Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
ILiang Chern
Department of Mathematics
National Taiwan University
2
Contents
1 Introduction 1
1.1 Financial Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Financial Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Payoff functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.5 Other kinds of options . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.6 Types of traders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.7 Basic assumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Asset Price Model 7
2.1 Efﬁcient market hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 The asset price model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.1 The discrete asset price model . . . . . . . . . . . . . . . . . . . . 7
2.2.2 The continuous asset price model . . . . . . . . . . . . . . . . . . 8
2.3 Random walk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 The solution of the discrete asset price model . . . . . . . . . . . . . . . . 10
2.5 The Brownian motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.5.1 The deﬁnition of a Brownian motion . . . . . . . . . . . . . . . . 10
2.5.2 The Brownian as a limit of random walk . . . . . . . . . . . . . . 11
2.5.3 Properties of Brownian motion . . . . . . . . . . . . . . . . . . . 12
2.6 It ˆ o’s formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.7 The solution of the continuous asset price model . . . . . . . . . . . . . . 13
2.8 Continuous model as a limit of the discrete model . . . . . . . . . . . . . . 14
2.9 Simulation of asset price model . . . . . . . . . . . . . . . . . . . . . . . 16
3 BlackScholes Analysis 19
3.1 The hypothesis of noarbitrageopportunities . . . . . . . . . . . . . . . . . 19
3.2 Basic properties of option prices . . . . . . . . . . . . . . . . . . . . . . . 20
3.2.1 The relation between payoff and options . . . . . . . . . . . . . . . 20
3.2.2 European options . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2.3 Basic properties of American options . . . . . . . . . . . . . . . . 22
3.2.4 Dividend Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.3 The BlackScholes Equation . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3.1 BlackScholes Equation . . . . . . . . . . . . . . . . . . . . . . . 26
3.3.2 Boundary and Final condition for European options . . . . . . . . . 27
3
4 CONTENTS
3.4 Exact solution for the BS equation for European options . . . . . . . . . . 28
3.4.1 Reduction to parabolic equation with constant coefﬁcients . . . . . 28
3.4.2 Further reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.4.3 BlackScholes formula . . . . . . . . . . . . . . . . . . . . . . . . 30
3.4.4 Special cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.5 Risk Neutrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.6 The delta hedging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.6.1 TimeDependent r, σ, µ . . . . . . . . . . . . . . . . . . . . . . . 35
3.7 Trading strategy involving options . . . . . . . . . . . . . . . . . . . . . . 36
3.7.1 Strategies involving a single option and stock . . . . . . . . . . . . 37
3.7.2 Bull spreads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.7.3 Bear spreads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.7.4 Butterﬂy spread . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4 Variations on BlackScholes models 41
4.1 Options on dividendpaying assets . . . . . . . . . . . . . . . . . . . . . . 41
4.1.1 Constant dividend yield . . . . . . . . . . . . . . . . . . . . . . . 41
4.1.2 Discrete dividend payments . . . . . . . . . . . . . . . . . . . . . 43
4.2 Warrants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3 Futures and futures options . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.3.1 Forward contracts . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.3.2 Futures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3.3 Futures options . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.3.4 BlackScholes analysis on futures options . . . . . . . . . . . . . . 47
5 Numerical Methods 51
5.1 Monte Carlo method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.2 Binomial Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.2.1 Binomial method for asset price model . . . . . . . . . . . . . . . 52
5.2.2 Binomial method for option . . . . . . . . . . . . . . . . . . . . . 52
5.3 Finite difference methods (for the modiﬁed BS eq.) . . . . . . . . . . . . . 53
5.3.1 Discretization methods . . . . . . . . . . . . . . . . . . . . . . . . 54
5.3.2 Binomial method is a forward Euler ﬁnite difference method . . . . 55
5.3.3 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.3.4 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.3.5 Boundary condition . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.4 Converting the BS equation to ﬁnite domain . . . . . . . . . . . . . . . . 62
5.5 Fast algorithms for solving linear systems . . . . . . . . . . . . . . . . . . 63
5.5.1 Direct methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.5.2 Iterative methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6 American Option 69
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.2 American options as a free boundary value problem . . . . . . . . . . . . . 70
6.2.1 American put option . . . . . . . . . . . . . . . . . . . . . . . . . 70
CONTENTS 1
6.2.2 American call option on a dividendpaying asset . . . . . . . . . . 72
6.3 American option as a linear complementary problem . . . . . . . . . . . . 72
6.4 Numerical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6.4.1 Projective method for American put . . . . . . . . . . . . . . . . . 74
6.4.2 Projective method for American call . . . . . . . . . . . . . . . . . 75
6.4.3 Implicit method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.5 Converting American option to a ﬁxed domain problem . . . . . . . . . . . 76
6.5.1 American call option with dividend paying asset . . . . . . . . . . 76
6.5.2 American put option . . . . . . . . . . . . . . . . . . . . . . . . . 78
7 Exotic Options 79
7.1 Binaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
7.2 Compounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
7.3 Chooser options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
7.4 Barrier option . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
7.4.1 downandout call(knockout) . . . . . . . . . . . . . . . . . . . . . 81
7.4.2 downandin(knockin) option . . . . . . . . . . . . . . . . . . . . 82
7.5 Asian options and lookback options . . . . . . . . . . . . . . . . . . . . . 83
8 PathDependent Options 85
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
8.2 General Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
8.3 Average strike options . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
8.3.1 European calls . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
8.3.2 American call options . . . . . . . . . . . . . . . . . . . . . . . . 87
8.3.3 Putcall parity for average strike option . . . . . . . . . . . . . . . 87
8.4 Lookback Option . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
8.4.1 A lookback put with European exercise feature . . . . . . . . . . . 89
8.4.2 Lookback put option with American exercise feature . . . . . . . . 90
9 Bonds and Interest Rate Derivatives 91
9.1 Bond Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
9.1.1 Deterministic bond model . . . . . . . . . . . . . . . . . . . . . . 91
9.1.2 Stochastic bond model . . . . . . . . . . . . . . . . . . . . . . . . 91
9.2 Interest models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
9.2.1 A functional approach for interest rate model . . . . . . . . . . . . 93
9.3 Convertible Bonds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
A Basic theory of stochastic calculus 97
A.1 Brownian motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
A.2 Stochastic integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
A.3 Stochastic differential equation . . . . . . . . . . . . . . . . . . . . . . . . 100
A.4 Diffusion process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
2 CONTENTS
Chapter 1
Introduction
1.1 Financial Markets
A society improves its welfare through investment. The ﬁnancial market provides a link
between saving and investment. Savers can earn high returns from their saving and bor
rowers can execute their investment plans to earn future proﬁts. In ﬁnancial markets, assets
are traded in. There are many kinds of ﬁnancial markets:
• stock markets,
• bond markets,
• currency markets, foreign exchange markets,
• commodity markets (oil, wheat, gold),
• futures and options markets.
In futures or options, more complex contracts than simple buy/sell trades have been intro
duced. These are called ﬁnancial derivatives.
1.2 Financial Derivatives
1. Forwards contract: A forward contract is an agreement which allows the holder of
the contract to buy or sell a certain asset at or by a certain day at a certain price. Here,
• the certain day—maturity or expiration date,
• the certain price—delivery price,
• the person who write the contract (has the asset) is called in short position,
• the person who holds the contract is called in long position.
2. Futures (futures contracts): A future contract, like a forward contract, except,
• it is normally traded in an exchange;
1
2 CHAPTER 1. INTRODUCTION
• it has standard features (including contract size, quality, delivery arrangement,
price quotes, daily price movement, position limit, etc.);
• it is a margin trading (certain minimal amount of money should be maintained
in a margin account);
• clearinghouse.
3. Options: There are two kinds of options — call options and put options. A call
(put) option is a contract between two parties, in which the holder has the right to
buy (sell) and the writer has the obligation to sell (buy) an asset at certain time in
the future at a certain price. The price is called the exercise price (or strike price).
The holder is called in long position, while the writer is called in short position.
The underlying assets of an option can be commodity, stocks, stock indices, foreign
currencies, or future contracts.
There are two kinds of exercise features:
• European options : Options can only be exercised at the maturity date.
• American options : Options can be exercised any time up to the maturity date.
1.3 Examples
Notation
• t current time
• T maturity date
• S current asset price
• S
T
asset price at time T
• E strike price
• c premium, the price of call option
• r bank interest rate
1. An investor buys 100 European call options on IBM stock with strike price $140.
Suppose
E = 140,
S
t
= 138,
T = 2 months,
c = 5 (the price of one call option).
If at time T, S
T
> E, then he should exercise this option. The payoff is 100 (S
T
−
E) = 100 (146 − 140) = 600, The premium is 5 100 = 500. Hence, he earns
$100. If S
T
≤ T, then he should not exercise his call contracts. The payoff is 0.
1.4. PAYOFF FUNCTIONS 3
The payoff function for a call option is Λ = max¦S
T
− E, 0¦. One needs to pay
premium (c
t
) to buy the options. Thus the net proﬁt from buying this call is
Λ −c
t
e
r(T−t)
.
2. Suppose
today is t = 8/22/95,
expiration is T = 4/14/96 ,
the strike price E = 250
for some stock. If S
T
= 270 at expiration, which is smaller than the strike price, we
should exercise this call option, then buy the share for 250, and sell it in the market
immediately for 270. The payoff Λ = 270 − 250 = 20. If S
T
= 230, we should
give up our option, and the payoff is 0. Suppose the share take 230 or 270 with equal
probability. Then the expected proﬁt is
1
2
0 +
1
2
20 = 10.
Ignoring the interest of bank, then a reasonable price for this call option should be
10. If S
T
= 270, then the net proﬁt= 20 − 10 = 10. This means that the proﬁts is
100% (He paid 10 for the option). If S
T
= 230 the loss is 10 for the premium. The
loss is also 100%. On the other hand, if the investor had instead purchased the share
for 250 at t, then the corresponding proﬁt or loss at T is ±20. Which is only ±8% of
the original investment. Thus, option is of high risk and with high return.
1.4 Payoff functions
At the expiration day, the payoff of a future or an option is the follows.
1. The payoff function of a future is
Λ = S
T
−E.
K
S
T
Λ
future (long)
K
S
T
Λ
future (short)
4 CHAPTER 1. INTRODUCTION
Payoff of a future, long position (left) and short position (right)
2. The payoff function of a call option is
Λ = max¦S
T
−E, 0¦.
K
S
T
Λ
call option (long):max{S
T
−K,0}
K
S
T
Λ
call option (short):−max{S
T
−K,0}=min{K−S
T
,0}
Payoff of a call, long position (left) and short position (right)
3. The payoff function of a put option is
Λ = max¦E −S
T
, 0¦.
K
S
T
Λ
put option (long):max{K−S
T
,0}
K
S
T
Λ
put option (short):−max{K−S
T
,0}=min{S
T
−K,0}
Payoff of a put, long position (left) and short position (right)
4. Below is a portion of a call option copied from the Financial Times.
the current time t = Feb 3
the expiration T = end of Feb,
T −t ≈ 10 days
S
t
= 2872
1.5. OTHER KINDS OF OPTIONS 5
E = 2650, 2700, 2750, 2800, 2850, 2900, 2950, 3000
c = 233, 183, 135, 89, 50, 24, 9, 3
2600 2650 2700 2750 2800 2850 2900 2950 3000 3050
−50
0
50
100
150
200
250
K
c
The FTSE index call option values versus exercise price.
1.5 Other kinds of options
• Barrier option: The option only exists when the underlying asset price is in some
prescribed value before expiry.
• Asian option: It is a contract giving the holder the right to buy or sell an asset for its
average price over some prescribed period.
• Lookback option: The payoff depends not only on the asset price at expiry but
also its maximum or minimum over some period price to expiry. For example, Λ =
max¦J −S
0
, 0¦, J = max
0≤τ≤T
S(τ).
1.6 Types of traders
1. Speculators (high risk, high rewards)
2. Hedgers (to make the outcomes more certain)
3. Arbitrageurs (Working on more than one markets, p12, p13, p14, Hull).
6 CHAPTER 1. INTRODUCTION
1.7 Basic assumption
Arbitrage opportunities cannot last for long. Only small arbitrage opportunities are ob
served in ﬁnancial markets. Our arguments concerning future prices and option prices will
be based on the assumption that “there is no arbitrage opportunities”.
Chapter 2
Asset Price Model
2.1 Efﬁcient market hypothesis
The asset prices move randomly because of the following efﬁcient market hypothesis:
1. The past history is fully reﬂected in the present price, which does not hold any future
information. This means the future price of the asset only depends on its current
value and does not depends on its value one month ago, or one year ago. If this were
not true, technical analysis could make aboveaverage return by interpreting chart of
the past history of the asset price. This contradicts to the hypothesis of no arbitrage
opportunities. In fact, there is very little evidence that they are able to do so.
2. Market reponds immediately to any new information about an asset.
2.2 The asset price model
We shall introduce a discrete model and a continuous model. We will show that the contin
uous model is the continuous limit of the discrete model.
2.2.1 The discrete asset price model
The time is discrete in this model. The time sequence is n∆t, n ∈ N. Let us denote the
asset price at time step n by S
n
. We model the asset price by
S
n+1
S
n
=
u with probability p
d with probability 1 −p.
(2.1)
Here, 0 < d < 1 < u. The information we are looking for is the following transition
probability P(S
n
= S[S
0
), the probability that the asset price is S at time step n with
initial price S
0
. We shall ﬁnd this transition probability later.
7
8 CHAPTER 2. ASSET PRICE MODEL
2.2.2 The continuous asset price model
Let us denote the asset price at time t by S(t). The meaningful quantity for the change of
an asset price is its relative change
dS
S
,
which is called the return. The change
dS
S
can be decomposed into two parts: one is
deterministic, the other is random.
• Deterministic part: This can be modeled by
dS
S
= µdt.
Here, µ is a measure of the growth rate of the asset. We may think µ is a constant
during the life of an option.
• Random part: this part is a random change in response to external effects, such as
unexpected news. It is modeled by a Brownian motion
σdz,
the σ is the order of ﬂuctuations or the variance of the return and is called the volatil
ity. The quantity dz is sampled from a normal distribution which we shall discuss
below.
The overall asset price model is then given by
dS
S
= µdt + σdz. (2.2)
We shall look for the transition probability density function {(S(t) = S[S(0) = S
0
). Or
equivalently, the integral
b
a
{(S(t) = S[S(0) = S
0
) dS
is the probability that the asset price S(t) lies in (a, b) at time t and is S
0
initially.
2.3 Random walk
To study the discrete asset price model, we study a simple model—the random walk in one
dimension—ﬁrst. Consider a particle moving randomly on a uniformly distributed grid
points on the real lines. Suppose the grid points are located at m∆x, m ∈ Z. In each time
step, the particle moves to its left adjacent grid point or right adjacent grid point with equal
probability. Suppose the particle is located at 0 initially. Let Z
n
denote the location of this
particle at time step n. Let w(m∆x, n∆t) denotes for the probability that the particle is
2.3. RANDOM WALK 9
located at the m∆x cell at the time n∆t. That is, w(m∆x, n∆t) = P(Z
n
= m∆x[Z
0
= 0).
By our rule,
Z
n+1
−Z
n
=
∆x with probability
1
2
−∆x with probability
1
2
and
w(m∆x, (n + 1)∆t) =
1
2
w((m−1)∆, n∆t) +
1
2
w((m + 1)∆, n∆t). (2.3)
Suppose in n times, the particle moves p times toward right and n − p time toward left.
Then
m = p −(n −p) = 2p −n or p =
1
2
(n + m).
Notice that m is even(odd), when n is even(odd). There is a onetoone correspondence
between ¦p [ 0 ≤ p ≤ n¦ and ¦m[ − n ≤ m ≤ n, m + n is even¦. Notice also that the
number of choices in n steps that the particle moves p times toward right is
n
p
:=
n!
(n−p)!p!
.
When p =
1
2
(n + m), we have
w(m∆x, n∆t) =
0, if m +n is odd,
n
p
(
1
2
)
n
, if m +n is even.
We may check that w(m∆x, n∆t) is a probability density function. Namely,
1. w(m∆x, n∆t) ≥ 0.
2.
¸
m
w(m∆x, n∆t) = 1.
Given any function f(m), we deﬁne its expectation value at n∆t by
< f(m) >:=
¸
m
f(m)w(m∆x, n∆t).
The moments < m
k
>, k ∈ N are particularly important. The ﬁrst moment < m > is
called the mean, while the second moment of the variation from mean < (m− < m >)
2
>
is called the variance. They can be found by computing < p
k
>, which in turn can be
computed through the help of the following generating function:
G(u) :=
¸
p
u
p
1
2
n
n
p
=
1 + u
2
n
.
Hence
< p >= G
(1) =
¸
p
p
1
2
n
n
p
=
n
2
.
From m = 2p −n, we have
< m >= 2 < p > −n = 0.
10 CHAPTER 2. ASSET PRICE MODEL
To compute the second moment < m
2
>, from m = 2p −n, we have
< m
2
>= 4 < p
2
> −4n < p > +n
2
.
With the help of the generating function,
G
(1) =
n
¸
p=0
p(p −1)
n
p
1
2
n
= < p
2
> − < p >
= < p
2
> −
n
2
On the other hand, from G(u) =
1+u
2
n
, we obtain G
(1) =
n(n−1)
4
. Hence, < p
2
>=
n
2
4
+
n
4
and
< m
2
>= 4 < p
2
> −4n < p > +n
2
= n.
The mean of this random walk is < m >= 0, while its variance is < (m− < m >)
2
>= n.
Exercise
1. Find the transition probability, mean and variance for the case
Z
n+1
−Z
n
=
∆x with probability p
−∆x with probability 1 −p
2. One can also ﬁnd the transition probability w by solving the difference equation (2.3).
2.4 The solution of the discrete asset price model
Let us consider the case
S
n+1
S
n
=
u with probability
1
2
d with probability
1
2
.
for simplicity. In n movements of the asset price, if the price goes up p times, then the price
at time step n∆t is S
n
= S
0
u
p
d
n−p
. Since there are
n
p
such choices, we then obtain the
transition probability of the asset:
P(S
n
= S[S
0
) =
n
p
1
2
n
if S = S
0
u
p
d
n−p
,
0 otherwise.
(2.4)
2.5 The Brownian motion
2.5.1 The deﬁnition of a Brownian motion
The deﬁnition of the (standard) Brownian motion z(t) is the following:
2.5. THE BROWNIAN MOTION 11
1. ∀t, z(t) is a random variable.
2. The increment z(t + s) −z(t), z(t) −z(t −u), u > 0, s > 0 are independent.
3. z(t) is continuous in t.
4. ∀s > 0, z
t+s
− z
t
is normally distributed with mean zero and variance s, i.e., its
probability density is ^(0, s)(i.e.,
1
√
2πs
e
−x
2
2s
).
2.5.2 The Brownian motion as a limit of random walk
We may realize the Brownian motion as the limit of the random walk in the previous sec
tion. Namely, Z
n
→z(t) as n →∞with m∆x →x, n∆t →t and
(∆x)
2
∆t
= σ ﬁxed. This
can be proved by the Stirling formula:
n! ≈
√
2πn
n+
1
2
e
−n
.
Recall that the probability
P(Z
n
= m∆x[Z
0
= 0) =
n
1
2
(m + n)
1
2
n
.
Using the Stirling formula, we have for n, p, n −p >> 1,
n
1
2
(m + n)
1
2
n
= (
1
2
)
n
n!
(
1
2
(n + m))!(
1
2
(n −m))!
≈ (
1
2
)
n
√
2πn
n+
1
2
e
−n
√
2π(
1
2
(n + m))
1
2
(n+m)+
1
2
√
2π(
1
2
(n −m))
1
2
(n−m)+
1
2
= (
2
πn
)
1
2
(1 +
m
n
)
−
1
2
(n+m)−
1
2
(1 −
m
n
)
−
1
2
(n−m)−
1
2
≈ (
2
πn
)
1
2
(1 −(
m
n
)
2
)
−
1
2
n
= (
2
πn
)
1
2
(1 −(
m
n
)
2
)
(
n
m
)
2
−
m
2
2n
≈ (
2
πn
)
1
2
exp(−
m
2
2n
)
As m∆ →x, n∆t →t, (∆x)
2
/∆t = σ ﬁxed, we obtain
P(Z
n
= m∆x[Z
0
= 0)/2∆x ≈ (
2
πn
)
1
2
1
2∆x
e
−
m
2
2n
= (
1
2πn∆t
)
1
2
e
−
(m∆x)
2
n∆tσ
2
√
∆t
2(∆x)
→
1
√
2πσ
2
t
e
−
x
2
2σ
2
t
This means that Z
n
→z(t).
12 CHAPTER 2. ASSET PRICE MODEL
2.5.3 Properties of Brownian motion
By deﬁnition
{(z(t) = x[z(0) = 0) =
1
√
2πt
e
−
x
2
2t
.
We can check
1. < z(t) >= 0
2. < z(t)
2
>= t
3. Independence of disjoint increments
{(z(t) = x[z(0) = 0) =
∞
−∞
{(z(t) = x[z(s) = y) {(z(s) = y[z(0) = 0) dy.
(2.5)
In particular, let us deﬁne an inﬁnitesimal increment
dz = z(t + dt) −z(t)
We have
1. < dz >= 0
2. < (dz)
2
>= dt
In fact we have more, we may think
dz =
√
dt (2.6)
where is a random variable with standard Gaussian distribution ^(0, 1) (i.e. mean is 0
and variance is 1). And we have
(dz)
2
= dt with probability 1. (2.7)
Exercise
1. Check (2.5).
2.6 Itˆ o’s formula
In this section, we shall study differential equations which consist of deterministic part:
˙ x = b(x), and stochastic part σ ˙ z(t). Here, z(t) is the Brownian motion. We call such an
equation a stochastic differential equation and expressed as
dx(t) = b(x(t))dt + σ(x(t))dz(t). (2.8)
An important lemma for ﬁnding their solution is the following It ˆ o’s lemma.
2.7. THE SOLUTION OF THE CONTINUOUS ASSET PRICE MODEL 13
Lemma 2.1 Suppose x(t) satisﬁes the stochastic differential equation (2.8), and f(x, t) is
a smooth function. Then f(x(t), t) satisﬁes the following stochastic differential equation:
df =
f
t
+ bf
x
+
1
2
σ
2
f
xx
dt + σf
x
dz (2.9)
Proof. This is not a proof, rather an intuition why (2.9) is true. According to the Taylor
expansion,
df = f
t
dt + f
x
dx +
1
2
f
tt
(dt)
2
+f
xt
dxdt +
1
2
f
xx
(dx)
2
+ .
Plug (2.8) into this equation. We recall that dz =
√
dt, where is a random variable with
standard Gaussian distribution ^(0, 1). In the Taylor expansion of df(x(t), t), the terms
(dt)
2
, dt dz are relative unimportant as comparing with the dt term and dz term. Using
(2.8) and noting (dz)
2
= dt with probability 1, we obtain (2.9).
A simple application of It ˆ o’s lemma is to ﬁnd the transition probability density function for
the s.d.e.
dx = adt +σdz
where a and σ are constants. By letting y = x−at, from It ˆ o’s lemma, y satisﬁes dy = σdz.
Thus, the transition probability density function for y is
{(y(t) = y[y(0) = y
0
) =
1
√
2πσ
2
t
e
−(y−y
0
)
2
/2σ
2
t
.
Or equivalently, the transition probability density function for x is
{(x(t) = x[x(0) = x
0
) =
1
√
2πσ
2
t
e
−(x−at−x
0
)
2
/2σ
2
t
.
2.7 The solution of the continuous asset price model
In this section, we want to ﬁnd the transition probability density function for the continuous
asset price model:
dS = µS dt + σS dz. (2.10)
with initial data S(0) = S
0
. We apply It ˆ o’s lemma with x = f(S) = log S. Then x satisﬁes
the s.d.e.
dx =
µ −
σ
2
2
dt + σ dz,
and x(0) = x
0
:= log S
0
. From the discussion of the previous section, we obtain
{(x(t) = x[x(0) = x
0
) =
1
√
2πσ
2
t
e
−(x−x
0
−(µ−
σ
2
2
)t)
2
/2σt
.
From
{(x(t) = x[x(0) = x
0
)dx = {(x(t) = x[x(0) = x
0
)dS/S
= {(S(t) = S[S(0) = S
0
)dS,
14 CHAPTER 2. ASSET PRICE MODEL
we obtain that the transition probability density function for S(t) is
{(S(t) = S[S(0) = S
0
) =
1
√
2πσ
2
tS
e
−(log
S
S
0
−(µ−
σ
2
2
)t)
2
/2σ
2
t
. (2.11)
This is called the lognormal distribution.
Exercise
1. Find the mean and variance of the lognormal distribution.
S
S
0
p
2.8 Continuous model as a limit of the discrete model
We want to show that the continuous model (2.10) is the limit of the discrete model (2.1).
The parameters in (2.10) are µ and σ. The parameters in (2.1) are u, d and p. We may
assume p = 1/2. First, we relate (µ, σ) and (u, d). Both models should have the same
mean and variance. For the continuous model, we compute its mean under the condition
S((n −1)∆t) = S
n−1
. Then
E(S(n∆t)[S
n−1
) =
S{(S, n∆t[S((n −1)∆t) = S
n−1
)dS
=
S
1
√
2πσ
2
∆tS
e
−(log
S
S
n−1
−(µ−
1
2
σ
2
)∆t)
2
/2σ
2
∆t
dS
2.8. CONTINUOUS MODEL AS A LIMIT OF THE DISCRETE MODEL 15
= S
n−1
1
√
2πσ
2
∆t
e
−(x−(µ−
1
2
σ
2
)∆t)
2
/2σ
2
∆t
e
x
dx,
= S
n−1
e
µ∆t
1
√
2πσ
2
∆t
e
−(
x
√
2σ
2
∆t
−
√
2σ
2
∆t
2
)
2
dx
= e
µ∆t
S
n−1
.
Here, we have used the changeofvariable: x = log
S
S
n−1
. For the second moment for the
continuous model, we have
E(S(n∆t)
2
[S
n−1
) =
S
2
{(S, ∆t[S
n−1
) dS
= e
(2µ+σ
2
)∆t
S
2
n−1
.
On the other hand, the mean and the second moment for the discrete model in one time step
∆t are
1
2
u +
1
2
d
S
n−1
1
2
u
2
+
1
2
d
2
S
2
n−1
.
In order to have the same means and variances in one time step in both models, we should
require
1
2
u
2
+
1
2
d
2
= e
(2µ+σ
2
)∆t
1
2
u +
1
2
d = e
µ∆t
.
Or
u = e
µ∆t
(1 +
e
σ
2
∆t
−1), (2.12)
d = e
µ∆t
(1 −
e
σ
2
∆t
−1). (2.13)
These relate (u, d) and (µ, σ).
Theorem 2.1 Let us ﬁx (µ, σ). Let us choose a ∆t and a ∆x with (∆x)
2
/∆t = σ
2
. Deﬁne
(u, d) by (2.12) and (2.13). Then
P(S
0
u
p
d
n−p
[S
0
)/2∆x −→{(S(t) = S[S(0) = S
0
)
as n∆t →t, n →∞and S
0
u
p
d
n−p
→S.
Proof. Let us deﬁne x = log S, x
0
= log S
0
. Then
log S
0
u
p
d
n−p
= x
0
+p log u + (n −p) log d.
Thus, what we want to show is equivalent to
P(x = x
0
+ p log u + (n −p)[x
0
)/2∆x →{(x(t) = x[x(0) = x
0
)
16 CHAPTER 2. ASSET PRICE MODEL
where
{(x(t) = x[x(0) = x
0
) =
1
√
2πσ
2
t
e
−(x−x
0
−(µ−
σ
2
2
)t)
2
/2σ
2
t
.
To show this, we deﬁne m = 2p −n. Then p =
1
2
(n + m), n −p =
1
2
(n −m). Hence
p log u + (n −p) log d =
1
2
(n + m) log u +
1
2
(n −m) log d
=
1
2
nlog(ud) +
1
2
mlog(
u
d
).
From (2.12) and (2.13),
u d = e
2µ∆t
(2 −e
σ
2
∆t
) ≈ e
2µ∆t
e
−σ
2
∆t
,
u
d
= 1 + 2σ
√
∆t + σ
2
∆t ≈ e
2σ
√
∆t
.
Hence
1
2
nlog ud +
1
2
mlog(
u
d
) ≈ n(µ −
1
2
σ
2
)∆t +mσ
√
∆t
= n(µ −
1
2
σ
2
)∆t +m∆x.
Deﬁne ∆x such that
(∆x)
2
∆t
= σ
2
. Then
p log u + (n −p) log d = n∆t(µ −
1
2
σ
2
) + m∆x.
Recall that the probability that the price moves up p times is
n
p
(
1
2
)
n
. Then the density is
n
p
1
2
n
/2∆x ≈ (
2
nπ
)
1
2
e
−
m
2
2n
−→
1
√
2πσ
2
t
e
−(x−x
0
−(µ−
σ
2
2
)t)
2
/2σ
2
t
2.9 Simulation of asset price model
Typically, µ = 0.16, σ is 0.20 ∼ 0.40 for a stock. To simulate the model
dS
S
= µdt + σdz
S(0) = S
0
We perform N sample paths ω
1
, , ω
N
. In each path, we choose time step ∆t = 0.01, for
instance. We obtain S
k+1
from S
k
by discretizing the s.d.e. and sample a number ξ from
the normalized Gaussian distribution ^(0, 1):
S
k+1
−S
k
S
k
= µ∆t + σξ
√
∆t
= 0.16 0.01 + 0.2 0.5 0.1
2.9. SIMULATION OF ASSET PRICE MODEL 17
Here, ξ = 0.5 is the sampled number. Then the transition probability density function
b
a
{(S(t) = S[S(0) = S
0
) dS ≈ #¦ω [ a ≤ S
n
(ω) ≤ b¦/N
18 CHAPTER 2. ASSET PRICE MODEL
Chapter 3
BlackScholes Analysis
3.1 The hypothesis of noarbitrageopportunities
The option pricing theory was introduced by Black and Scholes. The fundamental hypoth
esis of their analysis is that ”there is no arbitrage opportunities in ﬁnancial markets”.
For simplicity, we shall also assume
1. There exists a riskfree investment that gives a guaranteed return with interest rate r.
( e.g. government bond, bank.)
2. Borrowing or lending at such riskless interest rate is always possible.
3. There is no transaction costs.
4. All trading proﬁts are subject to the same tax rate.
We will use the following notations:
S current asset price
E exercise price
T expiry time
t current time
µ growth rate of an asset
σ volatility of an asset
S
T
asset price at T
r riskfree interest rate
c value of European call option
C value of American call option
p value of European put option
P value of American put option
Λ the payoff function
19
20 CHAPTER 3. BLACKSCHOLES ANALYSIS
3.2 Basic properties of option prices
3.2.1 The relation between payoff and options
1. Recall that
Λ(t) = max(S
t
−E, 0) for call option
Λ(t) = max(E −S
t
, 0) for put option
2. c(S
T
, T) = Λ(T).
Otherwise, there is a chance of arbitrage. For instance, if c(S
T
, T) < Λ, then we can
buy a call on price c, exercise it immediately. If S
T
> E, then Λ = S
T
− E > 0
and c < Λ by our assumption. Hence we have an immediate net proﬁt S
T
− E − c.
This contradicts to our hypothesis. If c(S
T
, T) > Λ, we can short a call and earn c.
If the person who buy the call does not claim, then we have net proﬁt c. If he does
exercise his call, then we can buy an asset from the market on price S
T
and sell to
that person with price E. The cost to us is S
T
−E. By doing so, the net proﬁt we get
is c −(S
T
−E) > 0. Again, this is a contradiction.
3. Similarly, we have
p(S
T
, T) = Λ(T)
C(S
t
, t) = Λ(t)
P(S
t
, t) = Λ(t)
3.2.2 European options
Lemma 3.2 We have the following for European options
max¦S −Ee
−r(T−t)
, 0¦ ≤ c ≤ S (3.1)
max¦Ee
−r(T−t)
−S, 0¦ ≤ p ≤ Ee
−r(T−t)
(3.2)
and the putcall parity
p +S = c + Ee
−r(T−t)
(3.3)
To show these, we need the following deﬁnition and lemmae.
Deﬁnition 2.1 A portfolio is a collection of investments.
For instance, a portfolio I = c − ∆S means that we long a call and short ∆ amount of an
asset S.
Lemma 3.3 Suppose I(t) and J(t) are two portfolios containing no American options
such that I(T) ≤ J(T). Then under the hypothesis of noarbitrageopportunities, we can
conclude that I(t) ≤ J(t), ∀t ≤ T.
3.2. BASIC PROPERTIES OF OPTION PRICES 21
Proof. Suppose the conclusion is false, i.e., there exists a time t ≤ T such that I(t) > J(t).
An arbitrageur can buy (long) J(t) and short I(t) and immediately gain a proﬁt I(t)−J(t).
Since I and J containing no American options, nothing can be exercised before T. At time
T, since I(T) ≤ J(T), he can use J(T) (what he has) to cover I(T) (what he shorts) and
gains a proﬁt J(T)−I(T). This contradicts to the hypothesis of noarbitrageopportunities.
As a corollary, we have
Corollary 2.1 If I(T) = J(T), then I(t) = J(t), ∀t ≤ T.
Now, we can prove the basic properties of European options 15.
Proof of Lemma 3.2.
1. Let I = c and J = S. At T, we have
I(T) = c
T
= max¦S
T
−E, 0¦ ≤ max¦S
T
, 0¦ = S
T
= J(T).
Hence, I(t) ≤ J(t) holds for all t ≤ T.
Remark. The equality holds when E = 0. In this case c = S
2. Consider I = c + Ee
−r(T−t)
and J = S. At time T,
I(T) = max¦S
T
−E, 0¦ + E = max¦S
T
, E¦ ≥ S
T
= J(T).
This implies I(t) ≥ J(t).
3. Let I = p and J = Ee
−r(T−t)
. At time T,
I(T) = max¦E −S
T
, 0¦ ≤ E = J(T).
Hence, I(t) ≤ J(t).
4. Consider I = p + S and J = Ee
−r(T−t)
. At time T,
I(T) = max¦S
T
, E¦ ≥ E = J(T)
. Hence, I(t) ≥ J(t).
5. Consider I = c + Ee
−r(T−t)
and J = p + S. At time T,
I(T) = c +E = max¦S
T
−E, 0¦ + E = max¦S
T
, E¦,
J(T) = p + S = max¦E −S
T
, 0¦ + S
T
= max¦E, S
T
¦
Hence, I(t) = J(t).
22 CHAPTER 3. BLACKSCHOLES ANALYSIS
3.2.3 Basic properties of American options
Lemma 3.4 For American options, we have
(i) The optimal exercise time for American call option is T and we have C = c.
(ii) The optimal exercise time for American put option is as earlier as possible, i.e. t,
and we have P ≥ p.
(iii) The putcall parity for American option:
S −E < C −P < S −Ee
−r(T−t)
(3.4)
As a consequence, P ≤ E.
To prove these properties, we need the following lemma.
Lemma 3.5 Let I or J be two portfolios that contain American options. Suppose I(τ) ≤
J(τ) at some τ ≤ T. Then I(t) ≤ J(t), for all t ≤ τ.
Proof. Suppose I(t) > J(t) at some t ≤ τ. An arbitrageur can long J(t) and short I(t) at
time t to make proﬁt I(t)−J(t) immediately. At later time τ, he can use J(τ) to cover I(τ)
with additional proﬁt J(τ) − I(τ), in case the person who owns I exercises his American
option.
Remark. The equality also holds if I(τ) = J(τ).
Proof of Lemma 3.4.
1. Firstly, we show C ≥ c. If not, then c(τ) > C(τ) for some time τ ≤ T, we can buy
C and sell c at time τ to make a proﬁt c(τ) −C(τ). The right of C is even more than
that of c. This is an arbitrage opportunity which is a contradiction.
Secondly, we show c ≥ C. Consider two portfolios I = C + Ee
−r(T−t)
and J = S.
Suppose we exercise C at some time τ ≤ T, then I(τ) = max¦S
τ
− E, 0¦ +
Ee
−r(T−τ)
and J(τ) = S
τ
. This implies I(τ) ≤ J(τ). By our lemma, I(t) ≤ J(t)
for all t ≤ τ. Since τ ≤ T arbitrary, we conclude I(t) ≤ J(t) for all t ≤ T. Combine
this inequality with the inequality of 2) of section 3.2, we conclude c = C. Further,
early exercise results C(τ) + Ee
−r(T−τ)
< S(τ). Hence, the optimal exercise time
for American option is T.
2. Example. Suppose S = 50, E = 40. If C is exercised before expiration, then the
investor needs to pay 40 to buy the share. However, he can instead invest $40 into
the bank to earn interest and there is a chance that the stock price may go up.
3. Suppose p(t) > P(t). Then we can make an immediate proﬁt by selling p and buying
P. We earn p −P and gain more right. This is a contradiction.
3.2. BASIC PROPERTIES OF OPTION PRICES 23
Next, we show that if we have a P, we should exercise it immediately. We consider
two portfolios I = P + S and J = Ee
−r(T−t)
. If we exercise P at some time τ,
t ≤ τ ≤ T, then
I(τ) = max¦E −S
τ
, 0¦ + S
τ
= max¦E, S
τ
¦ = E.
Putting this money into bank we will receive Ee
r(T−τ)
at time T. On the other hand,
J(τ) = E
−r(T−τ)
. Hence, I(τ) ≥ J(τ). Therefore, I(t) ≥ J(t). Further, we see
that if we exercise P at t, then I(T) = Ee
r(T−t)
is the maximum. Hence we should
exercise P as early as possible.
4. The second inequality follows from the putcall parity (3.3) and the facts that c = C
and P ≥ p. To show the ﬁrst inequality, we consider two portfolios: I = C +E and
J = P +S. Suppose P is exercised at some time τ, t ≤ τ ≤ T. Then we must have
E ≥ S
τ
(otherwise, we should not exercise our put option). Therefore,
J(τ) = max¦E −S
τ
, 0¦ + S
τ
= E
I(τ) = C(τ) + Ee
r(τ−t)
= max¦S
τ
−E, 0¦ + Ee
r(τ−t)
= Ee
r(τ−t)
.
From lemma, we have I(t) > J(t). Hence C + E > P +S.
Examples.
1. Suppose S(t) = 31, E = 30, r = 10%, T −t = 0.25 year, c = 3, p = 2.25. Consider
two portfolios:
I = c + Ee
−r(T−t)
= 3 + 30 e
−0.1×0.25
= 32.26,
J = p + S = 2.25 + 31 = 33.25.
We ﬁnd J(t) > I(t).
Strategy : long the security in portfolio I and short the security in portfolio J. This
results a cashﬂow: −3 + 2.25 + 31 = 30.25. Put this cash into a bank. We will get
30.25 e
0.1×0.25
= 31.02 at time T. Suppose at time T, S
T
> E, we can exercise
c, also we should buy a share for E to close our short position of the stock. Suppose
S
T
< E, the put option will be exercised. This means that we need to buy the share
for E to close our short position. In both cases, we need to buy a share for E to close
the short position. Thus, the net proﬁt is
31.02 −30 = 1.02.
2. Consider the same situation but c = 3 and p = 1. In this case
I = c +Ee
−r(T−t)
= 32.25
J = p + S = 1 + 31 = 32.
24 CHAPTER 3. BLACKSCHOLES ANALYSIS
and we see that J is cheaper.
Strategy: We long J and short I. To long J, we need an initial investment 31 + 1,
to short c, we gain 3. Thus, the net investment is 31 + 1 − 3 = 29 initially. We can
ﬁnance it from the bank, and we need to pay 29 e
0.1×0.25
= 29.73 to the bank at
time T. Now, at T, we must have that either c or p will be exercised. If S
T
> E,
then c is exercised. We need to sell the share for E to close our short position for c.
If S
T
< E, we exercise p. That is, we sell the share for E. In both cases, we sell the
share for E. Thus, the net proﬁt is 30 −29.73 = 0.27.
Remark. P −p is called the time value of a put. The maximal time value is E−Ee
−r(T−t)
.
3.2.4 Dividend Case
Many stocks pay out dividends. These are payments to shareholders out of the proﬁts made
by the company. Since the company’s wealth does not change after paying the dividends,
the stock price, the strike prices fall as the dividends being paid. If a company declared
a cash dividend, the strike price for options was reduced on the exdividend day by the
amount of the dividend.
Lemma 3.6 Suppose a dividend D will be paid during the life of an option. Then we have
for European option
S −D −Ee
−r(T−t)
< c ≤ S (3.5)
−S + D +Ee
−r(T−t)
< p ≤ Ee
−r(T−t)
. (3.6)
and the putcall parity:
c + Ee
−r(T−t)
= p +S −D. (3.7)
For the American options, we have (i)
S −D −E < C −P < S −Ee
−r(T−t)
(3.8)
provided the dividend is paid before exercising the put option, or (ii)
S −E < C −P < S −Ee
−r(T−t)
(3.9)
if the put is exercised before the dividend being paid.
Proof. We consider two portfolios:
I = c + D +Ee
−r(T−t)
,
J = S.
Then at time T,
I(T) = max¦S
T
−E, 0¦ + D + E = max¦S
T
, E¦ + D
J(T) = S
T
+D.
3.2. BASIC PROPERTIES OF OPTION PRICES 25
Hence I(T) ≥ J(T). This yields I(t) ≥ J(t) for all t ≤ T. This proves
c ≥ S −D −Ee
−r(T−t)
.
In other word, c is reduced by an amount D. Similarly, we have
p ≥ D +Ee
−r(T−t)
−S.
That is, p increases by an amount D.
For the putcall parity, we consider
I = c + D +Ee
−r(T−t)
J = S +p.
At time T,
I = J = max¦S
T
, E¦ + D.
This yields the putcall parity for all time.
When there is no dividend, we have shown that
C −P < S −Ee
−r(T−t)
.
When there is dividend payment, we know that
C
D
< C, P
D
> P
Hence,
C
D
−P
D
< C −P < S −Ee
−r(T−t)
.
For the American call option, we should not exercise it early, because the dividend will
cause the stock price to jump down, making the option less attractive. We should exercise
it immediately prior to an exdividend date.
For the American put option, we consider
I = C + D + E, J = P +S.
If we exercise P at τ ≤ T, then S
τ
< E and
I(τ) = D + Ee
r(τ−t)
,
J(τ) = E + D.
We have J(τ) ≤ I(τ). Hence J(t) ≤ I(t) for all t ≤ τ.
If the put option is exercised before the dividend being paid, then we should consider
I = C + E and J = P + S. At τ,
I(τ) = Ee
r(τ−t)
,
J(τ) = E.
Again, we have J(τ) ≤ I(τ). Hence J(t) ≤ I(t) for all t ≤ τ.
26 CHAPTER 3. BLACKSCHOLES ANALYSIS
3.3 The BlackScholes Equation
3.3.1 BlackScholes Equation
The fundamental hypothesis of the BlackScholes analysis is that there is no arbitrage
opportunities. Besides, we make the following additional assumptions:
(1) The asset price follows the lognormal distribution.
(2) There exists a riskfree interest rate r.
(3) No transaction costs.
(4) No dividend paid.
(5) Shorting selling is permitted.
Our purpose is to value the price of an option (call or put). Let V (S, t) denotes for the
price of an option. The randomness of V (S(t), t) would be fully correlated to that S(t).
Thus, we consider a portfolio which contains only S and V , but in opposite position in
order to cancel out the randomness. Then this portfolio becomes deterministic. To be more
precise, let the portfolio be
Π = V −∆S.
In one time step, the change of the portfolio is
dΠ = dV −∆dS.
Here ∆ is held ﬁxed during the time step. From It ˆ os lemma
dΠ = σS
∂V
∂S
−∆
dz
+
µS
∂V
∂S
−µ∆S +
∂V
∂t
+
1
2
σ
2
S
2
∂
2
V
∂S
2
dt. (3.10)
Now, we can eliminate the randomness by choosing
∆ =
∂V
∂S
at the starting time of each time step. The resulting portfolio
dΠ =
∂V
∂t
+
1
2
σ
2
S
2
∂
2
V
∂S
2
dt
is wholly deterministic. From the hypothesis of no arbitrage opportunities, the return,
dΠ
Π
,
should be the same as Π being invested in a riskless bank with interest rate r, i.e.
dΠ
Π
= rdt.
3.3. THE BLACKSCHOLES EQUATION 27
Otherwise, there would be either a net loss or an arbitrage opportunity. Hence we must
have
rΠdt =
µS
∂V
∂S
−µ∆S +
∂V
∂t
+
1
2
σ
2
S
2
∂
2
V
∂S
2
dt
=
∂V
∂t
+
1
2
σ
2
S
2
∂
2
V
∂S
2
dt,
or
∂V
∂t
+
1
2
σ
2
S
2
∂
2
V
∂S
2
= r
V −S
∂V
∂S
. (3.11)
This is the BlackScholes partial differential equation (P.D.E.) for option pricing. Its left
hand side is the return from the hedged portfolio, while its righthand side is the return
from bank deposit. Note that the equation is independent of µ.
Remark. Notice that the BlackScholes equation is invariant under the change of variable
S →λS.
3.3.2 Boundary and Final condition for European options
• Final condition:
c(S, T) = max¦S −E, 0¦
p(S, T) = max¦E −S, 0¦.
In general, the ﬁnal condition is
V (S, T) = Λ(S),
where Λ is the payoff function.
• Boundary conditions:
(i) On S = 0:
c(0, τ) = 0, ∀t ≤ τ ≤ T.
This means that you wouldn’t want to buy a right whose underlying asset costs
nothing.
(ii) On S = 0:
p(0, τ) = Ee
−r(T−τ)
.
This follows from the putcall parity and c(0, t) = 0.
(iii) For call option, at S = ∞:
c(S, t) ∼ S −Ee
−r(T−t)
, as S →∞.
Since S → ∞, the call option must be exercised, and the price of the option
must be closed to S −Ee
−r(T−t)
.
(iv) For put option, at S = ∞:
p(S, t) →0, as S →∞
As S → ∞, the payoff function Λ = max¦E − S, 0¦ is zero. Thus, the put
option is unlikely to be exercised. Hence p(S, T) →0 as S →∞.
28 CHAPTER 3. BLACKSCHOLES ANALYSIS
3.4 Exact solution for the BS equation for European op
tions
3.4.1 Reduction to parabolic equation with constant coefﬁcients
Let us recall the BlackScholes equation
∂V
∂t
+
1
2
σ
2
S
2
∂
2
V
∂S
2
+ rS
∂V
∂S
−rV = 0. (3.12)
This P.D.E. is a parabolic equation with variable coefﬁcients. Notice that this equation is
invariant under S →λS. That is, it is homogeneous in S with degree 0. We therefore make
the following changeofvariable:
dx =
dS
S
,
or equivalently,
x = log
S
E
The fraction S/E makes x dimensionless. The domain S ∈ (0, ∞) becomes x ∈ (−∞, ∞)
and
∂V
∂x
=
∂S
∂x
∂V
∂S
= S
∂V
∂S
,
∂
2
V
∂x
2
=
∂
∂x
S
∂V
∂S
=
∂S
∂x
∂V
∂S
+S
∂S
∂x
∂
2
V
∂S
2
= S
∂V
∂S
+ S
2
∂
2
V
∂S
2
=
∂V
∂x
+ S
2
∂
2
V
∂S
2
.
Next, let us reverse the time by letting
τ = T −t.
Then the BlackScholes equation becomes
∂V
∂τ
=
1
2
σ
2
∂
2
V
∂x
2
+
r −
1
2
σ
2
∂V
∂x
−rV.
We can also make V dimensionless by setting v = V/E. Then v satisﬁes
∂v
∂τ
=
1
2
σ
2
∂
2
v
∂x
2
+
r −
1
2
σ
2
∂v
∂x
−rv. (3.13)
3.4. EXACT SOLUTION FOR THE BS EQUATION FOR EUROPEAN OPTIONS 29
The initial and boundary conditions for v become
c(x, 0) = max¦e
x
−1, 0¦
p(x, 0) = max¦1 −e
x
, 0¦
c(−∞, τ) = 0,
p(−∞, τ) = e
−rτ
,
c(x, τ) → e
x
−e
−rτ
as x →∞
p(x, τ) → 0 as x →∞.
Our goal is to solve v for 0 ≤ τ ≤ T.
3.4.2 Further reduction
In investigating the equation (5.3), it is of the following form:
v
t
+ av
x
+ bv = v
xx
(3.14)
The part, v
t
+ av
x
is call the advection part of (3.14). The term bv is called the source
term, and the tern v
xx
is called the diffusion term. Here , we have absorbed the diffusion
coefﬁcient
1
2
σ
2
in to time by setting t = τ/(
1
2
σ
2
). (We somewhat abuse the notation here.
The new t here is different from the t we used before.)
The advection part:
v
t
+av
x
= (∂
t
+ a∂
x
) v
is a direction derivative along the curve (called characteristic curve)
dx
dt
= a.
This suggests the following changeofvariable:
y = x −at
s = t.
Then the direction derivative become
∂
s
= ∂
t
+ a∂
x
∂
y
= ∂
x
Hence the equation is reduced to
v
s
+ bv = v
yy
.
Next, the equation v
s
+ bv suggests that v behaves like e
bs
along the characteristic curves.
Thus, it is natural to make the following changeofvariable
v = e
bs
u.
30 CHAPTER 3. BLACKSCHOLES ANALYSIS
Then the equation is reduced to
u
s
= u
yy
.
This is the standard heat equation. Its solution can be expressed as
u(y, s) =
1
√
4πs
e
−
(y−z)
2
4s
f(z) dz
where f is the initial data. A simple derivation of this solution is given in the Appendix of
this chapter.
3.4.3 BlackScholes formula
Lert us return to the BlackScholes equation (5.3). Let us denote the rescaled payoff func
tion by
¯
Λ(x). That is,
¯
Λ(x) = Λ(Ee
x
)/E.
The changeofvariables above gives
s = τ/(
1
2
σ
2
)
y = x −as
a = 1 −r/(
1
2
σ
2
)
b = r/(
1
2
σ
2
)
u = e
rτ
v
Then
v(x, τ) = e
−rτ
1
√
2πσ
2
τ
e
−
(x−z(r−
1
2
σ
2
)τ)
2
2σ
2
τ
¯
Λ(z) dz (3.15)
In terms of the original variables, we have the following BlackScholes formula:
V (S, t) = e
−r(T−t)
1
2πσ
2
(T −t)S
e
−
(log(
S
S
)−(r−
1
2
σ
2
)(T−t))
2
2σ
2
(T−t)
Λ(S
) dS
(3.16)
We may express it as
V (S, t) = e
−r(T−t)
{(S
, T, S, t)Λ(S
) dS
(3.17)
Here,
{(S
, T, S, t) :=
1
2πσ
2
(T −t)S
e
−
(log(
S
S
)−(r−
1
2
σ
2
)(T−t))
2
2σ
2
(T−t)
. (3.18)
This is the transition probability density of an asset price model with growth rate r and
volatility σ. In other words, V is the present value of the expectation of the payoff under
an asset price model whose volatility is σ and whose growth rate is r. We shall come back
to this point later.
3.4. EXACT SOLUTION FOR THE BS EQUATION FOR EUROPEAN OPTIONS 31
3.4.4 Special cases
1. European call option. The rescaled payoff function for a European call option is
¯
Λ(z) = max¦e
z
−1, 0¦.
Then
v(x, τ) = e
−rτ
∞
0
1
√
2πσ
2
τ
e
−(x−z−(
1
2
σ
2
−r)τ)
2
/(2σ
2
τ)
(e
z
−1) dz.
This can be integrated. Finally, we get the exact solution for the European call option
c (S, t) = S^(d
1
) −Ee
−r(T−t)
^(d
2
), (3.19)
^(y) =
1
√
2π
y
−∞
e
−
z
2
2
dz, (3.20)
d
1
=
log(
S
E
) + (r +
1
2
σ
2
)(T −t)
σ
√
T −t
, (3.21)
d
2
=
log(
S
E
) + (r −
1
2
σ
2
)(T −t)
σ
√
T −t
. (3.22)
Exercise. Prove the formula (3.19).
2. European put option. Recall the putcall parity
c +Ee
−r(T−t)
= p + S.
We can obtain the price for p from c:
p(S, t) = Ee
−r(T−t)
^(−d
2
) −S^(−d
1
). (3.23)
Exercise. Show that ^(d
1
) −1 = ^(−d
1
). Use this to prove (3.23).
3. Forward contract Recall that a forward contract is an agreement between two par
ties to buy or sell an asset at certain time in the future for certain price. The payoff
function for such a forward contract is
Λ(S) = S −E.
The value V for this contract also satisﬁes the BS equation. Thus, its solution is
given by
V = Ee
−rτ
u,
where
u(x, τ) =
1
√
2πσ
2
τ
∞
−∞
e
−
(y−z−(r−
1
2
σ
2
)τ)
2
2σ
2
τ
(e
z
−1) dz
= e
x+rτ
−1.
32 CHAPTER 3. BLACKSCHOLES ANALYSIS
Hence,
V (S, t) = S −Ee
−r(T−t)
. (3.24)
This means that the current value of a forward contract is nothing but the difference
of S and the discounted E. Notice that this value is independent of the volatility σ of
the underlying asset.
Exercise. Show that the payoff function of a portfolio c −p is S −E. From this and
the BlackScholes formula (3.16), show the formula of the putcall parity.
4. Cashornothing. A contact with cashornothing is just like a bet. If S
T
> E, then
the reward is B. Otherwise, you get nothing. The payoff function is
Λ(S) =
B if S > E
0 otherwise.
Using the BlackScholes formula (3.16), we obtain the value of a cashornothing
contract to be
V (S, t) = Be
−r(T−t)
^(d
2
). (3.25)
5. Supershare. Supershare is a binary option whose payoff function is deﬁned to be
Λ(S) =
B if E
1
< S < E
2
0 otherwise.
One can show that the value for this binary option is
V (S, t) = Be
−r(T−t)
(^(d
2
(E
1
)) −^(d
2
(E
2
)))
where d
2
(E) is given by (3.22).
6. Deterministic case (σ = 0). In this case, the BlackScholes equation is reduced to
V
t
+ rSV
s
−rV = 0.
Or in τ, x and u variables:
u
τ
−ru
x
= 0
with initial data
u(x, 0) = Λ(Ee
x
),
Thus,
u(x, τ) = Λ(Se
x+rτ)
.
Or
V (S, t) = e
−r(T−t)
Λ(Se
r(T−t)
).
This means that when the process is deterministic, the value of the option is the
payoff function evaluated at the future price of S at T (that is Se
r(T−t)
), and then
discounted by the factor e
−r(T−t)
.
3.5. RISK NEUTRALITY 33
3.5 Risk Neutrality
Notice that the growth rate µ does not appear in the BlackScholes equation. The option
may be valued as if all random walks involved are risk neutral. This means that the drift
term (growth rate) µ in the asset pricing model can be replaced by r. The option is then
valued by calculating the present value of its expected return at expiry. Recall the lognormal
probability density function with growth rate r, volatility σ is
{(S
, T, S, t) :=
1
2πσ
2
(T −t)S
e
−
(log(
S
S
)−(r−
1
2
σ
2
)(T−t))
2
2σ
2
(T−t)
. (3.26)
This is the transition probability density of an asset price model in a riskneutral world:
dS
S
= rdt + σdz. (3.27)
The expected return at time T in this riskneutral world is
{(S
, T, s, t)Λ(S
)dS
.
At time t, this value should be discounted by e
−r(T−T)
:
V (S, t) = e
−r(T−t)
{(S
, T, S, t)Λ(S
)dS
.
We may reinvestigate the function ^ and the parameters d
i
in the BlackScholes formula.
After some calculation, we ﬁnd
^(d
2
) =
∞
E
{(S
, T, S, t)dS
. (3.28)
This is the probability of the event ¦
˜
S ≥ E¦, where
˜
S obeys the riskneutral pricing model:
d
˜
S
˜
S
= rdt + σdz.
Similarly, one can show that
^(d
1
) =
∞
E
{(S
, T, S, t)S
dS
Se
r(T−t)
. (3.29)
is the expectation of
˜
S at T when S = 1 at t and under the condition that
˜
S ≥ E at T.
3.6 The delta hedging
Hedging is the reduction of sensitivity of a portfolio to the movement of the underlying
of asset by taking opposite position in different ﬁnancial instruments. The BlackScholes
34 CHAPTER 3. BLACKSCHOLES ANALYSIS
analysis is a dynamical strategy. The delta hedge is instantaneously risk free. It requires
a continuous rebalancing of the portfolio and the ratio of the holdings in the asset and the
derivative product. The delta for a whole portfolio is ∆ =
∂Π
∂S
. This is the sensitivity of Π
against the change of S. By taking dΠ−∆ dS, the sensitivity of the portfolio to the asset
price change is instantaneously zero.
Besides the delta helge, there are more sophisticated trading strategies such as:
Gamma: Γ =
∂
2
Π
∂
2
S
2
,
Theta: θ = −
∂Π
∂t
,
Vega: =
∂Π
∂σ
,
rho: ρ =
∂Π
∂r
.
Hedging against any of these dependencies requires the use of another option as well as the
asset itself. With a suitable balance of the underlying asset and other derivatives, hedgers
can eliminate the shortterm dependence of the portfolio on the movement in t, S, σ, r.
For the Deltahedge for the European call and put options, we have the following propo
sitions.
Proposition 1 For European call options, its ∆ hedge is given by
∆ = ^(d
1
).
Proof. By deﬁnition,
∂c
∂S
= ^(d
1
) + S ^
(d
1
) d
1S
−Ee
−r(T−t)
^
(d
2
)d
2S
.
Since
d
1
=
log(
S
E
) + (r +
σ
2
2
)(T −t)
σ
√
T −t
,
we have
d
1S
=
1
Sσ
(T −t)
,
d
2S
=
1
Sσ
(T −t)
,
^
(d
i
) =
1
√
2π
e
−d
2
i
.
Hence,
∂c
∂S
= ^(d
1
) +
S^
(d
1
) −Ee
−r(T−t)
^
(d
2
)
/(Sσ
√
T −t)
≡ ^(d
1
) +1/(Sσ
√
T −t).
3.6. THE DELTA HEDGING 35
We claim that 1 = 0. Or equivalently,
S
E
^
(d
1
)
^
(d
2
)
= e
−rτ
This follows from the computation below.
S
E
^
(d
1
)
^
(d
2
)
= e
x
e
−(d
2
1
−d
2
2
)/2
.
From (3.21)(3.22),
d
2
1
−d
2
2
=
1
σ
2
τ
(x + rτ +
σ
2
τ)
2
−(x + rτ −
σ
2
τ)
2
= 2(x + rτ)
Hence,
S
E
^
(d
1
)
^
(d
2
)
= e
x
e
−x−rτ
= e
−rτ
.
Proposition 2 For European put options, its ∆ hedge is given by
∆ = ^(−d
1
).
Proof. From the putcall parity,
∆ =
∂p
∂S
=
∂c
∂S
−1 = ^(d
1
) −1 = −^(−d
1
),
3.6.1 TimeDependent r, σ, µ
Suppose r, σ, µ are functions of r, but also deterministic. The BlackScholes remains the
same. We use the changeofvariables:
S = Ee
x
, V = Ev, τ = T −t.
The BlackScholes equation is converted to
v
τ
=
σ
2
(τ)
2
v
xx
+ (r(τ) −
σ
2
(τ)
2
)v
x
−r(τ)v (3.30)
We look for a new time variable ˆ τ such that
dˆ τ = σ
2
(τ)dτ
36 CHAPTER 3. BLACKSCHOLES ANALYSIS
For instance, we can choose
ˆ τ =
τ
0
σ
2
(τ) dτ.
Then the equation becomes
v
ˆ τ
=
1
2
v
xx
+ a(ˆ τ)v
x
−b(ˆ τ)v. (3.31)
To eliminate a(ˆ τ), we consider the characteristic equation:
dx
dˆ τ
= −a(ˆ τ)
This can be integrated and yields
x = −
ˆ τ
0
a(τ
)dτ
+ y,
Or equivalently,
y = x +
ˆ τ
0
a(τ
)dτ
≡ x +A(ˆ τ).
Now, we consider the changeofvariable:
x
ˆ τ
→
y
ˆ τ
1
.
Then,
∂
∂x
[
ˆ τ
=
∂y
∂x
[
ˆ τ
∂
∂y
=
∂
∂y
,
and
∂
∂ˆ τ
1
[
y
=
∂ˆ τ
∂ˆ τ
1
∂
∂ˆ τ
+
∂x
∂ˆ τ
1
∂
∂x
=
∂
∂ˆ τ
−a(ˆ τ)
∂
∂x
.
The equation (3.31)is transformed to
v
ˆ τ
1
=
1
2
v
yy
−b(ˆ τ
1
)v.
Let B(ˆ τ
1
) =
ˆ τ
1
0
b(τ
)dτ
, and u = e
B(ˆ τ
1
)
v, then u
ˆ τ
1
=
1
2
u
yy
. And we can solve this heat
equation explicitly.
3.7 Trading strategy involving options
The options whose payoff are max¦S
T
− E, 0¦ or max¦E − S
T
, 0¦ are called vanilla
option. In this section, we shall discuss more general payoff functions. The goal is to
design a portfolio involving vanilla option with a designed payoff function.
3.7. TRADING STRATEGY INVOLVING OPTIONS 37
3.7.1 Strategies involving a single option and stock
There are four cases:
a. Π = S − c (writing a covered call option). In this strategy, we short a call, long a
share to cover c. The payoff of Π is Λ = S −max(S −E, 0) = min¦S, E¦. In this
case, we anticipate the stock price will increase.
b. Π = c − S (reverse of a covered call). In this strategy, we anticipate the stock price
will decrease. And Λ = −min¦S, E¦.
c. Π = p +S (protective put). In this portfolio, we long a p and buy a share to cover p.
We anticipate the stock price will increase. The payoff is Λ = S +max¦E−S, 0¦ =
max¦S, E¦.
d. Π = −p − S (reverse of a protective put). We do not anticipate the stock price will
increase. The payoff is −max¦S, E¦.
Below are the payoff functions for the above four cases.
E S
Λ
E S
Λ
(a) (b)
E S
Λ
E S
Λ
38 CHAPTER 3. BLACKSCHOLES ANALYSIS
(c) (d)
3.7.2 Bull spreads
In this strategy, an investor anticipates the stock price will increase. However, he would
like to give up some of his right if the price goes beyond certain price, say E
2
. Indeed, he
does not anticipate the stock price will increase beyond E
2
. Hence he does want to own a
right beyond E
2
. Such a portfolio can be designed as
Π = C
E
1
−C
E
2
, E
1
< E
2
,
where C
E
i
is a European call option with exercise price E
i
and C
E
1
, C
E
2
have the same
expiry. The payoff
Λ = max¦S
T
−E
1
, 0¦ −max¦S
T
−E
2
, 0¦
=
0 if S
T
< E
S
T
−E
1
if E
1
< S
T
< E
2
E
2
−E
1
if S
T
> E
2
E
1
E
2 S
Λ
E
2
−E
1
Since E
1
< E
2
, we have C
E
1
> C
E
2
. A bull spread, when created from C
E
1
− C
E
2
,
requires an initial investment. We can describe the strategy by saying that the investor has a
call option with a strike price E
1
and has chosen to give up some upside potential by selling
a call option with strike price E
2
> E
1
. In return, the investor gets E
2
− E
1
if the price
goes up beyond E
2
.
Example: C
E
1
= 3, C
E
2
= 1 and E
1
= 30, E
2
= 35. The cost of the strategy is 2. The
payoff
0 if S
T
≤ 30
S
T
−30 if 30 < S
T
< 35
5 if S
T
≥ 35
The bull spread can also be created by using put options
Π = P
E
1
−P
E
2
, E
1
< E
2
.
3.7. TRADING STRATEGY INVOLVING OPTIONS 39
3.7.3 Bear spreads
An investor entering into a bull spread is hoping that the stock price will increase. By
contrast, an investor entering into a bear spread is expecting the stock price will go down.
The bear spread is
Π = C
E
2
−C
E
1
, E
1
< E
2
.
There is cash ﬂow entered (C
E
2
−C
E
1
). The payoff is
3.7.4 Butterﬂy spread
If an investor anticipate the stock price will stay in certain region, say, E
1
< S
T
< E
3
, he
or she can have a butterﬂy spread such that the payoff function is positive in that region and
he or she gives up the return outside that region.
1. Butterﬂy spread using calls: Deﬁne the portfolio:
Π = C
E
1
−2C
E
2
+ C
E
3
, with E
1
< E
2
< E
3
.
where E
3
= E
2
+ (E
2
− E
1
). Its payoff function is a piecewise linear function and
is determined by Λ(E
1
) = Λ(E
3
) = 0, Λ(E
2
) = E
2
−E
1
. Below is the graph of its
payoff function.
E
1
S−E
1
E
2
−E
1
E
2
E
3
E
2
−S
S
Λ
Example: Suppose a certain stock is currently worth 61. A investor who feels that it
is unlikely that there will be signiﬁcant price move in the next 6 month. Suppose the
market of 6 month calls are
E C
55 10
60 7
65 5
The investor creates a butterﬂy spread by
Π = C
E
1
−2C
E
2
+ C
E
3
.
40 CHAPTER 3. BLACKSCHOLES ANALYSIS
The cost is 10 + 5 −2 7 = 1. The payoff is
55
5
60 65 S
Λ
2. Butterﬂy spread using puts.
P
E
1
+ P
E
3
−2P
E
2
, E
1
< E
3
, E
2
=
E
1
+ E
3
2
.
ϕ
E
2
=
linear
∆E if S = E
2
0 if S < E
2
−∆E, or S > E
2
−∆E
Remark 1. Suppose European options were available for every possible strike price E,
then any payoff function could be created theoretically:
Λ(S) =
¸
Λ
i
∆E
ϕ
E
i
where E
i
= i∆E, Λ
i
is constant. Then Λ(E
i
) = Λ
i
and Λ is linear on every interval
(E
i
, E
i+1
) and Λ is continuous. As ∆E → 0, we can approximate any payoff function by
using butterﬂy spreads.
Remark 2. One can also use cashornothing to create any payoff function:
Λ(S) =
¸
Λ
i
∆E
ψS −E
i
,
where
ψ(S) := H(S) −H(S −∆E).
The value for such a portfolio is
V = e
−r(T−t)
{(S
, T, S, t)Λ(S
)dS
,
= e
−r(T−t)
ΣΛ
i
{(E
i
≤ S ≤ E
i+1
).
Chapter 4
Variations on BlackScholes models
4.1 Options on dividendpaying assets
Dividends are payments to the shareholders out of the proﬁts made by the company. We
will consider two “deterministic” models for dividend. One has constant dividend yield.
The other has discrete dividend payments.
4.1.1 Constant dividend yield
Suppose that in a short time dt, the underlying asset pays out a dividend D
0
Sdt, where D
0
is a constant, called the dividend yield. This continuous dividend structure is a good model
for index options and for shortdated currency options. In the latter case, D
0
= r
f
, the
foreign interest rate.
As the dividend is paid, the return
dS
S
must fall by the amount of the dividend payment
D
0
dt. It follows the s.d.e. for the asset price is
dS
S
= (µ −D
0
)dt + σdz.
For a portfolio : Π = V −∆S, we choose ∆ =
∂V
∂S
in order to eliminate the randomness of
dΠ. In one time step, the change of portfolio is
dΠ = dV −∆dS −∆D
0
Sdt,
the last term −∆D
0
Sdt is the dividend our assets received. Thus
dΠ = dV −∆(dS + D
0
Sdt)
= σS
∂V
∂S
−∆
dz
+
(µ −D
0
)S
∂V
∂S
+
1
2
σ
2
S
2
∂
2
V
∂S
2
+ V
t
−(µ −D
0
)∆S −∆D
0
S
dt
=
V
t
−D
0
S
∂V
∂S
+
1
2
σ
2
S
2
∂
2
V
∂S
2
dt,
41
42 CHAPTER 4. VARIATIONS ON BLACKSCHOLES MODELS
Here, we have chosen ∆ =
∂V
∂S
to eliminate the random term. From the absence of arbitrage
opportunities, we must have
dΠ = rΠdt.
Thus,
V
t
−D
0
S
∂V
∂S
+
1
2
σ
2
S
2
∂
2
V
∂S
2
= r
V −S
∂V
∂S
.
i.e.,
V
t
+
1
2
σ
2
S
2
∂
2
V
∂S
2
+ (r −D
0
)S
∂V
∂S
−rV = 0
This is the BlackScholes equation when there is a continuous dividend payment.
The boundary conditions are:
c(0, t) = 0,
c(S, t) ∼ Se
−D
0
(T−t)
The latter is the asset price S discounted by e
−D
0
(T−t)
from the payment of the dividend.
The payoff function c(S, T) = Λ(S) = max¦S −E, 0¦.
To ﬁnd the solution, let us consider
c(S, t) = e
−D
0
(T−t)
c
1
(S, t).
Then c
1
satisﬁes the original BlackScholes equation with r replaced by r − D
0
and the
same ﬁnal condition. The boundary conditions for c
1
are
c
1
(0, t) = 0,
c
1
(S, t) ∼ S as S →∞
Hence,
c
1
(S, t) = S^(d
1,0
) −Ee
−(r−D
0
)(T−t)
^(d
2,0
)
or
c(S, t) = Se
−D
0
(T−t)
^(d
1,0
) −Ee
−r(T−t)
^(d
2,0
)
where
d
1,0
=
ln
S
E
+ (r −D
0
+
1
2
σ
2
)(T −t)
σ
√
T −t
,
d
2,0
= d
1,0
−σ
√
T −t
Remark. c `as D
0
.
Exercise. Derive the putcall parity for the European options on dividendpaying assets.
4.2. WARRANTS 43
4.1.2 Discrete dividend payments
Suppose our asset pays just one dividend during the life time of the option, say at time t
d
.
The dividend yield is a constant. At t
d
+, the asset holder receiver a payment d
y
S(t
d
−).
Hence,
S(t
d
+) = S(t
d
−) −d
y
S(t
d
−) = (1 −d
y
)S(t
d
−).
We claim that across the jumps, V should be continuous, i.e.,
V (S(t
d
−), t
d
−) = V (S(t
d
+), t
d
+).
Reason : Otherwise, there is a net loss or gain from buying V before t
d
then sell it right
after t
d
. To ﬁnd V (S, t), here is a procedure.
1. Solve the BlackScholes from T to T
d
+ to obtain V (S, t
d
+) (using the payoff func
tion Λ)
2. Adjusting V by
V (S, t
d
−) = V ((1 −d
y
)S, t
d
+)
3. Solve BlackScholes equation fromt
d
to t with the ﬁnal condition V ((1−d
y
)S, t
d
+).
Let c
d
be the European option for this dividendpaying asset. Then
c
d
(S, t) = c(S, t, E) for t
d
+ ≤ t ≤ T
c
d
(S, t
d
−) = c
d
(S(1 −d
y
), t
d
+) = c(S(1 −d
y
), t, E)
Note that
c(S(1 −d
y
), T, E) = max¦S(1 −d
y
) −E, 0¦ = (1 −d
y
) max¦S −(1 −d
y
)
−1
E, 0¦
and the linearity of the BlackScholes equation, we obtain
c(S(1 −d
y
), t, E) = (1 −d
y
)c(S, t, (1 −d
y
)
−1
E).
4.2 Warrants
An European warrant is a right to purchase an underlying stock at price X at expiry. We
want to determine the price of a warrant. Suppose a company has N outstanding shares and
M outstanding European warrants. Suppose each warrant entitles the holder to purchase
γ share from the company at time T at price X per share. Let V
T
be the value of the
company’s equity at T. If the warrant holders exercise, then the company received a cash
inﬂow MγX and the company’s equity increases to V
T
+ MγX. This value is distributed
to N + Mγ shares. Hence the share price becomes
V
T
+ MγX
N + γM
.
44 CHAPTER 4. VARIATIONS ON BLACKSCHOLES MODELS
The payoff to the warrant holder is
max
γ
¸
V
T
+MγX
N + Mγ
−X
, 0
=
Nγ
N + Mγ
max
V
T
N
−X, 0
.
This is exactly the payoff function for a European call. Thus, The value of the warrant at
time t should be
w =
Nγ
N + Mγ
c(
V
N
, t, X),
where V is the value of the company’s equity at time t, c(S, t, X) is the value of a European
call with strike price X. Since V = NS +Mw, i.e.,
V
N
= S +
M
N
w. We obtain a nonlinear
algebraic equation for w:
w =
Nγ
N + Mγ
c(S +
M
N
w, t, X)
=
Nγ
N + Mγ
(S +
M
N
w)^(d
1
) −Xe
−r(T−t)
^(d
2
)
where
d
1
=
log((S +
M
N
w)/X) + (r +
1
2
σ
2
)(T −t)
σ
√
T −t
d
2
= d
1
−σ
√
T −t.
This algebraic equation can be solved numerically.
4.3 Futures and futures options
4.3.1 Forward contracts
Recall that a forward contract is an agreement between two parties to buy or sell an under
lying asset on a certain price E at a certain future time T. Here, E is called the delivery
price. The payoff function for this forward contract is Λ = S
T
− E. Based on the no
arbitrage opportunity, the price for this forward contract is
f = S −Ee
−r(T−t)
.
Deﬁnition 3.2 The forward price F for a forward contract is deﬁned to be the delivery
price which would make that contract have zero value, i.e.,
F
t
= S
t
e
r(T−t)
.
One can take another point of view. Consider a party who is short the contract. He can
borrow an amount of money S
t
at time t to buy an asset and use it to close his short
position at T. The money he received at expiry, F, is used to pay the loan. If no arbitrage
opportunities, then
F = S
t
e
r(T−t)
.
4.3. FUTURES AND FUTURES OPTIONS 45
4.3.2 Futures
Futures are very similar to the forward contracts, except they are traded in an exchange,
thus, they are required to be standardized. This includes size, quality, price, expiry,. . . etc.
Let us explain the characters of a future by the following example.
1. Trading future contracts
• Suppose you call your broker to buy one July corn futures contract (5,000
bushels) on the Chicago Board of Trade (CBOT) at current market price.
• The broker send this signal to traders on the ﬂoor of the exchange.
• The trader signal this to ask other traders to sell, if no one want to sell, the trader
who represents you will raise the price and eventually ﬁnd someone to sell
• Conﬁrmation: Price obtained are sent back to you.
2. Speciﬁcation of the futures: In the above example, the speciﬁcation of this future is
• Asset : quality
• Contract size: 5,000 bushel
• Delivery arrangement: delivery month is on December
• price quotes
• Daily price movement limits: these are speciﬁed by the exchange.
• Position limits: the maximum number of contracts that a speculator may hold.
3. Operation of margins
• Marking to market: Suppose an investor who contacts his or her broker on June
1, 1992, to buy two December 1992 gold futures contracts on New York Com
modity Exchange. We suppose that the current future price is $400 per ounce.
The contract size is $100 ounces, the investor want to buy $200 ounces at this
price. The broker will require the investor to deposit funds in a “margin ac
count”. The initial margin, say is $2,000 per contract. As the futures prices
move everyday, the amount of money in the margin account also changes. Sup
pose, for example, by the end of June 1, the futures price has dropped from
$400 to $397. The investor has a loss of $2003=600. This balance in the mar
gin account would therefore be reduced by $600. Maintaining margin needs to
deposit. Certain account of money to keep that futures contract.
• Maintenance margin: To insure the balance in the margin account never be
comes negative, a maintenance margin, which is usually lower than the initial
margin, is set.
Theorem 4.2 Forward price and futures price are equal when the interest rates are con
stant.
46 CHAPTER 4. VARIATIONS ON BLACKSCHOLES MODELS
Proof. Suppose a futures contract lasts for n days. Let the future prices are
F
0
, , F
n
at the end of each business day. Let δ be the riskfree interest rate per day. Consider the
following two strategies:
1. Invest G
0
in a riskfree bond and take a long position of amount e
nδ
forward contract.
At day n, G
0
e
nδ
is used to buy the underlying asset at price S
T
e
nδ
.
2. • Invest F
0
amount of money in a riskfree bond.
• Take a long position of future e
δ
amount of at the end of day 0.
• Take a long position of future e
2δ
amount of at the end of day 1.
Day 0 1 2 n −1 n
futures price F
0
F
1
F
2
F
n−1
F
n
position e
δ
e
2δ
e
3δ
e
nδ
0
gain/loss 0 e
δ
(F
1
−F
0
) e
2δ
(F
2
−F
1
) e
nδ
(F
n
−F
n−1
)
compound e
δ
(F
1
−F
0
)e
(n−1)δ
e
2δ
(F
2
−F
0
)e
(n−2)δ
(F
n
−F
n−1
)e
nδ
The total gain/loss from the long position of the futures is
n
¸
i=1
(F
i
−F
i−1
)e
iδ
e
(n−i)δ
= (F
n
−F
0
)e
nδ
= (S
T
−F
0
)e
nδ
.
If we invest F
0
initially, at T, we received F
0
e
nδ
, N
0
investment is required for all
the long future positions. The payoff of strategy 2 is
F
0
e
nδ
+ (S
T
−F
0
)e
nδ
= S
T
e
nδ
.
Since both strategies have the same payoff, we conclude their initial investments must be
the same, i.e., F
0
= G
0
= S
T
e
r(T−t)
.
4.3.3 Futures options
Options on futures are traded in many different exchanges. They require the delivery of an
underlying futures contract when exercised. When a call futures option is exercised, the
holder acquires a long position in the underlying futures contract plus a cash amount equal
to the current futures price minus the exercise price.
Example. An investor who has a September futures call option on 25,000 pounds of cop
per with exercise price E = 70 cents/pound. Suppose the current future price of copper
4.3. FUTURES AND FUTURES OPTIONS 47
for delivery in September is 80 cents/pound. If the option is exercised, the investor re
ceived 10 cents 25, 000+long position in futures contract to buy 25,000 pound of copper
in September at price 80 cents/pound.
The maturity date of futures option is generally on, or a few days before, the earlist
delivery date of the underlying futures contract.
Futures options are more attractive to investors than options on the underlying assets
when it is cheaper or more convenient to deliver futures contracts rather than the asset itself.
Futures options are usually more liquid and involved lower transaction costs.
4.3.4 BlackScholes analysis on futures options
As we have seen that the futures price is identical to the forward price when the interest
rate is a constant, i.e., F = Se
r(T−t)
. From It ˆ os lemma, we obtain a pricing model for F:
dF = (F
t
+
1
2
F
SS
σ
2
S
2
)dt + F
S
dS
= (−re
r(T−t)
S)dt + Se
r(T−t)
dS
S
= (−rF)dt + F(µdt + σdz)
= ((µ −r)F)dt + Fσdz.
Hence,
dF
F
= (µ −r)dt + σdz. (4.1)
This means that the futures price is the same as a stock paying a dividend yield at rate r.
Next, we study the value V of a futures option. It is a function of F, t. Consider a
portfolio
Π = V −∆F
We choose ∆ =
∂V
∂F
to eliminate randomness of dΠ. Then
dΠ = dV −∆dF
= (
∂V
∂F
µ
F
F +
∂V
∂t
+
1
2
∂
2
V
∂F
2
σ
2
F
2
)dt +
∂V
∂F
σFdz −
∂V
∂F
(µ
F
Fdt + σFdz)
= (
∂V
∂t
+
1
2
∂
2
V
∂F
2
σ
2
F
2
)dt.
Since it costs nothing to enter into a future contract, the cost of setting up the above portfolio
is just V . Thus based on the no arbitrage opportunity,
dΠ = rV dt,
Thus, we obtain
V
t
+
1
2
σ
2
F
2
∂
2
V
∂F
2
= rV
The payoff function for a call option is Λ = max¦F − E, 0¦. This is because at time T,
S
T
= F
T
.
48 CHAPTER 4. VARIATIONS ON BLACKSCHOLES MODELS
To solve this equation, we recall the option price equation for stock paying dividend is
V
t
+
1
2
σ
2
S
2
∂
2
V
∂S
2
+ (r −D
0
)S
∂V
∂S
−rV = 0
In our case, D
0
= r, so the futures call option
c(F, t) = e
−r(T−t)
c
1
(F, t)
where c
1
satisﬁes BlackScholes equation with r replaced by r −r = 0. This gives
c
1
(F, t) = F^(d
1
) −E^(d
2
),
d
2
=
ln
F
E
−
1
2
σ
2
(T −t)
σ
√
T −t
,
d
1
= d
2
+ σ
√
T −t
Notice that this V is the same as
˜
V (S(F, t), t), where
˜
V is the solution of the option corre
sponding to the underlying asset S. That is
˜
V (S, t) = ^(d
1
) −Ee
−r(T−t)
^(d
2
)
d
2
=
ln
S
E
+ (r −
1
2
σ
2
)(T −t)
σ
√
T −t
,
d
1
= d
2
+ σ
√
T −t
We can write this
˜
V in terms of F by S = Fe
−r(T−t)
. Plug this into the above equation to
obtain
d
2
=
ln
Fe
−r(T−t)
E
−
1
2
σ
2
(T −t)
σ
√
T −t
=
ln
F
E
−
1
2
σ
2
(T −t)
σ
√
T −t
,
˜
V (S(F, t), t) = S^(d
1
) −Ee
−r(T−t)
^(d
2
)
= Fe
−r(T−t)
^(d
1
) −Ee
−r(T−t)
^(d
2
)
= V (F, t)
Conclusion:
1. Futures price F = S
t
e
r(T−t)
.
2. Futures price is the same as a stock paying dividend at yield rate r.
3. The price for future options is the same as the price for options on the underlying
assets.
Finally, let us ﬁnd the putcall parity for futures options.
4.3. FUTURES AND FUTURES OPTIONS 49
Proposition 3
c +Ee
−r(T−t)
= p + Fe
−r(T−t)
(4.2)
Proof. Consider two portfolios:
A = c + Ee
−r(T−t)
B = p +Fe
−r(T−t)
+ a futures contract
At time T,
Λ
A
= max¦F
T
−E, 0¦ + E = max¦F
T
, E¦,
Λ
B
= max¦E −F
T
, 0¦ + F + (F
T
−F) = max¦E, F
T
¦.
Hence we obtain A = B.
50 CHAPTER 4. VARIATIONS ON BLACKSCHOLES MODELS
Chapter 5
Numerical Methods
5.1 Monte Carlo method
We recall that the value of a European option is given by
V (S, t) = e
−r(T−t)
{(
˜
S, T, S, t)Λ(
˜
S)d
˜
S
where Λ is the payoff function, { is the transition probability density function of
˜
S which
satisﬁes
d
˜
S
˜
S
= rdt +σdz, ( initial state S(t) = S) (5.1)
i.e., it is the asset price model in the riskneutral world. The Monte Carlo simulation is a
numerical procedure to estimate V based on this formula.
To ﬁnd V , we sample, say, 10,000 paths from(5.1). We obtain S
i
(T), i = 1, . . . , 10000.
We then approximate V by
V ≈ e
−r(T−t)
1
N
N
¸
i=1
Λ(S
i
(T)).
To sample a path from (5.1), we divide the interval [t, T] into M subinterval with equal
length ∆t =
T−t
M
. We sample M random numbers
k
, k = 1, . . . , M with distribution
N(0, 1)(i.e., the normal distribution with mean 0, variance 1). We then deﬁne S
i
(t + k∆t)
by
S(t + k∆t) −S(t + (k −1)∆t)
S(t + (k −1)∆t)
= r∆t + σ
k
√
∆t.
Remark. The error of a MonteCarlo method is O
1
√
N
. If there is only one underlying
asset, the Monte Carlo does not have any advantage. However, if there are many under
lying assets, say more than three, the corresponding Black Scholes equation is a diffusion
equation in high dimensions. In this case, ﬁnite difference method is very difﬁcult and the
Monte Carlo method wins.
51
52 CHAPTER 5. NUMERICAL METHODS
5.2 Binomial Methods
In binomial method, we ﬁrst simulate a riskneutral asset price model forward in time by
a binomial model, then we determine the option price from the expectation of the payoff
function according to the price distribution of the asset in the riskneutral world.
5.2.1 Binomial method for asset price model
We consider the underlying asset is riskneutral, i,e.,
dS
S
= rdt + σdz (5.2)
We shall approximate this continuous model by the following discrete model.
First, we assume our discrete asset prices only take discrete values S
j
= S
0
e
j∆x
, where
S
0
is the asset price at current time t, and ∆x is a parameter to be determined later. We
want to ﬁnd the probability distribution of the asset price in a riskneutral world at time T
whose current price is S
0
.
Next, we discrete the continuous model in time, namely, we partition [t, T] into N
subintervals with equal length ∆t = (T −t)/N. The discrete asset price model is:
if the asset price is at S
j
at time step n, then the asset price will move up to S
j+1
=
S
j
u with probability p and move down to S
j−1
= S
j
d with probability 1 − p. Here,
u = e
∆x
and d = e
−∆x
.
Let us denote the probability that the price is at S
j
at time step n by P
n
j
. Then P
0
0
= 1 and
P
n
j
is exactly the binomial distribution:
P
n
j
=
n
r
p
r
(1 −p)
n−r
n + j = 2r
0 otherwise.
This discrete model depends on two parameters: u and p. ( The down ratio d = 1/u.) They
are determined by the conditions so that the discrete model and the continuous model have
the same mean and variance in one time step ∆t. We recall that these conditions are
pu + (1 −p)d = e
r∆t
pu
2
+ (1 −p)d
2
= e
(2r+σ
2
)∆t
Thus, p and u can be expressed in terms of r, σ and ∆t. A simple calculation gives
u = 1 + σ
√
∆t + O(∆t)
p =
1
2
+ O(
√
∆t)
We should require ∆t is chosen so that 0 ≤ p ≤ 1.
Remark. If we denote log S/E by x, then the movement of S on the discrete values S
j
corresponds to a movement of x on x
j
= j∆x. This movement is exactly the random walk
we introduced in Chapter 2.
5.3. FINITE DIFFERENCE METHODS (FOR THE MODIFIED BS EQ.) 53
5.2.2 Binomial method for option
Since the asset price only takes discrete values S
j
, we shall approximate V (S
j
, t +n∆t) by
V
n
j
. We recall that V
n
is the expected value of the option at (n+1)∆t discounted by e
−r∆t
.
If S takes value at S
j
at time step n, then S takes values S
j+1
with probability p and S
j−1
with probability 1 −p. Therefore, the expected value of V at time step n should satisﬁes
e
r∆t
V
n
j
= pV
n+1
j+1
+ (1 −p)V
n+1
j−1
.
Example. For put option,
T = 5 months = 0.4167 year
∆t = 1 months = 0.0833 year
r = 0.1, σ = 0.4
S = $50, E = $50
u = e
σ
√
∆t
= 1.1224
d = 0.8909
p = 0.5076
e
r∆t
= 1.0084
We begin to generate a binomial tree from S = 50 consisting of S
n
j
= Su
r
d
n−r
, where
n + j = 2r, −n ≤ j ≤ n. Then we compute V
n
j
inductively from n = N to n = 0 by
e
r∆t
V
n
j
= pV
n+1
j+1
+ (1 −p)V
n+1
j−1
.
with V
N
j
being the payoff function. The value V
0
0
is our answer.
5.3 Finite difference methods (for the modiﬁed BS eq.)
In this section, we shall solve the BlackScholes equation by ﬁnite difference methods.
Recall that the BlackScholes equation is
V
t
+
1
2
σ
2
S
2
∂
2
V
∂S
2
= r(V −S
∂V
∂S
).
Using dimensionless variables S = Ee
x
, V = Ev, τ = T −t, we have
v
τ
=
1
2
σ
2
∂
2
v
∂x
2
+ (r −
1
2
σ
2
)
∂v
∂x
−rv.
Let v = e
−rτ
u, then u satisﬁes
u
τ
=
1
2
σ
2
∂
2
u
∂x
2
+ (r −
1
2
σ
2
)
∂u
∂x
. (5.3)
The initial condition for u is
u(x, 0) =
max¦e
x
−1, 0¦ for call option
max¦1 −e
x
, 0¦ for put option
54 CHAPTER 5. NUMERICAL METHODS
The far ﬁeld boundary condition for u is
u(−∞, t) = 0, u(x, t) = e
x
e
rτ
as x →∞
for a call option, and
u(−∞, t) = 1, u(x, t) = 0 as x →∞
for a put option.
5.3.1 Discretization methods
To solve (5.3) numerically, we follow the following procedure:
1. Discretize space and time. We choose a proper ﬁnite domain (x
L
, x
R
), discretize it
into
x
j
= j∆x, j = −N, ..., N, where ∆x =
x
R
−x
L
N
.
Similarly, we discretize [t, T] into N steps, ∆t =
T−t
M
.
We shall approximate u(x
j
, n∆τ) by U
n
j
, V (x
j
, n∆τ) by V
n
j
. From v = e
−rτ
u, we
have
V
n
j
= e
−rn∆τ
U
n
j
.
2. Spatial discretization. We replace the spatial derivatives by ﬁnite differences:
(a) u
x
is replaced by one of the following three:
u
x
←
u
j+1
−u
j−1
2∆x
u
j
−u
j−1
∆x
if
1
2
σ
2
−r > 0
u
j+1
−u
j
∆x
if
1
2
σ
2
−r ≤ 0.
(b) u
xx
←
u
j+1
−2u
j
+u
j−1
(∆x)
2
Then the righthandside of (5.3) is discretized into
(QU)
j
≡ (
σ
2
2
)
U
j+1
−2U
j
+ U
j−1
(∆x)
2
+ (r −
σ
2
2
)
U
j+1
−U
j−1
2∆x
3. Temporal discretization. For the temporal discretization, we introduce the following
three methods:
(a) Forward Euler method:
U
n+1
j
−U
n
j
∆t
= (QU
n
)
j
.
(b) Backward Euler method:
U
n+1
j
−U
n
j
∆t
= (QU
n+1
)
j
.
(c) CrankNicolson method:
U
n+1
j
−U
n
j
∆t
=
1
2
[(QU
n+1
)
j
+ (QU
n
)
j
].
5.3. FINITE DIFFERENCE METHODS (FOR THE MODIFIED BS EQ.) 55
5.3.2 Binomial method is a forward Euler ﬁnite difference method
We choose x = S/S
0
, where S
0
is the current asset price value. Let x
j
= j∆x, j =
−N, ..., N, where ∆x is a small parameter satisfying some stability constraint to be shown
below. Let us partition the time interval [t, T] into N subintervals uniformally, and let
∆t = (T −t)/N.
For the forward Euler method, we rewrite it as
U
n+1
j
= U
n
j
+ ∆t(QU
n
)
j
= aU
n
j−1
+ bU
n
j
+ cU
n
j+1
In terms of V
n
j
, we have
e
r∆t
V
n+1
j
= V
n
j
+
1
2
∆t
(∆x)
2
σ
2
(V
n
j+1
−2V
n
j
+ V
n
j−1
)
+(r −
σ
2
2
)
∆t
2∆x
(V
n
j+1
−V
n
j−1
)
=
¸
1
2
∆t
(∆x)
2
σ
2
+ (r −
σ
2
2
)
∆t
2∆x
V
n
j+1
+ (1 −
∆t
(∆x)
2
σ
2
)V
n
j
+
¸
1
2
∆t
(∆x)
2
σ
2
−(r −
σ
2
2
)
∆t
2∆x
V
n
j−1
.
≡ aV
n
j+1
+bV
n
j
+ cV
n
j−1
We should require a, b, c ≥ 0 for stability reason. This will be discussed later. Notice that
a + b + c = 1. Thus V
n+1
j
is the “average” of V
n
j−1
, V
n
j
, V
n
j+1
with weight a, b, c, then
discounted by e
−r∆t
.
The stability condition a, b, c ≥ 0 reads
r −
σ
2
2
∆t
∆x
≤
∆t
(∆x)
2
σ
2
≤ 1
Next, let us consider a special case: b = 0. We can choose ∆τ and ∆x properly so that
b = 0. i.e., 1 =
∆t
(∆x)
2
σ
2
. In this case,
e
r∆t
V
n+1
j
= pV
n
j+1
+ (1 −p)V
n
j−1
where p = a =
1
2
∆t
(∆x)
2
σ
2
+ (r −
σ
2
2
)
∆t
2∆x
. The stability condition is satisﬁed if and only if
0 ≤ p ≤ 1. (5.4)
We see that this ﬁnite difference is identical to the binomial method in the previous section.
5.3.3 Stability
Deﬁnition 3.3 A ﬁnite difference method is called consistent to the corresponding P.D.E.
if for any solution of the corresponding P.D.E., it satisﬁes
F.D.E (ﬁnite difference equation) + (∆x, ∆t)
and →0 as ∆t, ∆x →0
56 CHAPTER 5. NUMERICAL METHODS
Deﬁnition 3.4 The truncation error of a ﬁnite difference method is deﬁned to be the func
tion (∆x, ∆t) in the previous deﬁnition.
For instance, the truncation error for central difference is
Qu =
σ
2
2
u
xx
+ (r −
σ
2
2
)u
x
+ O((∆x)
2
).
And the truncation for various temporal discretizations are
1. Forward Euler:
u(j∆x, (n + 1)∆t) −u(j∆x, n∆t)
∆t
−(Qu)(j∆x, n∆t) = O((∆x)
2
) + O(∆t).
2. Backward Euler:
u(j∆x, (n + 1)∆t) −u(j∆x, n∆t)
∆t
−(Qu)(j∆x, (n+1)∆t) = O((∆x)
2
)+O(∆t).
3. CrankNicolson method
u(j∆x, (n + 1)∆t) −u(j∆x, n∆t)
∆t
−
1
2
[(Qu)(j∆x, (n + 1)∆t) + (Qu)(j∆x, n∆t)]
= O((∆x)
2
) + O((∆t)
2
).
The true error U
n
j
−u(j∆x, n∆t) is usually estimated in terms of the truncation error.
Deﬁnition 3.5 A ﬁnite difference equation is said to be (L
2
−)stable if the norm
U
n

2
:= Σ
j
[U
n
j
[
2
∆x
is bounded for all n ≥ 0.
Deﬁnition 3.6 A ﬁnite difference method for a P.D.E. is convergent if its solution U
n
j
con
verges to the solution u(j∆x, n∆t) of the corresponding P.D.E..
Theorem 5.3 (Lax) : For linear partial differential equations, a ﬁnite difference method
is convergent if and only if it is consistent and stable.
This theorem is standard and its proof can be found in most numerical analysis text book.
We therefore omit it here.
Since the consistency is easily to achieve, we shall focus on the stability issue. A
standard method to analyze stability issue is the von Neumann stability analysis. It works
for P.D.E. with constant coefﬁcients. It also works “locally” and serves as a necessary
condition for linear P.D.E. with variable coefﬁcients and nonlinear P.D.E.. We describe his
method below.
5.3. FINITE DIFFERENCE METHODS (FOR THE MODIFIED BS EQ.) 57
We take Fourier transform of ¦U
j
¦
∞
j=−∞
by deﬁning
ˆ
U(ξ) =
∞
¸
j=−∞
U
j
e
−ijξ
It is a wellknown fact that
¸
j
[U
j
[
2
=
1
√
2π
π
−π
[
ˆ
U(ξ)[
2
dξ
≡ 
ˆ
U
2
Thus, the boundedness of
¸
j
[U
j
[
2
can be estimated by using 
ˆ
U
2
. The advantage of
using
ˆ
U is that the ﬁnite difference operation becomes a multiplier in terms of
ˆ
U. Namely,
¯
DU(ξ) =
¸
j
U
j+1
−U
j−1
2
e
−ijξ
=
¸
j
U
j
e
i(j+1)ξ
−U
j
e
i(j−1)ξ
2
=
e
iξ
−e
−iξ
2
¸
j
U
j
e
−ijξ
= (2i sin ξ)
ˆ
U(ξ)
For the ﬁnite difference operator QU,m we have
(QU) =
¸
j
(Qu)
j
e
−ijξ
= [
σ
2
2
1
(∆x)
2
(2cosξ −2) + (r −
σ
2
2
)
1
∆x
(2isinξ)]
ˆ
U
≡
ˆ
Q(ξ)
ˆ
U(ξ).
For forward Euler method,
U
n+1
(ξ) = (1 + ∆t
ˆ
Q(ξ))
´
U
n
= G(ξ)
´
U
n
= G(ξ)
n+1
´
U
n
We observe that
π
−pi
[
´
U
n
(ξ)[
2
dξ =
[G(ξ)[
2n
´
U
0
(ξ)[
2
dξ
≤ max
ξ∈(−π,π)
[G(ξ)[
2n
[
´
U
0
(ξ)[
2
dξ
58 CHAPTER 5. NUMERICAL METHODS
If [G(ξ)[ ≤ 1, ∀ξ ∈ (−π, π), then stability condition holds. On the other hand, if [G(ξ)[ >
1 at some point ξ
0
, then by the continuity of G, we have that
[G(ξ)[ ≥ 1 +
for some small > 0 and for all ξ with [ξ −ξ
0
[ ≤ δ for some δ > 0. Let consider an initial
condition such that
´
U
0
(ξ) =
1 [ξ −ξ
0
[ ≤ δ
0 otherwise.
Then the corresponding
´
U
n
will have
π
−pi
[
´
U
n
(ξ)[
2
dξ =
[G(ξ)[
2n
´
U
0
(ξ)[
2
dξ
→ ∞
as n →∞. We conclude the above discussion by the following theorem.
Theorem 5.4 For a ﬁnite difference equation with constant coefﬁcients, suppose its fourier
transform satisﬁes’
U
n+1
(ξ) = G(ξ)
´
U
n
(ξ)
Then the ﬁnite difference equation is stable if and only if
[G(ξ)[ ≤ 1 ∀ξ ∈ (−π, π].
Example: Let apply the forward Euler method for the heat equation: u
t
= u
xx
. Then
U
n+1
j
−U
n
j
∆t
=
1
(∆x)
2
(U
n
j+1
−2U
n
j
+ U
n
j−1
), =⇒U
n+1
= U
n
+
∆t
(∆x)
2
D
2
U
n
From von Neumann analysis:
U
n+1
= [1 +
∆t
(∆x)
2
(2 cos ξ −2)]
´
U
n
= (1 −4
∆t
(∆x)
2
sin
2
ξ
2
)
´
U
n
≡ G(ξ)
´
U
n
.
Hence
[G(ξ)[ ≤ 1 =⇒
∆t
(∆x)
2
sin
2
ξ
2
≤
1
2
=⇒
∆t
(∆x)
2
≤
1
2
(stability condition)
If we rewrite the ﬁnite difference scheme by
U
n+1
j
=
∆t
(∆x)
2
U
n
j+1
+ (1 −2
∆t
(∆x)
2
)U
n
j
+
∆t
(∆x)
2
U
n
j−1
≡ aU
n
j+1
+ bU
n
j
+ cU
n
j−1
.
5.3. FINITE DIFFERENCE METHODS (FOR THE MODIFIED BS EQ.) 59
Then the stability condition is equivalent to
a, b, c ≥ 0.
Since we have a + b + c = 1 from the deﬁnition, thus we see that the ﬁnite difference
scheme is nothing but saying U
n+1
j
is the average of U
n
j+1
, U
n
j
and U
n
j−1
with weights
a, b, c. In particular, if we choose
∆t
(∆x)
2
=
1
2
, then b = 0. If we rename a = p, c = 1 − p,
then U
n+1
j
= pU
n
j+1
+ (1 −p)U
n
j−1
. This can be related to the random walk as the follows.
Consider a particle move randomly on the grid points j∆x. In one time step, the particle
moves toward right with probability p and left with probability 1 − p. Let U
n
j
be the
probability of the particle at j∆x at time step n for a random walk.
U
n+1
j
= pU
n
j−1
+ (1 −p)U
n
j+1
.
We can also apply the above stability analysis to backward Euler method and the Crank
Nicolson method. Let us only demonstrate the analysis for the heat equation. We left the
analysis for the BlackScholes equations as exercises.
1. For backward Euler method,
U
n+1
j
−U
n
j
∆t
=
1
(∆x)
2
(U
n+1
j−1
−2U
n+1
j
+ U
n+1
j+1
).
Then
(1 + (4 sin
2
ξ
2
)
∆t
(∆x)
2
)
U
n+1
=
´
U
n
=⇒
U
n+1
= G(ξ)
´
U
n
,
where
G(ξ) =
1
1 + 4
∆t
(∆x)
2
(sin
2 ξ
2
)
.
We ﬁnd that [G(ξ)[ ≤ 1, for all ξ. Hence, the backward Euler method is always
stable.
2. For CrankNicolson method,
U
n+1
j
−U
n
j
∆t
=
1
2(∆x)
2
(U
n+1
j+1
−2U
n+1
j
+ U
n+1
j−1
) + (U
n
j+1
−2U
n
j
+ U
n
j−1
)
.
Its Fourier transform satisﬁes
U
n+1
−
´
U
n
∆t
=
1
2(∆x)
2
¸
−(4 sin
2
ξ
2
)
U
n+1
−(4 sin
2
ξ
2
)
´
U
n
.
We have
(1 + 2
∆t
(∆x)
2
(sin
2
ξ
2
))
U
n+1
= (1 −2
∆t
(∆x)
2
(sin
2
ξ
2
))
´
U
n
and hence
U
n+1
=
1 −2
∆t
(∆x)
2
sin
2 ξ
2
1 + 2
∆t
(∆x)
2
sin
2 ξ
2
´
U
n
.
Let α = 2
∆t
(∆x)
2
sin
2 ξ
2
, then G(ξ) =
1−α
1+α
. We ﬁnd that for all α ≥ 0, [G(ξ)[ ≤ 1,
hence CrankNicolson method is always stable for all ∆t, ∆x > 0.
60 CHAPTER 5. NUMERICAL METHODS
Exercise study the stability criterion for the modiﬁed BlackScholes equation
u
τ
=
σ
2
2
u
xx
+ (r −
σ
2
2
)u
x
,
for the forward Euler method, back Euler method and CrankNicolson method.
5.3.4 Convergence
Let us study the convergence for ﬁnite difference schemes for the modiﬁed BlackScholes
equation. Let us take the forward Euler scheme as our example. The method below can
also be applied to other scheme.
The forward Euler scheme is given by:
U
n+1
j
−U
n
j
∆t
= (QU
n
)
j
We have known that it has ﬁrstorder truncation error, namely, suppose u
n
j
:= u(j∆x, n∆t),
where u is the solution of the modiﬁed BlackScholes equation, then
u
n+1
j
−u
n
j
∆t
= (Qu
n
)
j
+ O(∆t) + O((∆x)
2
).
We subtract the above two equations, and let e
n
j
denotes for u
n
j
−U
n
j
and
n
j
denotes for the
truncation error. Then we obtain
e
n+1
j
−e
n
j
∆t
= (Qe
n
)
j
+
n
j
Or equivalently,
e
n+1
j
= ae
n
j+1
+ be
n
j
+ ce
n
j−1
+ ∆t
n
j
(5.5)
Here, a, b, c ≥ 0 and a+b +c = 1. We can take Fourier transformation
´
e
n
of e
n
. It satisﬁes
¯
e
n+1
(ξ) = G(ξ)
´
e
n
(ξ) + ∆t
´
n
(ξ)
where
G(ξ) = ae
iξ
+ b + ce
−iξ
.
Recall that the stability [G(ξ)[ ≤ 1 is equivalent to a, b, c ≥ 0. Thus, by applying the above
recursive formula, we obtain

´
e
n
 ≤ 
¯
e
n−1
 + ∆t
¯
n−1

≤ 
¯
e
n−2
 + ∆t
G
¯
n−2
 +
¯
n−1

≤ 
¯
e
n−2
 + ∆t

¯
n−2
 +
¯
n−1

≤ 
´
e
0
 + ∆t
n−1
¸
k=0

´
k

≤ O(∆t) + O((∆x)
2
).
5.3. FINITE DIFFERENCE METHODS (FOR THE MODIFIED BS EQ.) 61
Here, we have used the estimate for the truncation error

n
 = O(∆t) + O((∆x)
2
,
and that n∆t = O(1). We conclude the error analysis as the following theorem.
Theorem 5.5 The error e
n
j
:= u(j∆x, n∆t) −U
n
j
for the Euler method has the following
convergence rate estimate:
(
¸
j
[e
n
j
[
2
∆x)
1/2
≤ O(∆t) + O((∆x)
2
), for all n.
It is simpler to ontain the maximum norm estimate. Let E(n) := max
j
[e
n
j
[ be the maxi
mum error. From (5.5), we have
[e
n+1
j
[ ≤ a[e
n
j+1
[ + b[e
n
j
[ +c[e
n
j−1
[ + ∆t[
n
j
[
≤ aE(n) + bE(n) + cE(n) + ∆t
= E(n) + ∆t
where
:= max
j,n
[
n
j
[ = O(∆t) + O((∆x)
2
).
Hence,
E(n + 1) ≤ E(n) + ∆t.
Since we take U
0
j
= u
0
j
, there is no error initially. Hence, we have
E(n) ≤
n−1
¸
k=0
∆t
≤ n∆t
Since n∆t is a ﬁxed number, as we take the limit n → ∞, we obtain the error is bounded
by the truncation error. We summarize the above discussion as the following theorem.
Theorem 5.6 The error e
n
j
:= u(j∆x, n∆t) −U
n
j
for the Euler method has the following
convergence rate estimate:
max
j
[e
n
j
[ ≤ O(∆t) + O((∆x)
2
).
Exercise. Prove that the true error of the CrankNicolson scheme is O((∆t)
2
)+O((∆x)
2
).
5.3.5 Boundary condition
For the modiﬁed BlackScholes equation, we have
u(−∞, t) = 0, u(x, t) = e
x
e
rτ
as x →∞
62 CHAPTER 5. NUMERICAL METHODS
for a call option, and
u(−∞, t) = 1, u(x, t) = 0 as x →∞
In computation, we can choose a ﬁnite domain (x
L
, x
R
) with x
L
<< −1 and x
R
>> 1.
The boundary condition at the boundary points are an approximation to the above far ﬁeld
boundary condition.
In practice, we don’t even use this boundary condition. Indeed, if we want to know
u(x
k
, n∆t), we can ﬁnd the numerical domain of this quantity, which is the triangle
¦(j∆x, m∆t) [ [j −k[ ≤ n −m¦
We only need to compute u in this domain, which needs no boundary data.
5.4 Converting the BS equation to ﬁnite domain
The transformation x = log(S/E) converts the BS equation to a heat equation. However,
the domain of x is the whole real line. For numerical computation, it is desirable to have
a ﬁnite computation domain. The transformation in this section converts S to ξ with ξ ∈
(0, 1). The price is that the resulting equation has variable coefﬁcients. But this is not a
problem for numerical computation.
We deﬁne the transformation:
ξ =
S
S + E
(5.6)
V =
V (S, t)
S + E
(5.7)
τ = T −t. (5.8)
Notice that ξ is dimensionless and important values of ξ are near 1/2. With this, the inverse
transformation is
S =
Eξ
1 −ξ
,
dξ
dS
=
(1 −ξ)
2
E
We plug this transformation to the BS equation. We allow σ depend on S. Deﬁne ¯ σ(ξ) =
σ(Eξ/1 −ξ). Then the resulting equation is
∂V
∂τ
=
1
2
¯ σ
2
(ξ)ξ
2
(1 −ξ)
2
∂
2
V
∂ξ
2
+rξ(1 −ξ)
∂V
∂ξ
−r(1 −ξ)V , (5.9)
for 0 ≤ ξ ≤ 1 and τ > 0. The initial data reads
V (ξ, 0) =
1 −ξ
E
Λ
Eξ
1 −ξ
. (5.10)
For a call option, the payoff is Λ(S) = max(S −E, 0). The corresponding
V (ξ, 0) = max(S −E, 0)(1 −ξ)/E
= max(2ξ −1, 0).
5.5. FAST ALGORITHMS FOR SOLVING LINEAR SYSTEMS 63
Similarly, V (ξ, 0) = max(1 −2ξ, 0) for a put option.
On the boundaries ξ = 0 and ξ = 1, the diffusion coefﬁcients are degenerate. If the
solution is smooth up to the boundaries, then on the boundary, the equation is degenerate
to the following ordinary differential equations:
∂V (0, τ)
∂τ
= −rV (0, τ)
∂V (1, τ)
∂τ
= 0.
The corresponding solutions are
V (0, τ) = V (0, 0)e
rτ
(5.11)
V (1, τ) = V (1, 0). (5.12)
We can discretize equation (5.9) by ﬁnite difference method. Let ∆ξ and ∆τ are the
spatial and temporal mesh sizes, respectively. Let ξ
j
= j∆ξ, τ
n
= n∆τ. The boundaries
points are ξ
0
and ξ
M
. We use central difference for ∂
2
V /∂ξ
2
and ∂V /∂ξ. The resulting
ﬁnite difference equation reads
dv
j
dτ
=
1
2
σ
2
j
ξ
2
j
(1 −ξ
j
)
2
v
j+1
−2v
j
+ v
j−1
∆ξ
2
+rξ
j
(1 −ξ
j
)
v
j+1
−v
j−1
2∆ξ
−r(1 −ξ
j
)v
j
We can discretize this equation in the time direction by forward Euler method. The stability
constraint is
rξ
j
(1 −ξ
j
) −
1
2
σ
2
j
ξ
2
j
(1 −ξ
j
)
2
∆τ
∆ξ
≤ σ
2
j
ξ
2
j
(1 −ξ
j
)
2
∆τ
2∆ξ
2
≤ 1. (5.13)
Remark. Many options have nonsmooth payoff functions. This causes low order ac
curacy for ﬁnite difference scheme. Fortunately, many simple payoff function has exact
solution. For instance, the European call option. For general payoff function, we may sub
tract its nonsmooth part for which an exact solution is available. The remainder is smooth,
and a ﬁnite difference scheme can yield highorder accuracy.
5.5 Fast algorithms for solving linear systems
In the backward Euler method and the CrankNicolson method, we need to solve linear
systems of the form
AU = F.
For the backward Euler scheme,
A = diag (−a, 1 + a +c, −c)
64 CHAPTER 5. NUMERICAL METHODS
:=
¸
¸
¸
¸
¸
¸
¸
¸
¸
1 + a +c −c 0
−a 1 + a + c −c 0
0 −a 1 + a + c −c 0
.
.
.
.
.
.
.
.
.
.
.
.
0 −a 1 + a + c −c 0
0 −a 1 + a + c −c
0 −a 1 + a + c
and
U =
¸
U
n
j
L
+1
.
.
.
U
n
j
R
−1
, F =
¸
aU
n+1
j
L
.
.
.
cU
n+1
j
R
.
For the CrankNicolson scheme We have
AU
n+1
= BU
n
+ b
n+1/2
where A = diag (−
a
2
, 1 +
a
2
+
c
2
, −
c
2
) B = diag (
a
2
, 1 −
a
2
−
c
2
,
c
2
), and
b
n+1/2
=
¸
¸
¸
¸
¸
¸
a
U
n+1
j
L
+U
n
j
L
2
0
.
.
.
0
c
U
n+1
j
R
+U
n
j
R
2
.
Now, we concentrate on solving the linear system
Ax = f.
The matrix A is tridiagonal and diagonally dominant. Let us rewrite A = diag (a, b, c).
Here, the constants a, b, c are different from the average weights we had before. We may
assume b > 0. We say that A is diagonally dominant if b > [a[ + [c[. More generally, A
may takes the form A = diag (a
j
, b
j
, c
j
). and [b
j
[ > [a[ + [c
j
[. Without loss of generality,
we may normalize the j −th so that b
j
= 1.
There are two classes of methods to solve the above linear systems. One is called direct
methods, the other is called iterative methods. For onedimensional case as we have here,
direct method is usually better. However, for highdimensional cases, iterative methods are
better.
5.5.1 Direct methods
Gaussian elimination
Let us illustrate this method by the simple example: A = diag (a, 1, c). We multiple the
ﬁrst equation by −a and add it into the second equation to eliminate the term x
j
L
+1
in the
5.5. FAST ALGORITHMS FOR SOLVING LINEAR SYSTEMS 65
second equation. Then the resulting equation becomes
¸
¸
¸
¸
¸
1 c 0 0 0
0 1 −ac c 0 0
0 a 1 c 0
.
.
. 0
1
¸
¸
¸
¸
¸
x
j
L
+1
x
j
L
+2
x
j
L
+3
.
.
.
x
j
R
−1
=
¸
¸
¸
¸
¸
b
j
L
+1
−ab
j
L
+1
+ b
j
L
+2
b
j
L
+3
.
.
.
b
j
R
−1
We continue to eliminate the term a in the third equation, and so on. Finally, we arrive
¸
¸
¸
¸
¸
1 c 0 0 0
0 1 −ac c 0 0
0 0 1 −c/(1 −ac) c 0
0 0 0
.
.
. 0
0 0 0 0 1
¸
¸
¸
¸
¸
x
j
L
+1
x
j
L
+2
x
j
L
+3
.
.
.
x
j
R
−1
=
¸
¸
¸
¸
¸
b
j
L
+1
b
j
L
+2
b
j
L
+3
.
.
.
b
j
R
−1
Then x
j
can be solved easily. The diagonal dominance condition guarantee that the reduced
matrix is also diagonally dominant. Thus, this scheme is numerical stable.
LU decomposition
We decompose A = LU, where
L =
¸
¸
¸
¸
¸
¸
1 0 0 0
j
L
+2
1
.
.
.
.
.
.
0
.
.
.
.
.
.
.
.
.
0
.
.
.
.
.
.
.
.
.
0
0 0
j
R
−1
1
, U =
¸
¸
¸
¸
¸
¸
u
j
L
+1
v
j
L
+1
0 0
0 u
j
L
+2
.
.
.
.
.
.
0
.
.
.
.
.
.
.
.
.
0
.
.
.
.
.
.
.
.
.
v
j
R
−2
0 0 0 u
j
R
−1
It is easy to ﬁnd a recursion formula to ﬁnd the coefﬁcients , u and v’s. Once these are
found, we can ﬁnd x by solving
Ly = b, Ux = y.
These two equations are easy to solve. One can show that both L and U are diagonally
dominant if A is.
If we watch carefully, LUdecomposition is equivalent to the Gaussian elimination.
Cyclic reduction method
Let us take the case A = diag (a, 1, c) to illustrate this method. Consider three consecutive
equations
ax
2j−2
+ x
2j−1
+ cx
2j
= b
2j−1
ax
2j−1
+ x
2j
+cx
2j+1
= b
2j
ax
2j
+ x
2j+1
+cx
2j+2
= b
2j+1
66 CHAPTER 5. NUMERICAL METHODS
We can eliminate the oddindex terms x
2j−1
and x
2j+1
. Namely, −a (2j −1)−eq +(2j)
eq −c (2j + 1)eq: After normalization, we obtain
a
x
2j−2
+x
2j
+ c
x
2j+2
= b
j
Here,
a
= −
a
2
1 −2ac
, c
= −
c
2
1 −2ac
,
b
j
= (b
2j
−ab
2j−1
−cb
2j+1
)/(1 −2ac).
If we rename x
j
= x
2j
. Then we have A
x
= b
, where A
= diag (a
, 1, c
). Notice
that the system is reduced to half and with the same form. One can show that the iterative
mapping
a
c
→
a
c
converges to (0, 0)
t
quadratically fast, provided [a[ + [c[ < 1 initially. Thus, for few
iteration, the matrix A is almost an identity matrix. We can invert it trivially. Once x
2j
are
found, the oddindex x + 2j + 1 can be found from the equation:
ax
2j
+ x
2j+1
+ cx
2j+2
= b
2j+1
.
A careful reader should ﬁnd that the cyclic reduction is also a version of the Gaussian
elimination method.
5.5.2 Iterative methods
Most iterative methods can be viewed as a proper decomposition of A, then solve an im
portant and treat the rest as a perturbation term.
Jacob method
In Jacobi method, wer decompose
A = D + B
where D is the diagonal part and B is the off diagonal part. Since Ais diagonally dominant,
we may approximate x by the sequence x
n
, where x
n
is deﬁned by the following iteration
scheme:
Dx
n+1
+ Bx
n
= b.
Let the error e
n
:= x
n+1
−x
n
. Then
De
n
= −Be
n−1
Or
e
n
= −D
−1
Be
n−1
.
5.5. FAST ALGORITHMS FOR SOLVING LINEAR SYSTEMS 67
Let us deﬁne the maximum norm
e
n
 := max
j
[e
n
j
[
Then
[e
n
j
[ = [ −
a
b
e
n−1
j−1
−
c
b
e
n−1
j+1
[
≤ [
[a[
[b[
e
n−1
 +
[c[
[b[
e
n−1

=
[a[ +[c[
[b[
e
n−1

Hence,
e
n
 ≤ ρe
n−1

where ρ =
a+c
b
< 1 from the fact that A is diagonally dominant. This yields the conver
gence of the sequence x
n
. The limit x satisﬁes the equation Ax = b.
GaussSeidel method
In GaussSeidel method, A is decomposed into A = (D+L) +U, where D is the diagonal
part, L, the lower triangular part, and U, the upper triangular part of A. The approximate
solution sequence is given by
(D +L)x
n+1
+Ux
n
= b.
As before, the error e
n
:= x
n+1
−x
n
satisﬁes
e
n
= −(D + L)
−1
Ue
n−1
To analyze the decay of e
n
, we use Fourier method. Let
´
e
n
(ξ) :=
¸
j
e
n
j
e
−ijξ
.
Then we have
´
e
n
(ξ) = G(ξ)
¯
e
n−1
,
G(ξ) = −
ce
iξ
b + ae
−iξ
It is easy to see that the ampliﬁcation matrix G satisﬁes
max
ξ
[G(ξ)[ := ρ < 1, provided [b[ > [a[ +[c[. (5.14)
This shows that the GaussSeidel method also converges for diagonally dominant matrix.
Exercise. Show the above statement (5.14).
68 CHAPTER 5. NUMERICAL METHODS
Successive overrelaxation method (SOR)
In the methods of Jacobi and GaussSeidel, the approximate sequence x
n
is usually con
vergent monotonely. We therefore have a chance to speed them up by an extrapolation
procedure described below.
y
n+1
= −D
−1
(L +U)x
n
+ b
x
n+1
= x
n
+ ω(y
n+1
−x
n
).
Here, ω is a parameter. In order to speed up, we require ω > 1. We also need to require
ω < 2 for stability. The optimal ω is chosen to minimize the ampliﬁcation matrix G
ω
(ξ).
Exercise. Find the ampliﬁcation matrix G
ω
and the optimal ω for the matrix A = diag (a, b, c).
Also, determine the rate
ρ := min
ω
max
ξ
[G
ω
(ξ)[.
Multigrid method
Probably the most powerful method in higher dimension is the multigrid method.
Chapter 6
American Option
6.1 Introduction
An American option has the right to exercise any time during the life of the option. The
ﬁrst important thing we should note is that the value of an American option is greater
than or equal to the payoff function: V (S, t) ≥ Λ(S, t). Otherwise, there is an arbitrage
opportunity because we can buy the American option then sell it immediately to gain a net
proﬁt V −Λ.
We recall that the value of an American call option is equal to that of a European call
option. However, for other cases like the the American put option or the American call
option on dividendpaying asset, the American options do cost more. We explain why it is
so below. The ﬁgure below is the value of a European put.
S
P
x
E
Ee
−r(T−t)
x
x
E
69
70 CHAPTER 6. AMERICAN OPTION
Notice that P(S, t) < max¦E − S, 0¦ in some region in the St plane. In this region,
the corresponding American option must be higher than the European option, otherwise
for S, we can buy a put P(S, t), then exercise it immediately. We make a riskless proﬁt:
E −P −S > 0. Another example is the American call option on a dividendpaying asset.
Its value is shown in the Figure below.
S
P
x
E
C(S, t) ∼ Se
−D
0
(T−t)
for S >> 1, hence C(S, t) < max(S −E, 0) for some S
f
(t).
Since C(S, t) ∼ Se
−D
0
(T−t)
for large S, there is a region in (S, t)plane where C(S, t) <
max¦S −E, 0¦. In this region, if we could exercise the call option, then based on the same
argument above, there would be an arbitrage opportunity. Hence the corresponding Amer
ican call option should also satisfy
C(S, t) ≥ max¦S −E, 0¦.
6.2 American options as a free boundary value problem
6.2.1 American put option
We can view an American option as a free boundary value problem. Let us take the Amer
ican put option as an example.
First, there must be some value of S for which it is optimal from the holder’s point
of view to exercise the American option. Otherwise, we should hold the option for all
possible S. Then this option is identical to a European option. But we have seen that this
is not the case. In other words, there is a S
f
(t), if S < S
f
(t), one should exercise the put
option, which maximize the payoff function E−S. And for S > S
f
(t), we should hold the
6.2. AMERICAN OPTIONS AS A FREE BOUNDARY VALUE PROBLEM 71
option. This S
f
(t) is referred as the optimal exercise price. In other word, we should have
P(S, t) ≡ max(E − S, 0) for S < S
f
(t), and P(S, t) satisﬁes the BlackSholes equation
for S > S
f
(t).
However, we do not know S
f
(t) a priori. We should treat S
f
(t) as a new unknown
(called free boundary and we should impose boundary condition to determine it.) We claim
that the proper boundary condition on S
f
(t) are
1. P(S, t) is continuous across (S
f
(t), t),
2. ∂P(S, t)/∂S is also continuous across (S
f
(t), t).
Remark. For S < S
f
(t), we should exercise the American put option because the corre
sponding payoff Λ = max¦E − S, 0¦ is higher. Thus, for S < S
f
(t), the value of the put
option should be the payoff function Λ = E−S. Its derivative in S is −1. Thus, the second
boundary condition is equivalent to saying that
∂P
∂S
(S
f
(t), t) is continuous across (S
f
(t), t).
Reasons.
1. If S(t) = S
f
(t) then S(t + ∆t) > S
f
(t) with probability 1. This follows from
dS
S
=
µdt +σdz and µ > 0. If P is discontinuous across (S
f
(t), t), then P(S(t + ∆t), t +
∆t) = P(S(t), t) with probability 1. This would make an arbitrage opportunity by
buying P at t then selling it at t + ∆t.
2. We prove this by contradiction.
(a) If
∂P
∂S
(S
f
(t), t) < −1 then as S increases from S
f
(t), P(S, t) drops below
the payoff E − S, this contradicts to P(S, t) ≥ max¦E − S, 0¦. At S
f
(t),
P(S
f
(t), t) = E −S
f
(t).
(b) Suppose
∂P
∂S
(S
f
(t), t) > −1. First, for S > S
f
(t), P satisﬁes the BlackScholes
equation (for put option) and its solution curve should lie above the payoff
function E −S on the (S, P)plane with P(S
f
(t), t) = E −S
f
(t). This curve
moves up as S
f
(t) moves down, and the corresponding
∂P
∂S
(S
f
(t), t) decreases.
If
∂P
∂S
(S
f
(t), t) > −1, then we can move down S
f
(t) to another
˜
S
f
(t) < S
f
(t)
where
∂P
∂S
(
˜
S
f
(t), t) = −1. In this movement, the curve stays above the payoff
function. Now, if we exercise the put option for S ≤
˜
S
f
(t), the payoff E−
˜
S
f
(t)
is higher than E−S
f
(t). This means that S
f
(t) is not the optimal exercise price.
This is a contradiction.
Thus, we treat the American put option as the following free boundary value problem.
There exists an optimal exercise price S
f
(t) such that
1. for S < S
f
(t), early exercise is optimal, and P(S, t) = E −S;
2. for S > S
f
(t), one should hold the put option and P satisﬁes the BlackScholes
equation:
∂P
∂t
+
1
2
σ
2
S
2
∂
2
P
∂S
2
+rS
∂P
∂S
−rP = 0;
3. across the free boundary (S
f
(t), t), both P and
∂P
∂S
are continuous.
72 CHAPTER 6. AMERICAN OPTION
6.2.2 American call option on a dividendpaying asset
As we have seen in the introduction of this chapter that an American call option C(S, t) on
a dividendpaying asset has asymptotic value C(S, t) ∼ Se
−D
0
(T−t)
for large S. This value
is below the payoff function Λ ≡ max(S − E, 0). Therefore, there must an optimal S
f
(t)
such that we should exercise this call option when S > S
f
(t) and hold it when S < S
f
(t).
On the free boundary S = S
f
(t), based on the noarbitragy hypothesis, we should have
both C(S, t) and ∂C(S, t)/∂S are continuous the free boundary (S(t), t) for 0 < t < T.
We summarize this by the following equations
C
t
+
σ
2
2
S
2
∂
2
C
∂S
2
+ (r −D
0
)S
∂C
∂S
−rC = 0, 0 < S < S
f
(t)
C(S, t) ≡ Λ(S) ≡ max¦S −E, 0¦, S > S
f
(t).
On the free boundary S = S
f
(t), the boundary condition is required
C(S
f
(t), t) = S
f
(t) −E,
∂C
∂S
(S
f
(t), t) = 1, 0 ≤ t ≤ T.
6.3 American option as a linear complementary problem
The American option can also be formulated as a linear complementary problem, where
the free boundary is treated implicitly. To illustrate this linear complementary problem,
ﬁrst we notice that an American option should satisfy the following conditions:
(i) V ≥ Λ,
(ii) V
t
+
1
2
σ
2
S
2 ∂
2
V
∂S
2
≤ r(V −S
∂V
∂S
),
(iii) either V = Λ, or V
t
+
1
2
σ
2
S
2 ∂
2
V
∂S
2
= r(V −S
∂V
∂S
) should hold,
(iv) both V and
∂V
∂S
are continuous.
Here, V is the value of the American option, Λ is the corresponding payoff function.
We have seen the reasons for (i), and (iv). We explain the reasons of (ii) below. Let us
consider the portfolio, Π = V − ∆S. As we have seen that the Delta hedge eliminate the
randomness of Π and yields
dΠ = V
t
+
1
2
σ
2
S
2
∂
2
V
∂S
2
.
When it is optimal to hold the option, then
dΠ = r(V −
∂V
∂S
S).
Otherwise, we should have
dΠ ≤ r(V −
∂V
∂S
S),
6.3. AMERICAN OPTION AS A LINEAR COMPLEMENTARY PROBLEM 73
based on no arbitrage opportunities. Thus, the BlackScholes is replaced by the Black
Scholes inequality.
To show (iii), we know that if we exercise the option, then V = Λ, otherwise, we hold
the option and its value should satisfy the BlackScholes equation.
Properties (i)(iv) can be formulated as the following linear complementary problem:
(i) V −Λ ≥ 0,
(ii) V
t
+
1
2
σ
2
S
2 ∂
2
V
∂S
2
−r(V −S
∂V
∂S
) ≤ 0,
(iii) (V −Λ)
V
t
+
1
2
σ
2
S
2 ∂
2
V
∂S
2
−r(V −S
∂V
∂S
)
= 0,
(iv) both V and
∂V
∂S
are continuous.
Such a problem is called a linear complementary problem. The advantage of this formula
tion is that the free boundary is treated implicitly.
We can reformulate this problem in terms of x variable. As before, we use the following
change of variables: V = Ev, S = Ee
x
, τ = T −t. The free boundary now in xvariable
is x
f
(t). The free boundary value problem is formulated as
(i) for −∞< x < x
f
(t), v = 1 −e
x
, and
v
τ
−
σ
2
2
v
xx
−(r −
σ
2
2
)v
x
+ rv ≥ 0.
(ii) for x
f
(t) < x < ∞, v > 1 −e
x
, and
v
τ
−
σ
2
2
v
xx
−(r −
σ
2
2
)v
x
+ rv = 0,
(iii) both v and
∂v
∂x
are continuous.
The linear complementary problem is formulated as:
(i) v −(1 −e
x
) ≥ 0,
(ii) v
τ
−
σ
2
2
v
xx
−(r −
σ
2
2
)v
x
+rv ≥ 0,
(iii) (v
τ
−
σ
2
2
v
xx
−(r −
σ
2
2
)v
x
+ rv)(v −(1 −e
x
)) = 0,
(iv) v and v
x
are continuous.
with initial condition v(x, 0) = Λ(x) = 1 −e
x
.
We may replace v by ue
−rτ
to eliminate the term rv. Then we have
u
τ
−
σ
2
2
u
xx
−(r −
σ
2
2
)u
x
(u −g) = 0
u
τ
−
σ
2
2
u
xx
−(r −
σ
2
2
)u
x
≥ 0
74 CHAPTER 6. AMERICAN OPTION
u −g ≥ 0,
u(x, 0) = g(x, 0),
with u, u
x
being continuous. Here g(x, τ) = max¦e
rτ
(1 −e
x
), 0¦. The far ﬁeld boundary
conditions are
u(x, τ) →0, as x →∞, u(x, τ) →e
rτ
, as x →−∞.
A mathematical theory called parabolic variational inequality gives construction, existence,
uniqueness of the solution. (see reference: A. Friedman,Variational Inequality). Let us
demonstrate this theory brieﬂy. The method we shall use is called the penalty method. Let
us consider the following penalty function:
φ
N
(v) = −e
−Nv
.
It has the properties: (i) φ
N
> 0, (ii) φ
N
(v) → 0 whenever v > 0.. We consider the
following penalized P.D.E.:
u
τ
−
1
2
σ
2
u
xx
−(r −
σ
2
2
)u
x
+ φ
N
(u −g) = 0
u(x, 0) = g(x, 0),
with u →0 as x →+∞, u →e
rτ
as x →−∞.
From a standard theory of nonlinear P.D.E. (by monotone method, for instance), one
can show that the solution u
N
exists for all N > 0. We then need an estimate for u
N
and
∂u
N
∂x
. The boundedness of these two gives that u
N
has a convergent subsequence, say u
N
i
such that
u
N
i
→u,
with u
N
i
, u,
∂
∂x
u
N
i
, u
x
being continuous. Moreover,
[φ
N
i
(u
N
i
−g)[ ≤ constant
As N
i
→∞, we conclude u −g ≥ 0. Further, on the set ¦u −g > 0¦, φ
N
i
(u
N
i
−g) →0.
Hence we have
u
τ
−
1
2
u
xx
−(r −
σ
2
2
)u
x
= 0 on ¦u −g > 0¦.
6.4 Numerical Methods
6.4.1 Projective method for American put
We recall that the solution of the BlackScholes equation can be discretized by the follow
ing binomial method (or the forward Euler method):
e
r∆
V
n
j
= pV
n+1
j+1
+qV
n+1
j−1
, p, q ≥ 0, p +q = 1.
Similarly, the linear complementary problem can be discretized as
e
r∆t
V
n
j
−pV
n+1
j+1
−qV
n+1
j−1
V
n
j
−h
n
j
= 0,
6.4. NUMERICAL METHODS 75
e
r∆t
V
n
j
−pV
n+1
j+1
−qV
n+1
j−1
≥ 0, V
n
j
−Λ
n
j
≥ 0,
V
N
j
= Λ
N
j
.
Here, Λ
n
j
is the discretized payoff function after changing variable.
This discretized linear complementary problem can be solved by the following pro
jected forward Euler method. Deﬁne
V
n
j
= max¦e
−r∆t
(pV
n+1
j+1
+ qV
n+1
j−1
), Λ
n
j
¦.
One can showthat this method converges. (The main tool to prove this is a theory for mono
tone operator. One can show that the scheme is monotone, V 
∞
, V
x

∞
are bounded.
Reference. Majda & Crandell, Math. Comp..) Furthermore, from the construction, we have
V
n
j
≥ h
n
j
. Thus the limiting function satisﬁes V ≥ Λ. At those points V (S, t) > Λ(S, t),
we have V
n
j
> Λ
n
j
, for large n, where n∆t ∼ t and Ee
j∆x
∼ S. In this case, we always
have
V
n
j
= e
−r∆t
(pV
n+1
j+1
+ qV
n+1
j−1
).
Hence, the limiting function satisﬁes the BlackScholes equation whenever V (S, t) >
Λ(S, t). The regularity result (i.e. continuity of V and V
S
) follows from the theory of
monotone operator.
6.4.2 Projective method for American call
The linear complementary problem for this American call option is
(i) V (S, t) ≥ Λ(S) ≡ max¦S −E, 0¦
(ii) V
t
+
σ
2
2
S
2 ∂
2
V
∂S
2
+ (r −D
0
)S
∂V
∂S
−rV ≤ 0,
(iii)
V
t
+
σ
2
2
S
2 ∂
2
V
∂S
2
+ (r −D
0
)S
∂V
∂S
−rV
(V −Λ) = 0.
(iv) V and V
S
are continuous.
The binomial approximation for the BS equation is
e
r∆t
V
n
j
= pV
n+1
j+1
+qV
n+1
j−1
,
where
p =
σ
2
2
∆t
(∆x)
2
+ (r −D
0
−
σ
2
2
)
∆t
2∆x
, q =
σ
2
2
∆t
(∆x)
2
−(r −D
0
−
σ
2
2
)
∆t
2∆x
,
and p + q = 1. We choose ∆t and ∆x so that p > 0 and q > 0. For American option, V
has to be greater than Λ(S, t), the payoff function at time t. Hence, we should require
V
n
j
= max¦e
−r∆t
(pV
n+1
j+1
+ qV
n+1
j−1
), Λ
n
j
¦.
The above is the projective forward Euler method.
76 CHAPTER 6. AMERICAN OPTION
For the corresponding binomial model, ﬁrst we determine the up/down ratios u and d
for the riskless asset price by
pu + (1 −p)d = e
(r−D
0
)∆t
, pu
2
+ (1 −p)d
2
= e
(2(r−D
0
)+σ
2
)∆t
,
or equivalently,
u = A +
√
A
2
−1, d = 1/u,
A =
1
2
(e
−(r−D
0
)∆t
+ e
(r−D
0
+σ
2
)∆t
),
p =
e
(r−D
0
)∆t
−d
u −d
, q = 1 −p.
Then the binomial model is given by
S
n
j
=
uS
n−1
j−1
, with probability p
dS
n−1
j+1
, with probability q,
S
0
0
= S.
and
V
n
j
= max¦e
−r∆t
(pV
n+1
j+1
+ qV
n+1
j−1
), S
n
j
−E¦.
6.4.3 Implicit method
For implicit method like backward Euler or CrankNicolson method, we need to add the
constraints u
n+1
≥ g
n+1
for American option. It is important to know that if an iterative
method is used, then we should require this condition hold in each iteration steps. For
instance, in the SOR iteration method,
V
n,(k+1)
= max¦V
n,(k)
+ ω(y
n,(k+1)
−V
n,(k)
), Λ
n
¦,
y
n,(k+1)
= (D + L)
−1
(−UV
n,(k)
+f
n+1
), k = 0, K
V
n
≡ V
n,(K)
This guarantees that V
n+1
≥ Λ
n+1
.
6.5 Converting American option to a ﬁxed domain prob
lem
6.5.1 American call option with dividend paying asset
We consider the American call option on a dividend paying asset:
V
t
+
σ
2
2
S
2
∂
2
V
∂S
2
+ (r −D
0
)S
∂V
∂S
−rV = 0,
V (S, t) ≥ Λ(S) ≡ max¦S −E, 0¦,
6.5. CONVERTING AMERICAN OPTION TO A FIXED DOMAIN PROBLEM 77
0 ≤ S ≤ S
f
(t), 0 ≤ t ≤ T.
where S
f
(t) is the free boundary. On this free boundary, the boundary condition is required
V (S
f
(t), t) = S
f
(t) −E,
∂V
∂S
(S
f
(t), t) = 1, 0 ≤ t ≤ T.
We also need a condition for S
f
at ﬁnal time
S
f
(T) = max(E, rE/D
0
)
Before converting the problem, we ﬁrst remove the singularity of the ﬁnal data (i.e. non
smoothness of the payoff function) as the follows. We may substract V by an European
call option c with the same payoff data. Notice that c(S, t) has exact solution. The new
variable V −c satisﬁes the same equation, yet it has smooth ﬁnal data.
To convert the free boundary problem to a ﬁxed domain problem, we introduce the
following changeofvariables:
ξ = S/S
f
(t)
τ = T −t
u(ξ, τ) = (V (S, t) −c(S, t))/E
s
f
(τ) = S
f
(t)/E
The new equations for these new variables are
∂u
∂τ
=
σ
2
2
ξ
2 ∂
2
u
∂ξ
2
+
(r −D
0
) +
1
s
f
ds
f
dτ
ξ
∂u
∂ξ
−ru, 0 ≤ ξ ≤ 1,
0 ≤ τ ≤ T,
u(ξ, 0) = 0, 0 ≤ ξ ≤ 1,
u(1, τ) = g(s
f
(τ), τ), 0 ≤ τ ≤ T
∂u
∂ξ
(1, τ) = h(s
f
(τ), τ), 0 ≤ τ ≤ T
s
f
(0) = max(1, r/D
0
),
(6.1)
where
g(s
f
(τ), τ) = s
f
(τ) −1 −c(Es
f
(τ), T −τ),
h(s
f
(τ), τ) = s
f
(τ)
¸
1 −
∂c
∂S
(Es
f
(τ), T −τ)
.
At the boundary ξ = 0, the BlackSholes equation is degenerate to
∂u
∂τ
= −ru.
With the trivial initial condition yields
u(0, τ) = 0, 0 ≤ τ ≤ T.
78 CHAPTER 6. AMERICAN OPTION
In practice, we can solve the modiﬁed BlackSholes equation (6.1) with the boundary con
ditions
u(0, τ) = 0 0 ≤ τ ≤ T
∂u
∂ξ
(1, τ) = h(s
f
(τ), τ), 0 ≤ τ ≤ T
(6.2)
We can differentiate the other boundary condition in τ and yield an ODE for the free bound
ary:
∂u
∂τ
(1, τ) =
∂g
∂ξ
ds
f
dτ
.
with s
f
(0) = max(1, r/D
0
).
6.5.2 American put option
For American put option P, the BS equation is on the inﬁnite domain S > S
f
(t), 0 ≤ t ≤
T. Through the changeofvariable
η =
E
2
S
u(η, t) =
EP(S,t)
S
the inﬁnity domain problem is converted to a ﬁnite domain problem:
∂u
∂t
+
1
2
σ
2
η
2 ∂
2
u
∂η
2
−rη
∂u
∂η
= 0, 0 ≤ η ≤ η
f
(t),
0 ≤ t ≤ T
u(η, T) = max(η −E, 0), 0 ≤ η ≤ η
f
(T),
u(η
f
(t), t) = η
f
(t) −E, 0 ≤ t ≤ T,
∂u
∂η
(η
f
(t), t) = 1, 0 ≤ t ≤ T,
η
f
(T) = max(E, 0)
We can further convert it to a ﬁxed domain problem as that in the last section.
Chapter 7
Exotic Options
Option with more complicated payoff then the standard European or American calls and
puts are called exotic options. They are usually traded over the counter. Their prices are
usually not quoted on an exchange. We list some common exotic options below.
1. Binary options
2. compound options
3. chooser options
4. barrier options
5. Asian options
6. Lookback options
In the last two, the payoff depends on the history of the asset prices, for instance, the aver
ages, the maximum, etc., we shall call these kinds of options, the pathdependent options,
and will be discussed in the next Chapter.
7.1 Binaries
The payoff function Λ(S) is an arbitrary function. One particular binary option is the cash
ornothing call, whose payoff is
Λ(S) = BH(S −E).
This option can be interpreted as a simple bet on an asset price: if S > E at expiry the
payoff is B, otherwise zero. We have seen its value is
V = e
−r(T−t)
∞
0
{(S
, T, S, t)Λ(S
)dS
= e
−r(T−t)
B^(d
2
),
where
{(S
, T, S, t) =
1
2πσ
2
(T −t)
e
−
log(
S
S
)−(r−
σ
2
2
)(T−t)
2σ
2
(T−t)
is the transition probability density for asset price in riskneutral world.
79
80 CHAPTER 7. EXOTIC OPTIONS
7.2 Compounds
A compound option may be described as an option on an option. We consider the case
where the underlying option is a vanilla put or call and the compound option is vanilla
put or call on the underlying option. The extension to more complicated option on more
complicated option is relatively straightforward. There are four different classes of basic
compound options:
1. calloncall,
2. callonput,
3. putoncall,
4. putonput.
Let us investigate the case calloncall. Other cases can be treated similarly. The underlying
option is
Expiry : T
2
, Strike price : E
2
.
The compound option on this option
Expiry : T
1
< T
2
, Strike price : E
1
.
The underlying option has value C(S, t, T
2
, E
2
). At time T
1
, its value C(S, T
1
, T
2
, E
2
).
The payoff for the compound call option is max¦C(S, T
1
, T
2
, E
2
) − E
1
, 0¦. Because the
compound options value is governed only by the randomness of S, according to the Black
Scholes analysis, it also must satisfy the same BlackScholes equation. We then solve the
BlackScholes equation with payoff
max¦C(S, T
1
, T
2
, E
2
) −E
1
, 0¦.
7.3 Chooser options
A regular chooser option gives its owner the right to purchase, for an amount E
1
at time T
1
,
either a call or a put with exercise price E
2
at time T
2
. Thus, it is a “call on a call or put”.
Certainly, we have T
1
< T
2
. The payoff at T
1
for this callonacallorput” is
Λ = max¦C(S, T
1
) −E
1
, P(S, T
1
) −E
1
, 0¦.
The compound option also satisﬁes the BlackScholes equation for the same reason as
above. From this and payoff function at T
1
, we can value V at t. The contract can be made
more general by having the underlying call and put with different exercise prices and expiry
dates, or by allowing the right to sell the vanilla put or call. By using the BlackScholes
formula for vanilla option, there is no difﬁculty to value these complex chooser options.
7.4. BARRIER OPTION 81
7.4 Barrier option
Barrier options differ from vanilla options in that part of the option contract is triggered if
the asset price hits some barrier, S = X, say at some time prior to T. As well as being
either calls or puts, barrier options are categorized as follows.
1. upandin: the option expires worthless unless S reaches X from below before ex
piry.
2. downandin: the option expires worthless unless S reaches X from above before
expiry.
3. upandout: the option expires worthless if S reaches X from below before expiry.
4. downandout: the option expires worthless if S reaches X from above before ex
piry.
7.4.1 downandout call(knockout)
A European option whose value becomes zero if S ever goes as low as S = X. Sometimes
in the knockout options, one can have boundary of time, or one can have rebate if the
barrier is crossed. In the latter case, the option holder receives a speciﬁc amount Z for
compensation.
Let us consider the case of a European style downandout option without relate. We
assume X < E. The boundary conditions are
V (X, t) = 0, (boundary condition), V (S, t) ∼ S as S →∞.
The ﬁnal condition, V (S, T) = max¦S −E, 0¦. For S > X, the option becomes a vanilla
call, it satisﬁes the BlackScholes equation.
Let us ﬁnd its explicit solution. Let S = Ee
x
, t = T −
τ
(σ
2
/2)
, V = Ev. The Black
Scholes equation is transformed into
v
τ
= v
xx
+ (k −1)v
x
−kv,
where k =
r
σ
2
/2
. We make another change of variable :
v = e
αx+βτ
u.
We choose α, β to eliminate the lower order terms in the derivatives of x:
βe
αx+βτ
u +e
αx+βτ
u
τ
= α
2
e
αx+βτ
u + 2αe
αx+βτ
u
x
+ e
αx+βτ
u
xx
+(k −1)(αe
αx+βτ
u + e
αx+βτ
u
x
) −ke
αx+βτ
u.
This implies that α = −
1
2
(k −1) and β = −
1
4
(k + 1)
2
and equation becomes u
τ
= u
xx
.
Let x
0
= log(
x
E
), or X = Ee
x
0
. The boundary condition becomes u(x
0
, τ) = 0 and
u(x, τ) ∼ e
(1−α)x−βτ
as x →∞. The initial condition becomes
u(x, 0) = u
0
(x) = max¦e
1
2
(k+1)x
−e
1
2
(k−1)x
, 0¦.
82 CHAPTER 7. EXOTIC OPTIONS
This follows from the payoff function being
Λ = max¦S −E, 0¦ = E max¦
S
E
−1, 0¦ = E max¦e
x
−1, 0¦,
and V = e
αx+βτ
u(x, T −
τ
(σ
2
/2)
), with u(x, 0) = u
0
(x) = e
−αx
max¦e
x
− 1, 0¦ =
max¦e
(−α+1)x
−e
−αx
, 0¦. Notice that because X < E, we have x
0
< 0, and
u
0
(x) =
0 for x
0
< x < 0,
max¦e
(−α+1)x
−e
−αx
, 0¦ otherwise.
We use methodofreﬂection to solve above heat equation with zero boundary condition.
We reﬂect the initial condition about x
0
as
u(x, 0) =
u
0
(x), for x
0
< x < ∞
−u
0
(2x
0
−x), for −∞< x < x
0
.
The equation and the initial condition are unchanged under the changeofvariable: x →
2x
0
−x, u →−u. From the uniqueness of the solution, the solution has the property:
u(2x
0
−x, t) = −u(x, t).
From this, we can obtain that u(x
0
, t) = −u(x
0
, t) = 0.
Since C = Ee
αx+βτ
u
1
is the vanilla call, where u
1
satisﬁes the heat equation with the
initial condition:
u
1
(x, 0) =
e
1
2
(k+1)x
−e
1
2
(k−1)x
for x > 0
0, for x ≤ 0
Using this and the method of reﬂection, we may express V in terms of C as the follows.
First, we may write V = Ee
αx+βτ
(u
1
+u
2
), where the initial condition for u
2
is the reﬂected
condition from u
1
:
u
2
(x, 0) =
u
0
(2x
0
−x, 0) for x ≤ 0
0 for x > 0
The solution u
1
corresponds to C(S, t). The solution u
2
is corresponds to
e
2α(x−x
0
)
C(x
2
/S, t). We conclude
V = C(S, t) −(
S
X
)
−(k−1)
C(X
2
/S, t).
7.4.2 downandin(knockin) option
An “in” option becomes worthless unless the asset price reaches the barrier before expiry.
If S crosses the line S = X at some time prior to expiry, then the option becomes a vanilla
option. It is common for intype barrier option to give a rebate, usually a ﬁxed amount, if
the barrier is not hit. This compensates the holder for the loss of the option.
The boundary condition for an “in” option is the follows. The option is worthless as
S → ∞, i.e., V (S, t) → 0 as S → ∞. At T, if S > X, then V (S, T) = 0. For t < T,
V (X, t) = C(X, t). Since the option immediately turns into a vanilla call and must have
7.5. ASIAN OPTIONS AND LOOKBACK OPTIONS 83
the same value of this vanilla call. For S ≤ X, V (S, t) = C(X, t). We only need to solve
V for S > X. V still satisﬁes the same BlackScholes equation for all S, t, because its
randomness is fully correlated to the randomness of S.
We may write V = c − V , where c is the value of a vanilla call. Then the boundary
condition for V is V (S, t) = c −V ∼ S −0 = S as S →∞. And
V (X, t) = c(X, t) −V (X, t) = 0, V (S, T) = c(S, T) −V (S, T) = c(S, T) = Λ(S).
We observe that V is indeed a “downandout” barrier option. In other words, 1(downand
in) plus 1(downandout) equal to 1 vanilla call. This is because one and only one of the
two barrier options can be active at expiry and whichever it is, its value is the value of a
vanilla call.
7.5 Asian options and lookback options
In Asian options and lookback options, their payoff functions depend on the history of the
underlying asset. For example,
1. a Europeantype average strike option has the following payoff function
max¦S
T
−
1
T
T
0
S(τ)dτ, 0¦.
2. an Americantype average strike option,
Λ(S, t) = max¦S −
1
t
t
0
S(τ)dτ, 0¦.
3. geometric mean,
Λ(S, T) = max¦S −e
R
T
0
log S(τ)dτ
, 0¦.
4. Lookback call,
Λ(S, T) = max¦S −J, 0¦, J = max
0≤τ≤T
S(τ)
In general, the payoff depends on I, which is deﬁned by
I =
t
0
f(S(τ), τ)dτ,
where f is a smooth function. The payoff function is Λ(S, I). It is important to notice that
I(t) is independent of S(t). The value of an asian option should depend on S, t s well as
I. Indeed, we shall see in the next chapter that dI = fdt. The only randomness is through
S, therefore V can be valued through a delta hedge.
For the lookback option, it will be treated as a limiting case of an asian option. We shall
discuss this in the next chapter.
84 CHAPTER 7. EXOTIC OPTIONS
Chapter 8
PathDependent Options
8.1 Introduction
If the payoff depends on the history of the underlying asset price, such an option is called a
pathdependent option. The Asian options and the Russian options (Lookback options) are
the typical examples. The payoff functions for these options are, for example,
1. average strike call option: Λ = max¦S −
1
T
T
0
S(τ)dτ, 0¦,
2. average rate call option: Λ = max¦
1
T
T
0
S(τ)dτ −E, 0¦
3. geometric mean: the arithmetic mean
1
T
T
0
S(τ)dτ above is replaced by e
R
T
0
log(S(τ))dτ
.
4. lookback strike put: Λ = max¦max
0≤τ≤T
S(τ) −S, 0¦.
5. lookback rate put: Λ = max¦E −max
0≤τ≤T
S(τ), 0¦.
8.2 General Method
Let f be a smooth function, deﬁne
I(t) =
t
0
f(S(τ), τ)dτ.
In previous examples, f(S(τ), τ) = S(τ) for arithmetic mean and f(S(τ), τ) = log S(τ)
for geometric mean.
Notice that I(t) is a random variable and is independent of S(t). (This is because I(t)
is the sum of increment of functions of S before time t, and each increment of S(τ), τ < t,
is independent of S(t).) Therefore, we should introduce another independent variable I
besides S to value the derivative V (S, I, t).
The stochastic differential equations governed by S and I are
dS
S
= µdt + σdz,
dI(t) = f(S(t), t)dt.
85
86 CHAPTER 8. PATHDEPENDENT OPTIONS
Notice that there is no noize term in dI. The only randomness is from dS. Therefore, we
can use delta hedge to eliminate this randomness. Namely, we consider the portfolio
Π = V −∆S,
as before. We have
dΠ = dV −∆dS = (V
t
+
1
2
σ
2
S
2
∂
2
S
∂V
2
)dt + V
I
dI +V
S
dS −∆dS.
We choose ∆ =
∂V
∂S
to eliminate the randomness in dΠ. From the arbitrage assumption, we
arrive
V
t
+
1
2
σ
2
S
2
∂
2
V
∂S
2
+ f
∂V
∂I
= r(V −S
∂V
∂S
).
8.3 Average strike options
8.3.1 European calls
Let us consider an average strike call option with European exercise feature. Its payoff
function is deﬁned by
max¦S −
1
T
T
0
S(τ)dτ, 0).
Or, in terms of I, Λ(S, I, T) = max¦S −
I
T
, 0¦.
We notice that the modiﬁed BlackScholes equation
V
t
+S
∂V
∂I
+
σ
2
2
S
2
∂
2
V
∂S
2
+rS
∂V
∂S
−rV = 0,
and the initial data for the average strike options are invariant under the transformation:
(S, I) →λ(S, I). Therefore, we expect that its solution is a function of the scaleinvariant
variable R = I/S. Notice that this is also reﬂected in that
dR = (1 + (σ
2
−µ)R)dt −σRdz,
depends on R only. Since the initial data can be expressed as Λ = S max¦1 −R/T, 0¦, we
may also expect that V = SH(R, t). This reduces one independent variable.
Plug V = SH into the above modiﬁed BlackScholes equation:
SH
t
+
1
2
σ
2
S
2
(2
∂H
∂S
+ S
∂
2
H
∂S
2
) + S S
∂H
∂I
+rS
∂
∂S
(SH) −r(SH) = 0.
From
∂
∂S
=
∂R
∂S
,
∂
∂R
= −
R
S
∂
∂R
∂
2
∂S
2
=
2I
S
3
∂
∂R
+
I
2
S
4
∂
2
∂R
2
=
1
S
2
(2R
∂
∂R
+ R
2
∂
2
∂R
2
),
8.3. AVERAGE STRIKE OPTIONS 87
we obtain
SH
t
+
1
2
σ
2
S
2
(2(−
R
S
∂H
∂R
) + S
1
S
2
(2R
∂H
∂R
+ R
2
∂
2
H
∂R
2
))
+S
2
1
S
H
R
+ rSH + rS
2
(−
R
S
H
R
−rSH) = 0
Finally, we arrive
H
t
+
σ
2
2
R
2
∂
2
H
∂R
2
+ (1 −rR)
∂H
∂R
= 0.
The payoff function
Λ(R, T) = max¦1 −
R
T
, 0¦ = H(R, T), (ﬁnal condition) .
we should require the boundary conditions.
• H(∞, t) = 0. Since as R →∞implies S →0, then V →0, then H →0.
• H(0, t) is ﬁnite. When R = 0, we have S(τ) = 0, for all τ with probability 1. That
implies that Λ = 0, and consequently, H is ﬁnte at (0, t).
Next, we expect that the solution is smooth up to R =. This implies that H
R
, H
RR
are
ﬁnite at (0, t). We have the following two cases: (i) If R
2 ∂
2
H
∂R
2
= O(1) = 0 as R → 0 then
H = O(log R) for R → 0. Or (ii) if R
∂H
∂R
= O(1) = 0 as R → 0, then H = O(log R).
Both cases contradict to H(0, t) being ﬁnite. Hence, we have RH
R
(0, t), R
2
H
RR
(0, t) are
zeros as R →0. Hence the boundary condition H(0, t) is ﬁnite is equivalent to H
t
(0, t) +
H
R
(0, t) = 0.
This equation with boundary condition can be solved by using the hypergeometric func
tions. However, in practice, we solve it by numerical method.
8.3.2 American call options
We consider the average strike call option with American exercise feature. In this case,
H
t
+
1
2
σ
2
R
2
∂
2
H
∂R
2
+ (1 −rR)
∂H
∂R
≤ 0
H −Λ ≥ 0
(H
t
+
1
2
σ
2
R
2
∂
2
H
∂R
2
+ (1 −rR)
∂H
∂R
)(H −Λ) = 0,
where Λ(R, t) = max¦1 −
R
t
, 0¦, R(t) = I(t)/S(t), I(t) =
t
0
S(τ)dτ.
8.3.3 Putcall parity for average strike option
We study the putcall parity for average strike options with European exercise feature.
Consider a portfolio is C −P. The corresponding payoff function is
S max¦1 −
R
T
, 0¦ −S max¦
R
T
−1, 0¦ = S(1 −
R
T
) ≡ S H(R, T).
88 CHAPTER 8. PATHDEPENDENT OPTIONS
Since the BlackScholes equation in linear in R, we only need to solve the equation with
ﬁnal condition (i) H(R, T) = 1, (ii) H(R, T) = −
R
T
. For (i), H(R, t) ≡ 1. For (ii), since
the ﬁnal condition and the P.D.E. is linear in R, we expect that the solution is also linear in
R. Thus, we consider H is of the following form
H(t, R) = a(t) + b(t)R
Plug this into the equation, we obtain
d
dt
a +
d
dt
bR + (1 −rR)b = 0.
and
d
dt
a + b = 0,
d
dt
b −rb = 0.
with a(T) = 0, b(T) = −
1
T
. This differential equation can be solved easily:
b(t) = −
1
T
e
−r(T−t)
, a(t) = −
1
rT
(1 −e
−r(T−t)
).
Consequently, we obtain the putcall parity:
C −P = S(1 −
1
rT
(1 −e
−r(T−t)
−
1
T
e
−r(T−t)
1
S
t
0
S(τ)dτ)
= S −
S
rT
(1 −e
−r(T−t)
) −e
−r(T−t)
1
T
t
0
S(τ)dτ
8.4 Lookback Option
A lookback option is a derivate product whose payoff depends on the maximum or mini
mum of its underlying asset price. For instance, the payoff function for a lookback option
with European exercise feature is
Λ = max
0≤τ≤T
S(τ) −S(T).
Such an option is relatively expansive because it gives the holder an extremely advanta
geous payoff.
As before, let us introduce J(t) = max
0≤τ<t
S(τ). Since S(τ), τ < t are independent
of S(t), we see that J(t) is independent of S(t). This suggest that we should introduce
another independent variable J to value the lookback option in addition to S and t. We can
derive a stochastic differential equation for J as before. Indeed, it is dJ = 0. However,
we shall give a more careful approach. We shall use the fact that for a continuous function
S(),
max
0≤τ≤t
[S(τ)[ = lim
n→∞
t
0
[S(τ)[
n
dτ
1
n
.
8.4. LOOKBACK OPTION 89
We leave its proof as an exercise.
Remark. For the minimum, we have
lim
n→−∞
t
0
(S(τ))
n
dτ
1
n
= min
0≤τ≤t
S(τ).
Let us introduce
I
n
=
t
0
(S(τ))
n
dτ, J
n
= I
1
n
n
.
The s.d.e. for J
n
,
dJ
n
= (
t+dt
0
(S(τ))
n
dτ)
1
n
−(
t
0
(S(τ))
n
dτ)
1
n
=
1
n
S(t)
n
J
n−1
n
dt.
Now, as before we consider the delta hedge:
Π = P −∆S.
From the arbitrage assumption, we can derive the equation for P(S, J
n
, t):
dΠ = P
t
dt +
1
n
S
n
J
n−1
n
∂P
∂J
n
dt +
1
2
σ
2
S
2
∂
2
P
∂S
2
dt
= r(P −
∂P
∂S
S)dt
Taking n →∞, using the facts that J
n
→J and
S
J
≤ 1, we arrive
P
t
+
1
2
σ
2
S
2
∂
2
P
∂S
2
+ rS
∂P
∂S
−rP = 0.
This is the usual BlackScholes equation. The role of J here is only a parameter. This is
consistent to the fact that
dJ = 0.
8.4.1 A lookback put with European exercise feature
The range for S is 0 ≤ S ≤ J. This is because S ≤ J, for 0 ≤ t ≤ T. We claim that
P(0, J, t) = Je
−r(T−t)
.
Firstly, we have that Λ(0, J, T) = max¦J − S, 0¦ = J. Secondly, if S(t) = 0, then
S(τ) = 0 for t ≤ τ ≤ T. The asset price process becomes deterministic. Therefore, the
value of P is the discounted payoff: P(0, J, t) = Je
−r(T−t)
.
Next, we claim that
∂P
∂J
(J, J, t) = 0.
90 CHAPTER 8. PATHDEPENDENT OPTIONS
From µ > 0, the current maximum cannot be the ﬁnal maximum with probability 1. The
value of P must be insensitive to a small change of J.
We can use method of image to solve this problem. Its solution is given by
P = S(−1 + N(d
7
)(1 + k
−1
)) + Je
−r(T−t)
N(d
5
) −k
−1
(
S
J
)
1−k
N(d
6
),
where
d
5
= [ln(
J
S
) −(r −
σ
2
2
)(T −t)]/σ
√
T −t
d
6
= [ln(
S
J
) −(r −
σ
2
2
)(T −t)]/σ
√
T −t
d
7
= [ln(
J
S
) −(r +
σ
2
2
)(T −t)]/σ
√
T −t
k =
r
σ
2
/2
8.4.2 Lookback put option with American exercise feature
We have the following linear complementary equation,
L
BS
P ≤ 0, P −Λ ≥ 0, (L
BS
P)(P −Λ) = 0,
where
L
BS
=
∂
∂t
+
σ
2
S
2
2
∂
2
∂S
2
+rS
∂
∂S
−r.
The ﬁnal condition
P(S, J, T) = Λ(S, J, T) = J −S.
The boundary condition
∂P
∂J
(J, J, t) = 0.
We require P,
∂P
∂S
∂P
∂J
are continuous.
For lookback call option, we simply replace max
0≤τ≤t
S(τ) by min
0≤τ≤t
S(τ). For
instance, its payoff is Λ = S(T) −min
0≤τ≤T
S(τ).
Chapter 9
Bonds and Interest Rate Derivatives
9.1 Bond Models
A bond is a longterm contract under which the issuer promises to pay the bondholder
coupon payment (usually periodically) and principal (at the maturity dates). If there is no
coupon payment, the bond is called a zerocoupon bond. The principal of a bond is called
its face value.
9.1.1 Deterministic bond model
The value of a bond certainly depends on the interest rate. Let us ﬁrst assume that the
interest rate is deterministic temporarily, say r(τ), t ≤ τ ≤ T, is known. Let B(t, T) be
the bond value at t with maturity date T, k(t) be its coupon rate. This means that in a small
dt, the holder receives coupon payment k(t) dt. From the noarbitrage argument,
dB + k(t) dt = r(t)Bdt.
together with the ﬁnal condition:
B(T, T) = F( face value),
the bond value can be solved and has the following expression:
B(t, T) = e
−
R
T
t
r(τ) dτ
¸
F +
T
t
k(τ)e
R
T
τ
r(s) ds
dτ
.
Thus, the bond value is the sum of the present face value and the coupon stream.
However, the life span of a bond is long (usually 10 years or longer), it is unrealistic
to assume that the interest rate is deterministic. In the next subsection, we shall provide a
stochastic model.
9.1.2 Stochastic bond model
Let us assume that the interest rate satisﬁes the following s.d.e.:
dr = u(r, t)dt + w(r, t)dz,
91
92 CHAPTER 9. BONDS AND INTEREST RATE DERIVATIVES
where dz is the standard Wiener process. The drift u and the variance w
2
are proposed
by many researchers. We shall discuss these issues later. To ﬁnd the equation for B with
stochastic property of r, we consider a portfolio containing bonds with different maturity
dates:
Π = V (t, r, T
1
) −∆V (t, r, T
2
) ≡ V
1
−∆V
2
.
The change dΠ in a small time step dt is
dΠ = dV
1
−∆dV
2
,
where
dV
i
V
i
= µ
i
dt + σ
i
dz,
µ
i
=
1
V
i
V
i,t
+uV
i,r
+
1
2
w
2
V
i,rr
σ
i
=
1
V
i
wV
i,r
.
We we choose ∆ = V
1,r
/V
2,r
, then the random term is canceled in dΠ. From the no
arbitrage argument, dΠ = rΠdt. We obtain
µ
1
V
1
dt −∆µ
2
V
2
dt = r(V
1
−∆V
2
) dt.
This yields
(µ
1
−r)V
1
/V
1,r
= (µ
2
−r)V
2
/V
2,r
,
or equivalently
µ
1
−r
σ
1
=
µ
2
−r
σ
2
.
Since the lefthand side is a function of T
1
, while the righthand side is a function of T
2
.
Therefore, it is independent of T. Let us express it as a known function λ(r, t):
µ −r
σ
= λ(r, t).
Plug µ
i
and σ
i
back to this equation, and drop the index i, we obtain
V
t
+
1
2
w
2
V
rr
+ (u −λw)V
r
−rV = 0.
The function λ(r, t) =
µ−r
σ
is called the market price, since it gives the extra increase in
expected instantaneous rate of return on a bond per an additional unit of risk.
This stochastic bond model depends on three parameter functions u(r, t), w(r, t) and
λ(r, t). In the next section, we shall provide some model to determine them.
9.2. INTEREST MODELS 93
9.2 Interest models
There are many interest rate models. We list some of them below.
• Merton (1973): dr = αdt +σdz.
• Vasicek (1977): dr = β(α −r)dt + σdz.
• Dothan (1978): dr = σrdz.
• MarshRosenfeld (1983): dr = (αr
δ−1
+ βr)dt + σr
δ/2
dz.
• CoxIngersollRoss (1985) dr = β(α −r)dt +σr
1
2
dz.
• HoLee (1986): dr = α(t)dt + σdz.
• BlackKarasinski (1991): d ln r = (a(t) −b(t) ln r) dt + σdz.
The main requirements for an interest rate model are
• positivity: r(t) ≥ 0 almost surely,
• mean reversion: r should tends to increase (or to decrease) and toward a mean.
The CIR and BK models have these properties.
Below, we shall illustrate a uniﬁed approach proposed by Luo, Yen and Zhang.
9.2.1 A functional approach for interest rate model
The idea is to design r to be a function of x(t) and t, (i.e. r = r(x(t), t)) with x(t)
governed by a simple stochastic process. We notice that the OrnsteinUhlenbeck process
dx = −ηxdt + σdz has the property to tend to its mean (which is 0) time asymptotically.
While the Bessel process dx = /xdt + σdz has positive property. We then design the
underlying basic process is the sum of these two processes:
dx =
−ηx +
x
dt + σdz.
In general, we allow η, , and σ are given functions of t. With this simple process, we can
choose r = r(x(t), t). Then all interest models mentioned above correspond to different
choices of r(x, t), η, and .
• Merton’s model: we choose = η = 0, r = x + αt.
• Dothan’s model: = η = 0, r = e
x−σ
2
/2t
.
• HoLee: = η = 0, r = x +
t
0
α(s) ds.
• Vasicek: = 0, η = β, r = x +α.
• CIR: β = 2η, α = (σ
2
+ 2)/(8η), r =
1
4
x
2
.
94 CHAPTER 9. BONDS AND INTEREST RATE DERIVATIVES
• BlackKarasinski: = 0, r = exp(g(t, x)), where g = x +
t
0
a(s) ds, η = b(t).
With the interest rate model, the zerocoupon bond price V is given by
V (x, t) = E
e
R
T
t
r(s,x(s)) ds
[ x
t
= x
, t < T.
From the FeymannKac formula, V satisﬁes
¸
∂
∂t
+
1
σ
2
∂
2
∂x
2
+
−ηx +
x
∂
∂x
−r
V = 0.
This model depends three parameter functions (t), η(t), σ(t), and r(x, t). There is no
uniﬁed theory available yet with this approach and the approach of the previous subsection.
9.3 Convertible Bonds
A convertible bond is a bond plus a call option under which the bond holder has the right
to convert the bond into a common shares. Thus, it is a function of r, S, t and T. Let the
stochastic processes governed by S and r are
dS
S
= µdt + σdz
S
,
dr = udt + wdz
r
.
Suppose the correlation between dz
S
and dz
R
is
dz
S
dz
r
= ρ(S, r, t)dt.
The ﬁnal value of the convertible bond V (r, S, T) = F, the face value of the bond. Suppose
the bond can be converted to nS at any time priori to T. Then we have
V (r, S, t) ≥ nS.
We also have the boundary conditions:
lim
S→∞
V (r, S, t) = nS
lim
r→∞
V (r, S, t) = 0.
At S = 0 or r = 0, we should require V (r, 0, t) or V (0, S, t) to be ﬁnite.
The BlackScholes analysis for a convertible bond is similar to the analysis for a bond.
Let V
i
= V (r, S, t, T
i
), i = 1, 2. Consider a portfolio
Π = ∆
1
V
1
+ ∆
2
V
2
+ ∆
S
S.
In a small time step dt, the change of dΠ is
dΠ = ∆
1
dV
1
+ ∆
2
dV
2
+ ∆
S
dS,
9.3. CONVERTIBLE BONDS 95
where
dV
i
V
i
= µ
i
dt + σ
r,i
dz
r
+ σ
S,i
dz
S
,
µ
i
=
1
V
i
V
i,t
+
σ
2
2
S
2
V
i,SS
+ ρSwV
i,Sr
+
w
2
2
V
i,rr
+ µSV
i,S
+ uV
i,r
,
σ
S,i
=
1
V
i
σSV
i,S
σ
r,i
=
1
V
i
wV
i,r
.
This implies
dΠ = (∆
1
µ
1
V
1
+ ∆
2
µ
2
V
2
+ ∆µS) dt
+(∆
1
σ
S,1
V
1
+ ∆
2
σ
S,2
V
2
+ ∆σS) dz
S
+(∆
1
σ
r,1
V
1
+ ∆
2
σ
r,2
V
2
) dz
r
.
We choose ∆
1
, ∆
2
and ∆
S
to cancel the randomness terms dz
r
and dz
S
. This means that
(∆
1
σ
S,1
V
1
+ ∆
2
σ
S,2
V
2
+ ∆σS) dz
S
= 0
(∆
1
σ
r,1
V
1
+ ∆
2
σ
r,2
V
2
) dz
r
= 0.
And it yields
dΠ = (∆
1
µ
1
V
1
+ ∆
2
µ
2
V
2
+ ∆µS) dt
= r(∆
1
V
1
+ ∆
2
V
2
+ ∆S) dt.
Or equivalently,
∆
1
(µ
1
−r)V
1
+ ∆
2
(µ
2
−r)V
2
+ ∆(µ −r)S = 0.
This equality together with the previous two give that there exist λ
r
and λ
S
such that
(µ
1
−r) = λ
S
σ
S,1
+ λ
r
σ
r,1
(µ
2
−r) = λ
S
σ
S,2
+ λ
r
σ
r,2
(µ −r) = λ
S
σ
The functions λ
r
and λ
S
are called the market prices of risk with respect to r and S, respec
tively. Plug the formulae for µ
i
, σ
r,i
and σ
S,i
, we obtain the BlackScholes equation for a
convertible bond:
¸
∂
∂t
+
σ
2
2
S
2
∂
2
∂S
2
+ ρSw
∂
2
∂S∂r
+
w
2
2
∂
2
∂r
2
+ rS
∂
∂S
+ (u −λ
r
w)
∂
∂r
−r
V = 0.
96 CHAPTER 9. BONDS AND INTEREST RATE DERIVATIVES
Appendix A
Basic theory of stochastic calculus
A.1 Brownian motion
Let (Ω, T, P) be a probability space. A process is a function X : [0, ∞) (Ω, T) →
(E, B), such that for each t ≥ 0, X(t) is a random variable. Here, E = R
d
, B is the Borel
sets. Let
T
t
= σ¦X(s), s ≤ t¦
T
t
= σ¦X(s), s ≥ t¦
A process is called Markov if
P(A [ T
t
) = P(A [ X(t)), ∀A ∈ T
t
.
This is equivalent to
P¦X(r) ∈ B[T
t
¦ = P¦X(r) ∈ B[X(t)¦∀r > t,
A markov process is characterized by its transition probability:
P(t, x, s, B) := P¦X(s) ∈ B [ X(t) = x¦
with initial distribution
P¦X(0) ∈ B¦ = ν(B).
Theorem 1.7 If X is a Markov process, then the corresponding transition probability P
satisﬁes ChapmanKolmogorov equation:
P(t
0
, x
0
, t
1
, dx
1
)P(t
1
, x
1
, t
2
, B) = P(t
0
, x
0
, t
2
, B).
Conversely, if P is a function satisﬁes ChapmanKolmogorov equation, then there is a
Markov process whose transition probability is P.
97
98 APPENDIX A. BASIC THEORY OF STOCHASTIC CALCULUS
Two standard Markov processes are the Wiener process and the Poison process. The
Wiener process has the transition probability density function
p(t, x, s, y) =
1
(2π(t −s))
d/2
exp(−
[x −y[
2
s −t
).
Such a distribution is called a normal distribution with mean x and variance s −t.
Deﬁnition 1.7 Brownian motion: A process is called a Brownian motion(or a Wiener pro
cess) if
1. B
t
−B
s
has normal distribution with mean 0 and variance t −s,
2. it has independent increments: that is B
t
−B
s
is independent of B
u
for all u ≤ s < t,
3. B
t
is continuous in t.
It is easy to see that B
t
is markovian and its transition probability is
p(t, x, s, y) =
1
(2π(t −s))
d/2
exp(−
[x −y[
2
s −t
).
Deﬁnition 1.8 A process ¦X
t
[ t ≥ 0¦ is called martingale if
(i) EX
t
< ∞,
(ii) E(X
u
[ X
s
, 0 ≤ s ≤ t¦ = X
t
.
This means that if we know the value of the process up to time t and X
t
= x, then the
future expectation of X
u
is x.
Theorem 1.8 1. B
t
is a martingale.
2. V
2
t
−t is a martingale.
Proof.
1. E(B
t
) = 0. From the fact that B
t+s
− B
t
is independent of B
t
, for all s > 0, we
obtain E(B
t+s
−B
t
[ B
t
) = 0, for all s > 0. Hence,
E(B
t+s
[ B
t
) = E((B
t+s
−B
t
) + B
t
[ B
t
)
= E((B
t+s
−B
t
) [ B
t
) + E(B
t
[ B
t
)
= B
t
2. E(B
2
t
) = t < ∞.
A.1. BROWNIAN MOTION 99
3. Use
B
2
t+s
= ((B
t+s
−B
t
) + B
t
)
2
= (B
t+s
−B
t
)
2
+ 2B
t
(B
t+s
−B
t
) + B
2
t
and the fact that B
t+s
−B
t
is independent of B
t
, we obtain
E(B
2
t+s
[ B
t
) = s +B
2
t
.
Hence,
E(B
2
t+s
−(t + s) [ B
t
) = B
2
t
−t.
We can show that the Brownian motion has inﬁnite total variation in any interval. This
means that
lim
¸
[B(t
i
) −B(t
i−1
)[ = ∞.
However, its quadratic variation, deﬁned by
[B, B](a, b) := lim
¸
[B(t
i
) −B(t
i−1
)[
2
is ﬁnite. Here, a = t
0
< t
1
< < t
n
= b is a partition of (a, b), and the limit is taken to
be max(t
i
−t
i−1
) →0.
Theorem 1.9 We have [B, B](0, t) = t almost surely.
Proof. Let us partition (0, t) evenly into 2
n
subintervals. Let T
n
=
¸
2
n
i=1
[B(t
i
)−B(t
i−1
)[
2
.
We see that
ET
n
=
¸
E([B(t
i
) −B(t
i−1
)[
2
)) =
¸
[t
i
−t
i−1
[ = t.
V ar(T
n
) = V ar(
¸
[B(t
i
) −B(t
i−1
)[
2
)
=
¸
var((B(t
i
) −B(t
i−1
))
2
)
= 2
¸
(t
i
−t
i−1
)
2
= 2t
2
2
−n
Hence,
¸
∞
n=1
V ar(T
n
) < ∞. From Fubini theorem,
E
∞
¸
n=1
(T
n
−ET
n
)
2
< ∞.
This implies E((T
n
−ET
n
)
2
) →0 and hence, T
n
→ET
n
almost surely.
100 APPENDIX A. BASIC THEORY OF STOCHASTIC CALCULUS
A.2 Stochastic integral
We shall deﬁne the integral
t
0
f(s)dB(s).
The method can be applied with B replaced by a martingale, or a martingale plus a function
with ﬁnite total variation. We shall require that f ∈ H
2
, which means:
(i) f(t) depends only on the history T
t
of B
s
, for s ≤ t,
(ii)
t
0
E[f[
2
< ∞.
For this kind of functions, we can deﬁne its It ˆ o’s integral as the follows.
1. We deﬁne It ˆ o’s integral for f ∈ H
2
and being step functions. Its It ˆ o’s integral is
deﬁne by
t
0
f(s) dB(s) =
n
¸
i=0
f(t
i−1
)(B(t
i
) −B(t
i−1
)).
2. We use above step functions f
n
to approximate general function f ∈ H
2
. Using the
fact that, for step functions g ∈ H
2
,
E
t
0
g(s) dB(s)
2
=
t
0
E[g(s)[
2
ds,
one can show that
lim
n→∞
t
0
f
n
(s) dB(s)
almost surely.
An typical example is
t
0
B(s) dB(s) =
1
2
B(t)
2
−
t
2
.
From the deﬁnition, the integral can be approximated by
I
n
=
n
¸
i=1
B(t
i−1
)(B(t
i
) −B(t
i−1
).
We have
I
n
=
1
2
n
¸
i=1
B
2
(t
i
) −B
2
(t
i−1
) −(B(t
i
) −B(t
i−1
))
2
=
1
2
B
2
(t) −
1
2
n
¸
i=1
(B(t
i
) −B(t
i−1
))
2
We have seen that the second on the righthand side tends to
1
2
t almost surely.
A.3. STOCHASTIC DIFFERENTIAL EQUATION 101
A.3 Stochastic differential equation
A stochastic differential equation has the form
dX
t
= a(X
t
, t)dt + b(X
t
, t)dB(t) (A.1)
A Markov process X is said to be a strong solution of this s.d.e. if it satisﬁes
X
t
−X
0
=
t
0
a(X
s
, s) ds +
t
0
b(X
s
, s)dB(s).
Theorem 1.10 (Itˆ o’s formula) If X satisﬁes the s.d.e. dX = adt + bdB, then
df(X(t)) = (af
(X(t)) +
1
2
b
2
f
(X(t)))dt + f
(X(t))bdB(t).
We shall give a brief idea of the proof. In a small time step (t
i−1
, t
i
), let ∆B = B(t
i
) −
B(t
i−1
), ∆t = t
i
−t
i−1
. We have
f(X(t
i−1
) + ∆X) −f(X(t
i−1
)) = f
∆X +
1
2
f
(∆X)
2
+o((∆X)
2
).
We notice that
(∆X)
2
= (a∆t + b∆B)
2
= a
2
(∆t)
2
+ 2ab∆t +b
2
(∆B)
2
≈ b
2
∆t + o(∆t).
Plug ∆X and (∆X)
2
into the Taylor expansion formula for f. This yields the Inˆ o’s for
mula.
A.4 Diffusion process
For a s.d.e.(A.1), we deﬁne the associated semigroup T
t
by
T
t
f = E
x,t
(f(X(s)), s > t
From the Markovian property, one can show that T
t
is a semigroup. Indeed, in terms of
the transition probability density function p(t, x, s, y),
T
t
f =
p(t, x, s, y)f(y) dy.
For a semigroup T
t
, we deﬁne its generator as
Lf := lim
t→0
T
t
f −f
t
.
From It ˆ o’s formula,
Lf := af
+
1
2
b
2
f
.
This follows from It ˆ o’s formula and E(
t
0
g(s) dB(s) = 0).
With a ﬁxed s > t and f, we deﬁne
u(x, t) := (T
t
f)(x, t) = E
x,t
(f(X(s)).
102 APPENDIX A. BASIC THEORY OF STOCHASTIC CALCULUS
Theorem 1.11 If X satisﬁes the s.d.e. (A.1), then the associate u(x, t) := E
x,t
(f(X(s)),
s > t, satisﬁes the backward diffusion equation:
u
t
+Lu = 0,
and has ﬁnal condition u(s, x) = f(x).
Proof. First, we notice that
u(x, t −h) = E
x,t−h
(f(X(s))
= E
x,t−h
E
X(t),t
(f(X(s))
= E
x,t−h
(u(X(t), t))
This shifts ﬁnal time from s to t. Now, we consider
u(x, t −h) −u(x, t)
h
=
1
h
E
x,t−h
(u(X(t), t) −u(X(t −h), t))
=
1
h
t
t−h
Lu(X(s), t) ds
→ Lu(x, t)
as h → 0+. Next, u(x, s − h) = E
x,s−h
f(X(s)), we have X(s) → x as h → 0+ almost
surely. Hence u(x, s −h) →f(x) as h →0+.
Since u can be represent as
u(x, t) =
p(x, t, s, y)f(y) dy,
we obtain that p satisﬁes
p
t
+ L
x
p = 0,
and
p(s, x, s, y) = δ(x −y).
For diffusion equation with source term, its solution can be represented by the following
FeymannKac formula.
Theorem 1.12 (FeymannKac) If X satisﬁes the s.d.e. (A.1), then the associate
u(x, t) := E
x,t
¸
f(X(s)) exp
s
t
g(X(τ), τ) dτ
, s > t (A.2)
solves the backward diffusion equation:
u
t
+ Lu + gu = 0,
with ﬁnal condition u(s, x) = f(x).
A.4. DIFFUSION PROCESS 103
Proof. As before, we shift ﬁnal time from s to t:
u(x, t −h) = E
x,t−h
¸
f(X(s)) exp
s
t−h
g(X(τ), τ) dτ
= E
x,t−h
E
X(t),t
¸
f(X(s)) exp
s
t−h
g(X(τ), τ) dτ
= E
x,t−h
¸
E
X(t),t
¸
f(X(s)) exp
s
t
g(X(τ), τ) dτ
E
X(t),t
¸
exp
t
t−h
g(X(τ), τ) dτ
= E
x,t−h
¸
u(X(t), t) exp
t
t−h
g(X(τ), τ) dτ
.
Here, we have used independence of X in the regions (t −h, t) and (t, s). Now, from It ˆ o’s
formula, we have
u(x, t −h) −u(x, t) = E
x,t−h
¸
u(X(t), t) exp
t
t−h
g(X(τ), τ) dτ −u(X(t −h), t)
= E
x,t−h
t
t−h
d
u(X(s), t) exp
s
t−h
g(X(τ), τ) dτ
= E
x,t−h
t
t−h
(Lu(X(s), t) + u(X(s), t)g(X(s), s)) ds.
From this, it is easy to see that
lim
h→0+
u(x, t −h) −u(x, t)
h
= Lu(x, t) + g(x, t)u(x, t).
2
Contents
1 Introduction 1.1 Financial Markets . . 1.2 Financial Derivatives . 1.3 Examples . . . . . . . 1.4 Payoff functions . . . 1.5 Other kinds of options 1.6 Types of traders . . . . 1.7 Basic assumption . . . 1 1 1 2 3 5 5 6 7 7 7 7 8 8 10 10 10 11 12 12 13 14 16 19 19 20 20 20 22 24 26 26 27
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Asset Price Model 2.1 Efﬁcient market hypothesis . . . . . . . . . . . . 2.2 The asset price model . . . . . . . . . . . . . . . 2.2.1 The discrete asset price model . . . . . . . 2.2.2 The continuous asset price model . . . . . 2.3 Random walk . . . . . . . . . . . . . . . . . . . . 2.4 The solution of the discrete asset price model . . . 2.5 The Brownian motion . . . . . . . . . . . . . . . 2.5.1 The deﬁnition of a Brownian motion . . . 2.5.2 The Brownian as a limit of random walk . 2.5.3 Properties of Brownian motion . . . . . . 2.6 Itˆ ’s formula . . . . . . . . . . . . . . . . . . . . o 2.7 The solution of the continuous asset price model . 2.8 Continuous model as a limit of the discrete model . 2.9 Simulation of asset price model . . . . . . . . . .
3 BlackScholes Analysis 3.1 The hypothesis of noarbitrageopportunities . . . . . . . . 3.2 Basic properties of option prices . . . . . . . . . . . . . . 3.2.1 The relation between payoff and options . . . . . . 3.2.2 European options . . . . . . . . . . . . . . . . . . 3.2.3 Basic properties of American options . . . . . . . 3.2.4 Dividend Case . . . . . . . . . . . . . . . . . . . 3.3 The BlackScholes Equation . . . . . . . . . . . . . . . . 3.3.1 BlackScholes Equation . . . . . . . . . . . . . . 3.3.2 Boundary and Final condition for European options 3
.5 Fast algorithms for solving linear systems . . . . . . . . 5. . .5. . . . . . . . . . . . .7. . . . . . .4 Butterﬂy spread . . . . . . . . . . . . . . . . . .3 Futures options . . . . . . . . . . .1 Forward contracts . . 4. . . . . . . . . . .7. . . . . . . . . . .4. . . . . 3. . . . . . . .4. . . . . .3. . . 3. . . . . . 5. . . . . . . . . . . . . . . . . . . . 70 6. . . . . . . . . . . . . . . .3. . .4 BlackScholes analysis on futures options 5 Numerical Methods 5. . . . . . . . . . . . . . . . . .3. . . . . . . . The delta hedging .2 Iterative methods . . .2 Futures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Binomial method for asset price model . . 4. . . . . . . . . . . . . 3. . .7. . . . . . . . . 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 American Option 69 6. .6. . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Discretization methods . . . . . 70 . . . . . . . . . . . . . . . . . . .2. . . . . . . . . . . . . . . . . . .7 4 Variations on BlackScholes models 4. . . . . . . . . . . . . . . . . . . . . . .5 3. . . . . .1 Reduction to parabolic equation with constant coefﬁcients 3.3. . . .3. . . . . . . . .4 Special cases . . . . . . .1 TimeDependent r. . . . 4. . .3. . . 5. . . . . .2 Discrete dividend payments . . 5. . . . . . . . . . . . . . . . . .1 Introduction . 28 28 29 30 31 33 33 35 36 37 38 39 39 41 41 41 43 43 44 44 45 46 47 51 51 51 52 52 53 54 55 55 60 61 62 63 64 66 3. . . . . Risk Neutrality . . . . . . . . . . . . . . . .2. . . . . . . . . . .2. . . . . . . . .3.5 Boundary condition . . . . . . . . . . . 4. .3 Finite difference methods (for the modiﬁed BS eq. . . . 5. . . . 3. . . .1 Options on dividendpaying assets . . . . σ. . . . .2 Binomial method is a forward Euler ﬁnite difference method 5. . . . . . . . . . . . . . . . . . . .1 Strategies involving a single option and stock .1. . . . . . . . . . . .2 Binomial method for option . . . . .6 3. . 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5. . Trading strategy involving options . . .) . . . . . . . . . . .1 Monte Carlo method . .3. . . . . . . . . . . . . . . . . . . . . . .2 Warrants .5. . . . . . . . . . . . . 3. . . . . . . . . . . .7. . .1 American put option . . . . . . . . . . . . . . CONTENTS . . . . . . . . . 4. . . . . . . . . . . . . 4. .1 Direct methods . . . . . . . . . . . . . . . . . . . .1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5. . . . . . . . 3. . . . . .4 Convergence . . . .3 Stability . .2 Further reduction . . . . . . . . . . . . . . . . . . . . . . . . . 4.4. . . . . . .2 Binomial Methods . . . . . . . . 5. . 5. . .4 Exact solution for the BS equation for European options . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Bear spreads . . . . . . . .2 American options as a free boundary value problem . µ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4.3 BlackScholes formula . . . . . . . . . . . . . . . . . . . . . . 3. . . . . . .4 3.3. . . 5.2 Bull spreads . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Futures and futures options . . . . . . 4. . .1 Constant dividend yield . . . . . . . 69 6. . . . .4 Converting the BS equation to ﬁnite domain . . . . . .
. . . . . . .3 Average strike options . . . . 6. .3 Chooser options . . . . . 7. . . . . .3. . . . . . . . .1 Introduction . . . . . . . . . .1 American call option with dividend paying asset 6. . . . . .2 Compounds . Numerical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. . . . . . . . . .1 Projective method for American put . . . . . . . . A. . .3.2. . . . .4 Lookback Option . . . . .3 Stochastic differential equation A. . . . .5. . . . . 6. . . . . . . . . . . . . . . . . . . . . . .3 Putcall parity for average strike option . . . . . . . . . . . . . . .2 General Method . . . . . . . .3 6. . . . 9. . . . . . . . . . . . . . . . . . . . .5. . . . . . . . . . . . . . . . . . . . . . . . 1 72 72 74 74 75 76 76 76 78 79 79 80 80 81 81 82 83 85 85 85 86 86 87 87 88 89 90 91 91 91 91 93 93 94 97 97 99 100 101 6. . . . . . .4. . .2 downandin(knockin) option 7. . . . . . . . . .5 7 Exotic Options 7. . 8. . . . . . . . . . . . 8. . . . .2 Stochastic bond model . . . . . . . . . . . . . . . . . . . . . . 7. . . . . . . . . . . . . . . . . . . . . . . . . .5 Asian options and lookback options . . . . . . . . . . . . .1 Brownian motion . . . . . . . . . . . . . . . . . . .2 American call options . . . 7. 6. . . . . . . . . .1 Binaries . . . . . .1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A Basic theory of stochastic calculus A. . . . . . . . . . . . .1 downandout call(knockout) .1 A functional approach for interest rate model 9. . . . . . . . .4. . . . . . . . .2 Projective method for American call . . Converting American option to a ﬁxed domain problem . . . . . . . .1 Deterministic bond model . . . . . . . . . . . 8. . . 8. . . . . . . . . .4. . .2 Lookback put option with American exercise feature 9 Bonds and Interest Rate Derivatives 9. . . . . . . . . . . . . . . . . . . . . . . . . 8. . . . . . . . . .4 Barrier option . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. . . .3 Implicit method . . . . .2 American call option on a dividendpaying asset American option as a linear complementary problem . . . .4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8. . . . . . . 8. . . .4. . . . . . . . . . .3. . . 8 PathDependent Options 8. . . . .1 Bond Models . . . . . . . . . . . . . . . . . 9. . . . . . . . . . 8. 7.3 Convertible Bonds . . . . . . . . . . . . . . 9. . . . . . . . . . . . .2. . . .4 6. . . . . 7. . . .CONTENTS 6. . . . . . . . . . . . . . . . . . .2 American put option . . . . . . . . . . . . . . . . . . . . .1 European calls . . . . . . . . . . . . . . . . . . . . . . . 6. . . . . . .1 A lookback put with European exercise feature .2 Stochastic integral . . . . . . . . . . . .1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 Diffusion process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Interest models . . . . . . . A. . . 9. . . . . . . . . . . . . . . . . . . .
2 CONTENTS .
These are called ﬁnancial derivatives. In futures or options. • the certain price—delivery price.Chapter 1 Introduction 1. • the person who holds the contract is called in long position.2 Financial Derivatives 1. Forwards contract: A forward contract is an agreement which allows the holder of the contract to buy or sell a certain asset at or by a certain day at a certain price. • bond markets. except. The ﬁnancial market provides a link between saving and investment. gold). Futures (futures contracts): A future contract. • the certain day—maturity or expiration date.1 Financial Markets A society improves its welfare through investment. 1 . • futures and options markets. • the person who write the contract (has the asset) is called in short position. like a forward contract. assets are traded in. In ﬁnancial markets. There are many kinds of ﬁnancial markets: • stock markets. Here. • currency markets. • it is normally traded in an exchange. 1. Savers can earn high returns from their saving and borrowers can execute their investment plans to earn future proﬁts. wheat. 2. foreign exchange markets. more complex contracts than simple buy/sell trades have been introduced. • commodity markets (oil.
stock indices. The price is called the exercise price (or strike price). while the writer is called in short position. . quality.2 CHAPTER 1. An investor buys 100 European call options on IBM stock with strike price $140. • American options : Options can be exercised any time up to the maturity date. • it is a margin trading (certain minimal amount of money should be maintained in a margin account). A call (put) option is a contract between two parties. foreign currencies. then he should not exercise his call contracts. The premium is 5 × 100 = 500. daily price movement.). 3. INTRODUCTION • it has standard features (including contract size. ST > E. in which the holder has the right to buy (sell) and the writer has the obligation to sell (buy) an asset at certain time in the future at a certain price. The holder is called in long position. 5 (the price of one call option). The payoff is 0. If ST ≤ T . 1. The payoff is 100 × (ST − E) = 100 × (146 − 140) = 600. If at time T . then he should exercise this option. the price of call option • r bank interest rate 1. stocks. price quotes.3 Examples Notation • t current time • T maturity date • S current asset price • ST asset price at time T • E strike price • c premium. 2 months. delivery arrangement. etc. 138. Suppose E St T c = = = = 140. Options: There are two kinds of options — call options and put options. There are two kinds of exercise features: • European options : Options can only be exercised at the maturity date. position limit. or future contracts. Hence. he earns $100. The underlying assets of an option can be commodity. • clearinghouse.
If ST = 230. 2. PAYOFF FUNCTIONS 3 The payoff function for a call option is Λ = max{ST − E. we should exercise this call option. 2 2 Ignoring the interest of bank. and the payoff is 0. The payoff function of a future is Λ = ST − E. then a reasonable price for this call option should be 10. Thus. expiration is T = 4/14/96 . the strike price E = 250 for some stock. If ST = 270 at expiration. Which is only ±8% of the original investment. which is smaller than the strike price. Suppose the share take 230 or 270 with equal probability. if the investor had instead purchased the share for 250 at t. 1. This means that the proﬁts is 100% (He paid 10 for the option).1. Then the expected proﬁt is 1 1 × 0 + × 20 = 10. The payoff Λ = 270 − 250 = 20. Λ Λ K ST K ST future (long) future (short) . we should give up our option.4. then the corresponding proﬁt or loss at T is ±20. the payoff of a future or an option is the follows.4 Payoff functions At the expiration day. The loss is also 100%. One needs to pay premium (ct ) to buy the options. then the net proﬁt= 20 − 10 = 10. If ST = 270. option is of high risk and with high return. On the other hand. 0}. Suppose today is t = 8/22/95. 1. If ST = 230 the loss is 10 for the premium. Thus the net proﬁt from buying this call is Λ − ct er(T −t) . then buy the share for 250. and sell it in the market immediately for 270.
4
CHAPTER 1. INTRODUCTION Payoff of a future, long position (left) and short position (right) 2. The payoff function of a call option is Λ = max{ST − E, 0}.
Λ Λ
K S
T
K S
T
call option (long):max{ST−K,0}
call option (short):−max{ST−K,0}=min{K−ST,0}
Payoff of a call, long position (left) and short position (right) 3. The payoff function of a put option is Λ = max{E − ST , 0}.
Λ Λ
K ST
K ST
put option (long):max{K−ST,0}
put option (short):−max{K−ST,0}=min{ST−K,0}
Payoff of a put, long position (left) and short position (right) 4. Below is a portion of a call option copied from the Financial Times. the current time t the expiration T T −t St = Feb 3 = end of Feb, ≈ 10 days = 2872
1.5. OTHER KINDS OF OPTIONS E= c=
250 c
5 2900, 2950, 3000 24, 9, 3
2650, 2700, 2750, 2800, 2850, 233, 183, 135, 89, 50,
200
150
100
50
0 K −50 2600
2650
2700
2750
2800
2850
2900
2950
3000
3050
The FTSE index call option values versus exercise price.
1.5
Other kinds of options
• Barrier option: The option only exists when the underlying asset price is in some prescribed value before expiry. • Asian option: It is a contract giving the holder the right to buy or sell an asset for its average price over some prescribed period. • Lookback option: The payoff depends not only on the asset price at expiry but also its maximum or minimum over some period price to expiry. For example, Λ = max{J − S0 , 0}, J = max0≤τ ≤T S(τ ).
1.6 Types of traders
1. Speculators (high risk, high rewards) 2. Hedgers (to make the outcomes more certain) 3. Arbitrageurs (Working on more than one markets, p12, p13, p14, Hull).
6
CHAPTER 1. INTRODUCTION
1.7
Basic assumption
Arbitrage opportunities cannot last for long. Only small arbitrage opportunities are observed in ﬁnancial markets. Our arguments concerning future prices and option prices will be based on the assumption that “there is no arbitrage opportunities”.
7 . This means the future price of the asset only depends on its current value and does not depends on its value one month ago.1 Efﬁcient market hypothesis The asset prices move randomly because of the following efﬁcient market hypothesis: 1. This contradicts to the hypothesis of no arbitrage opportunities. We model the asset price by Sn+1 = Sn u d with probability p with probability 1 − p. We will show that the continuous model is the continuous limit of the discrete model. We shall ﬁnd this transition probability later. technical analysis could make aboveaverage return by interpreting chart of the past history of the asset price. n ∈ N. Market reponds immediately to any new information about an asset.2 The asset price model We shall introduce a discrete model and a continuous model. 0 < d < 1 < u. the probability that the asset price is S at time step n with initial price S0 . which does not hold any future information. or one year ago.1 The discrete asset price model The time is discrete in this model. The information we are looking for is the following transition probability P (Sn = SS0 ). there is very little evidence that they are able to do so. 2.Chapter 2 Asset Price Model 2.2. 2. (2.1) Here. The past history is fully reﬂected in the present price. If this were not true. 2. In fact. Let us denote the asset price at time step n by Sn . The time sequence is n∆t.
We may think µ is a constant during the life of an option.3 Random walk To study the discrete asset price model. m ∈ Z. It is modeled by a Brownian motion σdz. the integral b P(S(t) = SS(0) = S0 ) dS a is the probability that the asset price S(t) lies in (a. Consider a particle moving randomly on a uniformly distributed grid points on the real lines. the particle moves to its left adjacent grid point or right adjacent grid point with equal probability. 2. S which is called the return. Suppose the grid points are located at m∆x.2 The continuous asset price model Let us denote the asset price at time t by S(t). Or equivalently. S (2. The overall asset price model is then given by dS = µdt + σdz.2. dS S can be decomposed into two parts: one is • Deterministic part: This can be modeled by dS = µdt. the σ is the order of ﬂuctuations or the variance of the return and is called the volatility.2) We shall look for the transition probability density function P(S(t) = SS(0) = S0 ). • Random part: this part is a random change in response to external effects. The quantity dz is sampled from a normal distribution which we shall discuss below. µ is a measure of the growth rate of the asset. In each time step. Suppose the particle is located at 0 initially. the other is random. Let w(m∆x. Let Zn denote the location of this particle at time step n. The change deterministic. such as unexpected news. The meaningful quantity for the change of an asset price is its relative change dS .8 CHAPTER 2. we study a simple model—the random walk in one dimension—ﬁrst. b) at time t and is S0 initially. n∆t) denotes for the probability that the particle is . ASSET PRICE MODEL 2. S Here.
(n + 1)∆t) = w((m − 1)∆. the particle moves p times toward right and n − p time toward left. and if m + n is odd. The moments < mk >. n∆t) is a probability density function. There is a onetoone correspondence between {p  0 ≤ p ≤ n} and {m  − n ≤ m ≤ n. n∆t) ≥ 0. 2 From m = 2p − n. They can be found by computing < pk >. Notice also that the n! number of choices in n steps that the particle moves p times toward right is n := (n−p)!p! . w(m∆x. n∆t) = P (Zn = m∆xZ0 = 0). n∆t). n∆t) + w((m + 1)∆. ∆x with probability 1 2 Zn+1 − Zn = −∆x with probability 1 2 1 1 w(m∆x. n p 1 ( 2 )n .3) 2 2 Suppose in n times.3. p When p = 1 (n + m). n∆t) = 0. n∆t) = 1.2. 2. Namely. By our rule. . if m + n is even. we have 2 w(m∆x. 1. we have < m >= 2 < p > −n = 0. m w(m∆x. k ∈ N are particularly important. m + n is even}. n∆t). RANDOM WALK 9 located at the m∆x cell at the time n∆t. we deﬁne its expectation value at n∆t by < f (m) >:= m f (m)w(m∆x. Given any function f (m). Then 1 m = p − (n − p) = 2p − n or p = (n + m). n p p 1 2 n p = n . The ﬁrst moment < m > is called the mean. That is. (2. We may check that w(m∆x. 2 Notice that m is even(odd). when n is even(odd). which in turn can be computed through the help of the following generating function: G(u) := p up 1+u 2 1 2 n n n p = Hence < p >= G (1) = . w(m∆x. while the second moment of the variation from mean < (m− < m >)2 > is called the variance.
Hence.3). otherwise. if the price goes up p times. < p2 >= 2 4 n2 + n and 4 4 < m2 >= 4 < p2 > −4n < p > +n2 = n.4 The solution of the discrete asset price model Sn+1 = Sn u d with probability 1 2 with probability 1 . we have < m2 >= 4 < p2 > −4n < p > +n2 . we obtain G (1) = n(n−1) . The mean of this random walk is < m >= 0. Exercise 1.5 2. One can also ﬁnd the transition probability w by solving the difference equation (2. 2.4) 2. In n movements of the asset price. 2 Let us consider the case for simplicity. Since there are n such choices. from m = 2p − n. then the price at time step n∆t is Sn = S0 up dn−p . while its variance is < (m− < m >)2 >= n. n G (1) = p=0 p(p − 1) n p 1 2 n = < p2 > − < p > n = < p2 > − 2 On the other hand. we then obtain the p transition probability of the asset: P (Sn = SS0 ) = 0 n p 1 n 2 if S = S0 up dn−p . mean and variance for the case Zn+1 − Zn = ∆x −∆x with probability p with probability 1 − p n 2.5. With the help of the generating function.1 The Brownian motion The deﬁnition of a Brownian motion The deﬁnition of the (standard) Brownian motion z(t) is the following: . from G(u) = 1+u . (2. Find the transition probability.10 CHAPTER 2. ASSET PRICE MODEL To compute the second moment < m2 >.
2.5. THE BROWNIAN MOTION 1. ∀t, z(t) is a random variable. 2. The increment z(t + s) − z(t), z(t) − z(t − u), u > 0, s > 0 are independent. 3. z(t) is continuous in t.
11
4. ∀s > 0, zt+s − zt is normally distributed with mean zero and variance s, i.e., its −x2 1 probability density is N (0, s)(i.e., √2πs e 2s ).
2.5.2
The Brownian motion as a limit of random walk
We may realize the Brownian motion as the limit of the random walk in the previous sec2 tion. Namely, Zn → z(t) as n → ∞ with m∆x → x, n∆t → t and (∆x) = σ ﬁxed. This ∆t can be proved by the Stirling formula: √ 1 n! ≈ 2πnn+ 2 e−n . Recall that the probability P (Zn = m∆xZ0 = 0) = n 1 (m + n) 2 1 2
n
.
Using the Stirling formula, we have for n, p, n − p >> 1, n 1 (m + n) 2 1 2
n
1 n! = ( )n 1 2 ( 2 (n + m))!( 1 (n − m))! 2 √ 1 1 n 2πnn+ 2 e−n ≈ ( ) √ 1√ 1 1 1 2 2π( 1 (n + m)) 2 (n+m)+ 2 2π( 1 (n − m)) 2 (n−m)+ 2 2 2 1 1 2 1 m 1 m 1 = ( ) 2 (1 + )− 2 (n+m)− 2 (1 − )− 2 (n−m)− 2 πn n n 2 1 m 2 −1n 2 (1 − ( 2 ≈ ( ) )) πn n 2 2 1 m 2 ( n )2 − m 2n = ( ) 2 (1 − ( ) ) m πn n m2 2 1 ≈ ( ) 2 exp(− ) πn 2n
As m∆ → x, n∆t → t, (∆x)2 /∆t = σ ﬁxed, we obtain P (Zn = m∆xZ0 = 0)/2∆x ≈ ( 2 1 1 − m2 )2 e 2n πn 2∆x √ 2 1 − (m∆x) 1 ∆t = ( ) 2 e n∆tσ2 · 2πn∆t 2(∆x) x2 1 → √ e− 2σ2 t 2t 2πσ
This means that Zn → z(t).
12
CHAPTER 2. ASSET PRICE MODEL
2.5.3
Properties of Brownian motion
P(z(t) = xz(0) = 0) = √ 1 − x2 e 2t . 2πt
By deﬁnition
We can check 1. < z(t) >= 0 2. < z(t)2 >= t 3. Independence of disjoint increments
∞
P(z(t) = xz(0) = 0) =
−∞
P(z(t) = xz(s) = y) P(z(s) = yz(0) = 0) dy. (2.5)
In particular, let us deﬁne an inﬁnitesimal increment dz = z(t + dt) − z(t) We have 1. < dz >= 0 2. < (dz)2 >= dt In fact we have more, we may think dz = √ dt (2.6)
where is a random variable with standard Gaussian distribution N (0, 1) (i.e. mean is 0 and variance is 1). And we have (dz)2 = dt with probability 1. Exercise 1. Check (2.5). (2.7)
2.6
Itˆ ’s formula o
In this section, we shall study differential equations which consist of deterministic part: x = b(x), and stochastic part σ z(t). Here, z(t) is the Brownian motion. We call such an ˙ ˙ equation a stochastic differential equation and expressed as dx(t) = b(x(t))dt + σ(x(t))dz(t). An important lemma for ﬁnding their solution is the following Itˆ ’s lemma. o (2.8)
2.7. THE SOLUTION OF THE CONTINUOUS ASSET PRICE MODEL
13
Lemma 2.1 Suppose x(t) satisﬁes the stochastic differential equation (2.8), and f (x, t) is a smooth function. Then f (x(t), t) satisﬁes the following stochastic differential equation: df = 1 ft + bfx + σ 2 fxx dt + σfx dz 2 (2.9)
Proof. This is not a proof, rather an intuition why (2.9) is true. According to the Taylor expansion, 1 1 df = ft dt + fx dx + ftt (dt)2 + fxt dx dt + fxx (dx)2 + · · · . 2 2 √ Plug (2.8) into this equation. We recall that dz = dt, where is a random variable with standard Gaussian distribution N (0, 1). In the Taylor expansion of df (x(t), t), the terms (dt)2 , dt · dz are relative unimportant as comparing with the dt term and dz term. Using (2.8) and noting (dz)2 = dt with probability 1, we obtain (2.9). A simple application of Itˆ ’s lemma is to ﬁnd the transition probability density function for o the s.d.e. dx = adt + σdz where a and σ are constants. By letting y = x − at, from Itˆ ’s lemma, y satisﬁes dy = σdz. o Thus, the transition probability density function for y is P(y(t) = yy(0) = y0 ) = √ 1 2πσ 2 t 1 2πσ 2 t e−(y−y0 )
2 /2σ 2 t
.
Or equivalently, the transition probability density function for x is P(x(t) = xx(0) = x0 ) = √ e−(x−at−x0 )
2 /2σ 2 t
.
2.7
The solution of the continuous asset price model
In this section, we want to ﬁnd the transition probability density function for the continuous asset price model: dS = µS dt + σS dz. (2.10) with initial data S(0) = S0 . We apply Itˆ ’s lemma with x = f (S) = log S. Then x satisﬁes o the s.d.e. σ2 dx = µ − dt + σ dz, 2 and x(0) = x0 := log S0 . From the discussion of the previous section, we obtain P(x(t) = xx(0) = x0 ) = √ From P(x(t) = xx(0) = x0 )dx = P(x(t) = xx(0) = x0 )dS/S = P(S(t) = SS(0) = S0 )dS, 1 2πσ 2 t e−(x−x0 −(µ−
σ2 )t)2 /2σt 2
.
The parameters in (2. Exercise 1. n∆tS((n − 1)∆t) = Sn−1 )dS S √ 1 2πσ 2 ∆tS e −(log S −(µ− 1 σ 2 )∆t)2 /2σ 2 ∆t Sn−1 2 dS . The parameters in (2. Find the mean and variance of the lognormal distribution.11) p S S 0 2. σ) and (u. 1 2πσ 2 tS e −(log 2 S −(µ− σ )t)2 /2σ 2 t S0 2 .10) is the limit of the discrete model (2.1). For the continuous model. we compute its mean under the condition S((n − 1)∆t) = Sn−1 .8 Continuous model as a limit of the discrete model We want to show that the continuous model (2. Both models should have the same mean and variance. ASSET PRICE MODEL we obtain that the transition probability density function for S(t) is P(S(t) = SS(0) = S0 ) = √ This is called the lognormal distribution. (2.14 CHAPTER 2. First. We may assume p = 1/2. d and p. we relate (µ.1) are u.10) are µ and σ. Then E(S(n∆t)Sn−1 ) = = SP(S. d).
8. Proof. 2 2 In order to have the same means and variances in one time step in both models. dx √ x − 2σ 2 ∆t 2σ 2 ∆t 2 ) 2 = eµ∆t Sn−1 . Theorem 2. Then P (S0 up dn−p S0 )/2∆x −→ P(S(t) = SS(0) = S0 ) as n∆t → t.12) (2. σ). Let us choose a ∆t and a ∆x with (∆x)2 /∆t = σ 2 . x0 = log S0 .12) and (2.1 Let us ﬁx (µ. we have used the changeofvariable: x = log Sn−1 .13). d) and (µ.13) − 1). 2πσ 2 ∆t S Here. n → ∞ and S0 up dn−p → S. . we should require 1 2 1 2 2 u + d = e(2µ+σ )∆t 2 2 1 1 u + d = eµ∆t . On the other hand. 2 2 Or u = eµ∆t (1 + d = eµ∆t (1 − These relate (u. Then log S0 up dn−p = x0 + p log u + (n − p) log d.2. Let us deﬁne x = log S. CONTINUOUS MODEL AS A LIMIT OF THE DISCRETE MODEL = Sn−1 = Sn−1 e √ µ∆t 15 1 2πσ 2 ∆t √ 1 e−(x−(µ− 2 σ e −( √ 1 2 )∆t)2 /2σ 2 ∆t ex dx. e σ 2 ∆t (2. the mean and the second moment for the discrete model in one time step ∆t are 1 1 u + d Sn−1 2 2 1 2 1 2 2 u + d Sn−1 . d) by (2. For the second moment for the continuous model. σ). Deﬁne (u. ∆tSn−1 ) dS 2 )∆t = e(2µ+σ 2 Sn−1 . Thus. what we want to show is equivalent to P (x = x0 + p log u + (n − p)x0 )/2∆x → P(x(t) = xx(0) = x0 ) eσ2 ∆t − 1). we have E(S(n∆t)2 Sn−1 ) = S 2 P(S.
we choose time step ∆t = 0.01. σ is 0.d.9 Simulation of asset price model dS = µdt + σdz S S(0) = S0 Typically. we deﬁne m = 2p − n. and sample a number ξ from the normalized Gaussian distribution N (0. d Hence 1 1 u n log ud + m log( ) ≈ n(µ − 2 2 d = n(µ − Deﬁne ∆x such that (∆x)2 ∆t 2 2 √ 1 2 σ )∆t + mσ ∆t 2 1 2 σ )∆t + m∆x. Then the density is /2∆x ≈ ( σ2 2 1 − m2 1 2 2 ) 2 e 2n −→ √ e−(x−x0 −(µ− 2 )t) /2σ t 2t nπ 2πσ 2.2 × 0. Then 1 p log u + (n − p) log d = n∆t(µ − σ 2 ) + m∆x. for instance.5 × 0. n − p = 2 (n − m).13).1 . ωN .12) and (2. 2 = σ 2 .01 + 0. 2πσ 2 t 1 To show this. In each path. We obtain Sk+1 from Sk by discretizing the s. √ √ u = 1 + 2σ ∆t + σ 2 ∆t ≈ e2σ ∆t . 2 2 d e−(x−x0 −(µ− σ2 )t)2 /2σ 2 t 2 From (2.40 for a stock.16. · · · . µ = 0. Hence 2 p log u + (n − p) log d = 1 1 (n + m) log u + (n − m) log d 2 2 1 1 u = n log(ud) + m log( ). 1): √ Sk+1 − Sk = µ∆t + σξ ∆t Sk = 0. 2 Recall that the probability that the price moves up p times is n p 1 2 n n p 1 ( 2 )n .16 where P(x(t) = xx(0) = x0 ) = √ 1 CHAPTER 2. ASSET PRICE MODEL . u · d = e2µ∆t (2 − eσ ∆t ) ≈ e2µ∆t · e−σ ∆t . Then p = 1 (n + m).e.16 × 0.20 ∼ 0. To simulate the model We perform N sample paths ω1 .
2. ξ = 0.5 is the sampled number. SIMULATION OF ASSET PRICE MODEL Here. Then the transition probability density function b 17 P(S(t) = SS(0) = S0 ) dS ≈ #{ω  a ≤ Sn (ω) ≤ b}/N a .9.
18 CHAPTER 2. ASSET PRICE MODEL .
bank. we shall also assume 1.Chapter 3 BlackScholes Analysis 3. 4. All trading proﬁts are subject to the same tax rate.g. The fundamental hypothesis of their analysis is that ”there is no arbitrage opportunities in ﬁnancial markets”. government bond. 3.1 The hypothesis of noarbitrageopportunities The option pricing theory was introduced by Black and Scholes. Borrowing or lending at such riskless interest rate is always possible. There is no transaction costs. For simplicity. There exists a riskfree investment that gives a guaranteed return with interest rate r. We will use the following notations: S E T t µ σ ST r c C p P Λ current asset price exercise price expiry time current time growth rate of an asset volatility of an asset asset price at T riskfree interest rate value of European call option value of American call option value of European put option value of American put option the payoff function 19 . ( e.) 2.
This contradicts to our hypothesis. BLACKSCHOLES ANALYSIS 3. Lemma 3. t) = Λ(t) P (St .3 Suppose I(t) and J(t) are two portfolios containing no American options such that I(T ) ≤ J(T ). Recall that Λ(t) = max(St − E.2) .20 CHAPTER 3. 0} ≤ c ≤ S max{Ee−r(T −t) − S.2 3. The cost to us is ST − E. If c(ST .1) (3. If the person who buy the call does not claim. exercise it immediately. we have p(ST .2 European options Lemma 3. Similarly. 3.2. 0) Λ(t) = max(E − St . If he does exercise his call. By doing so. T ) = Λ(T ). we can conclude that I(t) ≤ J(t). there is a chance of arbitrage.3) (3.1 A portfolio is a collection of investments. 0) for call option for put option 2. For instance. T ) = Λ(T ) C(St . then Λ = ST − E > 0 and c < Λ by our assumption. we can short a call and earn c. ∀t ≤ T . a portfolio I = c − ∆S means that we long a call and short ∆ amount of an asset S.1 Basic properties of option prices The relation between payoff and options 1. Again. then we can buy a call on price c. Then under the hypothesis of noarbitrageopportunities. Deﬁnition 2. then we can buy an asset from the market on price ST and sell to that person with price E. if c(ST . then we have net proﬁt c. the net proﬁt we get is c − (ST − E) > 0.2. t) = Λ(t) 3. this is a contradiction. 0} ≤ p ≤ Ee−r(T −t) and the putcall parity p + S = c + Ee−r(T −t) To show these. T ) < Λ. we need the following deﬁnition and lemmae. c(ST . Hence we have an immediate net proﬁt ST − E − c. T ) > Λ. Otherwise. If ST > E. For instance.2 We have the following for European options max{S − Ee−r(T −t) . (3.
0} + E = max{ST . 0} + E = max{ST . The equality holds when E = 0. I(t) = J(t). 0} + ST = max{E. 0} ≤ E = J(T ). Proof of Lemma 3. ∀t ≤ T . Consider I = p + S and J = Ee−r(T −t) .2. I(T ) = c + E = max{ST − E. At time T . Hence. Consider I = c + Ee−r(T −t) and J = S. Now. E} ≥ E = J(T ) . there exists a time t ≤ T such that I(t) > J(t). I(t) ≤ J(t) holds for all t ≤ T . Hence.3.e.. nothing can be exercised before T . This implies I(t) ≥ J(t). 0} ≤ max{ST . since I(T ) ≤ J(T ). E} ≥ ST = J(T ). Hence. J(T ) = p + S = max{E − ST . BASIC PROPERTIES OF OPTION PRICES 21 Proof. he can use J(T ) (what he has) to cover I(T ) (what he shorts) and gains a proﬁt J(T )−I(T ). At time T . I(T ) = max{ST . Consider I = c + Ee−r(T −t) and J = p + S. At time T . 3.2. i. I(t) ≥ J(t). 4. . Let I = p and J = Ee−r(T −t) . At time T . This contradicts to the hypothesis of noarbitrageopportunities. As a corollary. E}. I(t) ≤ J(t). Remark. we can prove the basic properties of European options 15. Since I and J containing no American options. In this case c = S 2. An arbitrageur can buy (long) J(t) and short I(t) and immediately gain a proﬁt I(t)−J(t). 5. 1. we have I(T ) = cT = max{ST − E. ST } Hence. Let I = c and J = S. we have Corollary 2. At time T . 0} = ST = J(T ). Suppose the conclusion is false. I(T ) = max{E − ST . then I(t) = J(t). At T .1 If I(T ) = J(T ). I(T ) = max{ST − E.
An arbitrageur can long J(t) and short I(t) at time t to make proﬁt I(t)−J(t) immediately. To prove these properties. for all t ≤ τ . (ii) The optimal exercise time for American put option is as earlier as possible. Proof. (3. Hence. Remark.22 CHAPTER 3. Suppose we exercise C at some time τ ≤ T . Suppose S = 50. the optimal exercise time for American option is T . Firstly. we show C ≥ c. Suppose I(τ ) ≤ J(τ ) at some τ ≤ T . Example. This implies I(τ ) ≤ J(τ ). I(t) ≤ J(t) for all t ≤ τ . we can buy C and sell c at time τ to make a proﬁt c(τ ) − C(τ ). i. we need the following lemma. Suppose p(t) > P (t).4) .e. we show c ≥ C.4 For American options. he can use J(τ ) to cover I(τ ) with additional proﬁt J(τ ) − I(τ ). BLACKSCHOLES ANALYSIS 3.4. then the investor needs to pay 40 to buy the share. early exercise results C(τ ) + Ee−r(T −τ ) < S(τ ). If not. This is an arbitrage opportunity which is a contradiction. he can instead invest $40 into the bank to earn interest and there is a chance that the stock price may go up.2. Consider two portfolios I = C + Ee−r(T −t) and J = S. At later time τ . Further. By our lemma.3 Basic properties of American options Lemma 3. we conclude c = C. E = 40. Combine this inequality with the inequality of 2) of section 3. This is a contradiction. Then we can make an immediate proﬁt by selling p and buying P . 3. The equality also holds if I(τ ) = J(τ ). Since τ ≤ T arbitrary. The right of C is even more than that of c.5 Let I or J be two portfolios that contain American options. we conclude I(t) ≤ J(t) for all t ≤ T . We earn p − P and gain more right. However. P ≤ E. then c(τ ) > C(τ ) for some time τ ≤ T . Suppose I(t) > J(t) at some t ≤ τ . Proof of Lemma 3. 1. t. and we have P ≥ p. If C is exercised before expiration. then I(τ ) = max{Sτ − E. Lemma 3. Then I(t) ≤ J(t). 0} + Ee−r(T −τ ) and J(τ ) = Sτ . in case the person who owns I exercises his American option. Secondly. 2.2. (iii) The putcall parity for American option: S − E < C − P < S − Ee−r(T −t) As a consequence. we have (i) The optimal exercise time for American call option is T and we have C = c.
In both cases. Put this cash into a bank.2.25 = 32. We consider two portfolios I = P + S and J = Ee−r(T −t) . Therefore.3) and the facts that c = C and P ≥ p. t ≤ τ ≤ T .25. I(t) ≥ J(t).26. r = 10%. J = p + S = 2. the put option will be exercised. This results a cashﬂow: −3 + 2. To show the ﬁrst inequality.25. Suppose S(t) = 31. we can exercise c. Suppose P is exercised at some time τ . This means that we need to buy the share for E to close our short position. Hence. J(τ ) = E −r(T −τ ) . If we exercise P at some time τ .02.02 at time T . Examples.1×0. Therefore. ST > E. 0} + Sτ = E I(τ ) = C(τ ) + Eer(τ −t) = max{Sτ − E. Strategy : long the security in portfolio I and short the security in portfolio J. also we should buy a share for E to close our short position of the stock.25 + 31 = 30. we should exercise it immediately. 4. then I(τ ) = max{E − Sτ . I(τ ) ≥ J(τ ). we see that if we exercise P at t. . 0} + Eer(τ −t) = Eer(τ −t) . we show that if we have a P . Suppose at time T . T − t = 0. BASIC PROPERTIES OF OPTION PRICES 23 Next.25. we have I(t) > J(t). c = 3. In this case I = c + Ee−r(T −t) = 32.25 + 31 = 33. Thus.25 = 31. J(τ ) = max{E − Sτ . then I(T ) = Eer(T −t) is the maximum. p = 2. Consider two portfolios: I = c + Ee−r(T −t) = 3 + 30 × e−0. 2. From lemma. t ≤ τ ≤ T . 1. Then we must have E ≥ Sτ (otherwise. The second inequality follows from the putcall parity (3. Putting this money into bank we will receive Eer(T −τ ) at time T . 0} + Sτ = max{E. the net proﬁt is 31. We ﬁnd J(t) > I(t). we should not exercise our put option).1×0. E = 30. Further. Hence we should exercise P as early as possible.25 year. Hence C + E > P + S. On the other hand.02 − 30 = 1. Sτ } = E. we consider two portfolios: I = C + E and J = P + S. We will get 30.25 J = p + S = 1 + 31 = 32. Consider the same situation but c = 3 and p = 1.25 × e0. we need to buy a share for E to close the short position.3. Suppose ST < E.
Proof. We need to sell the share for E to close our short position for c. we must have that either c or p will be exercised. We consider two portfolios: I = c + D + Ee−r(T −t) . and the putcall parity: c + Ee−r(T −t) = p + S − D. If ST > E. I(T ) = max{ST − E.8) (3. The maximal time value is E − Ee−r(T −t) .2.5) −S + D + Ee−r(T −t) < p ≤ Ee−r(T −t) . we sell the share for E. These are payments to shareholders out of the proﬁts made by the company. Thus.73 to the bank at time T . the strike price for options was reduced on the exdividend day by the amount of the dividend. E} + D J(T ) = ST + D. Then at time T . then c is exercised.73 = 0.1×0. We can ﬁnance it from the bank. For the American options.9) (3. and we need to pay 29 × e0. the net proﬁt is 30 − 29.4 Dividend Case Many stocks pay out dividends. we need an initial investment 31 + 1.6 Suppose a dividend D will be paid during the life of an option. Thus. If ST < E. To long J. If a company declared a cash dividend. to short c. Lemma 3. 3. the strike prices fall as the dividends being paid. J = S. we gain 3. the stock price. 0} + D + E = max{ST . Since the company’s wealth does not change after paying the dividends. (3.24 CHAPTER 3. Then we have for European option S − D − Ee−r(T −t) < c ≤ S (3. Remark. Now.27. we have (i) S − D − E < C − P < S − Ee−r(T −t) provided the dividend is paid before exercising the put option. P − p is called the time value of a put. at T . BLACKSCHOLES ANALYSIS and we see that J is cheaper. Strategy: We long J and short I. we sell the share for E. That is. or (ii) S − E < C − P < S − Ee−r(T −t) if the put is exercised before the dividend being paid.6) .25 = 29.7) (3. we exercise p. the net investment is 31 + 1 − 3 = 29 initially. In both cases.
At time T . because the dividend will cause the stock price to jump down. then we should consider I = C + E and J = P + S. This yields the putcall parity for all time. . When there is no dividend. we know that CD < C.3. This proves c ≥ S − D − Ee−r(T −t) .2. we have J(τ ) ≤ I(τ ). Hence. BASIC PROPERTIES OF OPTION PRICES Hence I(T ) ≥ J(T ). We should exercise it immediately prior to an exdividend date. If the put option is exercised before the dividend being paid. J = P + S. making the option less attractive. I = J = max{ST . we have shown that C − P < S − Ee−r(T −t) . p increases by an amount D. J(τ ) = E. That is. At τ . we have p ≥ D + Ee−r(T −t) − S. PD > P 25 For the American call option. Again. CD − PD < C − P < S − Ee−r(T −t) . We have J(τ ) ≤ I(τ ). This yields I(t) ≥ J(t) for all t ≤ T . then Sτ < E and I(τ ) = D + Eer(τ −t) . E} + D. Hence J(t) ≤ I(t) for all t ≤ τ . we consider I = C + D + E. When there is dividend payment. Similarly. we consider I = c + D + Ee−r(T −t) J = S + p. Hence J(t) ≤ I(t) for all t ≤ τ . c is reduced by an amount D. If we exercise P at τ ≤ T . For the putcall parity. J(τ ) = E + D. For the American put option. I(τ ) = Eer(τ −t) . In other word. we should not exercise it early.
From Itˆ s lemma o dΠ = σS ∂V − ∆ dz ∂S ∂V 1 2 2 ∂ 2V ∂V + µS − µ∆S + + σ S ∂S ∂t 2 ∂S 2 dt. From the hypothesis of no arbitrage opportunities. t) denotes for the price of an option. (5) Shorting selling is permitted.e. The randomness of V (S(t). but in opposite position in order to cancel out the randomness. let the portfolio be Π = V − ∆S. we can eliminate the randomness by choosing ∆= ∂V ∂S at the starting time of each time step. Here ∆ is held ﬁxed during the time step. should be the same as Π being invested in a riskless bank with interest rate r. Then this portfolio becomes deterministic. (3.1 The BlackScholes Equation BlackScholes Equation The fundamental hypothesis of the BlackScholes analysis is that there is no arbitrage opportunities. Besides. In one time step. Thus. Our purpose is to value the price of an option (call or put). i. (2) There exists a riskfree interest rate r. BLACKSCHOLES ANALYSIS 3. To be more precise. we consider a portfolio which contains only S and V . (3) No transaction costs. Π . the change of the portfolio is dΠ = dV − ∆dS. Π is wholly deterministic.26 CHAPTER 3. the return. (4) No dividend paid. t) would be fully correlated to that S(t). we make the following additional assumptions: (1) The asset price follows the lognormal distribution. The resulting portfolio dΠ = 1 ∂2V ∂V + σ2S 2 2 ∂t 2 ∂S dt dΠ .3 3. dΠ = rdt.3.10) Now. Let V (S.
3. at S = ∞: p(S. Remark. T ) = max{S − E. the call option must be exercised. T ) → 0 as S → ∞. and the price of the option must be closed to S − Ee−r(T −t) . t) → 0. τ ) = Ee−r(T −τ ) . Its lefthand side is the return from the hedged portfolio. In general. Thus. (ii) On S = 0: p(0. the ﬁnal condition is V (S. . • Boundary conditions: (i) On S = 0: c(0. THE BLACKSCHOLES EQUATION 27 Otherwise. T ) = Λ(S). t) ∼ S − Ee−r(T −t) . Notice that the BlackScholes equation is invariant under the change of variable S → λS. T ) = max{E − S. 3. (iii) For call option. the put option is unlikely to be exercised. t) = 0. This means that you wouldn’t want to buy a right whose underlying asset costs nothing. there would be either a net loss or an arbitrage opportunity. (iv) For put option. (3.3. 0}. Since S → ∞.2 Boundary and Final condition for European options • Final condition: c(S. This follows from the putcall parity and c(0.E. while its righthand side is the return from bank deposit. Hence p(S. Note that the equation is independent of µ. Hence we must have ∂V 1 ∂ 2V ∂V rΠdt = µS − µ∆S + + σ 2 S 2 2 dt ∂S ∂t 2 ∂S 2 ∂V 1 ∂ V = + σ 2 S 2 2 dt. ∂t 2 ∂S or ∂V 1 ∂2V ∂V .) for option pricing.3. as S → ∞ As S → ∞. as S → ∞. at S = ∞: c(S. 0} is zero.11) + σ2S 2 2 = r V − S ∂t 2 ∂S ∂S This is the BlackScholes partial differential equation (P.D. τ ) = 0. where Λ is the payoff function. ∀t ≤ τ ≤ T. 0} p(S. the payoff function Λ = max{E − S.
13) . The domain S ∈ (0. ∞) becomes x ∈ (−∞.D. let us reverse the time by letting τ = T − t. Then the BlackScholes equation becomes ∂V 1 ∂ 2V 1 = σ2 2 + r − σ2 ∂τ 2 ∂x 2 ∂V − rV. Then v satisﬁes 1 ∂ 2v 1 ∂v = σ2 2 + r − σ2 ∂τ 2 ∂x 2 ∂v − rv. BLACKSCHOLES ANALYSIS 3. That is. S or equivalently. ∂t 2 ∂S ∂S (3. ∂x ∂S ∂V ∂V =S .E. We therefore make the following changeofvariable: dS dx = . Notice that this equation is invariant under S → λS. ∂x ∂S We can also make V dimensionless by setting v = V /E.12) This P.1 Exact solution for the BS equation for European options Reduction to parabolic equation with constant coefﬁcients Let us recall the BlackScholes equation ∂V 1 ∂2V ∂V + σ 2 S 2 2 + rS − rV = 0. ∂x ∂S ∂S ∂ ∂V S ∂x ∂S ∂S ∂V ∂S ∂ 2 V +S ∂x ∂S ∂x ∂S 2 ∂V ∂ 2V S + S2 2 ∂S ∂S ∂V ∂ 2V + S2 2 . it is homogeneous in S with degree 0.28 CHAPTER 3. is a parabolic equation with variable coefﬁcients. S x = log E The fraction S/E makes x dimensionless.4 3.4. ∞) and ∂V ∂x ∂ 2V ∂x2 = = = = = Next. ∂x (3.
vt + avx is call the advection part of (3.14) The part. ex − e−rτ as x → ∞ 0 as x → ∞. and the tern vxx is called the diffusion term. 0} max{1 − ex . we have absorbed the diffusion 1 coefﬁcient 2 σ 2 in to time by setting t = τ /( 1 σ 2 ). τ ) p(−∞.4. it is of the following form: vt + avx + bv = vxx (3. EXACT SOLUTION FOR THE BS EQUATION FOR EUROPEAN OPTIONS 29 The initial and boundary conditions for v become c(x. 0) p(x.) The advection part: vt + avx = (∂t + a∂x ) v is a direction derivative along the curve (called characteristic curve) dx = a. e−rτ . dt This suggests the following changeofvariable: y = x − at s = t. 3. τ ) p(x. τ ) Our goal is to solve v for 0 ≤ τ ≤ T . Thus. The term bv is called the source term. the equation vs + bv suggests that v behaves like ebs along the characteristic curves.3). . 2 The new t here is different from the t we used before.2 Further reduction In investigating the equation (5. 0} 0. it is natural to make the following changeofvariable v = ebs u. 0) c(−∞.4. Then the direction derivative become ∂s = ∂t + a∂x ∂y = ∂x Hence the equation is reduced to vs + bv = vyy . (We somewhat abuse the notation here. τ ) c(x.14).3. Next. Here . = = = = → → max{ex − 1.
S.3). ¯ Λ(x) = Λ(Eex )/E.4.16) We may express it as V (S. .18) This is the transition probability density of an asset price model with growth rate r and volatility σ. P(S . The changeofvariables above gives 1 s = τ /( σ 2 ) 2 y = x − as 1 a = 1 − r/( σ 2 ) 2 1 2 b = r/( σ ) 2 u = erτ v Then v(x. T.15) In terms of the original variables. BLACKSCHOLES ANALYSIS us = uyy . t)Λ(S ) dS (log( S )−(r− 1 σ 2 )(T −t))2 2 S 2σ 2 (T −t) (3. S. we have the following BlackScholes formula: V (S. A simple derivation of this solution is given in the Appendix of this chapter. Let us denote the rescaled payoff func¯ tion by Λ(x). (3. That is. t) = e −r(T −t) 1 2πσ 2 (T − t)S e − (log( S )−(r− 1 σ 2 )(T −t))2 2 S 2σ 2 (T −t) Λ(S ) dS (3. τ ) = e−rτ √ 1 2πσ 2 τ e− (x−z(r− 1 σ 2 )τ )2 2 2σ 2 τ ¯ Λ(z) dz (3. This is the standard heat equation. Its solution can be expressed as u(y. T. 3.3 BlackScholes formula Lert us return to the BlackScholes equation (5. In other words. We shall come back to this point later.30 Then the equation is reduced to CHAPTER 3. t) = e−r(T −t) Here. V is the present value of the expectation of the payoff under an asset price model whose volatility is σ and whose growth rate is r.17) . t) := 1 2πσ 2 (T − t)S e − P(S . s) = √ 1 − (y−z)2 e 4s f (z) dz 4πs where f is the initial data.
23). Thus. its solution is given by V = Ee−rτ u. Prove the formula (3.19) (3.20) (3. where u(x. τ ) = e−rτ 0 ∞ √ 1 2πσ 2 τ e−(x−z−( 2 σ 1 2 −r)τ )2 /(2σ 2 τ ) (ez − 1) dz. We can obtain the price for p from c: p(S. 2.4. t) = SN (d1 ) − Ee−r(T −t) N (d2 ). τ ) = √ = e 1 2πσ 2 τ − 1. Forward contract Recall that a forward contract is an agreement between two parties to buy or sell an asset at certain time in the future for certain price. t) = Ee−r(T −t) N (−d2 ) − SN (−d1 ). The value V for this contract also satisﬁes the BS equation. European put option. This can be integrated. y z2 1 √ N (y) = e− 2 dz. σ T −t Exercise. The rescaled payoff function for a European call option is ¯ Λ(z) = max{ez − 1.3. European call option. Exercise.4. we get the exact solution for the European call option c (S. Then v(x. Finally. ∞ (3. 2π −∞ S log( E ) + (r + 1 σ 2 )(T − t) √ 2 d1 = . Show that N (d1 ) − 1 = N (−d1 ).22) (3.4 Special cases 1.19). Use this to prove (3. The payoff function for such a forward contract is Λ(S) = S − E.23) e −∞ − (y−z−(r− 1 σ 2 )τ )2 2 2σ 2 τ (ez − 1) dz x+rτ . EXACT SOLUTION FOR THE BS EQUATION FOR EUROPEAN OPTIONS 31 3. 3.21) (3. σ T −t S log( E ) + (r − 1 σ 2 )(T − t) √ 2 d2 = . 0}. Recall the putcall parity c + Ee−r(T −t) = p + S.
Otherwise. Supershare is a binary option whose payoff function is deﬁned to be Λ(S) = B 0 otherwise. From this and the BlackScholes formula (3. . t) = Be−r(T −t) (N (d2 (E1 )) − N (d2 (E2 ))) where d2 (E) is given by (3. The payoff function is Λ(S) = B 0 if S > E otherwise. Or in τ. t) = Be−r(T −t) N (d2 ). the value of the option is the payoff function evaluated at the future price of S at T (that is Ser(T −t) ). the BlackScholes equation is reduced to Vt + rSVs − rV = 0.25) V (S. and then discounted by the factor e−r(T −t) . BLACKSCHOLES ANALYSIS V (S. Notice that this value is independent of the volatility σ of the underlying asset. you get nothing. 5. If ST > E.16).24) This means that the current value of a forward contract is nothing but the difference of S and the discounted E. Thus. 4. Deterministic case (σ = 0). Supershare. t) = S − Ee−r(T −t) . 0) = Λ(Eex ). u(x.32 Hence. show the formula of the putcall parity.22). Cashornothing. In this case. A contact with cashornothing is just like a bet. (3. Exercise. then the reward is B. Using the BlackScholes formula (3. we obtain the value of a cashornothing contract to be (3. CHAPTER 3. 6. x and u variables: uτ − rux = 0 with initial data u(x. if E1 < S < E2 One can show that the value for this binary option is V (S. This means that when the process is deterministic. t) = e−r(T −t) Λ(Ser(T −t) ). τ ) = Λ(Sex+rτ ) . Or V (S. Show that the payoff function of a portfolio c − p is S − E.16).
3. t) := 1 2πσ 2 (T − t)S e − (log( S )−(r− 1 σ 2 )(T −t))2 2 S 2σ 2 (T −t) . (3. t)S dS . Recall the lognormal probability density function with growth rate r. Ser(T −t) (3. (3. T.6 The delta hedging Hedging is the reduction of sensitivity of a portfolio to the movement of the underlying of asset by taking opposite position in different ﬁnancial instruments. T. where S obeys the riskneutral pricing model: ˜ dS = rdt + σdz. one can show that N (d1 ) = ∞ E P(S . The BlackScholes . s. this value should be discounted by e−r(T −T ) : V (S. we ﬁnd ∞ N (d2 ) = E P(S . S.26) This is the transition probability density of an asset price model in a riskneutral world: dS = rdt + σdz. (3. This means that the drift term (growth rate) µ in the asset pricing model can be replaced by r.5. The option is then valued by calculating the present value of its expected return at expiry.29) ˜ ˜ is the expectation of S at T when S = 1 at t and under the condition that S ≥ E at T . RISK NEUTRALITY 33 3. T. t)Λ(S )dS . t)dS . volatility σ is P(S . S. At time t. S The expected return at time T in this riskneutral world is P(S .28) ˜ ˜ This is the probability of the event {S ≥ E}. 3. ˜ S Similarly.27) We may reinvestigate the function N and the parameters di in the BlackScholes formula. After some calculation. S. S.5 Risk Neutrality Notice that the growth rate µ does not appear in the BlackScholes equation. t) = e−r(T −t) P(S . t)Λ(S )dS . T. T. The option may be valued as if all random walks involved are risk neutral.
we have the following propositions. r. With a suitable balance of the underlying asset and other derivatives. By taking dΠ − ∆ · dS. By deﬁnition. The delta for a whole portfolio is ∆ = ∂Π . ∂2S 2 ∂Π Theta: θ = − . This is the sensitivity of Π ∂S against the change of S. For the Deltahedge for the European call and put options. BLACKSCHOLES ANALYSIS analysis is a dynamical strategy. ∂σ ∂Π rho: ρ = .34 CHAPTER 3. Sσ (T − t) 1 2 N (di ) = √ e−di . ∂r Hedging against any of these dependencies requires the use of another option as well as the asset itself. 2π Hence. . The delta hedge is instantaneously risk free. σ. It requires a continuous rebalancing of the portfolio and the ratio of the holdings in the asset and the derivative product. ∂S Since S log( E ) + (r + σ2 )(T − t) √ d1 = . Proof. σ T −t 2 we have d1 S = d2 S = 1 Sσ (T − t) 1 . Besides the delta helge. . its ∆ hedge is given by ∆ = N (d1 ). ∂t ∂Π Vega: = . hedgers can eliminate the shortterm dependence of the portfolio on the movement in t. ∂c = N (d1 ) + S · N (d1 ) · d1 S − Ee−r(T −t) N (d2 )d2 S . the sensitivity of the portfolio to the asset price change is instantaneously zero. ∂c ∂S √ = N (d1 ) + SN (d1 ) − Ee−r(T −t) N (d2 ) /(Sσ T − t) √ ≡ N (d1 ) + I/(Sσ T − t). there are more sophisticated trading strategies such as: Gamma: Γ = ∂ 2Π . Proposition 1 For European call options. S.
22). µ Suppose r. We use the changeofvariables: S = Eex . ∆= ∂c ∂p = − 1 = N (d1 ) − 1 = −N (−d1 ).6. E N (d2 ) 35 Hence. d2 − d2 = 1 2 1 σ σ (x + rτ + τ )2 − (x + rτ − τ )2 σ2τ 2 2 = 2(x + rτ ) S N (d1 ) = ex · e−x−rτ = e−rτ . σ. THE DELTA HEDGING We claim that I = 0.3.1 TimeDependent r. V = Ev. its ∆ hedge is given by ∆ = N (−d1 ). ∂S ∂S 3.30) We look for a new time variable τ such that ˆ dˆ = σ 2 (τ )dτ τ . τ = T − t. Or equivalently.21)(3. but also deterministic. The BlackScholes remains the same. E N (d2 ) From (3.6. S N (d1 ) 2 2 = ex · e−(d1 −d2 )/2 . µ are functions of r. From the putcall parity. Proof. The BlackScholes equation is converted to vτ = σ 2 (τ ) σ 2 (τ ) vxx + (r(τ ) − )vx − r(τ )v 2 2 (3. Proposition 2 For European put options. S N (d1 ) = e−rτ E N (d2 ) This follows from the computation below. σ.
→ y τ1 ˆ . τ ˆ 2 and The equation (3. . The goal is to design a portfolio involving vanilla option with a designed payoff function. and u = eB(ˆ1 ) v. BLACKSCHOLES ANALYSIS τ 0 σ 2 (τ ) dτ. τ ˆ 3. ∂ ∂y ∂ ∂ τ = τ = . 0} or max{E − ST . Then the equation becomes 1 vτ = vxx + a(ˆ)vx − b(ˆ)v. Or equivalently.7 Trading strategy involving options The options whose payoff are max{ST − E. And we can solve this heat τ ˆ equation explicitly. τ τ ˆ 2 To eliminate a(ˆ). τ Now. ˆ ˆ ∂x ∂x ∂y ∂y ∂ ∂τ ∂ ˆ ∂x ∂ ∂ ∂ y = + = − a(ˆ) .31) x=− 0 a(τ )dτ + y. τ ˆ y =x+ 0 a(τ )dτ ≡ x + A(ˆ).31)is transformed to 1 τ Let B(ˆ1 ) = 0 1 b(τ )dτ . 0} are called vanilla option. then uτ1 = 2 uyy . In this section. we can choose τ= ˆ CHAPTER 3. we consider the changeofvariable: x τ ˆ Then. τ ∂ τ1 ˆ ∂ τ1 ∂ τ ˆ ˆ ∂ τ1 ∂x ˆ ∂τ ˆ ∂x 1 vτ1 = vyy − b(ˆ1 )v.36 For instance. we shall discuss more general payoff functions. we consider the characteristic equation: τ dx = −a(ˆ) τ dˆ τ This can be integrated and yields τ ˆ (3.
c. The payoff of Π is Λ = S − max(S − E. TRADING STRATEGY INVOLVING OPTIONS 37 3. And Λ = − min{S. Π = p + S (protective put). Π = c − S (reverse of a covered call). we anticipate the stock price will decrease. Π = −p − S (reverse of a protective put). we short a call. we anticipate the stock price will increase. Π = S − c (writing a covered call option). long a share to cover c. 0} = max{S.1 Strategies involving a single option and stock There are four cases: a. b. The payoff is Λ = S + max{E − S. In this strategy. E}. We anticipate the stock price will increase. E}.7. we long a p and buy a share to cover p. d. In this strategy. In this portfolio. E}. We do not anticipate the stock price will increase. 0) = min{S. In this case. E}.3. Below are the payoff functions for the above four cases. The payoff is − max{S. E Λ S Λ E S (a) (b) E S Λ Λ E S .7.
Example: CE1 = 3. requires an initial investment. 0} − max{ST − E2 . CE2 = 1 and E1 = 30. 0} if ST < E 0 = ST − E1 if E1 < ST < E2 E2 − E1 if ST > E2 Λ E −E 2 1 E1 E2 S Since E1 < E2 . he would like to give up some of his right if the price goes beyond certain price. E 1 < E2 .38 (c) CHAPTER 3. . We can describe the strategy by saying that the investor has a call option with a strike price E1 and has chosen to give up some upside potential by selling a call option with strike price E2 > E1 . the investor gets E2 − E1 if the price goes up beyond E2 . In return. Indeed. The payoff if ST ≤ 30 0 ST − 30 if 30 < ST < 35 5 if ST ≥ 35 The bull spread can also be created by using put options Π = PE1 − PE2 .2 Bull spreads In this strategy. Hence he does want to own a right beyond E2 . E 1 < E2 . The payoff Λ = max{ST − E1 . However. E2 = 35. an investor anticipates the stock price will increase.7. we have CE1 > CE2 . Such a portfolio can be designed as Π = CE1 − CE2 . A bull spread. The cost of the strategy is 2. he does not anticipate the stock price will increase beyond E2 . CE2 have the same expiry. where CEi is a European call option with exercise price Ei and CE1 . BLACKSCHOLES ANALYSIS (d) 3. when created from CE1 − CE2 . say E2 .
7. TRADING STRATEGY INVOLVING OPTIONS 39 3. Butterﬂy spread using calls: Deﬁne the portfolio: Π = CE1 − 2CE2 + CE3 . The payoff is 3.4 Butterﬂy spread If an investor anticipate the stock price will stay in certain region.7. Λ E −E 2 1 S−E1 E2−S E1 E2 E3 S Example: Suppose a certain stock is currently worth 61. Suppose the market of 6 month calls are E C 55 10 60 7 65 5 The investor creates a butterﬂy spread by Π = CE1 − 2CE2 + CE3 . E1 < E2 . . A investor who feels that it is unlikely that there will be signiﬁcant price move in the next 6 month. where E3 = E2 + (E2 − E1 ).3. E1 < ST < E3 . There is cash ﬂow entered (CE2 − CE1 ). Λ(E2 ) = E2 − E1 . say. he or she can have a butterﬂy spread such that the payoff function is positive in that region and he or she gives up the return outside that region.7. Below is the graph of its payoff function. The bear spread is Π = CE2 − CE1 . 1. with E1 < E2 < E3 . Its payoff function is a piecewise linear function and is determined by Λ(E1 ) = Λ(E3 ) = 0. an investor entering into a bear spread is expecting the stock price will go down. By contrast.3 Bear spreads An investor entering into a bull spread is hoping that the stock price will increase.
Butterﬂy spread using puts. BLACKSCHOLES ANALYSIS The cost is 10 + 5 − 2 × 7 = 1. Suppose European options were available for every possible strike price E. Λi is constant. The payoff is Λ 5 55 60 65 S 2. T. Λi ψS − Ei . One can also use cashornothing to create any payoff function: Λ(S) = where ψ(S) := H(S) − H(S − ∆E). or S > E2 − ∆E Remark 1. then any payoff function could be created theoretically: Λ(S) = Λi ϕE ∆E i where Ei = i∆E. we can approximate any payoff function by using butterﬂy spreads. S.40 CHAPTER 3. Ei+1 ) and Λ is continuous. Remark 2. . 2 ϕE2 if S = E2 if S < E2 − ∆E. t)Λ(S )dS . ∆E = e−r(T −t) ΣΛi P(Ei ≤ S ≤ Ei+1 ). linear = ∆E 0 E 1 < E3 . The value for such a portfolio is V = e−r(T −t) P(S . E2 = E1 + E3 . As ∆E → 0. PE1 + PE3 − 2PE2 . Then Λ(Ei ) = Λi and Λ is linear on every interval (Ei .
1 Constant dividend yield Suppose that in a short time dt. the change of portfolio is ∂V ∂S in order to eliminate the randomness of dΠ = dV − ∆dS − ∆D0 Sdt. where D0 is a constant. the return dS must fall by the amount of the dividend payment S D0 dt.1 Options on dividendpaying assets Dividends are payments to the shareholders out of the proﬁts made by the company. In the latter case. for the asset price is dS = (µ − D0 )dt + σdz. we choose ∆ = dΠ. 4. This continuous dividend structure is a good model for index options and for shortdated currency options. the last term −∆D0 Sdt is the dividend our assets received.Chapter 4 Variations on BlackScholes models 4. We will consider two “deterministic” models for dividend. In one time step. the foreign interest rate.1. The other has discrete dividend payments. S For a portfolio : Π = V − ∆S.e. ∂S 2 ∂S 41 . the underlying asset pays out a dividend D0 Sdt. called the dividend yield. It follows the s. One has constant dividend yield.d. As the dividend is paid. D0 = rf . Thus dΠ = dV − ∆(dS + D0 Sdt) ∂V − ∆ dz = σS ∂S ∂V 1 ∂ 2V + (µ − D0 )S + σ 2 S 2 2 + Vt − (µ − D0 )∆S − ∆D0 S dt ∂S 2 ∂S 2 ∂V 1 ∂ V = Vt − D0 S + σ 2 S 2 2 dt.
0 − σ T − t . t) ∼ Se−D0 (T −t) The latter is the asset price S discounted by e−D0 (T −t) from the payment of the dividend. 1 ∂V ∂ 2V Vt + σ 2 S 2 2 + (r − D0 )S − rV = 0 2 ∂S ∂S This is the BlackScholes equation when there is a continuous dividend payment. To ﬁnd the solution. VARIATIONS ON BLACKSCHOLES MODELS Here. Then c1 satisﬁes the original BlackScholes equation with r replaced by r − D0 and the same ﬁnal condition. t) = Se−D0 (T −t) N (d1. t) = 0.0 ) − Ee−r(T −t) N (d2. . From the absence of arbitrage ∂S opportunities. c as D0 . let us consider c(S.0 ) where d1. t) ∼ S as S → ∞ Hence. The payoff function c(S.0 S ln E + (r − D0 + 1 σ 2 )(T − t) 2 √ . t). ∂V 1 2 2 ∂ 2V ∂V Vt − D0 S + σ S =r V −S 2 ∂S 2 ∂S ∂S i.0 = d2.0 ) − Ee−(r−D0 )(T −t) N (d2. c1 (S.e. c(S. σ T −t √ = d1. Derive the putcall parity for the European options on dividendpaying assets. we must have dΠ = rΠdt.0 ) or c(S. T ) = Λ(S) = max{S − E. we have chosen ∆ = ∂V to eliminate the random term. t) = e−D0 (T −t) c1 (S.. Remark. The boundary conditions for c1 are c1 (0. 0}. t) = 0. t) = SN (d1. c1 (S. The boundary conditions are: c(0.42 CHAPTER 4. Exercise. Thus.
t. This value is distributed to N + M γ shares. Suppose each warrant entitles the holder to purchase γ share from the company at time T at price X per share. Suppose a company has N outstanding shares and M outstanding European warrants. td +) = c(S(1 − dy ). Then cd (S. T. td −) = cd (S(1 − dy ). 4. Solve the BlackScholes from T to Td + to obtain V (S.4. td +) (using the payoff function Λ) 2. t. N + γM . Adjusting V by V (S. V should be continuous. td −) = V (S(td +). we obtain c(S(1 − dy ). 0} = (1 − dy ) max{S − (1 − dy )−1 E. td +). The dividend yield is a constant. td +) 3. i. We claim that across the jumps.2 Discrete dividend payments Suppose our asset pays just one dividend during the life time of the option. E) = (1 − dy )c(S. At td +. To ﬁnd V (S.1. 0} and the linearity of the BlackScholes equation. t). td +). Hence the share price becomes VT + M γX . Reason : Otherwise. E) Note that c(S(1 − dy ). Hence. Let cd be the European option for this dividendpaying asset. (1 − dy )−1 E).2 Warrants An European warrant is a right to purchase an underlying stock at price X at expiry. WARRANTS 43 4..e. E) = max{S(1 − dy ) − E. We want to determine the price of a warrant. td −) = V ((1 − dy )S. If the warrant holders exercise. here is a procedure.2. t. E) for td + ≤ t ≤ T cd (S. t) = c(S. then the company received a cash inﬂow M γX and the company’s equity increases to VT + M γX. Let VT be the value of the company’s equity at T . the asset holder receiver a payment dy S(td −). S(td +) = S(td −) − dy S(td −) = (1 − dy )S(td −). there is a net loss or gain from buying V before td then sell it right after td . t. 1. V (S(td −). Solve BlackScholes equation from td to t with the ﬁnal condition V ((1−dy )S. say at time td .
i. N + Mγ N where V is the value of the company’s equity at time t.44 CHAPTER 4.. the price for this forward contract is f = S − Ee−r(T −t) .3 4. We obtain a nonlinear N algebraic equation for w: w = Nγ M c(S + w.0 N + Mγ = Nγ max N + Mγ VT − X. Based on the no arbitrage opportunity. E is called the delivery price. . X). Since V = N S + M w. The value of the warrant at time t should be Nγ V w= c( . VARIATIONS ON BLACKSCHOLES MODELS The payoff to the warrant holder is max γ VT + M γX − X . c(S. t. F . Consider a party who is short the contract.1 Futures and futures options Forward contracts Recall that a forward contract is an agreement between two parties to buy or sell an underlying asset on a certain price E at a certain future time T . Ft = St er(T −t) . t. then F = St er(T −t) .. t.e.e. The payoff function for this forward contract is Λ = ST − E.2 The forward price F for a forward contract is deﬁned to be the delivery price which would make that contract have zero value. X) is the value of a European V call with strike price X. 0 . 4. Here. i.3. N This is exactly the payoff function for a European call. Deﬁnition 3. X) N + Mγ N M Nγ (S + w)N (d1 ) − Xe−r(T −t) N (d2 ) = N + Mγ N where d1 = log((S + √ M w)/X) N σ T −t √ + (r + 1 σ 2 )(T − t) 2 d2 = d1 − σ T − t. This algebraic equation can be solved numerically. One can take another point of view. If no arbitrage opportunities. Thus. The money he received at expiry. is used to pay the loan. N = S + M w. He can borrow an amount of money St at time t to buy an asset and use it to close his short position at T .
except they are traded in an exchange. etc. a maintenance margin. the amount of money in the margin account also changes. thus. 2. Theorem 4. Suppose. • The trader signal this to ask other traders to sell.000 bushel • Delivery arrangement: delivery month is on December • price quotes • Daily price movement limits: these are speciﬁed by the exchange.000 bushels) on the Chicago Board of Trade (CBOT) at current market price. The initial margin. The broker will require the investor to deposit funds in a “margin account”. Let us explain the characters of a future by the following example. Trading future contracts • Suppose you call your broker to buy one July corn futures contract (5. .2 Forward price and futures price are equal when the interest rates are constant. 1992. the futures price has dropped from $400 to $397.000 per contract. which is usually lower than the initial margin. the investor want to buy $200 ounces at this price.2 Futures Futures are very similar to the forward contracts. quality. 1. Maintaining margin needs to deposit. is set.. expiry. • Position limits: the maximum number of contracts that a speculator may hold.3. 3. We suppose that the current future price is $400 per ounce. the trader who represents you will raise the price and eventually ﬁnd someone to sell • Conﬁrmation: Price obtained are sent back to you. As the futures prices move everyday. they are required to be standardized. for example. • Maintenance margin: To insure the balance in the margin account never becomes negative. This balance in the margin account would therefore be reduced by $600. . the speciﬁcation of this future is • Asset : quality • Contract size: 5. say is $2. to buy two December 1992 gold futures contracts on New York Commodity Exchange. FUTURES AND FUTURES OPTIONS 45 4. price. . • The broker send this signal to traders on the ﬂoor of the exchange. The contract size is $100 ounces. Certain account of money to keep that futures contract.3. The investor has a loss of $200×3=600.4. Operation of margins • Marking to market: Suppose an investor who contacts his or her broker on June 1. Speciﬁcation of the futures: In the above example. This includes size. by the end of June 1. if no one want to sell.
When a call futures option is exercised. Consider the following two strategies: 1. An investor who has a September futures call option on 25. Invest G0 in a riskfree bond and take a long position of amount enδ forward contract. Example. we received F0 enδ . Let δ be the riskfree interest rate per day. F0 = G0 = ST er(T −t) . we conclude their initial investments must be the same. They require the delivery of an underlying futures contract when exercised. Suppose the current future price of copper . Let the future prices are F0 . • Take a long position of future eδ amount of at the end of day 0. i. If we invest F0 initially. Suppose a futures contract lasts for n days.46 CHAPTER 4. Since both strategies have the same payoff.3. the holder acquires a long position in the underlying futures contract plus a cash amount equal to the current futures price minus the exercise price.000 pounds of copper with exercise price E = 70 cents/pound. 4. G0 enδ is used to buy the underlying asset at price ST enδ . • Take a long position of future e2δ amount of at the end of day 1. F n at the end of each business day. at T . VARIATIONS ON BLACKSCHOLES MODELS Proof. N0 investment is required for all the long future positions.e. • Invest F0 amount of money in a riskfree bond.. 2. The payoff of strategy 2 is F0 enδ + (ST − F0 )enδ = ST enδ . At day n. Day futures price position gain/loss compound 0 F0 eδ 0 δ e (F1 − F0 )e(n−1)δ 1 F1 e2δ eδ (F1 − F0 ) e2δ (F2 − F0 )e(n−2)δ 2 F2 e3δ e2δ (F2 − F1 ) ··· ··· ··· ··· ··· ··· n−1 Fn−1 enδ ··· ··· n Fn 0 enδ (Fn − Fn−1 ) (Fn − Fn−1 )enδ The total gain/loss from the long position of the futures is n (Fi − Fi−1 )eiδ · e(n−i)δ i=1 = (Fn − F0 )enδ = (ST − F0 )enδ . · · · .3 Futures options Options on futures are traded in many different exchanges.
1) F This means that the futures price is the same as a stock paying a dividend yield at rate r. dΠ = rV dt. 000+long position in futures contract to buy 25. or a few days before. Then dΠ = dV − ∆dF ∂V ∂V 1 ∂ 2V 2 2 ∂V ∂V = ( µF F + + σ F )dt + σF dz − (µF F dt + σF dz) 2 ∂F ∂t 2 ∂F ∂F ∂F 2 1∂ V 2 2 ∂V + σ F )dt. we obtain a pricing model for F : o 1 dF = (Ft + FSS σ 2 S 2 )dt + FS dS 2 dS r(T −t) = (−re S)dt + Ser(T −t) S = (−rF )dt + F (µdt + σdz) = ((µ − r)F )dt + F σdz. = ( ∂t 2 ∂F 2 Since it costs nothing to enter into a future contract. This is because at time T . Next. i. Thus. the investor received 10 cents ×25. 0}. we obtain 1 ∂2V Vt + σ 2 F 2 2 = rV 2 ∂F The payoff function for a call option is Λ = max{F − E.3. t.4. If the option is exercised. It is a function of F . to eliminate randomness of dΠ. dF = (µ − r)dt + σdz.e.3. The maturity date of futures option is generally on. we study the value V of a futures option. the earlist delivery date of the underlying futures contract. F = Ser(T −t) . 4.. ST = F T . Consider a portfolio Π = V − ∆F We choose ∆ = ∂V ∂F Hence. Futures options are usually more liquid and involved lower transaction costs. (4. the cost of setting up the above portfolio is just V . Futures options are more attractive to investors than options on the underlying assets when it is cheaper or more convenient to deliver futures contracts rather than the asset itself.4 BlackScholes analysis on futures options As we have seen that the futures price is identical to the forward price when the interest rate is a constant. Thus based on the no arbitrage opportunity. From Itˆ s lemma. .000 pound of copper in September at price 80 cents/pound. FUTURES AND FUTURES OPTIONS 47 for delivery in September is 80 cents/pound.
Futures price is the same as a stock paying dividend at yield rate r. t). The price for future options is the same as the price for options on the underlying assets. d2 = σ T −t √ d1 = d2 + σ T − t ˜ We can write this V in terms of F by S = F e−r(T −t) . Futures price F = St er(T −t) .48 CHAPTER 4. t) Conclusion: 1. let us ﬁnd the putcall parity for futures options. Plug this into the above equation to obtain d2 = = ln F e ln F E −r(T −t) E σ T −t − − t) √ . t) where c1 satisﬁes BlackScholes equation with r replaced by r − r = 0. where V is the solution of the option corresponding to the underlying asset S. σ T −t 1 2 σ (T 2 √ − 1 σ 2 (T − t) 2 ˜ V (S(F. σ √T − t d1 = d2 + σ T − t ˜ ˜ Notice that this V is the same as V (S(F. . t) = N (d1 ) − Ee−r(T −t) N (d2 ) 1 S ln E + (r − 2 σ 2 )(T − t) √ . This gives c1 (F. t) = F N (d1 ) − EN (d2 ). t) = SN (d1 ) − Ee−r(T −t) N (d2 ) = F e−r(T −t) N (d1 ) − Ee−r(T −t) N (d2 ) = V (F. so the futures call option c(F. t). we recall the option price equation for stock paying dividend is 1 ∂V ∂ 2V Vt + σ 2 S 2 2 + (r − D0 )S − rV = 0 2 ∂S ∂S In our case. That is ˜ V (S. t) = e−r(T −t) c1 (F. 2. t). D0 = r. F ln E − 1 σ 2 (T − t) 2 √ d2 = . VARIATIONS ON BLACKSCHOLES MODELS To solve this equation. Finally. 3.
FT }. ΛA = max{FT − E. Consider two portfolios: A = c + Ee−r(T −t) B = p + F e−r(T −t) + a futures contract At time T . 0} + E = max{FT .2) . FUTURES AND FUTURES OPTIONS Proposition 3 c + Ee−r(T −t) = p + F e−r(T −t) Proof. Hence we obtain A = B.4.3. 0} + F + (FT − F ) = max{E. ΛB = max{E − FT . E}. 49 (4.
VARIATIONS ON BLACKSCHOLES MODELS .50 CHAPTER 4.
we sample. if there are many underlying assets.Chapter 5 Numerical Methods 5. S(t + (k − 1)∆t) Remark. t)Λ(S)dS ˜ where Λ is the payoff function.1 Monte Carlo method We recall that the value of a European option is given by V (S. . 1)(i. . In this case. i = 1.1) ˜ S i. we divide the interval [t. P is the transition probability density function of S which satisﬁes ˜ dS = rdt + σdz. the Monte Carlo does not have any advantage. variance 1). S. say. . If there is only one underlying asset. .1). ﬁnite difference method is very difﬁcult and the Monte Carlo method wins. The Monte Carlo simulation is a numerical procedure to estimate V based on this formula. say more than three. 10. k = 1.e. M with distribution N (0. 51 . . . T. However. the normal distribution with mean 0. it is the asset price model in the riskneutral world. t) = e−r(T −t) ˜ ˜ ˜ P(S.. We then approximate V by V ≈ e−r(T −t) 1 N N Λ(Si (T )).. 10000. We sample M random numbers k . The error of a MonteCarlo method is O √1N . T ] into M subinterval with equal −t length ∆t = TM .e. . ( initial state S(t) = S) (5. To ﬁnd V .1). i=1 To sample a path from (5.000 paths from (5. . the corresponding Black Scholes equation is a diffusion equation in high dimensions. We obtain Si (T ). We then deﬁne Si (t + k∆t) by √ S(t + k∆t) − S(t + (k − 1)∆t) = r∆t + σ k ∆t.
p and u can be expressed in terms of r. Next. where S0 is the asset price at current time t. .2 Binomial Methods In binomial method. namely.. Remark. If we denote log S/E by x.) They are determined by the conditions so that the discrete model and the continuous model have the same mean and variance in one time step ∆t. We recall that these conditions are pu + (1 − p)d = er∆t pu2 + (1 − p)d2 = e(2r+σ 2 )∆t Thus. then we determine the option price from the expectation of the payoff function according to the price distribution of the asset in the riskneutral world.e. we ﬁrst simulate a riskneutral asset price model forward in time by a binomial model. NUMERICAL METHODS 5.52 CHAPTER 5. T ] into N subintervals with equal length ∆t = (T − t)/N . u = e∆x and d = e−∆x . we discrete the continuous model in time. n r This discrete model depends on two parameters: u and p.2) We shall approximate this continuous model by the following discrete model. and ∆x is a parameter to be determined later. then the asset price will move up to Sj+1 = Sj u with probability p and move down to Sj−1 = Sj d with probability 1 − p. σ and ∆t. (5. i. First. 5. A simple calculation gives √ u = 1 + σ ∆t + O(∆t) √ 1 p = + O( ∆t) 2 We should require ∆t is chosen so that 0 ≤ p ≤ 1. then the movement of S on the discrete values Sj corresponds to a movement of x on xj = j∆x. We want to ﬁnd the probability distribution of the asset price in a riskneutral world at time T whose current price is S0 . Here. Then P0 = 1 and Pjn is exactly the binomial distribution: Pjn = pr (1 − p)n−r n + j = 2r 0 otherwise. we assume our discrete asset prices only take discrete values Sj = S0 ej∆x . we partition [t.2.1 Binomial method for asset price model dS = rdt + σdz S We consider the underlying asset is riskneutral. This movement is exactly the random walk we introduced in Chapter 2. The discrete asset price model is: if the asset price is at Sj at time step n. ( The down ratio d = 1/u. 0 Let us denote the probability that the price is at Sj at time step n by Pjn .
Vt + σ S 2 ∂S 2 ∂S Using dimensionless variables S = Eex . we shall solve the BlackScholes equation by ﬁnite difference methods. then u satisﬁes 1 ∂ 2u 1 ∂u uτ = σ 2 2 + (r − σ 2 ) .3 Finite difference methods (for the modiﬁed BS eq. 2 ∂x 2 ∂x The initial condition for u is u(x. 0} max{1 − ex . Then we compute Vjn inductively from n = N to n = 0 by n+1 n+1 er∆t Vjn = pVj+1 + (1 − p)Vj−1 .1.1224 0. we have 1 ∂v 1 ∂2v − rv.4 $50. If S takes value at Sj at time step n. V = Ev. T ∆t r S u d p er∆t = = = = = = = = 5 months = 0. −n ≤ j ≤ n. Example. with VjN being the payoff function.0833 year 0.3.2 Binomial method for option Since the asset price only takes discrete values Sj . 5.0084 n We begin to generate a binomial tree from S = 50 consisting of Sj = Sur dn−r . vτ = σ 2 2 + (r − σ 2 ) 2 ∂x 2 ∂x Let v = e−rτ u. We recall that V n is the expected value of the option at (n + 1)∆t discounted by e−r∆t . we shall approximate V (Sj . Therefore. E = $50 √ eσ ∆t = 1.5076 1.4167 year 1 months = 0. 0} for call option for put option (5. the expected value of V at time step n should satisﬁes n+1 n+1 er∆t Vjn = pVj+1 + (1 − p)Vj−1 . where n + j = 2r. The value V00 is our answer. σ = 0.) 53 5. For put option.2.5. τ = T − t.3) .) In this section. t + n∆t) by Vjn . Recall that the BlackScholes equation is ∂V 1 2 2 ∂ 2V = r(V − S ). 0) = max{ex − 1. then S takes values Sj+1 with probability p and Sj−1 with probability 1 − p. FINITE DIFFERENCE METHODS (FOR THE MODIFIED BS EQ.8909 0.
CHAPTER 5. n∆τ ) by Ujn . We choose a proper ﬁnite domain (xL . where ∆x = . Spatial discretization. t) = 0. Temporal discretization. From v = e−rτ u.. t) = 1.54 The far ﬁeld boundary condition for u is u(−∞. xR ). . NUMERICAL METHODS u(x. N. ∆x (b) uxx ← uj+1 −2uj +uj−1 (∆x)2 Then the righthandside of (5.. t) = 0 as x → ∞ 5. we follow the following procedure: 1. For the temporal discretization. for a put option.. we have Vjn = e−rn∆τ Ujn . we introduce the following three methods: (a) Forward Euler method: Ujn+1 − Ujn = (QU n )j . We replace the spatial derivatives by ﬁnite differences: (a) ux is replaced by one of the following three: uj+1 −uj−1 2∆x uj −uj−1 1 ux ← if 2 σ 2 − r > 0 ∆x uj+1 −uj 1 if 2 σ 2 − r ≤ 0. V (xj . ∆t = TM . we discretize [t. for a call option.3) is discretized into (QU )j ≡ ( σ 2 Uj+1 − 2Uj + Uj−1 σ 2 Uj+1 − Uj−1 ) + (r − ) 2 (∆x)2 2 2∆x 3. Discretize space and time. ∆t 2 . t) = ex erτ as x → ∞ u(x. n∆τ ) by Vjn .3. N −t Similarly. ∆t (b) Backward Euler method: Ujn+1 − Ujn = (QU n+1 )j . and u(−∞. 2. ∆t (c) CrankNicolson method: Ujn+1 − Ujn 1 = [(QU n+1 )j + (QU n )j ].3) numerically. T ] into N steps. We shall approximate u(xj . j = −N. discretize it into xR − xL xj = j∆x.1 Discretization methods To solve (5.
3. we rewrite it as Ujn+1 = Ujn + ∆t(QU n )j n n = aUj−1 + bUjn + cUj+1 In terms of Vjn .E.e. We see that this ﬁnite difference is identical to the binomial method in the previous section. c. ∆t) and → 0 as ∆t. We can choose ∆τ and ∆x properly so that ∆t b = 0..5. n n er∆t Vjn+1 = pVj+1 + (1 − p)Vj−1 where p = a = 1 ∆t σ2 2 (∆x)2 + (r − σ 2 ∆t ) .D. In this case.E (ﬁnite difference equation) + (∆x. c ≥ 0 reads r− ∆t 2 σ 2 ∆t ≤ σ ≤1 2 ∆x (∆x)2 Next. 5. Notice that n n a + b + c = 1. i. j = −N. Vjn . if for any solution of the corresponding P.3. .3... and let ∆t = (T − t)/N . 2 2∆x The stability condition is satisﬁed if and only if (5.3 Stability Deﬁnition 3.. c ≥ 0 for stability reason. Thus Vjn+1 is the “average” of Vj−1 . For the forward Euler method.2 Binomial method is a forward Euler ﬁnite difference method We choose x = S/S0 .D. This will be discussed later. 2 (∆x)2 2 2∆x j−1 n n ≡ aVj+1 + bVjn + cVj−1 We should require a. The stability condition a. T ] into N subintervals uniformally. it satisﬁes F.. b. 1 = (∆x)2 σ 2 . Let us partition the time interval [t. ∆x → 0 . b.E. b. then −r∆t discounted by e .3 A ﬁnite difference method is called consistent to the corresponding P. FINITE DIFFERENCE METHODS (FOR THE MODIFIED BS EQ. where S0 is the current asset price value. Let xj = j∆x.) 55 5.D. where ∆x is a small parameter satisfying some stability constraint to be shown below. Vj+1 with weight a. let us consider a special case: b = 0. N .4) 0 ≤ p ≤ 1. we have er∆t Vjn+1 = Vjn + 1 ∆t 2 n n σ (Vj+1 − 2Vjn + Vj−1 ) 2 2 (∆x) σ 2 ∆t n +(r − ) (V n − Vj−1 ) 2 2∆x j+1 1 ∆t 2 σ 2 ∆t ∆t 2 n n = σ + (r − ) Vj+1 + (1 − σ )Vj 2 2 (∆x) 2 2∆x (∆x)2 1 ∆t 2 σ 2 ∆t + σ − (r − ) Vn .
We describe his method below. n∆t) − (Qu)(j∆x.E.E. n∆t) of the corresponding P. Backward Euler: u(j∆x. Deﬁnition 3.E. with constant coefﬁcients. with variable coefﬁcients and nonlinear P. ∆t) in the previous deﬁnition. is convergent if its solution Ujn converges to the solution u(j∆x. n∆t) −(Qu)(j∆x. ∆t 2. NUMERICAL METHODS Deﬁnition 3. It also works “locally” and serves as a necessary condition for linear P. (n + 1)∆t) + (Qu)(j∆x. The true error Ujn − u(j∆x. (n + 1)∆t) − u(j∆x. We therefore omit it here. we shall focus on the stability issue.D. CrankNicolson method u(j∆x.5 A ﬁnite difference equation is said to be (L2 −)stable if the norm Un is bounded for all n ≥ 0. n∆t) is usually estimated in terms of the truncation error. n∆t)] ∆t 2 = O((∆x)2 ) + O((∆t)2 ). n∆t) = O((∆x)2 ) + O(∆t). (n + 1)∆t) − u(j∆x. Deﬁnition 3. the truncation error for central difference is Qu = σ2 σ2 uxx + (r − )ux + O((∆x)2 ). a ﬁnite difference method is convergent if and only if it is consistent and stable. n∆t) 1 − [(Qu)(j∆x. 2 2 And the truncation for various temporal discretizations are 1. Theorem 5. (n+1)∆t) = O((∆x)2 )+O(∆t).. It works for P. Since the consistency is easily to achieve..3 (Lax) : For linear partial differential equations.E. (n + 1)∆t) − u(j∆x.D.E. For instance.D.D.4 The truncation error of a ﬁnite difference method is deﬁned to be the function (∆x. ∆t 3.56 CHAPTER 5.6 A ﬁnite difference method for a P.D. A standard method to analyze stability issue is the von Neumann stability analysis. 2 := Σj Ujn 2 ∆x . This theorem is standard and its proof can be found in most numerical analysis text book. Forward Euler: u(j∆x.
ˆ U n+1 (ξ) = (1 + ∆tQ(ξ))U n = G(ξ)U n = G(ξ)n+1 U n We observe that π −pi U n (ξ)2 dξ = ≤ G(ξ)2n U 0 (ξ)2 dξ max G(ξ)2n U 0 (ξ)2 dξ ξ∈(−π.m we have (QU ) = j (Qu)j e−ijξ σ2 1 σ2 1 ˆ (2cosξ − 2) + (r − ) (2isinξ)]U 2 (∆x)2 2 ∆x ˆ ˆ ≡ Q(ξ)U (ξ). DU (ξ) = j Uj+1 − Uj−1 2 e−ijξ = j Uj ei(j+1)ξ − Uj ei(j−1)ξ 2 Uj e−ijξ j = eiξ − e−iξ 2 ˆ = (2i sin ξ)U (ξ) For the ﬁnite difference operator QU . = [ For forward Euler method. FINITE DIFFERENCE METHODS (FOR THE MODIFIED BS EQ.) We take Fourier transform of {Uj }∞ j=−∞ by deﬁning ∞ 57 ˆ U (ξ) = j=−∞ Uj e−ijξ It is a wellknown fact that Uj  j 2 1 = √ 2π ≡ ˆ U 2 π −π ˆ U (ξ)2 dξ ˆ Thus.5. the boundedness of j Uj 2 can be estimated by using U 2 . Namely.π) .3. The advantage of ˆ ˆ using U is that the ﬁnite difference operation becomes a multiplier in terms of U .
NUMERICAL METHODS If G(ξ) ≤ 1.58 CHAPTER 5. Example: Let apply the forward Euler method for the heat equation: ut = uxx . suppose its fourier transform satisﬁes’ U n+1 (ξ) = G(ξ)U n (ξ) Then the ﬁnite difference equation is stable if and only if G(ξ) ≤ 1 ∀ξ ∈ (−π. Let consider an initial condition such that 1 ξ − ξ0  ≤ δ U 0 (ξ) = 0 otherwise. . we have that G(ξ) ≥ 1 + for some small > 0 and for all ξ with ξ − ξ0  ≤ δ for some δ > 0. π). π]. then stability condition holds. ∆t (∆x)2 j+1 From von Neumann analysis: U n+1 = [1 + ∆t (2 cos ξ − 2)]U n (∆x)2 ξ ∆t = (1 − 4 sin2 )U n 2 (∆x) 2 =⇒ U n+1 = U n + ∆t D2 U n (∆x)2 ≡ G(ξ)U n . then by the continuity of G. On the other hand. ∀ξ ∈ (−π. Hence G(ξ) ≤ 1 =⇒ ∆t ξ 1 ∆t 1 sin2 ≤ =⇒ ≤ (stability condition) (∆x)2 2 2 (∆x)2 2 If we rewrite the ﬁnite difference scheme by Ujn+1 = ∆t ∆t ∆t U n + (1 − 2 )Ujn + Un 2 j+1 2 (∆x) (∆x) (∆x)2 j−1 n n ≡ aUj+1 + bUjn + cUj−1 . We conclude the above discussion by the following theorem. Theorem 5.4 For a ﬁnite difference equation with constant coefﬁcients. Then the corresponding U n will have π −pi U n (ξ)2 dξ = → ∞ G(ξ)2n U 0 (ξ)2 dξ as n → ∞. Then Ujn+1 − Ujn 1 n = (U n − 2Ujn + Uj−1 ). if G(ξ) > 1 at some point ξ0 .
Let Ujn be the probability of the particle at j∆x at time step n for a random walk.) Then the stability condition is equivalent to a. Consider a particle move randomly on the grid points j∆x. 59 Since we have a + b + c = 1 from the deﬁnition. In one time step. Ujn+1 − Ujn 1 n+1 n+1 n n = (Uj+1 − 2Ujn+1 + Uj−1 ) + (Uj+1 − 2Ujn + Uj−1 ) . Ujn and Uj−1 with weights ∆t a. thus we see that the ﬁnite difference n n scheme is nothing but saying Ujn+1 is the average of Uj+1 . 2. We ﬁnd that G(ξ) ≤ 1. G(ξ) ≤ 1. 1. ∆x > 0. (1 + (4 sin2 ) 2 (∆x)2 1 ξ ∆t 4 (∆x)2 (sin2 2 ) . Ujn+1 − Ujn 1 n+1 = (U n+1 − 2Ujn+1 + Uj+1 ). 1+α hence CrankNicolson method is always stable for all ∆t. FINITE DIFFERENCE METHODS (FOR THE MODIFIED BS EQ. For CrankNicolson method. ∆t (∆x)2 j−1 Then where G(ξ) = 1+ ξ ∆t )U n+1 = U n =⇒ U n+1 = G(ξ)U n .3. If we rename a = p. ∆t 2(∆x)2 Its Fourier transform satisﬁes 1 U n+1 − U n ξ ξ = −(4 sin2 )U n+1 − (4 sin2 )U n . c. This can be related to the random walk as the follows. c = 1 − p. if we choose (∆x)2 = 1 . for all ξ. then G(ξ) = 1−α . .5. c ≥ 0. 2 n+1 n n then Uj = pUj+1 + (1 − p)Uj−1 . then b = 0. the particle moves toward right with probability p and left with probability 1 − p. 2 ∆t 2(∆x) 2 2 We have (1 + 2 and hence U n+1 = ∆t ξ ∆t ξ (sin2 ))U n+1 = (1 − 2 (sin2 ))U n 2 2 (∆x) 2 (∆x) 2 ∆t 1 − 2 (∆x)2 sin2 ξ 2 n U . Let us only demonstrate the analysis for the heat equation. b. We ﬁnd that for all α ≥ 0. We left the analysis for the BlackScholes equations as exercises. the backward Euler method is always stable. n n Ujn+1 = pUj−1 + (1 − p)Uj+1 . For backward Euler method. ξ ∆t 2 (∆x)2 sin2 2 1+ ξ ∆t Let α = 2 (∆x)2 sin2 2 . Hence. b. In particular. We can also apply the above stability analysis to backward Euler method and the CrankNicolson method.
we obtain en ≤ ≤ ≤ ≤ en−1 + ∆t en−2 + ∆t en−2 + ∆t n−1 n−1 G n−2 + n−1 n−1 n−2 + e0 + ∆t k=0 k ≤ O(∆t) + O((∆x)2 ). Let us take the forward Euler scheme as our example. Then we obtain en+1 − en j j = (Qen )j + ∆t Or equivalently. b. 2 2 for the forward Euler method. back Euler method and CrankNicolson method. c ≥ 0. . then un+1 − un j j = (Qun )j + O(∆t) + O((∆x)2 ). and let en denotes for un − Ujn and j j truncation error. Thus.60 CHAPTER 5. suppose un := u(j∆x.5) Here. c ≥ 0 and a + b + c = 1. n∆t). a. en+1 = aen + ben + cen + ∆t j+1 j j−1 j n j n j n j denotes for the (5. 5. The method below can also be applied to other scheme.4 Convergence Let us study the convergence for ﬁnite difference schemes for the modiﬁed BlackScholes equation. Recall that the stability G(ξ) ≤ 1 is equivalent to a. The forward Euler scheme is given by: Ujn+1 − Ujn = (QU n )j ∆t We have known that it has ﬁrstorder truncation error. b. ∆t We subtract the above two equations. We can take Fourier transformation en of en . It satisﬁes en+1 (ξ) = G(ξ)en (ξ) + ∆t n (ξ) where G(ξ) = aeiξ + b + ce−iξ . NUMERICAL METHODS Exercise study the stability criterion for the modiﬁed BlackScholes equation uτ = σ2 σ2 uxx + (r − )ux .3. j where u is the solution of the modiﬁed BlackScholes equation. by applying the above recursive formula. namely.
t) = 0. there is no error initially. n∆t) − Ujn for the Euler method has the following j convergence rate estimate: ( j en 2 ∆x)1/2 ≤ O(∆t) + O((∆x)2 ). j It is simpler to ontain the maximum norm estimate. E(n + 1) ≤ E(n) + ∆t . j j Exercise. and that n∆t = O(1). we obtain the error is bounded by the truncation error. for all n.5). FINITE DIFFERENCE METHODS (FOR THE MODIFIED BS EQ. Let E(n) := maxj en  be the maxij mum error. t) = ex erτ as x → ∞ .n Hence. Hence. 5. we have used the estimate for the truncation error n 61 = O(∆t) + O((∆x)2 . u(x.5. we have j n−1 E(n) ≤ k=0 ∆t ≤ n∆t Since n∆t is a ﬁxed number. Since we take Uj0 = u0 . we have en+1  ≤ aen  + ben  + cen  + ∆t n  j j j−1 j+1 j ≤ aE(n) + bE(n) + cE(n) + ∆t = E(n) + ∆t where := max  n  = O(∆t) + O((∆x)2 ). as we take the limit n → ∞. we have u(−∞. Theorem 5. Theorem 5.3.3. From (5. j j.6 The error en := u(j∆x. We summarize the above discussion as the following theorem. We conclude the error analysis as the following theorem.5 The error en := u(j∆x.) Here.5 Boundary condition For the modiﬁed BlackScholes equation. n∆t) − Ujn for the Euler method has the following j convergence rate estimate: max en  ≤ O(∆t) + O((∆x)2 ). Prove that the true error of the CrankNicolson scheme is O((∆t)2 ) + O((∆x)2 ).
the inverse transformation is Eξ dξ (1 − ξ)2 S= . ∂τ 2 ∂ξ ∂ξ for 0 ≤ ξ ≤ 1 and τ > 0. 0).9) For a call option. which needs no boundary data. if we want to know u(xk . In practice. (5. With this. We allow σ depend on S. it is desirable to have a ﬁnite computation domain. . t) = 1.4 Converting the BS equation to ﬁnite domain The transformation x = log(S/E) converts the BS equation to a heat equation. Indeed. The corresponding V (ξ. the payoff is Λ(S) = max(S − E. ξ = (5. We deﬁne the transformation: S S+E V (S. The price is that the resulting equation has variable coefﬁcients. 0)(1 − ξ)/E = max(2ξ − 1. we can ﬁnd the numerical domain of this quantity. we can choose a ﬁnite domain (xL . The boundary condition at the boundary points are an approximation to the above far ﬁeld boundary condition. For numerical computation. The transformation in this section converts S to ξ with ξ ∈ (0. 5. 0). = 1 − ξ dS E We plug this transformation to the BS equation. The initial data reads V (ξ. which is the triangle {(j∆x. the domain of x is the whole real line. t) V = S+E τ = T − t. NUMERICAL METHODS u(x. xR ) with xL << −1 and xR >> 1. 1). 0) = 1−ξ Λ E Eξ 1−ξ . m∆t)  j − k ≤ n − m} We only need to compute u in this domain. Then the resulting equation is 1 ∂ 2V ∂V ∂V = σ 2 (ξ)ξ 2 (1 − ξ)2 2 + rξ(1 − ξ) ¯ − r(1 − ξ)V . we don’t even use this boundary condition. n∆t).10) (5. Deﬁne σ (ξ) = ¯ σ(Eξ/1 − ξ).7) (5. CHAPTER 5. But this is not a problem for numerical computation. 0) = max(S − E.62 for a call option.8) Notice that ξ is dimensionless and important values of ξ are near 1/2. t) = 0 as x → ∞ In computation. and u(−∞.6) (5. However.
Let ξj = j∆ξ.5.11) (5. V (ξ. FAST ALGORITHMS FOR SOLVING LINEAR SYSTEMS 63 Similarly. Many options have nonsmooth payoff functions. 0) = max(1 − 2ξ. 5. The stability constraint is ∆τ 1 2 2 ∆τ 2 2 ≤ σj ξj (1 − ξj )2 ≤ 1. If the solution is smooth up to the boundaries. The resulting ﬁnite difference equation reads dvj dτ = 1 2 2 vj+1 − 2vj + vj−1 σj ξj (1 − ξj )2 2 ∆ξ 2 vj+1 − vj−1 +rξj (1 − ξj ) − r(1 − ξj )vj 2∆ξ We can discretize this equation in the time direction by forward Euler method. the equation is degenerate to the following ordinary differential equations: ∂V (0. we may subtract its nonsmooth part for which an exact solution is available. A = diag (−a. This causes low order accuracy for ﬁnite difference scheme. the diffusion coefﬁcients are degenerate. 0)erτ V (1. −c) . respectively.5. and a ﬁnite difference scheme can yield highorder accuracy.13) Remark. Let ∆ξ and ∆τ are the spatial and temporal mesh sizes. For the backward Euler scheme. Fortunately. τ ) = −rV (0. For instance. then on the boundary. (5. On the boundaries ξ = 0 and ξ = 1. The boundaries points are ξ0 and ξM . many simple payoff function has exact solution. we need to solve linear systems of the form AU = F. ∂τ The corresponding solutions are V (0.12) We can discretize equation (5. τ ) ∂τ ∂V (1. The remainder is smooth. τ ) = 0. rξj (1 − ξj ) − σj ξj (1 − ξj )2 2 ∆ξ 2∆ξ 2 (5. We use central difference for ∂ 2 V /∂ξ 2 and ∂V /∂ξ. the European call option. 0).9) by ﬁnite difference method. τ ) = V (1. τ ) = V (0. 0) for a put option. 1 + a + c. τ n = n∆τ . For general payoff function.5 Fast algorithms for solving linear systems In the backward Euler method and the CrankNicolson method.
the other is called iterative methods.... 5. 0 n+1 R 2 L n +Uj 2 R Now. The matrix A is tridiagonal and diagonally dominant. .. However. There are two classes of methods to solve the above linear systems. U = . . 1 − 2 a 2 c c − 2 . . We multiple the ﬁrst equation by −a and add it into the second equation to eliminate the term xjL +1 in the . NUMERICAL METHODS −c 0 ··· 1+a+c −c 0 −a 1 + a + c −c . cj ). the constants a. we concentrate on solving the linear system Ax = f.. . F = .64 1+a+c −a 0 := CHAPTER 5. c Uj 0 . and bn+1/2 = a n+1 n Uj +Uj L . 1 + 2 a 2 c c + 2 . Here. we may normalize the j − th so that bj = 1. − 2 ) B = diag ( a . iterative methods are better. . We say that A is diagonally dominant if b > a + c. c). We may assume b > 0. A may takes the form A = diag (aj . .5. direct method is usually better. . c). More generally. 1.. 2 ). . b. bj . One is called direct methods. for highdimensional cases. ··· 0 −a ··· 0 ··· ··· 0 . . c are different from the average weights we had before.. n UjR −1 cUjn+1 R For the CrankNicolson scheme We have AU n+1 = BU n + bn+1/2 where A = diag (− a . and bj  > a + cj .1 Direct methods Gaussian elimination Let us illustrate this method by the simple example: A = diag (a.. Without loss of generality. Let us rewrite A = diag (a. 1+a+c −a 0 ··· −c 1+a+c −a 0 −c 1+a+c and n+1 aUjL Ujn +1 L . For onedimensional case as we have here. b.
. . 0 .. . .. ··· bjL +1 xjL +1 0 0 xjL +2 bjL +2 0 xjL +3 = bjL +3 . Thus. . . . . . LUdecomposition is equivalent to the Gaussian elimination.. j +2 0 1 . we arrive 1 c 0 0 0 1 − ac c 0 0 0 1 − c/(1 − ac) c 0 0 0 0 0 0 0 ··· ··· ··· . u and v’s... 1 xjR −1 bjR −1 65 Then xj can be solved easily. . . Consider three consecutive equations ax2j−2 + x2j−1 + cx2j = b2j−1 ax2j−1 + x2j + cx2j+1 = b2j ax2j + x2j+1 + cx2j+2 = b2j+1 . If we watch carefully. U x = y. . . . . Finally. The diagonal dominance condition guarantee that the reduced matrix is also diagonally dominant. . . .. U = 0 . . vjR −2 ujR −1 It is easy to ﬁnd a recursion formula to ﬁnd the coefﬁcients .. .. . ··· 1 xjR −1 bjR −1 We continue to eliminate the term a in the third equation.. L= 0 0.. 0 . .0 . Once these are found.. this scheme is numerical stable. 0 vjL +1 ujL +2 . Then the resulting equation becomes xjL +1 bjL +1 1 c 0 0 ··· 0 0 1 − ac c 0 · · · 0 xjL +2 −abjL +1 + bjL +2 0 bjL +3 a 1 c · · · 0 xjL +3 = . These two equations are easy to solve. FAST ALGORITHMS FOR SOLVING LINEAR SYSTEMS second equation.. c) to illustrate this method. 0 ··· . .. . we can ﬁnd x by solving Ly = b.. One can show that both L and U are diagonally dominant if A is. . .5. . . and so on. LU decomposition We decompose A = LU .. . L . ···0 0 . . where ujL +1 1 0 0 ··· 0 .. . 1. 0 · · · 0 jR −1 1 0 0 . Cyclic reduction method Let us take the case A = diag (a.5.... ..
5. the matrix A is almost an identity matrix. Since A is diagonally dominant.5. NUMERICAL METHODS We can eliminate the oddindex terms x2j−1 and x2j+1 . −a × (2j − 1)−eq +(2j)eq −c × (2j + 1)eq: After normalization. wer decompose A=D+B where D is the diagonal part and B is the off diagonal part. 1 − 2ac 1 − 2ac bj = (b2j − ab2j−1 − cb2j+1 )/(1 − 2ac). Jacob method In Jacobi method. One can show that the iterative mapping a a → c c converges to (0. where xn is deﬁned by the following iteration scheme: Dxn+1 + Bxn = b. A careful reader should ﬁnd that the cyclic reduction is also a version of the Gaussian elimination method. for few iteration. the oddindex x + 2j + 1 can be found from the equation: ax2j + x2j+1 + cx2j+2 = b2j+1 . we may approximate x by the sequence xn . c =− . we obtain a x2j−2 + x2j + c x2j+2 = bj Here. Then we have A x = b . We can invert it trivially. where A = diag (a . Let the error en := xn+1 − xn .66 CHAPTER 5. provided a + c < 1 initially. then solve an important and treat the rest as a perturbation term. Notice that the system is reduced to half and with the same form. Thus. Once x2j are found. c ). 1. Namely. 0)t quadratically fast.2 Iterative methods Most iterative methods can be viewed as a proper decomposition of A. Then Den = −Ben−1 Or en = −D−1 Ben−1 . If we rename xj = x2j . a =− a2 c2 . .
FAST ALGORITHMS FOR SOLVING LINEAR SYSTEMS Let us deﬁne the maximum norm en := max en  j j 67 Then a c n−1 en  =  − en−1 − ej+1  j j−1 b b a n−1 c n−1 ≤  e + e b b a + c n−1 = e b Hence. en ≤ ρ en−1 where ρ = a+c < 1 from the fact that A is diagonally dominant. GaussSeidel method In GaussSeidel method. . Show the above statement (5. the upper triangular part of A. the error en := xn+1 − xn satisﬁes en = −(D + L)−1 U en−1 To analyze the decay of en .14) This shows that the GaussSeidel method also converges for diagonally dominant matrix. j Then we have en (ξ) = G(ξ)en−1 . Let en (ξ) := j en e−ijξ . As before.5. The approximate solution sequence is given by (D + L)xn+1 + U xn = b. This yields the converb gence of the sequence xn . Exercise. L. ceiξ G(ξ) = − b + ae−iξ It is easy to see that the ampliﬁcation matrix G satisﬁes max G(ξ) := ρ < 1. and U . The limit x satisﬁes the equation Ax = b. provided b > a + c. ξ (5. where D is the diagonal part.5. we use Fourier method. the lower triangular part. A is decomposed into A = (D + L) + U .14).
ω is a parameter. .68 Successive overrelaxation method (SOR) CHAPTER 5. NUMERICAL METHODS In the methods of Jacobi and GaussSeidel. Find the ampliﬁcation matrix Gω and the optimal ω for the matrix A = diag (a. determine the rate ρ := min max Gω (ξ). c). We also need to require ω < 2 for stability. The optimal ω is chosen to minimize the ampliﬁcation matrix Gω (ξ). y n+1 = −D−1 (L + U )xn + b xn+1 = xn + ω(y n+1 − xn ). In order to speed up. Exercise. b. We therefore have a chance to speed them up by an extrapolation procedure described below. ω ξ Multigrid method Probably the most powerful method in higher dimension is the multigrid method. Also. we require ω > 1. Here. the approximate sequence xn is usually convergent monotonely.
t).Chapter 6 American Option 6. We explain why it is so below. Otherwise. The ﬁgure below is the value of a European put. The ﬁrst important thing we should note is that the value of an American option is greater than or equal to the payoff function: V (S. We recall that the value of an American call option is equal to that of a European call option. P E Ee−r(T−t) x x x E S 69 . there is an arbitrage opportunity because we can buy the American option then sell it immediately to gain a net proﬁt V − Λ. the American options do cost more. However. t) ≥ Λ(S.1 Introduction An American option has the right to exercise any time during the life of the option. for other cases like the the American put option or the American call option on dividendpaying asset.
t) < max{E − S. t)plane where C(S. there would be an arbitrage opportunity. t) ≥ max{S − E. In this region. Otherwise. And for S > Sf (t). the corresponding American option must be higher than the European option. then exercise it immediately. We make a riskless proﬁt: E − P − S > 0. Its value is shown in the Figure below. one should exercise the put option. In other words. t). which maximize the payoff function E − S. 0}. if we could exercise the call option. 0) for some Sf (t). Let us take the American put option as an example. there is a Sf (t). Hence the corresponding American call option should also satisfy C(S. Another example is the American call option on a dividendpaying asset. we should hold the option for all possible S. t) ∼ Se−D0 (T −t) for S >> 1. t) ∼ Se−D0 (T −t) for large S. 0}. Since C(S. if S < Sf (t). P x E S C(S. 0} in some region in the St plane. t) < max(S − E. 6. AMERICAN OPTION Notice that P (S. Then this option is identical to a European option. hence C(S.1 American options as a free boundary value problem American put option We can view an American option as a free boundary value problem.2 6.70 CHAPTER 6. First.2. then based on the same argument above. But we have seen that this is not the case. we should hold the . otherwise for S. In this region. there must be some value of S for which it is optimal from the holder’s point of view to exercise the American option. we can buy a put P (S. there is a region in (S. t) < max{S − E.
we do not know Sf (t) a priori. then we can move down Sf (t) to another Sf (t) < Sf (t) ∂S ˜ where ∂P (Sf (t). This Sf (t) is referred as the optimal exercise price. At Sf (t). ∂S Reasons. First. t) > −1. t) is continuous across (Sf (t). We prove this by contradiction. Now. t) satisﬁes the BlackSholes equation for S > Sf (t). t) ≥ max{E − S. t) drops below ∂S the payoff E − S. This is a contradiction. There exists an optimal exercise price Sf (t) such that 1. both P and ∂P ∂S are continuous. 2. Remark. we should exercise the American put option because the corresponding payoff Λ = max{E − S. for S < Sf (t). for S < Sf (t). the second boundary condition is equivalent to saying that ∂P (Sf (t). ∂S ˜ If ∂P (Sf (t). For S < Sf (t). t) ≡ max(E − S. 2 ∂t 2 ∂S ∂S 3. 1. In this movement. we should have P (S. We should treat Sf (t) as a new unknown (called free boundary and we should impose boundary condition to determine it. t) > −1. t) = E − S. P )plane with P (Sf (t). 0) for S < Sf (t). (b) Suppose ∂P (Sf (t). 2. This curve moves up as Sf (t) moves down. Thus. t). t). Thus. the payoff E − Sf (t) is higher than E −Sf (t). Thus. for S > Sf (t). then P (S(t + ∆t). ∂P (S. (a) If ∂P (Sf (t). and the corresponding ∂P (Sf (t).6. the value of the put option should be the payoff function Λ = E −S. one should hold the put option and P satisﬁes the BlackScholes equation: ∂P 1 2 2 ∂ 2P ∂P + σ S + rS − rP = 0. t).2. t) = E − Sf (t). 2. 0} is higher. t) = E − Sf (t). Its derivative in S is −1. across the free boundary (Sf (t). P satisﬁes the BlackScholes ∂S equation (for put option) and its solution curve should lie above the payoff function E − S on the (S. AMERICAN OPTIONS AS A FREE BOUNDARY VALUE PROBLEM 71 option. t) decreases. t) is continuous across (Sf (t). and P (S.) We claim that the proper boundary condition on Sf (t) are 1. and P (S. t) < −1 then as S increases from Sf (t). the curve stays above the payoff ∂S ˜ ˜ function. . t + ∆t) = P (S(t). t) with probability 1. t) = −1. P (S. t)/∂S is also continuous across (Sf (t). P (Sf (t). t). P (S. t). This would make an arbitrage opportunity by buying P at t then selling it at t + ∆t. early exercise is optimal. If S(t) = Sf (t) then S(t + ∆t) > Sf (t) with probability 1. This means that Sf (t) is not the optimal exercise price. if we exercise the put option for S ≤ Sf (t). this contradicts to P (S. 0}. for S > Sf (t). However. In other word. we treat the American put option as the following free boundary value problem. This follows from dS = S µdt + σdz and µ > 0. If P is discontinuous across (Sf (t).
and (iv). the boundary condition is required C(Sf (t). ∂S ∂V S). Π = V − ∆S. Therefore. we should have both C(S. t) = Sf (t) − E. ∂C (Sf (t). ﬁrst we notice that an American option should satisfy the following conditions: (i) V ≥ Λ. 0}. We summarize this by the following equations Ct + σ2 2 ∂ 2C ∂C S + (r − D0 )S − rC = 0.72 CHAPTER 6. On the free boundary S = Sf (t). 2 ∂S 2 ∂S (iv) both V and ∂V ∂S 2 2 are continuous. we should have dΠ ≤ r(V − ∂V S). then dΠ = r(V − Otherwise. 0). or Vt + 1 σ 2 S 2 ∂ V = r(V − S ∂V ) should hold. This value is below the payoff function Λ ≡ max(S − E. t) for 0 < t < T . (ii) Vt + 1 σ 2 S 2 ∂ V ≤ r(V − S ∂V ). AMERICAN OPTION 6. ∂S . where the free boundary is treated implicitly. t) ≡ Λ(S) ≡ max{S − E. there must an optimal Sf (t) such that we should exercise this call option when S > Sf (t) and hold it when S < Sf (t). V is the value of the American option. 2 ∂S 2 ∂S (iii) either V = Λ. based on the noarbitragy hypothesis. 0 ≤ t ≤ T. 0 < S < Sf (t) 2 ∂S 2 ∂S C(S. 2 ∂S When it is optimal to hold the option. t) = 1. We explain the reasons of (ii) below. Λ is the corresponding payoff function. On the free boundary S = Sf (t).2. We have seen the reasons for (i).3 American option as a linear complementary problem The American option can also be formulated as a linear complementary problem. t) on a dividendpaying asset has asymptotic value C(S. Let us consider the portfolio. t)/∂S are continuous the free boundary (S(t). To illustrate this linear complementary problem. t) ∼ Se−D0 (T −t) for large S. As we have seen that the Delta hedge eliminate the randomness of Π and yields ∂ 2V 1 dΠ = Vt + σ 2 S 2 2 . t) and ∂C(S.2 American call option on a dividendpaying asset As we have seen in the introduction of this chapter that an American call option C(S. S > Sf (t). ∂S 6. Here.
v > 1 − ex . We may replace v by ue−rτ to eliminate the term rv. The free boundary now in xvariable is xf (t). 2 2 are continuous. We can reformulate this problem in terms of x variable. the BlackScholes is replaced by the BlackScholes inequality. 0) = Λ(x) = 1 − ex . (ii) vτ − σ2 v 2 xx − (r − σ2 )vx 2 + rv ≥ 0. 2 2 (ii) for xf (t) < x < ∞. 2 ∂S 2 ∂S (iv) both V and ∂V ∂S 2 2 are continuous. As before. and vτ − σ2 σ2 vxx − (r − )vx + rv ≥ 0. AMERICAN OPTION AS A LINEAR COMPLEMENTARY PROBLEM 73 based on no arbitrage opportunities. 2 ∂S 2 ∂S (iii) (V − Λ) Vt + 1 σ 2 S 2 ∂ V − r(V − S ∂V ) = 0. S = Eex . (iii) (vτ − σ2 v 2 xx − (r − σ2 )vx 2 (iv) v and vx are continuous. The advantage of this formulation is that the free boundary is treated implicitly. To show (iii). The free boundary value problem is formulated as (i) for −∞ < x < xf (t). (ii) Vt + 1 σ 2 S 2 ∂ V − r(V − S ∂V ) ≤ 0. then V = Λ.6. τ = T − t. Such a problem is called a linear complementary problem. we hold the option and its value should satisfy the BlackScholes equation.3. and vτ − (iii) both v and ∂v ∂x σ2 σ2 vxx − (r − )vx + rv = 0. Properties (i)(iv) can be formulated as the following linear complementary problem: (i) V − Λ ≥ 0. Then we have uτ − σ2 σ2 uxx − (r − )ux (u − g) = 0 2 2 σ2 σ2 uτ − uxx − (r − )ux ≥ 0 2 2 . + rv)(v − (1 − ex )) = 0. we know that if we exercise the option. we use the following change of variables: V = Ev. v = 1 − ex . otherwise. Thus. with initial condition v(x. The linear complementary problem is formulated as: (i) v − (1 − ex ) ≥ 0.
D. τ ) → erτ . . q ≥ 0. on the set {u − g > 0}. uniqueness of the solution. τ ) → 0. It has the properties: (i) φN > 0. The far ﬁeld boundary conditions are u(x.4.4 6. (see reference: A. Let us demonstrate this theory brieﬂy. ux being continuous. From a standard theory of nonlinear P.D. Here g(x. p + q = 1.Variational Inequality). as x → ∞.E. for instance). the linear complementary problem can be discretized as n+1 n+1 er∆t Vjn − pVj+1 − qVj−1 Vjn − hn j = 0. 0). with u → 0 as x → +∞. u(x.1 Numerical Methods Projective method for American put We recall that the solution of the BlackScholes equation can be discretized by the following binomial method (or the forward Euler method): n+1 n+1 er∆ Vjn = pVj+1 + qVj−1 . Let us consider the following penalty function: φN (v) = −e−N v . A mathematical theory called parabolic variational inequality gives construction. 2 2 6. The boundedness of these two gives that uN has a convergent subsequence. 0) = g(x. with uNi . The method we shall use is called the penalty method. as x → −∞. φNi (uNi − g) → 0. Hence we have 1 σ2 uτ − uxx − (r − )ux = 0 on {u − g > 0}. We then need an estimate for uN and ∂uN .. Further. Similarly. existence. we conclude u − g ≥ 0.74 CHAPTER 6. (ii) φN (v) → 0 whenever v > 0. AMERICAN OPTION u − g ≥ 0. 0). u(x. φNi (uNi − g) ≤ constant As Ni → ∞. u → erτ as x → −∞. ∂ u . τ ) = max{erτ (1 − ex ). say uNi ∂x such that uNi → u. with u. Friedman. p.E. 0}.: 1 σ2 uτ − σ 2 uxx − (r − )ux + φN (u − g) = 0 2 2 u(x. ∂x Ni ux being continuous. (by monotone method. Moreover. 0) = g(x. We consider the following penalized P. one can show that the solution uN exists for all N > 0. u.
(The main tool to prove this is a theory for monotone operator. ∂S Vt + σ2 2 ∂ 2 V S ∂S 2 2 (iv) V and VS are continuous. where n∆t ∼ t and Eej∆x ∼ S. for large n. One can show that the scheme is monotone. Reference.) Furthermore.4. the limiting function satisﬁes the BlackScholes equation whenever V (S. t). j we have Vjn > Λn . we should require n+1 n+1 Vjn = max{e−r∆t (pVj+1 + qVj−1 ). . ∂S + (r − D0 )S ∂V − rV (V − Λ) = 0. 2 (∆x)2 2 2∆x q= σ 2 ∆t σ 2 ∆t − (r − D0 − ) . Math. Majda & Crandell. Hence.4. In this case. where p= σ 2 ∆t σ 2 ∆t + (r − D0 − ) . t) > Λ(S. The binomial approximation for the BS equation is n+1 n+1 er∆t Vjn = pVj+1 + qVj−1 . The regularity result (i. V ∞ . j Here. Λn is the discretized payoff function after changing variable. j VjN = ΛN . t) ≥ Λ(S) ≡ max{S − E. the payoff function at time t. j One can show that this method converges. Deﬁne n+1 n+1 Vjn = max{e−r∆t (pVj+1 + qVj−1 ). Λn }.2 Projective method for American call The linear complementary problem for this American call option is (i) V (S. t) > Λ(S. For American option. t).. continuity of V and VS ) follows from the theory of monotone operator. 75 Vjn − Λn ≥ 0. t). Hence. we always j have n+1 n+1 Vjn = e−r∆t (pVj+1 + qVj−1 ). We choose ∆t and ∆x so that p > 0 and q > 0. Thus the limiting function satisﬁes V ≥ Λ. from the construction. Vx ∞ are bounded. NUMERICAL METHODS n+1 n+1 er∆t Vjn − pVj+1 − qVj−1 ≥ 0. j This discretized linear complementary problem can be solved by the following projected forward Euler method.6. Λn }. At those points V (S. we have Vjn ≥ hn . j The above is the projective forward Euler method. 2 (∆x)2 2 2∆x and p + q = 1. V has to be greater than Λ(S. 6.e. 0} (ii) Vt + (iii) σ2 2 ∂ 2 V S ∂S 2 2 + (r − D0 )S ∂V − rV ≤ 0. Comp.
76 CHAPTER 6. 6. k = 0. ﬁrst we determine the up/down ratios u and d for the riskless asset price by pu + (1 − p)d = e(r−D0 )∆t . t) ≥ Λ(S) ≡ max{S − E. V n.(k+1) − V n.(k) ). u−d Then the binomial model is given by n Sj = 0 S0 = S. Sj − E}.(k) + ω(y n. or equivalently.5. and n+1 n+1 n Vjn = max{e−r∆t (pVj+1 + qVj−1 ). Λn }.(k+1) = max{V n.4. 1 −(r−D0 )∆t 2 A = (e + e(r−D0 +σ )∆t ). 2 e(r−D0 )∆t − d p = .(k+1) = (D + L)−1 (−U V n. y n. 0}. then we should require this condition hold in each iteration steps. 6.(K) This guarantees that V n+1 ≥ Λn+1 . · · · K V n ≡ V n. 2 2 ∂S ∂S We consider the American call option on a dividend paying asset: Vt + V (S. in the SOR iteration method.(k) + f n+1 ). For instance. with probability p with probability q.5 6. n−1 dSj+1 . √ u = A + A2 − 1. . d = 1/u. It is important to know that if an iterative method is used.1 Converting American option to a ﬁxed domain problem American call option with dividend paying asset σ2 2 ∂ 2V ∂V S + (r − D0 )S − rV = 0. q = 1 − p. AMERICAN OPTION For the corresponding binomial model. n−1 uSj−1 . we need to add the constraints un+1 ≥ g n+1 for American option.3 Implicit method For implicit method like backward Euler or CrankNicolson method. pu2 + (1 − p)d2 = e(2(r−D0 )+σ 2 )∆t .
77 where Sf (t) is the free boundary. To convert the free boundary problem to a ﬁxed domain problem. τ ). rE/D0 ) Before converting the problem. t) = Sf (t) − E. τ ) = 0. ∂ξ f sf (0) = max(1. t) = 1. ≤T ≤T (6. τ ). ∂S We also need a condition for Sf at ﬁnal time Sf (T ) = max(E. 0≤ξ 0≤τ 0≤ξ 0≤τ 0≤τ ≤ 1. T − τ ). 0) = 0. 0 ≤ τ ≤ T. τ ) = sf (τ ) − 1 − c(Esf (τ ). CONVERTING AMERICAN OPTION TO A FIXED DOMAIN PROBLEM 0 ≤ S ≤ Sf (t). ∂τ With the trivial initial condition yields u(0. ∂V (Sf (t).5. ≤ T. t) has exact solution. ≤ 1. the BlackSholes equation is degenerate to ∂u = −ru. we ﬁrst remove the singularity of the ﬁnal data (i. On this free boundary. 0 ≤ t ≤ T. t) − c(S.1) . Notice that c(S. T − τ ) . r/D0 ). τ ) = sf (τ ) 1 − (Esf (τ ). We may substract V by an European call option c with the same payoff data. u(1. The new variable V − c satisﬁes the same equation. ∂τ 2 ∂ξ 2 sf dτ ∂ξ u(ξ.e. nonsmoothness of the payoff function) as the follows. ∂S At the boundary ξ = 0. τ ) = g(sf (τ ). ∂c h(sf (τ ). t))/E sf (τ ) = Sf (t)/E The new equations for these new variables are ∂u = σ2 ξ 2 ∂ 2 u + (r − D0 ) + 1 dsf ξ ∂u − ru. yet it has smooth ﬁnal data. τ ) = h(s (τ ). τ ) = (V (S. the boundary condition is required V (Sf (t).6. we introduce the following changeofvariables: = S/Sf (t) ξ τ = T −t u(ξ. 0 ≤ t ≤ T. where g(sf (τ ). ∂u (1.
AMERICAN OPTION In practice. 0 ≤ η ≤ ηf (T ).78 CHAPTER 6. t) = 1. 0 ≤ t ≤ T . ∂η 0 ≤ t ≤ T. we can solve the modiﬁed BlackSholes equation (6. τ ) = h(sf (τ ). 0 ≤ η ≤ ηf (t). τ ). 0 ≤ t ≤ T. t) = ηf (t) − E. Through the changeofvariable η=E S u(η.1) with the boundary conditions u(0. ∂τ ∂ξ dτ with sf (0) = max(1. τ ) = . t) = 2 EP (S. 0 ≤ τ ≤ T ∂ξ We can differentiate the other boundary condition in τ and yield an ODE for the free boundary: ∂u ∂g dsf (1. 0) We can further convert it to a ﬁxed domain problem as that in the last section. 0). .t) S the inﬁnity domain problem is converted to a ﬁnite domain problem: ∂u 1 2 2 ∂ 2 u ∂u ∂t + 2 σ η ∂η2 − rη ∂η = 0. ∂u (ηf (t). 6. T ) = max(η − E.5. the BS equation is on the inﬁnite domain S > Sf (t).2) ∂u (1. r/D0 ).2 American put option For American put option P . ηf (T ) = max(E. 0≤t≤T u(η. τ ) = 0 0≤τ ≤T (6. u(ηf (t).
One particular binary option is the cashornothing call. S. and will be discussed in the next Chapter. Asian options 6. the payoff depends on the history of the asset prices. S. 79 . They are usually traded over the counter.1 Binaries The payoff function Λ(S) is an arbitrary function. we shall call these kinds of options. 1 2 log( S )−(r− σ )(T −t) 2 S 2 (T −t) 2σ where P(S . Their prices are usually not quoted on an exchange. compound options 3. barrier options 5.Chapter 7 Exotic Options Option with more complicated payoff then the standard European or American calls and puts are called exotic options. 7. whose payoff is Λ(S) = BH(S − E). t)Λ(S )dS = e−r(T −t) BN (d2 ). 1. We list some common exotic options below. etc. chooser options 4. for instance. We have seen its value is V = e−r(T −t) 0 ∞ P(S . the pathdependent options. T. the maximum. Lookback options In the last two. Binary options 2.. T. t) = − e 2πσ 2 (T − t) is the transition probability density for asset price in riskneutral world. the averages. This option can be interpreted as a simple bet on an asset price: if S > E at expiry the payoff is B. otherwise zero.
The payoff for the compound call option is max{C(S. Let us investigate the case calloncall. either a call or a put with exercise price E2 at time T2 . We consider the case where the underlying option is a vanilla put or call and the compound option is vanilla put or call on the underlying option. calloncall. Strike price : E2 . P (S. T2 . E2 ). putoncall. The payoff at T1 for this callonacallorput” is Λ = max{C(S. From this and payoff function at T1 . 2. T2 .3 Chooser options A regular chooser option gives its owner the right to purchase. EXOTIC OPTIONS 7. for an amount E1 at time T1 . 0}. There are four different classes of basic compound options: 1. Thus. putonput. it also must satisfy the same BlackScholes equation. T2 . The underlying option has value C(S. T1 . . we can value V at t. Strike price : E1 . there is no difﬁculty to value these complex chooser options. 7. 3. E2 ). We then solve the BlackScholes equation with payoff max{C(S. E2 ) − E1 . we have T1 < T2 .80 CHAPTER 7. 4. Other cases can be treated similarly. or by allowing the right to sell the vanilla put or call. 0}. it is a “call on a call or put”. T2 . T1 . The extension to more complicated option on more complicated option is relatively straightforward. At time T1 . Because the compound options value is governed only by the randomness of S. T1 ) − E1 . Certainly. T1 ) − E1 . according to the BlackScholes analysis.2 Compounds A compound option may be described as an option on an option. The compound option on this option Expiry : T1 < T2 . The compound option also satisﬁes the BlackScholes equation for the same reason as above. its value C(S. The contract can be made more general by having the underlying call and put with different exercise prices and expiry dates. T1 . 0}. The underlying option is Expiry : T2 . E2 ) − E1 . callonput. t. By using the BlackScholes formula for vanilla option.
4 Barrier option Barrier options differ from vanilla options in that part of the option contract is triggered if the asset price hits some barrier. For S > X. BARRIER OPTION 81 7. In the latter case. V (S. 3. The boundary conditions are V (X. t) = 0. The initial condition becomes u(x. 4. τ ) ∼ e(1−α)x−βτ as x → ∞. Let us ﬁnd its explicit solution. 0}. downandout: the option expires worthless if S reaches X from above before expiry. V = Ev. As well as being either calls or puts. σ 2 /2 We make another change of variable : v = eαx+βτ u. the option holder receives a speciﬁc amount Z for compensation. Let S = Eex . τ ) = 0 and u(x. t = T − (σ2τ/2) . The boundary condition becomes u(x0 . 1. The BlackScholes equation is transformed into vτ = vxx + (k − 1)vx − kv. say at some time prior to T . the option becomes a vanilla call. We choose α. barrier options are categorized as follows. where k = r . We assume X < E. 0}. or X = Eex0 . Let us consider the case of a European style downandout option without relate.4. V (S. β to eliminate the lower order terms in the derivatives of x: βeαx+βτ u + eαx+βτ uτ = α2 eαx+βτ u + 2αeαx+βτ ux + eαx+βτ uxx +(k − 1)(αeαx+βτ u + eαx+βτ ux ) − keαx+βτ u. t) ∼ S as S → ∞. S = X. 1 1 . upandout: the option expires worthless if S reaches X from below before expiry. Sometimes in the knockout options. 0) = u0 (x) = max{e 2 (k+1)x − e 2 (k−1)x . downandin: the option expires worthless unless S reaches X from above before expiry. (boundary condition). upandin: the option expires worthless unless S reaches X from below before expiry. it satisﬁes the BlackScholes equation. This implies that α = − 1 (k − 1) and β = − 1 (k + 1)2 and equation becomes uτ = uxx .4. The ﬁnal condition. or one can have rebate if the barrier is crossed. 7. 2.7.1 downandout call(knockout) A European option whose value becomes zero if S ever goes as low as S = X. T ) = max{S − E. 2 4 x Let x0 = log( E ). one can have boundary of time.
82 This follows from the payoff function being Λ = max{S − E.4. where u1 satisﬁes the heat equation with the initial condition: 1 1 e 2 (k+1)x − e 2 (k−1)x for x > 0 u1 (x. we may express V in terms of C as the follows. t). e2α(x−x0 ) C(x2 /S. 0} = max{e(−α+1)x − e−αx . we have x0 < 0.e. 0}. E and V = eαx+βτ u(x. t). 0) for x ≤ 0 u2 (x. for x0 < x < ∞ for − ∞ < x < x0 . 0) = 0 for x > 0 The solution u1 corresponds to C(S. 0) = u0 (x). 0} = E max{ CHAPTER 7. It is common for intype barrier option to give a rebate. We use methodofreﬂection to solve above heat equation with zero boundary condition. t). We reﬂect the initial condition about x0 as u(x. From this. For t < T . V (S. usually a ﬁxed amount. u → −u. Since C = Eeαx+βτ u1 is the vanilla call. t) = −u(x0 . t). The boundary condition for an “in” option is the follows. 0) = u0 (x) = e−αx max{ex − 1. At T . 0}. t) − ( The solution u2 is corresponds to S −(k−1) ) C(X 2 /S. we can obtain that u(x0 . We conclude V = C(S. the solution has the property: u(2x0 − x. t) = C(X. If S crosses the line S = X at some time prior to expiry. 0} = E max{ex − 1. t) → 0 as S → ∞. t) = 0. −u0 (2x0 − x). EXOTIC OPTIONS S − 1.. otherwise. From the uniqueness of the solution. and u0 (x) = 0 max{e(−α+1)x − e−αx . where the initial condition for u2 is the reﬂected condition from u1 : u0 (2x0 − x. t). T ) = 0. t) = −u(x. T − (σ2τ/2) ). we may write V = Eeαx+βτ (u1 +u2 ). Notice that because X < E. for x ≤ 0 Using this and the method of reﬂection. with u(x. X 7. i. then the option becomes a vanilla option. The option is worthless as S → ∞. This compensates the holder for the loss of the option. The equation and the initial condition are unchanged under the changeofvariable: x → 2x0 − x. if S > X. Since the option immediately turns into a vanilla call and must have . V (X.2 downandin(knockin) option An “in” option becomes worthless unless the asset price reaches the barrier before expiry. if the barrier is not hit. 0) = 0. then V (S. 0} for x0 < x < 0. First.
V (S. t) = c(X. 0}. The value of an asian option should depend on S. Λ(S. The payoff function is Λ(S. We may write V = c − V . 1 Λ(S. T ) = max{S − e 4. 0}.7. Then the boundary condition for V is V (S. t) = c − V ∼ S − 0 = S as S → ∞. T ) = c(S. which is deﬁned by t t T S(τ )dτ. In other words. For the lookback option. And V (X. t) − V (X. J = max S(τ ) 0≤τ ≤T I= 0 f (S(τ ). We only need to solve V for S > X. t) = C(X. T ) = Λ(S). an Americantype average strike option. 1. For S ≤ X. Λ(S. T ) = c(S. the payoff depends on I. where c is the value of a vanilla call. The only randomness is through S. t. ASIAN OPTIONS AND LOOKBACK OPTIONS 83 the same value of this vanilla call. geometric mean. 7. t) = max{S − t 3. t) = 0. We shall discuss this in the next chapter. We observe that V is indeed a “downandout” barrier option. V still satisﬁes the same BlackScholes equation for all S. In general.5 Asian options and lookback options In Asian options and lookback options. a Europeantype average strike option has the following payoff function 1 max{ST − T 2. t). It is important to notice that I(t) is independent of S(t). where f is a smooth function. 0}. it will be treated as a limiting case of an asian option. we shall see in the next chapter that dI = f dt. t s well as I. Lookback call. because its randomness is fully correlated to the randomness of S.5. V (S. τ )dτ. T ) − V (S. 0}. I). 0 RT 0 log S(τ )dτ . therefore V can be valued through a delta hedge. its value is the value of a vanilla call. For example. This is because one and only one of the two barrier options can be active at expiry and whichever it is. Indeed. . their payoff functions depend on the history of the underlying asset. 0 S(τ )dτ. 1(downandin) plus 1(downandout) equal to 1 vanilla call. T ) = max{S − J.
84 CHAPTER 7. EXOTIC OPTIONS .
τ ) = log S(τ ) for geometric mean. for example. 8. Notice that I(t) is a random variable and is independent of S(t).2 General Method Let f be a smooth function. The Asian options and the Russian options (Lookback options) are the typical examples. 0}. 0}.) Therefore. S dI(t) = f (S(t). 5. 0}. 1. average rate call option: Λ = max{ T 1 3. and each increment of S(τ ). is independent of S(t).Chapter 8 PathDependent Options 8. The payoff functions for these options are. t). geometric mean: the arithmetic mean T T 0 1 T T 0 S(τ )dτ. The stochastic differential equations governed by S and I are dS = µdt + σdz. In previous examples.1 Introduction If the payoff depends on the history of the underlying asset price. lookback rate put: Λ = max{E − max0≤τ ≤T S(τ ). I. deﬁne t I(t) = 0 f (S(τ ). 4. 85 . f (S(τ ). we should introduce another independent variable I besides S to value the derivative V (S. τ )dτ. average strike call option: Λ = max{S − 1 2. (This is because I(t) is the sum of increment of functions of S before time t. 0} T 0 S(τ )dτ above is replaced by e RT 0 log(S(τ ))dτ . S(τ )dτ − E. τ < t. lookback strike put: Λ = max{max0≤τ ≤T S(τ ) − S. such an option is called a pathdependent option. τ ) = S(τ ) for arithmetic mean and f (S(τ ). t)dt.
3 8. We have 1 ∂ 2S dΠ = dV − ∆dS = (Vt + σ 2 S 2 )dt + VI dI + VS dS − ∆dS. I) → λ(S. we expect that its solution is a function of the scaleinvariant variable R = I/S. Notice that this is also reﬂected in that dR = (1 + (σ 2 − µ)R)dt − σRdz. T 0 I Or.3. Since the initial data can be expressed as Λ = S max{1 − R/T. PATHDEPENDENT OPTIONS Notice that there is no noize term in dI. This reduces one independent variable. The only randomness is from dS. 2 ∂S ∂I ∂S 8. Plug V = SH into the above modiﬁed BlackScholes equation: 1 ∂H ∂ 2H ∂H ∂ SHt + σ 2 S 2 (2 +S )+S·S + rS (SH) − r(SH) = 0. I). 0). Λ(S. Its payoff function is deﬁned by 1 T max{S − S(τ )dτ. 2 ∂S ∂S 2 ∂I ∂S From ∂R ∂ R ∂ ∂ = .86 CHAPTER 8. I. 0}. Therefore. we can use delta hedge to eliminate this randomness. we 1 ∂2V ∂V ∂V Vt + σ 2 S 2 2 + f = r(V − S ). =− ∂S ∂S ∂R S ∂R 2 2I ∂ I 2 ∂2 1 ∂ ∂2 2 ∂ = + = 2 (2R +R ). 2 ∂V 2 We choose ∆ = arrive ∂V ∂S to eliminate the randomness in dΠ. ∂S 2 S 3 ∂R S 4 ∂R2 S ∂R ∂R2 . Namely. We notice that the modiﬁed BlackScholes equation Vt + S ∂V σ2 ∂ 2V ∂V + S 2 2 + rS − rV = 0.1 Average strike options European calls Let us consider an average strike call option with European exercise feature. we consider the portfolio Π = V − ∆S. t). T ) = max{S − T . we may also expect that V = SH(R. as before. in terms of I. From the arbitrage assumption. ∂I 2 ∂S ∂S and the initial data for the average strike options are invariant under the transformation: (S. Therefore. depends on R only. 0}.
Hence. 2 ∂R2 ∂R where Λ(R. t) = max{1 − R . • H(0.3. This implies that HR .3. The corresponding payoff function is S max{1 − R R R . t) + HR (0. t) is ﬁnite is equivalent to Ht (0. However. we have S(τ ) = 0. T ) = max{1 − R . I(t) = t t 0 S(τ )dτ. Next. in practice. t) is ﬁnite. T ). R2 HRR (0. we arrive Ht + The payoff function Λ(R. In this case.8. T σ2 2 ∂ 2H ∂H + (1 − rR) R = 0. 2 ∂R2 ∂R 87 we should require the boundary conditions.2 American call options We consider the average strike call option with American exercise feature. t) = 0. then H = O(log R). 0} = S(1 − ) ≡ S · H(R. Or (ii) if R ∂H = O(1) = 0 as R → 0. then V → 0. R(t) = I(t)/S(t). • H(∞. 8. HRR are 2 ﬁnite at (0. ∂R Both cases contradict to H(0.3. H is ﬁnte at (0. t) being ﬁnite. (ﬁnal condition) . we expect that the solution is smooth up to R =. That implies that Λ = 0. Hence the boundary condition H(0. 0} = H(R. 1 2 2 ∂ 2H ∂H Ht + σ R + (1 − rR) ≤ 0 2 2 ∂R ∂R H −Λ ≥ 0 1 2 2 ∂ 2H ∂H (Ht + σ R + (1 − rR) )(H − Λ) = 0. We have the following two cases: (i) If R2 ∂ H = O(1) = 0 as R → 0 then ∂R2 H = O(log R) for R → 0. we solve it by numerical method. then H → 0. 8. Since as R → ∞ implies S → 0. Consider a portfolio is C − P . 0}. for all τ with probability 1. T ). AVERAGE STRIKE OPTIONS we obtain 1 R ∂H 1 ∂H ∂ 2H SHt + σ 2 S 2 (2(− ) + S · 2 (2R + R2 )) 2 S ∂R S ∂R ∂R2 1 R +S 2 HR + rSH + rS 2 (− HR − rSH) = 0 S S Finally. and consequently. t). This equation with boundary condition can be solved by using the hypergeometric functions. T T T . we have RHR (0. t) are zeros as R → 0.3 Putcall parity for average strike option We study the putcall parity for average strike options with European exercise feature. 0} − S max{ − 1. When R = 0. t) = 0. t). t).
we consider H is of the following form H(t. dt dt 1 with a(T ) = 0. let us introduce J(t) = max0≤τ <t S(τ ). we only need to solve the equation with ﬁnal condition (i) H(R. since T the ﬁnal condition and the P. τ < t are independent of S(t). PATHDEPENDENT OPTIONS Since the BlackScholes equation in linear in R. we obtain the putcall parity: C − P = S(1 − 1 1 1 t S(τ )dτ ) (1 − e−r(T −t) − e−r(T −t) rT T S 0 S 1 t = S− (1 − e−r(T −t) ) − e−r(T −t) S(τ )dτ rT T 0 8. T ) = 1. Indeed. As before. T ) = − R . This suggest that we should introduce another independent variable J to value the lookback option in addition to S and t. H(R. rT and Consequently. We shall use the fact that for a continuous function S(·). T a(t) = − 1 (1 − e−r(T −t) ). Thus. t) ≡ 1. t 0≤τ ≤t max S(τ ) = lim n→∞ S(τ ) dτ 0 n 1 n .D. is linear in R. This differential equation can be solved easily: 1 b(t) = − e−r(T −t) . For instance. R) = a(t) + b(t)R Plug this into the equation. . the payoff function for a lookback option with European exercise feature is Λ = max S(τ ) − S(T ). b(T ) = − T . we obtain d d a + bR + (1 − rR)b = 0. 0≤τ ≤T Such an option is relatively expansive because it gives the holder an extremely advantageous payoff. it is dJ = 0.88 CHAPTER 8. Since S(τ ). we shall give a more careful approach.4 Lookback Option A lookback option is a derivate product whose payoff depends on the maximum or minimum of its underlying asset price. For (i). we expect that the solution is also linear in R. However.E. dt dt d d a + b = 0. b − rb = 0. We can derive a stochastic differential equation for J as before. For (ii). we see that J(t) is independent of S(t). (ii) H(R.
n Jn = In . For the minimum. This is because S ≤ J. LOOKBACK OPTION We leave its proof as an exercise. using the facts that Jn → J and ≤ 1. we can derive the equation for P (S. we have t n→−∞ 89 lim (S(τ )) dτ 0 t n 1 n = min S(τ ). for 0 ≤ t ≤ T . Jn . J. if S(t) = 0. we arrive 1 ∂ 2P ∂P Pt + σ 2 S 2 2 + rS − rP = 0. J. 0≤τ ≤t Let us introduce In = 0 (S(τ ))n dτ. The range for S is 0 ≤ S ≤ J. n−1 n Jn Now. This is consistent to the fact that dJ = 0.1 A lookback put with European exercise feature P (0.4. Next.d. We claim that Firstly. The asset price process becomes deterministic. we claim that ∂P (J. 8. t) = Je−r(T −t) . we have that Λ(0. then S(τ ) = 0 for t ≤ τ ≤ T . as before we consider the delta hedge: Π = P − ∆S. J. t) = 0. for Jn . The role of J here is only a parameter.8. T ) = max{J − S. From the arbitrage assumption. 0} = J. t): dΠ = Pt dt + 1 S n ∂P 1 ∂ 2P dt + σ 2 S 2 2 dt n−1 n Jn ∂Jn 2 ∂S ∂P = r(P − S)dt ∂S S J Taking n → ∞. ∂J .4. t) = Je−r(T −t) . Secondly. J. Therefore. 1 The s. Remark.e. the value of P is the discounted payoff: P (0. 2 ∂S ∂S This is the usual BlackScholes equation. t+dt dJn = ( 0 (S(τ ))n dτ ) n − ( 0 1 t (S(τ ))n dτ ) n 1 1 S(t)n = dt.
∂J P − Λ ≥ 0.2 Lookback put option with American exercise feature We have the following linear complementary equation. 2 ∂t 2 ∂S ∂S We require P .90 CHAPTER 8. . The value of P must be insensitive to a small change of J. For instance. T ) = J − S. where LBS = The ﬁnal condition P (S. Its solution is given by S P = S(−1 + N (d7 )(1 + k −1 )) + Je−r(T −t) N (d5 ) − k −1 ( )1−k N (d6 ).4. ∂S ∂J For lookback call option. J. J. ∂ σ2S 2 ∂ 2 ∂ + + rS − r. its payoff is Λ = S(T ) − min0≤τ ≤T S(τ ). the current maximum cannot be the ﬁnal maximum with probability 1. LBS P ≤ 0. (LBS P )(P − Λ) = 0. we simply replace max0≤τ ≤t S(τ ) by min0≤τ ≤t S(τ ). J where J d5 = [ln( ) − (r − S S d6 = [ln( ) − (r − J J d7 = [ln( ) − (r + S r k = σ 2 /2 √ σ2 )(T − t)]/σ T − t 2 √ σ2 )(T − t)]/σ T − t 2 √ σ2 )(T − t)]/σ T − t 2 8. The boundary condition ∂P (J. We can use method of image to solve this problem. ∂P ∂P are continuous. PATHDEPENDENT OPTIONS From µ > 0. J. T ) = Λ(S. t) = 0.
1.1.e.1 Deterministic bond model The value of a bond certainly depends on the interest rate. t)dt + w(r.2 Stochastic bond model Let us assume that the interest rate satisﬁes the following s. say r(τ ). the bond is called a zerocoupon bond. T ) be the bond value at t with maturity date T . If there is no coupon payment. The principal of a bond is called its face value. dB + k(t) dt = r(t)Bdt. T ) = e− RT t r(τ ) dτ T F+ t k(τ )e RT τ r(s) ds dτ . t ≤ τ ≤ T . T ) = F ( face value). k(t) be its coupon rate. it is unrealistic to assume that the interest rate is deterministic.1 Bond Models A bond is a longterm contract under which the issuer promises to pay the bondholder coupon payment (usually periodically) and principal (at the maturity dates). Let us ﬁrst assume that the interest rate is deterministic temporarily. 9. the holder receives coupon payment k(t) dt. Let B(t. the bond value is the sum of the present face value and the coupon stream. From the noarbitrage argument.d.: dr = u(r. we shall provide a stochastic model. This means that in a small dt.Chapter 9 Bonds and Interest Rate Derivatives 9. Thus. However. t)dz. is known. the bond value can be solved and has the following expression: B(t. In the next subsection. 91 . the life span of a bond is long (usually 10 years or longer). together with the ﬁnal condition: B(T. 9.
.t + uVi. σ Plug µi and σi back to this equation.r .r .r .92 CHAPTER 9. Let us express it as a known function λ(r. dΠ = rΠdt. r. This yields (µ1 − r)V1 /V1. or equivalently µ1 − r µ2 − r = . and drop the index i. t): µ−r = λ(r. 2 The function λ(r. From the noarbitrage argument. The change dΠ in a small time step dt is dΠ = dV1 − ∆dV2 . T1 ) − ∆V (t.r = (µ2 − r)V2 /V2. it is independent of T . r. we obtain 1 Vt + w2 Vrr + (u − λw)Vr − rV = 0.r + w2 Vi. then the random term is canceled in dΠ. since it gives the extra increase in σ expected instantaneous rate of return on a bond per an additional unit of risk. T2 ) ≡ V1 − ∆V2 . We shall discuss these issues later. t).r /V2. In the next section. while the righthand side is a function of T2 . w(r. This stochastic bond model depends on three parameter functions u(r. σ1 σ2 Since the lefthand side is a function of T1 . BONDS AND INTEREST RATE DERIVATIVES where dz is the standard Wiener process. To ﬁnd the equation for B with stochastic property of r. The drift u and the variance w2 are proposed by many researchers. Therefore.rr Vi 2 1 wVi. t) = µ−r is called the market price. t). σi = Vi We we choose ∆ = V1. we consider a portfolio containing bonds with different maturity dates: Π = V (t. t). where dVi = µi dt + σi dz. Vi 1 1 µi = Vi. We obtain µ1 V1 dt − ∆µ2 V2 dt = r(V1 − ∆V2 ) dt. t) and λ(r. we shall provide some model to determine them.
• Dothan (1978): dr = σrdz. 4 . While the Bessel process dx = /xdt + σdz has positive property. INTEREST MODELS 93 9. we shall illustrate a uniﬁed approach proposed by Luo. We notice that the OrnsteinUhlenbeck process dx = −ηxdt + σdz has the property to tend to its mean (which is 0) time asymptotically. • Merton’s model: we choose = η = 0. • HoLee (1986): dr = α(t)dt + σdz.2. • CoxIngersollRoss (1985) dr = β(α − r)dt + σr 2 dz. . α = (σ 2 + 2 )/(8η). 1 9. • CIR: β = 2η. We then design the underlying basic process is the sum of these two processes: dx = −ηx + x dt + σdz. η. The CIR and BK models have these properties. • Vasicek: = 0.9. We list some of them below. Below.2 Interest models There are many interest rate models. • MarshRosenfeld (1983): dr = (αrδ−1 + βr)dt + σrδ/2 dz. α(s) ds. t). t)) with x(t) governed by a simple stochastic process. The main requirements for an interest rate model are • positivity: r(t) ≥ 0 almost surely. we allow η. r = ex−σ • HoLee: = η = 0. r = x + α. t). r = x + t 0 2 /2t . and .e. Yen and Zhang.2. • Merton (1973): dr = αdt + σdz. In general. r = r(x(t). Then all interest models mentioned above correspond to different choices of r(x. (i. • Vasicek (1977): dr = β(α − r)dt + σdz. and σ are given functions of t. we can choose r = r(x(t). r = x + αt. With this simple process. r = 1 x2 . • mean reversion: r should tends to increase (or to decrease) and toward a mean.1 A functional approach for interest rate model The idea is to design r to be a function of x(t) and t. η = β. • Dothan’s model: = η = 0. • BlackKarasinski (1991): d ln r = (a(t) − b(t) ln r) dt + σdz.
In a small time step dt. i = 1. r. 0.94 CHAPTER 9. At S = 0 or r = 0. Ti ). the zerocoupon bond price V is given by V (x. t) = E e RT t r(s. The BlackScholes analysis for a convertible bond is similar to the analysis for a bond. the face value of the bond. BONDS AND INTEREST RATE DERIVATIVES • BlackKarasinski: = 0. Thus. We also have the boundary conditions: S→∞ r→∞ lim V (r. r = exp(g(t. S dr = udt + wdzr . and r(x. t) = nS lim V (r. t < T. it is a function of r. t. η = b(t). t) = 0.x(s)) ds  xt = x . σ(t). η(t). S. From the FeymannKac formula. S. we should require V (r. where g = x + t 0 a(s) ds. S. t) ≥ nS. 2. Consider a portfolio Π = ∆1 V1 + ∆2 V2 + ∆S S. Then we have V (r. Suppose the bond can be converted to nS at any time priori to T . With the interest rate model. the change of dΠ is dΠ = ∆1 dV1 + ∆2 dV2 + ∆S dS. t) or V (0. t)dt.3 Convertible Bonds A convertible bond is a bond plus a call option under which the bond holder has the right to convert the bond into a common shares. . There is no uniﬁed theory available yet with this approach and the approach of the previous subsection. x)). S. t). 9. t and T . S. Let the stochastic processes governed by S and r are dS = µdt + σdzS . Let Vi = V (r. t) to be ﬁnite. Suppose the correlation between dzS and dzR is dzS dzr = ρ(S. The ﬁnal value of the convertible bond V (r. V satisﬁes ∂ 1 ∂2 ∂ + 2 2 + −ηx + − r V = 0. S. T ) = F . S. ∂t σ ∂x x ∂x This model depends three parameter functions (t).
i . σr.2 V2 ) dzr .3.i and σS.2 V2 + ∆σS) dzS = 0 (∆1 σr.2 V2 ) dzr = 0.1 V1 + ∆2 σr. Vi µi = σS. 2 ∂t 2 ∂S ∂S∂r 2 ∂r ∂S ∂r .1 V1 + ∆2 σr. This means that (∆1 σS.9.1 (µ2 − r) = λS σS.t + S 2 Vi.i σr. And it yields dΠ = (∆1 µ1 V1 + ∆2 µ2 V2 + ∆µS) dt = r(∆1 V1 + ∆2 V2 + ∆S) dt. 1 σ2 w2 Vi.1 V1 + ∆2 σS. respectively.r .2 V2 + ∆σS) dzS + (∆1 σr.r . ∆1 (µ1 − r)V1 + ∆2 (µ2 − r)V2 + ∆(µ − r)S = 0. CONVERTIBLE BONDS where dVi = µi dt + σr.2 (µ − r) = λS σ The functions λr and λS are called the market prices of risk with respect to r and S. Plug the formulae for µi . Vi.i dzr + σS.S Vi 1 wVi. This equality together with the previous two give that there exist λr and λS such that (µ1 − r) = λS σS.SS + ρSwVi. = Vi 95 We choose ∆1 . we obtain the BlackScholes equation for a convertible bond: σ2 ∂2 w2 ∂ 2 ∂ ∂ ∂ ∂2 + S 2 2 + ρSw + + rS + (u − λr w) − r V = 0.Sr + Vi 2 2 1 = σSVi.S + uVi.i This implies dΠ = (∆1 µ1 V1 + ∆2 µ2 V2 + ∆µS) dt + (∆1 σS. Or equivalently.2 + λr σr.i dzS .1 V1 + ∆2 σS.rr + µSVi. ∆2 and ∆S to cancel the randomness terms dzr and dzS .1 + λr σr.
BONDS AND INTEREST RATE DERIVATIVES .96 CHAPTER 9.
then there is a Markov process whose transition probability is P . P ) be a probability space. X(t) is a random variable. if P is a function satisﬁes ChapmanKolmogorov equation. E = Rd . such that for each t ≥ 0. 97 . A markov process is characterized by its transition probability: P (t. x0 . s ≤ t} F t = σ{X(s). t1 . s. B).Appendix A Basic theory of stochastic calculus A. t2 . Conversely.7 If X is a Markov process.1 Brownian motion Let (Ω. B) = P (t0 . Let Ft = σ{X(s). A process is a function X : [0. then the corresponding transition probability P satisﬁes ChapmanKolmogorov equation: P (t0 . ∞) × (Ω. B) := P {X(s) ∈ B  X(t) = x} with initial distribution P {X(0) ∈ B} = ν(B). x. B is the Borel sets. s ≥ t} A process is called Markov if P (A  Ft ) = P (A  X(t)). x1 . ∀A ∈ F t . dx1 )P (t1 . B). Here. t2 . x0 . F) → (E. Theorem 1. F. This is equivalent to P {X(r) ∈ BFt } = P {X(r) ∈ BX(t)}∀r > t.
y) = 1 x − y2 exp(− ). Hence. E(Bt ) = 0. x. Vt2 − t is a martingale. we obtain E(Bt+s − Bt  Bt ) = 0. Bt is continuous in t. 0 ≤ s ≤ t} = Xt . Bt − Bs has normal distribution with mean 0 and variance t − s. 2. This means that if we know the value of the process up to time t and Xt = x. (2π(t − s))d/2 s−t Deﬁnition 1. (ii) E(Xu  Xs . 3. x. Deﬁnition 1. s. . it has independent increments: that is Bt −Bs is independent of Bu for all u ≤ s < t.7 Brownian motion: A process is called a Brownian motion(or a Wiener process) if 1. 1. E(Bt+s  Bt ) = E((Bt+s − Bt ) + Bt  Bt ) = E((Bt+s − Bt )  Bt ) + E(Bt  Bt ) = Bt 2 2.98 APPENDIX A. then the future expectation of Xu is x. Proof. 2. E(Bt ) = t < ∞. Theorem 1. y) = x − y2 1 exp(− ). From the fact that Bt+s − Bt is independent of Bt . (2π(t − s))d/2 s−t Such a distribution is called a normal distribution with mean x and variance s − t. Bt is a martingale. The Wiener process has the transition probability density function p(t.8 A process {Xt  t ≥ 0} is called martingale if (i) EXt < ∞.8 1. s. BASIC THEORY OF STOCHASTIC CALCULUS Two standard Markov processes are the Wiener process and the Poison process. for all s > 0. for all s > 0. It is easy to see that Bt is markovian and its transition probability is p(t.
However. b). Proof. B](a. Let us partition (0. Hence. V ar(Tn ) = V ar( = = 2 B(ti ) − B(ti−1 )2 ) var((B(ti ) − B(ti−1 ))2 ) (ti − ti−1 )2 = 2t2 2−n Hence. ti − ti−1  = t. we obtain 2 2 E(Bt+s  Bt ) = s + Bt . This means that lim B(ti ) − B(ti−1 ) = ∞. deﬁned by [B. Let Tn = We see that ETn = E(B(ti ) − B(ti−1 )2 )) = 2n i=1 B(ti )−B(ti−1 )2 . Theorem 1. t) = t almost surely. BROWNIAN MOTION 3. We can show that the Brownian motion has inﬁnite total variation in any interval. This implies E((Tn − ETn )2 ) → 0 and hence. a = t0 < t1 < · · · < tn = b is a partition of (a. t) evenly into 2n subintervals. and the limit is taken to be max(ti − ti−1 ) → 0.1. Tn → ETn almost surely. B](0. . ∞ n=1 V ar(Tn ) < ∞.9 We have [B. its quadratic variation. ∞ E n=1 (Tn − ETn )2 < ∞. 2 2 E(Bt+s − (t + s)  Bt ) = Bt − t.A. Here. From Fubini theorem. b) := lim B(ti ) − B(ti−1 )2 is ﬁnite. Use 2 Bt+s = ((Bt+s − Bt ) + Bt )2 2 = (Bt+s − Bt )2 + 2Bt (Bt+s − Bt ) + Bt 99 and the fact that Bt+s − Bt is independent of Bt .
(ii) t 0 Ef 2 < ∞. We have In 1 = 2 n B 2 (ti ) − B 2 (ti−1 ) − (B(ti ) − B(ti−1 ))2 i=1 n 1 2 1 = B (t) − 2 2 (B(ti ) − B(ti−1 ))2 i=1 We have seen that the second on the righthand side tends to 1 t almost surely. one can show that n→∞ t lim fn (s) dB(s) 0 almost surely. We deﬁne Itˆ ’s integral for f ∈ H 2 and being step functions. We shall require that f ∈ H 2 .2 Stochastic integral t We shall deﬁne the integral 0 f (s)dB(s). we can deﬁne its Itˆ ’s integral as the follows. the integral can be approximated by n t In = i=1 B(ti−1 )(B(ti ) − B(ti−1 ). 2. for s ≤ t. Using the fact that. o 1. For this kind of functions. BASIC THEORY OF STOCHASTIC CALCULUS A. The method can be applied with B replaced by a martingale. t 2 t E 0 g(s) dB(s) = 0 Eg(s)2 ds. Its Itˆ ’s integral is o o deﬁne by t n f (s) dB(s) = 0 i=0 f (ti−1 )(B(ti ) − B(ti−1 )). An typical example is 1 t B(s) dB(s) = B(t)2 − . 2 . or a martingale plus a function with ﬁnite total variation.100 APPENDIX A. for step functions g ∈ H 2 . 2 2 0 From the deﬁnition. which means: (i) f (t) depends only on the history Ft of Bs . We use above step functions fn to approximate general function f ∈ H 2 .
e.e. t 1 Lf := af + b2 f . y). Theorem 1. s) ds + 0 b(Xs . o With a ﬁxed s > t and f . t) = Ex. y)f (y) dy. 2 We shall give a brief idea of the proof.1) Xt − X0 = 0 a(Xs .t (f (X(s)). o . From Itˆ ’s formula. Tt f = p(t. STOCHASTIC DIFFERENTIAL EQUATION 101 A.1).(A.A.10 (Itˆ ’s formula) If X satisﬁes the s. ∆t = ti − ti−1 . s > t From the Markovian property. we deﬁne its generator as Lf := lim t→0 Tt f − f . For a semigroup Tt . s. x. A.t (f (X(s)).d.d. We have 1 f (X(ti−1 ) + ∆X) − f (X(ti−1 )) = f ∆X + f (∆X)2 + o((∆X)2 ). in terms of the transition probability density function p(t.4 Diffusion process For a s. t) := (Tt f )(x. we deﬁne u(x. This yields the Inˆ ’s foro mula. if it satisﬁes t t (A. then o 1 df (X(t)) = (af (X(t)) + b2 f (X(t)))dt + f (X(t))bdB(t). dX = adt + bdB. we deﬁne the associated semigroup Tt by Tt f = Ex. ti ). 2 t This follows from Itˆ ’s formula and E( 0 g(s) dB(s) = 0). one can show that Tt is a semigroup. let ∆B = B(ti ) − B(ti−1 ). x. Indeed.d. s)dB(s). 2 We notice that (∆X)2 = (a∆t + b∆B)2 = a2 (∆t)2 + 2ab∆t + b2 (∆B)2 ≈ b2 ∆t + o(∆t).e. t)dB(t) A Markov process X is said to be a strong solution of this s. t)dt + b(Xt . Plug ∆X and (∆X)2 into the Taylor expansion formula for f .3 Stochastic differential equation A stochastic differential equation has the form dXt = a(Xt .3. s. In a small time step (ti−1 .
y) = δ(x − y).1). t.t−h (u(X(t). s. t) := Ex.2) solves the backward diffusion equation: ut + Lu + gu = 0. x) = f (x).1). y)f (y) dy. Since u can be represent as u(x. satisﬁes the backward diffusion equation: ut + Lu = 0. u(x. then the associate u(x. and has ﬁnal condition u(s. s − h) → f (x) as h → 0+. τ ) dτ . Theorem 1.102 APPENDIX A. t) := Ex. Proof. with ﬁnal condition u(s.t f (X(s)) exp t g(X(τ ). First. we have X(s) → x as h → 0+ almost surely.e. x. t) − u(X(t − h).d. t − h) − u(x. and p(s. t) = as h → 0+.s−h f (X(s)). Hence u(x.t−h EX(t).12 (FeymannKac) If X satisﬁes the s. then the associate s p(x. we consider u(x. t) ds h t−h → Lu(x. we notice that u(x. t) h 1 Ex. s > t (A.t−h (u(X(t).e. t) = we obtain that p satisﬁes pt + Lx p = 0. (A. Now. For diffusion equation with source term. u(x.t (f (X(s)). s. s > t. BASIC THEORY OF STOCHASTIC CALCULUS Theorem 1.11 If X satisﬁes the s. Next. its solution can be represented by the following FeymannKac formula. (A.t (f (X(s)) = Ex. t)) This shifts ﬁnal time from s to t. s − h) = Ex. . t − h) = Ex. t)) h 1 t = Lu(X(s). x) = f (x).t−h (f (X(s)) = Ex.d.
As before. from Itˆ ’s o formula. we have t u(x. t). DIFFUSION PROCESS Proof. t − h) = Ex. h→0+ h lim . t)u(x. τ ) dτ = Ex. t) exp t−h g(X(τ ). t) + g(x. t)g(X(s). s).A.t−h t−h t d u(X(s). τ ) dτ .4. τ ) dτ · EX(t). Now.t−h EX(t).t f (X(s)) exp = Ex. t) s = Ex.t−h EX(t).t−h f (X(s)) exp t−h g(X(τ ). t) + u(X(s). we shift ﬁnal time from s to t: s 103 u(x.t f (X(s)) exp t t g(X(τ ). τ ) dτ − u(X(t − h).t−h u(X(t). t − h) − u(x. t − h) − u(x. t) exp t−h g(X(τ ). From this. it is easy to see that u(x. Here. τ ) dτ t−h s t g(X(τ ). t) exp t−h t g(X(τ ). s)) ds. we have used independence of X in the regions (t − h.t−h u(X(t). τ ) dτ = Ex.t exp t−h g(X(τ ). τ ) dτ s = Ex. t) and (t. t) = Lu(x. t) = Ex.t−h t−h (Lu(X(s).
This action might not be possible to undo. Are you sure you want to continue?