NatureInspired Algorithms and Applications: Selected Papers from CIS2013
View this Special IssueResearch Article  Open Access
A ConvexRiskMeasure Based Model and Genetic Algorithm for Portfolio Selection
Abstract
A convex risk measure called weighted expected shortfall (briefly denoted as WES (Chen and Yang, 2011)) is adopted as the risk measure. This measure can reflect the reasonable risk in the stock markets. Then a portfolio optimization model based on this risk measure is set up. Furthermore, a genetic algorithm is proposed for this portfolio optimization model. At last, simulations are made on randomly chosen ten stocks for 60 days (during January 2, 2014 to April 2, 2014) from Wind database (CFD) in Shenzhen Stock Exchange, and the results indicate that the proposed model is reasonable and the proposed algorithm is effective.
1. Introduction
In 1952, Markowitz proposed the first quantitative risk measure (i.e., variance) for portfolio selection [1]. Because variance as the risk measure has essential drawbacks such as only being applicable to the cases in which the return obeys the normal distribution or elliptical distribution, it is rarely used as the risk measure currently. After that, value at risk (VaR) was proposed by Baumol [2] as a risk measure and it has been a widely used risk measure to manage or control risk for portfolio selection. For example, the Basle Committee on Banking Supervision allows banks to use VaR when determining their assetadequacy requirements arising from their exposure to market risk [3]. The reason for VaR to be a popular risk measure is that it is easily understood and can exactly answer the following question: under the normal market environment and given confidence level, what is the maximal potential loss of an investor in a certain period? Just as VaR definition mentioned, given some confidence level , the VaR of the portfolio at the confidence level is given by the smallest number such that the probability of the loss exceeding is not larger than ; that is,
Since VaR appears, many researchers have paid attention to use VaR in their portfolio models. For example, authors in [4] integrate GARCH model and VaR model and got the better results. However, VaR model has some serious drawbacks: (1) it usually gives preserved estimation on the risk of the markets [5]; (2) it does not satisfy subadditivity (i.e., the total risk of the sum of two investments is not larger than that of the sum of their risks [6]); and (3) it is not a coherent risk measure [7]. Thus, some researchers thought that VaR was not a good risk measure and even some researchers thought that it was seductive but dangerous (e.g., [8]). To get more reasonable risk measures, researchers discussed the conditions which a reasonable risk measure should satisfy (e.g., [6, 7, 9, 10]). The widely accepted and representative condition for a reasonable risk measure is coherent condition, which defines a risk measure to be a coherent measure if it satisfies the following properties:(I)monotonity: if , , for any two portfolios and ;(II)subadditivity: , for any two portfolios and ;(III)positive homogeneity: , for any portfolio , and ;(IV)translational invariance: If is a deterministic portfolio with guaranteed return , then for any portfolio .
Monotonity illustrates that if portfolio y always has better values than portfolio x under almost all scenarios, then the risk of y should be less than the risk of x. Subadditivity implies that the risk of two portfolios together cannot get any worse than adding the two risks separately; this is the diversification principle. Loosely speaking, positive homogeneity represents that if you double your portfolio then you will double your risk. Translational invariance indicates that after adding an amount of money to the original portfolio , the risk will be decreased by amount .
Based on concept of coherent risk measure, many new risk measures were proposed. For examples, Artzner et al. proposed a coherent risk measure WEC [7]. WEC is a good risk measure in theory, but it seriously depends on the distribution of random variable . To get easily computed coherent risk measure, Acerbi et al. [11] proposed the expected shortfall as the risk measure (ES) and Rockafellar and Uryasev [12] proposed a coherent risk measure: conditional value at risk (CVaR). Although the definitions of ES and CVaR are different, their key ideas are same. ES has an important property [13]: any coherent risk measure can be described by spectral risk measures and any coherent spectral risk measure can be represented by the linear combination of ES with different confidence levels. That means that ES is the basic components of coherent risk measures. CVaR also has attractive properties such as its formula can be easily obtained and thus can be easily used for portfolio selection. Recently, with the further deep study of coherent risk measures, some new coherent risk measures with good performance were proposed; for example, Rosazza Gianin [14] proposed a coherent risk measure based on gexpectation operator which can be applicable not only to the estimation of financial asserts, but also to the derivatives. Also, by choosing different gexpectation operator, this risk measure can satisfy the different prefers of investors to the risk. To handle the problems with nonnormal distribution and leptokurtosis (i.e., fat tails), Chen and Wang [15] proposed a new class of coherent risk measures based on norms. However, aforementioned coherent risk measures are only one side coherent risk measures [16]; that is, the distribution information on only one side of supply side and demand side is considered. From the point of view of competition, when an investor is going to buy a stock, he or she should also pay attention to the behavior of the seller. That is to say, to measure the risk more exactly, the behavior of investors to supply side and demand side should be simultaneously considered. In other words, twoside risk measures are needed. For example, Chen and Wang [16] proposed a twoside coherent risk measure which is easily applied.
Although there have been a lot of progresses in the research of coherent risk measures, there exist two serious problems in coherent risk measures. (1) Positive homogeneity is not reasonable in many cases and the condition is too strong [17, 18]; (2) translational invariance is not reasonable in many cases and more and more researchers do not adopt this condition in their risk measures (e.g., [19â€“21]). To relax the condition in positive homogeneity and get more reasonable risk measures, some authors suggested replacing subadditivity and positive homogeneity by the following convexity:and proposed convex risk measures (e.g., [22, 23]). From above analysis, one can see that a good and reasonable risk measure should satisfy monotonity and convexity. For this purpose, some risk measures satisfying monotonity and convexity were proposed (e.g., [24, 25]) and experiments indicate that the measures are reasonable.
In this paper, we adopt a convex risk measure called weighted expected shortfall (WES) as the risk measure [24] and propose a portfolio optimization model based on this risk measure. Then we design a genetic algorithm for this portfolio optimization model. At last, simulations are made on real data in the financial markets and the results indicate that the proposed model is reasonable and the proposed algorithm is effective.
The remaining parts are organized as follows. In Section 2, the portfolio optimization model is set up. The proposed genetic algorithm is proposed in Section 3. The computer simulations are made on real data in financial markets in Section 4, and the conclusions are made in Section 5.
2. Portfolio Optimization Model
In static state, risk can be seen as a random variable on a probability space , where represents uncertain rate of return of portfolio. Then, for a given confidence level , a new risk measure is defined in [24, 26] as follows.
Definition 1 (see [24, 26]). Suppose that ; for given confidence level , a new risk measure called weighted expected shortfall () is defined as follows:where is monotone nonincreasing function and positive convex when and nonnegative concave when . For example, when , , and when , where represents the risk aversion coefficient,and represents the probability of random event .
Also, in [24, 26], a computable formula (or an estimation) of was given. Suppose that there are risky assets and one riskfree asset and the portfolio for these assets is denoted by , where is the portfolio for risky assets, respectively, and is the portfolio for the riskfree asset. Let represent the rate of return of the th risky asset for and the rate of return of the riskfree asset for , respectively, in the period for . Let represent the rate of return of the th risky asset for and the rate of return of the riskfree asset for , respectively. It can be estimated by
Let represent the dividend yield of the th risky asset for in the period for . Let represent the dividend yield of the th risky asset for . It can be estimated by
Let be the initial portfolio, let be the per unit transaction cost of the th risky asset, let be the asset income marginal tax rate, let be ordinary marginal income tax rate, and let be the given target rate of return. Let
If we choose the proper confidence level and parameter such that is an integer, then, according to [26], the risk based on the new measure can be calculated bywhere is the smallest element of and is the th smallest element of .
Also, according to [26], an optimization portfolio model can be formulated as
Also, the condition is required to be satisfied in optimal solution in [26]. This is a nonlinear constraint, but this condition was not put in the model in order to make the model easily solved (although this is not reasonable). Thus the optimal solution obtained for the model (14)â€“(20) may not satisfy the condition . This will result in the obtained solution being not a true optimal solution.
Note that â€‰â€‰, , and â€‰â€‰ are variables in the above model; thus, there are total ( variables and the problem dimension is . But when we carefully check this model, we can find that these variables are not independent. In fact, from formulas (7) and (8), it can be seen that and are not independent to and can be completely presented by . Thus, variables and can be deleted for . Also note that when we delete variables and , constraints (17) and (18) as well as will be automatically satisfied. Furthermore, in order to estimate in formula (10) more precisely, we can use a large number of historic data (i.e., a large number ) in formulas (5), (6), and (10), but too large will result in the increasing of computation of risk function . In order to reduce the computation load, we have to choose a proper . In this way, we can simplify the model (14)â€“(20).
In summary, we can modify the above model by deleting variables and and constraints (17) and (18) as well as using proper value of to set up a new simplified optimization portfolio model as follows:
This optimization model has only dimensions which are much lower than those of the original model (14)â€“(20).
3. A New Genetic Algorithm for the Portfolio Model
For notation convenience, let denote the vector of all variables. Then is an dimensional vector. For notation convenience, let in this section.
The optimization portfolio model in previous section is a nonlinear optimization problem. It is very hard to get its global optimal solution using the traditional optimization methods. Genetic algorithms (briefly, GAs) are a new kind of intelligent optimization methods which are designed for these difficult optimization problems [27â€“29]. They exploit a set of potential solutions, named a population, and detect the optimal solution through cooperation and competition among individuals of the population. However, for GAs in global optimization, the major challenges are that an algorithm may be trapped in the local optima of the objective function and the convergent speed may be slow. These issues are particularly challenging when the dimension of the problem is high and there are numerous local optima. In order to improve the GAs, researchers have incorporated other techniques to enhance their performance. One important technique is to design more efficient crossover operators to enhance the local search ability of GAs [28, 29]. In this section, we first design an efficient crossover operator which can explore the search space efficiently. Then we design a mutation operator which can adaptively exploit the search space. Based on these, a new genetic algorithm is proposed.
3.1. Crossover Operator
In this section, the uniform design method [28â€“30] is used to design a new crossover operator. The main objective of uniform design is to sample a small set of points from a given set of points, such that the sampled points are uniformly scattered on the interested region. The crossover operator based on the uniform design is similar to a local search scheme; thus, it can effectively explore the search space. The detail is as follows.and let {} denote the decimal part of the real number . A widely used method to generate approximately uniformly distributed points in is as follows [30], where is a positive integer. Let be a positive prime, satisfying .and then is a set of approximately uniformly distributed points in .
Suppose that and are any two parents chosen for crossover. Let and for . Now a new crossover operator is designed to generate approximately uniformly distributed points in the following set:
Algorithm 2 (crossover operator). (1) Generate approximately uniformly distributed points in by aforementioned formulas, and denote the set of these points byIn simulations, the parameter value is taken as and .
(2) Generate uniformly distributed points in set byThen, the points in are offspring of and .
3.2. Mutation Operator
Suppose that is any parent chosen for mutation. Let denote the best individual in current population and let denote the best individual obtained up to now. Then the offspring of by mutation is defined by where is a random number in and with being a random number in .
3.3. The Proposed Evolutionary Algorithm
Algorithm 3. (1) (Initialization) Given population size , crossover probability and mutation probability . Randomly generate initial population . Let .
(2) (Crossover) Randomly choose pairs of parents from . For each pair, use Algorithm 2 to generate offspring. The set of all these offspring is denoted as .
(3) (Mutation) Randomly choose parents from . For each chosen individual , use mutation operator to get an offspring . The set of all these offspring is denoted as .
(4) (Selection) Select best individuals among to put into , and then randomly select individuals among to put into . Let .
(5) (Termination) If stop condition is satisfied, stop; otherwise, go to step (2).
4. Experimental Results for Real Stock Market
To evaluate the performance of the proposed optimization portfolio model based on a new risk measure called WES and the genetic algorithm, we conducted the experiments on randomly chosen ten stocks for 60 days (during January 2, 2014 to April 2, 2014) from Wind database (CFD) in Shenzhen Stock Exchange and the names and numbers of these ten stocks are shown in Table 1, where VKA represents Ashares of China Vanke Company Limited, PAB represents stock of Ping An Bank Company Limited, BLL represents stock of Baolilai Investment Company Limited, SZPRDA represents Ashares of Shenzhen Properties and Resources Development (Group) Limited, CSGA represents Ashares of CSG Holding Company Limited, SH represents stock of Shahe Industry, SZHA represents Ashares of Shenzhen Zhongheng Huafa Company Limited, SVOTA represents Ashares of Shenzhen Victor Onward Textile Industrial Company Limited, Konka A represents Ashares of Konka Group Company Limited, and SSIA represents Ashares of Shenzhen Shenbao Industrial Company Limited. We collect the daily returns and daily return rates from Shenzhen Stock Exchange in Wind database (CFD) in this period in the experiments.

In the experiments, the confidence level is chosen as 95% and the parameters in the proposed algorithm are taken as follows: , , , , and . and . We consider the case in which there is transaction cost of the risky asset for simplicity in the simulations; that is, let . We use the same values of parameters as those in [26]. That is, we suppose that investor only has currency initially; that is, , . Thus, and for . Let , , and . Also, we take weighted function as follows: when , , and when , where represents the risk aversion coefficient and .
We use the proposed algorithm to the optimization portfolio model (21). The results are given in Table 2. The rates of return of the 10 selected stocks for 60 days are given in Tables 3 and 4.

