Complexity in Economics and BusinessView this Special Issue
Research Article | Open Access
Lei Ge, Qiang Zhang, "Numerical Solutions to Optimal Portfolio Selection and Consumption Strategies under Stochastic Volatility", Complexity, vol. 2020, Article ID 9548060, 16 pages, 2020. https://doi.org/10.1155/2020/9548060
Numerical Solutions to Optimal Portfolio Selection and Consumption Strategies under Stochastic Volatility
Based on the method of dynamic programming, this paper uses analysis methods governed by the nonlinear and inhomogeneous partial differential equation to study modern portfolio management problems with stochastic volatility, incomplete markets, limited investment scope, and constant relative risk aversion (CRRA). In this paper, a three-level Crank–Nicolson finite difference scheme is used to determine numerical solutions under this general setting. One of the main contributions of this paper is to apply this three-level technology to solve the portfolio selection problem. In addition, we have used a technique to deal with the nonlinear term, which is another novelty in performing the Crank–Nicolson algorithm. The Crank–Nicolson algorithm has also been extended to third-order accuracy by performing Richardson’s extrapolation. The accuracy of the proposed algorithm is much higher than the traditional finite difference method. Lastly, experiments are conducted to show the performance of the proposed algorithm.
How to optimally allocate assets and optimally consume are extremely important and difficult topics in portfolio management [1–3]. These topics are important not only for theoretical consideration but also for applications in the financial industry. Early studies usually assumed the volatility of the risky asset to be a constant. However, in recent years, researchers found that volatility should be modeled as stochastic rather than deterministic [4–7]. This adds further complication to the problem. The optimal asset allocation and optimal consumption strategies are governed by the Hamilton–Jacobi–Bellman (HJB) equation. Due to the nonlinearity and inhomogeneity of this partial differential equation, no exact solution has been found. Furthermore, even numerical solutions are not available. In this paper, we present an accurate and efficient numerical method for solving this equation and generate the first set of accurate numerical solutions for this problem.
Due to the importance of portfolio selection under stochastic volatilities, several important theoretical works have been carried out, and exact solutions have been obtained under certain special settings, such as no consumption [8–10], complete markets which means that the stock movement and the volatility movement are either perfectly correlated or perfectly anticorrelated [9–12], or when investors have unit elasticity of intertemporal substitution of consumption .
In this paper, we consider this optimal stochastic control problem under a general setting: stochastic volatility, incomplete markets, finite investment horizons, and CRRA utility. Our numerical method combines a three-level Crank–Nicolson scheme and Richardson’s extrapolation technique. The Crank–Nicolson scheme has second-order accuracy in terms of discretization error, and Richardson’s extrapolation technique further improves the accuracy. We verify that our numerical method is accurate and efficient.
This paper is organized as follows. In Section 2, we describe the model for financial market, the stochastic control optimization procedure, and the governing HJB equation for the optimal asset allocation and consumption strategies. In Section 3, we present our numerical method for solving the HJB equation. In Section 4, we verify the accuracy and the efficiency of our numerical method and present accurate numerical solutions for the optimal asset allocation strategy and the optimal consumption strategy. In the last section, we present our conclusions.
2. Financial Market and Stochastic Control
We consider a market consisting of one riskless asset , whose price is governed bywith a constant risk-free interest rate and a risky asset modeled as
In (2), and are the return and the stochastic volatility of the stock price , respectively. is the stochastic variance of . Empirical studies show presence of mean reversion in the stock movements . Heston model  is selected for , namely,
Here, and are the increments of the Wiener processes under a probability . The correlation between and is , namely, . We assume is a constant. In (3), is the long-run average variance (i.e., as tends to infinity, the expected value of tends to ), is the rate at which reverts to , and is the volatility of the stock variance . The parameters are positive constants and need to satisfy the Feller condition, , to ensure that is strictly positive. The risk premia is defined as
Consider an investor who has an initial wealth and needs to determine strategies for asset allocation and consumption over an investment horizon . Let be the investor’s wealth at time . The strategies consist of an asset allocation rate and a consumption rate , which mean he/she allocates to the risky asset and to the riskless asset at time and consumes over the time interval . Thus, under the strategies and , the wealth process is governed by
The goal is to maximize the expected utilities over the investment horizon, namely,
In (6), and are control variables for this optimization problem. is the expectation operator under the probability . is the subjective discount rate, namely, the time preference of the investor. The larger is, the more weight the investor puts on the present than on the future. The parameter determines the relative importance between intertemporal consumption and the terminal wealth. and are the investor’s utility functions which measure the investor’s degree of satisfaction with the outcomes from intertemporal consumption and terminal wealth, respectively.
CRRA utility functions have been widely adopted for modeling investors’ behavior. Therefore, we adopt the CRRA utility function for and :where , , and are positive constants. Since and stand for the intertemporal consumption utility and the terminal wealth utility of the same investor, we use the same in and . However, and can be different since and have different dimensions.
Based on the HJB dynamic programming procedure, is governed bywith the optimal strategies and determined by
Equation (15) is a nonlinear and inhomogeneous partial differential equation. Since no closed-form solution is available for this equation, numerical computation plays a critical role for studying this important practical problem in modern finance. However, there are even no numerical solutions available in the literature.
3. Numerical Method
3.1. Crank–Nicolson Scheme and Richardson’s Extrapolation
We use a three-level Crank–Nicolson scheme (see [18–21]) of second-order accuracy to solve the nonlinear and inhomogeneous partial differential equation given by (19) and use Richardson’s extrapolation technique for further improving accuracy. Numerically, one can only solve (19) over a finite domain . Since the boundary conditions at and at are not known, we use one-sided difference method at these two numerical boundaries. Step sizes and are used to discretize and , respectively. Thus, and . We adopt the standard notation .
The three-level Crank–Nicolson scheme involves the levels , , and . It is straightforward to discretize all linear terms in (19) with second-order errors, namely,
The nonlinear term has two factors and . We discretize the factor at level and approximate the factor as an average between at level and that at level , namely,
Since (23) is not applicable to the boundaries at and , we used one-sided difference technique to discretize (19) and obtained the boundary equations in the following. It is straightforward to show that, at , we haveand at , we have
By substituting these expressions into (19), we havewherewith
After eliminating , , , and , (33) can be transformed into the following tridiagonal matrix form for :where
The method given by (21) and (22) is a two-step method, namely, depends on and . At the zeroth step, is given by initial condition (16), namely, , for . We now determine , the solution at the first step. Performing Taylor expansion on at giveswhere is given by (16) and , , and can be determined analytically from (19):
The details of derivations for , , and are given in Appendix A. From (36), is given bywith an error of .
In summary, our numerical solutions for and are determined by (16) and (38), respectively, and the numerical solutions for with are determined by (34). The numerical solution of obtained by the three-level Crank–Nicolson scheme has an accuracy of .
3.2. Performing Richardson’s Extrapolation
To further improve the accuracy of the numerical method, we apply Richardson’s extrapolation technique to . We will choose proportional to . Let represent obtained by (34) with a step size . Then,where is the exact value. We perform two computations with the step sizes and , respectively. Then, we have the following two equations:
For the purpose of giving a quick understanding of our method, Figure 1 presents a flowchart of the algorithm for solving (19) We also summarize the procedure in words in the following for obtaining the numerical solutions of , , , , , and . Here, we choose .(i)Step 1: initialize by initial condition (16), namely, , for .(ii)Step 2: initialize by (38) for .(iii)Step 3: for , knowing and , for , can be determined from (34), which is in a tridiagonal form and can be easily and efficiently solved. has an accuracy of .(iv)Step 4: to obtain and , we substitute from Steps 1–3 into (39) and (40). This provides the numerical solutions for optimal strategies and without extrapolation, which have accuracy of .(v)Step 5: to obtain , , and , we repeat Steps 1–3 with the step size to obtain . Then, from (45), (46), and (47), we obtain , , and , all of which have accuracy of .
In the next section, we will verify the accuracy of the numerical solutions without extrapolation and those with extrapolation.
4. Validation Study of the Numerical Method
Equation (19) is an inhomogeneous equation. However, since the inhomogeneous term affects neither the stability nor the accuracy of the three-level Crank–Nicolson method, it is sufficient to conduct validation studies for the corresponding homogeneous equation, namely, for the case of . Let be the solution of the homogeneous equation of (19), namely, the case of . Then, satisfieswith the initial condition . Following the procedure outlined in , the exact solution for can be obtained. After expressing asfrom (49), and are governed byand the solutions arewhere and . From (17) and (18), we obtain the exact solutions of optimal strategies and for the case of :where is given by (52).
The exact solutions , , and for the case of given by (50), (54), and (55) offer a benchmark for testing the accuracy of our numerical solutions. We show that our numerical solutions are accurate and efficient for . Since neither the inhomogeneous term nor the constant initial condition affects the stability or the accuracy of a numerical method, the accuracy and the stability of the method remain valid for . The numerical results for are presented at the end of this section.
4.1. Numerical Validation
We note that since determines the wealth scale and determines the temporal scale, without loss of generality, we choose in this study.
For the range of the state variables and , we consider . Based on the historical records of the Chicago Board Options Exchange Volatility Index, a popular measure of the implied volatility of S&P 500 index options, we examine the numerical solutions for the instantaneous volatility in the interval (to eliminate possible influence from the numerical boundary, the authors choose in their numerical computations). Since wealth does not appear in (19), (17), and (18), its value is irrelevant in our study.
In Table 2, we show the comparison between , , and . is the exact solution of (49). When , given by (34) is the numerical solution of (49) without performing Richardson’s extrapolation and given by (45) is the numerical solution of (49) after performing Richardson’s extrapolation. The relative errors in and , namely, and , are shown in the last two columns of Table 2.
In Table 3, we show the comparison between the exact solution given by (54), the numerical solution determined from (39) without performing Richardson’s extrapolation, and determined from (46) after performing Richardson’s extrapolation. The relative errors in and , namely, and , are shown in the last two columns of Table 3.