See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/339357269

Financial time series forecasting with deep learning : A systematic literature

review: 2005–2019

Article in Applied Soft Computing · February 2020

DOI: 10.1016/j.asoc.2020.106181

CITATIONS READS

1,214 6,330

3 authors:

Omer Berat Sezer Ugur Gudelek

TOBB University of Economics and Technology TOBB University of Economics and Technology
19 PUBLICATIONS 2,965 CITATIONS 16 PUBLICATIONS 1,969 CITATIONS

SEE PROFILE SEE PROFILE

Murat Ozbayoglu
TOBB University of Economics and Technology
143 PUBLICATIONS 5,151 CITATIONS

SEE PROFILE

All content following this page was uploaded by Omer Berat Sezer on 30 May 2020.

The user has requested enhancement of the downloaded file.

 Financial Time Series Forecasting with Deep Learning : A
Systematic Literature Review: 2005-2019
Omer Berat Sezera , M. Ugur Gudeleka , Ahmet Murat Ozbayoglua
a
Department of Computer Engineering, TOBB University of Economics and Technology, Ankara, Turkey

Abstract
Financial time series forecasting is undoubtedly the top choice of computational intelligence
for finance researchers in both academia and the finance industry due to its broad imple-
mentation areas and substantial impact. Machine Learning (ML) researchers have created
various models, and a vast number of studies have been published accordingly. As such, a
significant number of surveys exist covering ML studies on financial time series forecasting.
Lately, Deep Learning (DL) models have appeared within the field, with results that signifi-
cantly outperform their traditional ML counterparts. Even though there is a growing interest
in developing models for financial time series forecasting, there is a lack of review papers that
solely focus on DL for finance. Hence, the motivation of this paper is to provide a compre-
hensive literature review of DL studies on financial time series forecasting implementation.
We not only categorized the studies according to their intended forecasting implementation
areas, such as index, forex, and commodity forecasting, but we also grouped them based
on their DL model choices, such as Convolutional Neural Networks (CNNs), Deep Belief
Networks (DBNs), and Long-Short Term Memory (LSTM). We also tried to envision the
future of the field by highlighting its possible setbacks and opportunities for the benefit of
interested researchers.
Keywords: deep learning, finance, computational intelligence, machine learning, time
series forecasting, CNN, LSTM, RNN

1. Introduction
The finance industry has always been interested in the successful prediction of financial
time series data. Numerous studies have been published on ML models with relatively better
performances than classical time series forecasting techniques. Meanwhile, the widespread
application of automated electronic trading systems coupled with increasing demand for
higher yields keeps forcing researchers and practitioners to continue working on implementing
better models. Hence, new publications and implementations keep adding to the finance and
computational intelligence literature.
In the last few years, DL has strongly emerged as the best performing predictor class
within the ML field in various implementation areas. Financial time series forecasting is
no exception, and as such, an increasing number of prediction models based on various DL
techniques have been introduced in the appropriate conferences and journals in recent years.
Preprint submitted to Applied Soft Computing May 30, 2020
Despite the vast number of survey papers covering financial time series forecasting and
trading systems using traditional soft computing techniques, to the best of our knowledge,
no reviews have been performed on the literature for DL. Hence, we decided to work on such
a comprehensive study, focusing on DL implementations of financial time series forecasting.
Our motivation is two-fold; we not only aimed at providing a state-of-the-art snapshot
of academic and industry perspectives of developed DL models, but we also pinpoint the
important and distinctive characteristics of each studied model to prevent researchers and
practitioners from making unsatisfactory choices during their system development. We also
wanted to envision where the industry is heading by indicating possible future directions.
Our fundamental motivation was to answer the following research questions:

• Which DL models are used for financial time series forecasting ?

• How does the performance of DL models compare with that of their traditional ML
counterparts ?

• What is the future direction of DL research for financial time series forecasting ?

Our focus was solely on DL implementations for financial time series forecasting. For
other DL-based financial applications, such as risk assessment and portfolio management,
interested readers can refer to another recent survey paper [1]. Because we wanted to
single out financial time series prediction studies in our survey, we omitted other time series
forecasting studies that were not focused on financial data. Meanwhile, we included time-
series research papers that had financial use cases or examples, even if the papers themselves
were not directly concerned with financial time series forecasting. Also, we decided to include
algorithmic trading papers that were based on financial forecasting but ignore the ones that
did not have a time series forecasting component.
We mainly reviewed journals and conferences for our survey, but we also included Masters
and PhD theses, book chapters, arXiv papers, and noteworthy technical publications that
came up in web searches. We decided to only include articles published in English language.
During our survey, we realized that most of the papers using the term “deep learning" in
their description were published in the past five years. However, we also encountered some
older studies that implemented deep models, such as Recurrent Neural Networks (RNNs)
and Jordan-Elman networks. However, at their time of publication, the term “deep learning"
was not in common usage. Therefore, we decided to also include those papers.
According to our findings, this will be one of the first comprehensive “financial time
series forecasting" survey papers focusing on DL. Many ML reviews for financial time series
forecasting exist in the literature, but we have not encountered any study on DL. Hence, we
wanted to fill this gap by analyzing the developed models and applications accordingly. We
hope that as a result of this paper, researchers and model developers will have a better idea
of how they can implement DL models in their studies.
The remainder of this paper is structured as follows. In Section 2, existing surveys focused
on ML and soft computing studies for financial time series forecasting are mentioned. In
Section 3, we will cover existing DL models that are used, such as CNN, LSTM, and Deep
2
Reinforcement Learning (DRL). Section 4 will focus on the various financial time series
forecasting implementation areas using DL, namely stock forecasting, index forecasting,
trend forecasting, commodity forecasting, volatility forecasting, foreign exchange forecasting,
and cryptocurrency forecasting. In each subsection, the problem definition will be given,
followed by the particular DL implementations. In Section 5, overall statistical results about
our findings will be presented, including histograms related to the annual distributions of
different subfields, models, publication types, etc. A state-of-the-art snapshot of financial
time series forecasting studies will be given through these statistics. At the same time,
they will also show the areas that are already mature in comparison with promising or new
areas that still have room for improvement. Section 6 discusses the academic and industrial
achievements that have been accomplished and future expectations. The section will include
highlights of open areas that require further research. Finally, we conclude this paper in
Section 7 by summarizing our findings.

2. Financial Time Series Forecasting with ML

Financial time series forecasting and associated applications have been studied exten-
sively for many years. When ML started gaining popularity, financial prediction applica-
tions based on soft computing models accordingly also became available. Even though our
particular focus is on DL implementations of financial time series prediction studies, it will
be beneficial to briefly mention existing surveys covering ML-based financial time series
forecasting studies to provide some historical perspective.
In our study, we did not include any survey papers that were focused on specific finan-
cial application areas other than forecasting studies. However, we were faced with some
review publications that included a mix of financial time-series studies and other financial
applications. We decided to include those papers to maintain the comprehensiveness of our
coverage.
Examples of these aforementioned publications are provided here. There were published
books on stock market forecasting [2], trading system development [3], practical examples
of forex and market forecasting applications [4] using ML models, such as Artificial Neural
Networks (ANNs), Evolutionary Computations (ECs), and Genetic Programming (GP), and
Agent-based models [5].
There were also some existing journal and conference surveys. Bahrammirzaee et al. [6]
surveyed financial prediction and planning studies along with other financial applications
using various Artificial Intelligence (AI) techniques such as ANN, Expert Systems, and
hybrid models. Zhang et al. [7] also compared ML methods in different financial applications,
including stock market prediction studies. In Mochon et al. [8], soft computing models for
the market, forex prediction, and trading systems were analyzed. Mullainathan and Spies
[9] surveyed the prediction process in general from an econometric perspective.
There were also a number of survey papers concentrated on one particular ML model.
Even though these papers focused on one technique, the implementation areas generally
spanned various financial applications, including financial time series forecasting. Among
those soft computing methods, EC and ANN have had the most overall interest.
3
 In terms of EC studies, Chen wrote a book on Genetic Algorithms (GAs) and GP in
Computational Finance [10]. Later, Multiobjective Evolutionary Algorithms (MOEAs) were
extensively surveyed for various financial applications including financial time series predic-
tion [11, 12, 13]. Meanwhile, Rada reviewed EC applications along with Expert Systems for
financial investing models [14].
In terms of ANN studies, Li and Ma reviewed implementations of ANN for stock price
forecasting and some other financial applications [15]. Tkac et al. [16] surveyed different im-
plementations of ANN in financial applications, including stock price forecasting. Recently,
Elmsili and Outtaj surveyed ANN applications in economics and management research,
including economic time series forecasting [17].
There have also been several text mining surveys focused on financial applications, in-
cluding financial time series forecasting. Mittermayer and Knolmayer compared various text
mining implementations that extract market responses to news for prediction [18]. Mitra
et al. [19] focused on news analytics studies for prediction of abnormal returns for trading
strategies in their survey. Nassirtoussi et al. reviewed text mining studies for stock or forex
market prediction [20]. Kearney et al. [21] also surveyed text mining-based time series
forecasting and trading strategies using textual sentiment. Similarly, Kumar and Ravi [22]
reviewed text mining studies for forex and stock market prediction. Lately, Xing et al. [23]
surveyed natural language-based financial forecasting studies.
Finally, there were application-specific survey papers that focused on particular financial
time series forecasting implementations. Among these studies, stock market forecasting
had the most interest. A number of surveys were published for stock market forecasting
studies based on various soft computing methods at different times [24, 25, 26, 27, 28, 29,
30, 31]. Chatterjee et al. [32] and Katarya and Mahajan [33] concentrated on ANN-based
financial market prediction studies, whereas Hu et al. [34] focused on EC implementations
for stock forecasting and algorithmic trading models. In a different time series forecasting
application, researchers surveyed forex prediction studies using ANN [35] and various other
soft computing techniques [36].
Although many surveys exist for ML implementations of financial time series forecast-
ing, DL has not yet been surveyed comprehensively despite the emergence of various DL
implementations in recent years. This was the main motivation for our survey. In the next
section, we cover the various DL models used in financial time series forecasting studies.

3. Deep Learning
DL is a type of ANN that consists of multiple processing layers and enables high-level
abstraction to model data. The key advantage of DL models is extracting the good features
of input data automatically using a general-purpose learning procedure. Therefore, DL
models have been proposed for many applications such as: image, speech, video, and audio
reconstruction, natural language understanding (particularly topic classification), sentiment
analysis, question answering, and language translation [37]. The historical improvements of
DL models are surveyed in Schmidhuber et al. [38].

4
 Financial time series forecasting has been very popular among ML researchers for more
than 40 years. The financial community has been boosted by the recent introduction of DL
models for financial prediction and their accompanying publications. The success of DL over
ML models is the major attractive point for finance researchers. With more financial time
series data and different deep architectures, new DL methods will be proposed. In our survey,
the vast majority of studies found DL models to be better than their ML counterparts.
In the literature, there are different kinds of DL models: Deep Multilayer Perceptron
(DMLP), RNN, LSTM, CNN, Restricted Boltzmann Machines (RBMs), DBN, Autoencoder
(AE), and DRL [37, 38]. Throughout the literature, financial time series forecasting is mostly
considered as a regression problem. However, there is also a significant number of studies,
particularly on trend prediction, that use classification models to tackle financial forecasting
problems. In Section 4, different DL implementations are presented along with their model
choices.

3.1. Deep Multi Layer Perceptron (DMLP)

DMLP was one of the first developed ANNs. Its difference from shallow nets is that
DMLP contains more layers. Even though particular model architectures might have vari-
ations depending on different problem requirements, DMLP models consist mainly of three
layers: input, hidden, and output. The number of neurons in each layer and the number of
layers are the hyperparameters of the network. In general, each neuron in the hidden layers
has input (x), weight (w), and bias (b) terms. In addition, each neuron has a nonlinear acti-
vation function, which produces a cumulative output of the preceding neurons. Equation 1
[39] illustrates the output of a single neuron in the Neural Network (NN). There are different
types of nonlinear activation functions. The most commonly used nonlinear activation func-
tions are sigmoid (Equation 2) [40], hyperbolic tangent (Equation 3) [41], Rectified Linear
Unit (ReLU) (Equation 4) [42], leaky-ReLU (Equation 5) [43], swish (Equation 6) [44], and
softmax (Equation 7) [39]. Non-linear activations were compared in [44].
X
yi = σ( Wi xi + bi ) (1)
i

1
σ(z) = (2)
1 + e−z
ez − e−z
tanh(z) = (3)
ez + e−z

R(z) = max (0, z) (4)

R(z) = 1(x < 0)(αx) + 1(x >= 0)(x) (5)

f (x) = xσ(βx) (6)

5
 exp zi
softmax(zi ) = P (7)
exp zj
j

DMLP models have appeared in various application areas [45, 37] . Using a DMLP
model has advantages and disadvantages depending on the problem requirements. Through
DMLP models, problems such as regression and classification can be solved by modeling
the input data [46]. However, if the number of input features is increased (e.g., image as
input), the parameter size in the network will increase accordingly due to the fully connected
nature of the model, which will jeopardize the computational performance and create storage
problems. To overcome this issue, different types of Deep Neural Network (DNN) methods
have been proposed (such as CNN) [37]. With DMLP, much more efficient classification and
regression processes can be performed. In Figure 1, a DMLP model’s layers, neurons, and
weights between neurons are illustrated.

Figure 1: Deep Multi Layer Neural Network Forward Pass and Backpropagation [37]

The DMLP learning stage is implemented through backpropagation. The error in the
neurons in the output layer is propagated back to the preceeding layers. Optimization al-
gorithms are used to find the optimum parameters/variables of the NNs. They are used to
update the weights of the connections between the layers. Different optimization algorithms
have been developed: Stochastic Gradient Descent (SGD), SGD with Momentum, Adaptive
Gradient Algorithm (AdaGrad), Root Mean Square Propagation (RMSProp), and Adaptive
Moment Estimation (ADAM) [47, 48, 49, 50, 51]. Gradient descent is an iterative method
to find optimum parameters of the function that minimizes the cost function. SGD is an
algorithm that randomly selects a few samples for each iteration instead of the whole data
set [47]. SGD with Momentum remembers the update in each iteration, which accelerates
gradient descent [48]. AdaGrad is a modified SGD that improves on the convergence per-
formance of the standard SGD algorithm [49]. RMSProp is an optimization algorithm that
6
adapts the learning rate for each parameter. In RMSProp, the learning rate is divided by
a running average of the magnitudes of recent gradients for that weight [50]. ADAM is an
updated version of RMSProp that uses running averages of both the gradients and second
moments of the gradients. ADAM combines the advantages of RMSProp (works well in
online and non-stationary settings) and AdaGrad (works well with sparse gradients) [51].
As shown in Figure 1, the effect of backpropagation is transferred to the previous lay-
ers. If the effect of SGD is gradually lost when the effect reaches the early layers during
backpropagation, this problem is called the vanishing gradient problem [52]. In this case, up-
dates between the early layers become unavailable and the learning process stops. The high
number of layers in the neural network and the increasing complexity cause the vanishing
gradient problem.
The important issue in the DMLP are the hyperparameters of the networks and method
of tuning these hyperparameters. Hyperparameters are the variables of the network that af-
fect the network architecture and performance of the networks. The number of hidden layers,
number of units in each layer, regularization techniques (dropout, L1, L2), network weight
initialization (zero, random, He [53], Xavier [54]), activation functions (Sigmoid, ReLU, hy-
perbolic tangent, etc.), learning rate, decay rate, momentum values, number of epochs, batch
size (minibatch size), and optimization algorithms (SGD, AdaGrad, RMSProp, ADAM, etc.)
are the hyperparameters of DMLP. Choosing better hyperparameter values/variables for the
network results in better performance. Therefore, finding the best hyperparameters for the
network is a significant issue. In the literature, there are different methods to find best hy-
perparameters: Manual Search (MS), Grid Search (GS), RandomSearch (RS), and Bayesian
Methods (Sequential Model-Based Global Optimization (SMBGO), The Gaussian Process
Approach (GPA), Tree-structured Parzen Estimator Approach (TSPEA)) [55, 56].

3.2. Recurrent Neural Network (RNN)

The RNN is another type of DL network used for time series or sequential data, such
as language and speech. RNNs are also used in traditional ML models (Back Propagation
Through Time (BPTT), Jordan-Elman networks, etc.); however, the time periods in such
models are generally less than those used in deep RNN models. Deep RNNs are preferred
due to their ability to include longer time periods. Unlike Fully Connected Neural Networks
(FNNs), RNNs use internal memory to process incoming inputs. RNNs are used to analyze
time series data in various fields (handwriting recognition, speech recognition, etc.). As
stated in the literature, RNNs are good at predicting the next character in text, language
translation applications, and sequential data processing [45, 37].
The RNN model architecture consists of different numbers of layers and different types of
units in each layer. The main difference between RNN and FNN is that each RNN unit takes
the current and previous input data at the same time. The output depends on the previous
data in the RNN model. RNNs process input sequences one by one at any given time during
their operation. The units in the hidden layer hold information about the history of the
input in the “state vector". When the output of the units in the hidden layer is divided
into different discrete time steps, an RNN is converted into a DMLP [37]. In Figure 2, the
information flow in the RNN’s hidden layer is divided into discrete times. The status of the
7
node S at different times of t is shown as st , the input value x at different times is xt , and
the output value o at different times is shown as ot . Parameter values (U, W, V ) are always
used in the same step.

Figure 2: RNN cell through time[37]

RNNs can be trained using the BPTT algorithm. Optimization algorithms (SGD, RM-
SProp, ADAM) are used for the weight adjustment process. With the BPTT learning
method, the error change at time t is reflected in the input and weights of the previous t
times. The difficulty of training an RNN is that the RNN structure has a backward de-
pendence over time. Therefore, RNNs become increasingly complex as the learning period
increases. Although the main aim of using an RNN is to learn long-term dependencies,
studies in the literature show that when knowledge is stored for long time periods, it is not
easy to learn with an RNN [57]. To solve this particular problem, LSTMs with different
structures of ANN have been developed [37]. Equations 8 and 9 illustrate simpler RNN
formulations. Equation 10 shows the total error, which is the sum of the errors from each
time iteration1 .

ht = W f (ht−1 ) + W (hx) x[t] (8)

yt = W (S) f (ht ) (9)

T
∂E X ∂Et
= (10)
∂W t=1
∂W
Hyperparameters of the RNN also define the network architecture, and the performance
of the network is affected by the parameter choices, as in DMLP case. The number of
hidden layers, number of units in each layer, regularization techniques, network weight
initialization, activation functions, learning rate, momentum values, number of epochs, batch
size (minibatch size), decay rate, optimization algorithms, model (Vanilla RNN, Gated-
Recurrent Unit (GRU), LSTM), and sequence length are the hyperparameters of RNN.
Finding the best hyperparameters for the network is a significant issue. In the literature,
there are different methods to find the best hyperparameters: MS, GS, RS, and Bayesian
Methods (SMBGO, GPA, TSPEA) [55, 56].

1
Richard Socher, CS224d: Deep Learning for Natural Language Processing, Lecture Notes
8
3.3. Long Short Term Memory (LSTM)
LSTM [58] is a type of RNN where the network can remember both short term and long
term values. LSTM networks are the preferred choice of many DL model developers when
tackling complex problems such as automatic speech and handwritten character recognition.
LSTM models are mostly used with time-series data. Their applications include Natural
Language Processing (NLP), language modeling, language translation, speech recognition,
sentiment analysis, predictive analysis, and financial time series analysis [59, 60]. With
attention modules and AE structures, LSTM networks can be more successful in time series
data analysis, such as language translation [59].
LSTM networks consist of LSTM units. LSTM units merge to form an LSTM layer. An
LSTM unit is composed of cells, each with an input gate, output gate, and forget gate. These
gates regulate the information flow. With these features, each cell remembers the desired
values over arbitrary time intervals. Equations 11-15 show the form of the forward pass of
the LSTM unit [58] (xt : input vector to the LSTM unit, ft : forget gate’s activation vector,
it : input gate’s activation vector, ot : output gate’s activation vector, ht : output vector of the
LSTM unit, ct : cell state vector, σg : sigmoid function, σc , σh : hyperbolic tangent function,
∗: element-wise (Hadamard) product, W , U : weight matrices to be learned, b: bias vector
parameters to be learned) [60].

ft = σg (Wf xt + Uf ht−1 + bf ) (11)

it = σg (Wi xt + Ui ht−1 + bi ) (12)

ot = σg (Wo xt + Uo ht−1 + bo ) (13)

ct = ft ∗ ct−1 + it ∗ σc (Wc xt + Uc ht−1 + bc ) (14)

ht = ot ∗ σh (ct ) (15)
LSTM is a specialized version of RNN. Therefore, the weight updates and preferred op-
timization methods are the same. In addition, the hyperparameters of LSTM are just like
those of RNN: number of hidden layers, number of units in each layer, network weight ini-
tialization, activation functions, learning rate, momentum values, number of epochs, batch
size (minibatch size), decay rate, optimization algorithms, sequence length for LSTM, gradi-
ent clipping, gradient normalization, and dropout[60, 61]. To find the best hyperparameters
of LSTM, the hyperparameter optimization methods used for RNN are also applicable to
LSTM [55, 56].

9
3.4. Convolutional Neural Networks (CNNs)
The CNN is a type of DNN that consists of convolutional layers based on the convo-
lutional operation. It is the most common model used for vision and image processing-
based classification problems (image classification, object detection, image segmentation,
etc.) [62, 63, 64]. The advantage of the CNN is the number of parameters compared to
vanilla DL models, such as DMLP. Filtering with the kernel window function gives the ad-
vantage of image processing to CNN architectures with fewer parameters, which is beneficial
for computing and storage. In CNN architectures, there are different layers: convolutional,
max-pooling, dropout, and fully connected Multilayer Perceptron (MLP) layer. The convo-
lutional layer consists of a convolution (filtering) operation. A basic convolution operation
is shown in Equation 16, where t denotes time, s denotes feature map, w denotes kernel,
x denotes input, and a denotes variable. In addition, the convolution operation is imple-
mented on two-dimensional images. Equation 17 shows the convolution operation for a
two-dimensional image, where I denotes input image, K denotes the kernel, (m, n) denotes
image dimensions, and i and j denote variables. Consecutive convolutional and max-pooling
layers construct the deep network. Equation 18 describes the NN architecture, where W
denotes weights, x denotes input, b denotes bias, and z denotes the output of neurons. At
the end of the network, the softmax function is used to obtain the output. Equations 19
and 20 illustrate the softmax function, where y denotes output [39].
∞
X
s(t) = (x ∗ w)(t) = x(a)w(t − a) (16)
a=−∞
XX
S(i, j) = (I ∗ K)(i, j) = I(m, n)K(i − m, j − n). (17)
m n
X
zi = W i , j x j + bi . (18)
j

y = softmax(z) (19)

exp(zi )
softmax(zi ) = P (20)
exp(zj )
j

The backpropagation process is used for CNN model learning. Most commonly used
optimization algorithms (SGD, RMSProp) are used to find optimum CNN parameters. Hy-
perparameters of the CNN are similar to other DL model hyperparameters: number of
hidden layers, number of units in each layer, network weight initialization, activation func-
tions, learning rate, momentum values, number of epochs, batch size (minibatch size), decay
rate, optimization algorithms, dropout, kernel size, and filter size. To find the best CNN
hyperparameters, the following search algorithms are commonly used: MS, GS, RS, and
Bayesian methods. [55, 56].

10
3.5. Restricted Boltzmann Machines (RBMs)
An RBM is a productive stochastic ANN that can learn a probability distribution on
the input set [65]. RBMs are mostly used for unsupervised learning [66]. RBMs are used in
applications such as dimension reduction, classification, feature learning, and collaborative
filtering [67]. The advantage of RBMs is their ability to find hidden patterns in an unsu-
pervised manner. The disadvantage of RBMs is its difficult training process. “RBMs are
tricky because although there are good estimators of the log-likelihood gradient, there are
no known cheap ways of estimating the log-likelihood itself" [68].

Figure 3: RBM Visible and Hidden Layers [65]

An RBM is a two-layer, bipartite, and undirected graphical model that consists of two
layers: visible and hidden (Figure 3). The layers are not connected among themselves. Each
cell is a computational point that processes the input and makes stochastic decisions about
whether this nerve node will transmit the input. Inputs are multiplied by specific weights,
certain threshold values (bias) are added to input values, and then the calculated values
are passed through an activation function. In the reconstruction stage, the results from the
outputs re-enter the network as input before finally exiting the visible layer as output. The
values of the previous input and values after the processes are compared. The purpose of
this comparison is to reduce the difference.
Equation 21 illustrates the probabilistic semantics for an RBM using its energy function,
where P denotes the probabilistic semantics for an RBM, Z denotes the partition function, E
denotes the energy function, h denotes hidden units, and v denotes visible units. Equation 22
illustrates the partition function or normalizing constant. Equation 23 shows the energy of a
configuration (in matrix notation) of a standard RBM with binary-valued hidden and visible
units, where a denotes bias weights (offsets) for the visible units, b denotes bias weights for
the hidden units, W denotes matrix weight of the connection between hidden and visible
units, T denotes the transpose of matrix, v denotes visible units, and h denotes hidden units
[69, 70].
1
exp(−E(v, h))
P (v, h) = (21)
Z
XX
Z= exp(−E(v, h)) (22)
v h

11
 E(v, h) = −aT v − bT h − v T W h (23)
The learning is performed multiple times on the network [65]. The training of RBMs is
implemented by minimizing the negative log-likelihood of the model and data. The Con-
trastive Divergence (CD) algorithm is used as the stochastic approximation algorithm, which
replaces the model expectation using an estimation using Gibbs Sampling with a limited
number of iterations [66]. In the CD algorithm, the Kullback Leibler Divergence (KL-
Divergence) algorithm is used to measure the distance between its reconstructed probability
distribution and the original probability distribution of the input [71].
Momentum, learning rate, weight-cost (decay rate), batch size (minibatch size), reg-
ularization method, number of epochs, number of layers, initialization of weights, size of
visible units, size of hidden units, type of activation units (sigmoid, softmax, ReLU, Gaus-
sian units), loss function, and optimization algorithms are the hyperparameters of RBMs.
Similar to other deep networks, the hyperparameters are searched with MS, GS, RS, and
Bayesian methods (Gaussian process). In addition to these, Annealed Importance Sampling
(AIS) is used to estimate the partition function. The CD algorithm is also used for the
optimization of RBMs [55, 56, 72, 73].

3.6. Deep Belief Networks (DBNs)

A DBN is a type of deep ANN consisting of a stack of RBM networks (Figure 4). A
DBN is a probabilistic generative model that consists of latent variables. In a DBN, there
is no link between units in each layer. DBNs are used to find discriminate independent
features in the input set using unsupervised learning [69]. The ability to encode higher-
order network structures and fast inference are the advantages of DBNs [74]. DBNs have
the same disadvantages as RBMs because DBNs are composed of RBMs.

Figure 4: Deep Belief Network [65]

When a DBN is trained in an unsupervised manner, it can learn to reconstruct the input
set in a probabilistic manner. Then, the layers in the network begin to detect discriminating
features in the input. After this learning step, supervised learning is conducted for classifi-
cation [75]. Equation 24 illustrates the probability of generating a visible vector (W : matrix
12
weight of connection between hidden unit h and visible unit v, p(h|W ): prior distribution
over hidden vectors) [69].
X
p(v) = p(h|W )p(v|h, W ) (24)
h

The DBN training process can be divided into two steps: stacked RBM learning and
backpropagation learning. In stacked RBM learning, an iterative CD algorithm is used
[66]. In backpropagation learning, optimization algorithms (SGD, RMSProp, ADAM) are
used to train the network [74]. The hyperparameters of a DBNs are similar to thoseof an
RBM. Momentum, learning rate, weight-cost (decay rate), regularization method, batch size
(minibatch size), number of epochs, number of layers, initialization of weights, number of
RBM stacks, size of visible units in RBMs’ layers, size of hidden units in RBMs’ layers,
type of units (sigmoid, softmax, rectified, Gaussian units, etc.), network weight initializa-
tion, and optimization algorithms are the hyperparameters of DBNs. Similar to other deep
networks, the hyperparameters are searched with MS, GS, RS, and Bayesian methods. The
CD algorithm is also used for the optimization of DBNs [55, 56, 72, 73].

3.7. Autoencoders (AEs)

AE networks are ANNs used as unsupervised learning models. In addition, AE networks
are commonly used in DL models, wherein they remap the inputs (features) such that the
inputs are more representative for classification. In other words, AE networks perform
an unsupervised feature learning process, which fits very well with the DL framework. A
representation of a data set is learned by reducing the dimensionality with AEs. AEs are
similar to Feedforward Neural Networks (FFNNs)’ in their architecture. They consist of an
input layer, an output layer, and one or more hidden layers that connect them together. The
number of nodes in the input layer and the number of nodes in the output layer are equal to
each other in AEs, and they have a symmetrical structure. The most notable advantages of
AEs are dimensionality reduction and feature learning. Meanwhile, reducing dimensionality
and feature extraction in AEs cause some drawbacks. Focusing on minimizing the loss of
data relationship in the encoding of AEs causes the loss of some significant data relationships.
Hence, this may be considered as a drawback of AEs[76].
In general, AEs contain two components: encoder and decoder. The input x ∈ [0, 1]d
is converted through function f (x) (W1 denotes a weight matrix, b1 denotes a bias vector,
σ1 element-wise sigmoid activation function of the encoder). Output h is the encoded part
of the AEs (code), latent variables, or latent representation. The inverse of function f (x),
called function g(h), produces the reconstruction of output r (W2 denotes a weight matrix, b2
denotes a bias vector, and σ2 is an element-wise sigmoid activation function of the decoder).
Equations 25 and 26 illustrate the simple AE process [77]. Equation 27 shows the loss
function of the AE, the Mean Squared Error (MSE). In the literature, AEs have been used
for feature extraction and dimensionality reduction [39, 77].

h = f (x) = σ1 (W1 x + b1 ) (25)

13
 r = g(h) = σ2 (W2 h + b2 ) (26)

L(x, r) = ||x − r||2 (27)

AEs are a specialized version of FFNNs. The backpropagation learning is used for up-
dating the weights in the network [39]. Optimization algorithms (SGD, RMSProp, ADAM)
are used for the learning process of AEs. MSE is used as a loss function in AEs. In addition,
recirculation algorithms may also be used for the training of AEs [39]. AEs’ hyperparam-
eters are similar to those of DL hyperparameters. Learning rate, weight-cost (decay rate),
dropout fraction, batch size (minibatch size), number of epochs, number of layers, number
of nodes in each encoder layer, type of activation functions, number of nodes in each decoder
layers, network weight initialization, optimization algorithms, and number of nodes in the
code layer (size of latent representation) are the hyperparameters of AEs. Similar to other
deep networks, the hyperparameters are searched with MS, GS, RS, and Bayesian methods
[55, 56].

3.8. Deep Reinforcement Learning (DRL)

Reinforcement learning (RL) is a type of learning that differs from supervised and unsu-
pervised learning models. It does not require a preliminary data set that has been labeled
or clustered before. RL is an ML approach inspired by learning action/behavior, which
deals with what actions should be taken by subjects to achieve the highest reward in an
environment. There are different areas in which it is used: game theory, control theory,
multi-agent systems, operations research, robotics, information theory, investment portfolio
management, simulation-based optimization, playing Atari games, and statistics [78]. Some
advantages of using RL for control problems are that an agent can be easily re-trained to
adapt to changes in the environment and that the system is continually improved while
training is constantly performed. An RL agent learns by interacting with its surroundings
and observing the results of these interactions. This learning method mimics the basics of
how humans learn.
RL is mainly based on a Markov Decision Process (MDP). A MDP is used to formalize
the RL environment. A MDP consists of five tuples: state (finite set of states), action (finite
set of actions), reward function (scalar feedback signal), state transition probability matrix
(p(s0 , r|s, a), where s0 denotes next state, r denotes reward function, s denotes state, and a
denotes action), and discount factor (γ, present value of future rewards). The aim of the
agent is to maximize the cumulative reward. The return (Gt ) is the total discounted reward.
Equation 28 illustrates the total return, where R denotes rewards, t denotes time, and k
denotes a variable in time.
∞
X
2
Gt = Rt+1 + γRt+2 + γ Rt+3 + ... = γ k Rt+k+1 (28)
k=0
The value function is the prediction of future values. It provides information on how
good the state/action is. Equation 29 illustrates the formulation of the value function,
14
where v(s) denotes the value function, E[.] denotes the expectation function, Gt denotes the
total discounted reward, s denotes the given state, R denotes the rewards, S denotes the set
of states, and t denotes time.

v(s) = E[Gt |St = s] = E[Rt+1 + γv(St+1 )|St = s] (29)

Policy (π) is the agent’s behavior strategy. It is like a map from state to action. There
are two types of value functions to express the actions in the policy: state-value function
(vπ (s)) and action-value function (qπ (s, a)). The state-value function (Equation 30) is the
expected return of starting from s to following policy π (Eπ [.] denotes expectation function).
The action-value function (Equation 31) is the expected return of starting from s and taking
action a to following policy π (A denotes the set of actions and a denotes the given action).
X∞
vπ (s) = Eπ [Gt |St = s] = Eπ [ γ k Rt+k+1 |St = s] (30)
k=0

qπ (s, a) = Eπ [Gt |St = s, At = a] (31)

The optimal state-value function (Equation 32) is the maximum value function over all
policies. The optimal action-value function (Equation 33) is the maximum action-value
function over all policies.

v∗ (s) = max(vπ (s)) (32)

q∗ (s, a) = max(qπ (s, a)) (33)

The RL solutions and methods in the literature are too broad to review in this pa-
per. Therefore, we summarize the important issues of RL and important RL solutions and
methods. RL methods can mainly be divided into two types: model-based methods and
model-free methods. Model-based methods use a model that is known by the agent before,
value/policy, and experience. The experience can be real (sample from the environment) or
simulated (sample from the model). Model-based methods are mostly used in the applica-
tions of robotics and control algorithms [79]. Model-free methods are mainly divided into
two groups: value-based and policy-based. In value-based methods, a policy is produced
directly from the value function (e.g., epsilon-greedy). In policy-based methods, the policy
is parametrized directly. In value-based methods, there are three main solutions for MDP
problems: Dynamic Programming (DP), Monte Carlo (MC), and Temporal Difference (TD).
In the DP method, problems are solved with optimal substructure and overlapping sub-
problems. The full model is known and is used for planning in MDP. There are two iterations
(learning algorithms) in DP: policy iteration and value iteration. The MC method learns
experience directly by running an episode of game/simulation. MC is a type of model-free
method that does not require MDP transitions/rewards. It collects states and takes the
mean of returns for the value function. TD is also a model-free method that learns the
experience directly by running the episode. In addition, TD learns incomplete episodes
15
like the DP method by using bootstrapping. The TD method combines the MC and DP
methods. SARSA (state, action, reward, state, action; St , At , Rt , St+1 , At+1 ) is a type of
TD control algorithm. Q-value (action-value function) is updated with the agent actions.
It is an on-policy learning model that learns from actions according to the current policy
π. Equation 34 illustrates the update of the action-value function in the SARSA algorithm,
where St denotes current state, At denotes current action, t denotes time, R denotes re-
ward, α denotes learning rate, γ denotes discount factor. Q-learning is another TD control
algorithm. It is an off-policy learning model that learns from different actions that do not
require the policy π at all. Equation 35 illustrates the update of the action-value function
in the Q-Learning algorithm (the whole algorithm is described in [78], a0 denotes action).

Q(St , At ) = Q(St , At ) + α[R(t + 1) + γQ(St+1 , At+1 ) − Q(St , At )] (34)

Q(St , At ) = Q(St , At ) + α[R(t + 1) + γmaxa0 Q(St+1 , a0 ) − Q(St , At )] (35)

In the value-based methods, a policy can be generated directly from the value function
(e.g., using epsilon-greedy). Policy-based methods use the policy directly instead of using the
value function. It has both advantages and disadvantages over value-based methods. The
policy-based methods are more effective in high-dimensional or continuous action spaces and
have better convergence properties than value-based methods. It can also learn stochastic
policies. On the other hand, policy-based methods evaluate a policy that is typically ineffi-
cient and has high variance. It typically converges to a local rather than global optimum.
In the policy-based methods, there are also different solutions: Policy gradient, Reinforce
(Monte-Carlo Policy Gradient), and Actor-Critic [78] (details of policy-based methods can
be found in [78]).
DRL methods contain NNs. Therefore, DRL hyperparameters are similar to the DL
hyperparameters. Learning rate, weight-cost (decay rate), dropout fraction, regularization
method, batch size (minibatch size), number of epochs, number of layers, number of nodes
in each layer, type of activation functions, network weight initialization, optimization algo-
rithms, discount factor, and number of episodes are the hyperparameters of DRL. Similar
to other deep networks, the hyperparameters are searched with MS, GS, RS and Bayesian
methods [55, 56].

4. Financial Time Series Forecasting

The most widely studied financial application area is forecasting of a given financial
time series, particularly asset price forecasting. Even though some variations exist, the
main focus is on predicting the next movement of the underlying asset. More than half
of the existing implementations of DL are focused on this area. Even though there are
several subtopics of this general problem, including stock price forecasting, index prediction,
forex price prediction, commodity (oil, gold, etc.) price prediction, bond price forecasting,
volatility forecasting, cryptocurrency price forecasting, the underlying dynamics are the
same in all of these applications.
16
 Studies can also be clustered into two main groups based on their expected outputs:
price prediction and price movement (trend) prediction. Although price forecasting is es-
sentially a regression problem, in most financial time series forecasting applications, correct
price prediction of the price is not perceived to be as important as correctly identifying the
directional movement. As a result, researchers consider trend prediction, i.e., forecasting
which way the price will change, a more crucial study area compared with exact price pre-
diction. In that sense, trend prediction becomes a classification problem. In some studies,
only up or down movements are taken into consideration (2-class problem), although 3-class
problems also exist (up, down, or neutral movements).
LSTM and its variations along with some hybrid models dominate the financial time
series forecasting domain. LSTM, by its nature, utilizes the temporal characteristics of any
time series signal; hence, forecasting financial time series is a well-studied and successful
implementation of LSTM. However, some researchers prefer to either extract appropriate
features from the time series or transform the time series such that the resulting financial
data become stationary from a temporal perspective, meaning even if we shuffle the data
order, we will still be able to properly train the model and achieve successful out-of-sample
test performance. For those implementations, CNN and Deep Feedforward Neural Network
(DFNN) are the most commonly chosen DL models.
Various financial time series forecasting implementations using DL models exist in lit-
erature. We will cover each of them in the following subsections. In this survey paper, we
examine the papers using the following criteria: First, we group articles according to their
subjects. Then, we group related papers according to their feature set. Finally, we group
each subgroup according to DL models/methods.
For each implementation area, the related papers are subgrouped and tabulated. Each
table contains the following fields to provide information about the implementation details
for the papers within the group: Article (Art.) and Data Set are trivial, Period refers to
the time period for training and testing. Feature Set lists the input features used in the
study. Lag is the time length of the input vector (e.g., 30d means the input vector has
a 30 day window), and horizon shows how far into the future the model predicts. Some
abbreviations are used for the two aforementioned fields: min is minutes, h is hours, d is
days, w is weeks, m is months, y is years, s is steps, and * is mixed. Method shows the
DL models that are used in the study. Performance criteria provides the evaluation metrics,
and Environment (Env.) lists the development framework/software/tools. Some column
values might be empty, indicating there was no relevant information in the paper for the
corresponding field.

4.1. Stock Price Forecasting

Price prediction of any given stock is the most studied financial application of all. We
observed the same trend within DL implementations. Depending on the prediction time
horizon, different input parameters are chosen, varying from High Frequency Trading (HFT)
and intraday price movements to daily, weekly, or even monthly stock close prices. Also,
technical, fundamental analysis, social media feeds, and sentiment are among the different
parameters used for the prediction models.
17
 Table 1: Stock Price Forecasting Using Only Raw Time Series Data

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[80] 38 stocks in KOSPI 2010-2014 Lagged stock re- 50min 5min DNN NMSE, RMSE, -
turns MAE, MI
[81] China stock market, 1990-2015 OCHLV 30d 3d LSTM Accuracy Theano,
3049 Stocks Keras
[82] Daily returns of 2001-2016 OCHLV - 1d LSTM RMSE, MAE Python,
‘BRD’ stock in Theano
Romanian Market
[83] 297 listed compa- 2012-2013 OCHLV 2d 1d LSTM, MAD, MAPE Keras
nies of CSE SRNN, GRU
[84] 5 stock in NSE 1997-2016 OCHLV, Price 200d 1..10d LSTM, RNN, MAPE -
data, turnover CNN, MLP
and number of
trades.
[85] Stocks of Infosys, 2014 Price data - - RNN, LSTM Accuracy -
TCS and CIPLA and CNN
from NSE
[86] 10 stocks in S&P500 1997-2016 OCHLV, Price 36m 1m RNN, LSTM, Accuracy, Keras,
data GRU Monthly return Tensorflow
[87] Stocks data from 2011-2016 OCHLV 1d 1d DBN MSE, norm- -
S&P500 RMSE, MAE
[88] High-frequency 2017 Price data - 1min DNN, ELM, RMSE, MAPE, Matlab
transaction data of RBF Accuracy
the CSI300 futures
[89] Stocks in the 1990-2015 Price data 240d 1d DNN, GBT, Mean return, H2O
S&P500 RF MDD, Calmar
ratio
[90] ACI Worldwide, 2006-2010 Daily closing 17d 1d RNN, ANN RMSE -
Staples, and Sea- prices
gate in NASDAQ
[91] Chinese Stocks 2007-2017 OCHLV 30d 1..5d CNN + Annualized Python
LSTM Return, Mxm
Retracement
[92] 20 stocks in S&P500 2010-2015 Price data - - AE + LSTM Weekly Returns -
[93] S&P500 1985-2006 Monthly and * 1d DBN+MLP Validation, Test Theano,
daily log-returns Error Python,
Matlab
[94] 12 stocks from SSE 2000-2017 OCHLV 60d 1..7d DWNN MSE Tensorflow
Composite Index
[95] 50 stocks from 2007-2016 Price data - 1d, 3d, SFM MSE -
NYSE 5d

In this survey, we grouped first stock price forecasting articles according to their feature
sets, such as studies using only the raw time series data (price data, Open, Close, High,
Low, Volume (OCHLV)) for price prediction; studies using various other data, and studies
using text mining techniques. Regarding the first group, the corresponding DL models were
directly implemented using raw time series for price prediction. Table 1 tabulates the stock
price forecasting studies that used only raw time series data in the literature. In Table 1,
different methods/models are also listed based on four sub-groups: DNN (networks that
are deep but without any given topology details) and LSTM models, multi models, hybrid
models, novel methods.
DNN and LSTM models were solely used in 3 papers. In Chong et al. [80], DNN and
lagged stock returns were used to predict the stock prices in The Korea Composite Stock
Price Index (KOSPI). Chen et al. [81], and Dezsi and Nistor [82] applied raw price data as
the input to LSTM models.
Meanwhile, some studies implement multiple DL models for performance comparison
18
using only raw price (OCHLV) data for forecasting. Among the noteworthy studies, Sama-
rawickrama et al. [83] compared RNN, Stacked Recurrent Neural Network (SRNN), LSTM,
and GRU. Hiransha et. al. [84] compared LSTM, RNN, CNN, and MLP, whereas in Selvin
et al. [85], RNN, LSTM, CNN, and Autoregressive Integrated Moving Average (ARIMA)
were preferred. Lee and Yoo [86] compared 3 RNN models (SRNN, LSTM, GRU) for stock
price prediction and then constructed a threshold-based portfolio selecting stocks according
to predictions. Li et. al. [87] implemented DBN. Finally, the authors of [88] compared 4
different ML models for next price prediction in 1-minute price data: a 1 DL model (AE and
RBM), MLP, Radial Basis Function Neural Network (RBF) and Extreme Learning Machine
(ELM). They also compared the results for different sized datasets. The authors of [89] used
price data and DNN, Gradient Boosted Trees (GBT), and Random Forest (RF) methods
for the prediction of stocks in the Standard’s & Poor’s 500 Index (S&P500). Chandra and
Chan [90] used co-operative neuro-evolution, RNN (Elman network), and DFNN for the pre-
diction of stock prices in National Association of Securities Dealers Automated Quotations
(NASDAQ) (ACI Worldwide, Staples, and Seagate).
Meanwhile, hybrid models were used in some papers. Liu et al. [91] applied CNN+LSTM.
Heaton et al. [92] implemented smart indexing with AE. Batres et al. [93] combined DBN
and MLP to construct a stock portfolio by predicting each stock’s monthly log-return and
choosing only stocks that were expected to perform better than the median stock.
In addition, novel approaches were adapted in some studies. Yuan et al. [94] proposed
the novel Deep and Wide Neural Network (DWNN), which is combination of RNN and
CNN. Zhang et al. [95] implemented a State Frequency Memory (SFM) recurrent network.
In another group of studies, some researchers again focused on LSTM-based models.
However, their input parameters came from various sources including raw price data, tech-
nical and/or fundamental analysis, macroeconomic data, financial statements, news, and
investor sentiment. Table 2 summarizes these stock price forecasting papers. In Table 2,
different methods/models are also listed based on five sub-groups: DNN model; LSTM and
RNN models; multiple and hybrid models; CNN model; and novel methods.
DNN models were used in some stock price forecasting papers within this group. In Abe
et al. [96], a DNN model and 25 fundamental features were used for prediction of Japan
Index constituents. Feng et al. [97] also used fundamental features and a DNN model for
prediction. A DNN model and macro economic data, such as GDP, unemployment rate, and
inventories, were used by the authors of [98] for the prediction of U.S. low-level disaggregated
macroeconomic time series.
LSTM and RNN models were chosen in some studies. Kraus and Feuerriegel [99] imple-
mented LSTM with transfer learning using text mining through financial news and stock
market data. Similarly, Minami et al. [100] used LSTM to predict stock’s next day price
using corporate action events and macro-economic index. Zhang and Tan [101] implemented
DeepStockRanker, an LSTM-based model for stock ranking using 11 technical indicators.
In Zhuge et al. [102], the authors used the price time series and emotional data from text
posts to predict the opening stock price of the next day with an LSTM network. Akita et
al. [103] used textual information and stock prices through Paragraph Vector + LSTM for
forecasting prices and the comparisons were provided with different classifiers. Ozbayoglu
19
[104] used technical indicators along with stock data on a Jordan-Elman network for price
prediction.
There were also multiple and hybrid models that used mostly technical analysis features
as their inputs to the DL model. Several technical indicators were fed into LSTM and MLP
networks in Khare et al. [105] for intraday price prediction. Recently, Zhou et al. [106]
used a GAN for minimizing Forecast error loss and Direction prediction loss (GAN-FD)
model for stock price prediction and compared their model performances against ARIMA,
ANN and Support Vector Machine (SVM). Singh et al. [107] used several technical indicator
features and time series data with Principal Component Analysis (PCA) for dimensionality
reduction cascaded with a DNN (2-layer FFNN) for stock price prediction. Karaoglu et al.
[108] used market microstructure-based trade indicators as inputs into an RNN with Graves
LSTM detecting the buy-sell pressure of movements in the Istanbul Stock Exchange Index
(BIST) to perform price prediction for intelligent stock trading. In Zhou et al. [109], next
month’s return was predicted, and next-to-be-performed portfolios were constructed. Good
monthly returns were achieved with LSTM and LSTM-MLP models.
Meanwhile, in some papers, CNN models were preferred. Abroyan et al. [110] used 250
features, including order details, for the prediction of a private brokerage company’s real data
of risky transactions. They used CNN and LSTM for stock price forecasting. The authors
of [111] used a CNN model and fundamental, technical, and market data for prediction.
Novel methods were also developed in some studies. In Tran et al. [112], with the FI-
2010 dataset, bid/ask and volume were used as the feature set for forecasting. In the study,
they proposed Weighted Multichannel Time-series Regression (WMTR), and Multilinear
Discriminant Analysis (MDA). Feng et al. [113] used 57 characteristic features, including
Market equity, Market Beta, Industry momentum, and Asset growth, as inputs to a Fama-
French n-factor DL for predicting monthly US equity returns in New York Stock Exchange
(NYSE), American Stock Exchange (AMEX), or NASDAQ.
Table 2: Stock Price Forecasting Using Various Data

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[96] Japan Index con- 1990-2016 25 Fundamental 10d 1d DNN Correlation, Ac- Tensorflow
stituents from Features curacy, MSE
WorldScope
[97] Return of S&P500 1926-2016 Fundamental - 1s DNN MSPE Tensorflow
Features:
2
[98] U.S. low-level disag- 1959-2008 GDP, Unemploy- - - DNN R -
gregated macroeco- ment rate, Inven-
nomic time series tories, etc.
[99] CDAX stock market 2010-2013 Financial news, 20d 1d LSTM MSE, RMSE, TensorFlow,
data stock market MAE, Accuracy, Theano,
data AUC Python,
Scikit-
Learn
[100] Stock of Tsugami 2013 Price data - - LSTM RMSE Keras,
Corporation Tensorflow
[101] Stocks in China’s A- 2006-2007 11 technical indi- - 1d LSTM AR, IR, IC -
share cators
[102] SCI prices 2008-2015 OCHL of change 7d - EmotionalAnalysis
MSE -
rate, price + LSTM

20
 Table 2: Stock Price Forecasting Using Various Data

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[103] 10 stocks in Nikkei 2001-2008 Textual informa- 10d - Paragraph Profit -
225 and news tion and Stock Vector +
prices LSTM
[104] TKC stock in NYSE 1999-2006 Technical indica- 50d 1d RNN Profit, MSE Java
and QQQQ ETF tors, Price (Jordan-
Elman)
[105] 10 Stocks in NYSE - Price data, Tech- 20min 1min LSTM, MLP RMSE -
nical indicators
[106] 42 stocks in China’s 2016 OCHLV, Techni- 242min 1min GAN RMSRE, DPA, -
SSE cal Indicators (LSTM, GAN-F, GAN-D
CNN)
[107] Google’s daily stock 2004-2015 OCHLV, Techni- 20d 1d (2D)2 PCA SMAPE, PCD, R, Matlab
data cal indicators + DNN MAPE, RMSE,
2
HR, TR, R
[108] GarantiBank in 2016 OCHLV, Volatil- - - PLR, Graves MSE, RMSE, Spark
BIST, Turkey ity, etc. LSTM MAE, RSE, R2
[109] Stocks in NYSE, 1993-2017 Price, 15 firm 80d 1d LSTM+MLP Monthly return, Python,Keras,
AMEX, NASDAQ, characteristics SR Tensorflow
TAQ intraday trade in AWS
[110] Private brokerage - 250 features: or- - - CNN, LSTM F1-Score Keras,
company’s real data der details, etc. Tensorflow
of risky transactions
[111] Fundamental and - Fundamental , - - CNN - -
Technical Data, technical and
Economic Data market informa-
tion
[112] The LOB of 5 stocks 2010 FI-2010 dataset: - * WMTR, Accuracy, Preci- -
of Finnish Stock bid/ask and vol- MDA sion, Recall, F1-
Market ume Score
[113] Returns in NYSE, 1975-2017 57 firm character- * - Fama-French R2 , RMSE Tensorflow
AMEX, NASDAQ istics n-factor
model DL

A number of research papers have also used text mining techniques for feature extraction
but used non-LSTM models for stock price prediction. Table 3 summarizes the stock price
forecasting papers that used text mining techniques. In Table 3, different methods/models
are clustered into three sub-groups: CNN and LSTM models; GRU, LSTM, and RNN
models; and novel methods.
CNN and LSTM models were adapted in some of the papers. In Ding et al. [114], events
were detected from Reuters and Bloomberg news through text mining, and that information
was used for price prediction and stock trading through the CNN model. Vargas et al. [115]
used text mining on S&P500 index news from Reuters through an LSTM+CNN hybrid
model for price prediction and intraday directional movement estimation together. Lee et
al. [116] used financial news data and implemented word embedding with Word2vec along
with MA and stochastic oscillator to create inputs for a Recurrent CNN (RCNN) for stock
price prediction. Iwasaki et al. [117] also used sentiment analyses through text mining and
word embeddings from analyst reports and used sentiment features as inputs to a DFNN
model for stock price prediction. Then, different portfolio selections were implemented based
on the projected stock returns.
GRU, LSTM, and RNN models were preferred in the next group of papers. Das et
al. [118] implemented sentiment analysis on Twitter posts along with stock data for price
forecasting using an RNN. Similarly, the authors of [119] used sentiment classification (neu-
21
tral, positive, and negative) for opening or closing stock price prediction with various LSTM
models. They compared their results with SVM and achieved higher overall performance. In
Zhongshengz et al. [120], text and price data were used for the prediction of SSE Composite
Index (SCI) prices.
Novel approaches were reported in some papers. Nascimento et al. [121] used word
embeddings for extracting information from web pages and then combined it with stock
price data for stock price prediction. They compared the Autoregressive (AR) model and
RF with and without news. The results showed embedding news information improved the
performance. Han et al. [122] used financial news and the ACE2005 Chinese corpus. Dif-
ferent event types of Chinese companies were classified based on a novel event-type pattern
classification algorithm in Han et al. [122], and also next day stock price change was also
predicted using additional inputs.
Table 3: Stock Price Forecasting Using Text Mining Techniques for Feature Extrac-
tion

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[114] S&P500 Index, 15 2006-2013 News from - - CNN Accuracy, MCC -
stocks in S&P500 Reuters and
Bloomberg
[115] S&P500 index news 2006-2013 Financial news 1d 1d RCNN Accuracy -
from Reuters titles, Technical
indicators
[116] TWSE index, 4 2001-2017 Technical indica- 15d - CNN + RMSE, Profit Keras,
stocks in TWSE tors, Price data, LSTM Python,
News TALIB
[117] Analyst reports on 2016-2018 Text - - LSTM, CNN, Accuracy, R- R,
the TSE and Osaka Bi-LSTM squared Python,
Exchange MeCab
[118] Stocks of Google, 2016-2017 Twitter senti- - - RNN - Spark,
Microsoft and Apple ment and stock Flume,
prices Twitter
API,
[119] Stocks of CSI300 2009-2014 Sentiment Posts, 1d 1d Naive Bayes Precision, Recall, Python,
index, OCHLV of Price data + LSTM F1-score, Accu- Keras
CSI300 index racy
[120] SCI prices 2013-2016 Text data and 7d 1d LSTM Accuracy, F1- Python,
Price data Measure Keras
[121] Stocks from S&P500 2006-2013 Text (news) and 7d 1d LAR+News, MAPE, RMSE -
Price data RF+News
[122] News from 2012-2016 A set of news text - - Their unique Precision, Recall, -
Sina.com, ACE2005 algorithm F1-score
Chinese corpus

4.2. Index Forecasting

Instead of trying to forecast the price of a single stock, several researchers preferred to
predict the stock market index. Indexes are generally are less volatile than individual stocks
because they are composed of multiple stocks from different sectors and are more indicative
of the overall momentum and general state of the economy.
In the literature, different stock market index data have been used for experiments. The
most commonly used index data are as follows: S&P500, China Securities Index (CSI)300,
National Stock Exchange of India (NIFTY), Tokyo Nikkei Index (NIKKEI)225, Dow Jones
Industrial Average (DJIA), Shanghai Stock Exchange (SSE)180, Hong Kong Hang Seng
22
Index (HSI), Shenzhen Stock Exchange Composite Index (SZSE), London Financial Times
Stock Exchange Index (FTSE)100, Taiwan Capitalization Weighted Stock Index (TAIEX),
BIST, NASDAQ, Dow Jones Industrial Average 30 (DOW30), KOSPI, S&P500 Volatility
Index (VIX), NASDAQ100 Volatility Index (VXN), Brazilian Stock Exchange (Bovespa),
Stockholm Stock Exchange (OMX), and NYSE. The authors of the papers [123, 124, 125,
126, 127, 128, 129, 130, 131, 132, 133, 134, 114] used S&P500 as their dataset. The authors
of the studies [123, 124, 135, 136, 137] used NIKKEI as their dataset. KOSPI was used in
Li et al. [135], Jeong et al. [131], Baek et al. [132]. DJIA was used as the dataset in the
papers [123, 136, 137, 138, 139]. The authors of the articles [123, 135, 137, 131] used HSI as
the dataset in their stuedies, and SZSE was used in the studies [140, 135, 141, 142].
In addition, in the literature, there are different methods for the prediction of index data.
While some studies used only raw time series data, others used various other data such, as
technical indicators, index data, social media feeds, news from Reuters, and Bloomberg,
and statistical features of data (standard deviation, skewness, kurtosis, omega ratio, fund
alpha). In this survey, we first grouped the index forecasting articles according to their
feature sets such as studies using only raw time series data (price/index data, OCHLV);
then, we clustered the studies using various other data. Table 4 summarizes the index
forecasting papers using only raw time series data. Moreover, different methods (models)
were used for index forecasting. MLP, RNN, LSTM, and DNN (DFNN or DMLP) methods
were the most used methods for index forecasting. In Table 4, these various methods/models
are also listed as four sub-groups: ANN, DNN, MLP, and Fuzzy Deep Direct Reinforcement
Learning (FDDR) models; RL and DL models; LSTM and RNN models; and novel methods.
Table 4: Index Forecasting Using Only Raw Time Series Data

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[124] S&P500, Nikkei225, 2011-2015 Index data - 1d, LRNFIS RMSE, MAPE, -
USD Exchanges 5d, with Firefly- MAE
7d, Harmony
10d Search
[125] S&P500 Index 1989-2005 Index data, Vol- 240d 1d LSTM Return, STD, Python,
ume SR, Accuracy Tensor-
Flow,
Keras, R,
H2O
[127] S&P500, VIX 2005-2016 Index data * 1d uWN, cWN MASE, HIT, -
RMSE
[128] S&P500 Index 2010-2017 Index data 10d 1d, Stacked MAE, RMSE, R- Python,
30d LSTM, Bi- squared Keras,
LSTM Tensorflow
[131] S&P500, KOSPI, 1987-2017 200-days stock 200d 1d Deep Q- Total profit, Cor- -
HSI, and Eu- price Learning and relation
roStoxx50 DNN
[132] S&P500, 2000-2017 Index data 20d 1d ModAugNet: MSE, MAPE, Keras
KOSPI200, 10- LSTM MAE
stocks
[133] S&P500, 2009-2017 Autoregressive - 1d LSTM MSE, Accuracy Tensorflow,
Bovespa50, OMX30 part of the time Keras, R
series
[134] S&P500 2000-2017 Index data - 1..4d, GLM, MAE, RMSE Python
1w, LSTM+RNN
1..3m

23
 Table 4: Index Forecasting Using Only Raw Time Series Data

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[136] Nikkei225, IXIC, 1985-2018 OCHLV 5d 1d LSTM RMSE Python,
HSI, GSPC, DJIA Keras,
Theano
[138] DJIA - Index data - - Genetic Deep MSE Java
Neural Net-
work
[139] Log returns of the 1971-2002 Index data 20d 1d RNN TR, sign rate, -
DJIA PT/HM test,
MSFE, SR, profit
[140] Shanghai A-shares 2006-2016 OCHLV 10d - Embedded Accuracy, MSE Python,
composite index, layer + Matlab,
SZSE LSTM Theano
[141] 300 stocks from 2014-2015 Index data - - FDDR, DNN Profit, return, Keras
SZSE, Commodity + RL SR, profit-loss
curves
[142] Shanghai composite 1990-2016 OCHLV 20d 1d Ensembles of Accuracy -
index and SZSE ANN
[143] TUNINDEX 2013-2017 Log returns of in- - 5min DNN with hi- Accuracy, MSE Java
dex data erarchical in-
put
[144] Singapore Stock 2010-2017 OCHL of last 10 10d 3d Feed-forward RMSE, MAPE, -
Market Index days of index DNN Profit, SR
[145] BIST 1990-2002 Index data 7d 1d MLP, RNN, HIT, posi- -
MoE tive/negative
HIT, MSE, MAE
[146] SCI 2012-2017 OCHLV, Index - 1..10d Wavelet + MAPE, theil un- -
data LSTM equal coefficient
[147] S&P500 1950-2016 Index data 15d 1d LSTM RMSE Keras
[148] ISE100 1987-2008 Index data - 2d, TAR-VEC- RMSE -
4d, MLP, TAR-
8d, VEC-RBF,
12d, TAR-VEC-
18d RHE
[149] VIX, VXN, VXD 2002-2014 First five autore- 5d 1d, HAR- MAE, RMSE -
gressive lags 22d GASVR

ANN, DNN, MLP, and FDDR models were used in some studies. In Lachiheb et al.
[143], log returns of the index data were used with a DNN with hierarchical input for the
prediction of TUNINDEX data. Yong et al. [144] used a deep FFNN and Open, Close,
High, Low (OCHL) of the last 10 days of index data for prediction. In addition, MLP and
ANN were used for the prediction of index data. In Yumlu et al. [145], raw index data were
used with MLP, RNN, Mixture of Experts (MoE), and Exponential GARCH (EGARCH)
for forecasting. In Yang et al. [142], ensembles of ANN with OCHLV of data were used for
prediction of the Shanghai composite index.
Furthermore, RL and DL methods were used together for prediction of index data in
some studies. In Deng et al.[141], FDDR, DNN, and RL methods were used to predict 300
stocks from SZSE index data and commodity prices. In Jeong et al. [131], Deep Q-Learning
and DNN methods and a 200-day stock price dataset were used together for prediction of
the S&P500 index.
Most of the preferred methods for prediction of index data using raw time series data have
been based on LSTM and RNN. In Bekiros et al. [139], an RNN was used for prediction
of log returns of the DJIA index. In Fischer et al. [125], LSTM was used to predict
24
S&P500 index data. Althelaya et al. [128] used stacked LSTM and Bidirectional LSTM (Bi-
LSTM) methods for S&P500 index forecasting. Yan et al. [146] used an LSTM network to
predict the next day closing price of Shanghai stock index. In their study, they used wavelet
decomposition to reconstruct the financial time series for denoising and better learning. In
Pang et al. [140], LSTM was used for prediction of the Shanghai A-shares composite index.
Namini et al. [136] used LSTM to predict NIKKEI225, IXIC, HIS, GSPC and DJIA index
data. In Takahashi et al. [147] and Baek et al. [132], LSTM was also used for the prediction
of the S&P500 and KOSPI200 indexes. Baek et al. [132] developed an LSTM-based stock
index forecasting model called ModAugNet. The proposed method was able to beat Buy
and Hold (B&H) in the long term with an overfitting prevention mechanism. Elliot et al.
[134] compared different ML models (linear models), Generalized Linear Models (GMLs)
and several LSTM, and RNN models for stock index price prediction. In Hansson et al.
[133], LSTM and the autoregressive part of time series index data were used for prediction
of the S&P500, Bovespa50, OMX30 indexes.
Also, some studies adapted novel appraches. In Zhang et al. [138], a genetic DNN
was used for DJIA index forecasting. Borovykh et al. [127] proposed a new DNN model
called Wavenet convolutional net for time series forecasting. Bildirici et al. [148] proposed a
Threshold Autoregressive (TAR)-Vector Error Correction model (VEC)-Recurrent Hybrid
Elman (RHE) model for forex and stock index of return prediction and compared several
models. Parida et al. [124] proposed a method called Locally Recurrent Neuro-fuzzy Infor-
mation System (LRNFIS) with Firefly Harmony Search Optimization (FHSO) Evolutionary
Algorithm (EA) to predict the S&P500 and NIKKEI225 and USD Exchange price data.
Psaradellis et al. [149] proposed Heterogeneous Autoregressive Process (HAR) with GA
with a SVR (GASVR) model called HAR-GASVR for prediction of the VIX, VXN, Dow
Jones Industrial Average Volatility Index (VXD) indexes.
In the literature, some studies used various input data, such as technical indicators, index
data, social media news, news from Reuters, and Bloomberg, and statistical features of data
(standard deviation, skewness, kurtosis, omega ratio, fund alpha). Table 5 summarizes
the index forecasting papers using these aforementioned various data. DNN, RNN, LSTM,
and CNN methods were the most commonly used models in index forecasting. In Table 5,
different methods/models are also listed within four sub-groups: DNN model; RNN and
LSTM models; CNN model; and novel methods.
A DNN was used as the classification model in some papers. In Chen et al. [150], a
DNN and some features of the data (Return, Sharpe-ratio (SR), Standard Deviation (STD),
Skewness, Kurtosis, Omega ratio, Fund alpha) were used for prediction. In Widegren et
al. [126], DNN, RNN, and technical indicators were used for prediction of the FTSE100,
OMX30, S&P500 indexes.
In addition, RNN and LSTM models with various other data were also used for prediction
of the indexes. Hsieh et al. [137] used RNN and OCHLV of indexes and technical indicators
to predict the DJIA, FTSE, Nikkei, and TAIEX indexes. Mourelatos et al. [151] used
GASVR, and LSTM for forecasting. Chen et al. [152] used four LSTM models (technical
analysis, attention mechanism and market vector embedding) for prediction of the daily
return ratio of the HSI300 index. In Li et al. [135], LSTM with wavelet denoising and index
25
data, volume, and technical indicators were used for prediction of the HSI, SSE, SZSE,
TAIEX, NIKKEI, and KOSPI indexes. Si et al. [153] used a MODRL+LSTM method
to predict Chinese stock-IF-IH-IC contract indexes. Bao et al. [123] used stacked AEs to
generate deep features using OCHL of stock prices, technical indicators, and macroeconomic
conditions to feed LSTM to predict future stock prices.
Table 5: Index Forecasting Using Various Data

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[114] S&P500 Index, 15 2006-2013 News from - - CNN Accuracy, MCC -
stocks in S&P500 Reuters and
Bloomberg
[116] TWSE index, 4 2001-2017 Technical indica- 15d - CNN + RMSE, Profit Keras,
stocks in TWSE tors, Index data, LSTM Python,
News TALIB
[123] CSI300, NIFTY50, 2010-2016 OCHLV, Techni- - 1d WT, Stacked MAPE, Correla- -
HSI, NIKKEI225, cal Indicators autoen- tion coefficient,
S&P500, DJIA coders, THEIL-U
LSTM
[126] FTSE100, OMXS 1993-2017 Technical indica- 60d 1d DNN, RNN Accuracy, p-value -
30, SP500, Com- tors
modity, Forex
[129] S&P500, DOW30, 2003-2016 Index data, Tech- - 1w, CNN Accuracy Tensorflow
NASDAQ100, Com- nical indicators 1m
modity, Forex,
Bitcoin
[130] BSE, S&P500 2004-2012 Index data, tech- 5d 1d..1m PSO, HM- RMSE, MAPE -
nical indicators RPSO, DE,
RCEFLANN
[135] HSI, SSE, SZSE, 2010-2016 Index data, vol- 2d..512d 1d LSTM with Accuracy, MAPE -
TAIEX, NIKKEI, ume, technical wavelet
KOSPI indicators denoising
[137] DJIA, FTSE, 1997-2008 OCHLV, Techni- 26d 1d RNN RMSE, MAE, C
NIKKEI, TAIEX cal indicators MAPE, THEIL-
U
[150] Hedge fund monthly 1996-2015 Return, SR, 12m 3m, DNN Sharpe ratio, -
return data STD, Skewness, 6m, Annual return,
Kurtosis, Omega 12m Cum. return
ratio, Fund alpha
[151] Stock of National 2009-2014 FTSE100, 1d, 1d GASVR, Return, volatil- Tensorflow
Bank of Greece DJIA, GDAX, 2d, LSTM ity, SR, Accuracy
(ETE). NIKKEI225, 5d,
EUR/USD, Gold 10d
[152] Daily return ratio of 2004-2018 OCHLV, Techni- - - Market MSE, MAE Python,
HS300 index cal indicators Vector + Tensorflow
Tech. ind.
+ LSTM +
Attention
[153] Chinese stock-IF- 2016-2017 Decisions for in- 240min 1min MODRL+LSTMProfit and loss, -
IH-IC contract dex change SR
[154] HS300 2015-2017 Social media 1d 1d RNN-Boost Accuracy, MAE, Python,
news, Index data with LDA MAPE, RMSE Scikit-
learn

Besides, different CNN implementations with various data (technical indicators, news,
and index data) have been used in the literature. In Dingli et al. [129], CNN, and index
data, and technical indicators were used for the S&P500, DOW30, NASDAQ100 indexes
and Commodity, Forex, and Bitcoin prices. In Ding et al. [114], a CNN model with news
from Reuters and Bloomberg were used for prediction of the S&P500 index and 15 stocks’
26
prices in S&P500. In Lee et al. [116], CNN + LSTM and technical indicators, index data,
and news were used for forecasting of the Taiwan Stock Exchange (TWSE) index and 4
stocks’ prices in TWSE.
In addition, some novel methods have been proposed for index forecasting. Rout et
al. [130] used RNN models, Recurrent Computationally Efficient Functional Link Neural
Network (RCEFLANN), and Functional Link Neural network (FLANN), with their weights
optimized using various EAs like Particle Swarm Optimization (PSO), and Modified Version
of PSO (HMRPSO), for time series forecasting. Chen et al. [154] used social media news to
predict index price and index direction with RNN-Boost with Latent Dirichlet Allocation
(LDA) features.

4.3. Commodity Price Forecasting

A number of studies particularly focused on the price prediction of any given commodity,
such as gold, silver, oil, and copper. With the increasing number of commodities available
for public trading through online stock exchanges, interest in this topic will likely grow in
the following years.
In the literature, there have been different methods used for commodity price forecasting.
DNN, RNN, FDDR, and CNN are the most used models to predict commodity prices.
Table 6 lists the details of the commodity price forecasting studies with DL.
In Dingli et al. [129], the authors used a CNN for predicting the next week’s and next
month’s price directional movement. Meanwhile, RNN and LSTM models were used in
some commodity forecasting studies. In Dixon et al. [155], a DNN was used for commodity
forecasting. In Widegren et al. [126], forex, and index datasets were used. DNN and RNN
were used to predict the prices of time series data. Technical indicators were used as the fea-
ture set consisting of Relative Strength Index (RSI), Williams Percent Range (William%R),
Commodity Channel Index (CCI), Percentage Price Oscillator (PPOSC), momentum, and
Exponential Moving Average (EMA). In Lasheras et al. [156], the authors used an El-
man RNN to predict COMEX copper spot price (through New York Mercantile Exchange
(NYMEX)) from daily closing prices.
Hybrid and novel models have been adapted in some studies. In Zhao et al. [157], FNN
and Stacked Denoising Autoencoders (SDAE) deep models were compared against Support
Vector Regressor (SVR), Random Walk (RW), and Markov Regime Switching (MRS) models
for WTI oil price forecasting. As performance criteria, accuracy, Mean Absolute Percentage
Error (MAPE), and Root Mean Square Error (RMSE) were used. In Chen et al. [158], the
authors aimed to predict WTI crude oil prices using several models, including combinations
of DBN, LSTM, Autoregressive Moving Average (ARMA), and RW. MSE was used as the
performance criteria. In Deng et al. [141], the authors used FDDR for stock price prediction
and trading signal generation. They combined DNN and RL. Profit, return, SR, and profit-
loss curves were used as the performance criteria.

27
 Table 6: Commodity Price Forecasting

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[129] S&P500, DOW30, 2003-2016 Price data, Tech- - 1w, CNN Accuracy Tensorflow
NASDAQ100, Com- nical indicators 1m
modity, Forex,
Bitcoin
[155] Commodity, FX fu- 1991-2014 Price Data 100*5min 5min DNN SR, capability ra- C++,
ture, ETF tio, return Python
[126] FTSE100, OMX30, 1993-2017 Technical indica- 60d 1d DNN, RNN Accuracy, p-value -
S&P500, Commod- tors
ity, Forex
[156] Copper prices from 2002-2014 Price data - - Elman RNN RMSE R
NYMEX
[157] WTI crude oil price 1986-2016 Price data 1m 1m SDAE, Boot- Accuracy, Matlab
strap aggre- MAPE, RMSE
gation
[158] WTI Crude Oil 2007-2017 Price data - - ARMA + MSE Python,
Prices DBN, RW + Keras,
LSTM Tensorflow
[141] 300 stocks from 2014-2015 Price data - - FDDR, DNN Profit, return, Keras
SZSE, Commodity + RL SR, profit-loss
curves

4.4. Volatility Forecasting

Volatility is directly related to price variations in a given time period and is mostly used
for risk assesment and asset pricing. Some researchers implemented models for accurately
forecasting the underlying volatility of any given asset.
In the literature, there have been different methods used for volatility forecasting, includ-
ing LSTM, RNN, CNN, MM, and Generalised Auto-Regressive Conditional Heteroscedastic-
ity (GARCH) models. Table 7 summarizes the studies that focused on volatility forecasting.
In Table 7, different methods/models are also represented as three sub-groups: CNN; RNN
and LSTM models; and hybrid and novel models.
A CNN model was used in one volatility forecasting study based on HFT data [159].
Meanwhile, RNN and LSTM models were used in some studies. In Tino et al. [160], the
authors used financial time series data to predict volatility changes with Markov Models and
Elman RNN for profitable straddle options trading. Xiong et al. [161] used price data and
different types of Google Domestic trends with LSTM. Zhou et al. [162] used CSI300, and
28 words from the daily search volume based on Baidu as the dataset with LSTM to predict
index volatility. Kim et al. [163] developed several LSTM models integrated with GARCH
for volatility prediction.
Hybrid and novel approaches have also been adapted in some studies. In Nikolaev et
al. [164], an RMDN with GARCH (RMDN-GARCH) model was proposed. In addition,
several models, including traditional forecasting models and DL models, were compared for
volatility estimation. Psaradellis et al. [149] proposed a novel method called HAR with
GASVR (HAR-GASVR) for volatility index forecasting.

28
 Table 7: Volatility Forecasting

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[159] London Stock Ex- 2007-2008 Limit order book - - CNN Accuracy, kappa Caffe
change state, trades,
buy/sell orders,
order deletions
[160] DAX, FTSE100, 1991-1998 Price data * * MM, RNN Ewa-measure, -
call/put options iv, daily profits’
mean and std
[161] S&P500 2004-2015 Price data, 25 - 1d LSTM MAPE, RMSE -
Google Domestic
trend dimensions
[162] CSI 300, 28 words of 2006-2017 Price data and 5d 5d LSTM MSE, MAPE Python,
the daily search vol- text Keras
ume based on Baidu
[163] KOSPI200, Korea 2001-2011 Price data 22d 1d LSTM + MAE, MSE, -
Treasury Bond GARCH HMAE, HMSE
interest rate, AA-
grade corporate
bond interest rate,
gold, crude oil
[164] DEM/GBP ex- - Returns - - RMDN- NMSE, NMAE, -
change rate GARCH HR, WHR
[149] VIX, VXN, VXD 2002-2014 First five autore- 5d 1d, HAR- MAE, RMSE -
gressive lags 22d GASVR

4.5. Bond Price Forecasting

Some financial experts follow the changes in bond prices to analyze the state of the
economy, claiming bond prices represent the health of the economy better than the stock
market [165]. Historically, long term rates are higher than short term rates under normal
economic expansion, whereas immediately before recessions, short term rates pass long term
rates, i.e., an inverted yield curve. Hence, accurate bond price prediction is very useful.
However, DL implementations for bond price prediction are very scarce. In Bianchi et al.
[166], excess bond return was predicted using several ML models, including RF, AE, and
PCA networks and a 2-3-4-layer DFNN. 4-layer NN outperformed the other models.

4.6. Forex Price Forecasting

Foreign exchange markets have the highest volumes among all existing financial mar-
kets in the world. They are open 24/7, and trillions of dollars worth of foreign exhange
transactions happen in a single day. According to the Bank for International Settlements,
foreign-exchange trading has a volume of more than 5 trillion USD a day [167]. In addition,
there are a large number of online forex trading platforms that provide leveraged transaction
opportunities to their subscribers. As a result, there is huge interest in profitable trading
strategies by traders. Hence, there are a number of forex forecasting and trading studies
based on DL models. Because most of the global financial transactions are based on the US
Dollar, almost all forex prediction research papers include USD in their analyses. However,
depending on regional differences and intended research focus, various models have been
developed accordingly.
In the literature, different methods have been used for forex price forecasting, includ-
ing RNN, LSTM, CNN, DBN, DNN, AE, and MLP methods. Table 8 provides details
29
about these implementations. In Table 8, different methods/models are listed as four sub-
groups: Continuous-valued Deep Belief Networks (CDBN), DBN, DBN+RBM, and AE
models; DNN, RNN, Psi-Sigma Network (PSN), and LSTM models; CNN models; and
hybrid models.
CDBN, DBN, DBN+RBM, and AE models have been used in some studies. In [168],
Fuzzy information granulation integrated with CDBN was applied for predicting EUR/USD
and GBU/USD exchange rates. They extended a DBN with a Continuous Restricted Boltz-
man machine (CRBM) to improve performance. In Chao et al. [169], weekly GBP/USD and
INR/USD prices were predicted, whereas in Zheng et al. [170], CNY/USD and INR/USD
were the main focus. In both cases, DBN was compared with FFNN. Similarly, Shen et al.
[171] implemented several different DBN networks to predict weekly GBP/USD, BRL/USD
and INR/USD exchange rate returns. Shen et al. [172] combined Stacked AE and SVR
for predicting 28 normalized currency pairs using the time series data of USD, GBP, EUR,
JPY, AUD, CAD, and CHF.
DNN, RNN, PSN, and LSTM models were preferred in some studies. In Dixon et al.
[155], multiple DMLP models were developed for predicting AD and BP futures using 5-
minute data over in a 130 day period. Sermpinis et al. [173] used MLP, RNN, GP, and other
ML techniques along with traditional regression methods for also predicting EUR/USD time
series. They also integrated Kalman filter, LASSO operator, and other models to further
improve the results [174]. They further extended their analyses by including PSN and pro-
viding comparisons along with traditional forecasters ARIMA, RW, and STAR [175]. To
improve performance, they also integrated hybrid time-varying volatility leverage. Sun et
al. [176] implemented RMB exchange rate forecasting against JPY, HKB, EUR and USD by
comparing RW, RNN, and FFNN performances. Maknickiene et al. [177] predicted various
forex time series and created portfolios consisting of these investments. Each network used
LSTM (RNN EVOLINO), and different risk appetites for users have been tested. Maknick-
iene et al. [178] also used EVOLINO RNN + orthogonal input data for predicting USD/JPY
and XAU/USD prices over different periods.
Different CNN models were used in some studies. In Persio et al. [179], EUR/USD
was once again forecasted using multiple DL models, including MLP, CNN, RNN, and
Wavelet+CNN. Korczak et al. [180] implemented forex trading (GBP/PLN) using several
different input parameters in a multi-agent-based trading environment. One of the agents
used AE+CNN as the prediction model and outperformed all other models.
Hybrid models have also been adapted in some of the researches. Bildirici et al. [148]
developed several (TAR-VEC-RHE) models for predicting monthly returns for TRY/USD
and compared model performances. Nikolaev et al. [164] compared several models, including
traditional forecasting models and DL models, for DEM/GBP prediction. Parida et al.
[124] predicted AUD, CHF, MAX, and BRL against USD currency time series data using
LRNFIS and compared it with different models. Meanwhile, instead of using LMS-based
error minimization during learning, they used FHSO.

30
 Table 8: Forex Price Forecasting

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[168] EUR/USD, 2009-2012 Price data * 1d CDBN-FG Profit -
GBP/USD
[169] GBP/USD, 1976-2003 Price data 10w 1w DBN RMSE, MAE, -
INR/USD MAPE, DA,
PCC
[170] CNY/USD,INR/USD 1997-2016 Price data - 1w DBN MAPE, R- -
squared
[171] GBP/USD, 1976-2003 Price data 10w 1w DBN + RBM RMSE, MAE, -
BRL/USD, MAPE, accuracy,
INR/USD PCC
[172] Combination of 2009-2016 Price data - - Stacked AE MAE, MSE, Matlab
USD, GBP, EUR, + SVR RMSE
JPY, AUD, CAD,
CHF
[155] Commodity, FX fu- 1991-2014 Price Data 100*5min 5min DNN SR, capability ra- C++,
ture, ETF tio, return Python
[126] FTSE100, OMX30, 1993-2017 Technical indica- 60d 1d DNN, RNN Accuracy, p-value -
S&P500, Commod- tors
ity, Forex
[173] EUR/USD 2001-2010 Close data 11d 1d RNN and MAE, MAPE, -
more RMSE, THEIL-U
[174] EUR/USD 2002-2010 Price data 13d 1d RNN, MLP, MAE, MAPE, -
PSN RMSE, THEIL-U
[175] EUR/USD, 1999-2012 Price data 12d 1d RNN, MLP, MAE, MAPE, -
EUR/GBP, PSN RMSE, THEIL-U
EUR/JPY,
EUR/CHF
[176] RMB against USD, 2006-2008 Price data 10d 1d RNN, ANN RMSE, MAE, -
EUR, JPY, HKD MSE
[177] EUR/USD, 2011-2012 Price data - - Evolino RNN Correlation be- -
EUR/JPY, tween predicted,
USD/JPY, real values
EUR/CHF,
XAU/USD,
XAG/USD, QM,
QG
[178] USD/JPY 2009-2010 Price data, Gold - 5d EVOLINO RMSE -
RNN +
orthogonal
input data
[179] S&P500, EUR/USD 1950-2016 Price data 30d, 1d, Wavelet+CNN Accuracy, log- Keras
30d*min 1min loss
[180] USD/GBP, 2016 Price data - 5min AE + CNN SR, % volatility, H2O
S&P500, FTSE100, avg return/trans,
oil, gold rate of return
[148] ISE100, TRY/USD 1987-2008 Price data - 2d, TAR-VEC- RMSE -
4d, MLP, TAR-
8d, VEC-RBF,
12d, TAR-VEC-
18d RHE
[164] DEM/GBP ex- - Returns - - RMDN- NMSE, NMAE, -
change rate GARCH HR, WHR
[124] S&P500, 2011-2015 Price data - 1d, LRNFIS with RMSE, MAPE, -
NIKKEI225, USD 5d, FHSO MAE
Exchanges 7d,
10d

4.7. Cryptocurrency Price Forecasting

Since cryptocurrencies have become a hot topic in the finance industry in recent years,
many studies and implementations have been conducted. Most cryptocurrency studies have
31
focused on price forecasting.
The rise of Bitcoin from 1000 USD in January 2017 to 20,000 USD in January 2018 has
attracted much attention, not only from the financial industry, but also from the general
public. Recently, papers have been published on price prediction and trading strategy de-
velopment for Bitcoin and other cryptocurrencies. Given the attention that the underlying
technology has attracted, there is a strong chance that new studies will appear in the near
future.
In the literature, DNN, LSTM, GRU, RNN, and classical methods (ARMA, ARIMA,
Autoregressive Conditional Heteroscedasticity (ARCH), GARCH, etc.) have been used for
cryptocurrency price forecasting. Table 9 summarizes the studies that utilized these meth-
ods. Lopes [181] combined the opinion market and price prediction for cryptocurrency trad-
ing. Text mining combined with 2 models, CNN and LSTM, were used to extract opinion.
Bitcoin, Litecoin, and StockTwits were used as the dataset. OCHLV of prices, technical in-
dicators, and sentiment analysis were used as the feature set. McNally et al. [182] compared
Bayesian optimized RNN, LSTM, and ARIMA to predict Bitcoin price direction. Sensitivity,
specificity, precision, accuracy, and RMSE were used as the performance metrics.
Table 9: Cryptocurrency Price Prediction

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[181] Bitcoin, Litecoin, 2015-2018 OCHLV, tech- - 30min, CNN, LSTM, MSE Keras,
StockTwits nical indicators, 4h, 1d State Fre- Tensorflow
sentiment analy- quency
sis Model
[182] Bitcoin 2013-2016 Price data 100d 30d Bayesian Sensitivity, speci- Keras,
optimized ficity, precision, Python,
RNN, LSTM accuracy, RMSE Hyperas

4.8. Trend Forecasting

Although trend forecasting and price forecasting share the same input characteristics,
some researchers prefer to predict the price direction of an asset instead of its actual price.
This alters the nature of the problem from regression to classification, and the corresponding
performance metrics also change. However, it is worth noting that these two approaches are
still fundamentally the same; the difference is in the interpretation of the output.
In the literature, there are different methods for trend forecasting. In this survey, we
grouped the articles according to their feature sets, such as studies using only raw time
series data (only price data, OCHLV); studies using technical indicators, price data, and
fundamental data at the same time; studies using text mining techniques; and studies using
various other data. Table 10 summarizes the trend forecasting studies using only raw time
series data.

32
 Table 10: Trend Forecasting Using Only Raw Time Series Data

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[183] S&P500 stock in- 1963-2016 Price data 30d 1d NN Accuracy, preci- R, H2o,
dexes sion, recall, F1- Python,
score, AUROC Tensorflow
[184] SPY ETF, 10 stocks 2014-2016 Price data 60min 30min FNN Cumulative gain MatConvNet,
from S&P500 Matlab
[142] Shanghai composite 1990-2016 OCHLV 20d 1d Ensembles of Accuracy -
index and SZSE ANN
[185] 10 stocks from - Price data TDNN, Missed oppor- -
S&P500 RNN, PNN tunities, false
alarms ratio
[186] GOOGL stock daily 2012-2016 Time window 22d, * LSTM, GRU, Accuracy, Python,
price data of 30 days of 50d, RNN Logloss Keras
OCHLV 70d
[133] S&P500, 2009-2017 Autoregressive 30d 1..15d LSTM MSE, Accuracy Tensorflow,
Bovespa50, OMX30 part of the price Keras, R
data
[187] HSI, DAX, S&P500 1991-2017 Price data - 1d GRU, GRU- Daily return % Python,
SVM Tensorflow
[188] Taiwan Stock Index 2001-2015 OCHLV 240d 1..2d CNN with Accuracy Matlab
Futures GAF, MAM,
Candlestick
[189] ETF and Dow30 1997-2007 Price data CNN with Annualized Keras,
feature imag- return Tensorflow
ing
[190] SSEC, NASDAQ, 2007-2016 Price data 20min 7min EMD2FNN MAE, RMSE, -
S&P500 MAPE
[191] 23 cap stocks from 2000-2017 Price data and 30d * DBN MAE Python,
the OMX30 index in returns Theano
Nasdaq Stockholm

Different methods and models have been used for trend forecasting. In Table 10, these
are divided into three sub-groups: ANN, DNN, and FFNN models; LSTM, RNN, and Prob-
abilistic NN models; and novel methods. ANN, DNN, DFNN, and FFNN methods were used
in some studies. In Das et al. [183], NN with price data was used for trend prediction of
the S&P500 stock indexes. Navon et al. [184] combined deep FNN with a selective trading
strategy unit to predict the next price. Yang et al. [142] created an ensemble network of
several Backpropagation and ADAM models for trend prediction.
In the literature, LSTM, RNN, and Probabilistic Neural Network (PNN) methods with
raw time series data have also been used for trend forecasting. Saad et al. [185] compared
Timedelay Neural Network (TDNN), RNN, and PNN for trend detection using 10 stocks
from S&P500. Persio et al. [186] compared 3 different RNN models (basic RNN, LSTM,
and GRU) to predict the movement of Google stock prices. Hansson et al. [133] used LSTM
(and other classical forecasting techniques) to predict the trend of stocks prices. In Shen et
al. [187], GRU and GRU-SVM models were used for the trends of the HSI, The Deutscher
Aktienindex (DAX), and S&P500 indexes.
There are also novel methods that use only raw time series price/index data in the
literature. Chen et al. [188] proposed a method that used a CNN with Gramian Angular
Field (GAF), Moving Average Mapping (MAM), and Candlestick with converted image
data. In Sezer et al. [189], a novel method of CNN with feature imaging was proposed
for prediction of the buy/sell/hold positions of the Exchange-Traded Funds (ETFs)’ prices
33
and Dow30 stocks’ prices. Zhou et al. [190] proposed a method that uses Empirical Mode
Decomposition and Factorization Machine based Neural Network (EMD2FNN) models to
forecast the directions of stock closing prices accurately. In Ausmees et al. [191], DBN with
price data was used for trend prediction of 23 large cap stocks from the OMX30 index.
Table 11: Trend Forecasting Using Technical Indicators & Price Data & Fundamen-
tal Data

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[192] KSE100 index - Price data, sev- - - ANN, SLP, Accuracy -
eral fundamental MLP, RBF,
data DBN, SVM
[193] Stocks in Dow30 1997-2017 RSI (Technical 200d 1d DMLP with Annualized Spark ML-
Indicators) genetic algo- return lib, Java
rithm
[194] SSE Composite In- 1999-2016 Technical indi- 24d 1d RBM Accuracy -
dex, FTSE100, Pin- cators, OCHLV
gAnBank price
[195] Dow30 stocks 2012-2016 Price data, sev- 40d - LSTM Accuracy Python,
eral technical in- Keras,
dicators Tensor-
flow,
TALIB
[196] Stock price from 2008-2015 Technical indica- - 15min LSTM Accuracy, Preci- Keras
IBOVESPA index tors, OCHLV of sion, Recall, F1-
price score, % return,
Maximum draw-
down
[197] 20 stocks from NAS- 2010-2017 Price data, tech- 5d 1d LSTM, GRU, Accuracy Keras,
DAQ and NYSE nical indicators SVM, XG- Tensor-
Boost flow,
Python
[198] 17 ETF 2000-2016 Price data, tech- 28d 1d CNN Accuracy, MSE, Keras,
nical indicators Profit, AUROC Tensorflow
[199] Stocks in Dow30 1997-2017 Price data, tech- 20d 1d CNN with Recall, precision, Python,
and 9 Top Volume nical indicators feature imag- F1-score, annual- Keras,
ETF ing ized return Tensor-
flow, Java
[200] Borsa Istanbul 100 2011-2015 75 technical in- - 1h CNN Accuracy Keras
Stocks dicators, OCHLV
of price

Some studies have used technical indicators, price data, and fundamental data at the
same time. Table 11 summarizes the trend forecasting papers that used technical indicators,
price data, and fundamental data. In addition, these studies are clustered into three sub-
groups: ANN, MLP, DBN, and RBM models; LSTM and GRU models; and novel methods.
ANN, MLP, DBN, and RBM methods were used with technical indicators, price data, and
fundamental data in some studies. In Raza et al. [192], several classical and ML models
and DBN were compared for trend forecasting. In Sezer et al. [193], technical analysis
indicator’s (RSI) buy and sell limits were optimized with GA, which was used for buy-sell
signals. After optimization, DMLP was also used for function approximation. Liang et
al. [194] used technical analysis parameters, OCHLV of prices, and RBM for stock trend
prediction.
LSTM and GRU methods with technical indicators, price data, and fundamental data
were also used in some papers. In Troiano et al. [195], the crossover and Moving Average
34
Convergence and Divergence (MACD) signals were used to predict the trend of Dow 30 stock
prices. Nelson et al. [196] used LSTM for stock price movement estimation. Song et al. [197]
used stock prices, technical analysis features, and four different ML models (LSTM, GRU,
SVM and eXtreme Gradient Boosting (XGBoost)) to predict the trend of stock prices.
In addition, novel methods using CNN with the price data and technical indicators have
been proposed. Gudelek et al. [198] converted the time series of price data to 2-dimensional
images using technical analysis and classified them with a deep CNN. Similarly, Sezer et al.
[199] also proposed a novel technique that converted financial time series data consisting
of technical analysis indicator outputs to 2-dimensional images and classified these images
using a CNN to determine the trading signals. Gunduz et al. [200] proposed a method using
a CNN with correlated features combined to predict the trend of stock prices.
Besides, there have also been studies using text mining techniques. Table 12 summarizes
the trend forecasting papers using text mining techniques. Different methods/models are
represented by four sub-groups: DNN, DMLP, and CNN with text mining models; GRU
model; LSTM, CNN, and LSTM+CNN models; and novel methods. In the first group of
studies, DNN, DMLP, and CNN with text mining were used for trend forecasting. In Huang
et al. [201], the authors used different models, including Hidden Markov Model (HMM),
DMLP, and CNN using Twitter moods, to predict the next day’s movement. Peng et al.
[202] used the combination of text mining and word embeddings to extract information from
financial news and a DNN model for prediction of stock trends.
Moreover, GRU methods with text mining techniques have also been used for trend
forecasting. Huynh et al. [203] used financial news from Reuters and Bloomberg, stock
price data, and a Bidirectional Gated Recurrent Unit (Bi-GRU) model to predict future
stock movements. Dang et al. [204] used Stock2Vec and Two-stream GRU (TGRU) models
to generate input data from financial news and stock prices. Then, they used the sign
difference between the previous close and next open for the classification of stock prices.
The results were better than those of state-of-the-art models.
LSTM, CNN, and LSTM+CNN models were also used for trend forecasting. Verma et al.
[205] combined news data with financial data to classify stock price movement and assessed
them with certain factors. They used an LSTM model as the NN architecture. Pinheiro
et al. [206] proposed a novel method that used a character-based neural language model
using financial news and LSTM for trend prediction. In Prosky et al. [207], sentiment/mood
prediction and price prediction based on sentiment, price prediction with text mining, and
DL models (LSTM, NN, CNN) were used for trend forecasting. Liu et al. [208] proposed
a method that used two separate LSTM networks to construct an ensemble network. One
of the LSTM models was used for word embeddings with word2Vec to create a matrix
information input to the CNN. The other was used for price prediction using technical
analysis features and stock prices.
In the literature, there are also novel methods to predict the trend of time series data.
Yoshihara et al. [209] proposed a novel method that uses a combination of RBM, DBN,
and word embedding to create word vectors for an RNN-RBM-DBN network to predict the
trend of stock prices. Shi et al. [210] proposed a novel method called DeepClue that visually
interpretted text-based DL models in predicting stock price movements. In their proposed
35
method, financial news, charts, and social media tweets were used together to predict stock
price movement. Zhang et al. [211] proposed a method that performed information fusion
from several news and social media sources to predict the trend of stocks. Hu et al. [212]
proposed a novel method that used text mining techniques and Hybrid Attention Networks
based on financial news for trend forecasting of stocks. Wang et al. [213] combined technical
analysis and sentiment analysis of social media (related financial topics) and created a Deep
Random Subspace Ensembles (DRSE) method for classification. Matsubara et al. [214]
proposed a method that used a Deep Neural Generative Model (DGM) with news articles
using a Paragraph Vector algorithm to create the input vector for prediction of stock trends.
Li et al. [215] implemented intraday stock price direction classification using financial news
and stock prices.
Table 12: Trend Forecasting Using Text Mining Techniques

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[201] S&P500, NYSE 2009-2011 Twitter moods, 7d 1d DNN, CNN Error rate Keras,
Composite, DJIA, index data Theano
NASDAQ Compos-
ite
[202] News from Reuters 2006-2013 News, price data 5d 1d DNN Accuracy -
and Bloomberg,
Historical stock
security data
[203] News from Reuters, 2006-2013 Financial news, - 1d, 2d, Bi-GRU Accuracy Python,
Bloomberg price data 5d, 7d Keras
[204] News about Apple, 2006-2013 Price data, news, - - Two-stream Accuracy, preci- Keras,
Airbus, Amazon technical indica- GRU, sion, AUROC Python
from Reuters, tors stock2vec
Bloomberg, S&P500
stock prices
[205] NIFTY50 In- 2013-2017 Index data, news 1d, 2d, 1d LSTM MCC, Accuracy -
dex, NIFTY 5d
Bank/Auto/IT/Energy
Index, News
[206] News from Reuters, 2006-2013 News and sen- - 1h, 1d LSTM Accuracy -
Bloomberg, stock tences
price/index data
from S&P500
[207] 30 DJIA stocks, 2002-2016 Price data and 1m 1d LSTM, NN, Accuracy VADER
S&P500, DJI, news features from CNN and
from Reuters news articles word2vec
[208] APPL from S&P500 2011-2017 News, OCHLV, - 1d CNN + Accuracy, F1- Tensorflow
and news from Technical indica- LSTM, score
Reuters tors CNN+SVM
[209] News, Nikkei Stock 1999-2008 News, MACD - 1d RNN, Accuracy, P- -
Average and 10- RBM+DBN value
Nikkei companies
[210] News from Reuters 2006-2015 Financial news, 1d 1d DeepClue Accuracy Dynet
and Bloomberg for price data software
S&P500 stocks
[211] Price data, index 2015 Price data, news 1d 1d Coupled ma- Accuracy, MCC Jieba
data, news, social from articles and trix and ten-
media data social media sor
[212] News and Chinese 2014-2017 Selected words in 10d 1d HAN Accuracy, An- -
stock data a news nual return
[213] Sina Weibo, Stock 2012-2015 Technical indica- - - DRSE F1-score, pre- Python
market records tors, sentences cision, recall,
accuracy, AU-
ROC

36
 Table 12: Trend Forecasting Using Text Mining Techniques

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[214] Nikkei225, S&P500, 2001-2013 Price data and 1d 1d DGM Accuracy, MCC, -
news from Reuters news %profit
and Bloomberg
[215] News, stock prices 2001 Price data and 60min (1..6)*5min
ELM, DLR, Accuracy Matlab
from Hong Kong TF-IDF from PCA, BELM,
Stock Exchange news KELM, NN

Moreover, studies have also used different data variations. Table 13 summarizes the
trend forecasting papers using these various data clustered into two sub-groups: LSTM,
RNN, and GRU models and CNN models.
LSTM, RNN, and GRU methods with various data representations have been used in
some trend forecasting papers. Tsantekidis et al. [216] used limit order book time series data
and an LSTM method for trend prediction. Sirignano et al. [217] proposed a novel method
that used limit order book flow and history information to determine stock movements using
LSTM. The results of the proposed method were remarkably stationary. Chen et al. [154]
used social media news, LDA features, and an RNN model to predict the trend of index
prices. Buczkowski et al. [218] proposed a novel method that used expert recommendations
(Buy, Hold, or Sell), emsemble of GRU, and LSTM to predict the trend of stock prices.
CNN models with different data representations were also used for trend prediction.
Tsantekidis et al. [219] used the last 100 entries from the limit order book to create images
for stock price prediction using a CNN. Using the limit order book data to create a 2D
matrix-like format with a CNN for predicting directional movement was innovative. In
Doering et al. [159], HFT microstructure forecasting was implemented with a CNN.
Table 13: Trend Forecasting Using Various Data

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[216] Nasdaq Nordic 2010 Price and volume 100s 10s, LSTM Precision, Re- -
(Kesko Oyj, data in LOB 20s, call, F1-score,
Outokumpu 50s Cohen’s k
Oyj, Sampo,
Rautaruukki, Wart-
sila Oyj)
[217] High-frequency 2014-2017 Price data, 2h - LSTM Accuracy -
record of all orders record of all
orders, transac-
tions
[154] Chinese, The 2015-2017 Social media 1d 1d RNN-Boost Accuracy, MAE, Python,
Shanghai-Shenzhen news (Sina with LDA MAPE, RMSE Scikit
300 Stock Index Weibo), price learn
(HS300 data
[218] ISMIS 2017 Data - Expert identifier, - - LSTM + Accuracy -
Mining Competition class predicted by GRU +
dataset expert FCNN
[219] Nasdaq Nordic 2010 Price, Volume - - CNN Precision, Re- Theano,
(Kesko Oyj, data, 10 orders call, F1-score, Scikit
Outokumpu of the LOB Cohen’s k learn,
Oyj, Sampo, Python
Rautaruukki, Wart-
sila Oyj)

37
 Table 13: Trend Forecasting Using Various Data

Art. Data Set Period Feature Set Lag Horizon Method Performance Env.
Criteria
[159] London Stock Ex- 2007-2008 Limit order book - - CNN Accuracy, kappa Caffe
change state, trades,
buy/sell orders,
order deletions

5. Current Snaphot of The Field

The histogram of Publication Count in Topics

All-years count
40 Last-3-years count
Publication Count

30

20

10

0
sto

tr e

in

fo

co

vo

cr
de

y
re

la
nd
ck

pt
x
x

til

oc
fo
pr

pr
fo

od

ity

ur
re
ic

ic
re

ity

re
fo
ef

ca

ef
ca

nc
re
sti

pr
or

or
sti

ca

y
ic
n
ec

ec
n
g

pr
ef

sti
g
as

as

ic
n
or
tin

tin

ep
ec
g

re
as

Topic Name
di
tin

ct
g

io
n

Figure 5: The histogram of Publication Count in Topics

After reviewing all research papers specifically targeted at financial time series forecasting
implementations using DL models, we are now ready to provide some overall statistics about
the current state of the field. The number of papers included in our survey was 140. We
categorized the papers according to their forecasted asset type. We also analyzed the studies
based on their DL model choices, frameworks for the development environment, data sets,
comparable benchmarks, and some other differentiating criteria such as feature sets and
numbers of citations, which could not be included in this paper due to space constraints. We
will now summarize our notable observations to provide interested research with important
highlights within the field.
Figure 5 presents the various asset types that researchers developed their corresponding
forecasting models for. As expected, stock market-related prediction studies dominate the
field. Stock price forecasting, trend forecasting, and index forecasting were the top three
picks for financial time series forecasting research. So far, 46 papers have been published for
stock price forecasting, 38 for trend forecasting and 33 for index forecasting. These studies
38
 The Rate of Publication Count in Topics

stock price forecasting

index forecasting
Publication Count

15 commodity price forecasting

volatility forecasting
forex price forecasting
10 cryptocurrency price prediction
trend forecasting
forex industry

0
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
Year

Figure 6: The rate of Publication Count in Topics

constitute more than 70% of all studies, indicating high interest. Besides the above, there
were 19 papers on forex prediction and 7 on volatility forecasting. Meanwhile cryptocurrency
forecasting has started attracting researchers; however, only 3 papers on this topic have been
published yet, but this number is expected to increase in the coming years [220]. Figure 6
highlights the rate of publication counts for various implementation areas throughout the
years. Meanwhile, Figure 7 provides more details about the choice of DL models over various
implementation areas.
Figure 8 illustrates the increasing appetite of researchers to develop DL models for fi-
nancial time series implementations. Meanwhile, as Figure 9 indicates, most studies were
published in journals (57 of them) and conferences (49 papers), but a considerable number
of arXiv papers (11) and graduate theses (6) also exist.
One of the most important issues for a researcher is where they can publish their findings.
During our review, we also carefully investigated where each paper was published. We tabu-
lated our results for the top journals for financial time series forecasting in Fig 10. According
to these results, the journals with the most published papers include Expert Systems with
Applications, Neurocomputing, Applied Soft Computing, The Journal of Supercomputing,
Decision Support Systems, Knowledge-based Systems, European Journal of Operational Re-
search, and IEEE Access. The interested researchers should also consider the trends over
the last 3 years, as tendencies can vary depending on the particular implementation areas.
Carefully analyzing Figure 11 clearly validates the dominance of RNN-based models (65
papers) among all others for DL model choices, followed by DMLP (23 papers) and CNN
(20 papers). The inner-circle represents all years considered, while the outer circle provides
only the studies within the last 3 years. We should note that the RNN is a general model

39
 30
RNN 30 19 3 4 6 2 18 5
CNN 11 3 1 1 2 1 8 2 25
DMLP 13 9 2 0 4 1 6 3 20
DBN 2 0 0 0 4 0 3 4
15
AE 2 1 1 0 2 0 0 2
RL 0 2 1 0 0 0 0 0 10
RBM 1 0 0 0 1 0 2 1 5
Other 5 3 0 2 3 0 8 2
0
index forecasting

volatility forecasting

forex price forecasting

trend forecasting
stock price forecasting

cryptocurrency price prediction

forex industry
commodity price forecasting

Figure 7: Topic-Model Heatmap

with several versions, including LSTM and GRU. For RNN, researchers mostly prefer LSTM
due to its relatively simple model development phase; however, other types of RNN are also
common. Figure 12 provides a snapshot of the RNN model distribution. As mentioned
above, LSTM had the highest interest among all with 58 papers, while Vanilla RNN and
GRU had 27 and 10 papers, respectively. Hence, it is clear that LSTM is the most popular
DL model for financial time series forecasting and regression studies.
Meanwhile, DMLP and CNN were generally preferred for classification problems. Be-
cause time series data generally consist of temporal components, some data preprocessing
might be required before actual classification can occur. Hence, many of these implementa-
tions utilize feature extraction, selection techniques, and possible dimensionality reduction
methods. Many researchers mainly use DMLP due to the fact that its shallow MLP version
has been used extensively before and has a proven successful track record for many differ-
ent financial applications, including financial time series forecasting. Consistent with our
observations, DMLP was also mostly preferred in the stock, index, and particular trend fore-
casting because it is by definition, a classification problem with two (uptrend or downtrend)

40
 Histogram of Publication Count in Years

50
Publication Count in Year

40

30

20

10

0
19

19

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20
98

99

00

01

02

03

04

05

06

07

08

09

10

11

12

13

14

15

16

17

18
Year

Figure 8: The histogram of Publication Count in Years

The histogram of Publication Count in Publication Type

60
All-years count
Publication Count

Last-3-years count

40

20

0
Jo

Pr

Th

Bo
rx
u

oc

isc
es

ok
rn

iv
ee

is
al

Ch
A
di
A

r ti
ng

ap
r ti

cl
sA

te
cl

r
e

r ti
cl
e

Publication Type

Figure 9: The histogram of Publication Count in Publication Types

and three (uptrend, stationary, or downtrend) class instances.

In addition to DMLP, CNN is also a popular choice for classification-type financial time
series forecasting implementations. Most of these studies appeared within the last 3 years.
As mentioned before, to convert time-varying sequential data into a more stationary classi-
fiable form, some preprocessing might be necessary. Even though some 1-D representations

41
 The histogram of Top Journals

Expert Systems with Applications 4.292

Neurocomputing 4.072
Applied Soft Computing 4.873
Decision Support Systems 3.847
European Journal of Operational Research 3.806
IEEE Access 4.098
IEEE Transactions on Neural Networks 2.633
Knowledge-based Systems 5.101
Neural Computing and Applications 4.664
The Journal of Supercomputing 2.16
Algorithmic Finance -
Algorithms -
Concurrency and Computation Practice and Experience 1.167
Data & Knowledge Engineering 1.583
Electronic Commerce Research and Applications 2.911
Energy Economics 4.151
Engineering Letters -
Frontiers in Signal Processing -
IEEE Transactions on Industrial Informatics 7.377
Journal Name

IEEE Transactions on Knowledge and Data Engineering 3.857

IEEE Transactions on Neural Networks and Learning Systems 11.683
IEICE Transactions on Information and Systems -
Innovative Infotechnologies for Science Business and Education -
International Journal of Economics and Management Systems -
International Journal of Forecasting 3.386
International Journal of Intelligent Systems and Applications in Engineering -
International Journal of Machine Learning and Computing -
Journal of Business Economics and Management -
Journal of Computational Finance 0.758
Journal of Econometrics 1.949
Journal of International Financial Markets Institutions & Money 1.836
Journal of King Saud University – Computer and Information Sciences -
Journal of Mathematical Finance 0.39
Mathematical Problems in Engineering 1.179
Multimedia Tools and Applications 2.101
Pattern Recognition Letters 2.810
Plos One 2.776
Resources Policy 3.185
Review of Financial Economics -
Romanian Economic Business Review -
SSRN - Social Science Research Network -
Systems Engineering - Theory & Practice -
The Journal of Finance and Data Science -
Wireless Personal Communications 0.929
International Journal of Mathematics and Computers in Simulation 1.409
0 1 2 3 4 5

Journal Count
Last-3-years count Other-years count

Figure 10: Top Journals - corresponding numbers next to the bar graph are representing the impact factor
of the journals

exist, the 2-D implementation for CNN is more common, mostly inherited through image
recognition applications of CNN from computer vision implementations. In some stud-
ies [188, 189, 193, 199, 219], innovative transformations of financial time series data into
an image-like representation have been adapted, and impressive performances have been
achieved. As a result, CNN might increase its share of interest for financial time series

42
 Publication Count in Model Type

RNN
18.6%
DMLP
CNN
Other
20.8%
DBN
AE

Model Type 51%

52.5%
16.9% 14.1%

8.05%

6.04%
6.78%

5.08%

Figure 11: The Piechart of Publication Count in Model Types

Distribution of RNN Models

LSTM
Vanilla RNN
GRU

29.7%

RNN
60.4%

9.89%

Figure 12: Distribution of RNN Models

forecasting in the next few years.

As one final note, Figure 13 shows which frameworks and platforms the researchers and
developers used while implementing their work. We tried our best to extract this information
from the papers. However, we must keep in mind that not every publication provided their
development environment. Also, most papers did not give details, preventing us from a

43
 Frameworks

keras
python
other
22.2% 23.7% tensorflow
matlab
theano
r
h2o
Frameworks 2.22%
java
2.96%
17.8%
3.7%

5.19%

16.3% 5.93%

Figure 13: The Preferred Development Environments

more thorough comparison chart, i.e, some researchers claimed they used Python, but no
further information was given, while some others mentioned the use of Keras or TensorFlow,
providing more details. Also, within the “Other" section, the usage of Pytorch has increased
in the last year or so, even though it is not visible from the chart. Regardless, Python-
related tools were the most influential technologies behind the implementations covered in
this survey.

6. Discussion and Open Issues

From an application perspective, even though financial time series forecasting has a rela-
tively narrow focus, i.e., the implementations were mainly based on price or trend prediction,
depending on the underlying DL model, very different and versatile models exist in the lit-
erature. We must remember that even though financial time series forecasting is a subset of
time-series studies, due to the embedded profit-making expectations from successful predic-
tion models, some differences exist, such that higher prediction accuracy sometimes might
not reflect a profitable model. Hence, the risk and reward structure must also be taken
into consideration. At this point, we will try to elaborate on our observations about these
differences in various model designs and implementations.

6.1. DL Models for Financial Time Series Forecasting

According to the publication statistics, LSTM was the preferred choice of most re-
searchers for financial time series forecasting. LSTM and its variations utilized time-varying
data with feedback embedded representations, resulting in higher performances for time se-
ries prediction implementations. Because most financial data, one way or another, included
44
time-dependent components, LSTM was the natural choice in financial time series forecast-
ing problems. Meanwhile, LSTM is a special DL model derived from a more general classifier
family, namely RNN.
Careful analysis of Figure 11 illustrates the dominance of RNNs (which mainly consist of
LSTM). As a matter of fact, more than half of the published papers on time series forecasting
fall into the RNN model category. Regardless of its problem type, price, or trend prediction,
the ordinal nature of the data representation forced researchers to consider RNN, GRU, and
LSTM as viable preferences for their model choices. Hence, RNN models were chosen, at
least for benchmarking, in many studies for performance comparison with other developed
models.
Meanwhile, other models were also used for time series forecasting problems. Among
those, DMLP had the most interest due to the market dominance of its shallow cousin
(MLP) and its wide acceptance and long history within ML society. However, there is a
fundamental difference in how DMLP- and RNN-based models were used for financial time
series prediction problems.
DMLP fits well for both regression and classification problems. However, in general,
data order independence must be preserved to better utilize the internal working dynamics
of such networks, even though some adjustments can be made through the learning algorithm
configuration. In most cases, either trend components of the data need to be removed from
the underlying time series or some data transformations might be needed so that the resulting
data becomes stationary. Regardless, some careful preprocessing might be necessary for a
DMLP model to be successful. In contrast, RNN-based models can work directly with
time-varying data, making it easier for researchers to develop DL models.
As a result, most DMLP implementations had embedded data preprocessing before the
learning stage. However, this inconvenience did not prevent researchers from using DMLP
and its variations during their model development process. Instead, many versatile data
representations were attempted to achieve higher overall prediction performances. A combi-
nation of fundamental and/or technical analysis parameters along with other features, such
as financial sentiment through text mining, were embedded in such models. In most DMLP
studies, the corresponding problem was treated as classification, especially in trend predic-
tion models, whereas RNN-based models directly predicted the next value of the time series.
Both approaches had some success in outperforming the underlying benchmark; hence it is
not possible to claim superiority of one model type over another. However, as a general rule
of thumb, researchers prefer RNN-based models for time series regression and DMLP for
trend classification (or buy-sell point identification).
Another model that has increased in popularity recently is CNN. CNN also works bet-
ter for classification problems, and unlike RNN-based models, it is more suitable for either
non-time varying or static data representations. The comments made about DMLP are also
mostly valid for CNN. Furthermore, unlike DMLP, CNN mostly requires locality within the
data representation for better-performing classification results. One particular implementa-
tion area of CNN is image-based object recognition problems. In recent years, CNN-based
models have dominated this field, handily outperforming all other models. Meanwhile, most
financial data are time-varying, and it might not be easy to implement CNN directly for fi-
45
nancial applications. However, in some recent studies, various independent research groups
followed an innovative transformation of 1-D time-varying financial data into 2-D mostly
stationary image-like data so that they could utilize the power of CNN through adaptive
filtering and implicit dimensionality reduction. Hence, with that approach, they were able
to develop successful models.
There is also a rising trend of using deep RL-based financial algorithmic trading imple-
mentations; these are mostly associated with various agent-based models, where different
agents interact and learn from their interactions. This field has even more opportunities
to offer with advancements in financial sentiment analysis through text mining to capture
investor psychology; as a result, behavioral finance can benefit from these particular studies
associated with RL-based learning models coupled with agent-based studies.
Other models including DBN, AE, and RBM, were also used by several researchers, and
superior performances were reported in some of their work. However, interested readers
should check these studies case by case to see how they were modelled from data represen-
tation and learning perspectives.

6.2. Discussions on Selected Features

Regardless of the underlying forecasting problem, the raw time series data was almost
always somehow embedded directly or indirectly within the feature vector, which is particu-
larly valid for RNN-based models. However, in most other model types, other features were
also included. Fundamental analysis and technical analysis features were among the most
favorable choices for stock/index forecasting studies.
Meanwhile, in recent years, financial text mining has gained particular attention, mostly
for extracting investor/trader sentiment. The streaming flow of financial news, tweets, state-
ments, and blogs allowed researchers to build better and more versatile prediction and eval-
uation models, integrating numerical and textual data. The general methodology involves
extracting financial sentiment analysis through text mining and combining that information
with fundamental/technical analysis data to achieve better overall performance. It is logical
to assume that this trend will continue with the integration of more advanced text and NLP
techniques.

6.3. Discussions on Forecasted Asset Types

Although forex price forecasting is always popular among researchers and practitioners,
stock/index forecasting has always had the most interest among all asset groups. Regardless,
price/trend prediction and algo-trading models were mostly embedded with these prediction
studies.
These days, financial time series forecasting research on cryptocurrencies is a hot topic.
Cryptocurrency price prediction has growing demand from the financial community. Because
this topic is relatively new, we might see more studies and implementations in the near future
due to high expectations and promising rewards.
There were also a number of publications in commodity price forecasting research, par-
ticularly for predicting the price of oil. Oil price prediction is crucial due to its tremendous

46
effect on world economic activities and planning. Meanwhile, gold is considered a safe in-
vestment and almost every investor, at one time, considers allocating some portion of their
portfolio for gold-related investments. In times of political uncertaintly, many people turn
to gold to protect their savings. Although we did not encounter a noteworthy study for gold
price forecasting, due to its historical importance, there might be opportunities in this area
in years to come.

6.4. Open Issues and Future Work

Despite the general motivation for financial time series forecasting remaining fairly un-
changed, the means of achieving financial goals vary depending on the choices and trade-off
between traditional techniques and newly developed models. Because our fundamental focus
is on the application of DL for financial time series studies, we will try to assess the current
state of the research and extrapolate that into the future.

6.4.1. Model Choices for the Future

The dominance of RNN-based models for price/trend prediction will probably not disap-
pear anytime soon, mainly due to their easy adaptation to most asset forecasting problems.
Meanwhile, some enhanced versions of the original LSTM or RNN models, generally in-
tegrated with hybrid learning systems, are now becoming more common. Readers should
check individual studies and assess their performances to see which one fits the best for their
particular needs and domain requirements.
We have observed increasing interest in 2-D CNN implementations of financial forecast-
ing problems by converting the time series into an image-like data type. This innovative
methodology seems to work quite satisfactorily and provides promising opportunities. More
studies of this kind will probably continue in the near future.
Nowadays, new models are generated through older models via modifying or enhancing
existing models so that better performances can be achieved. Such topologies include Gen-
erative Adversarial Network (GAN), and Capsule networks. They have been used in various
non-financial studies; however, financial time series forecasting has not been investigated
for those models yet. As such, there can be exciting opportunities both from research and
practical points of view.
Another DL model that has been investigated thoroughly is Graph CNN. Graphs can be
used to represent portfolios, social networks of financial communities, fundamental analysis
data, etc. Even though graph algorithms can directly be applied to such configurations, dif-
ferent graph representations can also be implemented for time series forecasting problems.
Not much work has been done on this particular topic; however, through graph representa-
tions of time series data and implementing graph analysis algorithms or implementing CNN
through these graphs are among the possibilities that researchers can choose.
As a final note for future models, we believe that deep RL- and agent-based models
offer great opportunities for researchers. HFT algorithms and robo-advisory systems highly
depend on automated algorithmic trading systems that can decide what to buy and when
to buy without any human intervention. These aforementioned models can fit very well in
such challenging environments. The rise of the machines will also lead to a technological
47
(and algorithmic) arms race between Fintech companies and quant funds to be the best in
their neverending search for “achieving alpha". New research in these areas can be exactly
what is required.

6.4.2. Future Projections for Financial Time Series Forecasting

Most probably, for the foreseeable future, financial time series forecasting will have close
research cooperation with other financial application areas, such as algorithmic trading and
portfolio management, as was the case before. However, changes in the available data
characteristics and introduction of new asset classes might not only alter the forecasting
strategies of developers but also force developers to seek new or alternative techniques to
better adapt to these new challenging working conditions. In addition, metrics such as
Continuous Ranked Probability Score (CRPS) [221] for evaluating probability distributions
might be included for more thorough analysis.
One rising trend, not only for financial time series forecasting, but for all intelligent
decision support systems, is human-computer interaction and NLP research. Within that
field, text mining and financial sentiment analysis are of particular importance to financial
time series forecasting. Behavioral finance may benefit from new advancements in these
fields.
To utilize the power of text mining, researchers have started developing new data repre-
sentations, such as Stock2Vec [204], which can be useful for combining textual and numerical
data for better prediction models. Furthermore, NLP-based ensemble models that integrate
data semantics with time-series data might increase the accuracy of existing models.
One area that can significantly benefit from interconnected financial markets is auto-
mated statistical arbitrage trading model development. It has been used in forex and com-
modity markets before. In addition, many practitioners currently seek arbitrage opportuni-
ties in cryptocurrency markets [220] due to the huge number of coins available on various
marketplaces. Price disruptions, high volatility, and bid-ask spread variations cause arbi-
trage opportunities across different platforms. Some opportunists develop software models
that can track these price anomalies for instant materialization of profits. Also, it is pos-
sible to construct pairs of trading portfolios across different asset classes using appropriate
models. It is possible that DL models can learn (or predict) these opportunities faster and
more efficiently than classical rule-based systems. This will also benefit HFT studies, which
are constantly looking for faster and more efficient trading algorithms and embedded sys-
tems with minimum latency. To achieve that, Graphics Processing Unit (GPU)- or Field
Programmable Gate Array (FPGA)-based hardware solutions embedded with DL models
can be utilized. There is a lack of research accomplished in this hardware aspect of financial
time series forecasting and algorithmic trading. Given that there is sufficient computing
power available, it is worth investigating the possibility of better algorithms, because the
rewards are high.

6.5. Responses to our Initial Research Questions

We are now ready to address to our initially stated research questions. Based on our
observations, our answers to these questions are as follows:
48
 • Which DL models are used for financial time series forecasting ?
Response: RNN-based models (particularly LSTM) are the most commonly used mod-
els. Meanwhile, CNN and DMLP have been used extensively in classification-type
implementations (such as trend classification) as long as appropriate data processing
is applied to the raw data.

• How does the performance of DL models compare with that of their traditional machine
learning counterparts ?
Response: In the majority of studies, DL models were better than ML ones. However,
there were also many cases where their performances were comparable. There were
even two particular studies ([82, 175]) where ML models performed better than DL
models. Meanwhile, the preference of DL implementations over ML models is growing.
Advances in computing power, availability of big data, superior performance, implicit
feature learning capabilities, and user friendly model development environment for DL
models are among the main reasons for this migration.
One important issue that might be worth mentioning is the possibility of the publi-
cation bias of DL over ML models. Since DL is more recent than ML, a published
successful DL implementation might attract more audience than a comparable success-
ful ML model. Hence, the researchers implicitly might have an additional motivation
to develop DL models. However, this is probably a valid concern for every academic
publication regardless of the study area [222]. Meanwhile, in this survey, our aim was
to extract the published DL studies for financial forecasting without any prior assump-
tions, so the reader can decide which model works best for them through their own
experiences.

• What is the future direction of DL research for financial time series forecasting ?
Response: NLP, semantics, and text mining-based hybrid models ensembled with
time-series data might be more common in the near future.

7. Conclusions
Financial time series forecasting has been very popular among ML researchers for more
than 40 years. The financial community had a new boost lately with the introduction of DL
implementations for financial prediction research, and many new publications have appeared
accordingly. In our survey, we wanted to review existing studies to provide a snapshot of
the current research status of DL implementations for financial time series forecasting. We
grouped the studies according to their intended asset classes along with the preferred DL
model associated with the problem. Our findings indicate that although financial forecasting
has a long research history, overall interest within the DL community is on the rise through
utilization of new DL models; hence, many opportunities exist for researchers.

49
8. Acknowledgement
This work is supported by Scientific and Technological Research Council of Turkey
(TUBITAK) grant no 215E248.

Glossary
AdaGrad Adaptive Gradient Algorithm. 6, 7 CRPS Continuous Ranked Probability Score. 48
ADAM Adaptive Moment Estimation. 6–8, 13, CSE Colombo Stock Exchange. 18
14, 33 CSI China Securities Index. 18, 22, 26, 28, 29
AE Autoencoder. 5, 9, 13, 14, 18, 19, 26, 29–31, DA Direction Accuracy. 31
46 DAX The Deutscher Aktienindex. 29, 33
AI Artificial Intelligence. 3 DBN Deep Belief Network. 1, 5, 12, 13, 18, 19,
AIS Annealed Importance Sampling. 12 27–31, 33–36, 46
AMEX American Stock Exchange. 20, 21 DE Differential Evolution. 26
ANN Artificial Neural Network. 3–5, 8, 11–13, 18, DFNN Deep Feedforward Neural Network. 17, 19,
20, 23, 24, 31, 33, 34 21, 23, 29, 33
AR Active Return. 20 DGM Deep Neural Generative Model. 36, 37
AR Autoregressive. 22 DJI Dow Jones Index. 36
ARCH Autoregressive Conditional Heteroscedas- DJIA Dow Jones Industrial Average. 22–26, 36
ticity. 32 DL Deep Learning. 1–5, 7, 9, 10, 13, 14, 16–21, 23,
ARIMA Autoregressive Integrated Moving Aver- 24, 27–30, 35, 38–40, 44, 45, 47–49
age. 19, 20, 30, 32 DLR Deep Learning Representation. 37
ARMA Autoregressive Moving Average. 27, 28, DMLP Deep Multilayer Perceptron. 5–8, 10, 23,
32 30, 34, 35, 39–41, 45, 49
AUC Area Under the Curve. 20 DNN Deep Neural Network. 6, 10, 18–21, 23–33,
AUROC Area Under the Receiver Operating 35, 36
Characteristics. 33, 34, 36 DOW30 Dow Jones Industrial Average 30. 23, 26,
B&H Buy and Hold. 25 28
BELM Basic Extreme Learning Machine. 37 DP Dynamic Programming. 15, 16
Bi-GRU Bidirectional Gated Recurrent Unit. 35, DPA Direction Prediction Accuracy. 21
36 DRL Deep Reinforcement Learning. 2, 5, 16
Bi-LSTM Bidirectional LSTM. 22, 23, 25 DRSE Deep Random Subspace Ensembles. 36
BIST Istanbul Stock Exchange Index. 20, 21, 23, DWNN Deep and Wide Neural Network. 18, 19
24 EA Evolutionary Algorithm. 25, 27
Bovespa Brazilian Stock Exchange. 23, 25 EC Evolutionary Computation. 3, 4
BPTT Back Propagation Through Time. 7, 8 EGARCH Exponential GARCH. 24
BSE Bombay Stock Exchange. 26 ELM Extreme Learning Machine. 18, 19, 37
CCI Commodity Channel Index. 27 EMA Exponential Moving Average. 27
CD Contrastive Divergence. 12, 13 EMD2FNN Empirical Mode Decomposition and
CDAX German Stock Market Index Calculated by Factorization Machine based Neural Net-
Deutsche Börse. 20 work. 33, 34
CDBN Continuous-valued Deep Belief Networks. ETF Exchange-Traded Fund. 28, 31, 33, 34
30 FCNN Fully Connected Neural Network. 37
CDBN-FG Fuzzy Granulation with Continuous- FDDR Fuzzy Deep Direct Reinforcement Learn-
valued Deep Belief Networks. 31 ing. 23, 24, 27, 28
CNN Convolutional Neural Network. 1, 2, 5, 6, FFNN Feedforward Neural Network. 13, 14, 20,
10, 17–22, 25–42, 45–47, 49 24, 30, 33
CRBM Continuous Restricted Boltzman machine. FHSO Firefly Harmony Search Optimization. 25,
30 30, 31
50
FLANN Functional Link Neural network. 27 tion System. 23, 25, 31
FNN Fully Connected Neural Network. 7, 27, 33 LSTM Long-Short Term Memory. 1, 2, 5, 8, 9,
FTSE London Financial Times Stock Exchange In- 17–30, 32–37, 40, 44, 45, 47, 49
dex. 23, 25, 26, 28, 29, 31, 34 MACD Moving Average Convergence and Diver-
GA Genetic Algorithm. 4, 25, 34, 51 gence. 34, 36
GAF Gramian Angular Field. 33 MAD Mean Absolute Deviation. 18
GAN Generative Adversarial Network. 21, 47 MAE Mean Absolute Error. 18, 20, 21, 23, 24, 26,
GAN-FD GAN for minimizing Forecast error loss 29, 31, 33, 37
and Direction prediction loss. 20 MAM Moving Average Mapping. 33
GARCH Generalised Auto-Regressive Condi- MAPE Mean Absolute Percentage Error. 18, 21–
tional Heteroscedasticity. 28, 29, 31, 32, 24, 26–29, 31, 33, 37
52 MASE Mean Standard Deviation. 23
GASVR GA with a SVR. 25, 26, 28, 51 MC Monte Carlo. 15, 16
GBT Gradient Boosted Trees. 18, 19 MCC Matthew Correlation Coefficient. 22, 26, 36,
GDAX Global Digital Asset Exchange. 26 37
GLM Generalized Linear Model. 23 MDA Multilinear Discriminant Analysis. 20, 21
GML Generalized Linear Model. 25 MDD Maximum Drawdown. 18
GP Genetic Programming. 3, 4, 30 MDP Markov Decision Process. 14, 15
GPA The Gaussian Process Approach. 7, 8 MI Mutual Information. 18
GRU Gated-Recurrent Unit. 8, 18, 19, 21, 32–37, ML Machine Learning. 1–5, 7, 19, 25, 29, 30, 34,
40, 45 35, 45, 49
GS Grid Search. 7, 8, 10, 12–14, 16 MLP Multilayer Perceptron. 10, 18–21, 23, 24,
GSPC S&P500 Commodity Price Index. 24 29–31, 34, 40, 45
HAN Hybrid Attention Network. 36 MM Markov Model. 29
HAR Heterogeneous Autoregressive Process. 25, MODRL Multi-objective Deep Reinforcement
28, 51 Learning. 26
HAR-GASVR HAR with GASVR. 24, 28, 29 MoE Mixture of Experts. 24
HFT High Frequency Trading. 17, 28, 37, 47, 48 MOEA Multiobjective Evolutionary Algorithm. 4
HIT Hit Rate. 23, 24 MRS Markov Regime Switching. 27
HMAE Heteroscedasticity Adjusted MAE. 29 MS Manual Search. 7, 8, 10, 12–14, 16
HMM Hidden Markov Model. 35 MSE Mean Squared Error. 13, 14, 18, 20, 21, 23,
HMRPSO Modified Version of PSO. 26, 27 24, 26–29, 31–34
HMSE Heteroscedasticity Adjusted MSE. 29 MSFE Mean Squared Forecast Error. 24
HR Hit Rate. 21, 29, 31 MSPE Mean Squared Prediction Error. 20
HS China Shanghai Shenzhen Stock Index. 26, 37 NASDAQ National Association of Securities Deal-
HSI Hong Kong Hang Seng Index. 22–26, 33 ers Automated Quotations. 18–21, 23, 26,
IBOVESPA Indice Bolsa de Valores de Sao Paulo. 28, 33, 34, 36
34 NIFTY National Stock Exchange of India. 22, 26,
IC Information Coeffiencient. 20 36
IR Information Ratio. 20 NIKKEI Tokyo Nikkei Index. 22, 23, 25, 26, 31
ISE Istanbul Stock Exchange Index. 24, 31 NLP Natural Language Processing. 9, 46, 48, 49
IXIC NASDAQ Composite Index. 24 NMAE Normalized Mean Absolute Error. 29, 31
KELM Kernel Extreme Learning Machine. 37 NMSE Normalized Mean Square Error. 18, 29, 31
KL-Divergence Kullback Leibler Divergence. 12 NN Neural Network. 5, 6, 10, 16, 29, 33, 35–37
KOSPI The Korea Composite Stock Price Index. norm-RMSE Normalized RMSE. 18
18, 23, 25, 26, 29 NSE National Stock Exchange of India. 18
KSE Korea Stock Exchange. 34 NYMEX New York Mercantile Exchange. 27, 28
LAR Linear Auto-regression Predictor. 22 NYSE New York Stock Exchange. 18, 20, 21, 23,
LDA Latent Dirichlet Allocation. 26, 27, 37 34, 36
LOB Limit Order Book Data. 37 OCHL Open, Close, High, Low. 20, 24, 26
LRNFIS Locally Recurrent Neuro-fuzzy Informa- OCHLV Open, Close, High, Low, Volume. 18, 19,
51
 21–26, 32–34, 36 SGD Stochastic Gradient Descent. 6–8, 10, 13, 14
OMX Stockholm Stock Exchange. 23, 25, 28, 31, SLP Single Layer Perceptron. 34
33, 34 SMAPE Symmetric Mean Absolute Percentage
PCA Principal Component Analysis. 20, 21, 29, Error. 21
37 SMBGO Sequential Model-Based Global Opti-
PCC Pearson’s Correlation Coefficient. 31 mization. 7, 8
PCD Percentage of Correct Direction. 21 SPY SPDR S&P 500 ETF. 33
PLR Piecewise Linear Representation. 21 SR Sharpe-ratio. 21, 23–26, 28, 31
PNN Probabilistic Neural Network. 33 SRNN Stacked Recurrent Neural Network. 18, 19
PPOSC Percentage Price Oscillator. 27 SSE Shanghai Stock Exchange. 18, 21, 22, 26, 34
PSN Psi-Sigma Network. 30, 31 SSEC Shanghai Stock Exchange Composite. 33
PSO Particle Swarm Optimization. 26, 27 STD Standard Deviation. 23, 25, 26
R2 Squared correlation, Non-linear regression mul- SVM Support Vector Machine. 20, 22, 33–36
tiple correlation. 20, 21 SVR Support Vector Regressor. 25, 27, 30, 31, 51
RBF Radial Basis Function Neural Network. 18, SZSE Shenzhen Stock Exchange Composite Index.
19, 24, 31, 34 23, 24, 26, 28, 33
RBM Restricted Boltzmann Machine. 5, 11–13, TAIEX Taiwan Capitalization Weighted Stock In-
19, 30, 31, 34–36, 46 dex. 23, 25, 26
RCEFLANN Recurrent Computationally Effi- TALIB Technical Analysis Library Package. 26,
cient Functional Link Neural Network. 26, 34
27
TAQ Trade and Quote. 21
RCNN Recurrent CNN. 21, 22
TAR Threshold Autoregressive. 24, 25, 31
ReLU Rectified Linear Unit. 5, 7, 12
TD Temporal Difference. 15, 16
RF Random Forest. 18, 19, 22, 29
TDNN Timedelay Neural Network. 33
RHE Recurrent Hybrid Elman. 24, 25, 31
TF-IDF Term Frequency-Inverse Document Fre-
RL Reinforcement learning. 14, 15, 23, 24, 27, 28,
quency. 37
46, 47
TGRU Two-stream GRU. 35
RMDN Recurrent Mixture Density Network. 28,
31, 52 THEIL-U Theil’s inequality coefficient. 26, 31
RMDN-GARCH RMDN with GARCH. 28, 29 TR Total Return. 21, 24
RMSE Root Mean Square Error. 18, 20–24, 26–29, TSPEA Tree-structured Parzen Estimator Ap-
31–33, 37, 51 proach. 7, 8
RMSProp Root Mean Square Propagation. 6–8, TUNINDEX Tunisian Stock Market Index. 24
10, 13, 14 TWSE Taiwan Stock Exchange. 22, 26, 27
RMSRE Root Mean Square Relative Error. 21 VEC Vector Error Correction model. 24, 25, 31
RNN Recurrent Neural Network. 2, 5, 7–9, 18–33, VIX S&P500 Volatility Index. 23–25, 29
35–37, 39, 40, 45–47, 49 VXD Dow Jones Industrial Average Volatility In-
RS RandomSearch. 7, 8, 10, 12–14, 16 dex. 24, 25, 29
RSE Relative Squared Error. 21 VXN NASDAQ100 Volatility Index. 23–25, 29
RSI Relative Strength Index. 27, 34 WHR Weighted Hit Rate. 29, 31
RW Random Walk. 27, 28, 30 William%R Williams Percent Range. 27
S&P500 Standard’s & Poor’s 500 Index. 18–29, WMTR Weighted Multichannel Time-series Re-
31, 33, 36, 37 gression. 20, 21
SCI SSE Composite Index. 22 WT Wavelet Transforms. 26
SDAE Stacked Denoising Autoencoders. 27, 28 WTI West Texas Intermediate. 28
SFM State Frequency Memory. 18, 19 XGBoost eXtreme Gradient Boosting. 34, 35

52
References
[1] Ahmet Murat Ozbayoglu, Mehmet Ugur Gudelek, and Omer Berat Sezer. Deep learning for financial
applications : A survey. arXiv preprint arXiv:2002.05786, 2020.
[2] Rafik A. Aliev, Bijan Fazlollahi, and Rashad R. Aliev. Soft computing and its applications in business
and economics. In Studies in Fuzziness and Soft Computing, 2004.
[3] Ludmila Dymowa. Soft Computing in Economics and Finance. Springer Berlin Heidelberg, 2011.
[4] Boris Kovalerchuk and Evgenii Vityaev. Data Mining in Finance: Advances in Relational and Hybrid
Methods. Kluwer Academic Publishers, Norwell, MA, USA, 2000.
[5] Anthony Brabazon and Michael O’Neill, editors. Natural Computing in Computational Finance.
Springer Berlin Heidelberg, 2008.
[6] Arash Bahrammirzaee. A comparative survey of artificial intelligence applications in finance: artificial
neural networks, expert system and hybrid intelligent systems. Neural Computing and Applications,
19(8):1165–1195, June 2010.
[7] D. Zhang and L. Zhou. Discovering golden nuggets: Data mining in financial application. IEEE
Transactions on Systems, Man and Cybernetics, Part C (Applications and Reviews), 34(4):513–522,
November 2004.
[8] Asunción Mochón, David Quintana, Yago Sáez, and Pedro Isasi Viñuela. Soft computing techniques
applied to finance. Applied Intelligence, 29:111–115, 2007.
[9] Sendhil Mullainathan and Jann Spiess. Machine learning: An applied econometric approach. Journal
of Economic Perspectives, 31(2):87–106, May 2017.
[10] Shu-Heng Chen, editor. Genetic Algorithms and Genetic Programming in Computational Finance.
Springer US, 2002.
[11] Ma. Guadalupe Castillo Tapia and Carlos A. Coello Coello. Applications of multi-objective evolu-
tionary algorithms in economics and finance: A survey. In 2007 IEEE Congress on Evolutionary
Computation. IEEE, September 2007.
[12] Antonin Ponsich, Antonio Lopez Jaimes, and Carlos A. Coello Coello. A survey on multiobjective
evolutionary algorithms for the solution of the portfolio optimization problem and other finance and
economics applications. IEEE Transactions on Evolutionary Computation, 17(3):321–344, June 2013.
[13] Ruben Aguilar-Rivera, Manuel Valenzuela-Rendon, and J.J. Rodriguez-Ortiz. Genetic algorithms and
darwinian approaches in financial applications: A survey. Expert Systems with Applications, 42(21):
7684–7697, November 2015.
[14] Roy Rada. Expert systems and evolutionary computing for financial investing: A review. Expert
Systems with Applications, 34(4):2232–2240, 2008.
[15] Yuhong Li and Weihua Ma. Applications of artificial neural networks in financial economics: A survey.
In 2010 International Symposium on Computational Intelligence and Design. IEEE, October 2010.
[16] Michal Tkáč and Robert Verner. Artificial neural networks in business: Two decades of research.
Applied Soft Computing, 38:788 – 804, 2016.
[17] B. Elmsili and B. Outtaj. Artificial neural networks applications in economics and management
research: An exploratory literature review. In 2018 4th International Conference on Optimization
and Applications (ICOA), pages 1–6, April 2018.
[18] Marc-André Mittermayer and Gerhard F Knolmayer. Text mining systems for market response to
news: A survey. September 2006.
[19] Leela Mitra and Gautam Mitra. Applications of news analytics in finance: A review. In The Handbook
of News Analytics in Finance, pages 1–39. John Wiley & Sons, Ltd., May 2012.
[20] Arman Khadjeh Nassirtoussi, Saeed Aghabozorgi, Teh Ying Wah, and David Chek Ling Ngo. Text
mining for market prediction: A systematic review. Expert Systems with Applications, 41(16):7653–
7670, November 2014.
[21] Colm Kearney and Sha Liu. Textual sentiment in finance: A survey of methods and models. Interna-
tional Review of Financial Analysis, 33:171–185, May 2014.
[22] B. Shravan Kumar and Vadlamani Ravi. A survey of the applications of text mining in financial
domain. Knowledge-Based Systems, 114:128–147, December 2016.
53
[23] Frank Z. Xing, Erik Cambria, and Roy E. Welsch. Natural language based financial forecasting: a
survey. Artificial Intelligence Review, 50(1):49–73, October 2017.
[24] Bruce J Vanstone and Clarence Tan. A survey of the application of soft computing to investment and
financial trading. In Brian C Lovell, Duncan A Campbell, Clinton B Fookes, and Anthony J Maeder,
editors, Proceedings of the Eighth Australian and New Zealand Intelligent Information Systems Con-
ference (ANZIIS 2003), pages 211–216. The Australian Pattern Recognition Society, 2003. Copyright
The Australian Pattern Recognition Society 2003. All rights reserved. Permission granted.
[25] Ehsan Hajizadeh, H. Davari Ardakani, and Jamal Shahrabi. Application of data mining techniques in
stock markets : A survey. 2010.
[26] Binoy B. Nair and V.P. Mohandas. Artificial intelligence applications in financial forecasting – a
survey and some empirical results. Intelligent Decision Technologies, 9(2):99–140, December 2014.
[27] Rodolfo C. Cavalcante, Rodrigo C. Brasileiro, Victor L.F. Souza, Jarley P. Nobrega, and Adriano L.I.
Oliveira. Computational intelligence and financial markets: A survey and future directions. Expert
Systems with Applications, 55:194–211, August 2016.
[28] Bjoern Krollner, Bruce J. Vanstone, and Gavin R. Finnie. Financial time series forecasting with
machine learning techniques: a survey. In ESANN, 2010.
[29] P. D. Yoo, M. H. Kim, and T. Jan. Machine learning techniques and use of event information for
stock market prediction: A survey and evaluation. In International Conference on Computational In-
telligence for Modelling, Control and Automation and International Conference on Intelligent Agents,
Web Technologies and Internet Commerce (CIMCA-IAWTIC’06), volume 2, pages 835–841, November
2005.
[30] G Preethi and B Santhi. Stock market forecasting techniques: A survey. Journal of Theoretical and
Applied Information Technology, 46:24–30, December 2012.
[31] George S. Atsalakis and Kimon P. Valavanis. Surveying stock market forecasting techniques – part ii:
Soft computing methods. Expert Systems with Applications, 36(3):5932–5941, April 2009.
[32] Amitava Chatterjee, O.Felix Ayadi, and Bryan E. Boone. Artificial neural network and the financial
markets: A survey. Managerial Finance, 26(12):32–45, December 2000.
[33] R. Katarya and A. Mahajan. A survey of neural network techniques in market trend analysis. In 2017
International Conference on Intelligent Sustainable Systems (ICISS), pages 873–877, December 2017.
[34] Yong Hu, Kang Liu, Xiangzhou Zhang, Lijun Su, E.W.T. Ngai, and Mei Liu. Application of evolu-
tionary computation for rule discovery in stock algorithmic trading: A literature review. Applied Soft
Computing, 36:534–551, November 2015.
[35] Wei Huang, K. K. Lai, Y. Nakamori, and Shouyang Wang. Forecasting foreign exchange rates with
artificial neural networks: A review. International Journal of Information Technology & Decision
Making, 03(01):145–165, 2004.
[36] Dadabada Pradeepkumar and Vadlamani Ravi. Soft computing hybrids for forex rate prediction: A
comprehensive review. Computers & Operations Research, 99:262 – 284, 2018.
[37] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–444, 2015.
[38] Jürgen Schmidhuber. Deep learning in neural networks: An overview. Neural networks, 61:85–117,
2015.
[39] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016.
http://www.deeplearningbook.org.
[40] George Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of control,
signals and systems, 2(4):303–314, 1989.
[41] Barry L Kalman and Stan C Kwasny. Why tanh: choosing a sigmoidal function. In [Proceedings 1992]
IJCNN International Joint Conference on Neural Networks, volume 4, pages 578–581. IEEE, 1992.
[42] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines.
In Proceedings of the 27th international conference on machine learning (ICML-10), pages 807–814,
2010.
[43] Andrew L Maas, Awni Y Hannun, and Andrew Y Ng. Rectifier nonlinearities improve neural network
acoustic models. In Proc. icml, volume 30, page 3, 2013.

54
[44] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint
arXiv:1710.05941, 2017.
[45] Li Deng, Dong Yu, et al. Deep learning: methods and applications. Foundations and Trends R in
Signal Processing, 7(3–4):197–387, 2014.
[46] Matt W Gardner and SR Dorling. Artificial neural networks (the multilayer perceptron)—a review of
applications in the atmospheric sciences. Atmospheric environment, 32(14-15):2627–2636, 1998.
[47] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical
statistics, pages 400–407, 1951.
[48] Ilya Sutskever, James Martens, George Dahl, and Geoffrey Hinton. On the importance of initialization
and momentum in deep learning. In International conference on machine learning, pages 1139–1147,
2013.
[49] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and
stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011.
[50] Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average
of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26–31, 2012.
[51] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
[52] Yoshua Bengio, Patrice Simard, Paolo Frasconi, et al. Learning long-term dependencies with gradient
descent is difficult. IEEE transactions on neural networks, 5(2):157–166, 1994.
[53] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpass-
ing human-level performance on imagenet classification. In Proceedings of the IEEE international
conference on computer vision, pages 1026–1034, 2015.
[54] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neu-
ral networks. In Proceedings of the thirteenth international conference on artificial intelligence and
statistics, pages 249–256, 2010.
[55] James S Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl. Algorithms for hyper-parameter
optimization. In Advances in neural information processing systems, pages 2546–2554, 2011.
[56] James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. Journal of
Machine Learning Research, 13(Feb):281–305, 2012.
[57] Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural
networks. In International conference on machine learning, pages 1310–1318, 2013.
[58] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–
1780, 1997.
[59] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey,
Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation
system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144,
2016.
[60] Klaus Greff, Rupesh K Srivastava, Jan Koutník, Bas R Steunebrink, and Jürgen Schmidhuber. Lstm:
A search space odyssey. IEEE transactions on neural networks and learning systems, 28(10):2222–
2232, 2016.
[61] Nils Reimers and Iryna Gurevych. Optimal hyperparameters for deep lstm-networks for sequence
labeling tasks. arXiv preprint arXiv:1707.06799, 2017.
[62] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3d convolutional neural networks for human action
recognition. IEEE transactions on pattern analysis and machine intelligence, 35(1):221–231, 2012.
[63] Christian Szegedy, Alexander Toshev, and Dumitru Erhan. Deep neural networks for object detection.
In Advances in neural information processing systems, pages 2553–2561, 2013.
[64] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic
segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition,
pages 3431–3440, 2015.
[65] Xueheng Qiu, Le Zhang, Ye Ren, P. Suganthan, and Gehan Amaratunga. Ensemble deep learning
for regression and time series forecasting. In 2014 IEEE Symposium on Computational Intelligence in

55
 Ensemble Learning (CIEL), pages 1–6, 2014.
[66] Rafael Hrasko, André GC Pacheco, and Renato A Krohling. Time series prediction using restricted
boltzmann machines and backpropagation. Procedia Computer Science, 55:990–999, 2015.
[67] Ruslan Salakhutdinov, Andriy Mnih, and Geoffrey Hinton. Restricted boltzmann machines for col-
laborative filtering. In Proceedings of the 24th international conference on Machine learning, pages
791–798. ACM, 2007.
[68] Yoshua Bengio. Deep learning of representations for unsupervised and transfer learning. In Proceedings
of ICML workshop on unsupervised and transfer learning, pages 17–36, 2012.
[69] Abdel-rahman Mohamed, George Dahl, and Geoffrey Hinton. Deep belief networks for phone recog-
nition. In Nips workshop on deep learning for speech recognition and related applications, volume 1,
page 39. Vancouver, Canada, 2009.
[70] Honglak Lee, Roger Grosse, Rajesh Ranganath, and Andrew Y Ng. Convolutional deep belief networks
for scalable unsupervised learning of hierarchical representations. In Proceedings of the 26th annual
international conference on machine learning, pages 609–616. ACM, 2009.
[71] Laurens Van Der Maaten. Learning a parametric embedding by preserving local structure. In Artificial
Intelligence and Statistics, pages 384–391, 2009.
[72] Chengwei Yao and Gencai Chen. Hyperparameters adaptation for restricted boltzmann machines
based on free energy. In 2016 8th International Conference on Intelligent Human-Machine Systems
and Cybernetics (IHMSC), volume 2, pages 243–248. IEEE, 2016.
[73] Miguel A Carreira-Perpinan and Geoffrey E Hinton. On contrastive divergence learning. In Aistats,
volume 10, pages 33–40. Citeseer, 2005.
[74] Prasanna Tamilselvan and Pingfeng Wang. Failure diagnosis using deep belief learning based health
state classification. Reliability Engineering & System Safety, 115:124–135, 2013.
[75] Geoffrey E. Hinton, Simon Osindero, and Yee-Whye Teh. A fast learning algorithm for deep belief
nets. Neural Computation, 18(7):1527–1554, 2006.
[76] Qinxue Meng, Daniel Catchpoole, David Skillicom, and Paul J Kennedy. Relational autoencoder
for feature extraction. In 2017 International Joint Conference on Neural Networks (IJCNN), pages
364–371. IEEE, 2017.
[77] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and
composing robust features with denoising autoencoders. In Proceedings of the 25th international
conference on Machine learning, pages 1096–1103. ACM, 2008.
[78] Richard S Sutton and Andrew G Barto. Introduction to reinforcement learning, volume 135. MIT
press Cambridge, 1998.
[79] Duy Nguyen-Tuong and Jan Peters. Model learning for robot control: a survey. Cognitive processing,
12(4):319–340, 2011.
[80] Eunsuk Chong, Chulwoo Han, and Frank C. Park. Deep learning networks for stock market analysis
and prediction: Methodology, data representations, and case studies. Expert Systems with Applica-
tions, 83:187–205, October 2017.
[81] Kai Chen, Yi Zhou, and Fangyan Dai. A lstm-based method for stock returns prediction: A case
study of china stock market. In 2015 IEEE International Conference on Big Data (Big Data). IEEE,
October 2015.
[82] Eva Dezsi and Ioan Alin Nistor. Can deep machine learning outsmart the market? a comparison
between econometric modelling and long- short term memory. Romanian Economic Business Review,
11(4.1):54–73, December 2016.
[83] A.J.P. Samarawickrama and T.G.I. Fernando. A recurrent neural network approach in predicting daily
stock prices an application to the sri lankan stock market. In 2017 IEEE International Conference on
Industrial and Information Systems (ICIIS). IEEE, December 2017.
[84] M Hiransha, E.A. Gopalakrishnan, Vijay Krishna Menon, and K.P. Soman. Nse stock market predic-
tion using deep-learning models. Procedia Computer Science, 132:1351–1362, 2018.
[85] Sreelekshmy Selvin, R Vinayakumar, E. A Gopalakrishnan, Vijay Krishna Menon, and K. P. Soman.
Stock price prediction using lstm, rnn and cnn-sliding window model. In 2017 International Conference

56
 on Advances in Computing, Communications and Informatics (ICACCI). IEEE, September 2017.
[86] Sang Il Lee and Seong Joon Yoo. Threshold-based portfolio: the role of the threshold and its appli-
cations. The Journal of Supercomputing, September 2018.
[87] Xiumin Li, Lin Yang, Fangzheng Xue, and Hongjun Zhou. Time series prediction of stock price using
deep belief networks with intrinsic plasticity. In 2017 29th Chinese Control And Decision Conference
(CCDC). IEEE, May 2017.
[88] Lin Chen, Zhilin Qiao, Minggang Wang, Chao Wang, Ruijin Du, and Harry Eugene Stanley. Which
artificial intelligence algorithm better predicts the chinese stock market? IEEE Access, 6:48625–48633,
2018.
[89] Christopher Krauss, Xuan Anh Do, and Nicolas Huck. Deep neural networks, gradient-boosted trees,
random forests: Statistical arbitrage on the s&p 500. European Journal of Operational Research, 259
(2):689–702, June 2017.
[90] Rohitash Chandra and Shelvin Chand. Evaluation of co-evolutionary neural network architectures
for time series prediction with mobile application in finance. Applied Soft Computing, 49:462–473,
December 2016.
[91] Shuanglong Liu, Chao Zhang, and Jinwen Ma. Cnn-lstm neural network model for quantitative strat-
egy analysis in stock markets. In Neural Information Processing, pages 198–206. Springer International
Publishing, 2017.
[92] J. B. Heaton, N. G. Polson, and J. H. Witte. Deep learning in finance, 2016.
[93] Bilberto Batres-Estrada. Deep learning for multivariate financial time series. Master’s thesis, KTH,
Mathematical Statistics, 2015.
[94] Zhaozheng Yuan, Ruixun Zhang, and Xiuli Shao. Deep and wide neural networks on multiple sets of
temporal data with correlation. In Proceedings of the 2018 International Conference on Computing
and Data Engineering - ICCDE 2018. ACM Press, 2018.
[95] Liheng Zhang, Charu Aggarwal, and Guo-Jun Qi. Stock price prediction via discovering multi-
frequency trading patterns. In Proceedings of the 23rd ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining - KDD17. ACM Press, 2017.
[96] Masaya Abe and Hideki Nakayama. Deep learning for forecasting stock returns in the cross-section. In
Advances in Knowledge Discovery and Data Mining, pages 273–284. Springer International Publishing,
2018.
[97] Guanhao Feng, Jingyu He, and Nicholas G. Polson. Deep learning for predicting asset returns, 2018.
[98] Jianqing Fan, Lingzhou Xue, and Jiawei Yao. Sufficient forecasting using factor models. SSRN
Electronic Journal, 2014.
[99] Mathias Kraus and Stefan Feuerriegel. Decision support from financial disclosures with deep neural
networks and transfer learning. Decision Support Systems, 104:38–48, December 2017.
[100] Shotaro Minami. Predicting equity price with corporate action events using lstm-rnn. Journal of
Mathematical Finance, 08(01):58–63, 2018.
[101] Xiaolin Zhang and Ying Tan. Deep stock ranker: A lstm neural network model for stock selection. In
Data Mining and Big Data, pages 614–623. Springer International Publishing, 2018.
[102] Qun Zhuge, Lingyu Xu, and Gaowei Zhang. Lstm neural network with emotional analysis for prediction
of stock price. 2017.
[103] Ryo Akita, Akira Yoshihara, Takashi Matsubara, and Kuniaki Uehara. Deep learning for stock pre-
diction using numerical and textual information. In 2016 IEEE/ACIS 15th International Conference
on Computer and Information Science (ICIS). IEEE, June 2016.
[104] A. Ozbayoglu. Neural based technical analysis in stock market forecasting. In Intelligent Engineering
Systems Through Artificial Neural Networks, Volume 17, pages 261–266. ASME, 2007.
[105] Kaustubh Khare, Omkar Darekar, Prafull Gupta, and V. Z. Attar. Short term stock price prediction
using deep learning. In 2017 2nd IEEE International Conference on Recent Trends in Electronics,
Information & Communication Technology (RTEICT). IEEE, May 2017.
[106] Xingyu Zhou, Zhisong Pan, Guyu Hu, Siqi Tang, and Cheng Zhao. Stock market prediction on high-
frequency data using generative adversarial nets. Mathematical Problems in Engineering, 2018:1–11,

57
 2018.
[107] Ritika Singh and Shashi Srivastava. Stock prediction using deep learning. Multimedia Tools and
Applications, 76(18):18569–18584, December 2016.
[108] Sercan Karaoglu and Ugur Arpaci. A deep learning approach for optimization of systematic signal
detection in financial trading systems with big data. International Journal of Intelligent Systems and
Applications in Engineering, SpecialIssue(SpecialIssue):31–36, July 2017.
[109] Bo Zhou. Deep learning and the cross-section of stock returns: Neural networks combining price and
fundamental information. SSRN Electronic Journal, 2018.
[110] Narek Abroyan and. Neural networks for financial market risk classification. Frontiers in Signal
Processing, 1(2), August 2017.
[111] Google. System and method for computer managed funds to outperform benchmarks.
[112] Dat Thanh Tran, Martin Magris, Juho Kanniainen, Moncef Gabbouj, and Alexandros Iosifidis. Tensor
representation in high-frequency financial data for price change prediction. In 2017 IEEE Symposium
Series on Computational Intelligence (SSCI). IEEE, November 2017.
[113] Guanhao Feng, Nicholas G. Polson, and Jianeng Xu. Deep factor alpha, 2018.
[114] Xiao Ding, Yue Zhang, Ting Liu, and Junwen Duan. Deep learning for event-driven stock prediction.
In Proceedings of the 24th International Conference on Artificial Intelligence, IJCAI’15, pages 2327–
2333. AAAI Press, 2015.
[115] Manuel R. Vargas, Beatriz S. L. P. de Lima, and Alexandre G. Evsukoff. Deep learning for stock market
prediction from financial news articles. In 2017 IEEE International Conference on Computational
Intelligence and Virtual Environments for Measurement Systems and Applications (CIVEMSA). IEEE,
June 2017.
[116] Che-Yu Lee and Von-Wun Soo. Predict stock price with financial news based on recurrent convolu-
tional neural networks. In 2017 Conference on Technologies and Applications of Artificial Intelligence
(TAAI). IEEE, December 2017.
[117] Hitoshi Iwasaki and Ying Chen. Topic sentiment asset pricing with dnn supervised learning. SSRN
Electronic Journal, 2018.
[118] Sushree Das, Ranjan Kumar Behera, Mukesh Kumar, and Santanu Kumar Rath. Real-time sentiment
analysis of twitter streaming data for stock prediction. Procedia Computer Science, 132:956–964, 2018.
[119] Jiahong Li, Hui Bu, and Junjie Wu. Sentiment-aware stock market prediction: A deep learning
method. In 2017 International Conference on Service Systems and Service Management. IEEE, June
2017.
[120] Zhongshengz. Measuring financial crisis index for risk warning through analysis of social network.
Master’s thesis, 2018.
[121] Janderson B. Nascimento and Marco Cristo. The impact of structured event embeddings on scalable
stock forecasting models. In Proceedings of the 21st Brazilian Symposium on Multimedia and the Web
- WebMedia15. ACM Press, 2015.
[122] Songqiao Han, Xiaoling Hao, and Hailiang Huang. An event-extraction approach for business analysis
from online chinese news. Electronic Commerce Research and Applications, 28:244–260, March 2018.
[123] Wei Bao, Jun Yue, and Yulei Rao. A deep learning framework for financial time series using stacked
autoencoders and long-short term memory. PLOS ONE, 12(7):e0180944, July 2017.
[124] A.K. Parida, R. Bisoi, and P.K. Dash. Chebyshev polynomial functions based locally recurrent neuro-
fuzzy information system for prediction of financial and energy market data. The Journal of Finance
and Data Science, 2(3):202–223, September 2016.
[125] Thomas Fischer and Christopher Krauss. Deep learning with long short-term memory networks for
financial market predictions. European Journal of Operational Research, 270(2):654–669, October
2018.
[126] Philip Widegren. Deep learning-based forecasting of financial assets. Master’s thesis, KTH, Mathe-
matical Statistics, 2017.
[127] Anastasia Borovykh, Sander Bohte, and Cornelis W. Oosterlee. Dilated convolutional neural networks
for time series forecasting. Journal of Computational Finance, October 2018.

58
[128] Khaled A. Althelaya, El-Sayed M. El-Alfy, and Salahadin Mohammed. Evaluation of bidirectional lstm
for short-and long-term stock market prediction. In 2018 9th International Conference on Information
and Communication Systems (ICICS). IEEE, April 2018.
[129] Alexiei Dingli and Karl Sant Fournier. Financial time series forecasting–a deep learning approach.
Int. J. Mach. Learn. Comput, 7(5):118–122, 2017.
[130] Ajit Kumar Rout, P.K. Dash, Rajashree Dash, and Ranjeeta Bisoi. Forecasting financial time series
using a low complexity recurrent neural network and evolutionary learning approach. Journal of King
Saud University - Computer and Information Sciences, 29(4):536–552, October 2017.
[131] Gyeeun Jeong and Ha Young Kim. Improving financial trading decisions using deep q-learning: Pre-
dicting the number of shares, action strategies, and transfer learning. Expert Systems with Applications,
117:125–138, March 2019.
[132] Yujin Baek and Ha Young Kim. Modaugnet: A new forecasting framework for stock market index
value with an overfitting prevention lstm module and a prediction lstm module. Expert Systems with
Applications, 113:457–480, December 2018.
[133] Magnus Hansson. On stock return prediction with lstm networks. 2017.
[134] Aaron Elliot and Cheng Hua Hsu. Time series prediction : Predicting stock price, 2017.
[135] Zhixi Li and Vincent Tam. Combining the real-time wavelet denoising and long-short-term-memory
neural network for predicting stock indexes. In 2017 IEEE Symposium Series on Computational
Intelligence (SSCI). IEEE, November 2017.
[136] Sima Siami-Namini and Akbar Siami Namin. Forecasting economics and financial time series: Arima
vs. lstm, 2018.
[137] Tsung-Jung Hsieh, Hsiao-Fen Hsiao, and Wei-Chang Yeh. Forecasting stock markets using wavelet
transforms and recurrent neural networks: An integrated system based on artificial bee colony algo-
rithm. Applied Soft Computing, 11(2):2510–2525, March 2011.
[138] Luna M. Zhang. Genetic deep neural networks using different activation functions for financial data
mining. In 2015 IEEE International Conference on Big Data (Big Data). IEEE, October 2015.
[139] Stelios D. Bekiros. Irrational fads, short-term memory emulation, and asset predictability. Review of
Financial Economics, 22(4):213–219, November 2013.
[140] Xiongwen Pang, Yanqiang Zhou, Pan Wang, Weiwei Lin, and Victor Chang. An innovative neural
network approach for stock market prediction. The Journal of Supercomputing, January 2018.
[141] Yue Deng, Feng Bao, Youyong Kong, Zhiquan Ren, and Qionghai Dai. Deep direct reinforcement
learning for financial signal representation and trading. IEEE Transactions on Neural Networks and
Learning Systems, 28(3):653–664, March 2017.
[142] Bing Yang, Zi-Jia Gong, and Wenqi Yang. Stock market index prediction using deep neural network
ensemble. In 2017 36th Chinese Control Conference (CCC). IEEE, July 2017.
[143] Oussama Lachiheb and Mohamed Salah Gouider. A hierarchical deep neural network design for stock
returns prediction. Procedia Computer Science, 126:264–272, 2018.
[144] Bang Xiang Yong, Mohd Rozaini Abdul Rahim, and Ahmad Shahidan Abdullah. A stock market
trading system using deep neural network. In Communications in Computer and Information Science,
pages 356–364. Springer Singapore, 2017.
[145] Serdar Yümlü, Fikret S. Gürgen, and Nesrin Okay. A comparison of global, recurrent and smoothed-
piecewise neural models for istanbul stock exchange (ise) prediction. Pattern Recognition Letters, 26
(13):2093–2103, October 2005.
[146] Hongju Yan and Hongbing Ouyang. Financial time series prediction based on deep learning. Wireless
Personal Communications, 102(2):683–700, December 2017.
[147] Takahashi. Long memory and predictability in financial markets. Annual Conference of the Japanese
Society for Artificial Intelligence, 2017.
[148] Melike Bildirici, Elçin A. Alp, and Özgür Ö. Ersin. Tar-cointegration neural network model: An
empirical analysis of exchange rates and stock returns. Expert Systems with Applications, 37(1):2–11,
January 2010.
[149] Ioannis Psaradellis and Georgios Sermpinis. Modelling and trading the u.s. implied volatility indices.

59
 evidence from the vix, vxn and vxd indices. International Journal of Forecasting, 32(4):1268–1283,
October 2016.
[150] Jiaqi Chen, Wenbo Wu, and Michael Tindall. Hedge fund return prediction and fund selection: A
machine-learning approach. Occasional Papers 16-4, Federal Reserve Bank of Dallas, November 2016.
[151] Marios Mourelatos, Christos Alexakos, Thomas Amorgianiotis, and Spiridon Likothanassis. Financial
indices modelling and trading utilizing deep learning techniques: The athens se ftse/ase large cap use
case. In 2018 Innovations in Intelligent Systems and Applications (INISTA). IEEE, July 2018.
[152] Yuzhou Chen, Junji Wu, and Hui Bu. Stock market embedding and prediction: A deep learning
method. In 2018 15th International Conference on Service Systems and Service Management (IC-
SSSM). IEEE, July 2018.
[153] Weiyu Si, Jinke Li, Peng Ding, and Ruonan Rao. A multi-objective deep reinforcement learning
approach for stock index future’s intraday trading. In 2017 10th International Symposium on Com-
putational Intelligence and Design (ISCID). IEEE, December 2017.
[154] Weiling Chen, Chai Kiat Yeo, Chiew Tong Lau, and Bu Sung Lee. Leveraging social media news to
predict stock index movement using rnn-boost. Data & Knowledge Engineering, August 2018.
[155] Matthew Francis Dixon, Diego Klabjan, and Jin Hoon Bang. Classification-based financial markets
prediction using deep neural networks. SSRN Electronic Journal, 2016.
[156] Fernando Sánchez Lasheras, Francisco Javier de Cos Juez, Ana Suárez Sánchez, Alicja Krzemień, and
Pedro Riesgo Fernández. Forecasting the comex copper spot price by means of neural networks and
arima models. Resources Policy, 45:37–43, September 2015.
[157] Yang Zhao, Jianping Li, and Lean Yu. A deep learning ensemble approach for crude oil price fore-
casting. Energy Economics, 66:9–16, August 2017.
[158] Yanhui Chen, Kaijian He, and Geoffrey K.F. Tso. Forecasting crude oil prices: a deep learning based
model. Procedia Computer Science, 122:300–307, 2017.
[159] Jonathan Doering, Michael Fairbank, and Sheri Markose. Convolutional neural networks applied
to high-frequency market microstructure forecasting. In 2017 9th Computer Science and Electronic
Engineering (CEEC). IEEE, September 2017.
[160] P. Tino, C. Schittenkopf, and G. Dorffner. Financial volatility trading using recurrent neural networks.
IEEE Transactions on Neural Networks, 12(4):865–874, July 2001.
[161] Ruoxuan Xiong, Eric P. Nichols, and Yuan Shen. Deep learning stock volatility with google domestic
trends, 2015.
[162] Yu-Long Zhou, Ren-Jie Han, Qian Xu, and Wei-Ke Zhang. Long short-term memory networks for
csi300 volatility prediction with baidu search volume. 2018.
[163] Ha Young Kim and Chang Hyun Won. Forecasting the volatility of stock price index: A hybrid
model integrating lstm with multiple garch-type models. Expert Systems with Applications, 103:25–
37, August 2018.
[164] Nikolay Nikolaev, Peter Tino, and Evgueni Smirnov. Time-dependent series variance learning with
recurrent mixture density networks. Neurocomputing, 122:501–512, December 2013.
[165] Campbell R. Harvey. Forecasts of economic growth from the bond and stock markets. Financial
Analysts Journal, 45(5):38–45, 1989.
[166] Daniele Bianchi, Matthias Büchner, and Andrea Tamoni. Bond risk premia with machine learning.
SSRN Electronic Journal, September 2018.
[167] Venketas Warren. Forex market size: A traders advantage, 2019.
[168] Ren Zhang, Furao Shen, and Jinxi Zhao. A model with fuzzy granulation and deep belief networks
for exchange rate forecasting. In 2014 International Joint Conference on Neural Networks (IJCNN).
IEEE, July 2014.
[169] Jing Chao, Furao Shen, and Jinxi Zhao. Forecasting exchange rate with deep belief networks. In The
2011 International Joint Conference on Neural Networks. IEEE, July 2011.
[170] Jing Zheng, Xiao Fu, and Guijun Zhang. Research on exchange rate forecasting based on deep belief
network. Neural Computing and Applications, May 2017.
[171] Furao Shen, Jing Chao, and Jinxi Zhao. Forecasting exchange rate using deep belief networks and

60
 conjugate gradient method. Neurocomputing, 167:243–253, November 2015.
[172] Hua Shen and Xun Liang. A time series forecasting model based on deep learning integrated algorithm
with stacked autoencoders and svr for fx prediction. In ICANN, 2016.
[173] Georgios Sermpinis, Jason Laws, Andreas Karathanasopoulos, and Christian L. Dunis. Forecasting
and trading the eur/usd exchange rate with gene expression and psi sigma neural networks. Expert
Systems with Applications, 39(10):8865–8877, August 2012.
[174] Georgios Sermpinis, Christian Dunis, Jason Laws, and Charalampos Stasinakis. Forecasting and trad-
ing the eur/usd exchange rate with stochastic neural network combination and time-varying leverage.
Decision Support Systems, 54(1):316–329, December 2012.
[175] Georgios Sermpinis, Charalampos Stasinakis, and Christian Dunis. Stochastic and genetic neural
network combinations in trading and hybrid time-varying leverage effects. Journal of International
Financial Markets, Institutions and Money, 30:21–54, May 2014.
[176] Bo SUN and Chi XIE. Rmb exchange rate forecasting in the context of the financial crisis. Systems
Engineering - Theory & Practice, 29(12):53–64, December 2009.
[177] Nijolė Maknickienė and Algirdas Maknickas. Financial market prediction system with evolino neural
network and deplhi method. Journal of Business Economics and Management, 14(2):403–413, May
2013.
[178] Nijole Maknickiene, Aleksandras Vytautas Rutkauskas, and Algirdas Maknickas. Investigation of
financial market prediction by recurrent neural network. 2014.
[179] Luca Di Persio and Oleksandr Honchar. Artificial neural networks approach to the forecast of stock
market price movements. International Journal of Economics and Management Systems, (1):158–162,
2016.
[180] Jerzy Korczak and Marcin Hernes. Deep learning for financial time series forecasting in a-trader
system. In Proceedings of the 2017 Federated Conference on Computer Science and Information
Systems. IEEE, September 2017.
[181] Gonçalo Duarte Lima Freire Lopes. Deep learning for market forecasts. 2018.
[182] Sean McNally, Jason Roche, and Simon Caton. Predicting the price of bitcoin using machine learn-
ing. In 2018 26th Euromicro International Conference on Parallel, Distributed and Network-based
Processing (PDP). IEEE, March 2018.
[183] Sanjiv Das, Karthik Mokashi, and Robbie Culkin. Are markets truly efficient? experiments using
deep learning algorithms for market movement prediction. Algorithms, 11(9):138, September 2018.
[184] Ariel Navon and Yosi Keller. Financial time series prediction using deep learning, 2017.
[185] E.W. Saad, D.V. Prokhorov, and D.C. Wunsch. Comparative study of stock trend prediction using
time delay, recurrent and probabilistic neural networks. IEEE Transactions on Neural Networks, 9
(6):1456–1470, 1998.
[186] Luca Di Persio and Oleksandr Honchar. Recurrent neural networks approach to the financial forecast
of google assets. International Journal of Mathematics and Computers in Simulation, 11:713, 2017.
[187] Guizhu Shen, Qingping Tan, Haoyu Zhang, Ping Zeng, and Jianjun Xu. Deep learning with gated
recurrent unit networks for financial sequence predictions. Procedia Computer Science, 131:895–903,
2018.
[188] Jou-Fan Chen, Wei-Lun Chen, Chun-Ping Huang, Szu-Hao Huang, and An-Pin Chen. Financial time-
series data analysis using deep convolutional neural networks. In 2016 7th International Conference
on Cloud Computing and Big Data (CCBD). IEEE, November 2016.
[189] Omer Berat Sezer and Ahmet Murat Ozbayoglu. Financial trading model with stock bar chart image
time series with deep convolutional neural networks. arXiv preprint arXiv:1903.04610, 2019.
[190] Feng Zhou, Hao min Zhou, Zhihua Yang, and Lihua Yang. Emd2fnn: A strategy combining empirical
mode decomposition and factorization machine based neural network for stock market trend prediction.
Expert Systems with Applications, 115:136–151, 2019.
[191] Kristiina Ausmees, Slobodan Milovanovic, Fredrik Wrede, and Afshin Zafari. Taming deep belief
networks. 2017.
[192] Kamran Raza. Prediction of stock market performance by using machine learning techniques. In 2017

61
 International Conference on Innovations in Electrical Engineering and Computational Technologies
(ICIEECT). IEEE, April 2017.
[193] Omer Berat Sezer, Murat Ozbayoglu, and Erdogan Dogdu. A deep neural-network based stock trading
system based on evolutionary optimized technical analysis parameters. Procedia Computer Science,
114:473–480, 2017.
[194] Qiubin Liang, Wenge Rong, Jiayi Zhang, Jingshuang Liu, and Zhang Xiong. Restricted boltzmann
machine based stock market trend prediction. In 2017 International Joint Conference on Neural
Networks (IJCNN). IEEE, May 2017.
[195] Luigi Troiano, Elena Mejuto Villa, and Vincenzo Loia. Replicating a trading strategy by means of
lstm for financial industry applications. IEEE Transactions on Industrial Informatics, 14(7):3226–
3234, July 2018.
[196] David M. Q. Nelson, Adriano C. M. Pereira, and Renato A. de Oliveira. Stock markets price movement
prediction with lstm neural networks. In 2017 International Joint Conference on Neural Networks
(IJCNN). IEEE, May 2017.
[197] Yuan Song and Yingnian Wu. Stock trend prediction: Based on machine learning methods. Master’s
thesis, 2018.
[198] M. Ugur Gudelek, S. Arda Boluk, and A. Murat Ozbayoglu. A deep learning based stock trading
model with 2-d cnn trend detection. In 2017 IEEE Symposium Series on Computational Intelligence
(SSCI). IEEE, November 2017.
[199] Omer Berat Sezer and Ahmet Murat Ozbayoglu. Algorithmic financial trading with deep convolutional
neural networks: Time series to image conversion approach. Applied Soft Computing, 70:525–538,
September 2018.
[200] Hakan Gunduz, Yusuf Yaslan, and Zehra Cataltepe. Intraday prediction of borsa istanbul using convo-
lutional neural networks and feature correlations. Knowledge-Based Systems, 137:138–148, December
2017.
[201] Yifu Huang, Kai Huang, Yang Wang, Hao Zhang, Jihong Guan, and Shuigeng Zhou. Exploiting twitter
moods to boost financial trend prediction based on deep network models. In Intelligent Computing
Methodologies, pages 449–460. Springer International Publishing, 2016.
[202] Yangtuo Peng and Hui Jiang. Leverage financial news to predict stock price movements using word
embeddings and deep neural networks. In Proceedings of the 2016 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language Technologies. Association
for Computational Linguistics, 2016.
[203] Huy D. Huynh, L. Minh Dang, and Duc Duong. A new model for stock price movements prediction
using deep neural network. In Proceedings of the Eighth International Symposium on Information and
Communication Technology - SoICT 2017. ACM Press, 2017.
[204] L. Minh Dang, Abolghasem Sadeghi-Niaraki, Huy D. Huynh, Kyungbok Min, and Hyeonjoon Moon.
Deep learning approach for short-term stock trends prediction based on two-stream gated recurrent
unit network. IEEE Access, pages 1–1, 2018.
[205] Ishan Verma, Lipika Dey, and Hardik Meisheri. Detecting, quantifying and accessing impact of news
events on indian stock indices. In Proceedings of the International Conference on Web Intelligence -
WI17. ACM Press, 2017.
[206] Leonardo dos Santos Pinheiro and Mark Dras. Stock market prediction with deep learning: A
character-based neural language model for event-based trading. In Proceedings of the Australasian
Language Technology Association Workshop 2017, pages 6–15, 2017.
[207] Jordan Prosky, Xingyou Song, Andrew Tan, and Michael Zhao. Sentiment predictability for stocks.
CoRR, abs/1712.05785, 2017.
[208] Yang Liu, Qingguo Zeng, Huanrui Yang, and Adrian Carrio. Stock price movement prediction from
financial news with deep learning and knowledge graph embedding. In Knowledge Management and
Acquisition for Intelligent Systems, pages 102–113. Springer International Publishing, 2018.
[209] Akira Yoshihara, Kazuki Fujikawa, Kazuhiro Seki, and Kuniaki Uehara. Predicting stock market
trends by recurrent deep neural networks. In Lecture Notes in Computer Science, pages 759–769.

62
 Springer International Publishing, 2014.
[210] Lei Shi, Zhiyang Teng, Le Wang, Yue Zhang, and Alexander Binder. Deepclue: Visual interpretation
of text-based deep stock prediction. IEEE Transactions on Knowledge and Data Engineering, pages
1–1, 2018.
[211] Xi Zhang, Yunjia Zhang, Senzhang Wang, Yuntao Yao, Binxing Fang, and Philip S. Yu. Improving
stock market prediction via heterogeneous information fusion. Knowledge-Based Systems, 143:236–247,
March 2018.
[212] Ziniu Hu, Weiqing Liu, Jiang Bian, Xuanzhe Liu, and Tie-Yan Liu. Listening to chaotic whispers:
A deep learning framework for news-oriented stock trend prediction. In Proceedings of the Eleventh
ACM International Conference on Web Search and Data Mining, WSDM ’18, pages 261–269, New
York, NY, USA, 2018. ACM.
[213] Qili Wang, Wei Xu, and Han Zheng. Combining the wisdom of crowds and technical analysis for
financial market prediction using deep random subspace ensembles. Neurocomputing, 299:51–61, July
2018.
[214] Takashi Matsubara, Ryo Akita, and Kuniaki Uehara. Stock price prediction by deep neural generative
model of news articles. IEICE Transactions on Information and Systems, E101.D(4):901–908, 2018.
[215] Xiaodong Li, Jingjing Cao, and Zhaoqing Pan. Market impact analysis via deep learned architectures.
Neural Computing and Applications, March 2018.
[216] Avraam Tsantekidis, Nikolaos Passalis, Anastasios Tefas, Juho Kanniainen, Moncef Gabbouj, and
Alexandros Iosifidis. Using deep learning to detect price change indications in financial markets. In
2017 25th European Signal Processing Conference (EUSIPCO). IEEE, August 2017.
[217] Justin Sirignano and Rama Cont. Universal features of price formation in financial markets: Perspec-
tives from deep learning. SSRN Electronic Journal, 2018.
[218] Przemyslaw Buczkowski. Predicting stock trends based on expert recommendations using gru/lstm
neural networks. In Lecture Notes in Computer Science, pages 708–717. Springer International Pub-
lishing, 2017.
[219] Avraam Tsantekidis, Nikolaos Passalis, Anastasios Tefas, Juho Kanniainen, Moncef Gabbouj, and
Alexandros Iosifidis. Forecasting stock prices from the limit order book using convolutional neural
networks. In 2017 IEEE 19th Conference on Business Informatics (CBI). IEEE, July 2017.
[220] Thomas G Thomas Günter Fischer, Christopher Krauss, and Alexander Deinert. Statistical arbitrage
in cryptocurrency markets. Journal of Risk and Financial Management, 12, 2019.
[221] T Gneiting and A E Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the
American Statistical Association, 102(477):359–378, 2007.
[222] Hannah R Rothstein, Alex J Sutton, and Michael Borenstein. Publication bias in meta-analysis –
prevention, assessment and adjustment. John Wiley & Sons, Ltd, 2005.

63

View publication stats