Martingale invariance and utility maximization

Martingale invariance and utility maximization
Thorsten Rheinlander
Jena, June 2010
Thorsten Rheinlander ()
Martingale invariance
Jena, June 2010
1 / 27
Martingale invariance property
Consider two …ltrations F
G . F is immersed in G if every
F -martingale is a G -martingale.
Let S be some F -adapted locally bounded semi-martingale. Typical
choice: F = F S
Maximizing expected exponential utility from terminal wealth:
sup E
ϑ2Θ
exp
Z T
0
ϑt dSt
Is it possible for an agent to achieve higher expected utility by using
G -predictable strategies? What rôle does the immersion property play
here?
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
2 / 27
Duality approach to utility maximization
Under mild assumptions, the following key duality holds (Delbaen et
al (2002)):
sup E
ϑ2Θ
exp
Z T
0
ϑt dSt
=
exp
inf H (Q, P )
Q 2Me
Here Θ denotes a set of admissible strategies, Me the set of all
equivalent martingale measures for S, and H (Q, P ) is the relative
entropy.
The link between the two problems is provided by the minimal
entropy martingale measure Q E and its density representation
dQ E
= exp c +
dP
Z T
0
η t dSt
.
Here c = H Q E , P , and η is the optimal strategy for the primal
problem. Note that in general η depends on the time horizon T .
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
3 / 27
Proposition: Let S be a locally bounded F -adapted process such that the
entropy measure Q E (G) for S exists (and then Q E (F ) exists as well). If
E
E
E
F is immersed in G , then
h Z (Fi) = Z (G). Here Z (H) is the density
process, ZtE (H) = E
dQ E
dP
Ht , relative to …ltration H.
Proof: A martingale measure Q for S wrt. G induces a martingale
b wrt. F via d Qb = E dQ FT .
measure Q
dP
dP
b P
By the conditional Jensen’s inequality, we have H Q,
H (Q, P ).
E (G) has …nite relative entropy,
Applying this to Q E (G) we get that Q\
hence there is some martingale measure wrt. F which has …nite relative
entropy. Therefore Q E (F ) exists uniquely. As Z E (F ) S is a local
(P, F )-martingale, it is by immersion also a local (P, G)-martingale.
Hence Z E (F ) is the density process of a martingale measure wrt. G , with
terminal value dQ E (F ) /dP. As
H Q E (F ) , P
E (G), P
H Q\
H Q E (G) , P ,
the result follows by uniqueness of Q E (G ).
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
4 / 27
Remarks. 1) In particular, if Z E (G) is not F -adapted, then F cannot be
immersed into G . This can be useful in case it is complicated to determine
the structure of all F -martingales.
2) It is in general not true that Q E (F ) = Q E (G) implies
Z E (F ) = Z E (G), as will be seen later on. In particular, the density of
Q E (G) can be FT -measurable, whereas the density process Z E (G) is not
F -adapted.
3) Summing up, if F is immersed in G , then no additional expected utility
can be gained by employing G -predictable strategies. However, if the
immersion property fails, one can in general not conclude that this leads to
a utility gain.
4) To apply these results, the main task is to calculate Z E (G), the density
process of the entropy measure wrt. G .
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
5 / 27
Optimal martingale measure equation
We work with a …ltration G such that all G -martingales are
continuous. We know that the density of the minimal entropy
measure Q E can necessarily be written as
dQ E
= exp c +
dP
Z T
0
η t dSt
for some constant c and some G -predictable process η.
By the structure condition, there exists a local martingale M and a
G -predictable process λ such that
S
KT
Thorsten Rheinlander (London School of Economics)
= M+
=
Z T
0
Z
λ d [M ] ,
λ2t d [M ]t < ∞.
Martingale invariance
Jena, June 2010
6 / 27
Every martingale measure Q can be written as
dQ
dP
= E
= exp
= exp
Z
λ dM
Z
L
T
1
KT
2
Z
1
λt dSt + KT
2
λt dMt
LT
LT
1
[L]
2 T
1
,
[L]
2 T
where L is some local (P, G)-martingale strongly orthogonal to M.
As we look for a candidate measure with representation as above, we
would like to …nd such an L as well as a constant c and a
G -predictable process ψ such that we can decompose
1
KT = c +
2
Z T
0
ψt dSt + LT +
1
[L] .
2 T
We are looking for a solution (c, ψ, L) to this BSDE, and set
η = ψ λ.
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
7 / 27
We take as candidate
dQ
= exp c +
dP
Z T
0
η t dSt
=E
Z
λ dM
L
.
T
Veri…cation: we have to show that the stochastic exponential is a true
martingale, that H Q, P has …nite relative entropy, and that
Z T
0
η 2t d [S ]t 2 Lexp (P ).
Here Lexp (P ) is the Orlicz space generated by the Young function
exp ( ).
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
8 / 27
We want to show that the following stochastic exponential is a
martingale:
Z
Z =E
λ dM
L
This could be done e.g. by a criterion due to Liptser & Shiryaev
(1977), however, here we give another approach yielding …nite relative
entropy of the resulting measure as well.
De…ne a fundamental sequence of stopping times as
Tn = min f t j Kt _ [L]t
ng ^ T
By Novikov’s criterion, the ZT n are then densities of probability
measures Qn . We have
ZT = lim ZT n
n
Thorsten Rheinlander (London School of Economics)
Martingale invariance
a.s.
Jena, June 2010
9 / 27
Proposition: The following assertions are equivalent.
h
i
(i ) supn EQ n KT n + [L]T n < ∞
(ii ) supn H (Qn , P ) < ∞
(iii ) ZT is the density of a prob. meas. Q, and H (Q, P ) < ∞.
Proof:
(i ) () (ii ) This follows by a calculation based on Girsanov’s theorem.
(ii ) =) (iii ) The sequence (ZT n ) is u.i. by de la Vallée-Poussins criterion.
(iii ) =) (ii ) Follows from Barron’s inequality.
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
10 / 27
Stein & Stein model
The price process S is modelled as strong solution to the SDE
dSt = µVt2 St dt + σVt St dBt
and the volatility process V is given as
dVt = (m
αVt ) dt + β dWt .
Here µ, σ, m, α, β are positive non-zero constants and B, W are two
correlated Brownian motions with correlation coe¢ cient ρ 2 ( 1, +1).
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
11 / 27
Solving the optimal martingale measure equation and following the
veri…cation procedure yields that the entropy measure Q E for S
relative to the …ltration G generated by the BM (B, W ) is given by
dQ E
= exp c
dP
Z T
0
1
ρρ
1
φt + ψt Vt
1
St
1
dSt
.
Here φ, ψ are bounded deterministic functions which can be obtained
as system of two di¤erential equations involving one Riccatti-type
equation. In case the mean-reversion level m is non-zero, ψ will be
non-zero.
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
12 / 27
We choose F = F S , G = F B ,W . In case ρ 2 ( 1, +1), G is the
progressive enlargement of F S by sgn (V ) which is not F S -adapted.
As consequence, if the mean-reversion level m is non-zero and
ρ 2 ( 1, +1), then F is not immersed in G .
An agent who observes sgn (V ) can take advantage of this additional
information when maximizing her expected exponential utility from
terminal wealth, provided ρ, m 6= 0. In the uncorrelated case ρ = 0,
however, F is not immersed in G if m 6= 0, but since η is then equal
to 1/S, the optimal investment strategy does not use the additional
information and there is no bene…t for the insider. Note that in this
case we have Q E (G) = Q E (F ), but Z E (G) 6= Z E (F ).
It seems to be di¢ cult to calculate Z E wrt. F S in case m 6= 0
because of the rather complicated structure of F S .
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
13 / 27
Models with default
Let (Ω, F , P ) be a probability space, supporting a random variable
modeling the time of default τ > 0.
Ht = Iτ t is the counting process associated with τ, H the …ltration
generated by H. Let F be a Brownian …ltration, generated by some
BM W .
We assume that there is a deterministic intensity function µ such that
Z t
P (τ > t ) = exp
0
µs ds
.
Let G = H _ F . We assume that F is immersed into G . In
particular, W stays a Brownian motion in the larger …ltration G .
The G -martingale M associated with the one-jump process is given as
Mt = Ht
Thorsten Rheinlander (London School of Economics)
Z τ ^t
0
Martingale invariance
µs ds.
Jena, June 2010
14 / 27
e of a defaultable security under P as
We model the price process S
et /S
et
dS
= at dt + bt dWt + ct dMt
The functions a, b, and c are deterministic and bounded on [0, T ].
Finding a pricing measure by calibration is sometimes di¢ cult in this
framework. Therefore, we will choose martingale measures according
to optimality criteria.
Our main point is that the presence of two di¤erent noise terms dW
and dM makes the analysis much more complicated compared to a
scenario with only a Brownian driver.
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
15 / 27
Our goal is now to calculate several optimal martingale measures, i.e.
e is a local
probability measures Q such that the price process S
Q-martingale. This
is equivalent to the statement that the stochastic
R
e
e is a local Q-martingale.
logarithm S = d S /S
Assuming the existence of an equivalent martingale measure and a
structure condition, we can write the semi-martingale decomposition
of S uniquely in the form
S =N+
Z
λ d hN i
for a local martingale N and a predictable process λ; here hN i
denotes the predictable compensator of the quadratic variation
process [N ]. In our concrete market model, it is readily computed that
a
dN = b dW + c dM, λ = 2
.
b + µc 2
Moreover, we have
Z
λ dN
=
λ ∆N
=
Thorsten Rheinlander (London School of Economics)
Z
a
(b dW + c dM ) ,
+ µc 2
λc ∆M.
b2
Martingale invariance
Jena, June 2010
16 / 27
Minimal martingale measure
The minimal martingale measure (see Schweizer (1995)), commonly
b is characterized by the property that P-martingales
referred to as P,
b
strongly orthogonal to N under P remain martingales under P.
Its density is given by the Doléans-Dade stochastic exponential
E
Z
= exp
λ dN
T
Z T
0
∏
λt bt dWt
(1
1
2
λt ct ∆Mt ) .
Z T
0
λ2t bt2 dt
0 <t T
As there is one single jump of M with jump size one, the density of
the minimal martingale measure gets negative with non-zero
probability in case a jump occurs at τ T and λτ cτ > 1.
b is in general a signed measure, and we conclude that the
Therefore, P
minimal martingale measure is not a good choice in our situation.
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
17 / 27
Linear Esscher measure
R
For every strategy ϑ such that ϑ dS is exponentially special, we
S
de…ne the modi…ed
R Laplace cumulant process K (ϑ) as exponential
compensator of ϑ dS. In particular,
Z ϑ = exp
Z
ϑ dS
K S (ϑ )
is a local martingale with Z0ϑ = 1.
In case Z ϑ is a martingale on [0, T ], it is the density process of a
probability measure P ϑ . If we can …nd ϑ such that S is a local
P ϑ -martingale, P ϑ is called the linear Esscher measure.
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
18 / 27
By Kallsen & Shiryaev (2001), we have to solve the equation
DK X (ϑ ) = 0.
This translates into the equation
µ
a = b 2 ϑ +cµe ϑc
By the boundedness of the co-e¢ cients, a bounded solution always
exists and gives a martingale measure by the above exponential tilting
due to a veri…cation result by Protter & Shimbo (2008).
The linear Esscher measure is stable wrt. stopping, and coincides with
the minimal Hellinger measure as studied by Choulli & Stricker
(2005).
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
19 / 27
Minimal entropy martingale measure
Entropy measure Q E : Its density can always be written in the form
dQ E
= exp c +
dP
Z T
0
φt dSt
R
for some predictable process φ such that φ dS is a Q E -martingale.
In general, a martingale measure Q for S has a density of the form
dQ
=E
dP
Z
λ dN + L
,
T
where L, L0 = 0, is a local martingale strongly orthogonal to N.
According to martingale representation we can write L as
dL = bL dW + cL dM
for some predictable processes bL , cL . The orthogonality relation
hN, Li = 0 yields
bbL + ccL µ2 = 0.
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
20 / 27
Equating the two di¤erent representations, we obtain the optimal
martingale measure equation
c
=
Z
T
+
Z
1 2 2
λ b + µλc
2
φa
T
( φb
+ log (1
µcL
φµc
dt
λb + bL ) dW
( λ τ cτ
Thorsten Rheinlander (London School of Economics)
1 2
b
2 L
cL ( τ )
Martingale invariance
φτ cτ ))
Iτ
T.
Jena, June 2010
21 / 27
Assuming for now formally that there exists a smooth function u
(where t 2 [0, T ] and h 2 f0, 1g) such that
log (1
( λ τ cτ
cL ( τ )
φτ cτ )) = u (t, Ht )
u (T , h) = 0,
u (t, h)
u (t, Ht ) ,
h 2 f0, 1g ,
we write
log (1
(λτ cτ cL (τ ) φτ cτ ))
= u (τ, 1) u (τ, 0)
=
fu (T , 1) u (τ, 1) + u (τ, 0) u (0, 0)g
u (0, 0)
Z T
∂
u (t, Ht ) dt u (0, 0) .
=
0 ∂t
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
22 / 27
We work with the hypothesis that the integrals over the dt-terms as
well as the dW -terms vanish at maturity T . This yields the two
equations
1
φa + λ2 b 2
2
1
∂u
µλc + bL2 + µcL + φµc +
= 0,
2
∂t
φb + λb
bL = 0.
Inserting the orthogonality relation, we solve for φ as
Thorsten Rheinlander (London School of Economics)
φ=
λ+
ccL µ2
.
b2
Martingale invariance
Jena, June 2010
23 / 27
Moreover, we get, with the notation ∆t u := u (t, Ht )
that
cL = exp (∆u + φc ) + λc 1
c 2 µ2
= exp ∆u λc + 2 cL
b
+ λc
u (t, Ht ),
1.
Introducing
g (t, ut ) :=
1
φa + λ2 b 2
2
1
µλc + bL2 + µcL + φµc
2
,
t
we derive the system of two ordinary di¤erential equations,
∂u
+ g (t, ut ) = 0,
∂t
u (T , 0) = u (T , 1) = 0.
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
24 / 27
The density process Z E associated with Q E de…ned by
ZtE
=
dQ E
dP
= E
Gt
Z
0
λu bu
buL dWu +
Z
0+
λu cuS
cuL dMu
t
is the density process of the minimal entropy martingale measure.
Veri…cation proceeds along the lines of Rheinländer & Steiger (2006).
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
25 / 27
T. Choulli, C. Stricker (2005). Minimal entropy-Hellinger martingale
measure in incomplete markets. Mathematical Finance 15, 465–490
eamericonarticle
Delbaen, F., P. Grandits, T. Rheinländer, D. Samperi, M. Schweizer,
C. Stricker (2002). Exponential hedging and entropic penalties.
Mathematical Finance 12, 99-123
eamericonarticle
M. Frittelli (2000). The minimal entropy martingale measure and the
valuation problem in incomplete markets. Mathematical Finance 10,
39–52
eamericonarticle
P. Grandits, T. Rheinländer (2002). On the minimal entropy
martingale measure. Annals of Probability 30, 1003-1038
eamericonarticle
J. Kallsen, A. Shiryaev (2002). The cumulant process and Esscher’s
change of measure. Finance & Stochastics 6, 397–428
eamericonarticle
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
26 / 27
Y. Lee, T. Rheinländer (2010). Optimal martingale measures for
defaultable assets. Preprint
eamericonarticle
P. E. Protter, K. Shimbo (2008). No arbitrage and general
semimartingales. In: Markov Processes and Related Topics: A
Festschrift for Thomas G. Kurtz (Beachwood, Ohio, USA: Institute of
Mathematical Statistics), 267-283.
eamericonarticle
T. Rheinländer (2005). An entropy approach to the Stein and Stein
model with correlation. Finance and Stochastics 9, 399–413
eamericonarticle
T. Rheinländer, G. Steiger (2006). The minimal entropy martingale
measure for general Barndor¤-Nielsen/Shephard models. The Annals
of Applied Probability 16, 1319–1351.
eamericonarticle
M. Schweizer (1995). On the minimal martingale measure and the
Föllmer-Schweizer decomposition. Stochastic Analysis and
Applications 13, 573–599.
eamericonarticle
Thorsten Rheinlander (London School of Economics)
Martingale invariance
Jena, June 2010
27 / 27