Knowledge-Based Systems 198 (2020) 105944

Contents lists available at ScienceDirect

Knowledge-Based Systems
journal homepage: www.elsevier.com/locate/knosys

A parallel variable neighborhood search algorithm with quadratic

programming for cardinality constrained portfolio optimization
∗
Mehmet Anil Akbay, Can B. Kalayci , Olcay Polat
Pamukkale University, Faculty of Engineering, Department of Industrial Engineering, Kinikli, 20160, Denizli, Turkey

article info a b s t r a c t

Article history: Over the years, portfolio optimization remains an important decision-making strategy for investment.
Received 3 December 2019 The most familiar and widely used approach in the field of portfolio optimization is the mean–
Received in revised form 16 April 2020 variance framework introduced by Markowitz. Following this pioneering work, many researchers have
Accepted 18 April 2020
extended this model to make it more practical and adapt to real-life problems. In this study, one of
Available online 23 April 2020
these extensions, the cardinality constrained portfolio optimization problem, is considered. Cardinality
Keywords: constraints transform the quadratic optimization model into the mixed-integer quadratic programming
Metaheuristics problem, which is proved to be NP-Hard, making it harder to obtain an optimal solution within a
Variable neighborhood search reasonable time by using exact solution methodologies. Hence, the vast majority of the researchers
Asynchronous parallelization have taken advantage of approximate algorithms to overcome arising computational difficulties. To
Quadratic programming develop an efficient solution approach for cardinality constrained portfolio optimization, in this study,
Portfolio optimization a parallel variable neighborhood search algorithm combined with quadratic programming is proposed.
While the variable neighborhood search algorithm decides the combination of assets to be held in
the portfolio, quadratic programming quickly calculates the proportions of assets. The performance
of the proposed algorithm is tested on five well-known datasets and compared with other solution
approaches in the literature. Obtained results confirm that the proposed solution approach is very
efficient especially on the portfolios with low risk and highly competitive with state-of-the-art
algorithms.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction Although the mean–variance framework of Markowitz re-

mains a prominent strategy for the financial investment field
One of the key issues in investment is to make a higher profit explaining the trade-off between risk and return, because of
at a reasonable level of risk. In this context, one of the most the lack of real-life constraints the original model falls short to
common strategies is the ‘‘portfolio selection’’ defined as the generate realistic solutions for real-life problems. For instance,
compiling of the best asset allocation among various alternatives holding too many assets in a specific portfolio may not be log-
by considering the tradeoff between two conflicting objectives; ical for real-life cases because of additional costs. Therefore,
return maximization and risk minimization. The main idea behind an additional constraint, namely cardinality constraint (CC) is
the portfolio selection strategy is to spread the capital among introduced to the original model to restrict the number of assets
multiple securities efficiently rather than investing in a single se- [3]. It is also evident from the advancing literature that there
curity. In the early 50 s, Markowitz [1] proposed a mean–variance is an increasing interest in enhancing the original model with
framework as the pioneering work of the modern portfolio era real-life constraints, especially CC and boundary constraints (BC)
for the portfolio selection problem. Although diversification of [4]. Even though it is possible to solve the standard mean–
investment alternatives has been a long-standing practice, the variance model by using convex quadratic programming (QP),
mean–variance approach has become a revolutionary work that adding CC and BC transforms the structure of the original model
brings a quantitative strategy and modeling approach from the from quadratic optimization into the mixed-integer quadratic
operations research perspective for selecting best asset combina- programming problem, namely cardinality constrained portfolio
tion assuming that the historical prices of the market can reflect optimization (CCPO) problem, which is proved to be NP-Hard [5].
the future prices [2]. Since exact solution techniques may be ineffective to solve this
problem, there is an increasing interest in developing efficient
∗ Corresponding author. heuristic algorithms to obtain near-optimal solutions.
E-mail addresses: [email protected] (M.A. Akbay), In this study, an efficient two-phased solution methodology
[email protected] (C.B. Kalayci), [email protected] (O. Polat). combining a metaheuristic (variable neighborhood search) and

https://doi.org/10.1016/j.knosys.2020.105944
0950-7051/© 2020 Elsevier B.V. All rights reserved.
2 M.A. Akbay, C.B. Kalayci and O. Polat / Knowledge-Based Systems 198 (2020) 105944

exact (quadratic programming) approach is presented to solve the computational difficulty, there is a limited attempt to solve the
CCPO problem. While most of the researchers consider the stock problem by exact solution approaches with relaxation meth-
selection process with the determination of the asset weights ods and QP mechanisms. The publications utilizing various so-
simultaneously, in the proposed solution approach, selection of lution approaches for the CCPO problem along with the broad
the assets included in the portfolio and calculation of the optimal classification is presented in Fig. 1.
proportions are carried out separately, which contributes to a Overall, solution approaches developed for the CCPO problem
great algorithmic performance with simplicity and effectiveness. significantly increased in the last decade while population-based
Moreover, compared to other solution methodologies such as algorithms have the greatest number of implementations. Al-
population-based metaheuristic algorithms, VNS does not require though, the performance of the algorithm and its superiority over
long algorithm run times and relatively large memory resources. others may vary according to the type of the problem applied,
Therefore, these features enable it to be applied successfully on one of the most preferred approaches among population-based
CCPO as well as other combinatorial and global optimization solution approaches is evolutionary algorithms due to advantages
problems. Firstly, the variable neighborhood search (VNS) algo- such as conceptual simplicity, a wide range of application areas
rithm selects the assets to be included in the portfolio. Once and the possibility of hybridization with other algorithms. Chang
the assets are decided, the mixed-integer quadratic programming et al. [3] introduced cardinality constraints to the MVPO problem
problem is transformed back into the quadratic programming for the first time and analyzed the effect on the efficient frontier
model without the need for any other procedure to satisfy CC. as well as arising difficulty of calculating optimal solution by
Afterward, exact optimal proportions for the previously selected comparing the performance of genetic algorithms, tabu search,
assets are determined via QP. and simulated annealing algorithms. Based on the formulation
To the best of our knowledge, this study presents the first of Chang et al. [3], Woodside-Oriakhi et al. [20] using the same
implementation of VNS combined with an exact solution method heuristics, integrated subset optimization problem in which the
in the portfolio optimization (PO) literature. In addition to the desired return is introduced as inequality. Anagnostopoulos and
two-stage algorithm structure, the initial solution construction Mamanis [27] applied various evolutionary algorithms to success-
procedure, sequential neighborhood exchange, optimal pool size fully deal with the computational challenges of the constrained
selection mechanisms, and asynchronous parallelization strat- portfolio optimization model. Especially on large scale instances,
egy can be counted as critical algorithmic components of the the single objective algorithms are outperformed by the multi-
proposed solution approach which distinguish it from other al- objective ones. Chen et al. [42] proposed an extended algorithm
gorithms in the literature and provide superiority. Finally, the based on a multi-objective evolutionary structure combining local
proposed algorithm is tested on well-known benchmark datasets
search and non-dominated sorting to solve CCPO. Soleimani et al.
and compared with the other solution algorithms in the litera-
[25] introduced the market (sector) capitalization constraint in
ture. Experimental results are presented to show the effectiveness
addition to the transaction lots and cardinality constraint for the
of the developed algorithm along with the evaluation of the
first time in the literature and utilized a genetic algorithm to solve
competitiveness against other methodologies presented in the
the problem. Chang et al. [22] proposed GA to solve the CCPO
literature.
problem using different risk measures such as semi-variance,
The remainder of this paper is structured as follows: The
mean absolute deviation and variance with skewness. Besides, Pai
literature review on the CCPO problem is presented in Section 2.
and Michel [50] proposed an evolutionary optimization algorithm
The mathematical formulation of the problem is given and ex-
to solve the constrained PO problem and utilized a k-means
plained in Section 3. Detailed information about the proposed
clustering analysis to handle cardinality constraint. Because of the
solution approach is presented in Section 4. The results of the
limited capabilities of existing techniques in solving large-scale
experimental study are reported in Section 5 while Section 6
combinatorial problems, Liagkouras [31] introduced a new MOEA
presents the conclusion with future research directions.
incorporates a new three-dimensional encoding structure.
2. Literature review Besides the popularity of the evolutionary-based algorithms,
swarm-based solution approaches have also attracted consider-
Following Markowitz’s inspiring work, the portfolio optimiza- able attention in the last decade. Cura [32] applied particle swarm
tion (PO) problem attracted a lot of interest in the literature. optimization (PSO) approach for the first time on mean–variance
Surveys focusing on this problem from different perspectives CCPO problem and compared the performances of the proposed
are summarized as follows: Kolm et al. [6] addressed portfolio algorithm with a genetic algorithm (GA), tabu search (TS), and
optimization from a practical perspective considering some of the simulated annealing (SA) algorithms on well-known data sets.
challenges encountered with real-life constraints while Metaxio- Following, Zhu et al. [51] performed a comparative return analysis
tis and Liagkouras [7] reviewed the design and implementations between the proposed PSO algorithm and the other solution
of multi-objective evolutionary algorithms (MOEAs). Aouni et al. approaches such as GA and the Visual Basic Application Solver
[8] took into consideration of exact attempts along with various considering both restricted and unrestricted scenarios. Golmakani
risk measures. Then, Ertenlice and Kalayci [9] reviewed swarm and Fazel [34] added three more constraints such as bounds
intelligence algorithms adopted to solve in portfolio optimization. on holdings, minimum transaction lots and sector capitaliza-
Recently, Kalayci et al. [4] presented a comprehensive review tion to the model and proposed an improved PSO with various
on deterministic mean–variance portfolio optimization (MVPO) mutation strategies to overcome the complex structure of this
problem analyzing various solution approaches, models, perfor- constrained model. Deng et al. [33] proposed an improved PSO
mance measures and real-life constraints. They mainly focused algorithm by utilizing reflection, minimum hold and mutation
on exact and inexact solution approaches from the perspective strategies for the CCPO problem presenting a better performance
of a computational analysis. Doering et al. [10] on the other than the original PSO algorithm, especially on low-risk portfolios.
side, examined how well the metaheuristic algorithms devel- Pouya et al. [52] adapted invasive weed optimization algorithm
oped for NP-Hard class portfolio optimization problems generate to solve multi-objective portfolio optimization transforming into
high-quality solutions by considering the relationship between a single-objective programming model using fuzzy normalization
portfolio optimization and risk management. and uniform design method. After a thorough computational
While a vast majority of the researchers adapt approximate analysis, Kalayci et al. [38] determined that repairing infeasible
methods to obtain solutions for the CCPO problem due to its solutions may have a negative impact on the convergence of the
 M.A. Akbay, C.B. Kalayci and O. Polat / Knowledge-Based Systems 198 (2020) 105944 3

Fig. 1. Solution approaches proposed for the CCPO problem. (See Refs. [3,11–20,20–49]).

algorithm. Therefore, they enhanced their artificial bee colony search procedure (GRASP) together with quadratic programming
(ABC) algorithm by using infeasibility toleration along with the where stock indices are selected by GRASP and proportion de-
feasibility enforcement strategies. Silva et al. [36] proposed a termination is optimized with QP. Recently, Kalayci et al. [40]
unified PSO algorithm to deal with different versions of the PO developed a hybrid solution approach based on continuous ant
problem introducing an adaptive ranking procedure based on colony optimization enhanced with critical components from
non-dominated sorting, crowding distance and cost–benefit. artificial bee colony and genetic algorithms to specifically deal
It is evident from the number of studies in the PO literature with the CCPO problem.
[4], single solution-based solution approaches are less preferable
than the population-based algorithms. Although not preferred as 3. Portfolio optimization with Cardinality constraints
a stand-alone solution approach, they are combined with many
other solution algorithms due to their powerful exploitation ca- The mathematical model of the CCPO based on Chang et al. [3]
pabilities, simple structures and quick running times. Schaerf [19] is presented as follows:
improved the local search techniques proposed by Chang et al. Parameters
[3] with a broader set of possible neighborhood structures to
λ The trade-off parameter of variance and return
solve the CCPO problem. Bačević et al. [21] developed a VNS
based heuristic for portfolio optimization problem together with εi and δi The minimum and the maximum proportion of an asset i
additional non-convex constraints such as market capitalization N Number of securities in the index
and sector constraints and cardinality constraints along with a K The exact number of securities to be held in the portfolio
rebalancing strategy. In this strategy, they implemented a lo-
cal search strategy based on a continuous VNS algorithm since σij The covariance between a security i and a security j
problem variables are continuous. µi The expected return of a security i
In recent studies, various researchers preferred solution Variables
methodologies hybridizing multiple methods rather than using a
wi Proportion of securities
single heuristic algorithm to overcome deficiencies and improve {
exploration and exploitation capabilities of the relevant tech- 1
zi = If asset i is included in portfolio 1; otherwise 0.
niques. Maringer and Kellerer [44] proposed a hybrid algorithm 0
combining simulated annealing and an evolutionary strategy ⎡ ⎤ [ N ]
N N
for the CCPO problem. While the proposed algorithm uses a ∑ ∑ ∑
min λ ⎣ wi wj σij − (1 − λ)
⎦ wi µ i (1)
population of crystals, elimination of the population members
i=1 j=1 i=1
is based on evolutionary principles and acceptance criteria are
based on the SA algorithm. Gaspero et al. [48] developed a hybrid Subject to :
solution approach based on local search and QP with cardinality, N
∑
quantity, and pre-assignment constraints. While local search is wi = 1 (2)
i=1
used for selecting assets to be included in the portfolio, QP is
N
used for determining asset weights for portfolio optimization. ∑
zi = K (3)
Baykasoglu et al. [47] adopted a greedy randomized adaptive i=1
4 M.A. Akbay, C.B. Kalayci and O. Polat / Knowledge-Based Systems 198 (2020) 105944

εi zi ≤ wi ≤ δi zi i = 1, . . . , N (4) controlled by slave processors while the main algorithm is run

via the master processor. Crainic et al. [69] developed coopera-
zi ∈ {0, 1} i = 1, . . . , N (5)
tive neighborhood VNS, i.e., asynchronous master–slave relation-
0 ≤ wi ≤ 1 i = 1, . . . , N (6) ship, considering various search control mechanisms, information
0 ≤ εi ≤ δi ≤ 1 i = 1, . . . , N (7) exchange methods and search strategy rules.

Eq. (1) involves two conflicting objectives, return maximiza-

tion and risk minimization. As a weighting parameter, λ deter-
5. The proposed methodology
mines the trade-off between return and risk by linearly increasing
its value from 0 to 1. When λ is equal to 0, then the model
looks for the maximum return, while when λ is equal to 1, The pseudo-code describing the main structure of the pro-
the model looks for the minimum risk. Eq. (2) guarantees that posed methodology is presented in Fig. 3. The related hybrid
the security proportions add to one. Eq. (3), namely cardinality solution approach consists of two phases: selecting the assets
constraint, restricts the portfolio size and ensures that exactly K included in the portfolio and determining the asset weights, re-
number of securities are held in the portfolio. Thus, the continu- spectively. While the assets included in the portfolio are selected
ous structure of the efficient frontier turns into discontinuous by by the VNS algorithm, the asset proportions are determined by
introducing cardinality constraint to the model [3]. This transfor- using QP.
mation increases the complexity and computational difficulty of The algorithm starts with the construction of an initial solution
the problem. Eq. (4) puts a restriction on the asset proportions and asset selection pool. Afterward, the main part of the algo-
to be the pre-determined minimum and maximum limit values. rithm searches for the optimal portfolio by iteratively applying
Eqs. (5)–(7) presents decision variables and variable domains. shaking, local search, and quadratic programming procedures.
The constructed loop structure continues to run until the max-
4. Variable neighborhood search and parallelization strategies imum number of iterations is reached. If the obtained solution
after shaking procedure and subsequently applying QP is im-
VNS is a single solution and local search-based metaheuristic proved, shaking is continued with the improved one without
algorithm developed by Mladenovic and Hansen [53] to solve applying local search. Otherwise, the algorithm continues to the
combinatorial optimization problems. Unlike most of the local local search phase with the solution generated via the shaking
search-based algorithms, VNS uses a set of neighborhood struc- procedure. If the solution is improved in either shaking or local
tures instead of a single one. The main idea behind using multiple search phases supported with QP, the algorithm is continued
neighborhood structure is the fact that a local minimum within with the initial neighborhood structure. Otherwise, i.e., if both
a neighborhood may not be so for another one. Furthermore, solutions obtained from shaking and local search are failed to
VNS systematically changes the neighborhoods during the search. ensure improvement, neighborhood is increased, and the search
Thus, it ensures diversification in the search space and over- continues with the next shaking procedure.
comes the problem of getting stuck into local optima frequently If improvement is achieved (Fig. 3, Step 22), the variable
encountered in local search-based heuristics. neighborhood search algorithm is set to work back from the
Let Gk , (k = 1, . . . , kmax ) and Gk (x) signify the set of neighbor- first neighborhood (Fig. 3, Step 24), i.e. is restarted. Therefore, it
hoods and the set of solutions within the kth neighborhood of can be considered as a recursive algorithm that brings a higher
the solution x, respectively. The main structure of the basic VNS computational complexity. However, the maximum number of
is presented via the pseudo-code given in Fig. 2. evaluations (Fig. 3, Step 9 and Step 28) limits the algorithm
Compared to other heuristic algorithms, VNS provides sub- execution to a certain level of time. On the other hand, if an
stantial advantages over other algorithms due to its simple struc- improvement is not achieved, the algorithm can be considered
ture, integrability with different solution techniques and requir- as iterative, i.e., the time complexity of the proposed algorithm is
ing very few parameters [54]. Moreover, as with population- directly proportional to the number of assets held in the portfolio
based algorithms, it does not require a long algorithm run times (K ) and pool size used in the local search phase (poolsizeL ). In
and relatively large memory resources. These features enable it the local search step, utilizing the best improvement strategy
to be applied successfully on various combinatorial and global requires O(K × (poolsizeL − K )) of time complexity which results
optimization problems such as location problems [55], clustering in calling of quadratic programming function (K ×(poolsizeL − K ))
problems [56], graph problems [57], knapsack and packing prob- times. Hence, the worst-case time complexity of the algorithm is
lems [58], timetabling [59], various scheduling problems [60,61] proportional to the capability of the MATLAB Quadprog function.
and vehicle routing problems [62,63], assembly and disassembly The outer loop (Fig. 3, Step 12) brings a time complexity of (K 2 ×
line balancing [64–66] and portfolio optimization [21]. (poolsizeL − K )) for the entire algorithm. Moreover, the restriction
Although the basic VNS performs well in many different com- of the search pool and running the algorithm simultaneously on
binatorial optimization problem types, various extensions [67] the multiple processors significantly reduce the time complexity
have been developed to overcome problem-specific difficulties of the variable neighborhood search algorithm.
and improve the performance of the algorithm. Furthermore, with As the main components of the proposed approach; initial
the development of hardware technology, it is possible to run solution construction, neighborhood selection as well as shak-
various algorithm structures simultaneously using multiple pro- ing and local search procedures are presented in detail in the
cessors. Thus, distributing some part of the algorithm among the following sections.
available processors reduces computational time and enhances
the exploration capability of the relevant solution methodologies.
In the literature, several parallelization strategies are developed 5.1. Asset selection
for the VNS algorithm. García-López et al. [68] compared three
parallelization strategies: synchronous parallel VNS, i.e., paral-
lelizing the local search phase; replicated parallel VNS, i.e., run- The asset selection phase of the proposed solution approach
ning several VNS simultaneously applying multi-start strategy; consists of three phases: constructing an initial solution and the
and replicated shaking VNS, i.e., shaking and local search are search pool (i), shaking (ii) and local search (iii).
 M.A. Akbay, C.B. Kalayci and O. Polat / Knowledge-Based Systems 198 (2020) 105944 5

Fig. 2. Variable neighborhood search algorithm [54].

Fig. 3. Pseudo-code of the proposed solution approach.

5.1.1. Constructing initial solution and search pool risk levels. Since, negative values may cause some calculation
The initial solution can be generated randomly [3,33,38], or errors, Eqs. (10) and (11) are used to avoid miscalculations. Then,
through various solution construction strategies [32,45,47]. How- c values for each asset are calculated by using Eq. (12), and
ever, using methodologies that enable reaching (near) optimal finally, all assets are sorted based on calculated c values. After
solutions as quickly as possible matters since the algorithm per- sorting operation, assets with the highest c value represent the
formance and convergence to the efficient frontier may be di- most preferable assets, whereas assets with the lower c value are
rectly affected. Therefore, in this study, rather than starting with classified as relatively less preferable.
a random initial solution, the method proposed by Cura [32] has
been adapted to both construct an initial solution and to construct
the search pool to be used in the asset selection stage of the θi = 1 + (1 − λ)µi i = 1, . . . , N (8)
∑N
j=1 σij
proposed solution approach.
According to this methodology, each asset is sorted based on ρi = 1 + λ i = 1, . . . , N (9)
N
c values, which indicate the proportion between average return
and average risk according to the trade-off parameter (λ). In some Ω = −1 × min(0, θ1 , . . . , θN ) (10)
cases, the datasets also include assets with negative returns and ψ = −1 × min(0, ρ1 , . . . , ρN ) (11)
6 M.A. Akbay, C.B. Kalayci and O. Polat / Knowledge-Based Systems 198 (2020) 105944

θi + Ω 5.1.3. Local search

ci = i = 1, . . . , N (12)
ρi + ψ In the local search step of the proposed solution approach, the
best improvement strategy allowing the algorithm to run until
Let a(i) represent each asset in the index. An example ini- all possible movements are done is applied. Although it is known
tial solution and search pool representation belong to FTSE 100 that the best improvement is much more time consuming com-
dataset are presented in Fig. 4. While the first K assets which pared to the first improvement strategy, it is reported that better
indicate the assets to be kept in the portfolio are accepted as an sometimes faster results are obtained with the best improvement
initial solution after determining assets weights via QP, the rest when a good initial solution generated by a constructive heuristic
of the assets constitute the selection pool which is going to be or an initial solution construction strategy rather than the random
used in the shaking and local search steps later on. initial solution [70].
In some cases, narrowing the search pool to a certain extent In the local search phase, each asset in the current portfolio
may improve search performance and thus enable the optimal is removed, and an asset chosen from the search pool is in-
solution to be found in less time. Therefore, instead of including serted. The pseudocode of the local search procedure used in the
all the assets in the search pool, it should be restricted. Example proposed algorithm is presented in Fig. 7.
pool sizes for 5 levels determined based on the total number of
assets holding in the portfolio (K = 10) and the total number 5.1.4. Parallelization
of assets in the index (N) are also shown in Fig. 4. Details about In this study, a simple parallelization strategy is applied aim-
narrowing the search pool and finding the optimal pool size are ing to shorten the computational time. Based on the three-
given in the further sections. dimensional taxonomy proposed by Crainic et al. [71], applied
After the initial solution and the search-pool are constructed, strategy is categorized as pC /C /MPSS. While the first dimension
shaking and local search steps are systematically applied. of the taxonomy represents the search control cardinality, and
the second and third dimensions describe communication control
and search strategy, respectively. As seen in Fig. 8, search control
is maintained by running the algorithm independently and asyn-
5.1.2. Shaking chronously on different processors for each λ value rather than
One of the most critical parts of the VNS algorithm is the using a master–slave structure. Communication and information
mechanism of changing neighborhoods systematically which has transfer between the processors are not realized. Furthermore,
a crucial importance for the search performance. Therefore, de- while the same search strategy is adopted in each VNS algorithm
signing an effective search procedure considering both explo- distributed to the processors because the objective function with
ration and exploitation capability of the algorithm becomes one different λ values yields different results, different initial solu-
of the most important goals of this solution approach. Following tions are used for each algorithm in threads. Consequently, it
the aforementioned goals; two main components of the shaking becomes possible to collect optimal results of different λ values
procedure: determining of neighborhood structures and neigh- simultaneously from each thread of the processor.
borhood exchange mechanisms are of great importance. On the It is always possible to develop different parallelization struc-
one hand, the chosen neighborhood structure and change strategy tures to improve performance considering the local search phase
should effectively diversify searching to avoid a stuck into local as one of the most time-consuming components of the sequential
optima, on the other hand, it should ensure to reach global VNS algorithm or to improve exploration capability by simultane-
optimum as quickly as possible. ously starting multiple searches at different points of the search
In the developed algorithm, each neighborhood represents space. See Crainic and Hail [72] for the parallel applications for
how many of the assets will be removed from the current solution the various metaheuristics.
and how many new assets will be inserted. This information
is transferred to the shaking procedure via the variable k = 5.2. Determination of asset proportions
(1, . . . , kmax ) as shown in Figs. 5 and 6 and the number of assets
is removed and inserted based on this transferred information After selection of the assets to be included in the optimal
in the shaking step. Initially, the shaking phase starts with re- portfolio, cardinality constraint is eliminated, and the model was
moving/inserting only one asset by selecting randomly, and the transformed into a simple quadratic programming. After the elim-
number of assets to be removed/inserted is increased. K denotes ination of the cardinality constraint, the complexity of the prob-
the desired number of assets included in the portfolio. The maxi- lem significantly decreased. Furthermore, the search space was
mum number of assets can be removed and inserted is K (kmax = considerably narrowed because the number of assets evaluated
K ). In each iteration, the solution WS is generated randomly from reduced from N to K . After this phase, the weights of the assets
the incoming solution W by exchanging k assets. In other words, can be calculated using a simple heuristic algorithm. However,
the solution WS is within the neighborhood Gk (W ). since the problem is simplified to a quadratic optimization model,
Besides neighborhood structures to be used, neighborhood it can be easily solved with exact solution techniques, rather than
exchange mechanisms have great importance on diversification. adopting a heuristic algorithm to obtain near-optimal solutions.
Over a pre-determined set of neighborhoods, various strategies Therefore, QP is utilized to get exactly optimal results in the
can be implemented such as random, stochastic or sequential second phase [47]. In this study, MATLAB quadratic programming
selection. In this study, sequential neighborhood selection is used. solver (MATLAB R2019a, Optimization Toolbox, Version 8.3) is
According to this strategy, changing starts with the initial neigh- used to calculate asset proportions since VNS is also developed
borhood and continues by increasing the neighborhood in each in the same platform, thus no integration necessity required.
iteration. At any stage of shaking along with the current neigh-
borhood (Gk ), if no improvement is made, shaking continues with 6. Computational experiments
the next neighborhood (Gk+1 ); otherwise, it continues with the
first neighborhood (G1 ) for the next iteration. 6.1. Implementation
Basic representation of the sequential neighborhood exchange
mechanisms and shaking structure are presented in Fig. 5. and in The proposed algorithm has been programmed in MATLAB
the pseudo-codes showed in Figs. 3 and 6. R2019a and the statistical tests are performed with MINITAB 18.
 M.A. Akbay, C.B. Kalayci and O. Polat / Knowledge-Based Systems 198 (2020) 105944 7

Fig. 4. An example of search-pool and assets in the initial portfolio for the FTSE 100 dataset.

Fig. 5. Representation of sequential neighborhood change.

Fig. 6. Pseudo-code of shaking procedure.

Fig. 7. Pseudo-code of local search procedure.

8 M.A. Akbay, C.B. Kalayci and O. Polat / Knowledge-Based Systems 198 (2020) 105944

Fig. 8. Parallelization strategy for the variable neighborhood search algorithm.

The computational experiments were run on a workstation with affect the average results as much as in the local search phase,
an Intel Xeon E5-2650 2.0 GHz processor and 32 GB RAM. selecting larger pool size appears to worsen the performance
slightly when the pool size applied for shaking is adjusted to
6.2. Test problems cover relatively less assets indicated by the first 3 level tested
(see Fig. 9).
The proposed algorithm is tested on five publicly available
benchmark datasets, which are widely utilized in the literature
6.4. Performance measures
[73] to analyze performance and make a comparison with other
solution approaches proposed by Chang et al. [3], Cura [32], Deng
et al. [33], Lwin and Qu [43], Baykasoglu et al. [47], and Kalayci To compare the performance of the proposed solution ap-
et al. [38]. These datasets contain weekly prices of Hang Seng, proach and the other algorithms in the literature, mean per-
DAX 100, FTSE 100, S&P 100, and Nikkei 225 indices between centage error (MEAPE), median percentage error (MEDPE) min-
March 1992 and September 1997. The number of stocks (N) for imum percentage error (MINPE) and maximum percentage er-
each dataset is 31, 85, 89, 98, and 225, respectively. ror (MAXPE) performance measures introduced by Chang et al.
[3] and mean Euclidean distance (MEUCD), mean return error
6.3. Parameter settings (MRE) and variance of return error (VRE) performance measures
introduced by Cura [32] have been used.
It is needed to determine appropriate parameter values for Let (xi , yi ) represents each discrete point on the standard
the proposed solution approach to ensure the best algorithmic efficient frontier and (x∗ , y∗ ) represents each optimal solution
performance. In this context, parameter tests were performed on presenting variance and return values respectively produced by
each dataset in line with the following assumptions: Number of the proposed algorithm.
assets included in the portfolio (K = 10), εi = 0.01 (i = 1, . . . , N) Firstly, x∗∗ and y∗∗ values representing linearly interpolated
and a total number of evaluations = (N × 1000). The number of horizontal and vertical projections of each value on CCEF respec-
replications = 5. Levels tested for the parameters of the proposed tively are calculated as presented in the following formulations:
methodology is presented in Table 1. (y∗ − yk )
[ ]
One of the most critical parameters that may affect search x∗∗ = xk + (xj − xk ) (13)
(yj − yk )
performance and speed of the algorithm is the size of the pool
(x∗ − xk )
[ ]
from which assets included in the portfolio are chosen. Although y∗∗ = yk + (yj − yk ) (14)
choosing a large search pool is good because of the increased (xj − xk )
diversity in the search, the algorithm may lose extra time to
where x′k and x′j represent the variance values where yk =
reach optimal solutions. On the other hand, narrowing the search
max [yi |yi ≤ y∗ ], yj = min [yi |yi ≥ y∗ ], y′k andy′j represent the re-
pool may increase speed as well as exploitation capability of
turn values where xk = max [xi |xi ≤ x∗ ] and xj = min [xi |xi ≥ x∗ ],
the algorithm, while posing the problem of getting caught in
respectively. Afterward, percentage deviation errors are calcu-
local optimum points. Therefore, it is needed to search for an
lated in both directions:
optimal pool size for the developed algorithm. However, due to ⏐ ∗
⏐ y − y∗∗ ⏐
⏐
the different asset selection mechanisms applied in the shak-
ing and local search, the optimal pool size to be used in these
ϕj = 100 ⏐⏐ ⏐ (15)
y∗∗ ⏐
stages may be different. Therefore, the pool size parameter is ⏐ ∗
⏐ x − x∗∗ ⏐
⏐
set separately for both the shaking and local search steps. For ωj = 100 ⏐⏐ ∗∗
⏐ (16)
each parameter, 5 levels are determined based on the number x⏐
of assets (N) in the relevant dataset. As K denotes the number of While Eq. (18). defines the percentage deviation on return,
assets in the portfolio, which is fixed to 10, an example pool size Eq. (20). indicates the percentage deviation of variance.
representation for the FTSE 100 dataset is presented in Table 2. ∑E
Based on the levels determined, the optimal pool size was j min(ϕi , ψi )
MEAPE = (17)
searched by analyzing the interaction between shaking and local
⏐E√
search steps. Test results belong to the FTSE dataset are presented ⏐ X ∗∗ − √X ∗ ⏐
⏐
in Figs. 9 and 10. Where ψi = 100 ⏐ (18)
⏐ ⏐
X ∗∗
⏐
Fig. 9. represents the interaction plot and Fig. 10 demonstrates
⏐ ⏐
the main effect plot for pool size applied in the shaking and local MEDPE = min (median {ϕ1 , ϕ2 , . . . , ϕE } , median {ψ1 , ψ2 , . . . , ψE })
search steps based on the mean of the optimal solution. According (19)
to the test results, selecting a narrower pool in local search
MINPE = min (min {ϕ1 , ϕ2 , . . . , ϕE } , min {ϕ1 , ϕ2 , . . . , ϕE }) (20)
yielded better results and algorithmic performance. Although the
pool size selected during the shaking phase did not significantly MAXPE = max (max {ϕ1 , ϕ2 , . . . , ϕE } , max {ϕ1 , ϕ2 , . . . , ϕE }) (21)
 M.A. Akbay, C.B. Kalayci and O. Polat / Knowledge-Based Systems 198 (2020) 105944 9

Table 1
Levels tested for the parameters of the proposed methodology.
Parameter Denotes to Levels tested
∆ Threshold for improvement Fixed to: 1e − 16
MaxIt Maximum number of evaluations Fixed to: Nx1000
poolsizeS Pool size for shaking K + N /5 K + N /4 K + N /3 K + N /2 N
poolsizeL Pool size for local search K + N /5 K + N /4 K + N /3 K + N /2 N

Table 2
Example pool size representation for the FTSE 100 dataset.
Parameter Denotes to Levels tested
K + N /5 K + N /4 K + N /3 K + N /2 N
poolsizeS Pool size for shaking 28 32 40 55 89
poolsizeL Pool size for local search 28 32 40 55 89

Fig. 9. Interaction plot of the pool size for the shaking and local search phases.

Fig. 10. Main effect plot of the pool size for the shaking and local search phases.
10 M.A. Akbay, C.B. Kalayci and O. Polat / Knowledge-Based Systems 198 (2020) 105944

Fig. 11. Efficient frontiers for the Hang Seng dataset.

Fig. 12. Efficient frontiers for the DAX 100 dataset.

∑E √ ∗∗∗
j=1 (x − x∗ ) + (y∗∗∗ − y∗ ) 0.01 (i = 1, . . . , N ) and δi = 1 (i = 1, . . . , N) for the problem
MEUCD = (22)
E formulation.
∑E 100 |(x∗∗∗ −x∗ )| Let e = (1, . . . , 50) denotes each solution point generated
VRE =
j=1 x∗
(23) by different approaches corresponding to each λ value. The first
E experimental test is done to analyze the performance of the
∑E 100 |( y∗∗∗ −y∗ )| proposed algorithm for both the low values of λ(e < 45) and
j=1 y∗
MRE =
E
(24) high values of λ(e ≥ 45) because the linearity/nonlinearity of
the objective function is changed according to the λ value. For
instance, while λ = 0, objective function turns into almost
6.5. Computational results linear, increasing of λ value result in increasing non-linearity of
the objective function. Thus, the convergence of the generated
solutions by different algorithms may be affected based on λ
In this section, the computational results of the proposed
parameter.
algorithm are presented comparing with the other solution ap-
Figs. 11–15 shows the efficient frontiers obtained by utilizing
proaches in the literature by testing on 5 well-known benchmark
the solution procedure consisting of initial solution construction
instances. All tests were done by accepting K = 10, εi =
 M.A. Akbay, C.B. Kalayci and O. Polat / Knowledge-Based Systems 198 (2020) 105944 11

Fig. 13. Efficient frontiers for the FTSE 100 dataset.

Fig. 14. Efficient frontiers for the S&P 100 dataset.

procedure combining with quadratic programming (INIQP) and the performance indices outlined in Section 4. Due to the fact
the heuristic efficient frontier together with the unconstrained ef- that performance measures, as well as the number of λ values,
ficient frontier in concern with the CCPO problem for the XU030, used varies in the compared studies, results are grouped into
Hang Seng, DAX 100, FTSE 100, S&P 100 datasets. In general, three. To make an accurate comparison, the maximum number
it is seen that VNS combined with QP (VNSQP) performs better of evaluations is set to 1000xN for the proposed methodology.
convergence to the unconstrained efficient frontier (UCEF) in Firstly, the performance comparison is carried out for the
comparison with the INIQP. However, while the distance between proposed solution approach and PSO algorithm proposed by Deng
the obtained efficient frontiers are closer for the lower values of et al. [33] and INIQP considering the low values of λ(e < 45) and
λ (e < 45), the difference becomes much more significant for the high values of λ(e ≥ 45). Table 3 presents the mean percentage
higher values of λ (e ≥ 45) especially for the larger datasets. errors of the results obtained by the VNSQP algorithm, INIQP and
Computational results are obtained by the proposed solution PSO proposed by Deng et al. [33]. According to the results, it is
method and presented in Tables 3–5 with tuned parameters in significantly seen that the proposed solution approach shows a
comparison with the other algorithms in the literature based on superior performance comparing others especially for the higher
12 M.A. Akbay, C.B. Kalayci and O. Polat / Knowledge-Based Systems 198 (2020) 105944

Fig. 15. Efficient frontiers for the NIKKEI dataset.

values of λ(e ≥ 45) while giving comparable results for the low the variable neighborhood search algorithm. After determining
values of λ(e < 45). When DAX 100 and S&P 100 datasets are the assets to be included in the optimal portfolio, the problem
considered, even INIQP generates better solutions for the much turns into a simple quadratic programming model that seeks
more risk sensitive portfolios. what ratios available capital should be distributed on the pre-
Results of the second performance comparison which involves selected assets. Since the exact solution of the obtained model
mean, median, minimum and maximum percentage error scores can be found by using exact solution method, asset weights were
of the algorithms proposed by Chang et al. [3], Deng et al. [33] and determined by utilizing quadratic programming. The proposed
Lwin and Qu [43], INIQP and VNSQP is presented in Table 4. To methodology was compared with other solution methods in the
make a fair performance comparison, each measure is calculated literature testing on five well-known datasets. Obtained results
considering 50 different values of λ parameter. According to the revealed that the proposed methodology is highly efficient.
obtained results, it is significantly seen that the proposed solu- Some of the future research directions and discussions are
tion approach outperformed others for almost all performance summarized below:
measures.
Finally, Table 5 compares the test results of the proposed
• Since it is always possible to improve on the best-found so-
lutions, developing much more efficient solution approaches
methodology with Cura [32], Baykasoglu et al. [47], Kalayci et al.
and algorithm components which will answer problem-
[38] and INIQP considering performance measures such as
specific challenges will be always valuable for future re-
MEUCD, VRE and MRE proposed by Cura [32]. Comparing to the
search directions.
best results announced by relevant researchers, the proposed
solution approach performed better in terms of convergence to
• Testing the algorithm on datasets belonging to different
markets is important to see how well it responds to market-
the unconstrained efficient frontier based on VRE especially for
specific variability.
Hang Seng, DAX 100, FTSE 100 and S&P 100 datasets while
showing comparable performance on the NIKKEI datasets.
• Although the mean–variance framework is one of the most
widely used approaches for modeling portfolio optimization
These results confirm that the proposed methodology, variable
problem, developing new structures considering different
neighborhood search algorithm with quadratic programming, is
risk measures such as; value at risk, conditional value at
highly efficient in solving the CCPO. Moreover, the methodology
risk, mean absolute deviation and developing successful so-
combining initial solution construction strategy with quadratic
lution methodologies for enhanced models with additional
programming is also quite competitive results when compared
constraints are still a challenge to be handled.
to others. It even outperformed some of the competitors for some
datasets. This shows that the initial solution construction strategy • The testing of the proposed methodology on the multiple
used in this study provides a considerable advantage for the investment horizons is also worthy of investigation.
proposed solution approach.

7. Conclusion and discussion

In this study, a two-stage solution methodology was devel-

oped to solve the portfolio optimization problem with cardinal- Declaration of competing interest
ity constraints hybridizing variable neighborhood search algo-
rithm enhanced with asynchronous parallelization strategy and
quadratic programming. In the first stage, the assets included in
the portfolio were selected by using the VNS algorithm. A sorting The authors declare that they have no known competing finan-
strategy according to the risk coverage rate based on the λ param- cial interests or personal relationships that could have appeared
eter was used to determine the initial solution and search pool for to influence the work reported in this paper.
 M.A. Akbay, C.B. Kalayci and O. Polat / Knowledge-Based Systems 198 (2020) 105944 13

Table 3
The results compared with Deng et al. [33] based on mean percentage error based on 50 different values of λ parameter.

Table 4
The results based on 50 different values of λ parameter.
14 M.A. Akbay, C.B. Kalayci and O. Polat / Knowledge-Based Systems 198 (2020) 105944

Table 5
The results based on 51 different values of λ parameter.

CRediT authorship contribution statement [9] O. Ertenlice, C.B. Kalayci, A survey of swarm intelligence for portfolio
optimization: Algorithms and applications, Swarm Evol. Comput. (2018).
Mehmet Anil Akbay: Conceptualization, Methodology, Visu- [10] J. Doering, R. Kizys, A.A. Juan, À. Fitó, O. Polat, Metaheuristics for rich
portfolio optimisation and risk management: Current state and future
alization, Writing - review & editing. Can B. Kalayci: Conceptu- trends, Oper. Res. Perspect. 6 (2019) 100121.
alization, Methodology, Visualization, Writing - review & edit- [11] D. Bertsimas, R. Shioda, Algorithm for cardinality-constrained quadratic
ing. Olcay Polat: Conceptualization, Methodology, Visualization, optimization, Comput. Optim. Appl. 43 (2009) 1–22.
Writing - review & editing. [12] J. Gao, D. Li, Optimal Cardinality constrained portfolio selection, Oper. Res.
61 (2013) 745–761.
[13] F. Cesarone, A. Scozzari, F. Tardella, A new method for mean–variance
Acknowledgment
portfolio optimization with cardinality constraints, Ann. Oper. Res. 205
(2013) 213–234.
This research is funded by the Scientific and Technological [14] F. Cesarone, A. Scozzari, F. Tardella, Linear vs. quadratic portfolio selection
Research Council of Turkey (TUBITAK) with the grant number models with hard real-world constraints, Comput. Manag. Sci. 12 (2015)
214M224. 345–370.
[15] D.X. Shaw, S.C. Liu, L. Kopman, Lagrangian relaxation procedure for
cardinality-constrained portfolio optimization, Optim. Methods Softw. 23
References
(2008) 411–420.
[16] W. Murray, H. Shek, A local relaxation method for the cardinality con-
[1] H. Markowitz, Portfolio selection, J. Finance 7 (1952) 77–91. strained portfolio optimization problem, Comput. Optim. Appl. 53 (2012)
[2] H.M. Markowitz, The early history of portfolio theory: 1600–1960, Financ.
681–709.
Anal. J. 55 (1999) 5–16.
[17] X.T. Cui, X.J. Zheng, S.S. Zhu, X.L. Sun, Convex relaxations and MIQCQP
[3] T.J. Chang, N. Meade, J.E. Beasley, Y.M. Sharaiha, Heuristics for cardi-
reformulations for a class of cardinality-constrained portfolio selection
nality constrained portfolio optimisation, Comput. Oper. Res. 27 (2000)
problems, J. Global Optim. 56 (2013) 1409–1423.
1271–1302.
[18] I. Bavarsad Salehpoor, S. Molla-Alizadeh-Zavardehi, A constrained portfolio
[4] C.B. Kalayci, O. Ertenlice, M.A. Akbay, A comprehensive review of deter-
selection model at considering risk-adjusted measure by using hybrid
ministic models and applications for mean–variance portfolio optimization,
meta-heuristic algorithms, Appl. Soft Comput. 75 (2019) 233–253.
Expert Syst. Appl. 125 (2019) 345–368.
[5] R.M. Karp, Reducibility among combinatorial problems, in: R.E. Miller, J.W. [19] A. Schaerf, Local search techniques for constrained portfolio selection
Thatcher, J.D. Bohlinger (Eds.), Complexity of Computer Computations: problems, Comput. Econ. 20 (2002) 177–190.
Proceedings of a Symposium on the Complexity of Computer Compu- [20] M. Woodside-Oriakhi, C. Lucas, J.E. Beasley, Heuristic algorithms for the
tations, March 20–22, 1972, IBM Thomas J. Watson Research Center, cardinality constrained efficient frontier, European J. Oper. Res. 213 (2011)
Yorktown Heights, New York, and sponsored by the Office of Naval 538–550.
Research, Mathematics Program, IBM World Trade Corporation, and the [21] A. Bačević, N. Vilimonović, I. Dabić, J. Petrović, D. Damnjanović, D.
IBM Research Mathematical Sciences Department. Publishing, 1972, pp. Džamić, Variable neighborhood search heuristic for nonconvex portfolio
85–103. optimization, Eng. Econ. 64 (2019) 254–274.
[6] P.N. Kolm, R. Tutuncu, F.J. Fabozzi, 60 Years of portfolio optimization: [22] T.J. Chang, S.C. Yang, K.J. Chang, Portfolio optimization problems in differ-
Practical challenges and current trends, European J. Oper. Res. 234 (2014) ent risk measures using genetic algorithm, Expert Syst. Appl. 36 (2009)
356–371. 10529–10537.
[7] K. Metaxiotis, K. Liagkouras, Multiobjective evolutionary algorithms for [23] H. Jalota, M. Thakur, Genetic algorithm designed for solving portfolio
portfolio management: A comprehensive literature review, Expert Syst. optimization problems subjected to cardinality constraint, Int. J. Syst.
Appl. 39 (2012) 11685–11698. Assur. Eng. Manag. 9 (2018) 294–305.
[8] B. Aouni, M. Doumpos, B. Pérez-Gladish, R.E. Steuer, On the increasing [24] S.J. Sadjadi, M. Gharakhani, E. Safari, Robust optimization framework for
importance of multiple criteria decision aid methods for portfolio selection, cardinality constrained portfolio problem, Appl. Soft Comput. 12 (2012)
J. Oper. Res. Soc. (2018) 1–18. 91–99.
 M.A. Akbay, C.B. Kalayci and O. Polat / Knowledge-Based Systems 198 (2020) 105944 15

[25] H. Soleimani, H.R. Golmakani, M.H. Salimi, Markowitz-based portfolio [48] L.D. Gaspero, G.D. Tollo, A. Roli, A. Schaerf, Hybrid metaheuristics
selection with minimum transaction lots, cardinality constraints and re- for constrained portfolio selection problems, Quant. Finance 11 (2011)
garding sector capitalization using genetic algorithm, Expert Syst. Appl. 36 1473–1487.
(2009) 5058–5063. [49] Z. Zhou, X. Liu, H. Xiao, S. Wu, Y. Liu, A DEA-based MOEA/D algorithm for
[26] K.P. Anagnostopoulos, G. Mamanis, A portfolio optimization model with portfolio optimization, Cluster Comput. (2018).
three objectives and discrete variables, Comput. Oper. Res. 37 (2010) [50] G.A.V. Pai, T. Michel, Evolutionary optimization of constrained k-means
1285–1297. clustered assets for diversification in small portfolios, IEEE Trans. Evol.
[27] K.P. Anagnostopoulos, G. Mamanis, The mean–variance cardinality con- Comput. 13 (2009) 1030–1053.
strained portfolio optimization problem: An experimental evaluation of [51] H.H. Zhu, Y. Wang, K.S. Wang, Y. Chen, Particle Swarm Optimization (PSO)
five multiobjective evolutionary algorithms, Expert Syst. Appl. 38 (2011) for the constrained portfolio optimization problem, Expert Syst. Appl. 38
14208–14217. (2011) 10161–10169.
[28] K. Liagkouras, K. Metaxiotis, A new Probe Guided Mutation operator and [52] A.R. Pouya, M. Solimanpur, M.J. Rezaee, Solving multi-objective portfolio
its application for solving the cardinality constrained portfolio optimization optimization problem using invasive weed optimization, Swarm Evol.
problem, Expert Syst. Appl. 41 (2014) 6274–6290. Comput. 28 (2016) 42–57.
[29] R. Saborido, A.B. Ruiz, J.D. Bermudez, E. Vercher, M. Luque, Evolutionary [53] N. Mladenovic, P. Hansen, Variable neighborhood search, Comput. Oper.
multi-objective optimization algorithms for fuzzy portfolio selection, Appl. Res. 24 (1997) 1097–1100.
Soft Comput. 39 (2016) 48–63. [54] P. Hansen, N. Mladenovic, Variable neighborhood search: Principles and
[30] J. Branke, B. Scheckenbach, M. Stein, K. Deb, H. Schmeck, Portfolio opti- applications, European J. Oper. Res. 130 (2001) 449–467.
mization with an envelope-based multi-objective evolutionary algorithm, [55] P. Hansen, N. Mladenović, D. Perez-Britos, Variable neighborhood
European J. Oper. Res. 199 (2009) 684–693. decomposition search, J. Heuristics 7 (2001) 335–350.
[31] K. Liagkouras, A new three-dimensional encoding multiobjective evolu- [56] P. Hansen, N. Mladenović, J-Means: a new local search heuristic for
tionary algorithm with application to the portfolio optimization problem, minimum sum of squares clustering, Pattern Recognit. 34 (2001) 405–413.
Knowl.-Based Syst. 163 (2019) 186–203. [57] S.A. Canuto, M.G.C. Resende, C.C. Ribeiro, Local search with perturbations
[32] T. Cura, Particle swarm optimization approach to portfolio optimization, for the prize-collecting steiner tree problem in graphs, Networks 38 (2001)
Nonlinear Anal. RWA 10 (2009) 2396–2406. 50–58.
[33] G.F. Deng, W.T. Lin, C.C. Lo, Markowitz-based portfolio selection with [58] J. Puchinger, G.R. Raidl, U. Pferschy, The Core Concept for the Multi-
cardinality constraints using improved particle swarm optimization, Expert dimensional Knapsack Problem. Publishing, Berlin, Heidelberg, 2006, pp.
Syst. Appl. 39 (2012) 4558–4566. 195–208.
[34] H.R. Golmakani, M. Fazel, Constrained portfolio selection using particle [59] P. Côté, T. Wong, R. Sabourin, A Hybrid Multi-objective Evolutionary
swarm optimization, Expert Syst. Appl. 38 (2011) 8327–8335. Algorithm for the Uncapacitated Exam Proximity Problem. Publishing,
[35] Q.J. Ni, X.S. Yin, K.W. Tian, Y.Q. Zhai, Particle swarm optimization with Berlin, Heidelberg, 2005, pp. 294–312.
dynamic random population topology strategies for a generalized portfolio [60] P. Hansen, N. Mladenović, An introduction to variable neighborhood
selection problem, Nat. Comput. 16 (2017) 31–44. search, in: S. Voß, S. Martello, I.H. Osman & C. Roucairol (Eds.),
[36] Y.L.T.V. Silva, A.B. Herthel, A. Subramanian, A multi-objective evolutionary Meta-Heuristics: Advances and Trends in Local Search Paradigms for
algorithm for a class of mean–variance portfolio selection problems, Expert Optimization. Publishing, Boston, MA, 1999, pp. 433–458.
Syst. Appl. 133 (2019) 225–241. [61] P. Hansen, N. Mladenović, D. Urošević, Variable neighborhood search and
[37] W. Chen, Artificial bee colony algorithm for constrained possibilistic local branching, Comput. Oper. Res. 33 (2006) 3034–3045.
portfolio optimization problem, Physica A 429 (2015) 125–139. [62] C.B. Kalayci, C. Kaya, An ant colony system empowered variable neighbor-
[38] C.B. Kalayci, O. Ertenlice, H. Akyer, H. Aygoren, An artificial bee colony hood search algorithm for the vehicle routing problem with simultaneous
algorithm with feasibility enforcement and infeasibility toleration proce- pickup and delivery, Expert Syst. Appl. 66 (2016) 163–175.
dures for cardinality constrained portfolio optimization, Expert Syst. Appl. [63] O. Polat, C.B. Kalayci, O. Kulak, H.O. Gunther, A perturbation based variable
85 (2017) 61–75. neighborhood search heuristic for solving the Vehicle Routing Problem
[39] S.S. Bastiani, L. Cruz-Reyes, E. Fernandez, C. Gomez, Portfolio optimization with Simultaneous Pickup and Delivery with Time Limit, European J. Oper.
from a set of preference ordered projects using an ant colony based Res. 242 (2015) 369–382.
multi-objective approach, Int. J. Comput. Intell. Syst. 8 (2015) 41–53. [64] C.B. Kalayci, O. Polat, S.M. Gupta, A variable neighbourhood search al-
[40] C.B. Kalayci, O. Polat, M.A. Akbay, An efficient hybrid metaheuristic al- gorithm for disassembly lines, J. Manufact. Technol. Manag. 26 (2015)
gorithm for cardinality constrained portfolio optimization, Swarm Evol. 182–194.
Comput. 54 (2020) 100662. [65] C.B. Kalayci, O. Polat, S.M. Gupta, A hybrid genetic algorithm for sequence-
[41] S.K. Mishra, G. Panda, R. Majhi, Constrained portfolio asset selection using dependent disassembly line balancing problem, Ann. Oper. Res. 242 (2016)
multiobjective bacteria foraging optimization, Oper. Res. Int. J. 14 (2014) 321–354.
113–145. [66] O. Polat, C.B. Kalayci, Ö. Mutlu, S.M. Gupta, A two-phase variable neigh-
[42] B.L. Chen, Y.B. Lin, W.H. Zeng, H. Xu, D.F. Zhang, The mean–variance bourhood search algorithm for assembly line worker assignment and
cardinality constrained portfolio optimization problem using a local balancing problem type-II: an industrial case study, Int. J. Prod. Res. 54
search-based multi-objective evolutionary algorithm, Appl. Intell. 47 (2017) (2015) 722–741.
505–525. [67] P. Hansen, N. Mladenović, J.A. Moreno Pérez, Variable neighbourhood
[43] K. Lwin, R. Qu, A hybrid algorithm for constrained portfolio selection search: methods and applications, Ann. Oper. Res. 175 (2009) 367–407.
problems, Appl. Intell. 39 (2013) 251–266. [68] F. García-López, B. Melián-Batista, J.A. Moreno-Pérez, J.M. Moreno-Vega,
[44] D. Maringer, H. Kellerer, Optimization of cardinality constrained portfolios The parallel variable neighborhood search for the p-Median problem, J.
with a hybrid local search algorithm, OR Spectrum 25 (2003) 481–495. Heuristics 8 (2002) 375–388.
[45] A.N. Sadigh, H. Mokhtari, M. Iranpoor, S.M.T.F. Ghomi, Cardinality con- [69] T.G. Crainic, M. Gendreau, P. Hansen, N. Mladenović, Cooperative parallel
strained portfolio optimization using a hybrid approach based on particle variable neighborhood search for the p-Median, J. Heuristics 10 (2004)
swarm optimization and hopfield neural network, Adv. Sci. Lett. 17 (2012) 293–314.
11–20. [70] P. Hansen, N. Mladenović, First vs. best improvement: An empirical study,
[46] M. Tuba, N. Bacanin, Artificial bee colony algorithm hybridized with firefly Discrete Appl. Math. 154 (2006) 802–817.
algorithm for Cardinality constrained mean-variance portfolio selection [71] T.G. Crainic, M. Toulouse, M. Gendreau, Toward a taxonomy of parallel
problem, Appl. Math. Inf. Sci. 8 (2014) 2831–2844. tabu search heuristics, INFORMS J. Comput. 9 (1997) 61–72.
[47] A. Baykasoglu, M.G. Yunusoglu, F.B. Ozsoydan, A GRASP based solution [72] T.G. Crainic, N. Hail, Parallel metaheuristics applications in: E. Alba (Ed.),
approach to solve cardinality constrained portfolio optimization problems, Parallel Metaheuristics: A New Class of Algorithms. Publishing, Hoboken,
Comput. Ind. Eng. 90 (2015) 339–351. New Jersey, 2005.
[73] J.E. Beasley, Or-library - Distributing test problems by electronic mail, J.
Oper. Res. Soc. 41 (1990) 1069–1072.