CVaR Robust Mean-CVaR Portfolio Optimization

Abstract

One of the most important problems faced by every investor is asset allocation. An investor during making investment decisions has to search for equilibrium between risk and returns. Risk and return are uncertain parameters in the suggested portfolio optimization models and should be estimated to solve the problem. However, the estimation might lead to large error in the final decision. One of the widely used and effective approaches for optimization with data uncertainty is robust optimization. In this paper, we present a new robust portfolio optimization technique for mean-CVaR portfolio selection problem under the estimation risk in mean return. We additionally use CVaR as risk measure, to measure the estimation risk in mean return. To solve the model efficiently, we use the smoothing technique of Alexander et al. (2006). We compare the performance of the CVaR robust mean-CVaR model with robust mean-CVaR models using interval and ellipsoidal uncertainty sets. It is observed that the CVaR robust mean-CVaR portfolios are more diversified. Moreover, we study the impact of the value of confidence level on the conservatism level of a portfolio and also on the value of the maximum expected return of the portfolio.

1. Introduction

Portfolio optimization is one of the best known approaches in financial portfolio selection. The earliest technique to solve the portfolio selection problem is developed by Harry Markowitz in the 1952. In his so-called mean-variance (MV) portfolio optimization model, the portfolio return is measured by the expected return of the portfolio, and the associated risk is measured by the variance of portfolio returns [1].

Variance as the risk measure has its weaknesses. Controlling the variance does not only lead to low deviation from the expected return on the downside, but also on the upside [2]. Hence, alternative risk measures have been suggested to replace the variance such as Value at Risk (𝑉𝑎𝑅) that manage and control risk in terms of percentiles of loss distribution. Instead of regarding both upside and downside of the expected return, 𝑉𝑎𝑅 considers only the downside of the expected return as risk and represents the predicted maximum loss with a specified confidence level (e.g., 95%) over a certain period of time (e.g., one day) [3–5].

𝑉𝑎𝑅 is a popular risk measure. However, 𝑉𝑎𝑅 may have drawbacks and undesirable properties that limit its use [6–8], such as lack of subadditivity; that is, 𝑉𝑎𝑅 of two different investment portfolios may be greater than the sum of the individual 𝑉𝑎𝑅s. Also, 𝑉𝑎𝑅 is nonconvex and nonsmooth and has multiple local minimum, while we seek the global minimum [4, 8, 9]. So alternative risk measures were introduced such as Conditional Value at Risk (𝐶𝑉𝑎𝑅)—the conditional expected value of loss, under the condition that it exceeds the value at risk [3]. 𝑉𝑎𝑅 implies that “what is the maximum loss that we realize?” but 𝐶𝑉𝑎𝑅 asks: “How do we expect to incur losses when situation is undesirable?”. Numerical experiments show that minimum 𝐶𝑉𝑎𝑅 often leads to optimal solutions close to the minimum 𝑉𝑎𝑅, because 𝑉𝑎𝑅 never exceeds 𝐶𝑉𝑎𝑅 [5]. 𝐶𝑉𝑎𝑅 has better properties than 𝑉𝑎𝑅. It is a convex optimization problem, and thus it is easy to optimize [4]. It is demonstrated that linear programming techniques can be used for optimization of 𝐶𝑉𝑎𝑅 risk measure [5, 9].

Risk and return are uncertain parameters in portfolio optimization models, and estimating them might lead to large error in the final decision. To deal with such situation, one of the widely used and effective approaches is robust optimization technique. In this paper, we have applied this technique to give the robust counterpart of the mean-𝐶𝑉𝑎𝑅 portfolio selection problem under the estimation risk in mean return. Moreover, we have used 𝐶𝑉𝑎𝑅 as the risk measure to measure the estimation risk in mean return. The rest of the paper is organized as follows. In Section 2, we state the mean-𝐶𝑉𝑎𝑅 portfolio selection problem. Then because of the inevitable estimation error of the mean return of the assets, we present robust optimization by 𝐶𝑉𝑎𝑅 in Section 3. To solve the model efficiently, we use the smoothing technique of Alexander et al. [10]. Finally, in Section 4, we compare the performance of the 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 model with robust mean-𝐶𝑉𝑎𝑅 models using interval and ellipsoidal uncertainty sets on an example. We have observed that the 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 portfolios are more diversified and they are sensitive to initial data used to generate each set of 𝜇 samples. Moreover, we demonstrate that the value of confidence level affects the conservatism level, diversification, and also the value of the maximum expected return of the resulting portfolios.

2. Mean-Conditional Value at Risk

Consider assets 𝑆₁,…,𝑆_𝑛, 𝑛≥2, with random returns. Suppose 𝜇_𝑖 denotes the expected return of asset 𝑆_𝑖, and also consider 𝑥_𝑖 as the proportion of holding in the 𝑖th asset. We can represent the expected return of the resulting portfolio 𝑥 as follows: 𝐸[𝑥]=𝜇₁𝑥₁+⋯+𝜇_𝑛𝑥_𝑛=𝜇^𝑇𝑥. (1)

Also, we will assume that the set of feasible portfolios is a nonempty polyhedral set and represent that as Ω={𝑥∣𝐴𝑥=𝑏,𝐶𝑥≥𝑑} where 𝐴 is a 𝑚×𝑛 matrix, 𝑏 is an 𝑚-dimensional vector, 𝐶 is a 𝑝×𝑛 matrix, and 𝑑 is a 𝑝-dimensional vector [4]. In particular, one of the constraints in the set Ω is ∑^𝑛_𝑖=1𝑥_𝑖=1.

Let 𝑓(𝑥,𝑦) denote the loss function when we choose the portfolio 𝑥 from a set of feasible portfolios, and 𝑦 is the realization of the random events (the vector of the asset returns of 𝑛 assets). We consider the portfolio return loss, 𝑓(𝑥,𝑦), the negative of the portfolio return that is a convex (linear) function of the portfolio variables 𝑥: 𝑓𝑥,𝑦=−𝑦^𝑇𝑥=−𝑦₁𝑥₁+⋯+𝑦_𝑛𝑥_𝑛. (2)

We assume that the random vector 𝑦 has a probability density function denoted by 𝑝(𝑦). For a fixed decision vector 𝑥, the cumulative distribution function of the loss associated with that vector is computed as follows: 𝜓𝑥,𝛾=_{𝑓(𝑥,𝑦)≤𝛾}𝑝𝑦𝑑𝑦. (3)

Then, for a given confidence level 𝛼, the 𝛼-𝑉𝑎𝑅 associated with portfolio 𝑥 is represented as 𝑉𝑎𝑅_𝛼(𝑥)=min𝛾∈ℝ∣𝜓𝑥,𝛾≥𝛼. (4)

Also, we define the 𝛼-𝐶𝑉𝑎𝑅 associated with portfolio 𝑥 as 𝐶𝑉𝑎𝑅_𝛼(𝑥)=1(1−𝛼)_{𝑓(𝑥,𝑦)≥𝑉𝑎𝑅𝛼(𝑥)}𝑓𝑥,𝑦𝑝𝑦𝑑𝑦. (5)

Theorem 1. We always have 𝐶𝑉𝑎𝑅_𝛼(𝑥)≥𝑉𝑎𝑅_𝛼(𝑥), that means 𝐶𝑉𝑎𝑅 of a portfolio is always at least as big as its 𝑉𝑎𝑅. Consequently, portfolios with small 𝐶𝑉𝑎𝑅 also have small 𝑉𝑎𝑅. However, in general minimizing 𝐶𝑉𝑎𝑅 and 𝑉𝑎𝑅 are not equivalent.

Proof. See [4].

Since the definition of 𝐶𝑉𝑎𝑅 implies the 𝑉𝑎𝑅 function clearly, it is difficult to work with and optimize this function. Instead, the following simpler auxiliary function is considered [5]: 𝐹_𝛼1𝑥,𝛾=𝛾+(1−𝛼)_{𝑓(𝑥,𝑦)≥𝛾}𝑓𝑥,𝑦−𝛾𝑝𝑦𝑑𝑦, (6) and/or 𝐹_𝛼1𝑥,𝛾=𝛾+()𝑓1−𝛼𝑥,𝑦−𝛾⁺𝑝𝑦𝑑𝑦, (7) where 𝑎⁺=max{𝑎,0}. This function, considered as a function of 𝛾, has the following important properties that makes it useful for the computation of 𝑉𝑎𝑅 and 𝐶𝑉𝑎𝑅 [4]:(1)𝐹_𝛼 is a convex function of 𝛾.(2)𝑉𝑎𝑅_𝛼 is a minimizer over 𝛾 of 𝐹_𝛼.(3)The minimum value over 𝛾 of the function 𝐹_𝛼 is 𝐶𝑉𝑎𝑅_𝛼.

As a consequence of the listed properties, we immediately deduce that, in order to minimize 𝐶𝑉𝑎𝑅_𝛼(𝑥) over 𝑥, we need to minimize the function 𝐹_𝛼(𝑥,𝛾) with respect to 𝑥 and 𝛾 simultaneously min_𝑥𝐶𝑉𝑎𝑅_𝛼(𝑥)=min_𝑥,𝛾𝐹_𝛼𝑥,𝛾. (8)

Consequently, we can optimize 𝐶𝑉𝑎𝑅 directly, without needing to compute 𝑉𝑎𝑅 first. Since we assumed that the loss function 𝑓(𝑥,𝑦) is the convex (linear) function of the portfolio variables 𝑥, 𝐹_𝛼(𝑥,𝛾) is also a convex (linear) function of 𝑥. In this case, provided the feasible portfolio set Ω is also convex, the optimization problems in (8) are convex optimization problems that can be solved using well-known optimization techniques for such problems.

Instead of using the density function 𝑝(𝑦) of the random events in formulation (7) that it is often impossible or undesirable to compute it, we can use a number of scenarios in the names of 𝑦_𝑖 for 𝑖=1,…,𝑇. In this case, we consider the following approximation to the function 𝐹_𝛼(𝑥,𝛾)𝐹_𝛼1𝑥,𝛾=𝛾+(1−𝛼)𝑇^𝑇_𝑖=1𝑓𝑥,𝑦_𝑖−𝛾⁺. (9)

Now, in the problem min_𝑥𝐶𝑉𝑎𝑅_𝛼(𝑥), we replace 𝐹_𝛼(𝑥,𝛾) with 𝐹_𝛼(𝑥,𝛾)min_𝑥,𝛾1𝛾+(1−𝛼)𝑇^𝑇_𝑖=1𝑓𝑥,𝑦_𝑖−𝛾⁺. (10)

To solve this optimization problem, we introduce artificial variables 𝑧_𝑖 to replace (𝑓(𝑥,𝑦_𝑖)−𝛾)⁺. To do so, we add the constraints 𝑧_𝑖≥0 and 𝑧_𝑖≥𝑓(𝑥,𝑦_𝑖)−𝛾 to the problem [5]: min_{𝑥,𝑧,𝛾}1𝛾+(1−𝛼)𝑇^𝑇_𝑖=1𝑧_𝑖s.t.𝑧_𝑖𝑧≥0,𝑖=1,…,𝑇_𝑖≥𝑓𝑥,𝑦_𝑖−𝛾,𝑖=1,…,𝑇𝑥∈Ω. (11)

It should be noted that risk managers often try to optimize risk measure while expected return is more than a threshold value. In this case, we can represent mean-𝐶𝑉𝑎𝑅 model as follows: min_{𝑥,𝑧,𝛾}1𝛾+(1−𝛼)𝑇^𝑇_𝑖=1𝑧_𝑖s.t.𝜇^𝑇𝑧𝑥≥𝑅,_𝑖𝑧≥0,𝑖=1,…,𝑇_𝑖≥𝑓𝑥,𝑦_𝑖−𝛾,𝑖=1,…,𝑇𝑥∈Ω, (12) or min_{𝑥,𝑧,𝛾}−𝜇^𝑇1𝑥+𝜆𝛾+(1−𝛼)𝑇^𝑇_𝑖=1𝑧_𝑖s.t.𝑧_𝑖𝑧≥0,𝑖=1,…,𝑇_𝑖≥𝑓𝑥,𝑦_𝑖−𝛾,𝑖=1,…,𝑇𝑥∈Ω, (13) where the first constraint of problem (12) indicates that the expected return is no less than the target value 𝑅 and 𝜆≥0 used in problem (13) is risk aversion parameter that adapts the balance between expected return and 𝐶𝑉𝑎𝑅_𝛼(𝑥). It is important to note that there is an equivalence between 𝑅 and 𝜆 so that the problems (12) and (13) generate the same efficient frontiers. Since 𝑓(𝑥,𝑦) is linear in 𝑥, all the expressions 𝑧_𝑖≥𝑓(𝑥,𝑦_𝑖)−𝛾 represent linear constraints, and therefore the problem is a linear programming problem that can be efficiently solved using the simplex or interior point methods.

3. CVaR Robust Mean-CVaR Model

One of the uncertain parameters for mean-𝐶𝑉𝑎𝑅 model is 𝜇, and using estimations for this parameter leads to an estimation risk in portfolio selection. In particular, small differences in the estimations of 𝜇 can create large changes in the composition of an optimal portfolio. One way to reduce the sensitivity of mean-𝐶𝑉𝑎𝑅 model to the parameter estimations is using robust optimization to determine the optimal portfolio under the worst case scenario in the uncertainty set of the expected return. To this end, we represented robust mean-𝐶𝑉𝑎𝑅 models with interval and ellipsoidal uncertainty sets in the previous studies that have been demonstrated in formulations (14) and (15), respectively, min_{𝑥,𝑧,𝛾}1𝛾+(1−𝛼)𝑇^𝑇_𝑖=1𝑧_𝑖s.t.𝜇^𝐿𝑇𝑧𝑥≥𝑅,_𝑖𝑧≥0,𝑖=1,…,𝑇_𝑖≥𝑓𝑥,𝑦_𝑖−𝛾,𝑖=1,…,𝑇𝑥∈Ω,min_{𝑥,𝑧,𝛾}1𝛾+(1−𝛼)𝑇^𝑇_𝑖=1𝑧_𝑖s.t.𝑀^𝑇𝑥≤𝜇^𝑇𝑧𝑥−𝑅,_𝑖𝑧≥0,𝑖=1,…,𝑇_𝑖≥𝑓𝑥,𝑦_𝑖−𝛾,𝑖=1,…,𝑇𝑥∈Ω, (14)(15) where 𝜇^𝐿 is a given vector and 𝑀 is a 𝑛-dimensional matrix.

Now, we present 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 portfolio optimization problem that estimation risk in mean return is measured by 𝐶𝑉𝑎𝑅. The 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 model specifies that an optimal portfolio based on the tail of the mean loss distribution and the adjustment of the confidence level with regard to the preference of the investor corresponds to the adjustment of the conservative level, considering the uncertainty of the mean return [11].

In this model, 𝐶𝑉𝑎𝑅 is used to measure the risk of the portfolio return as before. In addition, when using the mean-𝐶𝑉𝑎𝑅 model, we consider the uncertainty of the expected return that can be considered as estimation risk and use 𝐶𝑉𝑎𝑅 to measure estimation risk. 𝐶𝑉𝑎𝑅 with this perspective is denoted as 𝐶𝑉𝑎𝑅^𝜇 (We use 𝐶𝑉𝑎𝑅^𝑦 to denote the 𝐶𝑉𝑎𝑅 risk measure discussed in Section 2 in order to differentiate it from 𝐶𝑉𝑎𝑅^𝜇 and also we use 𝐹^𝑦_𝛼(𝑥,𝛾) to denote its associated 𝐹_𝛼(𝑥,𝛾)). Thus, considering the problem (13), a 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 portfolio will be determined as the solution of the following optimization problem: min_{𝑥,𝑧,𝛾}𝐶𝑉𝑎𝑅^𝜇_𝛽−𝜇^𝑇𝑥1+𝜆𝛾+(1−𝛼)𝑇^𝑇_𝑖=1𝑧_𝑖s.t.𝑧_𝑖𝑧≥0,𝑖=1,…,𝑇_𝑖≥𝑓𝑥,𝑦_𝑖−𝛾,𝑖=1,…,𝑇𝑥∈Ω. (16)

For a portfolio of 𝑛 assets, we assume 𝜇∈ℝ^𝑛 is the random vector of the expected returns of the assets with a probability density function 𝑝(𝜇). To determine the mean loss of the portfolio, we define mean loss function, 𝑓(𝑥,𝜇), as follows [11]: 𝑓𝑥,𝜇=−𝜇^𝑇𝑥=−𝜇₁𝑥₁+⋯+𝜇_𝑛𝑥_𝑛. (17)

So, for confidence level 𝛽, 𝐶𝑉𝑎𝑅^𝜇_𝛽(−𝜇^𝑇𝑥) can be defined as follows: 𝐶𝑉𝑎𝑅^𝜇_𝛽−𝜇^𝑇𝑥=min_𝜂1𝜂+𝐸𝑓1−𝛽𝑥,𝜇−𝜂⁺. (18)

According to the definition of 𝐶𝑉𝑎𝑅^𝜇_𝛽 and 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 model, we find that 𝐶𝑉𝑎𝑅^𝜇_𝛽 will increase as the value of 𝛽 increases. This corresponds to taking more pessimism on the estimation risk in 𝜇 in the model and to optimizing the portfolio under worse cases of the mean loss. Thus, the resulting 𝐶𝑉𝑎𝑅 robust portfolio is more conservative. Conversely, conservatism of the portfolio is reduced as the value of 𝛽 decreases [11]. In Section 4, we will illustrate the impact of the value of 𝛽 on the conservatism level of a portfolio and also on the value of the maximum expected return of the portfolio.

As before, we can consider an auxiliary function to simplify the computations: 𝐹^𝜇_𝛽1𝑥,𝜂=𝜂+1−𝛽_𝜇∈ℝ^𝑛𝑓𝑥,𝜇−𝜂⁺𝑝𝜇𝑑𝜇, (19) and use the following approximation to the function 𝐹^𝜇_𝛽(𝑥,𝜂): 𝐹^𝜇_𝛽1𝑥,𝜂=𝜂+1−𝛽𝑚^𝑚_𝑖=1−𝜇^𝑇_𝑖𝑥−𝜂⁺, (20) where 𝜇₁,…,𝜇_𝑚 are a collection of 𝑚 independent samples for 𝜇 based on its density function 𝑝(𝜇). We can show that [5] min_𝑥∈Ω𝐶𝑉𝑎𝑅^𝜇_𝛽−𝜇^𝑇𝑥=min_{𝑥∈Ω,𝜂}𝐹^𝜇_𝛽𝑥,𝜂. (21)

So, with introducing artificial variables 𝑣_𝑖 to replace (𝑓(𝑥,𝜇_𝑖)−𝜂)⁺ in 𝐹^𝜇_𝛽(𝑥,𝜂) and adding the constraints 𝑣_𝑖≥0 and 𝑣_𝑖≥𝑓(𝑥,𝜇_𝑖)−𝜂 to the problem to do so, the 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 portfolio optimization problem becomes min_{𝑥,𝑧,𝑣,𝛾,𝜂}1𝜂+1−𝛽𝑚^𝑚_𝑖=1𝑣_𝑖1+𝜆𝛾+(1−𝛼)𝑇^𝑇_𝑖=1𝑧_𝑖s.t.𝑧_𝑖𝑧≥0,𝑖=1,…,𝑇_𝑖≥𝑓𝑥,𝑦_𝑖𝑣−𝛾,𝑖=1,…,𝑇_𝑖𝑣≥0,𝑖=1,…,𝑚_𝑖≥𝑓𝑥,𝜇_𝑖−𝜂,𝑖=1,…,𝑚𝑥∈Ω. (22)

This problem has 𝑂(𝑚+𝑛+𝑇) variables and 𝑂(𝑚+𝑛+𝑇) constraints that 𝑚 is the number of 𝜇-samples, 𝑛 is the number of assets, and 𝑇 is the number of 𝑦-scenarios. When the number of 𝑦-scenarios and 𝜇-samples increase, the approximations are getting closer to the exact values. But the computational cost significantly increases and thus makes the method inefficient.

Instead of this method, we can more efficiently determine the 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 portfolios using the smoothing method suggested by Alexander et al. [10]. Alexander presented the following function to approximate 𝐹^𝜇_𝛽(𝑥,𝜂): 𝐹^𝜇_𝛽1𝑥,𝜂=𝜂+1−𝛽𝑚^𝑚_𝑖=1𝜌_𝜀−𝜇^𝑇_𝑖𝑥−𝜂, (23) where 𝜌_𝜀(𝑎) is defined as follows: 𝜌_𝜀(𝑎)=𝑎𝑎𝑎≥𝜀,₂+14𝜀21𝑎+4𝜀−𝜀≤𝑎≤𝜀,0o.w. (24)

For a given resolution parameter 𝜀>0,𝜌_𝜀(𝑎) is continuous differentiable and approximates the piecewise linear function max(𝑎,0) [10]. We can also use this function to approximate 𝐹^𝑦_𝛼(𝑥,𝛾) as follows: 𝐹^𝑦_𝛼1𝑥,𝛾=𝛾+(1−𝛼)𝑇^𝑇_𝑖=1𝜌_𝜀𝑓𝑥,𝑦_𝑖−𝛾. (25)

Using smoothing method, the 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 model is as follows: min_{𝑥,𝛾,𝜂}1𝜂+1−𝛽𝑚^𝑚_𝑖=1𝜌_𝜀−𝜇^𝑇_𝑖1𝑥−𝜂+𝜆𝛾+(1−𝛼)𝑇^𝑇_𝑖=1𝜌_𝜀−𝑦^𝑇_𝑖𝑥−𝛾s.t.𝑥∈Ω. (26)

This formulation has 𝑂(𝑛) variables and 𝑂(𝑛) constraints. Thus, the number of variables and constraints does not change as the size of 𝜇-samples (𝑚) and 𝑦-scenarios (𝑇) increases. The efficiency of the smoothing approach is shown in Section 4.

4. Numerical Results

In this section, first we will compare the performance of the 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 model with robust mean-𝐶𝑉𝑎𝑅 models using interval and ellipsoidal uncertainty sets by actual data. Then, we will compare time required to compute the 𝐶𝑉𝑎𝑅 robust portfolios using problems (22) and (26). The dataset used here contains available returns for eight assets that expected return [12] and covariance matrix of the return of assets which have been given in Tables 1 and 2. In addition, the computations are based on 10,000 𝜇-samples generated from the Monte Carlo resampling (RS) method introduced in [13] and 96𝑦-scenarios obtained via computer simulation. It should be further noted that the computation is performed in MATLAB version 7.12, and run on a Core i5 CPU 2.40 GHz Laptop with 4 GB of RAM. Problems are solved using CVX [14] and function “fmincon” in Optimization Toolbox of MATLAB.

Table 1 

Mean return vector 𝜇.

Table 2 

Covariance matrix 𝑄.

4.1. Sensitivity to Initial Data

To show the sensitivity of the 𝐶𝑉𝑎𝑅 robust portfolios to initial data, we repeat RS sampling technique 100 times. Each of Figures 1, 2 and 3 display 100 𝐶𝑉𝑎𝑅 robust actual frontiers (actual frontiers are obtained by applying the true parameters on the portfolio weights derived from their estimated values [15]) for 𝛽=99%,90%,75%, respectively.

Figure 1 

100 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 actual frontiers (𝛽=99%).

Figure 2 

100 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 actual frontiers (𝛽=90%).

Figure 3 

100 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 actual frontiers (𝛽=75%).

As can be seen from figures, the 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 actual frontiers change with initial data used to generate 𝜇 samples. Also, this changes increase as the confidence level 𝛽 decreases. Thus, we can regard 𝛽 as an estimation risk aversion parameter. With these qualities, an investor who is more averse to estimation risk will choose a larger 𝛽. On the other hand, an investor who is more tolerant to estimation risk may choose a smaller 𝛽.

4.2. Portfolio Diversification

As we know, diversification decreases risk [4]. Portfolio diversification indicates distributing investment among assets in the portfolio. We illustrate in the following that compared with the robust mean-𝐶𝑉𝑎𝑅 portfolios with interval and ellipsoidal uncertainty sets, the 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 portfolios are more diversified. In addition, the diversification of the 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 portfolios decreases as the confidence level 𝛽 decreases.

To do so, we compute the 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 portfolios (for 𝛽=99%,90%,75%) and robust mean-𝐶𝑉𝑎𝑅 portfolios with interval and ellipsoidal uncertainty sets for the 8-asset example. The composition graphs of the resulting optimal portfolios are presented in Figures 4, 5, 6, 7, and 8. Considering these figures, when the expected return value increases from left to right, the allocated assets in the portfolios with minimum expected return are replaced by a composition of other assets, gently. Observing the right-most end of each graph, we can conclude the composition of the assets of the portfolio achieved from 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 model with 𝛽=99% is more diversified than that achieved from other models.

Figure 4 

Composition of robust mean-𝐶𝑉𝑎𝑅 portfolio weights with interval uncertainty set.

Figure 5 

Compositions of robust mean-𝐶𝑉𝑎𝑅 portfolio weights with ellipsoidal uncertainty set.

Figure 6 

Compositions of 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 portfolio weights (𝛽=99%).

Figure 7 

Compositions of 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 portfolio weights (𝛽=90%).

Figure 8 

Compositions of 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 portfolio weights (𝛽=75%).

In Figures 9, 10, and 11, the 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 actual frontiers for different values of 𝛽 are compared with robust mean-𝐶𝑉𝑎𝑅 actual frontiers with interval and ellipsoidal uncertainty sets and mean-𝐶𝑉𝑎𝑅 true efficient frontier. Since portfolios on the robust mean-𝐶𝑉𝑎𝑅 actual frontiers with interval and ellipsoidal uncertainty sets are less diversified, they should accept more risk for a given level of expected return and also achieve a lower maximum expected return. Consequently, in the figures, their actual frontiers are more right and lower than the other frontiers and this is one of the disadvantages of the low diversification in the portfolio. Seeing these frontiers, we also deduce that the maximum expected return and the associated return risk increase as the confidence level 𝛽 decreases. But in this case, the variations on the compositions of the resulting maximum-return portfolios might be large, and so the exact solution will not be always achieved. Instead, the maximum expected return of the portfolio is low for 𝛽=99%, and the variations will be low. So, the probability of having poor performance of the portfolio will be reduced when there is a big estimation risk of 𝜇. Thus, resulting robust portfolios will be too conservative. Consequently, an investor who is more risk averse to estimation risk selects a larger 𝛽 and obtains a more diversified portfolio. This justifies that it is reasonable to regard 𝛽 as a estimation risk aversion parameter.

Figure 9 

Robust mean-𝐶𝑉𝑎𝑅 with interval and ellipsoidal uncertainty sets and 𝐶𝑉𝑎𝑅 robust (𝛽=99%) actual frontiers.

Figure 10 

Robust mean-𝐶𝑉𝑎𝑅 with interval and ellipsoidal uncertainty sets and 𝐶𝑉𝑎𝑅 robust (𝛽=90%) actual frontiers.

Figure 11 

Robust mean-𝐶𝑉𝑎𝑅 with interval and ellipsoidal uncertainty sets and 𝐶𝑉𝑎𝑅 robust (𝛽=75%) actual frontiers.

4.3. Comparison of the Efficiency of Two Approaches for Computing 𝐶𝑉𝑎𝑅 Robust Portfolios

In Section 3, we introduced two formulations (22) and (26) to compute 𝐶𝑉𝑎𝑅 robust portfolios. Now, we show that computing 𝐶𝑉𝑎𝑅 robust portfolios via the smoothing approach (problem (26)) is more efficient. To do so, we compare the time required to solve problems (22) and (26) with different number of assets and different number of 𝜇-samples and 𝑦-scenarios. The results have been given in Table 3.

Table 3 

Time required to compute maximum-return (𝜆=0) portfolios for LP and smoothing approach (𝛽=99%, 𝜀=0.005).

As we see, the time required to compute 𝐶𝑉𝑎𝑅 robust portfolios via two approaches differs significantly when the sample size and the number of assets increase. For example, the time required to solve the 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 problem by two approaches for a problem with 8 assets and 5000 samples and 500 scenarios differs slightly. But, when the number of assets is more than 50, the number of scenarios is more than 500, and the sample size is more than 5000, differences become significant. A problem with 148 assets, 3000 𝑦-scenarios, and 25000 𝜇-samples is solved in less than 80 seconds using smoothing technique, while by (22) it took over 185 seconds. These comparisons show that when the number of scenarios and samples becomes larger, the smoothing approach is more computationally efficient to determine 𝐶𝑉𝑎𝑅 robust portfolios than other approach.

5. Conclusions

Since risk and return are uncertain parameters in the portfolio optimization, thus their estimation might lead to large error in the final decision. One of the widely used and effective approaches in optimization to deal with data uncertainty is robust optimization. In this paper, we have presented the robust counterpart of mean-𝐶𝑉𝑎𝑅 portfolio selection problem under the estimation risk in mean return. We have additionally used 𝐶𝑉𝑎𝑅 as risk measure to measure the estimation risk in mean return. To solve the model efficiently, the smoothing technique of Alexander et al. [10] is utilized. The performance of the 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 model with robust mean-𝐶𝑉𝑎𝑅 models for both interval and ellipsoidal uncertainty sets is compared. Our experiments have verified that the 𝐶𝑉𝑎𝑅 robust mean-𝐶𝑉𝑎𝑅 portfolios are more diversified. Applying the model for large and real data sets can be considered for future research.

Conflict of Interests

The authors declared that they have no conflict of interests.

Acknowledgment

The authors would like to thank office of vice president for research, University of Guilan for supporting this research.