Optimal Liquidation Problems in a RandomlyTerminated Horizon
Abstract
In this paper, we study optimal liquidation problems in a randomlyterminated horizon. We consider the liquidation of a large singleasset portfolio with the aim of minimizing a combination of volatility risk and transaction costs arising from permanent and temporary market impact. Three different scenarios are analyzed under AlmgrenChriss’s market impact model to explore the relation between optimal liquidation strategies and potential inventory risk arising from the uncertainty of the liquidation horizon. For cases where no closedform solutions can be obtained, we verify comparison principles for viscosity solutions and characterize the value function as the unique viscosity solution of the associated HamiltonJacobiBellman (HJB) equation.
Keywords: Dynamic Programming (DP) Principle; HamiltonJacobiBellman Equation; Randomlyterminated; Optimal Liquidation Strategies; Stochastic Control; Viscosity Solution.
AMS Subject Classifications: 35Q90, 49L20, 49L25, 91G80, 65M06.
1 Introduction
Understanding trade execution strategies is a key issue for financial market practitioners and has attracted growing attention from the academic researchers. An important problem faced by equity traders is how to liquidate large orders. Different from small orders, an immediate execution of large orders is often impossible or at a very high cost due to insufficient liquidity. A slow liquidation process, however, is often costly, since it may involve undesirable inventory risk. Almgren and Chriss [2] provided one of the early studies on the optimal execution strategy of large trades, taking into account the volatility risk and liquidation costs. In order to produce tractable and analytical results, they set the market impact cost per share to be linear in the rate of trading. Schied and Schoneborn [21] considered the infinitehorizon optimal portfolio liquidation problem for a von NeumannMorgenstern investor under the liquidity model of Almgren [1], in which a power law cost function was introduced to determine optimal trading strategies. However, most of the literature on optimal liquidation strategies mainly considered a known predetermined time horizon or infinite horizon. The case of unknown (or more precisely, randomlyterminated) time horizon is not fully addressed. In some situation, it is more realistic to assume that the liquidation horizon depends on some stochastic factors of the model. For example, some financial markets adopt the circuitbreaking mechanism, which makes the horizon of the investor subject to the stock price movement. Once the stock price hits the daily limits, all transactions of the stock will be suspended.
In this paper, we consider a randomlyterminated time horizon under three different scenarios that an agent might encounter in a financial market. AlmgrenChriss’s market impact model is employed to describe the underlying asset price:
where the constant is the absolute volatility of the asset price , is an onedimensional standard Brownian motion, is the actual transaction price, is an admissible control process, and represent, respectively, the permanent and temporary components of the market impact. We consider the liquidation problem of a large singleasset portfolio with the aim of minimizing a combination of volatility risk and transaction costs arising from permanent and temporary market impact.
We first consider the case with a predetermined time horizon , which can be used as a benchmark for other cases with randomlyterminated time horizon. In general, it is required that a liquidation strategy should satisfy the handsclean condition:
where is the number of shares held by the trader at the time . We first work on a subclass of deterministic controls, which do not allow for intertemporal updating, satisfying the handsclean condition. Obviously, the deterministic strategy obtained in the subclass might be no longer optimal when taking into account the entire class of admissible controls. We then temporarily relax the handsclean condition, and allow an immediate final liquidation (if necessary) so that the number of shares owned at the time is . We employ the dynamic programming (DP) approach to solve the stochastic control problems and prove that the optimal liquidation strategy actually converges to the deterministic strategy when the transaction cost involved by liquidating the outstanding position approaches to infinity.
We then move to analyze the randomlyterminated cases. Two different scenarios are analyzed to shed light on the relationship between liquidation strategies and potential position risk arising from the uncertainty of the time horizon. First, we consider the scenario where the liquidation process is terminated by an exogenous trigger event. We model the occurrence time of a trigger event to be random and its hazard rate process is given by . Once this event occurs, all liquidation processes will be forced to suspend. Compared with the case without trigger event, agents facing the scenario that an exogenous trigger event might occur during the trading horizon would like to accelerate the rate of liquidating to reduce their exposure to potential position risk and eventually in a smaller position when the trigger event occurs. Their strategy has a steeper gradient and is more “convex” when compared with those who are not threatened by this trigger event. Second, we consider the case when the liquidation process is subject to counterparty risk. Different from the exogenous trigger event setting, information set available to the counterparty risk modeler is more refined in terms of predictability. To model counterparty risk, we adopt the structural firm value approach, originated from Black and Scholes [3], and Merton [14], and let the firm’s asset value follow a geometric Brownian motion:
The incorporation of counterparty risk into the study of optimal liquidation does not come without cost. In order to examine its impact on optimal trading strategies, we have to introduce and employ viscosity solutions. By verifying the comparison principles for viscosity solutions, we characterize the value function as the unique viscosity solution of the associated HamiltonJacobiBellman (HJB) equation. This equation can be numerically solved. We further analyze the effectiveness of the numerical method and illustrate that the computational error is sufficiently small.
The remainder of this paper is structured as follows. The background and basic models of an agent’s liquidation problem are introduced in Section 2. Section 3 discusses typical liquidating problems under the benchmark model. In Sections 4–5, we discuss two different scenarios with randomlyterminated time horizons. Viscosity solution approach is adopted in these sections to study in great generality stochastic control problems. By combining these results with comparison principles, we characterize the value function as the unique viscosity solution of the associated dynamic programming equation, and this can then be used to obtain further results. Finally, concluding remarks are given in Section 6. For the sake of selfcontainedness, we provide the technical proofs in the Appendix.
2 Problem Setup
In this section, we first describe the market environment of the agent. We then present a market impact model to discuss the optimal liquidating problem.
2.1 The Market Environment and Market Impact Model
The agent starts at and has to liquidate a large position in a risky asset by time . This terminal time can be either deterministic or random, depending on the scenario that the agent is facing. For simplicity, we assume that the agent withholds the liquidation proceeds. In other words, he/she does not deposit the liquidation proceeds in his/her money market account. At any time , we adopt the following notations for the agent’s portfolio:
 (i)

;
 (ii)

;
 (iii)

;
 (iv)

.
The initial conditions are , , and .
Suppose the risky asset can be continuously liquidated during the trading horizon, namely, there is always sufficient liquidity for their execution^{1}^{1}1For simplicity, the transaction fees will not be considered in this paper.. Let denote the liquidation process. The shares held by the trader at any time can be written as follows:
We consider a probability space endowed with a filtration .
Definition 1
A stochastic process is called an admissible control process if all of the following conditions hold:
 (i)

(Adaptivity) For each , is adapted;
 (ii)

(Nonnegativity) , where is the set of nonnegative real values;
 (iii)

(Consistency)
 (iv)

(Squareintegrability)
 (v)

(integrability)
Furthermore, denote as the collection of admissible controls with respect to the initial time and as the collection of controls only satisfying condition (i), (iv) and (v).
We assume that the risky asset exhibits a price impact due to the feedback effects of the agent’s liquidation strategy. For any given admissible control , the market midprice of the stock is assumed to follow the dynamics:
(1) 
where is a standard Brownian motion with filtration , the constant is the absolute volatility of the asset price, and is the permanent component of the market impact. For simplicity, we further assume that is time homogeneous, namely, is independent of .
Generally speaking, the actual transaction price is
not always the same as the market midprice , since the market is not perfectly liquid, see, for example, Almgren and Chriss [2].
We assume and call the temporary price impact.
Intuitively, the function captures quantitatively how the limit order books available in the market are eaten up at different levels of trading speeds.
Assumption 0. The price dynamics follow a simple AlmgrenChriss linear market impact model (see, Almgren and Chriss [2]):
where and are positive constants.
An agent who holds the stock receives the capital gain or loss due to stock price movements. Thus, if the agent’s position is marked to market using the book value, ignoring market impact that would be incurred in converting these shares into cash, at any time , the agent’s portfolio value satisfies
(2) 
At any time before the end of trading,
2.2 Handsclean Condition
Let us recall that our task is to liquidate a largesize position by the time . Generally speaking, it is required that the handsclean condition should be satisfied:
(3) 
This technical condition, however, introduces some unexpected properties to the stochastic control problem. To tackle this problem, we temporary relax the handsclean condition and allow an inmmediate final liquidation (if necessary) so that the number of shares owned at equals zero. That is, given the state variables at the instant before the end of trading , if , then we will have an immediate final liquidation so that . The liquidation proceeds after this final trade is
where , for some constant , is the cost involved from liquidating the outstanding position . Thus, we have
The gain/loss from liquidating the outstanding position, , is given by
(4) 
2.3 Performance Criterion
Under the normal circumstance, investors are risk averse and demand a higher return for a riskier investment. The meanvariance criterion is popular for taking both return and risk into account. However, the meanvariance criterion may induce a potential problem of timeinconsistency, namely, planned and implemented policies are different. As mentioned in Rudloff et al. [19], a major reason for developing dynamic models instead of static ones is the fact that one can incorporate the flexibility of dynamic decisions to improve the objective function. Timeinconsistent criteria are generally not favorable to introduce in the study, since the associated policies are suboptimal.
To take both return and risk into account, instead of adopting the meanvariance criterion, we are most interested in the meanquadratic optimal agency execution strategies, as they are proved to be timeconsistent in [2, 6, 22]. In this section, we will introduce the quadratic variation and the corresponding objective function as follows.
2.3.1 Quadratic Variation
Formally, the quadratic variation of the portfolio value on is defined to be
(5) 
From the interpretation of Eq. (5), minimizing quadratic variation corresponds to minimizing volatility in the portfolio value process.
2.3.2 Objective Function
Let be a constant corresponding to the risk aversion. Then the agent’s objective is to find the optimal control for
(6) 
3 The Benchmark Model for Optimal Liquidation (Model 1)
Assumption 1
The liquidation horizon is a finitevalued, predetermined, and positive constant.
In this section, we present our benchmark model under Assumption 0 for the optimal liquidation problem. We first work on a subclass of deterministic controls^{2}^{2}2Controls that do not allow for intertemporal updating. satisfying the handsclean condition (3), and then move to the dynamic programming (DP) approach considering over the entire class of admissible controls. We prove that when the transaction cost involved by liquidating the outstanding position approaches to infinity, the optimal liquidation strategy obtained from DP approach converges to the deterministic one.
3.1 Deterministic Control
Let us first consider the case in which ranges only over the subclass of deterministic strategies in satisfying the handsclean condition
That is, , and the agent’s objective is to find the optimal strategy for
(7) 
The cost function of the deterministic control problem (7) is
where is the Lagrange multiplier (also called the adjoint state). The differential equation for the deterministic system is:
We assume the Hamiltonian has continuous firstorder derivatives in state, adjoint state, and control variables, namely, . Then the necessary conditions (also called Hamilton’s equation) for having an interior optimum of the Hamiltonian at , are given by
(8) 
It follows from the critical conditions in Eq. (8) and Assumption 0 that
(9) 
An explicit solution, which is unique according to Lasota and Opial [13], is given by
(10) 
There is a very interesting phenomenon in the deterministic control problem: the solution (10) has nothing to do with the permanent price impact . If a position of size units with initial market price is fully liquidated by time , i.e. , the expected value of the resulting cash becomes
Clearly, the permanent impact cost is independent of the time taken or strategy used to execute the liquidation.
3.2 Dynamic Programming Approach
Obviously, if we are allowed to update dynamically, namely,
replacing by the entire class of admissible strategies ,
then one will be able to further improve his/her performance.
In this section, we consider a stochastic approach.
We employ the DP method to solve the stochastic control problem (6).
This approach yields a HamiltonJacobiBellman (HJB) equation.
When this HJB equation can be solved by an explicit smooth solution,
the verification theorem then validates the optimality of the candidate solution to the HJB equation.
For more details about the verification theorem, we refer interested readers to
Pham [18] (Chapter 3), Øksendal [15] (Chapter 11),
and Øksendal and Sulem [16] (Chapter 3).
Let be the optimal value function beginning at a time with initial value , namely^{3}^{3}3It is worth noting that the value function does not depend explicitly on the stock price .,
(11) 
Temporarily assuming that ^{4}^{4}4 is the space of functions which is continuously differentiable in , and twice continuously differentiable in . From the DP principle, must satisfy the following HJB equation:
(12) 
We remark that the optimization problem included in Eq. (12) is a constrained optimization problem with constraints: (a1) ; and (a2) . Generally speaking, there is no straightforward method to solve this kind of problems. One simple way to handle this problem is to consider the corresponding unconstrained optimization problem:
(13) 
and then verify that the obtained result indeed satisfies all the constraints. From the HJB equation, Eq. (13), the optimal trading strategy without constraints is given by
Thus the value function solves the following Ordinary Differential Equation (ODE):
(14) 
Theorem 1
There is at most one solution to Eq. (14).
Proof: Let and be two solutions to Eq. (14). Define . Then the new function satisfies the following Partial Differential Equation (PDE):
Since the evolution equation for is linear and firstorder, one can solve the above problem explicitly by the method of characteristics, and find that is the unique solution to this problem. As a result, .
To solve Eq. (14), we consider an ansatz that is quadratic in the variable :
According to Theorem 1, if the above ansatz is a solution of Eq. (14), then it must be the unique solution. Under this setting, the optimal liquidating strategy takes the following form:
A direct substitution yields that the coefficients , and must satisfy the following ODEs:
(15) 
with terminal conditions: , and . Since System (15) is partially decoupled, we can find the exact solution via direct integrations. As a result, they are given by
(16) 
where the constants and are given by
It is worth noting that
(17) 
and that . Therefore,
As to the results obtained in this section, we have the following proposition.
Proposition 1
It is assumed that model parameters satisfy the condition:
(18) 
That is, market liquidity risk dominates the potential arbitrage opportunity introduced by permanent impact and potential position risk involved by price fluctuations. Then, is a strictly decreasing function in and for any . Furthermore, we have that
 (b1)

, for any time ; and that
 (b2)

.
The obtained optimal trading strategy (17) is also the optimal trading strategy for the constrained problem.
Proof: Notice that the graph of the function depends on the coefficient , with . Under Assumption (18), , and hence . Therefore,
i.e., is a strictly decreasing function in , and always holds for any . Thus, we conclude that
for any time , and that
Let denote the value function of the optimization problem (11) with time horizon , then for any , we have
(19) 
provided that the condition (18) holds. This is consistent with the fact that an investor’s ability to bear risk relates to his/her time horizon for investment^{5}^{5}5The ability to bear risk is measured mainly in terms of objective factors, such as time horizon, expected income, and the level of wealth relative to liability..
3.3 Relation between Deterministic and Stochastic Control
Theorem 2
When the transaction fees involved by liquidating the outstanding position approaches to infinity, the limit of the optimal stochastic control process satisfies the handsclean condition and it converges (pointwise) to the optimal deterministic control process . Meanwhile, the optimal trajectory converges (pointwise) to the one determined in the deterministic system . That is, as , we have

;

pointwise;

pointwise.
Proof: We complete the proof by the following two steps:
 Step 1 (Handsclean condition)

We first prove that, as , . We note that
A simple calculation yields
(20) As , , and hence
 Step 2 (Convergence)

We then prove that as ,

pointwise; and

pointwise.
First, we have
For any time ,
Thus, we have

In Figure 1, we illustrate how the transaction fees involved by liquidating the outstanding position , , affects the agent’s liquidating speed. We chose the following values of the model parameters: , , , , and .
Figure 1 illustrates that the speed of liquidation which is free of handsclean condition is always slower than that under the constraint of handsclean condition. As the transaction fees involved by liquidating the outstanding position increases (namely, as increases), the agent’s liquidating speed increases, indicating that the optimal stochastic control moves closer to the optimal deterministic control. The embedded subfigures in Figure 1 show, respectively, the differences between the deterministic and stochastic liquidating strategies and the corresponding trajectories with . Both of them are of magnitude .
4 Optimal Liquidation Strategy Subject to an Exogenous Trigger Event (Model 2)
In this section, we extend our results to models with an exogenous event, which does not depend on the information structure .
Assumption 2
The liquidation process will be suspended, if an exogenous trigger event occurs.
We model the occurrence time of a trigger event, denoted by , to be random, and the hazard rate is given by . The survival probability at time is
(21) 
The liquidation horizon is then defined by
(22) 
where the constant is a predetermined time horizon. A direct computation yields the following proposition:
Proposition 2
For , the density function of is
The probability that takes the value of is .
Denote by the event . At any time , i.e., the trigger event has not occurred prior to time , the agent’s objective is to find the optimal control for
(23) 
where
and is the information structure available to the agent up to and including time . If the trigger event occurs at time , all market transactions will be suspended at that time. The agent will end up with an outstanding position .
It is worth noting that
(24) 
and that
Here, the indicator function takes the value 1 when its argument is true and the value , otherwise. The last equality in Eq. (24) follows from the assumption that the trigger event is exogenous and does not depend on information structure .
Therefore, we have
(25) 
That is, the optimal liquidating problem with a random horizon defined in Eq. (22) is equivalent to an optimal liquidating problem with a finite horizon , a consumption process , a discount process , and a terminal condition .
4.1 Deterministic Control
Let us first consider the case in which ranges only over the subclass of deterministic strategies in satisfying the handsclean condition (3)^{6}^{6}6The handsclean condition only makes sense for the equivalent problem (25). While considering the original optimization problem (23), where the terminal time is a stopping time, the handsclean condition is no longer valid., namely, . Thus, the agent’s objective, before the trigger event occurs, is to find the optimal control for
The cost function of the deterministic control problem is
where is the Lagrange multiplier, and is the survival probability defined in Eq. (21). The differential equation for the deterministic system dynamics is
We assume that the Hamiltonian has continuous firstorder derivatives in the state, adjoint state, and the control variable, namely, . Then the necessary conditions for having an interior point optimum of the Hamiltonian at are given by
(26) 
It follows from Eq. (26) that
(27) 
Regarding this linear secondorder boundary value problem (BVP), its existence and uniqueness are standard. Interested readers can refer to, for example Hwang [11], for more details.
Consider the case when , which corresponds to the case of constant hazard rate, an explicit solution is given by
where
It is worth noting that (i) when , the model degenerates to Model 1; (ii) as , and , for all ; and . That is, as the final liquidation fee, per share, approaches infinity, the trader would immediately complete the transaction at the beginning of the trading horizon.
4.2 Dynamic Programming Approach
Let us consider the case of allowing dynamic updating, i.e., replacing by the entire class of admissible strategies . Let denote the optimal value function of Eq. (25) at any time prior to the occurrence of the trigger event. Under appropriate regularity assumptions, satisfies the following HJB equation:
(28) 
subject to the terminal condition: . Here, is the given hazard rate. Similarly, we consider relaxing the constraints associated with the HJB equation and solve the unconstrained optimization problem. We then prove that the obtained optimal control does satisfy all the constraints. The associated optimal trading strategy is
and hence the value function satisfies
(29) 
Regarding Eq. (29), we have the following theorem for the uniqueness of classical solutions.
Theorem 3
There is at most one solution to Eq. (29).
Proof: Suppose and are two solutions to Eq. (29). Define . Then the new function satisfies the following problem:
Since the evolution equation for is linear and firstorder, one can solve the above problem explicitly by the method of characteristics, and find that is the unique solution to this problem. As a result, .
Similar to Section 3.2, we consider an ansatz that is quadratic in the variable :
Substituting the ansatz into Eq. (29), we know that the coefficients , and must satisfy the following partially decoupled system:
(30) 
with terminal conditions: , and .
It is straightforward to verify that
However, the equation satisfied by is a Riccati equation, which can be reduced to a secondorder linear ODE:
(31) 
where is defined implicitly via . For this secondorder linear ODE, its existence and uniqueness are standard. Even though we know the existence and uniqueness of the solution, it is still difficult to solve it in a closedform for a general hazard rate . The above secondorder linear ODE can be easily solved in two cases: (i) its coefficients are constant; or (ii) its coefficients adopt particular forms.
If closedform solutions cannot be obtained, finite difference method can be applied to solving the BVP numerically. For more details, see, for example, Hwang [11].
Theorem 4
(Constant hazard rate). When the hazard rate is a constant, i.e., , the unknown function can be explicitly solved. It is given by