SF2975 Financial Derivatives - Exercise session 6

Chapter 15: Change of numeraire1

A numeraire is essentially a reference unit of value or account used to measure financial in-
struments By the FTAP we know that the risk neutral martingale measure Q, with the money
account B as numeraire, has the property of martingalizing all processes of the form 𝑆𝑡 /𝐵𝑡 where
S is the arbitrage free price process of any (non-dividend-paying) traded asset.
In many concrete situations the computational work needed for the determination of arbitrage
free prices can be drastically reduced by a clever change of numeraire, from the bank acount B
to some other numeraire S, and we will now analyze such change.

Consider an arbitrage free market model with asset prices 𝑆 0, 𝑆 1, ..., 𝑆𝑛 with 𝑆 0 > 0.

Recall the First Fundamental Theorem and the corresponding pricing formula:

Theorem The following hold with 𝑆 0 as the numeraire:

• The market model is free of arbitrage if and only if there exists a martingale measure,
𝑄 0 ∼ 𝑃 such that the processes
𝑆𝑡0 𝑆𝑡1 𝑆𝑡𝑛
, , . . .
𝑆𝑡0 𝑆𝑡0 𝑆𝑡0
are (local) martingales under 𝑄 0 .
• An arbitrage free price system for all T-claims X is given by the formula
" #
𝑋
Π𝑡 [𝑋 ] = 𝑆𝑡0 𝐸 0 0 |F𝑡 (1)
𝑆𝑇

where 𝐸 0 denotes expectation under 𝑄 0 .

1 Bjork, T. (2020). Arbitrage Theory in Continuous Time. Oxford University Press, USA.

1
But how to change from one choice of numeraire to another? i.e. how to determine the appro-
priate Girsanov transformation?

Suppose that we want to change the numeraire from 𝑆 0 to, say, 𝑆 1 . Our immediate problem is
to find a martingale measure 𝑄 1 corresponding to the numeraire 𝑆 1 , while keeping the fixed
price system defined by (1) and we also want to find the appropriate Girsanov transformation
which will take us from 𝑄 0 to 𝑄 1 .

Assume that 𝑄 0 is a martingale measure for the numeraire 𝑆 0 (on F𝑇 ) and assume that 𝑆 1 is
a positive asset price process such that 𝑆𝑡1 /𝑆𝑡0 is a true 𝑄 0 -martingale (and not just a local one).
Define 𝑄 1 on F𝑇 by the likelihood process

𝑑𝑄 1 𝑆𝑡1 𝑆 00
= 0 · 1, 0 ≤ 𝑡 ≤ 𝑇.
𝑑𝑄 0 F𝑡 𝑆𝑡 𝑆 0

Then 𝑄 1 is a martingale measure for 𝑆 1 , and it generates the same price system as 𝑄 0 .
In the case in which 𝑆 0 = 𝐵 and consider a numeraire 𝑆 we get

𝑑𝑄 𝑆 𝑆𝑡 1
= · , 0 ≤ 𝑡 ≤ 𝑇. (2)
𝑑𝑄 0 F𝑡 𝐵𝑡 𝑆 0
Example 15.7 (An Asset-or-Nothing Option)
Consider a standard Black-Scholes model with the usual P-dynamics:

𝑑𝑆𝑡 = 𝜇𝑆𝑡 𝑑𝑡 + 𝜎𝑆𝑡 𝑑𝑊𝑡 , 𝑑𝐵𝑡 = 𝑟𝐵𝑡 𝑑𝑡 .

The claim 𝑋 under consideration is an "asset-or-nothing", defined by

𝑋 = 𝑆𝑇 · I{𝑆𝑇 > 𝐾 },

where the strike price 𝐾 is a positive constant, and I is the indicator function. This option gives
one unit of the underlying asset if 𝑆𝑇 > 𝐾, and nothing if 𝑆𝑇 ≤ 𝐾.
We restrict ourselves to computing the price at time 𝑡 = 0. The form of 𝑋 clearly suggests
that we should use the underlying asset 𝑆 as the numeraire, and from by the Girsanov change
of measure we have
1 Q 1 S 𝑑𝑄 0  
E [𝑋 ] = E [𝑋 𝑆 ] = 𝑆 0 E [I{𝑆𝑇 > 𝐾 }] = 𝑆 0 1 − E [I{𝑆𝑇 < 𝐾 }] = 𝑆 0 − 𝑆 0 QS (𝑆𝑇 < 𝐾).
S S
𝐵𝑡 𝐵𝑡 𝑑𝑄

It thus remains to compute QS (𝑆𝑇 < 𝐾), and to this end, we use the Radon-Nykodim derivative
in (2) given by the Girsanov theorem. The Q-dynamics of 𝑆 are, as always,

𝑑𝑆𝑡 = 𝑟𝑆𝑡 𝑑𝑡 + 𝜎𝑆𝑡 𝑑𝑊𝑡Q,

2
substituing in (2), we get

𝑆𝑡 1 𝑆 0 exp((𝑟 − 21 𝜎 2 )𝑡 + 𝜎𝑊𝑡 ) 𝐵 0 
1 2

· = · = exp (− 2 𝜎 )𝑡 + 𝜎𝑊𝑡 .
𝐵𝑡 𝑆 0 𝐵 0𝑒 𝑟𝑡 𝑆0
That is, the Girsanov Theorem implies that we can write

𝑑𝑊𝑡Q = 𝜎𝑑𝑡 + 𝑑𝑊𝑡S,

where 𝑊𝑡S is a QS -Wiener process. We thus obtain the QS -dynamics of 𝑆 as

𝑑𝑆𝑡 = (𝑟 + 𝜎 2 )𝑆𝑡 𝑑𝑡 + 𝜎𝑆𝑡 𝑑𝑊𝑡S .

This is a geometric Brownian motion (GBM), so we have

𝜎2
 
𝑆𝑇 = 𝑆 0 exp (𝑟 + )𝑇 + 𝜎𝑊𝑇 ,
S
2
which we can write as
𝜎2
𝑆𝑇 = 𝑆 0𝑒 (𝑟 + 2 )𝑇 +𝜎𝑊𝑇 ,
S

2
 
where 𝑌 has the distribution N (𝑟 + 𝜎2 )𝑇 , 𝜎 2𝑇 under QS . A standard calculation shows that

QS (𝑆𝑇 < 𝐾) = 𝑁 (𝑑 1 ),

where 𝑁 is the cumulative distribution function (CDF) of a standard normal distribution, and
2
 
ln(𝐾/𝑆 0 ) − 𝑟 + 𝜎2 𝑇
𝑑1 = √ .
𝜎 𝑇
We have thus computed the price at 𝑡 = 0, and for a general 𝑡, we simply replace 𝑇 by 𝑇 − 𝑡.
The price of the asset-or-nothing call option is given by Π(𝑋 ) = 𝑆 0 (1 − 𝑁 (𝑑)).
Example 15.8 (Linearly Homogeneous Contracts)
A typical example when a change of numeraire is useful occurs when dealing with derivatives
defined in terms of several underlying assets. Assume, for example, that we are given two asset
prices 𝑆 1 and 𝑆 2 , and that the contract 𝑋 to be priced is of the form 𝑋 = Φ(𝑆 1, 𝑆 2 ), where Φ is a
given linearly homogeneous function, i.e.,

Φ(𝜆𝑥, 𝜆𝑦) = 𝜆Φ(𝑥, 𝑦)

for all 𝑥, 𝑦, and all 𝜆 > 0. Using the standard risk-neutral machinery with 𝐵 as the numeraire,
and denoting the risk-neutral martingale measure by Q, we would have to compute the price
as
1
Π𝑡 [𝑋 ] = EQ Φ(𝑆𝑇1 , 𝑆𝑇2 )|F𝑡 ,
 
𝐵𝑡

3
which essentially amounts to the calculation of a triple integral. If we instead use 𝑆 1 as the
numeraire, with martingale measure Q1 , we have

Π𝑡 [𝑋 ] = 𝑆𝑡1 EQ1 Φ(𝑆𝑇1 , 𝑆𝑇2 )/𝑆𝑇1 |F𝑡 = 𝑆 1 EQ1 [𝜑 (𝑍𝑇 )|F𝑡 ]

 

𝑆2
where 𝜑 (𝑧) = Φ(1, 𝑧) and 𝑍𝑡 = 𝑆𝑡1 . In this formula, we note that the factor 𝑆 1 is the price of the
𝑡
traded asset 𝑆 1 at time 𝑡, so this quantity does not have to be computed—it can be directly ob-
served on the market. Thus, the computational work is reduced to computing a single integral.
We also note the important fact that in the 𝑍 -economy we have a zero short rate. We now go
on to study a concrete example in this setting.
Example 15.9 (An Exchange Option) As an example of the reasoning above, assume that
we have two stocks, 𝑆 1 and 𝑆 2 , with price processes of the following form under the objective
probability measure P:

𝑑𝑆 1 = 𝜇 1𝑆 1𝑑𝑡 + 𝑆 1𝜎1 · 𝑑𝑊𝑡 , 𝑑𝑆 2 = 𝜇 2𝑆 2𝑑𝑡 + 𝑆 2𝜎2 · 𝑑𝑊𝑡 ,

where 𝜇 1, 𝜇2 ∈ R are deterministic real numbers, and 𝜎1, 𝜎2 ∈ R2 are deterministic row vectors.
𝑊𝑡 = (𝑊𝑡1,𝑊𝑡2 ) is a two-dimensional standard Wiener process under P, and we assume absence
of arbitrage.
The 𝑇 -claim to be priced is an exchange option, which gives the holder the right, but not
the obligation, to exchange one 𝑆 2 share for one 𝑆 1 share at time 𝑇 . Formally, this means that
the claim is given by
𝑋 = max(𝑆 2 − 𝑆 1, 0),
and we note that we have a linearly homogeneous contract function. It is thus natural to use
one of the assets as the numeraire, and we choose 𝑆 1 .
Using homogeneity, the price is given by
1
Π𝑡 [𝑋 ] = 𝑆𝑡1 EQ1 Φ(𝑆𝑇1 , 𝑆𝑇2 )/𝑆𝑇1 |F𝑡 = 𝑆𝑡1 EQ [max(𝑍𝑇 − 1, 0)|F𝑡 ],
 

𝑆2 1
with 𝑍𝑇 = 𝑆𝑇1 as before, and EQ denoting expectation under Q1 . Note that the upper case
𝑇
indices for 𝑆 2 , and 𝑆 1 are not powers but merely indices. We now see that the expectation above
is in fact the value of a European call option on 𝑍𝑇 , with strike price 𝐾 = 1 in the 𝑍 -economy,
where (as in all normalized asset price systems) the short rate 𝑟 = 0.
We need to compute the Q1 -dynamics of 𝑍𝑇 , but let’s first see its dynamics under the real-
world probability measure P. By Ito formula,
1 2 2
2
 1 1 1 1 2

𝑑𝑍𝑡 = 1 𝜇 2𝑆𝑡 𝑑𝑡 + 𝑆𝑡 𝜎2 · 𝑑𝑊𝑡 + 𝑆𝑡 − 1 2 (𝜇1𝑆𝑡 𝑑𝑡 + 𝑆𝑡 𝜎1 · 𝑑𝑊𝑡 ) + 1 ∥𝜎1 ∥ 𝑑𝑡
𝑆𝑡 (𝑆𝑡 ) 𝑆𝑡
− 𝑍𝑡 (𝜎1 ·𝜎2 ) 𝑑𝑡
= 𝑍𝑡 𝜇2𝑑𝑡 + 𝜎2 · 𝑑𝑊𝑡 − 𝑍𝑡 𝜇 1𝑑𝑡 + 𝜎1 · 𝑑𝑊𝑡 + 𝑍𝑡 ∥𝜎1 ∥ 2𝑑𝑡 − 𝑍𝑡 (𝜎1 ·𝜎2 ) 𝑑𝑡



 
= 𝑍𝑡 𝜇2 − 𝜇 1 + ∥𝜎1 ∥ 2 − (𝜎1 ·𝜎2 ) 𝑑𝑡 + 𝑍𝑡 (𝜎2 − 𝜎1 ) · 𝑑𝑊𝑡 .

4
We now want to find a Girsanov shift 𝜃 𝑡 = (𝜃 𝑡1, 𝜃 𝑡2 ) ⊤ such that, under the new measure Q defined
by the density
𝑑Q1 ∫ 𝑡 ∫ 𝑡 
= exp 𝜃 𝑠 ·𝑑𝑊𝑠 − 2 1
∥𝜃 𝑠 ∥ 2𝑑𝑠 ,
𝑑P F𝑡 0 0
the process 𝑍 becomes a Q-martingale (i.e. its drift vanishes).
By Girsanov’s theorem the Q1 -Brownian motion is
∫ 𝑡
Q1
𝑊𝑡 := 𝑊𝑡 − 𝜃 𝑠 𝑑𝑠,
0
1
and substituting 𝑑𝑊𝑡 = 𝑑𝑊𝑡Q + 𝜃 𝑡 𝑑𝑡 into the P-dynamics of 𝑍 gives

𝑑𝑍𝑡 = 𝑍𝑡 𝜇2 − 𝜇 1 + ∥𝜎1 ∥ 2 − (𝜎1 ·𝜎2 ) + (𝜎2 − 𝜎1 ) ·𝜃 𝑡 𝑑𝑡 + 𝑍𝑡 (𝜎2 − 𝜎1 ) · 𝑑𝑊𝑡Q .

To make 𝑍 a Q-martingale, we need the drift term to vanish. That is, we choose 𝜃 𝑡 such that

𝜇 2 − 𝜇1 + ∥𝜎1 ∥ 2 − (𝜎1 ·𝜎2 ) + (𝜎2 − 𝜎1 ) ·𝜃 𝑡 = 0.

Equivalently, this is a linear equation in the two unknowns (the components of the vector
𝜃 𝑡 ∈ R2 ):
(𝜎2 − 𝜎1 ) · 𝜃 𝑡 = − 𝜇2 − 𝜇 1 + ∥𝜎1 ∥ 2 − (𝜎1 · 𝜎2 ) .

A solution always exists as long as 𝜎2 ≠ 𝜎1 , but it is not unique, because there are infinitely
many vectors 𝜃 𝑡 satisfying this scalar equation:

(𝜎21 − 𝜎11 )𝜃 1 + (𝜎22 − 𝜎12 )𝜃 2 = − 𝜇2 − 𝜇 1 + ∥𝜎1 ∥ 2 − 𝜎1 · 𝜎2

and for each 𝜃 1 , we can write

−(𝜇 2 − 𝜇1 + ∥𝜎1 ∥ 2 − 𝜎1 · 𝜎2 ) − 𝜎21𝜃 1
𝜃2 = , if (𝜎22 − 𝜎12 ) ≠ 0,
(𝜎22 − 𝜎12 )

With this choice, the SDE of 𝑍 under Q1 becomes

1
𝑑𝑍𝑡 = 𝑍𝑡 (𝜎2 − 𝜎1 ) · 𝑑𝑊𝑡Q .
1
Notice that 𝜎2 − 𝜎1 ∈ R2 is a vector and 𝑊𝑡Q is a 2-dimensional Brownian motion. The dot
product produces a scalar, so 𝑍𝑡 is scalar-valued.
To rewrite this in the standard 1-dimensional Black-Scholes form, define
1
𝑋𝑡 := (𝜎2 − 𝜎1 ) · 𝑊𝑡Q .

Its variance is
Var(𝑋𝑡 ) = (𝜎2 − 𝜎1 ) · (𝜎2 − 𝜎1 ) 𝑡 = ∥𝜎2 − 𝜎1 ∥ 2𝑡 .

5
So we get the standard brownian motion:
𝑋𝑡
𝑊𝑡1 := ,
∥𝜎2 − 𝜎1 ∥
Substituting back, we obtain the equivalent 1-dimensional SDE:

𝑑𝑍𝑡 = 𝑍𝑡 ∥𝜎2 − 𝜎1 ∥ 𝑑𝑊𝑡1 .

We now conclude by using the Black-Scholes formula with zero short rate, unit strike price,
and volatility 𝜎 = ∥𝜎2 − 𝜎1 ∥. The price of the exchange option is thus given by the formula

Π𝑡 [𝑋 ] = 𝑆𝑡1 {𝑍𝑡 𝑁 (𝑑 1 ) − 𝑁 (𝑑 2 )}

where
1 1 √
 
𝑑1 = √ ln(𝑆𝑡2 /𝑆𝑡1 ) + 𝜎 2 (𝑇 − 𝑡) 𝑑2 = 𝑑1 − 𝜎 𝑇 − 𝑡 .
𝜎 𝑇 −𝑡 2