7720 0.4293 0.2324

(60,32,66) 2.0760 1.3292 0.7908 0.4442 0.2405
(120,64,130) 2.0775 1.3320 0.7940 0.4467 0.2419
(240,128,258) 2.0779 1.3329 0.7951 0.4476 0.2424

Componentwise (40,16,8) 2.0676 1.3094 0.7646 0.4232 0.2297

splitting (60,32,66) 2.0758 1.3287 0.7900 0.4435 0.2401
(120,64,130) 2.0774 1.3317 0.7936 0.4463 0.2417
(240,128,258) 2.0780 1.3328 0.7949 0.4474 0.2423

Transformation (40,16,8) 2.0691 1.3140 0.7721 0.4294 0.2325

procedure (60,32,66) 2.0760 1.3291 0.7908 0.4442 0.2405
(120,64,130) 2.0775 1.3319 0.7940 0.4467 0.2419
(240,128,258) 2.0780 1.3329 0.7951 0.4476 0.2424

True Value 2.0784 1.3336 0.7960 0.4483 0.2428

Table 4 Option prices at y = 0.25

3
10
240x128x258 Transformation procedure
Componentwise splitting
PSOR
2
10
CPU time in seconds

120x64x130

1
10
60x32x66

0
10

40x16x8

−1
10

−3 −2
10 10
RMSE

Figure 13 A comparison of RMSE and computing time

The speed and accuracy of option prices obtained using the transformation procedure are highly
dependent on the choice of the fixed-boundary problem solver. As Figure 13 shows however, even
Chockalingam and Muthuraman: Stochastic Volatility
23

with simple finite differences on a uniform grid, the scheme performs well, obtaining accurate
option prices quickly. In our C++ implementation, we have used the Generalized Minimal Resid-
ual Method (GMRES) provided by the GMM++ library to solve the resulting system of linear
equations. The use of specialized solvers such as finite element methods would increase the speed
and accuracy of the scheme by a significant factor, however by complicating implementation if
off-the shelf packages are not taken advantage of. This is illustrated in Muthuraman (2008) when
comparing American option pricing methodologies in a Black-Scholes setting.

5. Concluding Remarks
Much of the literature on option pricing under stochastic volatility focuses on European options. We
considered in this paper the harder problem of pricing American options in a stochastic volatility
setting. It was shown that computing the price of an American option under stochastic volatility
is only as difficult as computing the price of a series of European-type options when the exercise
policies are predetermined. A computational procedure for calculating the price as well as the
optimal-exercise boundary was developed. Using this procedure, we have sought insights into the
dependence of American option prices, exercise policies and implied volatilities on factors such as
the market price of volatility risk and the correlation between stock price and the volatility process.
The method was demonstrated to be as accurate as the PSOR method, while having better speeds
than other existing methods in our numerical experiments. An avenue for future research would
be to extend such a pricing methodology for other American-type derivatives. Specifically, pricing
options on multiple assets would be interesting since there are also inherently multi-dimensional
problems. The theoretical guarantees established for the proposed method critically depend on
the maximum principle (Theorem A.1). This begs the question whether the methodology can be
extended to higher dimensions if one can establish the maximum principle. For a problem in higher
dimensions, if one can determine the boundary update conditions and also prove the maximum
principle for the difference in value between two iterations, then certainly the critical part of
establishing convergence is in place. The other details would be problem dependent. Empirical
studies that explore questions of investor exercise behaviors against those computed by the optimal-
exercise policy would be interesting as well.

References
Andersen, L., M. Broadie. 2004. Primal-Dual Simulation Algorithm for Pricing Multidimensional American
Options. Management Science 50(9) 1222 – 1234.

Bakshi, G., N. Kapadia. 2003. Delta-Hedged Gains and the Negative Market Volatility Risk Premium.
Review of Financial Studies 16(2) 527–566.
 Chockalingam and Muthuraman: Stochastic Volatility
24

Ball, C. A., A. Roma. 1994. Stochastic Volatility Option Pricing. J. of Financial and Quantitative Analysis
29(4) 589–607.

Björk, T. 2004. Arbitrage Theory in Continuous Time. Oxford University Press.

Black, F., M. Scholes. 1973. The Pricing of Options and Corporate Liabilities. J. Political Economy 81(3)
637–654.

Bollerslev, Tim, Ray Y. Chou, Kenneth F. Kroner. 1992. Arch modeling in finance : A review of the theory
and empirical evidence. Journal of Econometrics 52(1-2) 5–59.

Brennan, M. J., E. S. Schwartz. 1977. The Valuation of American Put Options. J. of Finance 32(2) 449–462.

Broadie, M., J. Detemple. 2004. Option pricing: Valuation models and applications. Management Science
50(9) 1145–1177.

Broadie, M., J. Detemple, E. Ghysels, O. Torrés. 2000. American Options with Stochastic Dividends and
Volatility: A Nonparametric Investigation. Journal of Econometrics 94 53 – 92.

Broadie, M., P. Glasserman. 1997. Pricing American-Style Securities by Simulation. J. Economic Dynamics
and Control 21 1323–1352.

Broadie, M., P. Glasserman. 2004. A Stochastic Mesh Method for Pricing High-Dimensional American
Options. J. of Computational Finance 7(4) 35 – 72.

Carr, P., R. Jarrow, R. Myneni. 1992. Alternative Characterizations of American Put Options. Mathematical
Finance 2 87–106.

Clarke, N., K. Parrott. 1999. Multigrid for American Option Pricing with Stochastic Volatility. Applied
Mathematical Finance 6 177–195.

Cox, J. C., S. A. Ross, M. Rubinstein. 1979. Option Pricing: A Simplified Approach. J. of Financial
Economics 7 229–263.

Cryer, C. W. 1971. The Solution of a Quadratic Programming Problem Using Systematic Overrelaxation.
SIAM J. on Control 9 385 – 392.

Dragulescu, A.A., V.M. Yakovenko. 2002. Probability Distribution of Returns in the Heston Model with
Stochastic Volatility. Quantitative Finance 2(6) 443 – 458.

Fouque, J., G. Papanicolaou, K. Sircar. 2000. Derivatives in Financial Markets with Stochastic Volatility.
Cambridge University Press.

Henderson, V. 2005. Analytical Comparisons of Option Prices in Stochastic Volatility Models. Mathematical
Finance 15(1) 49–59.

Heston, S. L. 1993. A Closed-form Solution for Options with Stochastic Volatility and Applications to Bond
and Currency Options. Review of Financial Studies 6(2) 327–343.

Hull, J., A. White. 1987. The Pricing of Options on Assets with Stochastic Volatilities. J. of Finance 42(2)
281–300.
Chockalingam and Muthuraman: Stochastic Volatility
25

Ikonen, S., J. Toivanen. 2004. Operator Splitting Methods for Pricing American Options with Stochastic
Volatility. Applied Mathematics Letters 17(7) 809–814.

Ikonen, S., J. Toivanen. 2007a. Componentwise Splitting Methods for Pricing American Options Under
Stochastic Volatility. International Journal of Theoretical and Applied Finance 10(2) 331–361.

Ikonen, S., J. Toivanen. 2007b. Efficient Numerical Methods for Pricing American Options Under Stochastic
Volatility. Numerical Methods for Partial Differential Equations 24(1) 104 – 126.

Jacka, S. D. 1991. Optimal Stopping and the American Put. Mathematical Finance 1 1–14.

Karatzas, I., S. E. Shreve. 1998. Methods of Mathematical Finance. Springer-Verlag.

Kim, I. J. 1990. The Analytic Valuation of American Options. Review of Financial Studies 3 547–572.

Kumar, S., K. Muthuraman. 2004. A numerical method for solving stochastic singular control problems.
Operations Research 52(4) 563–582.

Longstaff, F. A., E. S. Schwartz. 2001. Valuing American Options by Simulation: Simple Least-Squares
Approach. Review of Financial Studies 14 113–147.

Muthuraman, K. 2008. A Moving Boundary Approach to American Option Pricing. J. Economic Dynamics
and Control 32(11) 3520 – 3537.

Muthuraman, K., S. Kumar. 2006. Multi-dimensional portfolio optimization with proportional transaction
costs. Mathematical Finance 16 301–335.

Myneni, R. 1992. The Pricing of the American Option. The Annals of Applied Probability 2(1) 1–23.

Nielsen, B. F., O. Skavhaug, A. Tveito. 2002. Penalty and Front-fixing Methods for the Numerical Solution
of American Option Problems. J. of Computational Finance 5(4) 69–97.

Oosterlee, C. W. 2003. On Multigrid for Linear Complementarity Problems with Application to American-
style Options. Electronic Transactions on Numerical Analysis 15 165–185.

Pantazopoulos, K. N., E. N. Houstis, S. Kortesis. 1998. Front-Tracking Finite Difference Methods for the
Valuation of American Options. Computational Economics 12(3) 255–273.

Rubinstein, M. 1994. Implied Binomial Trees. J. of Finance 49(3) 771–818.

Scott, Louis O. 1987. Option Pricing when the Variance Changes Randomly: Theory, Estimation, and an
Application. J. of Financial and Quantitative Analysis 22(4) 419–438.

Stein, E. M., J. C. Stein. 1991. Stock Price Distributions with Stochastic Volatility: An Analytic Approach.
Review of Financial Studies 4(4) 727–752.

Wu, L. X., Y. K. Kwok. 1997. A Front-fixing Finite Difference Method for the Valuation of American
Options. J. of Financial Engineering 6(2) 83–97.

Zvan, R., P. A. Forsyth, K. R. Vetzal. 1998. Penalty Methods for American Options with Stochastic Volatility.
J. Comput. Appl. Math. 91(2) 199–218.
 Chockalingam and Muthuraman: Stochastic Volatility
26

Appendix
A. Proofs
The proofs of theorems and propositions are collected here. We begin with the statement and proof of
Theorem A.1, which establishes the maximum principle that is needed for subsequent proofs.

Theorem A.1. For a given T̂ ∈ (0, ∞) and a continuous g(τ, y) > 0 for all (τ, y) ∈ (0, T̂ ) × R+ , let Cˆ =
{(τ, x, y) ∈ (0, T̂ ) × R2+ | x > g(τ, y)}. Also, let h be the solution to

Lh = 0 ˆ
in C, (15)

(L is defined in Equation (4)), with boundary conditions given by,

h(0, x, y) = 0, (16)
h(τ, g(τ, y), y) = F (τ, y), (17)
∂h
lim = 0, (18)
x→∞ ∂x
∂h
lim =0 and, (19)
y→∞ ∂y
∂h ∂h ∂h
rx + κm′ − rh − =0 at y = 0. (20)
∂x ∂y ∂τ
If r > 0 and F (τ, y) > 0 for all (τ, y) ∈ (0, T̂ ) × R+ , then the maxima of h are attained only on the boundary
(τ, g(τ, y), y) and the minima of h are attained on the boundary (0, x, y).
∂2h ∂2h ∂2 h
Proof: For notational convenience, let A, B and C represent ∂x2
, ∂x∂y
and ∂y 2
respectively.
We show that the maxima of h is attained only on the boundary (τ, g(τ, y), y) by ruling out other possibili-
ties. First, say an internal maxima exists and is attained at some (τ ′ , x′ , y ′ ). Now, this maxima will be no less
than h(τ, x, y) for all τ , x and y, meaning that h(τ ′ , x′ , y ′ ) ≥ h(τ, g(τ, y), y). This implies that h(τ ′ , x′ , y ′ ) > 0
∂h ∂h ∂h
since F (τ, y) > 0. Also, by the necessary conditions for an internal maxima, we have that ∂x
= ∂y
= ∂τ
= 0.
Substitution into Equation (15) yields
1 1
rh = Ax2 y + Bρυxy + Cυ 2 y. (21)
2 2
The Hessian needs to be a negative semidefinite matrix. This implies that the determinant of the leading
n × n principal minor of the Hessian is non-negative (non-positive) when n is even (odd). Hence at (τ ′ , x′ , y ′ ),
AC − B 2 ≥ 0 and A ≤ 0. Thus, at this point, if A 6= 0, the second-order differential terms in Equation (21)
can be rearranged as
Cυ 2
 
1 2 1 2 1 2 2Bρυ
Ax y + Bρυyx + Cυ y = Ay x + x+
2 2 2 " A A
2   #
1 Bρυ AC − B 2 ρ2
= Ay x + + υ2 . (22)
2 A A2

Now, since ρ2 ≤ 1, we have AC − B 2 ρ2 > AC − B 2 > 0. Hence, from Equation (21), we have rh < 0 implying
h(τ ′ , x′ , y ′ ) < 0, a contradiction. If A = 0, it follows that B = 0, leading to a similar contradiction. Therefore,
the maxima cannot be attained in the interior.
Chockalingam and Muthuraman: Stochastic Volatility
27

Now assume that the maximum is attained at some (T̂ , x′ , y ′ ), i.e., on the boundary τ = T̂ . We must have
∂h ∂h ∂h
that h(T̂ , x′ , y ′ ) ≥ maxτ,y F (τ, y) > 0, with ∂x
= ∂y
= 0 and ∂τ
≥ 0. Substitution into Equation (15) now
yields
1 1 ∂h
rh = Ax2 y + Bρυxy + Cυ 2 y − . (23)
2 2 ∂τ
Considering the function h along the cut taken at T̂ , we must again have A ≤ 0 and AC − B 2 ≥ 0. Rearranging
(if A 6= 0) we have " #!
2
AC − B 2 ρ2
 
1 1 Bρυ 1 ∂h
h= Ay x+ + υ2 − . (24)
r 2 A A2 r ∂τ

Using similar arguments as earlier, even when A = 0, we have h(T̂ , x′ , y ′ ) < 0, again a contradiction. The
maxima cannot be attained on the boundary τ = T̂ either.
Next, say that the maximum is attained at some (τ ′ , x′ , 0), i.e., on the boundary y = 0. Again, at (τ ′ , x′ , 0),
∂h ∂h ∂h
by conditions of maxima, we have ∂x
= ∂τ
= 0 and ∂y
≤ 0. Substitution into Equation (20) gives

1 ∂h
h = κm′ , (25)
r ∂y

leading to h(τ ′ , x′ , 0) ≤ 0, another contradiction.

Finally, we need to rule out the possibility of a maxima being reached as y → ∞. Consider a boundary
h
2 i
at a Ŷ > max Λ2κυ , m′ . Now assume that the maxima is achieved at some (τ ′ , x′ , Ŷ ) on the boundary
y = Ŷ . At this point, the coefficient of hy is negative, and since this is a maxima, the Hessian is negative
semidefinite, implying a negative LHS, therefore resulting in a contradiction.
Now since the maxima clearly cannot be attained on the boundary τ = 0, this leaves us with the result.
Arguments and reasoning as in the above will establish that the minima of h is achieved on the boundary
(0, x, y).

ˆ
Proposition A.1. If F (τ, y) = 0 for all (τ, y) ∈ (0, T̂ ) × R+ , then h(τ, x, y) = 0 for all (τ, x, y) ∈ C.

Proof: The proof follows directly from the proof of Theorem A.1. If F (τ, y) = 0 for all (τ, y) ∈ (0, T̂ ) ×
R+ , the maximum of h is also 0, since the maxima of h is attained at the boundary (τ, g(τ, y), y), where
ˆ
h(τ, g(τ, y), y) = F (τ, y). Since the minima of h is also 0, this must mean that h = 0 for all (τ, x, y) ∈ C. 
Proof of Proposition 3.1: Assuming there exists two solutions h1 and h2 , considering the equations
solved by h̄ = h1 − h2 and using Theorem A.1 directly gives the result.
Proof of Proposition 3.2: Let Cˆ0 = {(τ, x, y) ∈ (0, T̂ ) × R2+ |x > b0 (τ, y)} and
Cˆb = {(τ, x, y) ∈ (0, T̂ ) × R2+ |x > b(τ, y)}. In the region Cˆ0 − Cˆb , it is optimal to exercise, but the exercise policy
dictated by b0 chooses to sub-optimally hold. Hence in Cˆ0 − Cˆb , p0 < p = (K − x)+ . On the boundary b0 , we
have that p0 = (K − x)+ . Also, by Theorem A.1, the maxima of p0 is attained on b0 . Since p0 = K − x on b0

0
and p0 < K − x in Cˆ0 − Cˆb , we must have that ∂p
∂x 
< −1. 
0 (τ,b (τ,y )+,y )
For the rest of this section, subscripts denote derivatives.
 Chockalingam and Muthuraman: Stochastic Volatility
28

Proof of Theorem 3.1: Since pnx (τ, bn (τ, y)+, y) < −1 and lim pnx (τ, x, y) = 0, bn+1 exists by the
x→∞

continuity of pnx in Cˆn , where Cˆn = {(τ, x, y) ∈ (0, T̂ ) × R2+ |x > bn (τ, y)}. The definition of bn+1 also implies
bn+1 > bn .
From the definition of bn+1 , for all x ∈ (bn (τ, y), bn+1 (τ, y)), τ and y, we have pnx (τ, x, y) < −1. This implies

pn (τ, bn+1 (τ, y), y) − pn (τ, bn (τ, y), y) < −(bn+1 (τ, y) − bn (τ, y))
pn (τ, bn+1 (τ, y), y) < pn (τ, bn (τ, y), y) + bn (τ, y) − bn+1 (τ, y)
= K − bn+1 (τ, y)
= pn+1 (τ, bn+1 (τ, y), y)

Thus, pn+1 > pn on bn+1 . Now, the difference p̂ = pn+1 − pn solves

Lp̂ = 0 in Cˆn+1 ,
p̂(0, x, y) = 0,
p̂(τ, bn+1 (τ, y), y) > 0,
lim p̂x (τ, x, y) = 0,
x→∞
lim p̂y (τ, x, y) = 0, and
y→∞

Ap̂ = 0.

By Theorem A.1, p̂ attains its maxima on bn+1 and its minima of 0 on the boundary (0, x, y) for (x, y) ∈
(bn+1 , ∞) × R+ . This implies that p̂ > 0 in Cˆn+1 , i.e., pn+1 > pn .
Finally, we show that pnx +1 (τ, bn+1 (τ, y)+, y) < −1. Since pnx +1 = pnx + p̂x and pnx (τ, bn+1 (τ, y), y) = −1 by
the definition of bn+1 , it suffices to show that p̂x (τ, bn+1 (τ, y)+, y) < 0. Assume that p̂x (τ, bn+1 (τ, y)+, y) ≥ 0
instead. Now since lim p̂x (τ, x, y) = 0, p̂x (τ, bn+1 (τ, y)+, y) ≥ 0 implies that the maxima of p̂ is attained in
x→∞
Cˆn+1 . But this contradicts Theorem A.1, which states that the maxima of p̂ is attained on bn+1 . Therefore,
we must have that p̂x (τ, bn+1 (τ, y)+, y) < 0 ⇒ pnx +1 (τ, bn+1 (τ, y)+, y) < −1. 
B. Finite Difference Implementation
The fixed-boundary problem defined by Equations (4) - (10) can be solved using the finite difference method.
In this section, we discuss relevant implementation issues using the finite difference scheme for the Heston
stochastic volatility model.
For the sake of numerical implementation, the time axis, asset-price axis and variance axis are truncated
to [0, T̂ ], [0, X̂] and [0, Ŷ ] respectively, for some large enough T̂ , X̂, Ŷ ∈ R+ . The boundary conditions for
x → ∞ and y → ∞ are applied at X̂ and Ŷ respectively.
The time axis, asset-price axis and variance axis are discretized into l, m and n pieces yielding grid steps
T̂ X̂ Ŷ
δτ = l
, δx = m
and δy = n
respectively. For k = 0, . . . , l, i = 0, . . . , m and j = 0, . . . , n, the price of the
option at node (k, i, j) is denoted by p(τk , xi , yj ) = p(δτ k, δx i, δy j) = pki,j . Using this notation, a finite difference
discretization based on central differences can be obtained as follows:

∂p/∂x = (pki+1,j − pki−1,j )/2δx ,

Chockalingam and Muthuraman: Stochastic Volatility
29

∂p/∂y = (pki,j+1 − pki,j−1 )/2δy ,

∂ 2 p/∂x2 = (pki+1,j − 2pki,j + pki−1,j )/δx2 ,
∂ 2 p/∂y 2 = (pki,j+1 − 2pki,j + pki,j−1 )/δy2 ,
∂ 2 p/∂x∂y = (pki+1,j+1 − pki−1,j+1 − pki+1,j−1 + pki−1,j−1 )/4δx δy .

We obtain from Equation (4)

1 k
Di,j pi−1,j−1 + Dj2 pki,j−1 + Di,j
3 k 4 k
pi+1,j−1 + Di,j 5 k
pi−1,j + Di,j pi,j
6 k 7
+Di,j pi+1,j Di,j pki−1,j+1 + Dj8 pki,j+1 + Di,j
9 k
pi+1,j+1 = pk− 1
i,j , (26)

where

1 δτ ρυji
Di,j =− ,
4′ √
δτ κm δτ κj δτ υ 2 j δτ Λυ j
Dj2 = − − − p ,
2δy 2 2δy 2 δy
3 δτ ρυji
Di,j = ,
4
4 δτ ri δτ δy ji2
Di,j = − ,
2 2
5 δτ υ 2 j
Di,j = 1 + δτ δy ji2 + rδτ + ,
δy
2
6 δτ δy ji δτ ri
Di,j = − − ,
2 2
7 δτ ρυji
Di,j = ,
4 √
δτ κj δτ Λυ j δτ υ 2 j δτ κm′
Dj8 = + p − − and
2 2 δy 2δy 2δy
9 δτ ρυji
Di,j = − .
4
Similarly, Ap = 0 is discretized to yield

1 δτ ri k κm′ δτ k δτ ri k κm′ δτ k
pk−
i,j = pi−1,j + (1 + rδτ + )pi,j − pi+1,j − pi,j+1 .
2 δy 2 δy
Given an exercise policy b, the remaining boundary conditions are represented as follows:

p0i,j = (K − δx i)+ for i = 0, . . . , m, (27)

pki,j = (K − δx i)+ if xi ≤ b(τk , yj ), (28)
k k
p i,j −p i−1,j =0 if i = m (xi = X̂), and (29)
pki,j − pki,j−1 = 0 if j = n (yj = Ŷ ). (30)

1
Given the price for a time to expiry τk−1 , i.e, pk−
i,j ∀i, j, the price at time τk can be obtained by solving

a system of linear equations Dpk = pk−1 , where pk is a mn-vector that represents the option prices for all
asset prices and volatilities at time step τk and the mn × mn matrix D is assembled using Equations (26)
- (30) at each time step. This set of equations is solved for each k = 1, . . . , l. The resulting matrix p is then
the value function associated with the exercise policy b.