Financial Engineering∗

Craig Pirrong
Spring, 2008

January 7, 2013

∗
c CraigPirrong, 2008. Do not reproduce or distribute
without express written permission of copyright holder.

1
 Financial Engineering

• Since the publication of the Black-Scholes-

Merton model in 1973 there has been a
revolution in financial markets.

• This is the “Financial Engineering” Revo-

lution.

• Financial Engineering has been defined as

“the diagnosis, analysis, design, produc-
tion, pricing, and customization of solu-
tions” to corporate financial problems.

• Most notably, financial engineering involves

the creation of new derivative securities,
futures, forwards, swaps, and options of
various types.

2
 The Goal of This Course

• Financial engineering is a broad and diverse

subject. This course will focus on valuation
issues. That is, it will focus on how to
value derivative securities of all types.

• Accurate valuation requires you to under-

stand and use three important, related (but
distinct) disciplines.

• Stochastic calculus.

• Numerical analysis.

• Statistics and Econometrics.

3
Understanding the Strengths and Weaknesses
of Received Analytical Techniques

• The tools pioneered by Black-Scholes-Merton

are incredibly powerful, but they are not
perfect.

• Some very smart people–including Scholes

and Merton–have lost immense sums–billions
and billions of dollars–by putting too much
faith in these models.

• A good practitioner must know how the

models work, their strengths, and their weak-
nesses.

4
 Strengths and Weaknesses

• The strength of existing valuation method-

ology are its rigor and flexibility.

• The weakness is that these methods utilize

mathematical tools that make assumptions
about the behavior of financial prices that
are inconsistent with their real world be-
havior.

• “When your only tool is a hammer, every-

thing looks like a nail.”

• “The drunk looked under the streetlight for

his car keys because the light was better
there.”

• This course will teach you how to use a

hammer, but to recognize when you’re not
driving a nail.

5
 Stochastic Calculus

• Stochastic calculus is the fundamental tool

in financial engineering because the focus
of our interest is on random financial prices
such as stock or commodity prices.

• Stochastic calculus allows us to determine

how functions of random variables behave.

• Stochastic calculus works in continuous time.

It is usually easier to derive results in con-
tinuous time even though we have to dis-
cretize time in order to find solutions to
the equations that result from these deriva-
tions.

6
 Brownian Motion

• Brownian motion is the workhorse of stochas-

tic calculus. It is the way that randomness
is represented.

• Due to the nice properties of Brownian mo-

tion, it is the “hammer” that financial en-
gineers apply to virtually every problem

• Brownian motion is a mathematical rep-

resentation of a continuous time random
walk.

7
• Some important properties of Brownian mo-
tion are (a) continuity (all sample paths are
continuous); (b) the Markov property–the
time τ probability distribution of X(t) for
t > τ depends only on X(τ ) (i.e., no path
dependence), (c) it is a “Martingale,”[i.e.,
Eτ [X(t)] = X(τ ), (d) it is of quadratic
variation. That is, defining ti = it/n, as
n → ∞:
n

[X(tj ) − X(tj−1)]2 → t
j=1
(e) over finite time increments ti−1 to ti,
X(ti )−X(ti−1 ) has mean zero and variance
ti − ti−1.
 Stochastic Integration

• Stochastic integration is different from tra-

ditional integration.

• A stochastic integral of a function f (.) is

defined as:
 t
W (t) = f (τ )dX(τ ) =
0
n

lim f (tj−1)[X(tj ) − X(tj−1 )]
n→∞
j=1
where tj = jt/n.

• The key thing to note about this expres-

sion is that the function f (.) must be “non-
anticipatory.” That is, the function in the
summation is evaluated at the left hand
point of the integration interval, tj−1.

8
• Note that this integral (an “Ito integral”)
is defined as a limit in the mean-square-
sense. That is,
n

lim E{ [f (tj−1)(X(tj ) − X(tj−1)] −
n→∞
j=1
 t
f (τ )dX(τ )}2 = 0
0

• In essence, this states that the variance

of the difference between the summation
term and the integral vanishes as n goes
to infinity.
 Ito’s Lemma

• Ito’s lemma is the key tool we will use to

derive valuation formulae.

• Ito’s lemma describes how functions of Brow-

nian motions behave.

• The exact derivation of Ito’s lemma is ex-

tremely technical. A heuristic approach
sufficies for our purposes.

• Consider a function of a Brownian motion

F (., .).

9
• Divide the time between 0 and t into N
equal increments δt in length. Use Taylor’s
Theorem to approximate F (.):
N
 N

F (Xt , t) − F (X0, 0) = Ftδt + FxΔXt+j
j=1 j=1
N
 N

2
+ .5 Fxx(ΔXt+j ) + Ftt δt2
j=1 j=1
N

+ Ftx (δtΔXt+j )
j=1
where ΔXt+j = X(tt+j+1 − Xt+j ).
• Taking the mean-square-limit of this ex-
pression as N → ∞ implies:
 t
F (Xt ) = F (X0, 0) + Fx(Xτ , τ )dX(τ )
0
 t
+ [Ft + .5Fxx(Xτ , τ )]dτ
0

• The key trick is getting rid of the dX(τ )2

and replacing it with dτ . We can do this
because the mean-square-limit of X 2 is t.
 More on Ito’s Lemma

• By specifying a relatively general form for

dX, we can rewrite Ito’s lemma.

• Specifically, an “Ito Process” is:

dX = μ(X, t)dt + σ(X, t)dWt

• In this expression μ(X, t) is the “drift” in X

and σ(X, t) is the volatility. Moreover, dWt
is a Brownian motion.

• Ito’s lemma therefore becomes:

F (X(t), t) = F (X(0), 0)
 t  t
+ [Fτ + Fxμ + .5Fxxσ 2 ]dτ + FxσdWt
0 0

10
• We will see the term Fτ + Fxμ + .5Fxxσ 2
repeatedly.

• Ito’s lemma is more usually seen in a stochas-

tic differential equation form rather than its
stochastic integral equation form:

dF = FxdX + Ftdt + .5Fxx σ 2dt

dF = [Fτ + Fxμ + .5Fxxσ 2]dt + FxσdWt

 Multi-dimensional Ito’s Lemma

• If we have a function of multipe stochastic

variables xi, i = 1, . . . , N , there is a multi-
dimensional version of the Ito equation:
N
 N 
 N
dF = [Ft + μiFi + .5 Fij σij ]dt
i=1 i=1 j=1
N

+ FiσidWi
i=1
where σij = E(dWidWj ).

11
 Contingent Claims Pricing:
Arbitrage Derivation a la Black-Scholes

• If we make certain assumptions about the

stochastic process that an underlying claim
follows can use the stochastic calculus tools
to determine the value of a contingent claim
on this underlying (such as a forward or an
option).

• In particular, if the underlying follows an

Ito Process, we can show that the contin-
gent claim’s value solves a particular partial
differential equation. This can be shown in
two ways, both of which center on the con-
cept of arbitrage.

12
• We will derive this PDE both ways. For
simplicity, we will assume the underlying
asset price follows a so-called “geometric
Brownian Motion” (GBM):

dSt = μStdt + σSt dWt

where St is the underlying price at t, and
μ and σ are constants. This is called a
“stochastic differential equation” (SDE).

• In a GBM, the asset price can never be-

come negative, and percentage changes in
the asset price (returns) are normally dis-
tributed.
 Key Assumptions

• We will make some key assumptions in our

analysis. Specifically:

• Zero taxes.

• Zero transactions costs.

• Continuous trading. That is, markets are

always open and you can trade every in-
stant of time without impacting price.

• Constant risk free interest rate (we will

loosen this assumption later.)

13
 Deriving the Valuation PDE

• Consider any contingent claim on S. This

could be a futures, a forward, an option,
or something more exotic.

• Posit that the value of the contingent claim

is a function V (St , t).

• Form a portfolio consisting of Δ units of

the underlying and one unit of the contin-
gent claim. The value of this portfolio is
Π ≡ StΔ + V (St , t).

• By Ito’s lemma, the dynamics of this claim

are:

ΔdSt + VsdSt + Vt dt + .5VssdSt2

14
• Substituting from the SDE for dSt ,

dΠ = (μStΔ + μSt Vs + Vt )dt

+ σSt (Δ + Vs)dWt + .5σ 2 St2dt

• Note that the last term comes from the

fact that dWt2 = dt.

• We can make this portfolio riskless by set-

ting Δ = −Vs. Our equation then be-
comes:

dΠ = Vt dt + .5σ 2St2Vssdt

• Since this portfolio is riskless it must earn

the risk free interest rate. That is, dΠ =
rΠdt = rSt Δdt + rV (St , t)dt = −rSt Vsdt +
rV dt.
• This implies:

−rSt Vsdt + rV dt = Vt dt + .5σ 2VSS St2dt

• This is a second order parabolic PDE.

• This implies our valuation equation:

rV = Vt + rSt Vs + .5σ 2 St2Vss

• Note that this equation holds for any claim

on S.
 Boundary Conditions

• The only thing that differentiates distinct

contingent claims on the same underlying
(e.g., on the same stock) is their boundary
conditions.

• For instance, for a call expiring at T with

strike price X, we know that V (ST , T ) =
max[ST − X, 0]. For a put, we know that
V (ST , T ) = max[X − ST , 0]. For a forward
contract, we know that V (ST , T ) = ST .

• These payoff conditions are not sufficient

to determine V . We need two additional
boundary conditions. These are also driven
by the nature of the problem. For instance,
with a European call we know that V (0, t) =
0 and limS→∞ = S − e−r(T −t) X.
15
• Other examples. For an American call we
know that a smooth pasting condition must
hold. For knock-out option (say a knock-
out call) the value of the claim has to be
zero at the barrier.

• Given the PDE and the barrier conditions,

we can solve for the value of the claim.
 A Remarkable Feature

• Note that the value of the contingent claim

does not depend on the “true” drift of S.
That is, μ does not appear in the PDE.
The true drift is the expected return, and
depends on the risk aversion (i.e., the pref-
erences) of investors.

• For this reason, this is sometimes called

“preference free” pricing.

• Instead of the true drift, the risk free in-

terest rate appears in the valuation PDE.

16
• That is, you can value the contingent claim
as if the expected return on the undelrying
equals the risk free rate. This would be
true in general if and only if all investors are
risk neutral. That is why this is sometimes
referred to as “risk neutral” pricing.

• This result is an artifact of hedging. Due to

continuous trading and the absence of ar-
bitrage, an investor can hedge away all risk
of holding the contingent claim by trading
the underlying.

• The hedge must be adjusted dynamically

and continuously because Vs changes with
the underlying price.

• Contingent claims are valued by replication

(i.e., hedging) not by expectation. The
drift affects expected payoffs, but this does
not matter for valuation purposes.
• Thus, if the underlying is traded, we can
pretend that we are in a risk neutral world.

• This is a wonderful feature because it is

notoriously hard to estimate the true drift
of asset prices.

• The alternative derivation of the pricing

equation arrives at the same conclusion us-
ing a different mathematical approach.
 A Convenient Change in Variables

• The foregoing PDE is somewhat cumber-

some because it has non-constant coeffi-
cients (that is, the coefficients depend on
St ). Consider the following change of vari-
ables Zt = ln(St ).

• By Ito’s lemma,

dZ = (μ − .5σ 2 )dt + σdWt

• Using this change in variables

rV = Vt + (r − .5σ 2 )Vz + .5σ 2 Vzz

• Note that this equation has constant co-

efficients if σ is a constant. This makes
numerical solution easier.

17
What Happens When the Underlying is Not
Traded?

• The foregoing analysis depends on the as-

sumption that the investor can hedge away
the dWt risk by trading the underlying–there
is a nice linear relationship between dSt and
dWt . What if the underlying itself is not
traded?

• This is relevant in many contexts. For in-

stance, if the underlying risk factor is an
interest rate, the interest rate itself is not
traded. As another example (that we will
explore in more detail later), if volatility σ is
not a constant or a function of St , but is in-
stead a stochastic process, then an option
price is a function of a non-traded asset–
the volatility.
18
• Other examples: weather derivatives and
power derivatives. Weather is obviously
not a traded asset, but there are deriva-
tives written on weather. Similarly, since
power is not storable, you cannot create a
hedge portfolio that involves holding a po-
sition in spot power–I must consume power
the instant I purchase it, and cannot re-sell
it even an instant later.

• We can use our hedge derivation even if

the underlying isn’t traded, but we will no
longer be able to derive preference free re-
sults. Instead, our pricing equations will
depend on the true drift. Equivalently, any
pricing equation will have a “market price
of risk.”
 Pricing with a Non-Traded Underlying

• Consider two contingent claims on some

non-traded underlying x. The claim with
value denoted by V expires at time T . The
claim denoted by G expires at time T  > T .

• The SDE for x is:

dxt = φdt + σx dZt

where Zt is a Brownian motion. and σx and
φ are functions of x and t.

• Form a portfolio consisting of one unit of

V and Δ units of G. By Ito’s lemma, the
dynamics of this portfolio are:

dΠ = (ΔφGx + φVx + Vt + ΔGt + .5Δσx2G2

xx
+ .5σx2Vxx
2 )dt + σ (V + ΔG )dZ
x x x t

19
• Choose Δ to make the portfolio riskless.
This requires Δ = −Vx /Gx.

• Since the portfolio is riskless, and requires

initial investment V − ΔG,

rV −rΔG = Vt +ΔGt +.5Δσx2G2

xx +.5σ 2V 2
x xx

• Collecting all V terms on the lhs and all G

terms on the rhs:
rV − Vt − .5σx2 Vxx rG − Gt − .5σx2 Gxx
=
Vx Gx
• Note that we only have one equation but
two unknowns (V and G). Note, however,
that the lhs is a function of T (and not
T  ) whereas the reverse is true of the rhs.
This is possible if and only if both sides are
independent of the maturity date. Thus,
there exists some function a(x, t) such that
rV − Vt − .5σx2 Vxx
= a(x, t)
Vx

• This is conventionally rewritten:

a(x, t) = φ − σx λ(x, t)

• The function λ is referred to as the “mar-

ket price of x risk.”
• Using this definition, we derive the follow-
ing PDE:

rV = Vt + .5σx2 Vxx + (φ − σx λ)Vx

• Note that this equation depends on the
true drift of the x process–we can’t em-
ploy the convenient assumption that the
drift of the underlying process is equal to
the risk free rate.

• Note that this derivation is more general

than the earlier one because it implies the
basic valuation equation for a traded asset.
Note that a traded asset itself must satisfy
the PDE. Thus:
rS = (φ − σxλ)

• When φ = μS and σx = σS, this implies:

μ−r
λ=
σ

• Plugging this for λ in the PDE involving

(φ − σxλ) returns the basic valuation for-
mula derived earlier. Note that we cannot
use this trick unless x is a traded asset.
 Solving PDEs Using Finite Difference
Methods

• Solution of a PDE requires determination

of a function that satisfies the relevant equa-
tion at every possible value.

• Most PDEs have no closed form solution.

Even the heat equation and Black-Scholes-
Merton equations require numerical approx-
imation.

• More complicated numerical approximation

schemes are required to solve PDEs with
boundary conditions that are more compli-
cated that BSM.

• Finite difference methods are the most com-

mon means to solve PDEs.

20
 Alternative Approaches

• There are two basic finite difference schemes–

explicit and implicit. The binomial model
is an example of an explicit scheme.

• Explicit schemes are simpler, but (a) can

face stability problems, and (b) don’t con-
verge as quickly.

• Explicit schemes are only conditionally sta-

ble. That is, for a given choice of δt, the
method is stable only if δZ is sufficiently
small. If you choose too big a δZ, you get
numerical garbage. And I mean garbage.

• Implicit schemes are generally superior. They

are unconditionally stable. For any choice
of δt, you will get a non-garbage answer
regardless of the coarseness of the δZ grid.

21
 The Explicit Approach

• Although the explicit approach is deficient

in many ways, it is worthwhile to discuss it
to illustrate the basics of finite difference
methods. It is also a flexible and easy to
code approach that is useful when you need
something fast.

• All finite difference schemes start with a

grid. That is, the state variables and time
variable are discretized.

• Consider a stock price model in which the

natural log of the stock price Z is used
as the state variable. Assuming constant
σ and r, we know that the value of any
contingent claim on this stock must satisfy
the following PDE:
∂V ∂V ∂ 2V
+ (r − .5σ 2) + .5σ 2 = rV
∂t ∂Z ∂Z 2
22
• Solution of the PDE via finite differences
requires approximation of the relevant par-
tial derivatives on the grid.

• Step 1: Create a grid. There are I + 1 log

stock price points i = 0, . . . I. Although
it is not necessary, assume that the grid
points are evenly spaced, with each one δZ
apart. There are K time steps. Each step
is δt in length. The notation is that time
step k = 0 at expiration, and today is k =
Kδ. This notation is used because solution
involves moving backwards through time in
the grid, going from values we know (at
expiration) to those we don’t.

• Step 2: Estimate the partial derivatives.

Different schemes use different estimates.
 In the explict scheme, at node i, k + 1 of
the grid:
k −Vk
Vi+1
∂Vi i−1
≈
∂Z 2δZ
where i indicates the stock price step and k
indicates the time step, and Vik is the value
of the contingent claim at node i, k.
k − 2V k + V k
∂ 2Vi Vi+1 i i−1
≈
∂Z 2 (δZ)2

∂Vi Vik − Vik+1

≈
∂t δt

• Note well: at time step k + 1 we are using

values from time step k to calculate the
partial derivatives.
• This has one plus and one minus. The plus
is (as we will see in a bit) that the explicit
method requires no solution of a system of
equations.

• The minus is that we really aren’t calculat-

ing a partial derivative because we aren’t
holding time constant. This is what in-
troduces the instability in the explicit ap-
proach.

• We now plug our estimates for the partials

into our PDE to get:

Vik+1 = AVi−1
k + BV k + CV k
i i+1
δt + .5σ 2 δt , B =
where A = −(r − .5σ 2) 2δZ δZ 2
δt , and C = (r − .5σ 2) δt +
1 − rδt − σ 2 δZ 2 2δZ
δt .
.5σ 2 δZ 2
• We start at k = 0, and solve for the value
of the option at time step 1, which is one
time step prior to expiration. At this time
step, we know the value of the option at
k = 0 as a function of the price in the grid;
that is given by the contractual features of
the option. So we know everything on the
right hand side of our expression, so we can
solve for Vi2 for i = 1, . . . I − 1.

• For i = 0 and i = I, we need to use bound-

ary conditions because for these values of
i we need to know the value of the op-
tion at points outside our grid to solve our
equation. I’ll talk more about boundary
conditions later.
• Since we now know the value of the option
for all i at time step 1, we can proceed
to time step 2. Solve the equations again,
only using the values from time step 1 on
the right hand side. Then go to time step
3, and so on, until we reach the valuation
date.

• Couldn’t be easier, eh? Just solve I + 1

equations (one at a time) for each time
step.
 The Implicit Approach

• Again start with a grid.

• In the implicit scheme, at node i, k + 1 of

the grid:
k+1 k+1
∂Vi Vi+1 − Vi−1
≈
∂Z 2δZ
where i indicates the stock price step and k
indicates the time step, and Vik is the value
of the contingent claim at node i, k.
k+1
∂ 2Vi Vi+1 − 2Vik+1 + Vi−1
k+1

2
≈
∂Z (δZ)2

∂Vi Vik − Vik+1

≈
∂t δt

• Using this approximation, the valuation PDE

can be rewritten as:
k+1
AVi−1 + BVik+1 + CVi+1
k+1
= Vik
23
 where A = 2δZ δt (r − .5σ 2) − δt σ 2, B =
2δZ 2
1 + δZδt σ 2 + rδt, and C = − δt (r − .5σ 2 ) −
2 2δZ
δt σ 2.
2δZ 2

• Note: V−1 and VI+1 are outside the

grid, so we can only write these equa-
tions for i = 1, . . . , I − 1.

• In matrix form, we observe:

ML vk+1 = vk
where
ML =
⎛ ⎞
A B C 0 . . .
⎜ ⎟
⎜ 0 A B . . . . ⎟
⎜ ⎟
⎜ . 0 . . . 0 . ⎟
⎜ ⎟
⎜ . . . . B C 0 ⎟
⎝ ⎠
. . . 0 A B C
• This matrix has I−1 rows and I+1 columns.
⎛ k+1
⎞
V
⎜ 0k+1 ⎟
⎜ V ⎟
⎜ 1 ⎟
⎜ . ⎟
⎜ ⎟
⎜ ⎟
⎜ . ⎟
k 1 ⎜ ⎟
v + ⎜
=⎜ . ⎟
⎟
⎜ . ⎟
⎜ ⎟
⎜ ⎟
⎜ . ⎟
⎜ k+1 ⎟
⎜ V ⎟
⎝ I−1 ⎠
VIk+1
and
⎛ ⎞
V1k
⎜ ⎟
⎜ V2k ⎟
⎜ ⎟
⎜ . ⎟
⎜ ⎟
⎜ ⎟
⎜ . ⎟
⎜ ⎟
vk = ⎜
⎜ . ⎟
⎟
⎜ . ⎟
⎜ ⎟
⎜ ⎟
⎜ . ⎟
⎜ k ⎟
⎜ V ⎟
⎝ I−2 ⎠
k
VI−1
 Solving The Equation System

• The above shows that the solution of the

PDE involves the solution of a set of linear
equations.

• Unfortunately, as written, we have I − 1

equations (the I − 1 rows of the matrices)
in I + 1 unknowns (note there are I + 1
columns. We need additional equations.
These come from the boundary conditions.
These give us the value of V at the top and
the bottom of the grid.

• Boundary conditions are defined by the prob-

lem. For a European put struck at X,
for instance, we know that when S = 0,
the put value is Xe−r(T −t) . Thus, V0k+1 =
Xe−r(k+1)δt . Similarly, when S is very large,
the value of the put is zero. Thus, VIk = 0.
24
• We use the boundary conditions to add two
additional rows to our matrix and vector–
one row corresponding to the upper bound-
ary condition and another corresponding to
the lower boundary conditon.

• Denote the new ML matrix that includes

the additonal rows as M̂L:

M̂L =
⎛ ⎞
A0 B0 C0 D0 . .
⎜ ⎟
⎜ A B C 0 . . . ⎟
⎜ ⎟
⎜ 0 A B . . . . ⎟
⎜ ⎟
⎜ . 0 . . . 0 . ⎟
⎜ ⎟
⎜ ⎟
⎜ . . . . B C 0 ⎟
⎜ ⎟
⎝ . . . 0 A B C ⎠
. . 0 AI BI CI DI

• The A0, B0, C0, D0, AI , BI , CI , DI come

from the boundary conditions.
• Finally, we need to add two rows to the
vector on the right-hand-side. This also
reflects the boundary conditions. Call this
new RHS vector v̂k.

• Our equation system now becomes:

M̂L vk+1 = v̂k

• We’ll see how to adjust the matrices to

reflect the boundary conditions.
 Dirichlet Boundary Conditions

• Dirichlet boundary conditions set the value

of the claim equal to some known number
at the top and bottom of the grid.

• For instance, the value of a European put

is 0 as the stock or futures price approaches
infinity. The value of the Euro put is Xe−r(T −t)
when S or F equals 0.

• Therefore, when using a Dirichlet condition

for the lower boundary, set A0 = 1, B0 = 0,
C0 = 0, D0 = 0, and the first row of v̂k
equal to the value of the claim when the
underlying price is its lowest value on the
grid. For instance, for a European put, set
this value equal to Xe−r(T −t) if the lowest
value of the underlying in your grid is zero.
25
 If the lowest value in your grid is not zero
(which will be the case if you use the log
transform, for instance) set this value equal
to Xe−r(T −t) − S0 where S0 is the lowest
value of the underlying price in your grid.

• Similarly, for a Dirichlet upper boundary

condition, set AI = 0, BI = 0, CI = 0,
D0 = 1, and the last row of v̂k equal to
the value of the claim when the underlying
price is its lowest value on the grid. For in-
stance, for a European put, set this value
to zero.
 Von Neumann Conditions

• Von Neumann conditions fix the shape of

the price function at the boundaries. For
example, von Neumann conditions can set
the slope of the function at the bound-
ary equal to 1, or to zero. Alternatively,
they many set the second derivative at the
boundary to some value, such as zero.

• As an example, for a European call, the

slope of the value function should approach
1 as the stock price approaches infinity, and
the slope should approach zero as the stock
price approaches zero.

• Warning: To ease the notation what

follows assumes you are NOT using the
log transform. I’ll show you how to im-
plement this in the log transform later.
26
• Consider the upper boundary condition. We
have:

∂V
=1
∂S

• Our first order approximation of this, using

a one sided estimate of the derivative is:

VIk+1 − VI−1
k+1
=1
δS

• So, our upper boundary condition is:

VIk+1 = δS + VI−1
k+1

or

k+1
−VI−1 + VIk+1 = −δS

• We can implement this in our matrix by

setting AI = 0, BI = 0, CI = −1, DI = 1,
and the last row of vˆk equal to −δS.
• Our lower boundary condition can be ex-
pressed:

V1k+1 − V0k+1 = 0

• So we implement this by setting A0 = −1,

B0 = 1, C0 = 0, D0 = 0, and the first row
of vk equal to 0.

• Many derivatives are nearly linear at the

boundaries. Puts and calls are examples of
this. So another common type of boundary
condition is to set the second derivative
equal to zero at the boundary.

• Our approximation of this is at the upper

boundary is:

VIk+1 − 2VI−1
k+1 k+1
+ VI−2
=0
δS 2
 or
k+1 k+1
VI−2 − 2VI−1 + VIk+1 = 0

• To implement this, set AI = 0, BI = 1,

CI = −2, DI = 1, and set the last row of
v̂k equal to 0.

• Similarly, at the lower boundary, A0 = 1,

B0 = −2, C0 = 1, D0 = 0, and the first
row v̂k equal to 0.
 The Crank-Nicolson Appraoch

• A method that combines implicit and ex-

plicit methods–the Crank-Nicolson routine–
has very desirable convergence and stability
properties.

• There are some stability problems with C-

N that lead some (like Duffie) to favor
implicit schemes combined with extrapo-
lation.

27
 Implementing Crank-Nicolson

• Again start with our grid.

• In Crank-Nicolson, at node i, k + 1 of the

grid:
k+1 k+1
∂Vi Vi+1 − Vi−1
≈ .5[
∂Z 2δZ
k −Vk
Vi+1 i−1
+ ]
2δZ
where i indicates the stock price step and k
indicates the time step, and Vik is the value
of the contingent claim at node i, k.
k+1
∂ 2Vi Vi+1 − 2Vik+1 + Vi−1
k+1

2
≈ .5
∂Z (δZ)2
k − 2V k + V k
Vi+1 i i−1
+.5
(δZ)2

∂Vi Vik − Vik+1

≈
∂t δt
28
• Note that the C-N approximations are av-
erages of the implicit and explicit approxi-
mations for the partials.

• Using this approximation, the valuation PDE

can be rewritten as:
k+1
−AVi−1 + (1 − B)Vik+1 − CVi+1
k+1
=
k + (1 + B)V k + CV k
AVi−1 i i+1
where A = − 4δZ δt (r − .5σ 2) + δt σ 2, B =
4δZ 2
δt σ 2 − .5rδt, and C = δt (r − .5σ 2) +
− 2δZ 2 4δZ
δt σ 2.
4δZ 2
 Matrix Form

• In matrix form, we observe:

ML vk+1 = MR vk
where
ML =
⎛ ⎞
−A (1 − B) −C 0 . . .
⎜ ⎟
⎜
⎜
0 −A (1 − B) . . . . ⎟
⎟
⎜ . 0 . . . 0 . ⎟
⎜ ⎟
⎜ . . . . 1−B −C 0 ⎟
⎝ ⎠
. . . 0 −A (1 − B) −C
⎛ k+1
⎞
V
⎜ 0k+1 ⎟
⎜ V ⎟
⎜ 1 ⎟
⎜ . ⎟
⎜ ⎟
⎜ ⎟
⎜ . ⎟
⎜ ⎟
v k +1 = ⎜
⎜ . ⎟
⎟
⎜ . ⎟
⎜ ⎟
⎜ ⎟
⎜ . ⎟
⎜ k+1 ⎟
⎜ V ⎟
⎝ I−1 ⎠
VIk+1
29
 MR =
⎛ ⎞
A (1 + B) C 0 . . .
⎜ ⎟
⎜ 0 A (1 + B) . . . . ⎟
⎜ ⎟
⎜ . 0 . . . 0 . ⎟
⎜ ⎟
⎜ . . . . 1+B C 0 ⎟
⎝ ⎠
. . . 0 A (1 + B) C
and
⎛ ⎞
V0k
⎜ ⎟
⎜ V1k ⎟
⎜ ⎟
⎜ ⎟
⎜ . ⎟
⎜ ⎟
⎜ . ⎟
⎜ ⎟
vk = ⎜
⎜ . ⎟
⎟
⎜ . ⎟
⎜ ⎟
⎜ ⎟
⎜ . ⎟
⎜ k ⎟
⎜ V ⎟
⎝ I−1 ⎠
VIk
 Solving The Equation System

• The above shows that the solution of the

PDE involves the solution of a set of linear
equations.

• Unfortunately, as written, we have I − 1

equations (the I − 1 rows of the matrices)
in I + 1 unknowns (note there are I + 1
columns. We need additional equations.
These come from the boundary conditions.
These give us the value of V at the top and
the bottom of the grid.

• Boundary conditions are defined by the prob-

lem. For a European put struck at X,
for instance, we know that when S = 0,
the put value is Xe−r(T −t) . Thus, V0k+1 =
Xe−r(k+1)δt . Similarly, when S is very large,
the value of the put is zero. Thus, VIk = 0.
30
• We use the boundary conditions to add two
additional rows to our matrix and vector–
one row corresponding to the upper bound-
ary condition and another corresponding to
the lower boundary conditon.

• Denote the new ML matrix that includes

the additonal rows as M̂L:

M̂L =
⎛ ⎞
A0 B0 C0 D0 . .
⎜
⎜ −A (1 − B) −C 0 . . .
⎜
⎜
⎜
0 −A (1 − B) . . . .
⎜ . 0 . . . 0 .
⎜
⎜ . 1−B −C
⎜ . . . 0
⎜
⎝ . . . 0 −A (1 − B) −C ⎠
. . 0 AI BI CI DI

• The A0, B0, C0, D0, AI , BI , CI , DI come

from the boundary conditions.
• Moreover we need to add two rows to the
MR matrix. Call this MR . The new top
and bottom rows are all zeros.

• Finally, we need to add a new I+1 vector rk

to the right-hand-side. This vector takes
the boundary conditions into account. It
has zeros in rows i = 2, . . . I −1. It may have
non-zero values for i = 0 and i = I. These
values depend on the boudnary conditions.

• Therefore, our new system becomes:

M̂Lvk+1 = M̂R vk + rk
 Solving the PDE

• The essence of the solution technique should

now be easy to understand.

• Start with what you know–the payoff to the

option at expiration. For a put with strike
price X, for instance, at expiration Vi0 =
0
max[SeZi − X, 0], where S is the current
stock price.

• Given Vi0, ∀i, you know v0 . Then you can

solve the linear equation system for v1.

• Now you know v1, you can solve for v2 .

Continue this process until you get to vK .

31
 Dirichlet Boundary Conditions for C-N

• For Crank-Nicolson, when using a Dirichlet

condition for the lower boundary, set A0 =
1, B0 = 0, C0 = 0, D0 = 0, and the first
row of rk equal to the value of the claim
when the underlying price is its lowest value
on the grid. For instance, for a European
put, set this value equal to Xe−r(T −t) if
the lowest value of the underlying in your
grid is zero. If the lowest value in your grid
is not zero (which will be the case if you
use the log transform, for instance) set this
value equal to Xe−r(T −t) − S0 where S0 is
the lowest value of the underlying price in
your grid.

32
• Similarly, for a Dirichlet upper boundary
condition, set AI = 0, BI = 0, CI = 0,
D0 = 1, and the last row of rk equal to
the value of the claim when the underlying
price is its lowest value on the grid. For in-
stance, for a European put, set this value
to zero.
 Von Neumann Conditions

• Warning: To ease the notation what

follows assumes you are NOT using the
log transform. I’ll show you how to im-
plement this in the log transform later.

• Consider the upper boundary condition. We

have:
∂V
=1
∂S

• Our first order approximation of this, using

a one sided estimate of the derivative is:

VIk+1 − VI−1
k+1
=1
δS

• So, our upper boundary condition is:

VIk+1 = δS + VI−1
k+1

33
 or

k+1
−VI−1 + VIk+1 = −δS

• We can implement this in our matrix by

setting AI = 0, BI = 0, CI = −1, DI = 1,
and rkI = −δS.

• Our lower boundary condition can be ex-

pressed:

V1k+1 − V0k+1 = 0

• So we implement this by setting A0 = −1,

B0 = 1, C0 = 0, D0 = 0, and rk0 = 0.
• Many derivatives are nearly linear at the
boundaries. Puts and calls are examples of
this. So another common type of boundary
condition is to set the second derivative
equal to zero at the boundary.

• Our approximation of this is at the upper

boundary is:

VIk+1 − 2VI−1
k+1 k+1
+ VI−2
=0
δS 2
or
k+1 k+1
VI−2 − 2VI−1 + VIk+1 = 0

• To implement this, set AI = 0, BI = 1,

CI = −2, DI = 1, and set rk
I = 0.

• Similarly, at the lower boundary, A0 = 1,

B0 = −2, C0 = 1, D0 = 0, and rk 0 = 0.
 Solving the linear equation system:
The LU Decomposition

• We want to solve the following equation

for vk+1:
ˆk+1 = M vk + rk
ML v R

• The brute force way to solve this equation

is to invert ML . This is computationally
expensive.

• There are other solution techniques that

are much more efficient computationally.

• For European options, the LU decomposi-

tion is the best approach.

34
• You can decompose the square matrix M̂L into
two other matrices, one of which has non-
zero elements only on the diagonal and the
sub-diagonal (L) and another which has
non-zero elements only on the diagonal and
the superdiagonal (U) such that M̂L = LU .
Also, you can scale things so that L has
ones on the diagonal.

• To apply the LU decomposition, define q =

MR vk + rk.

• Then LUv = q. One can exploit the sparse-

ness of L and U to solve this equation for
v.

• This is computationally more efficient be-

cause due to the diagonality of L and U ,
they can be inverted using back-substitution.
 This involves sequential solution of N equa-
tions (where N is the number of rows and
columns in the matrix of interest), each
with a single unknown.

• Matlab (and some other programs) use LU

decomposition to invert matrices. There-
fore, if you are using Matlab (or one of
these other programs) you don’t need to
do the decomposition yourself. Just use
inv(ML ) or ML−1 .

• So far we have assumed that σ and r don’t

vary over time. In this case, you only have
to calculate [LU]−1 once, and apply it at
each time step.

• In more complicated problems, σ and r may

depend on time and the state variable. In
this case, A, B, and C will depend on k and
i, and the LU decomposition and inverstion
must be done at each time step.
 Solving the linear equation system:
The SOR Method

• The LU method is quick (especially with

time and state independent coefficients),
but is not readily applicable to American
options. The Successive Over-Relaxation
(SOR) method is somewhat slower, but
can handle American option problems when
combined with a projection step (to get
PSOR, “Projected Successive Overrelax-
ation.)

• This method is a modification of the Gauss-

Seidel method. Like G-S, it is an iterative
method. One makes an initial guess, and
then modifies that guess iteratively.

35
• On each iteration, the new value is the old
value plus a correction. One iterates until
the change between the new and old values
becomes very small.

• SOR solves for vk+1 iteratively. For Euro-

pean options the procedure is as follows.
First, one chooses an “over-relaxation pa-
rameter” ω, 1 ≤ ω ≤ 2. For n = 1, choose
ω = 1. Then define matrices D, L, and
U so that ML = D + L + U. Define Mω =
D + ω L and Nω = (1 − ω)D − ω U. Given an
k +1
initial guess of vn , solve:
k +1 −1N vk+1 + ω M −1q
vn +1 = M ω ω n ω

• In this expresion, the q is derived from MR vk + rk.

It is in essence the target value.

• In this expression n is the iteration number.

• In practice, this is pretty easy. For row
i, take the i’th row of MR vk + rk. Call
this value qi . It is the “target” that you
are trying to hit. Take your initial guess
for vi. From that subtract the difference
between this qi and the i’th row of ML v
and multiply the difference by the overre-
laxation parameter ω. The difference is es-
sentially an “error”–it is the difference be-
tween the value at that iteration and your
target value. So, under SOR new value
equals old value minus ω times error. In
the G-S method, ω = 1.

• Formally:

i−1
 I
n+1 n ω n+1  n )
vik = vik + (qi− Mij vkj − Mij vkj
Mii j=0 j=i

• Note you have to loop through each row

i = 1, . . . , I − 1 and do this for each row.
 So you are double looping; you are looping
through the i’s, and then iterating on each
row.

• Usually the initial guess v0k+1 is the option

value at the previous time-step.

• Continue to iterate until you achieve con-

vergence (within some user-specified error
tolerance). Convergence means that the
change in v from iteration n to iteration
n + 1 is small. Formally, calculate:
I

(vin+1 − vin )2
i=1
and stop when this sum gets sufficiently
small.

• Proceed to the next time step.

• Always keep track of the number of iter-
ations until convergence. After finishing
one time step, increase ω by a little bit
(say .05). If the number of iterations re-
quired for convergence for that time step is
smaller than for the previous step, increase
ω a little more for the next time step. If
the number of iterations increases, use the
ω from the previous time step for the re-
mainder of your analysis.
• For American options, at each iteration
step, you can’t use a matrix operation be-
cause it is necessary to take into account
the possibility of early exercise.

• With an American option, at a given time

step k for each stock price step i (starting
with i = 2) calculate:
i−1
 I

n+1 n+ ω n+1 n )
vik = vik (qi− Mij vkj − Mij vkj
Mii j=0 j=i

• In this expression Mij is the element in row

i and column j of the ML matrix and qi is
the element in row i of the q vector.
• Note that in this method, to solve for the
option value at stock price node i you use
the option values for nodes 1, 2, . . . , i − 1.
Immediately after solving for the value of
n+1
vik using this method, compare this to
the exercise value of the option. If the
n+1
exercise value of the option exceeds vik ,
n+1
replace vik with the option exercise value
n+1
for use in calcuating v(i+1)k . If the exercise
n+1
value is smaller, use the vik calcuated us-
ing the above formula.
 Improving Accuracy:
Richardson Extrapolation

• The greater the number of time steps and

asset steps, the more accurate your solu-
tion. The Crank-Nicolson method is accu-
rate O(δt2, δZ 2 ).

• Increasing accuracy in this way is compu-

tationally expensive, because the number
of calculations is proportional to 1/δtδZ.

• Richardson extrapolation allows you to get

accuracy O(δt2, δS 3).

36
• To implement RE, first solve the problem
for a given number of asset steps (e.g.,
20). Call the value of the option given this
approach V1 and the asset step δS1 . Then
increase the number of asset steps (e.g., to
30). Call the option value using this grid
V2 and the asset step δS2 . The RE value
of the option is:

∗ δS22V1 − δS12V2
V =
δS22 − δS12
Richardson Extrapolation and the Implicit
Method

• One virtue of the C-N method is that it

is second-order accurate even without ex-
trapolation, whereas the implicit method is
only first order accurate.

• However, the C-N method frequently ex-

hibits some instabilities in valuations for
at-the-money strikes. These instabilities
are especially evident when graphing the
Greeks–the Deltas and Gammas of the op-
tion.

37
• The implicit method does not exhibit these
spurious oscillations. So how can we can
our cake and eat it too? (Or as a famous
trader I know always says–how can we have
our cake and cookie too?) That is, how
can we get second order accuracy and no
spurious oscillations?

• It’s easy. Just do the implicit method with

Richardson Extrapolation on the time step.

• First estimate the value using N1 time steps.

Call this value V1 , Then estimate the value
using 2N1 time steps. Call this value V2 .
The extrapolated value is VE = 2V2 − V1 .
VE is second-order accurate and exhibits no
spurious oscillations.
 Jump Conditions

• Handling discrete cash flows (e.g., dividends)

and some other conditions (e.g., periodic
rather than continuous monitoring of a bar-
rier for a barrier option) requires use of so-
called “Jump Conditions.” They are called
this because key variables (e.g., the stock
price) jumps when something happens (e.g.,
a dividend is paid.

• Assume a dividend of size D will be paid at

time td. Note that the value of the option
must not change as a result of the dividend
payment (everyone knows the dividend will
be paid). Immediately before the dividend
is paid, the option is worth V (S, t−
d ). Imme-
diately afterwards, it is worth V (S − D, t+
d ).
Thus, V (S, t−
d ) = V (S − D, t +
d ).
38
• We address this problem as follows. Pro-
ceed backwards through the grid in the
usual fashion. When you reach td, solve the
value of the option in the usual way. Then
implement the jump condition. At each as-
set step i at time step k (corresponding to
td), define Ẑi = ln[exp(Zi ) − D].

• Next define:
∗ Ẑi − Z0
i = Int[ ]
δZ
Zi∗ +1 − Ẑ
μ=
δZ

V (Zi , t−
d ) = μV (Zi
+ +
∗ , t )+(1−μ)V (Zi∗ +1 , t )
d d

• You have to be careful when i is small, as

in this case î < 0. In this case, just use
i∗ = 0 and μ = 1.
• You have to be clever about setting up your
time steps. In certain cases, you can just
move the dividend to the closest payment
date (taking care to adjust by the time
value of money involved in displacing the
timing of the dividend). Alternatively, you
can create a new time step corresponding
exactly to the divident payment. Here you
have to be careful when constructing your
δt to make sure that at that new time step
you are using the right δt when calcuating
your A, B, and C coefficients.

• If the option is an American call, you have

to take into account the possibility of early
exercise. In this case, you need to utilize
the SOR technique discussed earlier.
 Martingale Methods

• There is an alternative (and equivalent)

way to value derivatives. This involves use
of Martingale Methods.

• In essence, Martingale Methods imply that

any derivative can be valued by calculating
its discounted expected cash flows under
some probability measure.

• Calculating an expectation involves inte-

grating over the relevant probability mea-
sure.

• Thus, integration methods (Gaussian, Monte

Carlo) can be used to value derivatives.

39
• Some elegant mathematical theory–notably
Kolmogorov’s backward equation and the
Feynman-Kac formula–show that the value
function implied by calculating the expec-
tation under the relevant probability mea-
sure must satisfy the same valuation PDE
we derived using the Black-Scholes arbi-
trage approach. Thus, the two methods
are different ways of skinning the same cat
(apologies to PETA members).

• The fact that the methods are equivalent

derives from the fact that the absence of
arbitrage is a necessary (and sometimes
sufficient) condition for the existence of
the relevant probability measure required
to calculate the expectation.
 Martingales

• A martingale is a “zero drift” stochastic

process. That is, Wt is a martingale if

E[WT |Wt ] = Wt ∀T ≥ t

• Martingales have very desirable properties

that facilitate solution of valuation prob-
lems.

40
 Probability Measure

• A probability measure assigns probabilities

to events.

• Formally, define a state space Ω that de-

fines all the possible states of the world–all
the things that can happen. An event is a
group of states of the world.

• A σ-field allows specification of sets of events

to which probabilities can be assigned. A
σ-field A on Ω has the following properties
(a) Ω ∈ A, (b) if Ai ∈ A, then the comple-
ment of Ai, AC i ∈ A, and (c)if Ai ∈ A, i =
1, . . . , n then ∪n
i=1Ai ∈ A.

41
• The elements of A are called measurable
sets. A probability measure associates to
each measurable set a real number in [0, 1].
The probability measure P has several prop-
erties: (a) P(Ω) = 1; (b) if Ai ∩ Aj =
∅ ∀i = j, then P(∪ni=1Ai ) = i=1 P(Ai );
n

and (c) P(∅) = 0.

• A probability measure quantifies the likeli-

hood of events. A triplet {Ω, A, P} is called
a proability space.
• In financial mathematics and engineering
the probability measure is almost always
the Gaussian (or normal) distribution. The
Gaussian distribution is like the streetlight
under which the drunk looks for his keys–
we use it not necessarily because it is ap-
propriate, but because the light is better
there.
 Equivalent Measures

• The concept of an equivalent measure is

key in modern valuation theory.

• Consider a probability measure P that de-

fines the probability that a particular ran-
dom variable Z will take a particular value.
An equivalent measure Q (a) has same null
sets, but (b) a different mean. That is, if
P(Z = Z ∗) = 0, then Q(Z = Z ∗) = 0, but
EP [Z] = EQ [Z].

• We care about probability measures for stochas-

tic processes. Consider a stochastic pro-
cess Xt with associated probability mea-
sure P. To be specific, assume dXt =
γ(X, t)dt + σ(X, t)dWt . dWt is a Brownian
motion under the probability measure P.
42
• Define processes u(X, t) and α(X, t) such
that

σ(X, t)u(X, t) = γ(X, t) − α(X, t)

Assume E[exp{.5 0t u(x, s)2ds}] < ∞. De-

fine a process Mt as follows:
 t  t
Mt = exp{− u(s)dWs − .5 u(s)2ds}
0 0

• Given these preliminaries, we can state an

important result that eases valuation prob-
lems. This is called Girsanov’s theorem.
 Girsanov’s Theorem

• Girsanov’s theorem states that if we start

with a process Xt, we can create another
process that has an arbitrary drift α(X, t)
under an equivalent measure Q. Specifi-
cally,

dXt = α(X, t)dt + σ(X, t)dW̃t

where dW̃t is a Brownian motion under the
probability measure

dQ = Mt dP
Mt is referred to as the Radon-Nikodym
derivative.

• Equivalently,

dW̃t = u(X, t) + dWt

43
• In essence, Girsanov’s theorem states that
given an initial process, we can create an
equivalent process with an arbitrary mean
by adjusting simultaneously the probability
measure. This adjusted probability mea-
sure is called an equivalent measure.

• Note that W̃t is a martingale under the al-

ternative measure Q, but not under the
original measure P. Similarly, Wt is a mar-
tingale under P, but not under Q.

• There are arbitrarily many equivalent mea-

sures. One special equivalent measure is an
equivalent martingale measure. Under an
equivalent martingale measure, discounted
prices of securities are martingales. That
is, for an asset price S, if interest rates are
constant, under the equivalent martingale
measure Q,
EQ[e−r(T −t) ST |St ] = St
If interest rates are stochastic,
− tT r(s)ds
EQ[e ST |St ] = St
The Link Between the Absence of Aribtrage
and the
Existence of an Equivalent Martingale
Measure

• Girsanov’s theorem is extremely useful be-

cause there is a link between the absence
of arbitrage and the existence of an equiv-
alent martingale measure. There are two
key results.

• First, if there exists an “equivalent martin-

gale measure” Q such that all discounted
prices processes are martingales under Q,
then there are no arbitrage opportunities.

• Second, under certain technical conditions,

the absence of arbitrage implies the ex-
istence of a unique equivalent martingale
measure.
44
• In finite state space models, the absence of
arbitrage always implies the existence of an
equivalent measure. In continuous models
(with an infinite number of states), addi-
tional technical conditions are required to
ensure the existence of an equivalent mea-
sure.

• This is where Girsanov’s theorem comes

in. The theorem tells us how to create an
equivalent measure so that asset prices rise
at the risk free rate.
 An Example

• Assume interest rates are constant. Con-

sider a stock that follows the stochastic
process

dSt = μSt dt + σSt dZ

where dZ is a Brownian motion.

• The solution to this SDE under the “true”

measure is:

EP [ST ] = S0eμT =
 ∞ √ −.5Z 2
(μ−.5σ 2 )T +σ T Z e√
S0 e
−∞ 2π

√
• (Prove that ST = S0 e (μ−.5σ 2 )T +σ
T Z sat-

isfies the above SDE–Use Ito’s Lemma.)

45
• Referring back to the statement of Gir-
sanov’s theorem, I can choose an arbitrary
drift by adjusting the probability measure.

• I want a drift rSt . Thus, if I choose:

μ−r
u(S, t) =
σ
and a new Brownian motion:

dZ̃t = u(S, t) + dZt

I get:
dSt = rSt + σSt dZ̃t
where Z̃t is a martingale under the equiva-
lent measure Q:

e−.5Z̃
2
dQ = √
2π
 Using the Martingale Approach

• The martingale approach allows us to de-

termine the value of any contingent claim
expiring at T as:
− 0T r(s)ds
V = EQ[Pt(St )e ]
where Pt (St ) is the payoff to the claim at
t ≤ T.

• Once we have established the equivalent

measure, we can value any contingent claim
by calculating an expectation over this mea-
sure.

• Expectations involve calculations of inte-

grals. Thus, implementing the equivalent
martingal approach entails use of numerical
integration. Numerical integration meth-
ods include Gaussian quadrature and Monte
Carlo techniques.

46
 Numeraires

• A “numeraire” is an asset used to discount

cash flows.

• So far we have implicitly worked using a

particular “numeraire” to discount cash flows
in our valuation formulae. This is the “money
market account” numeraire. This is an ac-
count that grows at the risk free rate of
interest.

• Although the money market account is the

most common numeraire, it is not the only
one. Moreover, sometimes clever choice of
a different numeraire can make valuation
easier.

• The Numeraire Irrelevance Theorem tells

us that we should get the same value for
a contingent claim regardless of the nu-
meraire we use.
47
• The theorem states that if P and Q are
two numeraires, the value of any contin-
gent claim is given by V :

Vt = PtEP [VT /PT ] = QtEQ[VT /QT ]

• In this expression, the notation EQ (resp.
EP ) means that the expectation is taken
under a measure in which VT /QT (resp.
VT /PT ) is a martingale.

• This expression suggests that when we change

numeraires, we must change probability mea-
sures. Girsanov’s theorem tells us how to
do this
 Determining the Drift Under a Given
Numeraire

• Assume initially that we use the money

market account Bt = ert as the numeraire.
There is another asset, S1t , with dS1t =
μ1S1t dt + σ1S1t dW1, that we want to use
as a numeraire. Consider the numeraire
ratio Nt = S1t /Bt . By Ito’s lemma:

dNt dS1t dBt dS1t dBt

= − −
Nt S1t Bt S1t Bt

• Note that dBt /Bt = rdt, and that the last

term in the above expression is therefore
of o(dt). Thus,

dNt
= −rdt + μ1dt + σ1 dW1
Nt

48
• Now Girsanov’s theorem comes into play.
A corollary to the theorem (there are many
ways to express GT) implies that if dWi is
a Brownian motion under the measure im-
plied by the money market numeraire, then
dW̃i = dWi −dWi(dNt /Nt) is a Brownian mo-
tion under the measure implied by the new
numeraire.

• Note that dWi−dWi (dNt/Nt) = σ1ρdt, where

ρ = E(dW1 dWi).

• We can apply this to the numeraire as-

set. Recall that when the money market
account is the numeraire, dS1t = rS1t dt +
σ1 S1t dW1. Thus, when S1t is the numeraire,
dS1t = rS1t dt + σ12S1y dt + σ1 S1t dW̃1.

• We can apply this to other assets. Con-

sider asset two, such that when the money
market account is the numeraire dS2t =
rS2t dt + σ2S2t dW2 . Thus, when S1t is the
numeraire, dS2t = rS2t dt + σ1 σ2ρS2t dt +
σ2 S2t dW̃2.
 An Example: The Quanto Forward

• A “quanto” is a contract with a payoff de-

termined by an asset with a price expressed
in currency A, but paid in currency B. For
instance, consider a contract with a payoff
equal to the FTSE 100 stock index paid in
dollars. If the index is 4000 at expiration,
the quanto holder gets 4000 dollars.

• It is easy to figure out the value of a for-

ward on the FTSE paid in pounds. As-
suming no dividends for simplicity, using a
pound sterling money market account as
the numeraire, Ft+τ = euτ St where u is the
sterling riskless rate and St is the FTSE
spot price.

49
• To price the quanto, let’s change numeraires
from the sterling money market to the dol-
lar money market. In doing so, we have to
be careful about converting the sterling to
dollars. Calling Dt the value of the ster-
ling money market at t, and Ct the value
of the dollar/sterling exchange rate, the
dollar value of the old numeraire is Dt Ct,
where dCt = μC Ctdt + σC CtdWC . Our nu-
meraire ratio Nt is therefore Bt/Dt Ct. By
Ito’s lemma:

dNt 2
= rdt − udt − μC dt − σC dWC + σC dt
Nt
• Under the old numeraire dSt = uSt dt +
σS St dWS . By GT, under the new measure:

dSt = uSt dt − σC σS ρCS dt + σS St dW̃S

where W̃S is a Brownian motion under the
new measure implied by the new numeraire.

• Call the quanto forward price k. This price

sets the value of the forward contract to
zero. Under the new measure, for a quanto
expiring at T > t

0 = E[(ST − k)/BT ] = e−r(T −t)E[ST − k]

• Given the process for S under the new nu-

meraire,

E[ST ] = St e(u−ρσC σS )(T −t) St

• This implies that k = St e(u−ρσC σS )(T −t) St .

The Links Between the PDE and Martingale
Approaches:
Komogorov’s Backward Equation and
Feynman-Kac

• The martingale and arbitrage approaches

to contingent claim valuation seem extremely
different. In fact, though, they give the ex-
act same answer.

• Two remarkable results prove that these

approaches are equivalent. The first theorem–
Komogorov’s backward equation–is relevant
when interest rates are constant. The second–
the Feynman-Kac formula–is relevant when
interest rates are stochastic.

50
• Kolmogorov’s equations states that if one
defines erT u(x, t) = Ex[f (x, T )], where x is
the Ito proces

dx = μ(x, t)dt + σ(x, t)dW

then

ru = ut + μ(x, t)ux + .5σ 2(x, t)uxx

subject to the conditon u(T, x) = f (x).
Under the equivalent measure, μ(x, t) = r.
This implies:

ru = ut + rux + .5σ 2(x, t)uxx

• This is identical to the equation we used

to derive the Black-Scholes formula.

• The Feynman-Kac formula extends this re-

sult to the case of stochastic interest rates.
Another Way of Showing the Equivalence of
the Approaches

• We can also use the Girsanov theorem to

illustrate the equivalence between the ar-
bitrage portfolio and equivalent martingale
approaches.

• Consider an asset with SDE dSt = μSt dt +

σSt dWt where dWt is a Brownian motion
under the true probablity measure P.

• We know that under the EMM e−rtSt is a

martingale. Note:

de−rt St = −re−rtSt + e−rtdSt =

e−rt[−rSt + μStdt + σSt dWt]

51
• The Girsanov theorem implies we can find
another process dW̃t such that

dW̃t = dWt + dXt

in which W̃t is a martingale under the prob-
ability measure Q with
t t 2
Xu dWu −.5 0 Xu dudP
dQ = e 0

• Subsituting, we get:

de−rt St = e−rt[−rSt + μStdt

+ σSt dW̃t − σSt dXt ]
• For this to be a martingale, it must be the
case that the drift is zero. Hence:

St (−r + μ + σdXt ) = 0
This requires:
μ−r
dXt = dt
σ

• Now consider a contingent claim V . Its

discounted value must also be a martin-
gale under the equivalent measure. Ito’s
lemma implies the dynamics of the dis-
counted value are:

de−rt V = e−rt [−rV + Vt + VsμSt + .5σ 2 St2Vss]dt

+ e−rt σSt VsdWt

• Next substitute dW̃t = μ−r

σ + dWt. This
implies:

de−rt V = e−rt [−rV + Vt + VsrSt + .5σ 2St2Vss]dt

+ e−rt σSt VsdW̃t
• This must be a martingale, which implies:

−rV + Vt + VsrSt + .5σ 2St2Vss = 0

• Again–the Black-Scholes equation!

 Valuing Contingent Claims Through
Integration

• The Martingale approach implies that we

can value contingent claims by taking an
expectation. Taking an expectation involves
integrating over a probability distribution.

• There are two basic integration techniques–

explicit integration, and Monte Carlo inte-
gration.

• Which is appropriate depends on circum-

stances.

52
 Explicit Integration

• Let’s take the simplest case–a European

call option with time τ to expiration struck
at K. The payoff to this option is max[S −
K, 0]. The Martingale approach implies that
the value of this call is:

C(S, K, τ ) =
∞ √
−rτ (r−.5σ 2 )τ −σ τ Z
e max[Se − K, 0]
−∞
e−.5Z
2
· √
2Π

• We note that we can solve for a critical

value of Z, Z ∗ such that S − K = 0. This
is:
ln S + (r − .5σ 2 )τ
Z∗ = K √
σ τ
53
• Thus
 Z∗ √
−rτ (r−.5σ 2 )τ −σ τ
C(S, K, τ ) = e [Se Z − K]
−∞
e−.5Z dZ
2
· √
2Π

• Consider the term

 Z∗ √ −.5Z 2
Se −.5σ 2 τ −σ τ e √
dZ
−∞ 2Π

√
• The exponent is: −.5σ 2τ − σ τ − .5Z 2 . We
can use the trick of “completing the square
to simplify this:

2 √ 2 √ 2
−.5(σ τ + 2σ τ + Z ) = −.5(Z + σ τ )

√
• Define a new variable Y = Z + σ τ . Since
√
Z ≤ Z ∗, Y ≤ Z ∗ + σ τ ≡ Y ∗ Thus,
 Z∗ √ −.5Z  Y∗
e−.5Y dY
2 2
Se −.5σ 2 τ −σ τ e √
dZ
= S √
−∞ 2Π −∞ 2Π
• This is SN (Y ∗), where N (.) is the cumu-
lative normal distribution. It can be es-
timated with arbitrary accuracy using nu-
merical techniques.

• Now consider the term

 Z∗ −.5Z 2
e dZ
K √
−∞ 2Π

• This is just KN (Z ∗).

• Therefore,

C(S, K, τ ) = SN (Y ∗) − e−rτ KN (Z ∗)

• This is the Black-Scholes-Merton formula.

 Monte Carlo Integration

• Monte Carlo techniques are extremely com-

mon in financial applications.

• In a Monte Carlo approach, one simulates

the behavior of the price of the underlying
using large numbers of random draws, and
then calculates the option value as a sam-
ple average of the payoffs of the option.

• Consider an option with 100 days to matu-

rity. First draw 100 standard normal vari-
ables z1, . . . , z100. This implies a series of
100 stock prices. The stock price on day
j is
√ j
(r−.5σ 2 )(j/365)+(σ/ 365) i=1 zi
Sj = S0e

54
• On day 100, the value of the option is:
max[S100 − K, 0].

• Save the value of the option at expira-

tion. Repeat this process many times (e.g.,
10,000 times) and take the average of all
the values you get. Discount the average
to reflect the time value of money. The
discounted average is the value of the op-
tion.
 Path Dependent Options

• One advantage of Monte Carlo is that you

can value so-called path-dependent options.

• A path-dependent option has a payoff that

depends on the path the underlying price
follows, not just the value of the underlying
at expiration.

• An example of a path-dependent option

is a “knock-out” call. This option be-
comes valueless if the value of the underly-
ing reaches some level prior to expiration.

• In a Monte Carlo valuation approach, you

can set the value of the option equal to
zero on any path that crosses the “knock-
out” barrier.

55
 Increasing the Accuracy of MC

• There are a variety of techniques to im-

prove the accuracy and speed of Monte
Carlo.

• In the antithetic variable technique, in your

first sample use zi, in your second sample
use −zi.

56
• The control variable technique requires the
existence of a related option that has an
analytical solution. For instance, consider
an “Asian” option. An Asian option has
a payoff that equals the average of the
underlying price over some time interval.
There is no analytical solution for an Asian
option with a payoff given by an arithmetic
average of prices. There is, however, a an-
alytical solution for the value of an Asian
option based on a geometric average.

• In the control variate technique, calcualate

the value of both the geometric and arith-
metic Asian using MC. Call the value of
the arithmetic option value given by the
MC approximation as A. Call the value of
the geometric value given by the MC as G.
Call the analytical value of the geometric
option G∗. Then estimate the value of the
arithmetic Asian as A + (G − G∗).
• Quasi-random sequences use a special al-
gorithm to choose the zi. An example is
the Sobol’ sequence. By insuring that there
are fewer “gaps” and “clumps” in the zi,
quasi-random sequences allows you to use
fewer random samples when estimating your
option value. Due to this, the accuracy of
the MC approximation is proportional to
√
1/M using qrs, instead of 1/ M using a
basic random number generator, where M
is the sample size.
 Which Valuation Method is Best?

• Explicit integration is preferable for Euro-

pean options.

• Monte Carlo is frequently the best approach

when (a) valuing options that have payoffs
that depend on several underlying variables
(e.g., stochastic volatility) and (b) certain
path dependent options.

• Monte Carlo cannot handle American op-

tions.

• PDE techniques are best for American op-

tions. Moreover, PDE approaches are fre-
quently flexible enough to value path-dependent
options (such as Asians). This requires
increasing the number of state variables,
which raises computation (and program-
ming) costs.

57
 How Well Do These Approaches Work?

• These valuation approaches work well when

its assumptions closely approximate reality.

• The key assumption is that the behavior of

financial prices is well-approximated by an
Ito Process.

• Unfortunately, there’s a lot of evidence that

financial prices don’t behave like Ito Pro-
cesses.

58
 The Volatility Smile

• If volatility is constant, then all options on

the same underlying should have the same
“implied volatility.” An implied vol is the
volatility parameter that sets the option
value given by the BSM model equal to
the market price of the option.

• If volatility is a function of S and t, as is

permissable with an Ito Process, then day
after day, the function σ(S, t) that best fits
options prices shouldn’t change.

59
• We know that implied volatility is not the
same for all options. Indeed, there is some-
thing called the volatility smile, or volatility
“smirk.” That is, implied volatilities are a
function of the strike price. The implied
vol for at-the-money options is lower than
the implied vol for options with lower strike
prices, and is sometimes lower than the
implied vol for options with higher strike
prices.
 Financial Orthodentia:
Fixing the Smile ;-)

• On a given day, you can choose (using

complex mathematical techniques) a func-
tion σ(S, t) that fits the volatility smile.

• Derman and Kani first derived a method

for implementing this approach based on
binomial trees. Although this is popular,
and widely used, it is numerically crude.
It faces what as known as an “overfitting
problem.” There is an infinite number of
σ(S, t) functions that can fit a finite num-
ber of option prices exactly. Which one to
choose? The DK approach always gives
you one such function, but if you change
the data (e.g., the options prices) by the
tiniest amount, the DK approach will give
60
 you a completely different function. That
is, the DK approach is not stable.

• There are other, very advanced approaches

that avoid overfitting. These are called
“inverse techniques.”

• Even if you use inverse techniques, you may

run into problems. This technique is good
if and only if the underlying price process
is an Ito process. If it isn’t, this approach
will not work.

• There is evidence that stock, bond, and

commodity prices are not Ito processes.
• With an Ito process, using inverse tech-
niques you should get the same σ(S, t) func-
tion every day. In fact, you don’t. More-
over, an Ito process implies that volatility
depends only on the underlying price. In
fact, there is a lot of evidence that volatil-
ity changes a lot and that these changes
cannot be attributed solely to changes in
the underlying price.

• Indeed, there is an immense body of ev-

idence showing that volatility varies ran-
domly over time, and that this variation is
largely independent of movements in the
underlying price.

• Random volatility can explain the smile.

• If volatility is random–that is, stochastic–

then the Ito process-based approach will
give incorrect option valuations.
 Valuation With Stochastic Volatility

• Volatility is not a traded asset. Therefore,

if we want to use volatility as a state vari-
able then (a) we now have two state vari-
ables (the underlying and the volatility) and
(b) we have to use a formula with a market
price of risk.

• Let’s specify a fairly general S volatility

process:

dσ = μσ + νdWσ

where the parameters μσ and ν are poten-

tially functions of σ and t.

• Then, calling λ the market price of volatil-

ity risk, our valuation PDE for a contingent
claim with value V becomes:
61
rV = Vt + .5ν 2 Vσσ + (μσ − νλ)Vσ
+ rSt VS + .5St2σ 2 VSS + St σνVSσ
 Implementation

• This is a two-dimensional parobolic PDE.

It can be solve using a variety of meth-
ods, including finite differences (especially
useful for American options), Monte Carlo
integration, or Fourier techniques.

62
• Solving the model also requires specifica-
tion of the volatlity process parameters.
There are several standard processes that
are tractible (but perhaps not completely
realistic) used for this purpose.

• Solving the model also requires knowledge

of the current value of σ and estimation
of the volatility process parameters. Since
σ is not observed directly, these are hard
problems. That is, σ is a “latent” process
that requires some fancy statistics to es-
timate. One approach is to use discrete
time (e.g., daily data) to estimate the pa-
rameters. This is feasible if the discrete
time model (e.g., a GARCH model) has a
continuous time limit.
• Another problem is that it is necessary to
estimate a market price of risk.

• Initially, practitioners assumed λ = 0. Much

empirical evidence shows that this is incor-
rect. If it were true, delta hedged options
positions would earn the riskless rate. In
fact, returns on such positions deviate sub-
stantially from the riskless rate. This im-
plies that there is another risk premium af-
fecting options prices. A likely candidate is
volatility risk.

• λ is also an unobserved function.

• The theoretically purest way to address all

these problems is to use historical options
price data as well as underlying price data
to estimate μσ , ν, and λ, and these parame-
ters plus current options prices to estimate
σ. This is a demanding process.
 Jumps

• Eyeballing any financial time series one sees

instances where prices seem to change dis-
continuously. That is, they “jump” or “gap.”

• Remember that Brownian motions have con-

tinuous sample paths–they do not exhibit
jumps. Thus, Ito processes cannot capture
the jumpiness observed in prices.

• One way to address this issue is to utilize

so-called “jump-diffusion” models. These
marry a Brownian motion process and a
Poisson process.

63
• A Poisson process is one that exhibits no
change with probability 1−λdt and changes
with probability λdt. That is, the poisson
process q is dq = 0 with probability 1 − λdt
and dq = 1 with probability λdt. λ is re-
ferred to as the intensity of the jump pro-
cess.

• a stock price model that incorporates the

jump is:

dS = μSdt + σSdW + (J − 1)Sdq

• In this expression, J gives the magnitude

of the jump. J can be deterministic (e.g.,
J = .9, indicating that the stock prices falls
10 percent during a jump) or J can be a
random variable.
• Although the jump process can be specified
quite generally to describe observed data
accurately, jumps pose acute problems for
valuation.

• Remember that the valuation methods that

we have used so far rely on hedging and
replication arguments. The jump compo-
nent cannot be hedged using the underly-
ing, however.

• If the magnitude of the jump is known and

takes a single value (i.e, J is a constant)
then we can construct a hedge portfolio
consisting of two options and the under-
lying. This introduces a market price of
risk.
• If the jump magnitude is stochastic things
get even more complicated. If J takes on K
values, we need a portfolio of K +1 options
and the underlying to hedge all relevant
risks. This injects K risk prices. Note that
if J is a continuous random variable, and
hence K = ∞, we have an infinite number
of risk prices!

• Various models have been proposed to fi-

nesse this issue. Most implicitly or explic-
itly assume that the jump risk is diversi-
fiable, or can be priced using the CAPM.
Such models calculate values through in-
tegration. These approaches are further
examples of looking for our keys under the
lamppost.

• Jump models also pose serious estimation

issues. That is, it is not a trivial task to
estimate the parameters of a jump process.
• Jump models have become especially pop-
ular for pricing electricity derivatives be-
cause electricity prices are very jumpy. This
points out a serious issue–descriptively ac-
curate characterizations of price processes
may not lead to reliable pricing models.
Even if we know the probability and in-
tensity of jumps, unless we know how the
market prices these jumps we cannot de-
rive reliable pricing models. Even if we can
take expectations under the true measure
(because we have characterized the statis-
tical properties of the jumpy price series)
we can’t value unless we know the relevant
probabilities under the equivalent measure.
• Jump models also pose problems in elec-
tricity because λ is time dependent–a jump
is more likely in the summer than the fall,
for instance. Moreover, the distribution of
J is almost certainly time dependent–big
jumps are more likely in the summer.

• Jump models provide a great illustration of

the dilemmas of derivative pricing; descrip-
tively accurate models seldom can be in-
corporated in the available valuation frame-
work. Practitioners therefore face a trade-
off between descriptive accuracy and valu-
ation feasibility.
 Interest Rate Models

• One of the most important areas of deriva-

tives modeling involves fixed income mar-
kets. Modeling derivatives on fixed income
products requires modeling interest rates.

• There are two basic “flavors” of interest

rate models: (a) spot rate models, and (b)
Heath-Jarrow-Morton forward rate models.

• The spot rate models are much more tractible.

We will focus on those.

64
 Spot Rate Models

• Spot rate models characterize the dynam-

ics of the instantaneous interest rate–the
interest rate at which you can borrow or
lend over the next instant.

• There is in fact no real world analogue to

the instantaneous spot rate–it is merely a
modeling convenience.

• The generic spot rate model is a single fac-

tor model of the type:

drt = θ(r, t)dt + σ(r, t)dz

• Different models make different assump-

tions about θ(., .) and σ(., .).

65
• The earliest model is the Vasicek model:

dr = a(b − r)dt + σdz

• in the Vasicek model, a > 0, b > 0, and σ

are constants. This model exhibits “mean
reversion.” That is, when r > b, the spot
rate tends to fall; when r < b, the interest
rate tends to rise. Thus, the interest rate
reverts to the long range mean value of b.

• Interest rates can become negative in this

model. This is a potentially serious prob-
lem (especially when dealing in low interest
rate environments).

• The Cox-Ingersoll-Ross (CIR) model:

√
dr = a(b − r)dt + σ rdz
• This “square root process” model rules out
negative interest rates. It implies that in-
terest rate volatility rises with interest rates.

• The Ho-Lee model:

dr = θ(t) + σdz

• The Hull-White model:

dr = (θ(t) − ar)dt + σdz

• This model adds mean reversion to Ho-

Lee.

• Negative interest rates are possible in both

Ho-Lee and Hull-White.
A Generalized Overview to Pricing Interest
Rate Derivatives

• Given a spot rate model, we can use our

standard pricing techniques to value any
fixed income derivative.

• Call V (r, t) the value of an interest sensitive

contingent claim. Then:

rV = Vt +.5σ(r, t)2Vrr +[θ(r, t)−σ(r, t)λ(r, t)]Vr

• Note that we have to include a market price

of risk because the spot rate is not a traded
asset.

• The spot rate models discussed earlier are

used largely because they are tractible, and
allow closed-form solutions of this equation
for certain instruments.
66
• In particular, it is possible to solve analyti-
cally for zero coupon bond prices using the
spot rate models discussed above. A zero
coupon bond that matures at T has the
boundary condition V (r, T ) = 1.

• For the spot rate models, the zero coupon

bond price PT is of the form

PT (r, t) = A(t, T )e−B(t,T )r(t)

• In the Vasicek model:

1 − e−a(T −t)
B(t, T ) =
a
and
(B(t, T ) − T + t)(a2b − .5σ 2)
A(t, T ) = exp[
a2
σ 2B(t, T )2
− ]
4a
• There is also a closed form solution for
a European call or put option on a zero
coupon bond in the Vasicek model.

• In the CIR model:

2(eγ(T −t) − 1)
B(t, T ) =
(γ + a)(eγ(T −t) − 1) + 2γ

2γe.5(T −t)(a+γ) 2
A(t, T ) = [ ]2ab/σ
(γ + a)(eγ(T −t) − 1) + 2γ

where γ = a2 + 2σ 2.

• In the Ho-Lee model, B(t, T ) = 1 and

P (0, T ) ∂P (0, t)
ln A(t, T ) = ln − (T − t)
P (0, t) ∂t
− .5σ 2t(T − t)2

• In Hull-White:
1 − e−a(T −t)
B(t, T ) =
a
 P (0, T ) ∂P (0, t)
ln A(t, T ) = ln − B(t, T )
P (0, t) ∂t
1 2 −aT −at 2 2at
− 3
σ (e − e ) (e − 1)
4a
 Term Structures

• Each spot rate model implies a term struc-

ture of interest rates. The term structure
is a curve that gives the interest rate as a
function of maturity. The T -period inter-
est rate is the yield on a zero coupon bond
with maturity T .

• The Vasicek model allows an upward slop-

ing, downward sloping, or “humped” term
structure. This is somewhat limiting.

• The Ho-Lee and Hull-White models allow

“exact” fitting of the term structure be-
cause of the “fudge factor” θ(t). One can
choose θ(t) to fit term structures exactly.

67
• Note that since you are fitting the models
to market prices, the θ(t) function implic-
itly includes a market price of risk.

• To calculate θ(t) in HL, collect zero prices

for every maturity. Use these prices (or the
prices from Eurodollar futures or FRAs) to
calculate instantaneous forward rates. An
instantaneous forward rate for time t is the
rate that I can lock in today for borrowing
at t for repayment at t+dt. Call the forward
rate for maturity t F (0, t). Then, in HL:

θ(t) = Ft(0, t) + σ 2 t

• This is sometimes approximated as θ(t) =

Ft(0, t), The partial derivative is estimated
using finite differences.
• In HW, follow the same data collection pro-
cedure, exept use:

σ2
θ(t) = Ft(0, t) + aF (0, t) + (1 − e−2at)
2a

• This is often approximated as

θ(t) = Ft(0, t) + aF (0, t)

 Calibration

• This process of fitting θ(t) is called “cali-

bration.”

• Calibration gives an exact fit to the term

structure.

• This exact fit seems comforting, but in fact

is dangerous. ALWAYS BEWARE A PER-
FECT FIT. Perfect fitting in fact implies
“overfitting.”

• The problem here is similar to the problem

with Derman-Kani discussed earlier. If I fit
θ(t) using one set of data, and then change
the data only slightly, I’ll get a completely
different θ(t).

• Inverse problem techniques are more ap-

propriate for this analysis.

68
 How Well do Short Rate Models Work?

• Short rate models are extremely popular.

They have some serious problems, though.

• For one thing, single factor models are of

dubious validity. Principal components anal-
ysis suggests that there are multiple factors–
at least three, perhaps as many as 10–
driving interest rates. The three strongest
factors appear to be “shift,” “twist,” and
“hump.” Thus, single factor models can’t
mimic the dynamics of real world interest
rates.

• Multi-factor models seem to be a desirable

alternative, but they are considerably more
complicated to implement.

69
• Relatedly, if the models were correct we
should see the same θ(t) functions day af-
ter day. In fact, we don’t. This reflects the
fact that the calibration papers over serious
limitations in the models’ ability to capture
real world interest rate dynamics. In partic-
ular, the models cannot reliably capture the
extreme steepness of the term structure at
the very short-maturity section. The fit-
ted θ(t) functions imply that the short end
should “flatten out” but it usually doesn’t.

• There is also substantial evidence of stochas-

tic volatility in interest rates. The standard
models don’t capture this.