FEbook 1 MAIN

THE FACTOR APPROACH TO DERIVATIVE PRICING
The BIGPicture in a LITTLE Book

James A. Primbs
September 17, 2010
2
Contents
1 Basic Building Blocks and Stochastic Dierential Equation Models 1
1.1 Brownian Motion and Poisson Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Gaussian Random Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.3 Poisson Random Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.4 Poisson Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.5 Increments of Brownian Motion and Poisson Processes . . . . . . . . . . . . . . . . . . 3
1.2 Stochastic Dierential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 Dierentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.2 The Dierential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.3 Compound Poisson Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.4 Ito Stochastic dierential equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.5 Poisson Driven Dierential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2 Itos Lemma 11
2.1 Itos Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.1 The chain rule of ordinary calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.2 Itos lemma for Brownian motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.3 Replacing dz
2
by dt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.4 Discussion of Itos lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Itos lemma for Poisson Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.1 Interpretation of Itos lemma for Poisson . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 More versions of Itos Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.1 Itos Lemma for Compound Poisson Processes . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.2 Itos Lemma for Brownian and Compound Poisson Processes . . . . . . . . . . . . . . 16
2.3.3 Itos Lemma for vector processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4 Itos lemma, the product rule, and a rectangle . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3 Standard Stochastic Dierential Equations with Solutions 21
3.1 Geometric Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1.1 Stock Price Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Geometric Poisson Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2.1 A conditional lognormal version of geometric Poisson Motion . . . . . . . . . . . . . . 23
3.3 A jump-diusion model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.4 A more general SDE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.4.1 The Ornstein-Uhlenbeck Process and Mean Reversion . . . . . . . . . . . . . . . . . . 24
i
ii CONTENTS
3.5 Cox-Ingersoll-Ross Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4 The Factor Approach to Arbitrage Pricing 29
4.1 The Factor Approach to Arbitrage Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Returns and Factors Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2.1 Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2.2 Stochastic dierential equations and factor models . . . . . . . . . . . . . . . . . . . . 30
4.3 The Factor Approach to Arbitrage using Returns . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3.1 Arbitrage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3.2 Null and Range Space Relationship . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.3.3 A Useful Absence of Arbitrage Condition . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.3.4 Interpretations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.3.5 A Problem with Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.4 The Factor Approach using Price Changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.4.1 Price Changes and Arbitrage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.4.2 Prot/Loss and Arbitrage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.5 Two standard examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.5.1 Stocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.5.2 Futures contracts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5 Constructing a Factor Pricing Framework 41
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.2 A Classication of Quantities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.2.1 Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.2.2 Underlying Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.2.3 Tradables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.2.4 A Derivative is a Tradable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.3 Factor Models for Underlying Variables and Tradables . . . . . . . . . . . . . . . . . . . . . . 43
5.3.1 Direct Factor Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.3.2 Factor Models via Itos Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.4 Tradables tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.5 Applying the Price APT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.5.1 Relative Pricing and Marketed Tradables . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.5.2 Pricing the Derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.5.3 Underdetermined and Overdetermined Systems . . . . . . . . . . . . . . . . . . . . . . 45
5.6 Three Step Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6 Application of the Factor Form: Equity Derivatives 49
6.1 Examples from Equity Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6.1.1 Black-Scholes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6.1.2 Dividend Paying Stocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.1.3 Cash Dividends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.1.4 Poisson Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.1.5 Options on Futures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
6.1.6 Jump diusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.1.7 Exchange one asset for another . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
CONTENTS iii
6.1.8 Stochastic volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
6.2 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
7 Application of the Factor Form:
Interest Rate and Credit Derivatives 63
7.1 Notation and the Money Market Account . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.2 Interest Rate Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
7.2.1 Single Factor Short Rate Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
7.2.2 Multi-Factor Short Rate Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.2.3 Heath-Jarrow-Morton . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
7.2.4 The LIBOR Market Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
7.3 Credit Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
7.3.1 Defaultable Bonds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
7.3.2 Defaultable Bonds with Random Intensity of Default . . . . . . . . . . . . . . . . . . . 73
7.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
8 Hedging 77
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
8.2 Hedging from a Factor Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
8.2.1 Description Using a Tradables Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
8.2.2 The Relationship Between Hedging and Arbitrage . . . . . . . . . . . . . . . . . . . . 78
8.2.3 Hedging Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
8.2.4 Hedging under Incompleteness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
8.2.5 A Question of Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
8.3 Hedging from an Underlying Variable Sensitivity Perspective . . . . . . . . . . . . . . . . . . 83
8.3.1 Black-Scholes Hedging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
8.3.2 Hedging Bonds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
8.3.3 Derivatives imply Small Changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
8.4 Higher Order Approximations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
8.4.1 The Greeks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
8.4.2 A Delta-Gamma Hedge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
8.4.3 Determining what the error looks like . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
8.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
8.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
9 The Road to Risk Neutrality 89
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
9.2 Do the Factors Matter? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
9.2.1 Brownian Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
9.2.2 Poisson Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
9.3 Risk Neutral Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
9.3.1 Brownian Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
9.3.2 Poisson Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
9.4 Pricing as an Expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
9.5 Applications of Risk Neutral Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
9.5.1 How to Apply Risk Neutral Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
9.5.2 Black-Scholes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
9.5.3 Poisson Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
9.5.4 HJM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
9.5.5 Libor Market Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
9.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
9.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
iv CONTENTS
Notation
R - the real line (, ).
R
+
- the nonegative real line [0, ).
S(t)- Stock price at time t.
S
t
-
S
t
Partial derivative of S with respect to t.
S
x
-
S
x
Partial derivative of S with respect to x.
S
xx
-

2
S
x
2
Second partial derivative of S with respect to t.
z(t) - Brownian motion.
(t; ) - Poisson Process with intensity .
x
T
- the transpose of a vector x.
A
T
- the transpose of a matrix A.
B
0
(t) - the time t value of the money market account.
B(t|T) - time t price of a zero coupon bond with maturity T and face value $1 at time t
B(t|T
1
; T
2
) - time t forward price of a zero coupon bond with maturity T
2
and face value $1 when
delivery of the forward contract is at time T
1
.
r - a vector of returns.
r
0
- a constant risk free rate of interest (when allowed to be a function of time, see the short rate
process below).
r
0
(t) - the instantaneous short rate process at time t.
r(t|s) - the instantaneous forward rate at time t between times s and s +ds.
R(t|T) - spot rate for time T t at the current time t.
R(t|T
1
; T
2
) - forward interest rate between time T
1
and T
2
at time t.
F(t|T) - time t forward price for contract with delivery at time T.
f(t|T) - time t futures price for contract with delivery at time T.
S(t|{T
i
}) - time t swap rate for swap dates {T
i
}.
x(t) - limit from the left: x(t) = lim
ht
x(h).
x
- notation for limit from the left: x(t).

- market price of risk.
v
vi CONTENTS
Chapter 1
Basic Building Blocks and Stochastic
Dierential Equation Models
This chapter contains an introduction to the basic mathematics required for derivative pricing and nancial
engineering. We provide building blocks for modeling assets in the form of Brownian motion and Poisson
processes. With these two building blocks we create more complicated models by using Brownian motion and
Poisson processes to drive dierential equations (which are then known as stochastic dierential equations).
The presentation here is tutorial and heuristic. However, dont let that fool you. If you gain intuition from
it, then you have received a powerful tool to add to your toolbox for problem solving.
1.1 Brownian Motion and Poisson Processes
Brownian motion and Poisson processes are our fundamental building blocks for creating models of asset
prices. The key features are that Brownian motion has continuous sample paths (with probability 1), and
Poisson processes jump! We begin with Brownian motion which is built on the Gaussian random variable.
1.1.1 Gaussian Random Variables
An n-dimensional Gaussian (Normal) random variable is a random variable with density function:
X f
X
(x) =
1
(2)
n/2
||
1/2
exp
1
2
(x )
T
1
(x )
(1.1)
where R
n
is the mean and R
nn
is the covariance matrix:
= E[X], (1.2)
= E[(X )(X )
T
]. (1.3)
1.1.2 Brownian Motion
Brownian motion (also known as a Wiener Processes) is a stochastic process built upon the Gaussian random
variable as follows. A real-valued stochastic process z(t) : t R
+
is a Brownian Motion if:
1. z(0) = 0.
2. z(t) z(s) N(0, t s) for t > s.
3. z(t
2
) z(t
1
), z(t
3
) z(t
2
), . . . , z(t
n
) z(t
n1
) are independent for t
1
t
2
t
n
.
1
2CHAPTER 1. BASIC BUILDINGBLOCKS ANDSTOCHASTIC DIFFERENTIAL EQUATIONMODELS
You should remember the following facts about Brownian motion, as they make Brownian motion an ideal
building block for unpredictable but continuous asset price movements:
There exists a version of Brownian motion that has continuous sample paths.
Brownian motion is nowhere dierentiable with probability 1.
The rst property says that Brownian motion is appropriate for price processes that dont jump. In many
cases, price processes do jump, hence we will need to introduce the Poisson process next to model jumps.
The second property can be interpreted in the context of predictability. If a curve is dierentiable at a
point, then that means that locally it can be approximated by a line, with the slope of the line being the
derivative of the curve at that point. But this means that we can predict (to order dt) the future value of
the curve. In nance, we often want to assume that we cannot predict future prices. Non-dierentiability
indicates that in the sense mentioned above, future prices are not predictable.
Therefore, Brownian motion is an ideal building block upon which to build asset price processes. A
sample path of Brownian motion is given in Figure 1.1.
0 0.2 0.4 0.6 0.8 1
1
0.8
0.6
0.4
0.2
0
0.2
0.4
time
Sample Path of Brownian Motion
z
(
t
)
Figure 1.1: A typical sample path of Brownian motion.
Just as there are vector Gaussian random variables, we can dene a vector Brownian motion as follows.
A vector Brownian motion z(t) R
n
with covariance structure R
nn
is a stochastic process satisfying
1. z(0) = 0.
2. z(t) z(s) N(0, (t s)) for t > s.
3. z(t
2
) z(t
1
), z(t
3
) z(t
2
), . . . , z(t
n
) z(t
n1
1
t
2
t
n
.
Thus a vector Brownian motion is build upon the vector Gaussian random variable.
1.1. BROWNIAN MOTION AND POISSON PROCESSES 3
Brownian motion has continuous sample paths. That is too well behaved for some events we would like
to model. For instance, market crashes, bankruptcy, etc. are often discontinuous price movements. Hence,
we need a process that jumps! Poisson processes, which are built on the Poisson random variable, are what
we are looking for.
1.1.3 Poisson Random Variables
A discrete random variable X taking values in the whole numbers is Poisson with parameter > 0 if
P(X = k) =

k
k!
exp() k = 0, 1, ... (1.4)
The mean of a Poisson random variable is E[X] = and the variance is V ar(X) = .
1.1.4 Poisson Process
A Poisson process is a stochastic process built on Poisson random variables as follows. A Poisson process
with parameter (intensity) is a stochastic process (t; ) : t R
+
that satises
1. (0) = 0.
2. (t) (s) is Poisson distributed with parameter (t s) for t > s.
3. (t
2
) (t
1
), (t
3
) (t
2
), . . . , (t
n
) (t
n1
1
t
2
t
n
.
For us, the most important property of Poisson processes is that they jump! Hence, they are good models for
market crashes, jumps, bankruptcy, and other unexpected discontinuous price movements. A typical sample
path from a Poisson process with intensity = 1 is given in Figure 1.2.
The parameter is often called the intensity (or sometimes the propensity) of the Poisson process. You
can think of it as the expected number of jumps in a single time period. Alternatively, you expect to see a
single jump every
1
time periods. Therefore, the larger the intensity, the more frequent the jumps.
We will assume that a Poisson process is continuous from the right, and not the left. That is, at the
exact time that a Poisson process jumps, it takes on the new value that it jumped to. Functions that are
right-continuous and have left-limits are called rcll functions (or cadlag or R-functions, etc). In a Poisson
process, it is important to remember that at a jump time it takes on the new value, thus making sample
paths of a Poisson process rcll functions.
1.1.5 Increments of Brownian Motion and Poisson Processes
Here are the intuitive pictures that I keep in mind when thinking of Brownian motion and Poisson Processes.
Over a small t Brownian motion and Poisson Processes can be thought of in simple and intuitive ways.
We intuitively think of t as a small increment in t. When dealing with a stochastic process X(t), we will
also think of X(t) as the change the occurs in X over a small time period t. That is
X(t) = X(t + t) X(t). (1.5)
This notion of an increment of a stochastic process will guide our intuition. In this way, we can look at
increments of Brownian motion and Poisson processes.
Brownian Motion
Over a small time t, Brownian motion looks like
z(t) = z(t + t) z(t) N(0, t) (1.6)
0 1 2 3 4 5
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
time
(
t
)
Sample Path of a Poisson Processes with Intensity 1
Figure 1.2: A typical sample path of a Poisson process.
or, written slightly dierently
z(t) =
t where N(0, 1) (1.7)

where this follows from the second dening property of Brownian motion. That is, a Brownian motion
dierential looks like a standard Gaussian multiplied by
t. Thus, we will often use E[z] = 0 and

E[(z)
2
] = t.
An even simpler picture arises from a binary approximation
z

t w.p. 1/2
t w.p. 1/2
(1.8)
where w.p. stands for with probability. This is depicted in gure 1.3.
Figure 1.3: Binary model of an increment in Brownian motion .
1.2. STOCHASTIC DIFFERENTIAL EQUATIONS 5
Poisson Process
A Poisson Process can also be approximated over a small time period t. Over t it is
(t; ) = (t + t; ) (t; ) = X where X Poisson(t). (1.9)
A binary approximation to a Poisson process is
()
1 w.p. t
0 w.p. 1 t
(1.10)
A simple picture of this heuristic increment model is given in Figure 1.4. Note that for a Poisson process,
Figure 1.4: Binary model of an increment of a Poisson process.
is either 0 or 1 to order t.
Thus, note the key dierence between a Brownian motion and a Poisson process. From the simple binary
model approximations, we see that in a Brownian motion the size of the move scales with the square root of
t. Hence over short periods of time the move in a Brownian motion is also small. This is why Brownian
motion has continuous paths. On the other hand, in a Poisson process, from the binary model we see that
the move size can always be 1, regardless of how small t is. On the other hand, the probability of having
a jump of size 1 scales with t and is small if t is small. Thus, Poisson processes jump when they move.
Over small periods of time the probability of a jump is also small. This is the essential dierence between
Brownian motion and the Poisson process.
1.2 Stochastic Dierential Equations
A simple way to think of a stochastic dierential equation is as a dierential equation that is driven by a
stochastic process. We will use this point of view here. Also, to avoid the technicalities of stochastic calculus,
we will present a simple intuitive approach to stochastic dierential equations and stochastic dierentials.
1.2.1 Dierentials
Roughly speaking, the notion of a dierential or innitesimal of a process is the idea that in an increment
X(t) = X(t + t) X(t) (1.11)
we can take t to be innitesimally small. In such a case we would write
dX(t) = X(t +dt) X(t) (1.12)
where dt is just a little bit of t, to quote Gillespie [8]. However, (1.12) has a problem when it comes to
processes that jump that are assumed to be right continuous. The Poisson process is a good example.
The Problem with Jumps
We have to be very careful when a process has jumps and we assume right continuity of paths. Here is the
problem. Assume that a Poisson process is currently at 0. That is (t) = 0. Now, assume that a jump
occurs at time s. Our convention will be to assume that at the exact time of the jump, the Poisson process
jumps to 1. That is (s) = 1. This means that for us, Poisson processes, and all other processes with jumps,
will be continuous from the right. Hence,
lim
hs
(h) = (s) (1.13)
A picture of this situation is shown in Figure 1.5.
Figure 1.5: A jump at time s.
Since this is our convention, now lets consider dening the dierential of a Poisson process as
d(t) = (t +dt) (t) (1.14)
where dt > 0.
Now, we know that a jump occurred at time s, so intuitively we should have d(s) = 1. However, let us
take the limit of d(t) for any t (including s) as dt 0. We obtain
lim
dt0
d(t) = lim
dt0
(t +dt) (t) = (t) (t) = 0 (1.15)
where this calculation followed by right continuity as dened by equation (1.13). But this indicates that
never jumps! Something must be wrong!!
What is wrong is that we have assumed that is right continuous, and then when we add dt to the
current time, we are implicitly taking the limit from the right. Hence, we are guaranteed never to capture
the jump!
This is purely a problem that arises from our convention to assume that Poisson processes are right
continuous. If we had assumed left continuity, then we wouldnt have any problem. However, it is common
in the literature to assume processes are right continuous, so we have adopted this convention. Therefore,
we need to adjust our notion of a dierential of a stochastic process slightly to account for our convention.
1.2.2 The Dierential
The solution to the above problem is that for a dierential, we should think of the following
dX(t) = X(t +dt) X(t) (1.16)
where X(t) = lim
ht
X(h) is the limit from the left of X at time t. By using the limit from the left, we
make sure to capture jumps of the process, no matter how small dt is made.
1.2. STOCHASTIC DIFFERENTIAL EQUATIONS 7
We will develop this point of view (which unfortunately cant be made rigorous, but provides the proper
intuition). Hence, reviewing from above, with Brownian motion we would have:
dz(t) = z(t +dt) z(t) N(0, dt). (1.17)
Note that since Brownian motion has continuous sample paths, z(t) = z(t). However, for a Poisson process,
we should think of dierentials as
d(t; ) = (t +dt; ) (t; ) Poisson(dt) (1.18)
in order to make sure that we capture jumps.
Dont forget that we also have the binary model approximations of Figures 1.3 and 1.4. Those binary
models provide the proper intuition, and in both cases, sums of them will limit as Brownian motion or a
Poisson process. For the dierential, we simply replace t by dt in (1.8) and (1.10), giving
dz

dt w.p. 1/2
dt w.p. 1/2
(1.19)
and
d()
1 w.p. dt
0 w.p. 1 dt
. (1.20)
1.2.3 Compound Poisson Process
When Poisson processes jump, they jump up by 1. We can generalize this and allow them to jump randomly.
Let (t; ) be a Poisson process with jump times t
1
, t
2
, .... Construct a new process
Y
(t; ), by assigning
jump Y
1
at time t
1
, Y
2
at time t
2
, etc. where Y
1
, Y
2
, ... are iid random variables. This process can be
written as
Y
(t; ) =
(t;)
i=0
Y
i
. (1.21)
That is, at time t it is the sum of (t; ) iid copies of Y , where (t; ) is a standard Poisson process. Processes
of this form can also conveniently be written as integrals,
Y
(t, ) =
(t;)
i=0
Y
i
=
t
0
Y
s
d(s; ). (1.22)
For this reason, we represent the dierential form of a compound Poisson process by Y d(t, ). That is, we
may write
d
Y
= Y d. (1.23)
Following along the lines of the binary approximation to a Poisson process as in Figure 1.4, an innitesimal
model of a compound Poisson process can be thought of as
Y d()
Y
i
w.p. dt
0 w.p. 1 dt
(1.24)
and a heuristic innitesimal picture of this is given in Figure 1.6.
1.2.4 Ito Stochastic dierential equations
Stochastic integrals can be dened in dierent ways. The most useful for us is the Ito stochastic integral. At
this point, I will not delve into the depths of the stochastic integral (because often people are never able to
return!), but merely provide the intuition that you should take away when considering stochastic dierential
equations.
Figure 1.6: Innitesimal model of a compound Poisson process.
A stochastic dierential equation will be written as:
dx(t) = a(x(t), t)dt +b(x(t), t)dz(t) (1.25)
where in this case, it is being driven by Brownian motion z(t). (At this stage, I will ignore the technical
conditions that must be placed on a and b in order to make such an equation well dened.) We will interpret
this equation as follows:
x(t +dt) x(t) = a(x(t), t)dt +b(x(t), t)(z(t +dt) z(t)). (1.26)
Since z(t) has independent increments, and a(x(t), t) and b(x(t), t) are evaluated at time t, they are inde-
pendent of dz(t) = z(t + dt) z(t). This is important! It allows us to do the following simple calculations
of the instantaneous drift and variance.
Instantaneous Drift and Variance
We can interpret a(x(t), t) as related to the instantaneous drift and b(x(t), t) as related to the instantaneous
volatility as follows:
E[dx|x(t)] = E[a(x(t), t)dt +b(x(t), t)dz(t)|x(t)] (1.27)
= a(x(t), t)dt +b(x(t), t)E[dz(t)|x(t)] (1.28)
= a(x(t), t)dt. (1.29)
Therefore, a(x(t), t) determines the instantaneous drift. On the other hand, we can compute the instanta-
neous variance of x as follows
E[(dx a(x(t), t)dt)
2
|x(t)] = E[b(x(t), t)
2
dz(t)
2
|x(t)] (1.30)
= b
2
(x(t), t)E[dz(t)
2
|x(t)] (1.31)
= b
2
(x(t), t)dt (1.32)
Hence, b
2
(x(t), t) determines the instantaneous variance of x.
1.2.5 Poisson Driven Dierential Equations
We can also drive a dierential equation by a Poisson process
dx(t) = a(x(t), t)dt +b(x(t), t)Y d(t; ) (1.33)
where note that we have written x(t) in the arguments of a and b. By x(t) we mean x(t) = lim
h0
x(th).
That is, x(t) is the limit from the left at time t. We will assume that a and b are left continuous in the t
1.3. SUMMARY 9
argument so that we may use t instead of t in the second argument of a and b. We will also sometimes use
the notation x
when we want to suppress the argument t, or even a
when suppressing the arguments of

a. The reason for using limits from the left is that in a Poisson process, we interpret our dierential as
d(t) = (t +dt) (t)
and for the Ito integral, we assume that the coecients a and b are evaluated at the point in time that the
dierential starts from. This is t.
This limit from the left is also important in a and b because we want a and b to be independent of d.
The only way we can do this is to make sure that we use left limits. Note that this means that if (t) jumps
at time t, which also causes a jump in x at time t, we evaluate x(t) in a and b which immediately preceeds
the jump.
With that established, once again, we can compute the instantaneous mean and variance:
E[dx(t)|x(t)] = E[a(x(t), t)dt +b(x(t), t)Y d(t)|x(t)] (1.34)
= a(x(t), t)dt +b(x(t), t)E[Y d(t)|x(t)] (1.35)
= a(x(t), t)dt +b(x(t), t)E[Y ]dt (1.36)
Hence, in this case, the d(t) term can contribute to the instantaneous mean. This can make things messy!
It is often nicer to think of the rst term as the mean term, and the second as the risk term. To do
this, we would like the second term to have zero instantaneous mean. Hence, we will often compensate the
Poisson process to give it zero mean. This is done by simply subtracting o the instantaneous mean from
the second term and adding it to the rst.
dx(t) = (a(x(t), t) +b(x(t), t)E[Y ])dt +b(x(t), t)(Y d(t) E[Y ]dt) (1.37)
Then we can also compute the instantaneous variance:
E[(dx(t) (a(x(t), t) +b(x(t), t)E[Y ])dt)
2
|x(t)]
= E[b
2
(x(t), t)(Y d(t) E[Y ]dt)
2
|x(t)]
= b
2
(x(t), t)V ar(Y d(t))
= b
2
(x(t), t)(E[Y
2
d(t)
2
) E[Y ]
2
E[d(t)]
2
)
= b
2
(x(t), t)(E[Y
2
]E[d(t)
2
] E[Y ]
2
2
dt
2
)
= b
2
(x(t), t)(E[Y
2
](dt +
2
dt
2
) E[Y ]
2
2
dt
2
)
= b
2
(x(t), t)E[Y
2
]dt +O(dt
2
)
Hence, to order dt, the instantaneous variance is given by b
2
(x(t), t)E[Y
2
].
1.3 Summary
Brownian motion (built upon the Gaussian random variable) and the Poisson Process (built upon the Poisson
random variable) are the basic building blocks used to create models of prices. In particular, we use these two
processes to drive dierential equations and that will allow us to capture a wide range of price phenomena.
Due to the continuity of Brownian motion, it is good for modeling price paths and variables that do not
jump. On the other hand, Poisson processes are an essential building block for modeling jumps in price
processes or variables.
Much intuition can be gained from simple incremental and dierential models of processes and
stochastic dierential equations. The simple binary approximations to Brownian motion and Poisson pro-
cesses are enough to correctly guide your intuition in the vast majority of cases. Thus, for modeling purposes,
make sure you have a solid understanding of these two building block processes.
1.4 Problems
Problem 1.4.1 Verify that our innitesimal model of a Poisson process over small time dt:
d =
1 wp. dt
0 wp. 1 dt
(1.38)
has a mean and variance that agree with a Poisson random variable with parameter dt to order dt.
Problem 1.4.2 Poisson Processes
Consider the time interval [0, 1]. Chop this time interval into n parts of equal length. Over each interval
dene the independent and identically distributed random variables X
i
where
X
i
=
1 w.p. /n
0 w.p. 1 /n
(1.39)
Let
Y =
n
i=1
X
i
(1.40)
(a) What is Pr(Y = 0)?
(b) In your answer in (a), take the limit as n . What do you get?
(b) What is Pr(Y = 1)?
(c) Again take the limit. What is your answer?
(d) Now consider an arbitrary but xed k with k < n. What is Pr(Y = k).
(e) Again take the limit as n , and show that this converges to the Poisson random variable. (You
will probably want to use Stirlings formula n!
2e
n
n
n+
1
2
. This calculation is a bit tricky!)
(Note: In this problem we converge to a Poisson random variable with parameter since we took the time
interval to be 1. If the time interval is t, we will converge to a Poisson random variable with parameter t.
As a function of t, we arrive at a Poisson process.)
Problem 1.4.3 Poisson Processes again.
Consider the following Markov chain. Let the state space be the whole numbers x = 0, 1, 2, .... Consider
the following transition probabilities over the time instant dt:
Pr(x(t +dt) = n|x(t) = n) = 1 dt (1.41)
Pr(x(t +dt) = n + 1|x(t) = n) = dt (1.42)
Let p
n
(t) = Pr(x(t) = n).
(a) Write down a dierential equation for p
0
(t). (hint: to derive a dierential equation, consider the
amount of probability that ows into and out of the state x = 0 over time dt.)
(b) Assume p
0
(0) = 1 (that is, at time zero, x = 0 with probability 1). Solve the dierential equation for
p
0
(t).
(c) Derive a dierential equation for p
n
(t), n > 0. Given your answer in (a), solve for p
1
(t). Explain how
you could solve for p
n
(t) for any n.
(Note: again we have arrived at a Poisson process, but this time through Markov chain theory. A Poisson
process is an example of a continuous time Markov process, and the set of dierential equations you derived
is the forward equation for this process.)
Chapter 2
Itos Lemma
2.1 Itos Lemma
Itos lemma is the chain rule for stochastic calculus. In this chapter we present Itos lemma for Brownian
motion and Poisson processes. It has been said that all of math nance can be done with just the knowledge
of Itos lemma. To you, this means that you should make sure that you know (and understand) Itos lemma.
In what follows, we will present versions of Itos lemma for Brownian motion and Poisson processes.
2.1.1 The chain rule of ordinary calculus
In ordinary calculus, here is how the chain rule works in conjunction with a dierential equation. Let x(t)
follow the dierential equation
dx
dt
= a(x, t). (2.1)
Now consider a function of x(t) and t. Lets call this function f(x(t), t). Assuming that f is dierentiable,
we can ask what the derivative of f is. To calculate it, we simply apply the chain rule
df(x, t)
dt
=
f
x
dx
dt
+
f
t
= f
x
dx
dt
+f
t
(2.2)
where we are using the notation f
x
=
f
x
and f
t
=
f
t
. Finally, we can substitute in for
dx
dt
from (2.1), giving
df(x, t)
dt
=
f
x
a(x, t) +
f
t
= a(x, t)f
x
+f
t
. (2.3)
This is a fairly straightforward calculation. However, when dealing with stochastic dierential equations,
the simple chain rule of ordinary calculus does not work. The reason is simple. Brownian motion is not
dierentiable so we cant really take its derivative or the derivative of any function of Brownian motion.
Also, Poisson processes jump, and at these jumps it is not even continuous, let alone dierentiable. Thus,
for stochastic dierential equations we need to develop the correct mathematics for dealing with a function
of a variable that follows a stochastic dierential equation. The guiding result is known as Itos lemma.
Multivariables Taylor Series Expansions
Before diving into Itos lemma, you should make sure that you recall your multivariables Taylor series
expansions to second order. This is extremely important! Itos lemma for Brownian motion is basically just
a modied Taylor expansion to second order. So, lets recall up to second order, the Taylor series expansion
11
12 CHAPTER 2. ITOS LEMMA
of a function f(x, t) around a point (x
0
, t
0
),
f(x, t) = f(x
0
, t
0
) +f
t
(x
0
, t
0
)(t t
0
) +f
x
(x
0
, t
0
)(x x
0
) (2.4)
+
1
2
f
tt
(x
0
, t
0
)(t t
0
)
2
+f
xt
(x
0
, t
0
)(x x
0
)(t t
0
) +
1
2
f
xx
(x
0
, t
0
)(x x
0
)
2
+. . . (2.5)
Now, we will typically denote dx = xx
0
, dt = t t
0
, and df = f(x, t) f(x
0
, t
0
), so that a Taylor series
expansion is
df = f
t
dt +f
x
dx +
1
2
f
tt
dt
2
+f
xt
dxdt +
1
2
f
xx
dx
2
+. . . (2.6)
and the arguments of f
t
, f
x
, ... have been supressed.
In the multivariable case when x R
n
, the Taylor series expansion is
df = f
t
dt +f
x
dx +
1
2
f
tt
dt
2
+dx
T
f
xt
dt +
1
2
dx
T
f
xx
dx +. . . (2.7)
where
f
x
= [f
x1
, . . . , f
xn
], f
xx
=
f
x1x1
. . . f
x1xn
.
.
.
.
.
.
.
.
.
f
xnx1
. . . f
xnxn
(2.8)
A useful trick in the multivariable case is to note that
dx
T
f
xx
dx = Tr(dx
T
f
xx
dx) = Tr(f
xx
dxdx
T
) (2.9)
where Tr() is the Trace of a matrix (i.e. the sum of the diagonal elements).
Operationally, Itos lemma for Brownian motion boils down to nothing more than substituting dx =
adt +bdz in the Taylor expansion of (2.6) (or (2.7) in the multivariable case) for dx, interpreting terms such
as dz
2
, and then throwing away terms of order higher than dt.
If that sounds simple, then you are right. Lets see how it works in more detail.
2.1.2 Itos lemma for Brownian motion
Given the dierential of x(t), Itos lemma allows us to compute the dierential of a function of x(t) and t.
Hence, it is the chain rule for stochastic dierential equations. The following result is Itos lemma when
x(t) is a process governed by a stochastic dierential equation driven by Brownian motion.
() Itos Lemma for Brownian Motion:
Consider the stochastic dierential equation (SDE)
dx = a(x, t)dt +b(x, t)dz (2.10)
and let f(x, t) be a twice continuously dierentiable function of x and t. Then
df(x, t) = (f
t
+a(x, t)f
x
+
1
2
b
2
(x, t)f
xx
)dt +b(x, t)f
x
dz. (2.11)
Heuristic Proof: I will suppress the arguments of a and b for convenience. Consider writing the Taylor
expansion of df.
df = f(x(t +dt), t +dt) f(x(t), t)
= f
t
dt +f
x
dx +
1
2
f
tt
(dt)
2
+
1
2
f
xx
(dx)
2
+f
xt
dxdt +...
2.1. ITOS LEMMA 13
Next, we will substitute in for dx using dx = adt +bdz which gives
df = f
t
dt +f
x
(adt +bdz) +
1
2
f
tt
(dt)
2
+
1
2
f
xx
(adt +bdz)
2
+f
xt
(adt +bdz)dt +...
= f
t
dt +f
x
adt +f
x
bdz +
1
2
f
tt
(dt)
2
+
1
2
f
xx
(a
2
dt
2
+ 2abdtdz +b
2
dz
2
) +f
xt
(adt
2
+bdzdt) +...
Now we take a crucial step, and only keep terms up to order dt using the following logic. The standard
deviation of dz is of order
dt. Hence, we think of dz as being of order dt

1/2
and only keep terms up to
order dt yielding
df = f
t
dt +f
x
adt +f
x
bdz +
1
2
f
xx
b
2
dz
2
...
Finally we replace dz
2
by its expectation dt which leads to Itos lemma
df = (f
t
+af
x
+
1
2
b
2
f
xx
)dt +bf
x
dz. (2.12)
2.
In this derivation there were a couple of dubious steps. The most glaring was replacing dz
2
by its expec-
tation dt. Lets see why this was a reasonable thing to do ...
2.1.3 Replacing dz
2
by dt
Here is a simple argument as to why it is reasonable to replace dz
2
by dt.
If we can say that dz
2
= dt, then by integrating we would have
T
0
dz
2
=
T
0
dt = T. (2.13)
Hence, let us see if this makes sense. Lets approximate the integral above by the sum
S(T, t) =
T
t
1
i=0
(z((i + 1)t) z(it))
2
T
0
dz
2
. (2.14)
Now, the claim is that as t 0, then S(T, t) T. But note that S(T, t) is a random variable since it
involves z(t). Therefore, we are trying to show that the random variable S(T, t) converges to the constant
T. To do this, we will show that the mean of S(T, t) is equal to T, and its variance approaches zero. This
does the trick, since a random variable with zero variance must be a constant equal to its mean. (When we
show that the variance approaches zero, we are proving convergence in mean square, or L
2
(P).)
Computing the mean:
Okay, lets rst compute the mean of S(T, t):
E[S(T, t)] = E
T
t
1
i=0
(z((i + 1)t) z(it))
2
=
T
t
1
i=0
E
(z((i + 1)t) z(it))

2
=
T
t
1
i=0
t = T.
(2.15)
Hence, the mean is T.
Computing the variance:
Now lets compute the variance of S(T, t). By independent increments
V ar(S(T), t) = V ar
T
t
1
i=0
(z((i + 1)t) z(it))
2
=
T
t
1
i=0
V ar((z((i + 1)t) z(it))
2
). (2.16)
But since z((i + 1)t) z(it) is Gaussian with mean zero and variance t, we have
V ar((z((i + 1)t) z(it))
2
) = E[(z((i + 1)t) z(it))
4
] E[(z((i + 1)t) z(it))
2
]
2
= 3(t)
2
(t)
2
= 2(t)
2
.
Therefore as t 0, we have
V ar(S(T, t)) =
T
t
1
i=0
V ar((z((i + 1)t) z(it))
2
) =
T
t
1
i=0
2(t)
2
= 2Tt 0. (2.17)
Hence, the limit of the variance of S(T, t) is zero. That means that the limit (in L
2
(P)) is a constant and
equal to the mean T. Note that the above argument has much of the avor of the weak law of large numbers.
This is the essential argument that allows us to use dz
2
= dt and a simplied version of the argument behind
a real derivation of Itos lemma.
Lets see an example of Itos lemma applied to so-called geometric Brownian motion.
Example 2.1.1 Let dx = axdt + bxdz and consider f(x) = ln(x). Then, according to Itos lemma, f(x)
satises
df = (f
t
+axf
x
+
1
2
b
2
x
2
f
xx
)dt +bxf
x
dz (2.18)
= (a
1
2
b
2
)dt +bdz (2.19)
where we have used that f
t
= 0, f
x
=
1
x
, and f
xx
=
1
x
2
.
2.1.4 Discussion of Itos lemma
At the risk of overdoing an attempt to provide intuition behind Itos lemma, I will leave you with the
following thoughts.
In Itos lemma, we kept terms up to order dt in the Taylor expansion. We are used to doing things like
this from ordinary calculus, but now that we are dealing with stochastic processes, we might question this
step. In particular, the standard deviation of dz is dt
1/2
. This means that moves in dz are usually much
larger than dt. Why doesnt this dz term completely dominate and even allow us to ignore terms of order
dt?
Again, I appeal to the wisdom of Gillespie [8]. He gave the following explanation, which is based on the
story of the tortoise and the hare. As the story goes, the tortoise and the hare race each other. The tortoise
being slow, starts the race and steadily works his way toward the nish line. The hare, on the other hand,
is quick and jumpy. He is much faster than the hare, but runs forward and backwards and easily gets o
track. In the end of the story, the tortoise wins the race.
Now your asking, What does this have to do with Itos lemma and stochastic dierential equations?
Well, the deterministic dt drift term is like the tortoise. It marches forward at a constant dt rate. On the
other hand, the dz term is like the hare. It is quick, and jumps around like order dt
1/2
. But its direction
is random. Some of the time it jumps forward and other times, backwards. Together, both the dt and dz
terms contribute to the stochastic dierential equation and neither term is guaranteed to dominate, just like
we dont know whether the tortoise or the hare will win the race!
2.2 Itos lemma for Poisson Processes
Itos lemma for Brownian motion is more subtle that Itos lemma for Poisson Processes. (The key dierence
is that Poisson processes have sample paths of nite variation, and this allows us to dene the stochastic
integral pathwise. Hence, Itos lemma in this case is merely an application of the Lebesgue-Stieljies calculus.)
2.2. ITOS LEMMA FOR POISSON PROCESSES 15
() Itos Lemma for Poisson Processes:
Given a Poisson stochastic dierential equation (SDE)
dx = a(x
, t)dt +b(x
, t)d (2.20)
let f(x, t) be a continuously dierentiable function of x and t. Then
df(x, t) = (f
t
+a(x
, t)f
x
)dt + (f(x
+b(x
, t), t) f(x
, t))d. (2.21)
The real derivation of this is using Lebesgue-Stieljies calculus, but once again I will provide a nice heuristic
argument. This time when we consider the dierential df we have to be careful because x
t
jumps! Taylor
expansions work well as an approximation when dx is small, however with jumps dx can be large! Note how
this plays into our derivation.
Heuristic Proof: We start by writing out the dierential and substituting in for dx.
df = f(x(t +dt), t +dt) f(x
, t) (2.22)
= f(x(t) +a
dt +b
d, t +dt) f(x
, t) (2.23)
where a
= a(x(t), t) and b
= b(x(t), t). Now, we note that possible jumps come from d. I dont like
this term, so I will add and subtract a term that doesnt contain the jump
df = f(x
+a
dt +b
d, t +dt) f(x
+a
dt, t +dt) +f(x
+a
dt, t +dt) f(x
, t) (2.24)
Now the nal two terms dont have a jump in them, so I will approximate them using ordinary calculus.
df = f(x
+a
dt +b
d, t +dt) f(x
+a
dt, t +dt) + (f
t
+a
f
x
)dt +O(dt
2
) (2.25)
Now lets analyze the rst two terms
f(x
+a
dt +b
d, t +dt) f(x
+a
dt, t +dt) (2.26)

When their is no jump in d, this term is zero. No jumps occur with probability 1 dt. When there is a
jump we have d = 1 and
f(x
+a
dt +b
, t +dt) f(x
+a
dt, t +dt). (2.27)

This occurs with probability dt. Since the a
dt term in the argument is of order dt, then overall, we can

think of the eect of the a
dt term as overall being of order dt

2
. Therefore, as dt 0, to order dt overall
this term is replaced by
f(x
+a
dt +b
, t +dt) f(x
+a
dt, t +dt) f(x
+b
, t) f(x
, t). (2.28)
Hence, combining the above arguments gives
f(x
+a
dt +b
d, t +dt) f(x
+a
dt, t +dt) (f(x
+b
, t) f(x
, t))d (2.29)
which completes the derivation of Itos lemma. 2
2.2.1 Interpretation of Itos lemma for Poisson
Itos lemma for Poisson processes simply says that when the Poisson process doesnt jump, use ordinary
calculus. When it jumps, the move in f is determined completely by the jump. That is it! Lets see an
example of this.
Example 2.2.1 Let dx = axdt + (b 1)xd and consider f(x) = ln(x). By Itos lemma for Poisson
processes, f(x) satises
df = (f
t
+f
x
ax)dt + (f(x
+ (b 1)x
) f(x
))d (2.30)
= adt + (ln(bx
) ln(x
))d (2.31)
= adt + ln(b)d (2.32)
where we have used that f
t
= 0 and f
x
=
1
x
.
2.3 More versions of Itos Lemma
In this section I will present other useful versions of Itos lemma. Heuristic derivations follow along the lines
of those for Brownian motion and the Poisson process. It is worthwhile working your way through a couple
of them, and through some of the problems at the end.
2.3.1 Itos Lemma for Compound Poisson Processes
When we are using a compound Poisson process, Itos lemma is modied slightly as follows.
() Itos Lemma for Compound Poisson Processes:
Given an SDE
dx = a(x
, t)dt +b(x
, t)Y d (2.33)
Let f(x, t) be a continuously dierentiable function of x and t. Then
df(x, t) = (f
t
+a
f
x
)dt + (f(x
+Y b
, t) f(x
, t))d (2.34)
This version of Itos lemma can be derived in a manner similar to Itos lemma for Poisson processes.
2.3.2 Itos Lemma for Brownian and Compound Poisson Processes
If we combine Brownian motion and compound Poisson processes, then we have the following result.
() Itos for Brownian and Poisson
Given a Poisson stochastic dierential equation (SDE)
dx = a(x
, t)dt +b(x
, t)dz +Y d (2.35)
let f(x, t) be a twice continuously dierentiable function of x and t. Then
df(x, t) = (f
t
+a(x
, t)f
x
+
1
2
b(x
, t)
2
f
xx
)dt +b(x
, t)f
x
dz + (f(x
+Y, t) f(x
, t))d (2.36)
Again, this can be shown with arguments similar to those given above.
2.3.3 Itos Lemma for vector processes
If we have multiple Brownian motions, then we have the following results
() Itos for two correlated Brownians
Given the stochastic dierential equations
dx
1
= a
1
dt +b
1
dz
1
(2.37)
dx
2
= a
2
dt +b
2
dz
2
(2.38)
2.4. ITOS LEMMA, THE PRODUCT RULE, AND A RECTANGLE 17
with a
1
= a
1
(x
1
, x
2
, t), a
2
= a
2
(x
1
, x
2
, t), b
1
= b
1
(x
1
, x
2
, t), b
2
= b
2
(x
1
, x
2
, t) where z
1
and z
2
are two corre-
lated Brownian motions with instantaneous correlation coecient (i.e. E[dz
1
dz
2
] = dt). Let f(x
1
, x
2
, t)
be twice continuously dierentiable function of x
1
, x
2
and t. Then
df(x
1
, x
2
, t) =
f
t
+a
1
f
x1
+a
2
f
x2
+
1
2
b
2
1
f
x1x1
+
1
2
b
2
2
f
x2x2
+b
1
b
2
f
x1x2
dt +b
1
f
x1
dz
1
+b
2
f
x2
dz
2
(2.39)
Using vector notation we can generalize the above result
() Itos for vectors
Given a vector stochastic dierential equation
dx = adt +Bdz (2.40)
(2.41)
where x R
n
, a = a(x, t) R
n
, B = B(x, t) R
nm
and z R
m
a vector Brownian motion with
instantaneous covariance matrix E[dzdz
T
] = dt. Let f(x, t) be twice continuously dierentiable function of
x and t. Then
df =
f
t
+f
x
a +
1
2
Tr(f
xx
BB
T
)
dt +f
x
Bdz (2.42)
where
f
x
= [f
x1
, f
x2
, ..., f
xn
] and f
xx
=
f
x1x1
f
x1xn
.
.
.
.
.
.
.
.
.
f
xnx1
f
xnxn
.
2.4 Itos lemma, the product rule, and a rectangle
In Itos lemma for Brownian motion, the mysterious term is
1
2
b
2
f
xx
dt, which enters because we are forced
to keep the dz
2
term which is of order dt. Without this term, Itos lemma would follow from the intuition
of ordinary calculus.
Let me give a little argument (which appears in Rogers and Williams [12]) to try to convince you that
Itos lemma is actually more intuitive than ordinary calculus. The example I will use is the product rule. In
ordinary calculus, we have the familiar formula
d(uv) = udv +vdu. (2.43)
In fact, this is the beginnings of the integration by parts formula. Lets try to derive this formula by a
simple rectangle argument. Here it is.
Consider the quantity d(uv). This is a change in the product uv, that is
d(uv) = (u +du)(v +dv) uv (2.44)
We can think of this as the area of a rectangle with sides of length u + du and v + dv minus the area of a
rectangle with sides of u and v. The rectangle in Figure 2.1 shows a picture of this. Now, it is obvious the
area of d(uv) is equal to the sum of the three colored rectangles in the gure. That is
d(uv) = udv +vdu +dudv. (2.45)
Hence, it is natural to expect to see a term related to dudv! In fact, it is more mysterious that in ordinary
calculus we are able to ignore this term. Of course, that is because in ordinary calculus, roughly speaking,
the terms du and dv are of order dt and hence dudv is a higher order term.
Figure 2.1: Itos Rectangle
Let us see how far we can get from this simple rectangle. Let us take u = v = z(t). Then our rectangle
formula (product rule) says that
d(z
2
) = d(z z) = zdz +zdz +dz
2
. (2.46)
Note that the formula above is exact! Now, recalling that dz
2
should be replaced by dt as in the argument
of Section 2.1.3, we have
d(z
2
) = 2zdz +dt (2.47)
which is Itos lemma for f(z) = z
2
. This gives us a formula for d(z
2
). Now we can proceed and choose u = z
and v = z
2
. From our rectangle and the formula for d(z
2
) we have
d(z
3
) = d(z z
2
) = zd(z
2
) +z
2
dz +dzd(z
2
) (2.48)
= z(2zdz +dt) +z
2
dz +dz(2zdz +dt) (2.49)
= 3z
2
dz + 3zdt +o(t) (2.50)
where in the last step we have placed higher order terms in o(t) and replaced dz
2
by dt. This is Itos lemma
for f(z) = z
3
.
Continuing on in this manner, we can easily derive Itos lemma for any polynomial of any order in z(t)!
At that point, we are not a far cry from Itos lemma for C
2
functions, as they can be approximated by limits
of polynomials. Hence, you see that Itos lemma is actually quite natural if you just remember the rectangle.
Furthermore, now you should have an easy time remembering Itos product rule d(uv) = udv +vdu +dudv.
2.5 Summary
Itos lemma is the most important result in stochastic calculus for derivative pricing. There are dierent
versions for Brownian motion and for Poisson processes. You should be familiar with both. Roughly speaking,
for Brownian motion, since the standard deviation of dz is of order
dt, second order terms in dz are of

order dt and cannot be ignored. This gives an extra term in Itos lemma compared to the ordinary chain
rule of calculus. For Poisson processes, most of the time no jumps are occurring and ordinary calculus is
ne. However, when a jump occurs, it causes a corresponding jump in any function of the Poisson process.
Thus Itos lemma for Poisson processes is simply a combination of the ordinary chain rule plus noting that
when the Poisson process jumps, any function of it has a corresponding jump.
2.6. PROBLEMS 19
2.6 Problems
Problem 2.6.1 Consider the Stochastic Dierential Equation:
dx = a(x, t)dt +b(x, t)dz +Y d (2.51)
where z is Brownian motion and is a generalized Poisson process with intensity and random jumps of
size Y . Let f(x, t) be a twice continuously dierentiable function of x and t. Find a stochastic dierential
equation for df. (i.e. what is Itos lemma in this case).
Problem 2.6.2 Multidimensional Ito Formula for Brownian Motion:
Consider the following vector Ito process:
dx = dt +Kdz (2.52)
where x R
n
, R
n
, K R
nm
and z R
m
where z is a vector Brownian motion of m uncorrelated
one-dimensional Brownian motions. Let f(x, t) : R
n
R R be a twice continuously dierentiable function
of x and t. What is df? (Hint: Use the same Taylor series argument used in class, but this time you can
use E[dzdz
T
] = Idt where I R
mm
is the identity matrix.)
Problem 2.6.3 Let
dx = adt +bdz
1
(2.53)
and
dy = fdt +gdz
2
(2.54)
where z
1
and z
2
are correlated Brownian motions with correlation coecient . (i.e. E(dz
1
dz
2
) = dt)
(a) Use Itos lemma to nd d(xy).
(b) Use Itos lemma to nd d(x/y).
Problem 2.6.4 Itos Lemma Practice
(a) Given dx = xdt +
xdz, use Itos lemma to derive df where f(x, t) = x

2
+t.
(b) Given dx = xdt +d
, use Itos lemma to derive df for f(x, t) = t

2
x.
(c) Given
dx = (x +y)dt +xdz
1
(2.55)
dy = ydt +xdz
2
(2.56)
where E[dz
1
dz
2
] = dt. Use Itos lemma to derive df for a generic twice continuously dierentiable
f(x, y, t).
Problem 2.6.5 Compute
T
0
z
t
dz
t
(2.57)
where z
t
is Brownian motion. (Hint: consider Itos lemma for z
2
.)
Problem 2.6.6 (Counter-intuition) Let
dx = axdt +bxdz (2.58)
Assume that a > 0. That is, the growth rate of E[x
t
] is positive. Now consider 1/x
t
. Is it possible for
E[1/x
t
] to also have a positive growth rate? Give conditions on a and b for when this is possible. Intuitively,
provide an explanation for this.
Problem 2.6.7 Intuition behind Itos lemma for Brownian motion. (This intuitive look at Itos lemma was
communicated to me by Muruhan Rathinam.)
Let
dx = adt +bdz (2.59)
then from Itos lemma f(x, t) satises
df = (f
t
+af
x
+
1
2
b
2
f
xx
)dt +bf
x
dz. (2.60)
Recall our heuristic derivation based on a Taylor expansion:
df = f
t
dt +f
x
dx +
1
2
f
xx
(dx)
2
+... (2.61)
or
df = f
t
dt +f
x
(adt +bdz) +
1
2
f
xx
(a
2
(dt)
2
+b
2
(dz)
2
+ 2abdtdz) +... (2.62)
(a) Compute the mean of df to lowest order in dt using the terms of the Taylor expansion shown above.
(b) Compute the standard deviation of the terms that contain randomness in the above expansion. That
is, compute the standard deviation of the dz, dtdz, (dz)
2
terms separately. Which term has standard
deviation of lowest order in dt? (Hint: The fourth moment of a N(0,
2
) is 3
4
.)
(c) Use this analysis to argue for the plausibility of Itos lemma, (2.60).
Problem 2.6.8 Show that (2.39) follows from (2.42).
Chapter 3
Standard Stochastic Dierential
Equations with Solutions
In this chapter we review some stochastic dierential equations that have closed form solutions. These are
also some of the stochastic dierential equation models used for modeling asset prices and other relevant
nancial variables. In these solutions, note the important role that Itos lemma plays. Most importantly,
not many stochastic dierential equations have closed form solutions. Thus, these are stochastic dierential
equations that everyone should know!
3.1 Geometric Brownian Motion
Geometric Brownian motion is the following stochastic dierential equation where a and b are constants.
dx = axdt +bxdz. (3.1)
A closed form solution to geometric Brownian motion can be found as follows. By Itos lemma with f(x) =
ln(x) we have
d ln(x) = (a
1
2
b
2
)dt +bdz. (3.2)
Since a and b are constants, integrating gives
ln(x
t
) ln(x
0
) = (a
1
2
b
2
)t +bz
t
(3.3)
which implies that
x(t) = e
(a
1
2
b
2
)t+bzt
x(0). (3.4)
Note that geometric Brownian motion is a log-normal process. That is, from (3.3) in log-coordinates x(t) is
a Gaussian process with drift a
1
2
b
2
and volatility b. Figure 3.1 shows a typical sample path of geometric
Brownian motion.
3.1.1 Stock Price Interpretation
Geometric Brownian motion is the standard model for continuous asset price movements. It comes from the
following. Let x(t) be the price of a stock. Then over the time period dt, the return r on the stock x(t) is
given by
r =
x(t +dt) x(t)
x(t)
=
dx
x
. (3.5)
21
22 CHAPTER 3. STANDARD STOCHASTIC DIFFERENTIAL EQUATIONS WITH SOLUTIONS
Figure 3.1: Typical Sample path of geometric Brownian motion.
Geometric Brownian motion models this instantaneous return as
dx
x
= adt +bdz (3.6)
where a and b are constants. Thus, the instantaneous return r over the next dt is Gaussian with
E[r] = adt, V ar(r) = b
2
dt. (3.7)
Thus, this is a very natural model of asset prices.
3.2 Geometric Poisson Motion
Geometric Poisson motion is the equivalent of geometric Brownian motion, but being driven by a Poisson
process,
dx = ax
dt + (b 1)x
d. (3.8)
The reason for the term b 1 is as follows. If the current value is x(t) and a jump occurs, then we will jump
by an amount (b 1)x(t). Thus, we will jump to x(t) + (b 1)x(t) = bx(t). Hence, by writing (b 1) we
think of b as indicating the multiple of the current state that we will jump to if a jump occurs. That is,
a jump leads to the transition x(t) bx(t). Note that if we didnt use this convention, it would be a bit
messier.
Again, there exists a closed form solution that is obtained by changing to log coordinates. By Itos lemma
with f(x) = ln(x)
d ln(x) = adt + ln(b)d. (3.9)
Integrating leads to
ln(x(t)) ln(x(0)) = at +ln(b)(t) (3.10)
or
x(t) = e
at+ln(b)(t)
x(0) = e
at
b
(t)
x(0). (3.11)
Thus, this process is a Poisson process plus drift in log-coordinates.
3.3. A JUMP-DIFFUSION MODEL 23
3.2.1 A conditional lognormal version of geometric Poisson Motion
The following stochastic dierential equation uses a compound Poisson process with a log normal jump size.
This results in a conditional log-normal process that is dierent from geometric Brownian motion. It is
modeled as
dx = ax
dt + (Y 1)x
d (3.12)
where Y = e
Z
is a lognormal random variable.
Once again, a change to log coordinates facilitates nding a closed form solution. Using Itos lemma with
f(x) = ln(x) gives
x(t) =
e
at
(t)
i=1
Y
i
x(0). (3.13)
But since Y is lognormal, we can write it as
t
i=1
Y
i
= e
(t)
i=1
Zi
where Z
i
is normal. Hence we have
x(t) =
e
at+
(t)
i=1
Zi
x(0) (3.14)
which, conditional on (t), is a lognormal process.
3.3 A jump-diusion model
The following stochastic dierential equation uses a generalized Poisson process with a log normal jump size
and a Brownian motion. This results in a process that involves jumps and a diusion term. It is known as
a jump-diusion model.
dx = ax
dt +bx
dz + (Y 1)x
d (3.15)
where Y is a lognormal random variable. Again, using Itos lemma with f(x) = ln(x) gives the solution,
which is
x(t) = e
(a
1
2
b
2
)t+bz(t)
(t)
i=1
Y
i
x(0). (3.16)
But since Y is lognormal, we can write it as
(t)
i=1
Y
i
= e
(t)
i=1
Zi
where Z
i
is normal. Hence we have
x(t) =
e
(a
1
2
b
2
)t+bz(t)+
(t)
i=1
Zi
x(0) (3.17)
which, conditioned on (t), follows the lognormal distribution.
This model is nice because it can produce distributions that have heavier tails (i.e. more extreme price
movements) than the log-normal distribution. In this way, it can be used to create implied volatility smiles
and smirks.
3.4 A more general SDE
The following SDE contains a slightly more general description of an SDE driven only by Brownian motion.
The intuition behind our route to the solution comes from the use of integrating factors in basic ODEs.
dx = (a +bx)dt + (c +fx)dz. (3.18)
To solve this, rst write the SDE as follows:
dx bxdt fxdz = adt +cdz. (3.19)
In standard ordinary dierential equations, we would try to make the left hand side an exact dierential
by using an integrating factor. The integrating factor is usually related to the solution to the dierential
equation if the right hand side of (3.19) is set to zero. That is,
dx bxdt fxdz = 0. (3.20)
The solution to this is
x(t) = x(0)e
(b
1
2
f
2
)t+fz(t)
. (3.21)
Hence, we will try an integrating factor of the form
e
(b
1
2
f
2
)tfz(t)
. (3.22)
Therefore, let us compute
d(e
(b
1
2
f
2
)tfz(t)
x) = e
(b
1
2
f
2
)tfz(t)
(b
1
2
f
2
)xdt fxdz +
1
2
f
2
xdt
+(a +bx)dt + (c +fx)dz f(c +fx) dt
= e
(b
1
2
f
2
)tfz(t)
((a fc)dt +cdz) . (3.23)
Integrating both sides leads to
e
((b
1
2
f
2
)tfz(t))
x(t) x(0) =
t
0
e
((b
1
2
f
2
)sfz(s))
((a fc)ds +cdz(s)) (3.24)
and rearrangement gives the nal form
x(t) = e
((b
1
2
f
2
)t+fz(t))
x(0) +
t
0
e
((b
1
2
f
2
)(ts)+f(z(t)z(s))
((a fc)ds +cdz(s)). (3.25)
3.4.1 The Ornstein-Uhlenbeck Process and Mean Reversion
A very useful model is one in which a price or nancial variable is mean reverting. That is, the process is
pulled toward some value (in this case, its long term mean level). A standard model for this is a Gaussian
model called the Ornstein-Uhlenbeck process [14],
dx = b(x a)dt +cdz (3.26)
where a is the level that x(t) reverts to, and b is the mean reversion rate. Thus, x(t) is drawn toward a at a
rate of b. The Brownian driven term cdz just adds noise.
A sample path of a mean reverting process is depicted in Figure 3.2. Mean reverting processes of this
specic form are sometimes also called a Vasicek model. Vasicek used a process of this form to model the
short rate process in term structure modeling [15].
Since a mean reverting process is of the form of (3.18), it has a closed form solution. The solution is
given by
x(t) = e
bt
x(0) +
t
0
e
b(ts)
(abds +cdz(s)). (3.27)
Thus, observe that x(t) is a Gaussian process as well. One disadvantage of this process is that values of x(t)
can become negative because the Gaussian distribution always has some probability of being negative. This
is undesirable because prices and quantities such as interest rates should not be negative.
3.5. COX-INGERSOLL-ROSS PROCESS 25
0 1 2 3 4 5
1.6
1.8
2
2.2
2.4
2.6
2.8
3
t
x
Simulation of MeanReversion Dynamics
Figure 3.2: Simulation of Mean-Reverting Dynamics, a = 2, b = 5, c = 0.5, x(0) = 3.
3.5 Cox-Ingersoll-Ross Process
Another version of a mean reverting process is the Cox-Ingersoll-Ross (CIR) process [4]. It is given by
dx = b(x a)dt +c
xdz (3.28)
and used often in short rate models or stochastic volatility type models. Note that it is very similar to the
Ornstein-Uhlenbeck process, except that the driving Brownin motion is multiplied by

x. This makes the
noise dependent on the size of x. Furthermore, with correct parameter values, this process will always be
positive.
This process is related to the Ornstein-Uhlenbeck process as follows. Consider n Ornstein Uhlenbeck
Processes
dy
1
=
1
2
y
1
dt +
1
2
dz
1
(3.29)
.
.
. (3.30)
dy
n
=
1
2
y
n
dt +
1
2
dz
n
(3.31)
where the Brownian motions z
1
, . . . , z
n
are uncorrelated. Now, consider the process
x(t) = y
2
1
(t) +. . . +y
2
n
(t) (3.32)
By Itos lemma, x(t) follows
dx =
n
i=1
2y
i
(t)
1
2
y
i
dt +
1
2
dz
i
+
n
i=1
2
4
dt (3.33)
=
n
i=1
y
2
i
+

2
4
dt +
n
i=1
(y
i
dz
i
) (3.34)
=
x(t) +n
2
4
dt +
n
i=1
(y
i
dz
i
) (3.35)
Now, we perform a little trick by writing the last term as
x(t)
n
i=1
y
i
x(t)
dz
i
. (3.36)
Now, it turns out that
n
i=1
y
i
x(t)
dz
i
(3.37)
is actually a Brownian motion!
How can we see this? Well, note that if we interpret each Brownian increment as being normally dis-
tributed with mean zero and variance dt, i.e. dz
i
N(0, dt), then (3.37) is just the sum of Gaussians. But,
it is a well known property that the sum of Gaussians is Gaussian. Thus, (3.37) is normally distributed. For
it to be the increment of Brownian motion, heuristically, we only need to show that the mean is zero and
the variance is dt. To compute the mean we have
E
i=1
y
i
x(t)
dz
i
=
n
i=1
y
i
x(t)
E[dz
i
]
= 0 (3.38)
and the variance is computed as
V ar
i=1
y
i
x(t)
dz
i
=
n
i=1
y
2
i
x(t)
V ar(dz
i
)
=
n
i=1
y
2
i
x(t)
dt
=
dt
x(t)
n
i=1
y
2
i
= dt (3.39)
where we used that the dz
i
s are independent and equation (3.32). Thus we can actually use the replacement
dz =
n
i=1
y
i
x(t)
dz
i
(3.40)
to convert (3.35) to
dx =
x(t) +n
2
4
dt +
x(t)dz (3.41)
Now, it turns out that for n 2, we are guaranteed to have a positive process (otherwise the process can
and will reach zero at times!). Hence, we should always make sure that our parameter choices correspond
to selecting n 2.
This is easily done since we can match coecients in (3.28) and (3.41). This gives
ab =
n
2
4
, b = , c = (3.42)
From these relationships, we get that n =
4ab
c
2
.
For n corresponding to an integer, we have the solution as the sum of Ornstein-Uhlenbeck processes.
Thus, overall, the process x(t) will be Chi-Squared distributed. A sample path of a CIR process is depicted
in Figure 3.3.
3.6 Summary
In this chapter we have explored some of the standard stochastic dierential equation models used in nance.
It is important to have a feel for these processes what their parameters correspond to. Note that most of them
correspond to some transformation of simple Gaussian processes such as the Ornstein-Uhlenbeck process or
the Poisson process. You can think of these models as building more and more complicated models from our
basic building blocks of Brownian motion and the Poisson process.
3.7. PROBLEMS 27
0 1 2 3 4 5
1.5
2
2.5
3
t
x
Simulation of CIR Dynamics
Figure 3.3: Simulation of a sample path of a CIR process, a = 2, b = 5, c = 0.5, x(0) = 3.
3.7 Problems
Problem 3.7.1 Solve the following stochastic dierential equations:
(a) dx = dt +xdz
(b) dx = xdt + xdz + x(Y 1)d where is a Poisson process with intensity and Y is a random
variable.
(c) What does your answer in (b) look like if Y is log-normally distributed.
Problem 3.7.2 Consider the process
dx = b(x a)dt +cdz (3.43)
Using Itos lemma, derive a stochastic dierential equation for y = e
x
in terms of only y. This model is
log-normal due to the fact that x is normal, but has a mean reverting property in log coordinates.
Problem 3.7.3 Verify equation (3.23). (Hint: Let f(x, z, t) = e
(b
1
2
f
2
)tfz
x where z has the trivial sde
dz = dz and use the multidimensional Itos lemma.)
Chapter 4
The Factor Approach to Arbitrage
Pricing
4.1 The Factor Approach to Arbitrage Pricing
In this chapter we present absence of arbitrage conditions when returns are described by linear factor models.
In doing so, we rely heavily on Ross Arbitrage Pricing Theory [13]. However, we do so in light of the
application to derivative pricing. We will nd in the following chapters that when coupled with Itos lemma,
a simple absence of arbitrage equation results in perhaps the simplest way to obtain the vast majority of
Black-Scholes type pdes that appear in derivative pricing. To back up this claim we will derive pdes in a
great number of situations. But rst we must understand the underlying principle, which is in fact quite
simple. Lets begin...
4.2 Returns and Factors Models
Most of the modeling in asset pricing theory is done using linear factor models. Hence, we adopt that
paradigm here. That is, a return is assumed to be of the form
r = +f (4.1)
where a and b are constants and f is a random variable called a factor. This is a one factor model. We will
also consider multifactor models of the form
r = +
n
i=1
i
f
i
(4.2)
where each f
i
, i = 1, ..., n is a random factor.
4.2.1 Returns
Above I said we would model returns as factor models. However, returns are also related to price changes.
Hence, factor models should arise from price changes as well.
Given a period of time t, the return on an asset is dened as
r =
P(t + t) P(t)
P(t)
(4.3)
where P(t) is the amount invested at time t and P(t + t) is the amount received at time t + t. Figure
4.1 shows the cash ow diagram corresponding to this where the time increment it denoted as dt.
29
30 CHAPTER 4. THE FACTOR APPROACH TO ARBITRAGE PRICING
Figure 4.1: Cash Flow Diagram
4.2.2 Stochastic dierential equations and factor models
In nance, we like to model asset prices as stochastic dierential equations. For instance, a stock price could
be
dS = Sdt +Sdz. (4.4)
This is a model of a price change over time dt. Hence, I should be able to use it to compute the return of
an asset over time dt. In fact, it is given by:
r =
S(t +dt) S(t)
S(t)
=
dS(t)
S(t)
= dt +dz. (4.5)
This is also an example of a factor model. In the above formula, if we associate = dt, = , and f = dz
then it is a standard linear factor model.
In general, if I am given a stochastic dierential equation
dx(t) = a(x(t), t)dt +b(x(t), t)dz (4.6)
then it is describing price changes over time period dt, and I can use it to determine a model of returns:
r =
dx(t)
x(t)
=
a(x(t), t)
x(t)
dt +
b(x(t), t)
x(t)
dz (4.7)
This also looks like a factor model with the factor being dz. In fact, I like to write it in the form:
r = dt +dz (4.8)
where =
a(x(t),t)
x(t)
and =
b(x(t),t)
x(t)
(and arguments are being suppressed). This is slightly dierent from
the standard factor model in that I explicitly write out dt in the rst term. I like to go even a little further
and interpret dt as a special factor which is non-random. This is purely for convenience, but I will use this
convention throughout.
Another note is that and dont seem to be constant! However, we should view a factor model from
an SDE as being valid only over dt and conditioned upon information at time t, including x(t) (or t and
x(t) if appropriate). Therefore, conditioned upon information at time t, x(t) is known and and are
then known constants over time t to t +dt.
Therefore, SDEs lead to instantaneous factor models that apply over time increments of dt. In what
follows, I will use notation that is indicative of SDEs in that factors will be denoted by dz and I will include
a special factor dt when I describe returns. However, for the sections that follow, you may think of the factor
dz as being any random variable, not just Brownian motion, and dt can be any time increment, not just an
instantaneous change in time.
4.3. THE FACTOR APPROACH TO ARBITRAGE USING RETURNS 31
4.3 The Factor Approach to Arbitrage using Returns
We consider the returns of tradable assets. By a tradable asset, I mean an asset that you can actually buy.
For instance, you can purchase a share of stock. If this stock pays a dividend, then when you purchase the
share of stock, you have purchased the dividend stream as well. Note that you cannot just purchase the
price of the stock, only a share, and hence you are stuck with everything that comes along with purchasing
a share (such as dividends). So, lets consider the returns of tradable assets.
Lets list the returns of tradable assets that are driven by the factor dz
r = dt +dz (4.9)
where r R
n
, R
n
are vectors and R
nm
is a matrix. The factors are contained in the vector
dz R
m
. That is, I have stacked up all the returns in a vector.
Now consider a portfolio of these assets. We will represent the portfolio by a vector x R
n
which denotes
the dollar amount invested in each tradable asset. Hence, the total cost of our portfolio is
cost =
n
i=1
x
i
= x
T
1 (4.10)
where 1 is a vector of ones.
Over the time period dt, the returns in (4.9) indicate the change in value of each tradable asset. Hence,
the total change in our portfolio is given by:
prot/loss =
n
i=1
x
i
r
i
= x
T
r. (4.11)
We can analyze this return in a little more detail by writing out what each r
i
is
x
T
r = x
T
(dt +dz) = (x
T
)dt + (x
T
)dz. (4.12)
We can now see that this return will be riskless as long as
x
T
= 0 (4.13)
since that is the coecient of the dz term which is providing the randomness. If x
T
= 0 then the return
on this portfolio is deterministic and given by x
T
dt.
4.3.1 Arbitrage
Let us consider a simple portfolio. I will choose this portfolio so that it costs nothing and has no risk. That
is
x
T
1 = 0 No cost (4.14)
x
T
= 0 No risk (4.15)
But, if something costs nothing and has no risk, then it better have no return! If it had a positive return, this
would be an arbitrage, since I would be guaranteed to make money (no risk), and I didnt use any money to
do it (no cost). If it had a negative return, then the portfolio x would be an arbitrage. Therefore, to have
no arbitrage, I must have no return as well. This leads us to a key condition which I state as the following
implication
() A (not very useful) Necessary Absence of Arbitrage Condition
x
T
1 = 0 No cost
x
T
= 0 No risk
x
T
= 0 No return (4.16)
which can be written in matrix form as
1
T
x = 0
T
x = 0. (4.17)
As the name indicates, the above condition for absence of arbitrage is not terribly useful. However, an equiv-
alent condition is extremely useful and will provide the foundation for our derivations of partial dierential
equations. In fact, the condition which is stated next is astonishingly useful! But to derive it, we need to
rst recall some linear algebra relationships.
4.3.2 Null and Range Space Relationship
Note that the condition (4.17) can be written as
Ax = 0
T
x = 0 where A =
1
T
. (4.18)
Now, the set of all vectors x such that Ax = 0 is known as the null space of the matrix N(A). That is
N(A) = {x|Ax = 0} . (4.19)
On the other hand, the set of all vectors y such that there exists an x with y = Ax is known as the range
space of the matrix A. That is
R(A) = {y | x such that Ax = y } . (4.20)
Finally, we recall the notion of the perpendicular set of a given set of vectors. Let M be a set of vectors,
then M
is a set of all vectors z such that z is orthogonal to all vectors in M. That is

M
z | z
T
x = 0, x M
(4.21)
In order to derive a useful condition from (4.17), we will use the following relationship between the null
and range space of a matrix.
() Null and Range Space Relationship: N(A)
= R(A
T
).
Proof: The proof of this is rather simple.
x N(A) Ax = 0 y
T
Ax = 0 for all y. (4.22)
Then
y
T
Ax = (A
T
y)
T
x = 0 (4.23)
which means that A
T
y is orthogonal to the null space of A for all y. That is exactly saying that
N(A)
= R(A
T
). (4.24)
2.
4.3.3 A Useful Absence of Arbitrage Condition
Using the null and range space relationship, we can convert (4.17) into a very useful condition. Its importance
cannot be overstated. Hence it receives two stars!!
( Return APT) A Useful Necessary Absence of Arbitrage Condition
4.3. THE FACTOR APPROACH TO ARBITRAGE USING RETURNS 33
A necessary and sucient condition for the implication (4.16) to be true is for there to exist a vector
R
m+1
such that
=
0
+ = (4.25)
where
(4.26)
with
0
R and R
m
.
Many of you will hopefully recognize this as nothing more than the simple version of Ross 1976 arbitrage
pricing theory [13]. Its derivation is given below.
Proof: Note that the condition in (4.17) indicates that if a portfolio x is in the null space of the matrix
1
T
then it also must be orthogonal to . Another way to put this is that

1
T
(4.27)
where N() is the null space. But, using the null and range space relationship, this means that
R
1
T
(4.28)
Thus, there exists a vector

R
m+1
such that [ 1 ]
= . 2
4.3.4 Interpretations
Lets try to get a bit of intuition into the mystical s.
Market Price of Risk
The APT equation for a single factor model r = dt +dz is
= [ 1 ]
=
0
+
1
. (4.29)
Lets look at this equation. On the left side is the expected return (), and on the right side is the volatility
(). Hence, this equation relates volatility to expected return.
A better way to think of is not as volatility, but as the amount of the risk factor dz that you have. If
is zero, then you are not exposed to the risk dz at all. If it is large, then you have purchased a lot of that
risk. Therefore,
1
tells you how much your expected return is increased (assuming
1
is positive) for each
unit of the risk dz that you take on. For this reason,
1
is called the market price of risk. In this case,
1
is the market price of the risk factor dz. This is a nice interpretation for
1
.
The Market Price of Time
But, we also have this pesky
0
to deal with. It is not tied to any risk. In fact, if we dont take on any risk
(i.e. = 0), then =
0
. Intuitively, if we dont have any risk, then we should be earning the risk free rate.
Hence, we might guess that
0
= r
0
, where r
0
denotes the risk free rate of interest. This is in fact correct.
But, lets justify this through a slightly dierent argument that will also get us a nice interpretation for
0
.
Let us consider a risk free asset. Its factor model is given by:
dB
0
B
0
= r
0
dt (4.30)
where r
0
is a constant risk free rate of interest. Since it must satisfy our return relationship, we have:
r
0
=
0
+ 0(
1
) =
0
(4.31)
which tells us that
0
= r
0
. If the s are the market prices of factors, then we may interpret
0
as the
reward for the time factor dt or the market price of time. Intuitively, this makes perfect sense, since the
risk free rate is the amount we are rewarded for taking on time and nothing else.
Now all the s should make sense. They relate the dierent factors dt, dz, etc. to how much we are
rewarded for taking on those factors. In a market with no arbitrage, every factor has a market price. Your
return is just given by looking at how much of each factor you have taken on, and multiplying them by their
market price and adding them up. Quite simple, right?
This is the basic interpretation of the concept of market price of risk. Dont forget it. It is very useful
to have this intuition.
4.3.5 A Problem with Returns
Using returns to model assets has a disadvantage. There are contracts that involve no up-front cost or price.
Hence, they dont have a well dened return, since by denition a return involves dividing by the price,
which can be zero. A futures contract is an example of this, since the mark-to-market mechanism resets the
value of a contract to zero every day. Therefore, in the above framework, a futures contract must be dealt
with as a special case. I dont like special cases, so below we will reformulate absence of arbitrage conditions
but in terms of price changes rather than returns. This eliminates the need for special cases.
4.4 The Factor Approach using Price Changes
I just mentioned that there can be some problems in dealing with returns. Hence, in this section we will
reformulate the factor approach to arbitrage pricing by working with prices rather than returns. In this case,
it will be okay to have an asset with a zero price.
4.4.1 Price Changes and Arbitrage
Lets return to our arbitrage portfolio and reformulate it in terms of price changes. In our original argument,
we let r R
n
be the returns of assets and specied the dollar amount invested in each asset by a vector
x R
n
. This time we will specify a vector of prices per unit of tradables P R
n
, cash ows resulting from
the changes in value of the tradables per unit dV R
n
, and shares or units of each asset purchased y R
n
.
The cash ow from changes in value over the period will satisfy the factor model
dV = Adt +Bdz (4.32)
where A R
n
, B R
nm
, and dz R
m
. The simple cash ow diagram is given in gure 4.2.
4.4.2 Prot/Loss and Arbitrage
In this case, the prot/loss on the portfolio is given by:
y
T
(dV) = y
T
(Adt +Bdz) = (y
T
A)dt + (y
T
B)dz (4.33)
Therefore, an arbitrage portfolio is one that has:
4.4. THE FACTOR APPROACH USING PRICE CHANGES 35
Figure 4.2: Cash Flow Diagram with Prices P and Value Changes dV
No Cost: y
T
P = 0,
No Risk: y
T
B = 0,
but has a prot
Prot: y
T
A = 0.
Of course, we want to eliminate arbitrages, so we would like the following implication to hold.
() A (not very useful) Necessary Absence of Arbitrage Condition
y
T
P = 0 No cost
y
T
B = 0 No risk
y
T
A = 0 No return (4.34)
which can be written in matrix form as
P
T
B
T
y = 0 A
T
y = 0. (4.35)
Once again, we can convert this to a dual condition that is useful:
(Price APT) A Useful Necessary Absence of Arbitrage Condition
A necessary condition for no arbitrage is for there to exist a vector

R
m+1
such that
P B
= P
0
+B = A (4.36)
where
(4.37)
with
0
R and R
m
.
What is the relationship between the return approach and the price change approach. Well, of course
they are highly related. The s in both cases are the same, and have the same interpretation. They are
market prices of risk. The main dierence is that the price approach uses shares or units of the asset to
describe the portfolio, not dollar amount. Furthermore, it describes prot/loss in terms of the cash ow, not
returns. These are really supercial dierences, but it is easier to understand some pricing situations, such
as futures contracts, in the price approach rather than in terms of returns.
4.5 Two standard examples
In this section we will give some little examples to help us gain some intuition into the approach. Lets start
with a stock and a bond.
4.5.1 Stocks
Assume that a stock follows a geometric Brownian motion (GBM) and there is a bond that earns the risk
free rate, r
0
. Their dynamics are given below:
dB = r
0
Bdt (4.38)
dS = Sdt +Sdz. (4.39)
What do our absence of arbitrage conditions say about these assets?
Returns
In terms of returns, we can write that:
dB
B
= r
0
dt (4.40)
dS
S
= dt +dz (4.41)
Hence, the APT equation says [ 1 ]
= or
1 0
1

0
r
0
(4.42)
Solving this equation for
0
and
1
gives
0
= r
0
(4.43)
1
=
r
(4.44)
Note that the market price of risk for dz is like an instantaneous Sharpe ratio. We will see this equation pop
up many times.
Prices
We can derive the same results using prices. In terms of prices and changes in value, the absence of arbitrage
equation says that [ P B ]
= A or
B 0
S S

0
r
0
B
S
(4.45)
0
and
1
gives
0
= r
0
(4.46)
1
=
r
, (4.47)
as expected.
4.6. SUMMARY 37
4.5.2 Futures contracts
One can enter into a futures contract without paying any money. This means that a futures contract is a
special case in our setup, and is dicult to understand in the context of returns because its return is not
dened (it has zero price!). It is much more naturally considered in terms of prices and value changes. Below
is the cash ow diagram for a futures contract.
The critical dierence between futures contracts and forward contracts is that futures contracts are
marked to market and settled daily. This means that they always begin the day with a zero price. At the
end of the day, the price change is settled. Note that this price change is not discounted, but rather the
change in the futures price. Hence, the change in value of the portfolio is equal to the change in the futures
price dV = df!
Figure 4.3: Cash Flow Diagram for Futures Contract
Therefore, a futures contract can have zero price, but the cash ow from the change in value will be given
by
df = fdt +fdz. (4.48)
Hence, in this case, we may consider a market with a bond and a futures contract. The relevant quantities
can be written as
prices changes factors
B dB = r
0
Bdt
0 df = fdt +fdz.
(4.49)
Now lets look at the price based arbitrage equation [ P B ] = A or
B 0
0 S

0
r
0
B
S
. (4.50)
0
and
1
gives
0
= r
0
(4.51)
1
=

. (4.52)
Note that a futures contract refers to the price at a xed time in the future. Hence, time is xed, and a
futures contract is purely a bet on the outcome of a random factor. Time is not in the mix. Hence, we see
that the equations separate. The bond is purely time, and the market price of time is r
0
. On the other hand,
the futures contract is purely a bet on the outcome of the factor. Hence time is not mixed into it, and the
market price of risk does not involve the risk free rate r
0
.
4.6 Summary
We have derived forms of an aribtrage pricing theory based either on linear factor models of returns or
value changes. These simple results will allow us to derive many of the partial dierential (or dier-
ence/integral/etc.) equations that arise in derivative pricing theory. But before doing so, we develop a
derivative pricing framework based upon our APT factor based approach. Then, in following chapters we
tackle examples in equity derivatives, interest rate derivatives, and beyond. What you should take away is
that simple arbitrage ideas that we derived in this chapter are the underpinnings of derivative pricing theory.
4.7 Problems
Problem 4.7.1 (How to construct an arbitrage)
Assume that there does not exist a such that = [1 ]. Then this means that an arbitrage opportunity
exists. That is, there exists a vector x specifying dollar amounts invested such that
T
x > 0 and 1
T
x = 0,
T
x = 0. (4.53)
One way to select a specic arbitrage would be to solve an optimization problem of the form
max
x
T
x subject to 1
T
x = 0,
T
x = 0, x
1
= 1. (4.54)
where the constraint x
1
limits the total dollar amount of long and short positions to equal $1. However,
in practice an alternative but equivalent approach is often used, as follows.
Choose such that [1 ] has minimum 2-norm. That is
min
[1 ]
2
. (4.55)
If
is this optimal , then we can dene

= [1 ]
. (4.56)
One could then solve the optimization problem
max
x
T
x subject to 1
T
x = 0,
T
x = 0, x
1
= 1. (4.57)
Show that this optimization problem is equivalent to the optimization problem in (4.54).
Problem 4.7.2 (More Linear Algebra and Dualities)
The Farkas Alternative [7] of linear algebra says the following
(i) Ax = b has a solution x 0
or (exclusive)
(ii) y
T
A 0, y
T
b < 0 has a solution y.
(4.58)
We can use the Farkas Alternative for implications as well. Consider the implication:
Ax = 0 b
T
x = 0 (4.59)
Prove by use of the Farkas alternative that this implication is true if and only if there exists a such that
b = A
T
. (4.60)
(Hint: Note that the implication is true if and only if there is no solution to Ax = 0 and b
T
x > 0.)
Problem 4.7.3 (Early Exercise and American Options)
American options allow the holder of the option to exercise at any time prior to expiration. Lets explore
how this would aect the absense of arbitrage ideas.
4.7. PROBLEMS 39
(a) Assume that the factor model for the option if it is not exercised is
r
1
=
c
dt +
c
dz. (4.61)
Explain why the factor model for the return on an American option, where E is the early exercise value and
c is the price of the option, is either
r
1
=
c
dt +
c
dz or r
1
=
E c
cdt
dt (4.62)
depending on whether the option is exercised or not.
(b) Assume that you are long the American option. As the holder of an American option, you get to decide
when to exercise the option. Let and correspond to factor models for tradable assets, where the rst
elements of these vectors
1
and
1
corresponds to the American option. Finally, let x be a vector of dollar
amounts invested in each asset (with x
1
corresponding to the American option).
Argue that a necessary condition for you not to be able to arbitrage is that the following implication is
true regardless of whether you are exercising the American option or not,
e
1
T
x = 1, 1
T
x = 0,
T
x = 0
T
x 0. (4.63)
(I.e., the implication must hold regardless of which return in (4.62) is being chosen.)
(c) Assume that the vectors e
1
, 1, are linearly independent. Using the Farkas Alternative from Problem
4.7.2, derive that an alternative absence of arbitrage condition is that there exists scalars 0, and
0
and
1
such that
=
0
+
1
e
1
. (4.64)
(d) Finally, using (4.64) and Part (a) of this problem, argue that for the American option, both of the
following conditions must hold
c

0
+
c
1
(4.65)
E c. (4.66)
Problem 4.7.4 (Modeling the binomial lattice as a factor model)
Consider a stock that can move on a binomial lattice, where at each step the stock price S(k) can either
move up to S(k + 1) = uS(k) or down to S(k + 1) = dS(k) with the probability structure
S(k + 1) =
uS(k) w.p. p
dS(k) w.p. 1 p
(4.67)
Let Z(k) be a standard binary random variable
Z(k) =
1 w.p. p
1 w.p. 1 p
(4.68)
Show that S(k) = S(k + 1) S(k) can be written in the form
S(k) = A(k) +B(k)Z(k) (4.69)
by nding A(k) and B(k).
Chapter 5
Constructing a Factor Pricing
Framework
5.1 Introduction
This chapter lays the framework for derivative pricing. That is, we try to provide the structure behind the
factor approach. Thus, the goal is to clarify the structure of the modeling involved in the factor approach,
and to provide an almost step by step method for attacking any problem.
As I proceed through this chapter, I will use two derivative pricing examples to illustrate points. The
examples are a call option on a non-dividend paying stock, and an absence of arbitrage zero coupon bond
pricing model. In particular, I will also emphasize the relative pricing nature of derivative pricing. That is,
derivative pricing determines an appropriate price relative to other securities that have already been priced
in a market. Before jumping into any of these ideas of derivative pricing, lets start by classifying the relevant
quantities in a model of any market.
5.2 A Classication of Quantities
In this book, we will be interested in the pricing of derivative securities. However, to understand that pricing
theory, it is best to rst understand the general modeling paradigm.
In our modeling framework, we will classify all quantities into three (possibly overlapping) categories.
They are factors, underlying variables, and tradables.
5.2.1 Factors
Factors are the most basic source of randomness in our models. In general, they are the driving Brownian
motions (z(t)) and Poisson processes ((t)). Furthermore, I like to think of time as a special factor. Thus,
in our models, the factors will show up as the dz, d, and dt terms.
5.2.2 Underlying Variables
Underlying variables are often quantities of interest that are functions of the factors. For example, an interest
rate could be an underlying variable. A stock price S(t) could also be an underlying variable. In general,
underlying variables are nancially relevant quantities that are functions of the factors, and used in the
modeling process.
41
42 CHAPTER 5. CONSTRUCTING A FACTOR PRICING FRAMEWORK
5.2.3 Tradables
Tradables are the quantities that you can actually trade and include in a portfolio. They are modeled
as being functions of the underlying variables. Some of the time, the functional relationship between the
tradable and an underlying variable is trivial. For example, a stock price S(t) can be an underlying variable,
and also a tradable. In other cases, the tradable is a more complicated function of an underlying variable.
For example, an interest rate r(t) can be an underlying variable, but it is not tradable. Instead, a bond
B(r(t), t) is tradable, and represented as a function of the underlying variable.
Figure 5.1: Picture of the Modeling Paradigm
It is extremely important to be able to separate quantities that are tradable from those that are not.
Why? Because tradables, and only tradables, are the quantities that must satisfy the absence of arbitrage
relationships from the previous chapter. And those absence of arbitrage relationships will lead to our pricing
formulas.
Lets consider an example to clarify the notion of what is tradable.
Example: A stock paying a dividend
To make sure we understand what a tradable is, lets consider the example of a stock that pays a dividend.
The stock has price S(t), and lets assume that it pays a continuous dividend at a rate of q. What this means
is that by reinvesting the dividend back into the stock, if you started by purchasing 1 share of stock at time
0, by time t you would have e
qt
shares.
What is the tradable quantity? You might be tempted to say that the price of the stock S(t) is tradable.
However, note that you cannot just purchase the price of the stock. Instead, you must purchase a share
of stock, and with this share you not only get the price of the stock but also the dividend. The point is that
you cant decouple the price from the dividend. They come together and that is the tradable.
5.2.4 A Derivative is a Tradable
In this book we are interested in pricing derivative securities. So, it is important to begin by dening what
we mean by a derivative security.
Derivative Security: A derivative security is a tradable whose value depends upon (or is a function of)
other underlying variables. In this case, we say that the derivative security is derivative to the underlying
variables.
Note that this denition of a derivative does not particularly distinguish it from any other tradable! We
typically model all tradables as being a function of some underlying variables (although often a tradable is
an underlying variable and hence trivially a function of itself).
In principle, there is no real distinction between a derivative and any other tradable. You will see that
the most important distinction is that a derivative is what you want to price and that other tradables are
already priced in the market.
The name derivative comes from the fact that the value of a derivative is derived from some underlying
variables. In order to have something to keep in mind, lets use the following example:
5.3. FACTOR MODELS FOR UNDERLYING VARIABLES AND TRADABLES 43
Example: European Call Option
Consider a European call option, which is the option to purchase a specied stock, say Coca-Cola, for a
specied price (the strike price), at a specied date (the expiration date). In this case, the option is a
derivative, because its value depends upon the stock price of Coca-Cola. Thus, the stock price of Coca-Cola
is the underlying variable.
In this example, Coca-Cola is an underlying variable and also a tradable. However, it is not always the
case that the underlying variable must also be tradable. In general, derivative securities are quite exible, and
so are the possible underlying variables. In fact, there are derivatives that depend on underlying variables
such as temperature or even wind speed which clearly are not tradable.
5.3 Factor Models for Underlying Variables and Tradables
To use the APT equations from the previous chapter, what we need are linear factor models for the price
changes of tradables. In the previous chapter we saw that we can interpret SDEs as instantaneous factor
models. Hence, in continuous time models, we would like SDE models for underlying variables and tradables
in particular.
In general, there are two ways that we arrive at factor models for tradables. The rst is that we directly
model the tradable as a factor model. The second (and most important) is Itos lemma.
5.3.1 Direct Factor Models
In many cases, we begin the modeling process by writing down a factor model or SDE. Examples of this
include the geometric Brownian motion model of stock price movement.
dS = Sdt +Sdz (5.1)
or a model of an interest rate (most likely the instantaneous short rate) as an SDE
dr
0
= adt +bdz. (5.2)
Thus, in many case, we begin the modeling process by writing down SDEs (or factor models) for underlying
variables and tradables.
5.3.2 Factor Models via Itos Lemma
The second method to obtain factor models is via Itos lemma. Since tradables, and derivatives in particular,
are functions of underlying variables, if we have an SDE model for an underlying then we can use Itos lemma
to obtain an SDE for the tradable.
Consider the example of a European call option on a stock following geometric Brownian motion as in
(5.1). The call option is a function of the underlying stock price S(t) and time t. We write this as c(S(t), t).
Then, by Itos lemma we have
dc = (c
t
+Sc
S
+
1
2
2
S
2
c
SS
)dt +Sc
S
dz (5.3)
which is a factor model for the call option c.
In some cases, we will know explicitly the functional relationship between the tradable or derivative and
the underlying variable. In that case, in Itos lemma we would be able to explicitly compute the partial
derivative terms c
t
, c
S
, etc.
A picture of the route to SDEs for tradables is given in Figure 5.2
Figure 5.2: Obtaining factor models or SDEs for tradables.
5.4 Tradables tables
We will approach the derivative pricing problem using the Price APT equation (4.36), rather than the Return
APT. But note that everything in this book can also be done use the Return APT equation.
To use the Price APT equation, we need two things: prices and factor models for value changes. The
previous section indicated how to obtain the factor model for value changes. Thus, we can tabulate prices
and factor models for value changes for all tradables. When we list these in a table, we call this a tradable
table. For instance, in a market with a stock, a bond, and a call option on the stock, the tradable table is
Prices
B
S
c
Changes
d
B
S
c
=
Factor Model
r
0
B
S
(c
t
+Sc
S
+
1
2
2
S
2
c
SS
)
dt +
0
S
Sc
S
dz
(5.4)
This is called a tradable table to emphasize the fact that only tradables need satisfy the absence of arbitrage
conditions. It supplies all the basic information that we need to extract from the market when applying the
Price APT.
5.5 Applying the Price APT
The nal step to a derivative pricing equation is to apply the Price APT equations. However, before applying
the Price APT equation, we rst separate marketed tradables from the derivative the we would like to price.
5.5.1 Relative Pricing and Marketed Tradables
Once we have set up the tradables table, we designate certain tradables as marketed. What we mean by a
marketed tradables is a tradable that is already priced by the market.
You should contrast the marketed tradables with the derivative that we would like to price. The derivative
is what we want to price using the information given from the marketed assets. This emphasizes the relative
pricing nature of derivative pricing. We price a derivative security relative to the marketed tradables. That
is, derivative pricing is a method of pricing a new tradable (the derivative) consistently with the existing
price of marketed tradables.
To see how this works, we separate the marketed tradables from the derivative, and put them all in a
tradables table.
Prices
P
m
P
Changes
d
V
m
V
=
Factor Model
A
m
A
dt +
B
m
B
dz
(5.5)
5.5. APPLYING THE PRICE APT 45
The tradables with the subscript m are the marketed tradables, and the tradable without any subscript is
the derivative that we are pricing.
5.5.2 Pricing the Derivative
Now we can use the Price APT to price the derivative. The Price APT equation says there needs to exist
0
and such that
A
m
= P
m
0
+B
m
(5.6)
A = P
0
+B. (5.7)
To obtain the market prices of risk
0
and , we only use the marketed tradables. Thus, we solve (5.6) to nd
the market prices of risk (often in terms of information from the marketed tradables). The market prices of
risk are then plugged into (5.7) which is the pricing equation for the derivative.
One may think of this procedure as rst calibration to market data, followed by pricing the new derivative.
Using marketed tradables to determine the market prices of risk is like calibrating parameters (market prices
of risk) in our absence of arbitrage pricing model to known data (marketed tradables). Once our model is
calibrated (the market prices of risk are determined), we can apply our absence of arbitrage model to price
other tradables (derivatives) that also must satisfy the same absence of arbitrage relationship.
This discussion has been a little abstract, so lets see an example
Example: Pricing a European Call Option
Lets take the tradables table in (5.4) and use it to illustrate pricing. Assume that c is a European call
option on the stock S.
First, we designate the bond B and stock S as being the marketed assets, and c is the derivative that we
would like to price. Thus, the price APT implies that
B
S
c
0
+
0
S
Sc
S
1
=
r
0
B
S
c
t
+Sc
S
+
1
2
2
S
2
c
SS
. (5.8)
where the top two rows correspond to the marketed tradables. Then, we use the marketed tradables to
determine the market prices of risk. Solving these equations gives
0
= r
0
and
1
=
r0
.
Finally, we use these market prices of risk in the third equation for our derivative to obtain
c
t
+Sc
S
+
1
2
2
S
2
c
SS
= r
0
c +
r
0
Sc
S
(5.9)
Rearranging this gives
c
t
+r
0
Sc
S
+
1
2
2
S
2
c
SS
= r
0
c (5.10)
which is the Black-Scholes partial dierential equation for the price of an option. If we specify the boundary
condition for a European call option as
c(S, T) = (S K)
+
, c(0, t) = 0 (5.11)
then this completely describes the price of the option.
A picture of the application of the Price APT is given in Figure 5.3.
5.5.3 Underdetermined and Overdetermined Systems
In the example above everything was perfect! I had two equations (one for the bond and one for the stock),
and two unknowns (the market price of time
0
and the market price of risk
1
) corresponding to the
marketed tradables. Thus, I could solve the equations for
0
and
1
. It was perfect!
Now, what if things arent so perfect. That is, what if the system of equations arising from the Price
APT for the marketed tradables is either underdetermined or overdetermined. Let use an example.
Figure 5.3: A picture of the application of the Price APT.
Underdetermined and Incompleteness
Assume that our assets are given by
dB = r
0
Bdt (5.12)
dS = Sdt +
1
Sdz
1
+
2
Sdz
2
. (5.13)
In this case, the price APT becomes
B 0 0
S
1
S
2
S
r
0
B
S
(5.14)
And we have three unknowns (
0
,
1
, and
2
), but only two equations! Thus, we cant uniquely solve for
the market prices of risk! (In this case, we can solve for
0
= r
0
, but not uniquely for
1
or
2
.)
What can we say in this situation and how should we think of this? Well, rst, we can say that if any
solution exists (it doesnt have to be unique, we just need for at least one solution to exist), then there is no
arbitrage. This is guaranteed by the APT equations.
Lets assume that many solutions exists. For example in the above equations, there are multiple possible
values for
1
and
2
. Thus, there are many possible market prices of risk that satisfy the no arbitrage
condition. This just means that from the tradable assets in the market (B and S), we cannot uniquely infer
the market prices of risk for dz
1
and dz
2
. There are many possibilities, and all are arbitrage free.
The practical consequence of this is that if we are asked to price a new security that depends on dz
1
and/or dz
2
, we will not be able to assign it a unique absence of arbitrage price. This is because the APT
equation (in either return or price form) acts as a pricing equation. (This use will become clear in the
following chapters.) However, it only provides a unique price if we have unique values for the market prices
of risk. This situation is called an incomplete market.
Incomplete markets are common in practice, and you will see in subsequent chapters that in order to
price derivative securities in incomplete markets, we must select values for market prices of risk that are not
uniquely dened. Since market prices of risk relate risk to reward for various factors, selecting a value for
5.6. THREE STEP PROCEDURE 47
a market price of risk is essentially the same as specifying how investors in the market trade o risk and
return. Thus, in incomplete markets some specication of the risk preferences of investors is needed to assign
a unique price to derivative securities. Furthermore, this specication of risk preferences is captured by the
selection of the market price of risk.
The above discussion might be a little abstract at this point, but in subsequent chapters you might
want to refer back to it when faced with pricing of derivatives in incomplete markets (see for example,
jump-diusion models or stochastic volaility).
Overdetermined and Calibration
Now lets consider the opposite situation. That is, when the set of equations is overdetermined. For example,
let the tradable assets be
dB = r
0
Bdt (5.15)
dS
1
=
1
S
1
dt +
1
S
1
dz (5.16)
dS
2
=
2
S
2
dt +
2
S
2
dz. (5.17)
In this case, the price APT becomes
B 0
S
1

1
S
S
2

2
S
r
0
B
1
S
1
2
S
2
(5.18)
In this case there are three equations and only two unknowns (
1
and
2
)! This system looks to be overde-
termined!
Now, we know from the price APT that for no arbitrage to exist there must exist a solution to this set
of equations. However, in general, for an overdetermined system of equations no solution will exist! What
does this mean?
Well, the rst thing it means is that strictly speaking, there is an arbitrage opportunity! But, the way
this situation often plays out in practice is usually slightly dierent. In practice this situation often leads to
some sort of calibration procedure.
Instead of declaring that an arbitrage exists, a trader will often just assume that the models being used
for B, S
1
and S
2
are not perfect, and that is the reason that no solution exists. Thus, the trader will search
for the
1
and
2
that best t the absence of arbitrage equations. They often call this process calibration,
and in practice it may be dicult to recognize directly as searching for best t s. (Watch for it in situations
such as term structure modeling where a single factor dz is used, but many tradables (bonds of dierent
maturities) exist, or when certain models are t to volatility smiles and smirks.) Again, you might want to
return to this discussion after reading subsequent chapters.
5.6 Three Step Procedure
To summarize, lets present a three step procedure to deriving a derivative pricing equation. The three steps
are:
1. Identify the tradable assets, underlying variables, and factors in a model. (See Figure 5.1.)
2. Write factor models (SDEs) for each tradable asset (this may involve Itos lemma), and construct a
tradables table. (See Figure 5.2.)
3. Apply the Price APT equation, rst to the marketed tradables in order to solve for the market prices
of risk (calibration), and then to the derivative using those market prices of risk (this usually results
in a partial dierential equation for the price). (See Figure 5.3.)
In this book, I will consider the limited scope of just deriving partial dierential equations for pricing, and I
will sometimes leave out the important step of actually solving the pde!! Furthermore, often many dierent
derivative securities satisfy the same basic pde and only dier due to their boundary condition. I will also
often sweep that under the rug, and merely note that a particular derivative security will correspond to the
solution of the pde with a particular boundary condition.
5.7 Summary
In this chapter, we have provided a three step procedure for deriving the pricing equation for a derivative
security based on the linear factor approach. This cookie cutter approach will allow us to derive many of
the partial dierential (or dierence/integral/etc.) equations that arise in derivative pricing theory. In the
next chapter we tackle examples in equity derivatives, and in the following chapter we address interest rate
derivatives. What you should take away is that simple arbitrage ideas that we derived in this chapter are
the underpinnings of derivative pricing theory.
5.8 Problems
Problem 5.8.1 (A Stock Paying Continuous Dividends)
Consider a stock S(t) following
dS = Sdt +Sdz (5.19)
that pays a continuous dividend at a rate of q. Also assume that a bond
dB = r
0
Bdt (5.20)
exists. If c(S, t) is a European call option on the stock, then identify the factors, underlying variables, and
create a tradables table.
Problem 5.8.2 (A Single Factor Short Rate Model)
Let r
0
(t) denote the short rate of interest, and assume that it follows
dr
0
= adt +bdz (5.21)
Assume that zero coupon bonds of maturity T are tradables and the price at time t is denoted by B(r, t|T).
Let these bonds be a function of the short rate and time t. Furthermore, assume that a money market account
exists that satises
dB
0
= r
0
(t)dB
0
dt (5.22)
In this model, then identify the factors, underlying variables, and create a tradables table.
Chapter 6
Application of the Factor Form:
Equity Derivatives
The factor approach to absence of arbitrage pricing is one of the quickest and most direct routes to deriving
pdes for derivatives. In this chapter we will see that it can be used to derive almost every Black-Scholes type
pde that occurs in derivative pricing.
6.1 Examples from Equity Derivatives
6.1.1 Black-Scholes
Black and Scholes started everything with this model [2]. The standard Black-Scholes set-up involves a bond
earning a risk free rate and a non-dividend paying stock that follows a GBM:
dB = r
0
Bdt (6.1)
dS = Sdt +Sdz. (6.2)
Furthermore, we consider a derivative on the stock. Generically we call this derivative c and assume that
its price process depends on the stock and time c(S, t) and is twice continuously dierentiable in both its
arguments. By Itos lemma,
dc = (c
t
+Sc
S
+
1
2
2
S
2
c
SS
)dt +Sc
S
dz. (6.3)
Therefore, the tradable assets are the stock S, bond B, and the derivative c. We can now move to Step 2
and construct a tradable table
B
S
c
B
S
c
r
0
B
S
c
t
+Sc
S
+
1
2
2
S
2
c
SS
dt +
0
S
Sc
S
dz. (6.4)
The tradable table contains all the information we need to apply the Price APT equation. Hence, this allows
us to move on to Step 3, which is to solve the Price APT equation P
0
+B
1
= A for absense of arbitrage
conditions.

B
S
c
0
+
0
S
Sc
S
1
=
r
0
B
S
c
t
+Sc
S
+
1
2
2
S
2
c
SS
. (6.5)
Solving for
0
and
1
using the rst two equations gives:
0
= r
0

1
=
r
0
. (6.6)
49
50 CHAPTER 6. APPLICATION OF THE FACTOR FORM: EQUITY DERIVATIVES
Plugging
0
and
1
into the last equation yields
r
0
c +
r
0
Sc
S
= c
t
+Sc
S
+
1
2
2
S
2
c
SS
. (6.7)
Finally, rearranging leads to the Black-Scholes equation:
c
t
+r
0
Sc
S
+
1
2
2
S
2
c
SS
= r
0
c. (6.8)
European Call and Put Options
For a European call option, the boundary condition is that at expiration T, the option is worth c(S, T) =
max(S K, 0) where K is the strike price. We also have the boundary conditions that c(0, t) = 0 for all t
and lim
S
c(S, t) = .
c(S, T) = max(S K, 0) (6.9)
c(0, t) = 0 t [0, T] (6.10)
c(, t) = t [0, T] (6.11)
The solution is
d
1
=
ln(S/K) + (r
0
+
1
2
2
)(T t)
T t
(6.12)
d
2
= d
1
T t (6.13)
c(S, t) = SN(d
1
) Ke
r0(Tt)
N(d
2
) (6.14)
A European put option is a derivative security p(S, t) with boundary conditions
p(S, T) = max(K S, 0) (6.15)
p(0, t) = 0 Ke
r0(Tt)
t [0, T] (6.16)
p(, t) = 0 t [0, T] (6.17)
The solution to the Black-Scholes equation under these boundary conditions is
d
1
=
ln(S/K) + (r
0
+
1
2
2
)(T t)
T t
(6.18)
d
2
= d
1
T t (6.19)
p(S, t) = Ke
r0(Tt)
N(d
2
) SN(d
1
) (6.20)
6.1.2 Dividend Paying Stocks
Lets assume that the stock is paying a continuous dividend at a rate of q. Then the stock and its dividend
stream is a tradable asset. That is, when we purchase the stock, we are also purchasing its dividend stream.
Hence we must consider them together as a tradable asset. Lets assume the stock price follows
dS = Sdt +Sdz. (6.21)
What we purchase is a single share of this stock. Let N(t) denote the number of shares held of this stock at
time t. The assumption of a continuous dividend at a rate of q is equivalent to assuming that the number of
shares N(t) grows according to the equation
dN = qNdt. (6.22)
6.1. EXAMPLES FROM EQUITY DERIVATIVES 51
Then over time dt, the portfolio with value N(t)S(t) changes to
N(t)S(t) (N(t) +dN)(S(t) +dS) (6.23)
= N(t)S(t) +dNS(t) +N(t)dS +dSdN (6.24)
= N(t)S(t) +qN(t)S(t)dt +N(t)S(t)dt +N(t)S(t)dz +o(dt). (6.25)
If we denote the value of the share with its dividend stream by v(t) = N(t)S(t) then we have
dv = v(t +dt) v(t) = ( +q)v(t)dt +v(t)dz. (6.26)
Hence, we consider v(t) the tradable.
Finally, we consider the option. Now, the payo of the option depends on the price of the stock alone,
and does not depend on the dividend stream. Therefore, we assume c = c(S, t) and apply Itos lemma to c
to obtain
dc = (c
t
+Sc
S
+
1
2
2
S
2
c
SS
)dt +Sc
S
dz. (6.27)
The next step is to write the tradable table
B
v
c
B
v
c
r
0
B
( +q)v
c
t
+Sc
S
+
1
2
2
S
2
c
SS
dt +
0
v
Sc
S
dz. (6.28)
Note that the stock price alone S(t) does not appear in the tradable table since it is not tradable. Only
tradable quantities appear in the tradables table, and only those quantities need to satisfy the absence of
arbitrage conditions.
The nal step is to solve the Price APT equation P
0
+B
1
= A,
B
v
c
0
+
0
v
Sc
S
1
=
r
0
B
v
c
t
+Sc
S
+
1
2
2
S
2
c
SS
. (6.29)
Solving for
0
and
1
gives:
0
= r
0

1
=
+q r
0
, (6.30)
and plugging
0
and
1
into the last equation gives
r
0
c +
+q r
0
Sc
S
= c
t
+Sc
S
+
1
2
2
S
2
c
SS
. (6.31)
c
t
+ (r
0
q)Sc
S
+
1
2
2
S
2
c
SS
= r
0
c. (6.32)
What does this apply to?
At rst thought, a continuous model for a dividend does not seem too realistic. However, a moment of
thought more, we realize that it is a decent approximation for a number of nancial situations. For instance,
a stock index with many stock that pay dividends at dierent times can be approximated as a continuous
dividend. So can a commodity with a convenience yield. Moreover, foreign currencies that are invested in a
money market account are essentially a security that earns a continuous dividend. So, the point is, dont let
the notation of calling S a stock limit your thinking about how these models can be applied.
European Call Option
Once again, we can consider the value of a European call option, but this time on a stock that pays a
continuous dividend. The solution is
d
1
=
ln(S/K) + (r
0
q +
1
2
2
)(T t)
T t
(6.33)
d
2
= d
1
T t (6.34)
c(S, t) = Se
q(Tt)
N(d
1
) Ke
r0(Tt)
N(d
2
) (6.35)
Note that this is very similar to the formula for a European call option on a non-dividend paying stock, except
that q shows up in a couple of places. This is too be expected, since (6.42) diers from the non-dividend
Black-Scholes equation only slightly.
6.1.3 Cash Dividends
Most individual stocks pay a prespecied dividend at a prespecied time. We often call this either a cash or
lump dividend. To compute the equation satised by an option on a stock that pays a cash dividend prior
to expiration, we can once again use our three step procedure.
Of course, the bond is a tradable. Now, we can purchase a share of the stock, and with this we will get
the stock price plus the dividend. This entire stream together it what is tradable.
To model this, we consider that the value corresponding to a single share is continuous and follows a
geometric Brownian motion. Call this value v(t), where it satises
dv = vdt +vdz. (6.36)
Now, we assume that the stock pays a dividend at the time in the amount of D. Therefore, we model the
stock price as dropping by D exactly at the time . To model this drop, we will use dirac delta notation
and write it as
dS = (S D(t ))dt +dz (6.37)
where (t ) is a dirac delta function at time . This notation is just to indicate that the price drops by
exactly D at time .
Now, we assume that c(S, t) is a derivative security, and we apply Itos lemma to obtain
dc = (c
t
+Sc
S
+
1
2
2
S
2
c
SS
+ (c(S
D, t) c(S
, t))(t ))dt +Sc

S
dz (6.38)
I have not described how to apply Itos lemma in this situation before. However, you may recognize that
this situation is very similar to a process driven by a Poisson process, since they have discrete jumps. In this
case, it is even simpler, since we know exactly when the jump will occur. The above equation just indicates
that we use Itos lemma for Brownian motion up to, and after the jump time . However, at time the
jump occurs, which causes c to drop by exactly D. Once again, the delta function (t ) is really just used
to indicate that a jump occurs at time .
Therefore, we have identied the tradables B, v, and c, and now we can write the tradables table.
B
v
c
B
v
c
r
0
B
v
c
t
+Sc
S
+
1
2
2
S
2
c
SS
+ (c(S
D, t) c(S
, t))(t )
dt +
0
v
Sc
S
dz
The nal step is to solve the Price APT equation
B
v
c
0
+
0
v
Sc
S
1
=
r
0
B
v
c
t
+Sc
S
+
1
2
2
S
2
c
SS
+ (c(S
D, t) c(S
, t))(t )
(6.39)
Solving for
0
and
1
gives:
0
= r
0

1
=
r
0
(6.40)
Plugging
0
and
1
r
0
c +
r
0
Sc
S
= c
t
+Sc
S
+
1
2
2
S
2
c
SS
+ (c(S
D, t) c(S
, t))(t ) (6.41)
c
t
+r
0
Sc
S
+
1
2
2
S
2
c
SS
+ (c(S
D, t) c(S
, t))(t ) = r
0
c. (6.42)
6.1.4 Poisson Processes
This was done by Cox and Ross [5]. This set-up involves a bond earning a risk free rate and a non-dividend
paying stock that follows a geometric Poisson motion (GPM):
dB = r
0
Bdt (6.43)
dS = S
dt + (k 1)S
d() (6.44)
where is the intensity of the Poisson process (t; ). Furthermore, we consider a derivative on the stock.
Generically we call this derivative c and assume that its price process depends on the stock and time c(S, t)
and is continuously dierentiable in both its arguments. By Itos lemma,
dc = (c
t
+Sc
S
)dt + (c(kS
) c(S
))d(). (6.45)
This is a case where the random factor does not have zero mean. I would like to write my factor equations
such that the factors are pure risk and dont have any mean drift. Hence, I will compensate the factor in
order to give it zero drift. In this case, my tradables are
dB = r
0
Bdt (6.46)
dS = (S +(k 1)S)dt + (k 1)S(d() dt) (6.47)
dc = (c
t
+Sc
S
+(c(kS
) c(S
)))dt + (c(kS
) c(S
))(d() dt) (6.48)

We can now construct a tradables table
B
S
c
B
S
c
r
0
B
S +(k 1)S
c
t
+Sc
S
+(c(kS
) c(S
))
dt +
0
(k 1)S
c(kS
) c(S
(d() dt)
This allows us to move on to Step 3, which is to solve the Price APT equation P
0
+B
1
= A for absence
of arbitrage
B
S
c
0
+
0
(k 1)S
c(kS
) c(S
1
=
r
0
B
S +(k 1)S
c
t
+Sc
S
+(c(kS
) c(S
))
.
Solving for
0
and
1
gives:
0
= r
0

1
=
r
0
k 1
+. (6.49)
Plugging
0
and
1
r
0
c +
r
0
k 1
+
(c(kS
) c(S
)) = c
t
+Sc
S
+(c(kS
) c(S
)) (6.50)
Finally, rearranging leads to the equation:
( r
0
) (c(kS
) c(S
)) = (k 1) (c
t
+Sc
S
r
0
c) . (6.51)
One can actually nd a closed form solution for European call and put options in this model. I will give the
solution for a call option with strike K and expiration T:
c(S, t) = S(x, y) Ke
r0(Tt)
(x, y/k) (6.52)
where
(, ) =
i=
e
i
i!
, y =
(r
0
)(T t)k
k 1
(6.53)
and x is the smallest non-negative integer greater than
ln(K/S)(Tt)
ln(k)
.
Note that this solution has a structure that is very similar to the Black-Scholes fomula. In fact, one can
interpret it as the Black-Scholes formula, but with a Poisson random variable replacing the Gaussian random
variable.
6.1.5 Options on Futures
This was essentially done by Black in [1]. Lets derive the pde satised by an option on a futures contract.
Assume there exists a bond and a futures contract with futures price f given as
dB = r
0
Bdt (6.54)
df = fdt +fdz. (6.55)
Furthermore, we consider a derivative on the futures price. Generically we call this derivative c and assume
that its price process depends on the futures price and time c(f, t) and is twice continuously dierentiable
in both its arguments. By Itos lemma,
dc = (c
t
+fc
f
+
1
2
2
f
2
c
ff
)dt +fc
f
dz. (6.56)
We can now complete Step 2 and construct a tradables table
B
0
c
B
f
c
r
0
B
f
c
t
+fc
f
+
1
2
2
f
2
c
ff
dt +
0
f
fc
f
dz. (6.57)
Recall that the futures contract was the special case that motivated us to consider the price approach
to absence of arbitrage, rather than working with returns. The key point was that the mark-to-market
mechanism always sets the price of a futures contract to zero while the change in value of the contract over
period dt is given by the change in the futures price df. This corresponds to the tradable table we have
written above.
This allows us to move on to Step 3, which is to solve the Price APT equation P
0
+B
1
= A for absense
of arbitrage.
B
0
c
0
+
0
f
fc
f
1
=
r
0
B
f
c
t
+fc
f
+
1
2
2
f
2
c
ff
(6.58)
Solving for
0
and
1
gives:
0
= r
0

1
=

. (6.59)
Plugging
0
and
1
r
0
c +

fc
f
= c
t
+fc
f
+
1
2
2
f
2
c
ff
. (6.60)
Finally, rearranging leads to the partial dierential equation:
c
t
+
1
2
2
f
2
c
ff
= r
0
c. (6.61)
Note that the Black-Scholes equation in this case (6.61), looks just like the Black-Scholes equation on a stock
paying a continuous dividend, however the continuous dividend rate is q = r. Therefore, we can just plug
into that formula to obtain the price of a European call option. The solution is
d
1
=
ln(f/K) + (
1
2
2
)(T t)
T t
(6.62)
d
2
= d
1
T t (6.63)
c(f, t) = fe
r0(Tt)
N(d
1
) Ke
r0(Tt)
N(d
2
). (6.64)
6.1.6 Jump diusion
A jump diusion model was rst solved by Merton [11]. This model is nice because it is related to many
other models in equity and interest rate derivatives. The model includes a risk free bond and an underlying
asset that has a diusion portion and a lognormal jump portion. This model has a closed form solution for
the derivative price which Merton computed in his original paper [11].
We will also nd that this model creates a problem for our factor approach. Merton originally solved this
problem using a similar technique, and we will bypass the problem in the same manner that Merton did.
Lets get started. Here are the basic assets
dB = r
0
Bdt (6.65)
dS = ( +E[Y 1])Sdt +Sdz +S
((Y 1)d() E[Y 1]dt) (6.66)

where the jump portion of the stock has been compensated. A derivative on the stock c(S, t) is a function
of S and t. By Itos lemma we have
dc = Lcdt +Sc
S
dz +
(c(Y S
) c(S
))d E[(c(Y S
) c(S
))]dt
(6.67)
where
Lc = c
t
+Sc
S
+
1
2
2
S
2
c
SS
+E[(c(Y S
) c(S
))] (6.68)
Now we come to the point that we need to write a tradables table. This is where we run into a problem.
The problem is how to identify factors. The randomness associated with the jump term does not enter in
the same way to S and c. We would like to be able to write down a linear factor model in the factors dz
and Y d or a compensated version of of Y d. However, in this case that is not possible unless we consider
c(Y S
) c(S
)d a new factor. However, since this risk is driven by the same Poisson process and jump
size Y , we would rather not do this. Yet, at this point we have no choice.
So, lets proceed by considering ((c(Y S
) c(S
))d E[(c(Y S
) c(S
))]dt) and (Y 1)d()

E[Y 1]dt as two dierent factors. For notational convenience, lets call them
d
1
= (Y 1)d() E[Y 1]dt (6.69)
d
2
=
(c(Y S
) c(S
))d E[(c(Y S
) c(S
))]dt
(6.70)
Then our tradables table is
B
S
c
B
S
c
r
0
B
( +E[Y 1])S
Lc
dt +
0 0 0
S S 0
Sc
S
0 1
dz
d
1
d
2
(6.71)
Finally, we would solve the Price APT equations
B
S
c
0
+
0 0 0
S S 0
Sc
S
0 1
r
0
B
( +E[Y 1])S
Lc
(6.72)
Now, we can write out the full pde which will have a number of unknown market prices of risk. This is
fairly messy and leaves a lot of degrees of freedom. Merton doesnt do this. Instead he make the assumption
that all jump risk is diversiable. That is, the market price of risk of any risk associated with the jump term
is zero! This is the same as setting
2
=
3
= 0. This is a strong assumption! But lets follow Merton and
see where this leads us.
With
2
=
3
= 0, the Price APT equations P
0
+B = A become
B
S
c
0
+
0
S
Sc
S
r
0
B
( +E[Y 1])S
Lc
. (6.73)
Now we can solve for the market prices of risk as,
0
= r
0
,
1
=
+E[Y 1]
, (6.74)
and the nal equation becomes
r
0
c +
+E[Y 1] r
0
Sc
S
= Lc (6.75)
which can be rewritten as
c
t
+ (r
0
E[Y 1])Sc
S
+
1
2
2
S
2
c
SS
= r
0
c E[(c(Y S
) c(S
))]. (6.76)
Bankruptcy
Lets consider a special case of complete bankruptcy. That is, Y = 0. Then the above equation becomes
c
t
+ (r
0
+)Sc
S
+
1
2
2
S
2
c
SS
= (r
0
+)c. (6.77)
This looks like standard Black-Scholes but the interest rate has been increased by the default probability!
We will see that this same relationship will also appear in defaultable bonds. In fact, this model is really
the prototype for defaultable bonds. Of course, we can give a closed form solution in this case based on the
Black-Scholes formula.
Note the following counterintuitive observation. In the Black-Scholes formula, the value of a European
call option increases with the risk free rate. This means that according to our model above, if the rate
of bankruptcy increases, then the value of call option will actually increase! You should think about this
carefully to understand why that is true...
Lognormal Jumps!
When Y the jump size is lognormal, then conditional on the number of jumps that have occurred before
expiration, stock distribution at expiration is lognormal, and there is a closed form solution to the pde for
European call and put options. For a European call option with strike K and expiration T, it is given by
c(S, t) =
n=0
(Tt)
(
(T t))
n
n!
c
BS
S, T t, K,
2
+
n
2
T t
, r
0
k +
n
T t
(6.78)
where c
BS
(S, T, K, , r
0
) is the Black-Scholes formula for a European call option with strike K and expiration
T on a non-dividend paying stock with current price S, volatility , and with risk free rate r
0
. We also have
that k = E[Y 1],
= (1 +k), and = ln(1 +k). This formula can be found in Mertons work [11].
Note that it is basically a combination of the Black-Scholes formula and the solution under Poisson
dynamics. The key is that under this jump diusion model, conditional on the number of jumps, the price
distribution at expiration is log-normal, indicating that the solution should look like Black-Scholes. The
only question is how many jumps have occurred. Therefore, the solution is basically that conditioned on
the number of jumps that have occurred, the answer is Black-Scholes. Hence, for each possible number of
jumps, we have a Black-Scholes formula, then we need to weight them by the probability of that number of
jumps, which is Poisson. Simple!
6.1.7 Exchange one asset for another
Margrabes [10] exchange one asset for another is one of my favorite derivatives. Why? because many other
derivatives can be thought of as exchanging two dierent assets. For instance, can a vanilla call option be
thought of as exchanging one asset for another? What are the two assets being exchanged?
Anyway, in this case, tradable assets will be called the bond B and two other assets S
1
and S
2
. Their
dynamics are given by:
dB = r
0
Bdt (6.79)
dS
1
=
1
S
1
dt +
1
S
1
dz
1
(6.80)
dS
2
=
2
S
2
dt +
2
S
2
dz
2
(6.81)
dc = L
1
cdt + +S
1
c
S1
dz
1
+S
2
c
S2
dz
2
(6.82)
where
L
1
c = (c
t
+
1
S
1
c
S1
+
2
S
2
c
S2
+
1
2
2
1
S
2
1
c
S1S1
+
1
2
2
2
S
2
2
c
S2S2
+
1
2
S
1
S
2
c
S1S2
) (6.83)
and is the correlation coecient between z
1
and z
2
. Thus, the tradable table is
B
S
1
S
2
c
B
S
1
S
2
c
r
0
B
1
S
1
2
S
2
Lc
dt +
0 0
1
S
1
0
0
2
S
2
1
S
1
c
S1

2
S
2
c
S2
dz
1
dz
2
. (6.84)
Finally, we solve the Price APT equations
B
S
1
S
2
c
0
+
0 0
1
S
1
0
0
2
S
2
1
S
1
c
S1

2
S
2
c
S2
r
0
B
1
S
1
2
S
2
Lc
. (6.85)
In this case we have four equations and only three unknowns. Lets use the rst three equations to solve
for
0
,
1
and
2
as
0
= r
0

1
=

1
r
0
2
=

2
r
0
2
. (6.86)
If we plug these into the nal equation, we obtain
Lc = r
0
c +

1
r
0
1
S
1
c
S1
+

1
r
0
1
S
1
c
S1
(6.87)
which upon rearrangement becomes
c
t
+r
0
S
1
c
s1
+r
0
S
2
c
s2
+
1
2
2
1
S
2
1
c
s1s1
+
1
2
2
2
S
2
2
c
s2s2
+
1
2
S
1
S
2
c
s1s2
= r
0
c. (6.88)
Closed form solution
There is a closed form solution for a European exchange one asset for another. Assume that you have the
option to exchange asset S
1
for asset S
2
at time T. Hence, the boundary condition is
c(S
1
, S
2
, T) = max(S
2
S
1
, 0), c(0, S
2
, t) = S
2
, c(S
1
, 0, t) = 0. (6.89)
The closed form solution for this option is
c(S
1
, S
2
, t) = S
2
N(d
1
) S
1
N(d
2
) (6.90)
where
d
1
=
ln(S
2
/S
1
) +
1
2
2
(T t)
T t
(6.91)
d
2
= d
1
T t (6.92)
=
2
1
+
2
2
2
1
2
. (6.93)
Note that this solution looks very much like Black-Scholes. Here is some intuition into why. Lets look at
the payo
c(S
1
, S
2
, T) = max(S
2
S
1
, 0) = S
1
max(
S
2
S
1
1, 0).
Now, the portion max(
S2
S1
1, 0) looks like the payo of a call option with strike 1 on the asset
S2
S1
! In fact,
dividing S
2
by S
1
is essentially changing units in order to value S
2
in units of S
1
. Therefore, in units of S
1
,
this is just a call option on S
2
with strike 1! The multiplication on the left by S
1
just converts the call option
back to units of dollars. Therefore, we shouldnt be surprised that the solution looks like Black-Scholes,
because this is really just a call option in a dierent set of units.
When we change units by denominating everything in terms of another asset (such as S
1
in this case),
we call that a change of numeraire, and we call the asset S
1
the numeraire asset. Needless to say, some
problems (such as this one) are easier to solve in a convenient set of units.
6.1.8 Stochastic volatility
First, we can list the relevant equations
dB = r
0
Bdt (6.94)
dS = Sdt +
vSdz
1
(6.95)
dv = adt +bdz
2
(6.96)
and we assume that z
1
and z
2
are correlated E[dz
1
dz
2
] = dt.
Our derivative can depend on S, t, and v. Hence, we write c(S, v, t). (What if I didnt assume the
derivative was a function of v? This would have been a bad assumption since I know in the Black-Scholes
case that c is a function of the volatility. Would I end up getting the wrong answer?)
By Itos lemma we have:
dc = Lcdt +
vSc
S
dz
1
+bc
v
dz
2
(6.97)
where
Lc = (c
t
+Sc
S
+ac
v
+
1
2
vS
2
c
SS
+
1
2
b
2
c
vv
+b
vSc
Sv
) (6.98)
In this case the tradables are
dB = r
0
Bdt (6.99)
dS = Sdt +
vSdz
1
(6.100)
dc = Lcdt +
vSc
S
dz
1
+bc
v
dz
2
(6.101)
6.2. PROBLEMS 59
So the tradable table is
B
S
c
B
S
c
r
0
B
S
Lc
dt +
0 0
vS 0
vSc
S
bc
v
dz
1
dz
2
. (6.102)
B
S
c
0
+
0 0
vS 0
vSc
S
bc
v
r
0
B
S
Lc
. (6.103)
Solving for the market prices of risks
0
and
1
gives
0
= r
0

1
=
r
0
v
. (6.104)
Note that we leave
2
as an unknown. Plugging these into the nal equations leads to
c
t
+r
0
Sc
S
+ (a
2
b)c
v
+
1
2
vS
2
c
SS
+
1
2
b
2
c
vv
+b
vSc
Sv
= r
0
c (6.105)
Under specic choices of a and b, for a European call option, this pde has a fairly convenient solution via
transform methods. I wont cover this in detail here. See Heston [9] if interested.
6.2 Problems
Problem 6.2.1 Show that the Black-Scholes formula (6.14) is a special case of exchange one asset for
another where one asset is the stock S(t) and the other is the bond B(t). In particular, show that the
exchange one asset for another formula (6.90) reduces to the Black-Scholes formula (6.14) in this case.
Problem 6.2.2 (The power of linearity (Breeden and Litzenberger)) We will use linearity to derive the
formula for a European digital option in terms of a European call option price. A digital option is an option
that pays o 1 if the underlying stock ends up above the strike price K, and pays nothing if it ends up below
the strike price.
(a) Let T be the expiration date for the digital option. Draw a picture of the payo as a function of S
T
.
(b) Assume that you know the price of European call options for every strike K for the expiration date
T (i.e. c(S, K, t)). Using only this information, derive the price of a digital option with strike K in
terms of c(S, K, t). (Hint: think of how you can construct the payo of a digital option by buying and
selling call options.)
(c) Taking the above argument one step further, derive the price of a security whose payo is a dirac
delta function at K at expiration T.
(d) Use the result of (c) to derive a general pricing formula for an arbitrary payo at time T. Note that if
you have made it this far, you have derived the risk neutral pricing formula in terms of c(S, K, t).
(If you plug in the Black-Scholes formula for c(S, K, t) then you have the standard risk neutral pricing
formula when the underlying asset follows a geometric Brownian motion.)
(e) As one nal task, price a security whose payo is (S
T
K)
2
for S
T
K and 0 for S
T
< K.
(Here is a big hint if you are having trouble with this problem. We can think of the payo of a European
call option as a ramp function. A digital option looks like a step function, and nally we price a dirac delta
function. What is the relationship between a ramp, step and delta function?)
Problem 6.2.3 All about Put-Call parity and Early Exercise.
(a) Derive put-call parity for European call and put options on a non-dividend paying stock.
(b) Early exercise is not optimal for a standard American call option on a non-dividend paying stock.
Derive this.
(c) Assume the payo of a derivative security on a non-dividend paying underlying stock is a convex
function that is non-positive when the stock is zero. Prove that if this is an American option, it is
never optimal to exercise early.
(d) Consider the European option to exchange one asset, S
2
, for another S
1
(assume they are non-dividend
paying) at expiration T. That is, if you currently hold S
2
, this option would give you the right to
exchange it for S
1
at time T. This is a derivative security with payo c(S
1
, S
2
, T) = max[S
1
S
2
, 0].
Derive a parity relating S
1
, S
2
and the options to exchange S
1
for S
2
and vice-versa.
(e) (Margrabe) Consider an American option to exchange one asset for another (non-dividend paying
again). Is it ever optimal to exercise this early?
(f ) Merton showed that it can be optimal to exercise a put option early, even on a non-dividend paying
stock. However, we can think of a European put option as exchanging a stock for a bond, in which case,
the American counterpart would never be exercised early according to Margrabe. Is there a contradiction
here?
Problem 6.2.4 (Exchange one asset for another) In this problem you will derive the formula for an option
to exhange one asset for another by reducing the pde to a standard Black-Scholes pde for a call option on
geometric Brownian motion.
(a) Let S
1
and S
2
be two assets whose dynamics are given by
dS
1
=
1
S
1
dt +
1
S
1
dz
1
(6.106)
dS
2
=
2
S
2
dt +
2
S
2
dz
2
(6.107)
where E[dz
1
dz
2
] = dt and assume that a risk free asset exists with constant interest rate r. Write down
the pde for an option to exchange S
2
for S
1
at expiration T. (Hence, the payo is max(S
1
S
2
, 0).)
(b) Consider the change of variable v = S
1
/S
2
and assume that the solution of the above pde is of the form
c(S
1
, S
2
, t) = S
2
f(v, t). Using these substitutions, write the pde from (a) in terms of S
2
, t, f, and v.
What equation must f(v, t) satisfy? What is the appropriate boundary condition for a European option
to exchange S
2
for S
1
in terms of the new variables.
(c) Using the Black-Scholes formula, write down a formula for the value of an option to exchange S
1
for
S
2
.
Problem 6.2.5 Consider a European call option on a non-dividend paying stock with stochastic volatility.
But, this time we model the instantaneous variance with a Poisson process. The idea is that most of the time
the instantaneous variance is
2
l
, but at random times the market goes wild and the instantaneous variance
jumps by an amount b. After the jump, the instantaneous variance exponentially decays back to its normal
level
l
. First write down the relevant dynamics for this problem, then derive a pde for the price of the
option. Again, this pde can be in terms of a market price of risk.
(Hint: your dynamics should look like:
dS = Sdt +
vSdz
dv = a(
2
l
v)dt +bd
where is a Poisson process with intensity .)
6.2. PROBLEMS 61
Problem 6.2.6 (Matlab Exercise) This exercise will introduce you to the implied volatility curve. If c
m
is
the market value of a call option with strike K, expiration T, and if c
BS
(S, K, T, r
0
, ) (assuming the stock is
non-dividend paying) is the Black-Scholes formula for a European call option. Then the value of the volatility
parameter
impl
that satises
c
m
= c
BS
(S, K, T, r
0
,
impl
) (6.108)
is known as the implied volatility. In this exercise, you will generate prices under Mertons jump-diusion
model, and compute implied volatility curves.
a. Use an initial stock price of S(0) = 1, expiration of T = 0.3, and a risk free rate of r = 0.05. For
a range of strike prices from K = [0.8, 1.2], rst compute the Black-Scholes price of options using
= 0.3, then assume that they represent market prices and plot the resulting implied volatility curve
as a function of K.
b. This time, use Mertons jump-diusion model to to generate the market prices. Use the same parameter
values as in part (a) (i.e. S(0) = 1, T = 0.3, r = 0.05, and = 0.3), and also jump intensity of = 2,
jump mean of
J
= 0 and jump standard deviation of
J
= 0.1. Once again plot the resulting implied
volatility curve. You should see a slight implied volatility smile!
Problem 6.2.7 (American Options) Use the results of Chapter 4, Problem 3 to derive the pde conditions
followed by an American call or put option where the underlying stock (non-dividend paying) and bond follow
dS = Sdt +Sdz
dB = r
0
Bdt.
Chapter 7
Application of the Factor Form:
Interest Rate and Credit Derivatives
In this chapter, we apply the factor approach to interest rate and credit derivatives. Again, the emphasis is
showing that the pdes describing derivative pricing can all be easily derived from the simple factor approach
equations.
Before getting started, lets make a few comments about interest rate modeling and derivatives. For any
derivative pricing problem, the rst challenge before being able to price a derivative is the calibration phase
as in Figure 5.3 of Chapter 5. For many equity derivatives, the calibration phase does not play prominently
into the analysis. In fact, in many equity derivative models the market is either complete in which case the
calibration phase of determining is simple, or the market is incomplete in which case the APT equations
are underdetermined and one is left with a degree of freedom in choosing .
Interest rate and term structure models tend to be the opposite. That is, we may use a single factor, but
have many, many marketed tradables. This makes the APT equations overdetermined! Thus, the calibration
phase is in determining values that best t all the marketed tradables (In theory if there isnt a perfect
t, then an arbitrage is available. However in practice we simply recognize that our model is too simple
and try to nd a best-t ). Once calibration is done, then pricing a derivative just proceeds by solving the
appropriate pde with boundary conditions using the calibrated . Because calibration is so important in
interest rate and term structure models, in what follows, we will often bring the models up to the point of
calibration, but not consider specic derivatives after that point.
Beyond this, the truth is that some of these models (such as HJM) are better suited for the risk neutral
approach to derivative pricing (A subject that is touched upon in Chapter 9) simply because the pdes that
we obtain are often too large and dicult to solve in any reasonable manner. Nevertheless, it is extremely
instructive to see that they can be understood via the factor approach. Since I have mentioned risk neutral
pricing, I should also mention that calibration in risk neutral pricing often looks dierent from what I
have called the calibration phase (which is just determining the market prices of risk from the marketed
tradables). However, they are really the same thing, just described using a dierent language. Hopefully,
at the end of your studies of derivative pricing you will be able to translate between them.
Okay, lets get to pricing...
7.1 Notation and the Money Market Account
In dealing with bonds and interest rate derivatives, we will need to establish some notation. Lets start with
the short rate process. The short rate r
0
(t) is the interest rate earned from time t to time t + dt quoted
using continuous compounding. (Note that the notation is similar to the constant risk free rate r
0
used in
the previous chapter. This is because the short rate plays the same crucial role as the constant risk free rate
in that it denes the market price of time,
0
.)
63
64CHAPTER 7. APPLICATIONOF THE FACTOR FORM:INTEREST RATE ANDCREDIT DERIVATIVES
On a plot of the term structure, the short rate is where is spot rate curve intersects the y-axis as in
Figure 7.1. The short rate process is important because it can be used to drive the entire spot rate curve.
Figure 7.1: The Short Rate Process
Furthermore, the short rate r
0
(t) describes the instantaneous return on the money market account. That
is, we dene the money market account B
0
(t) as the value of an account that is continuously rolled over at
the instantaneous short rate. That is, you invest your money at the rate r
0
(t) over the next dt, then you
reinvest that amount over the next dt at the rate r
0
(t + dt), and continue this. Hence, the money market
account B
0
(t) follows the dynamics
dB
0
(t) = r
0
(t)B
0
(t)dt (7.1)
The money market account often plays a special role in interest rate derivatives because its dynamics are
not explicitly driven by a random factor. Note that this is the case even though r
0
(t) itself can be random
and follow a stochastic dierential equation of its own. Pay careful attention to the role that the money
market account plays in what follows.
7.2 Interest Rate Derivatives
7.2.1 Single Factor Short Rate Models
The simpliest interest rate derivative model has a single underlying variable: the short rate. Furthermore,
this variable is driven by a single factor. Hence, these models are typically called single factor short rate
models. Vasicek [15] was the rst to recognize that pricing with absence of arbitrage was possible by modeling
the instantaneous short rate. Hence, we will model this short rate variable as
dr
0
= adt +bdz (7.2)
where the time dependence is supressed in r
0
= r
0
(t), and a = a(r
0
, t), b = b(r
0
, t) can be functions of r
0
and
t. When convenient, we will suppress these arguments, so dont always assume that a variable is a constant
if no arguments are explicitly listed.
The tradables are a money market account, which we denote by B
0
and zero coupon bonds with face
value of $1 of varying maturity that we denote by B(t|T) where t is the current time and T is the maturity
date. This is pictured in Figure 7.2. We assume that the zero coupon bonds are functions of the short rate,
B(r
0
, t|T). Then suppressing arguments, we can write
dB
0
= r
0
B
0
dt (7.3)
dB(T) = (B
t
(T) +aB
r
(T) +
1
2
b
2
B
rr
(T))dt +bB
r
(T)dz (7.4)
7.2. INTEREST RATE DERIVATIVES 65
Figure 7.2: Notation for Zero Coupon Bond Prices
where Itos lemma was used to derive the equation for dB(T), and B
r
, B
rr
, represent the rst and second
partial derivatives with respect to the short rate r
0
, respectively. In particular, not that for convenience
when using this partial derivative notation we supress the subscript r
0
and write B
r
= B
r0
.
Now lets write down the tradable table:
B
0
B(T)
B
0
B(T)
r
0
B
0
(B
t
(T) +aB
r
(T) +
1
2
b
2
B
rr
(T))
dt +
0
bB
r
(T)
dz (7.5)
B
0
B(T)
0
+
0
bB
r
(T)
1
=
r
0
B
0
(B
t
(T) +aB
r
(T) +
1
2
b
2
B
rr
(T))
(7.6)
The rst equation gives
0
= r
0
which is the random short rate. Then the second equation gives
(B
t
(T) + (a
1
b)B
r
(T) +
1
2
b
2
B
rr
(T)) = r
0
B(T) (7.7)
The boundary condition is of course that B(T|T) = 1. This equation must hold for zero coupon bonds of
any maturity. Note that it is in terms of a market price of risk. In fact, since in general we have many bonds,
any one of them should allow us to solve for
1
. In practice, dierent bonds will often give dierent values
of
1
, indicating that the model is not exactly correct.
When one tries to determine a single best t
1
from the bond data, we are treating all the bonds as
being marketed tradables, and thus this is interpreted as a calibration phase. After this calibration phase,
one could then use this model to price other interest rate derivatives such as caps, oors, bond options, etc.
7.2.2 Multi-Factor Short Rate Models
The single factor short rate models are usually not good enough to describe the term structure well. So,
instead of using a single factor model, we can use a multifactor model as follows.
Let X be a vector in R
n
of underlying variables aecting the term structure. We assume that these
variables follow a stochastic dierential equation model
dX = f(X, t)dt +g(X, t)dz (7.8)
where z R
n
is a vector of uncorrelated Brownian motions. (Thus, g(X, t) R
nn
and f(X, t) R
n
.)
We then assume that the short rate is a function of the underlying variables X. That is, r
0
(X, t). Note
that one possibility is to have the short rate be one of the variables in X. Thus, it is possible to choose
r
0
(X, t) = X
i
, where X
i
is the i th factor.
Now, to derive the pde that zero coupon bonds would follow, we just note that bonds are a function of
X and t. Hence, we have B(X, t|T). Via Itos lemma, our tradables become
B
0
B(T)
B
0
B(T)
r
0
(X, t)B
0
B
t
(T)+B
X
(T)f(X, t)+
1
2
Tr[B
XX
(T)g(X, t)g
T
(X, t)]
dt +
0
B
r
(T)g(X, t)
dz
where I have suppressed the X and t arguments in B. Now, by the price APT, we have
B
0
B(T)
0
+
0
B
X
(T)g(X, t)
r
0
(X, t)B
0
(B
t
(T) +B
X
(T)f(X, t) +
1
2
Tr[B
XX
(T)g(X, t)g
T
(X, t)])
(7.9)
where R
n
. The rst equation gives
0
= r
0
(X, t) which is the random short rate. Then the second
equation gives
(B
t
(T) +B
X
(T)(f(X, t) g(X, t)) +
1
2
Tr[B
XX
(T)g(X, t)g
T
(X, t)]) = r
0
(X, t)B(T) (7.10)
The boundary condition is of course that B(T|T) = 1. This can be a pretty complicated pde, and given that
it describes bonds as a function of , we could use the current term structure (or equivalently, bond prices)
to calibrate . This calibration can be quite dicult because bond prices are described by a pde, and to test
the t for dierent values of , we might have to solve the pde every time. It is much easier if the solution
to the pde is known in closed form as a function of , then we can quickly adjust to best t the market
data of bonds.
7.2.3 Heath-Jarrow-Morton
In the Heath-Jarrow-Morton [6] framework, instead of modeling the instantaneous short rate as driving
the term structure, they decided to model instantaneous forward rates. That is, let r(t|s) denote the
instantaneous forward rate seen from time t of the forward interest rate between time s and s + ds. With
this notational convention, we have that the instantaneous short rate is given by r(t|t) = r(t), and a zero
coupon bond with maturity T seen from time t will be denoted by B(t|T) and related to the instantaneous
short rates by
B(t|T) = exp
T
t
r(t|s)ds
. (7.11)
Hence, we also have that
r(t|T) =

T
ln(B(t|T)). (7.12)
HJM takes the instantaneous forward rates as the underlying variables, and models them as
dr(t|s) = (t|s)dt +(t|s)dz(t). (7.13)
This model is a bit more complicated than the previous single and multi-factor short rate models because
we see through equation (7.11) that B(t|T) is a function of an innite number of underlying variables r(t|s)
for s [t, T]! Because of this, if we were to write pricing formulas for derivatives, they would appear as
innite dimensional pdes!
Nevertheless, this model has an advantage over the previous single and multi-factor short rate models
in that equation (7.11) gives us an explicit relationship between the tradables and the (ininite) underlying
variables. Thus, instead of using a generic Itos lemma relationship between B(t|T) and the underlying
factors r(t|s), we will be able to make that relationship concrete by plugging in explicit partial derivatives
into Itos lemma.
Why is this helpful? Because this will allow us to pull the Price APT relationship for the marketed
tradables B(t|T) back to the underlying variables r(t|s) which opens up the possibility of doing calibration
Figure 7.3: Notation for Zero Coupon Bond Prices
directly on the underlying factors r(t|s) instead of the marketed tradables. If it is easier to work with market
data about instantaneous forward rates, then this can simplify the calibration process. To understand the
value of this, you should consider the calibration procedure that has to be done for a single or multi-factor
short rate model if no closed form bond pricing formula is known (i.e. no closed form solution to the pricing
pde is known), and compare that with using forward rate data to calibrate in the HJM model that we will
derive.
Okay, with those preliminaries out of the way, lets dive into the details. We will derive the Price APT
equations marketed tradables which are the bonds B(t|T). However, using the explicit relationship between
the bonds B(t|T) and the underlying variabels r(t|s) will allow us to pull the Price APT relationship back
to the underlying variables r(t|s) which would then allow us to calibrate using r(t|s) directly.
In this setup, we will take zero coupon bonds as our marketed tradables. That is, B(t|T) are tradables,
which means that we need to calculate dB(t|T) for our tradables table. This is a bit of a tricky calculation
because B(t|T) is really a function of an innite number of Ito processes r(t|s) for s [t, T] through the
equation
B(t|T) = exp
T
t
r(t|s)ds
(7.14)
To simplify our thinking, lets begin by assuming that s is indexed by k = 1...n so that B(t; T) will only
depend on n Ito processes. That is, let us think
B(t|T) = B({r(t|s
k
) : k = 1...n}, t|T) (7.15)
Then, by Itos lemma, we would have
dB =
B
t
(t|s) +
n
i=1
(t, s
i
)B
r(t|si)
+
n
i=1
n
j=1
1
2
(t|s
i
)(t|s
j
)B
r(t|si)r(t|sj)
dt
+
i=1
(t|s
i
)B
r(t|si)
dz
Now, to complete our computation of Itos lemma, we need to compute B
t
, B
r(t|si)
, and B
r(t|si)r(t|sj)
using
the discretization of (7.14):
B(t|T) = exp
k=1
r(t|s
k
)s
k
. (7.16)
This gives
B
t
exp
k=1
r(t|s
k
)s
k
r(t|t) = exp
T
t
r(t|s)ds
r(t|t) (7.17)
B
r(t|si)
exp
k=1
r(t|s
k
)s
k
s
i
= exp
T
t
r(t|s)ds
s
i
(7.18)
B
r(t|si)r(t|sj)
exp
k=1
r(t|s
k
)s
k
s
i
s
j
= exp
T
t
r(t|s)ds
s
i
s
j
. (7.19)
Plugging these into Itos lemma gives
dB(t|T) =
exp
k=1
r(t|s
k
)s
k
r(t|t)
n
i=1
(t|s
i
) exp
k=1
r(t|s
k
)s
k
s
i
+
1
2
n
i=1
n
j=1
(t|s
i
)(t|s
j
) exp
k=1
r(t|s
k
)s
k
s
i
s
j
dt
i=1
(t|s
i
) exp
k=1
r(t|s
k
)s
k
s
i
dz.
Taking continuous limits yields
dB(t|T) =
B(t|T)r(t|t) B(t|T)
T
t
(t|s)ds +
1
2
B(t|T)
T
t
T
t
(t|s)(t|r)drds
dt
B(t|T)
T
t
(t|s)ds
dz.
Now, applying the Price APT equation leads to
r(t|t)
T
t
(t|s)ds +
1
2
T
t
T
t
(t|s)(t|r)drds = r(t|t)
T
t
(t|s)ds
1
(7.20)
or
T
t
(t|s)ds
1
2
T
t
T
t
(t|s)(t|r)drds =
T
t
(t|s)ds
1
. (7.21)
The upshot of this calculation is that the mean and the volatilities of the forward rates r(t|s) must be related
through this equation. In fact, taking the partial derivative with respect to T gives
(t|T) (t|T)
T
t
(t|s)ds = (t|T)
1
(7.22)
which makes the relationship a little more explicit. Thus, we have pulled back the Price APT relationship
onto the underlying factors.
Again, to remind you, the point of this is that now we can calibrate from forward rate market date r(t|s)
rather than having to go through bond prices B(t|T).
This is just the calibration phase. To price some derivative we would not have an explicit relationship
between the derivative and the underlying variables r(t|s), and thus Itos lemma (to create the tradables
table) combined with the Price APT would lead to an innite dimensional pde. The point is that for actual
pricing of a new derivative, using the factor approach and pdes is not easy. This is one case in which risk
neutral pricing (Chapter 9) can help quite a bit.
7.2.4 The LIBOR Market Model
This model is similar to the HJM model except that it uses discrete forward rates instead of instantaneous
forward rates. Again, we will be able to use an explicit relationship between the marketed tradables of bonds
and the underlying variables of discrete forward rates to pull the Price APT relationship onto the underlying
variables. Once again, the idea is that this can make calibration easier. Due to its similarity with HJM, this
is also a model that is easier to price with using the risk neutral approach. Nevertheless, the factor approach
can provide us with an important perspective on this model. For further reading on this approach see [3].
Let R(t|T
1
; T
2
) be the forward interest rate between time T
1
and T
2
as seen at time t. For simplicity, let
T
i+1
= T
i
+ where is a xed amount of time. Furthermore, we use an interest rate convention so that
the price of a zero coupon bond is
B(t|T
1
) =
1
(1 +R(t|T
1
))
(7.23)
and for bond prices and forward rates we have
B(t|T
2
) =
B(t|T
1
)
(1 +R(t|T
1
; T
2
))
(7.24)
and more generally
B(t|T
i
) =
B(t|T
i1
)
(1 +R(t|T
i1
; T
i
))
(7.25)
etc for i = 1, ..., n.
Now, we will model the forward rates as stochastic dierential equations. For simplicity, we use a
geometric Brownian motion model.
dR(t|T
1
; T
2
) = a
1
R(t|T
1
; T
2
)dt +b
1
R(t|T
1
; T
2
)dz
2
(7.26)
For notational simplicity, lets write R
i
= R(t|T
i
, T
i+1
) and B
i
= B(t|T
i
) so that
dR
i
= a
i
R
i
dt +b
i
R
i
dz
i
(7.27)
Then prices of tradables satisfy
B
i
=
B
i1
1 +R
i1
(7.28)
Now, it should be clear that B
i
depends on R
i1
and hence the factor z
i1
, but also B
i1
. Now since B
i1
depends on R
i2
and B
i2
, etc., then B
i
ultimately depends on the factors z
i1
, z
i2
, z
i3
, ..., z
1
. Thus,
let us write generically that
dB
i
=
i
B
i
dt +B
i
i1
j=1
ij
dz
j
. (7.29)
The relationship above leads to a recursive dependence
B
i
= (1 +R
i
)B
i+1
. (7.30)
Using Itos lemma gives
dB
i
= dR
i
B
i+1
+ (1 +R
i
)dB
i+1
+dR
i
dB
i+1
(7.31)
= B
i+1
(a
i
R
i
dt +b
i
R
i
dz
i
) + (1 +R
i
)
i+1
B
i+1
dt +B
i+1
i
j=1
i+1,j
dz
j
(7.32)
+b
i
R
i
B
i+1
j=1
i+1,j
ij
dt (7.33)
=
B
i+1
(a
i
R
i
) +b
i
R
i
B
i+1
i
j=1
i+1,j
ij
+ (1 +R
i
)(
i+1
B
i+1
)
dt
+B
i+1
b
i
R
i
dz
i
+ (1 +R
i
)
B
i+1
i
j=1
i+1,j
dz
j
. (7.34)
This is the expression for our tradables. From this we can get two things. First, we note that by comparing
this to (7.29) we should have that
B
i+1
b
i
R
i
+ (1 +R
i
)B
i+1
i+1,i
= 0. (7.35)
Since the coecient of z
i
in (7.29) is zero. This allows us to solve for
i+1,i
as
i+1,i
=
b
i
R
i
(1 +R
i
)
. (7.36)
Futhermore, equating the coecients of z
j
for j < i in (7.29) and (7.34) gives
(1 +R
i
)B
i+1
i+1,j
= B
i
i,j
. (7.37)
Now, noting that B
i
= (1 +R
i
)B
i+1
gives
i+1,j
=
i,j
. (7.38)
Thus, by combining (7.36) and (7.38) we obtain
i+1,j
=
b
j
R
j
(1 +R
j
)
, i > j. (7.39)
Once we have established these relationships, we can move to constructing the tradables table and
applying the Price APT. Returning to (7.34), the tradables look like
dB
i
=
B
i+1
(a
i
R
i
) +b
i
R
i
B
i+1
i
j=1
i+1,j
ij
+ (1 +R
i
)(
i+1
B
i+1
)
dt
+B
i+1
b
i
R
i
dz
i
+ (1 +R
i
)
B
i+1
i
j=1
i+1,j
dz
j
(7.40)
= B
i
i+1
+
a
i
R
i
+b
i
R
i
i
j=1
i+1,j
ij
1 +R
i
dt
+B
i
i1
j=1
i+1,j
dz
j
(7.41)
where (7.39) was used in the last equality.
The Price APT then tells us that
i
= r
0
+
i1
j=1
i+1,j
(7.42)
where r
0
is the short rate process, and from (7.41),
i
=
i+1
+
a
i
R
i
+b
i
R
i
i
j=1
i+1,j
ij
1 +R
i
(7.43)
These equations mix parameters from the underlying variables (a
i
and b
i
), and the marketed tradables (
i
and
i,j
). We want conditions that dont involve anything related to the marketed tradables. Thus, we
would like to eliminate
i
and
i,j
terms.
To do this, lets substitute (7.42) into (7.43), giving
i1
j=1
i+1,j
=
i2
j=1
i,j
+
a
i
R
i
+b
i
R
i
i
j=1
i+1,j
ij
1 +R
i
(7.44)
This still contains
i,j
terms. But, we can substitute in using (7.39) to get rid of those terms and have
i1
j=1
b
j
R
j
(1 +R
j
)
i2
j=1
b
j
R
j
(1 +R
j
)
+
a
i
R
i
b
i
R
i
i
j=1
bjRj
(1+Rj)
ij
1 +R
i
(7.45)
or
i1
b
i1
R
i1
(1 +R
i1
)
a
i
R
i
b
i
R
i
i
j=1
bjRj
(1+Rj)
ij
1 +R
i
(7.46)
which is the calibration relationship pulled onto the underlying variables of the discrete forward rates R
i
.
(Note that with n tradables as bond prices, we only have n1 factors z
i
. Thus, there are only n1 market
prices of risk, and that is why the above market price of risk is
i1
and not
i
.)
Relations to HJM
The Libor Market Model is, of course, highly related to the HJM model. However, there are a couple of
dierence that are helpful to point out. The rst is that we derived that HJM as a single factor model (a
multifactor model is easy to do as well). That is, all the instantaneous forward rates were driven by the
same factor dz. In the Libor Market Model, each discrete forward rate is driven by a distinct factor dz
i
.
The next thing to note is that in the Libor Market Model, we used a recursive relationship to relate the
underlying variables to the marketed tradables
B
i
= (1 +R
i
)
B
i1
(1 +R
i1
)
(7.47)
which can be expanded to give
B
i
= (1 +R
i
)
B
i1
(1 +R
i1
)
=
i1
j=0
1
(1 +R
i1
)
(7.48)
This is, in fact, the discrete analog to the continuous equation used in HJM
B(t|T) = exp
T
t
r(t|s)ds
(7.49)
Thus, the two models (HJM and Libor Market Model) really are brothers.
Finally, the most important thing to note about the HJM and Libor Market Model is that they can
both be viewed and understood from the factor point of view. This is typically not done because the risk
neutral framework turns out to be better suited for these models. However, it is important to understand
that these models all come from just a simple application of the factor approach, despite their looking quite
complicated. I hope you have found this exercise valuable... lets move on to credit.
7.3 Credit Derivatives
Credit derivatives usually refer to derivatives that pay o depending on whether a bankruptcy has occured.
Since a bankruptcy is a sudden event, credit model rely heavily on Poisson processes. In what follows, I will
present the simplest models of a Defaultable Bond when the intensity of default is a constant, and then we
will generalize that to allow the intesity to be random. You should note the similarity between these models
and Mertons jump diusion model in Chapter 6, Section 6.1.6. In fact, if you understand Mertons model
then there isnt too much new here...
7.3.1 Defaultable Bonds
In deriving equations for defaultable bonds, we need two factors. The rst is the short rate, and the second
is a default factor. We model these as
dr
0
= adt +bdz (7.50)
dN = d() (7.51)
In this case, a defaultable bond of maturity T is a function or r
0
, N, and t: B(r
0
, N, t|T). Since N(t) is a
pure Poisson process, we can start it at N = 0 which is the no default state, and let N = 1 be the default
state. By Itos lemma we can write, after suppressing all arguments except N:
dB(N) = LB(N)dt +bB
r
(N)dz + (B(N + 1) B(N))(d() dt) (7.52)
where I have compensated the Poisson process and
LB(N) = (B
t
(N) +aB
r
(N) +
1
2
b
2
B
rr
(N) +(B(N) B(N + 1))) (7.53)
Now lets write down the tradables table, which includes the money market account:
B
0
B(N)
B
0
B(N)
r
0
B
0
LB(N)
dt +
0 0
bB
r
(N) (B(N + 1) B(N))

dz
d dt
(7.54)
B
0
B(N)
0
+
0 0
bB
r
(N) (B(N + 1) B(N))

1
r
0
B
0
LB(N)
(7.55)
From the rst equation we have
0
= r
0
the short rate. The second equation gives
r
0
B(N) +bB
r
(N)
1
+ (B(N + 1) B(N))
2
= LB(N) (7.56)
B
t
(N) +aB
r
(N) +
1
2
b
2
B
rr
+(B(N) B(N +1)) = r
0
B(N) +
1
bB
r
(N) +
2
(B(N +1) B(N)). (7.57)
But, we assume that we start with N = 0 and default is N = 1, so we have
B
t
(0) + (a
1
b)B
r
(0) +
1
2
b
2
B
rr
(0) = (r
0
+
2
)B(0) + (
2
)B(1). (7.58)
7.3. CREDIT DERIVATIVES 73
Dierent Types of Recovery
But, since B(0) is the no default state, it is the same as B. Furthermore, B(1) means that we are in default.
Hence, we can assume dierent recoveries in default such as no recovery B(1) = 0, or fractional recovery
B(1) = xB(0), or even xed recovery B(1) = X.
If we assume no recovery we obtain
B
t
(0) + (a
1
b)B
r
(0) +
1
2
b
2
B
rr
(0) = (r
0
+
2
)B(0) (7.59)
This looks exactly like the standard model for a bond, except that the short rate is increased by
2
.
That is, the default rate and its market price of risk just adjust the short rate of interest. Otherwise, the
price is the same. This means that if there is a closed form solution for bond prices under a single factor
short rate model, then there will also be a closed form solution for defaultable bonds!
7.3.2 Defaultable Bonds with Random Intensity of Default
This time we will allow the default intensity to follow a stochastic dierential equation. That is, we have
dr
0
= adt +bdz
1
(7.60)
dN = d() (7.61)
d = fdt +gdz
2
(7.62)
where E[dz
1
dz
2
] = dt. In this case, a defaultable bond of maturity T is a function of r
0
, N, and t:
B(r
0
, N, , t|T). Again, we start with N = 0 which is the no default state, and let N = 1 be the default
state. By Itos lemma we can write:
dB(N) = LB(N)dt +bB
r
(N)dz
1
+gB
(N)dz
2
+ (B(N + 1) B(N))(d() dt) (7.63)
where again I have compensated the Poisson process and this time
LB(N) = (B
t
(N)+aB
r
(N)+fB
(N)+
1
2
b
2
B
rr
(N)+
1
2
g
2
B
(N)+bgB
r
(N)+(B(N)B(N+1))) (7.64)
Now lets write down the tradables table:
B
0
B(N)
B
0
B(N)
r
0
B
0
LB(N)
dt +
0 0 0
bB
r
(N) gB
(N) (B(N + 1) B(N))
dz
1
dz
2
d dt
(7.65)
B
0
B(N)
0
+
0 0 0
bB
r
(N) gB
(N) (B(N + 1) B(N))
r
0
B
0
LB(N)
(7.66)
From the rst equation we have
0
= r
0
the short rate. The second equation gives
r
0
B(N) +bB
r
(N)
1
+gB
(N)
2
+ (B(N + 1) B(N))
3
= LB(N) (7.67)
B
t
(N) +aB
r
(N) +fB
(N) +
1
2
b
2
B
rr
(N) +
1
2
g
2
B
(N) +gbB
r
(N) +(B(N) B(N + 1))
= r
0
B(N) +
1
bB
r
(N) +
2
gB
(N) +
3
(B(N + 1) B(N)).
But, we assume that we start with N = 0 and default is N = 1, so we have
B
t
(0)+(a
1
b)B
r
(0)+(f
2
g)B
(0)+
1
2
b
2
B
rr
(0)+
1
2
g
2
B
(0)+gbB
r
(0) = (r
0
+
3
)B(0)+(
3
)B(1).
(7.68)
But, since B(0) is the no default state, it is the same as B(0). Furthermore, B(1) means that we are in
default. Hence, we can assume dierent recoveries in default such as no recovery B(1) = 0, or fractional
recovery B(1) = xB(0), or even xed recovery B(1) = X.
7.4 Problems
Problem 7.4.1 Derive a pde for the price of a European call option on a non-dividend paying stock when
interest rates are random, and the short rate follows Cox-Ingersoll-Ross dynamics. That is:
dS = Sdt +Sdz
1
(7.69)
dr
0
= a(b r
0
)dt +c
r
0
dz
2
(7.70)
where dz
1
and dz
2
are correlated E[dz
1
dz
2
] = dt. Your answer may contain a market price of risk.
Problem 7.4.2 (Options on Forwards and Futures)
(a) Assume interest rates are stochastic and driven by a single factor short rate model. Derive the pde for a
derivative on a futures contract and on a forward contract.
Assume the following
df = fdt +fdz
dr
0
= adt +bdz
B
where f is either the forward price or futures price and E[dzdz
B
] = dt.
Hint 1: Recall that futures prices are marked to market (See Section 6.1.5 in Chapter 6, whereas forward
prices are not marked to market.
Hint 2: Note that a forward price is not tradable. However, you may assume that the asset that the
forward contract is on is tradable on a spot market and there are standard relationships between the spot
price of an asset and its forward price.
Hint 3: My recommendation is to let your underlying variables be the future and forward price, re-
spectively, and the short rate. Additionally, it might be convenient to write bond dynamics generically as
dB =
B
Bdt +
B
Bdz
B
, where the short rate process is being driven by z
B
.
(b) What if interest rates are not stochastic? Do the pdes for options on forwards and futures look the same?
Problem 7.4.3 (PDE for a derivative on a dividend paying stock under stochastic interest rates)
Consider a market with a stock price process S(t) satisfying
dS = Sdt +Sdz
1
(7.71)
where z
1
(t) is a standard Brownian motion. Assume that this stock also pays a continuous dividend at a rate
of q. Thus, over time dt, the dividend amount is qSdt.
Additionally, assume that interest rates are random, and the term structure is driven by a single factor
short rate model
dr
0
= a(b r
0
)dt +fdz
2
(7.72)
where z
2
(t) is a standard Brownian motion and E[dz
1
dz
2
] = dt. Assume that a money market account
exists that satises
dB
0
= r
0
B
0
dt. (7.73)
Consider a third asset whose price at time t can be written as a twice dierentiable function of S(t), r
0
(t)
and t. Call this price function c(S(t), r
0
(t), t). If no arbitrage exists in the market, derive the pde that
c(S(t), r
0
(t), t) must satisfy. (Your answer may contain a market price of risk.)
Problem 7.4.4 (Flat Term Structure)
Consider a term structure model where the term structure of interest rates is at, but moves up and down
randomly. That is, let zero coupon bond prices with face value $1 and maturity T be denoted by B(t|T) and
satisfy
B(t|T) = exp(r(T t)) (7.74)
7.4. PROBLEMS 75
where r is modeled as a stochastic dierential equation
dr = a(r, t)dt +b(r, t)dz. (7.75)
Note that this model applies for all maturities T.
(a) Write down models for the instantaneous return of the money market account and for a generic zero
coupon bond with maturity T (i.e. for B(t|T)).
(b) Derive the absence of arbitrage condition in this case, and derive restrictions on a(r, t) and b(r, t). What
does this imply about the allowable dynamics for this term structure?
Chapter 8
Hedging
8.1 Introduction
We present two main approaches to hedging. In the rst, we recognize that risk comes from the factors.
Thus, hedging is based upon eliminating factor risk. In the second approach, we consider a derivative to be
a function of underlying variables. We are then hedged against variations in these underlying variables if
the derivative of the value of a portfolio with respect to the underlying variables is zero. Thus, our portfolio
is not sensitive to small variations in the underlying variables.
8.2 Hedging from a Factor Perspective
In abstract terms, hedging involves eliminating factor risk. Once again, we can either work with returns or
prices. In this chapter, we will work with prices.
Let dV
u
be the value change of the asset that we would like to hedge. Furthermore, we will assume that
we hold one unit of this asset. Thus a factor model for our unhedged portfolio prot/loss over the next time
instance is
dV
u
= A
u
dt +B
u
dz (8.1)
and we would like to hedge this position.
By hedging, we actually just mean that we will form a portfolio (that may be dynamically traded over
time) so that the factor model of the portfolio eliminates all factor risk. In this case, we have entirely hedged
out the risk.
Let us consider having various tradable assets to hedge with, and assume that these assets have a value
prot/loss change vector of
dV
h
= A
h
dt +B
h
dz (8.2)
Now, let y be the holdings of the assets used to hedge with and let dV
0
be the value change of the hedged
portfolio. This change in value is given by
dV
0
= dV
u
+ y
T
dV
h
= (A
u
+ y
T
A
h
)dt + (B
u
+ y
T
B
h
)dz (8.3)
To eliminate the risk, we need to eliminate the coecients of the factors. Thus, we need to choose y such
that
B
u
+ y
T
B = 0 (8.4)
That is really all there is to hedging!
77
78 CHAPTER 8. HEDGING
8.2.1 Description Using a Tradables Table
I like to use a tradables table description, so lets rework the hedging analysis. Consider a tradables table
description, but add our holdings variable y
Holdings
Prices
Changes
d
=
Factor Model
dt +
dz
(8.5)
Let us assume that the last tradable is the derivative that we would like to hedge and we assume that
we hold one unit of that derivative. Hence, our holdings vector is actually
y =
y
1
(8.6)
and the value change in the hedged portfolio is
dV
0
= y
T
dV = y
T
Adt +y
T
Bdz (8.7)
Thus, to hedge, we select y so that y
T
B = 0 (recalling of course that y is given by (8.6)).
8.2.2 The Relationship Between Hedging and Arbitrage
Hedging is actually quite closely related to arbitrage. In fact, a standard derivation of the Black-Scholes
equation begins from a hedging argument. In this section, lets see how hedging and arbitrage are related.
Recall the arbitrage price implication
y
T
P = 0 No cost
y
T
B = 0 No risk
y
T
A = 0 No return (8.8)
Note that y
T
B = 0 is one of the conditions of the arbitrage price implication. Thus, from the line following
equation (8.7), a hedged portfolio will automatically satisfy this condition.
The other condition that needs to be satised to set up the arbitrage is that the total cost or price must
be zero, y
T
P = 0. Lets see how we can alter our hedged portfolio to satisfy this no cost condition.
Creating a No Cost Hedge
Lets break out the tradables table in a little more detail. First, lets assume that there is a tradable with no
direct factor risk
dB
0
= r
0
B
0
dt (8.9)
Additionally, using the notation above, the asset to be hedged satises
dV
u
= A
u
dt +B
u
dz (8.10)
and the tradables used to hedge this asset are
dV
h
= A
h
dt +B
h
dz (8.11)
We can write this in a tradables table form as
y
0
y
1
B
0
P
h
P
u
B
0
V
h
V
u
r
0
B
0
A
h
A
u
dt +
0
B
h
B
u
dz. (8.12)
where y
0
is the holding of B
0
, y is the holdings of the assets with value changes of dV
h
, and we hold a single
unit of the asset to be hedged.
8.2. HEDGING FROM A FACTOR PERSPECTIVE 79
Now, we make the following important observation. The holding in the rst asset y
0
has no eect on
the hedge! Since it has no factor risk, including it in a portfolio has no eect on the factors! Hence, we can
choose its holding y
0
arbitrarily without aecting the hedge.
This recognition allows us to select y
0
so that the hedge has zero cost. That is, let y be the holdings of
a hedged portfolio, and then select y
0
so that the total cost is zero
y
0
B
0
+ y
T
P
h
+P
u
= 0 (8.13)
Thus, any completely hedged portfolio can be altered to satisfy the price APT implication. Thus, for no
arbitrage to exist, we must have zero prot/loss in this hedged portfolio. This means
y
0
rB
0
+ y
T
A
h
+A
u
= 0 (8.14)
This prot/loss condition is actually the Black-Scholes equation! Thus, a hedging strategy can be used to
derive the Black-Scholes equation via the above relationships.
A Simple Explanation
Here is the simple explanation of what we have done above. We hold an asset that we would like to hedge.
First, we use y to hedge away all the factor risk in our portfolio. Thus, the hedged portfolio has no direct
factor risk. Since it has no factor risk, for no arbitrage to exist it must be the same as the rst tradable
that also has no factor risk. In the Black-Scholes setting, this rst asset is the risk free asset, and all we
are saying is the the hedged portfolio must earn the risk free rate. In an interest rate derivative setting, the
tradable without direct factor risk is the money-market account and we are saying that the hedged portfolio
must earn the short rate. That is really all that we are doing with the above arguments!
8.2.3 Hedging Examples
Lets see how hedging is done in some examples.
Hedging in Black-Scholes
In the Black-Scholes set-up of Chapter 6, Section 6.1.1 we have the following tradables table with the rst
column being the holdings. We have assumed that we have one option c and that we are hedging with the
stock S. Thus, we assume the holding of the stock is y.
0
y
1
B
0
S
c
B
0
S
c
r
0
B
0
S
c
t
+Sc
S
+
1
2
2
S
2
c
SS
dt +
0
S
Sc
S
dz. (8.15)
Now, this portfolio is hedged if y
T
B = 0 which in this case is
yS +Sc
S
= 0 (8.16)
Solving for y gives y = c
S
. Thus, we have a hedged portfolio if we hold c
S
shares of the stock for every
option. The quantity c
S
is typically called the delta of the option.
Pricing
The above portfolio is hedged, but we can convert it to satisfy the arbitrage implication by choosing y
0
so
that the total cost is zero
y
0
B
0
c
S
S +c = 0 (8.17)
This is now a potential arbitrage portfolio. For no arbitrage to be present, we must have no prot/loss which
means that the drift term of our portfolio must be zero
y
0
r
0
B
0
c
s
S + (c
t
+Sc
S
+
1
2
2
S
2
c
SS
) = 0 (8.18)
Substituting in (8.18) for y
0
B
0
= c
S
S c from (8.17) gives
r
0
(c
S
S c) c
s
S + (c
t
+Sc
S
+
1
2
2
S
2
c
SS
) = 0 (8.19)
or
c
t
+r
0
Sc
S
+
1
2
2
S
2
c
SS
= r
0
c (8.20)
which is the Black-Scholes equation.
Once again, the simpler explanation of this derivation is that the hedged portfolio is risk free and hence
must earn the risk free rate. I presented the more structured derivation because I believe that it is always
important to have structure to fall back on when your intuition fails. However, the intuition is important as
well!
Hedging in Bond Pricing
Consider the single factor short rate models of Chapter 7, Section 7.2.1. Following this model, lets write
down the tradable table with holdings:
0
y
1
B
0
B
1
(T)
B
2
(T
0
)
B
0
B
1
(T)
B
2
(T
0
)
r
0
B
0
(B
1
t
(T) +aB
1
r
(T) +
1
2
b
2
B
1
rr
(T))
(B
2
t
(T
0
) +aB
2
r
(T
0
) +
1
2
b
2
B
2
rr
(T
0
))
dt +
0
bB
1
r
(T)
bB
2
r
(T
0
)
dz
(8.21)
Now, this portfolio is hedged if y
T
B = 0 which in this case is
ybB
1
r
(T) +bB
2
r
(T
0
) = 0 (8.22)
Solving for y gives y =
B
2
r
(T0)
B
1
r
(T)
. Note that hedged portfolios of bonds are sometimes called immunized.
Pricing
Using the same procedure as in the Black-Scholes case, we can convert this hedged portfolio into a potential
arbitrage portfolio and derive a pricing equation. Again, we choose y
0
so that the total cost is zero
y
0
B
0
B
2
r
(T
0
)
B
1
r
(T)
B
1
(T) +B
2
(T
0
) = 0 (8.23)
and note that for no arbitrage to exist, we must have no prot/loss
y
0
r
0
B
0
B
2
r
(T
0
)
B
1
r
(T)
(B
1
t
(T) +aB
1
r
(T) +
1
2
b
2
B
1
rr
(T)) + (B
2
t
(T
0
) +aB
2
r
(T
0
) +
1
2
b
2
B
2
rr
(T
0
)) = 0 (8.24)
Using y
0
B
0
=
B
2
r
(T0)
B
1
r
(T)
B
1
(T) B
2
(T
0
) from (8.23) and substituting this into the rst term in (8.24) gives
r
0
B
2
r
(T
0
)
B
1
r
(T)
B
1
(T) B
2
(T
0
)
B
2
r
(T
0
)
B
1
r
(T)
(B
1
t
(T)+aB
1
r
(T)+
1
2
b
2
B
1
rr
(T))+(B
2
t
(T
0
)+aB
2
r
(T
0
)+
1
2
b
2
B
2
rr
(T
0
)) = 0
(8.25)
B
2
t
(T
0
) +aB
2
r
(T
0
) +
1
2
b
2
B
2
rr
(T
0
) r
0
B
2
(T
0
)
B
2
r
(T
0
)
=
B
1
t
(T) +aB
1
r
(T) +
1
2
b
2
B
1
rr
(T) r
0
B
1
(T)
B
1
r
(T)
(8.26)
8.2. HEDGING FROM A FACTOR PERSPECTIVE 81
Now, since the left hand side only depends on B
2
and the right hand side only depends on B
1
and B
1
was
a bond of arbitrary maturity T, we must have that
B
1
t
(T) +aB
1
r
(T) +
1
2
b
2
B
1
rr
(T) r
0
B
1
(T)
B
1
r
(T)
=
(8.27)
for some constant
. To make this look like our previously derived bond pricing equation in Chapter 7
Section 7.2.1, lets let
= b (8.28)
so that (8.27) can be written as
B
1
t
(T) +aB
1
r
(T) +
1
2
b
2
B
1
rr
(T) r
0
B
1
(T)
B
1
r
(T)
= b (8.29)
Upon rearranging we have
B
1
t
(T) + (a b)B
1
r
(T) +
1
2
b
2
B
1
rr
(T) = r
0
B
1
(T) (8.30)
which is the bond pricing equation where is the market price of risk.
The intuitive explanation of this derivation is that rst we hedge out the factor risk by the choice of y.
Since this hedged portfolio has no factor risk, it must earn the same return as the tradable with no factor
risk, which is the money market account. This is where the pricing pde comes from!
8.2.4 Hedging under Incompleteness
In some cases it is impossible to eliminate all of the factor risk (this is true in incomplete markets). In this
case, we may still attempt to reduce the eect of the factors on our portfolio. But, we will have choices in
terms of how we want to best try to reduce risk from the factors.
To see how this is done, lets consider an example. In the jump diusion model of Chapter 6 Section
6.1.6, we had the following tradables table.
0
y
1
B
S
c
B
S
c
r
0
B
( +E[Y 1])S
Lc
dt +
0 0 0
S S 0
Sc
S
0 1
dz
d
1
d
2
(8.31)
where L is the dierential operator corresponding to the drift term in Itos lemma, and
d
1
= (Y 1)d() E[Y 1]dt (8.32)
d
2
=
(c(Y S
) c(S
))d E[(c(Y S
) c(S
))]dt
(8.33)
I have added the vector of holdings on the left side of the tradables table. I have assumed that I hold 1 unit
of the derivative c, and that I will hedge with the stock S. Since the bond B is not driven by any of the
risky factors, I do not need to consider it in the hedge.
So, the hedged portfolio takes the form
dV
h
= ydS +dc (8.34)
= y(( +E[Y 1])Sdt +Sdz +Sd
1
) +Lcdt +Sc
S
dz +d
2
(8.35)
= [y( +E[Y 1])S +Lc]dt + [yS +Sc
S
]dz +ySd
1
+d
2
(8.36)
One can clearly see that it is not possible to eliminate all the factor risk by choosing y. Thus, we are in an
incomplete market.
However, we can select y to eliminate some of the risk over the next dt. In fact, what Merton [11] did in
his pricing formula is equivalent to eliminating the Brownian risk dz. In that case, we would choose
y = c
S
(8.37)
where c
S
is the partial derivative of Mertons pricing formula.
But, lets consider a dierent alternative. Lets choose y to eliminate the variance of the portfolio over
the next dt. That is
V ar(dV
h
)
=
E[A
2
] 2yE[AB] +y
2
E[B
2
]
dt +
V ar(A) 2yCov(A, B) +y
2
V ar(B)
2
dt
2
+ [Sc
S
yS]
2
dt
(8.38)
where
A = c(Y S(t
), t) c(S(t
), t
) (8.39)
B = B(Y 1)S. (8.40)
Minimizing over y gives
y =
E[AB] +
2
dtCov(A, B) +
2
S
2
c
S
E[B
2
] +
2
dtV ar(B) +
2
S
2
(8.41)
and sending dt 0 leads to
y =
E[AB] +
2
S
2
c
S
E[B
2
] +
2
S
2
. (8.42)
8.2.5 A Question of Consistency
Note that the hedges we have derived above often involve knowledge of a pricing formula for the option c.
This can be seen by noting that y is often a function of c
S
for instance. Thus, our hedging analysis has
presupposed that the derivative c follows a specied formula that we know.
This presupposition of knowledge can bring up a question of consistency between the hedge and the
assumption of the formula for c. Consider the following example.
Lets consider a stochastic volatility model:
dB = r
0
Bdt (8.43)
dS = Sdt +
vSdz
1
(8.44)
dv = adt +bdz
2
(8.45)
with the tradables table
B
S
c
B
S
c
r
0
B
S
Lc
dt +
0 0
vS 0
vSc
S
bc
v
dz
1
dz
2
. (8.46)
with Lc = c
t
+Sc
S
+ac
v
+
1
2
vS
2
c
SS
+
1
2
b
2
c
SS
+
vbSc
Sv
.
Now, this is an incomplete market, so we cannot perfectly hedge away the risk. Nevertheless, lets consider
a hedged portfolio
dV
h
= y(Sdt+
vSdz
1
)+(c
t
+Sc
S
+ac
v
+
1
2
vS
2
c
SS
+
1
2
b
2
c
vv
+
vbSc
Sv
)dt+
vSc
S
dz
1
+bc
v
dz
2
(8.47)
Clearly, with y, we cannot eliminate all the risk. Lets consider a simple hedge where we only eliminate the
dz
1
risk. Thus, we choose y
vS +
vSc
S
= 0 or y = c
S
.
Thus, our hedge is to hold c
S
shares of the stock. But, what is c
S
?
Well, how about using the Black-Scholes formula for c(S, t) and computing c
S
from that? But, this is
not consistent because to derive the Black-Scholes formula we assume that volatility is not stochastic. In
our setting above, we are assuming that volatility is stochastic and thus there is no reason to believe that c
would follow the Black-Scholes equation. This is where the inconsistency in hedging often arises.
8.3. HEDGING FROM AN UNDERLYING VARIABLE SENSITIVITY PERSPECTIVE 83
What formula for c?
So, you ask, then what formula should I use for c. The answer is that you should use the formula that best
corresponds to the actual price and movements of c in the market. Thus, if there is stochastic volatility,
then you should use a model that most accurately captures that stochastic volatility and how it is reected
in the movement of the derivative c.
Now, some might argue that you should use a pricing formula that comes from the hedge that you are
doing. As we saw above, pricing and hedging are intimately related. This idea is both right and wrong.
Lets see why.
Here is the correct part. Any pricing formula should be related to a reasonable hedging scheme via
our analysis above. Why? Because hedging is the mechanism that enforces pricing. If there is no hedging
justication for a pricing formula, then there is no good reason to expect that pricing formula to hold.
However, if a pricing formula is tightly related to a hedging strategy, then deviations from that pricing
strategy can be exploited by turning the hedging strategy into an arbitrage (or almost arbitrage) opportunity
as shown in the previous section.
Here is the wrong part. Sometimes, you want to hedge for a reason unrelated to pricing. For example, in
the stochastic volatility case I might just want to hedge out the stock risk S via dz
1
and leave my portfolio
exposed to volatility risk. If I were interested in using some trading strategy that depended on volatility v
but not the stock price S, then this would be perfectly reasonable. But it probably would not be reasonable
to claim that I should be using a pricing formula related to this hedging strategy. The pricing formula c
that I should be using is the one that best reects the actual movement of c in the market.
To summarize, every pricing formula should be intimately connected to a hedging strategy. However,
the reverse is not true. Not every hedging strategy should be turned into a pricing formula. Furthermore,
hedging strategies often rely on a pricing formula for the derivative c. This pricing formula should be the
best pricing formula that reects the actual pricing and movement of the derivative c in the actual market.
In practice, many times that is not done (for many reasons, including computational, ease of use, etc).
Thus, often hedging analysis is inconsistent. For example, people like to rely on the Black-Scholes formula
even when its assumptions are not valid. Due to standard nance conventions, in places below we will fall
into this trap as well, so keep a sharp eye out for where it may be occurring.
8.3 Hedging from an Underlying Variable Sensitivity Perspective
In some cases, there is a simpler and faster way to derive hedges that doesnt involve an explicit use of the
factors. In this approach, we consider the asset that we are trying to hedge to be a function of underlying
variables, and hedge against those variables by eliminating the sensitivity to those underlying variables. In
the simplest terms, sensitivity is measured by the derivative of the hedged portfolio with respect to the
underlying variable of concern.
8.3.1 Black-Scholes Hedging
For example, in the Black-Scholes setup, we assume that an option is a function of the underlying stock S
and time t. That is, the price of an option is a function of the variables S and t which we write as c(S, t).
Now if we would like to hedge out the risk in our option, we note that all the risk in the price of an option
comes from its dependence on the stock variable. Hence, we form a hedged portfolio
P
h
= c(S, t) + yS (8.48)
and want to choose y so that this portfolio is hedged. But, in this case, since all the risk comes from the
stock variable, we will be hedged if our portfolio has no sensitivity to changes in the stock. Another way of
saying this is that we want the derivative of P
h
with respect to S to be zero. That means that small changes
in S will cause no change in the value of the hedged portfolio P
h
. Thus, we will be hedged.
This condition can be written as
P
h
S
= c
S
+ y = 0 (8.49)
Solving for y yields y = c
S
which is the same answer that we arrived at using factors. This hedge is known
as a delta hedge and the quantity c
S
is commonly referred to as the delta of the option.
There is a simple graphical interpretation of this hedge. Consider a plot with the underlying variable
on the x-axis, and the value of the tradable as a function of the underlying factor on the y-axis. In this
case, the underlying variable is the stock price S and the tradables are the stock itself S and the option c.
For simplicity, lets assume the option is a European call option following Black-Scholes. Figure 8.1 shows
a plot of the stock, call option, and delta hedged portfolio for an option with strike K = 10 and expiration
T = 0.3 on a stock with volatility = 0.3 and an interest rate of r
0
= 0.05. The hedged portfolio on the
right in Figure 8.1 is a portfolio of the stock (left) and call option (middle) so that at the current value of
the underlying variable (S = 10), the derivative of the portfolio value yS +c is zero (right plot).
0 5 10 15
10
5
0
5
10
15
Stock Value
Underlying Variable: S
T
r
a
d
a
b
le
: S
0 5 10 15
10
5
0
5
10
15
Option Value
T
r
a
d
a
b
le
: c
0 5 10 15
10
5
0
5
10
15
Hedged Portfolio
T
r
a
d
a
b
le
: H
e
d
g
e
d
P
o
r
tfo
lio
Figure 8.1: Delta Hedge: Left - Plot of the Stock, Middle - Plot of the option, Right - Plot of the hedged
portfolio assuming that the current price of the stock is $10.
To make this a potential arbitrage portfolio, we could add a position in the bond in order to make the
current price of the portfolio (at S = $10) be zero.
8.3.2 Hedging Bonds
We can apply this same underlying variables approach to our hedging of bonds. In that case, we hold a
bond B
2
(r
0
, t|T
0
) that is a function of the short rate r
0
(t) and time t. We wish to hedge this bond with
another bond B
1
(r
0
, t|T) that is also a function of the short rate r
0
(t) and time t. In this setup, the risk
in the price of a bond comes from the short rate r
0
(t) which is our underlying variable. Thus, a portfolio
P
h
= B
2
(r
0
, t|T
0
) + yB
1
(r
0
, t|T) is hedged if its derivative with respect to r
0
is zero. Thus
P
h
r
= B
2
r
(T
0
) + yB
1
r
(T) = 0 (8.50)
Solving for y yields y =
B
2
r
(T0)
B
1
r
(T)
which is the same answer that we arrived at using factors.
8.3.3 Derivatives imply Small Changes
Note that this underlying variable approach uses the derivative as a measure of sensitivity. Recall that the
derivative is the change in a function for a small change in the variable. Thus, when we say that a portfolio
is hedged against stock price movements because its derivative with respect to the stock is zero, this means
that the value of the portfolio change is approximately zero for small changes in the price of the stock. But,
the value of the portfolio change could be quite large if the stock change is large. Hence, this approach only
works if we know that the stock price changes will be small. This is the case for stock price movements
driven by Brownian motion since Brownian motion is continuous. However, if the stock price is driven by a
8.4. HIGHER ORDER APPROXIMATIONS 85
Poisson process, then we would expect it to have large jumps at times. Thus, in this case using the derivative
is not a good approach since the stock price change would be large and the derivative would likely not be
a good approximation to the change in the portfolio value. This was the situation in the Jump-Diusion
model above, and note that the hedging strategy that we derived was not that same one that would result
from this derivative approach.
8.4 Higher Order Approximations
Using the derivative to model the change in a portfolio due to the change in a variable is a linear approxi-
mation of portfolio value as a function of the underlying variable. Of course, one can also use a higher order
approximation to the portfolio value. An easy way to do this is to use the terms of a Taylor expansion. This
leads to the so-called Greeks, presented next.
8.4.1 The Greeks
Now, in general, a call option is not just a function of the price of the stock and time. From the Black-
Scholes formula, we see that it depends also on the risk free rate r
0
and the volatility of the option . In
our derivation of the Black-Scholes formula, we assumed that r
0
and were constant (not random factors).
However, recognizing that that is only an approximation, we can allow them to be underlying variables and
change. Then we can ask how their changes might cause the price of an option to change assuming that the
price follows the Black-Scholes formula. To do this, we would just construct a multivariable Taylor series
expansion of c(S, t, r
0
, ) as
c(S, t, r
0
, ) = c
t
t +c
S
S +c
r
r
0
+c
+
1
2
c
SS
(S)
2
+... (8.51)
I only included a single second order term (there are many second order terms) because is it the only named
second order term.
By looking at this Taylor expansion, we see that the various derivatives tell us the sensitivity of the price
of an option to changes in those variables. Each of these derivatives is given a Greek letter name as in Table
8.1.
theta c
t
delta c
S
rho c
r
vega c
gamma c
SS
Table 8.1: The Greeks
If you have a bad memory, like I do, then you can remember some of the Greeks by noting that the rst
letter of the Greek is the same as the partial derivative. That is, theta starts with t and is the partial
with respect to t. Similarly, rho starts with r and is with respect to r
0
, and vega starts with v and
is the partial with respect to volatility. Unfortunately, with delta and gamma you are on your own.
Note that in the typical dynamic hedging assumption, we are able to trade continuously. Thus, only
terms of order dt and lower matter. However, this Taylor expansion approach assumes that we are not
trading continuously. Thus, we use a t instead of a dt and furthermore, higher order terms will enter. This
is actually more practical than the typical continuous time trading assumptions.
8.4.2 A Delta-Gamma Hedge
One can think of a delta hedge as eliminate the rst order term in the Taylor expansion. Of course, one can
go further and ask if higher order terms can be eliminated as well. If you elminate the rst order term in
S and the second order term (S)
2
than this is called a Delta-Gamma hedge. For such a Delta-Gamma
hedge you need more than just the underlying stock, and furthermore, you need a tradable that depends on
the second order term (S)
2
. Lets show how by using another call option c
(2)
and the underlying stock,
we can Delta-Gamma hedge an option c.
In this case, the hedged portfolio is
P = c +y
1
S +y
2
c
(2)
(8.52)
and lets Taylor expand this to obtain
P(S, t) = c(S, t) +y
1
S +y
2
c
(2)
(8.53)
=
c
t
t +c
S
S +
1
2
c
SS
(S)
2
+y
1
S +y
2
c
(2)
t
t +c
(2)
S
S +
1
2
c
(2)
SS
(S)
2
(8.54)
=
c
t
+y
2
c
(2)
t
t +
c
S
+y
1
+y
2
c
(2)
S
S +
1
2
c
SS
+y
2
1
2
c
(2)
SS
(S)
2
(8.55)
To Delta-Gamma hedge, we have to eliminate the coecients of S and (S)
2
by choosing y
1
and y
2
. Thus,
we must solve
c
S
+y
1
+y
2
c
(2)
S
= 0 (8.56)
1
2
c
SS
+y
2
1
2
c
(2)
SS
= 0 (8.57)
The solution is
y
1
=
c
SS
c
(2)
SS
c
(2)
S
c
S
, y
2
=
c
SS
c
(2)
SS
(8.58)
In the plots of Figure 8.1, this would create a portfolio that has the rst and second derivative equal to zero
at the current value of the underlying variable.
In practice, Delta-Gamma hedges (and other hedges as well) are dicult because transaction costs can
make it expensive to trade too many assets.
8.4.3 Determining what the error looks like
We can use the Taylor expansion to see what the error looks like in a Delta hedge under Black-Scholes
assumptions. Consider the hedged portfolio
P(S, t) = c(S, t) c
S
S (8.59)
and lets perform a Taylor expansion of this in S and t
P(S, t) = c(S, t) c
S
S (8.60)
= c
t
t +
1
2
c
SS
(S)
2
dt +... (8.61)
= c
t
t +
1
2
c
SS
2
S
2
dt +... (8.62)
where I have only kept terms up to order t in (8.62). However, (8.61) is also a very useful equation for
intuituion. So we will use it as well. Note that the only dierence between (8.61) and (8.62) is that (S)
2
is
2
S
2
dt + higher order terms. For the analysis that we are looking at, the higher order terms dont matter.
Now lets assume that in the market, c satises the Black-Scholes equation corresponding to a volatility
value of
i
. Thus,
c
t
+r
0
Sc
S
+
1
2
2
i
S
2
c
SS
= r
0
c (8.63)
I use the subscript i on the volatility
i
to denote what is known as implied volatility. Implied volatility
is the value of volatility that when plugged into the Black-Scholes formula will make it equal the current
8.5. SUMMARY 87
market price of an option. To be more concrete, let c
m
be the current market price of an option where
the stock price is S
0
, strike price is K, time to expiration is T, and risk free rate is r
0
. Furthermore, let
c
BS
(S, K, T, r
0
, ) denote the Black-Scholes formula. Then the implied volatility is dened as the value of
i
that solves
c
m
= c
BS
(S
0
, K, T, r
0
,
i
) (8.64)
Thus, it answers the question: If the market is following Black-Scholes, what volatility value are they pluggin
in the Black-Scholes formula?
Thus, (8.63) assumes that the market is pricing the option using the Black-Scholes formula with a
volatility value of
i
. Thus, we can use (8.63) to substitute for c
t
in (8.61) which gives
P(S, t) = r
0
(c c
S
S)dt +
1
2
c
SS
(S)
2
2
i
S
2
dt (8.65)
and nally noting that P = c c
S
S gives
P(S, t) = r
0
Pdt +
1
2
c
SS
(S)
2
2
i
S
2
dt (8.66)
or purely to order dt using (8.62) instead of (8.61),
P(S, t) = r
0
Pdt +
1
2
c
SS
2
i
S
2
dt. (8.67)
Equation (8.66) shows that the gain or loss of our hedged portfolio relative to the risk free rate depends on
the actual change in the stock price over the next t in the term (S)
2
relative to the implied volatility
i
in Black-Scholes that is being used to price the option in the market. Equation (8.67) shows the same
thing, but in terms of the volatility of the stock rather than the stock move S. In particular, for this
hedge where we are long the option and short delta of the stock, we make money if the stock moves more
than implied volatility estimate, and we lose money if it moves less.
8.5 Summary
Hedging can be approach from two dierent points of view. In the rst point of view, we recognize the fact
that risk comes from the factors. Thus, in hedging we try to eliminate the factor risk. This point of view
is appropriate regardless of what the risky factors are. In the second point of view, tradables are viewed as
functions of underlying variables, and we hedge by constructing a portfolio that eliminates the sensitivity to
moves in the underlying variables. Usually we do this by setting the derivative of the hedged portfolio with
respect to the underlying variable to zero at the current value of the underlying variable. Since we only set
the derivative to zero, and the derivative reects a local approximation to the portfolio, this works as long
as there are only small moves in the underlying variable between rehedging opportunities. Furthermore, if
we have more tradables to place in our portfolio, we can set higher order derivatives of our portfolio to zero
as well and create better hedges. Really, if you underlying Taylor expansions, then you are on your way to
understanding the majority of hedging methods.
8.6 Problems
Problem 8.6.1 Verify equation (8.38).
Chapter 9
The Road to Risk Neutrality
9.1 Introduction
Risk neutral absence of arbitrage pricing is a widely used approach to derivative pricing. However, it tends
to be one of the most confusing, misunderstood, and misused pricing approaches, especially for those new
to derivative pricing. Nevertheless, when understood correctly, it is an extremely powerful approach. Thus,
in this chapter, I will explain the risk neutral pricing principle. This introduction to risk neutral pricing
does not follow the standard probability-heavy route, but instead motivates risk neutral pricing in a simple
manner from the factor approach.
9.2 Do the Factors Matter?
In our tradables table, we have
Prices
Changes
d
=
Factor Model
dt +
dz
(9.1)
and the absence of arbitrage condition is
A = P
0
+B (9.2)
Thus, absence of arbitrage only depends on the values of P, A, and B, and not on what the exact factors
dz are. Therefore, the prices P are absence of arbitrage for any value change with A, and B, regardless of
what the driving factors are. This leads us to ponder the following.
Hypothesis: Perhaps by changing the factors from dz to some other random factors d, where the A and
B representation of prices changes is the same, I can compute the absence of arbitrage prices P easier.
To see whether something can be made of this hypothesis, we have to start with a more basic question.
Question: Given a set of factors dz, what other set of factors d will have the same A and B representation?
At rst glance, you might be tempted to say that I can choose d to be anything I want. It doesnt have
to relate to the original dz at all because the absence of arbitrage condition does not see dz. However,
you would soon realize that this is not the case.
Why? Because when we deal with derivative securities, their factor model is determined by the application
of Itos lemma. Thus, the A and B representation of price changes does see the factors via Itos lemma.
Hence, Itos lemma puts a constraint on which factors d are consistent with the original factors dz. Lets
be a little more concrete about this.
89
90 CHAPTER 9. THE ROAD TO RISK NEUTRALITY
Let S(t) be a stock price with model
dS = Sdt +Sdz (9.3)
and let c(S, t) be a derivative security. By Itos lemma we have
dc =
c
t
+Sc
S
+
1
2
2
S
2
c
SS
dt +Sc
S
dz. (9.4)
Thus, the tradables table is
Prices
S
c
Changes
d
S
c
=
Factor Model
c
t
+Sc
S
+
1
2
2
S
2
c
SS
dt +
S
Sc
S
dz.
(9.5)
9.2.1 Brownian Factors
Changing the Mean Works
Lets try new factors given by the replacement
dz d = d z +dt (9.6)
That is, I will replace the original Brownian factors dz by a new factor that is a Brownian d z plus a drift
dt. It is important to consider the new factor to be the entire term d = d z +dt and not just d z!
We can ask whether this will lead to a representation consistent with the original dz. We would have
dS = Sdt +S(d) (9.7)
= Sdt +S(d z +dt) (9.8)
= ( +S)dt +Sd z (9.9)
and then by Itos lemma we have for c(S, t),
dc =
c
t
+ ( +)Sc
S
+
1
2
2
S
2
c
SS
dt +Sc
S
d z (9.10)
=
c
t
+Sc
S
+
1
2
2
S
2
c
SS
dt +Sc
S
(d z +dt) (9.11)
=
c
t
+Sc
S
+
1
2
2
S
2
c
SS
dt +Sc
S
(d) (9.12)
And thus we see that dz and d = d z + dt are consistent in that they produce the same A and B
representation regardless of whether dz or d = d z +dt is the factor!
The upshot is that we can replace any Brownian factors dz by Brownians plus a drift d z +dt, and the
same absence of arbitrage prices hold!
Changing the Variance Does Not Work
Lets try new factors given by the replacement
dz d = d z (9.13)
That is, I will replace the original Brownian factors by Brownians with a dierent variance. Again, the new
factor should be considered to be the entire term d = d z and not just the d z. We can ask whether this
factor d = d z is consistent with the original dz. We would have
dS = Sdt +S (d) (9.14)
= dt +S(d z) (9.15)
9.2. DO THE FACTORS MATTER? 91
and then by Itos lemma we have for c(S, t),
dc =
c
t
+Sc
S
+
1
2
2
S
2
c
SS
dt +Sc
S
(d z) (9.16)
=
c
t
+Sc
S
+
1
2
2
S
2
c
SS
dt +Sc
S
(d) (9.17)
(9.18)
Thus, we see that A is changed when we change the factor to another Brownian with a dierent variance.
Therefore, this substitution is not allowed.
Similarly, in the case of multiple Brownian factors, changes in the correlations or the covariance structure
is not allowed. The above results lead us to the following principle.
() Arbitrage Invariance Principle for Brownian Motion: If a set of prices P is absence of arbitrage
under Brownian factors dz R
n
with E[dzdz
T
] = dt, then P is also absence of arbitrage if the factors dz are
replaced by dz d = d z +dt with arbitrary and where d z are Brownian factors with E[d zd z
T
] = dt.
9.2.2 Poisson Factors
Changing the Intensity Works
Assume that the original factor is a Poisson Process d(t; ). Under d(t; ) we have
dS = Sdt + ( 1)Sd(t; ) (9.19)
dc = (c
t
+Sc
S
)dt + (c(S) c(S))d(t; ). (9.20)
Now, lets consider changing the factor to a Poisson with an altered intensity d = d (t; + ) where
+ > 0. In this case, we have
dS = Sdt + ( 1)Sd (t; +) (9.21)
dc = (c
t
+Sc
S
)dt + (c(S) c(S))d (t; +) (9.22)
and we see that the A and B representations remain the same under d = d (t; +).
Adding a Drift Doesnt Work
Again assume that the original factor is a Poisson Process d(t; ). Under d(t; ) we have
dS = Sdt + ( 1)Sd(t; ) (9.23)
dc = (c
t
+Sc
S
)dt + (c(S) c(S))d(t; ) (9.24)
Now, lets consider changing the factor d(t; ) d = d (t; ) +dt. In this case, we have
dS = Sdt + ( 1)Sd = Sdt + ( 1)S(d (t; ) +dt) (9.25)
dc = (c
t
+ (S + ( 1)S)c
S
)dt + (c(S) c(S))d (t; ) (9.26)
and we see that there is no way to recover the original A and B representations under d.
() Arbitrage Invariance Principle for Poisson: If a set of prices P is absence of arbitrage under
Poisson factors d
i
(
i
), then P is also absence of arbitrage under a dierent set of factors d = d
i
(
i
+
i
)
where
i
+
i
is the new intensity.
9.3 Risk Neutral Representations
The Brownian and Poisson factor invariance principles tell us that it is okay to alter or replace the factors in
certain ways. Why is this helpful? Because in some situations it is easier to price a derivative if we replace
the factors by something dierent from the original factors. In fact, in this section we show that we can
always replace the factors and put all tradables in what we will call a risk neutral representation. The fact
that we can do this will ultimately lead us to the risk neutral pricing principle. The basic idea is that for any
pricing problem, we can replace the factors to create the risk neutral representation which leads to simplied
pricing formulas. But we are getting ahead of ourselves. First, lets see what the risk neutral representations
are.
9.3.1 Brownian Factors
Consider the tradables table with Brownian factors dz
Prices
Changes
d
=
Factor Model
dt +
dz
(9.27)
and the absence of arbitrage condition is A = P
0
+ B. Now, we can substitute this into the tradables
table to obtain
Prices
Changes
d
=
Factor Model
P
0
+B
dt +
dz
(9.28)
Finally, we groups terms dierently to obtain
Prices
Changes
d
=
Factor Model
P
0
dt +
(dz +dt)
(9.29)
Now, by the Brownian factor invariance under changes to the drift, we can replace dz d = d z dt,
which leads to
Prices
Changes
d
=
Factor Model
P
0
dt +
d z
(9.30)
This is the risk neutral representation. It states that we can nd a replacement set of factors so that all
tradables have a drift equal to the market price of time
0
, and absence of arbitrage prices under this
representation will be the same as under the original representation using the actual Brownian factors.
Note that to obtain this risk neutral representation, the factors were replaced by new factors that con-
tained the market prices of risk! Lets formalize this notion of a risk neutral representation.
() Risk Neutral Representation for Brownians: Let
Prices
Changes
d
=
Factor Model
dt +
dz
(9.31)
be a tradables table. Then, under arbitrage invariant substitutions of the factors dz d = d z dt, the
following tradable table will produce the same absence of arbitrage prices
Prices
Changes
d
=
Factor Model
P
0
dt +
d z
(9.32)
Equation (9.32) is called the risk neutral representation because all tradables have a drift equal to the market
price of time, regardless of how risky they really are.
9.3. RISK NEUTRAL REPRESENTATIONS 93
9.3.2 Poisson Factors
Consider the tradables table with Poisson factors d(t, ).
Prices
Changes
d
=
Factor Model
dt +
d(t; )
(9.33)
and the absence of arbitrage condition is
A = P
0
+B (9.34)
Before proceeding, lets think about the market price of risk for a Poisson factor d. Since a Poisson process
either does nothing or jumps up by 1, it is always good to hold a positive amount of a Poisson factor. All
the risk is on the upside. On the other hand, being short a Poisson factor is adding real (downside) risk.
Thus, we would expect the market price of risk for a Poisson factor to be negative. That is, if you are short
a Poisson factor (B is negative), then you should be rewarded for taking on that risk, and a negative market
price of risk would reect that.
One can show this in a much more rigorous fashion, but for our purposes, lets just note that < 0 for a
Poisson factor. I dont like dealing with a negative quantity, so lets dene
= and rewrite the absence

of arbitrage condition in terms of
as
A = P
0
+B (9.35)
= P
0
+B(
) (9.36)
= P
0
B
(9.37)
where
> 0.
Now, we can substitute this into the tradables table to obtain
Prices
Changes
d
=
Factor Model
P
0
B

dt +
d(t; )
(9.38)
Now, by the Poisson factor invariance under changes to the intensity, we can replace d(t; ) d =
d (t; + ). Lets choose =
so that d = d (t;
) (here is where it is important that
> 0, so
that it can be an intensity!), which leads to
Prices
Changes
d
=
Factor Model
P
0
B

dt +
d (t;
)
(9.39)
The nal step is to compensate the Poisson process. That is, we subtract o the mean of the Poisson process
so that the random factor has mean zero.
Prices
Changes
d
=
Factor Model
P
0
dt +
(d(t;
dt)
(9.40)
Since the random factor term now has zero mean, we can see that the drift of the value changes V for all
tradables is equal to the market price of time
0
. This was the same as in Brownian case.
Thus, for Poisson processes, we can replace the original Poisson factors by new Poisson factors with
dierent intensities so that the drift of all tradables is the market price of time
0
(dont forget to subtract
o the mean of the Poisson factors so that they have zero mean!). This is the risk neutral representation
under Poisson factors!
Note that in this case, the new intensity of the factors in the risk neutral representation is equal to
(minus) the market price of risk! Lets make this formal.
() Risk Neutral Representation for Poissons: Let
Prices
Changes
d
=
Factor Model
dt +
d(t; )
(9.41)
be a tradables table. Then, under arbitrage invariant substitutions of the factors d(t; ) d = d (t;
),
the following tradable table will produce the same absence of arbitrage prices
Prices
Changes
d
=
Factor Model
P
0
dt +
(d (t;
dt) .
(9.42)
Equation (9.42) is called the risk neutral representation because all tradables have a drift equal to the market
price of time, regardless of how risky they really are.
9.4 Pricing as an Expectation
The risk neutral representations lead to a powerful risk neutral pricing formula. This is because instead of
starting with the real tradables table, we can start with the risk neutral representation. The fact that the
drift of all tradables is equal to the market price of time leads us to a convenient new pricing formula. Lets
see how it works for the Brownian case.
From the risk neutral representaion equation (9.32) we have
dV =
0
Pdt +Bd z. (9.43)
Lets assume that we are not dealing with a futures contract, so that V = P. Thus, we have
dP =
0
Pdt +Bd z (9.44)
which is telling us that the drift of P is
0
when we use the factors based on d z.
Now, via Itos lemma one can verify that
d
t
0
0(s)ds
P
= e
t
0
0(s)ds
Bd z. (9.45)
Taking expectations of both sides gives
d

E
t
0
0(s)ds
P
= 0 (9.46)
since z is a Brownian motion. (We also switched the d and the expectation.) Finally, integration of both
sides says that
P(0) =

E
t
0
0(s)ds
P(t)
. (9.47)
We have arrived at risk neutral pricing.
() The Risk Neutral Pricing Principle: Absence of arbitrage prices are given by the formula
P(0) =

E
t
0
0(s)ds
P(t)
(9.48)
where the expectation

E() is taken under the risk neutral representation and
0
is the market price of time.
This risk neutral pricing principle applies in the Poisson case as well. Thus, this is a new pricing point
of view that follows from the factor approach! The presentation has been a little abstract to this point, so
lets see how it would work in practice.
9.5. APPLICATIONS OF RISK NEUTRAL PRICING 95
9.5 Applications of Risk Neutral Pricing
Lets see how the risk neutral pricing principle is used in a couple of familiar situations. But rst, lets
outline how it is applied.
9.5.1 How to Apply Risk Neutral Pricing
Risk neutral pricing is applied using the following steps.
1. Via an arbitrage invariant substitution of the factors, convert the tradables table to its risk neutral
representation.
2. Apply the Risk Neutral Pricing Formula to the derivative security that is being priced.
That is it! Lets clarify with some examples.
9.5.2 Black-Scholes
Lets see how risk neutral pricing applies to the Black-Scholes setup. Recall that the bond and stock follow
dB = r
0
Bdt (9.49)
dS = Sdt +Sdz (9.50)
dc = (c
t
+Sc
S
+
1
2
2
S
2
c
SS
)dt +Sc
S
dz. (9.51)
Now, according to the risk neutral pricing principle, we have the same absence of arbitrage prices if we set
all the drifts to the risk free rate (market price of time). Hence, we have
dB = r
0
Bdt (9.52)
dS = r
0
Sdt +Sd z (9.53)
dc = r
0
cdt +Sc
S
d z. (9.54)
Now, applying the risk neutral pricing formula to the call option c(S, t) gives
c(S(0), 0) =

E
e
r0T
c(S(T), T)
. (9.55)
But, if T is expiration, then we know that c(S(T), T) = (S(T) K)
+
, so the risk neutral pricing formula is
c(S(0), 0) = e
r0T

E
(S(T) K)
+
(9.56)
where the e
r0T
was pulled out of the expectation because it is not random.
The expectation

E() is taken under the risk neutral representation (i.e. S(t) follows (9.53), etc.). Per-
forming this expectation leads to
d
1
=
ln(S/K) + (r
0
+
1
2
2
)(T)
T
(9.57)
d
2
= d
1
T (9.58)
c(S, 0) = SN(d
1
) Ke
r0T
N(d
2
) (9.59)
which is the Black-Scholes formula! Thus, we were able to obtain a pricing formula without resorting to
partial dierential equations!
9.5.3 Poisson Model
Lets try it for a Poisson model of Section 6.1.4. To be consistent with the approach above in Section 9.3.2,
I wont compensate the original Poisson factor d(t; ) as was done in Section 6.1.4. The tradables table is
B
S
c
B
S
c
r
0
B
S
c
t
+Sc
S
dt +
0
(k 1)S
c(kS
) c(S
d()
and the market price of time is
0
= r
0
while the market price of risk is
1
=
r
0
k 1
. (9.60)
For the risk neutral representation we use minus the market price of risk
1
=
r
0
k 1
(9.61)
and the risk neutral representation is
B
S
c
B
S
c
r
0
B
r
0
S
r
0
c
dt +
0
(k 1)S
c(kS
) c(S
(d(
1
)
1
dt) . (9.62)
Finally, we can apply the risk neutral pricing formula
c(S(0), 0) =

E
e
r0T
c(S(T), T)
. (9.63)
But, if T is expiration, then we know that c(S(T), T) = (S(T) K)
+
, so the risk neutral pricing formula is
c(S(0), 0) = e
r0T

E
(S(T) K)
+
. (9.64)
where the e
r0T
was pulled out of the expectation because it is not random. In the risk neutral representation,
(9.62) determines the expectation

E(). Thus, computing the expectation in (9.64) gives
c(S, t) = S(x, y) Ke
r0(T)
(x, y/k) (9.65)
where
(, ) =
i=
e
i
i!
, y =
(r
0
)kT
k 1
(9.66)
and x is the smallest non-negative integer greater than
ln(K/S)(T)
ln(k)
.
9.5.4 HJM
Recall the HJM model of Section 7.2.3. In this model, the underlying variables are given by the instantaneous
forward rates
dr(t|s) = (t|s)dt +(t|s)dz(t). (9.67)
and the tradables are bonds that follow
dB(t|T) =
B(t|T)r(t|t) B(t|T)
T
t
(t|s)ds +
1
2
B(t|T)
T
t
T
t
(t|s)(t|r)drds
dt
B(t|T)
T
t
(t|s)ds
dz
9.5. APPLICATIONS OF RISK NEUTRAL PRICING 97
If you recall from Section 7.2.3 this was extremely messy to deal with. So, lets start over but with the risk
neutral perspective in mind.
Lets start with the tradables that are the bonds. Since this is a single factor model, the risk neutral
representation under Brownians tells us that we may write the tradables in the form
dB(t|T) = r(t|t)B(t|T)dt +(t|T)B(t|T)d z (9.68)
where d z is the risk neutral factor and r(t|t) = r
0
(t) is the instantaneous short rate. Now, we can ask what
this implies about the instantaneous forward rates in the risk neutral world. Well, the relationship between
the bonds B(t|T) and the instantaneous forward rates is
r(t|T) =

T
ln B(t|T). (9.69)
Then, by Itos lemma, we have
dr(t|T) =

T
1
2
2
(t|T) r(t|t)
dt

T
(t|T)d z (9.70)
=

T
1
2
2
(t|T)
dt

T
(t|T)d z. (9.71)
If one lets
(t|T) =

T
(t|T) (9.72)
then
(t|T) =

T
1
2
2
(t|T)
= (t|T)

T
(t|T) = (t|T)
T
t
(t|s)ds (9.73)
where
dr(t|T) = (t|T)dt +(t|T)dz(t). (9.74)
These equations tell us that under the risk neutral representation, if we know the volatility of the instanta-
neous forward rates ((t|T)), then we can compute what the drift terms must be by equation (9.73).
Another way to get to this same result is to revisit the results of Section 7.2.3 equation (7.22) that said:
(t|T) (t|T)
T
t
(t|s)ds = (t|T)
1
(9.75)
and note that in the risk neutral representation, we have a zero market price of risk
1
= 0 (because all
tradables earn the risk free rate). Thus, the above equation becomes
(t|T) = (t|T)
T
t
(t|s)ds (9.76)
which is what we were looking for.
Thus, if we want to use the risk neutral representation, we would rst estimate the volatilities of the
instantaneous forward rates (t|T) from market data, then use (9.73) to compute the risk neutral drifts.
Once we have this, pricing proceeds via expectations as in the risk neutral pricing formula. In the
HJM model, the expectation is often computed by Monte Carlo. That is, one simulates the risk neutral
representation of the instantaneous forward rates. Then, one can compute the payo value of a derivative,
and compute the expectation of it. This is the risk neutral priceing approach.
9.5.5 Libor Market Model
The LMM model of Section 7.2.4 is also made quite simple by the use of risk neutrality. Recall that the
LMM is similar to the HJM model, but deals with forward rates between times T
i
and T
i+1
denoted by
R(t|T
i
, T
i+1
) and that are assumed to follow
dR(t|T
1
; T
2
) = a
1
R(t|T
1
; T
2
)dt +b
1
R(t|T
1
; T
2
)dz
2
. (9.77)
For notational simplicity, we write R
i
= R(t|T
i
, T
i+1
) with
dR
i
= a
i
R
i
dt +b
i
R
i
dz
i
. (9.78)
Recall from equation (7.46) that the calibration relationship on the underlying R
i
variables is
i1
b
i1
R
i1
(1 +R
i1
)
a
i
R
i
b
i
R
i
i
j=1
bjRj
(1+Rj)
ij
1 +R
i
. (9.79)
Now, we can switch to the risk neutral representation by setting all the market prices of risk to zero
i1
= 0.
This reduces the above equation to
0 =
a
i
R
i
b
i
R
i
i
j=1
bjRj
(1+Rj)
ij
1 +R
i
(9.80)
or by simplifying further
a
i
R
i
b
i
R
i
i
j=1
b
j
R
j
(1 +R
j
)
ij
= 0 (9.81)
or nally
a
i
= b
i
i
j=1
b
j
R
j
(1 +R
j
)
ij
(9.82)
which is quite a bit simplier than what we started with.
Note that this is similar to what we had in the HJM risk neutral model. We can use market data to
estimate the b
i
terms (the volatility of the forward rates), and then use equation (9.82) to obtain their drifts
in the risk neutral representation. Pricing is then done by expectation.
9.6 Summary
The point of this chapter was to show that risk neutral pricing is a logical consequence of the factor approach
to derivative pricing. The idea was that details of the factors do not seem to appear in the factor APT
equations that we used throughout the book. That gave us the idea that perhaps we could change the
factors (as long as we didnt disturb the basic factor coecient structure) and still arive at the same absense
of arbitrage price. By using a dierent set of factors (that still preserved the same absense of arbitrage
prices) in many case we can simplify our calculations of the absense of arbitrage prices. This is the basic
notion of risk neutral pricing. In fact, there is quite a bit more that one can do when the full power and
generality of the risk neutral approach is explored. But that, my friends, is the subject of another little
book...
9.7 Problems
Problem 9.7.1 Verify equation (9.45).
Problem 9.7.2 Verify equatin (9.70).
Bibliography
[1] F. Black. The pricing of commodity contracts. Journal of Financial Economics, 3:167179, 1976.
[2] F. Black and M. Scholes. The pricing of options and corporate liabilities. Journal of Political Economy,
81:637659, 1973.
[3] A. Brace, D. Gatarek, and M. Musiela. The market model of interest rate dynamics. Mathematical
Finance, 7:127154, 1997.
[4] J.C. Cox, J. Ingersoll, and S.A. Ross. A theory of the term structure of inteest rates. Econometrica,
53:385467, 1985.
[5] J.C. Cox and S.A. Ross. The valuation of options for alternative stochastic processes. Journal of
Financial Economics, 3:145166, 1976.
[6] R. Jarrow D. Heath and A. Morton. Bond Pricing and the Term Structure of Interest Rates: A New
Methodology. Econometrica, 60(1):77105, 1992.
[7] J. Franklin. Methods of Mathematical Economics. Springer-Verlag, New York, 1980.
[8] Daniel T. Gillespie. Markov Processes: an introduction for physical scientists. Academic Press, 1997.
[9] S. L. Heston. A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bond
and Currency Options. The Review of Financial Studies, 6(2):327343, 1993.
[10] W. Margrabe. The value of an option to exchange one asset for another. Journal of Finance, 33(1):177
186, 1978.
[11] Robert C. Merton. Option Pricing when the Underlying Stock Returns are Discontinuous. Journal of
Financial Economics, 5:125144, 1976.
[12] L.C.G. Rogers and D. Williams. Diusions, Markov Processes and Martingales: Volume 2, Ito Calculus.
Cambridge, 2000.
[13] S. A. Ross. The Arbitrage Theory of Capital Asset Pricing. Journal of Economic Theory, 59:341360,
1976.
[14] G.E. Uhlenbeck and L.S. Ornstein. On the theory of brownian motion. Phys. Rev., 36:823841, 1930.
[15] O. Vasicek. An Equilibrium Characterization of the Term Structure. Journal of Financial Economics,
5:177188, 1977.
99
Index
APT, 32
Price Form, 35
Return Form, 32
Arbitrage
Price Implication, 35
Return Implication, 31
Black-Scholes, 49, 95
Formula, 95
Hedging, 80
Brownian Motion, 1
Increment, 3
Calibration, 47
Compound Poisson Process, 7
Cox-Ingersoll-Ross, 25
Delta, 79, 84
Delta Hedge, 84
Delta-Gamma Hedge, 85
Derivative, 42
Denition, 42
Dividends, 50
Factor Models
Via Itos Lemma, 43
Factors, 41
Futures
Derivative Pricing, 54
Gaussian Random Variable, 1
geometric Brownian Motion
Black-Scholes, 49
Greeks, 85
delta, 79, 85
gamma, 85
rho, 85
Taylor Expansion, 85
theta, 85
vega, 85
Heath-Jarrow-Morton, 66
Hedging, 77
Black-Scholes, 80
Delta, 84
Delta-Gamma, 85
Immunization, 80
Incomplete, 81
Immunization, 80
Implied Volatility, 86
Incomplete, 46, 81
Hedging, 81
intensity, 3
Itos Lemma
Obtaining Factor Models, 43
Jump Diusion
Merton, 55
Libor-Market-Model, 69
Market
Incomplete, 46
Market Price of Risk, 33
Market Price of Time, 33
Marketed Tradables, 44
Merton, 55
Money Market Account, 63, 64
Dynamics, 64
Normal Random Variable, 1
Null Space, 32
Relation to Range Space, 32
Option
European Call, 43
Ornstein-Uhlenbeck, 24, 25
Perpendicular Space, 32
Poisson Process, 1, 3
Compound, 7
intensity, 3
Poisson Processes
Poisson Random Variable, 3
Price APT
100
INDEX 101
Application to Pricing, 44
Range Space, 32
Relation to Perp of Null Space, 32
Relative Pricing, 44
Risk Neutral, 89
Pricing, 95
Principle, 94
Risk Neutral Pricing
Risk Neutral Representation, 92, 94
Risk Neutral Representation, 92, 94
Short Rate, 64
Single Factor Models, 64
Vasicek, 64
Stochastic Process
Compound Poisson Process, 7
Tradables, 42
Marketed, 44
Tradables Table, 44
Underlying Variables, 41
Vasicek, 24, 64
Volatility
Implied, 86

FEbook 1 MAIN

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

FEbook 1 MAIN

Uploaded by

Copyright:

Available Formats

THE FACTOR APPROACH TO DERIVATIVE PRICING

The BIGPicture in a LITTLE Book

- notation for limit from the left: x(t).

t where N(0, 1) (1.7)

t. Thus, we will often use E[z] = 0 and

when we want to suppress the argument t, or even a

when suppressing the arguments of

dt. Hence, we think of dz as being of order dt

(z((i + 1)t) z(it))

dt, t +dt) +f(x

dt, t +dt) f(x

dt, t +dt) (2.26)

dt, t +dt). (2.27)

dt term in the argument is of order dt, then overall, we can

dt term as overall being of order dt

dt, t +dt) f(x

dt, t +dt) (f(x

dt, second order terms in dz are of

xdz, use Itos lemma to derive df where f(x, t) = x

, use Itos lemma to derive df for f(x, t) = t

is a set of all vectors z such that z is orthogonal to all vectors in M. That is

then it also must be orthogonal to . Another way to put this is that

is this optimal , then we can dene

, t))(t ))dt +Sc

))(d() dt) (6.48)

((Y 1)d() E[Y 1]dt) (6.66)

))]dt) and (Y 1)d()

(N) (B(N + 1) B(N))

(N) (B(N + 1) B(N))

= and rewrite the absence

) (here is where it is important that

You might also like