Shiryaev’s problem Our model Preliminaries The value function Numerical results
Detecting the maximum of a mean-reverting scalar process
GE.Espinosa Joint work with N.Touzi
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
LetX a Brownian motion. We want to track the maximum ofX over a fixed period ([0,1]).
References :
I Quickest Detection Problems in the Technical Analysis of the Financial Data, A. Shiryaev
I Optimal Stopping of the Maximum Process: the Maximality Principle, G. Peskir
I Predicting the Time of the Ultimate Maximum for Brownian Motion with Drift, J. du Toit and G. Peskir
Only for Brownian motions: importance of the explicit law of (X,X∗)
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
LetX a Brownian motion. We want to track the maximum ofX over a fixed period ([0,1]).
minimizeE`(X1∗−Xθ) for a certain loss function `
→ horizon objective References :
I Quickest Detection Problems in the Technical Analysis of the Financial Data, A. Shiryaev
I Optimal Stopping of the Maximum Process: the Maximality Principle, G. Peskir
I Predicting the Time of the Ultimate Maximum for Brownian Motion with Drift, J. du Toit and G. Peskir
Only for Brownian motions: importance of the explicit law of (X,X∗)
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
LetX a Brownian motion. We want to track the maximum ofX over a fixed period ([0,1]).
minimizeE`(X1∗−Xθ) for a certain loss function `
→ horizon objective
Financial Data, A. Shiryaev
I Optimal Stopping of the Maximum Process: the Maximality Principle, G. Peskir
I Predicting the Time of the Ultimate Maximum for Brownian Motion with Drift, J. du Toit and G. Peskir
Only for Brownian motions: importance of the explicit law of (X,X∗)
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
LetX a Brownian motion. We want to track the maximum ofX over a fixed period ([0,1]).
minimizeE`(X1∗−Xθ) for a certain loss function `
→ horizon objective References :
I Quickest Detection Problems in the Technical Analysis of the Financial Data, A. Shiryaev
I Optimal Stopping of the Maximum Process: the Maximality Principle, G. Peskir
I Predicting the Time of the Ultimate Maximum for Brownian
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
I Maximum process: given z ≥x >0, Zt = sup
s≤t
Xs∨z
V(x) = inf
0≤τ≤1E[`(Z1−Xτ)|X0 =Z0 =x] =V(0)
I The value function would be: v(t,x,z) = inf
t≤τ≤1E[`(Z1−Xτ)|Xt =x,Zt=z] 0≤x ≤z
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
I Maximum process: given z ≥x >0, Zt = sup
s≤t
Xs∨z (Xt,Zt) is a Markov process
I Optimization problem: V(x) = inf
0≤τ≤1E[`(Z1−Xτ)|X0 =Z0 =x] =V(0)
I The value function would be: v(t,x,z) = inf
t≤τ≤1E[`(Z1−Xτ)|Xt =x,Zt=z] 0≤x ≤z
I Maximum process: given z ≥x >0, Zt = sup
s≤t
Xs∨z
(Xt,Zt) is a Markov process
I Optimization problem:
V(x) = inf
0≤τ≤1E[`(Z1−Xτ)|X0 =Z0 =x] =V(0)
I The value function would be:
v(t,x,z) = inf
t≤τ≤1E[`(Z1−Xτ)|Xt =x,Zt=z] 0≤x ≤z
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
Theorem
Let`(x) =xp,p >0,p 6= 1. Then there exists zp>0 such as the stopping time:
τp= inf{t ≤1, Zt−Xt ≥zp√ 1−t}
is optimal.
In the case p = 2,z2 is the unique solution of the equation:
4φ(z)−2zϕ(z)−3 = 0 whereϕ(z) = 1
√ 2πe−z
2
2 and φ(z) = Z z
−∞
ϕ(u)du. The value function is given byV = 2φ(z2)−1
Theorem
Let`(x) =xp,p >0,p 6= 1. Then there exists zp>0 such as the stopping time:
τp= inf{t ≤1, Zt−Xt ≥zp√ 1−t}
is optimal. In the casep = 2,z2 is the unique solution of the equation:
4φ(z)−2zϕ(z)−3 = 0 whereϕ(z) = 1
√ 2πe−z
2
2 and φ(z) = Z z
−∞
ϕ(u)du.
The value function is given byV = 2φ(z2)−1
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
Idea of the proof
I First step: use properties of the Brownian motion and its maximum to reduce the problem from 3 dimensions to 1.
I Second step: derive the HJB equation associated to the new problem and solve explicitly this ODE. Then use verification arguments.
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
First step
IfE|`(Z1−Xt)|<∞, then for any stopping time τ in [0,1]:
E(`(Z1−Xτ)|Fτ) =E(`(Zτ∨ max
τ≤r≤1Xr −Xτ)|Fτ)
=E(`(Zτ∨ { max
0≤r≤1−τBr +Xτ} −Xτ)|τ,Xτ,Zτ)
=E(`(max(Zτ−Xτ, ξ))|τ,Xτ,Zτ) whereξ has the following distribution density:
So that:
V = inf
τ E
"
`(Zτ −Xτ)(2φ(Zτ−Xτ
√1−τ )−1) + 2 Z ∞
Zτ√−Xτ 1−τ
`(y√
1−τ)ϕ(y)dy
#
Now asx =z andX is a Brownian motion,Z−X has the same law as|X|. So that (`(x) =xp):
V = inf
τ E
(1−τ)p2Hp( |Xτ|
√1−τ)
withHp(y) =yp+ 2p Z ∞
y
up−1(1−φ(u))du
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
Now we introduce the ”new” times ∈[0,∞) by:
s =−1
2ln(1−t) LetBs =esX1−e−2s = √X1−tt , we have:
dBs =Bsds+
√ 2dWs
whereWs = 1
√
Z 1−e−2s √dBu is a Brownian motion.
We have reduced the problem to the following one:
W(b) = inf
σ≥0Eze−pσHp(|Bσ|) whereB is given by B0=b and the previous dynamics.
And we will haveV =W(0).
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
Second step:
We try to solve this problem by considering the following free boundary problem: findzp andW such as:
W00(b) +bW0(b)−pW(b) = 0 forb ∈(−zp,zp) W(±zp) =Hp(zp)
W0(±zp) =±H0(zp)
And we can solve it explicitly: forb ∈(−zp,zp), W(b) =Hp(zp)e
z2 p−b2
2 M(p+12 ,12,b22) M(p+12 ,12,z2p2) whereM(a,b,x) = 1 + abx+b(b+1)a(a+1)x2!2 +...
Second step:
We try to solve this problem by considering the following free boundary problem: findzp andW such as:
W00(b) +bW0(b)−pW(b) = 0 forb ∈(−zp,zp) W(±zp) =Hp(zp)
W0(±zp) =±H0(zp)
And we can solve it explicitly: forb ∈(−zp,zp), W(b) =Hp(zp)e
z2 p−b2
2 M(p+12 ,12,b22) M(p+12 ,12,z2p2) whereM(a,b,x) = 1 +bax+b(b+1)a(a+1)x2!2 +...
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
Andzp is the unique positive solution of:
Hp0(z)
Hp(z) +z = (p+ 1)zM(p+32 ,32,z22) M(p+12 ,32,z22) Outside (−zp,zp), we will have W(b) =Hp(|b|).
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
I True only for a Brownian motion (possibly with drift as investigated by Peskir)
I How to choose the period (which T)? What is its financial meaning?
I A trader does not want to miss several crossings
I A trader does not want to be in a position for a time of infinite expectation
Several crossings:
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
Infinite expected time of holding:
I sell at timeτ∗, then wait (at least) until the process crosses 0 to buy back. For a Brownian motion, the average waiting time is ExT0 = +∞, which is a problem for a trader.
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
1. `(Z1−Xτ) : horizon objective →interesting but not a natural criteria for trading
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
1. `(Z1−Xτ) : horizon objective →interesting but not a natural criteria for trading
2. `(ZT0−Xτ) : gain objective: one defines the trading range (in
%) → natural framework→ never studied before
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Problem Results
Limits of this model
Shiryaev’s problem Our model Preliminaries The value function Numerical results
The process Loss function
Dynamic programming framework
Consider a process starting atx >0 given by the following dynamics:
dXt =µ(Xt)dt+σ(Xt)dWt whereµ(x)≤0 for x≥0
Xt = max
0≤s≤tXs The time of hitting 0 is:
T0 = inf{t≥0,Xt = 0}
Shiryaev’s problem Our model Preliminaries The value function Numerical results
The process Loss function
Dynamic programming framework
Consider a process starting atx >0 given by the following dynamics:
dXt =µ(Xt)dt+σ(Xt)dWt whereµ(x)≤0 for x≥0
We define the associated current maximum process by:
Xt∗ = max
0≤s≤tXs
The time of hitting 0 is:
T0 = inf{t≥0,Xt = 0}
Consider a process starting atx >0 given by the following dynamics:
dXt =µ(Xt)dt+σ(Xt)dWt whereµ(x)≤0 for x≥0
We define the associated current maximum process by:
Xt∗ = max
0≤s≤tXs The time of hitting 0 is:
T0 = inf{t≥0,Xt = 0}
Shiryaev’s problem Our model Preliminaries The value function Numerical results
The process Loss function
Dynamic programming framework
Consider a loss function`:R+ →R,C2, increasing and convex.
We try to solve the following optimization problem V0(x) = inf
θ≤T0
E`(XT∗0−Xθ)
Consider a loss function`:R+ →R,C2, increasing and convex.
We try to solve the following optimization problem V0(x) = inf
θ≤T0
E`(XT∗0−Xθ)
Shiryaev’s problem Our model Preliminaries The value function Numerical results
The process Loss function
Dynamic programming framework
Letz ≥x (the inherited maximum), we introduce the following process
Zt =z∨ max
0≤s≤tXs
The value function of our problem is given by: V(x,z) = inf
θ≤T0
Ex,z`(ZT0−Xθ) Notice that it is a stationary problem
Shiryaev’s problem Our model Preliminaries The value function Numerical results
The process Loss function
Dynamic programming framework
Letz ≥x (the inherited maximum), we introduce the following process
Zt =z∨ max
0≤s≤tXs
The value function of our problem is given by:
V(x,z) = inf
θ≤T0
Ex,z`(ZT0−Xθ)
Shiryaev’s problem Our model Preliminaries The value function Numerical results
The process Loss function
Dynamic programming framework
Letz ≥x (the inherited maximum), we introduce the following process
Zt =z∨ max
0≤s≤tXs
The value function of our problem is given by:
V(x,z) = inf
θ≤T0
Ex,z`(ZT0−Xθ) Notice that it is a stationary problem
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Scale function Obstacle function Finiteness ofV A degenerate case
We introduce the functionα(x) =−2µ(x)σ(x)2(≥0) and the scale function of the processX is defined by:
S(x) = Z x
0
e
Ru
0α(v)dvdu
Assume thatX0=x>0 and Z0 =z ≥x, then the law ofZT0 is given by:
Px,z[ZT0≤y] =
(0 ify <z 1−S(x)S(y) else In particular,Px,z[ZT0 =z] = 1−S(x)S(z)
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Scale function Obstacle function Finiteness ofV A degenerate case
We introduce the functionα(x) =−2µ(x)σ(x)2(≥0) and the scale function of the processX is defined by:
S(x) = Z x
0
e
Ru
0α(v)dvdu We can compute explicitly the law ofZT0: Proposition
Assume thatX0=x>0 and Z0 =z ≥x, then the law ofZT0 is given by:
P [Z ≤y] =
(0 ify <z
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Scale function Obstacle function Finiteness ofV A degenerate case
Define the obstacle function
g(x,z) =Ex,z`(ZT0−x)
g(x,z) =`(z−x) +S(x) Z +∞
z
`0(u−x) S(u) du We defineL(X) as the set of loss functions `such as `S0 is integrable in a neighborhood of +∞
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Scale function Obstacle function Finiteness ofV A degenerate case
Define the obstacle function
g(x,z) =Ex,z`(ZT0−x) From the law ofZT0, we get:
Proposition
g(x,z) =`(z−x) +S(x) Z +∞
z
`0(u−x) S(u) du
We defineL(X) as the set of loss functions `such as `S0 is integrable in a neighborhood of +∞
Define the obstacle function
g(x,z) =Ex,z`(ZT0−x) From the law ofZT0, we get:
Proposition
g(x,z) =`(z−x) +S(x) Z +∞
z
`0(u−x) S(u) du We defineL(X) as the set of loss functions `such as `S0 is integrable in a neighborhood of +∞
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Scale function Obstacle function Finiteness ofV A degenerate case
We have the following result:
Proposition
Assume that for allx≥0, sup
u≥x
`(u)
`(u−x <∞.
Then there is equivalence between the following assertions:
I V is finite everywhere
I g is finite everywhere
I `∈L(X)
Typically, the additional assumption removes`of the form
`(x) =ex2 (without it we still have (iii)⇒(ii)⇒(i)).
We have the following result:
Proposition
Assume that for allx≥0, sup
u≥x
`(u)
`(u−x <∞.
Then there is equivalence between the following assertions:
I V is finite everywhere
I g is finite everywhere
I `∈L(X)
Typically, the additional assumption removes`of the form
`(x) =ex2 (without it we still have (iii)⇒(ii)⇒(i)).
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Scale function Obstacle function Finiteness ofV A degenerate case
If`(x) =x then the problem can be solved directly.
V(x,z) = inf
θ Ex,z(ZT0−Xθ) =Ex,zZT0−sup
θ ExXθ
So we now consider the new value function: W(x) = sup
θ≤T0
ExXθ And we have the following result:
Proposition
If`(x) =x then V =g
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Scale function Obstacle function Finiteness ofV A degenerate case
If`(x) =x then the problem can be solved directly.
V(x,z) = inf
θ Ex,z(ZT0−Xθ) =Ex,zZT0−sup
θ ExXθ So we now consider the new value function:
W(x) = sup
θ≤T0
ExXθ
If`(x) =x then V =g
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Scale function Obstacle function Finiteness ofV A degenerate case
If`(x) =x then the problem can be solved directly.
V(x,z) = inf
θ Ex,z(ZT0−Xθ) =Ex,zZT0−sup
θ ExXθ So we now consider the new value function:
W(x) = sup
θ≤T0
ExXθ And we have the following result:
Proposition
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
From now on, we assume that`(x) = x22
The HJB equation of the problem is:
min{LV,g −V}= 0 V(0,z) =`(z) Vz(z,z) = 0
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
From now on, we assume that`(x) = x22 We writeL= ∂x∂22 −α(x)∂x∂
The HJB equation of the problem is:
min{LV,g −V}= 0 V(0,z) =`(z) Vz(z,z) = 0
From now on, we assume that`(x) = x22 We writeL= ∂x∂22 −α(x)∂x∂
The HJB equation of the problem is:
min{LV,g −V}= 0 V(0,z) =`(z) Vz(z,z) = 0
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
We guess that the solution should be of the following form: find γ:R+ →R+, andv such as:
γ(x)≥x And:
Lv = 0 if 0<z < γ(x)
v =g andLg(x,z)≥0 ifz ≥γ(x) vz(z,z) = 0
v(0,z) =`(z)
v(x, γ(x)) =g(x, γ(x))
We will use a verification argument in the end
We guess that the solution should be of the following form: find γ:R+ →R+, andv such as:
γ(x)≥x And:
Lv = 0 if 0<z < γ(x)
v =g andLg(x,z)≥0 ifz ≥γ(x) vz(z,z) = 0
v(0,z) =`(z)
v(x, γ(x)) =g(x, γ(x)) vx(x, γ(x)) =gx(x, γ(x)) We will use a verification argument in the end
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
In the area whereV =g, the HJB equation implies that Lg ≥0.
Therefore we need to study the set Γ+={(x,z), Lg(x,z)≥0}.
we have:
I Γ+={(x,z), Lg(x,z)≥0}is above the graph of Γ
I Γ−={(x,z), Lg(x,z)≤0}is below the graph of Γ
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
In the area whereV =g, the HJB equation implies that Lg ≥0.
Therefore we need to study the set Γ+={(x,z), Lg(x,z)≥0}.
We call Γ(x) = min{z ≥x, Lg(x,z)≥0}, which is a C0 and piecewiseC1 function, decreasing and then increasing. Moreover, we have:
I Γ+={(x,z), Lg(x,z)≥0} is above the graph of Γ
I Γ−={(x,z), Lg(x,z)≤0} is below the graph of Γ
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
We call:
I Γ1 the decreasing part and Γ2 the increasing part of Γ.
I z0 = Γ(0)
I z1 = sup{z >0, Lg(z,z)<0}(possibly infinite)
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
Recall thatL= ∂x∂22 −α(x)∂x∂ → only partial derivatives wrtx.
FromLv = 0 together withv(x, γ(x)) =g(x, γ(x)),
vx(x, γ(x)) =gx(x, γ(x)) and vz(x,x) = 0 we get the following ODE:
γ0(x) = Lg(x, γ(x)) 1−S(γ(x))S(x) And we want (x, γ(x))∈Γ+.
Problem: we do not have an a priori initial condition (not a Cauchy problem). →We must choose the right one that will allow us to derive a verification result.
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
Recall thatL= ∂x∂22 −α(x)∂x∂ → only partial derivatives wrtx.
FromLv = 0 together withv(x, γ(x)) =g(x, γ(x)),
vx(x, γ(x)) =gx(x, γ(x)) and vz(x,x) = 0 we get the following ODE:
γ0(x) = Lg(x, γ(x)) 1−S(γ(x))S(x)
Problem: we do not have an a priori initial condition (not a Cauchy problem). →We must choose the right one that will allow us to derive a verification result.
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
Recall thatL= ∂x∂22 −α(x)∂x∂ → only partial derivatives wrtx.
FromLv = 0 together withv(x, γ(x)) =g(x, γ(x)),
vx(x, γ(x)) =gx(x, γ(x)) and vz(x,x) = 0 we get the following ODE:
γ0(x) = Lg(x, γ(x)) 1−S(γ(x))S(x) And we want (x, γ(x))∈Γ+.
Problem: we do not have an a priori initial condition (not a Cauchy problem). →We must choose the right one that will allow us to derive a verification result.
Recall thatL= ∂x∂22 −α(x)∂x∂ → only partial derivatives wrtx.
FromLv = 0 together withv(x, γ(x)) =g(x, γ(x)),
vx(x, γ(x)) =gx(x, γ(x)) and vz(x,x) = 0 we get the following ODE:
γ0(x) = Lg(x, γ(x)) 1−S(γ(x))S(x) And we want (x, γ(x))∈Γ+.
Problem: we do not have an a priori initial condition (not a Cauchy problem). →We must choose the right one that will allow us to derive a verification result.
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
We show that it is possible to chooseγ to be the sup of the solutions that cross Γ2 and that we then have:
x→zlim1
γ(x)−x= 0 Moreover, thisγ may cross Γ1 but not Γ2.
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
In general,γ defined before might cross Γ1. Therefore, the boundary will be made of 2 distinct parts (as we need (x, γ(x))∈Γ+).
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
As before, but replacing the Neumann condition by the Dirichlet condition on{x = 0} we get the following equation:
z2
2 −g(x,z) +gx(x,z)S(x) S0(x) = 0
We have the following result:
There existsz∗ <z0 such as: for anyz ∈[z∗,z0] the solution of f(x,z) = 0 admits a unique solution x(z) that satisfies
- Γ−11 (z)≤x(z)≤Γ−12 (z)(≤z) - if z <z0,x(z)>0
- if z =z∗, then (x(z∗),z∗)∈Γ2
- x(z) is C1
As before, but replacing the Neumann condition by the Dirichlet condition on{x = 0} we get the following equation:
z2
2 −g(x,z) +gx(x,z)S(x) S0(x) = 0 We have the following result:
There existsz∗ <z0 such as: for any z ∈[z∗,z0] the solution of f(x,z) = 0 admits a unique solutionx(z) that satisfies
- Γ−11 (z)≤x(z)≤Γ−12 (z)(≤z) - if z <z0,x(z)>0
- if z =z∗, then (x(z∗),z∗)∈Γ2
- x(z) is C1
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
Consistency of areas 1 and 2
The previous results guarantee that we are in one of the following situations:
I the graph of γ defined in area 2 does not cross Γ
I the graphs of the two previous γ will intersect on a (unique) point (¯x,¯z)
We now call γ the following function:
γ=
γ1 on [0,¯x] γ2 on [¯x,z1]
id on [z1,+∞) ifz1 <∞
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
Consistency of areas 1 and 2
The previous results guarantee that we are in one of the following situations:
I the graph of γ defined in area 2 does not cross Γ then area 1 is degenerate, we write (¯x,¯z) = (0,z0)
I the graphs of the two previous γ will intersect on a (unique) point (¯x,¯z)
We now call γ the following function:
γ=
γ1 on [0,¯x] γ2 on [¯x,z1]
id on [z1,+∞) ifz1 <∞
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
Consistency of areas 1 and 2
The previous results guarantee that we are in one of the following situations:
I the graph of γ defined in area 2 does not cross Γ then area 1 is degenerate, we write (¯x,¯z) = (0,z0)
I the graphs of the two previous γ will intersect on a (unique) point (¯x,¯z)
γ=
γ1 on [0,¯x] γ2 on [¯x,z1]
id on [z1,+∞) ifz1 <∞
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
Consistency of areas 1 and 2
The previous results guarantee that we are in one of the following situations:
I the graph of γ defined in area 2 does not cross Γ then area 1 is degenerate, we write (¯x,¯z) = (0,z0)
I the graphs of the two previous γ will intersect on a (unique) point (¯x,¯z)
We now callγ the following function:
γ1 on [0,¯x]
Now thatγ, ¯x and ¯z are defined, we define v on areas 1 and 2 by:
ifx ≤¯x,z ∈[¯z, γ(x)]
v(x,z) =v1(x,z) = z2
2 +gx(γ1−1(z),z) S(x) S0(γ1−1(z)) ifx ≥¯x,z ∈[¯z, γ(x)]
v(x,z) =v2(x,z) =g(γ2−1(z),z)
+gx(γ2−1(z),z)S(x)−S(γ2−1(z)) S0(γ2−1(z))
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
Area 3
Here we want to solveLv = 0 with the Dirichlet condition on {x= 0}, the Neumann condition on {x =z}, and 2 Dirichlet conditions on{(x,¯z),0≤x ≤¯x} and{(x,¯z),¯x ≤x≤z}
We set:
v(x,z) =v3(x,z) = z2
2 +S(x)[
Z +∞
z
u
S(u)du−K]
+∞ g (¯x,¯z)
Area 4
Herev =g, and by construction,Lg ≥0
Shiryaev’s problem Our model Preliminaries The value function Numerical results
HJB equation
Reduction to a free boundary problem Study ofLg
Equation of the obstacle Other areas
Verification
Finally, we must show thatv defined previously satisfiesv ≤g and v≥0. Then we are able to claim the following result:
Theorem(Verification theorem)
The functionv defined previously is indeed the value function V of our initial problem andτ∗= inf{t ≥0,Zt≥γ(Xt)}is an optimal stopping time
γ has the following form for an OU process :
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Plot ofγ
Theoretical simulations Market data simulations
We simulate an OU (α(x) =x) and use the following strategy:
when we ”detect” the maximum, we sell 1 stock, when we cross 0, we close the position, we do the same for the minimum
We simulate an OU (α(x) =x) and use the following strategy:
when we ”detect” the maximum, we sell 1 stock, when we cross 0, we close the position, we do the same for the minimum
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Plot ofγ
Theoretical simulations Market data simulations
We compute the following ”mean-reverting” process:
X =
BNP Socgen
MA(SocgenBNP ) −1
We compute the following ”mean-reverting” process:
X =
BNP Socgen
MA(SocgenBNP ) −1
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Plot ofγ
Theoretical simulations Market data simulations
Extension
Inhomogeneous problem:
dXt =µ(t,Xt)dt+σ(t,Xt)dWt whereµ(t,x)≤0 forx ≥0,t ≥0
Problem: Hard to find a regular upper bound forV in this case (necessary for a comparison result).
But we can find one using the homogeneous case if the inhomogeneity is only on the drift:
dXt =µ(t,Xt)dt+σ(Xt)dWt
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Plot ofγ
Theoretical simulations Market data simulations
Extension
Inhomogeneous problem:
dXt =µ(t,Xt)dt+σ(t,Xt)dWt whereµ(t,x)≤0 forx ≥0,t ≥0
Problem: Hard to find a regular upper bound forV in this case (necessary for a comparison result).
dXt =µ(t,Xt)dt+σ(Xt)dWt
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Plot ofγ
Theoretical simulations Market data simulations
Extension
Inhomogeneous problem:
dXt =µ(t,Xt)dt+σ(t,Xt)dWt whereµ(t,x)≤0 forx ≥0,t ≥0
Problem: Hard to find a regular upper bound forV in this case (necessary for a comparison result).
But we can find one using the homogeneous case if the inhomogeneity is only on the drift:
Short Bibliography
- The Pricing of the American Option, R. Myneni, The Annals of Applied Probability (1992)
- Quickest Detection Problems in the Technical Analysis of the Financial Data, A. Shiryaev, Mathematical Finance - Bachelier Congress (2000), Springer Finance
- Optimal Stopping of the Maximum Process : the Maximality Principle, G. Peskir, Ann. Prob. (1998)
- Predicting the Time of the Ultimate Maximum for Brownian Motion with Drift, J. du Toit and G. Peskir, Proc. Math.
Control Theory Finance (2007), Springer
Shiryaev’s problem Our model Preliminaries The value function Numerical results
Plot ofγ
Theoretical simulations Market data simulations
Special thanks to J.Lebuchoux - Reech Aim