algorithms

Article
Unleashing the Power of Tweets and News in Stock-Price
Prediction Using Machine-Learning Techniques
Hossein Zolfagharinia 1, * , Mehdi Najafi 1 , Shamir Rizvi 1 and Aida Haghighi 2

1 Global Management Studies Department, Ted Rogers School of Management, Toronto Metropolitan
University, Toronto, ON M5B 2K3, Canada; najafi.mehdi@torontomu.ca (M.N.);
shamir.rizvi@torontomu.ca (S.R.)
2 School of Occupational and Public Health, Faculty of Community Services, Toronto Metropolitan University,
Toronto, ON M5B 2K3, Canada; aida.haghighi@torontomu.ca
* Correspondence: h.zolfagharinia@torontomu.ca; Tel.: +1-416-979-5000 (ext. 557532)

Abstract: Price prediction tools play a significant role in small investors’ behavior. As such, this
study aims to propose a method to more effectively predict stock prices in North America. Chiefly,
the study addresses crucial questions related to the relevance of news and tweets in stock-price
prediction and highlights the potential value of considering such parameters in algorithmic trading
strategies—particularly during times of market panic. To this end, we develop innovative multi-
layer perceptron (MLP) and long short-term memory (LSTM) neural networks to investigate the
influence of Twitter count (TC), and news count (NC) variables on stock-price prediction under both
normal and market-panic conditions. To capture the impact of these variables, we integrate technical
variables with TC and NC and evaluate the prediction accuracy across different model types. We
use Bloomberg Twitter count and news publication count variables in North American stock-price
prediction and integrate them into MLP and LSTM neural networks to evaluate their impact during
the market pandemic. The results showcase improved prediction accuracy, promising significant
benefits for traders and investors. This strategic integration reflects a nuanced understanding of the
market sentiment derived from public opinion on platforms like Twitter.

Keywords: stock-price prediction; neural network; LSTM; multi-layer perceptron; news count
Citation: Zolfagharinia, H.; Najafi, M.;
Rizvi, S.; Haghighi, A. Unleashing the
Power of Tweets and News in
Stock-Price Prediction Using
Machine-Learning Techniques. 1. Introduction
Algorithms 2024, 17, 234. https:// The dynamic landscape of the global stock market plays a significant role in shaping
doi.org/10.3390/a17060234 economies, influencing individual financial decisions, and driving continuous innovation
Academic Editor: Jesús Ángel in investment strategies. The indisputable significance of the stock market is underscored
Román Gallego by the growth in total global capitalizations, which exceeded USD 109 trillion in 2023—a
remarkable threefold increase from the 2009 figure of USD 25 trillion [1,2]. As per World
Received: 8 March 2024 Bank data from 2017, stock trading’s substantial impact on the American economy has been
Revised: 17 May 2024
evident since 2013, with the total value of stocks traded on US markets consistently surpass-
Accepted: 24 May 2024
ing 200% of the nation’s annual GDP. The New York Stock Exchange (NYSE)—a symbolic
Published: 28 May 2024
hub of financial activity—boasts an average market capitalization of approximately USD
29 trillion, highlighting its central role in global financial markets [3]. The daily average of
about USD 169 billion in stocks traded on the NYSE further emphasizes the fluidity and
Copyright: © 2024 by the authors.
dynamism of the market [4].
Licensee MDPI, Basel, Switzerland. Beyond its macroeconomic influence, the movement of stocks on a micro level is
This article is an open access article crucial in determining the financial market’s overall well-being. Notably, studies such
distributed under the terms and as that of Chan and Woo [5] reveal that stock-market price booms can drive long-term
conditions of the Creative Commons economic growth. The involvement of novice investors in the stock market adds another
Attribution (CC BY) license (https:// layer of complexity to the financial ecosystem. In particular, there are now over 54% of US
creativecommons.org/licenses/by/ adults having some form of investment in the stock market [6]. This increase in individual
4.0/). investors can be attributed to the post-2008 financial crisis [7]. In this evolving landscape,

Algorithms 2024, 17, 234. https://doi.org/10.3390/a17060234 https://www.mdpi.com/journal/algorithms

Algorithms 2024, 17, 234 2 of 29

the need for reliable stock-price prediction models increases, especially as financial products
and services become more accessible to smaller investors.
Moreover, price prediction tools play a significant role in small investors’ behavior. As
such, numerous research studies (e.g., [8–23]) have been conducted to develop appropriate
models and techniques that would widely be employed in algorithmic trading. Algorithmic
trading, in essence, leverages computational power and mathematical formulas to illumi-
nate buy or sell decisions for financial securities on an exchange and incorporates complex
formulas and models with human oversight [24–27]. These techniques are integral to insti-
tutional firms’ trading algorithms, as they aid in minimizing transaction costs and market
risks [28]. The rise of artificial intelligence (AI) in both the stock market and financial firms
has significantly contributed to the growth of the algorithmic trading market. Companies
like Sentient have developed AI-powered algorithmic traders, thereby showcasing the
potential for these advanced algorithms to function as standalone entities [29,30].
While numerous studies have developed models and techniques utilizing AI methods
in stock-price prediction, they often focus on technical indices and general data in a normal
market context. More specifically, these studies frequently overlook the impact of news on
the decisions of small traders—particularly in market-panic scenarios. The profound impact
of the media—most notably the news and Twitter—on investors’ decisions in relation to
stock buying and selling is evident. This has greatly shaped the modern financial landscape.
News that is disseminated through traditional media or online platforms can swiftly
influence investor sentiment by providing critical information about companies, industries,
and the broader economic landscape. For example, during the COVID-19 pandemic,
the demand for oil rapidly declined due to business closures and travel restrictions; this
had, inevitably, caused the futures price for West Texas Intermediate (WTI) to plummet
from USD 18 a barrel to around −USD 37 a barrel [31]. Real-time updates on corporate
earnings, geopolitical events, and market trends can trigger immediate reactions and
prompt investors to make rapid decisions in reassessing their positions. The social media
platform Twitter, in particular, has become a dynamic space for financial discussions and
for disseminating market-related information. Tweets from influential market analysts,
financial experts, and even company executives can rapidly circulate and thereby influence
investor perceptions and drive fluctuations in stock prices. The accessibility and speed of
information on both news outlets and Twitter have made it imperative for investors to stay
vigilant, as their decisions are now increasingly shaped by the instantaneous flow of news
and opinions.
Evidently, both the news and tweets can potentially impact small traders’ trading
decisions. As such, this study aims to investigate whether or not these parameters can
affect price prediction performance. More specifically, we aim to answer the following
research questions:
• Can considering the news and tweets improve the stock-price prediction accuracy?
• Does the impact of the news and tweets on the price prediction differ under normal
and market-panic conditions?
Given these research questions, the primary aim of the research is to investigate
whether considering news and tweets can enhance stock-price prediction accuracy. This
impact is crucial for investors, financial institutions, and algorithmic trading firms, as more
accurate predictions can lead to better investment decisions, reduced risks, and improved
profitability. Furthermore, by examining the impact of news and tweets on stock-price
prediction, the research seeks to deepen understanding of the dynamic interactions between
market information, investor sentiment, and stock-price movements. This understanding
can help develop more accurate prediction models that capture the complex interaction
of factors influencing financial markets. Following the understanding of these impacts,
this research aims to investigate the stability of these impacts under panic conditions. In
other words, it aims to consider whether and how these impacts may alter during periods
of market distress. To this end, we chose the COVID-19 pandemic as the case for our
analysis, given its status as one of the most significant instances of market panic in recent
Algorithms 2024, 17, 234 3 of 29

history [32–34]. Similarly, considering the widespread utilization of LSTM and MLP models
in stock-price prediction (e.g., [12,19,21,35–45]), we opted to evaluate these two methods
specifically in the context of incorporating tweets and news count as predictive variables,
particularly during market distress.
The remainder of this paper is organized as follows. In Section 2, we review the relevant
literature to identify extant knowledge gaps and highlight the contributions of our current
study. Next, in Section 3, we define the problem under investigation and outline the proposed
model for price prediction. Then, in Section 4, we develop the solution algorithms. Moreover,
in Section 5, we analyze the results to address the research questions. Lastly, in Section 6, we
offer concluding remarks and insights and suggest future avenues for research.

2. Literature Review
A stock market is a place for publicly listed companies to trade stocks and other financial
instruments; likewise, the price of shares is termed the stock price [46]. Initial studies (e.g., [47])
held a view of the stock market as stochastic and, hence, non-predictable. However, later
studies (e.g., [22,35,48,49]) argued that the stock market may be predictable to an extent when
it is examined from a behavioral economics and socioeconomic theory of finance point of view.
Therefore, many research studies have developed different models to predict stock prices and
parameters in the stock market. These various approaches and techniques can be categorized
into three main categories in terms of strategy: (i) technical analysis, (ii) fundamental analysis,
and (iii) sentiment-based analysis [20,22,30]. Technical analysis is the most popular approach.
It defines a couple of indicators, such as open price, close price, low price, high price, and
trading volume, and applies mathematical and statistical techniques to structured data to
predict the stock price based on trends in the past and present stock prices [22,48,50–52]. On
the contrary, the fundamental analysis is concerned with the company that underlies the stock
itself instead of the actual stock [53–55]. The data that are used by the fundamental analyst
are usually unstructured and, thus, pose some challenges. However, this type of data has
occasionally been shown to be a good predictor of stock-price movement [56]. Moreover,
this approach is utilized by financial analysts daily, as it incorporates various factors, such
as economic forecasts, the efficiency of management, business opportunities, and financial
statements [57]. Fundamental analysis can, in essence, be defined as a method of finding a
stock’s inherent value based on financial analysis. Lastly, sentiment-based analysis is based
on linguistic feature extraction (e.g., [58–60]). This approach has not been as popular, given
the difficulty of developing reliable and efficient sentiment analysis tools. This is mainly due
to the design complexity and relevant source selection (which is of utmost importance).
Given the popularity of the technical analysis approach, many studies have employed
different statistical and mathematical techniques to predict stock prices. These techniques
can be generally categorized into three groups: (i) statistical model (STAT), (ii) evolutionary
algorithm (EA), and (iii) machine learning (ML), including neural networks (NN). To utilize
these techniques, decision-makers should decide on the type of target variable the model
aims to predict. These variables can either be the stock price, stock direction, index price,
or index direction. The stock price refers to an individual numerical value that the model
believes the stock will be priced at soon. The stock direction refers to the direction (i.e., up or
down) in which the stock price will move. Furthermore, the index could be another target
variable. Unlike the case of stocks, where the data pertains to that of an individual stock,
an index measures a section of the stock market by combining price data from multiple
stocks. A few well-known indices are the S&P 500 and the Dow Jones Industrial Average.
Exploring the existing literature reveals that diverse techniques have been used for
stock-market prediction. Table 1 summarizes studies that have used various techniques
for stock-market prediction. To compile this table, we utilized two prominent research
databases, namely the Web of Science (WOS) and Scopus, which are known for their
comprehensive coverage across various disciplines. Additionally, we supplemented our
search by exploring the reference lists of selected studies not initially found in these
databases. Our survey focused primarily on papers published after 2000, particularly
Algorithms 2024, 17, 234 4 of 29

those published within the last decade. Subsequently, we meticulously reviewed the
available literature to identify studies specifically addressing stock-price prediction through
technical or sentimental analysis techniques. Each identified article underwent a thorough
investigation, enabling us to categorize them based on the type of analysis and techniques
employed for stock-price prediction. As shown in Table 1, individual stock outputs are
significantly more popular than index-based computations. Likewise, ML and NN are the
most popular techniques that have been used in recent years. Hence, given the popularity of
these techniques, we will focus on reviewing research studies using ML and NN techniques
in order to position our work in the extant literature.

Table 1. A summary of the related literature [61–130].

# Reference Appr. Technique # Reference Appr. Technique # Reference Appr. Technique

NN and
1 Armano et al. [8] TECH NN and EA 37 Leigh et al. [62] TECH 73 Nie and Jin [98] TECH ML (SVM)
EA
ML ML (SVM) and
2 Wang et al. [9] TECH NN and STAT 38 Kim [63] TECH 74 Chen and Pan [99] TECH
(SVM) EA
STAT ML (SVM) and
3 Kao et al. [10] TECH ML (SVM) 39 Wang [64] TECH 75 Lahmiri [100] TECH
(FUZZY) EA
4 Liu and Lv [11] TECH NN and EA 40 Chen and Leung [65] TECH NN 76 Qiu et al. [101] T/F NN
STAT
5 Guo et al. [12] TECH STAT 41 Pai and Lin [66] TECH 77 Qiu and Song [102] TECH NN
(ARIMA)
Mahmud and Enke and
6 TECH Fuzzy 42 FUND NN 78 An and Chan [103] TECH EA
Meesad [13] Thawornwong [67]
Genetic Shynkevich
7 Wang et al. [14] TECH EA 43 Kim et al. [68] TECH 79 TECH ML (KNN)
Alg et al. [104]
Schumaker and ML
8 Zhang et al. [15] TECH ML (RF) 44 SENT 80 Ouahilal et al. [105] TECH ML (SVM)
Chen [69] (SVM)
ML
9 Zhang et al. [15] TECH NN 45 Huang and Tsai [70] TECH 81 Rout et al. [106] TECH ML and EA
(SVM)
ML(SVM) and ML
10 Bisoi et al. [16] TECH 46 Kara et al. [71] TECH 82 Castelli et al. [107] TECH Genetic Alg.
EA (SVM)
Groth and
11 Ding and Qin [17] TECH Deep RNN 47 SENT STAT 83 Tao et al. [108] TECH STAT
Muntermann [72]
12 Qiu et al. [18] TECH LSTM 48 Schumaker et al. [73] SENT STAT 84 Chong et al. [109] TECH ML (ARIMA)
13 Jin et al. [19] SENT NN (LSTM) 49 Yolcu et al. [74] TECH NN 85 Weng et al. [110] TECH ML (RF)
14 Vijh et al. [20] TECH ML-NN 50 Hagenau et al. [75] SENT STAT 86 Zhuge et al. [111] SENT NN
Umoh and STAT Kraus
15 Lu et al. [21] TECH CNN-LSTM 51 TECH 87 SENT NN
Udosen [76] (FUZZY) and Feuerriegel [112]
NN and
16 Pang et al. [35] TECH LSTM 52 Nayak et al. [77] TECH 88 Jeon et al. [113] TECH NN and STAT
EA
Bommareddy NN and
17 TECH STAT 53 Tsai and Hsiao [78] FUND 89 Agustini et al. [114] TECH STAT
et al. [36] EA
NN and
18 Hiransha et al. [37] TECH NN 54 Li et al. [79] T/S 90 Matsubara et al. [115] SENT NN
STAT
SVM and Ebadati and
19 Nguyen et al. [39] TECH 55 Sun et al. [80] TECH NN 91 TECH NN and EA
LSTM Mortazavi [116]
20 Kumar et al. [40] TECH LSTM 56 Dash et al. [81] TECH EA 92 Kooli et al. [117] T/F NN
NN and
21 Mittal et al. [41] TECH LSTM-SVM 57 Adebiyi et al. [82] TECH 93 Lahmiri [118] TECH ML (SVM)
STAT
Perdana and
22 TECH LSTM 58 Lee et al. [83] SENT ML (RF) 94 Zhou et al. [119] TECH NN
Rokhim [42]
Shirata and Junqué de Fortuny ML
23 TECH LSTM 59 SENT 95 Shah et al. [120] TECH NN and EA
Harada [43] et al. [84] (SVM)
ML
24 60 Bisoi and Dash [85] T/F 96 Gocken et al. [121] TECH NN and EA
(SVM)
STAT
25 Liu et al. [44] T/S LSTM-CNN 61 Mondal et al. [86] TECH 97 Vantstone et al. [122] SENT NN
(ARIMA)
Ammer and
26 TECH LSTM 62 Geva and Zahavi [87] T/S NN 98 Zheng et al. [123] TECH ANN-SVR
Aldhyani [45]
27 Yu and Yan [49] TECH NN 63 Jiang et al. [88] SENT STAT 99 Mehtab et al. [126] TECH LSTM
Cheng and NN and
28 TECH STAT (FUZZY) 64 Hafezi et al. [89] TECH 100 Chen et al. [124] TECH ML (VAR)
Yang [50] EA
29 Rundo et al. [52] TECH NN 65 Ballings et al. [90] TECH ML (RF) 101 Jiang et al. [125] TECH ML (DT, RF)
ML
30 Khairi et al. [53] T/F/S ML and NN 66 Nguyen et al. [91] SENT 102 Antad et al. [126] TECH ML (LR)
(SVM)
Ghorbani and
31 TECH STAT (PCA) 67 Wang et al. [92] FUND STAT 103 Khan et al. [127] SENT Hybrid NNs
Chong [54]
Zhong and STAT
32 SENT LSTM 68 Sun et al. [93] TECH 104 Shaban et al. [128] SENT Hybrid
Hitchcock [55] (FUZZY)
Gupta and NN and
33 SENT ML (TF-IDF) 69 Gocken et al. [94] TECH 105 Belcastro et al. [129] T/S LSTM
Chen [58] EA
NN and
34 Jing et al. [59] SENT CNN-LSTM 70 Dash and Dash [95] TECH 106 Al-Nefaie et al. [130] TECH MLP and GRU
EA
35 Wu et al. [60] SENT CNN-LSTM 71 Zhou et al. [96] TECH NN LSTM and
107 Current Study T/S
ML MLP
36 Wang [61] TECH STAT (FUZZY) 72 Shynkevich et al. [97] SENT
(KNN)
Algorithms 2024, 17, 234 5 of 29

2.1. ML Techniques in Stock-Price Prediction

Several ML techniques have been employed in stock-price prediction. Among these
techniques, the K nearest neighbor (KNN), random forests (RF), support vector machines
(SVM), regression, support vector regression (SVR), and ARIMA are the most popular.
The KNN algorithm is one of the initial ML algorithms for stock-price prediction and is
a supervised algorithm used exclusively for classification and regression. The primary
operation involves identifying the closest neighbors to the queried data point. If the task is
classification, the most-occurring neighbor value is returned and used as the output value.
However, if regression is the objective, then the average of all neighbor values is returned
and used as the output value. Chen and Hao [131] applied the regression format of this
algorithm for stock-price prediction. Nevertheless, its application in stock-price prediction
has decreased in recent years due to its simplicity, as it results in poor performance when
compared to newer ML approaches. For instance, Shynkevich et al. [97] utilized the KNN
and SVM algorithms for stock-price prediction and compared their performance to show
the superiority of the SVM algorithm. Shynkevich et al. [104] also used ML algorithms to
investigate the impact of the forecasting window length on stock-price prediction. They
concluded that KNN-SVM results in appropriate prediction accuracy; however, the KNN
leads to poor performance.
Regression is another supervised ML technique used for stock-price prediction. For
instance, Jiang et al. [88] incorporated this algorithm to predict stock-price movements. To
this end, they divided users into stakeholder groups and then analyzed how stakeholder
group postings correlated with events in the company. They then used this information
to predict movements in stock prices. Furthermore, Gupta and Chen [58] investigated
the sentiments extracted from a vast repository of tweets sourced from StockTwits. They
employed three distinct machine-learning algorithms (i.e., Naïve Bayes, SVM, and logistic
regression) and additionally explored five different featurization techniques (i.e., bag
of words, bigram, trigram, term frequency–inverse document frequency (TF-IDF), and
latent semantic analysis (LSA)) in an attempt to comprehensively understand the nuanced
relationship between sentiment and stock-market dynamics. The work of Zheng et al. [123]
also utilized ML for stock-price prediction. However, they introduced the bat algorithm
to optimize the three free parameters of the SVR model to create the BA-SVR hybrid
model. This model was then employed to forecast the closing prices of 18 stock indexes
in the Chinese stock market. The empirical results demonstrated that the BA-SVR model
surpassed both the polynomial kernel SVR model and the sigmoid kernel SVR model,
particularly when compared to the latter models without the optimized initial parameters.
Moreover, random forests (RF) are another ML algorithm for stock-price prediction
(e.g., [15,83,110,132]). This algorithm aggregates the power of a large number of individual
decision-tree algorithms to improve the prediction performance [133]. More specifically,
it is based on the simple principle that many relatively simple models that exhibit low
correlation will outperform any of the individual models when they operate as a joint
group [134]. Lee et al., [83] examined whether textual data improves stock-price predic-
tion when RFs are employed in training their models. The authors’ proposed RF model
consisted of 2000 trees, and all the models tested were successfully trained. Their results
demonstrated that the incorporation of textual data improved next-day price-movement
prediction by 10%. Similarly, Weng et al. [110] examined the effectiveness of incorporating
both textual data and technical data in stock-price prediction. Applying decision trees
was one of their tested methods alongside SVM and NN. The authors also highlighted
that incorporating textual data can improve prediction results. Patel et al. [132] compared
RF to NN and SVM and concluded that RF outperforms the others. However, it is worth
noting that the applied NN was very simple, with only one hidden layer containing two
neurons. Later, Zhang et al. [15] utilized RF as a critical part of training in a proprietary
stock prediction model. They concluded that the relatively effective performance of their
prediction model (in terms of accuracy and returns) is due to the incorporation of RF as
one of the integrated models used as a learning method.
Algorithms 2024, 17, 234 6 of 29

Support vector machines (SVM) is another promising model in the domain of ML. This
type of model separates data into distinct classes via a decision boundary and then seeks to
maximize the margin. Due to the nonlinearity of stock-price data, SVM is an appropriate
technique for prediction, as it can project the data points into a higher dimensional space
by performing a function on the data that makes the classes linearly separable. As a result,
many studies (e.g., [100,135]) have utilized SVM for stock-price prediction. Additionally,
several studies have successfully incorporated textual data into their SVMs for stock predic-
tion [75,91]. Nevertheless, the SVM algorithm is computationally time-consuming. Thus,
some studies (e.g., [16]) utilized a kernel function to improve the efficiency of mapping data
into a higher dimensional space. Similar to SVM, support vector regression (SVR) employs
the same components and techniques; however, rather than classifying the mathematical
operations, these operations are tasked with regressing the data points. Li et al. [79] uti-
lized a multiple-kernel SVR to test whether or not the inclusion of news articles alongside
technical indicators improved the predictive power of the SVR model. The results revealed
that the multiple-kernel SVR outperformed a normal SVR model. The work of Schumaker
et al. [73] employed the AZfintext qualitative stock-price predictor and regressed stock
quotes and financial news article data as inputs into an SVR algorithm for stock-price
prediction. The authors examined whether incorporating sentiment-based analysis into
the AZfintext system would improve the stock direction prediction accuracy. The results
showed that incorporating sentiment analysis into the AZfintext system did not improve
the overall prediction accuracy.
Another widely employed method for stock-price prediction involves utilizing the
auto-regressive integrated moving average (ARIMA) model, which is also commonly
known as the Box–Jenkins model in finance. ARIMA is specifically designed for training
in forecasting time-series data [136]. Functioning as a generalized random walk model,
ARIMA is finely tuned to eliminate residual autocorrelation, which is a statistical measure
of the correlation between variables based on past values. As a generalized exponen-
tial smoothing model, ARIMA can incorporate long-term trends and seasonality into its
predictions [137]. In recent stock-price prediction research, ARIMA has frequently been
integrated either into other ML algorithms or used as a benchmark for comparison. An
early example is the work of Pai and Lin [66], who developed and tested a hybrid ARIMA
and SVM model. They recognized ARIMA’s declining popularity and demonstrated its
utility in enhancing ML models. The results revealed that the proposed hybrid ARIMA-
SVM model outperformed both the standalone ARIMA and SVM models. A subsequent
study conducted by Adebiyi et al. [82] compared ARIMA with a three-layer NN and found
that the NN consistently outperformed ARIMA in most cases. The graphical representation
of their ARIMA predictions indicated a linear pattern, thereby emphasizing its limitation
in providing value-based forecasting. Similarly, Chong et al. [109] discovered that NNs
significantly outperformed the benchmark autoregressive model in stock-price prediction.

2.2. NN Techniques in Stock-Price Prediction

Neural networks (NN) are an ML technique that has recently gained popularity
in the domain of stock-price prediction. The earliest application of NNs in predicting
stock-market trends dates back to the early 1990s [12,15]. More recently, NNs have be-
come widely adopted, given their proven practical and powerful attributes in stock-price
prediction [18–23,35,37,48,49,55,59,60,117,138].
Sun et al. [80] utilized data on stock movements to analyze whether trading behavior
could be mapped and used to predict stock prices. For each individual stock, they first
analyzed stock trading activities and mapped a network. Then, they classified the trading
relationships and grouped them into appropriate categories. Next, they employed Granger
causality analysis to prove that the stock prices were indeed correlated with the different
trading categories. Moreover, they used a simple three-layer feed-forward NN to test the
trading predictability power. The NN incorporated technical indicators as well as trading
indicators. The results revealed that the NN performed well overall. Lastly, it is worth
Algorithms 2024, 17, 234 7 of 29

emphasizing that this positive result can be considered relatively intuitive, as it is generally
well-known that the activities of one group of traders influence another.
Furthermore, Geva and Zahavi [87] investigated whether market data, simple news
item counts, business events, and sentiment scores could improve various ML algorithms
in stock-price prediction. The authors considered NNs, decision trees, and basic regression.
The results demonstrated that among the algorithms tested, only the NN could fully exploit
the more intricate nature of the proposed sentiment/news inputs. The other models could
not take advantage of these inputs, given the complicated relationship between price and
sentiment/news indicators. Zhuge et al. [111] utilized Shanghai Composite Index data and
emotional data. Emotional data, in this case, involved sentiment analysis from the news
and microblogs that were related to a specific company. The authors demonstrated that 15
input variables, comprised of sentiment and technical indicators, could successfully predict
a Chinese company’s stock opening prices.
Kooli et al. [117] proposed a simple NN to examine whether the inclusion of accounting
variables (generated from the release of accounting disclosures) improved the prediction
accuracy of the NNs. The results showed that combining 48,204 daily stock closing prices
of 39 companies with the respective accounting disclosure variables improved the NNs’
prediction quality. However, this level of improvement drastically dropped when the
NN predicted prices in 2011, a time of civil unrest in Tunisia. This extreme example is
noteworthy, as it portrays how an observed variable (i.e., one that was able to consistently
improve the model accuracy) could lose its impact when emotional events occurred.
Vantstone et al. [122] investigated if the prediction of the price of 20 Australian stocks by a
neural network autoregressive (NNAR) model could be improved with the inclusion of inputs
in the form of counts of both news articles and tweets. The sentiment-based indicators used
in this study were generated by Bloomberg. These types of sentiment-based indicators are
increasingly becoming available. Additionally, due to the overall improvement in data-mining
techniques, these indicators should theoretically be more reliable than ever. Their study found
that the NNAR that incorporated the Bloomberg-generated news and Twitter-based sentiment
indicators had a higher quality of stock-price predictions. Due to the indicators being created
and readily available by Bloomberg, no text/data-mining models had to be utilized by the
authors. The resulting ease of incorporating news and Twitter indicators into NNs was equal
to any other technical indicator.
As shown in Table 1, LSTM stands out as one of the most popular NN techniques in
the price prediction literature. This technique was introduced by Olah [139] in 2015 and
has since been utilized in numerous studies. For example, Jin et al. [19] introduced an
LSTM-based model that incorporated sentiment analysis and employed empirical modal
decomposition to break down stock-price sequences. The authors’ approach enhanced
prediction accuracy by leveraging LSTM’s capacity to analyze relationships among time-
series data through its memory function. Furthermore, Lu et al. [21] introduced the
CNN-BiLSTM-AM method, which combined convolutional neural networks, bidirectional
long short-term memory, and an attention mechanism. The authors’ model aimed to predict
the following-day stock closing prices by extracting features using CNN, using BiLSTM
for prediction, and employing an attention mechanism to capture feature influences at
different times. A comparative analysis against seven other methods for predicting stock
closing prices on the Shanghai Composite Index revealed the superior performance of the
CNN-BiLSTM-AM method in terms of MAE and RMSE.
Vijh et al. [20] employed artificial neural network and random-forest techniques to
predict next-day closing prices for stocks across various sectors. By utilizing financial data,
such as open, high, low, and closed prices, they created new variables as inputs in the
model. The evaluation based on standard indicators RMSE and MAPE highlighted the
efficiency of their models for predicting stock closing prices. Wu et al. [60] explored LSTM
for stock-price prediction by introducing the S_I_LSTM method in incorporating multiple
data sources and investor sentiment. The authors’ approach leveraged sentiment analysis
based on convolutional NNs for calculating investors’ sentiment index and combined this
Algorithms 2024, 17, 234 8 of 29

with technical indicators and historical transaction data as features for LSTM prediction.
The results indicated that the predicted stock closing prices aligned more closely with
the accurate closing prices compared to the traditional LSTM methods. Lastly, Kurani
et al. [23] presented a comprehensive study on the use of artificial neural networks (ANN)
and support vector machines (SVM) for stock forecasting to provide further insights into
the application of machine-learning techniques in the financial domain.
Khan et al. [127] applied algorithms to social media and financial news data to assess
their impact on stock-market prediction accuracy over a span of ten subsequent days.
They conducted feature selection and minimized spam tweets in the datasets to enhance
prediction quality. Furthermore, the study involved experiments to identify stock markets
that were challenging to predict and those heavily influenced by social media and financial
news. The researchers also compared the outcomes of various algorithms to determine a
reliable classifier. The results recommended random forest for stock-trend prediction due
to its consistent results in all the cases. Finally, deep-learning techniques were employed,
and classifiers were combined to maximize prediction accuracy. The experimental findings
revealed the highest prediction accuracies of 80.53% and 75.16% using social media and
financial news data, respectively. Shaban et al. [128] introduced a new system based on
deep learning to predict the stock price. They combined LSTM and bidirectional gated
recurrent unit (BiGRU) to predict the closing price of the stock market. Then, they applied
the proposed method to some stocks and predicted their close price 10 and 30 min before
the actual time. Liu et al. [44] have another study that considered news in the market
price prediction. They developed a model based on TrellisNet and a sentiment attention
mechanism (SA-TrellisNet) to predict stock-market prices. They integrated the LSTM and
CNN models for sentiment analysis while employing a sentiment attention mechanism
to allocate weights and a trellis network for stock prediction. The hybrid model includes
three components: sentiment analysis, sentiment attention mechanism, and the prediction
model. Finally, they compared the proposed model with general methods to demonstrate
its performance.
Recently, the substantial impact of cryptocurrencies on the global financial markets
has led to an increasing number of price prediction studies in academic research. Ammer
and Aldhyani [45] proposed an LSTM algorithm to forecast the values of four types of
cryptocurrencies: AMP, Ethereum, Electro-Optical System, and XRP. To overcome the
problem of price-fluctuation prediction, they proposed an LSTM that captures the time
dependency aspects of the prices of cryptocurrencies and proposed an embedding network
to capture the hidden representations from linked cryptocurrencies. They then employed
these two networks in conjunction with each other to predict price. In addition, Belcastro
et al. [129] introduced a methodology aimed at optimizing cryptocurrency trading decisions
to enhance profit margins. Their approach integrates various statistical, text analytics, and
deep-learning methodologies to support a recommendation trading algorithm. Notably,
the study leverages supplementary data points, such as the correlation between social
media activity and price movements, causal relationships within price trends, and the
sentiment analysis of cryptocurrency-related social media, to generate both buy and sell
signals. Finally, the researchers conducted numerous experiments utilizing historical data
to evaluate the efficacy of the trading algorithm, achieving an average gain of 194% without
factoring in transaction fees and 117% when accounting for fees. Lastly, Al-Nefaie et al. [130]
employed AI algorithms, including the gated recurrent unit (GRU) and MLP, for forecasting
Bitcoin prices. They evaluated their models using various metrics, such as mean square
error (MSE), root mean square error (RMSE), Pearson correlation (R), and R-squared (R2),
to assess performance. Their findings indicated that the MLP method outperformed the
GRU approach. Given these studies, the primary contributions of the current study to the
extant literature are as follows.
• We are the first to use Bloomberg Twitter and news publication count variables as
critical inputs for stock-price prediction within the North American context;
Algorithms 2024, 17, 234 9 of 29

• We use a novel approach in employing Twitter and news publication count variables
as inputs into multi-layer perceptron (MLP) and long short-term memory (LSTM)
NNs. This novel approach seeks to assess the influence of these variables on various
NN architectures, allowing us to concurrently evaluate and contrast the stock-price
prediction performance of both models;
• We focus on examining the existence of a potential notable decline in model perfor-
mance during periods rife with market panic (e.g., the COVID-19 pandemic). There-
fore, we seek to provide insights into the robustness of the proposed models under
stressful conditions in financial markets.

3. Problem Definition and Formulation

Given the significant influence of social media and news on investor behavior in
the stock market [140,141], we employ machine-learning (ML) techniques for leveraging
such information to predict stock prices in the North American context. Predicting stock
prices accurately, especially during periods of heightened market uncertainty, presents a
unique challenge for existing prediction models. Notably, stock-price volatility experiences
a substantial fluctuation during economic downturns [142] which, in turn, adds complexity
to accurate forecasting. Trade volume serves as a more effective predictor of panic-induced
volatility than the traditional inputs commonly used in the literature (e.g., [143]). Fur-
thermore, some studies (e.g., [32,144,145]) revealed that investors’ reactions to news differ
significantly during times of market panic. Therefore, beyond the development of an
ML-based prediction model, our study investigates the public panic’s impact on the price
prediction. Hence, we seek to compare the predictive efficacy of the developed model in
periods of public panic with its performance in more stable market conditions. This dual
focus aims to provide an understanding of the model’s reliability under varying market
conditions, which enhances its practical applicability in real-world scenarios.
Both the accuracy and applicability of NN models in predicting stock prices based on
technical variables have been explored extensively in recent years
(e.g., [21,23,38–41,49,55,59,60,89,124,146]). Building on this foundation, our study lever-
ages NN techniques to incorporate market-sentiment information into the stock-price
prediction process. According to the literature, NN techniques fall into five main categories:
(i) multiple-layer perceptron (MLP) (e.g., [15,37,94,102,113,116,121,138]), (ii) long short-term
memory (LSTM) (e.g., [52,147]), (iii) convolutional NN (CNN) (e.g., [37,119]),
(iv) recurrent NN (RNN) (e.g., [17,19]), and (v) heuristic (e.g., [89,96,138]).
Multiple-layer perceptron (MLP) and long short-term memory (LSTM) have emerged
as the most commonly applied techniques in stock-price prediction. Consequently, we
adopt these two prevalent network types for formulating an advanced stock-price predic-
tion model. To this end, several crucial design decisions on NN configuration should be
made. These decisions encompass determining the optimal number of layers, nodes, and
training epochs and selecting the appropriate cost functions and optimizers. It is worth not-
ing that, within the extensive body of research utilizing NN as the primary algorithm, there
is a notable absence of standardization concerning their design. Moreover, the predominant
focus in NN design research lies in hyper-parameter optimization through algorithms
rather than in establishing empirical standards. This lack of proven baselines hinders
researchers from effectively comparing and advancing various NN architectures [148].

3.1. Multiple-Layer Perceptron (MLP) Network

A multiple-layer perceptron (MLP) is a specific feed-forward NN designed to analyze
non-linear data effectively. It comprises multiple layers of perceptrons, each functioning
as an algorithm for supervised learning. MLP’s capacity to handle chaotic and non-linear
data is enhanced by increasing the number of perceptrons. A concise representation of the
Algorithms 2024, 17, 234 10 of 29

input variable processing within an MLP is depicted through the following mathematical
expression: ! !
Yn (t) = Φ ∑ wno .Φ ∑ wkhn Xk (t) + bnh + bno (1)
n k

where Xk denotes the value of the kth variable input into the perceptron, wk represents the
corresponding weight, b is the bias, and Φ is the activation function, including the sigmoid
and tanh functions ([23,49]). Furthermore, Kn denoted the number of neurons in the nth layer.
These layers fall into three categories: (i) input, (ii) output, and (iii) hidden layers. Finally, “o”
represents values associated with the output neurons, and “h” represents values associated
with the hidden neurons. While it has been established that a single hidden layer in a neural
network can approximate any univariate function [149], stock-price prediction inherently
involves multi-variate complexities. Given the successful approximation of multi-variate
functions with just two hidden layers in a simple feed-forward network [150,151], we adopt
an NN configuration with two hidden layers for the MLP network.
Although various cost functions have been explored for the MLP network in the
literature, we specifically opt for the mean squared error (MSE). The effectiveness of MSE in
optimizing NNs has previously been demonstrated [152], thereby showcasing its ability to
handle an extensive magnitude of training data [153]. Additionally, we employ the Adam
(Adaptive Moment) optimization algorithm. This algorithm was first introduced by Hasan
et al. [154] and has been tested by researchers at OpenAI and Google Deepmind [154,155].
Adam has been found to be successful in handling non-stationary data while also being
able to handle both sparse and noisy gradients.
The number of hidden nodes is another parameter that should be determined for NN
configuration. Unlike epoch iterative testing, a generally agreed-upon formula can guide
the establishment of a testing range for the optimal number of hidden nodes. This formula
is presented as follows:
Ns
Nh = (2)
α( Ni + No )
where Nh denotes the number of hidden nodes, Ni denotes the number of input neurons,
and No and Ns denote the number of output neurons and samples in the training dataset,
respectively. Finally, α is an arbitrary scaling factor, usually 2–10. With a training dataset
comprising 1000 samples, an input layer of six neurons, and an output layer of one neuron,
we have opted for 13 hidden neurons in the MLP. Another critical aspect of the design
process involves determining the number of epochs. Since the optimal number of epochs
should be established on a case-by-case basis, we have conducted iterative testing to
identify the point at which the loss function ceases to decrease. It is worth noting that,
despite the iterative approach, we have set a maximum limit of 5000 epochs for the MLP
for the sake of simplicity.

3.2. Long Short-Term Memory (LSTM) Networks

LSTM networks represent a prominent class of NN extensively employed in stock-
price prediction research. LSTMs, categorized as recurrent NNs, distinguish themselves
from their feed-forward counterparts by incorporating the previous output as an input for
the subsequent timestamp. Unlike feed-forward NNs like MLP, which treat the first and
1000th inputs similarly, recurrent networks process data sequentially. This characteristic
makes LSTMs particularly suitable for addressing the inherent linearity of MLPs. Despite
the resemblance of basic linear transformations within an LSTM to those in an MLP, the
pivotal feature contributing to the widespread adoption of LSTMs is the integration of
gates and states, which fundamentally alter the nature of the NN.
To construct the LSTM network, similar to the MLP, we specify two hidden layers
with the objective of minimizing the mean squared error (MSE) as the cost function. Addi-
tionally, based on the demonstrated effectiveness in prior studies [156,157], we have chosen
to combine LSTM and the Adam optimization algorithm for this LSTM NN. Moreover, con-
Algorithms 2024, 17, 234 11 of 29

sidering the network’s complexity, we set the number of hidden neurons and the maximum
number of epochs to 60 and 1000, respectively, when the training sample size is 1000.

4. Model Construction
The current study aims to examine the specific impact of Twitter and news count
variables on stock-price prediction, with a primary emphasis on the North American
context. To this end, as shown in Figure 1, we first selected a group of stocks for inves-
tigation. A subset of the chosen stocks is collectively known as “FAANG”, an acronym
representing Facebook Inc. (FB), Apple Inc. (AAPL), Amazon.com Inc. (AMZN), Netflix
Inc. (NFLX), and Alphabet Inc. (GOOG). Coined by former Goldman Sachs fund manager
Jim Cramer [158], FAANG stocks are particularly significant for North American investors
and the overall stock market and are publicly traded on the NASDAQ. As of July 2020,
these five companies boasted a combined market capitalization of USD 4.1 trillion, which
constituted 16.64% of the total S&P 500 market capitalization. The S&P 500, an index com-
prised of 500 companies, historically represents 70% to 80% of the total US stock-market
capitalization. The substantial contribution of FAANG stocks to the S&P 500 highlights
their broader importance in shaping the North American stock market. The movements
Algorithms 2024, 17, x FOR PEER REVIEW
of FAANG stocks directly influence North American investors’ perceptions of the 12 of 29
overall
market and thereby impact trading decisions.

Stock Selection Data Splitting

Analysis & Evaluation

Train Set

Data Gathering/Extraction
Normal Test Set Panic Test Set
Data Preparation
Modeling
Input Parameters Selection MLP LSTM
(Feature Selection & Normalization)

Figure
Figure 1.
1. The
The methodology
methodology used
used for
for analysis.
analysis.

4.1. Input Parameters

The choice Selectionas a focal point stems from the unique characteristics of the
of Walmart
Bloomberg Twitter and
To establish the input news count variables.
parameters These variables
for our analysis, we reviewed gauge
the the frequency
parameters of
com-
mentions a specific company receives on Twitter or on digital news platforms.
monly explored in other studies. As depicted in Table 2, the most frequently examined Given the
recent surge
inputs includein the
openpopularity of price,
price, high digitallow
andprice,
socialclose
media, particularly
price, moving amongst
average, aand
younger
trade
volume. Among these, the four most prevalent ones are open, high, low, and close.count
demographic, there is potential variation in the significance of Twitter and News Con-
variables for companies adhering to traditional business models versus
sequently, the moving average (representing the price averaged over a specified number those embracing
non-traditional
of periods) and models. To investigate
trade volume thisthe
(indicating distinction,
number of wetrades
optedexecuted
to conductinaacomparative
day) are in-
analysis between Walmart and Amazon. Our aim was to discern any
corporated into only half of the models. In contrast, the remaining half incorporates disparities in how
the
Twitter and news count variables manifested for a conventional brick-and-mortar retailer
four most common inputs alongside Bloomberg-generated Twitter and news count data.
like Walmart versus a more contemporary retailer like Amazon. Similarly, the selection
Several reasons justify the exclusion of the least utilized variables. Primarily, this choice
of Ford and Tesla was motivated by the desire to contrast a more traditional automobile
facilitates a more direct comparison of the individual impact of these variables on enhanc-
manufacturer with a purely electric one. Ford adheres to the traditional car dealership
ing stock-price prediction. Likewise, since the total number of variables remains constant,
model while Tesla’s showrooms are commonly situated in malls and have all transactions
any observed performance improvement cannot be attributed to an increase in data vol-
occurring online [159,160]. Exploring the potential impact of this dichotomy in business
ume. Additionally, in the context of NN, adjustments to design decisions, such as the
models on both the Twitter and news count variables, as well as their effectiveness as inputs
number of neurons, are influenced by changes in the input variable count. Hence, for a
in stock-price prediction models, adds an intriguing dimension to our analysis.
methodologically sound comparison, we maintain consistency in the NN’s design, irre-
spective
4.1. InputofParameters
the variable set in use. It is worth noting that all variables are sourced from
Selection
Bloomberg and are formatted in a comma-separated value structure. The dataset we used
To establish the input parameters for our analysis, we reviewed the parameters com-
spanned from January
monly explored 2015
in other to MayAs
studies. 2020. We chose
depicted this time
in Table 2, theperiod
most in order to consider
frequently examined a
more
inputscomprehensive coverage,
include open price, highthat is, to
price, include
low price,both
closenormal market conditions
price, moving (i.e.,trade
average, and pre-
COVID-19) as well as market-panic conditions (i.e., a few months post-COVID-19 out-
break). The definitions for each variable can be found in Table 3.

Table 2. The indicators that have been investigated in the literature.

No Reference OP CP HP LP TV 30MA BB TB MACD

Algorithms 2024, 17, 234 12 of 29

volume. Among these, the four most prevalent ones are open, high, low, and close. Con-
sequently, the moving average (representing the price averaged over a specified number
of periods) and trade volume (indicating the number of trades executed in a day) are
incorporated into only half of the models. In contrast, the remaining half incorporates
the four most common inputs alongside Bloomberg-generated Twitter and news count
data. Several reasons justify the exclusion of the least utilized variables. Primarily, this
choice facilitates a more direct comparison of the individual impact of these variables on
enhancing stock-price prediction. Likewise, since the total number of variables remains
constant, any observed performance improvement cannot be attributed to an increase in
data volume. Additionally, in the context of NN, adjustments to design decisions, such
as the number of neurons, are influenced by changes in the input variable count. Hence,
for a methodologically sound comparison, we maintain consistency in the NN’s design,
irrespective of the variable set in use. It is worth noting that all variables are sourced
from Bloomberg and are formatted in a comma-separated value structure. The dataset
we used spanned from January 2015 to May 2020. We chose this time period in order to
consider a more comprehensive coverage, that is, to include both normal market conditions
(i.e., pre-COVID-19) as well as market-panic conditions (i.e., a few months post-COVID-19
outbreak). The definitions for each variable can be found in Table 3.

Table 2. The indicators that have been investigated in the literature.

No Reference OP CP HP LP TV 30MA BB TB MACD

1 Zhang et al. [15] ■ ■ ■ ■ ■
2 Ding and Qin [17] ■ ■ ■
Hiransha et al. [37] ■
3 Rundo et al. [52] ■ ■ ■ ■
4 Wu et al. [60] ■ ■ ■ ■ ■
5 Hafezi et al. [89] ■ ■ ■
6 Gocken et al. [94] ■ ■ ■
7 Zhou et al. [96] ■
8 Qiu and Song [102] ■ ■
9 Jeon et al. [113] ■ ■ ■ ■
10 Ebadati and Mortazavi [116] ■
11 Zhou et al. [119] ■ ■ ■ ■ ■ ■ ■ ■
12
13 Gocken et al. [121] ■ ■ ■ ■ ■ ■
14 Moghaddam et al. [138] ■
15 Nelson et al. [147] ■ ■ ■ ■
16 Dey et al. [161] ■ ■ ■ ■ ■
OP: open price, CP: close price, HP: high price, LP: low price, TV: trade volume, MA: moving avg, BB: Bollinger
band, TB: turnover bias, MACD: MACD index.

Table 3. Definitions of input parameters.

Variable Name Definition

Open Price The dollar value of the first trade since the market opened
High Price The highest dollar value trade of the day
Low Price The lowest dollar value trade of the day
Close Price The dollar value of the last trade before the market closed
30-Day Moving Average The average dollar value of one share in the last 30 days
Trade Volume The total quantity of shares traded all day.
Represents the difference between the number of tweets expressing
Twitter Count (TC) positive sentiment and those expressing negative sentiment towards the
parent company over a 24-h period.
Represents the total number of news publications mentioning the parent
News Publication Count (NC)
company over a 24-h period.

As mentioned, the data utilized in our study was extracted from Bloomberg, a rep-
utable financial data provider widely used in academic and industry research. To ensure
Algorithms 2024, 17, 234 13 of 29

accuracy and reliability, we accessed Bloomberg’s database and retrieved the required
information using their data-export functionality. Bloomberg’s database is renowned for its
accuracy, timeliness, and depth of coverage, making it a preferred choice for researchers
and practitioners in the financial industry. Specifically, we employed Bloomberg’s Excel
API to extract the data directly into an Excel format. This API allowed us to access various
financial data, including stocks’ open prices, high prices, low prices, and close prices daily
from 1 January 2015 to 31 May 2020. It also includes trading volume, the 30-day moving
average, the number of tweets with positive sentiment, the number of tweets with negative
sentiment, the number of tweets with neutral sentiment, and the news publication count. To
capture the sentiments of tweets, we consider the difference between positive and negative
sentiment tweets as the number of Twitter counts (TC). However, we do not involve the
sentiment in the news publication count (NC) parameter and consider the total number of
news publications mentioning the parent company over a 24 h period. The dataset includes
1412 data items for each stock, and the attributes were selected based on their potential
significance for analyzing stock-market dynamics and sentiment and their availability
within the Bloomberg database. The daily average number of tweets and news publications
for the selected companies are Apple (4952 tweets, 2506 news), Amazon (3009 tweets,
757 news), Facebook (3912 tweets, 701 news), Netflix (1556 tweets, 637 news), Google
(3930 tweets, 1196 news), Walmart (743 tweets, 375 news), Tesla (2218 tweets, 594 news),
and Ford (192 tweets, 271 news). Furthermore, by utilizing the Excel API, we could in-
tegrate the data into our analysis workflow, facilitating further processing and analysis.
Additional details regarding the collected data can be found in Table A1 in Appendix A.

4.2. Data Splitting, Modeling, and Analysis

Following the collection of daily stock data from Bloomberg, the dataset is divided
into two subsets: (i) a technical set (T) and (ii) a technical-plus set incorporating TC and
NC (T+). To address skewness, a log transformation is applied to the TC and NC. After this
transformation, all TC and NC variables exhibit an acceptable level of skewness. Min–max
normalization is then applied to all the remaining variables. Subsequently, the dataset is
further categorized into three sets: (i) a training set, (ii) a normal test set, and (iii) a panic
test set. The training set encompasses daily variables from January 2015 to 2019, the normal
test set comprises variables from January 2019 to November 2019, and the panic test set
includes variables from January 2019 to May 2020. For each stock, four price prediction
models are constructed. Each model undergoes training and execution of the NN five
times and the predictions are averaged to obtain the final model prediction. This approach
addresses the stochastic nature of NNs, which can result in slight performance variances.
The first two models utilize LSTM, where one model incorporates OP, HP, LP, CP, TV, and
30MA as inputs and the other uses OP, HP, LP, CP, TC, and NC as inputs. The next day’s
CP is defined as the target variable in these models. The third and fourth models are built
using MLP with the respective input sets.
Each defined model undergoes testing on both the normal test set and the panic test set.
The accuracy measure for model evaluation is the root mean squared error (RMSE), where
lower RMSE values indicate better stock-price predictions. RMSE is selected over other
error measures due to its expression in the original unit being measured. This characteristic
makes RMSE useful for analyzing the error gap between the expected and predicted
values [162]. Moreover, RMSE assigns a higher weight to larger errors compared to other
measures, making it particularly suitable for domains where significant errors in accuracy
are undesirable [161]. Incorporating the magnitude of error into RMSE is pertinent in stock-
price prediction research, as larger errors in prediction can potentially lead to greater losses
in buy or sell decisions. Recent stock-price prediction research adopted RMSE as the error
measure for comparative analysis of different companies’ stock-price predictions [163].
Algorithms 2024, 17, 234 14 of 29

5. Results and Discussion

In this section, we investigate the potential enhancement of stock-price prediction
by including Twitter and news variables and explore whether their influence varies sig-
nificantly during periods of market panic. In this regard, we employ the constructed
models to forecast the selected stock prices and incorporate the identified input parameters.
To distinguish the distinct impact amongst various parameters that might influence the
prediction accuracy, we apply both MLP and LSTM models with a technical set (T) and a
technical-plus set (T+) separately. We calculated the root mean squared error (RMSE) as
the accuracy index for each model and test-set configuration. The results of the designed
experiments are presented in Table A2 of Appendix A.
To evaluate the improvement in stock-price prediction between T+ and T models’
performance, we assessed the impact of Twitter and news variables by subtracting the
former’s RMSE from the latter’s. Furthermore, we examined whether this impact differed
in a panic market by comparing the results obtained in a normal market with those achieved
during a panic market. These results were then subjected to a statistical test for model
comparison. Additionally, when exploring the impact of input-data type, test-data type,
and NN selection, our analysis emphasized group-level analysis rather than a focus on
individual performances. Concentrating on the average relative difference between T+
input data and T input data for the same test set as a group aimed to provide insights
independent of specific decisions (e.g., individual stock selection) that may influence stock-
price prediction performance. This approach holds the potential for increased practical
replicability and utility. Therefore, the impact of T+ variables is analyzed based on the
RMSE difference compared to models utilizing exclusively technical variables.

5.1. Comparing Predictive Models under Panic and Normal Circumstances

To investigate the models’ performance under panic and normal circumstances, we
created two separate test sets, the normal set (nor-set) and the panic set (pan-set), for each
stock. The reason behind creating two sets for each stock is due to the nature of Bloomberg,
which generates TC and NC variables relating to the number of times the public mentions
a company. Therefore, interesting insights can be gained by analyzing the performance
of all models in times of increased panic in the stock market. It is a general consensus
that, in the first half of 2020, North America experienced widespread panic related to the
performance of the stock market and the economy due to the impact of the global COVID-
19 pandemic outbreak [164]. The average RMSE for models tested on the pan-set was
0.0670, while the average RMSE for the nor-set was 0.0354. There is an apparent decrease
in performance when testing models on the nor-set compared to the pan-set. To test the
statistical significance of this difference, a Wilcox signed-rank test was conducted between
the RMSEs of the models tested on these sets. The results showed that this difference in
performance is significant, even at a 1% level (p-value~0.001).
To better understand this discrepancy in performance, an analysis of how the variables
of interest differed between the different test sets can be found in Appendix. This analysis
provides insights into the changes in mean, standard deviation, and coefficient of variation
(CV) of the T+ variable (including TC, NC, and close price) across the overall dataset,
training dataset, normal test sets, and panic test set. It is worth noting that the normal
test set spans from January 2019 to November 2019 (including 218 observations), whereas
the panic test set includes 151 observations, covering December 2019 to May 2020 (this
is mainly due to the date of a report issued by the World Health Organization which
resulted in panic in the US stock market [34]). Additionally, the training data encompasses
1043 observations, covering the period from January 2015 to December 2018. By splitting
our analysis into these three distinct sections, we allow ourselves a starting point from
which to theorize what may have caused an average decrease in performance for the tests
done on the pan-set versus the nor-set.
Table 4 shows a summary of the analysis of two T+ variables, the TC and the NC,
and the price differences between the pan-set and the training data. The companies are
Algorithms 2024, 17, 234 15 of 29

displayed in descending order by mean RMSE for the pan-set. As shown in Table 4, when
comparing the bottom 50% to the top 50% of the list, the CV for the TC and NC variables
exhibit a 19% and 27% absolute difference, respectively. In addition, the percentage change
of mean TC and NC variables shows an 18% and 19% absolute difference, respectively. The
difference in variance between the data the models are learning from and the additional
period added to the nor-set to make it the pan-set are critical to our analysis. The average
percentage change in mean for the price data, as well as the average CV of the price data,
exhibits a similar difference between the panic data and training data for all the stocks
being predicted. This discrepancy in variable variance between the training and test sets
can be the primary cause of the drop in performance when it comes to the pan-set compared
to the nor-set.

Table 4. T+ variables’ mean and coefficient of variation in the panic test set and training data.

TC NC Price
Train vs. Pan-Set Train vs. Pan-Set Train vs. Pan-Set
CV −7% −15% −14%
Google
Mean −46% 23% −49%
CV −19% −7% −37%
Amazon
Mean −55% −17% 114%
CV −27% −35% 6%
Walmart
Mean −26% 132% 129%
CV −48% −46% −16%
Facebook
Mean −10% −71% 47%
CV 22% 38% −45%
Netflix
Mean −69% −53% 117%
CV −59% −20% −16%
Apple
Mean −71% 0% 105%
CV −103% −2% 12%
Ford
Mean −14% 34% −44%
CV −35% −13% −10%
Tesla
Mean −57% 13% 52%
Top 50% −25% −26% −15%
CV
Bottom 50% −44% 1% −15%
Top 50% −35% 17% 60%
Mean
Bottom 50% −53% −2% 57%

5.2. Comparing the Impact of T and T+ Variables

To determine whether using T+ variables leads to a significant improvement in price
prediction, we analyzed instances where the utilization of T+ variables, as opposed to T
variables, resulted in improvements in terms of RMSE. As shown in Table 5, there are
25 instances where using T+ variables led to an RMSE reduction. Among those instances
that do not lead to improvement, four are associated with Walmart, two are related to Tesla,
and one is concerned with Facebook. Furthermore, among these seven instances, four are
associated with the LSTM model, and three are related to the MLP model. It is also worth
noting that Tesla had a model that improved by 42% and another that lowered by 89%. In
essence, this signifies an ample need for researchers and traders to test across a wide variety
of scenarios before deciding to use a variable as an input in a stock-price prediction model.
Algorithms 2024, 17, 234 16 of 29

Table 5. The RMSE change when utilizing T vs. T+ variables.

NN Test Set Company T RMSE T+ RMSE RMSE ± RMSE% ±

Amazon 0.06246 0.06018 0.00228 4.0%
Apple 0.12860 0.12711 0.00149 1.0%
Facebook 0.07699 0.07392 0.00306 4.0%

PANIC
Ford 0.13641 0.07553 0.06087 45.0%
Google 0.06271 0.04938 0.01333 21.0%
Netflix 0.11226 0.09216 0.02010 18.0%
Tesla 0.17632 0.18758 −0.01126 −6.0%
Walmart 0.05845 0.07278 −0.01434 −25.0%
LSTM

Amazon 0.04672 0.04210 0.00462 10.0%

Apple 0.05848 0.05299 0.00549 9.0%
Facebook 0.04404 0.03668 0.00737 17.0%
NORMAL Ford 0.04024 0.03007 0.01017 25.0%
Google 0.03487 0.02874 0.00613 18.0%
Netflix 0.04211 0.03706 0.00505 12.0%
Tesla 0.05344 0.10091 −0.04748 −89.0%
Walmart 0.04021 0.04251 −0.00231 −6.0%
Amazon 0.01950 0.01928 0.00022 1.0%
Apple 0.02188 0.02141 0.00047 2.0%
Facebook 0.02746 0.02407 0.00339 12.0%
PANIC

Ford 0.16027 0.03661 0.12366 77.0%

Google 0.02388 0.02069 0.00320 13.0%
Netflix 0.02553 0.02399 0.00154 6.0%
Tesla 0.08409 0.04863 0.03546 42.0%
Walmart 0.01753 0.01814 −0.00062 −4.0%
MLP

Amazon 0.02602 0.02426 0.00176 7.0%

Apple 0.02002 0.01995 0.00007 0.0%
Facebook 0.02070 0.02193 −0.00123 −6.0%
NORMAL

Ford 0.04886 0.02047 0.02839 58.0%

Google 0.02355 0.02147 0.00209 9.0%
Netflix 0.02954 0.02842 0.00113 4.0%
Tesla 0.03610 0.03531 0.00079 2.0%
Walmart 0.01278 0.01362 −0.00085 −7.0%

A summary of these impacts is presented in Table 6. As highlighted in Table 6, the

most considerable improvement belongs to Ford, which has a 36% improvement over the
stock prediction using T+ variables. Notably, both MLP and LSTM configurations for Ford
exhibit superior performance with the pan-test set compared to the normal test set. This
ranking will guide the analysis of input variables, test data, and NN selection in future
assessments. Notably, no clear trends are apparent concerning the magnitude of mean TC
and NC variables and their relationship to RMSE% improvement. Table 7, however, reveals
insights when comparing the average TC to the average NC ratios. The top 50% of the list
in terms of RMSE% reduction was associated with the data featuring an average ratio of
1.66 TC/NC. In contrast, the bottom 50% of the list exhibited an almost 57% increase in the
average TC to average NC ratio, reaching 2.92. A potential interpretation of these findings
suggests that neither the TC nor NC alone necessarily fortifies the variables, emphasizing
the nuanced interaction of factors in predictive modeling.
Algorithms 2024, 17, 234 17 of 29

Table 6. The average RMSE% improvement across all configurations.

Rank Company Average Improvement via T+

1 Ford 51%
2 Google 15%
3 Netflix 10%
4 Facebook 7%
5 Amazon 5%
6 Apple 3%
7 Walmart −10%
8 Tesla −13%
Mean 9%

Table 7. The mean Twitter count to mean news publication count ratio.

Company Avg TC Avg NC TC/NC

Ford 195.8240 274.7395 0.7128

Top 50%

Google 4005.3400 1213.2782 3.3013

Netflix 1587.4800 647.8841 2.4503
Facebook 710.2050 3983.5188 0.1783
Bottom 50%

Amazon 3058.1700 769.4627 3.9744

Apple 5060.5100 2545.8502 1.9877
Walmart 754.8970 380.7387 1.9827
Tesla 2254.9500 600.7312 3.7537

Top 50% Avg 1624.7100 1529.8552 1.6606

Bottom 50% Avg 2782.1300 1074.1957 2.9246

In general, within ML, a distinctive requirement involves training with a distinct dataset.
It is necessary to thoroughly examine input variables across various stages of both the training
and testing processes [165]. The objective is to assess the potential impact of the TC and
NC variables on enhancing stock-price prediction. To this end, we conducted comparisons
between these variables within the training data and across both test sets. The rationale behind
this comparative analysis is to ascertain whether the pre-emptive examination of input data
(prior to actual testing) could reveal the possibility of improving the accuracy of stock-price
prediction by incorporating T+ variables into the model. Table 8 provides an analysis of the
change in the CV between the training and the test sets. The CV for the variables exhibits
relative stability across both the training and the test data for all models. Although CV is
always non-negative, the change percentage can be positive or negative.
Table 9 presents an analysis of the mean T+ variables across the training data, normal
test set (nor-set), and panic test set (pan-set). Notably, there is a consistent improvement
in RMSE across all eight tests with the inclusion of T+ variables. Among the top 50% of
performers, there is an average 25% reduction in the mean TC between the training and
test sets, suggesting that, on average, the TC is 25% lower in both test sets compared to the
training set. Conversely, the bottom 50% of performers exhibit an almost 60% decrease in
the mean TC between the training data and the average of both test sets. The difference in
the mean NC between the training data and test data for the top 50% of models indicates a
40% decrease. In contrast, the bottom 50% of companies (based on the RMSE percentage
improvement due to the addition of T+ variables) show an average 25% increase in the
mean NC between the training and testing phases.
Algorithms 2024, 17, 234 18 of 29

Table 8. The variable CV change between the training and test sets.

Training NC vs

Training NC vs
Training TC vs

Training TC vs
% ± in CV of

% ± in CV of

% ± in CV of

% ± in CV of
Nor-Set NC
RMSE% +/-

Pan-Set NC
Nor-Set TC

Pan-Set TC
Company
Rank by
1 Ford −90% −93% −6% −4%
2 Google 22% 21% −2% −5%
3 Netflix −9% 2% −14% 9%
4 Facebook −42% −41% −37% −36%
5 Amazon −19% −18% −14% −7%
6 Apple −18% −31% 0% −5%
7 Walmart −31% −32% 7% 2%
8 Tesla −39% −31% −29% −29%
Top 50% Avg −30% −28% −15% −9%
Bottom 50% Avg −27% −28% −9% −10%

Table 9. The mean TC and NC in the training data, the normal test set, and the panic test set.

Mean TC Mean NC
Company
Train Nor-Test Pan-Test Train Nor-Test Pan-Test
1 Ford 212 131 149 253 334 336
2 Google 4320 3560 3140 1206 1109 1237
3 Netflix 1823 1077 907 620 956 730
4 Facebook 686 866 780 4758 1914 1741
5 Amazon 3515 1816 1734 727 1035 890
6 Apple 6229 1612 1675 2479 2864 2735
7 Walmart 881 394 388 364 439 429
8 Tesla 2499 1395 1548 488 823 930
Top 50% 1760 1408 1244 1709 1078 1011
Bottom 50% 3281 1304 1337 1015 1290 1246
Top 50% Train vs. Avg Test 75% Top 50% Train vs. Avg Test 61%
Bottom 50% Train vs. Avg Test 40% Bottom 50% Train vs. Avg Test 125%

5.3. Comparing the MLP and LSTM Models

To thoroughly analyze the MLP and LSTM models’ performance, we evaluate the
mean and variance of the RMSE across various groupings for each model. The correspond-
ing outcomes are presented in Table 10. Table 10 illustrates that the MLP consistently
outperformed the LSTM across all subsets, showcasing lower RMSE values. Additionally,
the MLP models demonstrated a lower CV, averaging 0.1693. Despite the LSTM being
more complex than the MLP, it failed to surpass the MLP in price prediction accuracy. This
finding is in line with the literature, such as the work of Hiransha et al. [37], who also found
that their LSTM model did not outperform their MLP model for stock-price prediction
over a 400-day period. However, when tested over a 10-year period, the LSTM exhibited
improved accuracy and outperformed the MLP model—this was mainly attributed to
its memory feature. In other words, the memory feature is a key advantage of LSTMs
that becomes more beneficial in larger test sets. The intuition behind this phenomenon
lies in the increased prediction period, which allows for more opportunities for memory
utilization. Consequently, the enhanced complexity of the LSTM proves advantageous in
longer prediction periods; however, the same is not true in shorter testing periods.
Algorithms 2024, 17, 234 19 of 29

Table 10. The RMSE analysis for T vs. T+ variables and the panic test set vs. the normal test set.

T Vars T+ Vars Nor-Set Pan-Set

Mean 0.073395 0.069357 0.045699 0.097053

MLP
Range 0.141446 0.158834 0.072171 0.138198
ST.Dev 0.040658 0.040633 0.016259 0.041585
CV 0.554 0.586 0.356 0.428
Mean 0.037357 0.024891 0.025188 0.037060

LSTM
Range 0.147494 0.035011 0.036088 0.142744
ST.Dev 0.035671 0.008374 0.008748 0.035683
CV 0.955 0.336 0.347 0.963
MLP vs. LSTM Average
+/−Mean −0.036038 −0.044466 −0.020511 −0.059993 −0.0403
+/−CV −0.401 0.249 0.008 −0.534 −0.1693

As discussed regarding leveraging MLP and LSTM techniques, we examined factors

influencing stock-price prediction accuracy, particularly non-technical variables such as
Twitter and news-related data. In other words, we investigated how incorporating these
variables impacts prediction performance in normal and panic market conditions. By calcu-
lating RMSE, a widely used metric in predictive modeling, we quantitatively assessed the
accuracy of these predictions. By comparing the RMSEs of models utilizing only technical
variables (T) with those incorporating additional Twitter and news variables (T+), we
analyzed how these supplementary factors contribute to improved prediction accuracy. In
addition, we conducted several statistical tests to validate observed differences in the MLP
and LSTM models’ performance. Despite the LSTM’s reputation for handling sequential
data and its inherent complexity, the study finds that MLP consistently outperforms LSTM
across various subsets. One possible reason for the superiority of MLP is the volume of
data. More specifically, the memory feature is a key advantage of LSTMs that becomes
more beneficial in larger test sets. One plausible explanation for MLP’s superiority lies
in the volume of data. Specifically, while the memory feature of LSTMs becomes more
advantageous in larger test sets due to increased prediction periods [166,167], it may not
include the same benefits in smaller datasets. This result highlights the sophisticated
relationship between model complexity and dataset size in predictive modeling.

6. Conclusions and Future Research

This study investigated the influence of Twitter count (TC) and news count (NC)
variables on stock-price prediction within both normal and market conditions. We incorpo-
rate Bloomberg Twitter and news publication count variables into MLP and LSTM neural
networks to assess their predictive influence. Additionally, we analyze these effects during
the market panic to evaluate their stability. Our methodology integrates MLP and LSTM
neural networks with technical variables (T variables), TC, and NC (creating T+ variables)
for price prediction. The models are trained on data from January 2015 to December 2018
and tested on normal (January 2019 to November 2019) and panic periods (December 2019
to May 2020). We applied statistical analyses to these results, which revealed a notable
enhancement in stock-price prediction accuracy across various model types when these
additional variables were incorporated. Furthermore, the comparison between the T and T+
variables indicated that both traders and researchers could derive substantial benefits from
including TC and NC variables as inputs in neural network-based stock-price prediction
models. This integration not only enhanced prediction accuracy and provided significant
value to traders and investors, but it also facilitated the seamless incorporation of public
opinion into prediction models. Given the escalating impact of social media on societal
perspectives, the inclusion of TC and NC variables allows traders to consider the public’s
perception of corporations and products in their analyses. This strategic utilization of
social media data empowers traders to make more informed decisions, reflecting a nuanced
understanding of market sentiment and public opinion.
Algorithms 2024, 17, 234 20 of 29

While the proposed models aimed to analyze the impact of news and tweets on stock-
price prediction accuracy, it is important to acknowledge certain limitations for further
research. First, this study focused primarily on the COVID-19 pandemic as a period
of market distress. The impact identified in this case might not capture the full range
of possible market behaviors under different types of panic conditions, crises, financial
crashes, or geopolitical events. In addition, this study used LSTM and MLP models,
which, while established, may not represent the cutting edge in predictive modeling.
More advanced techniques, like hybrid approaches, might provide better performance
or additional insights. Furthermore, the study used sentiment analysis tools to quantify
investor sentiment from news and tweets. The accuracy of these tools can vary, and errors
in sentiment classification could affect the overall findings. Finally, the impact of news and
tweets on stock prices might vary across different regions and industry sectors. The study
does not explicitly address whether the findings are consistent across various markets and
industries or if specific segments primarily drive the results.
There are several avenues to develop future studies in this area. The current config-
uration of TC and NC variables in T+ entails countable measures and shares similarities
with other variables. In contrast to more generalized Twitter and news-based indicators,
Bloomberg’s generated Twitter and news count variables employ clearly defined and repli-
cable terms linked to objective measures. Moreover, certain indicator providers, such as
Yahoo Finance, may yield identical values for TC and NC, which can be integrated into
price prediction models. However, further analysis is imperative to ascertain why these
variables demonstrate efficacy across the majority of scenarios but exhibit sub-optimal per-
formance in specific instances. Thus, exploring these cases and identifying their underlying
sources of impact may be a promising avenue for future research. Another prospective area
for future investigation involves considering whether the influence of TC and NC variables
correlates with the size of the company. More specifically, future studies are encouraged
to explore whether the inclusion of TC and NC variables equally enhances stock-price
prediction accuracy for small-, medium-, and large-sized companies.
Furthermore, the observed variations in prediction performance between MLP and
LSTM models can be attributed to the test-period length and the extent of data preprocess-
ing. Traders and researchers should be conscious of these factors when selecting stock-price
prediction models. To enhance model applicability across different scenarios, it is recom-
mended to standardize data-preparation techniques to ensure the optimal performance of
various model types. This standardization should be replicable for each use of a specific
model and thereby promote consistency in the analyses related to stock-price prediction.
Likewise, it should advance research aimed at refining neural networks. Additionally, ex-
ploring alternative neural network architectures beyond MLP and LSTM is advisable, in an
attempt to identify simpler models that may outperform stock-price prediction. Moreover,
the current study focused solely on the quantity of tweets and news (TC and NC) while
neglecting their sentiment and diverse impacts. Future research could cover the varying
impact weights of distinct tweets or news items for each stock and involve their dominant
sentiment. Such considerations can be promising areas for future research, as they may
refine prediction accuracy in the dynamic landscape of stock-market forecasting.

Author Contributions: Conceptualization, H.Z. and S.R.; methodology, H.Z., M.N. and S.R.; software,
S.R.; validation, M.N. and S.R.; formal analysis, M.N. and S.R.; investigation, H.Z. and S.R.; resources,
H.Z. and A.H.; data curation, S.R.; writing—original draft preparation, M.N.; writing—review and
editing, H.Z., S.R. and A.H.; visualization, M.N. and S.R.; supervision, H.Z.; project administration,
A.H. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Data Availability Statement: Data can be accessed on Bloomberg Market News website at https:
//www.bloomberg.com/. The dataset we used was daily and spanned from January 2015 to May
2020.
Conflicts of Interest: The authors declare no conflict of interest.
Algorithms 2024, 17, 234 21 of 29

Appendix A
Table A1. Comparing variables in the overall, training, normal test, pan-test, and panic-period test set.

Ford Google Netflix Facebook Amazon Apple Tesla Walmart

TC NC Price TC NC Price TC NC Price TC NC Price TC NC Price TC NC Price TC NC Price TC NC Price
Training Set
Mean 212 253 12.3 4320 1206 1061 1823 620 170 686 4758 133 3515 727 947 6229 2479 141 2499 488 268 881 364 78
Variance 159,345 28,773 3.52 12,068,276 378,508 34,340 3,092,910 161,330 9125 497,886 21,219,182 1281 8,044,743 190,822 207,865 50,038,010 1,775,513 1228 5,862,998 203,445 3227 913,234 58,531 134
St Dev 399.18 169.63 1.88 3473.94 615.23 185.31 1758.67 401.66 95.52 705.61 4606.43 35.79 2836.33 436.83 455.92 7073.76 1332.48 35.04 2421.36 451.05 56.81 955.63 241.93 11.58
CV 1.88 0.67 0.15 0.80 0.51 0.17 0.96 0.65 0.56 1.03 0.97 0.27 0.81 0.60 0.48 1.14 0.54 0.25 0.97 0.92 0.21 1.08 0.66 0.15
Normal Set
Mean 131 334 9 3560 1109 708 1077 956 330 866 1914 180 1816 1035 1790 1612 2864 202 1395 823 264 394 439 107
Variance 16,797 41,977 0.3966 13,358,463 295,951 3622 894,611 239,512 1255 275,603 1,315,295 238 1,266,840 222,560 11,013 2,371,030 2,337,762 825 643,150 273,027 1873 93,790 104,811 74
St Dev 129.60 204.88 0.63 3654.92 544.01 60.18 945.84 489.40 35.43 524.98 1146.86 15.43 1125.54 471.76 104.94 1539.81 1528.97 28.72 801.97 522.52 43.28 306.25 323.75 8.60
CV 0.99 0.61 0.07 1.03 0.49 0.09 0.88 0.51 0.11 0.61 0.60 0.09 0.62 0.46 0.06 0.96 0.53 0.14 0.57 0.63 0.16 0.78 0.74 0.08
Pan Set
Mean 149 336 8 3140 1237 653 907 730 343 780 1741 185 1734 890 1869 1675 2735 232 1548 930 382 388 429 111
Variance 19,947 44,343 2.31 10,181,305 323,122 8425 804,327 288,309 1773 235,199 1,103,842 376 1,171,956 223,752 37,428 1,895,009 1,812,073 2471 1,036,664 343,690 37,976 88,037 85,835 89
St Dev 141.23 210.58 1.52 3190.82 568.44 91.79 896.84 536.94 42.11 484.97 1050.64 19.39 1082.57 473.02 193.46 1376.59 1346.13 49.71 1018.17 586.25 194.87 296.71 292.98 9.43
CV 0.95 0.63 0.19 1.02 0.46 0.14 0.99 0.74 0.12 0.62 0.60 0.10 0.62 0.53 0.10 0.82 0.49 0.21 0.66 0.63 0.51 0.76 0.68 0.08
Overall
Mean 196 275 11 4005 1213 956 1587 648 214 710 3984 147 3058 769 1184 5061 2546 164 2255 601 297 755 381 87
Variance 124,300 34,101 6 11,703,649 363,891 59,469 2,666,375 196,237 12,979 432,228 17,787,737 1562 6,884,086 204,387 327,469 41,627,970 1,797,736 3117 4,799,559 275,483 14,664 747,508 66,393 331
St Dev 352.56 184.66 2.45 3421.06 603.23 243.86 1632.90 442.99 113.93 657.44 4217.55 39.52 2623.75 452.09 572.25 6451.97 1340.80 55.83 2190.79 524.86 121.10 864.59 257.67 18.19
CV 1.80 0.67 0.22 0.85 0.50 0.26 1.03 0.68 0.53 0.93 1.06 0.27 0.86 0.59 0.48 1.27 0.53 0.34 0.97 0.87 0.41 1.15 0.68 0.21
Ford Google Netflix Facebook Amazon Apple Tesla Walmart
TC NC Price TC NC Price TC NC Price TC NC Price TC NC Price TC NC Price TC NC Price TC NC Price
Training Set
Mean 212 253 12.3 4320 1206 1061 1823 620 170 686 4758 133 3515 727 947 6229 2479 141 2499 488 268 881 364 78
Variance 159345 28773 3.52 12068276 378508 34340 3092910 161330 9125 497886 21219182 1281 8044743 190822 207865 50038010 1775513 1228 5862998 203445 3227 913234 58531 134
St Dev 399.18 169.63 1.88 3473.94 615.23 185.31 1758.67 401.66 95.52 705.61 4606.43 35.79 2836.33 436.83 455.92 7073.76 1332.48 35.04 2421.36 451.05 56.81 955.63 241.93 11.58
CV 1.88 0.67 0.15 0.80 0.51 0.17 0.96 0.65 0.56 1.03 0.97 0.27 0.81 0.60 0.48 1.14 0.54 0.25 0.97 0.92 0.21 1.08 0.66 0.15
Normal Set
Mean 131 334 9 3560 1109 708 1077 956 330 866 1914 180 1816 1035 1790 1612 2864 202 1395 823 264 394 439 107
Variance 16797 41977 0.3966 13358463 295951 3622 894611 239512 1255 275603 1315295 238 1266840 222560 11013 2371030 2337762 825 643150 273027 1873 93790 104811 74
St Dev 129.60 204.88 0.63 3654.92 544.01 60.18 945.84 489.40 35.43 524.98 1146.86 15.43 1125.54 471.76 104.94 1539.81 1528.97 28.72 801.97 522.52 43.28 306.25 323.75 8.60
CV 0.99 0.61 0.07 1.03 0.49 0.09 0.88 0.51 0.11 0.61 0.60 0.09 0.62 0.46 0.06 0.96 0.53 0.14 0.57 0.63 0.16 0.78 0.74 0.08
Algorithms 2024, 17, 234 22 of 29

Table A1. Cont.

Ford Google Netflix Facebook Amazon Apple Tesla Walmart

TC NC Price TC NC Price TC NC Price TC NC Price TC NC Price TC NC Price TC NC Price TC NC Price
Pan Set
Mean 149 336 8 3140 1237 653 907 730 343 780 1741 185 1734 890 1869 1675 2735 232 1548 930 382 388 429 111
Variance 19947 44343 2.31 10181305 323122 8425 804327 288309 1773 235199 1103842 376 1171956 223752 37428 1895009 1812073 2471 1036664 343690 37976 88037 85835 89
St Dev 141.23 210.58 1.52 3190.82 568.44 91.79 896.84 536.94 42.11 484.97 1050.64 19.39 1082.57 473.02 193.46 1376.59 1346.13 49.71 1018.17 586.25 194.87 296.71 292.98 9.43
CV 0.95 0.63 0.19 1.02 0.46 0.14 0.99 0.74 0.12 0.62 0.60 0.10 0.62 0.53 0.10 0.82 0.49 0.21 0.66 0.63 0.51 0.76 0.68 0.08
Overall
Mean 196 275 11 4005 1213 956 1587 648 214 710 3984 147 3058 769 1184 5061 2546 164 2255 601 297 755 381 87
Variance 124300 34101 6 11703649 363891 59469 2666375 196237 12979 432228 17787737 1562 6884086 204387 327469 41627970 1797736 3117 4799559 275483 14664 747508 66393 331
St Dev 352.56 184.66 2.45 3421.06 603.23 243.86 1632.90 442.99 113.93 657.44 4217.55 39.52 2623.75 452.09 572.25 6451.97 1340.80 55.83 2190.79 524.86 121.10 864.59 257.67 18.19
CV 1.80 0.67 0.22 0.85 0.50 0.26 1.03 0.68 0.53 0.93 1.06 0.27 0.86 0.59 0.48 1.27 0.53 0.34 0.97 0.87 0.41 1.15 0.68 0.21

This table provides insights into the changes in mean, standard deviation, and coefficient of variation (CV) of the T+ variable (including TC, NC, and close price) across the overall
dataset, training dataset, normal test sets, and panic test set. It is worth noting that the normal test set spans from January 2019 to November 2019, whereas the panic test set covers
December 2019 to May 2020 (this is mainly due to the date of a report issued by the World Health Organization, which resulted in a panic in the US stock market [34]). Additionally, the
training data encompasses the period from January 2015 to December 2018.
Algorithms 2024, 17, 234 23 of 29

Table A2. Experimental results.

Company Test Set Model RMSE Company Test Set Model RMSE
LSTM (T) 0.06271 LSTM (T) 0.06246
LSTM (TS) 0.04938 LSTM (T+) 0.06018
PANIC PANIC
MLP (T) 0.02388 MLP (T) 0.01950
MLP (TS) 0.02069 MLP (T+) 0.01928
Google Amazon
LSTM (T) 0.03487 LSTM (T) 0.04672
LSTM (TS) 0.02874 LSTM (T+) 0.04210
NORMAL NORMAL
MLP (T) 0.02355 MLP (T) 0.02602
MLP (TS) 0.02147 MLP (T+) 0.02426
LSTM (T) 0.07699 LSTM (T) 0.17632
LSTM (TS) 0.07392 LSTM (T+) 0.18758
PANIC PANIC
MLP (T) 0.02746 MLP (T) 0.08409
MLP (TS) 0.02407 MLP (T+) 0.04863
Facebook Tesla
LSTM (T) 0.04404 LSTM (T) 0.05344
LSTM (TS) 0.03668 LSTM (T+) 0.10091
NORMAL NORMAL
MLP (T) 0.02070 MLP (T) 0.03610
MLP (TS) 0.02193 MLP (T+) 0.03531
LSTM (T) 0.05845 LSTM (T) 0.12860
LSTM (TS) 0.07278 LSTM (T+) 0.12711
PANIC PANIC
MLP (T) 0.01753 MLP (T) 0.02188
MLP (TS) 0.01814 MLP (T+) 0.02141
Walmart Apple
LSTM (T) 0.04021 LSTM (T) 0.05848
LSTM (TS) 0.04251 LSTM (T+) 0.05299
NORMAL NORMAL
MLP (T) 0.01278 MLP (T) 0.02002
MLP (TS) 0.01362 MLP (T+) 0.01995
LSTM (T) 0.13641 LSTM (T) 0.11226
LSTM (TS) 0.07553 LSTM (T+) 0.09216
PANIC PANIC
MLP (T) 0.16027 MLP (T) 0.02553
MLP (T+) 0.03661 MLP (T+) 0.02399
Ford Netflix
LSTM (T) 0.04024 LSTM (T) 0.04211
LSTM (T+) 0.03007 LSTM (T+) 0.03706
NORMAL NORMAL
MLP (T) 0.04886 MLP (T) 0.02954
MLP (T+) 0.02047 MLP (T+) 0.02842

References
1. Edwards, J. Global Market Cap Is Heading toward $100 Trillion and Goldman Sachs Thinks the Only Way Is down. 2017. Available
online: https://www.businessinsider.de/global-market-cap-is-about-to-hit-100-trillion-2017-12?r=UK&IR=T (accessed on 5
March 2023).
2. SIFMA. Research Quarterly: Equities. Available online: https://www.sifma.org/resources/research/research-quarterly-equities/
(accessed on 5 January 2024).
3. NYSE: New York Stock Exchange, New York Stock Exchange. NYSE Total Market Cap. 2018. Available online: https://www.
nyse.com/market-cap (accessed on 12 April 2023).
4. FXCM. New York Stock Exchange (NYSE). 2016. Available online: https://www.fxcm.com/uk/insights/new-york-stock-
exchange-nyse/ (accessed on 13 April 2023).
5. Chan, H.L.; Woo, K.Y. Studying the Dynamic Relationships between Residential Property Prices, Stock Prices, and GDP: Lessons
from Hong Kong. J. Hous. Res. 2013, 22, 75–89. [CrossRef]
6. Jones, J.U.S. Ownership down among All but Older Higher Income. 2017. Available online: https://news.gallup.com/poll/2110
52/stock-ownership-down-among-older-higher-income.aspx (accessed on 15 April 2023).
7. Lusardi, A.; Mitchell, O.S. The Economic Importance of Financial Literacy: Theory and Evidence. J. Econ. Lit. 2014, 52, 5–44.
[CrossRef]
8. Armano, G.; Marchesi, M.; Murru, A. A hybrid genetic-neural architecture for stock indexes forecasting. Inf. Sci. 2005, 170, 3–33.
[CrossRef]
9. Wang, J.-J.; Wang, J.-Z.; Zhang, Z.-G.; Guo, S.-P. Stock index forecasting based on a hybrid model. Omega 2012, 40, 758–766.
[CrossRef]
10. Kao, L.-J.; Chiu, C.-C.; Lu, C.-J.; Chang, C.-H. A hybrid approach by integrating wavelet-based feature extraction with MARS and
SVR for stock index forecasting. Decis. Support Syst. 2013, 54, 1228–1244. [CrossRef]
11. Liu, H.; Lv, X.Y. Stock Price Prediction Model Based on IWO Neural Network and its Applications. Adv. Mater. Res. 2014, 989,
1635–1640. [CrossRef]
Algorithms 2024, 17, 234 24 of 29

12. Guo, Z.; Wang, H.; Yang, J.; Miller, D.J. A Stock Market Forecasting Model Combining Two-Directional Two-Dimensional Principal
Component Analysis and Radial Basis Function Neural Network. PLoS ONE 2015, 10, e0122385. [CrossRef] [PubMed]
13. Mahmud, M.S.; Meesad, P. An innovative recurrent error-based neuro-fuzzy system with momentum for stock price prediction.
Soft Comput. 2016, 20, 4173–4191. [CrossRef]
14. Wang, S.; Wang, L.; Gao, S.; Bai, Z. Stock price prediction based on chaotic hybrid particle swarm optimisation-RBF neural
network. Int. J. Appl. Decis. Sci. 2017, 10, 89. [CrossRef]
15. Zhang, L.; Wang, F.; Xu, B.; Chi, W.; Wang, Q.; Sun, T. Prediction of stock prices based on LM-BP neural network and the
estimation of overfitting point by RDCI. Neural Comput. Appl. 2018, 30, 1425–1444. [CrossRef]
16. Bisoi, R.; Dash, P.K.; Parida, A.K. Hybrid Variational Mode Decomposition and evolutionary robust kernel extreme learning
machine for stock price and movement prediction on daily basis. Appl. Soft Comput. 2019, 74, 652–678. [CrossRef]
17. Ding, G.; Qin, L. Study on the prediction of stock price based on the associated network model of LSTM. Int. J. Mach. Learn.
Cybern. 2020, 11, 1307–1317. [CrossRef]
18. Qiu, J.; Wang, B.; Zhou, C. Forecasting stock prices with long-short term memory neural network based on attention mechanism.
PLoS ONE 2020, 15, e0227222. [CrossRef] [PubMed]
19. Jin, Z.; Yang, Y.; Liu, Y. Stock closing price prediction based on sentiment analysis and LSTM. Neural Comput. Appl. 2020, 32,
9713–9729. [CrossRef]
20. Vijh, M.; Chandola, D.; Tikkiwal, V.A.; Kumar, A. Stock Closing Price Prediction using Machine Learning Techniques. Procedia
Comput. Sci. 2020, 167, 599–606. [CrossRef]
21. Lu, W.; Li, J.; Wang, J.; Qin, L. A CNN-BiLSTM-AM method for stock price prediction. Neural Comput. Appl. 2021, 33, 4741–4753.
[CrossRef]
22. Hu, Z.; Zhao, Y.; Khushi, M. A Survey of Forex and Stock Price Prediction Using Deep Learning. Appl. Syst. Innov. 2021, 4, 9.
[CrossRef]
23. Kurani, A.; Doshi, P.; Vakharia, A.; Shah, M. A Comprehensive Comparative Study of Artificial Neural Network (ANN) and
Support Vector Machines (SVM) on Stock Forecasting. Ann. Data Sci. 2023, 10, 183–208. [CrossRef]
24. Kim, K. Electronic and Algorithmic Trading Technology: The Complete Guide; Academic Press: New York, NY, USA, 2010.
25. Ritter, G. Machine Learning for Trading. SSRN Electron. J. 2017. [CrossRef]
26. Chen, L.; Gao, Q. Application of Deep Reinforcement Learning on Automated Stock Trading. In Proceedings of the 2019 IEEE 10th
International Conference on Software Engineering and Service Science (ICSESS), Beijing, China, 18–20 October 2019. [CrossRef]
27. Zhang, Z.; Zohren, S.; Roberts, S.J. Deep Reinforcement Learning for Trading. J. Financ. Data Sci. 2020, 2, 25–40. [CrossRef]
28. Das, S.; Kadapakkam, P.-R. Machine over Mind? Stock price clustering in the era of algorithmic trading. N. Am. J. Econ. Financ.
2018, 51, 100831. [CrossRef]
29. Business Wire. Global Algorithmic Trading Market to Surpass US$ 21,685.53 Million by 2026. 2019. Available online: https:
//www.businesswire.com/news/home/20190205005634/en/Global-Algorithmic (accessed on 10 September 2020).
30. Pricope, T.V. Deep Reinforcement Learning in Quantitative Algorithmic Trading: A Review. arXiv 2024, arXiv:2106.00123v1.
31. Johnston, M. What Happened to Oil Prices in 2020. Investopedia. 2022. Available online: https://www.investopedia.com/
articles/investing/100615/will-oil-prices-go-2017.asp (accessed on 5 March 2024).
32. Cevik, E.; Altinkeski, B.K.; Cevik, E.I.; Dibooglu, S. Investor sentiments and stock markets during the COVID-19 pandemic.
Financ. Innov. 2022, 8, 69. [CrossRef] [PubMed]
33. Joseph, I.; Obini, N.; Sulaiman, A.; Loko, A. Comparative Model Profiles of COVID-19 Occurrence in Nigeria. Int. J. Math. Trends
Technol. 2020, 68, 297–310. [CrossRef]
34. Baig, A.S.; Butt, H.A.; Haroon, O.; Rizvi, S.A.R. Deaths, panic, lockdowns and US equity markets: The case of COVID-19
pandemic. Financ. Res. Lett. 2020, 38, 101701. [CrossRef]
35. Pang, X.; Zhou, Y.; Wang, P.; Lin, W.; Chang, V. An innovative neural network approach for stock market prediction. J. Supercomput.
2020, 76, 2098–2118. [CrossRef]
36. Bommareddy, S.R.; Reddy, K.S.S.; Kaushik, P.; Kumar, V.; Hulipalled, V.R. Predicting the stock price using linear regression. Int. J.
Adv. Res. Comput. Sci. 2018, 9, 81–85.
37. Hiransha, M.; Gopalakrishnan, E.A.; Menon, V.K.; Soman, K.P. NSE Stock Market Prediction Using Deep-Learning Models.
Procedia Comput. Sci. 2018, 132, 1351–1362. [CrossRef]
38. Mehtab, S.; Sen, J.; Dutta, A. Stock Price Prediction Using Machine Learning and LSTM-Based Deep Learning Models. In Machine
Learning and Metaheuristics Algorithms, and Applications; Springer: Berlin/Heidelberg, Germany, 2021. [CrossRef]
39. Nguyen, H.T.; Tran, T.B.; Bui, P.H.D. An effective way for Taiwanese stock price prediction: Boosting the performance with
machine learning techniques. Concurr. Comput. Pract. Exp. 2021, 35, e6437. [CrossRef]
40. Kumar, A.; Hooda, S.; Gill, R.; Ahlawat, D.; Srivastva, D.; Kumar, R. Stock Price Prediction Using Machine Learning. In
Proceedings of the 2023 International Conference on Computational Intelligence and Sustainable Engineering Solutions (CISES),
Greater Noida, India, 28–30 April 2023. [CrossRef]
41. Mittal, A.; Joshi, N.; Savani, V. Stock Price Prediction Using Machine Learning Algorithm with Web Interface (GUI). In Artificial
Intelligence and Communication Technologies; SCRS: New Delhi, India, 2023; pp. 361–377. [CrossRef]
42. Perdana, I.L.; Rokhim, R. Stock price index prediction using machine learning. AIP Conf. Proc. 2023, 2693, 020031. [CrossRef]
Algorithms 2024, 17, 234 25 of 29

43. Shirata, R.; Harada, T. A Proposal of a Method to Determine the Appropriate Learning Period in Stock Price Prediction Using
Machine Learning. IEEJ Trans. Electr. Electron. Eng. 2024, 19, 726–732. [CrossRef]
44. Liu, W.-J.; Ge, Y.-B.; Gu, Y.-C. News-driven stock market index prediction based on trellis network and sentiment attention
mechanism. Expert Syst. Appl. 2024, 250, 123966. [CrossRef]
45. Ammer, M.A.; Aldhyani, T.H.H. Deep Learning Algorithm to Predict Cryptocurrency Fluctuation Prices: Increasing Investment
Awareness. Electronics 2022, 11, 2349. [CrossRef]
46. Wanjawa, B.W.; Muchemi, L. ANN model to predict stock prices at stock exchange markets. arXiv 2014, arXiv:1502.06434.
47. Malkiel, B.G. A Random Walk down Wall Street: Including a Life-Cycle Guide to Personal Investing; WW Norton & Company: New
York, NY, USA, 1999.
48. Moghar, A.; Hamiche, M. Stock Market Prediction Using LSTM Recurrent Neural Network. Procedia Comput. Sci. 2020, 170,
1168–1173. [CrossRef]
49. Yu, P.; Yan, X. Stock price prediction based on deep neural networks. Neural Comput. Appl. 2020, 32, 1609–1628. [CrossRef]
50. Cheng, C.-H.; Yang, J.-H. Fuzzy time-series model based on rough set rule induction for forecasting stock price. Neurocomputing
2018, 302, 33–45. [CrossRef]
51. Nguyen, N. Hidden Markov model for stock trading. Int. J. Financ. Stud. 2018, 6, 36. [CrossRef]
52. Rundo, F.; Trenta, F.; Di Stallo, A.L.; Battiato, S. Advanced Markov-Based Machine Learning Framework for Making Adaptive
Trading System. Computation 2019, 7, 4. [CrossRef]
53. Khairi, T.W.A.; Zaki, R.M.; Mahmood, W.A. Stock Price Prediction using Technical, Fundamental and News based Approach. In
Proceedings of the 2019 2nd Scientific Conference of Computer Sciences (SCCS), Baghdad, Iraq, 27–28 March 2019. [CrossRef]
54. Ghorbani, M.; Chong, E.K.P. Stock price prediction using principal components. PLoS ONE 2020, 15, e0230124. [CrossRef]
55. Zhong, S.; Hitchcock, D. S&P 500 Stock Price Prediction Using Technical, Fundamental and Text Data. Stat. Optim. Inf. Comput.
2021, 9, 769–788. [CrossRef]
56. Sorto, M.; Aasheim, C.; Wimmer, H. Feeling the stock market: A study in the prediction of financial markets based on news
sentiment. In Proceedings of the Southern Association for Information Systems Conference, St. Simons Island, GA, USA, 25
March 2017; p. 19.
57. Wafi, A.S.; Hassan, H.; Mabrouk, A. Fundamental analysis models in financial markets—Review study. Procedia Econ. Financ.
2015, 30, 939–947. [CrossRef]
58. Gupta, R.; Chen, M. Sentiment Analysis for Stock Price Prediction. In Proceedings of the 2020 IEEE Conference on Multimedia
Information Processing and Retrieval (MIPR), Shenzhen, China, 6–8 August 2020. [CrossRef]
59. Jing, N.; Wu, Z.; Wang, H. A hybrid model integrating deep learning with investor sentiment analysis for stock price prediction.
Expert Syst. Appl. 2021, 178, 115019. [CrossRef]
60. Wu, S.; Liu, Y.; Zou, Z.; Weng, T.-H. S_I_LSTM: Stock price prediction based on multiple data sources and sentiment analysis.
Connect. Sci. 2022, 34, 44–62. [CrossRef]
61. Wang, Y.-F. Predicting stock price using fuzzy grey prediction system. Expert Syst. Appl. 2002, 22, 33–38. [CrossRef]
62. Leigh, W.; Purvis, R.; Ragusa, J.M. Forecasting the NYSE composite index with technical analysis, pattern recognizer, neural
network, and genetic algorithm: A case study in romantic decision support. Decis. Support Syst. 2002, 32, 361–377. [CrossRef]
63. Kim, K.-J. Financial time series forecasting using support vector machines. Neurocomputing 2003, 55, 307–319. [CrossRef]
64. Wang, Y.-F. Mining stock price using fuzzy rough set system. Expert Syst. Appl. 2003, 24, 13–23. [CrossRef]
65. Chen, A.-S.; Leung, M.T. Regression neural network for error correction in foreign exchange forecasting and trading. Comput.
Oper. Res. 2004, 31, 1049–1068. [CrossRef]
66. Pai, P.-F.; Lin, C.-S. A hybrid ARIMA and support vector machines model in stock price forecasting. Omega 2005, 33, 497–505.
[CrossRef]
67. Enke, D.; Thawornwong, S. The use of data mining and neural networks for forecasting stock market returns. Expert Syst. Appl.
2005, 29, 927–940. [CrossRef]
68. Kim, M.-J.; Min, S.-H.; Han, I. An evolutionary approach to the combination of multiple classifiers to predict a stock price index.
Expert Syst. Appl. 2006, 31, 241–247. [CrossRef]
69. Schumaker, R.P.; Chen, H. Textual analysis of stock market prediction using breaking financial news. ACM Trans. Inf. Syst. 2009,
27, 1–19. [CrossRef]
70. Huang, C.-L.; Tsai, C.-Y. A hybrid SOFM-SVR with a filter-based feature selection for stock market forecasting. Expert Syst. Appl.
2009, 36, 1529–1539. [CrossRef]
71. Kara, Y.; Boyacioglu, M.A.; Baykan, K. Predicting direction of stock price index movement using artificial neural networks and
support vector machines: The sample of the Istanbul Stock Exchange. Expert Syst. Appl. 2011, 38, 5311–5319. [CrossRef]
72. Groth, S.S.; Muntermann, J. An intraday market risk management approach based on textual analysis. Decis. Support Syst. 2011,
50, 680–691. [CrossRef]
73. Schumaker, R.P.; Zhang, Y.; Huang, C.-N.; Chen, H. Evaluating sentiment in financial news articles. Decis. Support Syst. 2012, 53,
458–464. [CrossRef]
74. Yolcu, U.; Egrioglu, E.; Aladag, C.H. A new linear & nonlinear artificial neural network model for time series forecasting. Decis.
Support Syst. 2013, 54, 1340–1347. [CrossRef]
Algorithms 2024, 17, 234 26 of 29

75. Hagenau, M.; Liebmann, M.; Neumann, D. Automated news reading: Stock price prediction based on financial news using
context-capturing features. Decis. Support Syst. 2013, 55, 685–697. [CrossRef]
76. Umoh, U.A.; Inyang, U.G. A FuzzFuzzy-Neural Intelligent Trading Model for Stock Price Prediction. Int. J. Comput. Sci. Issues
2014, 12, 36.
77. Nayak, S.C.; Misra, B.B.; Behera, H.S. Fluctuation prediction of stock market index by adaptive evolutionary higher order neural
networks. Int. J. Swarm Intell. 2016, 2, 229. [CrossRef]
78. Tsai, C.-F.; Quan, Z.-Y. Stock Prediction by Searching for Similarities in Candlestick Charts. ACM Trans. Manag. Inf. Syst. 2014, 5,
9.
79. Li, X.; Xie, H.; Chen, L.; Wang, J.; Deng, X. News impact on stock price return via sentiment analysis. Knowl.-Based Syst. 2014, 69,
14–23. [CrossRef]
80. Sun, X.-Q.; Shen, H.-W.; Cheng, X.-Q. Trading Network Predicts Stock Price. Sci. Rep. 2014, 4, 3711. [CrossRef]
81. Dash, R.; Dash, P.; Bisoi, R. A self adaptive differential harmony search based optimized extreme learning machine for financial
time series prediction. Swarm Evol. Comput. 2014, 19, 25–42. [CrossRef]
82. Adebiyi, A.A.; Adewumi, A.O.; Ayo, C.K. Comparison of ARIMA and Artificial Neural Networks Models for Stock Price
Prediction. J. Appl. Math. 2014, 2014, 614342. [CrossRef]
83. Lee, H.; Surdeanu, M.; MacCartney, B.; Jurafsky, D. On the Importance of Text Analysis for Stock Price Prediction. In Proceedings
of the LREC 2014, Ninth International Conference on Language Resources and Evaluation, Reykjavik, Iceland, 26–31 May 2014;
pp. 1170–1175.
84. de Fortuny, E.J.; De Smedt, T.; Martens, D.; Daelemans, W. Evaluating and understanding text-based stock price prediction
models. Inf. Process. Manag. 2014, 50, 426–441. [CrossRef]
85. Bisoi, R.; Dash, P. A hybrid evolutionary dynamic neural network for stock market trend analysis and prediction using unscented
Kalman filter. Appl. Soft Comput. 2014, 19, 41–56. [CrossRef]
86. Mondal, P.; Shit, L.; Goswami, S. Study of Effectiveness of Time Series Modeling (Arima) in Forecasting Stock Prices. Int. J.
Comput. Sci. Eng. Appl. 2014, 4, 13–29. [CrossRef]
87. Geva, T.; Zahavi, J. Empirical evaluation of an automated intraday stock recommendation system incorporating both market data
and textual news. Decis. Support Syst. 2014, 57, 212–223. [CrossRef]
88. Jiang, S.; Chen, H.; Nunamaker, J.F.; Zimbra, D. Analyzing firm-specific social media and market: A stakeholder-based event
analysis framework. Decis. Support Syst. 2014, 67, 30–39. [CrossRef]
89. Hafezi, R.; Shahrabi, J.; Hadavandi, E. A bat-neural network multi-agent system (BNNMAS) for stock price prediction: Case
study of DAX stock price. Appl. Soft Comput. 2015, 29, 196–210. [CrossRef]
90. Ballings, M.; Van den Poel, D.; Hespeels, N.; Gryp, R. Evaluating multiple classifiers for stock price direction prediction. Expert
Syst. Appl. 2015, 42, 7046–7056. [CrossRef]
91. Nguyen, T.H.; Shirai, K.; Velcin, J. Sentiment analysis on social media for stock movement prediction. Expert Syst. Appl. 2015, 42,
9603–9611. [CrossRef]
92. Wang, L.; Wang, Z.; Zhao, S.; Tan, S. Stock market trend prediction using dynamical Bayesian factor graph. Expert Syst. Appl.
2015, 42, 6267–6275. [CrossRef]
93. Sun, B.; Guo, H.; Karimi, H.R.; Ge, Y.; Xiong, S. Prediction of stock index futures prices based on fuzzy sets and multivariate
fuzzy time series. Neurocomputing 2015, 151, 1528–1536. [CrossRef]
94. Göçken, M.; Özçalıcı, M.; Boru, A.; Dosdoğru, A.T. Integrating metaheuristics and Artificial Neural Networks for improved stock
price prediction. Expert Syst. Appl. 2016, 44, 320–331. [CrossRef]
95. Dash, R.; Dash, P. Efficient stock price prediction using a self evolving recurrent neuro-fuzzy inference system optimized through
a modified differential harmony search technique. Expert Syst. Appl. 2016, 52, 75–90. [CrossRef]
96. Zhou, T.; Gao, S.; Wang, J.; Chu, C.; Todo, Y.; Tang, Z. Financial time series prediction using a dendritic neuron model. Knowl.
Based Syst. 2016, 105, 214–224. [CrossRef]
97. Shynkevich, Y.; McGinnity, T.; Coleman, S.A.; Belatreche, A. Forecasting movements of health-care stock prices based on different
categories of news articles using multiple kernel learning. Decis. Support Syst. 2016, 85, 74–83. [CrossRef]
98. Nie, C.-X.; Jin, X.-B. The Interval Slope Method for Long-Term Forecasting of Stock Price Trends. Adv. Math. Phys. 2016, 2016,
8045656. [CrossRef]
99. Chen, R.; Pan, B. Chinese Stock Index Futures Price Fluctuation Analysis and Prediction Based on Complementary Ensemble
Empirical Mode Decomposition. Math. Probl. Eng. 2016, 2016, 3791504. [CrossRef]
100. Lahmiri, S. Intraday stock price forecasting based on variational mode decomposition. J. Comput. Sci. 2016, 12, 23–27. [CrossRef]
101. Qiu, M.; Song, Y.; Akagi, F. Application of artificial neural network for the prediction of stock market returns: The case of the
Japanese stock market. Chaos Solitons Fractals 2016, 85, 1–7. [CrossRef]
102. Qiu, M.; Song, Y. Predicting the Direction of Stock Market Index Movement Using an Optimized Artificial Neural Network
Model. PLoS ONE 2016, 11, e0155133. [CrossRef]
103. An, Y.; Chan, N.H. Short-Term Stock Price Prediction Based on Limit Order Book Dynamics. J. Forecast. 2017, 36, 541–556.
[CrossRef]
104. Shynkevich, Y.; McGinnity, T.; Coleman, S.A.; Belatreche, A.; Li, Y. Forecasting price movements using technical indicators:
Investigating the impact of varying input window length. Neurocomputing 2017, 264, 71–88. [CrossRef]
Algorithms 2024, 17, 234 27 of 29

105. Ouahilal, M.; El Mohajir, M.; Chahhou, M.; El Mohajir, B.E. A novel hybrid model based on Hodrick–Prescott filter and support
vector regression algorithm for optimizing stock market price prediction. J. Big Data 2017, 4, 31. [CrossRef]
106. Rout, A.K.; Dash, P.; Dash, R.; Bisoi, R. Forecasting financial time series using a low complexity recurrent neural network and
evolutionary learning approach. J. King Saud Univ. Comput. Inf. Sci. 2017, 29, 536–552. [CrossRef]
107. Castelli, M.; Vanneschi, L.; Trujillo, L.; Popovič, A. Stock index return forecasting: Semantics-based genetic programming with
local search optimiser. Int. J. Bio-Inspired Comput. 2017, 10, 159–171. [CrossRef]
108. Tao, L.; Hao, Y.; Yijie, H.; Chunfeng, S. K-Line Patterns’ Predictive Power Analysis Using the Methods of Similarity Match and
Clustering. Math. Probl. Eng. 2017, 2017, 3096917. [CrossRef]
109. Chong, E.; Han, C.; Park, F.C. Deep learning networks for stock market analysis and prediction: Methodology, data representations,
and case studies. Expert Syst. Appl. 2017, 83, 187–205. [CrossRef]
110. Weng, B.; Ahmed, M.A.; Megahed, F.M. Stock market one-day ahead movement prediction using disparate data sources. Expert
Syst. Appl. 2017, 79, 153–163. [CrossRef]
111. Zhuge, Q.; Xu, L.; Zhang, G. LSTM Neural Network with Emotional Analysis for Prediction of Stock Price. Eng. Lett. 2017, 25,
1–9.
112. Kraus, M.; Feuerriegel, S. Decision support from financial disclosures with deep neural networks and transfer learning. Decis.
Support Syst. 2017, 104, 38–48. [CrossRef]
113. Jeon, S.; Hong, B.; Chang, V. Pattern graph tracking-based stock price prediction using big data. Futur. Gener. Comput. Syst. 2018,
80, 171–187. [CrossRef]
114. Agustini, W.F.; Affianti, I.R.; Putri, E.R. Stock price prediction using geometric Brownian motion. J. Phys. Conf. Ser. 2018, 974,
012047. [CrossRef]
115. Matsubara, T.; Akita, R.; Uehara, K. Stock Price Prediction by Deep Neural Generative Model of News Articles. IEICE Trans. Inf.
Syst. 2018, 101, 901–908. [CrossRef]
116. Ebadati, E.O.M.; Mortazavi, T.M. An efficient hybrid machine learning method for time series stock market forecasting. Neural
Netw. World 2018, 28, 41–55. [CrossRef]
117. Kooli, C.; Trabelsi, R.; Tlili, F. The impact of accounting disclosure on emerging stock market prediction in an unstable socio-
political context. J. Account. Manag. Inf. Syst. 2018, 17, 313–329. [CrossRef]
118. Lahmiri, S. Minute-ahead stock price forecasting based on singular spectrum analysis and support vector regression. Appl. Math.
Comput. 2018, 320, 444–451. [CrossRef]
119. Zhou, X.; Pan, Z.; Hu, G.; Tang, S.; Zhao, C. Stock Market Prediction on High-Frequency Data Using Generative Adversarial Nets.
Math. Probl. Eng. 2018, 2018, 1–11. [CrossRef]
120. Shah, H.; Tairan, N.; Garg, H.; Ghazali, R. A Quick Gbest Guided Artificial Bee Colony Algorithm for Stock Market Prices
Prediction. Symmetry 2018, 10, 292. [CrossRef]
121. Göçken, M.; Özçalıcı, M.; Boru, A.; Dosdoğru, A.T. Stock price prediction using hybrid soft computing models incorporating
parameter tuning and input variable selection. Neural Comput. Appl. 2019, 31, 577–592. [CrossRef]
122. Vanstone, B.J.; Gepp, A.; Harris, G. The effect of sentiment on stock price prediction. In International Conference on Industrial,
Engineering and Other Applications of Applied Intelligent Systems; Springer: Cham, Switzerland, 2019; pp. 551–559.
123. Zheng, J.; Wang, Y.; Li, S.; Chen, H. The Stock Index Prediction Based on SVR Model with Bat Optimization Algorithm. Algorithms
2021, 14, 299. [CrossRef]
124. Chen, J.; Wen, Y.; Nanehkaran, Y.; Suzauddola; Chen, W.; Zhang, D. Machine learning techniques for stock price prediction and
graphic signal recognition. Eng. Appl. Artif. Intell. 2023, 121, 106038. [CrossRef]
125. Jiang, Z.; Liu, J.; Yang, L. Comparison Analysis of Stock Price Prediction Based on Different Machine Learning Methods. In
Proceedings of the 2nd International Academic Conference on Blockchain, Information Technology and Smart Finance (ICBIS
2023), Hangzhou, China, 17–19 February 2023; pp. 59–67. [CrossRef]
126. Antad, S.; Khandelwal, S.; Khandelwal, A.; Khandare, R.; Khandave, P.; Khangar, D.; Khanke, R. Stock Price Prediction Website
Using Linear Regression—A Machine Learning Algorithm. ITM Web Conf. 2023, 56, 05016. [CrossRef]
127. Khan, W.; Ghazanfar, M.A.; Azam, M.A.; Karami, A.; Alyoubi, K.H.; Alfakeeh, A.S. Stock market prediction using machine
learning classifiers and social media, news. J. Ambient. Intell. Humaniz. Comput. 2022, 13, 3433–3456. [CrossRef]
128. Shaban, W.M.; Ashraf, E.; Slama, A.E. SMP-DL: A novel stock market prediction approach based on deep learning for effective
trend forecasting. Neural Comput. Appl. 2023, 36, 1849–1873. [CrossRef]
129. Belcastro, L.; Carbone, D.; Cosentino, C.; Marozzo, F.; Trunfio, P. Enhancing Cryptocurrency Price Forecasting by Integrating
Machine Learning with Social Media and Market Data. Algorithms 2023, 16, 542. [CrossRef]
130. Al-Nefaie, A.H.; Aldhyani, T.H.H. Bitcoin Price Forecasting and Trading: Data Analytics Approaches. Electronics 2022, 11, 4088.
[CrossRef]
131. Chen, Y.; Hao, Y. A feature weighted support vector machine and K-nearest neighbor algorithm for stock market indices
prediction. Expert Syst. Appl. 2017, 80, 340–355. [CrossRef]
132. Patel, J.; Shah, S.; Thakkar, P.; Kotecha, K. Predicting stock and stock price index movement using Trend Deterministic Data
Preparation and machine learning techniques. Expert Syst. Appl. 2015, 42, 259–268. [CrossRef]
133. Athey, S.; Tibshirani, J.; Wager, S. Generalized random forests. Ann. Stat. 2019, 47, 1148–1178. [CrossRef]
134. Scornet, E.; Biau, G.; Vert, J.-P. Consistency of random forests. Ann. Stat. 2015, 43, 1716–1741. [CrossRef]
Algorithms 2024, 17, 234 28 of 29

135. Fenghua, W.; Jihong, X.; Zhifang, H.; Xu, G. Stock Price Prediction based on SSA and SVM. Procedia Comput. Sci. 2014, 31, 625–631.
[CrossRef]
136. Babu, C.N.; Reddy, B.E. A moving-average filter based hybrid ARIMA–ANN model for forecasting time series data. Appl. Soft
Comput. 2014, 23, 27–38. [CrossRef]
137. Kumar, M.; Anand, M. An application of time series ARIMA forecasting model for predicting sugarcane production in India.
Stud. Bus. Econ. 2014, 9, 81–94.
138. Moghaddam, A.H.; Moghaddam, M.H.; Esfandyari, M. Stock market index prediction using artificial neural network. J. Econ.
Financ. Adm. Sci. 2016, 21, 89–93. [CrossRef]
139. Olah, C. Understanding lstm networks–colah’s blog. Colah. github. io. 2015. Available online: https://colah.github.io/posts/20
15-08-Understanding-LSTMs/ (accessed on 3 March 2023).
140. Bloomberg. Every Time Trump Tweets about the Stock Market. 2019. Available online: https://www.bloomberg.com/features/
trump-tweets-market (accessed on 30 September 2023).
141. Investopedia. Can Tweets and Facebook Posts Predict Stock Behavior? 2019. Available online: https://www.investopedia.com/
articles/markets/031814/can-tweets-and-facebook-posts-predict-stock-behavior-and-rt-if-you-think-so.asp (accessed on 15
September 2023).
142. Asteriou, D.; Pilbeam, K.; Sarantidis, A. The Behaviour of Banking Stocks During the Financial Crisis and Recessions. Evidence
from Changes-in-Changes Panel Data Estimations. Scott. J. Political Econ. 2019, 66, 154–179. [CrossRef]
143. Erdogan, O.; Bennett, P.; Ozyildirim, C. Recession Prediction Using Yield Curve and Stock Market Liquidity Deviation Measures.
Rev. Financ. 2014, 19, 407–422. [CrossRef]
144. Kleinnijenhuis, J.; Schultz, F.; Oegema, D.; van Atteveldt, W. Financial news and market panics in the age of high-frequency
sentiment trading algorithms. J. Theory Pract. Crit. 2013, 14, 271–291. [CrossRef]
145. Angelovska, J. Investors’ behaviour in regard to company earnings announcements during the recession period: Evidence from
the Macedonian stock exchange. Econ. Res. Istraz. 2017, 30, 647–660. [CrossRef]
146. Rath, S.; Das, N.R.; Pattanayak, B.K. An Analytic Review on Stock Market Price Prediction using Machine Learning and Deep
Learning Techniques. Recent Patents Eng. 2024, 18, 88–104. [CrossRef]
147. Nelson DM, Q.; Pereira AC, M.; De Oliveira, R.A. Stock market’s price movement prediction with LSTM neural networks. In
Proceedings of the 2017 International Joint Conference on Neural Networks (IJCNN), Anchorage, AK, USA, 14–19 May 2017.
[CrossRef]
148. Yang, L.; Shami, A. On hyperparameter optimization of machine learning algorithms: Theory and practice. arXiv 2020,
arXiv:2007.15745. [CrossRef]
149. Guliyev, N.J.; Ismailov, V.E. A Single Hidden Layer Feedforward Network with Only One Neuron in the Hidden Layer Can
Approximate Any Univariate Function. Neural Comput. 2016, 28, 1289–1304. [CrossRef]
150. Stathakis, D. How many hidden layers and nodes? Int. J. Remote Sens. 2009, 30, 2133–2147. [CrossRef]
151. Thomas, A.J.; Petridis, M.; Walters, S.D.; Gheytassi, S.M.; Morgan, R.E. Two Hidden Layers are Usually Better than One. In
Engineering Applications of Neural Networks Communications in Computer and Information Science; Springer: Berlin/Heidelberg,
Germany, 2017; pp. 279–290. [CrossRef]
152. Saleem, N.; Khattak, M.I. Deep Neural Networks for Speech Enhancement in Complex-Noisy Environments. Int. J. Interact.
Multimed. Artif. Intell. 2020, 6, 84. [CrossRef]
153. Zhang, N.; Shen, S.-L.; Zhou, A.-N.; Xu, Y.-S. Investigation on Performance of Neural Networks Using Quadratic Relative Error
Cost Function. IEEE Access 2019, 7, 106642–106652. [CrossRef]
154. Hasan, M.M.; Rahman, M.S.; Bell, A. Deep Reinforcement Learning for Optimization. In Handbook of Research on Deep Learning
Innovations and Trends; Research Anthology on Artificial Intelligence Applications in Security; IGI Global: Hershey, PA, USA, 2021.
155. Fangasadha, E.F.; Soeroredjo, S.; Gunawan, A.A.S. Literature Review of OpenAI Five’s Mechanisms in Dota 2’s Bot Player. In
Proceedings of the 2022 International Seminar on Application for Technology of Information and Communication (iSemantic),
Semarang, Indonesia, 17–18 September 2022. [CrossRef]
156. Jiang, S.; Chen, Y. Hand Gesture Recognition by Using 3DCNN and LSTM with Adam Optimizer. In Advances in Multimedia
Information Processing—PCM 2017 Lecture Notes in Computer Science; Springer: Berlin/Heidelberg, Germany, 2018; pp. 743–753.
[CrossRef]
157. Chang, Z.; Zhang, Y.; Chen, W. Electricity price prediction based on hybrid model of adam optimized LSTM neural network and
wavelet transform. Energy 2019, 187, 115804. [CrossRef]
158. Skehin, T.; Crane, M.; Bezbradica, M. Day ahead forecasting of FAANG stocks using ARIMA, LSTM networks and wavelets. In
Proceedings of the 26th AIAI Irish Conference on Artificial Intelligence and Cognitive Science, Dublin, Ireland, 6–7 December
2018.
159. Johnson, A.; Reed, A. Tesla in Texas: A Showdown Over Showrooms. SAM Adv. Manag. J. 2019, 84, 47–56.
160. Cuofano, G. Tesla Distribution Strategy—FourWeekMBA. Tesla-Distribution-Strategy. 2024. Available online: https:
//fourweekmba.com/tesla-distribution-strategy/ (accessed on 5 March 2024).
161. Dey, P.; Hossain, E.; Hossain, I.; Chowdhury, M.A.; Alam, S.; Hossain, M.S.; Andersson, K. Comparative Analysis of Recurrent
Neural Networks in Stock Price Prediction for Different Frequency Domains. Algorithms 2021, 14, 251. [CrossRef]
Algorithms 2024, 17, 234 29 of 29

162. Kumar, R.; Kumar, P.; Kumar, Y. Time Series Data Prediction using IoT and Machine Learning Technique. Procedia Comput. Sci.
2020, 167, 373–381. [CrossRef]
163. Thakkar, A.; Chaudhari, K. CREST: Cross-Reference to Exchange-based Stock Trend Prediction using Long Short-Term Memory.
Procedia Comput. Sci. 2020, 167, 616–625. [CrossRef]
164. Pallavi, D.; Mourani, S.; Prosenjit, P.; Sufia, Z.; Abhijit, M. Use of Non-Linear Autoregressive Model to Forecast the Future Health
of Shrimp Farm. J. Mech. Contin. Math. Sci. 2021, 16, 59–64.
165. Oyekale, J.; Oreko, B. Machine learning for design and optimization of organic Rankine cycle plants: A review of current status
and future perspectives. WIREs Energy Environ. 2023, 12, e474. [CrossRef]
166. Greff, K.; Srivastava, R.K.; Koutník, J.; Steunebrink, B.R.; Schmidhuber, J. LSTM: A Search Space Odyssey. IEEE Trans. Neural
Netw. Learn. Syst. 2017, 28, 2222–2232. [CrossRef]
167. Dong, J.; Chen, Y.; Guan, G. Cost Index Predictions for Construction Engineering Based on LSTM Neural Networks. Adv. Civ.
Eng. 2020, 2020, 6518147. [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.