Applied Soft Computing Journal 100 (2021) 106943

Contents lists available at ScienceDirect

Applied Soft Computing Journal

journal homepage: www.elsevier.com/locate/asoc

Mean–variance portfolio optimization using machine learning-based

stock price prediction
∗
Wei Chen a , , Haoyu Zhang a , Mukesh Kumar Mehlawat b , Lifen Jia a
a
School of Management and Engineering, Capital University of Economics and Business, Beijing, China
b
Department of Operational Research, University of Delhi, Delhi, India

article info a b s t r a c t

Article history: The success of portfolio construction depends primarily on the future performance of stock markets.
Received 16 June 2020 Recent developments in machine learning have brought significant opportunities to incorporate
Received in revised form 22 November 2020 prediction theory into portfolio selection. However, many studies show that a single prediction
Accepted 24 November 2020
model is insufficient to achieve very accurate predictions and affluent returns. In this paper, a novel
Available online 5 December 2020
portfolio construction approach is developed using a hybrid model based on machine learning for stock
Keywords: prediction and mean–variance (MV) model for portfolio selection. Specifically, two stages are involved
Portfolio selection in this model: stock prediction and portfolio selection. In the first stage, a hybrid model combining
Stock prediction eXtreme Gradient Boosting (XGBoost) with an improved firefly algorithm (IFA) is proposed to predict
eXtreme Gradient Boosting stock prices for the next period. The IFA is developed to optimize the hyperparameters of the XGBoost.
Firefly algorithm
In the second stage, stocks with higher potential returns are selected, and the MV model is employed
Mean–variance model
for portfolio selection. Using the Shanghai Stock Exchange as the study sample, the obtained results
demonstrate that the proposed method is superior to traditional ways (without stock prediction) and
benchmarks in terms of returns and risks.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction machine learning methods. Common methods are neural net-

works (NNs) [8,9], support vector regression (SVR) [10,11], en-
Constructing a portfolio by proper stock selection has been semble learning [12,13], etc. A detailed review of stock prediction
considered an essential task for individual and institutional in- approaches can be found in [14,15]. Some comparative stud-
vestors. In this context, the portfolio’s improvement and opti- ies highlighted that machine learning has a stronger ability to
mization have become one of the most concerning issues in mod- deal with non-linear and non-stationary problems than statistical
models [16].
ern financial research and investment decision-making [1]. As
In the machine learning communities (even in machine learn-
we know, portfolio selection’s success is significantly contingent
ing competitions like1 ), there are various ensemble learning al-
upon the performance of future stock markets. Reasonable and
gorithms, whose purpose is to reduce prediction bias and vari-
accurate forecasts have the potential to generate high investment
ance and achieve better predictive performance than a single
returns and hedge risks [2].
algorithm [17]. Ensemble learning algorithms mainly include
The efficient market hypothesis [3] is the leading theory of
Adaptive Boosting (AdaBoost) [18], Gradient Boosted Decision
finance and economics. This hypothesis put forward that it is Tree (GBDT) [13], and eXtreme Gradient Boosting (XGBoost) [19].
impossible to predict the price of financial assets. However, many Among them, XGBoost has received much attention for its low
studies confirmed that stock prices and returns are predictable computational complexity, high prediction accuracy, and out-
to some extent [4]. Up to now, existing studies have mainly standing efficiency. In fact, XGBoost is an improved GBDT, which
focused on (i) statistical methods. Statistical methods aim at is composed of many decision trees and is applied for classifica-
predicting by analyzing the past price characteristics, such as tion and regression. Recently, XGBoost has been applied to the
autoregressive conditional heteroscedasticity (ARCH) [5], autore- forecasting of financial markets. For instance, Nobre and Neves
gressive integrated moving average (ARIMA) [6], generalized au- [19] proposed an expert system using the XGBoost as a binary
toregressive conditional heteroscedasticity (GARCH) [7]; and (ii) classifier. The results demonstrated that the system could achieve
higher average returns in financial markets. Dey et al. [20] used
∗ Corresponding author.
E-mail address: [email protected] (W. Chen). 1 The website is www.kaggle.com.

https://doi.org/10.1016/j.asoc.2020.106943
1568-4946/© 2020 Elsevier B.V. All rights reserved.
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

the XGBoost to classify the stock trend in different periods, and the portfolio selection models. For example, Paiva et al. [50]
the superiority of the XGBoost is verified by comparing it with put forward an investment decision model that uses the SVM
non-ensemble algorithms. However, to the best of our knowledge, to classify assets and combines with the MV model to form an
XGBoost is widely applied as a classifier for predicting the finan- optimal portfolio. Wang et al. [16] proposed a hybrid method
cial markets, but few studies use it for regression problems. In combining long short-term memory with the MV model, which
this paper, an XGBoost-based model is developed for stock price optimizes the formation of the portfolio in combination with
forecasting. asset preselection. These researches showed that the combina-
Based on the previous research in the literature, it is indi- tion of stock prediction and portfolio selection might provide
cated that XGBoost has several hyperparameters that need to a new perspective for financial analysis. In this study, the MV
be set artificially, and the setting of hyperparameters has an model is adopted for portfolio selection to determine each asset’s
essential impact on the prediction accuracy [21]. Moreover, even proportion.
if researchers know each hyperparameter’s role, it is hard to In sum, this paper proposes a novel approach to portfolio con-
accurately determine which hyperparameters need to be adjusted struction by combining machine learning-based model for stock
and how to adjust it to obtain the optimal model [22]. There- prediction with the MV model for portfolio selection. Specifically,
fore, it is crucial to adopt an appropriate optimization approach two main phases are involved in this model: stock prediction and
to choose optimal hyperparameters. Traditional hyperparameter portfolio selection. In the first phase, a hybrid model combining
optimization methods, such as grid search, generally select the XGBoost with IFA is proposed to predict stock prices for the next
hyperparameters with optimal performance by comparing dif- period, where the IFA is developed to optimize the hyperparam-
ferent combinations. However, this type of method belongs to eters of the XGBoost. In the second phase, stocks with higher
the enumeration, which cannot be used for large-scale calcula- potential returns are selected according to each stock’s predicted
tions with high precision [23]. In recent years, meta-heuristic prices, and the MV model is employed for allocating the invest-
algorithms, such as particle swarm optimization (PSO) [24], ar- ment proportion of the portfolio. The proposed method provides
tificial bee colony (ABC) [25], salp swarm algorithm (SSA) [26] several new features compared with existing studies on portfolio
and firefly algorithm (FA) [27,28], have been used to optimize construction in general (see Table 1). The main contributions of
hyperparameters. At present, the standard FA, with few param- this study can be summarized as follows.
eters, simple operation, and strong robustness, has been shown
as a useful tool in hyperparameter optimization problems; see, • We develop a hybrid model named IFAXGBoost for stock
for instance, Kuo et al. [29], Ibrahim and Khatib [30], Chahnasir price prediction, in which an improved firefly algorithm
et al. [31], Payne et al. [32], Mehr et al. [33]. However, the (IFA) is designed to optimize the XGBoost. In the proposed
standard FA has the problem of quickly falling into local optimum IFA, we dynamically divide the whole firefly group into
when solving complex optimization problems. To deal with this, an elite subgroup and a normal subgroup, and the chaotic
several variants of the traditional FA have been developed to search strategy and the PSO-based search strategy are de-
optimize the hyperparameters. Kazem et al. [34] proposed a stock signed accordingly.
price prediction model, in which the hyperparameters of SVR • Unlike most studies for portfolio construction, we incorpo-
is optimized by the chaotic FA. Zhang et al. [35] developed a rate the hybrid model IFAXGBoost into the mean–variance
modified FA to optimize SVR’s hyperparameters for stock price model to capture the future features of stock markets, thus
prediction, in which the opposition-based chaotic strategy and forming a novel portfolio construction approach.
dynamic adjustment strategy are introduced. Unfortunately, few • Using historical data from the Shanghai Stock Exchange,
current studies utilize the FA and its variants to determine the we compare the proposed method’s performance with tra-
hyperparameters of the XGBoost. This paper thus proposes an
ditional ways (without stock prediction) and benchmarks
improved Firefly algorithm (IFA) to optimize the hyperparameters
(with other prediction models). The experimental results
of XGBoost.
show the effectiveness and superiority of this method.
Portfolio optimization is the distribution of wealth among
several assets, in which two parameters, namely, expected re- Section 2 introduces the forecasting methods we use. In Sec-
turns and risks, are very important. Investors usually tend to tion 3, the MV model, is presented. Section 4 covers the details of
maximize returns and minimize risks. However, high returns the experiments performed. In Section 5, we report the empirical
typically bring increased risks. The mean–variance (MV) model results. Finally, Section 6 summarizes our main work and outlines
proposed by Markowitz [36] is one of the best models to solve the future work.
portfolio optimization problem. In this model, returns and risks
are quantified by means and variances, respectively, facilitating
2. Prediction method
investors to make tradeoffs between maximizing expected return
and minimizing risk. After Markowitz’s work, the standard MV
2.1. XGBoost method
model has been improved and extended in the following aspects:
(i) Concerning multi-period portfolio selections [37–39]. (ii) In-
troducing the alternative risk measurements. For instance, safety- The XGBoost is the abbreviation of ‘‘eXtreme Gradient Boost-
first model [40], mean-semivariance model [41], mean absolute ing,’’ proposed by Chen [55] in 2016. This method’s characteristics
deviation model [42], mean semiabsolute deviation models [43], are low in computational complexity, fast in running speed, and
etc. (iii) Including several real-world constraints, such as cardi- high accuracy. The objective function of the XGBoost is to com-
nality constraints, transaction costs [44–47]. More researches can bine the standard penalty term with the loss function term to
be found in comprehensive surveys [48,49]. However, the above obtain the optimal solution, in which the regular penalty term can
studies pay more attention to the improvement and expansion reduce the variance of the model, thus preventing over-fitting.
of the MV model, but neglect the selection of high-quality assets The additive model of the tree ensemble model can be described
before forming the optimal portfolio. Actually, in the investment by:
process, high-quality asset input is a reliable guarantee for the K
∑
construction of an optimal portfolio. In recent years, some studies ŷi = fk (xi ), fk ∈F , (1)
have been conducted by integrating the asset preselection with k=1

2
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

Table 1
Comparison with existing work.
Attribute [19] [9] [51] [52] [53] [44] [45] [54] [28] [35] [47] [50] [16] Proposed
approach
Portfolio selection ✓ ✓ ✕ ✕ ✕ ✓ ✓ ✕ ✕ ✕ ✓ ✓ ✓ ✓
Environment Crisp Crisp Crisp Crisp Crisp Fuzzy Uncer- Crisp Crisp Crisp Crisp Crisp Crisp Crisp
tain
XGBoost ✓ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✓
Hyperparameter ✓ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✓
optimization
Sharpe ratio ✓ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✓ ✓
Cardinality ✕ ✕ ✕ ✕ ✕ ✕ ✓ ✕ ✕ ✕ ✓ ✓ ✓ ✓
Financial market BEM US stock S&P 500, SSE China – – NASDAQ S&P-500, SSE S&P-100 Ibovespa FTSE 100 SSE
market FTSE- stock FTSE-
100, NSE, market 100,
SGX, Nikkei-
Hang 225
Seng, SSE
Solution approach MOO-GA Neural Support Adaptive PSO FA-SA FA-GA Chaos FA-MSVR MFA NSGA-II Support LSTM IFA
network vector SVR al- based vector
machine gorithm -FA machine
Stock prediction ✓ ✓ ✓ ✓ ✓ ✕ ✕ ✓ ✓ ✓ ✕ ✓ ✓ ✓
Transaction cost ✓ ✕ ✕ ✕ ✕ ✓ ✓ ✕ ✓ ✕ ✓ ✓ ✓ ✓

Acronyms:- BEM: Brazile exchange market, NSE: National stock exchange (India), SGX: Singapore exchange limited, SSE: Shanghai stock exchange, MOO-GA:
Multiobjective optimization genetic algorithm, GA: Genetic algorithm, SVR: Support vector machine regression, PSO: Particle swarm optimization, FA-SA: Firefly
algorithm and simulated annealing algorithm, FA-GA: Firefly algorithm and simulated annealing algorithm, FA: Firefly algorithm, FA-MSVR: Firefly algorithm and
multi-output support vector regression, MFA: Modified firefly algorithm, NSGA-II: Non-dominated sorting genetic algorithm LSTM: long short-term memory IFA:
Improved firefly algorithm.

where F is the set of regression trees, f is a tree in the tree space gi = ∂ŷ(t −1) l(yi , ŷ(t −1) ), (10)
F , K is the number of trees and xi represents the ith eigenvector.
hi = ∂ 2
ŷ(t −1) l(yi , ŷ (t −1)
). (11)
The objective function to be optimized is as follows:
∑ ∑ 2.2. Hyperparameter optimization method
L(j) = l(yi , ŷi ) + Ω (fk ), (2)
i k
In traditional machine learning methods, hyperparameter se-
where l(yi , ŷi ) is the loss function, and Ω is regular punishment lection is based on experience, leading to unreliable results and
that can be calculated as follows: increasing the prediction’s randomness. For the sake of solving
1 this problem, an improved firefly algorithm (IFA) is developed to
Ω (fk ) = γ T + λ∥ω∥2 , (3) select optimal hyperparameters of the XGBoost. In the following,
2
the standard FA is first reviewed, and then the IFA is presented.
where:
γ = L1 regularization coefficient, 2.2.1. Standard firefly algorithm
λ = L2 regularization coefficient, In the standard FA, the brightness is used to assess the quality
T = number of leaves, of fireflies, which is influenced by the fitness value of the fire-
ω = leaf weight. fly’s position. In terms of the minimization problem, the bright-
It is difficult to learn all parameters of a tree at once, so ness is inversely proportional to the fitness value. Furthermore,
the additive strategy is adopted in Eq. (1), which is extended as fireflies update their positions according to attraction, which is
follows: generally proportional to the distance between them. Let Xi =
(0)
ŷi = 0, (4) (xi1 , xi2 , . . . , xiD ) be the ith firefly in the D-dimensional space,
(1) (0) where i = 1, 2, . . . , N and N is the population size. The distance
ŷi = f1 (xi ) = ŷi + f1 (xi ), (5) between any two fireflies Xi and Xj is given as follows:
(2) (1)
ŷi = f1 (xi ) + f2 (xi ) = ŷi + f2 (xi ), (6) 
 D
···   ∑
(xid − xjd )2 ,

rij = Xi − Xj = √
  (12)
t
(t)
∑ (t −1) d=1
ŷi = fk (xi ) = ŷi + ft (xi ). (7)
k=1 where xid and xjd are the positions of firefly i and firefly j under
th the dth dimension.
The t objective function is described as follows: The firefly’s attraction is defined as:
n
2
β (rij ) = β0 e(−γ rij ) ,
∑ (t −1)
L(t) = l(yi , ŷi + ft (xi )) + Ω (ft ). (8) (13)
i=1 where β0 is the attraction at r = 0 and γ is the optical absorption
The second Taylor expansion of Eq. (8) is performed to obtain coefficient.
Eq. (9), where the first derivative gi and the second derivative hi For two fireflies, the less bright one moves toward the other.
are Eqs. (10) and (11), respectively. Assume that xj is brighter (better) than xi . The updated position
n [ ] of firefly i to firefly j is defined as follows:
∑ (t −1) 1
L(t) ≈ l(yi , ŷi + gi ft (xi ) + hi ft 2 (xi )) + Ω (ft ), (9) 2 1
2 xid (t + 1) = xid (t) + β0 e−γ rij (xjd (t) − xid (t)) + α (rand − ), (14)
i=1 2
3
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

where t denotes the number of algorithm iterations, rand is a where ω is a coefficient that reduces as the number of iterations
uniform random number within [0, 1], and α is the step size factor increases:
between [0, 1]
t
ω = α (1 − ). (19)
tmax
2.2.2. Improved firefly algorithm
The standard FA has typical problems of quickly falling into Finally, the flowchart and pseudo code of the IFA are described
local optimum and premature convergence [56]. To overcome as Fig. 1 and Algorithm 1, respectively.
this demerit, we first dynamically divide the whole firefly group
into an elite subgroup and a normal subgroup. The chaotic search
strategy is then used for the elite subgroup to improve the local Algorithm 1 Improved firefly algorithm
searchability. At the same time, the PSO-based search strategy is 1: Calculate the fitness values of each firefly
2: Initialize the population of fireflies (solutions) and a set of parameters
employed for the normal subgroup to enhance the global search 3: while t < MaxGeneration do
capability and speed up the global convergence. 4: Sort all fireflies according to the value of the objective function
(1) Movement of the elite subgroup 5: Obtain threshold p according to Eq. (15)
6: for i = p to N do
The convergence speed and solution quality of the FA are 7: Update the positions of elite fireflies according to Eqs. (16) and (17)
affected by the population quality. The introduction of chaotic 8: Evaluate new solutions and update brightness
sequences can improve the swarm’s diversity and accelerate the 9: end for
10: for i = 1 to p do
convergence of the global solution [57,58]. However, if the chaotic 11: for j = 1 to p do
sequences are applied to all solutions, the complexity of calcula- 12: if Ij >Ii then
tion will increase. Therefore, to overcome the above drawbacks, 13: Move firefly i toward j in d-dimension according to Eqs. (18) and
(19)
we propose a new strategy by incorporating the chaotic se- 14: end if
quences into elite fireflies’ movement strategy. The proposed 15: Evaluate new solutions and update brightness
strategy is described in detail below. 16: end for
17: end for
(1) Initialize the position of fireflies randomly, and calculate 18: Rank the fireflies and find the current best
the value of objective function. 19: end while
(2) Determine the threshold p to obtain elite fireflies. Since
fireflies’ position changes significantly in the early stage of op-
timization; more attention should be paid to exploring to avoid Remark. The time complexity of the IFA is O(n2 ), which is
falling into local optima. While in the later stage of optimization, the same as the standard FA. Note that n represents population
after enough exploration, the search pays more attention to the size and has a significant influence on complexity, where the
exploitation to achieve a better solution. Based on this consider- complexity rises with the increase of the population size. The
ation, the threshold p for controlling elite fireflies is determined reason for the high complexity is that the IFA compares each
as:
⌊ ⌋ firefly to the whole population.
t
p = N ∗ e tmax
−1
, (15)

where ⌊ ⌋ represents rounding down, t and tmax are the cur-

rent number of iterations and the total number of iterations, 3. Mean–variance model
respectively.
(3) The elite fireflies conduct the chaotic search in the search
space, and the position is updated as: The MV model proposed by Markowitz [36] lays the basis
for portfolio selection. In this model, the investment return and
xelite (t + 1) = xmin + c1 (xmax − xmin ), (16) risk are quantified by expected return and variance, respectively.
c1 = µc2 (1 − c2 ), (17) Rational investors always pursue the lowest risk under a specific
where xmax and xmin indicate the upper and lower bounds of the expected return or the highest return under a particular risk,
elite fireflies, respectively, c2 is a randomly chosen number within choosing an appropriate portfolio to maximize expected utility.
[0, 1] and µ is the control parameter of chaotic state. The MV model aims to make a trade-off between maximizing
(2) Movement of the ordinary subgroup returns and minimizing risks, which is expressed by a typical
In the standard FA, fireflies’ movement is only based on the multi-objective optimization formula:
brightness, which causes fireflies to quickly lose the advantage of
the optimal solution obtained, thereby reducing the search effi-
n n
ciency. Therefore, it is necessary to introduce other information ∑ ∑
besides brightness. In the PSO [59], each particle’s movement min xi xj σij
takes advantage of its current best position pbest and the global i=1 j=1
n
best position gbest of the whole population. In other words, a ∑
particle moves toward not only its best previous position but also max xi µ i
(20)
the best particle. Inspired by the PSO’s exploitation mechanism, i=1
⎧ n
this paper introduces the global best solution gbest into the FA. In ∑
xi = 1 ,
⎪
this way, the movement of fireflies is based on both brightness
⎨
s.t.
and the global optimal solution. The movement of the ordinary ⎪ i=1
0 ≤ xi ≤ 1, ∀ i = 1, . . . , n,
⎩
subgroup is expressed as follows:

2 1 where σij is covariance between asset i and asset j, xi and xj rep-

xid (t + 1) = xid (t) +β0 e−γ rij (xjd (t) − xid (t)) +ω(rand − )(gbest − xi ),
2 resent the proportion of the initial value, and µi is the expected
(18) return on asset i.
4
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

Fig. 1. The flowchart of the IFA.

Chang et al. [60] introduced the risk aversion coefficient to 4. Experimental design
change the multi-objective formula into the mono-objective for-
mulation: 4.1. Data and input indicator
⎡ ⎤ [ n ]
n n 4.1.1. Data
∑ ∑ ∑
min λ⎣ xi xj σij ⎦ − (1 − λ) xi µi The predictability of stock prices is related to the volatility
i=1 j=1 i=1 over time. Naturally, stable stocks would be easier to predict
⎧ n
∑ (21) than relatively noisy ones. The Shanghai Stock Exchange (SSE) 50
xi = 1, index is selected according to the turnover and total market value,
⎪
⎨
s.t. with the characteristics of a large scale, relatively stable, and
i=1
adequate liquidity [61]. Therefore, this paper randomly selects 24
⎪
0 ≤ xi ≤ 1, ∀ i = 1, . . . , n.
⎩
stocks in the SSE 50 index as candidate assets, large enough for
Among them, the risk aversion coefficient is between 0 and individual investors to choose stocks before forming portfolios.
1. λ = 0 indicates that investor is very risk-averse and pursues The ticker symbol of 24 stocks are 600,000, 600,018, 600,028,
the maximization of return without considering risk. In contrast, 600,029, 600,031, 600,048, 600,050, 600,089, 600,100, 600,156,
600,547, 600,637, 600,690, 600,837, 600,887, 600,999, 601,006,
λ = 1 means that investors seek to minimize risk without con-
601,088, 601,111, 601,166, 601,186, 601,688, 601,766, 601,988.
sidering a return. The value between the two extremes balances Data from November 2009 to November 2019 are collected for
the expected return maximization and risk minimization. As a analysis and divided into a training set and a testing set as a
result, investors can choose one of these possible solutions based ratio of 8:2. The training set is used to train the model and adjust
on their risk preference to form the appropriate portfolio. the hyperparameters to get good generalization. The test set is
5
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

Table 2 4.2. Proposed model: Ifaxgboost+mv

Summary statistics for candidate assets.
Stock Mean Std. Max. Min. Range In the financial market, individual investors usually would like
600,000 12.69 3.68 24.19 7.11 17.08 to know which measures should be adopted to help them possess
600,018 5.14 1.67 10.65 2.38 8.27 the best portfolio of maximal potential return with minimal risk.
600,028 6.52 1.78 14.14 4.1 10.04
Therefore, this paper mainly focuses on selecting stocks with high
600,029 6.59 2.54 15.98 2.26 13.72
600,031 11.95 7.66 41.05 4.63 36.42
returns to form a portfolio from the perspective of prediction.
600,048 11.37 3.41 27.69 4.86 22.83 The process has two critical phases, namely stock prediction (for
600,050 5.16 1.40 10.54 2.98 7.56 obtaining the future prices of stocks) and portfolio selection (for
600,089 11.01 4.52 25.72 5.52 20.2 forming optimal portfolio), with the overall flowchart shown in
600,100 13.24 5.83 32.56 6.43 26.13 Fig. 2.
600,156 6.44 2.64 17.45 3.12 14.33
600,547 34.81 14.05 91.88 15.2 76.68
(1) Stock prediction: XGBoost is employed for predicting the
600,637 20.88 12.15 77.23 5.7 71.53 prices of all candidate assets for the next period. Furthermore,
600,690 16.09 5.60 32.2 7.48 24.72 hyperparameters of the XGBoost are optimized by the developed
600,837 12.72 4.03 30.22 7.08 23.14 IFA during the training process. Mean square error (MSE) is
600,887 26.41 7.52 51.29 12.84 38.45 calculated and input into the IFA as a fitness function. The main
600,999 16.80 5.72 38.3 8.19 30.11
hyperparameters, such as subsample, learning_rate, reg_alpha,
601,006 8.11 1.45 14.32 5.63 8.69
601,088 20.78 5.10 37.99 12.85 25.14 reg_lambda, colsample_bytree, colsample_bylevel, gamma,
601,111 8.06 2.75 16.86 3.23 13.63 min_child_weight, are updated in the minimization process until
601,166 17.13 6.30 41.73 8.65 33.08 the prediction error MSE meets the requirements. The prediction
601,186 8.85 3.96 27.09 3.7 23.39 model named IFAXGBoost is applied to predict the stock price for
601,688 15.20 5.35 33.29 7.27 26.02
the next period once the optimal hyperparameters are obtained.
601,766 7.83 3.86 35.88 3.34 32.54
601,988 3.43 0.56 5.6 2.45 3.15
(2) Portfolio selection: This phase aims to determine the
capital proportion allocated to each asset. In this paper, the MV
model is used to conduct the portfolio’s asset allocation based on
Table 3 the selected high-quality asset and establish the optimal portfolio
Input indicators summary.
with unequal proportions. To be specific, the Monte Carlo method
Attribute Details Attribute Details is applied to generate different portfolios. That is, randomly create
( ) ( )
1 ln
close pricei
close pricei−1
11 ln
high pricei−3
open pricei−3
a set of weights and then calculate the mean return and vari-
(
close pricei−1
) (
low pricei
) ance of portfolios under the weight. Repeat the process 50,000
2 ln 12 ln
( close pricei−2
) ( open pricei
) times. From a statistical point of view, 50,000 random portfolios
close pricei−2 low pricei−1
3 ln close pricei−3
13 ln open pricei−1 cover most possible portfolios with different weights and are
( ) ( )
4 ln
close pricei−3
14 ln
low pricei−2 sufficiently representative [16]. Besides, the best one is chosen
close pricei−4 open pricei−2
(
high pricei
) (
low pricei−3
) from these 50,000 portfolios, according to the MV model. To find
5 ln 15 ln
( open pricei
) open pricei−3 a balance between return and risk, the Sharpe ratio is used to
high pricei
6 ln open pricei−1
16 True range make better decisions, and available resources are allocated to the
portfolio with the largest Sharpe ratio.
( )
high pricei
7 ln open pricei−2
17 Average true range
(period = 14)
4.3. Benchmarks with other alternative models
( )
high pricei
8 ln open pricei−3
18 Momentum index
(period = 10)
In addition to the MV model, equal-weight portfolio (1/N)
( )
high pricei−1
9 ln open pricei−1
19 Relative strength
index has also been studied by some scholars [2,62]. The following
( ) (period = 14) alternative models are based on the IFAXGBoost+MV and used for
high pricei−2
10 ln open pricei−2
comparison.
(1) Alternative model: IFAXGBoost+1/N
The design of this model is similar to the structure of IFAXG-
Boost+MV, using the same prediction model to select assets, but
used to evaluate the performance of the final model. Table 2 the asset proportion is evenly distributed. Specifically, in the first
demonstrates the summary statistics of the close prices for the phase, the IFAXGBoost is used to predict the price of assets. In
24 stocks. the second phase, the top k assets with higher expected returns
are selected and allocated the same proportion. The goal of the
4.1.2. Input indicators alternative model is to test the effectiveness of the MV model
In existing stock forecasting studies, various transaction data with the same asset selection.
(especially price factors) are used as important factors, because (2) Alternative model: Machine learning+MV or 1/N
These models are designed to determine whether different
they can better reflect the stock market information [2]. Basak
prediction methods impact the construction of the optimal port-
et al. [4] suggested that the average true range (ATR), moving
folio. To be specific, stock prices are predicted using the XGBoost,
average (MA), on balance volume (OBV), etc. are correlated with
LSTM, or SVR, and stocks with higher potential returns are se-
changes in the stock market. Zhou et al. [13] selected relative
lected for the next stage. Then, the composition of the portfolio
strength index (RSI), momentum index (MTM), etc. as feature is determined by the MV or 1/N method.
subsets. Moreover, financial time-series forecasting is always ex- (3) Alternative model: Random+MV or 1/N
plained by the lagged observations. For example, Paiva et al. [50] Unlike the IFAXGBoost+MV in the stock prediction phase, the
used several lagged variables and technical indicators to predict Random+MV or 1/N model is carried out randomly and does not
future stock prices. In this paper, we select 19 indicators as the rely on any prediction. Specifically, we randomly select a number
input of stock prediction, including 15 lagged return observations of assets from all samples and then apply MV or 1/N method to
and 4 technical indicators. To reduce the prediction errors, all determine the portfolio. The purpose of these alternative models
indicators are uniformly mapped to the interval [−1, 1]. Table 3 is to examine the necessity for stock prediction using machine
summarizes the selected input indicators. learning.
6
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

Fig. 2. Overall flowchart of IFAXGBoost+MV.

5. Experimental results SSAXGBoost (hyperparameters of XGBoost optimized by SSA), and

single models including XGBoost, LSTM, ELM, SVR and ANN. To
5.1. Performance of the proposed IFA investigate the accuracy of stock prediction methods, as in [23]
and [35], four indexes, namely mean absolute percentage error
In this section, a set of classical functions are selected to (MAPE), mean square error (MSE), mean absolute error (MAE),
evaluate the performance of the IFA, including three unimodal and root mean square error (RMSE), are used in this paper. MSE
functions and three multimodal functions. Table 4 presents the and RMSE often represent the dispersion of results, while MAPE
detailed descriptions of these functions. In the following experi- and MAE indicate the deviation of results. These indexes are
ments, we compare the IFA with the standard FA, particle swarm defined as follows:
optimization (PSO) [59], salp swarm algorithm (SSA) [63], grav- n
⏐ ⏐
1 ∑ ⏐ŷi − yi ⏐
itational search algorithm (GSA) [64], and artificial fish swarm MAPE = , (22)
algorithm (AFSA) [65]. The detailed parameters are given in Ta- n yi
i=1
ble 5. n
1∑ 2
Table 6 shows the maximum, minimum, mean, standard devi- MSE = (yi − ŷi ) , (23)
ation, range, and average time of the different algorithms based n
i=1
on 20 independent runs. It can be seen that in all cases, the IFA is n
1 ∑⏐
⏐yi − ŷi ⏐,
⏐
superior to other algorithms in terms of the maximum, minimum, MAE = (24)
mean. In most cases, IFA is better than other algorithms according n
i=1
to the standard deviation and range. However, the IFA requires 
 n
more calculation time than the PSO, SSA, and GSA. Also, Fig. 3 1 ∑ 2
RMSE = √ (yi − ŷi ) , (25)
provides the convergence curves of the IFA, FA, PSO, SSA, GSA and n
i=1
AFSA. We can see that the IFA has faster convergence speed and
higher accuracy than other algorithms. To sum up, the conver- where ŷi represents the predicted price, yi indicates the actual
gence accuracy, stability, and robustness of the IFA are superior price, and n is the total number of samples. Moreover, as in [66],
to the FA, PSO, SSA, GSA and AFSA. the training time is used as an indicator to reflect the complexity
of different prediction models.
5.2. Stock price prediction results From Tables 7 to 9, we can see that the error indexes of the
IFAXGBoost are the smallest among all prediction models. For
Reasonable and accurate forecasts have the potential to gen- example, compared with the FAXGBoost, the MAPE, MSE, MAE,
erate high investment returns and hedge risks. We compare the and RMSE are reduced by 2.402, 0.0004, 0.0035, and 0.0047,
proposed IFAXGBoost with combined models, including FAXG- respectively, demonstrating the effectiveness of the improvement
Boost (hyperparameters of XGBoost optimized by FA), PSOXG- strategy of the IFA. Compared with the XGBoost, the MAPE, MSE,
Boost (hyperparameters of XGBoost optimized by PSO) and MAE, and RMSE are reduced by 6.01, 0.003, 0.0178, and 0.0215,
7
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

Table 4
The description of functions in the experiment.
Name Function Range Global optimum
D
xi 2
∑
Sphere f1 (x) = [−100, 100] 0
i=1
∑D D
∏
Schwefel 2.22 f2 (x) = |xi | + | xi | [−10, 10] 0
i=1 i=1
D
ixi 2
∑
Sum square f3 (x) = [−10, 10] 0
i=1
D
(xi 2 − 10 cos 2π xi + 10)
∑
Rastrigin f4 (x) = [−5.12, 5.12] 0
i=1 ( √ )
D D
f5 (x) = −20 exp −0.2 1
− exp( D1 cos 2π xi ) + 20 + e
∑ ∑
Ackley D
xi 2 [−32, 32] 0
i=1 i=1
D D
1 x
xi 2 −
∑ ∏
Griewank f6 (x) = 4000
cos( √i ) + 1 [−600, 600] 0
i
i=1 i=1

Table 5
Parameter settings of different algorithms.
Parameters Population size Dimension of function Maximum generation
60 30 100
IFA β0 γ α0
0.8 0.0001 0.3
FA β0 γ α0
0.8 0.0001 0.3
PSO Inertia constant ω Cognitive constant c1 Social constant c2
0.8 1.5 1.5
SSA Constant c2 Constant c3
[0, 1] [0, 1]
GSA Constant G0 Constant α
100 20
AFSA Visual parameters α Step parameters β Constant δ
0.01 0.001 40

Table 6
Comparison of benchmark functions.
Functions Algorithms Maximum Minimum Mean Std. Range Time (s)
Sphere IFA 6.02e−16 2.85e−19 8.33e−17 1.47e−16 6.02e−16 240
FA 0.56456 0.31269 0.45392 0.05742 0.25187 217.2
PSO 980.040 6.81399 183.380 255.916 973.226 7.2
SSA 1183.08 152.543 540.519 301.025 1030.54 11.4
GSA 68968.2 38954.6 57461.1 7627.61 30013.6 138
AFSA 6078.47 3124.79 5016.66 1161.53 2953.68 1432.2
Schwefel 2.22 IFA 3.29e−17 3.96e−20 7.18e−18 9.27e−18 3.29e−17 439.8
FA 3.14913 2.73415 2.94920 0.13482 0.41497 437.4
PSO 11.6648 0.03378 1.79326 3.00335 11.6310 18
SSA 15.6728 1.84997 11.7304 5.63977 13.8228 6.6
GSA 73.6206 41.8232 52.1662 8.26753 31.7974 156.0
AFSA 102.310 91.0685 97.0715 5.02428 11.2417 1864.2
Sum square IFA 2.31e−17 1.73e−21 3.34e−18 5.78e−18 2.31e−17 217.8
FA 1.91345 1.33919 1.63745 0.16477 0.57425 273.6
PSO 10.1369 0.00461 2.14215 2.67684 10.1323 7.8
SSA 47.3368 7.26818 21.0227 10.7745 40.0686 12.6
GSA 798.116 282.106 493.112 121.951 516.010 147.6
AFSA 2108.52 1360.18 1771.28 203.430 748.340 1554
Rastrigin IFA 1.06e−14 0 2.57e−15 2.19e−15 1.06e−14 246.6
FA 186.856 129.501 160.166 16.6047 57.3548 301.2
PSO 74.9568 0.18701 23.7953 21.9883 74.7697 10.2
SSA 4.36474 0.97336 2.05947 0.77977 3.39137 10.8
GSA 375.902 228.156 314.460 38.7628 147.746 143.4
AFSA 356.465 289.224 334.102 15.9569 67.2409 1655.4
Ackley IFA 1.54e−08 1.13e−11 2.16e−09 3.58e−09 1.54e−08 243
FA 16.7882 13.1109 15.5079 0.98239 3.67728 297.0
PSO 7.60448 0.03696 3.24591 2.37827 7.56752 9.6
SSA 9.11950 6.71744 7,42947 0.99412 2.40206 11.4
GSA 20.0653 17.8782 19.3777 0.58647 2.18706 145.8
AFSA 20.0474 19.9629 20.0226 0.03426 0.08449 978.6
Griewank IFA 0.94119 0.00015 0.30975 0.29299 0.94103 379.8
FA 646.786 463.599 572.193 51.0674 183.186 336.0
PSO 10.9518 0.09625 2.15254 2.65646 10.8555 10.2
SSA 11.2893 3.15138 8.64558 3.44704 8.13795 11.4
GSA 663.006 459.482 566.064 64.6618 203.523 141.6
AFSA 3.64344 3.16135 3.40205 0.20722 0.48208 976.2

8
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

Fig. 3. The convergence characteristics of different algorithms.

Table 7
Comparison of IFAXGBoost, FAXGBoost and PSOXGBoost.
Stock IFAXGBoost FAXGBoost PSOXGBoost
MAPE MSE MAE RMSE Time (s) MAPE MSE MAE RMSE Time (s) MAPE MSE MAE RMSE Time (s)
600,000 18.5540 0.0027 0.0418 0.0524 1155.51 19.3440 0.0028 0.0441 0.0538 1281.30 20.4087 0.0031 0.0456 0.0556 514.87
600,018 14.1082 0.0069 0.0623 0.0736 1690.66 14.4579 0.0073 0.0698 0.0850 1472.04 14.4657 0.0070 0.0687 0.0836 782.99
600,028 19.2996 0.0018 0.0324 0.0428 1449.16 20.7184 0.0024 0.0360 0.0491 1414.57 20.8317 0.0025 0.0372 0.0495 733.72
600,029 12.4610 0.0048 0.0537 0.0694 1585.17 13.2132 0.0054 0.0570 0.0736 1675.70 13.0103 0.0051 0.0555 0.0711 773.91
600,031 19.6052 0.0010 0.0264 0.0318 1577.08 20.3144 0.0011 0.0273 0.0342 1643.54 24.2381 0.0015 0.0319 0.0382 577.73
600,048 12.6359 0.0034 0.0471 0.0587 1271.01 13.2466 0.0037 0.0486 0.0608 1381.16 12.7312 0.0035 0.0477 0.0590 523.89
600,050 14.9606 0.0050 0.0576 0.0710 823.56 14.4151 0.0056 0.0575 0.0748 821.84 16.6843 0.0075 0.0670 0.0866 547.91
600,089 33.9649 0.0015 0.0300 0.0393 1585.02 47.3970 0.0024 0.0381 0.0493 1595.27 45.5177 0.0024 0.0374 0.0493 773.79
600,100 24.2758 0.0017 0.0310 0.0414 1581.76 28.8860 0.0026 0.0366 0.0516 1545.89 27.8571 0.0024 0.0357 0.0494 674.67
600,156 66.0393 0.0037 0.0411 0.0613 1514.53 73.3475 0.0041 0.0460 0.0646 1471.99 79.5371 0.0046 0.0454 0.0676 488.53
600,547 22.4554 0.0023 0.0378 0.0484 1108.28 26.1129 0.0026 0.0415 0.0516 1080.06 23.8798 0.0026 0.0402 0.0510 510.61
600,637 32.3155 0.0011 0.0256 0.0343 991.75 40.4817 0.0018 0.0313 0.0432 1006.04 34.8962 0.0013 0.0274 0.0366 593.03
600,690 13.8130 0.0041 0.0477 0.0643 1095.14 14.9120 0.0044 0.0516 0.0670 901.87 15.0307 0.0048 0.0533 0.0690 576.52
600,837 27.6460 0.0037 0.0474 0.0609 1124.58 31.3699 0.0043 0.0509 0.0663 1240.57 33.0029 0.0047 0.0545 0.0687 680.46
600,887 14.8002 0.0053 0.0579 0.0727 955.33 15.9756 0.0059 0.0616 0.0769 1303.24 15.6084 0.0057 0.0596 0.0751 671.01
600,999 24.9024 0.0039 0.0519 0.0620 1025.47 25.0075 0.0039 0.0529 0.0627 927.93 24.5283 0.0038 0.0522 0.0615 624.31
601,006 15.0479 0.0030 0.0463 0.0551 980.37 14.9320 0.0029 0.0453 0.0546 925.34 14.6751 0.0031 0.0453 0.0560 608.70
601,088 15.3223 0.0028 0.0412 0.0528 1175.61 16.6750 0.0032 0.0442 0.0565 1178.09 16.5083 0.0031 0.0437 0.0555 679.60
601,111 12.6768 0.0043 0.0532 0.0656 1218.12 12.6088 0.0041 0.0525 0.0645 1138.27 13.4588 0.0049 0.0566 0.0702 720.45
601,166 11.4587 0.0013 0.0302 0.0365 1046.07 13.9372 0.0017 0.0343 0.0421 955.04 14.1385 0.0016 0.0339 0.0399 670.69
601,186 7.7528 0.0007 0.0216 0.0270 892.46 10.6040 0.0014 0.0303 0.0386 800.23 9.2715 0.0012 0.0265 0.0342 582.84
601,688 12.4458 0.0044 0.0535 0.0668 1651.49 13.9070 0.0051 0.0580 0.0715 1592.63 12.5883 0.0045 0.0534 0.0670 729.27
601,766 12.2179 0.0006 0.0204 0.0261 1236.32 14.1649 0.0009 0.0252 0.0310 1152.85 16.7428 0.0013 0.0275 0.0353 644.54
601,988 10.2852 0.0029 0.0426 0.0546 1340.20 10.6571 0.0033 0.0447 0.0582 1317.20 10.7130 0.0034 0.0442 0.0581 701.72
Avg. 19.5435 0.0030 0.0416 0.0528 1253.11 21.9452 0.0035 0.0452 0.0576 1242.61 19.5995 0.0036 0.0454 0.0578 641.07

9
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

Table 8
Comparison of SSAXGBoost, XGBoost and SVR.
Stock SSAXGBoost XGBoost SVR
MAPE MSE MAE RMSE Time (s) MAPE MSE MAE RMSE Time (s) MAPE MSE MAE RMSE Time (s)
600,000 18.7252 0.0029 0.0438 0.0539 704.06 19.4847 0.0038 0.0476 0.0622 0.327 22.8989 0.0044 0.0530 0.0664 0.124
600,018 15.3385 0.0082 0.0736 0.0905 961.55 16.9369 0.0103 0.0845 0.1018 0.494 13.2312 0.0071 0.0642 0.0846 0.180
600,028 21.3667 0.0024 0.0373 0.0488 921.55 22.0892 0.0033 0.0417 0.0580 0.535 30.2968 0.0081 0.0492 0.0900 0.138
600,029 13.7526 0.0060 0.0605 0.0771 916.57 14.9087 0.0089 0.0712 0.0947 0.501 20.6076 0.0057 0.0662 0.0756 0.157
600,031 21.0408 0.0012 0.0282 0.0344 861.90 24.4162 0.0014 0.0316 0.0380 0.489 25.3572 0.0016 0.0329 0.0409 0.111
600,048 13.0461 0.0040 0.0485 0.0628 745.95 18.8528 0.0075 0.0722 0.0868 0.370 15.3858 0.0070 0.0576 0.0840 0.075
600,050 17.1647 0.0060 0.0640 0.0777 702.78 17.5527 0.0085 0.0699 0.0923 0.337 24.1904 0.0132 0.0980 0.1149 0.099
600,089 45.5801 0.0022 0.0374 0.0472 906.49 45.3379 0.0020 0.0371 0.0457 0.463 79.4496 0.0047 0.0612 0.0687 0.126
600,100 28.2045 0.0024 0.0370 0.0494 731.26 26.7376 0.0027 0.0362 0.0528 0.450 51.0036 0.0048 0.0601 0.0695 0.137
600,156 82.6897 0.0051 0.0473 0.0711 729.68 78.8494 0.0067 0.0591 0.0823 0.304 114.557 0.0080 0.0773 0.0897 0.081
600,547 24.9897 0.0026 0.0405 0.0511 659.87 22.8197 0.0038 0.0435 0.0619 0.342 28.1979 0.0053 0.0525 0.0734 0.086
600,637 40.3576 0.0016 0.0303 0.0401 644.12 36.4610 0.0014 0.0284 0.0374 0.280 93.4172 0.0049 0.0616 0.0705 0.045
600,690 15.3900 0.0048 0.0524 0.0694 722.27 24.5663 0.0106 0.0868 0.1032 0.302 14.4138 0.0058 0.0512 0.0768 0.089
600,837 36.6201 0.0052 0.0579 0.0724 793.80 32.2967 0.0049 0.0546 0.0702 0.312 51.5690 0.0054 0.0613 0.0738 0.094
600,887 17.7203 0.0062 0.0647 0.0787 734.81 20.2491 0.0100 0.0786 0.1002 0.332 17.4289 0.0081 0.0698 0.0904 0.087
600,999 25.1544 0.0041 0.0544 0.0640 744.73 23.3568 0.0050 0.0574 0.0707 0.305 18.0374 0.0025 0.0418 0.0502 0.092
601,006 14.9304 0.0031 0.0460 0.0554 737.14 28.8123 0.0102 0.0918 0.1011 0.300 26.6909 0.0091 0.0870 0.0957 0.090
601,088 17.0726 0.0034 0.0456 0.0585 769.07 25.1895 0.0074 0.0714 0.0862 0.301 16.5613 0.0052 0.0492 0.0727 0.109
601,111 14.3416 0.0052 0.0596 0.0718 732.09 14.6090 0.0053 0.0602 0.0733 0.299 11.4462 0.0042 0.0486 0.0649 0.091
601,166 14.2486 0.0018 0.0349 0.0428 842.65 24.8959 0.0054 0.0638 0.0738 0.275 20.1787 0.0036 0.0503 0.0602 0.072
601,186 12.5179 0.0017 0.0348 0.0411 806.90 16.4512 0.0034 0.0463 0.0588 0.297 23.0714 0.0058 0.0670 0.0764 0.084
601,688 15.5370 0.0057 0.0633 0.0755 1005.3 16.5816 0.0073 0.0700 0.0858 0.318 19.0236 0.0092 0.0823 0.0962 0.107
601,766 14.9853 0.0011 0.0260 0.0328 963.22 21.3001 0.0020 0.0365 0.0447 0.262 14.8056 0.0015 0.0290 0.0387 0.032
601,988 10.4294 0.0035 0.0445 0.0590 775.10 20.4609 0.0108 0.0853 0.1042 0.275 26.4989 0.0154 0.1086 0.1244 0.124
Avg. 22.9668 0.0038 0.0472 0.0594 796.37 25.5507 0.0060 0.0594 0.0744 0.353 32.4200 0.0063 0.0616 0.0770 0.101

Table 9
Comparison of LSTM, ELM and ANN.
Stock LSTM ELM ANN
MAPE MSE MAE RMSE Time (s) MAPE MSE MAE RMSE Time (s) MAPE MSE MAE RMSE Time (s)
600,000 25.5598 0.0081 0.0619 0.0897 21.756 25.6666 0.0055 0.0590 0.0747 0.094 24.3916 0.0065 0.0576 0.0805 11.412
600,018 12.3320 0.0055 0.0601 0.0743 30.156 17.9025 0.0131 0.0863 0.1147 0.129 23.7162 0.0171 0.1161 0.1308 16.408
600,028 29.8726 0.0080 0.0515 0.0895 29.256 30.2479 0.0115 0.0511 0.1077 0.135 30.2268 0.0052 0.0554 0.0725 17.425
600,029 11.6938 0.0046 0.0521 0.0678 29.173 13.2660 0.0055 0.0553 0.0747 0.117 19.5489 0.0132 0.0927 0.1149 16.957
600,031 22.5866 0.0021 0.0336 0.0462 24.143 20.8977 0.0014 0.0300 0.0384 0.117 32.1178 0.0032 0.0487 0.0569 17.235
600,048 14.1086 0.0051 0.0527 0.0711 27.195 13.8842 0.0044 0.0517 0.0668 0.135 27.1125 0.0133 0.1032 0.1154 13.081
600,050 14.7126 0.0062 0.0574 0.0790 23.006 18.3512 0.0095 0.0720 0.0979 0.104 29.1270 0.0175 0.1156 0.1325 13.901
600,089 49.7376 0.0031 0.0389 0.0556 28.443 50.4193 0.0030 0.0411 0.0552 0.088 52.1739 0.0023 0.0418 0.0489 15.833
600,100 23.2798 0.0016 0.0307 0.0395 32.226 38.4277 0.0033 0.0467 0.0580 0.117 37.1215 0.0034 0.0463 0.0584 17.759
600,156 86.3625 0.0056 0.0350 0.0751 26.096 88.9497 0.0063 0.0571 0.0795 0.086 54.6501 0.0022 0.0346 0.0478 13.439
600,547 25.9425 0.0067 0.0527 0.0817 25.513 26.1420 0.0077 0.0500 0.0880 0.086 31.8298 0.0061 0.0594 0.0778 15.148
600,637 50.3675 0.0032 0.0348 0.0564 24.502 87.5423 0.0053 0.0587 0.0728 0.072 36.5604 0.0014 0.0275 0.0385 12.310
600,690 18.0326 0.0072 0.0645 0.0849 26.969 14.9387 0.0053 0.0539 0.0731 0.071 20.7830 0.0082 0.0731 0.0908 13.378
600,837 36.6479 0.0056 0.0523 0.0747 27.224 38.4141 0.0042 0.0511 0.0655 0.080 40.8014 0.0068 0.0689 0.0828 14.407
600,887 17.0115 0.0073 0.0642 0.0855 28.204 16.2234 0.0071 0.0626 0.0846 0.079 21.1549 0.0101 0.0809 0.1007 14.953
600,999 17.9670 0.0040 0.0440 0.0629 29.324 18.7511 0.0034 0.0424 0.0587 0.070 27.6133 0.0062 0.0668 0.0789 14.961
601,006 30.0729 0.0110 0.0941 0.1050 29.506 24.7404 0.0079 0.0788 0.0894 0.082 35.0847 0.0136 0.1096 0.1169 15.147
601,088 22.0254 0.0072 0.0593 0.0850 29.825 17.7052 0.0060 0.0515 0.0777 0.083 17.9113 0.0045 0.0509 0.0676 14.932
601,111 12.6172 0.0058 0.0537 0.0764 30.685 12.9304 0.0065 0.0556 0.0810 0.080 12.6421 0.0046 0.0574 0.0684 15.286
601,166 23.0482 0.0049 0.0576 0.0699 35.006 16.3489 0.0026 0.0409 0.0510 0.087 28.7367 0.0065 0.0722 0.0811 16.110
601,186 16.0329 0.0033 0.0452 0.0570 31.664 19.5319 0.0045 0.0560 0.0673 0.082 17.9257 0.0039 0.0503 0.0627 16.208
601,688 12.4341 0.0049 0.0514 0.0699 30.727 14.3915 0.0062 0.0611 0.0788 0.083 15.1949 0.0060 0.0637 0.0779 15.804
601,766 27.5028 0.0027 0.0452 0.0520 31.024 15.1569 0.0012 0.0273 0.0346 0.076 19.1109 0.0016 0.0334 0.0409 16.194
601,988 30.0167 0.0187 0.1207 0.1367 32.795 25.6264 0.0142 0.1048 0.1195 0.086 19.8566 0.0088 0.0806 0.0943 17.094
Avg. 23.6349 0.0059 0.0547 0.0744 28.517 27.7690 0.0061 0.0560 0.0754 0.093 28.1413 0.0072 0.0669 0.0807 15.224

respectively, indicating the availability of hyperparameter opti- 5.3. Optimal portfolio selection results
mization using the IFA. Compared with the other hybrid models
(PSOXGBoost and SSAXGBoost), the prediction errors are also This section will determine the proper cardinality of the port-
reduced, which means that the prediction accuracy after hyper- folio and evaluate the effectiveness and superiority of the pro-
parameter tuning is better than before. Also, the hybrid models, posed IFAXGBoost+MV.
i.e., IFAXGBoost, FAXGBoost, PSOXGBoost, and SSAXGBoost per-
form better than the single XGBoost, which indicates that it 5.3.1. Determination the cardinality of portfolio
Holding too many different stocks is difficult to manage for
is useful to optimize the model by adjusting hyperparameters.
individual investors. As a result, many studies considered a port-
However, hybrid models need longer running times than single
folio with less than ten stocks. Paiva et al. [50] argued that an
models. The reason is that the hyperparameters of hybrid models average of seven assets per day is appropriate for each portfo-
are tuned using swarm intelligence algorithms, while those of lio. Wang et al. [16] discovered that a portfolio with ten assets
single models are set artificially. performs better than those with other amounts. Therefore, we
10
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

Fig. 4. Annualized performances of different portfolio’s cardinalities.

Table 10
Performance characteristics without transaction cost.
Model IFAXG- IFAXG- XG- XG- LSTM+MV LSTM+1/N SVR+MV SVR+1/N Random+MV Random+1/N
Boost+MV Boost+1/N Boost+MV Boost+1/N
Panel A: Daily return characteristics
Mean return 0.0014 0.0010 0.0009 0.0005 0.0011 0.0008 0.0007 0.0002 0.0004 0.0001
Maximum 0.0593 0.0527 0.0563 0.0637 0.0655 0.0691 0.0582 0.0458 0.0415 0.0444
Minimum −0.0698 −0.0530 −0.0744 −0.0663 −0.0574 −0.0572 −0.0592 −0.0662 −0.0613 −0.0674
Range 0.1292 0.1057 0.1307 0.1300 0.1229 0.1263 0.1050 0.1244 0.1029 0.1118
Skewness −0.0825 −0.0457 −0.4562 −0.3810 −0.2044 −0.0444 −0.2254 −0.2166 −0.6127 −0.4172
Kurtosis 4.5911 2.7562 5.3739 4.0368 2.5240 2.1481 1.6780 2.4564 2.6394 2.2962
Panel B: Daily risk characteristics
Standard dev. 0.0118 0.0122 0.0128 0.0133 0.0143 0.0145 0.0134 0.0138 0.0133 0.0135
Downside dev. 0.0083 0.0087 0.0094 0.0098 0.0103 0.0103 0.0096 0.0099 0.0098 0.0098
1-percent VaR 0.0260 0.0274 0.0288 0.0303 0.0320 0.0328 0.0303 0.0317 0.0304 0.0313
1-percent CVaR 0.0300 0.0316 0.0331 0.0348 0.0369 0.0377 0.0348 0.0364 0.0349 0.0359
5-percent VaR 0.0179 0.0191 0.0200 0.0212 0.0223 0.0229 0.0212 0.0223 0.0213 0.0220
5-percent CVaR 0.0229 0.0242 0.0254 0.0268 0.0283 0.0290 0.0268 0.0281 0.0269 0.0277
Panel C: Annualized risk-return metrics
Mean return 0.4354 0.3004 0.2767 0.1623 0.3156 0.2284 0.1826 0.0624 0.1026 0.0330
Standard dev. 0.1879 0.1950 0.2044 0.2120 0.2270 0.2305 0.2124 0.2189 0.2113 0.2154
Downside dev. 0.1332 0.1386 0.1501 0.1559 0.1628 0.1639 0.1522 0.1576 0.1561 0.1570
Sharpe ratio 2.1576 1.3862 1.2067 0.6240 1.2582 0.8608 0.7186 0.1478 0.3439 0.0140
Sortino ratio 3.0441 1.9503 1.6437 0.8488 1.7541 1.2106 1.0030 0.2053 0.4655 0.0193
Panel D: Statistical tests for daily returns
t-statistic 2.9190 2.0640 1.7544 0.9437 2.2995 1.4785 1.0094 0.1880 0.2610 N.A.
p-value 0.0037 0.0396 0.0800 0.3458 0.0219 0.1400 0.3133 0.8510 0.7942 N.A.

choose stocks with cardinality k = 6, 7, 8, 9, 10 to form port- risk. Also, the Sortino ratio is similar to the sharpe ratio, except
folios. Moreover, annualized mean return, annualized standard that it can distinguish between good and bad fluctuations. In this
deviation, annualized Sharpe ratio, and annualized Sortino ratio
are selected to evaluate portfolios’ performance. The Sharpe ratio paper, the return of a risk-free asset is set to 0.03, according to
is a comprehensive indicator that can consider both return and the China treasuring bill rate in recent 10 years.
11
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

Table 11
Performance characteristics with transaction cost (0.5‰).
Model IFAXGBoost+MV IFAXGBoost+1/N XGBoost+MV XGBoost+1/N LSTM+MV LSTM+1/N SVR+MV SVR+1/N Random+MV Random+1/N
Panel A: Daily return characteristics
Mean return 0.0009 0.0005 0.0005 0.00008 0.0006 0.0003 0.0002 −0.0003 0.00004 −0.0003
Maximum 0.0588 0.0522 0.0558 0.0632 0.0650 0.0686 0.0577 0.0453 0.0454 0.0439
Minimum −0.0703 −0.0535 −0.0749 −0.0668 −0.0579 −0.0577 −0.0597 −0.0667 −0.0618 −0.0679
Range 0.1292 0.1057 0.1307 0.1300 0.1229 0.1263 0.1050 0.1244 0.1072 0.1118
Skewness −0.0825 −0.0457 −0.4686 −0.3810 −0.2048 −0.0444 −0.2281 −0.2166 −0.4598 −0.4172
Kurtosis 4.5911 2.7562 5.4415 4.0368 2.5259 2.1481 1.7002 2.4564 2.6009 2.2962
Panel B: Daily risk characteristics
Standard dev. 0.0118 0.0122 0.0128 0.0133 0.0143 0.0145 0.0134 0.0138 0.0131 0.0135
Downside dev. 0.0083 0.0087 0.0094 0.0098 0.0103 0.0103 0.0096 0.0099 0.0096 0.0098
1-percent VaR 0.0265 0.0279 0.0293 0.0308 0.0325 0.0333 0.0308 0.0322 0.0304 0.0317
1-percent CVaR 0.0305 0.0321 0.0336 0.0353 0.0374 0.0382 0.0353 0.0369 0.0349 0.0364
5-percent VaR 0.0184 0.0196 0.0205 0.0217 0.0228 0.0234 0.0217 0.0228 0.02147 0.0225
5-percent CVaR 0.0234 0.0247 0.0259 0.0273 0.0288 0.0295 0.0273 0.0286 0.0269 0.0282
Panel C: Annualized risk-return metrics
Mean return 0.2650 0.1416 0.1229 0.0203 0.1607 0.0831 0.0467 −0.0634 0.0097 −0.0929
Standard dev. 0.1879 0.1948 0.2036 0.2118 0.2270 0.2305 0.2119 0.2188 0.2087 0.2151
Downside dev. 0.1332 0.1384 0.1498 0.1556 0.1628 0.1639 0.1518 0.1575 0.1524 0.1567
Sharpe ratio 1.2414 0.5731 0.4561 −0.0454 0.5756 0.2302 0.0790 −0.4272 −0.0967 −0.5714
Sortino ratio 1.7500 0.8068 0.6200 −0.0619 0.8025 0.3238 0.1103 −0.5934 −0.1325 −0.7840
Panel D: Statistical tests for daily returns
t-statistic 2.9218 2.0640 1.7625 0.9437 2.0042 1.4788 1.0453 0.1880 0.6891 N.A.
p-value 0.0037 0.0396 0.0786 0.3458 0.0456 0.1399 0.2964 0.8510 0.4911 N.A.

The different models’ annualized performances under cardi- Following Panel A of Table 10, it is evident that the IFAXG-
nality k = 6, 7, 8, 9, 10 are shown in Fig. 4. It can be seen that Boost+MV has the highest daily mean return: 0.0014, the
the IFAXGBoost+MV is superior to the alternative models in terms LSTM+MV follows, with a return of 0.0011, and then the IFAXG-
of mean return, standard deviation, Sharpe ratio and Sortino Boost+1/N, with 0.0010. After considering the transaction cost, as
ratio. Specifically, when the cardinality k = 7, the return of shown in Tables 11 and 12, the IFAXGBoost+MV achieves the first
the IFAXGBoost+MV is 0.43, compared to 0.32 for the LSTM+MV, place in terms of daily mean return.
0.30 for the IFAXGBoost+1/N, 0.27 for the XGBoost+MV, 0.23 for (2) Panel B: comparison of daily risk characteristics
the LSTM+1/N, 0.18 for the SVR+MV, 0.16 for the XGBoost+1/N, In Panel B of Tables 10 to 12, we use standard deviation,
0.10 for the Random+MV, 0.06 for the SVR+1/N and 0.03 for downside deviation, value at risk (VaR), and conditional value at
the Random+1/N. As the annualized standard deviation of risk risk (CVaR) to measure risk. We can see that: (1) When transac-
measurement, the IFAXGBoost+MV has the lowest risk when the tion costs are ignored, the IFAXGBoost+MV achieves the lowest
cardinality k = 6, 7, 8, 9, 10. Considering the sharpe ratio, it risk, with a standard deviation of 0.0118, downside deviation of
is clear that the IFAXGBoost+MV achieves the highest level of 0.0083, 1-percent VaR of 0.0260, 1-percent CVaR of 0.0300, 5-
1.98 for k = 8, with the IFAXGBoost+1/N coming in second with percent VaR of 0.0179, and 5-percent CVaR of 0.0229. (2) When
1.32. With regard to the sortino ratio, the IFAXGBoost provides transaction costs are considered, the IFAXGBoost+MV has the
favorable results. For example, when the cardinality k = 10, lowest standard deviation, downside deviation, 1-percent VaR,
the IFAXGBoost+MV model has the highest sortino ratio as 2.58, 1-percent CVaR, 5-percent VaR, 5-percent CVaR.
followed by IFAXGBoost+1/N (1.79), SVR+MV (1.77), LSTM+MV (3) Panel C: comparison of annualized risk return indicators
(1.69), LSTM+1/N (1.57), Random+MV (1.06), XGBoost+MV (0.85), From Panel C of Table 10, the IFAXGBoost+MV has the highest
SVR+1/N(0.77), XGBoost+1/N (0.63) and Random+1/N (0.21). annualized return of 0.44. In addition, the Sharpe ratio of the
In general, when the cardinality k = 7, most models perform IFAXGBoost+MV reaches the highest level as 2.1576, and the
well in terms of mean return, standard deviation, Sharpe ratio, IFAXGBoost+1/N ranks the second with 1.3862. As for the Sortino
and Sortino ratio. Hence, in this paper, we choose the portfolio ratio, the IFAXGBoost+MV performs better, followed by IFAXG-
with k = 7 for subsequent analysis, which is consistent with Paiva Boost+1/N. After considering transaction costs of 0.5h and 1h ,
et al.’s research [50]. as shown in Tables 11 and 12, the IFAXGBoost+MV has the highest
annualized return, Sharpe ratio and Sortino ratio.
5.3.2. Superiority over benchmark models (4) Panel D: statistical tests for daily returns
Transaction cost is an essential factor that cannot be ignored in To prove the superiority of the proposed portfolio selection
quantitative investment research. In China’s stock market, trans- model statistically, one-tailed t-tests are carried out, and the
action costs mainly include transfer fees, stamp duty, commis- Random+1/N model is taken as the benchmark model. The testing
sions, and other components. Stamp duty adopts single transac- results, where the null hypothesis is that the difference between
tions, and additional fees mostly take the bilateral transaction. each model and the benchmark model is negative in terms of
This paper uses unilateral transaction costs and sets the trans- daily returns, further confirm the superiority of the proposed
action costs at 0, 0.5h and 1h. Tables 10 to 12 summarize IFAXGBooost+MV model, as the p-value is far less than the sig-
the performance of each model for different transaction costs, nificance level of 5% (see Panel D in Tables 10 to 12).
in which Panel A, B, C, and D describe daily return characteris- Moreover, Figs. 5 to 7 are presented to visualize the results
tics, daily risk characteristics, annualized risk-return metrics, and of Tables 10 to 12 through box plots of daily returns. It is
statistical tests for daily returns, respectively. observed that, when transaction costs are not considered, the
(1) Panel A: comparison of daily return characteristics IFAXGBoost+MV shows the lowest volatility, followed by the
12
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

Table 12
Performance characteristics with transaction cost (1‰).
Model IFAXGBoost+MV IFAXGBoost+1/N XGBoost+MV XGBoost+1/N LSTM+MV LSTM+1/N SVR+MV SVR+1/N Random+MV Random+1/N
Panel A: Daily return characteristics
Mean return 0.0004 0.00004 −0.00003 −0.0004 0.0001 −0.0002 −0.0003 −0.0008 −0.0005 −0.0008
Maximum 0.0583 0.0517 0.0553 0.0627 0.0645 0.0680 0.0448 0.0572 0.0449 0.0434
Minimum −0.0708 −0.0540 −0.0754 −0.0673 −0.0584 −0.0581 −0.0602 −0.0672 −0.0544 −0.0684
Range 0.1292 0.1057 0.1307 0.1300 0.1229 0.1262 0.1050 0.1244 0.0993 0.1118
Skewness −0.0825 −0.0457 −0.4683 −0.2273 −0.2048 −0.0444 −0.3810 −0.2166 −0.3590 −0.4172
Kurtosis 4.5911 2.7562 5.4381 4.0368 2.5259 2.1481 1.6805 2.4564 1.8954 2.2962
Panel B: Daily risk characteristics
Standard dev. 0.0118 0.0122 0.0128 0.0133 0.0143 0.0145 0.0134 0.0138 0.0127 0.0135
Downside dev. 0.0083 0.0087 0.0094 0.0098 0.0103 0.0103 0.0096 0.0099 0.0092 0.0098
1-percent VaR 0.0270 0.0284 0.0298 0.0313 0.0330 0.0338 0.0313 0.0327 0.0301 0.0322
1-percent CVaR 0.0310 0.0326 0.0341 0.0358 0.0379 0.0387 0.0358 0.0374 0.0345 0.0369
5-percent VaR 0.0189 0.0201 0.0210 0.0222 0.0233 0.0239 0.0222 0.0233 0.0214 0.0230
5-percent CVaR 0.0239 0.0252 0.0264 0.0278 0.0293 0.0300 0.0278 0.0291 0.0267 0.0287
Panel C: Annualized risk-return metrics
Mean return 0.1177 0.0106 −0.0079 −0.0966 0.0233 −0.0453 −0.0813 −0.1744 −0.1327 −0.1972
Standard dev. 0.1879 0.1949 0.2038 0.2119 0.2270 0.2303 0.2121 0.2186 0.2027 0.2151
Downside dev. 0.1332 0.1385 0.1499 0.1557 0.1628 0.1638 0.1523 0.1577 0.1474 0.1569
Sharpe ratio 0.4667 −0.0993 −0.1863 −0.5978 −0.0295 −0.3271 −0.5247 −0.9349 −0.8029 −1.0556
Sortino ratio 0.6580 −0.1397 −0.2532 −0.8131 −0.0412 −0.4600 −0.7307 −1.2958 −1.1037 −1.4477
Panel D: Statistical tests for daily returns
t-statistic 2.9208 2.0640 1.7597 0.9437 1.9738 1.4785 1.0077 0.1880 0.4182 N.A.
p-value 0.0037 0.0396 0.0791 0.3458 0.0490 0.1400 0.3141 0.8510 0.6760 N.A.

Fig. 5. Box plot of the daily returns without transaction costs.

XGBoost+MV; and when transaction costs are considered, the 0.06 for the Random+1/N. After considering the transaction costs
range of Random+MV is the lowest, but it has more outliers. 0.5h and 1h, as can be seen from Figs. 9 and 10, respectively,
To sum up, the IFAXGBoost+MV performs significantly better the cumulative return of each model decreases significantly, but
than the models based on XGBoost, LSTM, SVR, and random the IFAXGBoost+MV maintains the highest cumulative return.
selection in terms of return characteristics, risk characteristics, Furthermore, inspired by the studies [16,50], we need to prove
and risk-return metrics. whether the good performance of the IFAXGBoost+MV only oc-
curs within a certain period. Therefore, we compare the return–
5.3.3. Visualization of models performance
risk ratio of each model every quarter. As shown in Fig. 11,
To better show the superiority of the proposed IFAXGBoost+
the IFAXGBoost+MV achieves positive return–risk ratio, except
MV, we visualize the results in different forms. Fig. 8 demon-
strates the accumulative returns of each model, excluding trans- for Q1 2018 and Q2 2018, which are heavily influenced by the
action costs. The cumulative return of the IFAXGBoost+MV is the decline of the stock market in China. Also, Figs. 12 and 13 show
highest compared with the benchmarks. For example, the cumu- the quarterly return–risk ratio with transaction costs 0.5h and
lative return of the IFAXGBoost+MV is 0.94, compared to 0.62 1h , respectively. From Fig. 12, we can observe that the IFAXG-
for the IFAXGBoost+1/N, 0.56 for the XGBoost+MV, 0.31 for the Boost+MV achieves positive returns in the six of nine survey
XGBoost+1/N, 0.65 for the LSTM+MV, 0.46 for the LSTM+1/N, 0.36 quarters. After considering transaction costs 1h, only five survey
for the SVR+MV, 0.12 for the SVR+1/N, 0.28 for the Random+MV, quarters show that the IFAXGBoost+MV has good performance.
13
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

Fig. 6. Box plot of the daily returns with transaction cost (0.5h).

Fig. 7. Box plot of the daily returns with transaction cost (1h).

Fig. 8. Cumulative returns excluding transaction costs.

14
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

Fig. 9. Cumulative returns including transaction cost (0.5h).

Fig. 10. Cumulative returns including transaction cost (1h).

Overall, the IFAXGBoost+MV achieves favorable return–risk ratio Secondly, although the XGBoost has received much atten-
in most quarters. tion for its outstanding efficiency, it is not easy to accurately
predict the complexity and diversity of the input data. To im-
6. Discussion and conclusions prove the prediction accuracy and avoid the negative influence
of hyperparameter selection, the IFAXGBoost is developed for
6.1. Discussion for key findings predicting the prices of candidate assets, which are used for
portfolio selection. Further, IFAXGBoost is evaluated statistically.
This paper proposes a novel portfolio construction approach The outcomes of the IFAXGBoost are compared with the FAXG-
based on stock forecasting. The prediction model is developed Boost, PSOXGBoost, SSAXGBoost, XGBoost, LSTM, SVR, ANN, and
using a hybrid method based on the XGBoost and IFA, and the ELM. We can find that: (1) the IFAXGBoost can provide higher-
MV model is used for portfolio selection. In this paper, we have quality asset input for the portfolio selection; and (2) the hybrid
several findings. models based hyperparameter optimization can achieve higher
First of all, to improve the FA’s optimization ability, the IFA prediction accuracy than the single models.
is developed, which dynamically divides the whole firefly group Finally, prediction results are incorporated into the optimal
into an elite subgroup and an ordinary subgroup, and the chaotic portfolio formation; stocks with higher potential returns are se-
search strategy and the PSO-based search strategy are designed lected for constructing the portfolio. The results show the follow-
accordingly. After comparing the IFA against FA, PSO, SSA, AFSA as ing: (1) for individual investors, holding seven assets is appro-
well as GSA, the advantage of IFA is verified by a set of unimodal priate, (2) the IFAXGBoost+MV, IFAXGBoost+1/N, XGBoost+MV,
and multimodal test functions. XGBoost+MV, LSTM+MV, LSTM+1/N, SVR+MV, and SVM+1/N are
15
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

Fig. 11. Return–risk ratio of each quarter without transaction costs.

Fig. 12. Return–risk ratio of each quarter with transaction cost (0.5h).

Fig. 13. Return–risk ratio of each quarter with transaction cost (1h).

superior to Random+MV and Random+1/N in terms of return, risk, on the MV can obtain better results than those based on the
and risk-return ratio, which proves the necessity of incorporating 1/N, which indicate that the MV model plays an essential role
stock prediction into portfolio selection; and (3) the models based in portfolio selection. We also visualize the results in different
16
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

forms, such as box plot of daily return, line plot of cumulative [6] G. Box, G. Jenkins, Time series analysis: forecasting and control, J. Time 31
return, and bar chart of return–risk ratio, which can intuitively (4) (1976) 238–242.
[7] T. Bollerslev, Generalized autoregressive conditional heteroskedasticity, J.
reflect the superiority of the IFAXGBoost+MV.
Econom. 31 (3) (1986) 307–327.
In summary, the proposed IFAXGBoost+MV is superior to tra- [8] R. Bisoi, P. Dash, A. Parida, Hybrid Variational Mode Decomposition and
ditional methods (without stock prediction) and benchmarks in evolutionary robust kernel extreme learning machine for stock price
terms of returns, risks, and return–risk ratio. The reasons for and movement prediction on daily basis, Appl. Soft. Comput. 74 (2019)
this advantage include the following: (1) capturing the future 652–678.
[9] Y. Liu, I. Yeh, Using mixture design and neural networks to build stock
characteristics of stock markets, (2) improving the accuracy of selection decision support systems, Neural Comput. Appl. 28 (3) (2017)
forecasts by the developed IFAXGBoost, and (3) introducing the 521–535.
MV model to improve the efficiency of asset allocation. [10] W.C. Hong, M.W. Li, J. Geng, Y. Zhang, Novel chaotic bat algorithm for
forecasting complex motion of floating platforms, Appl. Math. Model. 72
(2019) 425–443.
6.2. Future works [11] R. Chen, C.Y. Liang, W.C. Hong, D.X. Gu, Forecasting holiday daily tourist
flow based on seasonal support vector regression with adaptive genetic
Although this research provides useful insights, there are some algorithm, Appl. Soft. Comput. 26 (2015) 435–443.
limitations to this study. The proposed study can further be [12] S. Karasu, A. Altan, Recognition mode or solar radiation time series
based on random forest with feature selection approach, in: 2019 11th
improved and extended from the following aspects. First, only
International Conference on Electrical and Electronics Engineering, ELECO,
the Shanghai stock exchange asset data has been considered. 2019, pp. 8–11.
However, due to different political environments and economic [13] F. Zhou, Q. Zhang, D. Sornette, L. Jiang, Cascading logistic regression onto
backgrounds, it is necessary to apply the IFAXGBoost+MV to other gradient boosted decision trees forforecasting and trading stock indices,
Appl. Soft. Comput 84 (2019) 105747.
capital markets to examine its versatility and effectiveness. Be-
[14] D. Gandhmal, K. Kumar, Systematic analysis and review of stock market
sides, the proposed approach is only based on the MV for portfolio prediction techniques, Compu. Sci. Rev. 34 (2019) 100190.
selection. However, multiple objectives such as higher moments [15] O. Bustos, A. Pomares-Quimbaya, Stock market movement forecast: A
(skewness and/or kurtosis), semi-variance, VaR, CVaR can be used Systematic review, Expert Syst. Appl. 156 (2020) 113464.
for portfolio selection to better represent the real investment [16] W. Wang, W. Li, N. Zhang, K. Liu, Portfolio formation with preselection
using deep learning from long-term financial data, Expert Syst. Appl. 143
experience. (2020) 113042.
[17] O. Sagi, L. Rokach, Ensemble learning: A survey, Wiley Interdiscip.
CRediT authorship contribution statement Rev.-Data Mining Knowl. Discov. 8 (4) (2018) e1249.
[18] X. Zhang, A. Li, R. Pan, Stock trend prediction based on a new status box
method and AdaBoost probabilistic support vector machine, Appl. Soft.
Wei Chen: Conceptualization, Writing - review & editing, Su- Comput. 49 (2016) 385–398.
pervision, Funding acquisition. Haoyu Zhang: Conceptualization, [19] J. Nobre, R. Neves, Combining principal component analysis, discrete
Methodology, Software, Writing - original draft. Mukesh Kumar wavelet transform and xgboost to trade in the financial markets, Expert
Mehlawat: Methodology, Review & editing. Lifen Jia: Methodol- Syst. Appl. 125 (2019) 181–194.
[20] S. Dey, Y. Kumar, S. Saha, S. Basak, Forecasting to classification: Predicting
ogy, Writing - review & editing.
the direction of stock market price using xtreme gradient boosting,
Working paper, 2016.
Declaration of competing interest [21] K. Song, F. Yan, T. Ding, L. Gao, S. Lu, A steel property optimization model
based on the XGBoost algorithm and improved PSO, Comput. Mater. Sci.
174 (2020) 109472.
The authors declare that they have no known competing finan-
[22] S. Zhao, D. Zeng, W. Wang, X. Chen, Z. Zhang, F. Xu, X. Liu, Mutation grey
cial interests or personal relationships that could have appeared wolf elite PSO balanced XGBoost for radar emitter individual identification
to influence the work reported in this paper. based on measured signals, Measurement (2020) 107777.
[23] L. Li, X. Zhao, M. Tseng, R. Tan, Short-term wind power forecasting based
on support vector machine with improved dragonfly algorithm, J. Clean
Acknowledgments
Prod. 242 (2020) 118447.
[24] E. Hajizadeh, M. Mahootchi, A. Esfahanipour, M. Massahi, A new NN-PSO
This research was supported by the National Natural Science hybrid model for forecasting Euro/Dollar exchange rate volatility, Neural
Foundation of China (Nos. 72071134, 71720107002), the Special Comput. Appl. 31 (7) (2019) 2063–2071.
Fund for Basic Scientific Research Operating Expenses of Beijing [25] Y.Z. Wang, Y.L. Nim, S. Lu, J.G. Wang, X.Y. Zhang, Remaining useful
life prediction of lithium-ion batteries using support vector regression
Municipal Colleges and Universities of Capital University of Eco- optimized by artificial bee colony, IEEE Trans. Veh. Technol. 68 (10) (2019)
nomics and Business, Beijing, China (QNTD202002), the Project 9543–9553.
of High-level Teachers in Beijing Municipal Universities in the [26] S. Mirjalili, A. Gandomi, S. Mirjalili, S. Saremi, H. Faris, Salp swarm
Period of 13th Five-year Plan, China (CIT&TCD20190338), the algorithm: a bio-inspired optimizer for engineering design problems, Adv.
Eng. Softw. 114 (2019) 163–191.
Humanity and Social Science Foundation of Ministry of Education
[27] T. Xiong, Y. Bao, Z. Hu, Multiple-output support vector regression with
of China (No. 19YJAZH005). The third author, Mukesh Kumar a firefly algorithm for interval-valued stock price index forecasting,
Mehlawat, acknowledges the support through MATRICS Scheme Knowl.-Based Syst. 55 (2014) 87–100.
of DST-SERB, New Delhi, India. [28] M. Akhavan-Amjadi, Fetal electrocardiogram modeling using hybrid evolu-
tionary firefly algorithm and extreme learning machine, Multidimens. Syst.
Signal Process. 31 (2020) 117–133.
References [29] R.J. Kuo, P.S. Li, Taiwanese export trade forecasting using firefly algorithm
based K-means algorithm and SVR with wavelet transform, Comput. Ind.
[1] T. Bodnar, S. Mazur, Y. Okhrin, Bayesian estimation of the global minimum Eng. 99 (2016) 153–161.
variance portfolio, European J. Oper. Res. 256 (1) (2017) 292–307. [30] I. Ibrahim, T. Khatib, A novel hybrid model for hourly global solar radiation
[2] F. Yang, Z. Chen, J. Li, L. Tang, A novel hybrid stock selection method with prediction using random forests technique and firefly algorithm, Energy
stock prediction, Appl. Soft. Comput. 80 (2019) 820–831. Conv. Manag. 138 (2017) 413–425.
[3] M. Jensen, Some anomalous evidence regarding market efficiency, J. Financ. [31] E.S. Chahnasir, Y. Zandi, M. Shariati, E. Dehghani, A. Toghroli, E.T. Mo-
Econ. 6 (2–3) (1978) 95–101. hamed, A. Shariati, M. Safa, K. Wakil, M. Khorami, Application of support
[4] S. Basak, S. Kar, S. Saha, L. Khaidem, S. Dey, Predicting the direction of vector machine with firefly algorithm for investigation of the factors
stock market prices using tree-based classifiers, N. Am. J. Econ. Fianc. 47 affecting the shear strength of angle shear connectors, Smart. Struct. Syst.
(2018) 552–567. 22 (4) (2018) 413–424.
[5] R. Engle, Autoregressive conditional heteroscedasticity with estimates of [32] A. Payne, G. Avendano-Franco, E. Bousquet, A. Romero, Firefly algorithm
the variance of United Kingdom inflation, Econometrica 50 (4) (1982) applied to noncollinear magnetic phase materials prediction, J. Chem.
987–1007. Theory Comput. 14 (8) (2018) 4455–4466.

17
W. Chen, H. Zhang, M.K. Mehlawat et al. Applied Soft Computing Journal 100 (2021) 106943

[33] A.D. Mehr, V. Nourani, V.K. Khosrowshahi, M.A. Ghorbani, A hybrid support [52] Y. Guo, S. Han, C. Shen, Y. Li, X. Yin, Y. Bai, An adaptive SVR
vector regression-firefly model for monthly rainfall forecasting, Int. J. for high-frequency stock price forecasting, IEEE Access 99 (6) (2018)
Environ. Sci. Technol. 16 (1) (2019) 335–346. 11397–11404.
[34] A. Kazem, E. Sharifi, F. Hussain, M. Saberi, O. Hussain, Support vector [53] X. Li, Y. Sun, Stock intelligent investment strategy based on support vector
regression with chaos-based firefly algorithm for stock market price machine parameter optimization algorithm, Neural Comput. Appl. 32 (6)
forecasting, Appl. Soft. Comput. 13 (2) (2013) 947–958. (2020) 1765–1775.
[35] J. Zhang, Y. Teng, W. Chen, Support vector regression with modified firefly [54] A. Jaafari, S. Termeh, D. Bui, Genetic and firefly metaheuristic algorithms
algorithm for stock price forecasting, Appl. Intell. 49 (2019) 1658–1674. for an optimized neuro-fuzzy prediction modeling of wildfire probability,
[36] H. Markowitz, Portfolio selection, J. Financ. 7 (1952) 77–91. J. Environ. Manag. 243 (2019) 358–369.
[37] R. Merton, Lifetime portfolio selection under uncertainty: The continuous- [55] T. Chen, C. Guestrin, XGBoost:A scalable tree boosting system, in: Pro-
time case, Rev. Econ. Stat. 51 (3) (1969) 247–257. ceedings of the 22nd ACM SIGKDD International Conference on Knowledge
[38] E.F. Fama, Multiperiod consumption-investment decisions, Amer. Econ. Discovery and Data Mining, 2016, pp. 785–794.
Rev. 60 (1) (1970) 163–174.
[56] L. Lv, J. Zhao, The firefly algorithm with Gaussian disturbance and local
[39] W. Chen, D. Li, Y. Liu, A novel hybrid ICA-FA algorithm for multiperiod
search, J. Sign. Process. Syst. 90 (8–9) (2018) 1123–1131.
uncertain portfolio optimization model based on multiple criteria, IEEE
[57] A. Altan, S. Karasu, Recognition of COVID-19 disease from X-ray images
Trans. Fuzzy Syst. 27 (5) (2019) 1023–1036.
by hybrid model consisting of 2D curvelet transform, chaotic salp swarm
[40] D. Roy, Safety-first and the holding of assets, Econometrica 20 (1952)
algorithm and deep learning technique, Chaos Solitons Fractals 140 (2020)
431–449.
110071.
[41] H. Markowitz, Portfolio Selection: Efficient Doversification of Investments,
Wiley, 1959. [58] D. Oliva, M. Elaziz, An improved brainstorm optimization using chaotic
[42] H. Konno, H. Yamazaki, Mean-absolute deviation portfolio optimization opposite-based learning with disruption operator for global optimization
model and its applications to Tokyo stock market, Manage. Sci. 37 (1991) and feature selection, Soft Comput. 24 (18) (2020) 14051–14072.
519–531. [59] J. Kennedy, R. Eberhart, Particle swarm optimization, in: Proceedings
[43] M. Speranza, Linear programming models for portfolio optimization, of ICNN’95 - International Conference on Neural Networks, 1995, pp.
Finance 12 (1993) 107–123. 1942–1948.
[44] W. Chen, Y. Wang, S. Lu, M. Mehlawat, A hybrid FA-SA algorithm for fuzzy [60] T. Chang, S. Yang, K. Chang, Portfolio optimization problems in different
portfolio selection with transaction costs, Ann. Oper. Res. 269 (1–2) (2018) risk measures using genetic algorithm, Expert Syst. Appl. 36 (7) (2009)
129–147. 10529–10537.
[45] W. Chen, Y. Wang, P. Gupta, M. Mehlawat, A novel hybrid heuristic [61] B. Chen, J. Zhong, Y. Chen, A hybrid approach for portfolio selection with
algorithm for a new uncertain mean-variance-skewness portfolio selection higher-order moments: Empirical evidence from Shanghai Stock Exchange,
model with real constraints, Appl. Intell. 48 (9) (2018) 2996–3018. Expert Syst. Appl. 145 (2020) 113104.
[46] J. Zhou, X. Li, Mean-semi-entropy portfolio adjusting model with [62] T. Fischer, C. Krauss, Deep learning with long short-term memory networks
transaction costs, J. Data Inf. Manag. 2 (2020) 121–130. for financial market predictions, European J. Oper. Res. 270 (2) (2018)
[47] M. Akbay, C. Kalayci, O. Polat, A parallel variable neighborhood search al- 654–669.
gorithm with quadratic programming for cardinality constrained portfolio [63] Sayed, I. Gehad, Khoriba, Ghada, Haggag, H. Mohamed, A novel chaotic
optimization, Knowl.-Based Syst. 198 (2020) 105944. salp swarm algorithm for global optimization and feature selection, Appl.
[48] R. Mansini, W. Ogryczak, M. Speranza, Twenty years of linear programming Intell. 48 (10) (2018) 3462–3481.
based portfolio optimization, European J. Oper. Res. 234 (2014) 518–535. [64] E. Rashedi, N. Hossein, S. Saryazdi, GSA: A gravitational search algorithm,
[49] M. Masmoudi, F. Abdelaziz, Portfolio selection problem: A review of Inform. Sci. 179 (13) (2009) 2232–2248.
deterministic and stochastic multiple objective programming models, Ann. [65] X. Li, Z. Shao, J. Qian, An optimizing method based on autonomous
Oper. Res. 267 (2018) 335–352. animate: Fish swarm algorithm, Syst. Eng. Theory Pract. 22 (11) (2002)
[50] F. Paiva, R. Cardoso, G. Hanaoka, W. Duarte, Decision-making for financial
32–38.
trading: A fusion approach of machine learning and portfolio selection,
[66] M. Chen, G. Zeng, K. Lu, J. Wang, A two-layer nonlinear combination
Expert Syst. Appl. 115 (2019) 635–655.
method for short-term wind speed prediction based on ELM, ENN, and
[51] M. Thenmozhi, G. Chand, Forecasting stock returns based on information
LSTM, IEEE Internet Things J. 6 (4) (2019) 6997–7010.
transmission across global markets using support vector machines, Neural
Comput. Appl. 27 (4) (2016) 805–824.

18