Calculating Risk Neutral Probabilities and Optimal

Portfolio Policies in a Dynamic Investment Model with

Downside Risk Control

Yonggan Zhao

Nanyang Business School

Nanyang Technological University, Singapore 639798

E-mail: aygzhao@ntu.edu.sg

William T. Ziemba

Sauder School of Business

University of British Columbia, Canada V6T 1Z2

E-mail: ziemba@interchange.ubc.ca

June 25, 2004

 Calculating Risk Neutral Probabilities and Optimal

Portfolio Policies in a Dynamic Investment Model with

Downside Risk Control

Abstract

This paper presents a method for solving multiperiod investment models with down-

side risk control characterized by the portfolio’s worst outcome. The stochastic program-

ming problem is decomposed into two subproblems: a nonlinear optimization model

identifing the optimal terminal wealth and a stochastic linear programming model

replicating the identified optimal portfolio. The replicating portfolio coincides with the

optimal solution to the investor’s problem if the market is frictionless. The multiperiod

stochastic linear programming model is designed to test for the existence of arbitrage

opportunities and its dual solutions generate all risk neutral probability measures.

Keywords: Risk Neutral Probabilities, Dynamic Investment Model, Multiperiod Stochas-

tic Programming

1. Introduction

Intertemporal investment models in continuous time have been studied by Merton (1969,

1971), Karatzas et al (1986, 1987), Harrison and Pliska (1981), Pliska (1986), Karatzas

(1989), Cox and Huang (1989), among others. Generally speaking, two approaches are

considered for analyzing these problems: stochastic control and martingale analysis. The

key tool used for stochastic control is the Bellman equation and a nonlinear partial differ-

ential equation that the derived utility (or the value function) satisfies. This approach is

2
intractable if market prices are arbitrarily specified and the market is constrained. The

martingale analysis can reduce the complexity by decomposing the problem to a static prob-

lem of identifying the optimal terminal portfolio value from the attainable set assuming the

absence of arbitrage opportunities. Then, a replication model can be implemented to derive

the optimal strategy. Assuming that market prices follow geometric Brownian motions and

that the utility function is HARA, analytic solutions can be derived with both approaches.

However, for the markets with trading frictions, such as liquidity constraints or asset hold-

ing and trading constraints, closed form solutions are not available. Nonetheless, numer-

ical solutions are not readily obtainable for markets with trading frictions. Bertsimas,

Kogan and Lo (2001) studied a derivatives hedging problem for incomplete markets in a

continuous time framework. This paper considers a discrete time investment environment

with general asset return processes characterized by scenarios.

The investor is assumed to be trading periodically to maximize the expected utility of

terminal wealth with downside risk control characterized by the portfolio’s worst outcome

among all possible scenarios at the end of horizon. With a similar objective function, Zhao

and Ziemba (2000) studied a dynamic portfolio model which has a superior performance for

a given data to the standard portfolio strategies, such as, constant portfolio insurance, buy

and hold, and fixed mix. Zhao, Haussmann, and Ziemba (2003) studied a similar approach

and derived an analytic portfolio strategies, assuming the market is complete and friction-

less. In this paper, the discrete time stochastic control model is analytically intractable

for arbitrary asset returns. Even with suitable assumptions on the asset return processes,

such as normal distributions, the discrete time model is still not easy to solve because of its

feature as a nonlinear multiperiod stochastic optimization model. We must look for other

methodologies to tackle the problem. Cox and Huang (1989) provided a martingale method

for solving continuous time models in an unconstrained market. The wealth level is con-

strained to be nonnegative, but there are no other constraints on the portfolio weights.

Pliska (1998) proposed a solution method for multiperiod stochastic models using martin-

3
gale method in a discrete time version of its continuous time analog in a complete market

setting. We develop our approach along this line. The first task for the martingale method

is to identify a scenario set which excludes arbitrage opportunities. Finding a risk neutral

probability is the second task which is needed as inputs to the static model that identi-

fies the optimal terminal wealth. A multiperiod stochastic linear programming model is

developed to test for the existence of arbitrage opportunities, to obtain all risk neutral

probabilities by solving its dual, and to solve for the optimal portfolio strategy.

After identifying the optimal terminal wealth the implementation of this model must be

used to derive the multiperiod and scenario-wise optimal investment strategy. The theory

is well developed but the computations need procedures as developed here. Based on the

assumption that the wealth at each time is a function of the state variables in continuous

time models, partial differential equations can be derived to solve for the optimal portfolio

strategies. However, this method tends to obscure the role of the optimization methodology

in the discrete time framework. Multi-period stochastic linear programming is a useful

tool for implementing planning models under uncertainty and it has made major improve-

ments to the practice of investment management. Edirisinghe, Naik and Uppal (1993)

applied a stochastic programming model for option replication with transaction costs and

trading constraints by minimizing the initial costs of an European call option. Cariño,

Ziemba et al (1994, 1998ab) successfully developed a planning model for a Japanese insur-

ance company; see also Ziemba and Mulvey (1998) for a survey of additional applications.

In this paper, multiperiod stochastic programming is used for identifying the existence of

market arbitrage opportunities as well as the implementation of the replication model.

Trading periodically, we can obtain the optimal investment portfolio weights by replicating

the terminal portfolio value scenario by scenario and minimizing the expected downside

replicating error. It is proved that the replicating portfolio is exactly the optimal portfolio

identified in the first step if the market is unconstrained. With trading constraints, liq-

uidity constraints, shorting costs and transaction costs, etc., this optimal portfolio value is

4
generally not perfectly replicable. Hence, the replicating portfolio is not exactly the opti-

mal solution to the original problem in the presence of market frictions. Since the identified

portfolio is no worse than any optimal portfolio value with policy constraints, we call this

portfolio the “ideal” portfolio. The investor’s goal is to find a portfolio which is the closest

to the ideal portfolio in the sense of minimizing the expected downside deviation from the

ideal portfolio. This ideal portfolio is perfectly replicable in a frictionless market, but it is

not in a market with trading frictions. The discrepancy is characterized by the replicating

downside error.

In Section 2, we provide the basic characterization of arbitrage opportunities, risk neu-

tral probabilities and the derived stochastic linear programming for an unconstrained mar-

ket. Section 3 discusses market constraints and the heuristic solution methods. Section 4

provides a numerical example through which the comparisons are made between the opti-

mal solutions for the markets with frictions and without frictions. Section 5 concludes the

paper.

2. Model Setting and Formulation

2.1. The Objective Function and the Scenario Tree

We assume that the investment objective is to maximize the expected utility of consump-

tion and/or terminal wealth. The utility function reflects the investors’ risk attitudes. How-

ever, utility functions allow portfolio values to be zero or even negative. Therefore, they do

not have explicit control on the downside losses. This is not acceptable for investors who

have a stream of financial obligations. We assume that, to incorporate downside risk con-

trol to investment models, investors must pay attention to the portfolio’s worst case out-

come and to choose a better objective function for the optimization model. Knowing that

the objective is to maximize the expected utility while increasing the portfolio’s worst case

outcome, investors can adopt an approach that incorporates the portfolio’s worst outcome

5
as an additional endogenous decision variable. An objective function can be the convex

combination of the utility from the actual wealth and the worst case outcome, i.e.,

ρU (W ) + (1 − ρ)U (K), (1)

where U (·) is a standard utility function, W is the portfolio payoff, and K is the portfolio’s

worst case outcome; see Zhao and Ziemba (2001) for a discussion of this approach. This

objective function exhibits explicit downside control while maximizing utility of wealth.

If ρ = 1, (1) becomes a traditional utility maximization problem and ρ = 0 represents

the problem of risk free cash positions. The choice of ρ reflects the investors’ risk control

intensity. Consider an investor with a log utility function, initial wealth $100 and a gamble

that pays $1 with probability 0.8 and −$2 with probability 0.2. Without downside risk

control (ρ = 1), the investor will accept this gamble. However, if the investor is very

concerned about the downside losses and uses the objective function (1), the investor with

ρ < 0.8375 would reject the gamble.

The asset returns are modeled as a vector stochastic process, rt = (rt0 , · · · , rtn ), for

t = 1, · · · , T , which may be both serially and cross sectionally correlated. The filtration

generated by rt consists of the σ-fields, Ft = σ(r1 , · · · , rt ), t = 1, · · · , T . The market

uncertainty is described as a scenario tree which specifies the information structure con-

cerning the security returns revealed to the investor over time; see Figure 1.

An important issue which has not been discussed in academic papers on discrete time

stochastic optimization models is how to test for the nonexistence of arbitrage opportuni-

ties. Admitting of arbitrage opportunities can result in a poorly designed portfolio struc-

ture. Hence, it is crucial for developing the martingale method to have a non-arbitrage

scenario set. We will discuss this issue below.

6
 Stage 0 Stage 1 Stage 2 Stage T-1 Stage T

Figure 1: A Scenario Tree Over Time

2.2. A Model without Trading Frictions

Unlike some continuous time models where the portfolio is characterized as either the

number of units of assets or proportional amount of the total wealth in each period, the

portfolio weights are defined to be the amount of wealth allocated in each asset. Then the

replication model can be formulated as a problem of stochastic linear programming with

simple recourse; see Birge and Louveaux (1997). Let xt = (xt0 , · · · , xtn )> be the amount

of wealth held in the riskless asset (xt0 ) and other risky assets (xti , i = 1, · · · , n). Assume

that xt are Ft -measurable for t = 1, · · · , T − 1, so that xt is non-anticipative. Each

scenario is determined by a single path ω = (ω0 , · · · , ωT ), where ωt represents a single

path of the information up to time t.

Assume there are no trading, liquidity, shorting nor transaction costs. The goal of the

investor is to maximize the expected utility of the terminal wealth with control on downside

7
losses. The investment problem is

max E[ρU (W ) + (1 − ρ)U (K)]

K,W,xt

s.t. x>
0 · 1 ≤ W0

(1 + rt )> xt−1 − 1> · xt ≥ 0, ∀t = 1, 2, · · · , T

(2)
>
(1 + rT ) · xT −1 − W ≥ 0

W − K ≥ 0, ∀ω ∈ Ω.

xt is Ft measurable.

where W0 is the initial portfolio value. Model (2) is a nonlinear multiperiod stochastic

optimization problem whose solution is not readily obtainable. We decompose the problem

into a static nonlinear optimization and a multiperiod stochastic linear programming.

2.3. Arbitrage and Risk Neutral Probabilities

Solving the above optimization model is intractable in a discrete time framework. Cox and

Huang (1989) solved a continuous analog of the above model using martingale method,

assuming market is complete. The martingale method is applicable only if there exists a

risk neutral probability. Harrison and Kreps (1979) showed that an arbitrage-free market

implies the existence of a risk neutral probability measure such that all prices of securities

discounted at the risk free rate are martingales. However, in a discrete framework, risk

neutral probability measure is not guaranteed to uniquely exist and risk neutral proba-

bility measure is not readily obtainable. We must know first that the financial market

is exclusive of arbitrage. This leads us to the characterization of arbitrage free with a

stochastic linear programming model. The usual definition of arbitrage free is

Definition 1. Let W be the payoff of a portfolio with initial cost W0 . A market is called

arbitrage free if W ≥ 0 with Pr[W > 0] > 0 implies W0 > 0.

The above definition implies that there is no “free lunch” – if a portfolio’s payoff is no

less than zero and strictly greater than zero with a positive probability, then the initial cost

8
of such a portfolio must be strictly greater than zero. The following multiperiod stochastic

linear programming characterizes the nonexistence of arbitrage.

Lemma 1. Assume the existence of a riskless asset. A market is arbitrage free if and only if

the following stochastic linear programming model

min W0
W0 ,xt

s.t. W0 − 1> x0 ≥ 0

(1 + r1 )> x0 − 1> x1 ≥ 0 (3)

..
.

(1 + rT )> xT −1 ≥ 0

is solvable and all constraints are binding at optimality.

Proof. Assume the nonexistence of arbitrage. Since W0 = 0 and xt = 0 for all t constitute

a feasible solution, model (3) is feasible. Since the market is arbitrage free, W0 ≥ 0. Oth-

erwise, one can find a portfolio strategy over time, x0 , · · · , xT −1 , such that the final payoff

of the portfolio is greater than or equal to 0, but the initial cost W0 < 0. This contradicts

the market nonexistence of arbitrage. Hence, model (3) is bounded and therefore solvable.

Furthermore, all constraints must be binding at optimality by the assumption of arbitrage

free.

Conversely, assume model (3) is solvable and all constraints are binding at optimality.

Let x0 , · · · , xT −1 be a portfolio strategy over time, such that the terminal payoff of the

9
portfolio W = (1+rT )> xT −1 ≥ 0 with Pr[W > 0] > 0. The dual of model (3) is

max 0
y0 ,··· ,yT

s.t. y0 (ω0 ) = 1
X
(1 + r1 )y1 (ω1 ) − y0 (ω0 ) = 0
ω1 (4)
..
.
X
(1 + rT )yT (ω0 , · · · , ωT ) − yT −1 (ω0 , · · · , ωT −1 ) = 0
ωT

y0 (ω0 ), y1 (ω0 , ω1 ) ··· yT (ω0 , · · · , ωT ) ≥ 0

By the strong duality, the initial cost, W0 = 1> x0 , of the portfolio strategy, x0 , · · · , xT −1 ,

must be nonnegative. Since the constraints for the final period is not binding, the portfolio

strategy, x0 , · · · , xT −1 , is not optimal. Therefore, W0 > 0. Thus, the market is arbitrage

free by definition.

The solution to Model (3) can be interpreted as the optimal strategy for the portfolio of

payoff 0 in any scenario. Such a portfolio in the market will have a zero initial cost in a

frictionless market.

We now relate nonexistence of arbitrage to a risk neutral probability. A risk neutral

probability measure is defined as a probability measure under which all one-period condi-

tional asset returns are equal to the riskless rate. Consider the dual of the model (3). Let

q0 (ω0 ), q1 (ω0 , ω1 ), · · · , qT (ω0 , · · · , ωT ) be a feasible solution to the dual model (4) for a

scenario ω = (ω0 , · · · , ωT )> . Given the information up to t

X
qt (ω0 · · · , ωt ) · (1 + rt (ωt |Ft−1 )) = qt−1 (ω0 , · · · , ωt−1 ) · 1, 1≤t≤T (5)
ωt ∈Ωt

where Ωt is the set of possible scenarios up to time t for a given ωt−1 . Hence,

X qt (ω0 ), · · · , ωt )
(1 + rt0 ) = 1,
ωt ∈Ωt
qt−1 (ω0 , · · · , ωt−1 )

where rt0 is the risk free rate for the t-th period, and

X qt (ω0 ), · · · , ωt )
(1 + rt0 ) · rt = rt0 .
ωt ∈Ωt
qt−1 (ω0 , · · · , ωt−1 )

10
 qt (ωt )
Therefore, qt−1 (ωt−1 )
(1 + rt0 ) is the conditional risk neutral probability of the state ωt for

given state ωt−1 . So, let

Y
Q : (ω0 , · · · , ωT ) → (1 + rt0 ) · qT (ω0 , · · · , ωT ).
1≤t≤T

Then, Q defines a risk neutral probability measure on Ω. Conversely, it can be argued

that any risk neutral probability can induce a feasible solution to model (4). The following

lemma characterizes the relationship between a risk neutral probability measure and a

feasible solution to (4).

Lemma 2. The set of the feasible solutions to model (4) is 1-1 correspondent to the set of all

risk neutral probability measures.

Finding all feasible solutions to model (4) is equivalent to solving a large scale linear

system with nonnegativity restriction on the unknowns. The size of the problem is large,

but a decomposition method can be applied. One can first obtain the conditional risk neu-

tral probabilities at each node of the scenario tree. Then multiplying all conditional risk

neutral probabilities along each scenario path from the root to an ending node obtains a

risk neutral probability measure for the whole scenario tree. The feasible set of model

(4) is a convex polyhedron. That is, if Q1 and Q2 are two risk neutral probabilities in-

duced by two feasible solutions to (4), then so is λ1 Q1 + λ2 Q2 , where λ1 + λ2 = 1 and

λ1 ≥ 0, λ2 ≥ 0. Hence, all risk neutral probabilities can be generated by s subset which

are induced by the vertexes of the polyhedron. Denote by < the subset of the risk neutral

probabilities induced by the vertexes of the solutions to model (4).

2.4. The Martingale Method

Let Wt = (1 + rt )> xt−1 be the wealth at time t. Denote E Q [·] as the operator of expecta-

tion under a risk neutral probability Q generated by a dual feasible solution to model (4).

The wealth Wt at time t is equal to 1> xt . The conditional expectation of discounted wealth

11
is
     
 
Y Y
EQ  (1 + ri0 )−1 Wt  Ft−1  = E Q  (1 + ri0 )−1 · (1 + rt )> xt−1  Ft−1 
 

1≤i≤t  1≤i≤t 
Y
= (1 + ri0 )−1 · 1> xt−1 (6)
1≤i≤t−1
Y
= (1 + ri0 )−1 · Wt−1 .
1≤i≤t−1

−1
Q
Theorem 1. The discounted wealth process 1≤i≤t (1 + ri0 ) · Wt under the risk neutral

probability is a martingale. The optimal portfolio can be obtained by solving the following

static concave programming problem

max E[ρU (W ) + (1 − ρ)U (K)]

K,W
 
Y
s.t. EQ  (1 + ri0 )−1 W  = W0 , ∀Q ∈ <. (7)
1≤i≤T

W − K ≥ 0, ∀ω ∈ Ω

where K is constant for ∀ ω ∈ Ω and W is FT measurable.

Proof. Since the probability sample space Ω is finite, Equation (6) proves that the dis-

counted wealth process is a martingale under any risk neutral probability. Let (W, K)

be the optimal solution to model (7). We need to show that there exists a portfolio strat-

egy, (x0 , · · · , xT −1 ), such that (W, K) and (x0 , · · · , xT −1 ) are the optimal solution to

model (2). Consider the following stochastic linear programming problem that minimizes

the initial cost of (super) replicating the portfolio payoff W , assuming no trading frictions.

min W̄0
W̄0 ,xt

s.t. W̄0 − 1> x0 ≥ 0

(1 + r1 )> x0 − 1> x1 ≥ 0 (8)

..
.

(1 + rT )> xT −1 ≥ W.

12
The dual of Model (8) is
X
max q(ω0 , · · · , ωT )W
y0 ,··· ,yT
(ω0 ,··· ,ωT )

s.t. y0 (ω0 ) = 1
X
(1 + r1 )y1 (ω1 ) − y0 (ω0 ) = 0
ω1 (9)
..
.
X
(1 + rT )yT (ω0 , · · · , ωT ) − yT −1 (ω0 , · · · , ωT −1 ) = 0
ωT

y0 (ω0 ), y1 (ω0 , ω1 ) ··· yT (ω0 , · · · , ωT ) ≥ 0

which is feasible by the existence of a risk neutral probability. Since its feasible set is

the same as that of (4), the objective value of model (9) equals W0 . Hence, model (9) is

solvable. By the strong duality, model (8) is also solvable and the optimal W̄0 = W0 . Let

(x0 , · · · , xT −1 ) be the optimal portfolio solution to model (8). Then, it is easy to verify that

(W, K) and (x0 , · · · , xT −1 ) constitute the optimal solution to model (3).

From Theorem 1, we have decompose the original multiperiod stochastic nonlinear pro-

gramming model into a static nonlinear program and a multiperiod stochastic linear pro-

gram. The computational complexity has been greatly reduced. Assume that the utility

function is twice continuously differentiable, then the optimal (W, K) can be obtained di-

rectly using the Lagrange multiplier method. Let

qt (ωt )
ξt (ωt ) = , ∀ ωt ∈ Ft , (10)
pt (ωt )

where pt (ωt ) is the physical probability for scenario ω up to time t information. ξt is

usually called the state price density or the Arrow-Debreu price per unit probability pt of

one dollar in state ωt at time t.

Let λ0 and λ = λ(ω) be the Lagrange multipliers for the constraints of model (7). The

Lagrangian is

L(W, K, λ0 , λ) = E[ρU (W ) + (1 − ρ)U (K)] − λ0 (E [ξT W ] − W0 ) + E [λ(W − K)].

13
The extended Kuhn Tucker conditions (for proof, see Zhao, Haussmann and Ziemba (2003))

are

(i). ρUx (W ) − λ0 ξT + λ = 0, P - a.s.,

(ii). (1 − ρ)Ux (K) − E[λ] = 0,

(iii). E [ξT W ] − W0 = 0,

(iv). λ(W − K) = 0, P - a.s.,

(v). λ ≥ 0, W ≥ K > 0, P - a.s.

where Ux (·) is the first order derivative of U (·). Let Ux−1 (·) be the inverse function of

Ux (·). The optimal W and K are related through

+
W = K + Ux−1 ρ−1 λ0 ξT − K ,


(11)

where λ0 and K are determined by


h
i
(1 − ρ)Ux (K) − ρE ρ−1 λ0 ξT − Ux (K) + = 0



(12)
h
i
E[KξT ] + E ξT Ux−1 ρ−1 λ0 ξT − K + = W0 .





By definition, the state price is given by the risk neutral probability, the physical prob-

ability, and the riskless interest rate. Therefore, the optimal wealth for a given utility

function can be determined by the state prices upon obtaining the risk neutral probabil-

ities. To illustrate, assume that the investor has a log utility function. Figure 2 shows

the relation between terminal wealth and the state price with varying risk control inten-

sity ρ. As ρ decreases, which indicates the investor becomes more risk averse, the level of

the portfolio’s worst outcome K increases and the expected value of the wealth decreases

simultaneously.

For a given ρ, if W (ρ) attains the minimum wealth K(ρ), then there exists a ξ ∗ (ρ)

such that

λ0 (ρ) · ξ ∗ (ρ) = Ux−1 (K(ρ)) .

14
 3 Uncontrolled
Controlled
Replicated

Portfolio Return
1.5

0
0 1.5 3
State Prices

Figure 2: Relation between Wealth and State Prices for Varying ρ

If the state price is below ξ ∗ (ρ), the investor will follow a traditional utility maximization

as if there were no downside risk control. In that case, the downside risk control constraint

is not binding.

After identifying the optimal terminal portfolio, we must apply the stochastic linear

programming model formulated as in (3) to complete the implementation - to derive the

multiperiod optimal portfolio strategy. Theorem 1 and its proof implies that the replicating

portfolio is exactly the optimal portfolio if the market is frictionless.

Theorem 2. The optimal value of model (2) is given by equation (11) and the optimal

solution to model (2) is given by the optimal solutions (x0 , · · · , xT −1 ) to model (8).

Proof. Equation (11) can be derived using the Kuhn-Tucker conditions (i), (iv) and (v). The

proof of Theorem 1 implies that the optimal solution x0 , · · · , xT −1 to Model (8) is also the

optimal solution to Model (2).

Theorem 2 presents a method for solving the general investment model by decomposing

the original problem into two subproblems: a static model and a replication model. The

static (but usually large scale) nonlinear optimization model identifies the terminal wealth

15
which satisfies the downside risk constraint, while the replication model is a multiperiod

stochastic linear program. The IBM stochastic solution optimization routine library is a

useful package for solving such a large scale stochastic linear programming model.

3. A Model for the Market with Trading Frictions

Section 2 discussed how to solve an investment problem with downside risk control in an

unconstrained market. The static nonlinear optimization model is easy to solve after ob-

taining all risk neutral probabilities. These probabilities are given by the dual feasible

solutions of the stochastic linear programming model (3) which checks for the existence of

arbitrage opportunities. There are algorithms that can provide the primal and dual so-

lution simultaneously. However, real markets have constraints such as those related to

trading, liquidity, transactions costs, etc. With these constraints added to the investor’s

problem (Model (2)), the new optimization model can not be easily decomposed into such

two subproblems, because the investor’s wealth process is no longer a martingale. Per-

fect replication of such a terminal wealth is generally impossible. However, knowing that

the optimal terminal wealth for an unconstrained market is superior to a terminal wealth

subject to the market constraints, one can start to replicate such a portfolio while all con-

straints are satisfied and the downside replicating error is minimized. The error may not

be zero as it is for an unconstrained market. However, this portfolio is the closest to the

optimal portfolio with no constraints and, at the same time, it satisfies the necessary con-

straints of the original problem.

Trading constraints frequently require that a minimum and/or maximum amount is

required when trades occur. These constraints prevent arbitrage opportunities. Let yt =

(yt1 , · · · , ytn )> and zt = (zt1 , · · · , ztn )> represent the amounts bought and sold of the

16
risky assets, respectively. The trading constraints are

0 ≤ yt ≤ αt
(13)
0 ≤ zt ≤ β t ,

where αt = (αt1 , · · · , αtn )> , β t = (βt1 , · · · , βtn )> represent the buying and selling up-

per bounds of the amount for each asset traded in each period, respectively.

Liquidity is usually defined as the ability to transact immediately and with negligibly

small impact on the price of a security regardless the size of the transaction. One of the

distinguishing features inherent in illiquid markets is a frequent inability to buy or sell an

asset at its temporary equilibrium price. The reason is that not all the information avail-

able about the asset is fully reflected in its current return and hence the asset behavior

becomes locally predictable, i.e., an excess demand will result in the increase of the stock

return over the next time-step and likewise an excess supply will result in a decrease of the

return. As a result, if a stock return is exceeding or going to exceed the riskless interest

rate, the stock is unlikely to be available for purchase at its intrinsic price. Similarly, a

falling market leads to an inability to sell the stock at its current intrinsic price. To accom-

modate these liquidity constraints, we specify the following periodic holding constraint for

each asset.

γ t · x> >
t 1 ≤ xt ≤ Γt · xt 1, (14)

where γ t = (γt1 , · · · , γtn )> and Γt = (Γt1 , · · · , Γtn )> represent the limit percentages of

the portfolio wealth held in each asset for period t.

Transaction costs are of two types. The first are brokerage fees and the second is to

prevent frequent trades that might affect the equilibrium stock prices. For simplicity, we

model these costs as proportional amounts to the transaction volume from asset to asset.

Let θ t = (θt1 , · · · , θtn )> be the proportional transaction costs for the risky assets. Buying

and selling the same asset in each period is not optimal. At time t, the amount xt0 in the

17
riskless asset is

xt0 − (1 + rt0 )xt−1,0 + (1 + θ t )> yt − (1 − θ t )> zt = 0. (15)

The amount xti in risky assets at time t is

xti − (1 + rti )xt−1,i − yti + zti = 0, ∀ i = 1, · · · , n. (16)

The initial constraints are

x00 + (1 + θ0 )> y0 − (1 − θ0 )> z0 = W0 ,

(17)
x0i − y0i + z0i = 0.

The terminal wealth W in unit of riskless asset is

X
(1 + rT 0 ) · xT −1,0 + (1 − θT i )(1 + rT i ) · xT −1,i − W = 0. (18)
1≤i≤n

The utility maximization model with market trading frictions is

max E[ρU (W ) + (1 − ρ)U (K)]

xt ,yt ,zt ,W,K

s.t. ( 13 − 18) (19)

W − K ≥ 0.

The solution to (19), a multiperiod stochastic nonlinear programming model, is generally

intractable. This is so large a model that even sophisticated software packages are not

able to solve it. Our aim is to decompose the problem into two problems analogous to the

frictionless market. The static model characterizes the frictionless optimal portfolio. The

replication model is different from its frictionless market version. It is not guaranteed that

perfect replication can be achieved. The replicating portfolio minimizes the expected value

of the downside deviation from the frictionless optimal portfolio with the downside control.

Let (W, K) be the optimal solution to the static nonlinear optimization model and denote
 
X
Z = W − (1 + rT 0 ) · xT −1,0 + (1 − θT i )(1 + rT i ) · xT −1,i  . (20)
1≤i≤n

18
Then, the stochastic linear program for replicating the frictionless optimal portfolio W is

min E[Z]

s.t. (13 − 17), (20) (21)

Z≥0

It is crucial to know the possible deviation of replication with trading frictions. Let

(W, K) be the optimal solution for the market with trading frictions, (W1 , K1 ) be the

optimal solution for the frictionless market and (W ∗ , K ∗ ) be the solution obtained by the

optimal replicating portfolio. Then

E[U (W )] ≤ E[U (W1 )],

W1 − W ∗ ≤ Z, P − a.s. (22)

K ∗ ≥ max{W1 (ω) − Z(ω) : ∀ ω ∈ Ω}.

The difference of the optimal expected utilities between Model (19) and the replication with

market trading frictions satisfies

E[ρU (W ) + (1 − ρ)U (K)] − E[ρU (W ∗ ) + (1 − ρ)U (K ∗ ]

≤ ρ (E[U (W1 )] − E[U (W ∗ )]) + (1 − ρ) (U (K1 ) − U (K ∗ )) (23)

≤ M · E[Z] + N |K1 − K ∗ | ,

where M = ρUx (K1 ) and N = (1 − ρ) max{K1 , K ∗ }. Equation (23) indicates that the

difference between the optimal solution and the replicating portfolio is bounded by a linear

combination of the expected downside replication error and the difference of the worst case

outcomes of the optimal portfolio and the replicating portfolio. As the initial wealth level

increases, the marginal utility diminishes. So, the replicating portfolio converges to the

optimal portfolio when the initial wealth amount becomes large.

19
4. Example

The investment opportunity set consists of the following five assets: Cash earning a con-

stant monthly risk free interest rate 0.4% and four other risky assets, the Dow Jones

Industrial Average (DJIA), the Lehman Government Bond index (LEHM), the Nasdaq

Composite (NSDQ), and the Standard & Poor 500 (S&P500). Denote the asset returns

by rt = (Dt , Lt , Nt , St )> . Assume that the one-period returns of these four assets follow

an identical and independent process which are multivariate normally distributed. Fig-

ure 3 depicts the monthly data from 01/01/1998 to 12/01/2000 of these four assets. Using

DJIA
LEHM
0.3 NSDQ
S&P500
0.2

0.1
Return

-0.1

-0.2

-0.3
12/1/97 12/1/98 12/1/99 11/30/00
Date

Figure 3: Monthly Index Returns (01/01/1998 - 12/01/2000)

standard mean and variance estimation methods yields

     
 Dt  0.0086  0.0028 −0.0003 0.0022 0.0023 
     
     
 Lt  0.0030 −0.0003 0.0015 −0.0005 −0.0002
  → N  ,  . (24)
     
0.0188  0.0022 −0.0005 0.0105
     
N t  0.0033 
     
     
St 0.0095 0.0023 −0.0002 0.0033 0.0024

One of the important issues in applied stochastic programming is how to generate an

arbitrage free scenario set for model inputs. Generating an arbitrage free scenario is even

more crucial in the model developed here because both the existence and the calculation of

20
the risk neutral probability rely on the arbitrage free assumption. For large-scale models,

this is a tedious task. However, the test procedure for the existence of arbitrage opportuni-

ties can be implemented by solving the large-scale stochastic linear programming problem

developed here. For simplicity, we implement the model with a four period investment hori-

zon and assume cross period identical independent distributions for the four risky assets.

To generate an appropriate scenario set, we require that the sample means and variances

are matched to the estimated means and variances. At each node, we take a sample of size

five as the possible scenarios for the next period. Based on (4), we developed an efficient

algorithm for generating scenarios that are exclusive to arbitrage opportunities and for

calculating the risk neutral probabilities. The data in Table 1 allows us to generate all 625

scenarios.

Table 1: Asset Return Scenario Distributions

States DJIA LEHM NSDQ S&P500 Probability

Outcome 1 0.0017 -0.0195 -0.1538 -0.0247 0.2418

Outcome 2 0.056 -0.0221 0.1180 0.06346 0.3331

Outcome 3 0.0039 -0.0096 0.0361 -0.0638 0.0411

Outcome 4 -0.0849 -0.0067 0.0007 -0.0561 0.2302

Outcome 5 -0.0012 0.0911 -0.0244 0.0024 0.1538

Although the risk neutral probability can be obtained by solving the dual of Model (3),

the problem can also be divided into a sequence of subproblems because of its separability

to reduce the computational complexity. Thus, if we can calculate the conditional risk neu-

tral probability at each node, the risk neutral probability can be obtained by multiplying

the conditional probabilities along each scenario path. For this example, we need only to

solve 1 + 5 + 25 + 125 = 156 linear systems of size 5 × 5.

We compare the replication result for the case ρ = 0.9 with a standard utility maxi-

mization (ρ = 1). Although the expected return is 1.1128 for ρ = 1 and 1.0735 for ρ = 0.9,

21
the portfolio’s worst outcome is 0.5021 for ρ = 1 and 0.8980 for ρ = 0.9. With a 95% con-

fidence interval, the portfolio would lose 0.3499 for ρ = 1 and only 0.1020 for ρ = 0.9.

As to the initial asset allocation, the uncontrolled strategy requires a much more volatile

portfolio weights than the controlled one as shown in Table 2.

Table 2: Initial Portfolio Weights

Asset Controlled Uncontrolled

Cash 0.0787 -0.6131

DJIA 0.6286 1.2818

LEHM -0.3345 -0.3276

NSDQ 0.6272 1.3782

S&P500 0.0000 -0.7192

This statistical result indicates that the large portfolio returns in high state prices

occur with tiny probabilities. Considering portfolio constraints, investors are interested in

achieving a wealth return corresponding to a middle range of state prices and can achieve

a reasonable portfolio rate return. Figure 4 depicts the difference of a replicating portfolio

and the optimal portfolio return. Investors are interested in knowing how the portfolio will

change as time goes along. The solution of a multiperiod stochastic programming model

facilitates the possibility of observing the portfolio change at each time period. The investor

can change the asset allocation strategy according to the provided optimal portfolio weights

or resolve the model with new input. Table 3 describes how the portfolio value and the

portfolio weights evolve for a typical scenario.

22
 3 Controlled
Replicated

Portfolio Return
1.5

0
0 1.5 3
State Prices

Figure 4: The Difference between the Optimal Portfolio and its Replication

5. Concluding Remarks

The relationships derived in this paper concerning arbitrage opportunities, risk neutral

probabilities and stochastic linear programs arise from analysis of duality relationships.

A very efficient test for market arbitrage opportunities is characterized using a stochastic

linear programming model. The calculation of the risk neutral probability is reduced to the

calculation of the periodic conditional risk neutral probabilities. Multiplying together the

conditional risk neutral probabilities along a scenario yields the risk neutral probability of

the scenario.

The investor’s problem is formulated as a multiperiod stochastic nonlinear program-

ming model. This model is intractable because of the non-linearity of the objective function

and a multiperiod horizon. Utilizing a martingale analysis, the problem is divided into two

models. The first model is a static problem for identifying the scenario-wise optimal port-

folio for a given utility function. The second model is to replicate the identified portfolio by

minimizing the downside expected replicating deviation. It is proved that the replicating

portfolio is the optimal portfolio identified in the first model for the unconstrained market

condition. However, with a constrained market, this replication might not be perfect. A

23
 Table 3: Portfolio Performance and Evolution for a Typical Scenario
Scenario 200 Stage 1 Stage 2 Stage 3 Stage 4 Stage 5

Cash 0.0787 0.0069 0.0878 0.0471

DJIA 0.6286 0.8012 0.9321 0.9389

LEMN -0.3345 -0.3465 -0.4580 -0.4008

NADQ 0.6272 1.1702 1.4883 1.6732

S&P500 0.0000 -0.5000 -0.8163 -1.0798

Sale (D) 0.0000 0.0000 0.0000

Sale (L) 0.0184 0.1138 0.0000

Sale (N) 0.0000 0.0000 0.0000

Sale (S) 0.5000 0.3435 0.2537

Purchase (D) 0.1311 0.1209 0.0000

Purchase (L) 0.0000 0.0000 0.1003

Purchase (N) 0.4572 0.2539 0.1933

Purchase (S) 0.0000 0.0000 0.0000

Wealth 1.0000 1.1319 1.2339 1.1787 1.1257

numerical example investigates how a reasonable portfolio strategy can be obtained. The

replicating portfolio is not practical (both long and short positions are extremely large) for

unconstrained markets. The replicating portfolio for the constrained market makes more

sense in terms of portfolio return and the reality of its position in assets. Statistically, at

the 95% level the constrained model is superior to the unconstrained model.

This paper also shows the applicability of stochastic programming methodology for de-

veloping dynamic investment models. The method of replication is extremely useful in

pricing contingent claims. We employed this idea to solve a complex investment model via

two simpler models. Considering the identified portfolio as a “contingent claim”, one can

24
implement a replicating strategy using finite resources subject to constraints. With a con-

strained market, a replicating error may exist which implies that there exists a computable

bound on the investor’s expected utility losses. With the analysis and statistical results of

the numerical example, it is evident that including constraints can increase portfolio per-

formance when a risk measure, such as Value at Risk, is considered. Due to its versatility

of decomposing a complex problem into two simpler models that can be easily implemented,

this method should be considered to be a very efficient way of solving practical models, both

constrained and unconstrained.

Acknowledgements. This research was partially supported by NSERC.

25
References

Bertsimas, D., L. Kogan, and A. W. Lo, 2001, “Hedging Derivative Securities and Incom-

plete Markets: An -Arbitrage Approach,” Operations Research, 49, 372–397.

Birge, J., and F. Louveaux, 1997, Introduction to Stochastic Programming, Springer-Verlag,

New York, 2 edn.

Black, F., and M. Scholes, 1973, “The Pricing of Options and Corporate Liabilities,” Journal

of Political Economy, 81, 637–654.

Brennan, M., and R. Solanki, 1981, “Optimal Portfolio Insurance,” Journal of Financial

and Quantitative Analysis, 16, 279–299.

Cariño, D., D. Kent, D. Myers, C. Stacy, M. Sylvanus, A. Tuner, K. Wanatabe, and

W. Ziemba, 1994, “The Russell-Yasuda Kasai Model: an Asset Liability model for a

Japanese Insurance Company Using multistage Stochastic Programming,” Interfaces,

24, 29–49.

Cariño, D., D. Myers, and W. Ziemba, 1998, “Concepts, Technical Issues, and Uses of the

Russell-Yasuda Kasai Financial Planning Model,” Operations Research, 46, 449–462.

Cariño, D., and W. Ziemba, 1998, “Formulation of the Russell-Yasuda Kasai Financial Plan-

ning Model,” Operations Research, 46, 433–449.

Cox, J., and C.-f. Huang, 1989, “Optimal Consumption and Portfolio Polices When Asset

Prices Follow a Diffusion Process,” Journal of Economic Theory, 49, 33–83.

Edirisinghe, C., V. Naik, and R. Uppal, 1993, “Optimal Replications of Options with Trans-

actions Costs,” Journal of Financial and Quantitative Analysis, 28, 117–138.

Evnine, J., and R. Henriksson, 1987, “Asset Allocation and options,” Journal of Portfolio

Management, 14, 56–61.

26
Grossman, S., and Z. Zhou, 1996, “Equilibrium Analysis of Portfolio Insurance,” Journal of

Finance, 51, 1379–1403.

Harlow, W., 1991, “Asset Allocation in a Downside-Risk Framework,” Financial Analysts

Journal, 47, 28–40.

Harrison, M., and D. Kreps, 1979, “Martingale and Multiperiod Securities Markets,” Jour-

nal of Economic Theory, 20, 382–408.

Harrison, M., and S. Pliska, 1981, “Martingales and Stochastic Integrals in the Theory of

Continuous Trading,” Stochastic Process Appl., 11, 215–260.

Karatzas, I., 1989, “Optimization Problems in the Theory of Continuous Trading,” SIAM J.

of Control and Optimization, 27, 1221–1259.

Karatzas, I., J. Lehoczky, and S. Shreve, 1987, “Optimal Portfolio and Consumption Deci-

sions for a “Small Investor” on a Finite Horizon,” SIAM J. Control & Optimization, 25,

1557–1586.

Karatzas, I., and S. Shreve, 1998, Methods of Mathematical Finance, Springer-Verlag, New

York.

Leland, H., and M. Rubinstein, 1995, “Replicating Options with Positions in Stock and

Cash,” Financial Analysts Journal, 51, 113–120.

Merton, R., 1971, “Optimal Consumption and Portfolio Rules in a Continuous -Time

Model,” Journal of Economic Theory, 3, 373–413.

Pliska, S., 1986, “A Stochastic Calculus Model of Continuous Trading: Optimal Portfolio,”

Mathematics of Operations Research, 11, 371–382.

Pliska, S., 1998, Introduction to Discrete Mathematical Finance, Blackwell Publishers Inc.

27
Zhao, Y., U. Haussmann, and W. Ziemba, 2003, “A Dynamic Investment Model with Control

on the Portfolio’s Worst Case Outcome,” Mathematical Finance, 13, 481–501.

Zhao, Y., and W. Ziemba, 2000, “A Dynamic Asset Allocation Model with Downside Risk

Control,” Journal of Risk, 3, 91–113.

Zhao, Y., and W. Ziemba, 2001, “A Stochastic Programming Model Using an Endogenously

Determined Worst Case Risk Measure for Dynamic Asset Allocation,” Mathematical Pro-

gramming, 89, 293–309.

28