Afstudeerverslagliekedejong Finalversion

Delft University of Technology
Faculty of Electrical Engineering, Mathematics and Computer Science

Delft Institute of Applied Mathematics
Option pricing with perturbation methods
A thesis submitted to the
Delft Institute of Applied Mathematics
in partial fulfillment of the requirements
for the degree
MASTER OF SCIENCE
in
APPLIED MATHEMATICS
by
Lieke de Jong
Delft, the Netherlands
January, 2010
Copyright c _ 2009 by Lieke de Jong. All rights reserved.
MSc THESIS APPLIED MATHEMATICS
Option pricing with perturbation methods
Lieke de Jong
Delft University of Technology
Daily supervisors Responsible professor
Drs. M. Lutterot (Rabobank) Prof. dr. ir. A.W. Heemink (TU Delft)
Dr. ir. W.T. van Horssen (TU Delft)
Other thesis committee members
Drs. S. van Weeren (Rabobank)
Prof. dr. C.W. Oosterlee (TU Delft)
January, 2010 Delft, the Netherlands
Abstract
This thesis discusses the use of perturbation theory in the context of financial mathematics, in
particular on the use of matched asymptotic expansions in option pricing.
Our methods are applied to the ordinary Black-Scholes model for illustration. In this simple
example of the Black-Scholes model an exact solution is available, so it is in fact not neccessary
to apply the method of asymptotic expansions on this model. However, in case we do apply
the method, two artificial layers have to be constructed. Making smart choices for the local
variables leads to a transformation of the equations into a heat equation, which can easily be
solved. Finally, the results are compared to a Taylor expansion of the exact solution to see that
this method is very accurate.
After this first instructive model, the method of matched asymptotic expansions is applied to
two more advanced models based on papers by Howison [7] and Hagan et al. [5]. Here, different
choices for the scalings are made.
The former discusses a fast mean-reverting stochastic volatility model that turns out to have
many open ends. In Howisons paper [7] quite a lot of assumptions and simplifications are made.
Unfortunately, often the motivation for them is not explicitly given in the paper, and in some
cases we even think these assumptions and simplifications are incorrect.
The latter examines a new three-parameter stochastic volatility model that successfully prices
back the volatility smile as observed in the market nowadays, and that is commonly used. The
derivation of this model is the main focus of this thesis. The resulting expression for the implied
volatility under the SABR model is obtained by considering the forward and backward Kol-
mogorov equations per order in , making some smart choices for local variables and functions
in order to transform them into an equation that looks like a heat equation, which is easier to
solve.
Recommendations for further investigation on these models would be to consider several different
choices for the scalings and see which one works best.
Preface
This thesis is the result of my Master of Science research project at the Delft University of
Technology, The Netherlands. This project has been carried out at the faculty of Electrical
Engineering, Mathematics and Computer Science, abbreviated as EEMCS, at the chair of Math-
ematical Physics and at Rabobank International, at the department of Derivatives Research &
Validation.
Acknowledgements
I would like to express my gratitude to all those who gave me the possibility to complete this
thesis.
First, I would like to thank the department of Mathematical Physics of the TU Delft, especially
my supervisor dr. ir. W.T. van Horssen. He has helped me a lot, answering my mathematical
questions and giving comments. And I would like to thank the department for providing me a
place to work at the TU Delft.
Furthermore, I would like to thank the Modelling & Research department of Rabobank In-
ternational, in particular the department of Derivatives Research & Validation, for giving me
the opportunity to do an internship and for providing me a place to work at the Rabobank in
Utrecht. My gratitude also goes to my daily supervisor at Rabobank, drs. M. Lutterot, for help-
ing me, especially with the financial interpretation of the models, as also with understanding
the mathematical steps that are taken in the articles. Apart from that, he also helped me a lot
with correcting the English text in my report, and giving suggestions for certain sentences to
explain and interpret the mathematics in words.
Next, I would like to mention that I appreciate Patrick Hagans help a lot. He is one of the
coauthors of an article that had to be studied to complete this thesis and he was willing to
answer our questions via e-mail, giving very clear explanation for the steps and details we did
not immediately understand.
Also, I would like to thank Dennis den Ouden and Jeroen Wille for their help, especially with
LaTeX and Matlab. Because of them writing the report and programming in Matlab became
a lot less difficult. Also my colleagues at Rabobank helped me a lot by explaining financial
theory and helping me understand what happens in practice. I also thank my family and friends
for supporting me and dropping by several times, so I could get my mind off work for some time.
At last, my gratitude goes to prof. dr. ir. A.W. Heemink and prof. dr. C.W. Oosterlee from TU
Delft and drs. S. van Weeren from Rabobank, for being part of the graduation committee.
Delft, January 2010 Lieke de Jong
Contents
1 Introduction 1
2 Perturbation theory 3
3 Introduction to the f inancial framework 9
3.1 Bonds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 Brownian motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.3 Stocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.4 Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.5 Speculation and hedging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4 Financial framework 15
4.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.2 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.3 Itos formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.4 Feynman-Kac . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.5 Self-financing strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.6 Put-call parity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5 Black-Scholes model 19
5.1 Derivation of the Black-Scholes model . . . . . . . . . . . . . . . . . . . . . . . . 19
5.1.1 Hedging: Making a risk-free portfolio . . . . . . . . . . . . . . . . . . . . . 19
5.2 Heat or diffusion equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.3 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.4 Exact solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.5 Vanilla options near expiry: boundary layers and scalings . . . . . . . . . . . . . 23
5.6 Outer problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.7 Outer solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.8 Inner problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.9 Inner solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.10 Taylor expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.11 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6 CEV models 35
6.1 Derivation of the CEV model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
6.2 Scalings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6.3 Outer expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6.4 Inner expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
i
ii CONTENTS
7 Fast mean-reverting volatility 41
7.1 Derivation of the fast mean-reverting stochastic volatility model . . . . . . . . . . 41
7.2 Scalings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
7.3 Outer expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
7.4 Boundary layer analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
7.5 Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
7.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
8 The SABR model 63
8.1 Derivation of the differential equation . . . . . . . . . . . . . . . . . . . . . . . . 64
8.2 Scalings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
8.3 Application of perturbation theory to SABR model . . . . . . . . . . . . . . . . . 66
8.4 Normal model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
8.5 Lognormal model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
8.6 Stochastic model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
8.7 Special cases: = 0 and = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
8.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
9 Conclusions 91
A List of symbols 97
B Order symbols 99
C Example problem in perturbation theory: layer at x = 1 101
D Feynman-Kac 103
E Notes on L
2
functions 105
F Martingales 107
G Similarity solutions for PDEs 109
H Greens functions 113
I Hermitian adjoint 117
J Fredholm Alternative 119
K Dimensional analysis 121
L The Dirac Delta 123
M Near-identity transformation 125
Chapter 1
Introduction
Modern nancial research depends heavily on mathematics. This thesis focuses on the use
of matched asymptotic expansions in option pricing. It presents illustrations of the approach
in plain vanilla option valuation using the Black-Scholes model, using a fast mean-reverting
(stochastic) volatility model, and in the stochastic , , (SABR) model.
We begin in section 2 with explaining how to use the method of matched asymptotic
expansions to solve a general singularly perturbed problem. To explain this method, we
will study a simple (physical) example. Four steps have to be accomplished in order to complete
the application of this method:
First, an outer expansion has to be constructed. This leads to an outer solution of the
problem.
After that, the boundary layer has to be analized. For this, an inner expansion has to be
made, which leads to an inner solution of the problem.
Next, these two solutions have to be matched at their boundary, using matching conditions,
to determine the remaining constants.
Finally, we will combine these two solutions to form a composite expansion. This is done
by adding the expansions and then subtracting the part that is common to both.
Section 3 is an introduction to the computational f inance subjects in this thesis. Before
presenting a mathematical description, it is necessary to clarify some economical definitions and
nomenclature. Next, in section 4 the f inancial framework will be considered. Some formulas
and theorems that are commonly used for financial models will be explained here.
In section 5 we will consider the first financial model: the Black-Scholes model. We will
derive this basic model and show a reduction from the Black-Scholes equation to the heat equa-
tion. Because the Black-Scholes approach leads to partial dierential equations, many physical
applied mathematicians contributed to this eld. These people mainly focused on obtaining
exact solutions to certain boundary value problems representing the prices of options, or on
numerical methods. The purpose of this thesis is to illustrate another technique: asymptotic
analysis, with a particular emphasis on the use of matched asymptotic expansions. After show-
ing the exact Black-Scholes solution, perturbation theory will be applied to this model. Finally,
we will compare the results of applying perturbation theory with the exact solution.
1
2 CHAPTER 1. INTRODUCTION
The next financial model, in section 6, is the constant elasticity of variance (CEV) model.
This model will be studied following the steps taken in Howisons paper [7]. The CEV model is
a more realistic version of the ordinary Black-Scholes model in chapter 5, because studies have
shown that relative price variances do change as the stock price changes, while the Black-Scholes
model assumes a constant stock price volatility, regardless of the level of the security price.
Boundary-layer techniques can also be applied in the analysis of fast-mean-reverting stochas-
tic volatility models. We will look at this third financial model in section 7 and we will show
how to construct the boundary layer near expiry for European options. In these models the
volatility itself is assumed to follow a stochastic process while the asset price is assumed to
follow the lognormal process as before.
European options are often priced and hedged using Blacks model, or, equivalently, the Black-
Scholes model. In the Black-Scholes model there is a one-to-one correspondence between the
price of a European option and the volatility parameter
B
. Consequently, option prices are
often quoted by stating the implied volatility
B
, the unique value of the volatility which yields
the options dollar price when used in the Black-Scholes model. In theory, the volatility
B
in the
Black-Scholes model is a constant. In practice, options with dierent strikes K require dierent
volatilities
B
to match their market prices.
To resolve this problem, we derive the stochastic , , (SABR) model in section 8. The
SABR model is very well explained in the paper by Hagan et al. [5]. In reality, options with
different strikes require different volatilities to match their market prices. So the volatility is
assumed to follow a stochastic process again, and the asset price and volatility are correlated.
Singular perturbation techniques are used to obtain the prices of European options under the
SABR model. From these prices we obtain a closed-form algebraic formula for the implied vola-
tility as a function of todays forward price f and the strike K. This market volatility smile is
critical for hedging.
Chapter 2
Perturbation theory
In this chapter we will use the method of matched asymptotic expansions
1
to solve a
singularly perturbed problem. To explain this method, we will study the following example:
2
y
+ 2y
+ 2y = 0, for 0 < x < 1, (2.0.1)

with boundary conditions y(0) = 0 and y(1) = 1. Here y = y(x) and
y
=
dy
dx
and y
=
d
2
y
dx
2
.
To construct a first-term approximation of the solution for small we will proceed in four steps.
3
Step 1: Outer solution
To begin, we will assume that the solution of the above problem can be expanded in powers of
:
y
(x) =
m
n=0
n
y
n
(x) +O
_
m+1
_
= y
0
(x) +y
1
(x) +
2
y
2
(x) +. . . . (2.0.2)
Substituting this into problem (2.0.1), we obtain
_
y
0
+y
1
+. . .
_
+ 2
_
y
0
+y
1
+. . .
_
+ 2 (y
0
+y
1
+. . .) = 0. (2.0.3)
By looking at all terms without we obtain the O(1) equation y
0
+y
0
= 0. The general solution
of this O(1) equation is
y
0
(x) = ae
x
, (2.0.4)
where a is an arbitrary constant. Looking at this solution, we have a dilemma, because there is
only one arbitrary constant but we have two boundary conditions. So the solution (2.0.4) and
the expansion (2.0.2) are incapable of describing a solution over the entire interval 0 x 1.
At this moment we have no idea which boundary condition, if any, we should require y
0
(x) to
satisfy. Based on what is observed in Example 2 in Section 1.7 of [6], it is a reasonable working
hypothesis to assume that (2.0.4) describes a solution over most of the interval 0 x 1, but
there is a boundary layer at either x = 0 or at x = 1, where a different approximation must be
used. Sometimes its even possible to have several boundary layers at the same time.
1
To dene an asymptotic approximation, rst the order symbols need to be introduced. See appendix B.
2
This example is taken from [6], p.48-56.
3
See page 34 of [12] for a summary of those 4 steps.
3
4 CHAPTER 2. PERTURBATION THEORY
For now lets assume that we have a boundary layer at x = 0.
4
We will look for a solution in
that area in step 2. The solution (2.0.4) is the first term in the expansion of the outer solution.
Step 2: Boundary layer
Based on the assumption that there is a boundary layer at x = x
0
= 0, we rescale the variable
x by introducing a local variable given as
=
x x
0
()
=
x
()
. (2.0.5)
After changing variables from x to , we will take to be fixed when expanding the solution in
terms of . This has the effect of stretching the area near x
0
= 0 when becomes small, such
that we can analyze the problem locally. At this point we only know that () = o(1), and we
have no a priori knowledge of a suitable choice of ().
The equation with respect to this local variable becomes
()
2
2
y
2
+
2
()
y
+ 2y
= 0, (2.0.6)
with boundary condition y
(0) = 0. The boundary condition at x = 0 has been included here

because the boundary layer is at the left end of the interval.
Just as with the algebraic equations studied in section 1.4 of [6], it is now neccessary to determine
the correct balancing in the following equation:
()
2
2
y
+
2
()
+ 2y
= 0, (2.0.7)
which contains three terms. The balance between the second and third term was already con-
sidered in Step 1, so the following two possibilities remain:
term 1 term 3 and term 2 is of higher order in (and becomes smaller as becomes
small):
()
2
= 0 () =
.
Now term 3 is O(1) but term 2 is O(
1/2
). This violates our original assumption that
term 2 is higher order, and so this balance is not possible.
term 1 term 2 and term 3 is higher order (and becomes smaller as becomes small):
()
2
=
2
()
() = .
Now term 1 and 2 are O
_
1
_
and term 3 is O(1). In this case, the conclusions are
consistent with the original assumptions, and so this is the balancing we are looking for.
This is said to be a distinguished limit
5
for the equation.
4
In this case, choosing the boundary layer to be at x = 1 gives no solution at all, because the matching in step
3 is not possible. See appendix C.
5
Also known as maximum balance or signif icant degeneration.
5
Now the differential equation (2.0.7) becomes
2
y
+ 2

+ 2y
= 0. (2.0.8)
Because () = the appropriate expansion for the boundary-layer solution is given by
y
() y
0
() +y
1
() +. . . . (2.0.9)
Substituting this expansion (2.0.9) into equation (2.0.8) gives
2
(y
0
+y
1
+. . .) + 2

(y
0
+y
1
+. . .) + 2 (y
0
+y
1
+. . .) = 0. (2.0.10)
By taking a look at all terms of O(1), we have the following problem to solve:
_
y
+ 2y
=0, for 0 < < ,

y
0
(0) =0.
(2.0.11a)
(2.0.11b)
The general solution of this problem is y
0
() = A
_
1 e
2
_
, where A is an arbitrary constant.
In this case, it should be observed that the boundary-layer equation (2.0.11) contains at
least one term of the outer-layer equation (i.e., y
0
+ y
0
= 0) in Step 1, to have a successful
completion of the matching in Step 3. In general, this isnt always the case.
The boundary-layer expansion (2.0.9) is supposed to describe the solution in the immediate
neighbourhood of the endpoint x = 0. It is therefore not unreasonable to expect that the outer
solution (2.0.4) applies over the remainder of the interval (assuming there are no other layers).
This means that the outer solution (2.0.4) should satisfy the boundary condition at x = 1 (i.e.,
y(1) = 1). So we find that a = e
1
, which implies y
0
(x) = e
1x
.
Step 3: Matching
It remains to determine the constant A in the first-term approximation of the boundary-layer
solution y
0
() = A
_
1 e
2
_
. To do this, the approximations we have constructed so far are
summarised in figure 2.2 on page 52 of [6]. The important point here is that both the inner
(boundary-layer) and outer expansions are approximations of the same function. Therefore, in
the transition region between the inner and outer layers we should expect that the two expansions
give the same result. This is accomplished by requiring that the value of y
0
as one comes out of
the boundary-layer (as ) is equal to the value of y
0
as one comes into the boundary layer
(as x 0). Imposing the condition y
0
() = y
0
(0) yields A = e
1
and the solution becomes
y
0
() = e
_
1 e
2
_
= e
1
e
12
. (2.0.12)
This completes the derivation of the inner and outer approximations of the solution of the
example problem (2.0.1). A plot of these solutions can be seen in figure 2.1. The next step is
to combine them into a single expression (Step 4).
Figure 2.1: The outer (blue) and inner (red) solution of the example problem.
Step 4: Composite expansion
Our description of the solution consists of two pieces, which we will now combine to form a
so-called composite expansion. This is done by adding the expansions and then subtracting
the part that is common to both. Thus,
y y
0
(x) +y
0
_
x
_
y
0
(0),
e
1x
e
12x/
. (2.0.13)
It may seem strange that it is possible to combine expansions over different intervals and still
have an asymptotic approximation of the solution. However, note that the boundary-layer so-
lution y
0
(x/) is constant to first order inside the outer region. This constant is y
0
(0) and to
compensate for its contribution, the term y
0
(0) is added to the composite expansion. Simi-
larly, the outer solution y
0
(x) is also constant to first order inside the boundary-layer region.
However, the term y
0
(0) removes its contribution in this region. The fact that the adjustment
in each case involves the constant y
0
(0) is not a coincidence, since it is the first term in the
inner and outer expansions found from matching: y
0
(0) is called the common part of the
expansions. The composite expansion gives a very good approximation to the solution over
the entire interval, see figure 2.4 on page 54 of [6].
In figure 2.2 one can see the composite expansion: this is a combination of the outer (blue) and
inner (red) solution.
In the next chapters, we will apply perturbation theory to a couple of different financial models.
7
Figure 2.2: Solution of the example problem: combining the outer (blue) and inner (red) solution
to a composite expansion (bold black line).
Chapter 3
Introduction to the f inancial
framework
This chapter is an introduction to the computational finance subjects in this thesis. Before
presenting a mathematical description, it is necessary to clarify some economical definitions.
Financial market instruments can be divided into two distinct categories. There are the under-
lying stocks: shares, bonds, commodities, foreign currencies, etcetera, and their derivatives:
claims that promise some payment or delivery in the future depending on an underlying stocks
behaviour. Derivatives can reduce risk, or they can magnify it.
3.1 Bonds
First, we consider a risk-free interest rate, r, which represents the growth of money in time.
We should be able to lend at that rate, and borrow - and in arbitrary size. To model this, we
need something to model the time-value of money: a zero-coupon bond B
t
, which we can
buy or sell at time zero for some price, say B
0
. After a small time step dt it will be worth
B
dt
= B
0
+rB
0
dt = B
0
(1 +r dt), such that B
t+dt
= B
t
+rB
t
dt = B
t
(1 +r dt). Consequently,
we obtain the following differential equation for the bond:
dB
t
= rB
t
dt, (3.1.1)
such that
B
t
= B
0
e
rt
, (3.1.2)
see figure 3.1.
3.2 Brownian motion
To model more complicated financial products, a stochastic variable is needed. Brownian mo-
tions are often being used for this.
The stochastic process W = (W
t
: t 0) is a P-Brownian motion if and only if
1. W
t
is continuous and W
0
= 0,
2. the value of W
t
is distributed, under measure P, as a normal random variable N(0, t),
9
10 CHAPTER 3. INTRODUCTION TO THE FINANCIAL FRAMEWORK
Figure 3.1: Bondprice B
t
for different values of t. Here B
0
= 1 and r = 0.03.
3. the increment W
s+t
W
s
is distributed as a normal N(0, t), under P, and is independent
of T
s
, the history of what the process did up to time s.
These are all necessary and sufficient conditions for the process W to be Brownian motion.
Brownian motion with drift is called Wiener process, and is a (one-dimensional) Gaussian
process.
We know that if the increment dW
t
is a Wiener process, then dW
2
t
:= (dW
t
)
2
dt as dt 0.
1
This can be explained by proving that E[dW
1
dW
2
] = dt for dW
1
and dW
2
with correlation
.
Proof:
Suppose we have two independent normal distributed random variables Z
1
, Z
2
N(0, t). Now
we define
X
1
= Z
1
,
X
2
= Z
1
+
1 Z
2
,
such that
E[X
1
X
2
] = E
_
Z
1
(Z
1
+
_
1 Z
2
)
_
,
= E
_
Z
2
1
+
_
1 Z
1
Z
2
)
_
,
= E
_
Z
2
1
+
_
1 E[Z
1
Z
2
] ,
= Var(Z
1
) +
_
1 0,
= t. (3.2.1)
1
See also page 58-59 of [1].
3.3. STOCKS 11
So X
1
and X
2
are -correlated. In case X = X
1
= X
2
, we have = 1 and E
_
X
2
= t.
Substitution of X = dW
t
, and dW
t
N(0, dt), indeed implies that dW
2
t
dt as dt 0. And
in the more general case, we have E[dW
1
dW
2
] = dt for dW
1
and dW
2
with correlation .
3.3 Stocks
In business and finance, a share of stock (also referred to as equity share) means a share of
ownership in a corporation (company). The initial price of a stock, at time t = 0, is given by
S
0
, whereas at time t it is given by S
t
. In a small time interval dt, this price will change from S
t
to S
t
+ dS
t
. Next, consider the relative change in price:
dS
t
S
t
. We can split this relative change
into two parts:
A deterministic part
t
dt, because instead of investing it, one could also store the money
(price of the stock) at a bank, and receive interest.
An stochastic part
t
dW
t
, where dW
t
N(0, dt) is known as a Wiener process, which
we discussed in the previous section.
So we obtain the following stochastic differential equation (sde):
dS
t
S
t
=
t
dt +
t
dW
t
. (3.3.1)
The change in the stock price divided by its original value,
dS
t
S
t
, is called the return. Here
t
is the drift rate of the stock S
t
. The volatility
t
is related to the standard deviation of the
stock price of a share. It is an indication for the random behaviour of the market. The stock
price S
t
follows the lognormal distribution that arises from sde (3.3.1). See figure 3.2.
Dividend will be received by the owner of a share of some profit making company. Typically,
the stock price will decrease when dividend is paid. In all financial models described in this
thesis, we will assume absence of dividend: = 0.
3.4 Derivatives
Secondly, the main topic of this thesis option contracts and the pricing of options will be
discussed.
A stock derivative is any financial instrument that has a value that is dependent on the price
of some underlying stock. Futures and options are the two main types of stock derivatives.
A forward contract is an agreement to buy or sell an asset (i.e., a bond, stock or anything
else of value that is owned by a person or company) for a certain price at a certain time in the
future. The participants in forward contracts are the holder and the writer. The holder, who
buys the contract, is said to take a long position on the asset. The writer sells the option
and takes a short position. A forward contract is binding towards both parties: the holder is
obliged to buy the asset and the writer is obliged to sell the asset.
Figure 3.2: Stockprice S
t
for different values of t.
This is not the case in option contracts. Options give the holder the right to exercise the
option, so he is not obliged to buy the asset. The writer is however obliged to sell the asset.
There are two basic types of options: the call and the put option. A call option is an option
that gives the holder the right to buy an asset for a certain price. A put option gives the holder
the right to sell an asset for a certain price. The price mentioned in either option contract is
called the exercise price or strike K.
The most commonly traded types of calls and puts are European and American options,
which are often called vanilla options.
2
In the case of European options, the holder may only
exercise the option at the time of maturity or exercise time, T. American options may also be
exercised in the period before maturity.
3.5 Speculation and hedging
Options are used for several purposes. The two most important ones are speculation and
hedging.
Speculation is quite easy to understand: if the holder buys a call option V
call
t
at time t, he
expects the stock price to increase. The strike price is denoted by K. If the stock price S is
greater than the strike K, the call option will be exercised and the net profit of the option will
be its payof f P(S
T
) := max(S
T
K, 0) minus the option price V
call
T
:
P(S
T
) V
call
T
= S
T
K V
call
t
e
r(Tt)
. (3.5.1)
2
The opposite of a vanilla option is called an exotic option: any of a broad category of options that may
include complex financial structures.
3.5. SPECULATION AND HEDGING 13
This is called the time value of money.
The second purpose to use options is hedging. The collection of all shares, options and other
derivatives owned by a trader is called a portfolio. Hedging means using options to reduce
the risk of this portfolio. A risk-free portfolio can sometimes be constructed by choosing your
hedging parameters such that all stochastic terms are eliminated in the expression for the value
of the portfolio.
Example
Let us take a closer look at hedging and speculating by looking at an example taken from [1].
We have an interest-free bond and a stock, both initially priced at $1. At the end of the next
time interval, the stock is worth either $2 or $0.50.
Question: What is the value of a bet that pays $1 if the stock goes up?
Solution: Let B denote the bond price, S the stock price and X the payoff of the bet. Let us
dene p as the change that the stock goes up. In this example we take p =
2
3
. The picture in
figure 3.3 describes the situation.
Figure 3.3: Pricing a bet.
Buy a portfolio consisting of
2
3
of a unit of stock and a borrowing of
1
3
of a unit of bond. The
cost of this portfolio
0
at time zero is
0
=
2
3
$1
1
3
$1 = $0.33.
After an up-jump, the portfolio becomes worth
up
=
2
3
$2
1
3
$1 = $1. And after a down-
jump, it is worth
down
=
2
3
$0.5
1
3
$1 = $0.
The portfolio exactly simulates the bets payoff, and must thus be worth exactly the same as
the bet X. It must be that the portfolios initial value of $0.33 is also the bets initial value:
X = $0.33.
Restrictions
There are some restrictions in hedging, including:
Transaction costs, i.e., a cost incurred in making a nancial exchange. For example,
most people, when buying or selling a stock, must pay a commission to their broker
(a regulated professional who buys and sells shares and other securities through market
makers on behalf of investors): that commission is a transaction cost of doing the stock
deal.
Liquidity, i.e., the capacity of a market to withstand an unusual amount of buying or
selling without affecting the market substantially. It refers to an assets ability to be easily
converted through an act of buying or selling without causing a significant movement in
the price and with minimum impact on its price. Money, or cash on hand, is the most
liquid asset.
Business days, which means any day including Monday to Friday and does not include
holidays. Trading can only be done on business days. Hence, hedging cannot be done
continuously.
Chapter 4
Financial framework
In this chapter the basics of the f inancial framework, with which we can describe the prices
of options and other derivatives, will be considered. Here some formulas and theorems will be
explained. These are valid for all financial models in the next chapters.
4.1 Notation
Before starting with the formulas and theorems, we list a couple of important variables:
V
t
is the value of an option or other derivative. This value V
t
is a function of the stock
value S
t
and time t.
Later on we will consider V
call
(S
t
, t) and V
put
(S
t
, t) as the values of a call, resp. put option.
The value of the option V
t
is dependent of:

t
: the drift of the stock,

t
: the volatility of the stock,
K: strike of the option,
r
t
: (risk-free) interest rate,
S
t
: the price of the stock,
T: expiry time, t = T,
t: time.
From now on we will omit the subscripts, and write , , and r, because in most of the simple
financial models they are presumed to be constant. In case they appear to be time dependent,
this will explicitly be mentioned, or we will add the subscript again. Also, starting in the next
section, we write S instead of S
t
and V instead of V
t
, for notational convenience.
4.2 Assumptions
Arbitrage indicates that it is possible in a financial market to make risk-free profits beyond the
interest gained when placing money in a bank account. We assume that there is no arbitrage
and there are no transaction costs. Also we assume the absence of dividends. Furthermore,
hedging continuously in time is assumed to be possible.
These assumptions are valid for all financial models, unless explicitly stated otherwise.
15
16 CHAPTER 4. FINANCIAL FRAMEWORK
4.3 Itos formula
In order to calculate with stochastic processes we have to consider Itos Formula. The derivation
of Itos Formula can be found in section 3.3 in the book of Rennie and Baxter ([1]).
Denition of Itos formula
If X is a stochastic process, satisfying dX
t
= dt + dW
t
and the function f is a deterministic,
twice continuously differentiable function, then f(X
t
) is also a stochastic process. Letting
f
(X
t
) =
f
x
(X
t
), and f
(X
t
) =

2
f
x
2
(X
t
),
we obtain
df(X
t
) = f
(X
t
) dX
t
+
1
2
f
(X
t
) dX
2
t
,
= f
(X
t
) ( dt + dW
t
) +
1
2
2
f
(X
t
) dt.
Here, all terms up to and including O(dt) are taken into account, and we omit all higher order
terms. After rearranging terms, we nd that the stochastic process f(X
t
) is given by
df(X
t
) =
_
f
(X
t
) +
1
2
2
f
(X
t
)
_
dt +f
(X
t
) dW
t
. (4.3.1)
4.4 Feynman-Kac
From the sde dS = Sdt + SdW
t
, which was introduced in section 3.3, the Black-Scholes
equation
V
t
+
V
s
+
1
2
2
V
s
2
= 0 (4.4.1)
can be derived, with = (s, t), = (s, t) and V = V (s, t), and terminal condition V (s, T) =
P(s). Here, P(s) is the payof f received at expiry date t = T. This derivation will be considered
in section 5.1.
Meanwhile, we will prove that the expectation V (S
t
, t) = E
Q
_
P(S
T
)
S
t
_
is equivalent to the
solution of the Black-Scholes equation, by proving the Feynman-Kac formulas
1
.
Proof:
Using Ito calculus, first write
dV (S
t
, t) =
V (S
t
, t)
t
dt +
V (S
t
, t)
S
t
dS
t
+
1
2
2
V (S
t
, t)
S
2
t
dS
2
t
+. . . . (4.4.2)
The next step is to assume the stochastic differential equation (SDE)
dS
t
= (S
t
, t)S
t
dt +(S
t
, t)S
t
dW
t
, (4.4.3)
with dW
t
N(0, dt) a Wiener process, as introduced in section 3.2.
1
More information on the Feynman-Kac formulas can be found in appendix D.
4.5. SELF-FINANCING STRATEGIES 17
Substituting this into equation (4.4.2) yields:
dV (S
t
, t) =
_
V (S
t
, t)
t
+(S
t
, t)
V (S
t
, t)
S
t
+
1
2
2
(S
t
, t)
2
V (S
t
, t)
S
2
t
_
dt
+ (S
t
, t)
V (S
t
, t)
S
t
dW
t
. (4.4.4)
From the Black-Scholes equation (4.4.1) we know that
V (S
t
, t)
t
+(S
t
, t)
V (S
t
, t)
S
t
+
1
2
2
(S
t
, t)
2
V (S
t
, t)
S
2
t
= 0, (4.4.5)
such that in dV (S
t
, t) only the stochastic term (containing dW
t
) remains, i.e.,
dV (S
t
, t) = (S
t
, t)
V (S
t
, t)
S
t
dW
t
. (4.4.6)
For the price V (S
t
, t), we can write
V (S
t
, t) = V (S
T
, T)
_
T
t
=t
dV (S
t
, t
),
= V (S
T
, T)
_
T
t
=t
(S
t
, t
)
V (S
t
, t
)
S
t
dW
t
. (4.4.7)
Because f(W
t
) = (S
t
, t)
V (S
t
,t)
S
t
is an L
2
-function and dW
t
N(0, dt) is a Wiener process, we
have
2
E
__
T
t
f(W
t
) dW
t
S
t
= s
_
= 0. (4.4.8)
Hence, using equation (4.4.7), we can write
V (s, t) = E
_
V (S
t
, t)
S
t
= s
_
,
= E
_
V (S
T
, T)
S
t
= s
_
E
__
T
t
f(W
t
) dW
t
S
t
= s
_
,
= E
_
P(S
T
)
S
t
= s
_
. (4.4.9)
4.5 Self-f inancing strategies

Now consider a portfolio (
t
,
t
) with value
t
. At time t it contains
t
units of security, i.e.,
stock (stock price is S
t
), and also
t
units of bond (bond price B
t
). So the value of our portfolio
at time t is
t
=
t
S
t
+
t
B
t
. At the next time instance, two things happen: the old portfolio
changes value, due to the change in S
t
and B
t
, and the old portfolio has to be adjusted to give
a new portfolio as instructed by the trading strategy (
t
,
t
). If the cost of the adjustment is
perfectly matched by the profits or losses made by the portfolio then no extra money is required
from outside the portfolio is self-f inancing.
Self-f inancing property
If (
t
,
t
) is a portfolio with stock price S
t
and bond price B
t
, then:
(
t
,
t
) is self-financing d
t
=
t
dS
t
+
t
dB
t
, for all t T. (4.5.1)
2
See appendix E for notes on L
2
-functions.
18 CHAPTER 4. FINANCIAL FRAMEWORK
4.6 Put-call parity
In financial mathematics, put-call parity defines a relationship between the price of a call
option V
call
and a put option V
put
, which both have an identical strike price K and expiry time
T. For European options the put-call parity relationship will be derived. These options cannot
be exercised before expiry time T. Put-call parity can be derived in a manner that is largely
model independent.
Suppose we buy a stock S
t
and a put option V
put
, and also we sell a call option V
call
. At an
arbitrary time t our portfolio is worth
= S
t
+V
put
V
call
. (4.6.1)
At expiry time T, we make a profit of
S
T
+max(KS
T
, 0)max(S
T
K, 0) =
_
S
T
+ 0 (S
T
K) = K, for S
T
> K,
S
T
+ (K S
T
) 0 = K, for S
T
K.
(4.6.2)
Hence, the question is: What is the value of a portfolio at an arbitrary time t, if it has a
guaranteed prof it of K at t = T?
Again we assume the presence of a risk-free interest rate r. The money that we had to invest in
the portfolio could also be stored at a bank, such that we would have received the risk-free in-
terest rate r. Thats why the portfolio is worth Ke
r(Tt)
; else there would certainly be arbitrage.
Thus we have the following relationship between options and the underlying stock:
S
t
+V
put
V
call
= Ke
r(Tt)
, for all t T, (4.6.3)
which is called put-call parity.
Chapter 5
Black-Scholes model
In this chapter, we will take a look at the first financial model: the Black-Scholes model. First,
the model will be derived, and we will show that the Black-Scholes equation can be reduced into
a heat equation. After showing the exact solution, perturbation theory will be applied to the
model and finally we will compare the result with this exact solution.
5.1 Derivation of the Black-Scholes model
Substituting V (S, t) for f in Itos Formula (4.3.1) gives:
dV = S
V
S
dW
t
+
_
S
V
S
+
1
2
2
S
2
2
V
S
2
+
V
t
_
dt. (5.1.1)
Now we first need to construct a portfolio.
5.1.1 Hedging: Making a risk-free portfolio
Consider a portfolio that contains units of stock S at time t, i.e., we are short units of
stock, which means we have sold them. Also we buy one option V , i.e., we are long one option
V . The value of our portfolio at time t is = V S. If the value of our portfolio changes
d during a small time step dt, we have
d = dV dS S d
..
=0
= dV dS. (5.1.2)
Here d = 0, because the number of stocks in the portfolio can only be changed at the end of
every time step dt, not during the time step itself.
Substitution of equation (5.1.1) into (5.1.2), using Itos formula (4.3.1), we find that d satisfies:
d = S
_
V
S

_
dW
t
+
_
S
V
S
+
1
2
2
S
2
2
V
S
2
+
V
t
S
_
dt. (5.1.3)
To make sure we have a risk-free portfolio, we should eliminate the stochastic term: the one
containing dW
t
. So we take =
V
S
, such that the change in value of the portfolio becomes
d =
_
S
_
V
S

_
+
1
2
2
S
2
2
V
S
2
+
V
t
_
dt =
_
1
2
2
S
2
2
V
S
2
+
V
t
_
dt. (5.1.4)
By storing your money at a bank, instead of investing it, one can receive a risk-free interest rate
r, just like we have seen before, looking at the bond B
t
in Chapter 3.1. Then the amount of
19
20 CHAPTER 5. BLACK-SCHOLES MODEL
money (which could have been the value of your portfolio) will produce rdt after a small
time step dt, without any risk. So substitution of d = rdt into equation (5.1.4) yields
rdt =
_
1
2
2
S
2
2
V
S
2
+
V
t
_
dt. (5.1.5)
After dividing equation (5.1.5) by dt and substitution of = V S with =
V
S
into equation
(5.1.5), we obtain the well-known Black-Scholes equation
V
t
+
1
2
2
S
2
2
V
S
2
+rS
V
S
rV = 0. (5.1.6)
5.2 Heat or dif fusion equation
We can reduce the Black-Scholes equation (5.1.6) into an equation of the form
u
=

2
u
x
2
, (5.2.1)
which is known as the heat equation or dif fusion equation.
For the call option we have
V
call
t
+
1
2
2
S
2
2
V
call
S
2
+rS
V
call
S
rV
call
= 0, (5.2.2)
with boundary conditions: V
call
(0, t) = 0, V
call
(S, t) S as S and at t = T we have a
terminal condition: V
call
T
:= V
call
(S
T
, T) = max(S K, 0).
This is a backward equation, with non-constant coefficients. To make it a forward equation, we
measure time backwards from expiry and scale it with , writing
1
t = T

1
2
2
=
1
2
2
(T t). (5.2.3)
Also we substitute S = Ke
x
and V
call
= Kv(x, ). Now the stock price S
t
, which followed a
log-normal distribution, becomes x, which is normally distributed. Also we scale S and V
call
with K to make x and v(x, ) dimensionless. After these substitutions, we obtain
v
=

2
v
x
2
+ (k 1)
v
x
kv, (5.2.4)
with k = r/
1
2
2
and initial condition v(x, 0) = max(e
x
1, 0).
This already looks a lot more like the heat equation (5.2.1). The next step is to substitute
v(x, ) = e
x+
u(x, ) for some constants and .
Substituting this into differential equation (5.2.4) yields
u +
u
=
2
u + 2
u
x
+

2
u
x
2
+ (k 1)
_
u +
u
x
_
ku. (5.2.5)
1
Deviding t by
2
is done to make dimensionless. From the sde dS = S dt + S dW
t
it follows that
[] =
_
1
t
_
.
5.3. MODEL 21
In order to let all terms with u and
u
x
cancel, we have to solve the following system for and
:
_
=
2
+ (k 1) k,
0 = 2 + (k 1).
(5.2.6)
The solution of system (5.2.6) is given by
_
=
1
2
(k 1),
=
1
4
(k + 1)
2
,
such that v(x, ) = e
1
2
(k1)x
1
4
(k+1)
2
u(x, ) and this indeed gives us the heat equation

u
=

2
u
x
2
, for < x < , > 0, (5.2.7)
with initial condition u(x, 0) = u
0
(x) = max
_
e
1
2
(k+1)x
e
1
2
(k1)x
, 0
_
.
5.3 Model
In [7] Howison first considers an asset (for example: a stock) whose price is S. This stock price
S can be modeled as a function of time t by the stochastic differential equation (sde)
dS = S dt +S dW
t
, (5.3.1)
in which dW
t
is the increment of a standard Brownian Motion
2
and and are, respectively,
the drift and volatility of the asset (taken to be constant).
Just like we have seen in the previous sections, we can set up a hedge portfolio = V S,
where the choice =
V
S
makes the portfolio risk-free again.
In absence of arbitrage and transaction costs, the portfolio earns a (constant) risk-free rate r,
so that d = rdt. This leads to the well-known Black-Scholes equation
V
t
+
1
2
2
S
2
2
V
S
2
+rS
V
S
rV = 0. (5.3.2)
This backward parabolic equation is to be solved with a terminal condition V (S, T) = P(S),
where P(S) is the payoff received at the expiry date t = T.
An alternative view of the hedging strategy is that it entails pricing with respect to a probability
measure Q that is risk-neutral, rather than the objective (observed) measure P associated with
the sde for the asset price cited above in (5.3.1). That is, if the asset is assumed to follow the
sde (5.3.1), then the discount value of the option at time t is given by
V (S
t
, t)
B
t
= E
Q
_
P(S
T
)
B
T
S
t
_
, (5.3.3)
which is a martingale
3
.
2
See section 3.2.
3
See appendix F.
Here, we will posit the existence of a deterministic r, and , such that the bond price B
t
and
the stock price S
t
follow
B
t
= B
0
e
rt
, (5.3.4)
S
t
= S
0
e
t+W
t
, (5.3.5)
where r is the deterministic risk-free interest rate, is the stock volatility and is the stock drift.
For the option price V (S
t
, t) this yields
V (S
t
, t) = B
t
E
Q
_
P(S
T
)
B
T
S
t
_
,
=
B
t
B
T
E
Q
_
P(S
T
)
S
t
_
,
= e
r(Tt)
E
Q
_
P(S
T
)
S
t
_
. (5.3.6)
In practice, the hedging strategy above is impractical and in particular it is impossible to hedge
continuously in time, or at each t. For this reason we introduce the hedge parameter gamma,
=

2
V
S
2
. (5.3.7)
This gamma is a measure of the risk incurred in rehedging at non-infinitesimal time intervals
t. To see this, consider a portfolio = V
t
S, with
t
=
V
S
evaluated at (S, t), that is
perfectly hedged at time t. Assume that no trading takes place over the interval (t, t +t).
Using Taylors theorem, the change in the portfolio over this interval is = V
t
S. In
the infinitesimal limit where t dt, is equal to the risk-free return rdt. Over the finite
interval we have a difference between the return on the portfolio and the risk-free interest rate.
This difference is called the hedging error, and is given by
_
e
rt
1
_
=
V
t
t +
1
2
2
V
S
2
S
2
r
_
V
V
S
S
_
t +o (t) ,
=
1
2
2
S
2
_
W
2
t
_

2
V
S
2
+o (t) . (5.3.8)
To obtain this hedging error, we have used the Black-Scholes equation (5.1.6) and the fact that
to this order S
2
=
2
S
2
W
2
, where W N (0, t) is the small change in a Wiener process W.
4
The hedging error (5.3.8) seems to be proportional to the random variable W
2
t, whose
expectation is zero, multiplied by the options gamma =

2
V
S
2
.
For call options we have
(S, T) =
d
2
dS
2
max (S K, 0) = (S K) , (5.3.9)
where () is the Dirac delta function. For put options we have the same gamma.
For call options without dividends []
S=0
= 1 for all t, such that
_
0
(S, t) dS = 1. As t T
the gamma of such an option is an approximation of the delta function.
4
See section 3.2.
5.4. EXACT SOLUTION 23
5.4 Exact solution
As a first example we consider a call option (or, by put-call parity
5
, a put option) in the
standard Black-Scholes model. The payoff is given by P(S) = max(S K, 0) and there is a
famous explicit formula for the option value,
V (S, t) = SN(d
+
) Ke
r(Tt)
N(d
), (5.4.1)
where
6
d
=
log(
S
K
) +
_
r
1
2
2
_
(T t)
T t
, (5.4.2)
and
N(x) =
1
2
_
x
e
s
2
/2
ds (5.4.3)
is the standard normal cumulative density function.
5.5 Vanilla options near expiry: boundary layers and scalings
In this section a rst application of perturbation theory on a nancial model will be presented
to show the method and nd an approximation. Because an exact solution is available, we
are able to compare the results.
Boundary layers
For small time, we can derive an approximation to the exact solution (5.4.1) as follows. We
will construct two boundary layers: one when the stock price is near the strike price, i.e. near
S = K, and one when time t is close to expiry time T, i.e. near t = T.
These articial (boundary) layers are illustrated in figure 5.1. We have constructed them
to show the method of asymptotic expansions applied to this basic financial model.
Note that we will only consider the region inside the layer near maturity, i.e. when T t is
small. In this case, the other regions are not taken into account. In particular we will look
inside a layer around S = K. Here, we have an inner expansion and an inner solution. In
the rest of the layer near t = T, i.e., when the spot S is far from the strike K, we have an outer
expansion and outer solution.
7
Scalings
First of all, we measure time backwards from expiry and scale it with
2
to make it dimensionless,
writing
t = T
t
2
t
= (T t)
2
. (5.5.1)
After this scaling the Black-Scholes equation (5.1.6) transforms into
V
t
=
1
2
S
2
2
V
S
2
+S
V
S
V, (5.5.2)
5
See section 4.6.
6
Note that traditionally d
1
and d
2
are used instead of d
+
and d
.
7
Here, Howisons nomenclature is quite unusual, but practical: the outer solution is inside one of the boundary
layers. The inner solution is located in the overlap of the two layers.
with =
r
2
dimensionless. Suppose that = O(1) and scaled time is small, such that = t
with 0 < 1. This has the eect of stretching the area near t = T. At this point the value
of scaling parameter is not yet known; we will determine it later on.
The scaled Black-Scholes equation (5.5.2) now becomes
1
=
1
2
S
2
2
V
S
2
+S
V
S
V. (5.5.3)
Figure 5.1: In this case there are 2 artificial layers: around t = T and S = K.
5.6 Outer problem
In figure 5.1 it is shown that, when the spot S is far from the strike K, we have a regular
outer expansion
V
(S, ) =
m
n=0
n
V
n
(S, ) +O
_
(m+1)
_
= V
0
(S, ) +
V
1
(S, ) +. . . , (5.6.1)
which gives
(V
0
(S, ) +
V
1
(S, ) +. . .) =
1
2
S
2

2
S
2
_
V
0
(S, ) +
V
1
(S, ) +. . .
_
+
S

S
_
V
0
(S, ) +
V
1
(S, ) +. . .
_

_
V
0
(S, ) +
V
1
(S, ) +. . .
_
.
The O(1) equation becomes
V
0
= 0, and for O(
) we have
8
V
1
=
1
2
S
2

2
V
0
S
2
. .
=0
+
_
S
V
0
S
V
0
_
=
_
S
V
0
S
V
0
_
,
8
If the spot S is far from the strike K, note that in either region V (hence, also V
0
) is linear.
5.7. OUTER SOLUTION 25
i.e.,
V
1
=
_
(S (S K)) = K, for S K,
0, for S K.
The term
1
2
S
2
2
V
0
S
2
in the O(
) equation vanishes, because for [S K[
K (i.e., if the spot

is far from the strike) we have =

2
V
S
2
= 0. This can be seen in figure 5.2.
Figure 5.2: European call option payoff max(S
t
K, 0) at different times t < T before maturity
(blue lines) and at maturity (red line) t = T (here: strike K = 30).
5.7 Outer solution
The solution of the O(1) outer equation with final condition V
0
(S, T) = P(S) = max(S K, 0)
is given by V
0
(S, t) = max(S K, 0). Next, solving the O(
) equation with final condition

V
1
(S, T) = 0, yields
V
1
(S, t) =
_
K, for S K,
0, for S K,
such that the expansion becomes
V
= V
0
+
V
1
=
_
S K +
K, for S K,
0, for S K.
(5.7.1)
which we can rewrite as
V
0
+
V
1
=
_
S K(1
), for S K
K, far above the strike,

0, for S K
K, far below the strike.

(5.7.2)
Note that in this case, we have a layer of width
.
Later on, after applying maximum balance, turns out to be equal to 2. This implies that the
previous expression (5.7.2) becomes equivalent to the first two terms in the small time expansion
in unscaled variables of the function
_
S Ke
r(Tt)
, for S > Ke
r(Tt)
,
0, for S < Ke
r(Tt)
.
(5.7.3)
The components of the function (5.7.3) are the value of the forward contract in which the option
holder is compelled to buy the asset, corresponding to certain exercise, and zero corresponding
to no exercise.
5.8 Inner problem
However, as remarked earlier and as can be seen in figure 5.2, we expect large gamma near the
strike K. Hence, the term containing the second derivative with respect to S cannot be ignored.
We deal with this by rescaling if S is near the strike price K. So, let us introduce a local inner
variable
x =
S K
K
, (5.8.1)
such that we have scaled S K with
and we divide by K to make x dimensionless. The value

of will be determined later, after applying maximum balance.
Also we introduce a scaling
v(x, ) =
V (S, )
K
, (5.8.2)
such that we scale V (S, ), which is O(
), with
to have v = O(1) and again we devide by K

to make v(x, ) dimensionless.
After this second scaling, the scaled Black-Scholes equation (5.5.3) becomes the dimensionless
equation
1
=
1
2
2
(1 +
x)
2
2
v
x
2
+

(1 +
x)
v
x
v, (5.8.3)
and the payoff is
9
v(x, 0) =
V (S, 0)
K
=
max(S K, 0)
K
= max(x, 0). (5.8.4)
Next, we will calculate the solution to this inner problem, and after that we will try to match it
with the outer solution (5.7.2).
First, assume and to be integer, such that the solution can be expanded in integer powers:
v
(x, ) = v
0
(x, ) +v
1
(x, ) +O
_
2
_
, (5.8.5)
which gives
1
(v
0
(x, ) +v
1
(x, ) +. . .) =
1
2
2
(1 +
x)
2

2
x
2
(v
0
(x, ) +v
1
(x, ) +. . .)
+

(1 +
x)

x
(v
0
(x, ) +v
1
(x, ) +. . .)
(v
0
(x, ) +v
1
(x, ) +. . .) . (5.8.6)
9
On page 5 of [7] Howison claims v(x, 0) = max(x, 0), which is incorrect.
5.9. INNER SOLUTION 27
Maximum balance
10
over the 3 terms of equation (5.8.6) gives = 2. A nice integer solution
for this constraint is choosing = 2 and = 1, which yields
(v
0
(x, ) +v
1
(x, ) +. . .) =
1
2
(1 +x)
2

2
x
2
(v
0
(x, ) +v
1
(x, ) +. . .)
+ (1 +x)

x
(v
0
(x, ) +v
1
(x, ) +. . .)

2
(v
0
(x, ) +v
1
(x, ) +. . .) . (5.8.7)
The O(1) problem becomes
v
0
=
1
2
2
v
0
x
2
, with v
0
(x, 0) = max(x, 0). (5.8.8)
For x the conditions for this equation are, consistently with the payoff,
lim
x+
v
0
(x, )
x
= 1,
lim
x
v
0
(x, ) = 0.
If x , S will stay inside the layer, but it will go towards the lower boundary of the layer,
because
S
K
1 . In that case v(x, ) will go to zero.
5.9 Inner solution
The inner problem is much simpler than the original problem. We will use one of the five basic
invariance properties of the diffusion equation to choose a local variable correctly.
Five basic invariance properties of the dif fusion equation
Suppose we have the initial value problem
_
u
t
= ku
xx
,
u(x, 0) = (x).
(5.9.1a)
(5.9.1b)
Our method is tantamount to solving the initial value problem (5.9.1a)-(5.9.1b) for a particular
(x) and then build the general solution from this particular one. Therefore we will use the
following five basic invariance properties:
11
1. The translate u(x y, t) of any solution u(x, t) of the pde (5.9.1a) is another solution for
any fixed y.
Proof :
t
u(x y, t) =

t
u(x, t) and

2
x
2
u(x y, t) =

2
x
2
u(x, t).
2. Any derivative (u
x
or u
t
, or u
xx
, etc.) of a solution u(x, t) of the pde (5.9.1a) is again a
solution, because u(x, t) C
for t > 0.
Proof : Suppose that u(x, t) satises the heat equation u
t
= ku
xx
. Dene y := Tu(x, t),
where T is an operator that can be any derivative of u(x, t). Then, y
t
= (Tu)
t
= T(u
t
) =
T(ku
xx
) = k(Tu)
xx
= ky
xx
, because derivatives can be interchanged using u(x, t) C
.
Hence, any derivative y = Tu(x, t) of any solution u(x, t) also satises the heat equation.
10
See chapter 2. This is also referred to as signif icant degeneration or distinguished limit.
11
See section 2.4 of [11].
3. A linear combination of solutions u
i
(x, t) of the heat equation (5.9.1a) is again a solution
of (5.9.1a).
Proof : By linearity:
t
_
i
u
i
(x, t)
_
=
t
u
i
(x, t) and

2
x
2
_
i
u
i
(x, t)
_
=
2
x
2
u
i
(x, t).
4. An integral of solutions of the pde (5.9.1a) is again a solution.
Proof : If S(x, t) is a solution of problem (5.9.1a), then so is the translate S(x y
i
, t) for
any constant y
i
, by property 1. and so is S(x y
i
, t)g(y
i
). Multiplication by the constant
y := y
i+1
y
i
yields that also S(xy
i
, t)g(y
i
)y
i
is a solution of the pde (5.9.1a). Thus,
summation over all y
i
and letting y
i
0 implies that
v(x, t) =
_

S(x y, t)g(y) dy,

is a solution of the heat equation (5.9.1a) for any function g(y), as long as this improper
integral converges appropriately. In fact, property 4. is just a limiting form of property 3.
5. If u(x, t) is a solution of the pde (5.9.1a), so is the dilated function u(
ax, at), for any

a > 0.
Proof : We can prove this using the chain rule:
Let v(x, t) := u(
ax, at). Then we have v

t
=
_
(at)
t
_
u
t
= au
t
and v
x
=
_
(
ax)
x
_
u
x
and
v
xx
=
au
xx
= au
xx
.
Using the dilated function u(
ax, at) in property 5. we know that the pde in (5.8.8) has

similarity solution
12
f
_
x
_
:=
v
0
(x, )
. (5.9.2)
Also we introduce a local variable
=
x
. (5.9.3)
After substituting the transformations (5.9.2) and (5.9.3) into the O(1) heat equation (5.8.8),
and multiplying by 2
, we obtain the following ode in f():

f
+f
f = 0, (5.9.4)
with boundary conditions
lim
f = 0,
lim
= 1.
(5.9.5)
The solution can be found by differentiating equation (5.9.4) once with respect to ,
d
d
_
f
() +f
() f()
_
= 0,
d
d
_
f
() +f
()
_
= 0, (5.9.6)
12
See appendix G.
which gives us an ordinary differential equation (ode) in f
. The solution of ode (5.9.6) is

given by
f
() = Ae
1
2
2
= c
2
n(). (5.9.7)
This is again a differential equation, which has a solution that is given by
f() = c
2
N() +c
2
n() +c
1
+c
0
. (5.9.8)
From the boundary conditions (5.9.5) we find that c
1
= c
0
= 0 and c
2
= 1, such that
f() = N() +n(). (5.9.9)
In the original variables, the solution is given by
v
0
(x, ) = xN
_
x
_
+
n
_
x
_
, (5.9.10)
where N() is as above in (5.4.3) and n() is its derivative e
x
2
/
2.
This solution can also directly be obtained by application of Greens functions.
13
The approximation we found in equation (5.9.10) and equation (H.0.21) is valid in the inner
region, while in the outer region we have the outer expansion (5.7.2). In more complicated prob-
lems one can often find a uniformly valid expansion, holding in both inner and outer regions, by
calculating outer+innercommon, in which outer and inner are the expansions we already
found, and common is the intermediate limiting behavior of these expansions used in matching.
In our case the outer expansion is so simple that it and the common expansion coincide, and
so the inner expansion is in fact uniformly valid and can be used as an approximation for all S
and small t
.
So a 1-term approximation is given by
v
(x, ) = v
0
(x, ) +O(),
= x N
_
x
_
+
n
_
x
_
=
S/K 1
N
_
S/K 1
T t
_
+

T t n
_
S/K 1
T t
_
.
In the original variables, this expression is
V (S, t) V
0
(S, t) K v
0
(x, )
= (S K) N
_
S/K 1
T t
_
+
T tK n
_
S/K 1
T t
_
. (5.9.11)
Note that the parameter , which is artificial, does not appear in the expression.
In figure 5.3 one can see the approximation (1-term inner expansion) we found in equation
(5.9.11) compared to the exact solution. As can be seen, both approximations are very close to
the exact solution.
Figure 5.3: The approximation with the 1-term expansion (green) versus the exact solution
(red).
Figure 5.4: The difference between approximation with 1-term expansion and the exact solution.
To see how close this approximation exactly is, we can subtract the exact solution, and make a
plot of the discrepancy between the approximation and the exact solution, see figure 5.4.
Using equation (5.8.7) again, the O() problem becomes
v
1
=
1
2
2
v
1
x
2
+x
2
v
0
x
2
+
v
0
x
, with v
1
(x, 0) = 0. (5.9.12)
As v
0
satisfies the diffusion equation
t
v
0
1
2
2
x
v
0
= 0, again we will use some of the the five
basic invariance properties of the diffusion equation to find:
If u

1
2
u
xx
= 0 and v

1
2
v
xx
= u, then a particular solution is u
p
(x, ) = u.
If u is as above and v

1
2
v
xx
= xu, then a particular solution is v
p
(x, ) = xu +
1
2
2
u
x
.
N
() +N
() = 0.
So a particular solution is readily found and, as it vanishes at = 0, it is the solution we need:
v
1
(x, ) = x
2
v
0
x
2
+
1
2
3
v
0
x
3
+
v
0
x
,
=
1
2
x
n
_
x
_
+ N
_
x
_
. (5.9.13)
For the 2-term approximation of v, this gives
v(x, ) v
0
(x, ) +v
1
(x, ),
= x N
_
x
_
+
n
_
x
_
+
_
1
2
x
n
_
x
_
+ N
_
x
__
,
= (x +) N
_
x
_
+
(1 +
1
2
x) n
_
x
_
. (5.9.14)
Furthermore the 2-term inner expansion V (S, ) = Kv(x, ) K(v
0
+v
1
) is again uniformly
valid. In the original variables, the expression reads
V (S, t)
_
S K +rK(T t)
_
N
_
S/K 1
T t
_
+
T t(S +K) n
_
S/K 1
T t
_
. (5.9.15)
Figure 5.5 presents the approximation (with the 2-term inner expansion) we found in (5.9.15)
compared to the exact solution and the approximation (1-term inner expansion) we found in
(5.9.11). As can be seen, both approximations are very close to the exact solution.
To see how close the approximations exactly are, we can subtract the exact solution, and make
a plot of the error, see figure 5.6. This figure makes clear that the 2-term expansion (5.9.15)
indeed is a better approximation than the 1-term expansion (5.9.11), because the error (blue
line) is much closer to zero.
13
The steps we need to take for this, are being explained in [11], on the pages 47 52. See appendix H.
Figure 5.5: The approximation with the 2-term expansion (blue) versus 1-term expansion (green)
and the exact solution (red).
Figure 5.6: The difference between approximation with the 2-term expansion (blue) versus
1-term expansion (green) and the exact solution.
5.10. TAYLOR EXPANSION 33
5.10 Taylor expansion
After substitution of = T t we can show that these expressions (5.9.11) and (5.9.15) agree
to O
_
2
_
with the small-time expansion of the exact solution below:
V (S, ) = S N(d
+
) Ke
r
N(d
), (5.10.1)
with
d
=
log(
S
K
) +
_
r
1
2
2
_

, (5.10.2)
and
N(x) =
1
2
_
x
e
s
2
/2
ds. (5.10.3)
The Taylor expansion of V (S, ) around = 0 is given by
V
Taylor
(S, ) =
n=0

n
n!
n

n
V
Taylor
=0
, (5.10.4)
= V
Taylor
=0
+

V
Taylor
=0
+
1
2

2

2

2
V
Taylor
=0
+O
_

3
_
. (5.10.5)
For the first term, V
Taylor
=0
, we have
V
Taylor
=0
=
_
S K, for S > K,
0, for S K,
(5.10.6)
because this is just the payoff function at expiry.
For the second term,

V
Taylor
=0
, we find

V
Taylor
=0
=
V

(S, 0) =
V
d
+
d
+

+
V
d

,
=
_
rK, for S > K,
0, for S K.
Hence, the Taylor expansion (5.10.5) becomes
V
T
(S, ) = V
Taylor
=0
+

V
Taylor
=0
=
_
S K(1 r) for S > K,
0, for S K.
(5.10.7)
In original variables this is
V
Taylor
=0
(S, t) =
_
S K +rK(T t) for S > K,
0, for S K.
(5.10.8)
The Taylor expansion (5.10.8) agrees up to O(
2
) with the 1-term and 2-term expansions we
found in (5.9.11) and (5.9.15).
5.11 Discussion
In this section a rst application of perturbation theory on a nancial model has been presented
in order to show the techniques and complications of the method of asymptotic expansions in a
nancial context. In the simple example of the Black-Scholes model an exact solution is avail-
able, so that it is possible to compare results.
First, we have constructed two articial layers: a boundary layer at the option maturity date
and an internal layer at the strike price. Next, the method of asymptotic expansions has been
applied, in order to nd a solution of the Black-Scholes equation (5.3.2). Comparison between
the resulting asymptotic expansion and the exact solution is very useful: we can check if the
results are consistent. It also indicates how accurate the approximation is.
Here, applying perturbation theory actually gives the same result as making a Taylor series of
the exact solution around = T t = 0. The more accurate the expansion (i.e., higher order
terms are taken into account), the better the approximation will become.
On page 8 of his paper [7] Howison is showing some gures containing this sections results.
Here, he plots the approximate call value minus the exact value as a function of moneyness
S/K for dierent times to expiry. In general, the approximation is remarkably good for these
practical parameter values, because the error is very small. Near S = K (or S/K = 1) the
error is a bit larger. This can be explained by the size of the risk factor =

2
V
S
2
, which is a
Dirac delta function around S = K. Far from the strike, the option price is linear in S if the
option is far in the money (itm), and zero if the option is far out of the money (otm). In these
cases the second derivative is zero. At S = K, when the option is at the money (atm) we have
an innite . In this case, the second derivative

2
V
S
2
contributes to the pde for the option price V .
Our opinion
Generally speaking, we think that the result of applying perturbation theory on the Black-
Scholes model is not very useful. For this simple model an exact solution is available, so an
asymptotic approximation is not necessary. Besides, the approximation will never be as accu-
rate as the exact solution. However, constructing these articial layers is an instructive way
to show the method and to understand how an asymptotic expansion analysis can be done.
Moreover, the results can be compared, because this exact solution is available. Due to this it
is possible to observe the accuracy of the method.
This can be seen as a rst step in applying perturbation theory on nancial models. After
understanding this, more complicated models can be considered.
Chapter 6
CEV models
In this chapter, we will consider a second financial model: the constant elasticity of variance
(CEV) model, which is an improved version of the Black-Scholes model.
Using the CEV functions, we can calculate the theoretical price, sensitivities, and implied vola-
tility of options, by applying a valuation technique based on the constant elasticity of variance
option pricing model. With this model we consider the possibility that the volatility of the
underlying asset depends on the price of the underlying asset.
This model is more realistic than the ordinary Black-Scholes model, because studies have shown
that price variances do indeed change as the stock price changes, while the Black-Scholes model
assumes a constant stock price volatility, regardless of the level of the stock price.
6.1 Derivation of the CEV model
Let us change the stochastic differential equation (sde) in (5.3.1) into
dS = S dt + S
/2
dW
t
, where ,= 2. (6.1.1)
Here
= S
1
2
0
, (6.1.2)
with S
0
= S(0) the stock price at t = 0, to make sure that the dimensions of S
2
0
_
S
S
0
_
are
correct (euro
2
, instead of euro
).
So, for the change in the option price dV we have, using Itos formula (4.3.1),
dV = S
/2
V
S
+
_
S
V
S
+
1
2

2
S
2
V
S
2
+
V
t
_
dt. (6.1.3)
Again we construct a portfolio = V S, which - using sde (6.1.1) and equation (6.1.3) -
gives
d = dV dS,
= S
/2
V
S
+
_
S
V
S
+
1
2

2
S
2
V
S
2
+
V
t
_
dt Sdt S
/2
dW
t
,
= S
/2
_
V
S

_
dW
t
+
_
S
_
V
S

_
+
1
2

2
S
2
V
S
2
+
V
t
_
dt.
35
36 CHAPTER 6. CEV MODELS
Once again we choose =
V
S
to eliminate the stochastic dW
t
term. This gives
d =
_
1
2

2
S
2
V
S
2
+
V
t
_
dt. (6.1.4)
Storing money at a bank account gives a risk-free interest rate r, so d = rdt. Substituting
this into (6.1.4) yields
rdt =
_
1
2

2
S
2
V
S
2
+
V
t
_
dt. (6.1.5)
Using = V S with =
V
S
gives
r
_
V
V
S
S
_
dt =
_
1
2

2
S
2
V
S
2
+
V
t
_
dt, (6.1.6)
which, after division by dt and reordering the terms, becomes the CEV equation
V
t
+
1
2

2
S
2
V
S
2
+rS
V
S
rV = 0. (6.1.7)
In original variables, the CEV equation (6.1.7) is given by
V
t
+
1
2
2
S
2
0
_
S
S
0
_
2
V
S
2
+rS
V
S
rV = 0. (6.1.8)
The difference from the previous Black-Scholes model is that the volatility is now S-dependent
for ,= 2. The model is used to represent a leverage effect whereby the impact of a given
stochastic change dW
t
is assumed to be greater when the asset price is small than when it is
large, and thus is assumed to be less than the Black-Scholes value: < 2.
The parameters S
0
and are not independent, but writing volatility at a given price level (for
example: S
0
= K) gives the same at-the-money
1
volatility for options with strike K as varies.
Calculating explicit solutions is thus much less straightforward, but the asymptotic procedure
is virtually the same as for the case = 2 (the ordinary Black-Scholes model).
6.2 Scalings
As before we measure time backwards from expiry and scale it with
2
, writing t = T t
/
2
such that t
= T t
2
. After this scaling again we have
V
t
=
1
2
S
2
2
V
S
2
+S
V
S
V, (6.2.1)
with =
r
2
dimensionless. Suppose that = O(1) and scaled time is small, such that = t
/
2
with 0 < 1.
6.3 Outer expansion
For the CEV equation (6.2.1) above, this gives
V
=
2
1
2
S
2
2
V
S
2
+
2
S
V
S

2
V, (6.3.1)
which is exactly the same as the previously used outer expansion (see section 5.6).
1
An option is at the money if the strike price K of the option is equal to the market price S of the underlying
security.
6.4. INNER EXPANSION 37
6.4 Inner expansion
For the inner problem, we first introduce a local variable
x =
S K
K
, (6.4.1)
such that x is dimensionless, and S K is scaled by .
Also we introduce a time rescaling
v(x, ) =
V (S, )
K
, (6.4.2)
such that v is dimensionless, and V (S, ) is scaled with , where =
(Tt)
2
2
.
For the derivatives in the replaced CEV equation (6.1.8) this gives
V
t
=
V
t
=
2
V
2
(Kv)
2
K
,
V
S
=
V
x
x
S
=
1
K
V
x
=
1
K
(Kv)
x
=
1
K
v
x
,
S
=
1
K
x

2
V
S
2
=
1
K
2
v
x
2
.
Thus the replaced CEV equation (6.1.8) transforms into
2
K
=
1
2
2
S
2
0
(1 +x)
1
K
2
v
x
2
+r(1 +x)K
v
x
Krv,
=
1
2
_
S
0
K
_
2
(1 +x)
2
v
x
2
+
r
2
(1 +x)
v
x

2
r
2
v,
=
1
2
_
S
0
K
_
2
(1 +x)
2
v
x
2
+(1 +x)
v
x

2
v,
such that the inner equation becomes
v
=
1
2
2
(1 +x)
2
v
x
2
+(1 +x)
v
x

2
v. (6.4.3)
Here
2
= (S
0
/K)
2
, and we have the same payoff as before: v(x, 0) = max(S K, 0).
First we rescale time by setting
=
2
, and again we expand
v
(x,
) = v
0
(x,
) +v
1
(x,
) +O
_
2
_
, (6.4.4)
which gives
_
v
0
(x,
) +v
1
(x,
) +. . .
_
=
1
2
2
(1 +x)

2
x
2
_
v
0
(x,
) +v
1
(x,
) +. . .
_
+ (1 +x)

x
_
v
0
(x,
) +v
1
(x,
) +. . .
_

2
_
v
0
(x,
) +v
1
(x,
) +. . .
_
, (6.4.5)
Because this appears to be the same problem as the previous one for v
0
, in (5.9.12), we will
obtain an inner solution that is similar to the one in (5.9.13), namely,
v(x, ) v
0
(x, ) +v
1
(x, ) = (x +) N
_
x
_
+
_
1 +
1
4
x
_
n
_
x
_
. (6.4.6)
When = 2, the approximation (6.4.6) reduces to the previous expression (5.9.14).
In original variables, we have the approximation
V (S, t) = Kv(x, ) K
_
v
0
(x, ) +v
1
(x, )
_
,
= K
_
(x +) N
_
x
_
+
_
1 +
1
4
x
_
n
_
x
__
,
=
_
S K +rK(T t)
_
N
_
(S/K 1)
T t
_
,
+
T t
_
S + (4 )K
4
_
n
_
(S/K 1)
T t
_
. (6.4.7)
In gure 6.1 the CEV values for a call option are presented for different values of < 2. The
absolute and relative differences between the = 2 case (ordinary Black-Scholes model) and
CEV values with < 2 can be seen in gure 6.2 and 6.3 respectively.
Figure 6.1: CEV values (call option) for different < 2
6.4. INNER EXPANSION 39
Figure 6.2: Absolute difference between the = 2 case (ordinary Black-Scholes model) and
CEV values with < 2.
Figure 6.3: Relative difference (%) between the = 2 case (ordinary Black-Scholes model) and
CEV values with < 2.
Chapter 7
Fast mean-reverting volatility
Boundary-layer techniques can be applied in the analysis of fast mean-reverting stochastic
volatility models. In these models the volatility itself is assumed to follow a stochastic process
while the asset price is assumed to follow the lognormal process as before.
1
_
dS
t
=
t
S
t
dt +
t
S
t
dW
t
,
d
t
= M
t
dt +
t
d
W
t
.
(7.0.1a)
(7.0.1b)
Here M
t
and
t
are the drift and the volatility of the volatility
t
respectively. The instantaneous
correlation between the Brownian motions W
t
and

W
t
is denoted by , i.e.,
2
E
_
dW
t
d
W
t
_
= dt.
A well-known example of a stochastic volatility model is the Heston model, for which
d(
2
t
) = (
t

) dt +
t
d
W
t
, (7.0.2)
for constants > 0, and
.
Leaving out the stochastic part (containing d
W
t
), we have
d
2
t
dt
=
_
(
t

) 0, for
t

,
, , > 0, for
t
<
,
(7.0.3)
such that in the long term the volatility
t
will always return to its equilibrium value
. That is
why this model is called mean-reverting. The fast mean-reversion is caused by the parameter
, which we will assume to be quite large. This will be explained in terms of time scales in
section 7.2.
7.1 Derivation of the fast mean-reverting stochastic volatility
model
Suppose we have the following model:
3
_
dS = rS dt +S dW
t
,
d = (M ) dt + d
W
t
,
(7.1.1a)
(7.1.1b)
1
Therefore we write
t
instead of in this case.
2
See chapter 3.2.
3
Note that we omit the subscripts from here on, for notational convenience.
41
42 CHAPTER 7. FAST MEAN-REVERTING VOLATILITY
where is the market price of volatility risk
4
, and we know that E
_
dW
t
d
W
t
_
= dt.
The value V of an option is now a function of the stock price S, time t and volatility :
V = V (S, , t). Using Itos formula (4.3.1) we obtain up to and including O(dt) the differential
dV =
V
t
dt +
1
2
2
V
S
2
dS
2
+
1
2
2
V
2
d
2
+

2
V
S
ddS +
V
d +
V
S
dS. (7.1.2)
Substituting dS and d as given in definition (7.1.1a) and (7.1.1b) resp. into equation (7.1.2),
yields
dV =
_
V
t
+
1
2
2
S
2
2
V
S
2
+
1
2
2
V
2
+S

2
V
S
+ (M )
V
+rS
V
S
_
dt
+
V
W
t
+
V
S
SdW
t
. (7.1.3)
Again we will set up a portfolio .
5
Because now we have two stochastic parts (due to two
Brownian motions dW
t
and d
W
t
), we have two hedge parameters and
1
, such that
= V S
1
V
1
. (7.1.4)
So our portfolio contains the option whose value is V (S, , t), a quantity of the stock S,
and a quantity
1
of another asset whose value V
1
depends on the stock price S, volatility
and time t. Note that the quantities and
1
can also become negative in case you sell stock
(going short).
The change d in this portfolio in a time dt is given by
d = dV dS
1
dV
1
, (7.1.5)
=
_
V
t
+
1
2
2
S
2
2
V
S
2
+
1
2
2
V
2
+S

2
V
S
+ (M )
V
+rS
V
S
_
dt
+
V
W
t
+
V
S
SdW
t
(rSdt +SdW
t
)
1
_
V
1
W
t
+
V
1
S
SdW
t
_

1
_
V
1
t
+
1
2
2
S
2
2
V
1
S
2
+
1
2
2
V
1
2
+S

2
V
1
S
+ (M )
V
1
+rS
V
1
S
_
dt.
To make the portfolio instantaneously risk-free, all terms containing dW
t
and d
W
t
must vanish.
That is true if
V
S

1
V
1
S
= 0, and
V

1
V
1
= 0,
i.e.,
1
=
_
V
_
_
V
1
_, and (7.1.6a)
=
V
S

_
V
_
_
V
1
_
V
1
S
. (7.1.6b)
4
A definition and explanation of this concept will be given later on in this chapter.
5
Based on the method used in [3] on p. 4-6.
7.1. DERIVATIONOF THE FAST MEAN-REVERTINGSTOCHASTIC VOLATILITYMODEL43
Substituting
1
and as above into equation (7.1.5), yields
d =
_
V
t
+
1
2
2
S
2
2
V
S
2
+
1
2
2
V
2
+S

2
V
S
_
dt
_
V
_
_
V
1
_
_
V
1
t
+
1
2
2
S
2
2
V
1
S
2
+
1
2
2
V
1
2
+S

2
V
1
S
_
dt.
The fact that the return on a risk-free portfolio must equal the risk free rate r, which we assume
to be deterministic, implies that d = rdt. Or explicitely, following the steps taken in [3],
and collecting all V -dependent terms on the left-hand side and all V
1
-dependent terms on the
right-hand side, we obtain
V
t
+
1
2
2
S
2
2
V
S
2
+
1
2
2
2
V
2
+S

2
V
S
+rS
V
S
rV
V
=
V
1
t
+
1
2
2
S
2
2
V
1
S
2
+
1
2
2
2
V
1
2
+S
2
V
1
S
+rS
V
1
S
rV
1
V
1
. (7.1.7)
Because the left-hand side of equation (7.1.7) is explicitly independent of V
1
and the right-hand
side is explicitly independent of V , either side must be independent of both V and V
1
. The
only way that this can be true is for both sides to be equal to some function of the independent
variables S, and t. We deduce that
V
t
+
1
2
2
S
2
2
V
S
2
+
1
2
2
V
2
+S

2
V
S
+rS
V
S
rV
=
_
(M )
_
V
, (7.1.8)
where, without loss of generality, we have written the arbitrary function of S, and t as
_
(M )
_
, where (M ) and are the drift and volatility functions from the
sde (7.1.1b) for instantaneous variance. This approach is analougos to the one in [3]. Here,
= (S, , t) is an arbitrary function.
The market price of volatility risk
To see why is called the market price of volatility risk, we will consider the following
portfolio, consisting of a -hedged (but not
1
-hedged) option V :
1
= V
V
S
S. (7.1.9)
After applying Itos formula (4.3.1) we obtain
d
1
= dV
V
S
dS,
=
_
V
t
+
1
2
2
S
2
2
V
S
2
+
1
2
2
V
2
+S

2
V
S
+ (M )
V
_
dt
+
V
W
t
. (7.1.10)
Note that all terms with dW
t
vanish, because the option is -hedged.
Again we expect to have d
1
= r
1
dt, such that d
1
r
1
dt = 0. However, because we did
not
1
-hedge the portfolio, we have
d
1
r
1
dt =
_
V
t
+
1
2
2
S
2
2
V
S
2
+
1
2
2
V
2
+S

2
V
S
+ (M )
V
_
dt
+
V
W
t
r
_
V
V
S
S
_
dt,
=
_
V
t
+
1
2
2
S
2
2
V
S
2
+
1
2
2
V
2
+S

2
V
S
+ (M )
V
+rS
V
S
rV
_
dt +
V
W
t
. (7.1.11)
Using equation (7.1.8) we find
d
1
r
1
dt =
_
(M )
_
V
dt + (M )
V
dt +
V
W
t
,
=
V
dt +
V
W
t
=
V
_
dt + d
W
t
_
,= 0. (7.1.12)
Taking the expectation to get rid of the d
W
t
term yields
E(d
1
r
1
dt) =
dt. (7.1.13)
We see that the extra return per unit of volatility risk d
W
t
scales linearly with , which is known
as the market price of volatility risk.
Hence, the price V (S, , t) satisfies the backward parabolic equation
V
t
+
1
2
2
S
2
2
V
S
2
+S

2
V
S
+
1
2
2
V
2
+rS
V
S
+ (M )
V
rV = 0, (7.1.14)
of which the Black-Scholes equation is a special case.
7.2 Scalings
We assume for simplicity that the drift and volatility of the volatility are functions of only ,
not of S or t. Let us consider the commonly occurring situation in which the volatility process
is fast mean-reverting, which means that the timescale for mean-reversion is much shorter than
that for the evolution of the asset price, their ratio
r
M
being small, such that we can apply
perturbation theory.
We now introduce the scaled variables
m := M M =
m
and :=
, (7.2.1)
such that

t
= M is large when M =
m
is large.
6
This causes the fast mean-reversion.
6
Note that M = O(
1
), which implies m = O(1), and note that = O(

1
), which implies = O(1).

7.3. OUTER EXPANSION 45
The relative sizes of these coefficients are chosen such that has a nontrivial invariant distri-
bution. We denote this time-independent invariant distribution by
p
() := lim
t
p(, t[
0
, 0),
where p(, t[
0
, 0) is the transition density function for at time t, conditional on
0
at time
zero, which satisfies the forward Kolmogorov equation
7
p
t
=

_
(M )p
_
+
1
2
2
_
2
p
_
. (7.2.2)
In terms of the original variables, this reads
p
t
=

__
m
_
p
_
+
1
2
2
_
p
_
. (7.2.3)
After applying the above scalings, the pricing equation becomes
V
t
+
1
2
2
S
2
2
V
S
2
+

S

2
V
S
+
1
2
2
V
2
+rS
V
S
+
_
m
_
V
t
rV = 0. (7.2.4)
7.3 Outer expansion
First, we write the pricing equation (7.2.4) for the fast mean-reverting process in the form
_
L
0
+
L1
2
+L
1
_
V = 0, (7.3.1)
where
L
0
=
1
2
2

2
2
+m

,
L1
2
= S

2
S

,
L
1
=

t
+
1
2
2
S
2

2
S
2
+rS

S
r.
Note that L
1
is the Black-Scholes operator
8
with volatility . Because L
0
is the generator of the
backward Kolmogorov equation for , its adjoint L
0
is the generator of the forward equation.
These Kolmogorov equations govern the evolution in time of the transition density function p.
Hence, p
() satisfies the forward Kolmogorov equation, with

p
t
= 0, because p
() does
not depend on t. Therefore, p
() satisfies
L
0
p
=
d
2
d
2
_
1
2
2
p
d
d
(mp
) = 0. (7.3.2)
Integrating once yields
1
2
d
d
_
2
p
()
_
mp
() = c
0
. (7.3.3)
7
We found this forward Kolmogorov equation using the method on p. 291 of [10]. This is very different from
the equation Howison writes down on page 14 of [7].
8
The Black-Scholes equation reads L
1
V = 0.
The homogeneous version of this equation is given by
1
2
d
d
t
_
2
p
hom
(
t
)
_
mp
hom
(
t
) = 0, (7.3.4)
which can be rewritten as
d
d
_
2
p
hom
()
_
2m
2
_
2
p
hom
()
_
= 0, (7.3.5)
such that
p
hom
() =
2c
1
2
exp
__

0
2m(s)
2
(s)
ds
_
. (7.3.6)
As a particular solution we take p
to be constant:
p
part
() := c
2
, (7.3.7)
because it has to look like the inhomogeneous part of the integrated Kolmogorov equation (7.3.3).
After substituting this particular solution into equation (7.3.3), we find c
2
=
c
0
m
.
The general solution is thus given by
p
() = p
hom
() +p
part
() =
2c
1
2
exp
__

0
2m(s)
2
(s)
ds
_
+
c
0
m
, (7.3.8a)
with p
() satisfying the property

_

() d = 1, (7.3.8b)
because p
() is a transition density function.

Howison does not explicitly show his solution in [7], but he only remarks that it is proportional
to
1
2
()
exp
_
2
_

m(s)
2
(s)
ds
_
,
assuming that
2
and m are such that p
exists. This is very different, because we did not find

the minus sign in the exponent, and also the integration constant
c
0
m
is missing here.
We now expand
V (S, , t) V
0
(S, , t) +
1
2
V1
2
(S, , t) +V
1
(S, , t) +
3
2
V3
2
(S, , t) +
2
V
2
(S, , t) +. . . .
After substituting this into the pricing equation (7.3.1), and collecting coefficients of equal
powers of together, we obtain
O(1) : L
0
V
0
= 0,
O(
) : L
0
V1
2
+L1
2
V
0
= 0,
O() : L
0
V
1
+L1
2
V1
2
+L
1
V
0
= 0,
O
_
3
2
_
: L
0
V3
2
+L1
2
V
1
+L
1
V1
2
= 0,
O
_
2
_
: L
0
V
2
+L1
2
V3
2
+L
1
V
1
= 0.
We will first solve the O(1) equation L
0
V
0
= 0 for V
0
, i.e.,
1
2
2
V
0
2
+m
V
0
= 0, (7.3.9)
which we can rewrite as
2
V
0
2
=
2m
2
V
0
. (7.3.10)
Integrating once yields
V
0
= c(S, t) exp
__

0
2m(s)
2
(s)
ds
_
. (7.3.11)
After another integration we find
V
0
(S, , t) =
_

0
c(S, t) exp
__

0
2m(s)
2
(s)
ds
_
d +c
0
(S, t),
= c(S, t)
_

0
exp
__

0
2m(s)
2
(s)
ds
_
d +c
0
(S, t), (7.3.12)
such that we can write this solution V
0
as
V
0
(S, , t) = c(S, t)g() +c
0
(S, t). (7.3.13)
Here, V
0
(S, , t) satisfies the following terminal condition at t = T:
V
0
(S, , T) = max(S K, 0). (7.3.14)
Also, we know that the solution V
0
satisfies the following limits
9
lim
S
K
V
0
S
= 1, and lim
S
K
0
V
0
S
= 0. (7.3.15)
Next, Howison uses the disputable
10
assumption that V
0
is independent of , to obtain c(S, t) = 0
and thus V
0
(S, , t) = c
0
(S, t).
The O(
) equation is given by
L
0
V1
2
+L1
2
V
0
= 0, (7.3.16)
with initial condition V1
2
(S, , 0) = 0.
Since V
0
= c
0
(S, t) is a function of S and t only, we have L1
2
V
0
= 0, such that the O(
)
equation (7.3.16) becomes
L
0
V1
2
= 0. (7.3.17)
This implies that V1
2
is independent of , and therefore
V1
2
= c1
2
(S, t). (7.3.18)
Note that if Howison would not have done the above assumption about the -independence of
V
0
, the O(
) equation would be completely dierent, because then the term L1

2
V
0
would not
be zero. Consequently, it would have been a lot more complicated to obtain the O(
) solution
9
See [10], p. 289.
10
See discussion in section 7.6.
V1
2
(S, , t).
Because L1
2
contains derivatives with respect to only, the O() equation
L
0
V
1
+L1
2
V1
2
+L
1
V
0
= 0, (7.3.19)
thus becomes
L
0
V
1
= L
1
V
0
, (7.3.20)
with initial condition V
1
(S, , 0) = 0.
To solve the O() equation (7.3.20) we use the inner product with some arbitrary function f
that is not yet determined:
< L
0
V
1
, f >= < L
1
V
0
, f > . (7.3.21)
This can also be written as
< V
1
, L
0
f >= < L
1
V
0
, f >, (7.3.22)
where L
0
is the Hermitian adjoint
11
of L
0
.
Because we know the dynamics of p
and the fact that L
0
p
= 0, we choose f = p
to obtain
< V
1
, 0 >= < L
1
V
0
, f >, (7.3.23)
i.e.,
< L
1
V
0
, p
>= 0. (7.3.24)
Since L
1
is an operator, we can write down the eigenvalue equation
L
1
i
=
i
i
, (7.3.25)
where
i
is an eigenfunction, and
i
is its corresponding eigenvalue. In this case we know that
there is an eigenvalue
i
= 0, because we have < L
1
V
0
, p
>= 0. By the Fredholm alter-

native
12
, we know that there is either no solution V
0
possible, or there are infinitely many
solutions, which implies that V
0
cannot be determined uniquely.
Define L
i
V
j
:=< L
i
V
j
, p
> for i = 0,
1
2
, 1 and j = 0, 1, 2, . . .. For i = 1 and j = 0 and using
equation (7.3.24) we obtain
L
1
V
0
=
_

L
1
V
0
p
() d,
=
_

_
V
0
t
+
1
2
2
S
2
2
V
0
S
2
+rS
V
0
S
rV
0
_
p
() d = 0. (7.3.26)
Here, the only -dependence is in the volatility coefficient
2
, so in all other terms the integral
vanishes:
< 1, p
>=
_

() d = 1, (7.3.27)
because p
is a probability distribution function. Hence,

L
1
V
0
=
c
0
t
+
1
2
2
S
2
2
c
0
S
2
+rS
c
0
S
rc
0
= 0, (7.3.28)
11
For determining this Hermitian adjoint L
0
, see appendix I.
12
See appendix J.
with
2
:=<
2
, p
>=
_

2
p
() d, (7.3.29)
representing the average of
2
with respect to the invariant distribution p
.
Equation (7.3.28) is the Black-Scholes equation with averaged variance
2
. The solution can be
calculated by standard Black-Scholes techniques once a final condition is given.
First of all, if the volatility in the Black-Scholes equation is a given deterministic function (t)
of time, the option value can be calculated by replacing in the relevant constant-volatility
formula by
_
1
T t
_
T
t
2
(s) ds
_
1
2
. (7.3.30)
Next, we can calculate V
1
. Because L
1
V
0
= 0, we have, by subtraction,
L
1
V
0
= L
1
V
0
L
1
V
0
=
1
2
_
2
_
S
2
2
c
0
S
2
. (7.3.31)
Here we can eliminate the majority of the terms in the right-hand side of equation (7.3.20),
leaving
L
0
V
1
=
1
2
_
2
_
S
2
2
c
0
S
2
. (7.3.32)
Because S
2
2
c
0
S
2
and eigenfunction c
1
(S, t) are independent of , and L
0
contains only derivatives
with respect to , this can be written as
L
0
_
V
1
c
1
S
2
2
c
0
S
2
_
=
1
2
_
2
_
, (7.3.33)
Hence, we know that the solution has the form
V
1
(S, , t) = g
1
()S
2
c
0
S
2
+c
1
(S, t), (7.3.34)
where g
1
() satisfies
L
0
g
1
=
1
2
_
2
_
, (7.3.35)
i.e.,
1
2
2
g
1
2
+m
g
1
=
1
2
_
2
_
. (7.3.36)
This can be rewritten as
2
g
1
2
+h()
g
1
= j(), (7.3.37)
with
h() :=
2m
2
, and j() :=

2
2
. (7.3.38)
Writing G
1
:=
g
1
, we have the following differential equation in terms of G

1
:
G
1
+h()G
1
= j(). (7.3.39)
Multiplying both sides by e

0
h(s) ds
yields
e

0
h(s) ds
G
1
+h()G
1
e

0
h(s) ds
= j() e

0
h(s) ds
. (7.3.40)
This can be rewritten as
_
e

0
h(s) ds
G
1
_
= j() e

0
h(s) ds
. (7.3.41)
Solving this differential equation for e

0
h(s) ds
G
1
, gives
e

0
h(s) ds
G
1
=
_

0
j( ) e

0
h(s) ds
d +H
1
(S, t), (7.3.42)
such that
V
1
= G
1
= e

0
h(s) ds
_

0
j( ) e

0
h(s) ds
d +H
1
(S, t)e

0
h(s) ds
, (7.3.43)
After integrating once we find
V
1
(S, , t) =
_

0
_
e

0
h(s) ds
_

0
j( ) e

0
h(s) ds
d +H
1
(S, t)e

0
h(s) ds
_
d +H
2
(S, t).
(7.3.44)
This is the integral form of the solution. One of these two complementary solutions, namely
H
2
(S, t), does not depend on and can be absorbed into eigenfunction c
1
(S, t). The other one
is unbounded at infinity. Hence, we choose the particular solution v
1
(S, , t) to be equal to zero.
Before proceeding any further, we outline the pattern followed by succesive iterations of the
solution procedure. For all n 0 we have the n
th
equation
L
0
V
n
= L1
2
V
n+
1
2
L
1
V
n1
for V
n
(S, , t). (7.3.45)
The right-hand side is assumed to be known from earlier stages. Solving this, we obtain a par-
ticular solution v
n
(S, , t) and an eigenfunction c
n
(S, t). For the case n = 0 the former is zero
and the latter is c
0
(S, t).
Next, we repeat this process for V
n+
1
2
, obtaining a further particular solution v
n+
1
2
and a further
eigenfunction c
n+
1
2
. Note that the eigenfunction c
n
(S, t) does not depend on and is annihilated
by the operator L
0
. Finally, we substitute the functions just found into the right-hand side of
the equation for V
n+1
:
L
0
V
n+1
= L
n+
1
2
V
n+
1
2
L
1
V
n
for V
n+1
(S, , t). (7.3.46)
Because the eigenfunction c
n+
1
2
(S, t) does not depend on , it is annihilated by the operator
L1
2
. Therefore, the right-hand side is known in terms of the particular solutions just obtained.
Here, the solvability condition L
0
V
n+1
=< L
0
V
n+1
, p
>= 0 for existence of a solution can be

applied. From equation (7.3.46) it follows that
L
0
V
n+1
= L
n+
1
2
V
n+
1
2
L
1
V
n
, (7.3.47)
i.e.,
0 = L
n+
1
2
v
n+
1
2
L
1
v
n
L
1
c
n
. (7.3.48)
This can be written as
L
1
c
n
= L
n+
1
2
v
n+
1
2
L
1
v
n
. (7.3.49)
We now apply this procedure to the case n =
1
2
. We have already found V1
2
and V
1
up to
eigenfunctions c1
2
(S, t) resp. c
1
(S, t), so we need only apply the solvability condition to the
equation
L
0
V3
2
= L1
2
V
1
L
1
V1
2
for V1
2
(S, , t). (7.3.50)
This yields
L
0
V3
2
= L1
2
V
1
L
1
V1
2
, (7.3.51)
i.e.,
0 = L1
2
V
1
L
1
V1
2
. (7.3.52)
Using the fact that eigenfunction c
1
(S, t) is annihilated by operator L1
2
, we obtain
L
1
V1
2
= L1
2
v
1
. (7.3.53)
Substituting v
1
(S, , t) = g
1
()S
2
2
c
0
S
2
, we obtain
L
1
V1
2
= L1
2
g
1
()S
2
2
c
0
S
2
,
= S
dg
1
d
S
_
S
2
2
c
0
S
2
_
+
dg
1
d
S
2
2
c
0
S
2
. (7.3.54)
For ease of notation, we set
T := S

S
(7.3.55)
equivalent to using a logarithmic price variable.
13
Defining
A1
2
,1
:=
dg
1
d
and A1
2
,2
:=
dg
1
d
, (7.3.56)
we obtain
L
1
V1
2
=
_
A1
2
,1
+A1
2
,2
T
_
_
T
2
T
_
c
0
. (7.3.57)
Note that this is a partial differential equation with constant coefficients. Next, we use that
L
1
c
0
= 0 and the fact that the time derivative of a solution of the homogeneous equation is
again a solution of the homogeneous equation. Also, we assume pro tem that the correct final
condition (from matching) is V1
2
(S, T) = 0. We find
14
V1
2
(S, , t) = (T t)
_
A1
2
,1
+A1
2
,2
T
_
_
T
2
T
_
c
0
. (7.3.58)
where c
0
is already known. As noted earlier, V1
2
is independent of .
This result, and elaborations thereof, is an important practical consequence of the method, since
it allows calibration of the three constants
2
,
dg
1
d
and
dg
1
d
to market prices of options
(as represented by an implied volatility surface) in a simple manner: the key point is that only
these directly deducable constants are needed, rather than the unobservable functions M(, t)
13
This can be explained by setting x := log S. Then,

x
=
S
x
S
= e
x
S
= S

S
.
14
Howison made a small mistake here, forgetting the minus sign.
and (, t).
We first calculate
L
0
V3
2
= L1
2
V
1
L
1
V1
2
, (7.3.59a)
using
L1
2
V
1
=
_
T
_

_
g
1
()
_
T
2
T
_
c
0
(S, t) +c
1
(S, t)
_
,
=
dg
1
d
( T)
_
T
2
T
_
c
0
(S, t). (7.3.59b)
and
L
1
V1
2
= L
1
V1
2
+
1
2
_
2
_
_
T
2
T
_
V1
2
=
_
A1
2
,1
+A1
2
,2
T
_
_
T
2
T
_
c
0
+
1
2
_
2
_
_
T
2
T
_
c
0
. (7.3.59c)
From this, it follows that
L
0
V3
2
=
_
dg
1
d
( T) +
1
2
_
2
_
(T t)
_
A1
2
,1
+A1
2
,2
T
_
_
T
2
T
_
A1
2
,1
A1
2
,2
T
_
_
T
2
T
_
c
0
(S, t), (7.3.60)
such that V3
2
has the form
V3
2
(S, , t) =
_
g
2
() +g
3
()T +g
1
()(T t)
_
A1
2
,1
+A1
2
,2
T
_
_
T
2
T
_
_
_
T
2
T
_
c
0
. (7.3.61)
Here, g
1
(), g
2
() and g
3
() satisfy
1
2
2
d
2
g
1
d
2
+m
dg
1
d
=
1
2
_
2
_
, (7.3.62a)
1
2
2
d
2
g
2
d
2
+m
dg
2
d
=
1
2
_
2
_
+
dg
1
d
A1
2
,1
, (7.3.62b)
1
2
2
d
2
g
3
d
2
+m
dg
3
d
=
1
2
_
2
_
dg
1
d
A1
2
,2
. (7.3.62c)
Then, the solvability condition for V
2
, applied to the equation
L
0
V
2
= L1
2
V3
2
L
1
V
1
, (7.3.63)
gives
L
0
V
2
= L1
2
V3
2
L
1
V
1
, (7.3.64)
i.e.,
0 = L1
2
V3
2
L
1
V
1
, (7.3.65)
We know that V
1
can be written as a particular solution v
1
(S, , t), plus an eigenfunction c
1
(S, t).
L
1
c
1
= L1
2
V3
2
L
1
v
1
= L1
2
V3
2
, (7.3.66)
recalling that v
1
(S, , t) = 0.
7.4. BOUNDARY LAYER ANALYSIS 53
Hence,
L
1
c
1
= L1
2
V3
2
=
_
T
_
_
dg
2
d
+
dg
3
d
T +
dg
1
d
(T t)
_
A1
2
,1
+A1
2
,2
T
_
_
T
2
T
_
_
_
T
2
T
_
c
0
=
_
A
1,1
+A
1,2
T +
_
A1
2
,1
+A1
2
,2
T
_
(T t)
_
A1
2
,1
+A1
2
,2
T
_
_
T
2
T
_
_
_
T
2
T
_
c
0
,
where
15
A
1,1
:=
_
T
_
dg
2
d
and A
1,2
:=
_
T
_
dg
3
d
. (7.3.67)
The relevant particular solution c
1
(S, t) with zero payoff c
1
(S, T) = 0 can be obtained in a
similarly way. Again, we use that L
1
V
1
= 0 and the fact that the time derivative of a solution
of the homogeneous equation is again a solution of the homogeneous equation. This yields that
the particular solution c
1
(S, t) is equal to
_
(T t) (A
1,1
+A
1,2
T) +
1
2
(T t)
2
_
A1
2
,1
+A1
2
,2
T
_
T
2
T
_
__
A1
2
,1
+A1
2
,2
T
_
_
_
T
2
T
_
c
0
.
However, we leave open the possibility of adding a further solution c
1
(S, t), if the payoff, de-
termined by matching into the boundary layer, dictates that we should do so. Similarly, the
-dependence of the solution can only be resolved by matching.
7.4 Boundary layer analysis
Reasoning by means of analogies to chapter 5, let us introduce a boundary layer in t near t = T
of size O(), defining the inner time variable via
t = T + =
t T
, for < 0. (7.4.1)

Note that here < 0, opposite the t = T in the Black-Scholes calculations in chapter 5,
where > 0.
The pricing equation for V (S, , ) reads
1
+
1
2
2
S
2
2
V
S
2
+

S

2
V
S
+
1
2
2
V
2
+rS
V
S
+
_
m
_
V
rV = 0. (7.4.2)
Again we can write this in the form
_
L
0
+
L1
2
+
L
1
_
V (S, , ) = 0, (7.4.3)
using the operators
L
0
=

+L
0
=

+
1
2
2

2
2
+m

L1
2
= L1
2
= S

2
S

L
1
=
1
2
2
S
2

2
S
2
+rS

S
r.
15
Howison claims that the definitions should be A
1,1
:= g
2
()( ) and A
1,2
:= g
3
()( ).
Note that

L
0
, unlike L
0
, contains the time derivative

. Next, we expand
V
(S, , ) = V
0
+
V1
2
+V
1
+. . . . (7.4.4)
From this equation, the following O(1) equation can be determined:
L
0
V
0
= 0, (7.4.5a)
i.e.,
V
0
+
1
2
2
V
0
2
+m
V
0
= 0, (7.4.5b)
0
(S, , 0) = P(S).
Following the disputable assumption
16
in Howisons paper [7], that V
0
is independent of , only
V
0
= 0 (7.4.6)
is left, which yields that V
0
is a function of S only. The initial condition implies that this
function is the payoff. Hence,
V
0
(S, , ) = P(S). (7.4.7)
Note that for this first order inner approximation matches automatically with the
one-term outer solution V
0
(S, t) as t T. The procedure of matching will be discussed in the
next section.
The O(
) equation is given by
L
0
V1
2
+

L1
2
V
0
= 0, (7.4.8)
with initial condition V1
2
(S, , 0) = 0.
17
Using the fact that V
0
does not depend on yields
L1
2
V
0
= S

2
V
0
S

V
0
= 0, (7.4.9)
such that
L
0
V1
2
= 0, (7.4.10)
which is the same differential equation as the O(1) equation. This time the initial condition
V1
2
(S, , 0) = 0 gives
V1
2
(S, , t) = 0. (7.4.11)
The O() equation is given by
L
0
V
1
+

L1
2
V1
2
+

L
1
V
0
= 0, (7.4.12)
1
(S, , 0) = 0. Using the solutions we found above, i.e., V
0
= P(S) and
V1
2
= 0, this reduces to
L
0
V
1
=
L
1
P(S). (7.4.13)
16
This issue was already discussed in the previous section 7.3 of this thesis.
17
On page 20 of [7] Howison states that the initial condition is given by V1
2
(S, T) = 0, which must be a
typographical error, because we are working in (S, , )-coordinates now.
7.4. BOUNDARY LAYER ANALYSIS 55
Writing
L
0
V
1
=
V
1
+L
0
V
1
, (7.4.14)
gives
V
1
+L
0
V
1
=
1
2
2
S
2
2
P(S)
S
2
rS
P(S)
S
+rP(S). (7.4.15)
Because
P
= 0, we can write

1
2
2
S
2
2
P(S)
S
2
rS
P(S)
S
+rP(S) (7.4.16)
for the right hand side. This can also be written as
18
1
2
_
2
_
S
2
2
P
S
2
L
1
P. (7.4.18)
Equation (7.4.15) thus becomes
V
1
+L
0
V
1
=
1
2
_
2
_
S
2
2
P
S
2
L
1
P. (7.4.19)
Although it is not immediately clear which steps are taken to obtain this, according to Howison,
a particular solution is
V
1
= g
1
()S
2
2
P
S
2
L
1
P +1
1
(S), (7.4.20)
where 1
1
is an arbitrary function. This indeed is a solution of problem (7.4.19) with initial
condition V
1
(S, , 0) = 0. In fact, this is the correct form for the asymptotic behaviour of V
1
as , see figure 7.1. To see this, first note that applying the inner product to equation
(7.4.19) yields
<
V
1
+L
0
V
1
, p
>=<
1
2
_
2
_
S
2
2
P
S
2
L
1
P, p
> . (7.4.21)
Figure 7.1: As , it goes towards the left boundary of the layer (away from maturity T).
This gives
< V
1
, p
>= L
1
P, (7.4.22)
18
Recall that when computing the outer solution, we defined
L
1
V :=< L
1
V, p
>=
V
+
1
2
2
S
2
2
V
S
2
+rS
V
S
rV, with
2
:=<
2
, p
> . (7.4.17)
because < L
0
V
1
, p
>= 0, < L
1
P, p
>= L
1
P and
_

1
2
_
2
_
S
2
2
P
S
2
d =
1
2
S
2
2
P
S
2
_
_
_
_
_
2
_

d
. .
=1
2
p
d
. .
=:
2
_
_
_
_
_
= 0. (7.4.23)
And so, integrating equation (7.4.22) and using the initial condition V
1
(S, , 0) = 0,
< V
1
, p
>= L
1
P. (7.4.24)
Using the solution (7.4.20) found for V
1
, we also have that
< V
1
, p
> = < g
1
()S
2
2
P
S
2
L
1
P +1
1
(S), p
>,
= < g
1
(), p
> S
2
2
P
S
2
L
1
P +1
1
(S),
= g
1
()S
2
2
P
S
2
L
1
P +1
1
(S). (7.4.25)
Comparing these two expressions (7.4.24) and (7.4.25) for the inner product, we conclude that
1
1
(S) = g
1
()S
2
2
P
S
2
. (7.4.26)
Thus, as ,
V
1
V
1
=
_
g
1
() g
1
()
_
S
2
2
P
S
2
L
1
P, (7.4.27)
since what is left after subtracting the particular solution, V
1
V
1
, satisfies the homogeneous
version of the parabolic equation, has initial data that vanishes at large and small , and
therefore vanishes as . This is the first step in the matching, because here we consider
what happens if we go to the boundary of the layer ( ). Further mathching will be done
in the next section 7.5.
7.5 Matching
We can now complete the matching. From the outer expansion, written in inner variables, we
have
V
:= V
0
(S, , T +) +
1
2
V1
2
(S, , T +) +V
1
(S, , T +) +. . .
V
0
(S, , T) +
V
0
t
(S, , T) +
1
2
0 +V
1
(S, , T),
= V
0
(S, , T) +
_
V
0
t
(S, , T) +V
1
(S, , T)
_
,
= c
0
(S, T) +
_
c
0
t
(S, T) +g
1
()S
2
2
c
0
S
2
+c
1
(S, T)
_
. (7.5.1)
Note that the particular solutions in V1
2
and V
1
that are multiplied by T t do not contribute
to this, because they are O
_
3
2
_
.
7.5. MATCHING 57
One-term matching of the outer and inner expansion yields
lim
tT
c
0
(S, t) = lim
V
0
(S, , ) c
0
(S, T) = P(S) (7.5.2)
and
L
1
c
0
= 0
c
0
t
+
1
2
2
S
2
2
c
0
S
2
+rS
c
0
S
rc
0
= 0,
c
0
t
=
1
2
2
S
2
2
c
0
S
2
+rS
c
0
S
rc
0
. (7.5.3)
The inner expansion yields V
0
(S, , t) = P(S), such that one-term matching this with the outer
expansion gives c
0
(S, t) = P(S). This implies that
c
0
t
=
L
1
P(S), (7.5.4)
where

L
1
is the inner product of the operator as it was defined formulating the inner expansion:
L
1
P(S) :=

L
1
P(S) =<

L
1
P(S), p
() >=
1
2
2
S
2
2
P
S
2
+rS
P
S
rP. (7.5.5)
Replacing c
0
(S, T) by P(S) and
c
0
t
(S, T) by
L
1
P, the three-term outer expansion written in
inner variables (7.5.1) can be rewritten as
V
= P(S) +
_

L
1
P +g
1
()S
2
2
c
0
S
2
+c
1
(S, T)
_
. (7.5.6)
As demonstrated in section 7.4, the large- behaviour of the three-term inner expansion is
P(S) +
_
L
1
P +
_
g
1
() g
1
()
_
S
2
2
P
S
2
_
. (7.5.7)
Matching these two expressions, the missing final condition for c
1
(S, t) is
c
1
(S, T) = g
1
()S
2
2
P
S
2
. (7.5.8)
Let us define w := S
2
2
c
0
S
2
. We can rewrite

L
1
c
0
= 0 as
1
2
2
S
2
2
c
0
S
2
+rS
c
0
S
rc
0
= 0, (7.5.9)
i.e.,
S
2
2
c
0
S
2
=
2r
2
_
rS
c
0
S
rc
0
_
. (7.5.10)
So we can replace w by
2r
2
_
rS
c
0
S
rc
0
_
, such that
L
1
w =
1
2
2
S
2
2
w
S
2
+rS
w
S
rw,
=
1
2
2
S
2

2
S
2
_
2r
2
_
rS
c
0
S
rc
0
__
+rS
w
S
rw,
= rS
2

S
_
S
2
c
0
S
2
+
c
0
S

c
0
S
_
+rS
w
S
rw,
= rS
2

S
_
w
S
_
+rS
w
S
rw,
= rS
2
1
S
2
w rS
2
1
S
w
S
+rS
w
S
rw = 0.
Hence, as w = S
2
2
c
0
S
2
itself appears to be a solution of

L
1
V = 0, we have
c
1
(S, t) = g
1
()S
2
2
c
0
S
2
, (7.5.11)
and the complete outer expansion to the three lowest orders is
V (S, , t) c
0
(S, t) +
1
2
(T t)
_
A1
2
,1
+A1
2
,2
T
_
S
2
_
T
2
T
_
c
0
+
_
g
1
() g
1
() (T t) (A
1,1
+A
1,2
T)
1
2
(T t)
2
_
A1
2
,1
+A1
2
,2
T
_
T
2
T
_
__
A1
2
,1
+A1
2
,2
T
_
_
_
T
2
T
_
c
0
,
recalling that T := S

S
.
This seems to be a nice solution, however there are several occurences of c
0
(S, t) in the above
expression for V (S, , t), which is an undetermined function.
7.6 Discussion
In Howisons paper [7], some assumptions, simplifications and remarks are presented below.
After that, the method and solution will be discussed.
Method and assumptions
As a final example in his paper [7], Howison shows the applicability of boundary-layer techniques
in the analysis of fast mean-reverting stochastic volatility models. A well-known example of a
stochastic volatility model is the Heston model.
In these models the volatility itself is assumed to follow a stochastic process while the asset price
is assumed to follow the lognormal process as before in section 5, considering the Black-Scholes
model. That is,
_
dS
t
=
t
S
t
dt +
t
S
t
dW
t
,
d
t
= M
t
dt +
t
d
W
t
.
(7.6.1a)
(7.6.1b)
The two Wiener processes are assumed to be correlated by
E
_
dW
t
d
W
t
_
= dt. (7.6.2)
Here, the timescale for mean-reversion is assumed to be much shorter than that for the evolution
of the asset price. Therefore, their ratio
r
M
is small, such that we can apply perturbation
theory. This results in a scaling of M and by resp.

, such that

t
= M is large
when M =
m
is large.
We do know that the scaled variables m and depend on . Unfortunately, we do not know
their functional form.
First, we have made an outer expansion, where we have assumed pro tem that the correct final
condition (from matching) for the O
_
1
2
_
problem is V1
2
(S, T) = 0. Secondly, we have investi-
gated the boundary layer near t = T. At the end of the boundary layer analysis, Howison starts
7.6. DISCUSSION 59
matching already, by introducing a particular solution V
1
, which vanishes as . After
that, the matching is completed by one-term matching of the outer and inner expansion.
Finally, an expression for the option price V is obained. This solution will be discussed at the
end of this section.
Simplif ications
On page 16 of [7] Howison claims that the solution V
0
of the O(1) equation is a function of
the stock price S and time t only (i.e., V
0
= c
0
(S, t), which is independent of the volatility
). He argues that this choice follows naturally from the behaviour of V
0
in the limits for
small and large S. We, on the other hand, have found a general solution of the form
V
0
(S, , t) = c(S, t)g() +c
0
(S, t). (7.6.3)
However, this solution can satisfy the following two limits:
lim
S
K
V
0
S
= 1, and lim
S
K
0
V
0
S
= 0 (7.6.4)
and exhibit -dependence simultaneously. For example, if we would take c(S, t) = S
2
e
S
,
both limits would be satisfied, and V
0
still depends of in the middle part, because these
are only conditions for very large and very small stock prices S. So, for intermediate S, the
solution V
0
can still depend on . We have asked Howison if there is a model assumption
which says it cannot, but he did not answer that question.
Note that if Howison would not have done the above assumption about the -independence
of V
0
, the higher order equations would be completely dierent, because then the opera-
tor L1
2
would contribute and give some non-zero terms. Consequently, it would have been
a lot more complicated to obtain the higher order solutions V1
2
(S, , t), V
1
(S, , t), etcetera.
Remarks
On page 14 of [7] Howison presents a forward Kolmogorov equation
p
t
=

_
1
2
2
p
(Mp),
while we have found
p
t
=

_
(M )p
_
+
1
2
2
_
2
p
_
,
i.e.,
p
t
=

_
1
2
2
p
(Mp) +

_
p
_
+
__
. (7.6.5)
using p. 291 of [10]. After asking Howison for explanation of why this extra term vanishes,
he answered that we are right that there is a missing term in the FKE:
I forgot to include the market price of risk in that term (I think the f irst version of the
paper incorporated it into M). As a result I think that p
() should have an additional

term of O
_
1
2
_
and this will have consequences further on (i.e., will make the subsequent
formulae worse). Thank you for pointing out this error.
However, if we assume it to be incorporated into M, we would still have an extra term
left, namely,
_
p
_
. (7.6.6)
We have asked Howison to explain this, but unfortunately he did not reply to any of our
subsequent e-mails.
The solution of this forward Kolmogorov equation is also different from Howisons answer.
On page 15 of [7] Howison does not explicitly show his solution p
, but he only remarks

that if it exists, then it is proportional to
1
2
()
e
2

m(s)/
2
(s) ds
.
This is very different from the expression (7.3.8) we obtained, because we did not find the
minus sign in the exponent, and also the integration constant
c
0
m
is missing here.
In the expression for V1
2
on page 18 of [7] Howison lacks a minus sign.
On page 19 of [7] the definitions of A
1,1
and A
1,2
are probably incorrect: the operator T
is missing from both A
1,1
and A
1,2
.
The particular solution

V
1
= g
1
()S
2
2
P
S
2
L
1
P +
1
1
(S) on page 20 of [7] is indeed a
solution of the initial value problem
V
1
+L
0
V
1
=
1
2
_
2
_
S
2
2
P
2
S
L
1
P, (7.6.7a)
V
1
(S, , 0) = 0. (7.6.7b)
It is, however, not clear which steps are taken to obtain this solution.
Method and solution
Howison made certain choices for the scalings. However, often the motivation for them was not
explicitly given. There are many other ways to scale your parameters. For instance, an other
combination of parameters could be chosen in the maximum balance and/or the order of the
scaling in might be dierent. Due to lack of information, for example about how m and
depend on , we cannot determine an O(1) outer solution V
0
that depends on . Howison is
actually throwing away some information at this point by taking c(S, t) = 0.
After applying the method of matched asymptotic expansions to the fast mean-reverting stochas-
tic volatility model and following Howisons assumptions, we have obtained the following ex-
pression for the three lowest order terms of the solution:
V (S, , t) c
0
(S, t) +
1
2
(T t)
_
A1
2
,1
+A1
2
,2
T
_
S
2
_
T
2
T
_
c
0
+
_
g
1
() g
1
() (T t) (A
1,1
+A
1,2
T)
1
2
(T t)
2
_
A1
2
,1
+A1
2
,2
T
_
T
2
T
_
__
A1
2
,1
+A1
2
,2
T
_
_
_
T
2
T
_
c
0
,
where T := S

S
. However, since the function c
0
(S, t) is still undetermined, this result doesnt
seem to be very useful.
7.6. DISCUSSION 61
Our opinion
In the paper written by Howison [7], quite a lot of assumptions and simplications are made.
cases we even think these assumptions and simplications are incorrect.
Also, the choice of the scalings has not been explained very well. Howison presumably attempts
to apply the same procedure as he used for the Black-Scholes model, as discussed in section 5
of this thesis. However, this is not the only way to scale the parameters in this model. He has
chosen the derivative with respect to time to be O(), instead of O(1). So in the rst order
approach in equation (7.3.9), the time derivative has been left out of the problem, because

t
only occurs in the operator L
1
. The reason for this might be that in case the time derivative
would occur in the O(1) problem, it is likely that one would obtain a more complicated solution
V that does not only depend on t, but also on t.
Multiple-scale analysis is a global perturbation scheme that is useful in systems characterized
by disparate time scales. A two-scale expansion might work well for this model. The trick is
to introduce a new variable = t. This variable is called the slow time because it does not
become signicant until t
1
. Then take an expansion of the form

V (S, , t) = V
0
(S, , ) +V
1
(S, , ) +. . . . (7.6.8)
So, a recommendation for further investigation on this model would be to use multiple-scale
methods.
Chapter 8
The SABR model
In the classical Black-Scholes model, the volatility is assumed to be constant. But in reality,
options with different strikes require different volatilities for the underlyings to match their mar-
ket prices. This is called the market volatility smile. Handling these market smiles correctly
is critical for hedging.
The SABR model is a stochastic volatility model that attempts to capture this volatility smile
in derivatives markets. The name is an abbreviation of Stochastic Alpha, Beta, Rho,
referring to the three key parameters of the model.
The SABR model describes a single forward F, such as a forward interest rate, a forward swap
rate, or a forward stock price. The volatility of the forward F is described by a parameter .
Here the volatility is not constant, but is itself a random function of time. SABR is a dynamic
model in which both F and are stochastic state variables whose time evolution is given by the
following system of stochastic differential equations:
_
dF = F
dW
1
, F(0) = f
d = dW
2
, (0) = .
(8.0.1a)
(8.0.1b)
Here, f and are the forward and volatility resp. at time t = 0. On the exercise date t
ex
the
decision whether or not to exercise the option is made. On the settlement date t
set
all payments
are made and the forward contract matures. The period t
set
t
ex
is called the settlement delay.
W
1
and W
2
are two correlated Wiener processes with correlation coefficient 1 < < 1. The
constant parameter satisfies the condition 0 1. The above dynamics are a stochastic
version of the CEV model with the skewness parameter , because this CEV model is aug-
mented by stochastic volatility. In fact, it reduces to the CEV model if = 0.
Under the SABR model (8.0.1), the price of European options is given by Blacks formula
1
V
call
= D(t
set
) f^(d
+
) K^(d
) , (8.0.2)
V
put
= V
call
+D(t
set
) [K f] , (8.0.3)
1
Note that there is a difference between Blacks formula and the Black-Scholes formula:
In case we are considering a spot price as we did before, the spot measure has to be used. In this case, the Black-
Scholes formula can be derived. Here, we have a forward, which is a martingale in the forward measure because
its expectation is equal to its present value. This leads to Blacks formula. These two formulas are equivalent for
deterministic interest rates. When the interest rate is stochastic, an extra term occurs in Blacks formula.
63
64 CHAPTER 8. THE SABR MODEL
with D(t) the discount factor for date t, i.e., the value today of $1 to be delivered on date t, and
d
=
log (f/K)
1
2
2
B
t
ex
t
ex
, (8.0.4)
where the implied volatility
B
(f, K) is given by
B
(f, K) = (fK)
(1)/2
_
1 +
(1 )
2
24
log
2
_
f
K
_
+
(1 )
4
1920
log
4
_
f
K
_
+. . .
_
1
_
z
x(z)
_
_
1 +
_
(1 )
2
24
2
(fK)
1
+
1
4
(fK)
(1)/2
+
2 3
2
24

2
_
t
ex
+. . .
_
.
Here,
z =

(fK)
(1)/2
log
_
f
K
_
, (8.0.5)
and x(z) is defined by
x(z) = log
_
_
1 2z +z
2
+z
1
_
. (8.0.6)
In section 8.6 we will derive this result.
For the special case of at-the-money options, options struck at K = f, this formula reduces to
ATM
=
B
(f, f) =

f
(1)
_
1 +
_
(1 )
2
24
2
f
22
+
1
4
f
(1)
+
2 3
2
24

2
_
t
ex
+. . .
_
. (8.0.7)
8.1 Derivation of the differential equation
After applying Itos formula (4.3.1), we find the following expression for the infinitesimal incre-
ment in the option price:
dV =
V
t
dt +
V
F
dF +
V
d +
1
2
2
V
F
2
dF
2
+
1
2
2
V
d
2
+

2
V
F
dFd,
=
_
V
t
+
1
2
2
F
2
2
V
F
2
+
1
2
2
V
2
+
2
F

2
V
F
_
dt +F
V
F
dW
1
+
V
dW
2
.
Next we construct a portfolio = V F
1
V
1
, such that d = dV dF
1
dV
1
. The
observation that d = rdt, leads to
r (V F
1
V
1
) dt = dV dF
1
dV
1
. (8.1.1)
Substituting the formulas for dV , dF and dV
1
into equation (8.1.1) yields
_
V
t
+
1
2
2
F
2
2
V
F
2
+
1
2
2
V
2
+
2
F

2
V
F
rV
_
dt
+ F
V
F
dW
1
+
V
dW
2
F
dW
1
Fdt
=
1
_
V
1
t
+
1
2
2
F
2
2
V
1
F
2
+
1
2
2
V
1
2
+
2
F

2
V
1
F
rV
1
_
dt. (8.1.2)
8.1. DERIVATION OF THE DIFFERENTIAL EQUATION 65
Collecting all stochastic terms containing dW
1
and dW
2
together, this equation (8.1.2) trans-
forms into
_
V
t
+
1
2
2
F
2
2
V
F
2
+
1
2
2
V
2
+
2
F

2
V
F
rV
_
dt Fdt
=
1
_
V
1
t
+
1
2
2
F
2
2
V
1
F
2
+
1
2
2
V
1
2
+
2
F

2
V
1
F
rV
1
_
dt
+
_
1
F
V
1
F
F
V
F
+ F
_
dW
1
+
_
V
1
_
dW
2
. (8.1.3)
To obtain a risk-free portfolio, these stochastic terms have to be eliminated, as we did before in
section 7.1. Hence, we choose
_
1
=
_
V
_
_
V
1
_,
=
V
F

1
V
1
F
=
V
F

_
V
_
_
V
1
_
V
1
F
.
(8.1.4a)
(8.1.4b)
Following the steps taken in [3], substituting
1
and as above into equation (8.1.3), and
collecting all V -dependent terms on the left-hand side and all V
1
-dependent terms on the right-
hand side, we obtain
V
t
+
1
2
2
F
2
2
V
F
2
+
1
2
2
2
V
2
+
2
F

2
V
F
F
V
F
rV
_
V
_
=
V
1
t
+
1
2
2
F
2
2
V
1
F
2
+
1
2
2
2
V
1
2
+
2
F

2
V
1
F
F
V
1
F
rV
1
_
V
1
_ . (8.1.5)
Because the left-hand side of equation (8.1.5) is explicitly independent of V
1
and the right-hand
side is explicitly independent of V , either sides must be independent of both V and V
1
. The only
way that this can be is for both sides to be equal to some function of the independent variables
F, and t. We deduce that
V
t
+
1
2
2
F
2
2
V
F
2
+
1
2
2
V
2
+
2
F

2
V
F
F
V
F
rV =
V
, (8.1.6)
where, without loss of generality, we have written the arbitrary function of F, and t as
_
0
_
, where 0 and are the drift and volatility functions from the sde (8.0.1) for instan-
taneous variance. Note that here = (F, , t).
Hence, the price V (F, , t) satisfies the following partial differential equation (pde):
V
t
+
1
2
2
F
2
2
V
F
2
+
1
2
2
V
2
+
2
F

2
V
F
F
V
F

V
rV = 0. (8.1.7)
Scaling parameters in the above pde, using maximum balance, does not seem to work very well,
since after scaling we obtain pdes that are still dicult to solve. Hence, for the SABR model
we will follow the procedure of Hagan et al., applying the scalings on the system of stochastic
dierential equations and using Kolmogorov equations, see section 8.2 of this thesis.
8.2 Scalings
Consider the following model in which both F and are represented by stochastic state variables
whose time evolution is given by the following system of stochastic differential equations:
_
dF = C(F)dW
1
, F(0) =f,
d = dW
2
, (0) =,
(8.2.1a)
(8.2.1b)
under the forward measure, where the two Wiener processes are correlated by
E[dW
1
dW
2
] = dt. (8.2.2)
8.3 Application of perturbation theory to SABR model
Initially, we will analyze the model with a general function C(F), for notational convenience.
After that, the results are specialized to the power law F
.
Our analysis is based on a small volatility expansion, where we take both the volatility and
the volatility of volatility
2
to be small. To carry out this analysis systematically, we first scale
them by as follows:
=

, = , (8.3.1a)
=

, = , (8.3.1b)
such that and are O
s
(1). This yields
_
dF = C(F) dW
1
, F(0) =f,
d = dW
2
, (0) =,
(8.3.2a)
(8.3.2b)
Suppose that the market is in state F(t) = f, (t) = at date t. Define the transition density
function p by
p(t, f, ; T, F, A) dFdA = prob
_
F < F(T) < F + dF, A < (T) < A+ dA
F(t) = f, (t) =
_
.
According to [10], p. 291, the unique Forward Kolmogorov Equation (fke) of a stochastic
differential equation (sde)
df = D
1
dt +D
2
dW, (8.3.3)
where D
1
and D
2
are matrices, and f, dt and dW are vectors, is given by
f
T
=
N
i=1
x
i
_
D
1
i
(x
1
, . . . , x
N
)f
_
+
1
2
N
i=1
N
j=1
2
x
i
x
j
_
D
2
ij
(x
1
, . . . , x
N
)f
_
. (8.3.4)
Next, we can construct the fke for the transition density function p, which is given by
T
p(t, f, ; T, F, A) =
1
2
2
F
2
_
2
A
2
C
2
(F) p
_
+
1
2
2
A
2
_
2
A
2
p
_
+
1
2
2
FA
_
2
A
2
p
_
+
1
2
2
AF
_
2
A
2
p
_
,
=
1
2
2
A
2

2
F
2
_
C
2
(F) p
_
+
2

2
FA
_
A
2
C(F) p
_
+
1
2
2

2
A
2
_
A
2
p
_
,
2
This is commonly abbreviated as volvol.
8.3. APPLICATION OF PERTURBATION THEORY TO SABR MODEL 67
for t < T, with p = (F f)(A) at t = T.
3
Let V (t, f, ) be the value of a European call option at date t, when the economy is in state
F(t) = f, (t) = . Let t
ex
be the options exercise date, and let K be its strike. Omitting the
discount factor D(t
ex
), because we do a current valuation of the final payoff, the value of the
option is
V (t, f, ) = E[max(F(t
ex
) K, 0)[F(t) = f, (t) = ] ,
=
_

A=
_

F=K
(F K) p(t, f, ; t
ex
, F, A) dF dA. (8.3.5)
Since
p(t, f, ; t
ex
, F, A) = (F f)(A) +
_
t
ex
t
T
p(t, f, ; T, F, A) dT, (8.3.6)
we can rewrite equation (8.3.5) as
V (t, f, ) = max(f K, 0)+
_
t
ex
T=t
_

F=K
_

A=
(F K)

T
p(t, f, ; T, F, A) dA dF dT. (8.3.7)
Next, substitute the fke
T
p(t, f, ; T, F, A) =
1
2
2
A
2

2
F
2
_
C
2
(F) p
_
+
2

2
FA
_
A
2
C(F) p
_
+
1
2
2

2
A
2
_
A
2
p
_
into equation (8.3.7), where integrating the A derivatives, i.e.,

2
FA
_
A
2
C(F)p
_
and
1
2
2

2
A
2
_
A
2
p
_
(8.3.8)
over all A yields zero. This is caused by the fact that p and its derivatives with respect to A
resp. F go faster to zero than A resp. F and its powers go to , see [2]. Therefore our option
price reduces to
V (t, f, ) = max(f K, 0)
+
1
2
2
_
t
ex
T=t
_

A=
_

F=K
A
2
(F K)

2
F
2
_
C
2
(F) p
_
dF dA dT, (8.3.9)
where we have changed the order of integration. Integration by parts twice with respect to F
yields
_

K
A
2
(F K)

2
F
2
_
C
2
(F) p
_
dF =
_
A
2
(F K)

F
_
C
2
(F) p
_
_
_

K
A
2

F
_
C
2
(F) p
_
dF,
= A
2
_
C
2
(F) p
K
= A
2
C
2
(K)p,
Again, this is caused by the property that p goes faster to zero than C(F) and its powers go to
[2] such that
V (t, f, ) = max(f K, 0) +
1
2
2
C
2
(K)
_
t
ex
T=t
_

A=
A
2
p(t, f, ; T, K, A) dA dT. (8.3.10)
3
This terminal condition follows directly from the definition of the transition density function p, as given
above.
According to [10], p. 291, the unique Backward Kolmogorov Equation (bke) of a stochastic
differential equation (sde)
df = D
1
dt +D
2
dW, (8.3.11)
where D
1
and D
2
are matrices, and f, dt and dW are vectors, is given by
f
t
=
N
i=1
x
i
_
D
1
i
(x
1
, . . . , x
N
)f
_
+
1
2
N
i=1
N
j=1
2
x
i
x
j
_
D
2
ij
(x
1
, . . . , x
N
)f
_
. (8.3.12)
Before we will construct the bke for the transition density function p, let us define
P(t, f, ; T, K) :=
_

A
2
p(t, f, ; T, K, A) dA, (8.3.13)
to simplify the problem further. Then P satisfies the following Backward Kolmogorov Equa-
tion (bke):
P
t
+
1
2
2
C
2
(f)
2
P
f
2
+
2
2
C(f)

2
P
f
+
1
2
2
P
2
= 0, (8.3.14)
for t < T and P =
2
(f K) for t = T.
Since t does not appear explicitly in this equation, P depends only on the combination T t,
and not on t and T separately. So, define
:= T t and
ex
:= t
ex
t. (8.3.15)
Then our pricing equation becomes
V (t, f, ) = max(f K, 0) +
1
2
2
C
2
(K)
_

ex
0
P(, f, ; K)d, (8.3.16)
where P is the solution of the problem
_
_
P
=
1
2
2
C
2
(f)
2
P
f
2
+
2
2
C(f)

2
P
f
+
1
2
2
P
2
, for > 0,
P =
2
(f K), at = 0.
(8.3.17a)
(8.3.17b)
Since P starts out as a delta function, initially its derivatives will be large enough so that the
size of the

2
P
f
2
term offsets the smallness of .
4
To capture this limit, let us define the local variable
:=
f K
, (8.3.18)
such that f can be replaced by K +.
Then,
f

1
,

2
f
2

1
2
, (8.3.19)
4
This means, the product
2
2
P
f
2
is still large, even if is small, because

2
P
f
2
is the second derivative of the
delta function, which has a very steep slope.
and we have the Taylor expansion
C(f) = C(K +) = C
0
_
1 +
1
+
1
2
2
+. . .
_
, (8.3.20)
where
C
0
= C(K),
1
=
C
(K)
C(K)
,
2
=
C
(K)
C(K)
. (8.3.21)
Substituting this into the pde yields,
P
=
1
2
2
C
2
0
_
1 + 2
1
+
2
_
2
+
2
1
_
2
+. . .
_

2
P
2
+
2
2
C
0
1 +
1
+. . .

2
P
+
1
2
2
P
2
, (8.3.22a)
for > 0 with the initial condition
5
P =
2
() =

2
() as 0. (8.3.22b)
Expanding
P
(t, , ) =
1
P
0
(t, , ) +P
1
(t, , ) +P
2
(t, , ) +. . . , (8.3.23)
substituting this expansion into problem (8.3.22) and equating like powers of leads to the
following O(
1
) problem:
_
_
P
0
=
1
2
2
C
2
0
2
P
0
2
, for > 0,
P
0
=
2
(), as 0.
(8.3.24a)
(8.3.24b)
The solution of the above problem can be computed by applying the following transformations:
_
_
y :=

2
,
k :=
C
2
0
2
2
,
t := ,
u(t, y) := P
0
(, ).
(8.3.25a)
(8.3.25b)
(8.3.25c)
(8.3.25d)
Then,
2
y
2
=
4

2
2
, (8.3.26)
and the problem (8.3.31) transforms into
_
_
u
t
k
2
u
y
2
= 0,
u(t = 0, y) = (y),
(8.3.27a)
(8.3.27b)
which is the heat or diffusion equation, subject to a Dirac delta in the initial condition.
5
Here, () is the Dirac delta function. For a proof of the equality () =
()
, see Appendix L.
The solution to this problem (8.3.27) is given by the fundamental solution or heat kernel
u(t, y) =
1
4kt
e
y
2
4kt
. (8.3.28)
In terms of the original variables, this equals
P
0
(, ) =
1
_
4
C
2
0
2
2
exp
_
4
4
C
2
0
2
2
_
=
[[
_
2C
2
0
exp
_

2
2
2
C
2
0
_
, (8.3.29)
i.e.,
P
0
(, ) =
[[
_
2C
2
0
exp
_
(f K)
2
2
2
2
C
2
0
_
. (8.3.30)
The O(1) problem is given by
_
_
P
1

1
2
2
C
2
0
2
P
1
2
=
1
2
C
2
0
2
P
0
2
, for > 0,
P
1
= 0, as 0.
(8.3.31a)
(8.3.31b)
whereas the O() problem is given by
_
_
P
2

1
2
2
C
2
0
2
P
2
2
=
1
2
C
2
0
2
P
1
2
+C
0
2
P
0
, for > 0,
P
2
= 0, as 0.
(8.3.32a)
(8.3.32b)
Because the differential equations and initial conditions are independent of , all P
i
(i = 1, 2, . . .)
are O(1) so we can conclude that [4]
P =
1
P
0
+P
1
+P
2
+. . . =

_
2
2
2
C
2
0
exp
_
(f k)
2
2
2
2
C
2
0
_
1 +. . . . (8.3.33)
The expansion
C(f) =
n=0
C
(n)
(K)
()
n
n!
= C(K) +C
(K) +O
_
2
_
(8.3.34)
can be rewritten as
C(f) = C(K)
_
1 +
C
(K)
C(K)

_
= C(K)
_
1 +
C
(K)
C(K)
(f K)
_
, (8.3.35)
such that the difference C(f) = C(K) becomes
C(f) C(K) = C
(K)(f K). (8.3.36)

Since the +. . . in equation (8.3.33) involves powers of (f K)/C(K), this expansion would
become inaccurate as soon as (f K)C
(K)/C(K) becomes a significant fraction of 1, i.e., as

soon as C(f) and C(K) are significantly different.
Note that small changes in the exponent cause much greater changes in the probability density.
Therefore, a better approach is to re-cast the series as
P =

_
2
2
C
2
(K)
exp
_
(f K)
2
2
2
2
C
2
(K)
_
1 +. . . (8.3.37)
and expand the exponent, since one expects that only small changes to the exponent will be
needed to effect the much larger changes in the density.
We can refine this approach by noting that the exponential can be written in terms of an integral
as follows:
(f K)
2
2
2
2
C
2
(K)
1 +. . . =
1
2
_
1
_
f
K
df
C(f
)
_
2
1 +. . .. (8.3.38)
This can be explained by writing out the integral in the right-hand side as follows:
_
f
K
df
C(f
)
=
_

0
d
C(K +
)
, (8.3.39a)
=
_

0
d
C(K) +C
(K)
+O(
2
)
, (8.3.39b)
=

C(K)
_

0
d
1 +
C
(K)
C(K)

+O(
2
)
, (8.3.39c)
=

C(K)
_

0
d
1 +
1
+O(
2
)
, (8.3.39d)
=

C(K)
_

0
_
1
1
+O
_
2
__
d
, (8.3.39e)
=

C(K)
_

1
2
2
+O
_
2
_
_
, (8.3.39f)
=

C(K)
=
f K
C(K)
. (8.3.39g)
Here, the Mercator series
6
are used to perform the step from equation (8.3.39d) to equation
(8.3.39e), i.e.,
1
1 + +O(
2
)
= 1 +1 +O
_
2
_
= 1 +O
_
2
_
. (8.3.41)
Having found that
f K
C(K)
=
_
f
K
df
C(f
)
(8.3.42)
implies that
(f K)
2
C
2
(K)
=
__
f
K
df
C(f
)
_
2
(8.3.43)
and thus we can maken the substitution
(f K)
2
2
2
2
C
2
(K)
1 +. . . =
1
2
_
1
_
f
K
df
C(f
)
_
2
1 +. . ., (8.3.44)
which concludes the derivation of equation (8.3.38).
6
In mathematics, the Mercator series or Newton-Mercator series is the series
n=1
(1)
n+1
n
x
n
= x
x
2
2
+
x
3
3

x
4
4
+. . . . (8.3.40)
This is the Taylor series for the natural logarithm shifted by 1, i.e., log(1 + x).
Here, it is a natural choice to change variables from f to
z :=
1
_
f
K
df
C(f
)
, (8.3.45)
since f only occurs in combination with this integral. Also, we define
B(z) := C(f). (8.3.46)
Then,
f

1
C(f)
z
=
1
B(z)
z
and

z
, (8.3.47)
such that
2
f
2

1
2
B
2
(z)
_

2
z
2

B
(z)
B(z)
z
_
, (8.3.48a)
2
f

1
B(z)
_

2
z

z
2
z
2

1
z
_
, (8.3.48b)
2

2
2

2z
2
z
+
z
2
2
z
2
+
2z
z
. (8.3.48c)
Also,
7
(f K) = (zC(K)) =
1
C(K)
(z). (8.3.49)
Therefore, (8.3.16) through (8.3.17) become
V (t, f, ) = max(f K, 0) +
1
2
2
C
2
(K)
_

ex
0
P(, z, )d, (8.3.50)
where P(, z, ) is the solution of the boundary value problem
_
_
P
=
1
2
2
P
z
2

1
2
B
P
z
+

2
P
z
z
2
P
z
2

P
z
+
1
2
2
P
2

2
2
z
2

2
P
z
+
1
2
2
z
2
2
P
z
2
+
2
2
z
P
z
, for > 0,
P =

C(K)
(z), at = 0,
(8.3.51a)
(8.3.51b)
i.e.,
_
_
P
=
1
2
_
1 2z +
2
2
z
2
_

2
P
z
2

1
2
B
P
z
+
_

2
2
z
_
_

2
P
z

P
z
_
+
1
2
2
P
2
, for > 0,
P =

C(K)
(z), at = 0.
(8.3.52a)
(8.3.52b)
Accordingly, let us define
8
P(, z, ) :=

C(K)P, (8.3.53)
7
In integral form.
8
Here, we have adopted the notation of Hagan et al., containing C(K). Actually, in this case it would be
better to write

P(, z, ) :=

B(0)P. Recall that B(z) = C(f) and B(0) = C(K).

such that P can be replaced by

C(K)
P.
Then,
C(K)
P
f

C(K)
_

C(K)

P
_
=
1
P +

P
, and (8.3.54a)
C(K)
2
P
2

C(K)
2
_

C(K)

P
_
=
1
P +
_
,
=
1
+

P
2

P
2
_
=
2
+

2

P
2
. (8.3.54b)
In terms of

P we obtain
V (t, f, ) = max(f K, 0) +
1
2
C(K)
_

ex
0
P(, z, )d, (8.3.55)

where

P(, z, ) is the solution of
_
=
1
2
_
1 2z +
2
2
z
2
_

2

P
z
2

1
2

P
z
+
_

2
2
z
_

2

P
z
+
1
2
2
_
2

P
2
+ 2
_
, for > 0,
P = (z), at = 0.
(8.3.56a)
(8.3.56b)
To leading order,

P is the solution of the standard diffusion problem
=
1
2
2

P
z
2
, (8.3.57)
with

P = (z) at = 0. So it is a Gaussian to leading order. The next stage is to transform
the problem into the standard diffusion problem through O(), and then through O
_
2
_
, etc.
This is the near identify transform method
9
which has proven so powerful in near-Hamiltonian
systems
10
.
Note that the variable does not enter the problem for

P until O(), so
P(, z, ) = P
0
(, z) +P
1
(, z, ) +. . . . (8.3.59)
Consequently, the derivatives

2
P
z
,

2
P
2
and

P
are all O(). Recall that we are only solving

for

P through O
_
2
_
.
9
See appendix M.
10
A near-Hamiltonian system is of the form
_
_
dp
dt
=
H
q
+(. . .) =: F,
dq
dt
=
H
p
+(. . .)=: G,
(8.3.58a)
(8.3.58b)
with
F
p
=
G
q
.
So, through this order, we can re-write our boundary value problem as
_
=
1
2
_
1 2z +
2
2
z
2
_

2

P
z
2

1
2

P
z
+

2

P
z
, for > 0,
P = (z), at = 0.
(8.3.60a)
(8.3.60b)
Let us now eliminate the
1
2

P
z
term. Define H(, z, ) by
P =
_
C(f)
C(K)
_
n
H =
_
B(z)
B(0)
_
n
H, (8.3.61)
where n is an O(1) constant that will be determined in the next steps.
Then,

P
z
=
_
B(z)
B(0)
_
n
_
H
z
+n
B
B
H
_
, (8.3.62a)
2

P
z
2
=
_
B(z)
B(0)
_
n
_
2
H
z
2
+ 2n
B
B
H
z
+
2
2
_
n
B
B
+ (n
2
n)
B
2
B
2
_
H
_
, (8.3.62b)
2

P
z
=
_
B(z)
B(0)
_
n
_

2
H
z
+nz
B
B
H
z
+n
B
B
H
+n
B
B
H +O
_
2
_
_
. (8.3.62c)
The option price now becomes
V (t, f, ) = max(f K, 0) +nB(0)
_
B(z)
B(0)
_
n
_

ex
0
H(, z, ) d, (8.3.63)
where H(, z, ) is the solution of
H
=
1
2
_
1 2z +
2
2
z
2
_
_
2
H
z
2
+ 2n
B
B
H
z
+
2
2
_
n
B
B
+ (n
2
n)
B
2
B
2
_
H
_
+
_

2
H
z
+nz
B
B
H
z
+n
B
B
H
+n
B
B
H +O
_
2
_
_
1
2
B
_
H
z
+n
B
B
H
_
for > 0, with initial condition H = (z) at = 0. Omitting all O
_
3
_
terms, and combining
some terms yields
H
=
1
2
_
1 2z +
2
2
z
2
_

2
H
z
2
+n
B
B
H
z
+
1
2
2
_
n
B
B
+ (n
2
n)
B
2
B
2
_
H
2
n
B
B
_
z
H
z
H
_
1
2
B
H
z
1
2
2
n
2
B
2
B
2
H +
_

2
H
z
+n
B
B
H
_
. (8.3.64)
Next, we combine the third and sixth term of equation (8.3.64) and reorder terms to obtain
H
=
1
2
_
1 2z +
2
2
z
2
_

2
H
z
2
+
_
n
1
2
_
B
B
H
z
+
1
2
2
_
n
B
B
+ (n
2
2n)
B
2
B
2
_
H
2
n
B
B
_
z
H
z
H
_
+
_

2
H
z
+n
B
B
H
_
. (8.3.65)
The choice n =
1
2
, eliminates the term containing
H
z
. This yields
V (t, f, ) = max(f K, 0) +
1
2
B(0)
_
B(z)
B(0)
_1
2
_

ex
0
H(, z, ) d, (8.3.66)
for the option price, where H(, z, ) is the solution of
H
=
1
2
_
1 2z +
2
2
z
2
_

2
H
z
2

1
2
B
_
z
H
z
H
_
+
2
2
_
1
4
B
B

3
8
B
2
B
2
_
H +
_

2
H
z
+
1
2
B
H
_
, (8.3.67)
for > 0, with initial condition H = (z) at = 0.
Since this equation (8.3.67) is independent of to leading order, i.e.,
_
_
_
H
=
1
2
2
H
z
2
, for > 0,
H = (z), at = 0,
(8.3.68a)
(8.3.68b)
we can conclude that the -derivatives
H
and

2
H
2
are no larger than O().
At O() equation (8.3.67) depends on only through the last term
_

2
H
z
+
1
2
B
H
_
. (8.3.69)
Because the -derivatives are no larger than O(), this last term is actually only O
_
3
_
, and
can thus be neglected, since only the terms through O
s
_
2
_
are taken into account.
This yields
_
_
H
=
1
2
_
1 2z +
2
2
z
2
_

2
H
z
2

1
2
B
_
z
H
z
H
_
+
2
2
_
1
4
B
B

3
8
B
2
B
2
_
H, for > 0,
H = (z), at = 0.
(8.3.70a)
(8.3.70b)
There are no longer any -derivatives, so now can be treated as a parameter instead of as
an independent variable. This means we have succesfully reduced the problem to one dimension.
Let us now remove the
H
z
term through O
_
2
_
. To leading order the ratios
B
(z)
B(z)
and
B
(z)
B(z)
(8.3.71)
are constant. We can replace them by
b
1
:=
B
(z
0
)
B(z
0
)
and b
2
:=
B
(z
0
)
B(z
0
)
, (8.3.72)
committing only an O() error, where the constant z
0
will be chosen later on. This can be seen
by expanding z = z
0
+z
1
+. . ..
We now define

H by
H = exp
_
1
4
2
b
1
z
2
_

H, (8.3.73)
such that
H
exp
_
1
4
2
b
1
z
2
_

H
, (8.3.74a)
H
z
exp
_
1
4
2
b
1
z
2
_
_

H
z
+
1
2
2
b
1
z

H
_
, (8.3.74b)
2
H
z
2
exp
_
1
4
2
b
1
z
2
_
_
2

H
z
2
+
2
b
1
z

H
z
+
1
2
2
b
1

H
_
. (8.3.74c)
Again, only terms up to and including O
_
2
_
have to be taken into account.
Then our option price becomes
V (t, f, ) = max(f K, 0) +
1
2
_
B(0)B(z)e
2
b
1
z
2
/4
_

ex
0
H(, z, ) d, (8.3.75)
and equation (8.3.70a) transforms into
=
1
2
_
1 2z +
2
2
z
2
_
_
2

H
z
2
+
2
b
1
z

H
z
+
1
2
2
b
1

H
_
1
2
2
b
1
_
z
_

H
z
+
1
2
2
b
1
z

H
_

H
_
+
2
2
_
1
4
b
2
3
8
b
2
1
_

H. (8.3.76)
Neglecting all terms that are O
_
3
_
, this can be rewritten as
=
1
2
_
1 2z +
2
2
z
2
_

2

H
z
2
+
1
2
2
b
1
z

H
z
+
1
4
2
b
1

H
1
2
2
b
1
z

H
z
+
1
2
2
b
1

H +
2
2
_
1
4
b
2
3
8
b
2
1
_

H, (8.3.77)
i.e.,
=
1
2
_
1 2z +
2
2
z
2
_

2

H
z
2
+
2
2
_
1
4
b
2
3
8
b
2
1
_

H +
3
4
2
b
1

H, (8.3.78)
for > 0, with initial condition H(0, z, ) = (z).
Next, we define
x :=
1
_
z
0
d
_
1 2 +
2
, (8.3.79)
i.e.,
x =
1
_
z
0
d
_
1
2
+ ( )
2
=
1
_
z
0
d
_
1
2
_
1 +
()
2
1
2
. (8.3.80)
Define
:=

_
1
2
, (8.3.81)
such that x can be written as
x =
1
_
z
=0
d
1 +
2
. (8.3.82)
Pythagoras theorem states that in a right triangle with sides of length 1 and , the hypotenuse
(the side opposite the right angle) is equal to
1 +
2
. Also we define the angle as illustrated
in figure 8.1.
Figure 8.1: Right triangle with sides of length 1 and , and hypotenuse

1 +
2
.
From this we can derive that
1
1 +
2
= cos , (8.3.83a)
= tan() d = d tan =
d
cos
2
. (8.3.83b)
Substituting this into the expressions for x (8.3.82) yields
x =
1
_
z
=0
cos
cos
2
d =
1
_
z
=0
d
cos
. (8.3.84)
After computing this integral, we obtain
x =
1
_
log
1
cos
+ tan
_
z
=0
. (8.3.85)
Using the expressions of in terms of and the fact that

1 +
2
> , this can be written as
x =
1
_
log
_
_
1 +
2
+
__
z
=0
. (8.3.86)
The last step is to substitute the definition of back into this expression, to obtain
x =
1
_
log
_
_
1 2 +
2
_
1
2
+

_
1
2
__
z
=0
,
=
1
log
_
_
1 2z +
2
2
z
2
+z
_
1
2
_
log
_
1
_
1
2
_
. (8.3.87)
Hence, the definition of x (8.3.79) can also be written as
x :=
1
log
_
_
1 2z +
2
2
z
2
+z
1
_
. (8.3.88)
It is also possible to find an implicit expression for z in terms of x, by defining
y := z . (8.3.89)
Then,
y
2
= (z )
2
=
2
2
z
2
2z +
2
, (8.3.90)
such that equation (8.3.88) can be written as
x =
1
log
_
_
1
2
+y
2
+y
1
_
. (8.3.91)
(1 )e
x
=
_
1
2
+y
2
+y. (8.3.92)
We can also write
_
1
2
+y
2
y =
_
_
1
2
+y
2
y
_
_
1
2
+y
2
+y
_
1
2
+y
2
+y
=
1
2
+y
2
y
2
_
1
2
+y
2
+y
=
1
2
(1 )e
x
=
(1 )(1 +)
(1 )e
x
= (1 +)e
x
. (8.3.93)
Using these two expressions (8.3.92) and (8.3.93), we can write y as
y =
1
2
__
_
1
2
+y
2
+y
_
_
_
1
2
+y
2
y
__
,
=
1
2
_
(1 ) e
x
(1 +) e
x
,
=
1
2
__
e
x
e
x
_
_
e
x
+e
x
_
,
= sinh (x) cosh (x) . (8.3.94)
Substituting this into the definition of y (8.3.89) and reordering yields
z = y + = sinh (x) cosh (x) +. (8.3.95)
Hence, expression (8.3.88) can be written implicitly as
z = sinh (x) (cosh (x) 1) . (8.3.96)
After the transformation from z to x, we have

H
z

1
_
1 2z +
2
2
z
2

H
x
=
1
I(z)

H
x
, (8.3.97a)
2

H
z
2

1
I
2
(z)
_
2

H
x
2
I
(z)

H
x
_
,
=
1
1 2z +
2
2
z
2
_
2

H
x
2
I
(z)

H
x
_
. (8.3.97b)
Here,
I() =
_
1 2 +
2
. (8.3.98)
In terms of x, we have
V (t, f, ) = max(f K, 0) +
1
2
_
B(0)B(z) exp
_
1
4
2
b
1
z
2
__

ex
0
H(, x) d, (8.3.99)
with
=
1
2
2

H
x
2

1
2
I
(z)

H
x
+
2
2
_
1
4
b
2
3
8
b
2
1
_

H +
3
4
2
b
1

H, (8.3.100)
for > 0, with initial condition H = (x), at = 0.
The final step is to define Q(, x) by
H = I
1
2
(z(x)) Q =
_
1 2z +
2
2
z
2
_1
4
Q. (8.3.101)
Then,
11
I
1
2
(z)
Q
, (8.3.102a)

H
x
I
1
2
(z)
_
Q
x
+
1
2
I
(z)Q
_
, (8.3.102b)
2

H
x
2
I
1
2
_
2
Q
x
2
+I
(z)
Q
x
+
2
2
_
1
2
I
I +
1
4
I
2
_
Q
_
. (8.3.102c)
Hence, the option price becomes
V (t, f, ) = max(f K, 0) +
1
2
_
B(0)B(z)I
1
2
(z)e
1
4
2
b
1
z
2
_

ex
0
Q(, x) d, (8.3.103)
where Q is the solution of
Q
=
1
2
2
Q
x
2
+
1
2
I
Q
x
+
1
2
2
_
1
2
I
I +
1
4
I
2
_
Q
1
2
I
Q
x
1
4
2
I
2
Q+
2
2
_
1
4
b
2
3
8
b
2
1
_
Q+
3
4
2
b
1
Q, (8.3.104)
i.e.,
Q
=
1
2
2
Q
x
2
+
2
2
_
1
4
I
I
1
8
I
2
_
Q+
2
2
_
1
4
b
2
3
8
b
2
1
_
Q+
3
4
2
b
1
Q, (8.3.105)
for > 0, with initial condition Q = (x), at = 0.
As above, we can replace I(z), I
(z), I
(z) by the constants I(z

0
), I
(z
0
), I
(z
0
),
and commit only O() errors.
11
Here, the fact that
z
x
= I(z) is used to compute the derivatives with respect to x.
Define the constant by
:=
2
_
1
4
I
(z
0
)I(z
0
)
1
8
_
I
(z
0
)
_
2
_
+
2
_
1
4
b
2
3
8
b
2
1
_
+
3
4
b
1
, (8.3.106)
where z
0
will be chosen later.
Then, through O
_
2
_
, we can simplify our problem to
_
_
_
Q
=
1
2
2
Q
x
2
+
2
Q, for > 0,
Q = (x), at = 0.
(8.3.107a)
(8.3.107b)
The solution of system (8.3.107) is
Q(, x) =
1
2
e
x
2
2
e
. (8.3.108)
Expanding the last exponential e
in a Taylor series around = 0 yields 1 +

2
+ O
_
4
_
.
Note that this expanding
1
(1
2
3
2
+...)
3
2
in a Taylor series around 0 gives the same result. Hence,
Q(, x) =
1
2
e
x
2
2
1
_
1
2
3
2
+. . .
_3
2
, (8.3.109)
through O
_
2
_
.
This solution yields the option price
V (t, f, ) = max(f K, 0) +
1
2
_
B(0)B(z)I
1
2
(z)e
1
4
2
b
1
z
2
_

ex
0
1
2
e
x
2
2
e
d.
Observe that this can be written as
V (t, f, ) = max(f K, 0) +
1
2
f K
x
_

ex
0
1
2
e
x
2
2
e
d, (8.3.110a)
where
2
:= log
_
z
f K
_
B(0)B(z)
_
+ log
_
xI
1
2
(z)
z
_
+
1
4
2
b
1
z
2
. (8.3.110b)
Expanding
2
through O
_
2
_
yields [4]
2

_
1
12
b
2
1
8
b
2
1
_
2
z
2
+
_
_
1
12
I
(
1
2
z)
I(
1
2
z)

1
24
_
I
(
1
2
z)
I
0
(
1
2
z)
_
2
_
_
2
z
2
+
1
4
2
b
1
z
2
.
(8.3.111)
Using
z
x
= 1 +O() and I(z) = 1 +. . ., we note that

x
2
matches
3
=
1
3
2
_
1
4
I
(z
0
)I(z
0
)
1
8
_
I
(z
0
)
_
2
_
+
1
3
2
_
1
4
b
2
3
8
b
2
1
_
+
1
4
b
1
, (8.3.112)
and thus
e
=
1
_
1
2
3
_3
2
=
1
_
1 2
2

x
2
_
3
2
+O
_
4
_
, (8.3.113)
through O
_
2
_
.
Therefore, our option price is
V (t, f, ) = max(f K, 0) +
1
2
f K
x
_

ex
0
1
2
e
x
2
2
e
1
_
1 2
2

x
2
_
3
2
d. (8.3.114)
Changing integration variables to q :=
x
2
2
reduces this to
V (t, f, ) = max(f K, 0) +
[f K[
4
_

x
2
2
ex
e
q+
2
(q
2
)
3
2
dq. (8.3.115)
That is, the value of a European call option is given by
V (t, f, ) = max(f K, 0) +
[f K[
4
_

x
2
2
ex
3
2
e
q
dq, (8.3.116a)
with
2
:= log
_
z
f K
_
B(0)B(z)
_
+ log
_
xI
1
2
(z)
z
_
+
1
4
2
b
1
z
2
, (8.3.116b)
through O
_
2
_
.
Solving problem (8.3.17) for P(, f, ; K) and then substituting this into the pricing formula
(8.3.16) to obtain the option value V (t, f, ) under the SABR model, the resulting formulas
(8.3.116) we find are awkward and not very useful.
First comparison: Normal model
To cast the results in a more usable form, we re-compute the option price under the normal
model
dF =
N
dW, (8.3.117)
and then equate the two prices to determine which normal volatility
N
needs to be used to
reproduce the option price under the SABR model. That is, we find the implied normal
volatility of the option under the SABR model.
Second comparison: Lognormal model
By doing a second comparison between option prices under the lognormal model
dF =
B
FdW, (8.3.118)
and the normal model (8.4.1), we then convert the implied normal volatlity to the usual implied
lognormal (Black-Scholes) volatility, i.e., we quote the option price predicted by the SABR model
in terms of the options implied volatility.
8.4 Normal model
Equations (8.3.116a)-(8.3.116b) form a general formula for the dollar price of the call option
under the SABR model. The utility and beauty of this formula is not overwhelmingly apparent.
To obtain a useful formula, we convert this dollar price into the equivalent implied volatilities.
To cast the results in a more usable form, we re-compute the option price under the normal
model
dF =
N
dW, (8.4.1)
and then equate the two prices to determine which normal volatility
N
needs to be used to
reproduce the option price under the SABR model. That is, we find the implied normal
volatility of the option under the SABR model.
Suppose we repeated the above analysis for the ordinary normal model
dF = C(F)dW
1
, F(0) = f. (8.4.2)
Here, we set
C(f) = 1 B(z) = 1, B(0) = 1, b
1
=
B
B
= 0 and b
2
=
B
B
= 0. (8.4.3a)
=
N
, (8.4.3b)
= 0
N
is constant, not stochastic. (8.4.3c)
As 0 we have to apply LHopitals rule for x to avoid
0
0
. This yields
x = lim
0
1
z
_
1 2z +
2
2
z
2
= lim
0
z
_
1 2z +
2
2
z
2
= z. (8.4.4a)
Furthermore,
2
:= log
_
z
f K
_
B(0)B(z
_
+ log
_
xI
1
2
z
_
+
1
4
2
b
1
z
2
,
= log
_

N
f K
z
_
+ log
_
I
1
2
_
+ 0 = log
_

N
f K
zI
1
2
_
= log
_
I
1
2
_
= log
_
(1 2z +
2
2
z
2
)
1
4
_
=
1
4
log(1) = 0, (8.4.4b)
for = 0.
Hence, the lower limit of the integral in formula (8.3.116a)
x
2
2
ex
2
(8.4.5a)
can be replaced by
(f K)
2
2
2
N
ex
, (8.4.5b)
such that the option value for the normal model is exactly
V (t, f, ) = max(f K, 0) +
[f K[
4
_

(fK)
2
2
2
N
ex
q
3
2
e
q
dq. (8.4.6)
8.4. NORMAL MODEL 83
Working out the integral and applying the property N(a) = 1 N(a) for all a then yields the
exact European option price
V (t, f, ) = (f K)N
_
f K
ex
_
+
N
ex
n
_
f K
ex
_
(8.4.7)
for the normal model.
From equation (8.4.6) it is clear that the expression for the option price under the normal model
matches the general expression for the option price under the SABR model (8.3.116a)-(8.3.116b)
if and only if the lower integral limits are the same, i.e.,
x
2
2
ex
2
=
(f K)
2
2
2
N
ex
. (8.4.8)
This implies that the normal volatility
N
satisfies
1
2
N
=
_
x
2
2
ex
2
2
_
2
ex
(f K)
2
=
x
2
(f K)
2
_
1
2

x
2
ex
_
. (8.4.9)
Taking the square root and expanding
1
1
2

x
2
ex
(8.4.10)
in a Taylor series around = 0 shows that the options implied normal (absolute) volatility
is given by
N
=
f K
x
_
1 +
2

x
2
ex
+. . .
_
. (8.4.11)
Since x = z (1 +O()), we can rewrite the answer as
12
N
=
_
f K
z
__
z
x(z)
_
_
1 +
2
(
1
+
2
+
3
)
ex
+. . .
_
, (8.4.12)
where
f K
z
=
(f K)
_
f
K
df
C(f
)
=
_
1
f K
_
f
K
df
C(f
)
_
1
. (8.4.13)
On page 28 of their paper [5] Hagan and his coauthors state: This factor represents the average
dif f iculty in dif fusing from todays forward f to the strike K, and would be present even if the
volatility were not stochastic. This probably means that this factor is the main multiplier for
the volatility and its appearence does not depend on the value of the volvol .
The next factor is
z
x(z)
=
z
1
log
_
12z+
2
2
z
2
+z
1
_ =

log
_
12+
2
+
1
_, (8.4.14)
12
This can be seen by making a Taylor expansion of the definition of x (8.3.88) around = 0 up to and including
O(). This yields x = z
_
1 +
1
2
z
_
= z (1 +O()).
where
:= z =
1
_
f
K
df
C(f
)
=

_
f
K
df
C(f
)
=

f K
C (f
avg
)
_
1 +O
_
2
__
. (8.4.15)
Here, f
avg
:=

fK is the geometric average of f and K.
13
On page 28 of [5] Hagan and col-
leagues claim: This factor represents the main ef fect of the stochastic volatility. With this
they presumably mean that in the expression for the normal volatility (8.4.12) the lowest order
term in is driven by f
avg
.
The coefficients
1
,
2
and
3
provide relatively minor corrections. Through O
_
2
_
these
corrections are
1
=
1
x
2
log
_
z
f K
_
B(0)B(z)
_
, (8.4.16a)
2
=
1
x
2
log
_
xI
1
2
z
_
, (8.4.16b)
3
=
1
x
2
4B
. (8.4.16c)
i.e., using x = z (1 +O()) as we did before,
1
=
1
z
2
log
_
z
f K
_
C(f)C(K)
_
=
2
2
2
1
24

2
2
C
2
(f
avg
) +. . . , (8.4.17a)
2
=
1
z
2
log
_
x
z
_
1 2z +
2
2
z
2
_1
4
_
=
2 3
2
24

2
2
+. . . , (8.4.17b)
3
=
1
z
2
4B
=
1
4z
2
1
C (f
avg
) +. . . , (8.4.17c)
where
1
:=
C
(f
avg
)
C(f
avg
)
and
2
:=
C
(f
avg
)
C (f
avg
)
. (8.4.18)
Let us briefly summarize before continuing. Under the normal model, the value of a European
call option with strike K and exercise date
ex
is given by
V (t, f, ) = (f K)N
_
f K
ex
_
+
N
ex
_
f K
ex
_
. (8.4.19)
For the SABR model,
_
d

F = C(

F) dW
1
,

F(0) = f,
d = dW
2
, (0) = ,
(8.4.20a)
(8.4.20b)
where
E[dW
1
dW
2
] = dt. (8.4.20c)
The value of the call option is given by the same formula, at least through O
_
2
_
, provided we
use the implied normal volatility
N
(K) =
(f K)
_
f
K
df
C(f
_

x()
_
(8.4.21a)
_
1 +
_
2
2
2
1
24

2
C
2
(f
avg
) +
2 3
2
24

2
+
1
4
1
C (f
avg
)
_
ex
+. . .
_
.
13
Note that the arithmetic average
f+K
2
could have been used equally well at this order of accuracy.
8.5. LOGNORMAL MODEL 85
Here,
f
avg
=
_
fK,
1
=
C
(f
avg
)
C (f
avg
)
,
2
=
C
(f
avg
)
C (f
avg
)
, (8.4.21b)
=

f K
C (f
avg
)
, x() = log
_
_
1 +
2
+
1
_
. (8.4.21c)
The first two factors provide the dominant behaviour, with the remaining factor 1 + [. . .]
2
ex
usually providing small corrections
14
.
One can repeat the analysis for a European put option, or simply use put-call parity as we have
seen before in section 4.6. Because in this case we consider a forward f (instead of a stock S),
there is no discount factor e
r(Tt)
needed. Hence, the put-call parity is given by
V
put
= V
call
+K f, (8.4.22)
Using the fact that N(a) = 1 N(a) and n(a) = n(a) for all values of a, we obtain the
following expression for the value of a European put option with strike K and exercise date
ex
:
V
put
= (f K)N
_
f K
ex
_
+
N
ex
_
f K
ex
_
+K f,
= (K f)N
_
K f
ex
_
+
N
ex
_
K f
ex
_
, (8.4.23)
where the implied volatility
N
is given by the same formulas (8.4.21a)-(8.4.21c) as the call.
8.5 Lognormal model
By doing a second comparison between option prices under the lognormal model
dF =
B
FdW, (8.5.1)
and the normal model (8.4.1), we then convert the implied normal volatlity to the usual im-
plied lognormal (Black-Scholes) volatility, i.e., we quote the option price predicted by the SABR
model in terms of the options implied volatility. Here C(F) = F.
To derive the implied Black volatility, consider Blacks model
dF =
B
FdW, (8.5.2)
where we have written the volatility as
B
to stay consistent with the preceding analysis. For
Blacks model, the value of a European call with strike K and exercise date
ex
is
V
call
= fN(d
+
) KN(d
), (8.5.3a)
V
put
= V
call
+D(t
set
)(K f), (8.5.3b)
with
d
=
log(
S
K
)
1
2
2
B
ex
ex
, (8.5.3c)
where we are omitting the overall factor D(t
set
) as before.
14
Hagan et al. claim: These are corrections of around 1% or so.
We can obtain the implied normal volatility for Blacks model by repeating the preceding anal-
ysis for the SABR model with C(f) = f and = 0.
Setting C(f) = f and = 0 in equation (8.4.21a)-(8.4.21c) shows that
N
(K) =
(f K)
log(f/K)
_
1
1
24
2
B
ex
+. . .
_
(8.5.4)
through O
_
2
_
.
When we equate the two formulas for
N
(K) (8.5.4) and (8.4.21a), and multiply through by
_
log(f/K)
f K
_
1
_
1 +
1
24
2
B
ex
_
(8.5.5)
to clear up the left-hand side, we obtain
B
=
log(f/K)
_
f
K
df
C(f
_

x()
_
_
1 +
1
24
2
B
ex
_
(8.5.6)
_
1 +
_
2
2
2
1
24

2
C
2
(f
avg
) +
2 3
2
24

2
+
1
4
1
C (f
avg
)
_
ex
+. . .
_
.
Remember that we are only working through O
_
2
_
, so we can neglect any higher order terms
that arise. The final answer, which is an expression for
B
is obtained by substituting for
B
its first order approximation [4]
B
=
C(f
avg
)
f
avg
(1 +O()) , (8.5.7)
so we can replace
_
1 +
2
2
B
ex
(8.5.8)
on the right-hand side by
_
1 +
2
_
1
24
2
C(f
avg
)
2
1
f
2
avg
_
ex
_
(8.5.9)
which can be combined with the other
2
ex
terms. Through O
_
2
_
this yields
B
(K) =
log(f/K)
_
f
K
df
C(f
_

x()
_
(8.5.10)
_
1 +
_
2
2
2
1
+
1
f
2
av
24

2
C
2
(f
avg
) +
2 3
2
24

2
+
1
4
1
C (f
avg
)
_
ex
+. . .
_
.
This is the main result of the paper written by Hagan and colleagues [5].
8.6 Stochastic model
As originally stated, the SABR model consists of the special case C(f) = f
. The model then

becomes
_
d

F =

F
dW
1
,

F(0) = f,
d = dW
2
, (0) = ,
(8.6.1a)
(8.6.1b)
8.6. STOCHASTIC MODEL 87
where
E[dW
1
dW
2
] = dt. (8.6.1c)
For C(f) = f
, we have
1
=
C
(f
avg
)
C(f
avg
)
=
f
1
avg
f
avg
= f
1
avg
and (8.6.2a)
2
=
C
(f
avg
)
C(f
avg
)
= ( 1)
f
2
avg
f
avg
= ( 1)f
2
avg
. (8.6.2b)
Substituting this into equation (8.4.21a)-(8.4.21c) shows that the normal volatility for this model
is
N
(K) =
(f K)
_
f
K
f
df
_

x()
_
_
1 +
_
2( 1)f
2
avg

2
f
2
avg
24

2
f
2
avg
+
2 3
2
24

2
+
1
4
f
1
avg
f
avg
_
ex
+. . .
_
,
=
(1 )(f K)
f
1
K
1

_

x()
_
_
1 +
_
(2 )
2
24f
22
av
+
2 3
2
24

2
+
1
4
4f
1
av
_
ex
+. . .
_
, (8.6.3a)
through O
_
2
_
with f
avg
=
fK as before and
=

f K
f
av
and x() = log
_
1 2 +
1
_
. (8.6.3b)
This can be simplified by expanding
15
f K =
_
fK log
_
f
K
__
1 +
1
24
log
2
_
f
K
_
+
1
1920
log
4
_
f
K
_
+. . .
_
,
f
1
K
1
= (1 )(fK)
(1)/2
log
_
f
K
_
_
1 +
(1 )
2
24
log
2
_
f
K
_
+
(1 )
4
1920
log
4
_
f
K
_
+. . .
_
.
Here, terms higher than fourth order can be neglected, because f K = O() and therefore
log
_
f
K
_
is small.
15
Define x := log(f/K) and write
f K =
_
fK
_
_
f
K

_
K
f
_
= 2
_
fK
_
e
1
2
x
e
1
2
x
_
2
= 2
_
fK sinh
_
x
2
_
,
=
_
fKx
_
1 +
x
2
3! 2
3
+
x
4
5! 2
4
+. . .
_
=
_
fK log
_
f
K
__
1 +
1
24
log
2
_
f
K
_
+
1
1920
log
4
_
f
K
_
+. . .
_
.
A similar approach can be used to expand f
1
K
1
.
Then, the implied normal volatility (8.6.3a) reduces to
N
(K) = (fK)
/2
1 +
1
24
log
2
_
f
K
_
+. . .
1 +
(1)
2
24
log
2
_
f
K
_
+. . .
_

x()
_
(8.6.5a)
_
1 +
_
(2 )
2
24(fK)
1
+
2 3
2
24

2
+
1
4
(fK)
(1)/2
_
ex
+. . .
_
,
where
=

(fK)
(1)/2
log
_
f
K
_
and x() = log
_
_
1 2 +
2
+
1
_
. (8.6.5b)
Equating the above formula (8.6.5a) for
N
(K) for the SABR model and the expression for
N
(K) for Blacks model (8.5.4) and multiplying through by
_
log(f/K)
f K
_
1
_
1 +
1
24
2
B
ex
_
(8.6.6)
yields
B
=

(fK)
(1)/2
1 +
1
24
2
B
ex
1 +
(1)
2
24
log
2
_
f
K
_
+. . .

x()
(8.6.7)
_
1 +
_
(2 )
2
24(fK)
1
+
2 3
2
24

2
+
1
4
(fK)
(1)/2
_
ex
+. . .
_
.
Next
B
has to be replaced by
B
=
C(f
avg
)
f
avg
(1 +O()) (8.6.8)
as we did before, such that we can replace
1 +
1
24
2
B
ex
(8.6.9)
on the right-hand side by
1 +
1
24
2
f
21
avg

ex
(8.6.10)
Hence, the expression for the implied Black volatility for the SABR model (8.6.8) becomes
16
B
=

(fK)
(1)/2
1
1 +
(1)
2
24
log
2
_
f
K
_
+. . .

x()
(8.6.11)
_
1 +
_
(1 )
2
2
24(fK)
1
+
2 3
2
24

2
+
1
4
(fK)
(1)/2
_
ex
+. . .
_
.
through O
_
2
_
, where and x() are given by equation (8.6.5b) as before.
16
In equation (B.69c) on page 31 of [5], Hagan et al. claim that there should also be an in the numerator of
the first fraction. This must be a typographical mistake, because in equation (2.17a) on page 9 of the same paper
[5] no occurs.
8.7. SPECIAL CASES: = 0 AND = 1 89
8.7 Special cases: = 0 and = 1
There are two special cases we will take a special look at: the stochastic normal model ( = 0)
and the stochastic lognormal model ( = 1).
Normal model: = 0
For the stochastic normal model ( = 0), the implied volatilities of European call and put
options are
N
(K) =
_
1 +
2 3
2
24

2
ex
+. . .
_
, (8.7.1a)
B
(K) =
log
_
f
K
_
f K
_

x()
__
1 +
_

2
24fK
+
2 3
2
24

2
_
ex
+. . .
_
, (8.7.1b)
through O
_
4
_
, where
=

_
fK log
_
f
K
_
and x() = log
_
_
1 2 +
2
+
1
_
. (8.7.1c)
Lognormal model: = 1
For the stochastic lognormal model ( = 1), the implied volatilities are
N
(K) =
f K
log
_
f
K
_
_

x()
__
1 +
_
2
24
+

4
+
2 3
2
24

2
_
ex
+. . .
_
, (8.7.2a)
B
(K) =
_

x()
__
1 +
_
4
+
2 3
2
24

2
_
ex
+. . .
_
, (8.7.2b)
through O
_
4
_
, where
=

log
_
f
K
_
and x() = log
_
_
1 2 +
2
+
1
_
. (8.7.2c)
8.8 Discussion
The most important assumptions that are made in the paper by Hagan et al. [5] are presented
below. Furthermore, the method and solution will be discussed in this section.
Method and assumptions
Hagan and his colleagues use singular perturbation techniques to obtain the plain-vanilla op-
tion prices implied by the SABR model, and from these the associated implied volatilities. In
their book [9] Rebonato, McKay and White explain that Implied volatilities are just the wrong
number to put in the wrong formula to get the right price, so there is no great fundamental
meaning in obtaining implied volatilities rather than prices. However, for very good reasons,
these wrong numbers have become the common metric in the market place to communicate the
prices of options.
The SABR model describes a single forward F. The volatility of this forward F is described
by the parameter , which itself follows a stochastic process, while the forward is assumed to
follow the CEV process as before in section 6. That is,
_
dF = C(F)dW
1
, F(0) =f,
d = dW
2
, (0) =.
(8.8.1a)
(8.8.1b)
The two Wiener processes are assumed to be correlated by
E[dW
1
dW
2
] = dt. (8.8.2)
All the parameters of the model, , and , are assumed to be constants, not functions of time,
and there is no mean-reversion in the stochastic process for volatility.
Both the volatility and the volvol are assumed to be small. This results in the following
scalings:
=

, = , (8.8.3a)
=

, = . (8.8.3b)
The pde approach does not seem to work very well in this case, because after scaling we obtain
pdes that are dicult to solve. Hence, for the SABR model it is better to follow the procedure
of Hagan et al., who directly apply the scalings on the system of stochastic dierential equations
(8.8.1) and use Kolmogorov equations, as we did in section 8.2 of this thesis.
The resulting expression for the implied volatility under the SABR model is obtained by con-
sidering the forward and backward Kolmogorov equations per order in , making smart choices
for local variables and functions in order to attempt to transform them into an equation that
looks like a heat equation, which is easier to solve.
Solution
The SABR model is very well explained in the paper by Hagan et al. [5]. Singular perturbation
techniques are used to obtain the prices of European options under the SABR model. From
these prices we obtain the following closed-form algebraic formula for the implied volatility as a
function of todays forward price f and the strike K:
B
=

(fK)
(1)/2
1
1 +
(1)
2
24
log
2
_
f
K
_
+. . .

x()
(8.8.4a)
_
1 +
_
(1 )
2
2
24(fK)
1
+
2 3
2
24

2
+
1
4
(fK)
(1)/2
_
ex
+. . .
_
through O
_
2
_
, where and x() are given by
=

(fK)
(1)/2
log
_
f
K
_
and x() = log
_
_
1 2 +
2
+
1
_
. (8.8.4b)
Our opinion
Though the SABR model is very well explained in the paper by Hagan et al. [5], it took us
quite some time and eort to understand all derivations. Fortunately, Patrick Hagan was very
helpful in answering our questions via e-mail and explaining the steps we did not immediately
understand. This helped us a lot in comprehensing the details of this particular application of
perturbation theory.
Chapter 9
Conclusions
This section summarizes the main results and conclusions that we have obtained in this thesis.
Also some recommendations for further investigation will be presented below.
This thesis discusses the use of perturbation theory in the context of financial mathematics, in
particular on the use of matched asymptotic expansions in option pricing. Our methods are
applied to the ordinary Black-Scholes model for illustration, and two more advanced models
based on papers by Howison [7] and Hagan et al. [5].
Black-Scholes model
A first application of perturbation theory on a financial model has been presented in section
5 in order to show the techniques and complications of the method of asymptotic expansions
in a financial context. In this simple example of the Black-Scholes model an exact solution is
available, so it is in fact not neccessary to apply the method of asymptotic expansions on this
model. However, in case we do apply the method, we can construct two artificial layers, and
make smart choices for the local variables, in order to attempt to transform the equations into
a heat equation, which can be solved. A nice property of this model is that it is possible to
compare the results with the exact solution, to see that it is a very accurate method. Note
that this exact solution can be obtained by transforming the Black-Scholes equation into a heat
equation (as we have shown in section 5.2).
Fast mean-reverting stochastic volatility model
Howisons paper [7] discusses a fast mean-reverting stochastic volatility model that turns out to
have many open ends. In this paper quite a lot of assumptions and simplifications are made.
cases we even think these assumptions and simplifications are incorrect.
An important simplification is that Howison makes some assumptions about -independence
of the O(1) solution V
0
. He argues that this choice follows naturally from the behaviour of V
0
in the limits for small and large S, but he doesnt explain this. This simplification has large
consequences for the higher order equations and their solutions. If the O(1) solution would
depend on , these higher order equations would be completely different and it would make
solving them a lot more complicated.
Another important point of discussion is that Howison has chosen to make the derivative with
respect to time to be O(), instead of O(1). As a consequence, the remaining O(1) equation does
not look like a heat equation anymore, because in the first order approach in equation (7.3.9),
91
92 CHAPTER 9. CONCLUSIONS
the time derivative has been left out of the problem. The similarities with the application of
perturbation theory on the Black-Scholes model have thus dissappeared here. Howison does not
explicitely motivate his choice. A reason for this choice might be that it is likely that taking
the time derivative into account in the O(1) problem leads to a more complicated first order
solution V
0
, which does not only depend on t, but also on t. In that case, a totally different
approach, using multiple timescales, would be needed.
After applying the method of matched asymptotic expansions to the fast mean-reverting stochas-
tic volatility model and following Howisons assumptions, we have obtained the following ex-
pression for the three lowest order terms of the solution:
V (S, , t) c
0
(S, t) +
1
2
(T t)
_
A1
2
,1
+A1
2
,2
T
_
S
2
_
T
2
T
_
c
0
(9.0.1)
+
_
g
1
() g
1
() (T t) (A
1,1
+A
1,2
T)
1
2
(T t)
2
_
A1
2
,1
+A1
2
,2
T
_
T
2
T
_
__
A1
2
,1
+A1
2
,2
T
_
_
_
T
2
T
_
c
0
,
where T := S

S
. However, since the function c
0
(S, t) is still undetermined, this result doesnt
seem to be very useful.
Lognormal-normal model
The next model that has been considered, is a lognormal stock process with a normal volatility
process, also known as the Schobel-Zhu model, given by
_
dS = S dt +S dW, dW N(0, dt),
d = a dt +b d
W, d
W N(0, dt).
(9.0.2a)
(9.0.2b)
Here, a, b R are constants and the stochastic processes W and

W have correlation .
This model was supposed to be a first step to the SABR model. We expected it to be easier
than SABR, because the drift and volatility of volatility were chosen to be constant. However,
after trying to apply the method of asymptotic expansions to this model, unfortunately this
appeared to be not that easy. In particular, the inner equations were hard to solve, because
these partial differential equations contained coefficients that depend on the time and space
parameters. Therefore, we have decided to abandon this model and directly continue with the
SABR model.
SABR model
The paper written by Hagan and colleagues [5] examines a new three-parameter stochastic vola-
tility model (the SABR model) that successfully prices back the volatility smile as observed in
the market nowadays, and that is commonly used. This resulting expression for the implied
volatility under the SABR model is obtained by considering the forward and backward Kol-
mogorov equations per order in , making some smart choices for local variables and functions
in order to attempt to transform them into an equation that looks like a heat equation, which
is easier to solve. Note that, contrary to Howisons approach, Hagan et al. do take into account
the first derivative with respect to time in their O(1) equation.
The main result of this section is the following closed-form algebraic formula for the implied
93
volatility as a function of todays forward price f and the strike K:
B
=

(fK)
(1)/2
1
1 +
(1)
2
24
log
2
_
f
K
_
+. . .

x()
(9.0.3a)
_
1 +
_
(1 )
2
2
24(fK)
1
+
2 3
2
24

2
+
1
4
(fK)
(1)/2
_
ex
+. . .
_
through O
_
2
_
, where and x() are given by
=

(fK)
(1)/2
log
_
f
K
_
and x() = log
_
_
1 2 +
2
+
1
_
. (9.0.3b)
Recommendations
Recommendations for further investigation on these models would be to consider several different
scalings and see which one works best.
For example, using multiple-scale methods to find the correct O(1) equation and solution for
Howisons fast mean-reverting volatility model. Also, the -independence of the O(1) solution
V
0
should be reconsidered to see if it is possible to obtain a first order approach that does depend
on .
For the SABR model, also some choices are made for the scalings. Here, a nice suggestion for
further investigation would be to see if any other small parameters occur in this model. For
example, to examine what happens if (also) (1 ) and/or (
1
2
are small.
94 CHAPTER 9. CONCLUSIONS
Bibliography
[1] M. Baxter and A. Rennie. Financial Calculus, an introduction to derivative pricing. Cam-
bridge University Press, 1996.
[2] Giuseppe Da Prato. Kolmogorov Equations for Stochastic PDEs. Birkhauser, 2004.
[3] Jim Gatheral. The volatility surface, a practioners guide. John Wiley & Sons, Inc., 2006.
[4] Patrick S. Hagan. Private communication.
[5] Patrick S. Hagan, Deep Kumar, Andrew S. Lesniewski, and Diana E. Woodward. Managing
smile risk. 2002.
[6] M.H. Holmes. Introduction to Perturbation Methods. Springer-Verlag, 1995.
[7] S. Howison. Matched asymptotic expansions in nancial engineering. 2005.
[8] S. Howison and D. Bakstein. A non-arbitrage liquidity model with observable parameters
for derivatives. 2004.
[9] Riccardo Rebonato, Kenneth McKay, and Richard White. The SABR/LIBOR Market
Model. John Wiley & Sons, Inc., 2009.
[10] Steven E. Shreve. Stochastic calculus for nance. Springer Science + Business Media, LLC,
2004.
[11] Walter A. Strauss. Partial Dierential Equations - An Introduction, Second Edition. John
Wiley & Sons, Ltd, 2008.
[12] Ferdinand Verhulst. Methods and Applications of Singular Perturbations. Springer, 2005.
95
96 BIBLIOGRAPHY
Appendix A
List of symbols
Greek symbols
: Second order derivative of the option price, one of the Greeks.
: Parameter in the CEV model.
: First order derivative of the option price, one of the Greeks.
Used for heding, to determine the amount of stock that has to be in the portfolio.
: Dividend.
: Stretching parameter.
t
: Drift.
t
: Volatility.
Latin symbols
B
t
: Bond price.
D(t): Discount factor for time t.
K: Strike price.
N(): Standard normal cumulative distribution function, N(x) =
1
2
_
x
z
2
2
dz.
N(, ): Normal distribution with mean and variance
2
.
n(): Standard normal probability density function, n(x) =
1
2
e
x
2
2
.
P(S): Payoff.
r
t
: Risk-free interest rate.
S
t
: Stock price.
T: Expiry time.
t: Time.
V : General option price.
V
call
: Price of a call option.
V
put
: Price of a put option.
W
t
: Wiener Process.
97
98 APPENDIX A. LIST OF SYMBOLS
Appendix B
Order symbols
To define an asymptotic approximation, first the order symbols
1
need to be introduced. The
reason for this is that we will be interested in how functions behave as a parameter, typically ,
becomes small.
For example, the function f() = does not converge to zero as fast as g() =
2
when 0.
Therefore, a notation to denote this fact is needed.
Def initions
g = O(f) as
0
means that there are constants k
0
and
1
(independend of ), such that
[g()[ k
0
[f()[ for
0
< <
1
. (B.0.1)
We say that g is big Oh of f as
0
.
g = o(f) as
0
means that for every positive , there is an
2
(independent of ), such
that
[g()[ [f()[ for
0
< <
2
. (B.0.2)
We say that g is little oh of f as
0
.
f() = O
s
(g()) if f = O(g) and f ,= o(g) for 0.
Another useful way to make this determination involves the limit
l := lim
0
g()
f()
. (B.0.3)
If this limit l exists and is finite, then g = O(f) as
0
.
Similarly, if l = 0, then g = o(f) as
0
.
1
Also referred to as Landau symbols.
99
100 APPENDIX B. ORDER SYMBOLS
Appendix C
Example problem in perturbation
theory: layer at x = 1
In chapter 2 the following example problem is being considered:
y
+ 2y
+ 2y = 0, for 0 < x < 1, (C.0.1)

with boundary conditions y(0) = 0 and y(1) = 1.
Step 1: Outer solution
If the layer is chosen to be at x = 1, the O(1) outer solution y
0
(x) = ae
x
should satisfy the
boundary condition at x = 0, i.e., y(0) = 0. This is only the case if we choose the arbitrary
constant a to be equal to zero, such that the outer solution is given by y
0
(x) = 0.
Step 2: Boundary layer analysis
Inside the layer at x = 1 we have a local variable
=
x 1
()
. (C.0.2)
After substitution of this variable in problem (C.0.1), we obtain
()
2
2
y
2
+
2
()
y
+ 2y
= 0. (C.0.3)
Maximum balance yields () = , such that equation (C.0.3) transforms into
2
y
2
+ 2
y
+ 2y
= 0. (C.0.4)
Expanding y
= y
0
+y
1
+. . . gives the following O(1) equation:
2
y
2
+ 2
y
= 0. (C.0.5)
The general solution of the O(1) equation is y
0
() = Ae
2
+B. Here, the boundary condition
at x = 1 should be satisfied, i.e., at =
x1
= 0 we have y
0
(0) = 1. This implies that A+B = 1,
such that the inner solution becomes y
0
() = A
_
e
2
1
_
+ 1.
101
102APPENDIXC. EXAMPLE PROBLEMINPERTURBATIONTHEORY: LAYER ATX = 1
Step 3: Matching
To determine the arbitrary constant A in the inner solution, a matching condition is needed.
Letting , which means going towards the boundary of the layer, yields A = 0, to
avoid that the inner solution would become infinitely large. The inner solution thus becomes
y
0
() = 1. Unfortunately, in this case we are not able to match this with the outer solution
y(x) = 0.
Conclusion
The boundary layer cannot be located at x = 1. Therefore, let us asume it to be at x = 0.
The importance of matching cannot be overemphasized: It is one of the essential steps. If the
inner and outer solution do not match, it is necessary to go back and determine where the error
was made. The possibilities where this happens are almost endless. On page 56 of the book
written by Holmes [6] theres a list of useful places to start to look.
Appendix D
Feynman-Kac
Richard Feynman and Mark Kac have established a link between partial differential equations
(pdes) and stochastic processes. It offers a method of solving certain pdes by simulating random
paths of a stochastic process. Suppose we are given the pde:
f
t
+g(x, t)
f
x
+
1
2
2
(x, t)
2
f
x
2
= 0, (D.0.1)
subject to the boundary condition f(x, T) = (x), then the Feynman-Kac formula reads:
f(x, t) = E [(X
T
)[T
t
] . (D.0.2)
Here, X is an Ito process driven by the equation
dX = g(X, t) dt +(X, t) dW
t
, (D.0.3)
with W
t
a Wiener process and the initial value for X(t) is X(0) = x.
Proof of the Feynman-Kac formula
1
The pde for f(x, t) is given, so using Itos formula (4.3.1) on f we obtain:
df =
f
t
dt +
f
x
dx +
1
2
2
f
x
2
dx
2
,
=
_
f
t
+g(x, t)
f
x
+
1
2
2
(x, t)
2
f
x
2
_
dt +(x, t) dW
t
,
= (x, t) dW
t
.
Integrating both sides gives
_
T
t
df = f(X
T
, T) f(x, t) =
_
T
t
(x, t)
f
x
dW
t
. (D.0.4)
Taking the expectation, we nd
f(x, t) = E [f(X
T
, T)] = E [(X
T
)] . (D.0.5)
1
Source: Lecture notes of the TU Delft course Computational Finance (WI4154), lecture 04, by Lech A. Grze-
lak and C.W. Oosterlee.
103
104 APPENDIX D. FEYNMAN-KAC
Appendix E
Notes on L
2
functions
Consider an L
2
function f, and the expectation value of its Ito integral
E
__
T
t
f(W
t
) dW
t
_
, (E.0.1)
which is named after Kiyoshi Ito. See figure E.1 below.
Figure E.1: Ito integral
Using the fact that the increment W
t
i+1
W
t
i
is independent of W
t
i
, such that we can split up
the expectation, yields
E
__
T
t
f(W
t
) dW
t
_
= E
_
t
i
f(W
t
i
)
_
W
t
i+1
W
t
i
_
_
,
=
t
i
E[f(W
t
i
)] E
_
W
t
i+1
W
t
i
.
Because E[f(W
t
i
)] < and E
_
W
t
i+1
W
t
i
= 0, we conclude that
E
__
T
t
f(W
t
) dW
t
_
= 0. (E.0.2)
105
106 APPENDIX E. NOTES ON L
2
FUNCTIONS
Appendix F
Martingales
In probability theory, a martingale is a stochastic process
1
, such that the conditional expected
value of an observation at some time t, given all observations up to some earlier time s < t, is
equal to the observation at that earlier time s. The concept of martingale in probability theory
was introduced by Paul Pierre Levy.
Def inition of a martingale
A discrete-time martingale is a discrete-time stochastic process X
1
, X
2
, X
3
, . . . that satisfies
E([X
n
[) < , (F.0.1)
E(X
n+1
[X
1
, . . . , X
n
) = X
n
, (F.0.2)
for all n 0.
Example
Suppose X
n
is a gamblers fortune after n tosses of a fair coin, where the gambler wins $1 if the
coin comes up heads and loses $1 if the coin comes up tails. The gamblers conditional expected
fortune after the next trial, given the history, is equal to his present fortune. So this sequence
is a martingale.
Properties
Perfectly tradable goods, like shares of stock, are subject to the law of one price: they should
cost the same amount wherever they are bought. This law requires an ecient and liquid mar-
ket. Any discrepancy that may exist in pricing perfectly tradable goods, will lead to an arbitrage
opportunity. Goods that cannot be costlessly traded are not subject to this law.
Here, a nice property is that martingales are tradables, and non-martingales are non-tradables.
More information on this statement can be found on page 116-118 of the book writen by Baxter
and Rennie ([1]).
1
I.e., a sequence of random variables X
1
, X
2
, X
3
, . . ..
107
108 APPENDIX F. MARTINGALES
Appendix G
Similarity solutions for PDEs
The technique of similarity solutions is one of the various techniques for reducing a partial
differential equation (pde) into an ordinary differential equation (ode), or at least turn the
original pde into another pde, reducing the number of independent variables.
This is an approach that identifies equations for which the solution depends on certain group-
ings of the independent variables rather than depending on each of the independent variables
separately.
First consider the heat equation
u
t
D
2
u
x
2
= 0. (G.0.1)
We introduce the so-called dilation transformation
z =
a
x,
s =
b
t,
v(z, s) =
c
u
_
a
z,
b
s
_
.
(G.0.2)
So the heat equation (G.0.1) transforms into
bc
v
s
D
2ac
2
v
z
2
= 0. (G.0.3)
Hence, for b c = 2ac (i.e., b = 2a), this equation is invariant under this transformation. So if
u(x, t) solves the heat equation in the variables x, t, then for z, s, v(z, s) as given v(z, s) solves
the heat equation in the variables z, s.
Note that
vs
c/b
= (
c
u)
_
b
t
_
c/b
= ut
c/b
, and
z
s
a/b
=

a
x
(
b
t)
a/b
=
x
t
a/b
,
such that both groupings of variables are invariant under the transformation (G.0.2) for all
choices of a, b, c. This suggests that we look for a solution of the heat equation (G.0.1) that is
of the form
u = t
c/b
y() for =
x
t
a/b
=
x
t
, since
a
b
=
1
2
, (G.0.4)
109
110 APPENDIX G. SIMILARITY SOLUTIONS FOR PDES
which gives
u
t
= t
c/2a1
_
c
2a
y()

2
y
()
_
,
u
x
= t
c/2a1/2
y
(),
2
u
x
2
= t
c/2a1
y
(),
such that the heat equation (G.0.1) transforms into
t
c/2a1
_
Dy
()
c
2a
y() +

2
y
()
_
= 0. (G.0.5)
So the pde (G.0.1) has indeed been reduced to the ode (G.0.5). A non-zero solution y() sat-
isfying (G.0.5) is called a similarity solution of the heat equation (G.0.1).
If the heat equation (G.0.1) is satisfied for x > 0, t > 0 and if u(x, t) satisfies u(x, 0) = 0 for
x > 0, u(x, t) for x and
u
x
(0, t) = Q for t > 0, then it follows that for
x or t 0 and = 0 if x = 0, because y() = t
c/2a
u(x, t) and = x/
t.
For y this gives y() = 0 and
u
x
(0, t) = t
c/2a1/2
y
(0) = Q if and only if c = a, because Q is

a constant. So the initial boundary value problem for u(x, t) reduces to the following problem
for y():
_
_
_
Dy
()
1
2
y() +

2
y
() = 0,
y
(0) = Q,
y() 0 as .
(G.0.6)
If the boundary condition at x = 0 is given by u(0, t) = u
0
then u(0, t) = t
c/2a
y(0) = u
0
if and
only if c = 0, because u
0
is a constant. In this case the initial boundary value problem for u(x, t)
reduces to
_
_
_
Dy
() +

2
y
() = 0,
y(0) = u
0
,
y() 0 as .
(G.0.7)
We can integrate the equation Dy
() +

2
y
() = 0 once to obtain y
() = c
1
e
2
/4D
, such that
y() = c
1
_

0
e
2
/4D
d +c
2
= c
3
erf
_

4D
_
+c
2
, (G.0.8)
where erf(x) :=
2
_
x
0
e
2
d. This has the property that erf() = 1.
The boundary conditions now lead to
y() = u
0
u
0
erf
_

4D
_
,
u(x, t) = u
0
u
0
erf
_
x
4Dt
_
= u
0
erfc
_
x
4Dt
_
,
where erfc(s) = 1 erf(s).
111
The solution of problem (G.0.6) is given by
y() = C
1
+C
2
_
2
De
2
/4D
+ erf
_

4D
__
. (G.0.9)
And since y
() = C
1
+C
2
_
erf
_

4D
__
, the other conditions are satisfied by
y() = Q
Q
_
2
De
2
/4D
+ erf
_

4D
__
,
= Q erfc
_

4D
_
2Q
_
D
2
/4D
.
112 APPENDIX G. SIMILARITY SOLUTIONS FOR PDES
Appendix H
Greens functions
Greens functions are named after the British mathematician George Green. They are the basic
solutions to linear differential equations. A Greens function is a building block that can be
used to construct many useful solutions.
Consider differential equations of the form
Lu(x) = f(x), (H.0.1)
where L = L(x) is a linear differential operator acting on distributions over a subset of the
Euclidean space R
n
, at a point s.
Denition
A Greens function
1
G(x, s) of a linear differential operator L(x) is any solution of
LG(x, s) = (x s), (H.0.2)
where () is the Dirac delta function.
Motivation
If a function G(x, s) can be found for an operator L, then we multiply equation (H.0.2) by f(s)
and integrate with respect to s to obtain
_
LG(x, s)f(s)ds =
_
(x s)f(s) ds = f(x). (H.0.3)
By equation (H.0.1), the right hand side of equation (H.0.3) is equal to Lu(x), so (H.0.3) becomes
Lu(x) =
_
LG(x, s)f(s) ds. (H.0.4)
Because operator L(x) is linear and acts on the variable x only, we can take the operator L out
of the integration on the right hand side, and obtain
Lu(x) = L
__
G(x, s)f(s) ds
_
, (H.0.5)
which implies that the solution u(x) of differential equation (H.0.1) is given by
u(x) =
_
G(x, s)f(s) ds. (H.0.6)
1
Also referred to as source functions, fundamental solutions, gaussians or propagators of the diusion equation,
or simply diusion kernels.
113
114 APPENDIX H. GREENS FUNCTIONS
Application
Our goal is to find a particular solution of the heat equation
u
=
1
2
u
xx
, (H.0.7)
with initial condition u(x, 0) = (x) = max(x, 0), and then to construct all the other solutions,
using property 4. in section 5.9, which says that an integral of solutions of the heat equation is
again a solution.
The particular solution we will look for is the one, denoted Q(x, ), that satisfies the special
initial condition
Q(x, 0) =
_
1 for x > 0,
0 for x < 0.
(H.0.8)
This function for the initial condition is known as the Heaviside function. The reason for this
choice is that this initial condition does not change under dilation
2
.
First we will look for Q(x, ) of the form
Q(x, ) = g(p) where p =
x
2
, (H.0.9)
and g is a function of one variable (to be determined).
Clearly, the initial condition (H.0.8) does not change at all under the dilation x
ax, t at.
So we look for a Q that satisfies the heat equation (H.0.7) and initial condition (H.0.8) and has
the form (H.0.9).
Using (H.0.9), we can convert the heat equation (H.0.7) into an ode for g, by use of the chain
rule. This gives
0 = Q

1
2
Q
xx
,
=
1
2
x
2
g
(p)
1
2
1
2
g
(p),
=
1
4
_
2pg
(p) +g
(p)
_
, (H.0.10)
such that the ode becomes
g
(p) + 2pg
(p) = 0. (H.0.11)
The ODE (H.0.11) is solved using the integrating factor e
2p dp
= e
p
2
, such that we obtain
g
(p) = c
1
e
p
2
(H.0.12)
and
Q(x, ) = g(p) = c
1
_
x/
2
0
e
p
2
dp +c
2
, for > 0. (H.0.13)
Using initial condition (H.0.8), we can express the limits as follows.
If x > 0, 1 = lim
t0
Q(x, ) = c
1
_

0
e
p
2
dp +c
2
= c
1
2
+c
2
. (H.0.14)
If x < 0, 0 = lim
t0
Q(x, ) = c
1
_
0
e
p
2
dp +c
2
= c
1
2
+c
2
. (H.0.15)
2
See appendix G for the denition of a dilation transformation.
115
This determines the coefficients
_
c
1
=
1
,
c
2
=
1
2
.
Therefore Q(x, ) is the function
Q(x, ) =
1
2
+
1
_
x/
2
0
e
p
2
dp, for > 0. (H.0.16)
After checking, we find that solution (H.0.16) does indeed satisfy the heat equation (H.0.7),
initial condition (H.0.8) and has the form (H.0.9).
Now define
3
G(x, ) =
Q
x
=
1
2
k
e
x
2
/2
, for > 0. (H.0.17)
and
u(x, ) =
_

G(x y, )(y) dy, for > 0. (H.0.18)

By property 4. in section 5.9, the integral u(x, ) of solution G(x, ) is a dierent solution of the
heat equation (H.0.7). We even claim that u(x, ) is the unique solution of (H.0.7) satisfying
the initial condition u(x, 0) = (x).
Hence, after substituting (H.0.17) and the initial condition (x) = max(x, 0) into equation
(H.0.18) we find that the solution is given by
u(x, ) =
1
2
_

e
(xy)
2
/2
max(y, 0) dy, for > 0.
=
1
2
_

0
ye
(xy)
2
/2
dy, for > 0.
Now transform (x,y,) into (x,z,) by using
z =
x y
, (H.0.19)
such that we can replace y by y = x
z. Assuming x and to be constant with respect to

z, we have dz =
1
dy. This gives

u(x, ) =
1
2
_

x/
(x
z)e
z
2
/2
dz,
=
1
2
_
x/
(x
z)e
z
2
/2
dz,
=
x
2
_
x/
e
z
2
/2
dz
2
_
x/
ze
z
2
/2
dz,
=
x
2
_
x/
e
z
2
/2
dz
2
_
x/
ze
z
2
/2
dz, for > 0. (H.0.20)
3
Note that
Q
x
is the Dirac delta function, because Q(x, 0) is equal to the Heaviside function.
116 APPENDIX H. GREENS FUNCTIONS
The second part of u(x, ) we can compute by partial integration.
_
x/
ze
z
2
/2
dz =
_
x/
e
z
2
/2
d
_
z
2
2
_
=
_
e
z
2
/2
_
x/
.
So the solution (H.0.20) indeed becomes
u(x, ) =
x
2
_
x/
e
z
2
/2
dz
1
2
_
e
z
2
/2
_
x/
,
=
x
2
_
x/
e
z
2
/2
dz +
e
x
2
/2
,
= x N
_
x
_
+
n
_
x
_
, for > 0. (H.0.21)
Thus, using Greens functions we find exactly the same solution as we already found in (5.9.10).
Appendix I
Hermitian adjoint
In mathematics, specifically in functional analysis, each linear operator on a Hilbert space has
a corresponding adjoint operator. The adjoint of an operator A is also sometimes called the
Hermitian adjoint (after Charles Hermite) of A and is denoted by A
.
One can show that there exists a unique continuous linear operator A
: H H with the
following property:
< Ax, y > = < x, A
y > for all x, y H. (I.0.1)

Here < , > denotes the inner product, which is defined as follows:
< f, g >:=
_
b
a
f(t)g(t) dt. (I.0.2)
Immediate properties:
A
= A.
If A is invertible, so is A
. Then, (A
)
1
= (A
1
)
.
(A+B)
= A
+B
.
(A)
, where
denotes the complex conjugate of the complex number .

(AB)
= B
.
Application
Consider the operator
L
0
=
1
2
2

2
2
+m

. (I.0.3)
We will compute its adjoint by using the definition of the inner product
< L
0
u, v > =
_

_
1
2
2
u
2
+m
u
_
v dx. (I.0.4)
After applying integration by parts twice, this gives
< L
0
u, v > =
_

u
_
1
2
2
v
2
m
v
_
dx = < u, L
0
v > . (I.0.5)
Here,
L
0
=
1
2
2

2
2
m

(I.0.6)
is called the adjoint operator of L
0
.
117
118 APPENDIX I. HERMITIAN ADJOINT
Appendix J
Fredholm Alternative
Consider a differential equation
L(u) = f, (J.0.1)
where L is an operator that operates on u.
First we expand
u =
n
a
n
n
, (J.0.2)
where
n
are the eigenvectors of u and a
n
are coefficients that are yet undetermined.
Because
n
are the eigenvectors of u, we have
L(
n
) =
n
n
, (J.0.3)
with
n
are the corresponding eigenvalues.
Substitution into the differential equation yields
L(
n
a
n
n
) = f, (J.0.4)
which gives
n
a
n
n
= f. (J.0.5)
Multiplication by a function
m
, which is orthogonal to
n
, and integration over a domain D
results in
a
m
m
_
D
2
m
dV =
_
D
f
m
dV. (J.0.6)
Now the Fredholm Alternative (named after the Swedish mathematician Ivar Fredholm)
states that the following 3 situations are possible:
If
m
,= 0 for all m, then the coefficients a
m
are unique, and the solution can also be
determined uniquely.
In case there is a
m
= 0, then:
If
_
D
f
m
dV ,= 0, there is no solution.
If
_
D
f
m
dV = 0, then the coefficients a
m
are not determined, and there are infinitely
many solutions.
119
120 APPENDIX J. FREDHOLM ALTERNATIVE
Appendix K
Dimensional analysis
Dimensional analysis is a conceptual tool often applied to understand physical situations
involving certain (physical) quantities. It is routinely used by mathematicians, statisticians,
physical scientists and engineers to check the plausibility of derived equations and computations.
The Buckingham theorem is a key theorem in dimensional analysis. The theorem loosely
states that if we have a physically meaningful equation involving a certain number, n, of physical
variables, and these variables are expressible in terms of k independent fundamental physical
quantities, then the original expression is equivalent to an equation involving a set of p = n k
dimensionless variables constructed from the original variables: it is a scheme for nondimen-
sionalization. This provides a method for computing sets of dimensionless parameters from the
given variables, even if the form of the equation is still unknown. However, the choice of dimen-
sionless parameters is not unique: Buckinghams theorem only provides a way of generating sets
of dimensionless parameters, and will not choose the most physically meaningful.
Application
On page 11 of [7] Howison claims that the value of an option satisfies the following nonlinear
equation
V
t
+
1
2
2
S
2
2
V
S
2
+rS
V
S
rV +
2
S
2
_
2
V
S
2
_
2
+
1
2
2
S
4
_
2
V
S
2
_
3
= 0. (K.0.1)
If we define L
BS
as the Black-Scholes differential operator, given by
L
BS
=

t
+
1
2
2
S
2

2
S
2
+rS

S
r, (K.0.2)
we can write (K.0.1) as
L
BS
V +
2
S
2
_
2
V
S
2
_
2
+
1
2
2
S
4
_
2
V
S
2
_
3
= 0. (K.0.3)
Let us express all dimensions in the problem in terms of the dimensions of K and t, as follows:
_
_
[V ] = [K] ,
[S] = [K] ,
[r] =
_
t
1
,
[] =
_
t
1
2
_
,
[] = 1,
[] = 1,
(K.0.4)
121
122 APPENDIX K. DIMENSIONAL ANALYSIS
we find that
[L
BS
V ] =
_
K
t
_
+
_
1
t
K
3
1
K
2
_
+
_
1
t
K
1
K
_
_
K
t
_
=
_
K
t
_
, (K.0.5)
_
2
S
2
_
2
V
S
2
_
2
_
=
_
1
1
t
K
2
_
K
K
2
_
2
_
=
_
1
t
_
, (K.0.6)
_
1
2
2
S
4
_
2
V
S
2
_
3
_
=
_
1
1
t
K
4
_
K
K
2
_
3
_
=
_
K
t
_
. (K.0.7)
So the terms of (K.0.3) have the following dimensions
_
K
t
_
+
_
1
t
_
+
_
K
t
_
. (K.0.8)
Dimensionally, this is incorrect, we cannot sum up terms that have different dimensions.
Suggestion
If we replace equation (K.0.1) by
V
t
+
1
2
2
S
2
2
V
S
2
+rS
V
S
rV +
2
S
3
_
2
V
S
2
_
2
+
1
2
2
S
4
_
2
V
S
2
_
3
= 0, (K.0.9)
the dimensions are correct. The S
2
indeed turns out to be a typographical error when it was
copied from page 24 of [8]. If we replace it by S
3
, we obtain equation (K.0.9), that is correct.
Appendix L
The Dirac Delta
The Dirac delta is a mathematical construct introduced by theoretical physicist Paul Dirac.
Informally, it is a generalized function representing an innitely sharp peak bounding unit area:
a function
1
(x) that has the value zero everywhere, except at x = 0 where its value is innitely
large in such a way that its total integral is 1:
x=0
(x) = 0 and
_

(x) dx = 1. (L.0.1)
In section 8.3 we use the fact that
(x) =
(x)
, (L.0.2)
which will be proven in this section.
Proof :
Using the property
x=0
(x) = 0, also
x=0
[[(x) = 0. (L.0.3)
_

[[(x) dx =
_

(x) d(x) =
_

[[(y) dy = 1. (L.0.4)
Conclusion:
(x) = [[(x), (L.0.5)
and thus
(x) =
(x)
[[
. (L.0.6)
In section 8.3 we will use this property without taking the absolute value of in the denominator.
This is allowed, because we have > 0.
1
The Dirac delta is not strictly a function. While for many purposes it can be manipulated as such, formally
it can be dened as a distribution that is also a measure.
123
124 APPENDIX L. THE DIRAC DELTA
Appendix M
Near-identity transformation
Def inition
Suppose that a solution x
(t) can be approximated by y
(t) for t 0. If for t 0

x
(t) y
(t) = O(
m
) , 0 t
n
C, (M.0.1)
with m, n, C constants independent of , we call y
(t) an O(
m
) approximation of x
(t) on the
timescale
1
n
.
Application
Consider the n-dimensional equation in the standard form
x = f(t, x) +
2
g(t, x) +
3
R(t, x, ), (M.0.2)
in which the vector fields f and g are T-periodic in t with averages f
0
and g
0
.
Next, define the vectorfield
u(t, x) :=
_
t
0
_
f(s, x) f
0
(x)
_
ds a(x), (M.0.3)
where f(s, x) f
0
(x) has average zero, but this does not hold necessarily for the integral. The
function a(x) is chosen such that u
0
(x) (i.e., the average of u(t, x)) vanishes.
We now introduce the near-identity transformation
x(t) := w(t) +u(t, w(t)). (M.0.4)
We call this near identity as x(t) w(t) = O() for t 0 This will be used to simplify
equation (M.0.2). The near identity transformation is also called the averaging or normalising
transformation. Substituting this into the equation for x (M.0.2), we obtain
w(t) +
u
t
(t, w(t)) +
u
w
(t, w(t)) w(t)
= f ((t, w(t) +u(t, w(t))) +
2
g ((t, w(t) +u(t, w(t))) +
3
. . . . (M.0.5)
Using the definition of u
1
(t, x), the left-hand side of equation (M.0.5) becomes
LHS = w(t) + u(t, w(t)) w(t) +
_
f(t, w) f
0
(w)
_
,
= (I + u(t, w(t))) w(t) +f(t, w) f
0
(w). (M.0.6)
125
126 APPENDIX M. NEAR-IDENTITY TRANSFORMATION
Here, I is the n n identity matrix, and with f(t, x) we denote the derivative with respect
to x only, this is an n n matrix.
Because u is uniformly bounded, so is u(t, w(t)), and we may invert the matrix (I + u(t, w(t)))
to obtain
(I + u(t, w(t)))
1
= I u(t, w(t)) +O
_
2
_
, t 0. (M.0.7)
Expanding f and g, we obtain
f ((t, w(t) +u(t, w(t))) = f(t, w) + f(t, w)u(t, w) +. . . , (M.0.8a)
g ((t, w(t) +u(t, w(t))) = g(t, w) + g(t, w)u(t, w) +. . . , (M.0.8b)
such that
w(t) = f
0
(w) +
2
f(t, w)u(t, w) +
2
g(t, w) +
2
f
0
(w) u(t, w) +O
_
3
_
. (M.0.9)
Put f
1
(t, w) := f(t, w)u(t, w) f
0
(w) u(t, w), with average f
0
1
. This yields
w(t) = f
0
(w) +
2
f
1
(t, w) +
2
g(t, w) +O
_
3
_
, (M.0.10)
such that after averaging we have
v(t) = f
0
(v) +
2
f
0
1
(v) +
2
g
0
(v) +O
_
3
_
. (M.0.11)
The average of w(t) is denoted as v(t), for notational convenience.
We can prove that
x(t) = v(t) +u(t, v(t)) +O
_
2
_
(M.0.12)
on the timescale
1

Afstudeerverslagliekedejong Finalversion

Uploaded by

Copyright:

Available Formats

Afstudeerverslagliekedejong Finalversion

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Afstudeerverslagliekedejong Finalversion

Uploaded by

Copyright:

Available Formats

Delft University of Technology

Faculty of Electrical Engineering, Mathematics and Computer Science

+ 2y = 0, for 0 < x < 1, (2.0.1)

(0) = 0. The boundary condition at x = 0 has been included here

=0, for 0 < < ,

4.5 Self-f inancing strategies

u(x, ) and this indeed gives us the heat equation

) equation vanishes, because for [S K[

K (i.e., if the spot

) equation with final condition

K, far above the strike,

K, far below the strike.

and we divide by K to make x dimensionless. The value

to have v = O(1) and again we devide by K

S(x y, t)g(y) dy,

ax, at), for any

ax, at). Then we have v

ax, at) in property 5. we know that the pde in (5.8.8) has

, we obtain the following ode in f():

. The solution of ode (5.9.6) is

), which implies m = O(1), and note that = O(

), which implies = O(1).

() satisfies the forward Kolmogorov equation, with

() satisfying the property

() is a transition density function.

exists. This is very different, because we did not find

) equation would be completely dierent, because then the term L1

and the fact that L

>= 0. By the Fredholm alter-

is a probability distribution function. Hence,

, we have the following differential equation in terms of G

>= 0 for existence of a solution can be

, for < 0. (7.4.1)

() should have an additional

, but he only remarks

. Then take an expansion of the form

(K)(f K). (8.3.36)

(K)/C(K) becomes a significant fraction of 1, i.e., as

B(0)P. Recall that B(z) = C(f) and B(0) = C(K).

P(, z, )d, (8.3.55)

are all O(). Recall that we are only solving

(z) by the constants I(z

in a Taylor series around = 0 yields 1 +

. The model then

+ 2y = 0, for 0 < x < 1, (C.0.1)

(0) = Q if and only if c = a, because Q is

G(x y, )(y) dy, for > 0. (H.0.18)

z. Assuming x and to be constant with respect to

dy. This gives

y > for all x, y H. (I.0.1)

denotes the complex conjugate of the complex number .

(t) can be approximated by y

(t) for t 0. If for t 0

You might also like