Interest Rate Modelsirmuva.rogerlord.com/Slides.pdf · Suppose you are a trader in a bank or a...
Transcript of Interest Rate Modelsirmuva.rogerlord.com/Slides.pdf · Suppose you are a trader in a bank or a...
October – November 2014
Dr. ir. Roger Lord
(roger (dot) lord (at) gmail (dot) com)
Interest Rate Models
MSc in Stochastics and Financial Mathematics
VU, UvA, Utrecht and Leiden University
October - November 2014 2
Material
The slides, a selection of the book of Filipović,
chapter 2 of my PhD thesis, plus a selection of
scientific articles will be the material;
All material will be made available at:
http://irmuva.rogerlord.com/
October - November 2014 3
Material (2)
The material from Filipović will be:
HJM: 6.1 up to and including 6.4
Forward measures: 7.1 – 7.3 (7.1 until remark 7.1)
Market models: 11.1 – 11.7 (without 11.3 - 11.4)
For the last part (affine models) the material is:
Chapter 2 from Lord [2008], to be found here.
Finally, the papers of Brigo and Mercurio [2001],
Andreasen [2005], Mercurio [2009] are recommended
as additional reading material.
October - November 2014 4
Material (3)
Other good books on interest rate models:
October - November 2014 5
Outline
We will deal with the following topics:
1. Why do we need models?
2. Heath-Jarrow-Morton
3. Forward measures
4. Market models
5. Recent developments
6. Affine models
Three sets of homework exercises will be provided, and these will determine 50% of your grade. The other half will be determined by the exam.
October - November 2014 6
Schedule
Date Time Location
Thu 23-10-2014 15:00-18:00 SP G3.05
Thu 30-10-2014 15:00-18:00 SP G3.05
Thu 06-11-2014 09:00-11:00 D1.162
Thu 20-11-2014 09:00-13:00 SP G3.05
Thu 18-12-2014
(Exam)
13:00 -16:00 SP D1.112
October - November 2014 8
The need for models
Suppose you are a trader in a bank or a hedge fund,
and you only trade exchange-traded options. Do you
really need a model?
Suppose you are working in a pension fund, and
you can value your swaps and swaptions portfolio
relatively simply by bootstrapping the interest rate
curve and interpolating between available swaption
volatility quotes. Do you really need a model?
October - November 2014 9
The need for models (2)
We do not always need sophisticated pricing models.
However, they can be of great importance in the
following situations:
Valuing (exotic) contracts for which there is hardly
any market data available;
As a mean of a more sophisticated interpolation
technique – incorporating no-arbitrage restrictions
in volatility interpolation is already quite tricky;
To determine hedges.
October - November 2014 10
Exotic interest rate derivatives
Once we know how to price the plain vanilla options
(caps/floors and swaptions), the next step is of course
to consider more exotic contracts. Reasons for using
swaps and swaptions have been treated. Suppose
further that a company is uncertain about:
whether or not it will require the swap;
from when onwards it will require the swap.
In this situation it may be advisable for the company
to enter into a Bermudan swaption. With this product
the company can decide whether and when it enters
into a swap.
October - November 2014 11
Exotic interest rate derivatives (2)
Bermudan swaptions are characterised by:
Possible exercise dates ex = {t1, …, tM} ;
Typically there are two lockout dates: a date until
which we cannot call, and one specifying after
which date we cannot call anymore.
The most liquid Bermudans have a fixed maturity. If
the swaption is a payer swaption, and we choose to exercise at Ti ex , the payoff is:
ii,i Tat )T(PS
October - November 2014 12
Exotic interest rate derivatives (3)
The term fixed maturity refers to the fact that the
maturity date of the swap, i.e. the final payment date,
T, is fixed. Typically the final lockout date is one
period prior to the final payment date, i.e. T-1. With t1
being the first possible exercise date, the deal is often
referred to as a T no-call (nc) t1, or t1 into T
Bermudan swaption.
October - November 2014 13
Exotic interest rate derivatives (4)
An example of a Bermudan option in real life:
October - November 2014 15
Recap of short rate models
Short rate models model the short rate (note: this rate
is not directly observable in the market).
Mathematically it is equal to:
where f(t,T) is the instantaneous forward rate:
See also Section 2.2 of Filipović.
)t,t(f)T,t(flim)t(rTt
T
)T,t(Pln)T,t(f
October - November 2014 16
Recap of short rate models (2)
There are a wide variety of short-rate models, for
example the so-called equilibrium/endogenous
models, which try to explain the behaviour of bond
and option prices as a function of the short rate. Some
examples:
Model Specification
Merton (1973) )t(dWdt)t(dr
Vašíček (1977) )t(dWdt)r)t(r(a)t(dr
Dothan (1978) )t(dW)t(rdt)t(r)t(dr
Cox, Ingersoll & Ross (1985) )t(dW)t(rdtr)t(ra)t(dr
October - November 2014 17
Recap of short rate models (3)
Hereafter a number of authors considered no-arbitrage
models in which the term structure is a given that has
to be fitted in order to be able to price other, more
complex derivatives – coinciding with an increased
liquidity in the interest rate market:
Model Specification
Ho-Lee (1986)
Black-Derman-Toy (1990)
Hull-White (1990)
Hull-White Ext-CIR (1990)
Black-Karasinski (1991)
)t(dWdt)t()t(dr
)t(dW)t(r)t(dt)t(r)t()t(dr
)t(dW)t(dt)t()t(r)t(a)t(dr
)t(dW)t(r)t(dt)t()t(r)t(a)t(dr
)t(dW)t(dt)t()t(r)t(a)t(rlnd
October - November 2014 18
An extension of short rate models
There is a general way to come up with a no-arbitrage
extension of an otherwise tractable short rate model.
The idea builds on that of Hull and White, and is
formalised in e.g. Brigo and Mercurio [2001].
The ingredients are:
A short rate model in which we can calculate
Pmodel(t,T) as a function of the state variables;
Market prices Pmarket(0,T) for all T;
In general Pmodel(0,T) Pmarket(0,T)
(otherwise there would be no point to this exercise).
October - November 2014 19
An extension of short rate models (2)
By setting:
we obtain bond prices:
The objective is that the left-hand side equals the
market prices of zero-coupon bonds:
)t(r)t()t(r elmod
)T,t(Pdu)u(exp)T,t(P elmod
T
t
)T,0(Pdu)u(exp)T,0(P elmod
T
0market
October - November 2014 20
An extension of short rate models (3)
Rearranging the equation yields:
)T,0(f)T,0(f
T
)T,0(Pln
T
)T,0(Pln
T
ln)T(
elmodmarket
marketmodel
)T,0(P
)T,0(P
market
model
October - November 2014 21
Shortcomings of short rate models
Suppose we have a one dimensional affine short rate
model:
i.e. the zero bonds are exponentially affine in the short
rate r(t). We can already see that the model is a bit too
simple. Suppose the short rate is initially r and moves
to r+ε. Then the yield curve changes as follows:
)t(r)T,t(B)T,t(Ae)T,t(P
tT
)r()T,t(B)T,t(A
tT
r)T,t(B)T,t(A)T,t(R)T,t(R
October - November 2014 22
Shortcomings of short rate models (2)
The difference is therefore εB(t,T)/(T-t), which means
that only particular changes can be achieved across the
yield curve. For example, consider the Vašíček model:
In this time-homogeneous model we therefore have:
The mean reversion coefficient a is typically just a
couple of percent, so that the yield curve can roughly
only shift parallel.
)a(O)tT(a1)tT/()T,t(B 2
21
)tT(a
a1 e1)T,t(B
October - November 2014 23
Shortcomings of short rate models (3)
An example in the Vašíček model of what can happen
when the short rate moves from 2.5% to 2.7% (quite a
large move):
October - November 2014 24
Shortcomings of short rate models (4)
Furthermore, one can check that in one-factor models
that have an affine term structure zero yields and
instantaneous forward rates are perfectly correlated.
Empirically for zero yields and continuously
compounded annual forward rates:
2
4
6
8
2
4
6
8
10
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Co
rre
lati
on
Tenor (yr)
Tenor (yr)
2
4
6
8
2
4
6
8
10
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Co
rre
lati
on
Tenor (yr)
Tenor (yr)
October - November 2014 25
Heath-Jarrow-Morton
Before the previously described developments that
allowed short rate models to fit the initial term
structure exactly, the shortcomings of short rate
models were:
Initial term structure could not be fit exactly;
Fact that driving process is one-dimensional is not
very realistic.
For these reasons, Heath, Jarrow and Morton ventured
to investigate whether they could formulate a multi-
dimensional model for instantaneous forward rates.
October - November 2014 26
Heath-Jarrow-Morton (2)
In short rate models we model the short rate. In the
Heath-Jarrow-Morton (HJM) framework, we model
all instantaneous forward rates under Q:
where:
is a one-dimensional adapted stochastic process;
is a d-dimensional adapted stochastic process;
W is a d-dimensional Wiener process;
f(0, T) = fmarket(0, T).
)t(dW)T,t(dt)T,t()T,t(df
October - November 2014 27
Heath-Jarrow-Morton (3)
The question Heath, Jarrow and Morton asked, is:
What does the absence of arbitrage imply for (t,T)?
Remember we can write zero-coupon bond prices as:
and we know that, in absence of arbitrage, the
following must hold:
T
tdu)u,t(fexp)T,t(P
)t(dWOdt)t(r)T,t(P
)T,t(Pd
October - November 2014 28
Heath-Jarrow-Morton (4)
We should analyse the dynamics of P(t,T) and see
what this implies about its drift.
First of all notice that the integrated SDE reads:
implying the following for the short rate:
t
0
t
0)s(dW)T,s(ds)T,s()T,0(f)T,t(f
t
0
t
0)s(dW)t,s(ds)t,s()t,0(f)t,t(f)t(r
October - November 2014 29
Heath-Jarrow-Morton (5)
If we take the logarithm of P(t,T), we obtain:
We assume the conditions on and are such that
Fubini’s theorem (classical and stochastic) apply. See
also paragraph 7.5 of this course for a rigorous proof
of Fubini’s theorem. It follows that:
T
t
T
t
t
0
t
0
T
t
T
t
du)s(dW)u,s(duds)u,s(
du)u,0(fdu)u,t(f)T,t(Pln
t
0
t
s
t
0
T
s
T
t
t
0dsdu)u,s(dsdu)u,s(duds)u,s(
October - November 2014 30
Heath-Jarrow-Morton (6)
Collecting results, this yields:
where we defined:
t
0
t
0
T
s
)s(dW)T,s(v
ds)s(rdu)u,s(
)T,0(Pln)T,t(Pln
T
sdu)u,s()T,s(v
October - November 2014 31
Heath-Jarrow-Morton (7)
Applying Itō’s lemma to – ln P(t,T) yields:
Since the drift should equal r(t), we can answer Heath,
Jarrow and Morton’s question as:
)t(dW)T,t(v
dt)T,t(vdu)u,t()t(r)T,t(P
)T,t(dP 2
21
T
t
2
21
T
t)T,t(vdu)u,t(
October - November 2014 32
Heath-Jarrow-Morton (8)
When is P(t,T) a true martingale?
Either when the Novikov condition is satisfied:
for all maturities T, or:
If forward rates f(t,T) are nonnegative.
The latter implies 0 P(t, T) 1. Since P(t, T) divided
by the money market account is a local martingale,
and it is now also uniformly bounded (MMA 1), it is
by definition a true martingale.
T
0
2
21 dt)T,t(vexpE
October - November 2014 33
HJM and short rate models
Remember that the short rate process satisfies:
and that:
t
0
t
0)s(dW)t,s(ds)t,s()t,0(f)t,t(f)t(r
T
tdu)u,t()T,t(v
2
212
21
T
t)T,t(v
T)T,t()T,t(vdu)u,t(
October - November 2014 34
HJM and short rate models (2)
As an example, consider . Then we have:
We obtain:
)tT()T,t(v
)tT()T,t(vT
)T,t( 22
21
)T,t(
)t(Wt)t,0(f
)s(dWds)t,s()t,0(f)t,t(f)t(r
22
21
t
0
t
0
October - November 2014 35
HJM and short rate models (3)
In SDE form this becomes:
which is exactly the Ho-Lee model with the drift that
is required to fit the initial term structure exactly.
)t(dWdttdt)t,0(f)t(dr 2
t
October - November 2014 36
HJM and short rate models (4)
Let us look at the general case. We start with:
Assuming that (t, T) is differentiable w.r.t. T, as well
as the technical conditions to be allowed to use
Fubini’s theorem (see Proposition 6.1):
t
0
t
0)s(dW)t,s(ds)t,s()t,0(f)t,t(f)t(r
t
0
u
0 u
t
0
t
0
t
s u
t
0
t
0
t
0
t
0
du)s(dW)u,s()s(dW)s,s(
)s(dWdu)u,s()s(dW)s,s(
)s(dW)s,s()t,s()s(dW)s,s()s(dW)t,s(
October - November 2014 37
HJM and short rate models (5)
Similarly:
Hence the short rate is an Itō process:
where is defined accordingly.
t
0
u
0 u
t
0
t
0duds)u,s(ds)s,s(ds)t,s(
t
0 u
t
0 udu)u,0(f)0(rdu)u,0(f)0,0(f)t,0(f
t
0
t
0)s(dW)s,s(ds)s()0(r)t(r
October - November 2014 38
HJM – Existence and uniqueness
For modelling purposes one would prefer
The HJM SDE in this case equals:
If the volatility function is uniformly bounded, jointly
continuous and Lipschitz continuous in the last
argument, there exists a unique solution of f(t,T).
)t(dW)T,t(f,T,t
dtdu)u,t(f,u,t)T,t(f,T,t)T,t(dfT
t
)T,t(f,T,t)T,t(
October - November 2014 39
HJM and the curse of dimensionality
Consider a one-factor HJM model:
Whilst this is a one-factor model, its general form
suffers from the curse of dimensionality.
To see the curse of dimensionality in a one-factor
model, let us integrate the SDE to obtain:
)t(dW)T,t(dtdu)u,t()T,t()T,t(dfT
t
t
0)s(dW)T,s(...)T,t(f
October - November 2014 40
HJM and the curse of dimensionality (2)
Focussing on the correlation function:
One can immediately see that for general volatility
functions this leads to an infinite-dimensional process.
t
0
2t
0
2
t
0
0
ds)T,s(ds)S,s(
ds)T,s()S,s()T,t(f),S,t(fCorr
October - November 2014 41
HJM and the curse of dimensionality (3)
Jamshidian [1991], Cheyette [1992], Babbs [1993],
Ritchen and Subsankramanian [1993] and Carverhill
[1994], see also Andreasen [2005] for more
background, noticed that separable volatility functions
will lead to a Markovian model in a manageable
number of variables. In a one factor setting, this is
achieved by:
Then the stochastic integral immediately becomes:
t
0
t
0)u(dW)u(h)T(g)u(dW)T,u(
)t(h)T(g)T,t(
October - November 2014 42
HJM and the curse of dimensionality (4)
The correlation function then becomes:
which immediately reduces the dimensionality, and
makes the model amenable to a tree and finite
difference implementation, see also Andreasen [2005].
One can show that we can retrieve the Hull-White
model exactly if we choose the following separable
volatility function in the one factor setup:
1)T,t(f),S,t(fCorr0
)tT(ae)T,t(
October - November 2014 43
HJM – pros and cons
If we return to the initial goals of Heath, Jarrow and
Morton, we can see that they realised them:
The initial term structure can be fitted exactly;
More realistic (multi-dimensional) models of the
term structure can be constructed.
This can also be achieved within short rate models,
though within HJM we only have to worry about
volatilities (and correlations).
At the end of the day, however, neither short rates nor
instantaneous forward rates are market rates – this led
to the development of the so-called market models.
October - November 2014 45
Forward measures
So far the usual setup has been to price derivatives in
the risk-neutral measure Q, associated with the bank
account as the numeraire asset (we assume B(0) = 1):
The value at time t, (t), of a claim X at time T in this
measure is:
dt)t(B)t(r)t(dB
)T(B
X)t(B)t( t
QE
October - November 2014 46
Forward measures (2)
If in particular we choose X = 1 and t = 0, we obtain:
Since P(0,T) B(T) > 0, we can define an equivalent
probability measure QT Q as follows:
)T(B)T,0(P
)T,T(P)0(B1
)T(B
1)0(B)T,0(P 00
QQ EE
)T(B)T,0(P
1T
dQ
dQ
October - November 2014 47
Forward measures (3)
For t T we have:
Assume we are now in the HJM framework:
)t(B)T,0(P
)T,t(P
)T(B)T,0(P
1t
T
t
T
t
Q
Q
E
dQ
dQE
dQ
dQ
)t(dW)T,t(vdt)t(r)T,t(P
)T,t(dP
October - November 2014 48
Forward measures (4)
Hence we can write:
The Cameron-Martin-Girsanov theorem implies that:
is a Brownian motion under the T-forward measure.
W)T,(v
)u(dW)T,u(vdu)T,u(vexp
t
T
t
T
t
2
21
T
t
dQ
dQ
t
0
T ds)T,s(v)t(W)t(W �
October - November 2014 49
Forward measures (5)
Now consider the ratio of two bonds: P(t,S) / P(t,T). In
the T-forward measure this is a martingale, since it is
the ratio of a tradeable asset, divided by the numeraire.
In formulae, we have:
Q
QE
Q
QEE QQQ
d
d/
)T,t(P
)S,t(P
d
d
)T,t(P
)S,t(P T
u
T
uu
T
)t(B)T,0(P
)T,t(P/
)t(B)T,0(P
)S,t(Puu
QQ EE
)T,u(P
)s,u(P
October - November 2014 50
Forward measures (6)
In the HJM framework, we have:
where:
and WT is a Brownian motion under the T-forward
measure.
T
T,St W)T,0(P
)S,0(P
)T,t(P
)S,t(P
)T,t(v)S,t(v)t(T,S
October - November 2014 51
Forward measures (7)
The T- and S-forward measures are related by:
and WT is a Brownian motion under the T-forward
measure.
T
T,St
T
S
T
S
W
)S,0(P)T,t(P
)T,0(P)S,t(P
)T,t(P
)t(B)T,0(P
)t(B)S,0(P
)S,t(P
ttt
dQ
dQ
dQ
dQ
dQ
dQ
October - November 2014 52
Expectation hypothesis (2)
Changing measure we obtain:
Since we defined: , we obtain:
Hence, f(t, T) is a T-forward martingale.
t
0
Tt
0
t
0
T
s
)s(dW)T,s(ds)T,s(v)T,s(
dsdu)u,s()T,s()T,0(f)T,t(f
�
�
T
tdu)u,t()T,t(v
t
0
T )s(dW)T,s()T,0(f)T,t(f
October - November 2014 53
Forward measures (8)
Moving back to the original claim X:
Using the change of measure, we find:
)T(B
X)t(B)t( t
QE
X)T,t(P
d
d/
)T(B
X)t(B
d
d)t(
T
TT
t
TtTt
Q
E
Q
QE
Q
QE
October - November 2014 54
Forward measures (9)
This approach might greatly simplify our pricing
formula, since we now do not have to worry about the
covariance between X and the bank account B(T). The
change of measure effectively takes care of this for us.
This result holds in general for any valid numeraire N:
It makes sense to choose that numeraire that will allow
us to calculate the pricing formula as efficiently as
possible.
)T(N
X
)t(N
)t(t
NE
October - November 2014 55
Forward measures (10)
Kanō Jigorō:
In randori we learn to employ
the principle of maximum
efficiency even when we could
easily overpower an opponent.
Damien Hirst:
What I really like is minimum
effort for maximum efficiency.
October - November 2014 56
Expectation hypothesis
Back to the T-forward measure – we can show that the
expectation hypothesis holds:
The long-term rate is determined by the short-term
rate, plus a risk premium
For the proof within the HJM setting, we need the
following two elements:
t
0
T ds)T,s(v)t(W)t(W �
t
0
t
0
T
s)s(dW)T,s(dsdu)u,s()T,s()T,0(f)T,t(f �
October - November 2014 57
Expectation hypothesis (3)
Moreover, since:
we obtain the following version of the expectation
hypothesis:
T
0
T )s(dW)T,s()T,0(f)T,T(f)T(r
)T,t(f
)s(dW)T,s()T,0(f)T(rt
0
T
t
T
QE
October - November 2014 58
Bond option pricing
Let us consider a European option on a bond maturing
at time S. The maturity of the option is T < S, and
strike price K. The price of this option is:
Utilising the T-forward measure, this simplifies to:
)T(B
K)S,T(P)t(B)t( t
QE
K)S,T(P)T,t(P)t(T
t
QE
October - November 2014 59
Bond option pricing (2)
If P(T,S) would be lognormal in the T-forward
measure, this would lead to a Black-Scholes like
formula. Actually, recall Black’s ’76 formula. If:
X is lognormally distributed
and
Then the forward (undiscounted) option price equals:
)d(KN)d(FNKX 21t
E
FXt E )tT()X(lnVar 2
t
tT
)tT(K/Flnd
2
21
2,1
October - November 2014 60
Bond option pricing (3)
From the discussion on forward measures, we know:
Furthermore, in the HJM framework we have:
If we therefore assume that (t, T) is a vector with
non-stochastic entries (i.e. only time-dependent
volatilities), we have a lognormal bond ratio.
)T,t(P
)S,t(P
)T,T(P
)S,T(PT
t
QE
T
T,St W)T,0(P
)S,0(P
)T,t(P
)S,t(P
October - November 2014 61
Bond option pricing (4)
The variance of the logarithm of the bond ratio is:
leading to the following (forward) option price:
T
t
2
T,S
T
T,STt du)u(WlnVar
)d(KN)d(N)T,t(P
)S,t(PK)S,T(P 21t
T
QE
T
t
2
T,S
T
t
2
T,S21
)T,t(KP
)S,t(P
2,1
du)u(
du)u(lnd
October - November 2014 62
Bond option pricing (5)
Consider the Vašíček model:
From exercise 6.6 we know that we have:
leading to the result that we can easily apply Black’s
formula to the pricing of bond options within this
model.
)t(dWdt)r)t(r(a)t(rd
)t(dWedt)T,t()T,t(fd )tT(a
October - November 2014 63
Bond option pricing (6)
Let us consider the pricing of ATM caplets in the no-
arbitrage Hull-White model, with constant volatilities.
The Black price of a caplet is given by (when IV is
the quoted vol):
when the caplet is ATM, K = Li(0), which yields:
iIV
i2IV2
1i
iIV
i2IV2
1i
T
TK/)0(Lln
T
TK/)0(Lln
i1ii KNN)0(L)T,0(P
1)T(N2)T,0(P)T,0(P
)T(N)T(N)0(L)T,0(P
iIV21
1ii
iIV21
iIV21
i1ii
October - November 2014 64
Bond option pricing (7)
In the no-arbitrage Hull-White model the price of a
zero-coupon bond put option equals:
where:
where .
)d(NdKN)T,0(P
)T,T(PK)T,0(P
1)T,0(P
)T,0(P
2i
1ii0i
i
1i
iT
QE
)T,a2(B)TT,a(B)T,T(PlnVarT
Tddd
i
2
i1i
2
1ii
T
0i
2
HW
iHW12T
T))T,0(KP/()T,0(Pln
1
i
iHW
i2HW2
1i1i
a/)e1()t,a(B at
October - November 2014 65
Bond option pricing (8)
To price a caplet with strike K, we end up with:
and for an ATM caplet K = Li(0). Note that:
so that the price of a caplet ends up being:
)T,T(P)T,0(P)K1( 1iiK1
1T
0ii i
iE
1)T,0(P
)T,0(P
)T,0(P
)T,0(P
)T,0(P
))0(L1)(T,0(P
1i
i
i
1i
i
ii1i
1)T(N2)T,0(P
)T(NTN)T,0(P
iHW21
i
iHW21
iHW21
i
October - November 2014 66
Bond option pricing (9)
Equating the Black price and the Hull-White price:
Now, for small values of x we have:
which leads to the following approximation:
1)T(N2)T,0(P
1)T(N2)T,0(P)T,0(P
iHW21
i
iIV21
1ii
2
x21)x(N
HW
ii
iiIV
)0(L
)0(L1
October - November 2014 67
Bond option pricing (10)
Remember what we are plotting when we are plotting
the ATM caplet volatility curve:
The Black vol for the ith caplet (on Li, which
sets at Ti and pays at Ti+1), as a function of Ti;
For a fixed tenor, i.e. Ti+1 – Ti = constant (3m, 6m).
The Hull-White vol (HW) is equal to:
)T,a2(B)TT,a(B
)T,T(PlnVar
iT1
i1i
1ii
T
0T1
HW
i
i
i
Fixed
October - November 2014 68
Bond option pricing (11)
An example of a typical graph of ATM caplet vols is
the following (23 July 1999, USD):
October - November 2014 69
Bond option pricing (12)
If we look at at-the-money caplet vols we can clearly
see a hump at intermediate maturities. This means that
the implied volatility of intermediate interest rates is
higher. Why?
One possible (plausible) reason is:
Central banks control the short-term rates;
Traders determine the intermediate maturities, who
are myopic, hereby causing volatility;
Mean-reversion for longer-term rates pulls down
the volatilities of longer-term rates.
October - November 2014 70
Bond option pricing (13)
What we end up with is:
The first term is decreasing when the term structure is
increasing, and vice versa. The second is a purely
decreasing function in Ti (when a > 0). So when the
forward term structure is increasing (which usually is
the case), we can only have a decreasing ATM caplet
volatility structure. So the one-factor Vašíček/Hull-
White model is not rich enough. Note: when a = 0, the
volatilities are flat. So mean reversion is important!
)e1()0(L
)0(L1)T( i
i
aT2
aT21
ii
iiiIV
October - November 2014 72
A little bit of history
When it came to pricing bond options, Black’s 1976
model was considered. There are two obvious
problems that can be signalled immediately:
deterministic discounting: perhaps this is second-
order because the payoff and the discounting cancel
out to some extent?
lognormal asset values: for a bond this means that
the bond price may be larger than the face value, i.e.
negative yields are possible.
October - November 2014 73
A little bit of history (2)
There is however another problem: the pull-to-par
problem. Initially the bond price is known, then the
uncertainty increases gradually. However, at
maturity we know the value of the bond again.
October - November 2014 74
A little bit of history (3)
People were quite accustomed to assume that
underlyings were lognormally distributed, and
switched to using rates (LIBORs, swap rates) or yields
as the underlying assets, and assuming these were
lognormal. This got rid of two problems:
The pull-to-par problem;
Bonds were now no longer able to be larger than
their face value;
But rates and yields are not tradables…
October - November 2014 75
A little bit of history (4)
Nevertheless, people started using Black’s formula for
rates and yields. Certainly until the ’90s, people were
using Black’s formula as it was convenient, but it was
regarded as being theoretically unsound. Today, caps
and swaptions are still quoted using Black ’76.
First of all, remember that a forward LIBOR rate that
has Ti as its reset date, and Ti+1 as its payment date,
with daycount fraction i, is defined as:
1)T,t(P
)T,t(P1)t(L
1i
i
i
i
October - November 2014 76
A little bit of history (5)
The price of a caplet is, using Black’s formula:
Price of a caplet:
where i is the volatility of the ith LIBOR.
Secondly, remember that a swap rate is defined as:
tTddtT
)tT(K/)t(Llnd
)d(KN)d(N)t(L)T,t(P
i
2
i12
ii
i
2
i21
i1
21i1ii
1
i 1ii
,
)T,t(P
)T,t(P)T,t(P)t(S
October - November 2014 77
A little bit of history (6)
Price of a swaption:
Here we use the payoff of a payer swaption as a call
option on the swaprate:
We can see this as a security paying (S,(T)-K)+ at
times T+1, …, T. Black’s ’76 price is then:
K)T(S)T,T(P ,
1
i 1ii
tTddtT
)tT(K/)t(Slnd
)d(KN)d(N)t(S)T,t(P
,12
,
2
,21
,
1
1
i 21,1ii
October - November 2014 78
A little bit of history (7)
Example of EUR ATM swaption vols at 05-04-2013 1
YR
2 Y
R
3 Y
R
4 Y
R
5 Y
R
6 Y
R
7 Y
R
8 Y
R
9 Y
R
10
YR
12
YR
15
YR
20
YR
25
YR
30
YR
0.00%
20.00%
40.00%
60.00%
80.00%
100.00%
1 M
O
6 M
O
1 Y
R
3 Y
R
5 Y
R
7 Y
R
9 Y
R
12
YR
20
YR
30
YR
Swap tenor
Bla
ck v
ola
tilit
y (%
)
Expiry
80.00%-100.00%
60.00%-80.00%
40.00%-60.00%
20.00%-40.00%
0.00%-20.00%
October - November 2014 79
A little bit of history (8)
How about a theoretical justification of Black’s
formula? Consider a caplet. Its price at time t, using
the Ti+1-forward measure, can be written as:
One can show that:
so that the only assumption that is effectively being
made, is that the forward LIBOR rate is lognormally
distributed.
K)T(L)T,t(P ii
1i
t1ii E
)]T(L[)t(L ii
1i
ti
E
October - November 2014 80
A little bit of history (9)
For the swaption, we again noted that the payoff of a
payer swaption is, at the first reset date T:
The PVBP is a portfolio of zero-coupon bonds, and
hence is a tradable asset with a positive value. We can
use it as the numéraire asset, to find:
where we are working under the (,)-swap measure.
K)T(S)T,T(P ,
1
i 1ii
K)T(S)T,t(P ,
,
t
1
i 1ii E
October - November 2014 81
A little bit of history (10)
We know now that in the market:
Caplets are quoted using Black’s price, which in
effect assumes that Li(Ti) is a lognormal martingale
under the Ti+1-forward measure;
Swaptions are quoted using Black’s price, which in
effect assumes that S,(T) is a lognormal
martingale under the (,)-swap measure.
These are all models focussing on one quantity. Can
we however build up a consistent term structure model
that implies the above rates are all lognormal (in their
own measure)?
October - November 2014 82
Market models
The previous modelling approaches were all in terms
of rates that are not directly observable in the market:
Short rates (short rate models);
Instantaneous forward rates (HJM);
The rates which are directly observable, are:
LIBORs (also the underlyings in caps/floors);
Swap rates (also the underlyings in swaptions).
The market models focus on modelling these.
October - November 2014 83
Market models (2)
In 1997, Brace, Gątarek and Musiela, Jamshidian, and
Miltersen, Sandmann and Sondermann came up with
the market models. We will start with the LIBOR
market model (often just LMM or BGM/J model).
Note that a forward LIBOR rate equals:
Since P(t,Ti) / P(t,Ti+1) is a ratio of tradeables, the
quantity must be a martingale under the measure
induced by the denominator, the Ti+1-bond.
1)T,t(P
)T,t(P1)t(L
1i
i
i
i
October - November 2014 84
Market models (3)
The forward Li(t) is just an affine combination of this
ratio, which immediately yields, for s ≤ t:
i.e. the forward LIBOR rate is a martingale. This also
shows that the floating leg of a swap can be priced by
discounting the “expected” cash flows.
Consider a swap that has T as its first reset date, and
T as the final payment date. It pays Li(Ti) at Ti+1, for i
= , …, -1.
)s(L)t(L ii
1i
s E
October - November 2014 85
Market models (4)
The value of the swap then becomes:
Unfortunately one cannot always split seemingly
linear payments like that into “expected” cash flows
and discounting. Consider a Libor-in-arrears payment,
where the payment is made at the reset date.
)T,t(P)T,t(P
)T,t(P)T,t(P
)T(L)T,t(P)t(V
1
i 1ii
1
i ii
1i
t1iifloat
E
October - November 2014 86
Market models (5)
Hence, it pays iLi(Ti) at time Ti. We can reinvest this
in a deposit maturing at time Ti+1, and hence its
(undiscounted) value will be equal to:
which will not only depend on forward rates, but also
on the second moment of the Libor rate, which will
depend on its volatility.
)T(L1)T(L iiiiii
1i
t E
October - November 2014 87
Market models (6)
To price caplets/floorlets, Black’s model assumed the
Libor rate had lognormal dynamics. In the Ti+1-
forward measure we therefore should have:
Contrary to the book, we use a one-dimensional
Brownian motion here, and further assume:
A reduction of dimensionality can be achieved by
reducing the rank of the correlation matrix.
)t(dW)t(L)t()t(dL 1i
iiii
dt)t()t(dW)t(dW ij
k
j
k
i
October - November 2014 88
Market models (7)
What about Lj, for j < i? Clearly, Lj(t) cannot be a
martingale under this measure. But it should be under
its “own” measure:
Let us start with j = i-1. We know:
should also be a Ti+1-martingale. Remember:
)t(dW)t(L)t()t(dL 1j
jjjj
)t(L1)t(L1)T,t(P
)T,t(P
)T,t(P
)T,t(Pii1i1i
1i
i
i
1i
y,xddyxdxyxyd
October - November 2014 89
Market models (8)
Hence:
Its drift should be zero. If dWi(t) · dWj(t) = ij(t) dt
(under any measure), we find:
)t(dL)t(dL
)t(dL)t(L1
)t(dL)t(L1)T,t(P
)T,t(Pd
ii1i1i
1i1iii
ii1i1i
1i
1i
)t(L)t(L)t()t()t(
)t(LDrift)t(L1)T,t(P
)T,t(PDrift
i1ii1ii,1ii1i
1i
1i
ii1i
1i
1i1i
October - November 2014 90
Market models (9)
Solving for the drift of Li-1(t), we find:
and we thus end up with:
We can derive the correct dynamics recursively, see
section 11.2 in the book.
)t(L)t(L1
)t(L)t()t()t()t(LDrift 1i
ii
iii1ii,1i1i
1i
)t(dW)t(
dt)t(L1
)t(L)t()t()t(
)t(L
)t(dL
1i
1i1i
ii
iii1ii,1i
1i
1i
October - November 2014 91
Market models (10)
Under the Ti+1-forward measure, we find:
So now we have a consistent no arbitrage model in
which forward LIBORs are modelled. Each forward
LIBOR is a lognormal martingale in its own forward
measure. This means:
The LMM is consistent with Black’s caplet formula.
)t(dW)t(
dt)t(L1
)t(L)t()t()t(
)t(L
)t(dL
1i
jj
i
1jkkk
kkkjjk
j
j
October - November 2014 92
Market models (11)
The derivation till here was due to Jamshidian. Brace,
Gątarek and Musiela, and Miltersen, Sandmann and
Sondermann constructed an HJM setup in which Libor
rates were lognormal. The argument goes roughly as
follows (cf. 11.1 of the book). From Lemma 7.1:
implying the following for Li(t):
)t(dW)T,t(P
)T,t(P)t(
)T,t(P
)T,t(Pd 1i
1ii
T
1i
iT,T
1i
i
)t(dW)t(L1)t(1
)T,t(P
)T,t(Pd)t(dL 1i
1ii
T
iiT,T
i1ii
ii
October - November 2014 93
Market models (12)
Now if we suppose there is an Rd-valued deterministic
function (t,T), such that:
Then we have:
One can prove that such a function exists, see Chapter
11.1 for a sketch.
)T,t()t(L1
)t(L)t( i
ii
iiT,T 1ii
)t(dW)t(L)T,t()t(dL 1iT
iii
October - November 2014 94
Market models (13)
If we take a step back, we have found a way to derive
the dynamics of all Libor rates in one forward
measure. A common choice is the terminal measure,
so that we view all Libors L0, …, LN in the
TN+1-forward measure. Since we calculate terms like:
we numerically blow up each term. Whether this is a
major issue or not depends on the total variance of
each term.
)T,t(P/)t(X 1N
October - November 2014 95
Market models (14)
Numerically it seems to be more advantageous
however to sometimes use a measure that is similar to
the risk neutral measure. We define the spot Libor
measure by defining a discrete version of the money
market account:
What is happening economically is that we are rolling
all our investments into the next maturing zero-coupon
bond.
i
1j j1j
i
*
)T,T(P
1)T(B
October - November 2014 96
Market models (15)
Essentially this boils down to:
Using the T1-forward measure from T0 to T1;
The T2-forward measure from T1 to T2;
etc.
The relevant dynamics for each Libor rate can be
derived by using the appropriate dynamics on each
interval. See also Section 11.4.
October - November 2014 97
Market models (16)
How about pricing swaptions? We know the forward
swap rate is defined as:
and it should be a martingale under the (,)-forward
swap measure.
1
j jj
1
j1j
j
j
1
i 1ii
1jj
1
j 1
i 1ii
1jj
1
i 1ii
,
)t(L)t(w
1)T,t(P
)T,t(P1
)T,t(P
)T,t(P
)T,t(P
)T,t(P)T,t(P
)T,t(P
)T,t(P)T,t(P)t(S
October - November 2014 98
Market models (17)
So we are only interested in the diffusion of the
forward swaprate. Itō’s lemma yields:
The first approach to yield something tractable (by
BGM) first assumed:
This is clearly not the case, but let us start from here.
1
j j
j
,
, )t(dL)t(L
)t(S)dt(O)t(dS
)t(w)t(L
)t(Sj
j
,
October - November 2014 99
Market models (18)
Then:
Defining j(t) = wj(t)Lj(t)/S,(t), if S, would be
lognormal, we would be able to use Black’s formula
with the following total variance:
1
j
,
jj
,
jj
,
,)t(dW)t(
)t(S
)t(L)t(w)dt(O
)t(S
)t(dS
1
k,j
T
0kjjkkj
T
0 )t(S
)t(dS
)t(S
)t(dS2,
Black
dt)t()t()t()t()t(
dtT,
,
,
,
October - November 2014 100
Market models (19)
But the weights j(t) are not deterministic, but
stochastic… what happens to them if we bump the
forward curve up/down by 25 bp?
October - November 2014 101
Market models (20)
These are the percentage differences in i · j for an
upward move of 25 bp:
October - November 2014 102
Market models (21)
The products i· j do not seem too stochastic. So, let
us “freeze” them at their initial value, and obtain:
By calculating the real S,(t)/Li(t) and doing some
additional “freezing” we can get a more accurate
approximation that works in the same way (just a
slightly different, more complicated, ). The latter can
be found in Hull, and is due to Hull and White, and
Rebonato and Jäckel.
1
k,j
T
0kjjkkj
2,
Black dt)t()t()t()0()0(T)(
October - November 2014 103
Market models (22)
Actually, it’s not bad at all…
-35
-30
-25
-20
-15
-10
-5
0
5
10
0 5 10 15 20
Maturity
Dif
fere
nce in
Bla
ck v
ol. (
bp
)
Simple Shape-corrected
October - November 2014 104
Market models (23)
So where has this got us? In the LMM:
We work with N forward LIBOR rates; forward
rates, and hence the corresponding zero-coupon
bonds are input;
Caplets/floorlets are priced via Black’s formula:
Swaptions priced via Black’s formula:
iT
0
2
ii
2i
Black dt)t(T)(
1
k,j
T
0kjjkkj
2,
Black dt)t()t()t()0()0(T)(
October - November 2014 105
Market models (24)
A calibration to ATM caps/floors and swaptions is
therefore much more direct than in other models,
where we typically have a least squares calibration
model in terms of squared differences:
much easier calibration to at-the-money products
N
1i
2
market,ielmod,ii P)(Pwmin θθ
October - November 2014 106
Market models (25)
Some challenges with the LIBOR market model:
Not a model of the full term structure; e.g. if we
model 6m LIBORs, there is no direct way to
recover 3m LIBORs, see e.g. 11.8 for a solution;
Although each LIBOR is lognormal under its own
forward measure, other LIBORs are not. This
complicates a Monte Carlo simulation, see 11.6;
LMM only allows for a perfect fit to volatilities at
one strike; how about the skew/smile?
In many products LIBORs are not (directly) the
underlying rates, but swap rates are.
October - November 2014 107
Market models (26)
To understand the latter problem, consider a
Bermudan swaption. In a Bermudan swaption we can
decide at Ti to enter into a swap with payment dates
Ti+1 through TN. The most important drivers of the
product are therefore:
Si,N for all call dates Ti and their interdependence;
European swaptions with the same strike as the
Bermudan product.
We would certainly like to calibrate exactly to these
core European swaptions.
October - November 2014 108
Market models (27)
For these purposes more generic market models have
been developed. Jamshidian introduced the most
popular of these, the (co-terminal) swap market model
(SMM), which models co-terminal swap rates. Here
each swaprate is assumed to be a lognormal
martingale under its own forward swap measure:
Derivation proceeds much the same as in the LMM.
)t(dW)t()t(S
)t(dSN,i
N,iN,i
N,i
N,i
October - November 2014 111
Recent developments (2)
Remember that the value of a floating leg equals:
Clearly this is independent of how the date strip is
structured, but does depend on start date (T) and end
date (T).
)T,t(P)T,t(P
)T,t(P)T,t(P
)T(L)T,t(P)t(V
1
i 1ii
1
i ii
1i
t1iifloat
E
October - November 2014 112
Recent developments (3)
Again - what is the value of this structure?
In our setup, it must be zero.
October - November 2014 114
Recent developments (5)
Though the basis between deposit rates and overnight
(o/n) rates of the same maturity have exploded.
October - November 2014 115
Recent developments (6)
So how can we explain this? Remember that LIBOR is
the average unsecured funding rate at which a
contributor bank can obtain unsecured funding in the
London interbank market, for a given period, in a
given currency.
What happens if bank A has a cash deposit with bank
B, and bank B defaults? Since it is unsecured funding,
bank A will have a claim on the assets of bank B, but
will have a much larger loss than a secured creditor.
Hence, there is credit risk involved – something we
ignored till now.
October - November 2014 116
Recent developments (7)
Like Mercurio [2009] let us introduce the concept of
default-free zero coupon bonds (P(t,,T)), and
defaultable zero-coupon bonds (D(t, T)). If we define
as the default time of a counterparty, and assume:
That the default process and rates are independent;
We have a constant recovery rate;
Then the price of a defaultable bond follows as:
T)R1(R)T,t(P
1)R1(R)T,t(P)T,t(D
T
t
]T[t
T
Q
EQ
October - November 2014 117
Recent developments (8)
Returning to LIBOR rates – this was defined as the
return on buying the zero-coupon bond maturing at
Ti+1, at Ti, and selling this bond at its maturity. Using
this defaultable zero-coupon bond, the return is:
1
T)R1(R
1
)T,T(P
11
1)T,T(D
11)T(L
1i
T
T1iii
1iii
ii
1i
iQ
October - November 2014 118
Recent developments (9)
To value swaps, we need forward LIBOR rates, which
typically follow from FRAs or are implied from
swaps. Contrary to what many books state (also that of
Filipović), a FRA in practice pays the following
amount at Ti:
In a default-free setting this is equivalent to receiving
the amount of the numerator at Ti+1.
)T(L1
K)T(L
iii
iii
October - November 2014 119
Recent developments (10)
Assuming the FRA itself has no counterparty risk
(more on this later), we obtain as its value:
The FRA rate K which renders the value 0, equals:
1i
T
T0i1ii
1i
T
T1ii0ii
iii
iiiti
T)R1(RK1)T,0(P)T,0(P
T)R1(R)T,T(PK11)T,0(P
)T(L1
K)T(L)T,0(P
1i
i
iT
1i
i
iT
iT
QE
QE
E
Q
Q
Q
1T)R1(R)T,0(P
)T,0(P1K
1i
T
T01i
i
i 1i
i
iT
QEQ
October - November 2014 120
Recent developments (11)
Since 0 R 1, and , we have:
and thus:
The FRA rate is larger than the default-free one.
1T0 1i
T
T1i
i
Q
1T)R1(R 1i
T
T01i
i
iT
QEQ
1)T,0(P
)T,0(P1
1T)R1(R)T,0(P
)T,0(P1K
1i
i
i
1i
T
T01i
i
i 1i
i
iT
QEQ
October - November 2014 121
Recent developments (12)
So how do we proceed? Well, most setups still ignore
the direct credit risk element, and proceed as follows:
Assume the risk-free rate is the overnight rate – we
can estimate our term structure from the overnight
swap (OIS) market;
Strip 3m forward rates from FRAs and 3m swaps,
where the discount rate is the OIS curve;
Strip 6m forward rates from FRAs and 6m swaps,
where the discount rate is the OIS curve;
etc.
October - November 2014 122
Recent developments (13)
Overnight swaps by the way are slightly different than
regular swaps, in that all o/n rates are compounded up
and paid out in one go. In the larger markets this
happens annually. This does not complicate the
valuation greatly.
October - November 2014 123
Recent developments (14)
Clearly there is more to it – if we only pay the
compounded o/n rate on an annual basis, are we not
exposed to a similar credit risk? Yes, though most
trades in the market these days are fully collateralised.
How does collateralisation work?
Let party A and B enter into a par swap;
Initially this has zero market value;
At time t, the contract has a value of V(t); suppose
this is positive to party A;
Party A asks party B to post V(t) of cash or bonds;
October - November 2014 124
Recent developments (15)
How does collateralisation work (continued)?
Party B posts V(t) worth of collateral;
The net exposure is now 0;
This process continues till the end of the contract
(and is typically netted over all outstanding trades).
An interest rate is reimbursed over the posted
collateral – this depends on the credit support annex
(CSA) under which this contract falls, but is often
equal to the o/n rate;
October - November 2014 125
Recent developments (16)
Schematically the following roughly happens:
V(t)
V(t) V(t+)-V(t)
V(t+)
V(t+2)
V(t+2)-V(t+ )
…
Value of the contract
Collateral to be posted
Interest rate accrued over collateral
October - November 2014 126
Recent developments (17)
Should the collateral rate not feature in the pricing
formula? If we have to set some money aside that we
cannot use for other purposes, the cost of this has to be
taken into account. Since the collateral rate before the
credit crunch was close to Libor, and banks were able
to fund themselves at Libor, this was neglected.
How can we incorporate this in a mathematical
model? We refer the interested reader to Piterbarg
[2012], a paper called “Cooking with collateral”, and
only sketch the very rough thought process.
October - November 2014 127
Recent developments (18)
Suppose we have a collateral account C(t), on which
we earn a (possibly stochastic) rate c(t):
We assume that we can replicate this return in the
market. Now suppose we are at time t, and are pricing
a derivative that pays 1 unit of cash at t + dt. By
purchasing:
we can replicate this cash flow.
dt)t(C)t(c)t(dC
)t(cdtexp
October - November 2014 128
Recent developments (19)
Piterbarg has made this more rigorous, and eventually
ends up at the result that any payoff can be priced as:
where Q is the measure under which the collateral
account is used as the numeraire.
This ties in nicely with Mercurio, in whose work the
OIS curve was used as a proxy for the risk-free rate. It
is not because it is almost risk-free that this is correct,
it is because all quotes are based on fully (cash)
collateralised swaps.
)T(Vdt)t(cexp)t(V i
T
tti
QE
October - November 2014 129
Recent developments (20)
What if there is the option to choose which collateral
one can post? Think of EUR cash, or GBP cash?
It goes too far to describe the full model here, but it
suffices to say that the problem then becomes a lot
more complicated. The current value of such a
structure can be expressed as:
where Pd is a domestic zero-coupon bond, cf and cd the
foreign and domestic collateral rates, and rd,f the drift
of the foreign exchange rate.
)T(Ve)T,0(P)t(V
Tt df,df dt0),t(c)t(r)t(cmaxT,d
td E
October - November 2014 130
Recent developments (21)
This means that a simple EUR 6m swap, collateralised
with EUR and GBP cash (earning respectively the
Eonia and Sonia rates on collateral), depends on:
6m swap rates
Eonia swap rates
Sonia swap rates
EUR/GBP FX forward rates and xccy basis swaps
Potentially EUR and GBP swaptions
Pricing a simple swap has become complicated!
October - November 2014 131
Recent developments (22)
Other recent developments:
What about the option to post EUR or GBP cash
collateral?
CVA / DVA – charges due to expected loss on
portfolio in case of default of counterparty or of
oneself;
Move to central clearing or exchanges – this also
affects the pricing;
There are many, many new challenges in derivative
pricing!
October - November 2014 134
Affine models
We will now turn to affine models, a class of models
which is very popular thanks to its modelling
flexibility and analytical tractability.
Duffie and Kan [1996] pioneered these models in a
term structure context, and were subsequently treated
in more detail in Duffie, Pan and Singleton [2000] and
Duffie, Filipović and Schachermayer [2003].
Prominent members of this class are Black-Scholes
[1973], the Vašíček [1977] and Cox-Ingersoll-Ross
[1985] interest rate models, as well as the Heston
stochastic-volatility model [1993].
October - November 2014 135
Affine models (2)
The material will be Chapter 2 of Lord [2008], and we
will assume that the processes do not contain a Lévy
component. For the interested reader, Filipović’
Chapter 10 also focuses on affine processes.
Let X be a Markov process in D Rn:
where W is a standard Brownian motion on Rn, and:
: D n
: D nn
)t(d)t(dt)t()t(d WXσXμX
October - November 2014 136
The money market or bank account is defined as:
Now, the process X is called affine, if and only if:
, for (m0, m1) n nn
,
for (0, 1) nn nnn
where and 1x is interpreted
as a vector inner product
, for (r0, r1) n
Affine models (3)
dt)t(M)t(r)t(Md X
xmmx 10)(
xxx 10
n
1i ii10
T x)()(
n1111 ΣΣΣ
xrxT
10r)(r
October - November 2014 137
Heston stochastic volatility model
In Heston’s [1993] stochastic volatility model, the
stochastic volatility is modelled by the same square
root process as is used in the Cox-Ingersoll-Ross
interest rate model:
The drift is deterministic, and the variance process
v(t) follows a mean-reverting square root process, with
the rate of mean reversion, the long-term level of
variance, and the volatility of variance.
)t(dW)t(vωdtθ)t(vκ)t(vd
)t(dW)t(S)t(vdt)t(Sμ)t(Sd
v
S
October - November 2014 138
Heston stochastic volatility model (2)
Moreover, both Brownian motions are correlated:
Clearly, (S, v) is not an affine process. We can
however switch to (x, v), where x = ln S:
dtρ)t(dW)t(dW vS
)t(dW)t(v)ρ1(ω
)t(Wd)t(vρωdt)θ)t(v(κ)t(dv
)t(Wd)t(vdt)t(vμ)t(xd
2
2
1
121
October - November 2014 139
Heston stochastic volatility model (3)
Using the notation from before, the drift can be
identified as:
The variance-covariance matrix can be seen to equal:
0
021
10 mm
vωvρω
vρωv
v)ρ1(ωvρω
0v
v)ρ1(ωvρω
0v)()(
2
T
22
Txx
October - November 2014 140
Heston stochastic volatility model (4)
Leading to:
Since the interest rate is taken to be constant, this
finishes the classification of the Heston model as an
affine one.
210
1
00
00
00
00
October - November 2014 141
Affine models (4)
Duffie, Pan and Singleton [2000] studied the extended
characteristic function of affine processes, which is
defined as:
where u Cn. A quick recap on characteristic
functions will follow later, for now it suffices to
remember that the characteristic function determines
the full probability distribution.
X(T)iuexpT)P(t,
X(T)iuexpM(t)X(t)T,t,u,
TT
t
T
M(T)1Q
t
E
E
October - November 2014 142
Affine models (5)
Contrary to the real-valued characteristic function, the
extended version is not always defined. We have:
This equals the –Im(u)th moment, and unfortunately
not every moment is guaranteed to be finite. Hence we
define the strip of regularity as:
In the remainder we will drop the term “extended” and
just use the term characteristic function.
)Im(i]|e|[|]e[||)(| T)iT)i TT
uuX(uX(u EE
)Im(inuuX C
October - November 2014 143
Affine models (6)
The question we are going to pose now is how we can
find the characteristic function of an affine process.
The characteristic function, as it can be viewed as a
special type of payoff, should satisfy the Feynman-
Kac formula:
subject to the obvious boundary condition:
rtr
t
T2
21
T
σσxx
μx T
)T(iexp)t(,T,T, TXuXu
October - November 2014 144
Affine models (7)
Under certain technical regularity conditions, Duffie et
al. [2000] show that the solution is given by:
where A and B are respectively R and Rn-valued
functions. From the ‘ansatz’ we find:
xuBuxuT)T,t,()T,t,(Aexp),T,t,(
T
T
2T
td
d
td
Ad
tBB
xxB
xx
B
October - November 2014 145
Affine models (8)
Using tr(AB) = tr(BA) if AB and BA are well-defined:
Inserting this into the PDE, and dividing by we find:
xBBBBxBBxx
1
T
0
T
10
TT
T
2
)(trtr
xrxBBBB
xmmBxB
T
101
T
0
T
21
10
T
T
r
)(td
d
td
Ad
October - November 2014 146
Affine models (9)
Since the PDE should hold for every x, it must in
particular hold for x = 0, x = (1, 0, …, 0)T, etc. This
leads to the following system of ODEs:
where we denote BT1B for the vector in Cn with ith
element equal to BT1iB. The boundary conditions
carry over: A(u, T, T) = 0 and B(u, T, T) = iu.
11
T
21T
1
00
T
21
0
T
td
d
rtd
Ad
rBBBmB
BBmB
October - November 2014 147
Affine models (10)
For some models the system of ODEs (a system of
Ricatti equations) can be solved explicitly, though in
general they have to be solved numerically. This can
be done quite efficiently.
We return to the technical conditions. If we define:
then the conditions formulated by Duffie et al. are:
)t(M
),T,t,()t(
xu ))t(()t()t()t( T
XσBη
)t(E
T
0
T dt)t()t( ηηE
October - November 2014 148
Affine models (11)
In any model, the undiscounted (forward) value of a
call price on an asset S can be written as:
Very generally, we can write the call price as:
where S indicates a probability in the measure
associated with using the asset as the numeraire, and P
indicates the T-forward measure. Finally, F(t, T) is the
forward price of the underlying asset at time T, as seen
from time t – equal to: .
])K)T(S([T,K),t(SC T
t
E
K)T(SKK)T(S)T,t(FT,K),t(SC PS
)]T(S[)T,t(F T
tE
October - November 2014 149
Affine models (12)
This general form also holds in the Black-Scholes
model, where we can explicitly calculate the two
probabilities as N(d1) and N(d2).
A probability can be calculated by means of Fourier
inversion, an approach dating back to Lévy [1925], see
also Gurland [1948] and Gil-Pelaez [1951]:
where k = ln K, and is the characteristic function of
the logarithm of the underlying asset.
duiu
)u(eReK)T(S
0
iuk
π1
21
P
October - November 2014 150
To obtain the same probability in the stock-price
measure, we can use a measure transform:
In a model without dividends, the scaling is exactly
the inverse of the forward, i.e. 1 / (-i), so that the
full term evaluates to (u-i) / (-i).
Affine models (13)
))T(Sln)iu(iexp()t(S
)T,t(P
))T(Slniuexp()T,t(P/)T,T(P
)t(S/)T(S
))T(Slniuexp(d
d))T(Slniuexp(
t
t
tt
P
P
PS
E
E
P
SEE
October - November 2014 151
The S-probability can therefore be expressed as:
Though this is not the best way to value derivatives
via Fourier inversion (see Lord and Kahl [2007]), it
at least shows that if we can calculate the T-forward
characteristic function in closed-form (which we can
for affine models), we can evaluate European option
prices via numerical integration.
Affine models (14)
du)i(iu
)iu(eReK)T(S
0
iuk
π1
21
S