Quantum-Inspired and Hybrid Deep Reinforcement Learning Approaches

for Multi-Period Portfolio Optimization and Intelligent Asset Management

i
TABLE OF CONTENTS

ii
 Title Page

Declaration

Certificate

Acknowledgment

Table of Contents

List of Tables

List of Figures

Abstract

1. INTRODUCTION 1

2. MOTIVATION OF RESEARCH 2

3. OBJECTIVES 4

4. DESIGN PHASES 3

5. SCOPE OF RESEARCH 5

6. REVIEW OF LITERATURE 5

7. FINDINGS AND SUMMARY FROM LITERATURE 8

8. PROPOSED METHODOLOGY I: A CONSTRAINED MULTI-PERIOD 9

PORTFOLIO OPTIMIZATION MODEL BASED ON QUANTUM-INSPIRED

OPTIMIZATION

8.1 Functional Block of Framework 9

8.1.1 Pre-processing 10

8.2 Proposed methodology 11

8.2.1 Multi-constraint modelling 11

8.2.2 Objective function 12

8.2.3 Quantum whale optimization for MPO 12

8.3 Results and discussion 16

iii
 8.3.1 Dataset description 16

8.3.2 Performance indicators 17

8.3.3 Empirical analysis 17

8.3.4 Statistical test analysis 26

8.3.5 Discussion 27

9. PROPOSED METHODOLOGY II: AN INTELLIGENT PORTFOLIO 29

MANAGEMENT SCHEME BASED ON HYBRID DEEP REINFORCEMENT

LEARNING AND CUMULATIVE PROSPECTIVE APPROACH

9.1 Proposed Method 29

9.1.1 Functional blocks of the framework 29

9.2 Results 32

9.2.1 Discussion 40

41
10. LIMITATIONS AND DIRECTIONS FOR FUTURE RESEARCH

11.GOALS ACHIEVED BY RESEARCH/CONCLUSION 42

REFERENCES 44

LIST OF PUBLICATION

iv
 LIST OF TABLES

Table 1 Mathematical formulation of varied evaluation indicators 17

Table 2 Sharpe ratio comparison for the proposed and existing algorithms 19

( u i=0 . 05 )
Table 3 Sortino ratio comparison of proposed and existing algorithms 19

( u i=0 . 05 )
Table 4 STARR ratio comparison for the proposed and existing algorithms 20

( u i=0 . 05 )
Table 5 Information ratio comparison for the proposed and existing 21

algorithms ( i
u =0 . 05 )
Table 6 Shannon entropy comparison of both proposed and existing 21

algorithms ( i
u =0 . 05 )

Table 7 Downside deviation comparison for the proposed and existing 22

algorithms ( i
u =0 . 05 )

Table 8 Statistical analysis results 26

Table 9 Quantitative analysis of the proposed and existing PO schemes 27

Table 10 Performance analysis of Proposed SPPO model 33

Table 11 Evaluation of portfolio performance 37

Table 12 Performance comparison of Sharpe ratio 38

Table 13 Performance comparison of Sortino ratio 39

Table 14 Performance comparison of Information ratio 39

Table 15 Performance of proposed method 40

LIST OF FIGURES

Figure 1 Block diagram of proposed MPO model 10

v
Figure 2 The overall wealth secured by the portfolios based on period 18

Figure 3 Comparison analysis in terms of convergence for both proposed and 18

existing algorithms
Figure 4 Average returns of the optimal portfolio selected by the algorithms 24
u =0 . 05 (b) ui =0 . 07 and (c) ui =0 .10
for (a) i
Figure 5 Comparison of net return rates of the proposed and existing schemes 25
u =0 . 05 (b) ui =0 . 07 and (c) ui =0 .10
for (a) i
Figure 6 Proposed architecture 30

Figure 7 Actual Vs predicted price values 34

Figure 8 Error metrics Vs Epoch size 34

Figure 9 Comparison of Error metrics (a) MAE) (b) RMSE (c) MSE (d) 35
MAPE
Figure 10 Convergence analysis 36

Figure 11 Evaluation of portfolio performance 37

Figure 12 Time complexity analysis 38

ABSTRACT
In the context of portfolio optimization, selecting optimal portfolios that maximize returns
while minimizing risk is a critical challenge for investors, particularly in the dynamic and

vi
complex stock market environment. This study introduces two advanced models to address
this challenge. First, a multi-constraint multi-period portfolio optimization (MPO) model is
proposed, which incorporates various market risk factors along with higher-order moments
(HOM) such as skewness, kurtosis, transaction costs, diversification, and budget constraints.
To solve this multi-constrained problem, a novel quantum-inspired whale optimization
algorithm (QWOA) is developed, leveraging quantum entanglement to overcome the slow
convergence issues of traditional whale optimization. The empirical evaluation demonstrates
that the proposed QWOA model outperforms existing algorithms such as WOA, GWO, FOA,
PSO, and FA in optimizing portfolio performance.

Second, the study integrates stock prediction with portfolio optimization through the Stock
Prediction and Portfolio Optimization (SPPO) model. Initially, stock prices are predicted
using a hybrid deep reinforcement learning (DRL) model, where a Gated Recurrent Unit
Network (GRUN) simulates agent-environment interactions. The optimal actions in the
prediction process are determined using a Quantum Differential Evolution Algorithm (Q-
DEA). Subsequently, the Cumulative Prospect Theory (CPT) model is employed to select the
optimal portfolio based on the predicted best assets. This comprehensive approach not only
enhances portfolio returns but also addresses the complexity of asset selection in a volatile
market, providing significant advantages over traditional methods. The next work can involve
the exploration of other quantum-inspired algorithms and the integration of real-time data for
more adaptive portfolio optimization strategies.

vii
1. INTRODUCTION

Portfolio management had become a critical focus in the financial environment, aiming to
select and maintain a set of investments that aligned with the long-term financial goals and
risk tolerance of companies, clients, or institutions. Over the past few decades, there had been
a significant increase in the use of optimization tools and financial modeling within the equity
portfolio landscape [1]. Initially, many organizations started with a single product or service;
however, as they experienced success and identified new opportunities, they expanded their
offerings, creating a more diverse portfolio [2]. The concept of a portfolio referred to a
collection of various assets owned by an individual or organization to meet financial
requirements. Portfolio analysis was crucial for evaluating an organization's strengths and
limitations, which, in turn, guided strategic adjustments to optimize overall performance [3].
The analysis often dealt with complex systems and encountered risks due to undefined
potential choices [4].

Portfolio optimization (PO) was centered on finding the optimal combination of securities
and proportions to maximize returns while minimizing risks. This process involved assigning
financial assets in a manner that balanced the expected outcome and associated risks [5].
However, the lack of constraints, such as cardinality constraints (CC), boundary constraints
(BC), traction loss (TL), and traction cost (TC), posed significant challenges to the PO
models [6]. Multi-objective optimization frameworks were often employed to address these
issues, as they could minimize conflicting objectives, particularly the balance between return
and risk [7]. Quantum annealing (QA) emerged as a promising approach to solving
computational problems, enhancing performance, and reducing errors through quantum
dynamic tuning [8].

Historically, strategic asset allocation (SSA) had been determined using single-period
optimization, particularly the mean-variance framework introduced by Markowitz. However,
the Merton portfolio optimization (MPO) approach, which considered stochastic volatility
through asymptotic approximation, had provided more sophisticated insights by accounting
for time-scale fluctuations in volatility [9]. More recently, machine learning techniques, such
as regularization and cross-validation, had been applied to optimize portfolios, offering new
methods to address overfitting and improve predictive accuracy [10].

The challenge of multi-period portfolio optimization (MPO) had drawn significant attention
within the research community, as it required careful consideration of market fluctuations and

1
risk factors over time. Traditional algorithms often faced limitations, such as over fitting,
falling into multiple local optima, errors in parameter estimation, and reduced computational
efficiency [11]. The whale optimization algorithm (WOA) had shown promise in addressing
these issues by solving PO problems with higher-order moments (HOM). However, its
application had been limited to single-period PO, necessitating the development of multi-
period models that could account for varying market risk factors [12]. To overcome the slow
convergence problem inherent in WOA, researchers had introduced quantum-inspired whale
optimization algorithms (QWOA), which improved the exploration phase and identified
Pareto optimal solutions more efficiently [13].

In portfolio management, the ultimate goal was to achieve the best possible outcome by
strategically planning, evaluating performance, and maintaining policies [14]. The mean-
variance (MV) model had been widely used, though it had its limitations, particularly
concerning practical risk factors and the constraints faced by real-world investors, such as
portfolio size and trading restrictions [15]. In response to these challenges, researchers had
explored mathematical programming techniques and developed various optimization
algorithms to enhance portfolio selection and optimization processes [16]. These algorithms,
including heuristic methods like Genetic Algorithm (GA), Particle Swarm Optimization
(PSO), Ant Colony Optimization (ACO), Artificial Bee Colony (ABC) algorithm, and
Harmony Search Algorithm (HSA), had provided efficient solutions to complex portfolio
problems by identifying portfolios with the highest potential returns and the lowest risk of
loss [17].

Furthermore, machine learning had emerged as a powerful tool in portfolio optimization,

offering data-driven approaches that could analyze large volumes of financial data to predict
price trends and movements. This had led to improved performance in portfolio optimization
by enabling more accurate and timely decision-making [18]. As the financial landscape
continued to evolve, the integration of these advanced optimization algorithms and machine
learning techniques represented a promising direction for future research and application in
portfolio management.

2. MOTIVATION OF RESEARCH

(1) In order to increase convergence rates and accuracy, the paper presents a quantum-
inspired whale optimization algorithm (QWOA) for multi-period portfolio optimization.
It takes into account a number of restrictions, including risk factors, transaction costs,

2
 diversification, boundary, and budget. This strategy seeks to improve the optimization
process overall by maximizing profits while lowering risks.
(2) To Optimizes portfolios and increases stock prediction accuracy by combining
Cumulative Prospect Theory (CPT) with Deep Reinforcement Learning (DRL). The
Quantum Differential Evolution Algorithm (Q-DEA) determines the best course of action,
whereas the hybrid DRL model uses a Gated Recurrent Unit Network (GRUN) to mimic
agent-environment interactions. By precisely forecasting stock prices and choosing the
finest assets, the CPT technique improves portfolio management decision-making.
(3) Explores the use of quantum computing techniques to improve portfolio management
using Markowitz Mean-Variance Optimization. It uses quantum algorithms like the
Quantum Approximate Optimization Algorithm and Quantum Autoencoder to efficiently
process and analyze financial data, thereby enhancing the accuracy and efficiency of
portfolio optimization, thereby improving asset selection and risk management.

3. OBJECTIVES

 To achieve the highest net return rates, an MPO model must be developed and
subjected to several real-time and higher-order restrictions.
 To achieve more precision and efficiency, the restricted MPO problem can be solved
by introducing the idea of quantum entanglement.
 Conducting empirical analysis on the model that has been described will verify how
well it performs in addressing multi-constrained optimization issues when compared
to other current approaches.
 The goal is to accurately identify the best assets for the portfolio and achieve future
profits by integrating the asset pre-selection model into the portfolio optimization
model.
 In order to accurately forecast the stock prices for the upcoming time, a hybrid DRL-
based prediction model will be presented in the suggested framework.
 To include the price prediction model into a Markov decision process, whereby the
GRUN is used to mimic the interactions between the agent and environment.
 Using the Q-DEA technique, choose the best course of action for the interactions in
order to increase forecast accuracy.

4. DESIGN PHASES

I. Multi-Period Portfolio Optimization Using Quantum-Inspired Optimization

3
Phase 1: To develop a multi-period portfolio optimization (MPO) model incorporating
multiple real-time and higher-order constraints, ensuring maximum net return rates.

Phase 2: To introduce the concept of quantum entanglement in the whale optimization

algorithm (QWOA) to enhance accuracy and efficiency in solving constrained MPO
problems.

Phase 3: To conduct empirical evaluations of the proposed model and compare its
performance with existing optimization methods, such as WOA, GWO, FOA, PSO, and FA,
in optimizing constrained MPO.

II. Intelligent Portfolio Management Using Hybrid Deep Reinforcement Learning

Phase 4: To integrate an asset pre-selection model into the portfolio optimization framework
for accurate selection of optimal assets, ensuring higher profitability.

Phase 5: To implement a hybrid deep reinforcement learning (DRL)-based stock price

prediction model using the Gated Recurrent Unit Network (GRUN) and optimize prediction
accuracy using the Quantum Differential Evolution Algorithm (Q-DEA).

Phase 6: To apply the Cumulative Prospect Theory (CPT) model for selecting optimal
portfolios based on predicted stock prices, ensuring improved decision-making in asset
selection.

III. Portfolio Management Using Quantum Computing and Markowitz Mean-Variance

Optimization

Phase 7: To preprocess stock data using forward fill techniques and normalization to improve
training and testing feature sets.

Phase 8: To employ the Quantum Approximate Optimization Algorithm (QAOA) for

obtaining a binary solution vector representing selected features in the portfolio selection
process.

Phase 9: To implement a Quantum Autoencoder (QAE) model for learning a low-dimensional

representation of input features and train a Variational Quantum Regressor (VQR) for target
value prediction.

Phase 10: To perform portfolio selection using the Markowitz Mean-Variance Optimization
Model based on the predicted optimal assets, ensuring effective risk-return trade-offs.

4
5. SCOPE OF RESEARCH
1. To maximize net return rates and improve asset distribution, create a sophisticated
MPO model that incorporates several real-time and higher-order restrictions.
2. To improve the accuracy of asset selection and investment decision-making, combine
DRL stock prediction with CPT portfolio optimization.
3. By using QWOA, sluggish convergence difficulties in conventional optimization
models may be resolved, increasing the effectiveness of MPO problem solving.
4. To increase prediction accuracy and guarantee better asset selection in dynamic
market situations, use GRUN to simulate stock market behavior.
5. For effective portfolio selection that lowers computing complexity and enhances
decision-making, use QAOA, QAE, and VQR.
6. To confirm the suggested models' efficacy in portfolio management, compare them to
current optimization methods.

6. REVIEW OF LITERATURE

The literature introduces a number of strategies to address PO-related issues. Below is a

summary of several recent methods based on this principle:

Hossein Babazadeh and Akbar Esfahanipour [19] had created, a novel portfolio optimization
model was developed using Value at Risk (VaR) as a risk measure to account for extreme
risk, with VaR estimated through Extreme Value Theory (EVT). To enhance practicality, real
trading constraints such as cardinality, budget, floor, and ceiling constraints were considered,
resulting in a non-convex NP-hard problem. To solve this, a new design of the Non-
dominated Sorting Genetic Algorithm (NSGA-II) was proposed. The model's performance
was evaluated against three VaR estimation methods—historical simulation, GARCH, and t-
student GARCH—using S&P 100 index data. Experimental results showed that the enhanced
NSGA-II effectively solved the mean-VaR portfolio optimization problem with better
performance and faster solving time than the original NSGA-II. It also outperformed other
Non-dominated-based algorithms like SPEA-II, NSPSO, and NSACO, particularly in the
low-risk area of the Pareto front.

Q. H. Zhai et al. [20] had created , a multiconstraint portfolio optimization model was
developed that balances investment returns with fluctuations by incorporating capital asset
5
pricing model, arbitrage pricing theory, and the Fama–French three-factor model to quantify
the price of individual stocks and portfolios. The model was constructed based on the second-
order stochastic dominance rule, higher moments of return series, Shannon entropy, and
practical investment constraints. The whale optimization algorithm (WOA) was proposed to
optimize this multiconstraint model using FTSE100 index data, resulting in a significant
improvement in the rate of return compared to the simple diversified buy-and-hold strategy or
the FTSE100 index itself. Extensive experiments confirmed that WOA outperformed other
swarm intelligence optimization algorithms, such as gray wolf optimizer, fruit fly
optimization algorithm, particle swarm optimization, and firefly algorithm, across various
performance indicators, especially in environments with harsh constraints. Future work can
involve exploring other advanced optimization techniques and extending the model to
account for more complex market dynamics and further constraints.

Behera et al. [21] Had proposed two modified firefly algorithms, namely the crazy firefly
algorithm and the variable step size firefly algorithm, which were individually hybridized
with the standard particle swarm optimization (PSO) algorithm to improve the efficiency of
data clustering. Clustering, which involves grouping similar data based on certain
characteristics, benefited from the integration of these nature-inspired optimization
techniques. The proposed hybrid algorithms were evaluated using ten UCI Machine Learning
Repository datasets and eight Shape sets, with performance assessed using two clustering
validity measures: Compact-separated and David–Bouldin. The experimental results
demonstrated that both hybrid algorithms outperformed the existing hybridized firefly
particle swarm optimization algorithm, showcasing their enhanced ability in clustering tasks.
Future work can involve applying these hybrid algorithms to more complex datasets and
exploring further optimizations to improve clustering accuracy and efficiency across diverse
domains.

Tamara Teplova's [22] had studied on portfolio optimization used data from 10 international
ETFs from 2012 to 2022. She extended the Black-Litterman (BL) approach by incorporating
ARMA-GARCH-copula-based expected returns and the Conditional Value at Risk (CVaR)
metric. The BL approach, known for its Bayesian methodology, combines equilibrium returns
with investor views to produce refined expected returns. Teplova used the Regular Vine
copula for its flexibility in multivariate dependency modeling. Performance was compared
against Mean-Variance optimization and an equal weights portfolio. Key metrics like tail
risks, maximum drawdown, turnover, and break-even points were reported. The results

6
showed that the copula-BL portfolio outperformed in terms of lower tail risk and higher risk-
adjusted returns. Future work should focus on further refining copula models and exploring
the impact of additional constraints and market conditions.

Yaman et al. [23] concentrated on a hybrid strategy to solve the cardinality constraint
portfolio optimization model, which is based on the nonlinear neural network (NNN) and the
genetic algorithm. An essential model in finance theory is the portfolio optimization model,
which uses NNN to draw an efficient frontiers graph related to the portfolio selection issue.
The model was created to take on the difficulty of NP hard tasks. The Istanbul Stock
Exchange (ISE-30) dataset was utilized in the study to examine the efficacy of the suggested
methodology. The information ratio, Sharpe ratio, sortino ratio, and average return were the
model's performance measures. The suggested approach outperforms the current model when
the results are compared to the traditional model (Active set technique).

Sharma et al. [24] investigated the subject of portfolio optimization and provided an efficient
portfolio prediction and selection model based on a multi-serial cascaded network (MCNet)
with a hybrid meta-heuristic algorithm. The dataset was gathered from the portfolios of
several companies. MCNet, which is used in conjunction with autoencoders, 1D convolution
neural networks (1DCNN), and recurrent neural networks (RNNs), forecasts the firms'
advantages. Following the benefits prediction, the artificial rabbit and hummingbird
algorithm (IARHA) is integrated to pick the maximum profit. The implementation was
carried out using the Python platform. The constructed model performs better in experiments;
the final RMSE and MAE measure outcomes are 89% and 56%, respectively.

Wang et al. [25] introduced a deep neural network (DNN) based prediction-based portfolio
optimization algorithm. Three models are used by DNN for prediction: convolution neural
network (CNN), long short memory (LSTM) neural network, and deep multi-layer perceptron
(DMLP). Firstly stocks future returns were projected by applying DNN. Next, each stock's
risk was calculated using the DNN's prediction mistakes. The portfolio optimization model
was constructed by integrating the semi-absolute deviation of the prediction error and the
predicted returns. Component stocks of the China Securities 100 index on the Chinese stock
market were utilized as experimental data by the author for this work. The experiment
demonstrates that, under varying desired returns, the DMLP+MSAD models outperformed
other models, with DMLP having the lowest predicted errors.

7
Kizys et al. [26] created a model for portfolio optimization using the simheuristic algorithm
(SA) with random returns and noisy covariance. Using Monte Carlo simulation and a variable
neighbourhood search metaheuristic, SA addresses noisy covariance and stochastic returns
that are modeled as random variables. The paper's technique aims to solve real-sized
stochastic examples in a comparatively short amount of time. Three experiment results—
stochastic covariance, stochastic covariance and returns, and extra instances—are examined
for result analysis.

Asgari et al. [27] introduced a stochastic programming model with a harmony search method
for portfolio selection from a multi-objective stock market. The model's primary goal is to
offer a practical tool and a framework that tackles certain flaws. For numerical
demonstrations, the model uses actual stock data from the Iranian stock market. The
superiority of the suggested model has been demonstrated in terms of the algorithm's runtime
and the quality of the provided solution. The outcomes demonstrate that the suggested meta-
heuristic algorithm outperforms a genetic algorithm. A 90%, 99%, and 99.9% confidence
level has been reached by the suggested model for small, medium, and large size examples,
respectively.

Wu et al. [28] had developed a Portfolio Management System (PMS) using reinforcement
learning and two neural networks, Convolutional Neural Networks (CNN) and Recurrent
Neural Networks (RNN), to forecast stock prices. The system uses a novel reward function
based on the Sharpe ratio, which significantly outperforms traditional methods. The PMS
achieved an average return increase of 39.0% and a drawdown reduction of 13.7% compared
to models using a trading return reward function. The CNN model was more effective for
constructing reinforcement learning portfolios but presented 1.98 times more drawdown risk
than the RNN model. The system demonstrated profitability, reduced investment risk, and
enhanced resource allocation across various datasets. Future work could involve refining the
reinforcement learning models and testing the PMS on broader datasets to optimize risk
management and investment outcomes.

7. FINDINGS AND SUMMARY FROM LITERATURE

(1) A novel portfolio optimization model was developed using Value at Risk (VaR) as a risk
measure, enhancing the ability to account for extreme risk. However, the model's non-
convex NP-hard nature requires computationally expensive optimization methods, and its
efficiency could be improved further.

8
(2) A multiconstraint portfolio optimization model was proposed using several financial
theories and the Whale Optimization Algorithm (WOA). The model outperformed simpler
strategies, but the complexity of real-world market dynamics and additional constraints
could be better addressed through more advanced optimization techniques.
(3) Modified firefly algorithms, when hybridized with particle swarm optimization, improved
clustering efficiency. However, the algorithms could benefit from further optimization to
handle more complex datasets and achieve better clustering accuracy across diverse
domains.
(4) Portfolio optimization using the Black-Litterman approach and ARMA-GARCH-copula
models showed superior performance over traditional methods. Future improvements
could refine copula models and incorporate more constraints to address varying market
conditions.
(5) A hybrid strategy combining neural networks with genetic algorithms addressed the
cardinality constraint in portfolio optimization. While it showed promising results, the
model faced challenges in scalability and adapting to different market conditions.
(6) A multi-serial cascaded network (MCNet) based portfolio prediction model with hybrid
meta-heuristic algorithms demonstrated better performance in asset selection. Despite its
success, future research could explore more robust prediction models to reduce prediction
errors further.
(7) The deep neural network (DNN)-based portfolio optimization method using CNN,
LSTM, and DMLP models outperformed traditional models. However, improvements are
needed in handling market volatility and reducing prediction errors.
(8) A portfolio optimization model using the simheuristic algorithm addressed stochastic
returns and noisy covariance. The method worked well on smaller datasets, but challenges
in scaling the approach for larger, more complex datasets remain.
(9) A stochastic programming model with harmony search outperformed genetic algorithms
in portfolio selection, but its efficiency could be improved, especially in handling larger
market datasets and further refining runtime.
(10) A portfolio management system (PMS) using reinforcement learning and CNN-RNN
models achieved promising returns and reduced risk. However, the model requires further
refinement, especially in risk management strategies and adaptability to broader datasets
for optimization.

9
8. PROPOSED METHODOLOGY I: A CONSTRAINED MULTI-PERIOD
PORTFOLIO OPTIMIZATION MODEL BASED ON QUANTUM-INSPIRED
OPTIMIZATION

8.1 FUNCTIONAL BLOCK OF FRAMEWORK

Many strategies have been employed in the research that have already been conducted to
produce optimum portfolios; still, in order to lessen the impact of risks, it is necessary to pay
attention to a number of market risk indicators. In order to address the optimization problems,
the suggested model aims to take the higher-order restrictions into account as risk factors. A
QWOA is presented in this research to solve the multi-constrained optimization issue. In
order to solve the problem, the classical algorithm incorporates the behavior of quantum
entanglement. In addition to lowering transaction costs, the suggested method took into
account additional aspects including kurtosis and skewness as potential investment hazards.
Using the suggested approach, portfolio restrictions are lessened by maximizing the
aforementioned components. Figure 1 displays the planned MPO system's block diagram.

10
 Figure 1: Block diagram of proposed MPO model
First, the restrictions are established, and the optimization goal is specified as selecting the
best portfolio from the database. Subsequently, the suggested algorithm is run through several
iterations in order to determine the ideal portfolio.

8.1.1 Pre-processing

The Kenneth French data library is where the dataset used for PO was obtained. First, a data
mining approach is used to pre-process the investment portfolio dataset. Upon close
inspection, the dataset has no duplicate values, and the distribution of the data is uniform. To
address the missing data, the suggested model makes use of the missing value imputation
approach. Due to a lack of data throughout time and sophisticated measurement equipment,
the dataset contains some missing values. To improve the quality of inputs, a pre-processing
step is carried out in order to fill the available missing data. The mean value computation is
used to replace the missing value.
P
∑i −0 Z i
Mean ( Z )=
P (1)

where
Zi denotes the data displayed in a row and P refers to the total amount of data in the
dataset. To find any missing or crooked rows, the dataset is first analyzed row by row. The
missing data are then eliminated by looking at it column by column. To lessen the shortage,
null values are inserted into the spaces left by missing data. Following pre-processing,
optimization is applied to the dataset.

8.2 Proposed methodology

8.2.1 Multi-constraint modelling
Many restrictions must be taken into account for real-world settings, making the selection of
optimum portfolios with maximum returns a difficult challenge. In addition to the market risk
assessment, there are other limits that impact the assets in the portfolios. The MPO is the
focus of this study. As a result, once the intervals are examined, the optimization constraints
are established.

The portfolio's asset allocation should be risk-free and optimal for the investors to execute
appropriate investment plans. The investors are involved in the decision-making process. The

11
stock market fluctuates throughout time, therefore it's critical to develop an ideal portfolio
selection technique so that investors can comprehend these swings and adjust the asset
allocation appropriately. By breaking down the optimization issue according to the time
periods, it is possible to comprehend the fluctuations in the market risks. Between the bottom
and upper boundaries of the time interval, variations and risks are taken into account.

The following is the expression for the asset returns for a timet :
~
Ψ t , i =[ Ψ t , i , Ψ t , i ] ; 0≤Ψ t , i ≤Ψ t , i (2)

where
Ψ t , i denotes the upper bound's return value Ψ t , i W" the lower bound's return

value. For a given time t , the investment ratio vector is represented by the symbol
I t =( I t , 1 , I t , 2 , .....I t , n ) I ≥0
. If t , i , it means that no short selling constraint is applied.
8.2.2 Objective function
QWOA aims to maximize returns while adhering to many higher order restrictions. Thus, the
objective function is defined as follows:

[ ]
t n n
max Ν t =Ν ( t−1 ) ∏ ∑ It, i~
Ψ i, t −∑ κ (t , i )|I (t , i )−I ( t−1 , i )|
j=2 i=1 i=1
Subjected to ,
S ( I t ) > S ( I t−1 ) ,
Κ ( I t ) < Κ ( I t−1 ) ,
ℓi z i≤I i ≤ui zi , z i ∈ {0 , 1 } ,
D ( I t ) > D ( I t−1 ) ,
n
∑ I i , t =1,
i=1
I i ≥0 ; i=0 , 1 , .. .. n (3)
To determine the ideal portfolio from the data set, the goal function in equation (3) must be
maximized for each iteration by meeting all restrictions.
8.2.3 Quantum whale optimization for MPO

The objective function is solved by the method of QWOA, which this study suggests. From
the portfolio dataset, the optimum portfolio is identified. To improve on the traditional WOA
method of determining the ideal portfolio, this algorithm is presented. The intelligent
behavior of humpback whales (search agents) is used to create the conventional WOA [29].
These whales have a distinct bubble-net feeding habit that is optimized through mathematics.
This algorithm's main advantages are its ease of implementation, which requires fewer

12
parameters, and its potential to increase versatility. All of the individual whales or search
agents are taken to occupy places inside the prescribed solution space, which is first
established using the WOA algorithm. Subsequently, the search agents are presumed to
display actions like surrounding targets, investigating, and taking advantage of the search
area for every repetition. Every search phase updates the locations of the search agents as the
algorithm looks for the best effective solution during each mechanism. The fitness function
evaluation leads to the identification of the best possible solution in the search space. The
assets in this work serve as the search agents in the search space, and the optimum solution
identifies the best combinations of assets to create the portfolio. All of the individual whales
or search agents are taken to occupy places inside the prescribed solution space, which is first
established using the WOA algorithm. Subsequently, the search agents are presumed to
display actions like surrounding targets, investigating, and taking advantage of the search
area for every repetition. Every search phase updates the locations of the search agents as the
algorithm looks for the best effective solution during each mechanism. The fitness function
evaluation leads to the identification of the best possible solution in the search space. The
assets in this work serve as the search agents in the search space, and the optimum solution
identifies the best combinations of assets to create the portfolio.

(i) Mathematical descriptions for WOA

The algorithm's optimization process involves three key phases: encircling prey, exploration,
and exploitation. In the encircling phase, search agents identify and surround the prey, with
the agent closest to the prey initially considered the optimal solution. The positions of
neighboring agents are then updated based on this optimal position. The mathematical
formulation involves the coefficient vectors and the current position, which is adjusted
through iterations. During the exploration phase, agents update their positions randomly
rather than based on the best solution, allowing them to search the space more broadly. This
phase uses random vectors and coefficients to guide the agents to new positions in the search
space. Finally, in the exploitation phase, the algorithm simulates the bubble-net attacking
behavior of whales, combining a shrinking encircling process and a spiral position update.
This phase uses a probability-based approach to switch between these two methods, refining
the agents' positions to converge on the optimal solution.

(ii) QWOA modelling

13
The use of quantum computing in optimization algorithms aids in achieving desired
performance by giving a quadratic improvement over traditional techniques. WOA faces the
issue of local optima and attains a lower convergence rate, even if it can provide more
effective search techniques. Thus, the suggested study employs the quantum notion to prevent
early convergence and maintain a better balance between local and global exploration in
conventional WOA. The advent of the QWOA for PO is driven by the shortcomings of the
conventional WOA algorithm, which include delayed convergence, local optima stagnation,
and decreased accuracy. The QWOA for MPO is presented in Algorithm 1.

Algorithm 1: QWOA for MPO

Initialization of the search agents in the search space with multi-constraints, maximum

iterations
Max tr
Evaluation of the fitness function using below equation

[ ]
t n n
max Ν t =Ν ( t−1 ) ∏ ∑ It, i~
Ψ i, t −∑ κ (t , i )|I (t , i )−I ( t−1 , i )|
j=2 i=1 i=1
Subjected to ,
S ( I t ) > S ( I t−1 ) ,
Κ ( I t ) < Κ ( I t−1 ) ,
ℓi z i≤I i ≤ui zi , z i ∈ {0 , 1 } ,
D ( I t ) > D ( I t−1 ) ,
n
∑ I i , t =1,
i=1
I i ≥0 ; i=0 , 1 , .. .. n

Compute and identify the best search agent Η⃗ ∗¿ ¿

While ( τ < Max tr )

For every search agent in the search space,
Update b , δ , β , l and ρ
// Encircling behavior
if ( ρ<0 . 5 )

if (|δ|<1 )
Update the position of the search agent using equation

⃗ =|⃗β⋅Η⃗ ∗( τ )− Η⃗ ( τ )|.
℘

14
//exploration (quantum entanglement)

else if (|δ|≥1 )
Mimic the entanglement behavior on the search agents for
exploration using equations
ℑ=|ϕ|×W ij + ( 1−|ϕ|)×W ik 0<|ϕ|<1 and
⃗ ( τ +1 )= Η
Η ⃗ ( τ )+ Η
f.

Update the position of the search agent using equation

Γ 1
ℏ=|ℑ± ln
2 ω
| ( )
Identify the location of the search agent based on iterations using

equation
Η f =ℑ ±η|ℑ−ℏ|ln (1 ω )
Update the position of the search agent using equation
⃗ ( τ +1 )= Η
Η ⃗ ( τ )+ Η
f

end if
// Exploitation
else if ( ρ≥0 .5 )
Search agent’s position is updated using equation

Η⃗ ( τ+1)= ¿ { Η⃗ ∗( τ )−⃗δ⋅℘⃗ if ρ<0.5 ¿ ¿¿¿

end if
end for
Check for the search agents beyond the search boundary and amend it
Fitness evaluation using equation

15
 [ ]
t n n
max Ν t =Ν ( t−1 ) ∏ ∑ It, i~
Ψ i, t −∑ κ (t , i )|I (t , i )−I ( t−1 , i )|
j=2 i=1 i=1
Subjected to ,
S ( I t ) > S ( I t−1 ) ,
Κ ( I t ) < Κ ( I t−1 ) ,
ℓi z i≤I i ≤ui zi , z i ∈ {0 , 1 } ,
D ( I t ) > D ( I t−1 ) ,
n
∑ I i , t =1,
i=1
I i ≥0 ; i=0 , 1 , .. .. n

Evaluate the solutions obtained for global best and update Η⃗ ∗¿ ¿

τ =τ +1
end while
Return Η⃗ ∗¿ ¿

8.3 Results and discussion

To determine the ideal portfolio with the optimum asset distribution, a broad range of asset
sets can be used with the suggested MPO model. The approach that is being given assists
investors in managing market swings by determining the optimal asset allocation over various
time periods that might provide the lowest possible risks and losses. Higher-order restrictions
are taken into account in addition to other real-world limitations by taking into account the
stock market's risks and volatility. The model's main goal is to identify the asset distribution
that may offer a greater net return rate while posing the fewest risks. Empirical analysis has
been done to increase the usefulness of the suggested model, and the results are compared
with the current methods to show the improvement.

First, a detailed discussion of the computational setup and description of the financial dataset
used for study of the proposed model are given. The performance increase of the suggested
model is interpreted through analysis using key performance indicators like the Sortino ratio,
Sharpe ratio, information ratio, and other metrics. In order to confirm the superiority of the
suggested research, the obtained findings are also compared with the models that are
currently in use. The results are additionally expanded for other time intervals by modifying

16
the upper limits, as the suggested model operates with regard to the time intervals. The next
parts cover the general improvement and operation of the suggested model in MPO.

8.3.1 Dataset description

The 32_investment_portfolios with investment and operating profits (OPs) gathered from the
Kenneth French data set are used for all simulations. Utilizing the monthly results from July
1963 to July 2021, the 32 investment portfolios are built. Value-weighted returns for market
equity (ME)-based portfolios were included in the dataset. The dataset contains return figures
for thirty-two assets during the fiscal years 1963–2021. Every year, at the end of June, return
values are gathered and divided into two size groups: small and giant, according to the New
York Stock Exchange's (NYSE) breakpoint. Based on the NYSE quartile breakpoints, the
stock returns in each group are once more distributed across four distinct OP groups and four
investment groups. The June dataset's portfolios for the year t to t+1 include all NYSE,
AMEX, and National Association of Securities Dealers Automated Quotations (NASDAQ)
equities at various points in time. The dataset that is being used for simulation has 697 rows
and 33 columns in total.

8.3.2 Performance indicators

Numerous assessment metrics, including Shannon entropy, information ratio, Sortino, Sharpe,
and downside deviation, are used to assess overall performance. Table 1 displays the
mathematical expressions for the assessment indicators.

Table 1 Mathematical formulation of varied evaluation indicators

Sharpe ratio E p ( φa −φ b )
Sharpe=
σa
Sortino ratio φ a −φb
Sortino=
σa

{
STARR ratio αt
STARR=¿ , α t >0 ¿ ¿¿¿
CVaR ( φa−φb )

{(
Information ratio αt
Info=¿ , α t >0 ¿ ¿¿¿
σ φ a −φb )

17
 √
Downside deviation T
1
DD= ∑ min ( φ ja−φ jb ) , 0
T j=1

E p specifies the expected value, φ a specifies the asset return value, φ b specifies
where,
σ
the risk-free return value, a indicates the standard deviation, T specifies the time period,
φ jaspecifies the asset return within the time period and
φ jb indicates the risk free return
within the time period.

8.3.3 Empirical analysis

To measure the performance against the other current methods, the suggested MPO model is
simulated. The literature's meta-heuristic algorithms are used as a point of comparison.
Initially, the suggested model required pre-processing to account for missing information. It
is possible to use this general method on any portfolio dataset. Following pre-processing, the
limitations are assessed, and an objective evaluation is used to determine the portfolios' net
return rate. In order to choose the best portfolio from the dataset by maximizing objectives,
the QWOA is presented. Figure 2 shows an illustration of the total wealth values that the
algorithm produced at various points in time.

Figure 2: The overall wealth secured by the portfolios based on period

The model's secured wealth values are shown by the y-axis, while the alterations in the time
period are represented by the x-axis. According to the graph, when the time period is
extended, the wealth values grow linearly. The model's profit is first determined to be modest,

18
but as time goes on, risk factors become apparent, and the model reduces these elements to
provide better profits. The new portfolios produced bigger earnings than the ones the model
selected from the previous ones. The profit increases dramatically from 115 to 160 when the
time period is extended from 1 to 4. This suggests that, subject to a few limitations, the
suggested approach is ideal and can lead to increased wealth. Figure 3 compares the
convergence of proposed and current techniques and shows the findings visually.

Figure 3: Comparison analysis in terms of convergence for both proposed and existing
algorithms
By changing the time period from 2 to 10, the convergence of the model is assessed. The
suggested model converged more quickly than the other techniques, as can be seen in the
graph. This suggests that the suggested technique is more efficient for PO since it requires
less computing time. The rate of convergence increases as the time interval increases.
Therefore, the suggested approach can be understood to be able to identify the ideal portfolio
in the shortest amount of time. Table 2 displays the results of the comparison of the Sharpe

ratio for the upper bound ( i

u =0 . 05 )
.

Table 2: Sharpe ratio comparison for the proposed and existing algorithms ( i
u =0 . 05 )
Techniques Sharpe ratio
Mean Optimal
FA 1.3023 2.6582
PSO 1.5921 2.6265
FOA 0.7703 2.0340
GWO 2.3138 5.3432
CSO 2.9256 5.2673

19
 MBO 2.8634 5.4982
FSO 2.9952 5.4925
WOA 3.2811 5.7412
QWOA 4.101048 5.016254

The Sharpe ratio of the model was calculated and contrasted with current techniques. The
model sought to increase the ratio's value. A mean Sharpe ratio of 4.101048 and an ideal
Sharpe ratio of 5.016254 were attained by the suggested model. While the WOA and GWO
algorithms' mean and ideal ratios were lower than those of the suggested model, they still
exhibited excellent optimal Sharpe ratios. The ideal ratios and mean of the FOA algorithm
were extremely low. A greater Sharpe ratio, which indicates a better link between returns and
risk incurred, was offered by the suggested model. There were only minor differences found
between QWOA and regular WOA. The Sortino ratio comparison analysis is shown in Table
3.

Table 3: Sortino ratio comparison of proposed and existing algorithms ( i

u =0 . 05 )
Techniques Sortino ratio
Mean Optimal
FA 0.2746 0.2644
PSO 0.2817 0.2945
FOA 0.2732 0.2725
GWO 0.2930 0.3728
CSO 0.2986 0.3456
MBO 0.2973 0.3372
FSO 0.2817 0.2903
WOA 0.3004 0.3327
QWOA 1.01031 0.89327

At an aggregate mean of 1.01031, the suggested QWOA model outperformed current models
in terms of Sortino ratio, and its ideal Sharpe ratio was 0.89327. Although the GWO
algorithm produced a better ideal Sortino ratio, the WOA method guaranteed a higher average
Sortino ratio. The Sortino ratio values of the QWOA model were notably higher than those of
conventional algorithms, suggesting that it offered the best risk-adjusted returns. Table 4
displays the STARR ratio comparative study of the suggested and current methods.

20
 Table 4: STARR ratio comparison for the proposed and existing algorithms ( i
u =0 . 05 )
Techniques STARR ratio
Mean Optimal
FA −0.03521 −0.03737
PSO −0.03577 −0.03742
FOA −0.03458 −0.03663
GWO −0.03737 −0.04253
CSO -0.03658 -0.04163
MBO -0.03691 -0.04272
FSO -0.03719 -0.04288
WOA −0.03905 −0.04305
QWOA −0.04205 −0.04632

The STARR ratio is calculated using the suggested model and contrasted with traditional
approaches. The ideal STARR ratio is -0.04632, whereas the mean is -0.04205. The ideal
ratios are -0.03737, -0.03742, -0.03663, -0.04253, and -0.04305, whereas the current models
attain mean STARR ratios of -0.03521, -0.03577, -0.03458, -0.03737, and -0.03905. The
suggested model outperforms the current models in terms of outcomes, with lower values
being favored. The proposed and current algorithms' information ratio comparison findings
are displayed in Table 5.

Table 5: Information ratio comparison for the proposed and existing algorithms ( i
u =0 . 05 )
Techniques Information ratio
Mean Optimal
FA 0.1616 0.1536
PSO 0.1647 0.1695
FOA 0.1635 0.1617
GWO 0.1722 0.2060
CSO 0.1714 0.1663
MBO 0.1706 0.1659
FSO 0.1716 0.1667
WOA 0.1759 0.1921
QWOA 0.17630 0.19826

21
The mean and ideal ratios of 0.17630 and 0.19826, respectively, demonstrate the high
information ratio that the suggested model delivers. With an ideal value of 0.1921, and a
mean of 0.1759, the WOA algorithm yields superior results. At 0.2060, the GWO algorithm
attains a higher optimum ratio. In terms of information ratio, the QWOA model performs
better than previous models, suggesting that it can find the best portfolios with large excess
returns for longer investment plans. To assess how well the suggested and current models
compare, the Shannon entropy metric is calculated.

Table 6: Shannon entropy comparison of both proposed and existing algorithms ( i

u =0 . 05 )
Techniques Shannon entropy
Mean Optimal
FA 4.2213 4.1845
PSO 4.2087 4.2044
FOA 4.2235 4.2332
GWO 4.1639 4.0754
CSO 4.1547 4.0653
MBO 4.1665 4.0698
FSO 4.1765 4.0701
WOA 4.1047 4.0839
QWOA 4.3918 4.34452

Once the value is set to 0.05, the Shannon entropy is obtained. The portfolio's diversity is
assessed using the Shannon entropy metric, which is higher when it is high. For the suggested
model, the Shannon entropy mean and optimum values are 4.3918 and 4.34452, respectively.
The QWOA method among the examined models achieves a high mean and ideal value for
Shannon entropy, with values of 4.3918 and 4.34452, respectively. The suggested model
produced optimal values when compared to the current algorithms, suggesting that it may be
used to choose diversified portfolios that would result in greater net return rates. Table 7
presents a comparison of the suggested and current algorithms' downside deviations.

Table 7: Downside deviation comparison for the proposed and existing algorithms

( u i=0 . 05 )
Techniques Downside deviation

Mean Optimal

22
 FA 0.2273 0.2512

PSO 0.2252 0.2269

FOA 0.2232 0.2384

GWO 0.2257 0.2040

CSO 0.2231 0.2081

MBO 0.2217 0.2002

FSO 0.2208 0.2016

WOA 0.2306 0.2316

QWOA 0.2266 0.2154

This results in the downward deviation seen in Table 7 when the upper bounds are set to 0.05.
Minimizing the downside deviation is necessary to create optimum portfolios that yield
higher rewards. This metric aids in determining the total loss an investor may incur in
addition to the standard deviation loss. For the suggested QWOA model, the optimal
downside deviation is 0.2154, while the mean is 0.2266. When compared to traditional
algorithms, the QWO model fared better in terms of downside deviation. Therefore, it is
evident that the QWOA can find the best portfolio, even if it results in a minimal loss value
that is larger than the portfolios chosen by the current algorithms. Figure 4 displays the
comparison between the return ratios of the suggested method and the conventional
algorithms.

23
(a)

(b)

(c)

24
 u =0.05
Figure 4: Average returns of the optimal portfolio selected by the algorithms for (a) i
u =0.07 and (c) ui =0.10
(b) i

The suggested model's main goal is to raise net return rates while minimizing risk. Investors
expect higher rates of return prior to executing any investment strategy, and a greater return
value over a longer time horizon is required for the portfolio. Thus, the proposed QWOA
model is built for multiple time periods, and the optimal portfolio is selected by the model
that can survive extended time periods. The graph shows that the QWOA offered a high
return portfolio that was ideal. When compared to the portfolios selected by the other
algorithms, the returns offered by the QWOA are substantial. By changing the duration in
days, the portfolio's return values may be found. This suggests that the model's optimum
portfolio is resilient to extended investment plans. Figure 5 shows the results of equating the
net return rates produced by the portfolio chosen by the proposed QWOA with those of the
traditional algorithms.

(a)

25
 (b)

(c)

Figure 5: Comparison of net return rates of the proposed and existing schemes for (a)
ui =0.05 (b) ui =0.07 and (c) ui =0.10

Changes of the upper limits to 0.05, 0.07, and 0.10 provide the figures. The net return rates of
the suggested and WOA methods are nearly identical in Figure 5(a). As demonstrated by the
comparative study, the suggested model has a net return rate of 0.082%, which is greater than
the net return rates of the other algorithms. When the loss rates were compared, the QWOA
model produced the least amount of loss, suggesting that the suggested model is more
capable of handling market limitations than the other models. The box plots' error bars show
the loss values as a result of the risk variables. The net return rates obtained by the QWOA
are contrasted with those of other algorithms with the upper bound of 0.07 in Figure 5(b). The
suggested model's return rates are greater in the graphical depiction than in the case of the

26
current models. The ideal net return rate in this case cannot be reached by the conventional
WOA method.

Furthermore, the WOA model has large loss values, which results in reduced return values.
FA obtained the lowest net return rates among the studied algorithms. The net return rates
attained by the suggested model are significantly higher than those of the current methods in
Figure 5(c). The WOA algorithm's inability to produce larger net returns in this case suggests
that the model is unable to handle the market risks. Furthermore, when compared to the
suggested model, the FA and FOA algorithms had extremely poor net returns. Based on the
study conducted, it can be inferred that the QWOA model that has been provided has the
capability to determine the ideal portfolio for investors in order to provide greater net return
rates. This is made possible by advancements made to the conventional WOA algorithm. The
program found the global best solution with more efficiency and accuracy thanks to the idea
of quantum computing.

8.3.4 Statistical test analysis

A statistical analysis is done in this part to evaluate the quality of the suggested task. In this
case, the t-statistic, p, and standard deviation values are evaluated to complete the study. To
confirm the superiority of the suggested strategy, the obtained values are also compared with
those of other optimization techniques that are currently in use. The statistical analysis
findings for the suggested and current methodologies are displayed in Table 8.

Table 8: Statistical analysis results

Algorithm Standard T-statistic P-value

deviation

CSO 0.0051 5.9646 0.072

MBO 0.0047 5.0182 0.068

FSO 0.0042 4.6753 0.057

WOA 0.0003 3.1085 0.0031

QWOA (Proposed) 0.0002 2.8106 0.0028

The performance of the suggested strategy in MPO is shown by the acquired results, which
are displayed in Table 8. The suggested model's obtained P-value and T-statistics outperform
those of other conventional optimization techniques. In a similar vein, the robustness of the

27
suggested research is shown by the standard deviation of the suggested model in comparison
to alternative methods.

8.3.5 Discussion

The proposed model offers optimal outcomes compared to classical algorithms, enabling
effective search in the exploration phase and improved accuracy in selection due to the
adoption of quantum mechanics' entanglement behavior. This enhances computing efficiency
and performance of traditional WOA, overcoming slow convergence issues. Quantum
entanglement uses quantum systems to assess the superposition of states, allowing optimizers
to explore multiple solutions in a large search space. The algorithm solves multi-constrained
optimization problems, providing an optimal portfolio with higher net return rates and
minimum risks. Simulation outcomes show slight changes in return rates when upper bound
values are varied, but the model maintains balance in selecting the optimal portfolio. The
convergence rate of the proposed QWOA approach is enhanced at each entanglement level.
Empirical analysis using indicators like downside deviation, Shannon entropy, information
ratio, STARR ratio, Sortino ratio, and Sharpe ratio shows the better efficacy of the proposed
QWOA approach compared to other existing optimization algorithms. The proposed model
selects portfolios considering market fluctuations based on different time periods,
outperforming other algorithms.

Table 9: Quantitative analysis of the proposed and existing PO schemes

Authors Methods Datasets Sharpe ratio
Corsaro et Split Bregman DowJones, NASDAQ100, 0.460 – 0.933
al. [30] iteration FTSE100, SP500,
NASDAQComp, FF49
Nesaz et al. NSGA-II Stock portfolio from p-value: 0.04, t-
[31] NYSE static: 1.8 for 40
stocks
Khan et al. QBAS Shanghai Stock Exchange 9.210 for 20 stocks
[32] 50 Index
Oprisor and Hidden markov Data from Quotemedia 0.23 for confidence
Kwon [33] model, Black > 60%
Litterman model
Wang et al. Markowitz Mean– Real time data Optimized Sharpe

28
 [34] Variance ratio 5.624
optimization
Setiawan et GWO Data randomly collected 0.2455
al. [35] from LQ-45 index in
Indonesia Stock
Exchange.
Setiawan et GA Data randomly collected 0.24918
al. [35] from LQ-45 index in
Indonesia Stock
Exchange.
Setiawan et Grasshopper Data randomly collected 0.24551
al. [35] optimization (GO) from LQ-45 index in
Indonesia Stock
Exchange.
Shahid et al. Gradient based S&P BSE Sensex of 0.3969
[36] optimization (GBO) Indian stock exchange
Proposed QWOA 32 investment portfolios 5.016254

The quantitative analysis used the Sharpe ratio to analyze performance indicators in MPO
problems. Corsaro et al. provided a formula for allocating wealth values to assets within a
time horizon, but it was not suitable for long-term investment plans. The NSGA-II model
provided optimal results but faced risk factors. The QBAS model achieved a better Sharpe
ratio but was affected by market fluctuations over time. Oprisor and Kwon's model balancing
risk-adjusted returns performed better but required further evaluation for multiple investment
views. The proposed QWOA model is optimal and suitable for solving MPO problems with
higher-order constraints.

This paper introduces an improved optimization algorithm based on quantum entanglement

behavior to solve the Multi-Positional Optimization (MPO) problem with multiple
constraints. The goal is to mitigate real-time and higher-order constraints and achieve
maximum net return rates in portfolio management. The model performed well in identifying
optimal asset distributions that can withstand longer investment plans. Simulation results
showed that the quantum mechanics principle improved the algorithm's efficiency and
accuracy. The model achieved a higher convergence rate and total net returns above 0.85%

29
for different caps values. The model is suitable for determining optimal wealth distributions
with better performance rates than existing models. Future research aims to use different
quantum optimization algorithms and implement artificial intelligence-driven approaches to
predict return rates based on historical data.

9. PROPOSED METHODOLOGY II: AN INTELLIGENT PORTFOLIO

MANAGEMENT SCHEME BASED ON HYBRID DEEP REINFORCEMENT
LEARNING AND CUMULATIVE PROSPECTIVE APPROACH

9.1 Proposed Method

In order to build a trading plan and generate a sizable return, stock predictions play a crucial
part. In the case of active portfolio design and optimization, the successful stock prediction
outcomes serve as a requirement. The goal of portfolio creation is typically to maximize
projected return rate for a given risk.

9.1.1 Functional blocks of the framework:

Figure 6 provides a succinct illustration of the proposed approach's general schematic

architecture/ Functional blocks of the framework:.

30
 Figure 6: Proposed architecture

1) Hybrid Deep reinforcement learning:

Through a trial-and-error method, reinforcement learning—in which a software agent is

placed in an environment—progresses beyond learning to raise the cumulative reward
concept. DRL is a cutting-edge and quickly developing technique that uses neural networks
as function approximations for traditional learning processes such as Q-learning.

2) Markov Decision process:

The suggested model assigns the agent to conduct the necessary actions for suitable
interaction with the environment and regards the stock price prediction processes as MDP.
The MDP is a 4-tuple set that may be explained as;

MDP= { ST , AC ,TR AC ( ST , ST '), RE AC ( ST , ST ' ) } (4)

Whereas the collection of possible states is represented by ST , the action set that the agent
can take over the environment is described by AC . The reward function is represented by
RE, while the transition function (TR) is used to indicate the change from one state to
another.

31
3) Gated Recurrent Unit Network:

The GRUNs are a representation of the gating strategy used in recurrent neural networks,
which consists of gates that regulate both the input at that moment and previously concealed
states. The reset and update gates are part of the GRU.

 The update gate determines how much of the past memory must be carried over into the
future.
 The reset gate determines the overall quantity of historical data that needs to be erased.

In contrast to the GRUN, which has effective agent-environment interactions, future stock
prices are predictable.

4) Quantum Differential Evolution Algorithm:

To increase prediction accuracy, Q-DEA is used to choose the best course of action for each
contact. The individual set that represents a chromosome is the foundation of how the
traditional DEA method operates. The Q-DEA technique is used to choose the best assets
from the anticipated stocks while minimizing the number of chromosomes and processing
time. The primary distinction between Q-DEA and other DEAs is that the chromosomes
encode all potential solutions rather than just one.

5) Evaluation of fitness function:

The best asset that enhances the trading outcome is implied by the forecasted stock's
increased profit rate. Using the fitness function listed below, the best assets are chosen from
the anticipated stock.

FITNESS= MAX (Stock profit ) (5)

Quantum selection process: Using equation Z= S p − R f , which can be expressed as follows,

people are chosen based on fitness function in order to choose the finest assets from the
anticipated stock.

S=¿ {Best asset ; if stock=High profit¿¿¿ (6)

32
Based on increased profit rates, the Q-DEA technique selects the best assets from the
forecasted stock. The CPT model is used to carry out the best portfolio selection procedure
based on the optimal assets that were derived from the forecasted stock using Q-DEA.

6) Portfolio Optimization

A portfolio's optimal selection aids in achieving greater stock market gains. It is crucial to
choose the ideal portfolio with the ideal asset collection in order to optimize the return rate.
Better trading services are provided via portfolio optimization using the CPT model. Among
the presumptions covered by CPT are

Assumption 1: Regarding the return rate in relation to the end wealth, the CPT investors are
more concerned.

Assumption 2: In contrast to financial market participants, CPT investors do not incur

transaction costs while making investments.

9.2 Results

This section describes the performance results of portfolio optimization using CPT and stock
price prediction using DRL. The PYTHON simulation tool is utilized to evaluate the results
of the experiments. Twenty percent of the stock data is utilized for testing and eighty percent
is used for training in the suggested task. The Nation Stock Exchange (NSE), India's NIFTY
50 index values are used in the suggested research project to collect data for accurate stock
price forecast and stock market portfolio optimization. The data, which shows trade volumes
and price histories, was collected from 50 stock firms. The stock data of Adaniport.csv, which
was gathered between 2007 and 2021, has been taken into consideration in the suggested
study. You may obtain the NIFTY-50 Stock Market Data (2000-2021) from
https://www.kaggle.com/datasets/rohanrao/nifty50-stock-market-data.

A. Performance metrics

A variety of performance indicators, such as Mean Absolute Error (MAPE), Mean Square
Error (MSE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error
(MAPE), are taken into consideration while estimating the success of the suggested model,
which is based on stock price prediction and portfolio optimization.

33
B. Performance analysis and comparison

Better results in choosing the optimum portfolio with optimal return rates may be obtained by
putting the suggested portfolio optimization model into practice using the PYTHON
simulation tool. This section describes the several performance measures that were analyzed
by contrasting the suggested strategy with current methodologies. The performance analysis
of the suggested SPPO model is shown in Table 10.

Table 10: Performance analysis of Proposed SPPO model

Performance metrics Performance outcomes

MSE 0.167

MAE 0.393

RMSE 0.409

MAPE 0.515

Sharpe ratio 0.394

Sortino ratio 1.544

Information ratio 0.193

It is evident from the above table that the suggested SPPO model has produced better results
in terms of MSE, MAE, RMSE, MAPE, Sharpe ratio, Sortino ratio, and information ratio. In
the cases of MSE, MAE, RMSE, and MAPE, the suggested model evaluates a lesser error
rate of 0.130, 0.114, 0.148, and 0.153. Better results were also obtained by the information
ratio, sortino ratio, and Sharpe ratio.

34
1) Actual and predicted price analysis:

Figure 7 illustrates the process of comparing the profit or loss realized by past and present
stocks through the analysis of actual and expected prices.

Figure 7: Actual Vs predicted price values

The GRUN model predicts future stock prices for Adani Port data from 2018 to 2021,
providing valuable insights for traders. The analysis reveals that the anticipated price yields
superior results when compared to the actual price, indicating its potential for effective
decision-making in the stock market.

2) Evaluation of error metrics:

The optimum approach may be examined using minimal error rate performance analysis,
where the error metrics are assessed to determine the selection loss. The assessment of error
metrics by adjustment of the period size is shown in Figure 8.

Figure 8: Error metrics Vs Epoch size

35
The proposed model's MAE, MSE, RMSE, and MAPE performance are evaluated by
changing epoch sizes between 0 and 300. The model's error metrics decrease as epoch size
increases. If epoch size is 50, the model achieves 0.2-0.3 error metrics. When epoch size is
100, the model's error rate ranges from 0.05 to 0.15%. The SPPO model provides optimal
training, resulting in low error values.

Figure 9: Comparison of Error metrics (a) MAE) (b) RMSE (c) MSE (d) MAPE

The suggested SPPO model's MAE, RMSE, MSE, and RMSE performance are compared to
those of other techniques, including CNN, LSTM, and Bi-LSTM, in the research. When
compared to other methods like CNN, LSTM, and Bi-LSTM, the suggested model obtains
reduced RMSE values of 0.14, MAE values of 0.11, MSE values of 0.13, and MAPE values
of 0.15. The suggested model has lower error rates than the current approaches, which have
higher error rates as a result of improper stock data management and inadequate training
capacity. The performance of the suggested model is justified since it has lower error rates,
which are frequently brought on by improper management of stock data and inadequate
training capability.

36
3) Convergence curve analysis:

The suggested model analyzes the convergence performance of both current and new models
in relation to RMSE. Figure 10 describes the suggested SPPO model's convergence analysis.

Figure 10: Convergence analysis

The study examines the convergence performance of portfolio optimization by adjusting the
number of iterations based on the model's fitness function. Results show that the suggested
model converges faster over increasing iterations compared to current methods like PSO,
ABC, GA, and QBAS. This is due to less stock data used for training, more over fitting
problems, and longer selection processes, resulting in lower convergence rates.

4) Portfolio performance assessment:

An ideal portfolio should have minimal risk and a maximum net return rate. The portfolio
optimization performance based on return rates is displayed in Figure 11.

37
 Figure 11: Evaluation of portfolio performance

CPT is used to choose the best portfolio based on the chosen best assets in order to maximize
the created portfolio. The suggested model's return rate, produced by the 2021 portfolio,
exhibits superior performance. When comparing the suggested model to the current models,
the suggested model yields higher performance results. The return value that various
portfolios have achieved is displayed in Table 11.

Table 11: Evaluation of portfolio performance

Techniques 2018 (%) 2019 (%) 2020 (%) 2021 (%)

PSO 2.76 2.16 2.87 3.28

ABC 1.76 2.08 2.67 3.19

GA 1.24 1.98 2.45 3.09

QBAS 1.99 2.34 2.67 3.12

Proposed 3.28 3.19 3.09 3.59

5) Time complexity analysis:

38
The time complexity analysis in selecting an optimal portfolio based on predicted stocks of
proposed SPPO and existing optimization strategies are described in Figure 12.

Figure 12: Time complexity analysis

The proposed SPPO model outperforms existing methods in terms of run time, with the
optimal portfolio selection time taking 321 seconds. This is significantly shorter than existing
methods like PSO, ABC, GA, and QBAS, which took 478 seconds, 489 seconds, 422
seconds, and 400 seconds respectively. The difference is attributed to the accumulation of
large data, less convergence, and high storage space requirements of existing approaches.

The mean and optimal values obtained by evaluating Sharpe ratio is presented in Table 12.

Table 12: Performance comparison of Sharpe ratio

Techniques Sharpe ratio

Mean Optimal

PSO 0.93 1.73

ABC 1.65 2.26

GA 1.28 3.87

QBAS 3.39 4.63

Proposed 2.16 5.59

39
It can be concluded from the study that the suggested model offers a stronger relationship
between returns and risk due to its greater Sharpe ratio. The sortino ratio evaluation's mean
and ideal values are shown in Table 13.

Table 13: Performance comparison of Sortino ratio

Techniques Sortino ratio

Mean Optimal

PSO 0.33 0.42

ABC 0.75 0.76

GA 0.43 0.32

QBAS 0.35 0.29

Proposed 1.85 1.09

On comparison of proposed sortino values, the existing values are inferior. The mean value of
PSO is comparatively lower and optimal value of ABC is the second highest value. Hence, it
can be stated that the proposed model is capable of providing optimal return rates of
predicted stocks compared to other algorithms. Table 14 interprets the mean and optimal
values obtained by evaluating information ratio.

Table 14: Performance comparison of Information ratio

Techniques Information ratio

Mean Optimal

PSO 0.26 0.23

ABC 0.19 0.19

GA 0.18 0.21

QBAS 0.21 0.23

Proposed 0.27 0.25

9.2.1 Discussion

40
The performance of a proposed technique is evaluated using current methods such as PSO,
ABC, GA, and QBAS, which were previously used for stock prediction. The suggested
technique has lower computing complexity and a better sortino ratio and sharp ratio, despite
some restrictions. Table 15 displays the performance evaluation of the suggested technique
under methods.

Table 15: Performance of proposed method

Techniques Performance matrices

Sortino ratio Sharpe ratio Max Drawdown

SARL 1.18 1.54 -

DPM 0.86 1.95 -

CRP 1.21 0.97 44.36

OLMAR 0.16 1.89 99.00

EW 1.64 1.92 44.32

ONS -2.61 - 51.06

UP -2.63 - 44.31

WMAMR -16.35 - 99.80

Proposed 1.85 2.16 20.950

In the method uses State Augmented Reinforcement Learning (SARL), Deep Portfolio
Management (DPM), Constant rebalanced portfolio (CRP), On-Line Moving Average
Reversion (OLMAR) and Equal Weight (EW) are the techniques used in the method to
evaluate the performance. The high dimensionality, high computational cost, and difficulty in
interpretability are major challenges in SARL. The over fitting and high computational
complexity are the major challenges. Low security and high error rate are the major
challenges presented in CRP. High error rate and high model complexity makes the OLMAR
method more complex. From the resultant performance evaluation the proposed method is

41
more suitable in the stock prediction. Online Newton Step (ONS), Universal Portfolios (UP),
and Weighted Moving Average Mean Reversion (WMAMR) are also used in the proposed
method to evaluate the performance. From the resultant performance evaluation the proposed
performed is effectively performed in the sector.

The study focuses on the importance of accurate portfolio selection in the stock market. The
NIFTY-50 Stock Market Data (2000-2021) dataset is used to gather stock data. A hybrid DQL
model is developed to predict stock prices using MDP, while the GRUN model simulates
agent interactions with the environment. The Q-DEA model is used to select the best assets
from stocks, and the optimal portfolio is selected using the CPT approach. The model
achieved better performance in portfolio selection compared to other optimization strategies
in 2021. Future studies should consider more input features, risk metrics, and stock data to
improve the model's performance.

10. LIMITATIONS AND DIRECTIONS FOR FUTURE RESEARCH

Quantum-inspired algorithms, such as the Quantum Whale Optimization Algorithm

(QWOA), can be computationally intensive, especially when dealing with large datasets or
complex multi-period portfolio optimization (MPO) problems. This complexity may lead to
longer processing times compared to traditional optimization methods. The effectiveness of
the proposed models heavily relies on the quality and availability of financial data. Inaccurate
or incomplete data can adversely affect the accuracy of stock price predictions and,
consequently, the performance of the portfolio optimization models. While the models
demonstrate strong performance on specific datasets, their ability to generalize across
different market conditions and asset classes remains uncertain. The dynamic nature of
financial markets poses challenges in ensuring that the models perform consistently over
time. Implementing quantum-inspired optimization algorithms in real-world financial
systems may encounter integration challenges. These include compatibility issues with
existing infrastructure and the need for specialized expertise to manage and interpret quantum
computing outputs.

Future research in quantum-inspired portfolio optimization should focus on enhancing

algorithm efficiency to reduce computational complexity. Developing hybrid models that
integrate classical and quantum computing techniques can leverage the strengths of both,
potentially leading to more efficient and accurate optimization processes. Improving data

42
preprocessing and handling techniques is crucial for enhancing the reliability of stock price
predictions and overall portfolio optimization performance. Implementing robust methods to
manage missing or noisy data can significantly contribute to more accurate and dependable
models. Conducting extensive validation and testing of models across various market
scenarios and asset classes is essential to assess their robustness and adaptability. Backtesting
models with historical data can provide insights into their performance under different market
conditions, aiding in the refinement of optimization strategies. Exploring the practical
integration of quantum-inspired optimization algorithms into existing financial systems is
vital. Addressing challenges related to scalability, user-friendliness, and developing interfaces
that facilitate seamless interaction between quantum computing platforms and traditional
financial software will be key to successful implementation. Incorporating advanced risk
management strategies into portfolio optimization models can enhance their resilience.
Developing models that dynamically adjust to changing market conditions and effectively
manage various types of financial risks will contribute to more robust investment strategies.
Encouraging collaboration between quantum computing experts, financial analysts, and data
scientists can lead to the development of more sophisticated and effective portfolio
optimization models. Such interdisciplinary efforts can foster innovation and the creation of
models that are both theoretically sound and practically applicable.

11. GOALS ACHIEVED BY RESEARCH/CONCLUSION

This research presents a comprehensive approach to portfolio optimization using quantum-

inspired techniques, deep reinforcement learning (DRL), and the Markowitz Mean-Variance
Optimization Model. The first study introduces a multi-period portfolio optimization model
that maximizes net return rates by incorporating quantum-inspired whale optimization
algorithm (QWOA) and quantum entanglement. The second study proposes an intelligent
portfolio management scheme that combines stock prediction with portfolio optimization
using a hybrid DRL model, a gated recurrent unit network (GRUN), and the quantum
differential evolution algorithm (Q-DEA). The cumulative prospect theory (CPT) model is
used to select the optimal portfolio based on predicted best assets. The third study focuses on
developing an effective portfolio management system by integrating quantum computing
approaches with the Markowitz Mean-Variance Optimization Model. The Quantum
Approximate Optimization Algorithm (QAOA) is used to obtain binary solution vectors
representing selected features, while a Quantum Autoencoder (QAE) model learns a low-
dimensional representation of input features.

43
Collectively, these studies contribute to the field of portfolio optimization by:

(1) Developing models that address multiple real-time and higher-order constraints to achieve
maximum net return rates.
(2) Introducing quantum-inspired optimization algorithms, such as QWOA and QAOA, to
enhance traditional optimization models.
(3) Integrating deep reinforcement learning techniques, like GRUN and Q-DEA, for accurate
stock price prediction and optimal asset selection.
(4) Employing advanced pre-processing and dimensionality reduction methods, including
forward fill, normalization, and quantum autoencoders, to improve data representation
and prediction accuracy.

This research has been instrumental in developing intelligent, sophisticated, and robust
methods for portfolio optimization. These approaches are versatile and applicable to various
financial contexts, offering flexibility for both broad market analyses and specific asset
evaluations.

REFERENCE

[1] Philip TU, William EE, Rajunor BA (2012) Portfolio Analysis Models: A Review.
European Journal of Business and Management. 4(18):101-20.

44
[2] Ban GY, El Karoui N, Lim AE (2018) Machine learning and portfolio optimization.
Management Science. 64(3):1136-54.ss

[3] Chen C, Leydesdorff L (2014) Patterns of connections and movements in dual‐map

overlays: A new method of publication portfolio analysis. Journal of the association for
information science and technology. 65(2):334-51.

[4] Klavans R, Boyack KW (2017) Research portfolio analysis and topic prominence.
Journal of Informetrics. 11(4):1158-74.

[5] Dai M, Jin H, Kou S, Xu Y (2021) A dynamic mean-variance analysis for log returns.
Management Science. 67(2):1093-108.

[6] Kalayci CB, Ertenlice O, Akbay MA (2019) A comprehensive review of deterministic

models and applications for mean-variance portfolio optimization. Expert Systems with
Applications. 125:345-68.

[7] Lwin KT, Qu R, MacCarthy BL (2017) Mean-VaR portfolio optimization: A

nonparametric approach. European Journal of operational research. 260(2):751-66.

[8] Ertenlice O, Kalayci CB (2018) A survey of swarm intelligence for portfolio

optimization: Algorithms and applications. Swarm and evolutionary computation. 39:36-52.

[9] Fliege J, Werner R (2014) Robust multi-objective optimization & applications in

portfolio optimization. European Journal of Operational Research. 234(2):422-33.

[10] Grant E, Humble TS, Stump B (2021) Benchmarking quantum annealing controls with
portfolio optimization. Physical Review Applied. 15(1):014012.

[11] Nystrup P, Madsen H, Lindström E (2018) Dynamic portfolio optimization across hidden
market regimes. Quantitative Finance. 18(1):83-95.

[12] Cho C, Yang J, Jang B. Spectrum of Influence: Heterogeneous Macroeconomic Factors’

Effects on Stocks Based on Size, Style, and Sector in the South Korean Market. Style, and
Sector in the South Korean Market.

[13] Liu YJ, Zhang WG (2013) Fuzzy portfolio optimization model under real constraints.
Insurance: Mathematics and Economics. 53(3):704-11.

[14] Chen W, Li D, Lu S, Liu W (2019) Multi-period mean–semivariance portfolio

optimization based on uncertain measure. Soft Computing. 23:6231-47.

45
[15] Kamali R, Mahmoodi S, Jahandideh MT (2019) Optimization of multi-period portfolio
model after fitting best distribution. Finance Research Letters. 30:44-50.

[16] Li Z, Tsang KH, Wong HY (2020) Lasso-based simulation for high-dimensional multi-
period portfolio optimization. IMA Journal of Management Mathematics. 31(3):257-80.

[17] Dai Y, Qin Z (2021) Multi-period uncertain portfolio optimization model with minimum
transaction lots and dynamic risk preference. Applied Soft Computing. 109:107519.

[18] Nokhandan BP, Khalili-Damghani K, Hafezalkotob A, Didehkhani H (2021) A Nash

bargaining solution for a multi period competitive portfolio optimization problem: Co-
evolutionary approach. Expert Systems with Applications. 184:115509.

[19] Babazadeh H, Esfahanipour A (2019) A novel multi period mean-VaR portfolio

optimization model considering practical constraints and transaction cost. Journal of
Computational and Applied Mathematics. 361:313-42.

[20] Zhai QH, Ye T, Huang MX, Feng SL, Li H (2020) Whale optimization algorithm for
multiconstraint second-order stochastic dominance portfolio optimization. Computational
Intelligence and Neuroscience. 2020.

[21] Behera, Mandakini, Archana Sarangi, Debahuti Mishra, Pradeep Kumar Mallick, Jana
Shafi, Parvathaneni Naga Srinivasu, and Muhammad Fazal Ijaz. “Automatic data clustering
by hybrid enhanced firefly and particle swarm optimization algorithms.” Mathematics 10, no.
19 (2022): 3532.

[22] Teplova, Tamara, Mikova Evgeniia, Qaiser Munir, and Nataliya Pivnitskaya. “Black-
Litterman model with copula-based views in mean-CVaR portfolio optimization framework
with weight constraints.” Economic Change and Restructuring 56, no. 1 (2023): 515-535.

[23] Yaman I, Dalkılıç TE. A hybrid approach to cardinality constraint portfolio selection
problem based on nonlinear neural network and genetic algorithm. Expert Systems with
Applications. 2021 May 1; 169:114517.

[24] Sharma M, Sharma PK, Vijayvergia HK, Garg A, Agarwal SS, Saxena VP. MCN
portfolio: An efficient portfolio prediction and selection model using multiserial cascaded
network with hybrid meta-heuristic optimization algorithm. Network: Computation in Neural
Systems. 2024 May 9:1-38.

46
[25] Ma Y, Wang W, Ma Q. A novel prediction based portfolio optimization model using deep
learning. Computers & Industrial Engineering. 2023 Mar 1; 177:109023.

[26] Kizys R, Doering J, Juan AA, Polat O, Calvet L, Panadero J. A simheuristic algorithm
for the portfolio optimization problem with random returns and noisy covariances. Computers
& Operations Research. 2022 Mar 1; 139:105631.

[27] Asgari H, Behnamian J. Multi-objective stock market portfolio selection using multi-
stage stochastic programming with a harmony search algorithm. Neural Computing and
Applications. 2022 Dec; 34(24):22257-74.

[28] Wu, Mu-En, Jia-Hao Syu, Jerry Chun-Wei Lin, and Jan-Ming Ho. "Portfolio
management system in equity market neutral using reinforcement learning." Applied
Intelligence (2021): 1-13.

[29] Mirjalili S, Lewis A (2016) The whale optimization algorithm. Advances in engineering
software. 95:51-67.

[30] Corsaro S, De Simone V, Marino Z (2021) Split Bregman iteration for multi-period
mean variance portfolio optimization. Applied Mathematics and Computation. 392:125715.

[31] Nesaz HH, Jasemi M, Monplaisir L (2020) A new methodology for multi-period
portfolio selection based on the risk measure of lower partial moments. Expert Systems with
Applications. 144:113032.

[32] Khan AT, Cao X, Li S, Hu B, Katsikis VN (2021) Quantum beetle antennae search: a
novel technique for the constrained portfolio optimization problem. Science China
Information Sciences. 64:1-4.

[33] Oprisor R, Kwon R (2020) Multi-period portfolio optimization with investor views
under regime switching. Journal of Risk and Financial Management. 14(1):3.

[34] Wang, Lu, Ferhana Ahmad, Gong-li Luo, Muhammad Umar, and Dervis Kirikkaleli.
“Portfolio optimization of financial commodities with energy futures.” Annals of Operations
Research 313, no. 1 (2022): 401-439.

[35] Setiawan, E. P. “Comparing bio-inspired heuristic algorithm for the mean-CVaR

portfolio optimization.” In Journal of Physics: Conference Series, vol. 1581, no. 1, p. 012014.
IOP Publishing, 2020.

47
[36] Shahid, Mohammad, Zubair Ashraf, Mohd Shamim, and Mohd Shamim Ansari. “A
novel portfolio selection strategy using gradient-based optimizer.” In Proceedings of
International Conference on Data Science and Applications: ICDSA 2021, Volume 2, pp. 287-
297. Springer Singapore, 2022.

48