Artificial Intelligence Review (2023) 56:3847–3886

https://doi.org/10.1007/s10462-022-10273-7

A brief review of portfolio optimization techniques

Abhishek Gunjan1 · Siddhartha Bhattacharyya2

Published online: 15 September 2022

© The Author(s), under exclusive licence to Springer Nature B.V. 2022

Abstract
Portfolio optimization has always been a challenging proposition in finance and manage-
ment. Portfolio optimization facilitates in selection of portfolios in a volatile market situ-
ation. In this paper, different classical, statistical and intelligent approaches employed for
portfolio optimization and management are reviewed. A brief study is performed to under-
stand why portfolio is important for any organization and how recent advances in machine
learning and artificial intelligence can help portfolio managers to take right decisions
regarding allotment of portfolios. A comparative study of different techniques, first of its
kind, is presented in this paper. An effort is also made to compile classical, intelligent, and
quantum-inspired techniques that can be employed in portfolio optimization.

Keywords Portfolio optimization · Statistical measures · Machine learning · Deep

learning · Reinforcement learning · Evolutionary techniques · Quantum computing

1 Introduction

In financial terms, a portfolio is a collection of assets/investments. Due to changing mar-

ket dynamics, diversification of investment is a crucial mechanism used to reduce risk on
investment, as proposed by Markowitz et al. Markowitz (1959). Based on degree of risk
and return, portfolio can be further categorized as (i) Aggressive Portfolio (ii) Defensive
Portfolio (iii) Income Portfolio (iv) Speculative Portfolio and (v) Hybrid Portfolio. Table 1
provides a brief description on the different types of portfolios.
During the buying and selling of the assets, a transaction cost is imposed by the brokers
or financial institutes as commission or other expenses Mansini et al. (2015). Henceforth,
an optimum number of asset selection is important to reduce the transaction cost. In order
to handle this situation, a proximal approach involving time penalization is proposed in
García-Galicia et al. (2019) which involves a three-step process viz., (i) a dynamic stock
price process (ii) transformation of solution as a continuous-time discrete-state Markov

* Siddhartha Bhattacharyya
[email protected]
Abhishek Gunjan
[email protected]
1
Christ Deemed to be University, Bangalore, India
2
Rajnagar Mahavidyalaya, Birbhum, India

13
Vol.:(0123456789)
3848 A. Gunjan, S. Bhattacharyya

decision processes Bäuerle and Rieder (2011) and (iii) a time penalization technique on
transaction cost. Selection of limited number of assets makes the investment more manage-
able and explainable Babaei and Giudici (2021). A sparse and robust portfolio selection
process is proposed in Lee et al. (2020) based on two step process viz., (i) develop a sparse
mean-variance portfolio selection model using semi-definite relaxation and (ii) extend
the model by using L2-norm regularization and worst-case optimization to formulate two
sparse and robust portfolio selection models.
Most of the methods used in portfolio selection proces suffers from estimation error and
hence there is a need to regularize the portfolio process Bruder et al. (2013). A lagrangian
regularization method is suitable for continuous time markov chain as proposed in Vazquez
and Clempner (2020). Based on the complexity of the problem, it is quite evident that any
portfolio management techniques requires a dynamic decision making process and hence a
reinforcement based learning techniques Wu et al. (2021) are highly suitable to effectively
manage the portfolios.
Portfolio management Cooper et al. (2001), hence involves dynamic decision process
where the portfolios are evaluated, selected, and prioritized based on financial objec-
tive and risk tolerance setup by the organization or client or institute Jeffery and Leliv-
eld (2004). Portfolio management requires careful SWOT (Strength-Weakness-Opportu-
nity-Threat) analysis across all the investments. The choice of right stock is done based
on different category of portfolio as described in Table 1. Some of the key elements of
portfolio management are (i) Asset Allocation (ii) Diversification and (iii) Rebalancing. as
mentioned in Table 3. Portfolio Management can be of two types viz., (i) Passive Portfolio
Management and (ii) Active Portfolio Management as mentioned in Table 4.
Further, the types of portfolio can be extended to (i) Transaction Cost Portfolio (ii)
Robust Portfolios (iii) Regularized Portfolios and (iv) Reinforcement learning based Port-
folios as mentioned in Table 2.
Portfolio management also requires optimization techniques to select the best portfo-
lio. Over the years, portfolio optimization techniques have also evolved from techniques
like mean-variance (MV) Markowitz (1959), variance with skewness (VwS) Samuelson
(1975), Value-at-Risk (VaR) Jorion (1997), Conditional Value-at-Risk Rockafellar and
Uryasev (2000), Mean-absolute deviation (MAD) Konno and Yamazaki (1991) and Mini-
max (MM) Young (1998) to more advanced heuristic and meta-heuristic based methods.
In recent years, evolutionary algorithms (EA) and Swarm Intelligence (SI) have become
popular in the field of portfolio optimization. Some of the SI based techniques used are
particle swarm optimization (PSO) Kennedy and Eberhart (1942), ant colony optimiza-
tion (ACO) Coloni et al. (1996), bacterial foraging optimization (BFO) Passino (2002),
artificial bee colony (ABC) Karaboga (2005), cat swarm optimization (CWO) Chu et al.
(2006), firefly algorithm (FA) Yang (2009), invasive weed optimization (IWO) Karimkashi
and Kishk (2010), bat algorithm (BA) Yang (2010) and fireworks algorithm (FA) Tan and
Zhu (2010). Other than the above mentioned techniques, quantum inspired approaches
have also become popular in the field of portfolio optimization.
The objective of this paper is to list different portfolio optimization techniques along
with their advantages and limitations. This will help the audience to have a compiled view
of the different state-of-art techniques and related advancements which has taken place in
recent years.
The remainder of the paper is organized as follows. Section 2 briefly lists the motivation
behind this article. Section 3 briefly lists different types of portfolios. Section 4 provides a
brief definition and different types of measures used in portfolio management. Section 5
provides a brief introduction of portfolio optimization and its relevance. Section 6 lists

13
A brief review of portfolio optimization techniques 3849

out the classical approaches employed in portfolio optimization. Section 7 lists the intel-
ligent approaches. The paper is concluded in Section 8. The organization of this paper is
described in Fig. 1.

2 Motivation

In every time bond project management, success is determined by two factors viz., "doing
the right thing" and "doing things right" Rämö (2002). The portfolio management tech-
niques hence require better allocation of stock to maximize return while keeping the risk
low. Portfolio management and optimization have always been a challenging problem and
have aroused interest among the researchers in the technical as well as financial domains
Haugh and Lo (2001). Despite multiple advancements done in new theories and computa-
tion power, portfolio optimization still remains a challenging problem to solve. Also, as per
the "Bureau of Labor Statistics", the employment of financial analyst including portfolio
managers is expected to increase by 5% from 2019 to 2029. Henceforth, there is a need for
better tools and mechanisms to perform portfolio optimization. Some of the key factors
which make it an area of interest includes
(i) dynamically changing market factors (ii) dynamically changing constraints like regu-
larity constraints, liquidity, taxes, transaction costs, management fees, etc., and (iii) need
for high computation power to determine the right distribution of selection in a timely
bound manner. The main contribution of this paper is that it presents a brief review of clas-
sical and intelligent approaches that can be employed for portfolio optimization. A brief
study of quantum-inspired based approaches are also touched upon, which opens new areas
for efficient and faster computing in this direction.

Fig. 1  Paper organization flow

13
3850 A. Gunjan, S. Bhattacharyya

3 Types of portfolios

Based on the degree of risk and return, the portfolio can be categorized into five cat-
egories viz., (i) Aggressive Portfolio (ii) Defensive Portfolio (iii) Income Portfolio (iv)
Speculative Portfolio and (v) Hybrid Portfolio. Aggressive portfolio aims for higher
returns and often undertake higher risks to achieve the objective. The preference is
given to the companies that are in initial growth stages. On the contrary, defensive port-
folio aims for minimal risk. The preference is given to the companies that provide daily
need products to ensure the risk is minimum in adverse conditions. An income port-
folio is somewhat similar to a defensive portfolio, but focuses on gain from dividends
or other recurring benefits. Few examples of income portfolio includes real estates,
FMCG, and other stable industries. The speculative portfolio aims for extremely high
risk and is often termed as gambling. Few examples of speculative portfolio includes
initial public offers (IPOs), investments into initial level of product research, and takeo-
ver targets. The hybrid portfolio aims to provide optimum return with optimum degree
of risk. It uses an amalgamation of different types of assets based on the degree of risks
and returns.
Furthermore, based on behavior the types of portfolio can be extended into four cat-
egories viz., (i) Transaction cost portfolio (ii) Robust portfolio (iii) Regularized portfo-
lio and (iv) Reinforcement learning portfolio. During buying and selling of assets, there
is transaction cost involved as transaction fee and charges. The transaction cost portfolio
aim to reduce the transaction cost by employing time penalization techniques. Contrary
to this, robust portfolio aims to reduce transaction cost in sparse and robust portfolio
selection process. Typically, an optimization process often leads to estimation error that
are handled in regularized portfolio. Furthermore, a changing market dynamics requires
continuous learning of market conditions that are handled in reinforcement learning
portfolio. A detail optimization technique to handle reinforcement learning portfolio is
listed in Section 7.4. The types of portfolios are listed in Tables 1, 2

4 Portfolio management

Portfolio in literal terms means a collection of financial investments in the form of

stock, bond, cash, assets, and other forms of commodities which tends to give some
form of return in future. A portfolio can be managed by an individual or a group of
financial institutes specialized in managing portfolio. Hence, building portfolio is one
of the key parameters in growing business. Proper analysis of portfolio can help a com-
pany to analysis if the business would receive more or less investments. It also helps to
analyze the overall strategy and focus of a company. In few cases, it has been observed
that companies are opening multiple business models at once, despite their inability to
execute them successfully Aversa et al. (2017). These factors must be analyzed by port-
folio managers while selecting the right assets.
At any given time t, a portfolio can be represented as
n
∑
P(t) = wj (t)sj (1)
j=1

13
A brief review of portfolio optimization techniques 3851

where, s1 , s2 , ..., sn are the securities and w1 , w2 , ..., wn are the weights or quantities associ-
ated with the securities. The weight at time t is represented as
$ invested in securities sj
wtj =
Total $ invested in portfolio Pt

Let us assume a list of securities as GOOGL, AMZN, AAPL with optimum weights as
3, 2, 5, then the corresponding portfolio at time t is represented as
Pt = 3 × GOOGL + 2 × AMZN + 5 × AAPL

Based on the market dynamics there could be extended investment or withdrawal of exist-
ing stocks. The decision involves allocation or re-allocation of stock and requires dynamic
opportunity, multiple goals and objective consideration. A good analyst should consider
multiple factors before making any allocation and de-allocation decisions Cooper et al.
(2001). An efficient portfolio management involves dealing with problems like (i) future
events and opportunities (ii) dynamic decision making capabilities (iii) investments made
could be in different stages making decision making more complex and (iv) limitations
in resources which are allocated. A portfolio manager typically work with four goal viz.,
value maximization, balancing, strategic direction, and right number of projects. Further,
portfolio optimization can be performed based on multiple parameters, as mentioned in
Table 5.
The following subsections list out different measures used in portfolio management. Sub-
section 4.1 briefly describes mean variance risk measure while subsection 4.2 describes
mean-absolute deviation. Further minimax is described in subsection 4.3 while lower partial
moments are described in subsection 4.4.

4.1 Mean variance risk measure

A revolutionary technique for portfolio selection as proposed by Markowitz Markowitz (1959)

is based on mean-variance. As per the proposal, mean-variance can be used for effective port-
folio selection when the objective is to (i) minimize variance for a given expected return and
(ii) maximize expected return for a given variance Mean-Variance (MV) Markowitz (1959)
analysis is very useful and easy to implement when the distribution function is not quadratic
in nature. Above all, mean-variance methodology is very fast. MV can also be employed
while tracking errors Roll (1992); Feldstein (1969). One of the major drawbacks of mean-
variance analysis is that it cannot be employed in high complex problems where the distribu-
tion function is polynomial in nature. A comprehensive guide on mean-variance is available in
Markowitz and Todd (2000). The objective of the MV model is to find the weight of the assets
that will minimize the variance at a given rate of return. The mathematical model is given by
Markowitz and Todd (2000)
n n
∑ ∑
minimize 𝜎ij xi xj (2)
i=1 j=1

rj xj ≥ 𝜌M0 ,
n
∑
subject to (3)
j=1

13
3852 A. Gunjan, S. Bhattacharyya

Also, 0 ≤ xj ≥ uj , j = 1, 2, ..., n where, 𝜌ij is the covariance between assets i and j, xj is the
amount invested in an asset j, rj is the expected return of asset j per period, 𝜌 is a parameter
representing the minimal rate of return required by an investor, M0 is the total amount of
fund and uj is the maximum amount of money which can be invested in asset j. The chal-
lenge in the MV model arises when it tries to compute covariance and hence cannot be
employed in large scale portfolio optimization problems.
Further, in large implementation, data is generally not normally distributed and there lies
asymmetry in the probability distribution. A mean-variance-skewness based measure thus
becomes more effective in such scenarios Samuelson (1975). In the traditional approaches, the
trade-off between risk and return is not explicit leading to less intutive investment decisions.
A stochastic based optimization algorithm is presented in Yin et al. (2002), which has been
used for a class of stock liquidation problems that are based on hybrid geometric Brownian
motion models. It allows regime changes by using continuous-time finite-state Markov chain.
A controlled regime-switching system Yang et al. (2015) has been used where the observation
is noisy. In such scenarios, the focus has been to minimizing the variance subject to a fixed
terminal expectation, and thereby employ a Wonham filter Borisov (2011) to convert the par-
tially observable system into a completely observable system. A continuous-time penalization
García-Galicia et al. (2019) has been employed when the market is characterized as arbitrage-
free that involves transaction cost and no short-selling of stocks. A lagrange Ito and Kunisch
(2008) multiplier approach has been proposed in García-Galicia et al. (2019) to solve convex
non-linear problems.
A sparse mean-variance portfolio selection model Lee et al. (2020) can be employed when
the assets are sparse and robust at the same time. A suitable portfolio selection is performed
by using two convex portfolio selection models to handle both sparsity and robustness in
the assets with dynamically changing markets. The proposed technique in Lee et al. (2020)
assumes that (i) the portfolio should not contain too many assets and (ii) the portfolio should
be robust.

4.2 Mean‑absolute deviation

The mean-absolute deviation (MAD) model as proposed in Konno and Yamazaki (1991)
Konno and Koshizuka (2005) can be employed for large scale and highly diversified portfolio
selection problems. A long term asset liability management (ALM) Bauer et al. (2006) model
and mortgage backed security Hayre (2002) model can also be employed where the invest-
ments are done for a longer period and the investments are made in very diversified portfo-
lios. A robust portfolio optimization using MAD Moon and Yao (2011) was applied on differ-
ent stock market data NYSE (2003). It has been observed that the model takes advantages of
reduction in computational complexity and provides optimum results. MAD is also observed
to outperform the mean-variance Markowitz (1959) based models. This model suffers a draw-
back, as it penalizes both positive and negative deviations. Mathematically, the model can be
represented as Konno and Yamazaki (1991)

13
A brief review of portfolio optimization techniques 3853

n n
∑ ∑
minimize w(x) = E ∣ Rj xj − E Rj xj ∣
j=1 j=1

E ∣ Rj ∣ xj ≥ 𝜌M0 , and
n
∑
subject to (4)
j=1
n
∑
xj = M 0 ,
j=1

0 ≤ xj ≤ uj , j = 1, 2, ..., n where Rj is the return of asset j, xj is the amount invested in asset

j, 𝜌 is a parameter representing the minimal rate of return required by an investor, M0 is the
total amount of fund and uj is the maximum amount of money which can be invested in an
asset j.

4.3 Minimax

Minimax (MM) Cai et al. (2004) Li et al. (2019) uses the minimum return as a measure of
risk. In scenarios where the assets returns are multivariate and normally distributed, then
both Minimax Cai et al. (2004) and mean-variance Markowitz (1959) leads to the same
result. Minimax Cai et al. (2004)Li et al. (2019) has certain advantages when the returns
are not normally distributed. Minimax Cai et al. (2004) is fast due to its property of linear
programming and can be employed for more complex models and constraints. One of the
disadvantages of Minimax Cai et al. (2004) is its sensitivity to outliers, and hence it cannot
be employed when the historical data is missing. Mathematically, the model can be repre-
sented as Cai et al. (2004)
max Mp

wj yj − Mp ≥ 0, t = 1, 2, ..., T,
N
∑
subject to
j=1

wj yj ≥ G,
N
∑
(5)
j=1

wj ≤ W,
N
∑

wj ≥ 0, j = 1, 2, ..., N
j=1

where, yij is the return on one dollar invested in a security j in a time period t, yj is the aver-
age return on security j, wj is the portfolio allocation to security j, Mp is the minimum
return on portfolio, G is the minimum level of return and W is the total allocation.

4.4 Lower partial moment

Lower Partial Moment (LPM) Nawrocki (1992) Brogan and Stidham (2008) uses a set of
moments to estimate downside risk in a portfolio. Based on continuous study, there could
be multiple such kinds of moments. Hence, as a part of LPM Nawrocki (1992), multiple

13
3854 A. Gunjan, S. Bhattacharyya

N such kinds of lower moments are analyzed. It is further used to calculate sortino ratio,
omega ratio, and kappa ratio Chen (2016). The model can be defined as Nawrocki (1992)
𝜏

∫−∞ (6)
LPM𝛼 (𝜏, Ri) = (𝜏 − R)𝛼 𝜕F(R)

where, 𝛼 is the degree of LPM, 𝜏 is the target return, R is the return, and 𝜕F(R) is the cumu-
lative distribution function of the asset return R.

5 Portfolio optimization

Portfolio optimization is the process of selecting the best combination of assets based on
pre-defined objectives. The objectives could be maximization of return or minimization
of risk, or both. Each optimization technique uses different portfolio measures as fitness
functions. In fact, due to large number of assets to choose from, manual selection of right
set of assets is difficult and hence requires advanced techniques as mentioned in Sections 6
and 7. Further, each optimization technique uses different portfolio measures as objective
functions to come up with an optimum combination of assets as mentioned in Table 5.
The optimization techniques are based on heuristics and do not require historical data for
optimization. This paper presents several optimization techniques using different portfolio
measures as listed in Table 5. Several optimization techniques can be used based on their
complexities, limitations, and advantages as highlighted in Tables 6, 7, 8, 9.

6 Classical approaches to portfolio optimization

Portfolio optimization is inevitably an important process to minimize risk and maxi-

mize return on investments. Henceforth, it has always been a challenging task to have a
proper mix of risks associated along with expected returns with limited number of assets
to be considered. Almost all the classical approaches like mean-variance Markowitz
and Todd (2000), variance with skewness Samuelson (1975), mean absolute deviation
Konno and Yamazaki (1991); Konno and Koshizuka (2005), minimax Cai et al. (2004)
Li et al. (2019), and lower partial moments Nawrocki (1992) mentioned in Table 5 are
based on asset and cardinality constraints Cesarone et al. (2011). There are numerous
approaches mentioned in the literature McNeil et al. (2015)Galai et al. (2001), which
mainly incorporate a combination of portfolio states Ray and Bhattacharyya (2015).
Value-at-Risk (VaR) Dowd (2007)Holton (2003)Jorion (2007) is a measure of, how the
market value of an asset is likely to reduce over a period of time under certain market
conditions. VaR Jorion (2007) requires determination of three parameters viz., (i) (ii)
time horizon (iii) confidence level and unit of VaR. While VaR Jorion (2007) is fairly a
simplistic technique, it does not consider risk measures. As mentioned in Ray and Bhat-
tacharyya (2017), a conditional Value-at-Risk (CVaR) Rockafellar and Uryasev (2002)
based technique can be employed to consider risk measures. The condition associated
with value-at-risk is the weighted average of VaR and losses exceeding value-at-risk
Krokhmal et al. (2002). VaR Jorion (2007) and CVaR Rockafellar and Uryasev (2002)
focus mainly on maximizing the expected return. Both VaR Jorion (2007) and CVaR
Rockafellar and Uryasev (2002) are efficient when the underlying risk factors are nor-
mally distributed, but face challenges for discrete distribution, making it non-convex

13
A brief review of portfolio optimization techniques 3855

Rockafellar (1970). Another challenge while using VaR Jorion (2007) is that it lacks
an analytical functional form and is generally a nonlinear, non-convex and non-smooth
function.
There are multiple smoothing techniques presented in Xu et al. (2007)Xu (2003)Xu and
Ng (2006) which can be employed while using VaR Jorion (2007) and CVaR Rockafellar
and Uryasev (2002) to get optimum results. While working with long-term investments,
there is some amount of uncertainty involved, which can be handled by using Monte Carlo
simulation Eckhardt (1987). The scenario based analysis can be further done using Mul-
vey (2001) and Hibiki (2001). Furthermore, an uncertainty set can be employed with ran-
dom interval Chen et al. (2011) as an optimization technique. VaR Jorion (2007) requires
a perfect knowledge of data, but in reality it is prone to errors. Hence, an estimation based
on wrong data points may be misleading and inaccurate. The technique proposed in Chen
et al. (2011) can be employed in such scenarios having high performance and low cost.
Another hybrid technique which combines mean-variance portfolio optimization technique
and Shefrin Statman’s behavioral portfolio technique is proposed in Das et al. (2010). The
proposed mental accounting (MA) Das et al. (2010) framework includes structure of port-
folios, definition of risk, and attitude towards risks. The proposed methods in Das et al.
(2010) result in better connection between portfolio production and consumption goals.
Value at Risk (VaR) Dowd (2007)Holton (2003)Jorion (2007) is one of the primitive
statistical approaches used for portfolio optimization. Both risk and return are correlated
with each other, and at times inversely related. High risks may lead to high returns and vise
versa. VaR is easily interpretable and at times self explanatory. VaR requires estimation of
three parameters viz., (i) time period (ii) confidence level and (iii) unit of value at risk.
The time period denotes the time horizon that is considered for analysis. Typical
buying and selling can be intraday, weekly, monthly, or yearly. Also, the period can be
taken as 1 day, 10 days, or any other arbitrary value. The confidence level is the interval
in which VaR Jorion (2007) would not exceed the maximum loss. Typically, the confi-
dence value is taken as 95% or 99% depending on the objective set for portfolio optimi-
zation. The unit of VaR Jorion (2007) refer to the quantity of stock, bond, other types of
assets to be considered while building the portfolio. A mathematical representation of
portfolio is given in equation 1.
VaR Jorion (2007) could be misleading at times as it does not consider the worst case
loss and cannot be applied in large portfolios. VaR Jorion (2007) is discrete in nature
and difficult to simulate. Hence, there is a need to look at scenario based approaches
like the Conditional Value at Risk (CVaR) Lim et al. (2011). Suppose f(x, y) represents
the percentile of loss or reward based on decision vectors x = (x1 , x2 , ..., xn ) and random
vectors y = (y1 , y2 , ..., yn , then VaR can be represented as 𝛼 percentile. VaR calculated
considering loss distributed 𝛼 is termed as the conditional value at risk. (CVaR) can be
of two types, viz., “upper CVaR” (CVaR+ ) or “lower CVaR” (CVaR−). In CVaR+ , the
expected loss exceeds VaR. Thus, (CVaR) can be represented as Lim et al. (2011)
CVaR = 𝜆VaR + (1 − 𝜆) CVaR+ , where 0 ≤ 𝜆 ≥ 1 (7)
As denoted by the expression, CVaR Rockafellar and Uryasev (2002) can handle extreme
losses by using dynamic weight 𝛼 derived from f(x, y) based on the decision and random
vectors. Also, CVaR Rockafellar and Uryasev (2002) can be employed in volatile scenar-
ios and helps in managing risk more efficiently as compared to VaR Jorion (2007). CVaR
Rockafellar and Uryasev (2002) suffers a major disadvantage that it cannot indicate the
maximum loss that can be incurred.

13
3856 A. Gunjan, S. Bhattacharyya

Some of the traditional techniques like hidden markov model Rabiner and Juang (1986),
efficient frontier Best and Hlouskova (2000), and monte carlo simulation Pedersen (2014)
are other popular classical approaches used in portfolio optimization.
The mean-variance Markowitz (1959) portfolio optimization is proposed as a paramet-
ric quadratic programming (PQP) in Zl et al. (2008). A bagging along with boosting can
also be employed to construct an ensemble of classifiers Derbeko et al. (2002). Further-
more, bagging helps in reducing the variance Bauer and Kohavi (1999)Bühlmann and Yu
(2002)Friedman and Hall (2007) and hence is an ideal choice by portfolio managers.

7 Intelligent approaches to portfolio optimization

Machine Learning (ML) Alpaydin (2020) based techniques are now very popularly used
due to their ability to learn from historical data. A huge volume of dataset can be quickly
churned and analyzed using these intelligent techniques. This section contains a list of
intelligent approaches that can be applied for portfolio selection and optimization.
The following subsections mention different intelligent approaches used in portfolio
optimization.

7.1 Bayesian approaches

During the 1770s, Thomas Bayes introduced the “Bayes Theorem” which is still relevant
and used in solving multiple complex problems using its inferential technique. Bayes’ theo-
rem describes the probability of the occurrence of an event based on prior knowledge of
conditions which might be related to the event. In the context of portfolio management,
it represents the probability of risk of buying or selling based on prior knowledge of the
market conditions.
According to Bayes’ theorem, the conditional probability is represented as Alpaydin
(2020)
p(Ck )p(x|Ck )
p(Ck |x) = (8)
p(x)

where, p(CK |x) is the posterior probability or the updated probability after taking into con-
sideration other parameters; p(Ck ) is the prior probability or the probability that is assigned
before any relevant evidence is taken into account; p(x|Ck ) is the likelihood; and p(x) is the
evidence from historical observations.
Portfolio optimization as proposed in Shenoy and Shenoy (2000) uses a Bayesian net-
work to model risk and return. It has been employed to visualize the relationship between
different variables in the model. Similar to financial analysts, it combines historical data
with qualitative information to draw a relationship between different portfolios. The tech-
nique proposed in Shenoy and Shenoy (2000) is well suited in situation where qualita-
tive and quantitative information are combined. Further Bayesian approaches can also
be applied in new product development, as proposed in Yang and Xu (2017). It uses a
three step process viz., (i) (ii) identify portfolio management factors and determine per-
formance criteria (iii) model relationship among the factors within a similar time frame
and (iv) develop a Bayesian network model. A fuzzy based Bayesian approach as proposed
in Bai et al. (2020) can be employed to assess risk and devise risk-reduction strategies.
The proposed technique takes project interdependency into consideration to determine the

13
A brief review of portfolio optimization techniques 3857

probability of all the risks by employing fuzzy based techniques. The error caused due
to subjectivity of expert knowledge is also considered in the proposed approach Bai et al.
(2020).
Bayesian network can also be used to identify risk transfer Guan et al. (2014)Guan and
Guo (2014) and helps in constructing an independent network of risks. This particularly
overcomes the drawback of greedy algorithms, which require a starting structure and other
risk parameters. The results as proposed in Guan et al. (2014)Guan and Guo (2014) shows
that the interdependent network of risks is an effective measure to determine the risk trans-
fer in projects and helps in computing the value of portfolio risks.

7.2 Support vector machine based approaches

Support vector regression (SVR) Drucker et al. (1997) has been used to determine the
quantity to buy and sell. SVR is a machine learning based approach where historical data
is fed to the SVR Drucker et al. (1997) to learn the pattern from the historical path. In
modern times, machine learning based techniques have become more popular due to their
ability to run faster on huge datasets. They also have the ability to learn changing market
positions of the stocks and can be employed for dynamic selection and rejection problems.
Given a training set of instance-label pairs (xi , yi ), i = 1, 2, ..., m where xi 𝜖 Rn and
yi 𝜖−1, 1, for a linearly separable case, the data points can be correctly classified by
⟨w.xi ⟩ + b ≥ + 1, for yi = +1 (9)

⟨w.xi ⟩ + b ≤ − 1, for yi = −1 (10)

Combining the above two equations gives
yi (⟨w.xi ⟩ + b) − 1 ≥ 0, ∀i (11)
where, w is the normal vector of the hyperplane and b is the bias value. The objective of
SVM Cortes and Vapnik (1995) is to find the most optimum hyperplane. SVM Cortes and
Vapnik (1995) has certain advantages over statistical based techniques like (i) (ii) ability
to learn from historical data (iii) can be used to build feedback to continuously learn (iv)
very effective in high dimensional space (vi) memory efficiency. SVM Cortes and Vapnik
(1995) works very well when the data points are linearly separable, but in the scenario
when there is overlap, it underperforms.
Further, SVM Cortes and Vapnik (1995) is also known as regularization formulation
where the loss in prediction and penalty is traded off by one or more regularization param-
eters. Henceforth, it can be used to trace the solution paths for 𝜖-support vector regression,
uniclass SVM, and quadratic regression as mentioned in Zl et al. (2008).

7.3 Neural network based approaches

A neural network or a network of neurons, also referred to as the artificial neural network
Haykin (2010), is used for solving complex computational and learning problems. A typi-
cal neural network has input layers, hidden layers, and output layers as shown in Fig. 2.
The hidden layers help to refine learning from historical data.
In a neural network, the neuron is an information processing unit which forms the basis
of any neural network model Haykin (2010). A nonlinear model of a neural network is

13
3858 A. Gunjan, S. Bhattacharyya

Fig. 2  Neural network with two hidden layers

Fig. 3  Neural network with nonlinear model

shown in Fig. 3. As shown in Fig. 3, there are three basic elements of any neural network
model viz., (i) (ii) synapses or connecting links with weights (iii) an adder for adding the
input signals and (iv) an activation function for limiting the amplitude of the output. Bias
bk is also introduced to reduce or increase the net input to the activation function based on
amplitude. A neuron k described in Fig. 3 is represented as Haykin (2010)
m
∑
uk = wkj xj (12)
j=1

and
yk = 𝜙(uk + bk ) (13)
where, x1 , x2 , ..., xm are the input signals; w1 , w2 , ...wm are the synaptic weights of the neu-
ron k. uk is the linear combiner output; bk is the bias; 𝜙(.) is the activation function and yk is
the output signal. The activation potential is represented as
v k = uk + b k (14)
Portfolio optimization as proposed in Ban et al. (2018) uses a performance-based regu-
larization (PBR) function, which helps in reducing the estimation error. PBR is considered

13
A brief review of portfolio optimization techniques 3859

for both mean-variance and mean-CVaR problems. A quadratic polynomial constraint is

introduced in Ban et al. (2018) for two convex approximation. Further, the result was tested
using k-fold based validation. Other than L1 and L2 regularization, there is also a need
to have some mechanism to penalize incorrect predictions. A comprehensive penalization
method is demonstrated in Eckstein and Kupper (2019) which can be used as a coherent
measure for risk Artzner et al. (1999). It has been found that recurrent neural networks
(RNN) are very efficient in understanding temporal dependencies. A comprehensive study
with a hybrid model using long-short term memory (LSTM) as proposed by Schmidhuber
et al. Schmidhuber and Hochreiter (1997), and Wiener et al. Saboia (1977), is presented in
Choi (2018). In the proposed approach, the ARIMA model has been used to filter linear
tendencies in data and passed the residual to the LSTM model. The proposed technique
in Choi (2018) was tested on 150 S &P500 stocks and was compared with walk-forward
optimization method Ładyżyński et al. (2013). Further, in the series of hybrid techniques,
there has been works performed using ensemble techniques using deep learning and rein-
forcement learning. The objective is to generate financial-model-free reinforcement learn-
ing framework with deep learning techniques to come up with better portfolio manage-
ment. The proposed technique uses (i) ensemble of identical independent evaluators (EIIE)
topology (ii) online stochastic batch learning (OSBL) and (iii) reward function Jiang et al.
(2017). The proposed technique used cryptocurrency and is examined in three back-test
experiments with trading period of 30 minutes. In all the three back-tests, the proposed
techniques seem to perform better than the integrated CNN (iCNN) (Jian and Liang, 2017).
An optimal reward using reinforcement learning based techniques can be achieved in port-
folio optimization, as proposed in Chaouki et al. (2020). A comparison of Sharpe ratio of
major neural network based techniques are mentioned in figure 4.
Further, a risk-sensitive multi-agent network (RSMAN) Park et al. (2022) has shown
better performance over deep neural networks (DNNs). The proposed framework in
Park et al. (2022) uses (i) (ii) risk-sensitive agents (RSAs) and (iii) risk adaptive portfo-
lio generator (RAPG) and hence can understand the market risk better than DNNs.

Fig. 4  Sharpe Ratio Comparison of Neural Network based appraoches Jiang et al. (2017)

13
3860 A. Gunjan, S. Bhattacharyya

Fig. 5  Reinforcement Learning

7.4 Reinforcement approaches

Reinforcement follows a mechanism of dynamic learning or weight based on action

taken by the agents. There is a continuous observation based on user actions and the
learning weight is further defined as shown in Fig. 5. Reinforcement based learning is
very popularly used in computer gaming and can be extended to portfolio optimization
problems where the weights can be learned dynamically.
There are four major elements in reinforcement learning viz., (i) Policy (ii) Reward
(iii) Value and (iv) Model. A policy defines the way the agents behaves in a certain sce-
nario. It requires mapping of the current state with the actions taken by the agent. Fur-
ther, based on the objective set and the agent’s reaction, positive and negative rewards
are given. The objective is to maximize the total reward. The distinction of good and
bad rewards is defined by the value function. The environment behaviour is modelled by
using appropriate model.
Reinforcement learning involves learning procedure through an agent. The agent
interacts with the environment and takes appropriate action which modifies the state.
The action taken by the agent leads to either reward or penalty. At any given point of
time t (where t=0, 1, 2, ...), st represents the set of all possible states (where st ∈ S rep-
resents the state of the agent). Also at ∈ A(st ) represents the action taken by the agent at
t where A(st ) is the set of possible actions in state st . On every action at , reward rt+1 ∈ R
is generated, and the agent moves to the next state st+1 Alpaydin (2020).
A state function is given by Alpaydin (2020)
v𝜋 (s) = E𝜋 [Gt |St = s] (15)
where, v is the value function with state s of policy 𝜋 . Gt is the expected return from the
given state St.
The reinforcement based technique Sutton and Barto (2018) works on maximizing
award to the agent. This technique can be certainly applied in portfolio optimization,
where the weight can be dynamically computed against static weights used in tradi-
tional approaches. Portfolio Management System (PMS) using a convolutional neural
network (CNN) and recurrent neural network (RNN) as proposed in Wu et al. (2021)
uses a novel reward function involving Sharpe ratio and evaluation technique to measure
the performance of the model. The proposed reward function further enhances the per-
formance and supports resource allocation for empirical stock trading. With reinforce-
ment learning (RL) there are two major branches which are developed viz., (i) Q-learn-
ing and (ii) policy gradient. While Q-learning is used to learn the optimal reaction and

13
A brief review of portfolio optimization techniques 3861

establish Q-table using Markovian circumstances, policy gradient is used to model opti-
mal reward.
An adaptive technique, as proposed in Almahdi and Yang (2017), extends the work
using recurrent reinforcement learning (RRL) Moody and Saffell (2001) and builds a var-
iable weight allocation scheme to maximize expected drawdown. The proposed method
fetches buy and sell signal and asset allocation by employing a risk adjusted objective func-
tion and calmar ratio Young (1991). Further, a deep reinforcement technique where invest-
ment decisions and actions are made periodically based on the present global objective,
is proposed in Yu et al. (2019). An interactive system in real-time is proposed in Yu et al.
(2019) along with other list of novel modules viz., (i) infused prediction module (IPM)
(ii) data augmentation module (DAM) and (iii) behaviour cloning module (BCM) which
can be implemented for both on-policy and off-policy reinforcement learning (RL). Fur-
ther, there are different optimization techniques viz., (i) deep deterministic policy gradient
(DDPG) (ii) proximal policy optimization (PPO) and (iii) policy gradient (PG) explored in
Liang et al. (2018) that have been used on datasets of China stock market. The proposed
technique in Liang et al. (2018) considers at optimum variable weight based portfolio allo-
cation under expected maximum drawdown Almahdi and Yang (2017) and risk-adjusted
profile as proposed in Galai et al. (2001). A comprehensive feature engineering technique
can be found in Li (2017) with a comprehensive guide on deep Q learning (DQL) Achiam
et al. (2019) with changing structure and hyperparameters in Mnih et al. (2015). In recent
times, continuous time García-Galicia et al. (2019) and continuous control Aboussalah
and Lee (2020) based techniques have also become popular within reinforcement based
approaches. The approach mentioned in García-Galicia et al. (2019) can be applied in the
context of continuous time reinforcement learning. It assumes that the process is a con-
tinuous-time discrete-state Markov chain with simple and dynamic constraints. Based on
the assumptions, a memory-less model i.e., a continuous-time Markov process (CTMP)
is proposed which changes state over time. Further, a multidimensional state space can
be addressed by Stacked Deep Dynamic Recurrent Reinforcement Learning (SDDRRL)
as proposed in Aboussalah and Lee (2020). It captures the latest market conditions and
then tweaks the portfolio accordingly. In addition to that, it uses (i) (ii) Gaussian Process
(GP) and (iii) Expectation Improvement (EI) as acquisition functions which make it more
dynamic and enable it to learn continuously changing market parameters. The problem of
vanishing gradient caused due to presence of memory gate is also handled in the proposed
technique Aboussalah and Lee (2020).
In recent times, crypto-currencies have become very popular and are faced with differ-
ent challenge due to erratic changes in price. Further, crypto-currencies are decentralized
and are affected by government policies and actions Betancourt and Chen (2021). In addi-
tion to that, crypto-currencies show higher volatility and henceforth require a better and
advanced techniques to be applied. Such kind of market space requires rapid incorpora-
tion of new assets, adapt to changing state, and take action Betancourt and Chen (2021).
A proximal policy optimization (PPO) Schulman et al. (2017) has been shown to perform
well while dealing with crypto-currencies and has been validated in three different setups
with variable episode lengths Betancourt and Chen (2021). In scenarios where the transac-
tion costs are dynamic, there are different novel modules proposed in Betancourt and Chen
(2021) viz., (i) formulation of trading system with variable number of assets (ii) optimum
management of transaction cost (iii) implementation and validation of proposed technique
using dataset of crypto-currencies.
Further, the erratic and complex behavior of market is determined by endogenous (like
law of demand and supply, interest rates, income, etc.) and exogenous factors (like social

13
3862 A. Gunjan, S. Bhattacharyya

media, change in government policies, etc.). A combinatory approach using a restricted

stacked autoencoder and convolutional neural network (CNN) is proposed in Soleymani
and Paquet (2020) based on (i) high level feature extraction containing eleven correlated
features and (ii) feature reduction using restricted stacked autoencoder to reduce training
time. Due to reduction in the training time, the proposed technique Soleymani and Paquet
(2020) can be employed for designing online learning and online portfolio management.
An extension of the work from Soleymani and Paquet (2020) is proposed in Soleymani
and Paquet (2021) which is based on a graph convolutional reinforcement learning called
DeepPocket. DeepPocket uses a restricted stacked autoencoder for feature extraction and
focuses on exploiting the time-varying interrelations between the financial instruments.
The interrelation is represented by a graph where the nodes are financial instruments and
the edges represent the correlation between financial instruments. Recently, algorithmic-
based hedging has become popular and requires a multi-period portfolio selection model.
A dedicated multi-agent-based deep reinforcement learning technique is proposed in Lin
et al. (2022) using a two-level nested agent. The multi-agent-based deep reinforcement
learning (MABDRL) framework Lin et al. (2022) is employed in online portfolio selec-
tion problem where each agent is equipped with multiscale convolutional neural networks
(MCNNs) Shi et al. (2019) and the ensemble of identical independent evaluators (EIIE)
topology Jiang et al. (2017) to learn risk shift behaviour. Further, a graph convolutional
based neural network combined with reinforcement learning can provide better explain-
ability to the relationship between assets and their corresponding companies Shi et al.
(2022). The relationship between the companies is represented as a relational graph convo-
lutional network (R-GCN) Schlichtkrull et al. (2018), designed for heterogenous graph and
is efficient in handling multi-relational data (Fig. 6).

7.5 Evolutionary approaches

Evolutionary computing Eiben et al. (2003) is a computational intelligence technique

inspired from the biological evolution process. Similar to natural selection, the fitness of an
individual in any given population is determined by how well they succeed in adapting and
achieving their goals. In the context of portfolio selection problem, an individual candidate
is selected based on its ability to efficiently seed as a future candidate in the evolution
process.
There are two major forces that form the fundamentals of evolutionary systems viz., (i)
variation operators (recombination and mutation) and (ii) selection procedure Eiben et al.
(2003). The variation operators involve the process of recombining and mutation leading
to an improvement in the fitness parameters also denoted as the adaptation process. This

Fig. 6  Genetic technique

13
A brief review of portfolio optimization techniques 3863

leads to an increase in the viability of an individual, which is further reflected in the num-
ber of offspring. This process leads to better adaptability to the dynamic environment. It
should be noted that during the selection procedure, at times even weak individuals can
be selected. In other words, both recombination and mutation processes are stochastic in
nature.
Combinatorial Optimization Problems (COPs) Ausiello et al. (2012) have been identi-
fied as one of the NP-hard problems and remain an active area of research Wolsey and
Nemhauser (1999). Broadly there are two main approaches for solving COPs Cappart et al.
(2020) viz., (i) extract scenario based techniques and (ii) heuristic based techniques. As
mentioned in Section 7.4, reinforcement based techniques are best fit to handle such sce-
narios, but come with constraints while dealing with large datasets. This led to the explora-
tory works using evolutionary based approaches. An evolutionary based technique like the
Genetic Algorithm (GA) Whitley (1994) can be employed in optimization problems. Typi-
cally, any GA Whitley (1994) based technique works on the five states viz., (i) initialization
(ii) evaluation (iii) selection (iv) crossover, and (v) mutation Mirjalili (2019).
Particle swarm optimization (PSO) Kennedy and Eberhart (1942)Kennedy and Eberhart
(1995) first intended to simulate social behaviour as a representation of the movement of
organisms in a bird flock or fish school, can also be employed in volatile market condition.
PSO Kennedy and Eberhart (1995)El-Shorbagy and Hassanien (2018) is a heuristic based
technique that considers each member in the population as a particle. PSO Kennedy and
Eberhart (1995)El-Shorbagy and Hassanien (2018) uses the technique of “gbest neighbor-
hood topology” as proposed by Kennedy et al. Cura (2009). The underlying principle is
that each particle remembers its best previous position and the best previous position vis-
ited by any particle in the whole swarm of particles. In short, a particle moves towards the
best particle and the best previous position. Let us suppose that there are N dimensions,
where each dimension represents an asset for each portfolio. Then each particle includes
proportion variables xpi and decision variables zpi such that the proportion variables,
xpi (p = 1, 2, ..., P) where, P is the number of particles in the swarm. Also, the total number
of dimensions that a particle owns will be 2 × N . PSO Kennedy and Eberhart (1995)El-
Shorbagy and Hassanien (2018) is robust and can be applied for determining non-smooth
global optimization problems and seems to perform better than other heuristic based tech-
niques. It is also found to give high quality results and is less computationally intensive.
One of the challenges with PSO is that it lacks mathematical foundation and is difficult to
apply in real life scenarios.
The “cardinality constrained mean-variance (CCMV) Li et al. (2006) model" as pro-
posed by Chang et al. Chang et al. (2000) and Fernandex and Gomez Fernández and
Gómez (2007) has been employed for portfolio optimization by using PSO Kennedy and
Eberhart (1942). The results have been compared with genetic algorithm (GA) Chang
et al. (2000)Oh et al. (2005)Yang (2006), tabu search (TS) Chang et al. (2000), simulated
annealing (SA) Chang et al. (2000)Crama and Schyns (2003)Derigs and Nickel (2004),
neural networks Fernández and Gómez (2007), and other techniques like heuristic Mansini
and Speranza (1999), meta-heuristic Derigs and Nickel (2003), and hybrid heuristic based
approaches Schlottmann and Seese (2004). For comparison and benchmarking, the weekly
prices from March 1992 to September 1997 have been used from Seng (1992). Based
on the experiments conducted, it is observed that PSO Kennedy and Eberhart (1995)El-
Shorbagy and Hassanien (2018) performs better than other heuristic based approaches. A
further study on the cardinality constrained based portfolio optimization can be found in
Salahi et al. (2014) where the quadratic programming (QP) Frank and Wolfe (1956) prob-
lem Anagnostopoulos and Mamanis (2011a) is formulated as a mixed-integer quadratic

13
3864 A. Gunjan, S. Bhattacharyya

problem (MIQP) Lazimy (1982) by adding cardinality and quadratic constraints. On com-
paring integer and categorical particle swarm optimization (ICPSO) Goodman et al. (2016)
and improved harmony search (IHS) algorithm Mahdavi et al. (2007), is observed to per-
form much faster than ICPSO Goodman et al. (2016) on large data sets. For comparison
and benchmarking, the weekly prices from March 1992 to September 1997 have been used
from Seng (1992). A detailed experiment for constrained portfolio optimization using PSO
Kennedy and Eberhart (1995)El-Shorbagy and Hassanien (2018) is conducted in Zhu et al.
(2011) on the Shanghai Stock Exchange 50 Index (SSE 50 Index) and it is found that PSO
Kennedy and Eberhart (1995)El-Shorbagy and Hassanien (2018) outperforms GA and
VBA (Visual Basic Application). A detailed heterogeneous multiple population particle
swarm optimization (HMPPSO) algorithm is proposed in Yin et al. (2015) where the whole
population is divided into smaller sub-populations and different variants of PSO Kennedy
and Eberhart (1995)El-Shorbagy and Hassanien (2018) are applied on each sub-popula-
tion. For comparison and benchmarking, the weekly prices from March 1992 to September
1997 Seng (1992) have been used. On comparing with other PSO Kennedy and Eberhart
(1995)El-Shorbagy and Hassanien (2018) variants, HMPPSO Yin et al. (2015) seems to be
more effective and robust and could be employed on high dimensional problems.
On the other hand, ant colony optimization (ACO) Forqandoost Haqiqi and Kazemi
(2011) is a probabilistic model designed to select the best path designed based on pher-
omone-based communication of biological ants. ACO Forqandoost Haqiqi and Kazemi
(2011) can be employed to solve the minimum cost problem, a situation which involves
multiple nodes and non-directed arcs. The objective is to establish a minimum cost path
between the source and destination. ACO Forqandoost Haqiqi and Kazemi (2011) works
on both forward and backward modes and simulates forward movement of ants when they
move from nest to food and backward movement when they move from food to nest. The
forward movement typically employs probabilistic based selection while backward move-
ment is performed based on the deposited memory, which further helps to avoid or reduce
loops. The selection made in the forward movement is also done based on historical knowl-
edge deposited with multiple to and fro ant movements. Typically, the selection of frag-
ment I from K possible choices is done based on Forqandoost Haqiqi and Kazemi (2011)
𝜏
Probi = ∑k i (16)
1 𝜏k

where, 𝜏k is the amount of pheromone and k is the kth fragment. The evaluation is per-
formed based on the evaporation and pheromone deposit phases given by
(t + 1) = 𝜏i (t)(1 − 𝛾) + 𝛿i (17)
where, 𝛾 is the evaporation rate and 𝜏i is the amount of pheromone in the fragment i.
ACO Forqandoost Haqiqi and Kazemi (2011) is employed where there is a dynamic
change in path, distance, and other parameters. Since ACO is based on probabilistic distri-
bution, hence a significant change in topology can make the selection process slow.
An extension of ACO Forqandoost Haqiqi and Kazemi (2011) is proposed in Deng
and Lin (2010) to solve the cardinality constrained Markowitz mean-variance portfolio
optimization (CCMPO) problem. For comparison and benchmarking, the weekly prices
from March 1992 to September 1997 have been used from Seng (1992). The results
show that ACO Forqandoost Haqiqi and Kazemi (2011) is more robust and effective
than the standard PSO Kennedy and Eberhart (1995)El-Shorbagy and Hassanien (2018)
for low risk investment scenarios. The proposed technique uses three major components

13
A brief review of portfolio optimization techniques 3865

viz., (i) solution construction (ii) constraint satisfaction and (iii) pheromone update
Deng and Lin (2010). Further, a clustering based technique along with ACO Forqan-
doost Haqiqi and Kazemi (2011) is used to perform portfolio optimization Rezani et al.
(2020). Clustering using k-means++ Bahmani et al. (2012) is used to diversify the port-
folio and suitable stocks and their weights are chosen ACO Forqandoost Haqiqi and
Kazemi (2011). The three stages viz., (i) coefficient allocation (ii) quality assessment
and (iii) updating weight help in dynamically determining the right stock. Experiments
conducted using one year data from SP500 (2016) is found to give the best results on
cluster sizes of 8 where the proposed technique Rezani et al. (2020) gave the highest
average return and Sharpe ratio. An average entropy based ACO Forqandoost Haqiqi
and Kazemi (2011) is proposed in Li et al. (2012) where the parameters of the algo-
rithm have been adjusted adaptively. The adaptive behaviour tends to avoid stagnation
behaviour. ACO Forqandoost Haqiqi and Kazemi (2011) is further extended to solve
multi-criteria optimization problem with population-based ACO (PACO) Guntsch and
Middendorf (2003) based on the multi-colony ACO approach as proposed by Mari-
ano and Morales Mariano and Morales (1999). However, most of the above techniques
work on a single objective of either maximizing the return or minimizing the risk. Thus,
these techniques are unable to deal with conflicting objectives of both maximization of
returns and minimization of risks. Due to the above limitation, multi-objective evolu-
tionary techniques Van Veldhuizen and Lamont (1998) have gained popularity in the
recent past due to their ability to find multiple pareto-optimal solutions Lin (1976) out
of conflicting objectives.
A comprehensive study on Multi-objective Evolutionary Algorithms (MOEAs) Zhou
et al. (2011) viz., Non-dominated Sorting Genetic Algorithm II (NSGA-II) Yusoff et al.
(2011), Pareto Envelop-based Selection Algorithm (PESA) Corne et al. (2000) and
Strength Pareto Evolutionary Algorithm 2 (SPEA2) Zitzler et al. (2001), is available in
Anagnostopoulos and Mamanis (2011b). It is observed that the performance using these
techniques is independent of the risk function used in design. Both return and risk can
be simultaneously optimized using a multi-objective algorithm, which makes is it more
efficient than other techniques. A classification and genetic based technique is presented
in Guennoun and Hamza (2012). The proposed technique uses minimum value-at-risk
and maximum Value (MinVaRMaxVaL) using a two stage process using (i) k-means
Alpaydin (2020) and (ii) dynamic optimization algorithm. Further, two more variants
of the multi-objective particle swam optimization (MOPSO) Reyes-Sierra and Coe-
llo (2006) algorithm for portfolio optimization are reported in Babaei et al. (2015). A
comparative study between MOPSOs Reyes-Sierra and Coello (2006), NSGA-II Yusoff
et al. (2011), and SPEA2 Zitzler et al. (2001) is presented in Babaei et al. (2015). A
novel mean-VaR optimization technique using a univariate Generalized Auto Regressive
Conditional Heteroscedasticity (GARCH) model is proposed in Ranković et al. (2016).
The proposed work in Ranković et al. (2016) considers both portfolios with fixed weight
and portfolios with fixed holdings of assets. An efficient learning-guided hybrid multi-
objective evolutionary algorithm (MODE-GL) is proposed in Lwin et al. (2017), which
can be used in constraints like cardinality, quantity, reassignment, round-lot and class
constraints Lwin et al. (2017). A mutation based technique, using multi-objective evo-
lutionary algorithms (MOEAs) proposed in He and Aranha (2020), can be employed to
find more than one interesting solution in a single run. A comparison of mean return
error of different evolutionary techniques are mentioned in Fig. 7.

13
3866 A. Gunjan, S. Bhattacharyya

Fig. 7  Mean Return Error of different Evolutionary Techniques Cura (2009)Salahi et al. (2014)Yin et al.
(2015)

Table 1  Types of portfolios

Sl. No. Portfolio Description

1. Aggressive Portfolio Aims for higher returns and undertakes high risks
2. Defensive Portfolio Aims for minimum risk and gives minimum return
3. Income Portfolio Very similar to defensive portfolio, but focuses on gains from
dividends or other types of recurring benefits
4. Speculative Portfolio Aims for extremely high risk. Somewhat similar to gambling
5. Hybrid Portfolio Aims to provide optimum return with optimum degree of risk

7.6 Approaches based on quantum computing

In recent years, lots of effort have been put on trying hybrid based approaches to make
portfolio optimization more realistic. It has been found that quantum and quantum-inspired
computing techniques can help solve difficult optimization problems Orus et al. (2019).
During the early twentieth century, Newton and Maxwell dominated the laws of physics
and were assumed to be correct. Later it was however noted that there are certain flaws in
these techniques which led to the formulation of a new mathematical framework like quan-
tum computing Kaye et al. (2007).

13
 Table 2  Extended types of portfolio management
Sl. No. Portfolio Description
A brief review of portfolio optimization techniques

1. Transaction Cost Portfolio García-Galicia et al. (2019) Aims to reduce the transaction cost by using time penalization techniques.
2. Robust Portfolio Lee et al. (2020) Aims to reduce transaction cost in sparse and robust portfolio selection process.
3. Regularized Portfolio Vazquez and Clempner (2020) Aims to reduce estimation error
4. Reinforcement Learning Portfolio Aims to continuously learn the changing market condition by applying several
reinforcement learning approaches.
3867

13
 3868

13
Table 3  Key elements of portfolio management
Sl. No. Key element Description

1. Asset Allocation An effective portfolio management involves proper asset allocation, keeping short term and long term
objectives in mind. This requires proper understanding of asset, as all the assets do not move similarly.
This also requires proper mix of asset selection.
2. Diversification Distributing risk and return within an asset is an important factor. Due to changing market dynamics, it is
difficult to know exactly about any subset of an asset. Thus, diversification becomes an important key
element in portfolio management.
3. Rebalancing Rebalancing is the process of bringing back the allocated assets to the desired asset mix. The process of
rebalancing involves selling high-priced assets and putting that money in buying less-priced assets.
A. Gunjan, S. Bhattacharyya
A brief review of portfolio optimization techniques 3869

Table 4  Types of portfolio management

Sl. No. Type of portfolio management Description

1. Passive Management This involves long term strategy where the assets are allocated for
long term.
2. Active Management This involves active participation where the assets are bought and
sold regularly to gain maximum return.

Quantum computing (QC) is a part of quantum physics, where its inherent dynamics
are governed by the Schrödinger equation (SE) Talbi et al. (2006)McMahon (2007). On
multiple experiments, QC is shown to give better performance on complex and NP-hard
problems which require large solution space. Quantum machines are represented by a wave
function in the Hilbert space H. The Hilbert space, H can also be imagined as an extension
of a two-dimensional or a three-dimensional space to spaces with finite or infinite number
of dimensions.

7.6.1 Qubit

Qubit is the smallest unit of information in quantum computing McMahon (2007) repre-
sented as �0⟩ and �1⟩, also known as two-state qubit vectors in quantum computing. The
qubit vectors are also represented as McMahon (2007)
� � � �
1 0
�0⟩ = and�1⟩ =
0 1

Every notation in the above equation is referred to as the complex conjugate transpose of
each other.

7.6.2 Quantum superposition principle

The quantum superposition principle can be understood as the linear combination of quan-
tum vectors represented in two states as McMahon (2007)
�𝜓⟩ = 𝛼�0⟩ + 𝛽�1⟩

where, 𝛼 and 𝛽 are complex numbers which must satisfy the condition

∣ 𝛼 ∣2 + ∣ 𝛽 ∣2 = 1

where, 𝛼 2 and 𝛽 2 are the probabilities of measuring the basis states �0⟩ and �1⟩, respectively.
Thus, a qubit can be represented as a superposition of two basis states viz., �0⟩ and �1⟩
implemented using quantum gates (Q-gates). Thus, these basis states exist simultaneously
until a quantum measurement destroys the superposition. Hence, for n number of qubits,
the number of states in a quantum machine is 2nDiVincenzo (1998)Han and Kim (2002).

7.6.3 Quantum‑inspired metaheuristic algorithms

Metaheuristic-based techniques are widely perceived optimization techniques to solve

complex and large computational problems. Quantum-inspired metaheuristic techniques

13
 3870

13
Table 5  Portfolio optimization analysis models
Sl. No. Model Advantages Limitations

1. Mean-Variance Markowitz and Todd (2000) Should be used when the risk parameters are small. Cannot be used in case of high risk parameters due to
Used when the return distribution is compact and computational problems caused due to higher order
portfolio decisions are made frequently. polynomial functions. Can be applied when portfolio
returns are normally distributed.
2. Variance with skewness Samuelson (1975) Can be employed where the portfolio returns are not Optimization can ensure only local optimal solution.
normally distributed
3. Value-at-Risk Jorion (1997) Easy to understand the level of risk. Applicable to Could be misleading. Does not measure worst case
large number of assets. loss. Difficult to employ in large portfolios.
4. Conditional Value-at-Risk Rockafellar and Uryasev Can handle extreme loss by using weighted average. CVaR does not indicate the maximum loss that can be
(2000) Can be used in volatile scenario. Risk management incurred.
can be performed efficiently.
5. Mean-absolute deviation (MAD) Konno and MAD is a linear program and hence relatively fast, Can lead to calculation risks caused due to skipping
Yamazaki (1991) unlike mean-variance which is quadratic. Used of covariance matrix. Penalizes both negative and
when asset returns are multivariate and normally positive deviations.
distributed.
6. Minimax Young (1998) Has logical advantages when the returns are not Sensitive to outliers. Cannot be used when historical
normally distributed. Similar to MAD, it is a linear data is missing.
program and hence faster. Can be used for more
complex models and constraints.
7. Lower Partial Moment Nawrocki (1992)Brogan and Suits investors’ perception about risk. Undesirable Computation of portfolio risk is tedious. Sensitive to
Stidham (2008) downside and desirable upside deviations are sepa- outliers.
rated. Penalizes only downside deviations.
A. Gunjan, S. Bhattacharyya
 Table 6  Comparative study of classical and intelligent approaches
Sl. No. Methods Advantages Limitations

1. Statistical Based Approaches Easy to implement. Good for smaller data sets and time peri- Does not consider multiple features into account.
ods. Easily interpretable
2. Regression Based Approaches Simple, interpretable, and easy to implement. can learn itera- Efficiency is highly dependent on the learning rate chosen.
tively from historical data.
3. Bayesian Based Approaches Can be easily extended, unlike other machine learning based
No universal method to construct a network from data. While
techniques. Can be employed for reasoning based problems.
training, it is not assured that all the training patterns are used
A brief review of portfolio optimization techniques

while training.
4. Neural Network Based Approaches Better learning model as compared to other traditional tech- Should be used when dealing with large datasets. Not self-
niques. Fault tolerance, distributed memory and at times can explanatory and absence of specific rules to determine the
work with inadequate knowledge. structure of network.
5. Reinforcement Based Approaches Can solve complex optimization problems. Learning model is Should not be used for simple problem and on smaller data set.
very similar to learning patterns of humans. Can be used to Too much reinforcement can adversely affect learning. The
adaptively learn based on changing states. curse of dimensionality can limit the learning.
6. Evolutionary Based Approaches Conceptually simple, and outperforms classical techniques by Higher computation required. Results at times are not self-
imposing non-linear constraints and non-stationary condi- explanatory and reproducible.
tions. Ability to parallelism and robust to dynamic changes.
7. Quantum-based Approaches Adding qubits can increase the storage exponentially and are The energy required by quantum computer is much larger than
useful to solve very complex compute extensive problems. traditional computers. Still there is a lot of unknowns as this
Faster as compared to any other methods. is an ongoing area of research.
3871

13
 Table 7  Comparative study of different state-of-art approaches
3872

Sl. No. Methods Advantages Limitations

13
1. Value-at-Risk (VaR) Jorion (1997) Simple and easy to implement. Works well when the Difficult to use in real-life scenario as the return is not
return is normally distributed and there is the pres- always normally distributed.
ence of historical data.
2. Conditional Value-at-Risk (CVaR) Rockafellar and Risk management with CVaR can be done better CVaR mainly focus on risk and does not represent
Uryasev (2002) than VaR. CVaR also accounts for loss exceeding maximum loss that can be incurred.
VaR.
3. Particle swarm optimization (PSO) Kennedy and Simple to implement, robust, less computation Can be difficult to specify initial design parameters.
Eberhart (1942) power, can be used to run in parallel computing. Can converge prematurely, especially in case of
complex problem. Cannot handle discrete problem.
4. Integer and categorical particle swarm optimization Can be used for both integer and categorical values. Assumes that the variable are independent of each
(ICPSO) Goodman et al. (2016) Can be used when there is no natural ordering other. More sensitive to bias.
to solution value. Can be employed to handle a
discrete problem.
5. Heterogeneous multiple population particle swarm Can be employed with problem having high dimen- Requires large dataset to be applied. Can be applied
optimization (HMPPSO) Yin et al. (2015) sions. Efficient over PSO and multiple variants of when large population can be split as mutually
PSO can be employed in each sub-population. exclusive partitions.
6. Ant Colony Optimization (ACO) Deng and Lin Can adapt to dynamic changes. Can search optimum Time of convergence is not guaranteed. Probabilistic
(2010) solution in parallel. Convergence is guaranteed. distribution can change in very iteration.
7. Population-Based ACO (PACO) Guntsch and Mid- Combines the power of population and probabilistic Requires selection of prerequisite parameters. More
dendorf (2003) based approach. Better convergence as compared efficient when population is mutual exclusive.
to traditional ACO. Can be used to solve multi-
criteria problem.
8. Support Vector Machine (SVM) Alpaydin (2020) Easy to understand and implement. Effective in high Not suitable for large dataset. Works well when
dimensional space. Requires less computation there is a clear margin of separation. Performance
power. degrades when there is noise in data.
9. Bayesian-Based approach Karatzas and Zhao (2001) Good at combining prior information in building No correct way to choose prior. The result at times are
strong decision making framework. Inference not self-explanatory.
drawn are conditional based on circumstances.
10. Long-Short term memory (LSTM) Choi (2018) Can be employed to learn lengthy time period Hard to implement. Suffers from gradient exploding
dependencies. and vanishing problem.
A. Gunjan, S. Bhattacharyya
 Table 7  (continued)
Sl. No. Methods Advantages Limitations
11. Integrated CNN (iCNN) Chaouki et al. (2020) Model free approach. Can be employed in absence of Initial allocation could lead into loss before yielding
historical knowledge. profit.
A brief review of portfolio optimization techniques
3873

13
 3874

Table 8  Comparative study of different state-of-art approaches

Sl. No. Methods Advantages Limitations

13
12. Recurrent reinforcement learning (RRL) Moody and Used while optimizing performance criteria like Cannot be used in immediate estimation scenario.
Saffell (2001) differential Sharpe ratio and differential down- Initial allocation could lead into loss before yielding
side deviation ratio. Discover investment policies profit.
dynamically. Avoids Bellman’s curse of dimension-
ality. Used in discrete action space.
13. Stacked Deep Dynamic Recurrent Reinforcement Employed in continuous action is required in multi- A detailed and comprehensive study with neural
Learning (SDDRRL) Aboussalah and Lee (2020) dimensional state space. Requires updated market based technique is missing. Computationally large.
information to rebalance portfolio. Does not Requires current market state for better prediction.
require any time series predictions.
14. Multi-objective evolutionary algorithms (MOEAs) Could be used to solve a complex multi-objective Lack of reproducibility. Parameter tuning and control-
He and Aranha (2020) problem. Used in high dimensionality search space. ling is difficult.
Used with search space is non-linear, non-differen-
tial, non-continuous, and non-convex.
15. Non-Dominated Sorting Genetic Algorithm II Characteristics of fast non-dominated sorting Lack of reproducibility. Parameter tuning and con-
(NSGA-II) Yusoff et al. (2011) approach, fast crowded distance estimation proce- trolling is difficult. Requires careful selection of
dure and simple crowded comparison operator. parameters.
16. Multi-objective particle swam optimization Can be employed for multiple-objective scenario. Can be difficult to specify initial design parameters.
(MOPSO) Reyes-Sierra and Coello (2006) Uses less computation power as compared to Can converge prematurely, especially in case of
evolutionary techniques. complex problem. Cannot handle discrete problem.
17. Generalized auto regressive conditional heterosce- Used when volatility of return of groups of stocks More sensitive to the frequency of data used. Should
dasticity (GARCH) Ranković et al. (2016) with large number of observation need to be evalu- be used while dealing with very large time series
ated. data.
18. Quantum-inspired tabu search (QTS) Chou et al. Performs better than other heuristic algorithms in Not efficient with increase in quantity of items.
(2011)Chou et al. (2014) optimization problem without premature conver-
gence. Can be used to solve knapsack problem in
shorter time. Can avoid the problem of over-fitting.
Flexible, profitable and stable. Performs well in
upward trends.
A. Gunjan, S. Bhattacharyya
 Table 9  Comparative study of different state-of-art approaches
Sl. No. Methods Advantages Limitations

19. Multi-objective quantum-inspired tabu search Can be employed while dealing with multiple Need to be further evaluated with multiple different
(MOQTS) Chou et al. (2014) objectives. Flexible and profitable. Can run more financial parameters.
A brief review of portfolio optimization techniques

simulation with different parameters.

20. Quantum-inspired Firefly Algorithm with Particle Converges faster than PSO and other firefly based No experimental finding in portfolio optimization
Swarm Optimization (QIFAPSO) Zouache et al. techniques. space.
(2016)
21. Quantum-inspired acromyrmex evolutionary algo- Can be employed to find an efficient global optimiza- Cannot be used for multiple objective scenarios.
rithm (QIAEA) Montiel et al. (2019) tion method for complex systems. Performs better
than quantum-inspired evolutionary algorithms.
Has higher average accuracy and lower standard
deviation.
3875

13
3876 A. Gunjan, S. Bhattacharyya

combine the power of quantum computing and metaheuristics and have been shown to
perform better than the classical counterparts Karmakar et al. (2017) in terms of exhib-
iting faster convergence and having better exploration and exploitation capabilities. The
quantum-inspired metaheuristic algorithms emulate the principles of quantum mechan-
ics. These algorithms resort to qubit based encoding Nielsen and Chuang (2001)DiVin-
cenzo (1998) of the solution space and are characterized by quantum-inspired versions of
the underlying operators. Due to their ability to solve complex and large computational
problems, these quantum-inspired metaheuristics are widely used in both constrained and
unconstrained problems Rebentrost and Lloyd (2018)Orus et al. (2019).
Of late, several incarnations of these quantum-inspired metaheuristics have been
evolved to yield quantum-inspired versions of the classical metaheuristics including the
genetic algorithm Davis (1991), tabu search Glover and Laguna (1998), PSO Kennedy and
Eberhart (1995), and ACO Dorigo et al. (2006) to work in quantum space. Typical exam-
ples include the quantum-inspired genetic algorithm Narayanan and Moore (1996)Han
et al. (2001)Saad et al. (2021), quantum-inspired tabu search Chiang et al. (2014)Chou
et al. (2014)Kuo and Chou (2017), quantum-inspired PSO Alvarez-Alvarado et al. (2021)
Agrawal et al. (2021), quantum-inspired ACO Dey et al. (2018)Mohsin et al. (2021) to
name a few.

7.6.4 Quantum‑inspired metaheuristic‑based approaches

Portfolio optimization employs several techniques to maximize the profit and reduce risks.
Henceforth, trading rules can be designed which can be employed to take decisions on
when to sell and buy based on rules identified. A combination of trading rules can be
termed as a trading strategy, which can be formulated as an optimization problem. A novel
method of optimization is proposed in Kuo et al. (2013) where a quantum-inspired Tabu
Search (QTS) Chou et al. (2011)Chou et al. (2014) algorithm is used to find the optimum
combination of trading rules along with a sliding window to avoid over-fitting. The pro-
posed method is evaluated with the Buy & Hold method and is found to give better results.
Further, an extension of QTS Chou et al. (2011)Chou et al. (2014) called the multi-objec-
tive quantum-inspired tabu search (MOQTS) Chou et al. (2014) is proposed. While QTS
Chou et al. (2011)Chou et al. (2014) works on fewer rules, MOQTS can handle more num-
ber of rules and is observed to outperform QTS Chou et al. (2011)Chou et al. (2014) in
terms of profit and successful transaction rates.
A brief study on “Quantum-inspired Firefly Algorithm with Particle Swarm Opti-
mization (QIFAPSO)” is available in Zouache et al. (2016) applied on some discrete
optimization problems. The results from QIFAPSO Zouache et al. (2016) are compared
with two classes of algorithms viz., quantum inspired algorithms and PSO inspired
algorithms on high dimensional knapsack instance and is found to outperform by a big
margin. Further, a dynamic portfolio optimization technique as proposed in Mugel et al.
(2020) implemented (i) D-Wave Hybrid quantum annealing (ii) Variational Quantum
Eigensolver (VQE) (iii) VQE Constrained on IBM-Q with the quantum inspired Ten-
sor Network (TN) algorithm. Based on the tests conducted in Mugel et al. (2020), it is
found that the D-Wave hybrid quantum annealing, and tensor network are able to handle
large systems of approximately, 1272 fully-connected qubits without hitting computa-
tional limitations. It is also observed that D-Wave is very fast, whereas tensor network
provides better portfolio optimization at a cost of higher computation time. Further,
quantum-inspired acromyrmex evolutionary algorithm (QIAEA) as proposed in Montiel

13
A brief review of portfolio optimization techniques 3877

et al. (2019) is inspired from the acromyrmex ant species also known as leaf-cutter ants.
On several experiments, it is shown as QIAEA gives the best performance in terms
of precision and recall as compared to the classical GA Whitley (1994) or quantum-
inspired GA Narayanan and Moore (1996).
A bird’s eye view of the different classical and intelligent approaches to portfolio
management is provided in Table 6. A detailed comparative study of different state-of-
art approaches is mentioned in Tables 7, 8, 9.Further experimental results of different
approaches are mentioned in Tables 10, 11, 12, 13, 14, 15.

Table 10  Comparison of different state of art techniques from literature

Index Assets Errors GA TS SA PSO ICPSO

Hang Sang 31 Mean Euclidean 0.0040000 0.0040000 0.0040000 0.0049000 0.0000824

Distance
Variance of return 1.6441000 1.6578000 1.6628000 2.2421000 1.9003836
error (%)
Mean return error (%) 0.6072000 0.6107000 0.6238000 0.7427000 0.6409340
Time (s) 18 9 10 34 57
DAX 100 85 Mean Euclidean 0.0076000 0.0082000 0.0078000 0.0090000 0.0001298
Distance
Variance of return 7.2180000 9.0309000 8.5485000 6.8588000 7.2061918
error (%)
Mean return error (%) 1.2791000 1.9078000 1.2817000 1.5885000 1.1876764
Time (s) 99 42 52 179 254
FTSE 100 89 Mean Euclidean 0.0020000 0.0021000 0.0021000 0.0022000 0.0000408
Distance
Variance of return 2.8660000 4.0123000 3.8205000 3.0596000 3.3812060
error (%)
Mean return error (%) 0.3277000 0.3298000 0.3304000 0.3640000 0.3239293
Time (s) 106 42 55 190 269
S & P 100 98 Mean Euclidean 0.0041000 0.0041000 0.0041000 0.0052000 0.0000756
Distance
Variance of return 3.4802000 5.7139000 5.4247000 3.9136000 4.5894404
error (%)
Mean return error (%) 1.2258000 0.7125000 0.8416000 1.4040000 0.8963859
Time (s) 126 51 66 214 323
Nikkei 225 Mean Euclidean 0.0093000 0.0010000 0.0010000 0.0019000 0.0000220
Distance
Variance of return 1.2056000 1.2431000 1.2017000 2.4274000 1.8414523
error (%)
Mean return error (%) 5.3266000 0.4207000 0.4126000 0.7997000 0.4328426
Time (s) 742 234 286 919 2676

13
3878 A. Gunjan, S. Bhattacharyya

Table 11  Comparison of different state of art techniques from literature

Index Assets Errors HIS WPSO CPSO CLPSO HMPPSO

Hang Sang 31 Mean Euclidean 0.0000820 0.0000951 0.0001018 0.0001220 0.0000790

Distance
Variance of return 1.8044041 2.0118800 2.2931800 3.5785200 1.6758400
error (%)
Mean return error (%) 0.6483910 0.7484910 0.7959690 0.6352230 0.6212350
Time (s) 55 119 162 152 136
DAX 100 85 Mean Euclidean 0.0001296 0.0001746 0.0001863 0.0001912 0.0001547
Distance
Variance of return 7.3849655 8.0357600 11.0700000 8.5494500 7.0019500
error (%)
Mean return error (%) 1.0492997 1.5159100 1.5134200 1.4103400 1.3258100
Time (s) 159 459 502 548 463
FTSE 100 89 Mean Euclidean 0.0000400 0.0000591 0.0000653 0.0000414 0.0000390
Distance
Variance of return 3.2479355 4.0458500 5.3473600 2.9917200 2.8747300
error (%)
Mean return error (%) 0.3202570 0.4934480 0.5143340 0.3465540 0.3216470
Time (s) 168 475 511 586 499
S & P 100 98 Mean Euclidean 0.0000746 0.0000956 0.0001222 0.0000799 0.0000782
Distance
Variance of return 3.9023695 3.8469800 7.1698400 3.4338900 2.8476000
error (%)
Mean return error (%) 0.9480060 0.9200940 1.1214000 0.8896880 0.7726480
Time (s) 186 568 634 676 585
Nikkei 225 Mean Euclidean 0.0000206 0.0000518 0.0000978 0.0000261 0.0000205
Distance
Variance of return 1.6020636 3.2030600 7.5329400 1.3008500 1.1965100
error (%)
Mean return error (%) 0.4036726 1.1182400 1.8971600 0.7386130 0.4399000
Time (s) 659 2242 2479 2376 2190

8 Discussions and conclusion

In this study, different types of portfolios and portfolio measures used in portfolio manage-
ment and optimization have been discussed. Different portfolio measures are used depend-
ing on the different types of portfolios. A comparative study of different classical and intel-
ligent approaches for portfolio optimization has been presented in this paper.
It is observed that from the comparative studies that the evolutionary optimization
methods stand in good stead when it comes to optimized allocation of portfolios. Further-
more, the evolving quantum-inspired evolutionary techniques also promise to be better
viable alternatives.
However, evolution of a better portfolio model is a challenge for the overall performance
improvement of these portfolio optimization techniques. Moreover, robust and efficient
optimization techniques remain to be investigated for yielding better optimization scenar-
ios, including the optimization of the risk-return paradigm involving conflicting objectives.
Researchers are constantly investing their efforts in this direction.

13
A brief review of portfolio optimization techniques 3879

Table 12  Comparison of different state of art techniques from literature

2016-09-07 to 2016-10-28 2016-12-08 to 2017-01-28 2017-03-07 to 2017-
04-27

Algorithm MDD fAPV SR MDD fAPV SR MDD fAPV SR

CNN 0.224 29.695 0.087 0.216 8.026 0.059 0.406 31.747 0.076
bRNN 0.241 13.348 0.074 0.262 4.623 0.043 0.393 47.148 0.082
LSTM 0.280 6.692 0.053 0.319 4.073 0.038 0.487 21.173 0.060
iCNN 0.221 4.542 0.053 0.265 1.573 0.022 0.204 3.958 0.044
Best Stock 0.654 1.223 0.012 0.236 1.401 0.018 0.668 4.594 0.033
UCRP 0.265 0.867 -0.014 0.185 1.101 0.01 0.162 2.412 0.049
UBAH 0.324 0.821 -0.015 0.224 1.029 0.004 0.274 2.230 0.036
Anticor 0.265 0.867 -0.014 0.185 1.101 0.010 0.162 2.412 0.049
OLMAR 0.913 0.142 -0.039 0.897 0.123 -0.038 0.733 4.582 0.034
PAMR 0.997 0.003 -0.137 0.998 0.003 -0.121 0.981 0.021 -0.055
WMAMR 0.682 0.742 -0.0008 0.519 0.895 0.005 0.673 6.692 0.042
CWMR 0.999 0.001 -0.148 0.999 0.002 -0.127 0.987 0.013 -0.061
RMR 0.900 0.127 -0.043 0.929 0.090 -0.045 0.698 7.008 0.041
ONS 0.233 0.923 -0.006 0.295 1.188 0.012 0.170 1.609 0.027
UP 0.269 0.864 -0.014 0.188 1.094 0.009 0.165 2.407 0.049
EG 0.268 0.865 -0.014 0.187 1.097 0.010 0.163 2.412 0.049
Bk 0.436 0.758 -0.013 0.336 0.770 -0.012 0.390 2.070 0.027
CORN 0.999 0.001 -0.129 1.000 0.0001 -0.179 0.999 0.001 -0.125
M0 0.335 0.933 -0.001 0.308 1.106 -0.008 0.180 2.729 0.044

*MDD Maximum Drawdown, fAPV final Accumulated Portfolio Value, SR Sharpe Ratio

Table 13  Comparison of Function GA ACO PSO QGA AQGA QIAEA

different state of art techniques
from literature
1 87.50 100.00 67.50 70.00 92.50 100.00
2 35.00 0.00 10.00 20.00 10.00 92.50
3 57.50 20.00 12.50 57.50 60.00 90.00
4 5.00 7.50 20.00 25.00 25.00 52.50
5 57.00 100.00 40.00 35.00 55.00 100.00
6 45.00 87.50 45.00 57.50 80.00 92.50
7 17.50 0.00 2.50 30.00 17.50 52.50
8 40.00 0.00 5.00 52.00 60.00 57.50
9 22.50 10.00 25.00 22.50 52.50 40.00
10 100.00 92.50 45.00 97.50 97.50 100.00
11 92.50 97.50 52.50 52.50 67.50 82.50
12 70.00 100.00 100.00 50.00 40.00 100.00
13 35.00 100.00 50.00 20.00 10.00 35.00
14 95.00 100.00 100.00 57.50 77.00 100.00
15 47.50 0.00 25.00 35.00 25.00 52.50

*Hit accuracy, Population size=40, No. of generations=50, No. of

runs=40

13
3880 A. Gunjan, S. Bhattacharyya

Table 14  Comparison of Method Yearly

different state of art techniques Profit
from literature Return (%)

Buy and Hold -7.20

RSI in normal operating -12.00
RSI in abnormal operating 5.40
Genetic Logic Rule 2.02
Single Genetic Login Rule -0.05
Genetic Neural Networks 10.27
QTS trading system 15.96

*Experiment result performed with stock index of TAIEX on data

from 28-Jun-2000 to 2-Jul-2004

Table 15  Comparison of different state of art techniques from literature

Method / stock Nasdaq (%) DJI (%) NYSE (%) S &P 500 (%) Average (%)

Buy and Hold 55.18 42.03 29.29 45.33 42.96

QTS 34.13 32.20 26.70 47.32 31.78
Best QTS 58.39 38.15 38.45 54.00 47.25
23.39 10.49 11.85 10.71 14.11
Rates of QTS 70.00 100.00 86.67 100.00 89.17
MOQTS 70.15 14.54 23.93 30.75 34.44
Best MOQTS 77.12 38.45 54.00 31.11 50.17
of MOQTS 13.95 9.60 8.01 12.20 10.94
Rates of QTS 100.00 100.00 93.33 65.00 89.58

References
Aboussalah AM, Lee CG (2020) Continuous control with stacked deep dynamic recurrent reinforcement
learning for portfolio optimization. Expert Syst Appl 140(112):891
Achiam J, Knight E, Abbeel P (2019) Towards characterizing divergence in deep q-learning. arXiv preprint
arXiv:​1903.​08894
Agrawal R, Kaur B, Agarwal P (2021) Quantum inspired particle swarm optimization with guided explora-
tion for function optimization. Appl Soft Comput 102(107):122
Almahdi S, Yang SY (2017) An adaptive portfolio trading system: A risk-return portfolio optimization using
recurrent reinforcement learning with expected maximum drawdown. Expert Syst Appl 87:267–279
Alpaydin E (2020) Introduction to machine learning. MIT Press, Cambridge
Alvarez-Alvarado MS, Alban-Chacón FE, Lamilla-Rubio EA, Rodríguez-Gallegos CD, Velásquez W
(2021) Three novel quantum-inspired swarm optimization algorithms using different bounded poten-
tial fields. Sci Rep 11(1):1–22
Anagnostopoulos KP, Mamanis G (2011) The mean-variance cardinality constrained portfolio optimization
problem: an experimental evaluation of five multiobjective evolutionary algorithms. Expert Syst Appl
38(11):14208–14217
Anagnostopoulos KP, Mamanis G (2011) Multiobjective evolutionary algorithms for complex portfolio
optimization problems. CMS 8(3):259–279
Artzner P, Delbaen F, Eber JM, Heath D (1999) Coherent measures of risk. Math Financ 9(3):203–228
Ausiello G, Crescenzi P, Gambosi G, Kann V, Marchetti-Spaccamela A, Protasi M (2012) Complexity and
approximation: combinatorial optimization problems and their approximability properties. Springer,
New York

13
A brief review of portfolio optimization techniques 3881

Aversa P, Haefliger S, Reza DG (2017) Building a winning business model portfolio. MIT Sloan Manag
Rev 58(4):49–54
Babaei G, Giudici P (2021) Explainable artificial intelligence for crypto asset allocation. Available at
SSRN 3977051
Babaei S, Sepehri MM, Babaei E (2015) Multi-objective portfolio optimization considering the depend-
ence structure of asset returns. Eur J Oper Res 244(2):525–539
Bahmani B, Moseley B, Vattani A, Kumar R, Vassilvitskii S (2012) Scalable k-means++. arXiv preprint
arXiv:​1203.​6402
Bai L, Zhang K, Shi H, An M, Han X (2020) Project portfolio resource risk assessment considering
project interdependency by the fuzzy bayesian network. Complexity. https://​doi.​org/​10.​1155/​2020/​
54109​78
Ban GY, El Karoui N, Lim AE (2018) Machine learning and portfolio optimization. Manag Sci
64(3):1136–1154
Bauer E, Kohavi R (1999) An empirical comparison of voting classification algorithms: bagging, boost-
ing, and variants. Mach Learn 36(1):105–139
Bauer R, Hoevenaars R, Steenkamp T (2006) Asset liability management. The oxford handbook of pen-
sions and retirement income. Oxford University Press, Oxford, pp 417–440
Bäuerle N, Rieder U (2011) Markov decision processes with applications to finance. Springer, New York
Best MJ, Hlouskova J (2000) The efficient frontier for bounded assets. Math Methods Oper Res
52(2):195–212
Betancourt C, Chen WH (2021) Deep reinforcement learning for portfolio management of markets with
a dynamic number of assets. Expert Syst Appl 164:114002
Borisov AV (2011) The wonham filter under uncertainty: a game-theoretic approach. Automatica
47(5):1015–1019
Brogan AJ, Stidham S Jr (2008) Non-separation in the mean-lower-partial-moment portfolio optimiza-
tion problem. Eur J Oper Res 184(2):701–710
Bruder B, Gaussel N, Richard JC, Roncalli T (2013) Regularization of portfolio allocation. Available at
SSRN 2767358
Bühlmann P, Yu B (2002) Analyzing bagging. Ann Stat 30(4):927–961
Cai X, Teo KL, Yang X, Zhou XY (2004) Minimax portfolio optimization: empirical numerical study. J
Oper Res Soc 55(1):65–72
Cappart Q, Moisan T, Rousseau LM, Prémont-Schwarz I, Cire A (2020) Combining reinforcement learn-
ing and constraint programming for combinatorial optimization. arXiv preprint arXiv:​2006.​01610
Cesarone F, Scozzari A, Tardella F (2011) Portfolio selection problems in practice: a comparison
between linear and quadratic optimization models. arXiv preprint arXiv:​1105.​3594
Chang TJ, Meade N, Beasley JE, Sharaiha YM (2000) Heuristics for cardinality constrained portfolio
optimisation. Comput Oper Res 27(13):1271–1302
Chaouki A, Hardiman S, Schmidt C, Sérié E, De Lataillade J (2020) Deep deterministic portfolio opti-
mization. J Financ Data Sci 6:16–30
Chen JM (2016) Sortino, omega, kappa: the algebra of financial asymmetry. Postmodern portfolio the-
ory. Springer, New York, pp 79–105
Chen W, Tan S, Yang D (2011) Worst-case var and robust portfolio optimization with interval random
uncertainty set. Expert Syst Appl 38(1):64–70
Chiang HP, Chou YH, Chiu CH, Kuo SY, Huang YM (2014) A quantum-inspired tabu search algorithm
for solving combinatorial optimization problems. Soft Comput 18(9):1771–1781
Choi HK (2018) Stock price correlation coefficient prediction with arima-lstm hybrid model. arXiv pre-
print arXiv:​1808.​01560
Chou YH, Kuo SY, Chen CY, Chao HC (2014) A rule-based dynamic decision-making stock trading
system based on quantum-inspired tabu search algorithm. IEEE Access 2:883–896
Chou YH, Kuo SY, Kuo C (2014) A dynamic stock trading system based on a multi-objective quantum-
inspired tabu search algorithm. In: 2014 IEEE International Conference on Systems, Man, and
Cybernetics (SMC), pp 112–119. IEEE
Chou YH, Yang YJ, Chiu CH (2011) Classical and quantum-inspired tabu search for solving 0/1 knap-
sack problem. In: 2011 IEEE International Conference on Systems, Man, and Cybernetics, pp
1364–1369. IEEE
Chu SC, Tsai PW, Pan JS (2006) Cat swarm optimization. In: Pacific Rim International Conference on
Artificial Intelligence, pp 854–858. Springer
Coloni A, Dorigo M, Maniezzo V (1996) Ant system: optimization by a colony of cooperating agent.
IEEE Trans Syst Man Cybern Part B 26(1):29–41
Cooper RG, Edgett SJ, Kleinschmidt EJ (2001) Portfolio management. Pegasus, New York

13
3882 A. Gunjan, S. Bhattacharyya

Corne DW, Knowles JD, Oates MJ (2000) The pareto envelope-based selection algorithm for multiobjective
optimization. In: International Conference on Parallel Problem Solving from Nature, pp 839–848.
Springer
Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20(3):273–297
Crama Y, Schyns M (2003) Simulated annealing for complex portfolio selection problems. Eur J Oper Res
150(3):546–571
Cura T (2009) Particle swarm optimization approach to portfolio optimization. Nonlinear Anal Real World
Appl 10(4):2396–2406
Das S, Markowitz H, Scheid J, Statman M (2010) Portfolio optimization with mental accounts. J Financ
Quant Anal 45(2):311–334
Davis L (1991) Handbook of genetic algorithms. CumInCAD
Deng GF, Lin WT (2010) Ant colony optimization for markowitz mean-variance portfolio model. In: Inter-
national Conference on Swarm, Evolutionary, and Memetic Computing, pp 238–245. Springer
Derbeko P, El-Yaniv R, Meir R (2002) Variance optimized bagging. In: European Conference on Machine
Learning, pp 60–72. Springer
Derigs U, Nickel NH (2003) Meta-heuristic based decision support for portfolio optimization with a case
study on tracking error minimization in passive portfolio management. OR Spectrum 25(3):345–378
Derigs U, Nickel NH (2004) On a local-search heuristic for a class of tracking error minimization problems
in portfolio management. Ann Oper Res 131(1):45–77
Dey S, Bhattacharyya S, Maulik U (2018) Quantum-inspired automatic clustering technique using ant col-
ony optimization algorithm. In: Quantum-Inspired Intelligent Systems for Multimedia Data Analysis,
pp 27–54. IGI Global
DiVincenzo DP (1998) Quantum gates and circuits. Proc R Soc Lond Ser A 454(1969):261–276
Dorigo M, Birattari M, Stutzle T (2006) Ant colony optimization. IEEE Comput Intell Mag 1(4):28–39
Dowd K (2007) Measuring market risk. Wiley, New York
Drucker H, Burges CJ, Kaufman L, Smola A, Vapnik V et al (1997) Support vector regression machines.
Adv Neural Inf Process Syst 9:155–161
Eckhardt R (1987) Stan ulam, john von neumann, and the monte carlo method. Los Alamos Sci
15(30):131–136
Eckstein S, Kupper M (2019) Computation of optimal transport and related hedging problems via penaliza-
tion and neural networks. Appl Math Optim 83(2):639–667
Eiben AE, Smith JE et al (2003) Introduction to evolutionary computing, vol 53. Springer, New York
El-Shorbagy MA, Hassanien AE (2018) Particle swarm optimization from theory to applications. Int J
Rough Sets Data Anal 5(2):1–24
Feldstein MS (1969) Mean-variance analysis in the theory of liquidity preference and portfolio selection.
Rev Econ Stud 36(1):5–12
Fernández A, Gómez S (2007) Portfolio selection using neural networks. Comput Oper Res
34(4):1177–1191
Forqandoost Haqiqi K, Kazemi T (2011) Ant colony optimization approach to portfolio optimization. Tohid,
Ant Colony Optimization Approach to Portfolio Optimization (August 26, 2011)
Frank M, Wolfe P et al (1956) An algorithm for quadratic programming. Naval R Logist Q 3(1–2):95–110
Friedman JH, Hall P (2007) On bagging and nonlinear estimation. J Stat Plan Inference 137(3):669–683
Galai D, Mark R, Crouhy M (2001) Risk management: comprehensive chapters on market, credit, and oper-
ational risk. McGraw-Hill, New York
Galai D, Mark R, Crouhy M (2001) Risk management: comprehensive chapters on market, credit, and oper-
ational risk. McGraw-Hill, New York
García-Galicia M, Carsteanu AA, Clempner JB (2019) Continuous-time reinforcement learning approach
for portfolio management with time penalization. Expert Syst Appl 129:27–36
Glover F, Laguna M (1998) Tabu search. In: Handbook of combinatorial optimization, pp 2093–2229.
Springer
Goodman R, Thornton M, Strasser S, Sheppard JW (2016) Micpso: A method for incorporating dependen-
cies into discrete particle swarm optimization. In: 2016 IEEE Symposium Series on Computational
Intelligence (SSCI), pp 1–8. IEEE
Guan D, Hipel KW, Fang L, Guo P (2014) Assessing project portfolio risk based on bayesian network. In:
2014 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pp 1546–1551. IEEE
Guan Dj, Guo P (2014) Constructing interdependent risks network of project portfolio based on bayesian
network. In: 2014 International Conference on Management Science & Engineering 21th Annual
Conference Proceedings, pp 1587–1592. IEEE
Guennoun Z, Hamza F et al (2012) Stocks portfolio optimization using classification and genetic algo-
rithms. Appl Math Sci 6(94):4673–4684

13
A brief review of portfolio optimization techniques 3883

Guntsch M, Middendorf M (2003) Solving multi-criteria optimization problems with population-based aco.
In: International Conference on Evolutionary Multi-Criterion Optimization, pp 464–478. Springer
Han KH, Kim JH (2002) Quantum-inspired evolutionary algorithm for a class of combinatorial optimiza-
tion. IEEE Trans Evol Comput 6(6):580–593
Han KH, Park KH, Lee CH, Kim JH (2001) Parallel quantum-inspired genetic algorithm for combinatorial
optimization problem. In: Proceedings of the 2001 Congress on Evolutionary Computation (IEEE
Cat. No. 01TH8546), vol 2, pp 1422–1429. IEEE
Haugh MB, Lo AW (2001) Computational challenges in portfolio management. Comput Sci Eng 3(3):54–59
Haykin S (2010) Neural networks and learning machines, 3/E. Pearson Education India
Hayre L (2002) Salomon Smith Barney guide to mortgage-backed and asset-backed securities, vol 105.
Wiley, New York
He Y, Aranha C (2020) Solving portfolio optimization problems using moea/d and levy flight. arXiv pre-
print arXiv:​2003.​06737
Hibiki N (2001) Multi-period stochastic programming models using simulated paths for strategic asset allo-
cation. J Oper Res Soc Japan 2(44):193
Holton GA (2003) Value-at-risk. Academic Press, Washington
Ito K, Kunisch K (2008) Lagrange multiplier approach to variational problems and applications. SIAM
Jeffery M, Leliveld I (2004) Best practices in it portfolio management. MIT Sloan Manag Rev 45(3):41
Jiang Z, Xu D, Liang J (2017) A deep reinforcement learning framework for the financial portfolio manage-
ment problem. arXiv preprint arXiv:​1706.​10059
Jorion P (1997) Value at risk: the new benchmark for controlling market risk. Irwin Professional Pub
Jorion P (2007) Value at risk: the new benchmark for managing financial risk. The McGraw-Hill Companies
Inc, New York
Karaboga D (2005) An idea based on honey bee swarm for numerical optimization. Tech. Rep, Citeseer
Karatzas I, Zhao X (2001) Bayesian adaptive portfolio optimization. Option pricing, interest rates and risk
management, pp 632–669
Karimkashi S, Kishk AA (2010) Invasive weed optimization and its features in electromagnetics. IEEE
Trans Antennas Propag 58(4):1269–1278
Karmakar S, Dey A, Saha I (2017) Use of quantum-inspired metaheuristics during last two decades. In:
2017 7th International Conference on Communication Systems and Network Technologies (CSNT),
pp 272–278. IEEE
Kaye P, Laflamme R, Mosca M, et al. (2007) An introduction to quantum computing. Oxford University
Press on Demand
Kennedy J, Eberhart R (1942) Particle swarm optimization in: Neural networks. In: Proceedings IEEE Inter-
national Conference on 1995, pp 1942–1948
Kennedy J, Eberhart R (1995) Particle swarm optimization. In: Proceedings of ICNN’95-International Con-
ference on Neural Networks, vol. 4, pp 1942–1948. IEEE
Konno H, Koshizuka T (2005) Mean-absolute deviation model. Iie. Transactions 37(10):893–900
Konno H, Yamazaki H (1991) Mean-absolute deviation portfolio optimization model and its applications to
tokyo stock market. Manage Sci 37(5):519–531
Krokhmal P, Palmquist J, Uryasev S (2002) Portfolio optimization with conditional value-at-risk objective
and constraints. J Risk 4:43–68
Kuo SY, Chou YH (2017) Entanglement-enhanced quantum-inspired tabu search algorithm for function
optimization. IEEE Access 5:13236–13252
Kuo SY, Kuo C, Chou YH (2013) Dynamic stock trading system based on quantum-inspired tabu search
algorithm. In 2013 IEEE Congress on Evolutionary Computation, pp 1029–1036. IEEE
Ładyżyński P, Żbikowski K, Grzegorzewski P (2013) Stock trading with random forests, trend detection
tests and force index volume indicators. In: International Conference on Artificial Intelligence and
Soft Computing, pp 441–452. Springer
Lazimy R (1982) Mixed-integer quadratic programming. Math Program 22(1):332–349
Lee Y, Kim MJ, Kim JH, Jang JR, Chang Kim W (2020) Sparse and robust portfolio selection via semi-
definite relaxation. J Oper Res Soc 71(5):687–699
Li B, Sun Y, Aw G, Teo KL (2019) Uncertain portfolio optimization problem under a minimax risk meas-
ure. Appl Math Model 76:274–281
Li D, Sun X, Wang J (2006) Optimal lot solution to cardinality constrained mean-variance formulation for
portfolio selection. Math Financ 16(1):83–101
Li Y (2017) Deep reinforcement learning: an overview. arXiv preprint arXiv:​1701.​07274
Li Y, Heng B, Zhou S, Chen R, Liu S (2012) A novel aco algorithm based on average entropy for real
estate portfolio optimization. J Theor Appl Inf Technol 45(2):502–507

13
3884 A. Gunjan, S. Bhattacharyya

Liang Z, Chen H, Zhu J, Jiang K, Li Y (2018) Adversarial deep reinforcement learning in portfolio man-
agement. arXiv preprint arXiv:​1808.​09940
Lim AE, Shanthikumar JG, Vahn GY (2011) Conditional value-at-risk in portfolio optimization: coher-
ent but fragile. Oper Res Lett 39(3):163–171
Lin J (1976) Multiple-objective problems: pareto-optimal solutions by method of proper equality con-
straints. IEEE Trans Autom Control 21(5):641–650
Lin YC, Chen CT, Sang CY, Huang SH (2022) Multiagent-based deep reinforcement learning for risk-
shifting portfolio management. Appl Soft Comput 123(108):894
Lwin KT, Qu R, MacCarthy BL (2017) Mean-var portfolio optimization: a nonparametric approach. Eur
J Oper Res 260(2):751–766
Mahdavi M, Fesanghary M, Damangir E (2007) An improved harmony search algorithm for solving
optimization problems. Appl Math Comput 188(2):1567–1579
Mansini R, Ogryczak W, Speranza MG (2015) Portfolio optimization with transaction costs. In: Linear
and Mixed Integer Programming for Portfolio Optimization, pp 47–62. Springer
Mansini R, Speranza MG (1999) Heuristic algorithms for the portfolio selection problem with minimum
transaction lots. Eur J Oper Res 114(2):219–233
Mariano CE, Morales EM (1999) Moaq an ant-q algorithm for multiple objective optimization problems.
In: Proceedings of the 1st Annual Conference on Genetic and Evolutionary Computation, vol 1, pp
894–901
Markowitz H (1959) Portfolio selection, efficent diversification of investments. Wiley, New York
Markowitz HM, Todd GP (2000) Mean-variance analysis in portfolio choice and capital markets, vol 66.
Wiley, New York
McMahon D (2007) Quantum computing explained. Wiley, New York
McNeil AJ, Frey R, Embrechts P (2015) Quantitative risk management: concepts, techniques and tools-
revised edition. Princeton University Press, Princeton
Mirjalili S (2019) Genetic algorithm. Evolutionary algorithms and neural networks. Springer, New York,
pp 43–55
Mnih V, Kavukcuoglu K, Silver D, Rusu AA, Veness J, Bellemare MG, Graves A, Riedmiller M, Fid-
jeland AK, Ostrovski G et al (2015) Human-level control through deep reinforcement learning.
Nature 518(7540):529–533
Mohsin SA, Darwish SM, Younes A (2021) Qiaco: a quantum dynamic cost ant system for query optimi-
zation in distributed database. IEEE Access 9:15833–15846
Montiel O, Rubio Y, Olvera C, Rivera A (2019) Quantum-inspired acromyrmex evolutionary algorithm.
Sci Rep 9(1):1–10
Moody J, Saffell M (2001) Learning to trade via direct reinforcement. IEEE Trans Neural Networks
12(4):875–889
Moon Y, Yao T (2011) A robust mean absolute deviation model for portfolio optimization. Comput Oper
Res 38(9):1251–1258
Mugel S, Kuchkovsky C, Sanchez E, Fernandez-Lorenzo S, Luis-Hita J, Lizaso E, Orus R (2020)
Dynamic portfolio optimization with real datasets using quantum processors and quantum-inspired
tensor networks. arXiv preprint arXiv:​2007.​00017
Mulvey JM (2001) Multi-stage optimization for long-term investors. In: Quantitative Analysis In Finan-
cial Markets: Collected Papers of the New York University Mathematical Finance Seminar (Vol-
ume III), pp 66–85. World Scientific
Narayanan A, Moore M (1996) Quantum-inspired genetic algorithms. In: Proceedings of IEEE Interna-
tional Conference on Evolutionary Computation, pp 61–66. IEEE
Nawrocki DN (1992) The characteristics of portfolios selected by n-degree lower partial moment. Int
Rev Financ Anal 1(3):195–209
Nielsen MA, Chuang IL (2001) Quantum computation and quantum information. Phys Today 54(2):60
NYSE NASDAQ, A (2003) New york stock exchange (nyse), nasdaq, and amex data from january 1996
to september 2003 in the crsp/compustat merged database (Jan’1996–Sep’2003)
Oh KJ, Kim TY, Min S (2005) Using genetic algorithm to support portfolio optimization for index fund
management. Expert Syst Appl 28(2):371–379
Orus R, Mugel S, Lizaso E (2019) Quantum computing for finance: overview and prospects. Rev Phys
4(100):028
Park K, Jung HG, Eom TS, Lee SW (2022) Uncertainty-aware portfolio management with risk-sensitive
multiagent network. In: IEEE Transactions on Neural Networks and Learning Systems
Passino KM (2002) Biomimicry of bacterial foraging for distributed optimization and control. IEEE Control
Syst Mag 22(3):52–67
Pedersen M (2014) Portfolio optimization and monte carlo simulation. Available at SSRN 2438121

13
A brief review of portfolio optimization techniques 3885

Rabiner L, Juang B (1986) An introduction to hidden Markov models. IEEE ASSP Mag 3(1):4–16
Rämö H (2002) Doing things right and doing the right things time and timing in projects. Int J Project
Manag 20(7):569–574
Ranković V, Drenovak M, Urosevic B, Jelic R (2016) Mean-univariate garch var portfolio optimization:
actual portfolio approach. Comput Oper Res 72:83–92
Ray J, Bhattacharyya S (2015) Value-at-risk based portfolio allocation using particle swarm optimization.
Int J Comput Sci Eng (E-ISSN: 2347-2693) 3: 1–9
Ray J, Bhattacharyya S (2017) Particle swarm optimization technique for optimizing conditional value-at-
risk based portfolio. Int J Comput Sci Eng 5(2)
Rebentrost P, Lloyd S (2018) Quantum computational finance: quantum algorithm for portfolio optimiza-
tion. arXiv preprint arXiv:​1811.​03975
Reyes-Sierra M, Coello CC et al (2006) Multi-objective particle swarm optimizers: a survey of the state-of-
the-art. Int J Comput Intell Res 2(3):287–308
Rezani M, Hertono G, Handari B (2020) Implementation of iterative k-means-+ and ant colony optimiza-
tion (aco) in portfolio optimization problem. In: AIP Conference Proceedings, vol 2242, p 030022.
AIP Publishing LLC
Rockafellar RT (1970) Conjugate convex functions in optimal control and the calculus of variations. J Math
Anal Appl 32(1):174–222
Rockafellar RT, Uryasev S (2002) Conditional value-at-risk for general loss distributions. J Bank Financ
26(7):1443–1471
Rockafellar RT, Uryasev S et al (2000) Optimization of conditional value-at-risk. J Risk 2:21–42
Roll R (1992) A mean/variance analysis of tracking error. J Portfolio Manag 18(4):13–22
Saad HM, Chakrabortty RK, Elsayed S, Ryan MJ (2021) Quantum-inspired genetic algorithm for resource-
constrained project-scheduling. IEEE Access 9:38488–38502
Saboia JLM (1977) Autoregressive integrated moving average (arima) models for birth forecasting. J Am
Stat Assoc 72(358):264–270
Salahi M, Daemi M, Lotfi S, Jamalian A (2014) Pso and harmony search algorithms for cardinality con-
strained portfolio optimization problem. Adv Model Optim 16(3):559–573
Samuelson PA (1975) The fundamental approximation theorem of portfolio analysis in terms of means,
variances and higher moments. Stochastic optimization models in finance. Elsevier, Amsterdam, pp
215–220
Schlichtkrull M, Kipf TN, Bloem P, Berg Rvd, Titov I, Welling M (2018) Modeling relational data with
graph convolutional networks. In: European Semantic Web Conference, pp 593–607. Springer
Schlottmann F, Seese D (2004) A hybrid heuristic approach to discrete multi-objective optimization of
credit portfolios. Comput Stat Data Anal 47(2):373–399
Schmidhuber J, Hochreiter S et al (1997) Long short-term memory. Neural Comput 9(8):1735–1780
Schulman J, Wolski F, Dhariwal P, Radford A, Klimov O (2017) Proximal policy optimization algorithms.
arXiv preprint arXiv:​1707.​06347
Seng H (1992-1997) Weekly price from hang seng in Hong Kong, dax 100 in Germany, FTSE 100 in UK, s
&p 100 in USA and Nikkei in Japan
Shenoy C, Shenoy PP (2000) Bayesian network models of portfolio risk and return. The MIT Press,
Cambridge
Shi S, Li J, Li G, Pan P (2019) A multi-scale temporal feature aggregation convolutional neural network for
portfolio management. In: Proceedings of the 28th ACM International Conference on Information
and Knowledge Management, pp 1613–1622
Shi S, Li J, Li G, Pan P, Chen Q, Sun Q (2022) Gpm: a graph convolutional network based reinforcement
learning framework for portfolio management. Neurocomputing 498:14–27
Soleymani F, Paquet E (2020) Financial portfolio optimization with online deep reinforcement learning and
restricted stacked autoencoder-deepbreath. Expert Syst Appl 156(113):456
Soleymani F, Paquet E (2021) Deep graph convolutional reinforcement learning for financial portfolio man-
agement-deeppocket. Expert Syst Appl 182(115):127
SP500: One year data from s &p 500 from January 2016 to January 2017 (Jan’2016–Jan’2017)
Sutton RS, Barto AG (2018) Reinforcement learning: an introduction. MIT Press, Cambridge
Talbi H, Draa A, Batouche M (2006) A novel quantum-inspired evolutionary algorithm for multi-sensor
image registration. Int Arab J Inf Technol 3(1):9–15
Tan Y, Zhu Y (2010) Fireworks algorithm for optimization. In: International Conference in Swarm Intel-
ligence, pp 355–364. Springer
Van Veldhuizen DA, Lamont GB (1998) Multiobjective evolutionary algorithm research: a history and anal-
ysis. Tech. Rep., Citeseer

13
3886 A. Gunjan, S. Bhattacharyya

Vazquez E, Clempner JB (2020) Customer portfolio model driven by continuous-time markov chains: an l 2
Lagrangian regularization method. Econ Comput Econ Cybern Stud Res 54(2)
Whitley D (1994) A genetic algorithm tutorial. Stat Comput 4(2):65–85
Wolsey LA, Nemhauser GL (1999) Integer and combinatorial optimization, vol 55. Wiley, New York
Wu ME, Syu JH, Lin JCW, Ho JM (2021) Portfolio management system in equity market neutral using rein-
forcement learning. Appl Intell 51(11):8119–8131
Zl Wu, Zhang A, Li Ch, Sudjianto A (2008) Trace solution paths for svms via parametric quadratic pro-
gramming. Data Mining Using Matrices and Tensors, KDD Worskshop
Xu C (2003) Soft approach for solving hard optimization problems. In: Proceedings of of the Fall National
Conference of Japan Society for Management Information, pp 74–77
Xu C, Ng P (2006) A soft approach for hard continuous optimization. Eur J Oper Res 173(1):18–29
Xu C, Wang J, Shiba N (2007) Multistage portfolio optimization with var as risk measure. Int J Innov Com-
put Inf Control 3(3):709–724
Yang X (2006) Improving portfolio efficiency: a genetic algorithm approach. Comput Econ 28(1):1
Yang XS (2009) Firefly algorithms for multimodal optimization. In: International Symposium on Stochastic
Algorithms, pp 169–178. Springer
Yang XS (2010) A new metaheuristic bat-inspired algorithm. In: Nature Inspired Cooperative Strategies for
Optimization (NICSO 2010), pp 65–74. Springer
Yang Y, Xu DL (2017) A methodology for assessing the effect of portfolio management on npd perfor-
mance based on bayesian network scenarios. Expert Syst 34(2):e12186
Yang Z, Yin G, Zhang Q (2015) Mean-variance type controls involving a hidden markov chain: models and
numerical approximation. IMA J Math Control Inf 32(4):867–888
Yin G, Liu R, Zhang Q (2002) Recursive algorithms for stock liquidation: a stochastic optimization
approach. SIAM J Optim 13(1):240–263
Yin X, Ni Q, Zhai Y (2015) A novel pso for portfolio optimization based on heterogeneous multiple popula-
tion strategy. In: 2015 IEEE Congress on Evolutionary Computation (CEC), pp 1196–1203. IEEE
Young MR (1998) A minimax portfolio selection rule with linear programming solution. Manag Sci
44(5):673–683
Young TW (1991) Calmar ratio: a smoother tool. Futures 20(1):40
Yu P, Lee JS, Kulyatin I, Shi Z, Dasgupta S (2019) Model-based deep reinforcement learning for dynamic
portfolio optimization. arXiv preprint arXiv:​1901.​08740
Yusoff Y, Ngadiman MS, Zain AM (2011) Overview of nsga-ii for optimizing machining process param-
eters. Procedia Eng 15:3978–3983
Zhou A, Qu BY, Li H, Zhao SZ, Suganthan PN, Zhang Q (2011) Multiobjective evolutionary algorithms: a
survey of the state of the art. Swarm Evol Comput 1(1):32–49
Zhu H, Wang Y, Wang K, Chen Y (2011) Particle swarm optimization (pso) for the constrained portfolio
optimization problem. Expert Syst Appl 38(8):10161–10169
Zitzler E, Laumanns M, Thiele L (2001) Spea2: improving the strength pareto evolutionary algorithm. TIK-
Report 103
Zouache D, Nouioua F, Moussaoui A (2016) Quantum-inspired firefly algorithm with particle swarm opti-
mization for discrete optimization problems. Soft Comput 20(7):2781–2799

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Springer Nature or its licensor holds exclusive rights to this article under a publishing agreement with the
author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is
solely governed by the terms of such publishing agreement and applicable law.

13