# Optimal Investment under Dynamic Risk Constraints and Partial Information

Wolfgang Putschögl
Johann Radon Institute for Computational and Applied Mathematics (RICAM) Austrian Academy of Sciences www.ricam.oeaw.ac.at

20th September 2007
Joint work with J. Saß (RICAM), Supported by FWF, Project P17947-N12 Workshop and Mid-Term Conference on Advanced Mathematical Methods for Finance Vienna

Outline

Model Setup Problem formulation Time-Dependent Convex Constraints Dynamic Risk Constraints Gaussian Dynamics for the Drift A hidden Markov Model (HMM) for the Drift Example

2 / 34

Model Setup
Filtered probability space: (Ω, F = (Ft )t∈[0,T ] , P) Finite time horizon: T > 0 Money market: bond with stochastic interest rates r dSt
(0)

= St rt dt ,

(0)

S0 = 1 ,

(0)

i.e., St

(0)

“Z t ” = exp rs ds ,
0

r uniformly bounded and progressively measurable w.r.t. F Stock market: n stocks with price process St = (St , . . . , St ) , return Rt , and ˜ excess return Rt , where dSt = Diag(St )(µt dt + σt dWt ) , dRt = µt dt + σt dWt , ˜ dRt = dRt − rt dt .
(1) (n)

W n-dimensional standard Brownian motion w.r.t. F and P drift µt ∈ Rn Ft -adapted and independent of W volatility σt ∈ Rn×n progressively measurable w.r.t. FtS , σt non-singular, and σt−1 uniformly bounded.
3 / 34

Risk Neutral Probability Measure
We introduce the risk neutral probability measure ( → for ﬁltering and optimization). Deﬁnition Martingale density process „ Z t Z 1 t Zt = exp − θs dWs − θs 2 0 0 with θt = σt−1 (µt − rt 1n ) the market price of risk ˜ Risk neutral probability measure P deﬁned by ˜ dP := ZT dP ˜ ˜ E expectation operator under P Girsanov’s theorem: ˜ Wt := Wt + ˜ deﬁnes a P-Brownian motion
4 / 34

2

« ds

Z
0

t

θs ds

Partial Information
Remark We consider the case of partial information: → we can only observe interest rates and stock prices (F r ,S ) but not the drift The portfolio has to be adapted to F r ,S ˆ ˜ → we need the conditional density ζt = E Zt |FtS ˆ ˜ → we need the ﬁlter for the drift µt = E µt |FtS ˆ Assumption The interest rates r are F S -adapted → F r ,S = F S Z is a martingale w.r.t. F and P Lemma We have F S = F W = F R → the market is complete w.r.t. F S
5 / 34

˜

˜

Outline

Model Setup Problem formulation Time-Dependent Convex Constraints Dynamic Risk Constraints Gaussian Dynamics for the Drift A hidden Markov Model (HMM) for the Drift Example

6 / 34

Deﬁnition Trading strategy πt : n-dimensional, F S -adapted, measurable Initial capital x0 > 0 Wealth process X π satisﬁes dXtπ = πt (µt dt + σt dWt ) + (Xtπ − 1n πt )rt dt
π X0 = x0

A strategy is admissible if Xtπ ≥ 0 a.s. for all t ∈ [0, T ] πt represents the wealth invested in the stocks at time t ηtπ = πt /Xtπ denotes the corresponding fraction of wealth

7 / 34

Utility Functions

Deﬁnition U : [0, ∞) → R ∪ {−∞} is a utility function, if U is strictly increasing, strictly concave, twice continuously differentiable on (0, ∞), and satisﬁes the Inada conditions: U (∞) = lim U (x) = 0 ,
x→∞

U (0+) = lim U (x) = ∞ .
x↓0

I denotes the inverse function of U . Assumption I(y ) ≤ Ky a , |I (y )| ≤ Ky −b for all y ∈ (0, ∞) and a, b, K > 0

Example Logarithmic utility U(x) = log(x) Power utility U(x) = x α /α for α < 1, α = 0.

8 / 34

Optimization Problem

Optimization Problem We optimize under partial information! Objective: Maximize the expected utility from terminal wealth, i.e., ˆ ˜ maximize E U(XT ) under (risk) constraints we still have to specify. The optimization problem consists of two steps: 1. Find the optimal terminal wealth 2. Find the corresponding trading strategy

9 / 34

Outline

Model Setup Problem formulation Time-Dependent Convex Constraints Dynamic Risk Constraints Gaussian Dynamics for the Drift A hidden Markov Model (HMM) for the Drift Example

10 / 34

Time-Dependent Convex Constraints
We can write our model under full information with respect to F R as dRt = µt dt + σt dVt , ˆ t ∈ [0, T ] .

where the innovation process V = (Vt )t∈[0,T ] is a P-Brownian motion deﬁned by Z t Z t Z t −1 −1 −1 Vt = Wt + σs (µs − µs ) ds = ˆ σs dRs − σs µs ds . ˆ
0 0 0

Kt represents the constraints on portfolio proportions at time t → ηtπ ∈ Kt Kt is a Ft -progressively measurable closed convex set ∅ = Kt ⊆ Rn that contains 0 For each t we deﬁne the support function δt : Rn → R ∪ {+∞} of −Kt by δt (y ) = sup(−x y ) ,
x∈Kt

y ∈ Rn .

→ δt (y ) is Ft -progressively measurable → y → δt (y ) is a lower semicontinuous, proper, convex function on its effective ˜ domain Kt = {y ∈ Rn : δt (y ) < ∞}
11 / 34

Time-Dependent Convex Constraints
Deﬁnition A trading strategy η π is called Kt -admissible for initial capital x0 > 0 if Xtπ ≥ 0 a.s. and ηtπ ∈ Kt for all t ∈ [0, T ]. We denote the class of admissible trading strategies for initial capital x0 by AKt (x0 ). ˜ We introduce the set H of dual processes νt : [0, T ] × Ω → Kt which are ˆR T ` ´ ˜ R Ft -progressively measurable processes, satisfying E 0 νt 2 + δt (νt ) dt < ∞. For each dual process ν ∈ H we introduce a new interest rate process rtν = rt + δt (νt ). a new drift process µν = µt + νt + δt (νt )1n . ˆt ˆ a new market price of risk θtν = σt−1 (ˆt − rt + νt ) µ a new density process ζ ν given by dζtν = −θtν ζtν dVt Then: Solution under constraints = solution under no constraints with new market coefﬁcients! Problem: Find optimal ν!
12 / 34

Time-Dependent Convex Constraints
Proposition
η Suppose x0 > 0 and E[U − (XT )] < ∞ for all η π ∈ AK (x0 ).

A trading strategy η π ∈ AK (x0 ) is optimal, if for some y ∗ > 0, ν ∗ ∈ H
π ˜∗ XT = I(y ∗ ζT ) ,

X ν (y ∗ ) = x0 ,

˜∗ ˜ν where ζT = ζT . Further, η π and ν ∗ have to satisfy the complementary slackness condition δt (νt∗ ) + (ηtπ ) νt∗ = 0 , y ∗ , ν ∗ solve the dual problem ˆ ˜ ˜ ˜ ˜ν V (y ) = inf E U(y ζT ) ,
ν∈H

t ∈ [0, T ] .

˘ ¯ ˜ where U(y ) = supx>0 U(x) − xy , y > 0 is the convex dual function of U. If F R = F V holds, then an optimal trading strategy exists.
13 / 34

Outline

Model Setup Problem formulation Time-Dependent Convex Constraints Dynamic Risk Constraints Gaussian Dynamics for the Drift A hidden Markov Model (HMM) for the Drift Example

14 / 34

Limited Expected Loss & Limited Expected Shortfall

Suppose we cannot trade in [t, t + ∆t]. Then “Z π ∆Xtπ = Xt+∆t − Xtπ = Xtπ exp
t t+∆t

” “Z rs ds − Xtπ + exp
t

t+∆t

” rs ds (ηtπ ) Xtπ

Z t+∆t ” ” “ “ 1 Z t+∆t ˜ diag(σs σs ) ds + σs dWs − 1 . × exp − 2 t t Next, we impose the relative LEL constraint ˆ ˜ ˜ E (∆Xtπ )− |FtS < εt , with εt = LXtπ . Deﬁnition ˛ ˆ ˘ ˜ ¯ KtLEL := ηtπ ∈ Rn ˛˜ (∆Xtπ )− |FtS < εt E

15 / 34

Limited Expected Loss & Limited Expected Shortfall
We introduce the relative LES constraint as an extension to the LEL constraint ˆ ˜ ˜ E (∆Xtπ + qt )− |FtS < εt , with εt = L1 Xtπ and qt = L2 Xtπ . LES with L2 = 0 corresponds to LEL with L = L1 . LEL: any loss in [t, t + ∆t] can be hedged with L% of the portfolio value. LES: any loss greater L2 % of the portfolio value in [t, t + ∆t] can be hedged with L1 % of the portfolio value. LEL & LES: For hedging we can use standard European call and put options. Deﬁnition ˛ ˆ ˘ ˜ ¯ KtLES := ηtπ ∈ Rn ˛˜ (∆Xtπ + qt )− |FtS < εt E Lemma KtLEL and KtLES are convex. For n = 1 we obtain the interval KtLES = [ηtl , ηtu ].
16 / 34

bounds on η π for LEL and LES

10
Bounds on η π Bounds on η π

10

0

0

−10 2 1
L (in % of Wealth)

5

−10 2 1
L1

10 5 0 0
L2

0 0

∆t (in days)

17 / 34

bounds on η π for LEL

8 6 4 2 0 0

Upper bound on η π

L = 2% L = 1.2% L = 0.4%
Lower bound on η π

0 −2 −4 −6 −8 0
L = 2% L = 1.2% L = 0.4%

1

2 3 ∆t (in days)

4

5

1

2 3 ∆t (in days)

4

5

18 / 34

Other constraints
Value-at-Risk constraint: Under the original measure ∆Xtπ is given by “Z ∆Xtπ = Xtπ exp “ “Z × exp
t t t+∆t ` t+∆t

” rs ds − Xtπ + (ηtπ ) Xtπ ´ 1 diag(σs σs ) ds + 2 Z
t t+∆t

µs −

” “Z σs dWs − exp
t

t+∆t

rs ds

””

.

We impose for n = 1 the relative VaR constraint on the loss (∆Xtπ )− , ` ´ P (∆Xtπ )− > LXtπ |FtS , µt = µt < γ . ˆ VaR is computed under the original measure P. Under partial information we need the (unknown) value of the drift → use e.g. µt = µt . ˆ For n = 1 we obtain the interval K VaR = [ηtl , ηtu ]. If n > 2 then K VaR may not be convex! Possible to apply a large class of other risk constraints e.g. CVaR.
19 / 34

Strategy

Corollary (Logarithmic utility) U(x) = log(x), n = 1, no constraints: ηto := ηtπ = With constraints: 8 >ηtu > > < c π ηt := ηt = ηto > > > l :η
t

1 (ˆt − rt ) . µ σt2 if ηto > ηtu , ˆ ˜ if ηto ∈ ηtl , ηtu , if ηto < ηtl .

Hence, we cut off the strategy obtained under no constraints if it exceeds or falls below a certain threshold.

20 / 34

Outline

Model Setup Problem formulation Time-Dependent Convex Constraints Dynamic Risk Constraints Gaussian Dynamics for the Drift A hidden Markov Model (HMM) for the Drift Example

21 / 34

Gaussian Dynamics (GD) for the Drift
Drift: modeled as the solution of the stochastic differential equation (cf. Lakner ’98) ¯ dµt = κ(¯ − µt ) dt + υ dWt , µ µ0 ∼ N (ˆ0 , ρ0 ), n-dimensional, µ ¯ W is a n-dimensional Brownian motion with respect to (F , P), We are in the situation of Kalman-ﬁltering with signal µ, observation R, and ˆ ˛ ˜ ﬁlter µt = E µt ˛ FtS . ˆ Filter: µt is the unique F S -measurable solution of ˆ ˆ` ´ ˜ dˆt = −κ − ρt (σt σt )−1 µt + κ¯ dt + ρt (σt σt )−1 dRt , µ ˆ µ ρt = −ρt (σt σt )−1 ρt − κρt − ρt κ + υυ ˙ with initial condition (ˆ0 , ρ0 ). µ ˜ ζ −1 satisﬁes dζt−1 = ζt−1 (ˆt − rt 1n ) (σt )−1 dWt . µ Proposition F S = F R = F W = F V → an optimal trading strategy exists.
22 / 34

,

˜

Bayesian case

The Bayesian case is a special case of the Gaussian dynamics for the drift. Drift: µt ≡ µ0 = (µ0 , . . . , µ0 ) is an (unobservable) F0 -measurable Gaussian ˆ random variable with known mean vector µ0 and covariance matrix ρ0 . Filter: Explicit solution: Z t Z t ” “ ”−1 “ (σs σs )−1 dRs , µt = 1n×n + ρ0 ˆ (σs σs )−1 ds µ0 + ρ 0 ˆ
0 0 (1) (n)

Z t ”−1 “ (σs σs )−1 ds ρ0 . ρt = 1n×n + ρ0
0

23 / 34

Outline

Model Setup Problem formulation Time-Dependent Convex Constraints Dynamic Risk Constraints Gaussian Dynamics for the Drift A hidden Markov Model (HMM) for the Drift Example

24 / 34

HMM: The Drift
The drift process µ of the return, is a continuous time Markov chain given by µt = BYt , B ∈ Rn×d ,

where Y is a continuous time Markov chain with state space the standard unit vectors {e1 , . . . , ed } in Rd , and rate matrix Q ∈ Rd×d , where
Qkl is the jump rate or transition rate from ek to el , P λk = −Qkk = d l=1,l=k Qkl is the rate of leaving ek , the waiting time for the next jump is exponentially distributed with parameter λk and Qkl /λk is the probability that the chain jumps to el when leaving ek for l = k .

The different states of the drift are the columns of B. We can write the market price of risk as θt = σt−1 (µt − rt 1n ) = Θt Yt , where Θt := σt−1 (B − rt 1n×d ) .

25 / 34

HMM: Filtering

We are in the situation of HMM ﬁltering since Rt = We need

Rt
0

BYs ds +

Rt
0

σs dWs .

ˆ ˜ the conditional density ζ = (ζt )t∈[0,T ] = E Zt |FtS = 1 1E , d t ˆ ˜ ˜ −1 the unnormalized ﬁlter E = (Et )t∈[0,T ] = E ZT Yt |FtS , ˆ ˜ ˆ ˆ the normalized ﬁlter Y = (Yt )t∈[0,T ] = E Yt |FtS = 1 EtE = ζt Et .
d t

Theorem (Wonham/Elliott) Z Et = E[Y0 ] +
0 t

Z Q Es ds +
0

t

˜ Diag(Es )Θs dWs

Proposition F S = F R = F W = F V → an optimal trading strategy exists.
˜

26 / 34

Outline

Model Setup Problem formulation Time-Dependent Convex Constraints Dynamic Risk Constraints Gaussian Dynamics for the Drift A hidden Markov Model (HMM) for the Drift Example

27 / 34

Example (1/3)

120 110 100 90 80
S

0.4 0.2 0 −0.2 −0.4
µ ˆ BY

70 0

0.2

0.4 0.6 Time

0.8

1

0

0.2

0.4 0.6 Time

0.8

1

We consider the HMM for the drift.

28 / 34

Example ct’d (2/3)

2
σ

0.35 0.3 0.25 0.2 0

1 0 −1 −2
0.2 0.4 0.6 Time 0.8 1
η ηl ηu

0

0.2

0.4 0.6 Time

0.8

1

For the volatility we consider the Hobson-Rogers model.

29 / 34

Example ct’d (3/3)

1.6 1.4 1.2 1

0.8 0 0.2 0.4 0.6 Time 0.8 1

30 / 34

Numerical Results (1/2)

We consider 20 stocks of the Dow Jones Industrial Index We use daily prices (adjusted for dividends and splits) for 30 years, 1972–2001 Parameter estimates are based on ﬁve years with starting year 1972, 1973,..., 1996 using a Markov Chain Monte Carlo algorithm. We apply the strategy in the subsequent year → we perform 500 experiments whose outcomes we average. We consider LEL- and LES-constraint.

31 / 34

Numerical Results ct’d (2/2)
ˆ U(XT ) b&h Merton GD Bayes HMM GD Bayes HMM GD Bayes HMM mean 0.1188 0.0248 -1.2002 0.0143 -0.0346 0.0252 0.1002 0.1285 -0.0395 0.0950 0.1505 median unconstrained 0.1195 0.0826 -1.0000 0.0824 0.0277 0.0294 0.0988 0.1242 -0.0350 0.0968 0.1402 0.2297 0.4815 0.9580 0.5071 0.9247 0.1767 0.1595 0.2004 0.3086 0.2752 0.3434 0 2 79 2 13 0 0 0 0 0 0 st.dev. aborted

LEL risk constraint (L=0.5%)

LES risk constraint (L1=0.1%,L2=5%)

LEL and LES improve the performance of all models. With LEL and LES we don’t go bankrupt anymore. The HMM strategy with risk constraints outperforms all other strategies.
32 / 34

Conclusion & Outlook
Conclusion We show how to apply dynamic risk constraints using time-dependent convex constraints. We derive explicit trading strategies with dynamic risk constraints under partial information. The numerical results indicate that dynamic risk constraints can reduce the risk and improve the performance. Outlook Allow for consumption. More detailed analysis of the multidimensional case. Explicit strategies for general utility.

33 / 34