Received May 6, 2022, accepted May 22, 2022, date of publication May 30, 2022, date of current version

June 13, 2022.

Digital Object Identifier 10.1109/ACCESS.2022.3179364

A New Hybrid Cryptocurrency Returns

Forecasting Method Based on Multiscale
Decomposition and an Optimized Extreme
Learning Machine Using the Sparrow
Search Algorithm
XIAOXU DU 1, ZHENPENG TANG 2, JUNCHUAN WU3 , KAIJIE CHEN1 , AND YI CAI1
1 School of Economics and Management, Fuzhou University, Fuzhou 350108, China
2 School of Economics and Management, Fujian Agriculture and Forestry University, Fuzhou 350002, China
3 School of Economics and Management, Nanchang University, Nanchang 330031, China

Corresponding author: Zhenpeng Tang ([email protected])

This work was supported in part by the National Natural Science Foundation of China (NSFC) under Grant 71973028 and Grant 71573042.

ABSTRACT The return series of cryptocurrencies, which are emerging digital assets, exhibit nonstationarity,
nonlinearity, and volatility clustering compared to other traditional financial markets, making them excep-
tionally difficult to forecast. Therefore, accurate cryptocurrency price forecasting is essential for market
participants and regulators. It has been demonstrated that improved data forecasting accuracy can be achieved
through decomposition, but few researchers have performed information extraction on the residual series
generated by data decomposition. Based on the construction of a ‘‘decomposition-optimization-integration’’
hybrid model framework, in this paper, we propose a multi-scale hybrid forecasting model that combines the
residual components after primary decomposition for secondary decomposition and integration. This model
uses the variational modal decomposition (VMD) method to decompose the original return series into a finite
number of components and residual terms. Then, the residual terms are decomposed, and the features are
extracted using the completed ensemble empirical mode decomposition with adaptive noise (CEEMDAN)
method. The components are predicted by an extreme learning machine optimized by the sparrow search
algorithm, and the final predictions are summed to obtain the final results. Forecasts for the returns of
Bitcoin and Ethereum, which are significant cryptocurrency assets, are compared with other benchmark
models constructed based on different ideas. We find that the proposed quadratic decomposition VMD-Res.-
CEEMDAN-SSA-ELM hybrid model demonstrates the optimal and most stable forecasting performance in
both one-step and multi-step ahead prediction of the cryptocurrency return series.

INDEX TERMS Cryptocurrency, model selection, decomposition-ensemble, extreme learning machine,

sparrow search algorithm.

I. INTRODUCTION and digital hashing combined with smart contracts. Bitcoin

In recent years, the rise of blockchain technology in the con- (BTC), the first cryptocurrency, was introduced in 2008 [1].
text of the increasing integration of finance and the internet Since then, the variety of cryptocurrencies is growing by leaps
has led to the rapid development of cryptocurrency, a new and bounds, such as Ethereum (ETH) and Ripple (XRP). Cur-
type of virtual asset. Cryptocurrency uses the principles rently, there are more than 9,000 cryptocurrencies and more
of cryptography to secure transactions on a transaction-by- than 400 related exchanges, and the terms ‘‘coin wave’’ and
transaction basis using the encryption of virtual currencies ‘‘chain wave’’ are buzzwords. An important issue regarding
cryptocurrencies is price volatility.
The associate editor coordinating the review of this manuscript and As seen in Figure 1, the global cryptocurrency market
approving it for publication was Bo Pu . experienced significant volatility during the selected period,

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 10, 2022 60397
 X. Du et al.: New Hybrid Cryptocurrency Returns Forecasting Method

etc., as well as provide a reference for the government to

formulate relevant regulatory policies [11], [12].
According to existing research, asset price volatility in
financial markets is dynamic and highly nonlinear [4]. The
price forecasting problem of the cryptocurrency market is
similar to that of traditional stock and foreign exchange
markets and is also a financial time series forecasting prob-
lem. However, due to the special trading time system of the
cryptocurrency market, its price volatility is more obvious
and different from other financial markets [13]. Currently,
the methods involved in financial time series forecasting
mainly include the following traditional econometric models,
artificial intelligence methods, and hybrid models.
Traditional econometric forecasting models include lin-
ear multiple regression models [14], error correction mod-
els (ECMs)[15], autoregressive integrated moving average
FIGURE 1. Global cryptocurrency market capitalization and trading models (ARIMA) [16], [17]and vector autoregressive mod-
volume (2014.03-2022.03). els (VAR) [18]. But the econometric models have specific
assumptions, such as that the time series are trending and
repeatable and that the data are stable. For data that meet these
assumptions, a good prediction can be achieved. However,
especially in 2017, when a main wave of growth occurred such models have limited predictive power for time series
that represented a 3,175% year-on-year increase in mar- with nonlinear, nonstationary and volatile clustering charac-
ket capitalization, and in 2018, when the market fell by teristics [19].
78%. Then, during the global COVID-19 epidemic, the mar- With the rapid development of computer technology, many
ket experienced a nearly two-year oscillation period when artificial intelligence (AI) techniques have been used in
the cryptocurrency market was experiencing another major research related to time series forecasting. Common repre-
market [2]. Compared to traditional financial assets, whose sentative models include random forests [20], backpropaga-
valuation is based on fundamental information, cryptocur- tion neural networks (BPs) [21], artificial neural networks
rencies have a decentralized and virtual existence without (ANNs) [22], Bayesian neural networks (BNNs) [23], con-
a physical backing. Moreover, they are neither associated volutional neural networks (CNNs) [24], and support vector
with any commodity nor with a company, and governments regressions (SVRs) [25], [26]. Extreme learning machines
have no senior regulatory authority over them [3]. Due to (ELMs) [27], as an emerging learning framework for feed-
the specificity of cryptocurrencies, the complexity of their forward neural networks, can overcome the training dilemma
price fluctuations can be attributed to the multiple factors of backpropagation algorithms for single hidden layer feed-
and uncertainties that interact in the market, including the forward networks (SLFNs). Due to its advantages in learning
economic and political environment, and investor behavior convergence speed and parameter settings and noise resis-
that can lead to price instability. In addition, cryptocurrency tance, it has been gradually applied to classification and
trading rules are different from those of traditional financial prediction of various complex sequences, and a series of
markets. Since decentralized cryptocurrencies can be traded important results have been achieved [28]–[30]. Artificial
24 hours a day and 7 days a week, information and events intelligence algorithms are data-driven, have addressed some
generated at any time can immediately affect the price of limitations of traditional econometric models in forecasting
cryptocurrencies, rather than at specific market trading times to a certain extent, and have significantly improved forecast-
(such as the stock market) [4]. In summary, one of the most ing accuracy. However, such methods are more sensitive to
challenging areas of time series is the accurate prediction of parameters and model settings and are prone to local optima
the trends in cryptocurrency market quotes [5], [6]. Although and overfitting problems [31].
the cryptocurrency market is extremely complex and risky, In the field of financial time series forecasting, hybrid
it still represents an emerging alternative investment product models have become more popular forecasting methods,
with high returns and low correlation to other traditional and their constructed frameworks have been widely used
financial assets [7]. These characteristics make cryptocur- in many studies and proven to be effective in improv-
rency financial instruments that can be used to hedge against ing forecasting ability. Many researchers have built hybrid
uncertainty [8]–[10]. Therefore, an accurate cryptocurrency models to achieve effective forecasting for time series,
price prediction model can deepen the grasp of cryptocur- including Bitcoin price forecasting [32], exchange rate
rency market price fluctuation patterns, and provide a rea- forecasting [33], [34], and international crude oil price fore-
sonable basis for investors’ investment decisions in terms casting [35]. Generally, hybrid models are based on the idea
of optimal hedging, option pricing, portfolio diversification, of ‘‘decomposition-integration’’, which is divided into three

60398 VOLUME 10, 2022

X. Du et al.: New Hybrid Cryptocurrency Returns Forecasting Method

steps: data decomposition, modal forecasting, and integrated the prediction errors generated by the decomposition, thus
learning. Unlike traditional end-to-end price forecasting causing a certain degree of data distortion [36], [49], [50].
methods, hybrid algorithms are used to first decompose the In fact, the residual term in its complex and nonlinear form
original data through a decomposition algorithm that extracts may still carry valid predictive information, and further dis-
time-domain features of the time series [36]. Common section of the residual term is necessary [51]. As a single
decomposition algorithms, include empirical mode decom- forecasting model, the ELM model has advantages in terms
position (EMD) [37], ensemble empirical mode decomposi- of convergence speed and parameter settings, but the results
tion (EEMD) [38], and complete ensemble empirical mode obtained by forecasting using an unoptimized single learning
decomposition with adaptive noise (CEEMDAN) [39].With machine algorithm are unstable, and the forecasting accu-
the development of decomposition algorithm research, the racy is not high. Therefore, advanced related algorithms are
VMD [40] algorithm effectively separated components with needed to optimize the single forecasting model to improve
similar frequencies, thus enabled more efficient decompo- the accuracy and stability of the forecasting part. Many
sition of the original sequence and demonstrated its supe- existing time series forecasts only consider one-step ahead
rior performance in dealing with complex signals disturbed forecasting. However, investors are also very concerned about
by noise. After decomposing the data, each decomposed the short-term market in the actual investment process, espe-
component is predicted separately by the prediction models cially the cryptocurrency market, which is characterized by
mentioned above, such as econometric models and artificial more significant volatility clustering. Therefore, multi-day-
intelligence models. Finally, the components are summed to ahead forecasting is more important, and multi-step ahead
obtain the overall prediction results. forecasting can help investors provide a more comprehensive
Currently, hybrid models have achieved many results in and effective reference basis.
forecasting research in related fields. Sun et al. [41] pro- In view of the shortcomings of the existing research
posed a new carbon price prediction model based on EEMD- and based on the inheritance of the abovementioned model
IBA-ELM, tested the validity and stability of the model by construction idea and overcoming its limitations, a hybrid
examining the historical carbon prices, then concluded that model is constructed in this paper that is composed of a
the proposed model could significantly improve the predic- data decomposition algorithm, an optimization algorithm,
tion accuracy. CAO et al. [42] constructed CEEMDAN to and a forecasting model, namely, the proposed VMD-Res.-
decompose the stock price series and then predicted them CEEMDAN-SSA-ELM hybrid model, where Res. represents
by LSTM, empirically obtained that CEEMDAN was more the residual term after VMD. The innovation of the VMD-
thorough in decomposition than EMD. This hybrid prediction Res.-CEEMDAN-SSA-ELM model lies in the following
model showed superior performance in predicting the stock points.
price series. Zhu et al. [43] decomposed the carbon price (1) Focusing on the residual terms generated by the decom-
into multiple modes using the VMD model, further recon- position previous studies ignored in constructing the hybrid
structed the modes according to the evolutionary clustering model, we apply the quadratic decomposition technique to
algorithm proposed by CCI, and made predictions. The final combine VMD and CEEMDAN to a complex cryptocurrency
carbon price prediction results were obtained, proving that the return series. First, the VMD algorithm is adopted, the most
‘‘decomposition-clustering-prediction’’ method could better effective for processing complex signals [52]. The original
predict carbon prices. Jiang et al. [44] constructed new two- series is decomposed, and then the residual terms obtained
stage ensemble models by combining EMD (or VMD), ELM, after primary decomposition are taken into account. Next,
and improved harmony search (IHS) algorithm for stock the residual terms obtained after the VMD decomposition of
price prediction. The results show that the proposed model the original sequence are decomposed further by using the
has superior performance in terms of accuracy and stability CEEMDAN algorithm to extract the complex nonlinear infor-
compared with other models. mation. The overall data characteristics of the original time
The advantages of each model can be maximized through series can be better understood through the secondary decom-
model integration. As a result, the advantages of each model position technique, which is more accurate and complete for
can be used to overcome better the shortcomings of a single the decomposition of the original data. (2) A single prediction
model that obtains significant differences in prediction results model can vary in its predictive effectiveness in different
in different situations and under various prediction evaluation situations. In artificial intelligence algorithms, although the
criteria [45], [46]. Therefore, more prediction information is ELM model has some advantages in classification and pre-
used, which improves the prediction performance. However, diction studies, it depends on the input parameters. Therefore,
the existing hybrid model construction methods still have the hidden layer neuron parameters of the ELM are opti-
the following shortcomings. Most of the previous integrated mized by introducing the cluster intelligence optimization
hybrid models decompose the original sequence into a finite SSA algorithm [53], which has advantages in terms of search
number of modal components and residual terms through one accuracy, convergence rate, stability and avoidance of local
decomposition, then the resulting modalities are predicted optimization. The process achieves better stability of the
through the prediction model [47], [48]. The residual terms, prediction module and improves the prediction accuracy to
which are discarded as general components, will accumulate compensate for the deficiencies associated with a single

VOLUME 10, 2022 60399

 X. Du et al.: New Hybrid Cryptocurrency Returns Forecasting Method

prediction model. (3) While previous studies have focused is mixed and multiplied by the e−jwk t phase to adjust the
more on one-step ahead forecasting, this paper applies spectrum of each component signal to the fundamental fre-
the proposed VMD-Res.-CEEMDAN-SSA-ELM model to quency band. Then, the bandwidth is determined by estimat-
one-step ahead and multistep ahead forecasting of cryptocur- ing each component using the Gaussian smoothing method.
rency returns, and the results of multi-step ahead forecast- The corresponding constrained variational model can be
ing are more closely related to the market. Moreover, the described as
accuracy and robustness of the model in predicting com-  (   )
plex, nonlinear, and volatility-clustered time series is verified X j ∗ −jwk t 2
∂t (δ(t) +

 min ) uk (t) e
by comparing it with benchmark models. Finally, the pro- {uk },{wk } k πt 2 (1)
posed approach is entirely data-driven and does not require 
 P
s.t. k uk = f
excessive assumptions or consideration of exogenous influ-
ences that lead to market sentiment fluctuations, facilitat- In this equation, {uk } := {u1 , . . . , uK } is the mode
ing investors to make appropriate decisions that are more component obtained after VMD decomposition, and {wk } :=
realistic. {w1 , . . . , wK } is the center frequency of the mode compo-
The rest of this paper is presented as follows. In Section 2, nents, VMF. ∂t denotes the partial derivative of t, δ(t) refers
the individual components and details of the VMD-Res.- to the shock function, and ∗ denotes the convolution sign. f
CEEMDAN-SSA-ELM model are proposed. In Section 3, is the original input signal.
the daily closing price data of Bitcoin and Ethereum, which Step 2: To make the signal reconstruction accurate, it must
are representative among cryptocurrencies, are obtained from be constrained by introducing an incremental Lagrange func-
the CoinMarketCap website as empirical samples. The tra- tion to convert the original equation into an unconstrained
ditional single benchmark model, the integrated benchmark variational problem. As a result, the optimal solution is
approach, and the VMD-Res.-CEEMDAN-SSA-ELM model derived as follows:
constructed in this paper are compared based on one-step-
ahead and five-step-ahead forecasting using evaluation met- L ({uk } , {wk } , λ)
rics to test the performance of the proposed model. This   2
X j ∗
paper is concluded in Section 4, and a plan for future work =α ∂t (δ(t) + ) uk (t) e−jwk t
is presented. k πt 2
Xuk (t) 2 D X E
II. INTRODUCTION TO THE METHODOLOGY + f (t) − + λ(t), f (t) − uk (t) (2)
k k
2
A combined model (VMD-Res.-CEEMDAN-SSA-ELM)
was developed based on the idea of ‘‘decomposition- where α is the quadratic penalty factor introduced to guaran-
ensemble’’ and the combination of secondary decomposi- tee the accuracy of signal reconstruction when it occurs, and
tion techniques with machine learning methods, aiming to λ is the Lagrange multiplier used to control the strictness of
predict cryptocurrency returns more accurate. Because the the constraints.
prediction model proposed in this paper consists of sev- Step 3: Solve the variational problem by searching for the
eral models, the components of the model and the overall optimal solution of equation (2) using the alternating direc-
model are described below: the VMD algorithm, the CEEM- tion method of multipliers (ADMM). Equations (3) – (5) are
k (w), ŵk , and λ̂
iterated several times to obtain ûn+1
DAN algorithm, the extreme learning machine, the sparrow n+1 n+1 .
search algorithm, and the whole new hybrid model built in Futhermore, the optimal solution of the constrained varia-
this study. tional model is obtained until the iterative condition (6) is
satisfied.
A. VARIATIONAL MODE DECOMPOSITION (VMD)
λ̂(w)
fˆ (w) −
P
VMD is an adaptive, quasi-orthogonal, and completely i6 =k ubi (w) + 2
nonrecursive decomposition method proposed by ûn+1
k (w) = (3)
1 + 2α(w − wk )2
Dragomiretskiy and Zosso (2013). In the process of signal R∞
w|ûk (w)|2 dw
decomposition, the optimal center frequency and finite band- ŵn+1
k = R0 ∞ 2
(4)
width of each mode can be matched adaptively by searching 0 |ûkh(w)| dw i
X
and solving to achieve the effective separation and frequency λ̂n+1 = λ̂n + τ fˆ (w) − ûn+1
k (w) (5)
domain division of the characteristic mode components of the k

signal. Thus, the effective decomposition components of a

In these equations, ûnk (w), fˆ (w) and λ̂n (w) are the Fourier
given signal are obtained, and finally the optimal solution of
transforms of ûnk , f (t) and λn respectively.
the problem is obtained. The detailed VMD steps are shown
as follows. 2
ûn+1 − ûnk
P
Step 1: A Hilbert transform is implemented on every modal k k
signal to obtain a unilateral spectrum. The exponential term 2
2
<ε (6)
of the modal function corresponding to the center frequency ûnk 2

60400 VOLUME 10, 2022

X. Du et al.: New Hybrid Cryptocurrency Returns Forecasting Method

B. COMPLETE ENSEMBLE EMPIRICAL MODE unique optimal solution must be generated in ELM. The pro-
DECOMPOSITION WITH ADAPTIVE NOISE (CEEMDAN) cess is achieved by setting the number of hidden layer nodes
Torres et al. (2011) proposed a CEEMDAN algorithm based of the network. The input weights and biases do not need to be
on EMD and EEMD. The algorithm effectively suppresses adjusted during execution; therefore, the advantages of ELM
the mode mixing of EMD by adding finite times of adaptive are fast learning and good generalization performance. The
white noise at each stage. It can achieve a more thorough ELM network model is illustrated in Figure 2.
decomposition of the signal data with a more minor recon-
struction error by removing noise residuals with fewer aver-
aging times. The decomposition steps of CEEMDAN are as
follows:
Step 1: Let x (n) be the original signal sequence, ε be
the adaptive coefficient, wi (n) be the noise sequence added
for each decomposition, and xi (n) be the signal sequence
after i times of noise is added. The average value of N sub-
experiments of EMD decomposition is the first intrinsic mode
component IMF1 ,
xi (n) = x (n) + εwi (n) (7)
N
1 X
IMF1 (n) = IMF1i (n) (8)
N
i=1
FIGURE 2. Extreme learning machine process.
Step 2: Calculate the residual sequence r1 (n) of the first
stage and obtain a new r1i (n) for the N sub-experiments In Figure 2, x1 ∼ xn are the nodes of the input neuron,
until the EMD decomposition finishes its work on the IMF ω11 ∼ ωnk are the weights between the input layer and the
component. hidden layer, g (x) is the activation function, b1 ∼ bk are
r1 (n) = x (n) − IMF1 (n) (9) the hidden layer node thresholds, β11 ∼ βnk are the weights
between the hidden layer and the output network model layer,
Step 3: Based on Step 2, calculate the second intrinsic mode and y1 ∼ yn are the outputs of the model.
component IMF2 . Suppose there are N training samples {(xi , yi )}N i=1 , and
N
xi = [xi1 , xi2 , . . . , xin ]T ∈ Rn refers to the n-dimensional
IMF2 (n) =
1 X
IMF1 {r1 (n) + ε1 IMF1 [wi (n)]} (10) input data of the training set. t i = [ti1 , ti2 , . . . , tim ]T ∈ Rm
N is the m-dimensional ideal output value of the training set.
i=1
The ELM network model expression of the hidden layer
Step 4: Repeat the calculation to the stage k +1. We can obtain nodes K and assuming an activation function of gi (xi ) can
the residual sequence rk (n) at stage and the k + 1th intrinsic be expressed as
modal component IMFk+1 .
K
rk (n) = rk−1 (n) − IMFk (n)
X
(11) yj = βi gi (ωi xj + bi ), j = 1, 2 . . . , K (14)
N i=1
1 X
IMFk+1 (n) = IMF1 {r1 (n) + εk IMFk [wi (n)]} (12)
N In this equation, ωi is the input weight vector connected
i=1
between the input layer nodes and the first hidden layer node
Step 5: The above steps are repeated. If the number of extreme i. βi is the weight matrix between the ith hidden layer and
points of the residual sequence is ≤ 2, the EMD is stopped, the output layer. bi is the threshold value of the ith hidden
then the final residual sequence R(n) and intrinsic mode com- layer node. y
j is the actual output of the network model. And
ponent IMFk are obtained. Finally, the initial signal sequence gi ωi xj + bi is the activation function.
x (n) is decomposed as If the single-feeder neuron network in the hidden layer
XK can approach any training sample with zero error, it can be
x(n) = IMFk (n) + R (n) (13) expressed as
i=1
N
C. EXTREME LEARNING MACHINE OPTIMIZED BY THE X
SPARROW SEARCH ALGORITHM (SSA-ELM) kyi − ti k = 0 (15)
i=1
1) EXTREME LEARNING MACHINE (ELM)
Huang et al. (2004) proposed an ELM algorithm to solve Then we obtained
single hidden-layer feedforward neural networks. It mainly K
X
uses the generalized inverse theory of matrices. Compared tj = βi gi (ωi xj + bi ), j = 1, 2 . . . , N (16)
with the traditional neural network learning algorithm, only a i=1

VOLUME 10, 2022 60401

 X. Du et al.: New Hybrid Cryptocurrency Returns Forecasting Method

which can be a matrix represented as t

In this equation, Xworst is the worst position of discoverers
of the entire process in the tth iteration, XPt+1 is the best
Hβ = T (17) position of the discoverers in the t + 1th iteration, and A
In this equation, is an dimensional matrix with the same dimensions as the
input. Each element is randomly assigned 1 or −1, and A+ =
g1 (ω1 x1 + b1 ) · · · gi (ωK x1 + bK )
 
AT (AAT )−1 ; n is the number of sparrows.
.. .. ..
H = . . . (18)
 
 Usually, some foragers will act as warning sparrows to help
gi (ω1 xN + b1 ) · · · gi (ωK xN + bK ) N ×K the discoverer forage, and when in danger, they will counter
trap or withdraw, close to other sparrows.
H stands for the hidden-layer output matrix and T for the 
ideal output vector. t
Xbest
 + λ · Xi,d t − Xt
best fi > fg
Therefore, the optimal solution of Hβ = T is obtained, t+1
Xi,d = (22)
 t t

t +J X i,d −X worst
which is given by Xi,d

(fi −fw )+ε fi = fg

β̂ = H + T (19) t
In this equation, Xbest is the best position of warners in the
In this equation, H + refers to the augmentation matrix of the t-th iteration; as a step control parameter, λ is a random
matrix H. number that obeys a normal distribution with an average of
The entire training process needs to be run only once to 0 and variance of 1; J ∈ [−1, 1] is a random number; fi is the
obtain the optimal solution, making the ELM’s generalization current sparrow’s fitness value; fg and fw are the best fitness
ability very strong. value and the worst fitness value, respectively; ε is defined as
a tiny constant, which is mainly used to prevent the case of
2) SPARROW SEARCH ALGORITHM (SSA) the fi − fw = 0. The main steps of the sparrow algorithm are
The sparrow search algorithm (SSA) is a new intelli- as follows.
gent optimization algorithm proposed by Xue et al. in Step 1: Initialize the population, set the total number of
2020 that idealizes and formulates the corresponding rules sparrow population n, number of discoverers, number of
for the predatory behavior of sparrow groups. This algorithm warners, the maximum number of iterations T , and alarm
assumes two types of sparrows: discoverers and foragers. threshold R2 .
Discoverers actively search for food, and foragers obtain food Step 2: Use mean squared error (MSE) as the fitness func-
from discoverers. In addition, some predators can grab food. tion and then calculate the fitness value of each sparrow. Find
The role of the discoverer is to guide the entire sparrow and define the best and worst fitness values as fg and fw ,
population in searching and predating. And its position can respectively.
be expressed by an equation, Step 3: The new positions of the discoverers, foragers, and
(   warners are calculated using equations (20) – (22), and if the
t+1 X t · exp
i,d
−i
α·itermax R2 < ST fitness value of the new position is greater than that of the
Xi,d = (20) previous position, it is updated.
t
Xi,d + Q · L R2 ≥ ST
Step 4: Perform iterations and repeat step 3 to continuously
t+1 update the positions of sparrows, stopping when the number
In this equation, t is the current number of iterations; Xi,d
is the location of the ith sparrow in the t + 1th iteration of of iterations is T . Therefore, the position of the sparrow with
the dth dimension; T is the maximum number of iterations. the lowest fitness value in all iterations is the optimal solution.
α ∈ [0, 1] is a random number; Qis a random number that
obeys a normal distribution; R2 and ST are the warning and 3) THE EXTREME LEARNING MACHINE FOR SPARROW
safety values, respectively. And L is a matrix of size 1 × d SEARCH ALGORITHM OPTIMIZATION (SSA-ELM)
where each element is 1. ELM can be used for nonlinear function fitting and prediction
If R2 < ST , this means that predators are not nearby, so the problems with small-sample learning. However, the stability
discoverers can perform a wide-area search. If R2 ≥ ST , that of model training can be affected by its input weights and
is, the predators have been found, the rest of the sparrows implied layer thresholds. The SSA algorithm has the advan-
need to leave their present position. tages of high search accuracy, fast convergence, and good
Foragers should observe the discoverers during this pro- stability, so that the SSA algorithm can optimize the ELM
cess, and when an abundant food source is noted, the foragers input parameters and weights to improve the prediction effi-
will leave their location to compete for food. If the scramble ciency and obtain more stable prediction results. The detailed
is successful, they will receive food from the finder; thus, the operation flow of the SSA-ELM model is as follows:
foragers’ positions are updated as Step 1: The sparrow population is initialized as the discov-
  t t
 erer, forager, and predator. The corresponding fitness value
Q · exp Xworst2−Xi,j

i > n2 Pb of each sparrow is calculated separately, and the best
t+1 i
Xi,d = (21) fitness is defined along with the position of the corresponding
X t+1 + X t − X t+1 · A+ · L other situation

P i,d P sparrow as Xbest .

60402 VOLUME 10, 2022

X. Du et al.: New Hybrid Cryptocurrency Returns Forecasting Method

Step 2: Iterations are performed to determine the optimal SSA-ELM model can improve the prediction accuracy, con-
initial weights and thresholds. These data can be obtained by vergence speed, and stability of cryptocurrency returns to
comparing the value of fitness function MSE. More specif- a certain extent. The specific modeling steps are shown in
ically, when exercising the second iteration, the minimum Figure 4.
MSE of the current sparrow generations should be compared Step 1: First decomposition. Decompose the original series
with the optimal adaptive value of the previous optimal fitness into each mode component VMFi through the VMD decom-
value Pb . If it is less than Pb , the optimal fitness value Pb position technique, then subtract each mode component from
must be updated to the minimum MSE of the current gener- the original series to obtain the residual series (Res.).
ation of sparrows, and the position of this sparrow should be Step 2: Secondary decomposition. CEEMDAN was
updated to the optimal position Xbest . Otherwise, the optimal applied to decompose the residual series further to obtain
adaptation value Pb and the optimal position Xbest need not another set of subseries IMFi . Then we normalized the
be updated, and the next iteration can be performed. decomposed components VMFi and IMFi .
Step 3: The iteration should be stopped until it reaches the Step 3: Forecast process. The modal components obtained
set value itermax . Finally, the optimal weights and thresholds from the decomposition of the original series and the residual
obtained from the model optimized by the sparrow algorithm term decomposition are predicted by the SSA-ELM.
are used to construct a new ELM model for prediction. The Step 4: Ensemble. The predictions of the residuals are
specific steps of the SSA-ELM are shown in figure 3. superimposed with the predictions of each VMFi to obtain
the final forecast result.

III. EMPIRICAL ANALYSIS

A. DATA DESCRIPTION AND EVALUATION CRITERIA
Bitcoin and Ethereum account for nearly 66% of the mar-
ket capitalization in the cryptocurrency market and enjoy
the majority of the daily trading volume, even reaching
more than 70% of the whole market on June 30, 2021.
Therefore, in this study, the log returns of the daily closing
prices of Bitcoin and Ethereum were selected as the predic-
tion objects. The log returns of the tth trading day, rt =
[ln (pt ) − ln (pt−1 )] × 100, and the returns of the above two
virtual currencies are predicted using the proposed VMD-
Res.-CEEMDAN-SSA-ELM to verify the model’s effective-
ness, and the daily closing prices of BTC and ETH were
obtained from the web (http://www.CoinMarketCap.com/).
Combined with past data, awareness and interest in cryp-
tocurrencies was not high until 2017, after which cryptocur-
rency assets such as Bitcoin really caught the attention of
FIGURE 3. SSA-ELM process. investors and academics. Trading volume data also confirmed
this trend. Therefore, this study selects the return rates of BTC
and ETH from January 1, 2017, to June 30, 2021, with 1,642
D. VMD-RES.-CEEMDAN-SSA-ELM MODEL returns data each.
As previously mentioned, cryptocurrency return series have In the respective returns datasets of Bitcoin and Ethereum,
complex characteristics, such as typical nonstationary, non- the training and test sets were divided; the first 1,492 returns
linear, and volatility clustering, with limited accuracy when data were used as the training set, and the remaining 150 data
using a single forecasting method. Because the VMD decom- were used as the test set. Table 1 lists the descriptive infor-
position technique can decompose the complex signal into mation related to Bitcoin and Ethereum returns data. The
several mode components with much lower complexity, the empirical operation in this study was completed by MATLAB
prediction accuracy is substantially improved when each 2019b.
modal component obtained from the VMD decomposition is In this study, four evaluation metrics, mean absolute error
modeled separately through common forecasting methods. (MAE), normalized root mean square error (NRMSE), and
Therefore, we adopted the VMD algorithm to decompose symmetric mean absolute percentage error (SMAPE), coeffi-
the original sequence firstly. And then, based on the first cient of determination (R2 ) were selected to test the prediction
decomposition, a secondary decomposition of the residual effectiveness of the models. In addition, to more concisely
terms by CEEMDAN is considered, which in turn enables compare the differences in evaluation metrics between dif-
a more thorough decomposition of the original sequence. ferent benchmark models and the proposed model, we define
Finally, ELM is used to optimize the SSA. The advanced the following three evaluation metrics relative to the proposed

VOLUME 10, 2022 60403

 X. Du et al.: New Hybrid Cryptocurrency Returns Forecasting Method

FIGURE 4. Structure of the VMD-Res.-CEEMDAN-SSA-ELM model.

TABLE 1. Descriptive statistics for Bitcoin and Ethereum return data. determining parameter K, time series data decomposition was
performed.
Then, the sum of each mode component generated by
the VMD decomposition is subtracted from the original
sequence to obtain the residual sequence. Due to its complex-
ity, the residual series using a predictive algorithm is compli-
cated to predict accurately. Hence, previous studies typically
neglected this series. To a certain extent, this operation tends
to lead to a loss of information. Thus, the secondary decom-
position technique was adopted in this study to extract more
available information, and the complex residual series was
further decomposed using CEEMDAN technology. Taking
the decomposition process of the Bitcoin return sequence
as an example, the decomposition process is shown in
model. Table 2 presents the definitions and formulas for the Figure 5.
relevant evaluation indicators. In addition, due to the large span of values of individual
features, the differences in their units and magnitudes led to
features that are not comparable with each other. Therefore,
B. DATA PROCESSING the data needed to be normalized for each decomposed sub-
In practical applications, the optimal number of modal com- series of modal component data before being predicted using
ponents cannot be directly determined when decomposing the machine learning methods.
original time series through VMD decomposition because of In this study, the data were linearly altered using the
the admixture of noise in the original time series. Therefore, minimum-maximum deviation normalization method with
the average instantaneous frequency observation method was the following expressions:
used in this study to determine the optimal K value. For Bit-
x − xmin
coin returns series data, the average instantaneous frequency x̃ = (23)
decreases less at the end when the value of K is 11, and xmax − xmin
that is over decomposition. Therefore, the optimal number of where x denotes the original feature data, x̃ denotes the
VMD decomposition modes for the Bitcoin returns series was standardized subseries data, xmax represents the maximum
10. Similarly, the optimal number of components for VMD value in the original sequence, and xmin denotes the minimum
decomposition of the Ethereum returns series was 12. After value.

60404 VOLUME 10, 2022

X. Du et al.: New Hybrid Cryptocurrency Returns Forecasting Method

FIGURE 5. Mode components of the original bitcoin return series decomposed by VMD-Res.-CEEMDAN.

C. FORECAST RESULTS can be used to decompose the return series. At the same
To verify the validity of the proposed model, we compare the time, due to different modeling ideas, scholars have been
predictive validity and stability of the proposed model with able to improve the forecasting accuracy by decomposing
those of the basic models. the high-frequency components of the primary decompo-
First, we test the forecasting performance differences sition quadratically and integrating the forecasts, and we
between the hybrid and single forecasting models and select correspondingly incorporate such models. Finally, based
the optimal basic model for the forecasting module. We intro- on the residual terms generated after considering the pri-
duce traditional econometric models and artificial intelli- mary decomposition proposed in this paper, we construct
gence models as basic models to determine the differences in an integrated forecasting model combining the residual
forecasting performance between single forecasting models. terms.
The intelligent cluster SSA algorithm is also selected to Based on the above considerations, ten additional bench-
optimize the above better basic prediction models, so that the mark models are constructed in this paper compared with the
model with the best prediction performance can be selected proposed models. Specifically: (1) single prediction mod-
as the main component of the hybrid model. els: ARIMA, BP, SVR, ELM, SSA-ELM; (2) hybrid pre-
Second, to examine the difference in prediction perfor- diction models: EMD-SSA-ELM, CEEMDAN-SSA-ELM,
mance between the proposed quadratic decomposition model CEEMDAN-VMD-SSA-ELM, VMD-SSA-ELM, VMD-
considering residual terms compared with the commonly Res.-SSA-ELM.
used methods in the general decomposition-integration In addition, one-step and multistep ahead forecasting was
framework, we empirically modeled the commonly used performed in this paper by one-step-ahead and five-step-
methods in each possible combination. Specifically, in the ahead, i.e., the data of the first six trading days allow forecast-
modal decomposition stage, EMD, CEEMDAN, or VMD ing the return of the next 1 and 5 trading days, respectively.

VOLUME 10, 2022 60405

 X. Du et al.: New Hybrid Cryptocurrency Returns Forecasting Method

TABLE 2. Relevant evaluation indicators. SSA-ELM improve to a certain extent compared with other
single forecasting models when forecasting complex finan-
cial time series with high volatility, such as cryptocurrency
returns. Therefore, we adopt SSA-ELM as the primary model
for the forecasting module in the following construction of the
hybrid model.
However, in a comprehensive view, the R2 indicators of
the overall fit of the single forecasting model starting from
a data-driven approach do not perform well. They cannot
effectively capture the complex cryptocurrency return data
characteristics.

2) COMBINED MODEL WITHOUT CONSIDERING RESIDUAL

TERM DECOMPOSITION
By comparing the forecasting results of the com-
mon hybrid models EMD-SSA-ELM, CEEMDAN-SSA-
ELM, CEEMDAN-VMD-SSA-ELM and VMD-SSA-ELM
constructed under the decomposition-based integration
framework. It is easy to find that the combined model
has significantly improved in four evaluation metrics,
MAE, NRMSE, SMAPE and R2 , compared with the sin-
gle prediction model. The prediction accuracy and sta-
bility of the combined model with the decomposition
technique are better than those of the single prediction
model.
Furthermore, these combined models are viewed sepa-
rately. The evaluation indices of the prediction results of
CEEMDAN-SSA-ELM are generally better than those of
EMD-SSA-ELM, which shows that the CEEMDAN decom-
position is more complete than the EMD decomposition
for data decomposition, and thus makes the data features
extraction more adequate. The VMD-SSA-ELM achieves
the best prediction results in this stage. In this regard,
it can be shown that the VMD decomposition technique
has stronger decomposition ability for time series with high
The evaluation metrics of the prediction results are shown in complexity and high volatility like cryptocurrency prices, and
Table 3. can better extract serial data features and handle complex
Figures 6 (a) - (d) show the true values and prediction signals.
results of the different models for Bitcoin and Ethereum in
one-step and five-step ahead. 3) COMBINED MODEL CONSIDERING RESIDUAL
TERM DECOMPOSITION
D. MODEL COMPARISON AND ANALYSIS The VMD-Res.-SSA-ELM and the proposed model consider
1) NONCOMBINED MODELS the residual terms after the first decomposition and incor-
First, the common econometric and artificial intelligence porate the prediction of the resulting residual series. The
models ARIMA, BP, SVR, and ELM are used as benchmark difference is that the model proposed in this paper further
prediction models to predict the return data of BTC and ETH, decomposes the residual terms generated from the primary
as shown in Table 3. From the overall performance, under the decomposition to obtain the corresponding components, and
comparison of the four single models, it can be seen that ELM integrates the predictions to obtain the final prediction results.
performs the best in all evaluation metrics in the prediction of The prediction results of these two models are shown in
Bitcoin and Ethereum 1 day ahead and 5 days ahead. ARIMA, Table 3. The predictive evaluation metrics obtained from
on the other hand, has the worst prediction performance. The the model proposed in this paper are further improved than
possible reason is that ARIMA, as a classical linear model, the VMD-Res.-SSA-ELM. It can be seen that the residual
has difficulty capturing the pattern due to cryptocurrency series generated after the original return series decomposed
return data’s the non-linear and high volatility characteristics. by VMD also contains important and complex information,
Furthermore, we introduce the SSA optimization algo- and the direct use of the SSA-ELM model to forecast the
rithm combined with ELM. The four evaluation indicators of residual terms directly has a limited effect. For example,

60406 VOLUME 10, 2022

X. Du et al.: New Hybrid Cryptocurrency Returns Forecasting Method

TABLE 3. Comparison of the one-step and five-step ahead prediction results of different models for BTC and ETH returns return data.

FIGURE 6. True values, one-step and five-step ahead prediction results of the different models for BTC and ETH.

for the one-step ahead prediction of bitcoin returns, the In general, the proposed model performs best in all eval-
proposed model achieves an overall fit evaluation index R2 of uation metrics compared to all other benchmark models.
0.9744. Therefore, the quadratic decomposition technique proposed

VOLUME 10, 2022 60407

 X. Du et al.: New Hybrid Cryptocurrency Returns Forecasting Method

FIGURE 7. Error analysis of multi-step predictions for all models in cases of BTC and ETH.

60408 VOLUME 10, 2022

X. Du et al.: New Hybrid Cryptocurrency Returns Forecasting Method

in this paper for the residual term appears to be necessary. results compared to the other models. The proposed model
The original sequence is decomposed twice by VMD and has optimal prediction accuracy and stability by the positions
CEEMDAN, which can effectively combine the advantages of the three main statistical indicators.
of the two algorithms to grasp the original sequence’s charac-
teristics better, and thus can be combined with the SSA-ELM 6) BRIEF SUMMARY
prediction module to obtain more accurate prediction results Overall, the VMD-Res.-CEEMDAN-SSA-ELM multiscale
further. hybrid model can forecast cryptocurrency returns more accu-
rately. Compared to the benchmark model, the proposed
4) ANALYSIS OF FORWARDING MULTISCALE FORECASTING hybrid model performs best in the four evaluation metrics
A longitudinal comparison of multiscale forward forecasting shown in the one-step and five-step ahead forecasting pro-
shows that the forecasting performance of both the bench- cess for two cryptocurrency asset returns. Moreover, com-
mark model and the proposed multiscale hybrid model grad- pared with the single forecasting model, the hybrid model
ually decreased as the forecasting scale increased. The result improves the evaluation indicators significant without con-
implies that all models outperformed the one-step ahead sidering the residual term. Specifically, compared with the
prediction scenario in terms of predictive power over the common ARIMA model, the MAE of the proposed model
five-step ahead prediction. This is mainly because as the decreases by 82.49%, the NRMSE by 84.05%, and the
forecasting scale increases, the complexity and high volatility SMAPE by 66.02% in the context of a one-step ahead forecast
of the cryptocurrency return series increase. Therefore, the of bitcoin returns, and the R2 index reaches 0.9744., We found
forecasting accuracy 5 days in advance is lower than fore- that the proposed model had a more concentrated prediction
casting 1 day in advance. In addition, some data information error distribution and showed excellent and stable prediction
is not trained by the model in actual Bitcoin and Ethereum performance in comparing the prediction errors.
returns forecasting, which leads to the gradual weakening
of the model’s forecasting ability and explains the increased IV. CONCLUSION AND FUTURE WORK
difficulty of forecasting cryptocurrency return series multiple Based on the idea of ‘‘decomposition-integration’’, this paper
steps ahead compared to one day ahead. proposed a multi-scale hybrid model containing quadratic
decomposition, optimization and prediction algorithms and
5) FORECAST ERROR ANALYSIS applies it to the study of daily returns of mainstream cryp-
To further understand and compare the forecast error dis- tocurrencies. The empirical analysis led to the following
tribution between the models, we analyze the forecast conclusions.
errors generated by the benchmark models and the pro- (1) Based on complex systems methodology, decomposi-
posed model in the empirical study of one-step and five- tion and integration techniques were used to decompose the
step ahead prediction scenarios for major cryptocurrency cryptocurrency return series into subseries, then predict each
returns. Figures 7 (a) - (d) show the forecast error distribu- subseries individually, and finally integrate and reconstruct
tions and the corresponding fitted error distribution curves of the prediction results of each subseries to form the overall
the benchmark and proposed models, and then depict Taylor prediction results. This process could improve the forecasting
plots with the error data. It is worth mentioning that Taylor accuracy more effectively than a single model.
diagrams are used to facilitate the analysis of correlation (2) The VMD technique performed better when dealing
measures between different models by presenting statistical with highly complex time series data such as nonstationary
information such as standard deviation, correlation coeffi- and nonlinear data. Adapting VMD decomposition combined
cient, and root mean square deviation (RMSD) generated by with the SSA-ELM algorithm could substantially improve the
each different forecasting model in a single plot during the prediction results compared with CEEMDAN and EMD.
actual forecasting process [54]. The Taylor plots plotted in (3) The combined VMD-Res.-CEEMDAN-SSA-ELM
the paper are generated by normalizing the standard deviation quadratic decomposition model had a significantly stronger
and root mean square deviation to make the comparison forecasting ability than the combined single decomposition
between models more intuitive. model. Considering the residual terms, the VMD and CEEM-
We combine the results of the one-step and five-step ahead DAN quadratic decomposition techniques could be used to
forecasts of the two cryptocurrency assets returns for error decompose the nonstationary, nonlinear, and highly complex
analysis. First, the error distribution curves fitting the predic- financial time series effectively with clustered fluctuations
tion errors of the different models show that the hybrid model into several more regular smooth subseries. The combined
has a more minor error than the single model. In addition, model had significant advantages over the single and other
the latter three hybrid models using VMD decomposition combined models without considering the residual terms.
techniques have a smaller range of error point fluctuations Information and the optimal prediction results achieved in
and run more smoothly. The errors of the proposed models are both the 1-step-ahead and 5-step-ahead prediction studies
more distributed around zero, with the slightest correspond- proved the robustness of the model.
ing standard deviations. Taylor plots show that the proposed The proposed multiscale hybrid model, VMD-Res.-
model is closest to the actual values among all empirical CEEMDAN-SSA-ELM, conducted an empirical study based

VOLUME 10, 2022 60409

 X. Du et al.: New Hybrid Cryptocurrency Returns Forecasting Method

on the returns of Bitcoin and Ethereum. We found that the [18] R. Bohte and L. Rossini, ‘‘Comparing the forecasting of cryptocurrencies
proposed model could be used to improve the accuracy of by Bayesian time-varying volatility models,’’ J. Risk Financial Manage.,
vol. 12, no. 3, p. 150, Sep. 2019.
cryptocurrency return forecasting effectively. These forecasts [19] P. Jiang, Z. Liu, J. Wang, and L. Zhang, ‘‘Decomposition-selection-
can help short-term investors in the cryptocurrency market ensemble forecasting system for energy futures price forecasting based on
to more accurately understand and grasp the market’s price multi-objective version of chaos game optimization algorithm,’’ Resour.
Policy, vol. 73, Oct. 2021, Art. no. 102234.
trends. However, the empirical evidence in this paper for [20] J. Yoon, ‘‘Forecasting of real GDP growth using machine learning models:
cryptocurrency returns series also proves that the trends of Gradient boosting and random forest approach,’’ Comput. Econ., vol. 57,
price series can be affected by complex multidimensional no. 1, pp. 247–265, Jan. 2021.
[21] S. Qi, K. Jin, B. Li, and Y. Qian, ‘‘The exploration of internet finance
factors, sharp fluctuations still characterize the data, and the by using neural network,’’ J. Comput. Appl. Math., vol. 369, May 2020,
performance of the model is weakened in multistep fore- Art. no. 112630.
casting. Therefore, in future research, we plan to consider [22] O. Sohaib, W. Hussain, M. Asif, M. Ahmad, and M. Mazzara, ‘‘A PLS-
upgrading the proposed model by combining other mul- SEM neural network approach for understanding cryptocurrency adop-
tion,’’ IEEE Access, vol. 8, pp. 13138–13150, 2020.
tidimensional and complex influencing factors, and hence [23] L. Cocco, R. Tonelli, and M. Marchesi, ‘‘Predictions of bitcoin prices
considering higher frequencies (e.g., hours) and longer time through machine learning based frameworks,’’ PeerJ Comput. Sci., vol. 7,
horizons (e.g., months or years). We look forward to this work p. e413, Mar. 2021.
[24] S. Alonso-Monsalve, A. L. Suárez-Cetrulo, A. Cervantes, and D. Quintana,
can better grasp the financial time series characteristics and ‘‘Convolution on neural networks for high-frequency trend prediction of
apply them to the actual portfolio strategy design to provide cryptocurrency exchange rates using technical indicators,’’ Expert Syst.
reference for investors and regulators. Appl., vol. 149, Jul. 2020, Art. no. 113250.
[25] R. Gupta, S. Tanwar, S. Tyagi, and N. Kumar, ‘‘Machine learning models
for secure data analytics: A taxonomy and threat model,’’ Comput. Com-
REFERENCES mun., vol. 153, pp. 406–440, Mar. 2020.
[1] S. Nakamoto. (2008). Bitcoin: A Peer-to-Peer Electronic Cash System. [26] M. Poongodi, A. Sharma, V. Vijayakumar, A. P. Sharma, R. Iqbal,
[Online]. Available: https://bitcoin.org/bitcoin.pdf and R. Kumar, ‘‘Prediction of the price of ethereum blockchain cryp-
[2] S. Aras, ‘‘On improving GARCH volatility forecasts for bitcoin via tocurrency in an industrial finance system,’’ Comput. Electr. Eng., vol. 81,
a meta-learning approach,’’ Knowl.-Based Syst., vol. 230, Oct. 2021, Jan. 2020, Art. no. 106527.
Art. no. 107393. [27] G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, ‘‘Extreme learning machine: A
[3] D. Shen, A. Urquhart, and P. Wang, ‘‘Does Twitter predict bitcoin?’’ Econ. new learning scheme of feedforward neural networks,’’ in Proc. IEEE Int.
Lett., vol. 174, pp. 118–122, Jan. 2019. Joint Conf. Neural Netw., Jul. 2004, pp. 985–990.
[4] H. Guo, D. Zhang, S. Liu, L. Wang, and Y. Ding, ‘‘Bitcoin price forecast- [28] G.-B. Huang, D. H. Wang, and Y. Lan, ‘‘Extreme learning machines: A
ing: A perspective of underlying blockchain transactions,’’ Decis. Support survey,’’ Int. J. Mach. Learn. Cybern., vol. 2, no. 2, pp. 107–122, Jun. 2011.
Syst., vol. 151, Dec. 2021, Art. no. 113650. [29] M. Shariati, M. S. Mafipour, P. Mehrabi, Y. Zandi, D. Dehghani,
[5] R. Chowdhury, M. A. Rahman, M. S. Rahman, and M. R. C. Mahdy, A. Bahadori, A. Shariati, N. T. Trung, M. N. Salih, and S. Poi-Ngian,
‘‘An approach to predict and forecast the price of constituents and index ‘‘Application of extreme learning machine (ELM) and genetic pro-
of cryptocurrency using machine learning,’’ Phys. A, Stat. Mech. Appl., gramming (GP) to design steel-concrete composite floor systems at
vol. 551, Aug. 2020, Art. no. 124569. elevated temperatures,’’ Steel Compos. Struct., vol. 33, pp. 319–332,
[6] Z. Zhang, H.-N. Dai, J. Zhou, S. K. Mondal, M. M. García, and Nov. 2019.
H. Wang, ‘‘Forecasting cryptocurrency price using convolutional neural [30] I. Ahmad, M. Basheri, M. J. Iqbal, and A. Raheem, ‘‘Performance com-
networks with weighted and attentive memory channels,’’ Expert Syst. parison of support vector machine, random forest, and extreme learning
Appl., vol. 183, Nov. 2021, Art. no. 115378. machine for intrusion detection,’’ IEEE Access, vol. 6, pp. 33789–33795,
[7] S. F. Cheng, G. De Franco, H. B. Jiang, and P. K. Lin, ‘‘Riding the 2018.
blockchain mania: Public firms’ speculative 8-K disclosures,’’ Manage. [31] S. Sun, S. Wang, and Y. Wei, ‘‘A new ensemble deep learning approach
Sci., vol. 65, pp. 5901–5913, Dec. 2019. for exchange rates forecasting and trading,’’ Adv. Eng. Informat., vol. 46,
[8] Y. Ma, F. Ahmad, M. Liu, and Z. Wang, ‘‘Portfolio optimization in the Oct. 2020, Art. no. 101160.
era of digital financialization using cryptocurrencies,’’ Technological Fore- [32] D. Aggarwal, S. Chandrasekaran, and B. Annamalai, ‘‘A complete empir-
casting Social Change, vol. 161, Dec. 2020, Art. no. 120265. ical ensemble mode decomposition and support vector machine-based
[9] I. Shaikh, ‘‘Policy uncertainty and bitcoin returns,’’ Borsa Istanbul Rev., approach to predict bitcoin prices,’’ J. Behav. Experim. Finance, vol. 27,
vol. 20, no. 3, pp. 257–268, Sep. 2020. Sep. 2020, Art. no. 100335.
[10] D. Aggarwal, ‘‘Do bitcoins follow a random walk model?’’ Res. Econ., [33] H. Lin, Q. Sun, and S.-Q. Chen, ‘‘Reducing exchange rate risks in interna-
vol. 73, no. 1, pp. 15–22, Mar. 2019. tional trade: A hybrid forecasting approach of CEEMDAN and multilayer
[11] J. Li, Y. Yuan, and F.-Y. Wang, ‘‘A novel GSP auction mechanism for LSTM,’’ Sustainability, vol. 12, no. 6, p. 2451, Mar. 2020.
ranking bitcoin transactions in blockchain mining,’’ Decis. Support Syst., [34] S. Sun, S. Wang, Y. Wei, and G. Zhang, ‘‘A clustering-based nonlinear
vol. 124, Sep. 2019, Art. no. 113094. ensemble approach for exchange rates forecasting,’’ IEEE Trans. Syst.,
[12] R. Matkovskyy and A. Jalan, ‘‘From financial markets to bitcoin markets: Man, Cybern. Syst., vol. 50, no. 6, pp. 2284–2292, Jun. 2020.
A fresh look at the contagion effect,’’ Finance Res. Lett., vol. 31, pp. 93–97, [35] W. Liu, W. D. Liu, and J. Gu, ‘‘Forecasting oil production using ensemble
Dec. 2019. empirical model decomposition based long short-term memory neural
[13] M. Muzammal, Q. Qu, and B. Nasrulin, ‘‘Renovating blockchain with network,’’ J. Petroleum Sci. Eng., vol. 189, Jun. 2020, Art. no. 107013.
distributed databases: An open source system,’’ Future Gener. Comput. [36] Y. Li, S. Wang, Y. Wei, and Q. Zhu, ‘‘A new hybrid VMD-ICSS-
Syst., vol. 90, pp. 105–117, Jan. 2019. BiGRU approach for gold futures price forecasting and algorithmic trad-
[14] N. Uras, L. Marchesi, M. Marchesi, and R. Tonelli, ‘‘Forecasting bitcoin ing,’’ IEEE Trans. Computat. Social Syst., vol. 8, no. 6, pp. 1357–1368,
closing price series using linear regression and neural networks models,’’ Dec. 2021.
PeerJ Comput. Sci., vol. 6, p. e279, Jul. 2020. [37] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng,
[15] B. Kapar and J. Olmo, ‘‘Analysis of bitcoin prices using market and N.-C. Yen, C. C. Tung, and H. H. Liu, ‘‘The empirical mode decomposition
sentiment variables,’’ World Economy, vol. 44, no. 1, pp. 45–63, Jan. 2021. and the Hilbert spectrum for nonlinear and non-stationary time series
[16] Z. H. Munim, M. H. Shakil, and I. Alon, ‘‘Next-day bitcoin price forecast,’’ analysis,’’ Proc. Roy. Soc. London Ser. A, Math., Phys. Eng. Sci., vol. 454,
J. Risk Financial Manage., vol. 12, no. 2, p. 103, Jun. 2019. no. 1971, pp. 903–995, Mar. 1998.
[17] D. T. Nguyen and H. V. Le, ‘‘Predicting the price of bitcoin using hybrid [38] Z. Wu and N. E. Huang, ‘‘Ensemble empirical mode decomposition: A
ARIMA and machine learning,’’ in Proc. Int. Conf. Future Data Secur. noise-assisted data analysis method,’’ Adv. Adapt. Data Anal., vol. 1, p. 41,
Eng., Vietnam, 2019, pp. 696–704. Jan. 2009.

60410 VOLUME 10, 2022

X. Du et al.: New Hybrid Cryptocurrency Returns Forecasting Method

[39] M. E. Torres, M. A. Colominas, G. Schlotthauer, and P. Flandrin, ‘‘A com- ZHENPENG TANG received the Ph.D. degree
plete ensemble empirical mode decomposition with adaptive noise,’’ in management science and engineering from the
in Proc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), Wuhan University of Technology, China, in 2003.
May 2011, pp. 4144–4147. He is currently a Professor with the School of Eco-
[40] K. Dragomiretskiy and D. Zosso, ‘‘Variational mode decomposition,’’ nomics and Management, Fujian Agriculture and
IEEE Trans. Signal Process., vol. 62, no. 3, pp. 531–544, Feb. 2014. Forestry University, China. He is the author of four
[41] W. Sun and J. Zhang, ‘‘Carbon price prediction based on ensemble books and over 70 journal articles. His research
empirical mode decomposition and extreme learning machine optimized
interests include financial risk management, big
by improved bat algorithm considering energy price factors,’’ Energies,
data analysis, artificial intelligence, and predictive
vol. 13, no. 13, p. 3471, Jul. 2020.
[42] J. Cao, Z. Li, and J. Li, ‘‘Financial time series forecasting model based modeling.
on CEEMDAN and LSTM,’’ Phys. A, Stat. Mech. Appl., vol. 519,
pp. 127–139, Apr. 2019.
[43] J. Zhu, P. Wu, H. Chen, J. Liu, and L. Zhou, ‘‘Carbon price forecasting with
variational mode decomposition and optimal combined model,’’ Phys. A,
Stat. Mech. Appl., vol. 519, pp. 140–158, Apr. 2019.
[44] M. Jiang, L. Jia, Z. Chen, and W. Chen, ‘‘The two-stage machine learning
ensemble models for stock price prediction by combining mode decom-
position, extreme learning machine and improved harmony search algo-
rithm,’’ Ann. Oper. Res., vol. 309, no. 2, pp. 553–585, Feb. 2022. JUNCHUAN WU received the Ph.D. degree in
[45] Q. Qin, K. Xie, H. He, L. Li, X. Chu, Y.-M. Wei, and T. Wu, financial engineering from Fuzhou University,
‘‘An effective and robust decomposition-ensemble energy price fore- China, in 2020. He is currently a Lecturer with the
casting paradigm with local linear prediction,’’ Energy Econ., vol. 83, School of Economics and Management, Nanchang
pp. 402–414, Sep. 2019. University, China. His research interests include
[46] W. Sun and C. Xu, ‘‘Carbon price prediction based on modified wavelet financial risk management, big data analysis, data
least square support vector machine,’’ Sci. Total Environ., vol. 754, mining, and predictive modeling.
Feb. 2021, Art. no. 142052.
[47] G. Xing, S. Sun, and J. Guo, ‘‘A new decomposition ensemble learning
approach with intelligent optimization for PM2.5 concentration forecast-
ing,’’ Discrete Dyn. Nature Soc., vol. 2020, pp. 1–11, Mar. 2020.
[48] B. Wang and J. Wang, ‘‘Energy futures and spots prices forecasting by
hybrid SW-GRU with EMD and error evaluation,’’ Energy Econ., vol. 90,
Aug. 2020, Art. no. 104827.
[49] X. Sun, H. Zhang, J. Wang, C. Shi, D. Hua, and J. Li, ‘‘Ensemble stream-
flow forecasting based on variational mode decomposition and long short
term memory,’’ Sci. Rep., vol. 12, no. 1, pp. 1–19, Jan. 2022.
[50] Z. Xu, J. Zhou, L. Mo, B. Jia, Y. Yang, W. Fang, and Z. Qin, ‘‘A novel runoff KAIJIE CHEN is currently pursuing the Ph.D.
forecasting model based on the decomposition-integration-prediction degree with the School of Economics and Man-
framework,’’ Water, vol. 13, no. 23, p. 3390, Dec. 2021. agement, Fuzhou University, China. His main
[51] E. Gulay and O. Duru, ‘‘Hybrid modeling in the predictive analyt- research interests include financial risk manage-
ics of energy systems and prices,’’ Appl. Energy, vol. 268, Jun. 2020, ment, big data mining, model selection, and price
Art. no. 114985. forecasting.
[52] H. Huang, J. Chen, X. Huo, Y. Qiao, and L. Ma, ‘‘Effect of multi-scale
decomposition on performance of neural networks in short-term traffic
flow prediction,’’ IEEE Access, vol. 9, pp. 50994–51004, 2021.
[53] J. Xue and B. Shen, ‘‘A novel swarm intelligence optimization approach:
Sparrow search algorithm,’’ Syst. Sci. Control Eng., vol. 8, no. 1, pp. 22–34,
Jan. 2020.
[54] K. E. Taylor, ‘‘Summarizing multiple aspects of model performance
in a single diagram,’’ J. Geophys. Res., Atmos., vol. 106, no. D7,
pp. 7183–7192, Apr. 2001.

XIAOXU DU is currently pursuing the Ph.D. YI CAI is currently pursuing the master’s degree
degree with the School of Economics and with the School of Economics and Management,
Management, Fuzhou University, China. His main Fuzhou University, China. His main research inter-
research interests include financial market com- ests include financial risk management, big data
plexity, financial risk management, big data min- mining, model selection, and price forecasting.
ing, model selection, and price forecasting.

VOLUME 10, 2022 60411