electronics

Article
Predicting Economic Trends and Stock Market Prices with Deep
Learning and Advanced Machine Learning Techniques
Victor Chang 1, * , Qianwen Ariel Xu 1 , Anyamele Chidozie 2 and Hai Wang 3

1 Department of Operations and Information Management, Aston Business School, Aston University,
Birmingham B4 7ET, UK; [email protected]
2 School of Computing, Engineering and Digital Technologies, Teesside University,
Middlesbrough TS1 3BX, UK
3 School of Computer Science and Digital Technologies, Aston University, Birmingham B4 7ET, UK;
[email protected]
* Correspondence: [email protected] or [email protected]

Abstract: The volatile and non-linear nature of stock market data, particularly in the post-pandemic
era, poses significant challenges for accurate financial forecasting. To address these challenges, this
research develops advanced deep learning and machine learning algorithms to predict financial
trends, quantify risks, and forecast stock prices, focusing on the technology sector. Our study seeks to
answer the following question: “Which deep learning and supervised machine learning algorithms
are the most accurate and efficient in predicting economic trends and stock market prices, and
under what conditions do they perform best?” We focus on two advanced recurrent neural network
(RNN) models, long short-term memory (LSTM) and Gated Recurrent Unit (GRU), to evaluate
their efficiency in predicting technology industry stock prices. Additionally, we integrate statistical
methods such as autoregressive integrated moving average (ARIMA) and Facebook Prophet and
machine learning algorithms like Extreme Gradient Boosting (XGBoost) to enhance the robustness
of our predictions. Unlike classical statistical algorithms, LSTM and GRU models can identify and
retain important data sequences, enabling more accurate predictions. Our experimental results show
that the GRU model outperforms the LSTM model in terms of prediction accuracy and training
time across multiple metrics such as RMSE and MAE. This study offers crucial insights into the
Citation: Chang, V.; Xu, Q.A.;
predictive capabilities of deep learning models and advanced machine learning techniques for
Chidozie, A.; Wang, H. Predicting
Economic Trends and Stock Market
financial forecasting, highlighting the potential of GRU and XGBoost for more accurate and efficient
Prices with Deep Learning and stock price prediction in the technology sector.
Advanced Machine Learning
Techniques. Electronics 2024, 13, 3396. Keywords: stock prices; deep learning; artificial neural networks; recurrent neural networks; long
https://doi.org/10.3390/ short-term memory (LSTM); gated recurrent unit (GRU)
electronics13173396

Academic Editor: Simeone Marino

Received: 17 June 2024 1. Introduction

Revised: 15 August 2024
The finance sector is a crucial domain for applying advanced deep learning (DL)
Accepted: 22 August 2024
and machine learning (ML) models due to its dynamic nature and the significant stakes
Published: 26 August 2024
involved in financial decision-making. Accurate financial forecasting in this sector can
lead to substantial economic benefits, reduced risks, and more informed decisions. In the
complex and constantly evolving world of finance, forecasting has been a key focus for
Copyright: © 2024 by the authors.
many researchers over the years. The volatility and unpredictability of the stock market
Licensee MDPI, Basel, Switzerland. present significant challenges for investors. Predicting the future performance of companies
This article is an open access article through financial forecasting is one of the most extensively studied applications in the
distributed under the terms and finance industry. Accurate stock price predictions play a critical role in making profitable
conditions of the Creative Commons investment decisions, although the inherent complexities of the financial market make it a
Attribution (CC BY) license (https:// formidable task.
creativecommons.org/licenses/by/ Companies raise capital by dividing their ownership and selling shares, making stock
4.0/). price prediction a significant financial application. Stock prices fluctuate based on factors

Electronics 2024, 13, 3396. https://doi.org/10.3390/electronics13173396 https://www.mdpi.com/journal/electronics

Electronics 2024, 13, 3396 2 of 27

such as company performance, brand value, market activity, inflation, trends, and investor
sentiment. While some aspects, like sales and purchases, can be estimated, the complexities
add a layer of difficulty to developing accurate models that capture trends and forecast
future prices. Predicting future trends can be the difference between investment success
and failure for investors. Traditional methods, such as technical and fundamental analyses,
have been used to study patterns and predict future stock prices, but they often fall short
when dealing with the dynamic, non-stationary nature of stock markets influenced by
factors like announcement headlines, social media tweets, corporate news, and other mood
indicators [1,2].
Over the years, numerous statistical methods like regressions and time series models
(ARIMA, SARIMA [3], GARCH) have been employed to predict future stock prices. While
beneficial in some respects, these methods struggle with handling stock price data. For
example, the autoregressive integrated moving average (ARIMA) model has been applied
to predict the stock market using historical financial data; however, these statistical models
often fall short due to the non-linear structure of time series data [4].
To overcome the inefficiencies of statistical methods, various Artificial Intelligence (AI)
models have been developed and integrated into statistical analysis to predict future stock
market trends. These include classical machine learning (ML) algorithms such as Support
Vector Machines (SVMs) [5] and Random Forest (RF) [6], as well as deep learning (DL)
algorithms such as recurrent neural networks (RNNs) [7], Convolutional Neural Networks
(CNNs) [8], and other deep learning methods for multivariate time series data analysis. Rao
and Reimherr introduce a novel class of non-linear function-on-function regression models
specifically designed for functional data using neural networks. The authors propose
two model-fitting strategies: Function-on-Function Direct Neural Networks (FFDNNs)
and Function-on-Function Basis Neural Networks (FFBNNs). These strategies are tailored
to leverage the inherent structure of functional data and capture complex relationships
between functional predictors and responses [9]. These AI models, with their capacity to
learn from extensive datasets and continuously improve, offer promising potential for
automated and more accurate future stock price predictions.
Deep learning methods have been extensively used in the existing literature to predict
future stock prices, significantly contributing to improved model accuracy [10]. White
was a pioneer in implementing an artificial neural network (ANN) for financial market
forecasting, using the daily prices of IBM as a database [11]. Although this initial study did
not achieve the expected results, it highlighted several difficulties, such as the overfitting
problem and the low complexity of the neural network, which used only a few entries and
one hidden layer. This study highlighted possible future improvements, including adding
more features to the ANN, working with different forecasting horizons, and evaluating
model profitability. Over the years, deep learning capabilities have greatly improved, and
various parameter tuning methods have been developed to address the issues mentioned
by White [11]. A family of recurrent neural network (RNN) architectures, including vari-
ations of gated recursive units (GRUs) and long and short-term memory (LSTM), have
become popular methods for predicting stock market patterns. Recent studies highlight
the effectiveness of combining sentiment analysis with deep learning models. For instance,
Sonkiya et al. used BERT for sentiment analysis and GANs for stock price prediction, show-
ing improved performance over traditional methods like ARIMA and neural networks
such as LSTM and GRU [12]. Similarly, Maqsood et al. demonstrated that incorporating
sentiment from local and global events into deep learning models enhances prediction
accuracy, as evidenced by improved RMSE and MAE metrics [13]. Another innovative
approach by Patil et al. utilized graph theory to model the stock market as a complex
network. Their hybrid models, which combined graph-based structural information with
deep learning and traditional machine learning techniques, outperformed standard models
by leveraging the spatio-temporal relationships between stocks [14].
Despite these advancements, there is a notable gap in current research. Compara-
tive analyses of LSTM and GRU for predicting stock prices of technology companies are
Electronics 2024, 13, 3396 3 of 27

insufficient. Existing studies often lack the necessary industry specificity, resulting in un-
satisfactory predictions when models trained on general stock market data are applied
to specific industries or companies. This study aims to address this gap by focusing on
the technology industry and applying LSTM and GRU models to enhance the precision
of technology stock forecasts. By comparing these models, we aim to determine the more
effective method for predicting technology sector stock prices, ultimately aiding investors
in making data-driven decisions.
This study uniquely applies LSTM and GRU deep learning models, along with various
machine learning algorithms, to predict stock prices in the technology sector. We aim to
identify the more effective model among them, offering a crucial contribution to financial
forecasting. The objective is to better understand the patterns, trends, and volatility of
the tech stock market and develop an efficient model to bolster the accuracy of tech stock
forecasts, enabling data-driven decision-making for investors.
The remainder of this paper is structured as follows: Section 2 provides a brief intro-
duction to the various computational methods and data analysis techniques utilized in our
study. Section 3 covers the preliminaries, our approach, and illustrative examples. Section 4
presents the numerical results, while Section 5 details additional experiments and valida-
tions. Finally, Section 6 concludes the paper and highlights directions for future research.

2. Theoretical Background
This section outlines the various computational methods and data analysis techniques
employed in our study to predict stock price movements. The theoretical foundation of our
approach relies on both deep learning and traditional machine learning frameworks. Deep
learning is particularly adept at processing and learning from large datasets, making it ideal
for the complex patterns observed in stock market data. Machine learning algorithms like
XGBoost complement deep learning by providing efficient, scalable methods for regression
and classification.

2.1. Review of the LSTM and GRU Architecture

In traditional neural networks, the output of a neuron is rarely used as an input for
the next step. When we focus on a proven oddity, though, we observe that our ultimate
output is often influenced by both external inputs and prior produce. For example, while
reading a book, comprehension of each sentence is based on both the current flow of words
and the context set by previous sentences. Traditional neural networks do not have the
idea of ‘context’ or ‘constancy’.
A simple RNN with an input circle produces a result ht for some information xt at
time step t. It uses two bits of information, xt+1 and ht , to obtain the output ht+1 at the next
time step t + 1. Data may be transmitted from one network step to the next using a circle.
An RNN, on the other hand, is not without impediments. When the setting is from a long
time ago, it helps tremendously to get the intended result. However, RNNs face challenges
when required to rely on distant past information to produce the desired output. This RNN
stumbling block was extensively studied by Hochreiter [15] and Bengio et al. [16], who also
identified the underlying theories to determine why RNNs may not work in the long run.
Fortunately, LSTM models and GRUs are built to address these challenges.
The standard neural network is severely constrained without context-based reasoning.
To overcome this restriction, the concept of recurrent neural networks (RNNs) has been
developed. Figure 1 illustrates a simple RNN with a feedback loop on the left. X denotes
the input layer, and A is the middle layer consisting of multiple hidden layers that receive
X. The figure compares the simple RNN with a feedback loop to its equivalent unrolled
form on the right side. In a time series data sequence, if X0 is the input at the start time, and
h0 is the output, then h0 together with x1 will be the input for the next step, and this process
is repeated for all inputs from different time periods, allowing the network to remember
the context during training. In the next section, we summarize LSTM and GRU networks.
 Electronics 2024, 13, 3396 4 of 27
Electronics 2024, 13, 3396 4 of 27
Electronics 2024, 13, 3396 4 of 27

network
network to
to remember
remember the
the context
context during
during training.
training. In
In the
the next
next section,
section, we
we summarize
summarize
Electronics 2024, 13, 3396
LSTM and
network to GRU networks.
remember the
LSTM and GRU networks. context during training. In the next section, we summarize
4 of 27
LSTM and GRU networks.

Figure
Figure 1.
1. Simple
Simple RNN
RNN with
with an
an input
input circle
circle and
and its
its equivalent
equivalent unrolled
unrolled presentation.
presentation.
Figure 1. Simple RNN with an input circle and its equivalent unrolled presentation.
Figure 1. Simple RNN with an input circle and its equivalent unrolled presentation.
LSTM
LSTMand and GRU
andGRU Networks
GRUNetworks
Networks
LSTM
LSTMHochreiter
and GRU Networks
and Schmidhuber developed an exceptional type of RNN that can learn
Hochreiter and
Hochreiter and Schmidhuber
Schmidhuber developed
developed an an exceptional
exceptional type
type of
of RNN
RNN thatthat can
canlearn
learn
over
over long
long distances.
Hochreiter
distances. Various
and Schmidhuber
Various other
other researchers
developed
researchers later
anlater improved
exceptional
improved this
type leading
ofleading
this RNN that effort [17–19].
can[17–19].
effort learn
over
LSTM long
and distances.
GRUs Various
were other researchers
developed to solve thelater improved
protracted this leading effort [17–19].
dependency
over
LSTM longanddistances.
GRUs Various
were other researchers
developed to solve later
the improved
protracted this leadingproblem.
dependency problem. Sutton
effort [17–19].
Sutton
LSTM and GRUs were developed to solve the protracted dependency problem. Sutton and
and
LSTM
and Barto
and
Barto discussed
GRUs were
discussed the
the evolution
developed
evolution and
to
and refinement
solve the of LSTM
protracted
refinement of LSTM and GRUs
dependency
and GRUs from
problem.
from RNNs
RNNs [20].
Sutton
[20].
Barto discussed the evolution and refinement of LSTM and GRUs from RNNs [20]. RNNs
RNNs
and
RNNsBartoare made up
up of
discussed
areupmade theaa evolution
of series ofof repeating neural
and refinement network
of LSTMmodules.
and GRUs In aa standard
Infrom RNNsRNN, [20].
are made of a series of series
repeating repeating neural modules.
neural network network modules.
In a standard standard
RNN, RNN,
repeating
repeating
RNNs modules
are made
repeating up of
modules contain
a series
contain a simple computational
of repeating
a simple node,
neural network
computational represented
modules. Inby
node, represented by a single
a standard tanh
RNN, ac-
modules contain a simple computational node, represented by a single atanhsingle tanh
activationac-
tivation
repeating
tivation function,
modules
function, as
as shown
contain
shown a in
in Figure
simple
Figure 2.
computational
2. node, represented by a single tanh ac-
function, as shown in Figure 2.
tivation function, as shown in Figure 2.

Figure 2.
Figure2. The
2.The repeating
Therepeating module
repeatingmodule in
inaaastandard
modulein standard RNN
standardRNN contains
containsaaasingle
RNNcontains single layer.
singlelayer.
layer.
Figure
Figure 2. The repeating module in a standard RNN contains a single layer.
LSTM cells
LSTM cells
LSTM can
cells can track
cantrack information
trackinformation
information overover multiple
overmultiple time
multiple time steps.
timesteps. Information
steps. Information
Information is is added
is added
added or or
or
eliminated
eliminated
eliminated through
LSTM through
cells
through structures
canstructures
track
structures called
information gates.
over
calledgates.
called gates. Gates
Gates
Gates naturally
multiple
naturally
naturally allow
time allow
steps. information
Information
allowinformation
information through
isthrough
throughadded via
ora
viavia
aa sigmoid
eliminated
sigmoid
sigmoid neural
neural net net
through
neural layer
net layer
and and
structures
layer and aa pointwise
a called gates.
pointwise multiplication.
Gates naturally
multiplication.
pointwise multiplication. The
Theallow repeating
Theinformation
repeating module
repeating module
through
in an LSTM
module in an
invia
an
LSTM
aisLSTM is shown in Figure 3. LSTM models process the information by first forgetting ir-
sigmoid
shown is shown
inneural
Figure in Figure
net
3. layer
LSTM 3.
andLSTM
a
models models
pointwise
process process the information
multiplication.
the information The
by by
repeating
first first forgetting
module
forgetting in
irrelevant an
ir-
relevant
LSTM
parts of
relevant parts
is the
shown
parts of
ofinthe
previous
the previous
Figure 3. then
state,
previous state,
LSTM then
models
storing
state, storing
thenthe mostthe
process
storing the most
the relevant
parts ofparts
information
relevant
most relevant by of
of the
thefirst
parts new new
new infor-
forgetting
information
the ir-
infor-
mation
relevant
to to
the state
mation to the
parts
the state
of of
the the
state of
cell,
of the
previouscell,
thirdly
the thirdly
cell,state,
updating
thirdly updating
then storing
their
updating their
internal
their internal
the moststatus,
internal status,
relevant
and
status, and
parts
then
and ofthen
the new
finally
then finally pro-
infor-
producing
finally pro-
ducing
mation
the the
to
output.
ducing the output.
the state of the cell, thirdly updating their internal status, and then finally pro-
output.
ducing the output.

Figure
Figure 3. The
3.The
Figure3. repeating
Therepeating module
repeatingmodule in
modulein an
inan LSTM.
anLSTM.
LSTM.
Figure 3. The repeating module in an LSTM.
The forget gate in an LSTM unit determines which cell state information to exclude
from the model. The memory cell takes the previous instant ht−1 and the current input
information xt and transforms them into a long vector (ht−1 , xt ) to become
 
f t = σ W f ·[ht−1 , xt ] + b f (1)
 The forget gate in an LSTM unit determines which cell state information to exclude
from the model. The memory cell takes the previous instant ℎ and the current input
information 𝑥 and transforms them into a long vector (ℎ , 𝑥 ) to become

Electronics 2024, 13, 3396

𝑓 = 𝜎 𝑊 . [ℎ ,𝑥 ] + 𝑏 (1)
5 of 27
where 𝑊 is the weight matrix associated with the forget gate, 𝑏 is the bias term, and
σ is the sigmoid activation function. To determine how much of the current inputs 𝑥
where W is the weight matrix associated with the forget gate, b f is the bias term, and σ is
should bef allocated to the cell 𝐶 , an input gate is utilized, preventing non-essential in-
the sigmoid activation function. To determine how much of the current inputs xt should be
formation from accessing the memory cells:
allocated to the cell Ct , an input gate is utilized, preventing non-essential information from
accessing the memory cells: 𝑖 = 𝜎 (𝑊 . [ℎ , 𝑥 ] + 𝑏 ) (2)
it = σ (Wi ·[ht−1 , xt ] + bi ) (2)
𝐶 = tanh(𝑊 . [ℎ , 𝑥 ] + 𝑏 ) (3)
Ct = tan h(Wc ·[ht−1 , xt ] + bc ) (3)
where Wi and 𝑊 are the weight matrices for the input gate and candidate cell state, re-
where W and Wc are the weight matrices for the input gate and candidate cell state, respec-
spectively,i and 𝑏 and 𝑏 are their respective bias terms. The function tanh is the hyper-
tively, and bi and bc are their respective bias terms. The function tan h is the hyperbolic
bolic tangent activation function.
tangent activation function.
= f𝑓t ∗∗C𝐶t−1 +
C𝐶t = + i𝑖t ∗∗C𝐶t (4)
(4)
Theoutput
The outputgate
gatedetermines
determines how much of of the
thecurrent
currentcell
cellstate
stateisisincluded
includedininthe output.
the out-
TheThe
put. sigmoid layer
sigmoid processes
layer the the
processes output information
output informationfirst, followed
first, followed by by
thethe
tanh
tanhfunction,
func-
and and
tion, thenthen
multiplies it byitthe
multiplies by sigmoid layer
the sigmoid output
layer to get
output to the finalfinal
get the output component:
output component:

o𝑜t =
= σ𝜎((𝑊
Wo ·[. [ℎ
ht−1 ,, x𝑥t ]]+
+ b𝑏o ) ) (5)
(5)
where 𝑊 is the weight matrix for the output gate and 𝑏 is the bias term.
where Wo is the weight matrix for the output gate and bo is the bias term.
The final output value of the cell is defined as
The final output value of the cell is defined as
ℎ = 𝑜 ∗ tanh(𝐶 ) (6)
ht = ot ∗ tanh(Ct ) (6)
Cho created the Gated Recurrent Unit (GRU), a kind of RNN, in 2014 with the pur-
pose of fixing
Cho the vanishing
created the Gated gradient
Recurrentissue
Unitof(GRU),
RNNsa[21].
kindThe GRU in
of RNN, s key benefit
2014 over
with the other
purpose
structures is that
of fixing the it requires
vanishing fewerissue
gradient parameters,
of RNNstrains
[21]. quicker,
The GRU’sandkey
requires less
benefit data
over to
other
generalize.
structures The structure
is that of thefewer
it requires GRUparameters,
model is shown
trainsin quicker,
Figure 4.and requires less data to
generalize. The structure of the GRU model is shown in Figure 4.

Figure
Figure4.4.The
Theinternal
internalstructure
structureofofthe
theGRU
GRUmodel.
model.

Theupdate
The update and reset
reset gates
gatesproduce
produceintermediate
intermediate values
values 𝑧 and
zt and 𝑟 , respectively,
rt , respectively, while
the final
while the memory of the of
final memory general-purpose unit stores
the general-purpose unit the result
stores t [21]. h
thehresult The[21].
update
The specifies
update
the amount
specifies of prior of
the amount priorxtinput
input 𝑥 andhoutput
and output t−1 thatℎshould thatbe conveyed
should to the next
be conveyed cell,
to the
governed by the weight W
next cell, governed by the weightt . The 𝑊 . The reset gate determines how much data should
reset gate determines how much data should be erased
from
be memory.
erased from memory.
Thefollowing
The followingarearethe
the
most most essential
essential equations
equations thatthat characterize
characterize the operation
the operation of theof
the GRU:
GRU:
zt = σ(Wz ·[ht−1 , xt ]) (7)
rt = σ(Wr ·[ht−1 , xt ]) (8)
ht = tanh(Wh ·[rt ∗ ht−1 , xt ])
e (9)
ht = (1 − zt )∗ ht−1 + zt ∗ e
ht (10)
Electronics 2024, 13, 3396 6 of 27

where Wz , Wr , and Wh are the weight matrices for the update gate, reset gate, and candidate
activation, respectively. The operator · denotes matrix multiplication, while ∗ denotes
element-wise multiplication. The functions σ and tanh are the sigmoid and hyperbolic
tangent activation functions, respectively.
In this paper, we will use deep learning (DL) models to analyze selected technological
stock patterns as one-dimensional time series and attempt to forecast future stock prices by
examining past historical prices and the most critical technical indicators. This research
will compare the performance of the LSTM and the GRU ensemble models on selected
technology stock data to investigate stock price patterns.

2.2. Attention Mechanism

The attention mechanism has recently gained much traction in the context of time
series data. Self-attention, global attention, and local attention are examples of attention
approaches. In general, applications such as voice recognition, machine translation, and
part of speech tagging benefit greatly from the attention mechanism.
The concentrated attention is focused on a single element in the input, which is
picked information by maximal or random sampling, and it requires further training to get
exceptional outcomes. On the other hand, soft attention is a process that assigns weights to
all of the information to allow more effective information utilization. In the soft attention
mechanism, the attention score at time t (et) is computed using a weight matrix Wa and a
bias term b, acting on the input elements x1 , x2 , . . . , x T :

et = tanh(Wa { x1 , x2 , . . . , x T } + b) (11)

These scores are then normalized using the softmax function to produce the attention
weights (at ):
exp(et )
at = T (12)
∑k=1 exp(ek )
The attention mechanism generally involves two steps: the first phase involves cal-
culating the attention distribution, and the second step involves computing the weighted
average of the incoming information using the attention distribution as a guide. The process
is initiated with the attention scoring function S, passing the result to the softmax layer,
generating the attention weights (1, 2, . . . , n)· Following that, the softmax layer is handed
Electronics 2024, 13, 3396 7 of 27
the attention weights 1, 2, . . . Finally, the attention weight vector is weighted and averaged
against the input data to arrive at the final result. The attention process is shown in Figure 5.

Figure5.
Figure 5. The
The basic
basic structure
structureof
ofthe
theattention
attentionmodel.
model.

2.3. Time Series Forecasting Methods

Time series forecasting is a critical aspect of data analysis and prediction, particularly
when dealing with sequential data points recorded at regular intervals. Three methods
are introduced as follows.

2.3.1. Autoregressive Integrated Moving Averages (ARIMAs)

Electronics 2024, 13, 3396 7 of 27

2.3. Time Series Forecasting Methods

Time series forecasting is a critical aspect of data analysis and prediction, particularly
when dealing with sequential data points recorded at regular intervals. Three methods are
introduced as follows.

2.3.1. Autoregressive Integrated Moving Averages (ARIMAs)

Autoregressive integrated moving average (ARIMA) models are a popular choice
for stock price prediction due to their ability to handle the complex and dynamic nature
of financial time series data [22]. Stock prices often exhibit non-stationary behavior, and
ARIMA models excel at differencing the data to achieve stationarity, making them suitable
for modeling. Moreover, these models incorporate autoregressive and moving average
components, allowing them to capture dependencies on past stock prices and the impact of
past shocks, both of which are crucial factors in stock price movements. ARIMA models also
offer parameter tuning flexibility, making them adaptable to specific stock price datasets.
Their interpretability further aids in understanding the driving factors behind stock price
predictions. As a baseline model, ARIMA provides a solid foundation for assessing the
performance of more advanced forecasting techniques in our project, making it a valuable
choice for stock price prediction tasks [1,4].
The ARIMA model works by first differencing the time series data to achieve sta-
tionarity, removing trends and seasonality. Then, it utilizes autoregressive (AR) terms to
model the relationship between current and past values and moving average (MA) terms
to account for the impact of past shocks or white noise. The model’s order, represented
as (p, d, q), determines the number of AR and MA terms and the degree of differencing
needed. The ARIMA model estimates these parameters and fits the model to the data.
During forecasting, it uses past observations and model parameters to make predictions for
future data points. We employed the auto-ARIMA module from the ‘pmdarima’ package
for our analysis, leveraging its automatic selection of the optimal p, d, and q terms for
our time series model. This approach ensured that we obtained the best possible results,
streamlining the modeling process and enhancing forecast accuracy.

ŷt = µ + ϕ1 yt−1 + . . . + ϕp yt−p − θ1 et−1 − . . . − θq et−q (13)

where ŷt represents the forecasted value, µ is the mean term, ϕ1 , . . ., ϕp are the autoregres-
sive coefficients, yt−1 ,. . .,yt−p are the lagged values of the series, θ 1 , . . ., θ q are the moving
average coefficients, and et−1 , . . ., et−q are the lagged forecast errors.

2.3.2. XGBoost (Extreme Gradient Boost)

Extreme Gradient Boosting (XGBoost) is a powerful machine learning algorithm
renowned for its accuracy and robustness in predictive modeling [23]. In our stock price
prediction, XGBoost is a compelling choice for several reasons. First, it can handle complex,
non-linear relationships in financial time series data, making it well suited for capturing
intricate patterns in stock prices. Second, XGBoost can handle missing data, an occasional
issue in financial datasets, through its built-in handling mechanisms. Finally, XGBoost
offers flexibility in parameter tuning, enabling us to fine-tune the model’s performance for
our specific dataset. The XGBoost model works by building an ensemble of decision trees,
where each tree corrects the errors of the previous one. These trees are combined into a
strong predictive model. The algorithm assigns a weight to each tree and uses a gradient
descent optimization process to minimize the prediction errors. The final prediction is a sum
of predictions from all the trees [4,23]. Through this ensemble approach, XGBoost leverages
the strengths of multiple decision trees to provide accurate and reliable predictions, making
it a valuable asset in our stock price prediction project.
Electronics 2024, 13, 3396 8 of 27

2.3.3. Facebook Prophet

The Facebook Prophet algorithm is an open-source time series data prediction tool
developed by Facebook using the additive regression model. It is robust in identifying
the components of time series data like trend and seasonality and forecasting values by
combining them. It accepts only two columns (‘ds’ for date and ‘y’ for values) as the
input dataset. Implementing Facebook Prophet does not require an in-depth prerequisite
knowledge of time series data. It provides generalized parameters and automatically
uncovers seasonal movements. The performance of FB Prophet may vary based on the
dataset as it depends on seasonality and trends. Ref. [24] highlights the importance of
timing in enhancing forecasting accuracy, which is accomplished with the use of the
prophet algorithm. This journal uses the Facebook Prophet library to define three different
hyperparameters, namely seasonality, trend, and holidays.
Having established the theoretical foundations and the rationale behind the selection
of our computational models, the next section delves into the preliminary considerations
and the detailed development of our methodological approach. This includes data col-
lection, preprocessing techniques, model training, and evaluation metrics, providing a
comprehensive overview of how we operationalize the theoretical insights discussed here.

2.4. Other Recent Advancements in the Area

Yun et al. [25] improve stock price prediction by using genetic algorithms to optimize
feature subset selection. The authors maximize feature subset selection by combining ge-
netic algorithms with machine learning regressions, improving the interpretability and
precision of stock price predictions. This method stands out in particular for how well it
strikes a balance between interpretability and model complexity. There are a few restrictions
on the study, though. The arbitrary selection of external factors and technological indica-
tors may have impacted the accuracy of the prediction. Furthermore, the study does not
completely account for the social environment of stock market dynamics, which includes
market news and public opinion, and it lacks clear criteria for feature selection.
The application of Bi-Directional Long Short-Term Memory (Bi-LSTM) networks for
stock price prediction is examined by the authors in [26]. This method has the advantage
of analyzing data sequences both forward and backward, which may highlight patterns
and trends that conventional models would miss. The outcomes demonstrate that Bi-LSTM
models have the potential to outperform conventional LSTM models, particularly when
managing the volatility of stock market data. The authors do, however, also note that it is
possible that the Bi-LSTM model will not be able to adequately capture the intricate and
erratic character of market movements. They contend that in order to make the model more
reliable and strong for everyday application, more testing and modification are required.
In their publication, Zhao and Yang [27] provide a thorough method for predicting
changes in stock prices through the integration of many deep learning models. In order
to take advantage of the temporal and geographical characteristics of financial data, the
authors suggest a framework that blends many neural network designs, including CNNs
and LSTM models. The goal of this integrated strategy is to increase prediction accuracy
by identifying the intricate relationships that influence changes in stock prices. The study
shows that when it comes to stock price direction prediction, the integrated framework
performs better than conventional machine learning models. Though the study’s findings
are encouraging, it also draws attention to issues with computational complexity and the
requirement for huge datasets in order to properly train these deep learning models. The
practical use of the framework may be limited, particularly for smaller enterprises or indi-
vidual investors, due to its heavy reliance on large amounts of data and computer resources.
In the study [28], a multi-layered long short-term memory (LSTM) network is used
to present a novel technique for enhancing stock price prediction. In order to improve
the LSTM architecture’s capacity to identify long-term dependencies in financial time
series data, the authors concentrate on optimizing it. The study successfully tackles the
difficulties of predicting stock prices, which are essentially volatile and non-linear, by
Electronics 2024, 13, 3396 9 of 27

employing a sequential LSTM approach. The findings show that our improved LSTM
model outperforms more conventional approaches in terms of prediction, especially when
it comes to identifying the complex patterns of stock price fluctuations. The model’s reliance
on substantial computational resources and extensive, high-quality datasets is one of the
study’s acknowledged potential shortcomings.
A dedicated recurrent neural network (RNN) architecture designed for time series
data is introduced by Lu and Xu [29] with a focus on stock price prediction. The temporal
relationships and non-linear patterns of financial data, among other difficulties, are well
handled by the TRNN model. The TRNN outperforms conventional RNN architectures
in terms of predicted accuracy and processing overhead by incorporating particular ap-
proaches. This paper makes a strong argument for the use of sophisticated RNN models
in stock price prediction by highlighting the significance of customized neural network
designs in financial forecasting.

3. Preliminary Considerations and Development of the Approach

In this section, we lay the groundwork for our stock price prediction models, detailing
the data collection, preprocessing, and model evaluation methodologies.

3.1. Historical Context and Progression

Stock price prediction is one of the most challenging applications in financial stud-
ies due to the complex nature of stock price time series. Numerous factors, including
historical time series records, key technical indicators, macroeconomic variables, and in-
vestor sentiment, influence stock prices, leading to non-stationarity and non-linearity in
the data. Artificial neural networks (ANNs) and, particularly, deep learning (DL) meth-
ods can be advantageous in predicting future stock prices and aid investors in reducing
investment risk.
The pioneering study in applying ANNs for forecasting stock prices dates back to
White’s work in 1988, where he developed a standard feedforward single hidden layer
architecture to predict IBM’s stock prices [11]. Although this study had drawbacks, such as
the overfitting problem, it opened avenues for more advanced models, such as recurrent
neural networks (RNNs).
In machine learning models, input data points are transformed into outputs through a
learning process derived from exposure to existing input–output pairs. The main step in
ML or DL is to transform the data meaningfully. In ANNs, the learning process is done
by building a set of layers where information is fed to the first layer (input layer) and
passes through subsequent layers until purified information is produced. The depth of the
network refers to the number of layers contributing to the structure of the model. Deep
learning occurs when the number of layers is substantial.
Feedforward Neural Networks and recurrent neural networks are two main types of
neural networks. While the former involves information flowing from the input layer to
the output layer, the latter includes at least one cyclic path of synaptic connections. The
neurons in RNNs not only use the inputs to the neuron, but also use the outputs from
the previous time steps. Hence, RNNs are suitable for sequential data such as time series.
Long short-term memory (LSTM) and Gated Recurrent Units (GRUs) are two types of
RNNs designed to address the vanishing gradient problem during network training by the
backpropagation algorithm through time.

3.2. Data Collection, Exploration, and Preparation

Stock market data can be fascinating to study, and excellent predictive models can
result in significant financial gains. Finding a large, well-structured dataset on a diverse set
of companies can be challenging despite the seemingly limitless availability of financial data
on the internet. The dataset for this study is accessed from the API of Yahoo Finance, which
is often used as a reliable source of financial data. Yahoo Finance provides a comprehensive
collection of financial data that includes stock prices, indices, ETFs, mutual funds, bonds,
Electronics 2024, 13, 3396 10 of 27

and options worldwide. In addition, it offers extensive historical data, some going back
many decades. This is especially useful for long-term financial analysis and historical
research. We utilized stock data from Apple (AAPL), Amazon (AMZN), Google (GOOG),
and Microsoft (MSFT) from the past ten years, sourced from the Yahoo Finance database.
These stocks were chosen to leverage the findings of this study in building effective price
forecasting algorithms to aid investment decisions. Exploratory data analysis (EDA) will
be employed to gain a better understanding of the basic characteristics and nature of the
collected dataset, including data visualization.

3.2.1. Train and Test Split

The dataset will be split into a training set and a test set in an 80:20 ratio; the training
set will be used to train the models, while the test set will be used to evaluate their
performance. The 80/20 split is commonly used because it often provides a good balance,
allowing the model to learn from a large portion of the data while reserving enough data
for testing. The data were split into training and testing sets using k-fold cross-validation to
guarantee the models’ robustness and generalizability. This method offers a more accurate
approximation of the model’s performance on unknown data while also assisting in the
reduction of overfitting.

3.2.2. Data Shaping for LSTM and GRU Models

LSTM and GRU models require data structured into time steps or look-back periods.
For this study, both the training and testing datasets are structured with a 60-day look-back
period (60 time steps). Consequently, the models will use the last 60 days of data to predict
current or future stock prices.

3.3. Preprocessing and Normalization

Normalization is a common approach for preparing data for machine learning, often
included as part of data cleansing. The primary goal of normalization is to scale all
attributes consistently. This makes it easier to discuss the performance and training stability
of the model. When the ranges of the features differ, normalization is necessary. There are
several approaches to normalization, also known as rescaling, including the following:
(i) Min-Max normalization: this technique scales a feature to fit within a specific range,
usually 0 to 1, according to the following formula:

x − min( x )
x′ = (14)
max ( x ) − min( x )

where x is the original value, and min(x) and max(x) are the minimum and maximum
values in the dataset, respectively.
(ii) Mean normalization: this method adjusts the data based on the mean and can be
computed as the following:

x − average ( x )
x′ = (15)
max( x ) − min( x )

where the mean of the dataset (average(x)) is used.

(iii) Z-score normalization: also known as standardization, this approach uses the Z-
score or standard score and is often utilized in machine learning algorithms like
Support Vector Machines (SVMs) and logistic regression. It can be calculated using
the following formula:
x−µ
z= (16)
σ
where µ is the mean and σ is the standard deviation of the dataset.
Given the wide range and high volatility of the volume and turnover elements in this
study, we employ Min-Max normalization to scale all attributes between 0 and 1.
Electronics 2024, 13, 3396 11 of 27

3.4. Model Evaluation

The model’s performance will be assessed using the Mean Absolute Error (MAE),
Root Mean Square Error (RMSE), Mean Directional Accuracy (MDA), and the coefficient
of determination (R2 ) [30]. As the MAE and RMSE values decrease, it becomes easier to
predict how close the predicted value will be to the actual value. The model’s fit is expected
to be better as the coefficient of determination (R2 ) approaches one. Mean Directional
Accuracy (MDA) is generally used to evaluate the model’s ability to predict the direction of
change rather than the magnitude of the forecasting error. The formula for RMSE, MAE,
R2 , and MDA are shown below.
r
1 N
N ∑ i =1 i
RMSE = (y − ŷi )2 (17)

where yi and ŷi are the actual and forecasted values, respectively, and N is the total number
of observations.
1 N
N ∑ J =1 j
MAE = | x − x̂ j | (18)

where x j and x̂ j are the actual and forecasted values at time j.

2
1
∑N
i=1 (yi − ŷi )
R2 = N
2
(19)
1
N ∑N
i=1 (yi − y i )

where y is the mean of the actual values.

1 N
∑ j=1 1sign


MAD = x j − x j −1 (20)
N
where N is the total number of observations (trading days), and x j and x j−1 are the actual
and forecast values, respectively.
The methodology framework is developed as follows (Figure 6). In summary, we start
with data collection, gathering historical stock price data and financial indicators from Ya-
hoo Finance for companies such as Apple, Amazon, Google, and Microsoft. This is followed
by data exploration, where we perform exploratory data analysis (EDA) to understand the
dataset’s characteristics and trends. Next, we prepare the data by splitting it into training
(80%) and testing (20%) sets, employing k-fold cross-validation to ensure robustness. Pre-
processing and normalization are then applied, using techniques like Min-Max, Mean, and
Z-score normalization to make the data suitable for model training. For model construction,
we develop long short-term memory (LSTM) and Gated Recurrent Unit (GRU) models, as
well as XGBoost and Facebook Prophet, for machine learning approaches to predict future
stock prices. The models’ performance is evaluated using metrics such as Mean Absolute
Error (MAE), Root Mean Square Error (RMSE), Mean Directional Accuracy (MDA), and the
coefficient of determination (R2 ) to assess accuracy and effectiveness. Finally, we conduct
a risk–return tradeoff analysis to examine the predicted stock prices in terms of risk and
return, aiding investment decisions. This comprehensive and systematic approach ensures
the development and evaluation of effective stock price prediction models, enhancing the
accuracy of financial forecasts and supporting informed investment choices.
We now proceed to evaluate its performance through comprehensive numerical results
and analyses in the next sections.

3.5. The Architectural Diagram

The architectural diagram for processing and analyzing data is presented in Figure 7
with the explanations as follows.
 stockWeprices
now inproceed to risk
terms of evaluate its performance
and return, through comprehensive
aiding investment numerical re-
decisions. This comprehensive
sults and analyses in the next sections.
and systematic approach ensures the development and evaluation of effective stock price
prediction models, enhancing the accuracy of financial forecasts and supporting informed
investment choices.
Electronics 2024, 13, 3396 We now proceed to evaluate its performance through comprehensive numerical 12 of re-
27
sults and analyses in the next sections.

Figure 6. The methodology framework for this research.

3.5. The Architectural Diagram

The architectural diagram for processing and analyzing data is presented in Figure 7
with the6.6.explanations
Figure
Figure The as follows.
Themethodology
methodology frameworkfor
framework forthis
thisresearch.
research.

3.5. The Architectural Diagram

The architectural diagram for processing and analyzing data is presented in Figure 7
with the explanations as follows.

Thearchitecture
Figure7.7.The
Figure architecturediagram
diagramfor
forprocessing
processingand
andanalyzing
analyzingdata.
data.

Input Layer: Time series data, such as stock prices over a given look-back period
Input Layer: Time series data, such as stock prices over a given look-back period (e.g.,
(e.g., 60 time steps), make up the input data. A minimum of 4 GB memory is expected.
60 time steps), make up the input data. A minimum of 4 GB memory is expected.
Layer of LSTM (128 Units): With 128 units, the LSTM layer is the first hidden layer.
Layer
Figure
of architecture
7. The
LSTM (128 diagram
Units): With
for
128 units,
processing
the
and
LSTM layer
analyzing
is the first hidden layer.
data.time steps, which allows it
The LSTM layer keeps a memory of prior inputs over numerous
The LSTM layer keeps a memory of prior inputs over numerous time steps, which allows
to identify long-term dependencies in the time series data. This aids in seeing patterns that
it to identify long-term dependencies in the time prices
series data.aThis aids in seeing patterns
mightInput
not beLayer:
obviousTime at series data,
first but aresuch as stock
essential for preciseover given
forecasting. look-back period (e.g.,
that might
60 time not
steps), be obvious at first but are essential for precise forecasting.
GRU Layermake up the After
(128 Units): inputthat,
data.theA minimum
output of the of 4LSTM
GB memory is expected.
layer is sent to a GRU layer,
GRU Layer (128 Units): AfterWiththat,128 theunits,
output of the LSTM layerislayerthe is sent to a GRU
whichLayer
has 128of LSTM
units as(128
well.Units):
Similar to the LSTM, the
the LSTM
GRU layer ismore first hidden
computationallylayer.
layer,
The which
LSTM has
layer 128 units
keeps a as well.
memory Similar
of prior to the
inputs LSTM,
over the GRU
numerous
efficient overall yet may still capture dependencies across time. By combining the benefits layer
time is more
steps, computa-
which allows
tionally
it both
of efficient
to identify
recurrent overall
long-term yet
unit types, may still
dependencies
LSTM and capture dependencies
in thelayers
GRU time improve
series data.across
the This time.
model’s Byincombining
aidscapacity
seeing the
patterns
to identify
benefits of
that might
intricate both recurrent
not be patterns
temporal unit
obvious at types,
in first LSTM and GRU layers improve
but are essential for precise forecasting.
the data. the model s capacity
to identify
GRUintricate
Dense Layer
Layer (128temporal
(64 Units):patterns
Units): After
The in the
that,
following thedata.
output
layer of theofLSTM
consists 64 unitslayerandis is
sent to a(com-
dense GRU
layer, which
pletely has 128
connected). In units
order asto well.
createSimilar
a moreto the LSTM,
condensed the GRU layer
representation is more
that will becomputa-
utilized
tionally
to produce efficient overall
the final yet may
prediction, still
this capture
layer dependencies
processes the output across
fromtime. By combining
the GRU layer. the
Dense Layer (32 Units): The features retrieved by the earlier
benefits of both recurrent unit types, LSTM and GRU layers improve the model s capacity layers are further refined
by
to aidentify
secondintricate
dense layer with 32
temporal units, which
patterns reduces the amount of data that will go into
in the data.
the final forecast.
Layer of Output: The prediction, usually the stock price for the following time step or
day, is provided by the output layer, which is the last layer. This layer has a linear activation
function, which is common for regression tasks like stock price prediction and includes a
single unit for predicting a single value.

4. Numerical Results
4.1. Data Preprocessing and Exploratory Analysis
This study collected daily historical stock datasets for Apple, Google, Microsoft, and
Amazon stocks using the API of Yahoo Finance. The selected stocks are from international
public companies traded at both NASDAQ and the NYSE. The time series data range from
Electronics 2024, 13, 3396 13 of 27

1 January 2013 to 30 March 2022, encompassing 3775 trading days. The daily time series data
was downloaded automatically using Python’s connection to the Yahoo Finance API. Daily
open price, daily highest price, daily lowest price, daily close price, daily adjusted closing
price, and daily trading volume are all included in the dataset. Table 1 below presents the
description of the features provided in the datasets downloaded from Yahoo Finance.

Table 1. Description of the features provided in the datasets.

FEATURE DESCRIPTION
OPEN PRICE The price at which a stock was initially traded at the start of a trading day.
CLOSE PRICE The last price of a stock in the last transaction on a given trading day.
HIGH PRICE The highest price at which a stock traded on a specified trading day.
LOW PRICE The lowest price at which a stock traded on a specified trading day.
ADJUSTED CLOSE PRICE Adjusted close price based on the reflection of dividends and splits.
TRADING VOLUME A total number of shares/contracts traded on a given trading day.

The closing data were used to compute the daily returns for each technological stock
used to train the models. The most straightforward and obvious way to understand the
stock trend is through the characteristics of the stock price. Compared to the absolute value
of stock prices, price trend returns are more effective in stock forecasting. Different stocks
have different base prices, leading to large variations in absolute stock price values. Using
Electronics 2024, 13, 3396 14 of 27
daily returns reduces the prediction’s sensitivity to the price base.
For the training and testing of the model, the data were split into training and test
sets, with 80% of the total data used for training the model and the remaining 20% used
standard deviations of the features, especially for the adjusted closing price and trading
for testing.
volume, indicate
Figure thatshows
8 below the stock prices and
the closing trading
price volumes
line chart for theofselected
these companies werestocks,
technological more
volatile during
providing theoverview
a quick reportingofperiod.
the collected data.

Figure 8. The
Figure 8. The closing
closing price
price of
of the
the technological stock prices
technological stock prices selected.
selected.

4.2. Hyperparameter Selection

In order to develop Process
a better understanding of the technological stock price data used
in this study,
The summary
process statistics
of selecting for all
the best the features
collection were computedfor
of hyperparameters anda are presented
model in
is known
Appendix A.
as hyperparameter tuning. Variables that can be adjusted during this optimization process
Based
include the on the summary
number of units,statistics provided
batch size, inrate,
learning the tables above, for
and dropout all four companies,
rate.
all features appear to be right-skewed, as the means are consistently higher than the
1. Units: The optimization strategy sets the number of units in each LSTM and GRU
medians, suggesting an upward trend in stock prices over time. In addition, the high
model to 128 and 64 in the first and second layers, respectively.
standard deviations of the features, especially for the adjusted closing price and trading
2. Batch
volume, size: For
indicate tuning
that the model,
the stock the trading
prices and batch size is set toof1.these companies were more
volumes
3. Learning rate: The learning
volatile during the reporting period. rate of the Adam optimizer is set at 0.1.
4. Dropout layer: During model training, it is common to observe a pattern where the
model performs well on the training data but fails to replicate this success on the
testing and validation data. This discrepancy, often due to overfitting, is a major con-
cern, especially in deep learning models that require a substantial amount of data for
training. Dropout is a simple but effective regularization strategy used in neural net-
works to mitigate this overfitting problem. The cells of the recurrent neural network
Electronics 2024, 13, 3396 14 of 27

4.2. Hyperparameter Selection Process

The process of selecting the best collection of hyperparameters for a model is known
as hyperparameter tuning. Variables that can be adjusted during this optimization process
include the number of units, batch size, learning rate, and dropout rate.
1. Units: The optimization strategy sets the number of units in each LSTM and GRU
model to 128 and 64 in the first and second layers, respectively.
2. Batch size: For tuning the model, the batch size is set to 1.
3. Learning rate: The learning rate of the Adam optimizer is set at 0.1.
4. Dropout layer: During model training, it is common to observe a pattern where the
model performs well on the training data but fails to replicate this success on the
testing and validation data. This discrepancy, often due to overfitting, is a major
concern, especially in deep learning models that require a substantial amount of
data for training. Dropout is a simple but effective regularization strategy used in
neural networks to mitigate this overfitting problem. The cells of the recurrent neural
network are dropped at random. The dropout rate is around 0.2.
The model’s training may be excessive or insufficient. Early stopping criteria are often
used to prevent complications caused by having too many or too few epochs. These criteria
allow for the creation of a large number of training epochs and then stopping the training
when the model’s parameters no longer improve on the validation set.
The full specification of the parameters used to train the model is provided in Table 2.

Table 2. Model Training Parameter Specifications.

PARAMETER VALUES
NODES WITHIN INPUT LAYER look-back period × input features
STEPS 2026 with early stopping criteria of the patience of 1 epoch
BATCH SIZE 1
HIDDEN LAYER 1 LSTM/GRU layer with 128 units
DROP OUT LAYER 0.2 dropout rate
LOOK-BACK PERIOD 60
OUTPUT LAYER 1

4.3. Results of the Models

This section contains the performance of the deep learning models (LSTM and GRU)
for each of the stock prices considered, as well as the technical indicators (Table 3). The
Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE),
coefficient of determination (R2 ), and Mean Absolute Difference (MAD) are used to evaluate
the performance of the models. In addition, the performance of the models is compared
to the other deep learning and traditional forecasting methods reported in the literature
(Table 4). Our study showed that the GRU model generally outperformed the LSTM model
across multiple metrics such as RMSE and MAE, with the GRU achieving an RMSE of 3.43
and an MAE of 6.53 for Apple stock, which is notably lower than the LSTM’s RMSE of 9.15
and MAE of 7.81. When compared to other models from existing studies, such as the S-GAN
model, which reported an RMSE of 1.83 on Apple stock, our models still indicate room
for improvement. The consideration of investigator sentiment enhanced the prediction
capability of S-GAN [12]. Additionally, the ARIMA model from the literature showed
an RMSE of 18.25, indicating that our GRU model offers significant advancements over
traditional methods [12]. However, the performance of the LSTM model is less competitive
than other LSTM models in the literature [31,32]. These comparisons highlight that while
our GRU model is competitive, especially with respect to more traditional approaches,
there is potential for further performance enhancement by integrating more advanced
techniques and conducting extensive hyperparameter optimization.
Electronics 2024, 13, 3396 15 of 27

Table 3. Model performance for stocks.

TRAINING TIME
MODELS RMSE MAE R2 MAD STOCK
(SECS)
LSTM 9.1463 7.8058 0.7609 6.0689 54 Apple
GRU 3.4273 6.5298 0.8229 6.7607 51 Apple
LSTM 103.5552 61.7088 0.4679 107.5323 57 Google
GRU 67.4582 35.1966 0.7742 115.7959 52 Google
LSTM 32.3734 31.045 0.6998 13.8247 63 Microsoft
GRU 8.0805 5.2005 0.8319 6.6803 61 Microsoft
LSTM 116.948 5.8673 0.7479 191.7697 64 Amazon
GRU 82.9599 3.8673 0.8731 182.9617 61 Amazon

Table 4. Comparison of Model Performance Against State-of-the-Art Methods Reported in the Literature.

MODELS RMSE STOCK

LSTM 9.1463 Apple
GRU 3.4273 Apple
LSTM 103.5552 Google
GRU 67.4582 Google
LSTM 32.3734 Microsoft
GRU 8.0805 Microsoft
LSTM 116.948 Amazon
GRU 82.9599 Amazon
Apple (AAPL), Google (GOOG),
LSTM [25] 18.89.
Microsoft (MSFT), and Amazon (AMZN)
S-GAN [26] 1.827 APPLE
ARMIA [26] 18.2469 APPLE
Apple (AAPL), Google (GOOG),
LSTM [27] 6.59
Microsoft (MSFT), and Amazon (AMZN)
Apple (AAPL), Google (GOOG),
Ridge [27] 8.72
Microsoft (MSFT), and Amazon (AMZN)
Apple (AAPL), Google (GOOG),
Neural Network [27] 7.91
Microsoft (MSFT), and Amazon (AMZN)

4.3.1. Apple Stock Prediction

Using the daily historical stock datasets for Apple Inc. (Cupertino, CA, USA), along
with the technical indicators, the performance of LSTM and GRU models was assessed with
the Mean Absolute Error (MAE), the Root Mean Square Error (RMSE), the Mean Absolute
Deviation (MAD), and the coefficient of determination (R2 ). The results are presented
in Table 3.
From Table 3, it can be observed that the GRU model forecasts the Apple stock
price more accurately, as the RMSE and MAE values are considerably lower for the GRU
model (3.4273 and 6.5298, respectively) than LSTM model (9.1463 and 7.8058, respectively).
Additionally, the R2 value is higher for the GRU model (0.8229) than for the LSTM model
(0.7609), suggesting a better fit. It should also be observed that the GRU model has a shorter
training time than the LSTM model. Figure 9 below depicts the pattern of the actual closing
prices and predicted closing prices for the LSTM and GRU models.
 more accurately, as the RMSE and MAE values are considerably lower for the GRU model
(3.4273 and 6.5298, respectively) than LSTM model (9.1463 and 7.8058, respectively). Ad-
ditionally, the R2 value is higher for the GRU model (0.8229) than for the LSTM model
(0.7609), suggesting a better fit. It should also be observed that the GRU model has a
Electronics 2024, 13, 3396 shorter training time than the LSTM model. Figure 9 below depicts the pattern of the
16 ac-
of 27
tual closing prices and predicted closing prices for the LSTM and GRU models.

Figure 9. Apple stock: actual and predicted close price, LSTM and GRU models.
Figure 9. Apple stock: actual and predicted close price, LSTM and GRU models.
4.3.2. Google Stock Prediction
Using the daily historical stock datasets for Google Inc. (Mountain View, CA, USA),
along with technical indicators, the performance of the LSTM and GRU models was
assessed using MAE, RMSE, MAD, and R2 .
From Table 3, it can be observed that the GRU model makes more accurate stock
price predictions for Google, with lower RMSE and MAE values (67.4582 and 35.1966,
respectively), compared to the LSTM model (103.5552 and 61.7088, respectively). The R2
value is also higher for the GRU model (0.7742) than for the LSTM model (0.4679). The
GRU model also has a shorter training time. Figure 10 below depicts the pattern of the
actual closing prices and predicted closing prices for the LSTM and GRU models.

4.3.3. Microsoft Stock Prediction

The performance of the LSTM and GRU models for Microsoft stock, as measured by
MAE, RMSE, MAD, and R2 , is summarized in Table 3.
Table 3 indicates that the GRU model forecasts Microsoft stock prices more accurately,
with significantly lower RMSE and MAE values (8.0805 and 5.2005, respectively) compared
to the LSTM model (32.3734 and 31.0450, respectively). The GRU model also shows a
better fit (R2 = 0.8319) and has a shorter training time than the LSTM model. Figure 11
shows the evolution of actual and expected closing prices for the LSTM and GRU models.
Furthermore, the GRU model needs less training time than the LSTM model.

4.3.4. Amazon Stock Prediction

The performance of the LSTM and GRU models for Amazon stock, as measured by
MAE, RMSE, MAD, and R2 , is summarized in Table 3.
From Table 3, it is evident that the GRU model predicts Amazon stock prices more
accurately, with lower RMSE and MAE values (82.9599 and 3.8673, respectively) compared
 4.3.2. Google Stock Prediction
Using the daily historical stock datasets for Google Inc. (Mountain View, CA, USA),
along with technical indicators, the performance of the LSTM and GRU models was as-
Electronics 2024, 13, 3396 sessed using MAE, RMSE, MAD, and R2. 17 of 27
From Table 3, it can be observed that the GRU model makes more accurate stock price
predictions for Google, with lower RMSE and MAE values (67.4582 and 35.1966, respec-
tively), compared
to the LSTM to the(116.9480
model LSTM model and(103.5552 and 61.7088, respectively).
5.8673, respectively). The R2 valueThe R 2 value is
is higher for the
also
GRUhigher
model for(0.8731)
the GRUthanmodelfor(0.7742)
the LSTMthanmodel
for the(0.7479).
LSTM model
The(0.4679).
GRU modelThe GRU
also model
has a shorter
also has atime.
training shorter training
Figure time. Figure
12 below depicts10the
below depicts
actual theprices
closing pattern of the
and actual closing
predicted closing prices
prices and predicted closing prices for the LSTM and GRU models.
for the LSTM and GRU models.

Electronics 2024, 13, 3396 18 of 27

Figure10.
Figure Googlestock:
10.Google stock: actual
actual andand predicted
predicted close,
close, LSTMLSTM and GRU
and GRU models.
models.

4.3.3. Microsoft Stock Prediction

The performance of the LSTM and GRU models for Microsoft stock, as measured by
MAE, RMSE, MAD, and R2, is summarized in Table 3.
Table 3 indicates that the GRU model forecasts Microsoft stock prices more accu-
rately, with significantly lower RMSE and MAE values (8.0805 and 5.2005, respectively)
compared to the LSTM model (32.3734 and 31.0450, respectively). The GRU model also
shows a better fit (R2 = 0.8319) and has a shorter training time than the LSTM model. Figure
11 shows the evolution of actual and expected closing prices for the LSTM and GRU mod-
els. Furthermore, the GRU model needs less training time than the LSTM model.

Figure 11.
Figure Microsoft stock:
11. Microsoft stock: actual
actualand
andpredicted
predictedclose
closeprice,
price,LSTM
LSTMand GRU
and models.
GRU models.

4.3.4. Amazon Stock Prediction

The performance of the LSTM and GRU models for Amazon stock, as measured by
MAE, RMSE, MAD, and R2, is summarized in Table 3.
From Table 3, it is evident that the GRU model predicts Amazon stock prices more
accurately, with lower RMSE and MAE values (82.9599 and 3.8673, respectively) com-
2
 Electronics2024,
Electronics 2024, 13,
13, 3396
3396 19 of 27 18 of 27

Figure
Figure12.12.
Amazon stock:
Amazon actual
stock: andand
actual predicted closeclose
predicted price,price,
LSTMLSTM
and GRU
andmodels.
GRU models.

4.4. Predicted Risk–Return

In all the examined Tradeoff
cases, the Gated Recurrent Unit (GRU) model not only demon-
strated superior forecasting
A risk–return tradeoff plotaccuracy
was createdbuttoalso
linktrained faster.stock
the predicted Thisprices
dual advantage
from the GRU illustrates
the GRU
model withmodel’s
effectiveefficiency and effectiveness
decision-making. in predicting
This plot visually stockthe
represents prices,
modela crucial aspect of
s perfor-
market
mance byinvestments.
connecting the However,
risks from there is an interesting
predicted returns amongnuance worth
the stock considering.
prices. For Apple
It visualizes
and tradeoffs
these Google stock,
for thethe GRU
four model has
technology a slightly
stocks consideredhigher Mean
in this Absolute
study: Deviation
Apple, Google, Mi-(MAD),
crosoft,
which and Amazon.
indicates the The risk–return
average tradeoff
distance between ploteach
is presented in Figure
data point 13mean.
and the below.While the GRU
model’s predicted averages are closer to the actual values, some individual predictions
deviate more from the actual values compared to the long short-term memory (LSTM)
model. The slight increase in MAD may be due to the inherent variability of stock prices.
Different stocks have different characteristics resulting from a company’s market behaviors,
such as trading volume, market sentiment, market or company events, and financial perfor-
mance, resulting in higher volatility for certain stocks. Therefore, the slightly higher MAD
of the GRU model does not necessarily imply a lack of predictive power but may reflect
the nature of the data it is dealing with. To conclude, the GRU model appears to be more
suitable for firms to use as a stock price forecasting tool, given its overall advantages in
terms of forecasting accuracy and time efficiency. However, it is recommended to consider
individual forecasting biases when dealing with stocks with high price volatility.

4.4. Predicted Risk–Return Tradeoff

A risk–return tradeoff plot was created to link the predicted stock prices from the GRU
model with effective decision-making. This plot visually represents the model’s perfor-
mance by connecting the risks from predicted returns among the stock prices. It visualizes
these tradeoffs for the four technology stocks considered in this study: Apple, Google,
Microsoft, and Amazon. The risk–return tradeoff plot is presented in Figure 13 below.
As observed from the predicted risk–return tradeoff plot presented above, there is a
Figure 13. Predicted
positive risk–return
relationship tradeoff
between plot.
risk and expected returns for each of the four technology
stocks considered in this study, aligning with the foundational principle of finance that
As observed from the predicted risk–return tradeoff plot presented above, there is a
higher returns usually come at the cost of higher risk. However, there are disparities
positive relationship between risk and expected returns for each of the four technology
in the results of tradeoff analyses of these technology giants. Investing in Google stock
is the most conservative investment, with the lowest risk and lowest expected returns.
On the other hand, the risk associated with Amazon stock is higher than that of Apple
stock. However, Apple stock predicts more expected returns despite the lower risk than
Electronics 2024, 13, 3396 19 of 27

Amazon stock. This could be due to the high price of Amazon’s stock, which might result
in more price volatility. Apple stock’s counter-intuitive scenario might result from market
Figure 12. Amazon stock: actual and predicted close price, LSTM and GRU models.
sentiment, Apple’s strong financial performance, or the company’s potential for future
growth.
4.4. It suggests
Predicted that Apple
Risk–Return could provide an attractive risk–return tradeoff for investors.
Tradeoff
Microsoft stock is also associated with lower risk but considerably higher expected returns
whenAcompared
risk–returnwith
tradeoff plotstocks.
Google was created to linkreflect
This might the predicted
investorstock prices from
confidence the GRU
in Microsoft’s
model with effective decision-making. This plot visually represents the
business model, its diverse range of offerings, and its solid financial performance. model s perfor-
The
mance by connecting
risk–return the risks
tradeoff chart from
shows predicted
that the riskreturns among
and return the stock
profiles prices. Itstocks
of different visualizes
vary
these tradeoffs
even within thefor the four
same technology
industry. stocks considered
The analysis in this study: Apple,
provides decision-makers with Google, Mi-
an effective
crosoft, and Amazon. The risk–return tradeoff plot is presented in Figure 13 below.
tool to align their investment decisions with their risk appetite and return expectations.

Figure 13.
Figure Predicted risk–return
13. Predicted risk–return tradeoff
tradeoff plot.

5. Additional
As observed Experiments and Validations
from the predicted risk–return tradeoff plot presented above, there is a
Ourrelationship
positive objective is to identify
between theand
risk most accuratereturns
expected model for
for each
each of
of the
the four
four stock prices.
technology
The primary goal of this analysis is to construct reliable and precise forecasting models
specifically tailored for short- to medium-term predictions. To ensure the consistency and
reliability of these models, we conducted a validation exercise over a 30-day time horizon,
starting from 1 January 2023. This timeframe simulates the intended use of these models in
real-world scenarios, allowing us to assess their practical effectiveness and suitability for
forecasting stock prices in a reasonable timeframe. The results are summarized in Table 5.

Table 5. Performance of four selected models on stocks.

STOCK MODEL MSE MAE RMSE

Apple ARIMA 256.58 13.15 16.01
Apple XGBOOST 254.24 14.61 15.94
Apple LSTM 113.32 8.04 10.64
Apple FB PROPHET 1355.6 31.03 36.81
Apple GRU 95.23 7.41 9.76
Amazon ARIMA 1194.5 28.56 34.56
Amazon XGBOOST 426.92 17.69 20.66
Amazon LSTM 240.69 8.6 15.51
Electronics 2024, 13, 3396 20 of 27

Table 5. Cont.

STOCK MODEL MSE MAE RMSE

Amazon FB PROPHET 5819.41 70.9 76.28
Amazon GRU 210.47 7.96 14.51
Google ARIMA 813.05 24.17 28.51
Google XGBOOST 56.22 5.94 7.49
Google LSTM 82.59 4.84 9.08
Google FB PROPHET 4172.36 59.34 64.59
Google GRU 73.21 4.53 8.56
Microsoft ARIMA 1940.95 38.09 44.05
Microsoft XGBOOST 239.23 13 15.46
Microsoft LSTM 278.49 10.39 16.68
Microsoft FB PROPHET 9446.06 88.19 97.19
Microsoft GRU 249.92 9.87 15.81

5.1. Performance of Four Selected Models on Apple Stock

The Root Mean Square Error (RMSE) is our primary metric for assessing the accuracy
of forecasting models. As shown in Table 5, the RMSE scores were 10.64 for LSTM, 15.94 for
XGBoost, 16.01 for ARIMA, and 36.81 for Facebook Prophet. A lower RMSE signifies
better predictive performance, indicating that LSTM and XGBoost outperformed ARIMA
and Facebook Prophet. LSTM achieved the lowest RMSE, suggesting it was the most
accurate in capturing Apple’s stock price trends, followed closely by XGBoost. ARIMA and
Facebook Prophet had higher RMSE scores, implying they struggled to capture stock price
fluctuations effectively. Nevertheless, model selection should consider other factors like
computational complexity and suitability for the specific forecasting task.

5.2. Performance of Four Selected Models on Amazon Stock

In the case of Amazon stock predictions, the RMSE scores were 15.51 for LSTM,
20.66 for XGBoost, 34.56 for ARIMA, and a notably higher 76.28 for Facebook Prophet.
As shown in Table 5, a lower RMSE signifies better predictive accuracy, and here, LSTM
exhibited the lowest RMSE, indicating its superior ability to capture Amazon’s stock price
trends. XGBoost also performed well, with a relatively low RMSE. In contrast, ARIMA
had a higher RMSE, suggesting it struggled to effectively capture stock price movements.
Facebook Prophet, with the highest RMSE, appears to have had the most difficulty in
accurately forecasting Amazon’s stock prices.

5.3. Performance of Four Selected Models on Google Stock

In the context of Google stock predictions, the RMSE scores were 9.08 for LSTM,
7.49 for XGBoost, 28.51 for ARIMA, and a substantially higher 64.59 for Facebook Prophet.
As shown in Table 5, a lower RMSE score indicates better predictive accuracy, and in this
case, both LSTM and XGBoost delivered commendable results with low RMSE values,
suggesting their effectiveness in capturing Google’s stock price trends. In contrast, ARIMA
exhibited a higher RMSE, indicating it struggled to predict the stock price movements
accurately. Facebook Prophet, with the highest RMSE, seems to have faced significant
challenges in providing accurate forecasts for Google stock.

5.4. Performance of Four Selected Models on Microsoft Stock

The RMSE scores obtained were 16.68 for LSTM, 15.46 for XG Boost, 44.05 for ARIMA,
and the score was notably higher at 97.19 for Facebook Prophet. As shown in Table 5,
lower RMSE values indicate better predictive accuracy, and in this case, both LSTM and
Electronics 2024, 13, 3396 21 of 27

XGBoost demonstrated a relatively strong performance, with low RMSE scores, suggesting
their effectiveness in capturing Microsoft’s stock price trends. ARIMA, on the other hand,
exhibited a higher RMSE, indicating some difficulty in accurately predicting the stock price
movements. With the highest RMSE, Facebook Prophet has faced substantial challenges in
providing accurate forecasts for Microsoft stock.

5.5. Discussion: Forecasting Accuracy

Advanced machine learning techniques have yielded a collection of top-performing
models, each with unique strengths in providing predictive insights. These models con-
sistently demonstrate their expertise in delivering directional accuracy, enabling us to
grasp the general trends in stock price movements. Furthermore, they excel in generating
predicted values that closely mirror actual stock prices, highlighting their proficiency in
capturing the intricate patterns hidden in the financial markets.
Among these models, the XGBoost model stands out as the best model, with the high-
est accuracy, and it is a symbol of predictive power. Its exceptional precision in forecasting
Google’s stock price, in particular, underscores the robustness and finesse of the XGBoost
algorithm in decoding the complexities inherent in the stock market. It has earned its place
as the most accurate model in our extensive analysis, leaving an indelible mark on our
quest for precision in stock price forecasting.
Our models can predict the general direction of stock movements, but they have
trouble capturing unexpected, occasionally unanticipated market changes brought on
by uncertainties and speculations, which motivate us to investigate further under what
conditions we can rely on these models with certainty and be solely satisfied with their
accuracy. Interestingly, our expedition has revealed some issues due to the inconsistency
in various accuracy metrics. This variability serves as a reminder of the multifaceted
nature of stock market predictions and emphasizes the importance of evaluating results
from multiple angles. The absence of a uniform standard for assessment underscores the
necessity of a comprehensive evaluation approach.
Furthermore, it is crucial to note that achieving forecasting excellence yields a variety
of solutions. Each model is unique across all scenarios and within every scoring metric.
Choosing the most suitable model becomes a subjective decision, dependent on the user’s
specific objectives and preferences. For example, investors focusing solely on directional ac-
curacy may favor one model, while those concerned about short-term fluctuations may lean
toward a different one. Thus, the adaptability of model selection to align with predefined
goals becomes important in our quest for precision.
To sum it up, our exploration into forecasting accuracy has revealed valuable insights.
Our models are effective in indicating stock price trends and are generally accurate. How-
ever, they struggle with the unpredictability of the market. Hence, we suggest that users
take a balanced approach, consider models from different perspectives, and exercise caution
when using them in the ever-changing world of stock market predictions.

5.6. Implications of This Research

This work advances the field of deep learning-based stock price prediction in several
significant ways. First, it fills a vacuum in the existing literature by conducting a targeted
comparison analysis of LSTM and GRU models, specifically for the technology industry.
Second, the models built here offer more precision than those trained on general stock market
data, thanks to the incorporation of industry-specific knowledge. Finally, the study offers
investors useful advice by determining the best model for technology stock predictions.
These contributions build on earlier research by highlighting the significance of industry-
specific modeling techniques and their potential to enhance investment decision-making.
The implications of this study are not just theoretical but have significant economic
consequences. Robust stock price forecasting models, such as the GRU model discussed,
can empower investors to make more informed decisions, potentially improving portfolio
performance. By focusing on the technology sector, this study provides insights into a
Electronics 2024, 13, 3396 22 of 27

field that has been a key driver of economic growth. The potential for improved tech-
nology stock price forecasting to enhance market efficiency and capital allocation is a
compelling prospect.

6. Discussions
6.1. Contributions
In this paper, we clearly demonstrate that we have made three major contributions to
this research.
First, we developed machine learning (ML) frameworks for social, economic, and
demographic prediction, as we have developed ML models to perform accurate analysis
and predictions for selected stock prices and the risks associated with them. Modeling
stock prices provides crucial insights into the dynamics of financial markets, with profound
implications for the economy and society at large. Stock markets essentially represent public
expectations of corporate growth and economic health. Advanced forecasting fuels data-
driven decision-making, risk assessment, and policy actions that shape social outcomes [33].
A second contribution is our use of big data and data sources for digital and computa-
tional analysis, since we have used a big data approach to analyze stock market prices and
predictions and investigate their relations to the US market and its economy. This research
implemented machine learning on an extensive dataset of 3775 daily observations across
four major technology stocks over ten years. The data-intensive modeling approaches
demonstrate the power of modern computational statistics to uncover complex patterns in
economic time series data [34].
Third, we used deep learning for stock prediction, as the primary goal of this study
was to employ deep learning, AI, and machine learning methods like the recurrent neural
network (RNN) to accurately anticipate the pattern of future stock prices in the technology
sector. We used daily technology stock data and basic technical indicators and compared
LSTM and GRU models, which belong to the RNN family, to ascertain which of them is
more efficient in predicting stock prices of technology industries. To achieve this aim, this
study collected daily historical stock datasets for Apple, Google, Microsoft, and Amazon
stocks from the API of Yahoo Finance through the Python ‘pandas_datareader.data’ and
Yahoo Finance library. The stocks selected are for international public companies traded
at both NASDAQ and the NYSE. The time series data range from 1 January 2013 to
30 March 2022. Together, the series contains 3775 trading days. The daily time series data
were automatically downloaded because Python is connected to the Yahoo Finance API.
The dataset includes daily prices: open, highest, lowest, close, adjusted close, and daily
trading volume.
The study applied deep learning models to analyze selected technological stock pat-
terns as a one-dimensional time series and forecast future stock prices by examining past
historical prices and the most critical technical indicators. The analysis built a compari-
son system to examine the performance of the LSTM and the GRU ensemble models on
the selected technology stock data to identify a parsimonious model for the real-world
representation of the technology stock markets.
The performances of the LSTM and GRU models were assessed using the Mean
Absolute Error (MAE), the Root Mean Square Error (RMSE), the Mean Absolute Deviation
(MAD), and the coefficient of determination (R2 ). From the results, it is observed that the
GRU model makes it simpler to predict how close the predicted value of the Apple, Google,
Amazon, and Microsoft stocks are to the actual value, as the RMSE and MAE values are
considerably lower for the GRU model for all of the technology stocks than for the LSTM
model. Moreover, the model’s fit (R2 ) is observed to be better for the GRU model than
for the LSTM model. It was also observed from our analysis that the GRU model has a
shorter training time than the LSTM model. Therefore, the GRU model produced a better
forecasting system for predicting daily technology stock data and fundamental technical
indicators. It can be used to efficiently estimate the pattern of future stock prices within the
technology industry.
Electronics 2024, 13, 3396 23 of 27

Lastly, this study linked the predicted stock from the GRU model with effective
decision-making. The risk–return tradeoff plot was computed as a visual depiction of the
model performance to connect risks from predicted returns among the technological stock
prices, and it can be observed that there is a positive relationship between risk and expected
returns for each of the four technology stocks considered in this study. Investing in Google
stock is associated with the lowest risk and lowest expected returns. However, the risk
associated with Amazon stock is higher than that of Apple stock. However, Apple stock
predicted more expected returns despite the lower risk compared with Amazon stock.
Microsoft stock is also associated with lower risk, but considerably higher expected returns
compared with Google stocks.
The present study has several contributions. Firstly, our study focuses on the technol-
ogy sector, comparing LSTM and GRU models specifically for technology stocks like Apple,
Google, Microsoft, and Amazon. This sector-specific analysis reveals unique patterns that
are not seen in broader market studies, providing more useful insights for technology
investors. Additionally, we evaluate not only the accuracy but also the training efficiency
of the models, offering practical insights into their computational performance. Our study
also includes a risk–return analysis based on predicted stock prices, giving practical insights
into investment strategies. These points highlight the unique aspects of our approach and
the significant contributions of our work.

6.2. Limitations of the Study

There are two main limitations of this study. Firstly, the primary setback experienced
in the process of this study was the inadequacy of stock price data. As stated earlier in
this study, a number of factors influence stock market volatility, and building an efficient
machine learning model that predicts these stock prices with minimum error requires a
sizeable number of attributes that are not available for many of the stocks considered. The
Bureau of Labor Statistics reports that there are around 260 trading days every year, which
is considered insufficient if there is the need to go far in time for more examples, such as by
examining data from the last two to three years.
The other limitation is that building an effective system for stock market prediction
requires a denoising process that involves adding more technical indicators, such as the
daily sentiment polarity score, which will help remove human feelings for the proper
estimation of future stock prices. However, this process requires complicated computation
methods that could not be considered in this study due to time constraints.

6.3. Future Research

The study results show that the GRU model is an effective model for predicting
technology stock prices among the recurrent neural network models. However, this result
cannot be generalized for all other stock market data due to the lack of a sizeable amount of
stock data. Therefore, it can only be concluded arbitrarily that the GRU model outperforms
the LSTM model in stock forecasting. It is recommended that future studies compare these
two models on a larger quantity of datasets and extend the estimation to stock price data in
other industries.
Future studies should also consider focusing on building the stock price prediction
system through a deep neural network that considers historical financial data, technical
indicators, and financial news, and use a large volume of the training dataset to yield
less prediction error. The reason is that stock price data are very volatile and often show
noisy characteristics as well as non-stationary patterns. The inclusion of more technical
indicators, a large volume of the training dataset, financial news, and posts can be used to
denoise the data.
Finally, this study also recommends the utilization of stacked models, as this study
solely compared models with each other. Future researchers could discover more by
stacking models to see if they can improve prediction ability.
Electronics 2024, 13, 3396 24 of 27

While our study focuses on predicting stock prices in the technology sector, the ad-
vanced deep learning (DL) and machine learning (ML) models we employ have broad
applicability across various industries. For example, these models can predict patient out-
comes and optimize treatment plans in the healthcare sector. The energy sector can utilize
our solution to forecast consumption patterns and optimize grid operations. In retail, our
models can forecast sales and manage inventory. Financial institutions can leverage these
techniques for credit scoring, fraud detection, and risk management. By demonstrating the
versatility of our solution, we highlight its potential to address diverse challenges across
different sectors, underscoring the broad impact and utility of our approach.

7. Conclusions
This study has made important contributions to financial market analysis by utilizing
cutting-edge machine learning and deep learning techniques. We successfully designed ML
models that predict stock prices with precision and assess associated risks, offering critical
insights into the factors that shape financial markets and, by extension, the broader economy.
By applying big data approaches to analyze extensive historical stock data, we showcased
the effectiveness of modern computational methods in revealing intricate patterns within
economic datasets. Our comparison of LSTM and GRU models demonstrated that the GRU
model excels in both prediction accuracy and computational efficiency, particularly within
the technology sector. Additionally, the study’s analysis of the risk–return relationship
provided actionable insights for investors, highlighting the distinct behaviors of major
technology stocks. These findings not only enhance our understanding of stock market
dynamics but also provide a strong foundation for future research in financial forecasting
and investment strategy development.

Author Contributions: Conceptualization, V.C. and A.C.; methodology, V.C.; software, A.C.; valida-
tion, V.C., Q.A.X. and H.W.; formal analysis, V.C., A.C. and H.W.; investigation, V.C. and Q.A.X.; re-
sources, V.C.; data curation, A.C.; writing—original draft preparation, V.C. and A.C.; writing—review
and editing, V.C., Q.A.X. and H.W.; visualization, V.C. and A.C.; supervision, V.C.; project administra-
tion, V.C.; funding acquisition, V.C. All authors have read and agreed to the published version of
the manuscript.
Funding: This work is partly supported by VC Research (VCR 000221) and Leverhulme Trust
(VP1-2023-025).
Data Availability Statement: Data is available upon requests. Readers can also download data from
Yahoo Finance or Google Finance.
Acknowledgments: We thank Akash Prasad and Akram Dehnokhalaji for spending some time to
help improve part of this research project.
Conflicts of Interest: The authors declare no conflicts of interest.

Appendix A

Table A1. Summary statistics of Apple stock price features.

OPEN HIGH LOW CLOSE ADJ CLOSE VOLUME

COUNT 2268 2268 2268 2268 2268 2268
MEAN 57.2297 57.84394 56.63649 57.2662 55.63746 166,807,034
STD 44.14435 44.68723 43.6103 44.17467 44.78484 107,981,557
MIN 13.97714 14.29536 13.88821 14.06357 12.30091 41,000,000
25% 27.14937 27.35375 26.86813 27.14625 24.96922 95,148,700
50% 39.47 39.845 39.02875 39.36125 37.76188 133,553,600
75% 68.54375 69.69688 67.44 68.765 67.7272 202,644,200
MAX 182.63 182.94 179.12 182.01 181.7784 1,065,523,200
Electronics 2024, 13, 3396 25 of 27

Table A2. Summary statistics of Google stock price features.

OPEN HIGH LOW CLOSE ADJ CLOSE VOLUME

COUNT 2268 2268 2268 2268 2268 2268
MEAN 1139.757 1150.551 1128.898 1139.985 1139.985 1,932,454
STD 674.3742 681.2976 667.1945 674.1677 674.1677 1,261,224
MIN 398.8052 400.4789 386.053 398.5611 398.5611 7922
25% 607.7186 611.8685 602.0832 607.2103 607.2103 1,222,225
50% 986.725 990.855 976.655 984.065 984.065 1,557,583
75% 1307.397 1319.705 1304.069 1313.13 1313.13 2,191,964
MAX 3037.27 3042 2997.75 3014.18 3014.18 23,219,507

Table A3. Summary statistics of Microsoft stock price features.

OPEN HIGH LOW CLOSE ADJ CLOSE VOLUME

COUNT 2268 2268 2268 2268 2268 2268
MEAN 114.5614 115.6476 113.4146 114.5935 110.6238 31,954,477.16
STD 84.94871 85.80088 84.00433 84.94511 86.28801 16,886,826.19
MIN 30.3 30.9 30.27 30.6 25.62621 7,425,600
25% 47.1775 47.6675 46.695 47.26 41.86476 22,205,600
50% 77.63 77.9 77.36 77.78 73.41168 28,010,550
75% 157.185 158.8025 156.0725 157.6175 154.3496 36,433,425
MAX 344.62 349.67 342.2 343.11 342.402 248,428,500

Table A4. Summary statistics of Amazon stock price features.

OPEN HIGH LOW CLOSE ADJ CLOSE VOLUME

COUNT 2268 2268 2268 2268 2268 2268
MEAN 1454.869 1470.281 1437.496 1454.153 1454.153 4,047,869
STD 1075.819 1088.213 1061.821 1074.703 1074.703 2,153,351
MIN 248.94 252.93 245.75 248.23 248.23 881,300
25% 482.5175 489.3 474.9075 482.1525 482.1525 2,686,300
50% 1023.14 1032.22 1016.75 1026.27 1026.27 3,464,050
75% 1949 1975.377 1931.703 1954.335 1954.335 4,693,025
MAX 3744 3773.08 3696.79 3731.41 3731.41 23,856,100

References
1. Ariyo, A.A.; Adewumi, A.O.; Ayo, C.K. Stock Price Prediction Using the ARIMA Model. In Proceedings of the 2014 UKSim-AMSS
16th International Conference on Computer Modelling and Simulation, Cambridge, UK, 26–28 March 2014; pp. 106–112.
2. Nicholas Refenes, A.; Zapranis, A.; Francis, G. Stock performance modeling using neural networks: A comparative study with
regression models. Neural Netw. 1994, 7, 375–388. [CrossRef]
3. Malki, A.; Atlam, E.-S.; Hassanien, A.E.; Ewis, A.; Dagnew, G.; Gad, I. SARIMA model-based forecasting required number of
COVID-19 vaccines globally and empirical analysis of peoples’ view towards the vaccines. Alex. Eng. J. 2022, 61, 12091–12110.
[CrossRef]
4. Paliari, I.; Karanikola, A.; Kotsiantis, S. A comparison of the optimized LSTM, XGBOOST and ARIMA in Time Series forecasting.
In Proceedings of the 2021 12th International Conference on Information, Intelligence, Systems & Applications (IISA), Chania,
Crete, Greece, 12–14 July 2021; pp. 1–7.
Electronics 2024, 13, 3396 26 of 27

5. Cheng, Y.; Yi, J.; Yang, X.; Lai, K.K.; Seco, L. A CEEMD-ARIMA-SVM model with structural breaks to forecast the crude oil prices
linked with extreme events. Soft Comput. 2022, 26, 8537–8551. [CrossRef]
6. Chatterjee, A.; Bhowmick, H.; Sen, J. Stock Price Prediction Using Time Series, Econometric, Machine Learning, and Deep
Learning Models. In Proceedings of the 2021 IEEE Mysore Sub Section International Conference (MysuruCon), Hassan, India,
24–25 October 2021; pp. 289–296.
7. Escudero, P.; Alcocer, W.; Paredes, J. Recurrent Neural Networks and ARIMA Models for Euro/Dollar Exchange Rate Forecasting.
Appl. Sci. 2021, 11, 5658. [CrossRef]
8. Liang, F.; Liang, F.; Zhang, H.; Zhang, H.; Fang, Y.; Fang, Y. The Analysis of Global RMB Exchange Rate Forecasting and Risk
Early Warning Using ARIMA and CNN Model. J. Organ. End User Comput. (JOEUC) 2022, 34, 1–25. [CrossRef]
9. Rao, A.R.; Reimherr, M. Modern non-linear function-on-function regression. Stat. Comput. 2023, 33, 130. [CrossRef]
10. Thakkar, A.; Chaudhari, K. A comprehensive survey on deep neural networks for stock market: The need, challenges, and future
directions. Expert Syst. Appl. 2021, 177, 114800. [CrossRef]
11. White, H. Economic prediction using neural networks: The case of IBM daily stock returns. In Proceedings of the IEEE 1988
International Conference on Neural Networks, San Diego, CA, USA, 24–27 July 1988; Volume 452, pp. 451–458.
12. Sonkiya, P.; Bajpai, V.; Bansal, A. Stock price prediction using BERT and GAN. arXiv 2021, arXiv:2107.09055.
13. Maqsood, H.; Mehmood, I.; Maqsood, M.; Yasir, M.; Afzal, S.; Aadil, F.; Selim, M.M.; Muhammad, K. A local and global event
sentiment based efficient stock exchange forecasting using deep learning. Int. J. Inf. Manag. 2020, 50, 432–451. [CrossRef]
14. Patil, P.; Wu, C.-S.M.; Potika, K.; Orang, M. Stock Market Prediction Using Ensemble of Graph Theory, Machine Learning
and Deep Learning Models. In Proceedings of the 3rd International Conference on Software Engineering and Information
Management, Sydney, NSW, Australia, 12–15 January 2020; pp. 85–92.
15. Hochreiter, S. Untersuchungen zu Dynamischen Neuronalen Netzen. Bachelor’s Thesis, Technische Universität München,
München, Germany, 1991.
16. Bengio, Y.; Simard, P.; Frasconi, P. Learning long-term dependencies with gradient descent is difficult. IEEE Trans. Neural Netw.
1994, 5, 157–166. [CrossRef]
17. Hochreiter, S. The Vanishing Gradient Problem During Learning Recurrent Neural Nets and Problem Solutions. Int. J. Uncertain.
Fuzziness Knowl. Based Syst. 1998, 6, 107–116. [CrossRef]
18. Schmidhuber, J.; Wierstra, D.; Gagliolo, M.; Gomez, F. Training recurrent networks by Evolino. Neural Comput. 2007, 19, 757–779.
[CrossRef] [PubMed]
19. Chen, J.; Chaudhari, N.S. Segmented-Memory Recurrent Neural Networks. IEEE Trans. Neural Netw. 2009, 20, 1267–1280.
[CrossRef]
20. Sutton, R.S.; Barto, A.G. Reinforcement Learning: An Introduction; A Bradford Book; The MIT Press: Cambridge, MA, USA; London,
UK, 2018.
21. Cho, K.; van Merriënboer, B.; Gulcehre, C.; Bahdanau, D.; Bougares, F.; Schwenk, H.; Bengio, Y. Learning Phrase Representations
using RNN Encoder–Decoder for Statistical Machine Translation. In Proceedings of the 2014 Conference on Empirical Methods in
Natural Language Processing (EMNLP), Doha, Qatar, 25–29 October 2014; pp. 1724–1734.
22. Jiang, L.; Subramanian, P. Forecasting of Stock Price Using Autoregressive Integrated Moving Average Model. J. Comput. Theor.
Nanosci. 2019, 16, 3519–3524. [CrossRef]
23. Kavzoglu, T.; Teke, A. Predictive Performances of Ensemble Machine Learning Algorithms in Landslide Susceptibility Mapping
Using Random Forest, Extreme Gradient Boosting (XGBoost) and Natural Gradient Boosting (NGBoost). Arab. J. Sci. Eng. 2022,
47, 7367–7385. [CrossRef]
24. Lilly Sheeba, S.; Neha, G.; Anirudh Ragavender, R.M.; Divya, D. Time Series Model for Stock Market Prediction Utilising Prophet.
Turk. J. Comput. Math. Educ. (TURCOMAT) 2021, 12, 4529–4534.
25. Yun, K.K.; Yoon, S.W.; Won, D. Interpretable stock price forecasting model using genetic algorithm-machine learning regressions
and best feature subset selection. Expert Syst. Appl. 2023, 213, 118803. [CrossRef]
26. Han, C.; Fu, X. Challenge and Opportunity: Deep Learning-Based Stock Price Prediction by Using Bi-Directional LSTM Model.
Front. Bus. Econ. Manag. 2023, 8, 51–54. [CrossRef]
27. Zhao, Y.; Yang, G. Deep Learning-based Integrated Framework for stock price movement prediction. Appl. Soft Comput. 2023,
133, 109921. [CrossRef]
28. Quadir, A.; Kapoor, S.; Junni, A.V.C.; Sivaraman, A.K.; Tee, K.F.; Sabireen, H.; Janakiraman, N. Novel optimization approach for
stock price forecasting using multi-layered sequential LSTM. Appl. Soft Comput. 2023, 134, 109830. [CrossRef]
29. Lu, M.; Xu, X. TRNN: An efficient time-series recurrent neural network for stock price prediction. Inf. Sci. 2024, 657, 119951.
[CrossRef]
30. Salah, S.; Alsamamra, H.R.; Shoqeir, J.H. Exploring Wind Speed for Energy Considerations in Eastern Jerusalem-Palestine Using
Machine-Learning Algorithms. Energies 2022, 15, 2602. [CrossRef]
31. Li, Z.; Yu, H.; Xu, J.; Liu, J.; Mo, Y. Stock Market Analysis and Prediction Using LSTM: A Case Study on Technology Stocks. Innov.
Appl. Eng. Technol. 2023, 2, 1–6. [CrossRef]
Electronics 2024, 13, 3396 27 of 27

32. Liu, Y. Analysis and forecast of stock price based on LSTM algorithm. In Proceedings of the 2021 IEEE International Conference on
Computer Science, Electronic Information Engineering and Intelligent Control Technology (CEI), Fuzhou, China, 24–26 September
2021; pp. 76–79.
33. Jabeur, S.B.; Mefteh-Wali, S.; Viviani, J.-L. Forecasting gold price with the XGBoost algorithm and SHAP interaction values. Ann.
Oper. Res. 2024, 334, 679–699. [CrossRef]
34. Kumar, G.; Jain, S.; Singh, U.P. Stock Market Forecasting Using Computational Intelligence: A Survey. Arch. Comput. Methods Eng.
2020, 28, 1069–1101. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.