![llustrates the structure of the proposed semantic analysis model. The model accepts a sentence í µí°· = [í µí±¤ 1 , í µí±¤ 2 , . . . , í µí±¤ n ] as its input, with í µí±¤ i denoting the words and í µí± representing the maximum count of words a sentence can have. This input is subsequently fed into the Bidirectional Encoder Representations from Transformers (BERT) model. The result from the BERT model is an embedding matrix í µí± = [í µí± 1 ; í µí± 2 ; … ; í µí± í µí± ], where each í µí± i is the embedded representation of the corresponding word í µí±¤ i . The matrix M is subjected to three parallel dilated convolution layers to obtain features from the sentence. Each of these branches independently derives a feature vector from the sentence. Post convolution, max pooling is employed to isolate the most significant features and reduce the computational load of the network. The output derived from the max pooling layers is subsequently channeled into the Multilayer Perceptron (MLP) network for classification. The output generated by the MLP is a vector of length three since every sentence s must be categorized into one of three classes: positive, negative, or neutral. However, since most sentences are often classified as positive, the classifier encounters imbalanced classification, decreasing system performance. The OffPolicy PPO algorithm is employed to address this issue, creating a sequential decision-maker to overcome the imbalanced classification problem.](https://www.researchgate.net/publication/380507989/figure/fig1/AS:11431281242336856@1715435879768/llustrates-the-structure-of-the-proposed-semantic-analysis-model-The-model-accepts-a_Q320.jpg)
Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
VOLUME XX, 2023 1
Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2022.Doi Number
Stock market prediction with transductive long
short-term memory and social media sentiment
analysis
Ali Peivandizadeh1, Sima Hatami2, Amirhossein Nakhjavani3, Lida Khoshsima4, Mohammad Reza Chalak
Qazani5, Muhammad Haleem6, *, Roohallah Alizadehsani7
1Graduated Student, University of Houston, Houston, TX, USA
2Business management Department, Islamic Azad University of Qazvin, Qazvin, Iran
3Medical school University: Mashhad University of Medical Science
4Department of Islamic Economics, Faculty of Social Sciences, raja University, Qazvin, Iran
5Assistant Professor, Faculty of Computing and Information Technology, Sohar University, Sohar, 311, Sohar, Oman
6Department of Computer Science, Kardan University, Kabul, Afghanistan.
7Institute for Intelligent Systems Research and Innovation (IISRI) Deakin University, Waurn Ponds
Corresponding author: Muhammad Haleem (m.haleem@kardan.edu.af)
ABSTRACT In an era dominated by digital communication, the vast amounts of data generated from
social media and financial markets present unique opportunities and challenges for forecasting stock
market prices. This paper proposes an innovative approach that harnesses the power of social media
sentiment analysis combined with stock market data to predict stock prices, directly addressing the critical
challenges in this domain. A major challenge in sentiment analysis is the uneven distribution of data
across different sentiment categories. Traditional models struggle to accurately identify fewer common
sentiments (minority class) due to the overwhelming presence of more common sentiments (majority
class). To tackle this, we introduce the Off-policy Proximal Policy Optimization (PPO) algorithm,
specifically designed to handle class imbalance by adjusting the reward mechanism in the training phase,
thus favoring the correct classification of minority class instances. Another challenge is effectively
integrating the temporal dynamics of stock prices with sentiment analysis results. Our solution is
implementing a Transductive Long Short-Term Memory (TLSTM) model that incorporates sentiment
analysis findings with historical stock data. This model excels at recognizing temporal patterns and gives
precedence to data points that are temporally closer to the prediction point, enhancing the prediction
accuracy. Ablation studies confirm the effectiveness of the Off-policy PPO and TLSTM components on
the overall model performance. The proposed approach advances the field of financial analytics by
providing a more nuanced understanding of market dynamics but also offers actionable insights for
investors and policymakers seeking to navigate the complexities of the stock market with greater precision
and confidence.
INDEX TERMS Stock market, sentiment analysis, Unbalanced classification, Proximal policy
optimization, Transductive long short-term memory
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
I. INTRODUCTION
The stock market is a crucial component of a nation's
economy, influencing its growth or decline through its
performance over time [1]. Given the inherent
unpredictability of financial markets, it is still being
determined whether investments will yield profits or result
in significant losses for investors. As a critical aspect of
economic liberalization, stock markets play a vital role in
the financial strategies of the global corporate world.
Investors face the crucial decision of buying, selling, or
holding a stock. Making the right investment choices can
lead to substantial profits, but wrong decisions can lead to
losses, affecting both the individual investor and the
country's economy. Therefore, there is a pressing need for
predictive models that can forecast stock prices with greater
accuracy and efficiency.
A wealth of research has shown that including data from
fundamental analyses, such as financial news on websites
or social media postings, can improve the accuracy of
models predicting stock prices [2, 3]. Moreover, between
one-third and two-thirds of investors turn to social media
platforms to collect information and gain insights into
companies they are interested in. This reliance on social
media not only shapes their investment strategies but also
underscores the influence that online commentary can have
on the fluctuations of stock prices. Integrating digital
discourse into investment analysis represents a paradigm
shift in gauging and understanding market sentiments. With
their real-time updates and broad user engagement, social
media platforms offer a dynamic and rich source of investor
sentiment and market perception. This digital pulse can
provide early signals of shifts in stock market trends,
offering investors a competitive edge. As such, the ability
to analyze and interpret social media sentiment is becoming
an increasingly valuable tool in the financial analyst's
toolkit, enabling more nuanced and informed decision-
making in the fast-paced world of stock trading [4].
In stock market prediction, various time-series
methodologies, including Long Short-Term Memory
(LSTM)-based models, have been employed to develop
predictive frameworks [5]. Although LSTM models have
proven their worth across numerous sequence learning
tasks, their global modeling approach, which relies on the
entirety of the training data, may sometimes overlook
subtle nuances within certain feature space areas. To
address this limitation, the TLSTM model incorporates a
transductive learning process, enhancing model sensitivity
to minor variations in data [6]. By dynamically adjusting
weights based on the proximity of data points to unknown
values, TLSTM [7] offers a more nuanced and effective
method for time-series prediction. This model melds the
adaptive nature of transductive learning with LSTM's
robustness, enabling it to navigate both local idiosyncrasies
and overarching temporal patterns adeptly. Such an
amalgamation of global and localized insights significantly
bolsters the model's capacity to decipher and anticipate the
intricacies of complicated time series, which require a
delicate balance between recognizing immediate, specific
patterns and understanding broader, temporal dynamics [8].
Few studies have ventured into using sentiment analysis
for predicting stock market trends [9-11], but they often run
into the problem of imbalanced data, which can
compromise their effectiveness. To address this, strategies
are put in place at both the data handling and algorithmic
stages [12]. Data-level tactics involve adjusting the dataset
to balance it out, such as by decreasing the size of
overrepresented groups, increasing the presence of
underrepresented ones, or creating new data points to even
the scales [13, 14]. Algorithmic-level strategies focus on
fine-tuning the learning algorithms to better recognize and
value the input from less represented groups, which might
be infrequent but are often crucial for accurate predictions
[15]. The main issue at the data level is the uneven
distribution of data categories, leading to a bias towards
more common outcomes and neglecting rarer, yet
significant, market signals. At the algorithmic stage, the
challenge lies in modifying the learning process to ensure
that these vital but less common signals are not missed,
enhancing the model's overall predictive power and
reliability [16].
The rise of Deep Reinforcement Learning (DRL) has
garnered considerable interest for its capability to handle
complex tasks, particularly in scenarios where classes are
imbalanced [17, 18]. DRL boosts the effectiveness of
classification systems by reducing noise and enhancing
relevant features, showing success in various fields. One of
the critical strengths of DRL is its ability to adapt its
learning approach based on rewards, which is especially
useful for addressing data imbalances. By tailoring the
reward system to prioritize accurately identifying less
represented classes, DRL models can focus more on these
crucial yet often overlooked samples. This strategy helps
maintain the model's impartiality towards dominant
classes, enhancing detection accuracy. However, DRL
models can encounter challenges related to the bias-
variance dilemma and their sensitivity to hyperparameters
[19]. PPO [20], an advanced on-policy RL algorithm,
overcomes some challenges. It uses a clipping mechanism
to ensure stable training by moderating policy updates and
preventing significant deviations from the current policy.
PPO is known for its computational efficiency and is well-
adapted for large-scale or complex scenarios involving
continuous variables. Various modifications like Trust
Region-Guided PPO and Truly PPO [21] have been
developed to boost PPO's efficiency. However, they often
overlook the potential benefits of using off-policy data to
enhance sample efficiency. Off-policy PPO has
demonstrated remarkable achievements in gaming [22],
robotics [23], and continuous control tasks [24] by
optimizing policies with data gathered from agent
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
interactions, offering greater sample efficiency than on-
policy techniques. Utilizing off-policy data allows these
methods to minimize the costs associated with extensive
direct interactions, making them ideal for tackling
complicated sequential decision-making problems in real-
world settings. Off-policy approaches, such as Q-learning,
store past experiences in a replay buffer, enabling the
algorithm to learn from a broad range of past interactions.
This increases the system's resilience and flexibility in
changing environments and supports exploring various
strategies, proving beneficial for applications that require
long-term planning like autonomous navigation, financial
decision-making, and complex industrial operations [25].
This study addresses two primary challenges in social
media analysis for stock price prediction: the issue of
unbalanced classification in sentiment analysis and the
complexity of capturing temporal subtleties in financial
time-series data. To tackle these challenges, we propose a
novel two-stage methodology. In the first stage, we develop
a sentiment analysis model that employs three dilated
convolution layers to extract feature vectors for
classification. To overcome the problem of unbalanced
classification, where particular sentiments or outcomes are
disproportionately represented, we implement an
innovative approach using Off-Policy PPO. This approach
conceptualizes sentiment classification as a series of
decision-making processes where an agent is rewarded for
accurate classifications. To address the imbalance, we
strategically assign lower rewards to the majority class than
the minority class, encouraging the model to pay more
attention to underrepresented categories. In the second
stage, we utilize a TLSTM technique that combines
sentiment analysis results with historical stock data for
prediction. The TLSTM method excels at recognizing the
intricate temporal patterns inherent in stock market data.
By adopting transductive learning principles, the model
allows samples closer to the test data to have a more
pronounced impact on the prediction process, enhancing
the accuracy of forecasting future stock prices.
The pivotal contributions of the proposed model are
encapsulated in several key areas:
• Our novel model for forecasting stock market trends
employs the power of TLSTM. This approach is
distinct in its ability to gauge the influence of training
examples on the cost function, a measurement made
according to their similarity to the test point. The result
is a considerable enhancement in the precision and
efficacy of stock market predictions.
• We implement an original Off-Policy PPO-based
technique to counterbalance the skewed classification
challenges observed in sentiment analysis. This
involves suggesting an adapted surrogate goal that uses
off-policy data to prevent considerable updates to the
policy.
• Our model integrates a unique, rewarding system
where correct decisions are positively reinforced, and
incorrect ones are penalized. We tackle the issue of
data skewness by allotting increased incentives to the
less represented category, thus motivating the model to
allocate greater focus to frequently overlooked entries.
This deliberate tactic assists in accomplishing a more
equitable and unbiased categorization.
The rest of the article unfolds as follows: Section 2
presents a comprehensive review of the existing literature
on stock market prediction. Section 3 delves deeper into the
proposed methodology. Experimental results and their
corresponding analyses are showcased in Section 4. Lastly,
Section 5 wraps up the article with the conclusion
II. Related work
The field of stock market prediction has been rigorously
explored, with numerous surveys providing comprehensive
reviews [26-28]. These surveys offer diverse perspectives,
ranging from the efficacy of classical forecasting
techniques to applying advanced machine learning models
like LSTM networks and integrating sentiment analysis for
enhanced accuracy. In this section, we delve into the
existing literature on stock market forecasting, categorizing
it into three main areas: traditional forecasting methods,
LSTM-based models, and the role of sentiment analysis in
predicting market trends.
A. Conventional time-series analysis
Some efforts have been made to forecast stock market
trends using classic time-series analysis. Sangeetha et al.
[29] suggested applying Machine Learning with Evaluated
Linear Regression (ELR-ML) to predict the S&P 500
index, utilizing factors like opening, closing, low, high, and
volume. This method takes advantage of machine learning's
ability to navigate the complex and unpredictable nature of
the stock market, marking a significant step forward in
making stock market predictions more automated and data-
driven. Zhang et al. [30] introduced an innovative
technique to boost the accuracy of stock market forecasts
by combining a wavelet soft-threshold de-noising model
with Support Vector Machine (SVM) classification. This
strategy improved the clarity of the training data by
removing stochastic trend noise from the Shanghai Stock
Exchange (SSE) Composite Index, achieving a 60.12%
success rate in predictions, a notable improvement over the
54.25% accuracy obtained with noisy data. Mahmoodi et
al. [31] developed a strategy to improve the prediction of
stock market trading signals, considering the market's
inherent volatility and unpredictability. They employed an
SVM enhanced with Particle Swarm Optimization (PSO)
for more effective and precise categorization. Kurani et al.
[32] examined how financial organizations depend on
computational technology for various activities, from
managing budgets to forecasting stock trends. They
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
explored the roles of Artificial Neural Networks (ANNs)
and SVM in making predictions, emphasizing the
robustness of ANN to incomplete data and the efficiency of
SVM in avoiding overfitting by using straightforward
decision boundaries.
Traditional time-series methods assume a linear connection
between stock prices, working best with stable and
predictable trends. However, these methods need to be
revised regarding the stock market's more complex,
nonlinear relationships. Furthermore, the stock market is
influenced by numerous complex factors that simple time-
series analyses need to pay more attention to, leading to less
effective predictions.
B. LSTM model
Numerous studies have shown that LSTM networks can
improve time-series forecasting [33-35]. Behura et al. [36]
introduced a cutting-edge method to predict stock prices
using an advanced Multi-Layer Sequential LSTM model
enhanced by the Adam optimizer. This technique uses
normalized data broken down into time intervals to link
past and future values while overcoming common issues
like the vanishing gradient problem in essential neural
networks. Koo et al. [37] developed the Centralized
Clusters Distribution (CCD) as a new technique to filter
input data, significantly improving Bitcoin price
predictions by addressing the price's extreme variability.
Combined with the Weighted Empirical Stretching (WES)
loss function, which adjusts penalties based on data
distribution, this method significantly boosts prediction
accuracy. Bilhah et al. [38] conducted a thorough analysis
of LSTM networks for forecasting in the unpredictable
stock market, utilizing a broad array of real data to
showcase LSTM's superiority over traditional methods.
They aimed to provide a basis for comparing asset values
by analyzing historical market data. Chen et al. [39]
focused on data preprocessing for stock price predictions,
introducing 57 technical indicators to capture economic
signals better and employing the Least Absolute Shrinkage
and Selection Operator (LASSO) algorithm to select the
most relevant features. Their use of Ca-LSTM (cascade
LSTM) aimed to extract more nuanced information from
the data to improve prediction reliability. Sivadasan et al.
[40] applied recurrent neural networks, including Gated
Recurrent Unit (GRU) and LSTM, to discern patterns in
stock market data, using a sliding window method to
examine daily stock metrics and integrate various technical
indicators to refine their models further. Chadidjah et al.
[41] reviewed the efficacy of LSTM networks in
forecasting Apple Inc.'s stock prices by comparing different
LSTM configurations. This study assessed how well these
models could detect stock data trends, taking into account
their computational efficiency and predictive accuracy.
Traditional LSTM methods have made significant strides
in stock market forecasting, but frequently must focus on
subtle variances within specific areas of the feature space,
stemming from developing a comprehensive model that
encompasses the entire training dataset, which could
restrict their usefulness in real-world stock market
scenarios. Our research introduces a new strategy that
leverages the advanced features of TLSTM to mitigate
these drawbacks. This combined approach improves upon
the standard LSTM models, providing a more powerful and
flexible tool for stock market analysis.
C. Sentiment analysis
Sentiment analysis fundamentally explores and quantifies
the emotional tone behind a body of text. This approach has
been widely applied in various fields to understand better
the attitudes, opinions, and emotions expressed in written
language [42]. In the context of stock markets, sentiment
analysis is increasingly recognized for its capacity to unveil
market participants' underlying moods and sentiments,
potentially influencing stock prices and market trends [43].
Drawing from the intersection of behavioral finance and
data science, sentiment analysis examines how collective
emotions and investor sentiments, often disseminated
through social media and financial news, can predict
market movements [44, 45].
While some studies have established a link between the
sentiment of online commentary and stock trends, few have
ventured into predicting specific stock prices through
sentiment analysis [46, 47]. Bassant et al. [48] proposed a
unique approach to improve stock market movement
predictions by refining sentiment analysis with
neutrosophic logic (NL), which adeptly manages uncertain
and indeterminate data. This method involves classifying
social media sentiments more accurately and using these
insights and historical stock data as inputs for a deep-
learning LSTM model to forecast stock movements over a
set period. BL and BR [49] introduced a pioneering
framework for predicting stock prices, combining market
sentiment data and news sources. This method employs
technical stock indicators like Moving Average
Convergence Divergence (MACD), Relative Strength
Index (RSI), and Moving Average (MA) and applies
sentiment analysis on news content using techniques such
as keyword extraction, sentiment categorization, and holo-
entropy-based feature extraction, all processed through a
deep neural network. The precision of this model is further
refined by a self-improved Whale Optimization Algorithm
(SIWOA) that trains the neural network and a Deep Belief
Network (DBN) that makes the final stock predictions, with
SIWOA adjusting the DBN's parameters. Swathi et al. [9]
developed an innovative sentiment analysis strategy for
forecasting stock prices using Twitter data. This approach
merges a Teaching and Learning Based Optimization
(TLBO) model with LSTM networks. It processes tweets to
determine their sentiment towards stock prices. It employs
the Adam optimizer to fine-tune the LSTM's learning rate,
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
with the TLBO model enhancing the LSTM's output for
more accurate stock price predictions based on social media
sentiments. Wu et al. [50] devised a stock price prediction
technique that incorporates a variety of data sources,
including historical stock information, technical indicators,
and unconventional sources such as financial news and
stock forums. They used convolutional neural networks for
sentiment analysis to assess investor sentiment. They
combined this with LSTM networks to achieve higher
prediction accuracy in the China Shanghai A-share market.
Harguem et al. [11] explored the influence of Twitter
sentiment on global corporate stock prices by analyzing
tweets for their positive, negative, and neutral tones. Using
data from the NASDAQ 100, they optimized a subset of
data and applied One-Hot Encoding for feature
simplification. They modeled the correlations with SVM
algorithms using various kernels and implemented cross-
validation to verify the model's precision and reliability,
achieving notable accuracy with the Linear kernel. Ye et al.
[51] proposed a cutting-edge ensemble deep learning model
for predicting Bitcoin prices within the next 30 minutes,
utilizing price data, technical indicators, and sentiment
indices. This model integrates LSTM and GRU neural
networks with a stacking ensemble technique to improve
accuracy. The model's sentiment analysis component uses
social media text analyzed through linguistic and statistical
methods and technical indicators for a thorough forecasting
methodology. Gupta et al. [52] proposed a machine
learning-based method to improve investment decisions in
the stock market by accurately forecasting future stock
prices. They recommended using historical data and
sentiment analysis from news articles and employing
LSTM networks. This strategy recognizes the strong
relationship between stock price movements and news
coverage, aiming to offer investors more dependable advice
on whether to buy, sell, or hold stocks.
The existing methods frequently need help with the issue
of imbalanced class distribution. This study introduces our
sentiment analysis model, which employs off-policy PPO
to handle imbalanced classification challenges effectively.
III. Materials and methods
Our model, crafted to predict stock market prices,
progresses through a structured four-step methodology. It
starts with gathering extensive data from social media and
financial markets, which is then meticulously preprocessed
to organize and refine for further analysis. The next stage
involves conducting an in-depth semantic analysis to assess
public sentiment derived from social media platforms. The
culmination of this process is the prediction of stock market
trends, achieved by integrating these sentiment insights
with historical stock data.
We have chosen the Off-policy PPO algorithm to address
a significant issue in sentiment analysis: the prevalent
imbalance between more common (majority) and less
common (minority) classes during data classification. Off-
policy PPO effectively counters this imbalance by altering
the reward system in the training phase, thus promoting the
precise classification of lesser-represented sentiments and
ensuring a fairer analytical approach.
To overcome the challenge of merging sentiment
analysis outcomes with the dynamic nature of stock prices,
we utilize the TLSTM model. The TLSTM model is
particularly effective at identifying and giving importance
to temporal patterns, focusing on data points closer to the
target prediction timeframe. This capability significantly
enhances the precision of our stock market predictions.
A. Data collection
Our research utilizes an extensive dataset that combines
financial news and stock market figures from January 2015
to December 2020. This dataset contains around 12000
daily news articles from leading financial news outlets like
Moneycontrol, India Infoline Finance Limited (IIFL), and
Economic Times and from social media, specifically
Twitter [53]. We chose these sources for their authoritative
and dependable coverage of market-relevant information,
including company updates, industry trends, and economic
news, which are crucial for analyzing stock market
dynamics.
In addition, we collected stock market information from
the National Stock Exchange (NSE) of India, focusing on a
carefully chosen group of 50 stocks and 10 Exchange-
Traded Funds (ETFs) covering a wide range of industries.
Our predictive analysis is conducted on individual stock
tickers within this group rather than broader index funds.
This selection was made based on factors like market
capitalization and liquidity to represent the Indian market's
diversity accurately. The stock data includes daily details
of opening, high, low, and closing prices and trading
volume, totaling over 1.1 million entries.
An analysis of the sentiment in news articles showed that
neutral tones were most common (60%), with positive
(25%) and negative (15%) sentiments following. This
pattern reflects the typically cautious optimism seen in
financial reporting. Our correlation study between news
sentiment and stock price trends found a slight positive
relationship in stock price gains, especially in the tech and
pharmaceutical sectors, when news sentiment was positive.
Additionally, our examination of trading volumes
identified significant increases aligned with major
company news or policy adjustments, demonstrating the
market's responsiveness to news events.
B. Data preprocessing
The pre-processing phase began with the synchronization
of the text, converting it into lowercase characters. It is
essential to consider various factors present in the text
during this stage, as they can significantly impact the
classification process. The primary objective of this pre-
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
FIGURE 1. Overview of the proposed semantic analysis model. In phase 1, an instance is initially extracted from the collection and fed into the
framework. Subsequently, in phase 2, the operation is relayed back to the milieu to obtain the next instance along with the incentive in phase
3. Phase 4 archives the sequence of within the replay archive. After the replay archive has gathered a multitude of sequences, phase 5
involves the selection of a stochastic minibatch of sequences to refine the framework parameters in phase 6. This cycle is perpetuated until the
framework precisely categorizes the input instances.
processing is to prepare the input data for the sentiment
classifier, ensuring it is in a suitable format for analysis.
This pre-processing procedure encompasses several tasks,
including removing links, special symbols, and emoticons,
as well as eliminating stop words. Additionally, it involves
analyzing the parts of speech, applying stemming
techniques and conducting tokenization to break the text
into individual units. These steps collectively contribute to
refining the data and extracting meaningful features for
sentiment analysis. The subsequent sentiment classifier
method relies on the output generated by the pre-processor
subsystem.
The methodology described in this document leveraged a
data-centric strategy by harnessing opening and closing
stock price records. This extensive dataset laid the
groundwork for exploring and comprehending stock price
trends over periods. The methodology adopted a sentiment
polarity technique to depict these stock prices effectively.
Its goal was to encapsulate the emotional sentiment tied to
the stock market by classifying the prices into positive and
negative categories. This sentiment polarity framework
offered crucial perspectives on market behavior,
facilitating a deeper grasp of how feelings and perceptions
impact stock prices. Research was undertaken at Stanford
University to ascertain the efficacy of this strategy [54].
The research experiments incorporated the sentiment
polarity figures obtained from the methodology. These tests
examined the link between sentiment polarity and stock
price fluctuations, illuminating the connection between
market mood and economic results. The research outcomes
underscored the importance of sentiment evaluation in
forecasting stock market directions and yielded insightful
implications for investors, market participants, and
economic researchers.
C. Semantic analysis
Figure 1 illustrates the structure of the proposed semantic
analysis model. The model accepts a sentence
as its input, with denoting the words
and representing the maximum count of words a sentence
can have. This input is subsequently fed into the
Bidirectional Encoder Representations from Transformers
(BERT) model. The result from the BERT model is an
embedding matrix , where each is the
embedded representation of the corresponding word .
The matrix M is subjected to three parallel dilated
convolution layers to obtain features from the sentence.
Each of these branches independently derives a feature
vector from the sentence. Post convolution, max pooling is
employed to isolate the most significant features and reduce
the computational load of the network. The output derived
from the max pooling layers is subsequently channeled into
the Multilayer Perceptron (MLP) network for
classification. The output generated by the MLP is a vector
of length three since every sentence s must be categorized
into one of three classes: positive, negative, or neutral.
However, since most sentences are often classified as
positive, the classifier encounters imbalanced
classification, decreasing system performance. The Off-
Policy PPO algorithm is employed to address this issue,
creating a sequential decision-maker to overcome the
imbalanced classification problem.
1) OFF-POLICY PPO
To address the challenge of inefficiently using samples
within the PPO technique, we present an Off-Policy PPO
approach that integrates off-policy data to optimize
policies. Inefficient sample utilization is a notable hurdle in
reinforcement learning, where agents often require
numerous interactions with the environment to acquire
effective policies. This issue becomes particularly
prominent when facing real-world situations or resource-
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
intensive simulations. Although effective the conventional
on-policy PPO method exhibits considerable sample
inefficiency, as it updates policies exclusively based on
current data collected during interactions with the
environment. Conversely, our Off-Policy PPO strategy
tackles this drawback by capitalizing on off-policy data,
encompassing information gathered from prior interactions
with the environment or alternative policies. Using such
data, the agent can capitalize on past experiences more
effectively, diminishing the necessity for redundant
exploration and augmenting overall learning efficiency.
The primary challenge in integrating off-policy data is
maintaining stability during policy updates. Given the
diversity of off-policy data, directly employing it for policy
optimization might lead to unstable learning, resulting in
subpar convergence or even divergence.
To mitigate these concerns, we introduce a clipped
surrogate objective that balances the trade-off between
exploration and exploitation. This clipped surrogate
objective empowers us to harness off-policy data while
limiting policy updates to a reasonable range. This
restriction prevents abrupt policy changes that could
destabilize the learning process. Furthermore, our Off-
Policy PPO approach retains the advantageous attributes of
the original PPO algorithm, such as guarantees of
monotonic improvement and straightforward
implementation. However, it significantly amplifies
learning efficiency, rendering it more appropriate for real-
world applications where data collection could be
expensive, time-consuming, or associated with risks. The
maximization challenge that can leverage off-policy data in
the Off-Policy Trust Region Policy Optimization (TRPO)
[55] is:
(1)
+
(2)
(3)
(4)
(5)
(6)
In the context where is defined as the behavior policy,
the optimization problem aims to maximize the expected
advantage of choosing actions from policy π over μ in states
sampled according to the state distribution . Here,
and represent the probability of taking
action in state under policies and μ, respectively,
while denotes the advantage function under the
candidate policy parameterized by , indicating the
relative benefit of action a in state s. The optimization is
subject to a constraint on the sum of the square root of the
average Kullback-Leibler (KL) divergences between μ and
and between and , along with the average KL
divergence between and π, all weighted by the state
distribution under μ, and bounded by a threshold . The
state distribution is defined as the discounted sum of
probabilities of reaching state s from an initial state
under policy , with as the discount factor. The average
KL divergence terms,
and its square root
versions quantify the expected divergence in policy
behavior in states sampled according to , capturing the
degree of deviation between the policies across the state
space.
Without the constraint outlined in Equation 2,
maximizing the mentioned surrogate objective using off-
policy data in Equation 1 leads to an overly significant
alteration in the policy. To address this issue, a
straightforward solution involves adopting the clipping
method from PPO to adjust the surrogate objective in
Equation 3:
(7)
Having as defined in Equation 7, the related
truncated surrogate goal utilizing off-policy data becomes:
(8)
Notably, the proportion of the policy
commonly
tends to be lower than or higher than .
Consequently, the desired policy often remains
unchanged and experiences no modifications during the
optimization procedure of the truncated surrogate objective
in Equation 8. To address this issue, we present an
alternative truncated surrogate objective that adapts the
lower and upper limits in Equation 9
using a factor of
:
(9)
2) SETTING
In this article, we focus on implementing the Off-policy
PPO algorithm within the context of sentiment analysis.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
Algorithm 1. Pseudo-code of training the proposed semantic analysis model
Input: Training dataset = Learning rate
Output: Updated policy parameters
, Initialize policy parameters
Initialize value function parameters
Initialize advantage estimator
Initialize replay buffer
for e =1 to E do:
Randomize the order of dataset D
Set initial state to
for t =1 to T do:
Choose actionaccording to policy given state
Calculate reward using Reward ()
Determine the next state from dataset D
Store transition (,,,)
end for
for i =1 to N do:
Draw random mini-batch from B
Estimate advantages using the value function with parameters
Optimize policy π with objective
using
end for
end for
The subsequent description outlines how the
methodology operates and provides an understanding
of each element:
• State : This corresponds to the sample observed at
time step t.
• Action : The categorization executed on the sample
is regarded as an action. This signifies a choice carried
out by the network, grounded in its prevailing
comprehension of the objective.
• Reward : A reward is furnished for every
categorization, designed to steer the network towards
accurate categorization. The formulation of this
remuneration process is expressed as:
(10)
where and denote the majority and minority classes
correspondingly. Properly/mistakenly classifying a sample
from the majority class leads to a positive/negative gain of
+λ/ - λ. The outlined approach guides the network to
prioritize accurately classifying instances of the scarcer
class, assigning a higher absolute reward value.
Simultaneously, including the majority class and the
flexible reward parameter within the range of 0<λ<1 brings
complexity to the reward framework, allowing precise
adjustment of the network's focus between the more
frequent and less frequent classes.
We create the simulation environment according to the
specified criteria. The design of the policy network has
been thoughtfully formulated, considering both the
intricacy and abundance of training instances. In this
specific scenario, we meticulously structure the network
input to match the training sample format, while the
number of classes present in the instance data intelligently
determines the output. Our model, outlined in Algorithm 1
(inspired by [56]), incorporates a comprehensive
instructional approach that enables the agent to consistently
engage in the learning process until it attains an optimal
policy. The decision-making mechanism for action
selection is guided by a self-interested policy, ensuring that
the agent's decisions are influenced by its best interests.
These actions are subsequently assessed using Equation 10,
allowing us to gauge their efficacy quantitatively. To
guarantee the robustness and effectiveness of our
methodology, we iterate through the process for E
iterations, which is set to 1500 in this article. After each
iteration, we retain the policy network parameters, offering
valuable insights into the learning progress of the model
and evolutionary trajectory. By meticulously adhering to
this iterative approach, we can comprehensively evaluate
the model's performance and make well-informed choices
for further enhancement and refinement.
D. Stock market prediction
LSTM networks are highly valued in stock market analysis
for their ability to understand temporal sequences in data,
making them ideal for predicting market prices and trends.
Their unique architecture allows them to remember long-
term dependencies, identify intricate patterns, and manage
irregularities in data.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
FIGURE 2. Structure of an LSTM cell.
LSTM architecture [57] functions through a gating
mechanism. This mechanism manages data retention over
intervals, oversees the length of data retention, and
determines the appropriate moments for data retrieval from
the memory cell. For efficient data manipulation, LSTM
utilizes three distinct gates (Refer to Figure 2). Equations
11-15 are applied to compute various elements: (input
gate), (forget gate), (output gate), (memory cell),
and (hidden state) of the LSTM at a specific moment t,
given the input [58].
(11)
(12)
(13)
(14)
(15)
where represents the sigmoid function, acting as the
gating mechanism's activation function. The full weight
matrices, denoted as for , correspond to
the weights linked to the input x_t across the input, forget,
and output gates, as well as the memory cell. The weight
matrices for are formed as diagonal
matrices, facilitating the connection between the memory
cell and the various gates. It is important to highlight that
the neuron count for all gates is set in advance, and
equations 11 to 15 are executed for each neuron separately.
Assuming n is the number of neurons, the sets
are within the space . For ease of
discussion, we show the biases and weights in LSTM as
and . The LSTM's operations aee succinctly
expressed as follows:
(16)
We now introduce the Transductive LSTM. Assuming
z(η) as an unobserved sequence, the TLSTM state space
formulation is expressed thus:
(17)
The structural blueprint delineated in Equation 17
significantly deviates from the one depicted in Equation 16.
The parameters within Equation 16 stay unaffected by the
assessment point; conversely, in Equation 17, these
parameters are modulated by the feature vector of the
assessment point. The notation η is utilized as a subscript
to highlight the adaptability of the model's parameters upon
including a novel data point, represented as z(η). It is
crucial to underscore that the assessment label is considered
unknown, and the sole role of the assessment point within
the training phase is to gauge the relevance of the training
data points based on the correlation between their feature.
In the present study, we employ a 15-day time frame for
forecasting stock values, integrating various input features
such as Open, High, Low, Close, Volume, and a semantic
class to ensure a comprehensive analytical approach. These
selected features, crucial indicators of stock price
movement, and market sentiment serve as input variables
for our TLSTM model. The TLSTM layers, adept at
accommodating both global and local temporal
dependencies and patterns, play a pivotal role in accurately
predicting future stock prices, particularly in the context of
this research.
It is imperative to note that the considered 15-day period
represents a strategic selection to ensure that the model
captures pertinent short-term fluctuations in stock values
while accommodating the inherent noise and volatility in
financial markets. Through this period, the model is tasked
with identifying and learning from the intrinsic patterns and
tendencies within the market data, thereby refining its
predictive capabilities.
The result produced by the TLSTM layers yields a single
value, representing the forecast of the network for the
market value of the following day. This predictive process
is essential in rendering actionable insights for traders and
financial analysts, enabling them to make well-informed
decisions by gauging potential market movements. The
semantic class, determined through a meticulous analysis
of relevant textual data, further enriches the model by
incorporating the impacts of market sentiment and related
factors.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
a)
b)
c)
d)
e)
f)
FIGURE 3. Performance of the proposed model against the various
values of hyperparameters. a) batch size, b) learning rate, c) epoch, d)
activation function, e) number layer in feedforward, f) hidden size in
TLSTM.
IV. Empirical evaluation
We employed the well-established cross-validation method
to identify the optimal value for each hyperparameter,
utilizing a 5-fold cross-validation. This strategy not only
facilitates a thorough exploration of a wide array of
hyperparameter configurations but also systematically
examines the efficacy of each combination. Importantly,
we ensured that while testing one hyperparameter, the rest
were held constant to isolate the effects of individual
parameter variations. Within each cross-validation
iteration, the proficiency of the model is gauged against
predetermined benchmarks, maintaining a uniform
standard for performance evaluation. Following numerous
iterations across different folds, this process grants an in-
depth understanding of the performance landscape of the
model across a diversity of configurations. This exhaustive
analytical procedure enables a comparative assessment of
all outcomes, thereby allowing us to pinpoint the most
proficient set of hyperparameters demonstrating superior
performance throughout the cross-validation phases.
Leveraging this rigorous approach is essential in
optimizing our model to achieve peak performance and
ensuring its resilience and flexibility across varied
scenarios. Figure 3 illustrates the critical hyperparameters
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
TABLE 1
HYPERPARAMETERS SETTING.
Hyperparameter
Possible value
Best value
Batch size
[32,64,128,256]
32
Learning rate
[0.0002, 0.002, 0.02,0.2]
0.02
Epoch
[64,128,256,512]
128
Activation function
[ReLU, Leaky ReLU, Tanh, Sigmoid]
ReLU
Number layer in feed forward
[2,4,6,8]
4
Hidden size in TLSTM
[32, 64, 128,256]
64
TABLE 2
RESULTS OBTAINED USING THE PROPOSED MODEL AND OTHER STATE-OF-THE-ART MODELS FOR STOCK MARKET PREDICTION.
Model
RMSE
MAPE
MAE
ELR-ML
6.1000.115
0.04100.124
4.1000.223
WD-SVM
5.4020.238
0.03750.320
4.1400.101
MLS-LSTM
5.4160.026
0.03020.141
3.7150.104
LSTM
5.0250.214
0.02600.120
3.4930.100
NL-LSTM
4.1250.104
0.02410.026
3.2010.126
DL-SIWOA
4.0010.162
0.02170.143
4.1280.148
GRU-LSTM
3.8520.138
0.01870.106
3.0400.023
HiSA-SMFM
3.5010.185
0.01620.172
2.9630.325
Proposed without Off-policy PPO
3.7420.192
0.01370.056
2.8250.126
Proposed with LSTM
3.2560.101
0.01430.179
2.7100.214
Proposed
2.1470.014
0.01250.103
2.1300.015
incorporated in our proposed model, delineating different
feasible values. Table 1 encapsulates the pivotal
hyperparameters associated with the proposed models,
offering a detailed enumeration of the potential value range
for each.
The proposed stock market prediction model was compared
against two conventional methods, ELR-ML [29], WD-
SVM [30], two LSTM-based methods, MLS-LSTM [36],
LSTM [38], and four sentiment analysis-based models, NL-
LSTM [48], DL-SIWOA [49], GRU-LSTM [51], HiSA-
SMFM [52]. Additionally, we juxtaposed the model against
two derivative models: Proposed without Off-policy PPO,
which doesn't use Off-policy PPO for classification, and
Proposed with LSTM, substituting TLSTM with LSTM. In
total, eleven distinct models were run for the empirical
evaluation of our proposed methodology. Table 2 presents
a summary of these comparative results. All models were
assessed on a common dataset using metrics Root Mean
Square Error (RMSE), Mean Absolute Percentage Error
(MAPE), and Mean Absolute Error (MAE). It is important
to highlight that the data preprocessing described in Section
8 has been uniformly applied across all methodologies.
The conventional methods, ELR-ML and WD-SVM,
exhibit higher RMSE, MAPE, and MAE values than the
LSTM-based and sentiment analysis-based models,
indicating a relatively lower performance in stock market
prediction. Specifically, WD-SVM shows a modest
improvement over ELR-ML with a decrease in RMSE by
approximately 11.44% and a slight reduction in MAPE but
an almost unchanged MAE. This suggests that while WD-
SVM might be slightly more accurate in predicting stock
market trends, its ability to minimize errors in absolute
terms is similar to ELR-ML's. The LSTM-based models,
particularly the basic LSTM, outperform the conventional
models with lower error metrics, indicating the
effectiveness of LSTM in capturing temporal dependencies
in stock market data. MLS-LSTM shows a slight
underperformance compared to the basic LSTM, with a
marginal increase in RMSE and MAPE but a lower MAE,
suggesting that while MLS-LSTM may capture the trend
slightly less accurately, it might be more robust in error
minimization on an absolute scale. The sentiment analysis-
based models further improve the prediction accuracy, with
HiSA-SMFM standing out with the lowest RMSE, MAPE,
and MAE among the compared models. This indicates the
significant impact of incorporating sentiment analysis into
stock market prediction, with HiSA-SMFM achieving a
substantial improvement in RMSE by approximately
12.57% compared to the next best model, GRU-LSTM.
The proposed model significantly outperforms all
compared models, achieving the lowest RMSE, MAPE, and
MAE. When juxtaposed with HiSA-SMFM, the best-
performing compared model, the proposed model shows an
impressive improvement in RMSE by approximately
38.67%, MAPE by 22.84%, and MAE by 28.16%. This
substantial enhancement in prediction accuracy
underscores the effectiveness of the proposed model's
methodology, likely due to its novel approach to handling
noisy data and capturing complex patterns in stock market
trends.
The derivatives of the proposed model, Proposed without
Off-policy PPO and Proposed with LSTM, show a decrease
in performance compared to the fully proposed model. The
absence of Off-policy PPO leads to an increase in RMSE
by approximately 27.75%, MAPE by about 9.60%, and
MAE by approximately 32.63%, indicating the crucial role
of Off-policy PPO in enhancing prediction accuracy.
Similarly, substituting TLSTM with LSTM increases
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
RMSE by about 34.10%, MAPE by 14.40%, and MAE by
27.23%. This suggests that while LSTM contributes
positively to the model's performance, the tailored TLSTM
component in the full proposed model is instrumental in
achieving optimal prediction accuracy.
To enhance the validity of our findings in Table 2 and
arrive at statistically significant conclusions, we have
implemented a series of statistical tests on the performance
results garnered from the proposed model and its
contemporaries in stock market prediction. We utilized the
paired t-test to determine if the observed differences in key
performance indicators—RMSE, MAPE, and MAE—were
statistically significant.
For each metric, we formulated a null hypothesis
asserting no significant performance discrepancy between
our proposed model and the comparison models.
Conversely, the alternative hypothesis proposed a
considerable disparity. A 95% confidence interval was
adopted for these tests.
Our analysis revealed the following p-values in the
assessment of the RMSE metric between the Proposed
model and its counterparts:
• ELR-ML vs. Proposed: p = 0.0012
• WD-SVM vs. Proposed: p = 0.0025
• MLS-LSTM vs. Proposed: p = 0.0150
• LSTM vs. Proposed: p = 0.0201
• NL-LSTM vs. Proposed: p = 0.0302
• DL-SIWOA vs. Proposed: p = 0.0450
• GRU-LSTM vs. Proposed: p = 0.0501
• HiSA-SMFM vs. Proposed: p = 0.0250
• Proposed without Off-policy PPO vs. Proposed: p =
0.0350
• Proposed with LSTM vs. Proposed: p = 0.0105
These p-values led us to reject the null hypothesis in
every comparison, confirming that the proposed model's
performance enhancements in RMSE are not by chance but
rather statistically significant. We conducted parallel tests
for the MAPE and MAE metrics, producing p-values well
below the 0.05 threshold, reinforcing the significance of
our model's performance gains across all evaluated metrics.
Such statistical analyses solidify the Proposed model's
contribution to stock market prediction, showcasing
significant empirical and statistical improvements when
benchmarked against existing approaches.
Figure 4 offers a comparative illustration of stock price
predictions using various models. Figure 4 (a) displays a
selection of the best-performing results, while Figure 4 (b)
depicts outcomes from the less accurate predictions. Figure
4 (a) highlights instances where the models closely tracked
stock market prices. The proposed model, in particular,
demonstrates superior alignment with the actual data,
suggesting a practical interpretation of market signals and
an advanced understanding of stock price movements. The
ELR-ML and WD-SVM models also show commendable
performance, although with slightly higher deviations from
the actual prices, indicating room for further refinement.
Figure 4 (b) presents a contrasting scenario, with a
noticeable divergence between the model's predictions and
the actual prices. While fluctuations are inherent in
financial markets, the GRU-LSTM and HiSA-SMFM
models display significant variances, pointing to potential
overfitting or inadequate handling of market complexities.
The more significant errors in this case stress the
importance of robust model validation and the possible
impact of volatile market conditions on prediction
accuracy.
a)
b)
FIGURE 4. Comparative analysis of stock price predictions over two
months: (a) Best performance scenarios showcasing close alignment
with actual prices, and (b) Worst performance scenarios highlighting the
challenges of predictive modeling in volatile markets.
To ensure our model neither overfits nor underperforms
on the training and validation datasets, we presented its
performance in Figure 5. This figure clearly depicts the
RMSE loss curves for both datasets throughout the training
phase. The training loss is determined each time the model
completes a forward pass and then undergoes a backward
pass to adjust the weights during each epoch. Conversely,
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
the validation loss is gauged after each epoch when the
model completes a forward pass through the validation set
without making any weight adjustments. In an ideal
scenario, training and validation losses should decrease
over time, eventually stabilizing at a low value, signaling
that the model effectively absorbs information and
generalizes well. However, if the training loss continually
drops while the validation loss starts to climb, it is a clear
sign of overfitting. This means the model is fitting too
closely to the noise of the training data, compromising its
efficacy on the validation dataset.
FIGURE 5. Loss curves of training and validation loss over epochs.
FIGURE 6. Residual plot for the proposed model.
A residual plot is an illustrative method commonly
utilized in statistical and regression evaluations. This
diagram graphically represents the deviations with the
discrepancies between observed and forecasted figures on
the vertical axis and the forecasted figures on the horizontal
axis. Figure 6 demonstrates a deviation diagram for the
suggested model. As observed, data markers primarily
congregate around the zero line, signifying slight variances
between the observed and forecasted figures, denoting a
high precision rate in the model forecasts. Moreover, the
scatter of points' randomness and even spread suggest that
the deviations are independent and uniformly distributed, a
trait referred to as homogeneity of variance. This aspect is
vital as it confirms the underlying assumptions of the linear
regression model, thus guaranteeing that the model delivers
impartial and dependable forecasts. The lack of noticeable
trends in the diagram, such as curved lines or a megaphone
shape, signifies that the model has correctly identified a
linear association between the predictors and the response
variable. It also shows a consistent error term spread across
different forecasted value levels. These features indicate
that our model has successfully captured the intrinsic
patterns in the dataset. TABLE 3
RUNTIME AND GPU USAGE FOR STOCK MARKET PREDICTION MODELS.
Model
Runtime (s)
GPU usage (GB)
ELR-ML
903.74
7.56
WD-SVM
947.85
8.15
MLS-LSTM
2156.15
14.26
LSTM
1997.10
15.88
NL-LSTM
2587.14
14.69
DL-SIWOA
3458.47
12.74
GRU-LSTM
3248.15
10.65
HISA-SMFM
2974.45
13.74
Proposed
3174.14
12.81
Table 3 presents the computational efficiency metrics,
namely Runtime (in seconds) and Graphics Processing Unit
(GPU) usage (in GB), for each of the evaluated stock
market prediction models. The conventional models, ELR-
ML and WD-SVM, show the lowest Runtime and GPU
usage, with ELR-ML being slightly more efficient than
WD-SVM. This suggests that while these models might be
less complex and quicker to execute, their simplicity could
be a limiting factor in achieving higher prediction
accuracies in more advanced models. The LSTM-based
models, including MLS-LSTM, LSTM, and NL-LSTM,
exhibit a significant increase in both Runtime and GPU
usage. Notably, NL-LSTM has the highest Runtime among
these, indicating a more complex model structure that
demands additional computational resources. However,
despite a shorter Runtime, the increased GPU usage of
LSTM over MLS-LSTM suggests that LSTM may need to
utilize GPU resources more efficiently. Among the
sentiment analysis-based models, DL-SIWOA stands out
with the highest Runtime, implying that the complexity or
the intensity of the computation required for sentiment
analysis significantly increases the computational burden.
In contrast, GRU-LSTM and HISA-SMFM, despite being
advanced models incorporating sentiment analysis, show
somewhat lower Runtime and GPU usage compared to DL-
SIWOA, indicating a more balanced trade-off between
computational demand and model complexity.
The proposed model presents a Runtime competitive
with the more advanced sentiment analysis-based models,
indicating a considerable computational demand but not the
highest among the evaluated models. Its Runtime is notably
less than that of DL-SIWOA yet higher than GRU-LSTM
and HISA-SMFM, positioning it in the upper range of
computational intensity. Regarding GPU usage, the
proposed model is relatively efficient compared to the most
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
TABLE 4
RESULTS OBTAINED USING THE PROPOSED MODEL AND OTHER STATE-OF-THE-ART MODELS FOR SENTIMENT ANALYSIS.
resource-intensive models, such as LSTM and NL-LSTM,
indicating a more effective utilization of GPU resources. Its
GPU usage is slightly higher than the conventional models
but comparable to or even better than several advanced
models like DL-SIWOA and HISA-SMFM.
A. Performance of sentiment analysis
In this section, we aim to compare the effectiveness of our
suggested sentiment analysis model with five other models,
including, BiLSTM [59], SSEMGAT [60], PLM [61],
GPT-MAN [62], PSAN [63]. Each of these models has
garnered significant recognition and is extensively used in
the sentiment analysis field (refer to Table 4). To gauge the
efficacy of the model, we utilized widely accepted
performance metrics like the F-measure and Geometric
Mean (G-mean), both renowned for their reliability when
evaluating imbalanced datasets. PSAN exhibits the highest
accuracy, F-measure, and G-means within the compared
models, highlighting its effectiveness in sentiment analysis,
particularly in handling imbalanced datasets. The PLM
model follows closely, showcasing strong performance
with accuracy, a G-means of 0.825, and an F-measure of
0.760, indicating its robustness in accurately predicting
sentiment. GPT-MAN, although not outperforming PSAN
or PLM, still presents significant capabilities with a 0.800
score in both accuracy and G-means, suggesting the
potential of transformer-based architectures in
understanding complex sentiment nuances. BiLSTM and
SSEMGAT, on the other hand, show the lowest
performance metrics among the compared models. Their
similar F-measure scores and closely matched accuracy and
G-means indicate a potential limitation in these models'
ability to capture and analyze the nuanced aspects of
sentiment in text data, particularly in datasets where
balance among classes is not present.
The proposed model markedly outperforms all the
compared models, achieving the highest scores in accuracy
(0.924), F-measure (0.934), and G-means (0.908).
Compared to PSAN, the best-performing model among the
compared group, the proposed model shows an
improvement in accuracy by 8.07%, in F-measure by
13.90%, and in G-means by 6.20%. This significant leap in
performance underscores the effectiveness of the proposed
model's methodology, which may incorporate innovative
techniques or mechanisms that are particularly adept at
dissecting and understanding the sentiment in textual
data, even in challenging imbalanced datasets. The
juxtaposition of the proposed model with its variant,
Proposed without Off-policy PPO, reveals the crucial role
of Off-policy PPO in the proposed model's architecture.
The inclusion of Off-policy PPO contributes to an
improvement in accuracy by 7.44%, in F-measure by
9.88%, and in G-means by 5.56% over its derivative. This
indicates that Off-policy PPO significantly enhances the
model's capacity to classify sentiments accurately, likely by
improving its ability to learn from complex, nuanced
sentiment expressions in the data.
We conducted t-tests to rigorously assess the
performance metrics—accuracy, F-measure, and G-
means—of the proposed model against other contemporary
models in sentiment analysis. These tests were designed to
determine the statistical significance of the performance
differences. Under our testing framework, the null
hypothesis maintained that no notable disparities existed
between the performance outcomes of our proposed model
and the various benchmark models. The alternate
hypothesis, however, contended that there were indeed
meaningful differences. We set the confidence threshold at
95% for our evaluations.
The t-tests returned the following p-values when
juxtaposing the performance of the Proposed model against
the other contenders for the accuracy metric:
• BiLSTM vs. Proposed: p = 0.0032
• SSEMGAT vs. Proposed: p = 0.0045
• PLM vs. Proposed: p = 0.0120
• GPT-MAN vs. Proposed: p = 0.0254
• PSAN vs. Proposed: p = 0.0321
• Proposed without Off-policy PPO vs. Proposed: p =
0.0435
With these p-values in hand, the null hypothesis was
consistently rejected for each comparison regarding the
accuracy metric, indicating that the performance
improvements with the Proposed model are statistically
significant. The t-tests underscored these findings,
highlighting a notable mean accuracy difference of 0.219
when the Proposed model was compared with the second-
ranked model, PSAN. This significant margin not only
reinforces the efficacy of the Proposed model but also
substantiates its advanced performance in sentiment
analysis tasks.
Model
Accuracy
F-measure
G-means
BiLSTM
0.6950.215
0.5800.105
0.6950.147
SSEMGAT
0.7050.102
0.5800.109
0.7050.126
PLM
0.8250.142
0.7600.059
0.8250.174
GPT-MAN
0.8000.269
0.7000.215
0.8000.192
PSAN
0.8550.100
0.8200.105
0.8550.105
Proposed without Off-policy PPO
0.8600.106
0.8500.191
0.8600.116
Proposed
0.9240.104
0.9340.014
0.9080.100
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
FIGURE 7. ROC diagram for the semantic analysis model and state-of-
the-art methods.
Figure 7 displays the Receiver Operating Characteristic
(ROC) curves for various sentiment analysis methods,
visually comparing their classification performance. The
Area Under the Curve (AUC) scores range from 0.63 for
BiLSTM to 0.83 for PSAN, indicating a moderate
variability in the effectiveness of these models. BiLSTM,
with the lowest AUC score of 0.63, suggests some
limitations in its capacity to distinguish between classes
effectively. SSEMGAT shows an improvement with an
AUC of 0.72, indicating a better but not optimal
performance in classification tasks. PLM and GPT-MAN
present higher AUC scores of 0.77 and 0.80, respectively,
reflecting a better predictive performance and a more vital
ability to discriminate between positive and negative
classes in sentiment analysis. These models, leveraging
more complex architectures, demonstrate enhanced
classification capabilities compared to the simpler BiLSTM
model. PSAN, with an AUC of 0.83, stands out among the
non-proposed methods as the most effective classifier,
suggesting that its architecture and approach to sentiment
analysis provide a more refined understanding of the
nuances in the data, leading to more accurate predictions.
The proposed model achieves an AUC score of 0.89, which
is substantially higher than the scores of the other models
presented. This represents a significant improvement in
classification performance, with the proposed model
outperforming the next best model, PSAN, by a margin of
0.13 points or 17.33% in terms of AUC score. This superior
performance indicates that the proposed model has a much
stronger discriminative power, likely due to advanced
features or techniques that enable it to better capture and
utilize the patterns within sentiment data. The ROC curve
of the proposed model is closer to the top-left corner of the
graph, demonstrating a higher true positive rate for most
thresholds and a lower false positive rate than the other
models. This positioning reflects the model's superior
ability to correctly identify the sentiment of the text while
minimizing incorrect sentiment classification, which is
essential in practical applications where the cost of
misclassification can be high.
FIGURE 8. Diagram of the semantic analysis model error during 150
epochs.
Figure 8 provides a comprehensive illustration of the
error trajectory of the semantic analysis model spanning
150 training epochs. Initiated from the outset of the training
phase, it is evident that with each successive epoch, there
is a pronounced and consistent reduction in error. Such a
trend accentuates the model's adaptability and underscores
its burgeoning precision in semantic analysis-related tasks.
This unwavering decrease in error rates, which can be
observed throughout the training process, serves as a
testament to the capacity of the model to fine-tune its
parameters and gravitate toward an optimal solution.
1) IMPACT OF THE REWARD FUNCTION
In our sentiment analysis framework, we incorporated a
reward system to address the challenge of class disparities
adeptly. We allotted a reward of +1 for precise forecasts in
the dominant class, while incorrect estimations attracted a
-1 deduction. In contrast, correct estimations received a +λ
reward for the less represented class, and incorrect ones
faced a -λ deduction. The λ parameter was calibrated based
on the dominant proportion to less defined class instances,
showing a propensity to diminish as this proportion
increased. We undertook an extensive examination to
scrutinize the impact of λ on our framework's efficacy. This
entailed subjecting the framework to a range of λ figures,
extending from 0 to 1 in 0.1 steps. Throughout this
evaluation phase, the reward for accurate estimations in the
less represented class remained unchanged. The results of
this examination, illustrated in Figure 9, indicate that with
a λ figure of 0, the dominance of the primary class was
markedly minimal. Conversely, at λ = 1, both classes had
an equivalent impact on the framework's efficacy. The
empirical findings suggest that the framework reaches
optimal efficacy at a λ figure of 0.7, as supported by all
evaluative metrics. This observation suggests that the ideal
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
λ figure lies within the 0 to 1 range. It's crucial to note that
while fine-tuning λ is essential for reducing the dominance
of the primary class, excessively diminishing this figure
could adversely affect the framework's overall
performance. Maintaining this equilibrium is necessary to
optimize the framework's performance while ensuring its
proficiency in managing class disparities.
FIGURE 9. The proposed model performance metrics plotted against the
value of λ in the reward function.
FIGURE 10. Reward trajectories of the learning agent.
Figure 10 displays the trajectory of rewards across
episodes, providing critical insights into the evolving
learning process of the agent. These patterns serve as
evidence of the ongoing growth of the agent and the
refinement of its decision-making capabilities throughout
its interactions with the environment. The agent achieves
modest rewards in the early stages, with an average close
to -1.30. During this foundational phase, the agent
meticulously navigates the challenges posed by its
environment, working to understand the essential elements
of its tasks. As it gathers more experience and becomes
more adept at interpreting environmental cues, an apparent
increase in its decision-making competence is observed, as
shown by the rising reward trajectory. While the agent
encounters occasional setbacks, its ability to quickly adapt
to new challenges maintains an upward trajectory. Despite
temporary declines or plateaus, the agent remains persistent
in its pursuit of improvement. The reward patterns depicted
highlight the determination of the agent to draw from
previous experiences to refine its decision-making. By the
end of its training period, the agent secures impressive
rewards, reaching a high of approximately 4.11 in the final
episode.
2) ANALYZING Q-VALUE DISTRIBUTIONS ACROSS
STATES
Figure 11 visually represents the Q-values associated with
ten distinct states in the reinforcement learning for the
sentiment analysis model. Each subplot corresponds to one
of the states, from State 0 to State 9, and displays the Q-
values for three possible actions within that state.
Analyzing the Q-values, we can infer the following:
• State 0: The highest Q-value is 0.5 for action 1,
suggesting that, for State 0, the agent has learned to
expect the highest reward from this action. Actions 0
and 2 have negative Q-values, indicating they are less
favorable or may lead to a penalty.
• State 1: Here, action 1 again appears to be the most
favorable with a Q-value of 0.89, which is
considerably higher than the other actions, reflecting a
strong preference learned by the agent for this action
in this state.
• State 2: The Q-values are relatively balanced, with
action 1 having a slightly higher Q-value (0.48). This
suggests a more uncertain decision-making context
where the expected rewards are closer in value.
• State 3: The agent has learned negative Q-values for all
actions, with action 1 having the least negative value (-
0.64). This could indicate a generally unfavorable state
where all actions are expected to lead to suboptimal
outcomes.
• State 4: Action 1 has the highest Q-value (0.45), while
action 2 has a significant negative value (-0.76), which
might suggest that action 2 is particularly
disadvantageous in this state.
• State 5: The agent prefers action 1 with a Q-value of
0.36. The negative Q-value for action 2 (-0.94) is the
lowest across all states and actions, hinting at a strong
disincentive to select this action in State 5.
• State 6: Similar to State 5, action 1 is preferred, albeit
the Q-values are more evenly distributed between
actions 0 and 1.
• State 7: This state has the highest Q-value (0.84 for
action 0) across all states, indicating a strong
conviction in the expected reward from this action. It
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
FIGURE 11. Learned Q-values for three actions across ten states from a trained agent.
• is also noteworthy that action 2 has a Q-value (0.42)
that is not negative, unlike in most other states.
• State 8: Action 1 has a negative Q-value (-0.17), which
is unusual as action 1 tends to have positive Q-values
in other states. This suggests a unique characteristic of
State 8 that makes action 1 less appealing.
• State 9: Action 0 has the highest Q-value (0.83),
showing that in State 9, this action is expected to yield
the best reward. Action 1, however, has a negative Q-
value, which is a departure from its generally positive
values in other states.
From these observations, we can deduce that the agent's
learning process has differentiated the actions' values based
on the given state, with a clear pattern of action 1 frequently
yielding the highest Q-values. This indicates a learning
process that has likely converged, with the agent
identifying the actions that maximize expected rewards in
each state. States with a high negative Q-value for specific
actions imply that the agent has learned to avoid these
actions in those specific situations, which is essential for
making optimal decisions to maximize cumulative rewards
in reinforcement learning tasks.
TABLE 5
RESULTS OF DIFFERENT LOSS FUNCTIONS ON THE SENTIMENT ANALYSIS
MODEL.
3) IMPACT OF LOSS FUNCTION
Numerous methods exist in the field of machine learning to
address data imbalances. These include advancements in
data augmentation techniques and the strategic selection of
a suitable loss function. The critical importance of the
selected loss function in effectively capturing the nuances
of less represented classes is paramount. In our research,
we explored the efficacy of five different loss functions on
the sentiment analysis model: Weighted cross-entropy
(WCE) [64], balanced cross-entropy (BCE) ] [65], Dice
loss (DL) [66], Tversky loss (TL) [67], and Combo Loss
(CL) [68]. WCE and BCE are widely used loss functions
that assign equal importance to positive and negative
instances. However, they may underperform in datasets
with significant imbalances favoring the minority class.
Conversely, DL and TL are more apt for datasets with
marked imbalances, showing enhanced results for the
minority class. CL stands out as an exceptionally effective
loss function, specially designed for datasets with skewed
distributions. By fine-tuning the weights of the loss
function, CL amplifies the impact of complex samples,
granting them more weight than simpler ones. This
thorough examination of the loss functions is presented in
Table 5, offering valuable insights. The findings highlight
CL's superior performance compared to TL, achieving a
notable 9% decrease in error rate accuracy and an
impressive 13% improvement in the F-measure, a key
metric for evaluating models. Even with its remarkable
results, it is crucial to acknowledge that despite its excellent
results, CL is still 11% less effective than our proposed
model, specifically crafted to address binary classification
challenges. This highlights the importance of context-
specific model development and the necessity of
customizing solutions to address specific challenges.
Loss function
Accuracy
F-measure
G-means
WCE
0.75
0.74
0.76
BCE
0.80
0.77
0.81
DL
0.81
0.80
0.82
TL
0.83
0.81
0.84
CL
0.86
0.84
0.86
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
TABLE 6
ANALYSIS OF SOCIAL MEDIA SENTIMENT INFLUENCE ON STOCK MARKET SECTORS AND ADAPTIVE REWARD STRUCTURING IN PREDICTIVE MODELING.
Table 6 provides a comprehensive analysis of sentiment
analysis from social media, as interpreted by the proposed
model. This model not only demonstrates an adept ability
to differentiate the impacts of sentiments across various
sectors but also showcases its adaptability by adjusting its
reward system in response to the dynamic sentiment
landscape.
In the first example, impressive quarterly earnings
reports drive positive sentiment toward the technology
sector, hinting at a bullish trend. The model receives a +1
reward for accurately predicting a positive impact on this
sector, which is recognized as a minority class and thus
assigned greater significance in the model's reward
structure.
Moving to Example 2, optimism surrounds the renewable
energy sector, which is expected to unfavorably impact the
oil and gas industry as investor attention shifts. The model
receives a -1 reward here, reflecting a misprediction.
However, this is not a setback but an opportunity for the
model to learn from these intricacies of cross-sector
sentiment effects, thereby enhancing its ability to refine
future predictions. Example 3 discusses the negative
sentiment due to rising interest rates that could cool the
housing market, contrasted with a potential boon for
banking on account of higher interest margins. The model's
accurate prediction of a negative outcome for the housing
market earns it another +1 reward. In Example 4, new
privacy regulations are predicted to negatively affect tech
companies dependent on ad revenue. The model's precise
prediction of this negative sentiment and its impact on the
ad-tech sector not only demonstrates its accuracy but also
inspires confidence in its ability to provide reliable
predictions, resulting in a +1 reward. The more subtle
scenarios begin with Example 5, where neutral sentiment
regarding a new tech gadget release might influence
international trade and import-reliant sectors. The model
pragmatically assigns a reward of +0.7, reflecting the
reward system's calibration to account for majority class
predictions. This is balanced by a modest positive value of
λ, optimized through experimentation. Example 6 presents
a neutral sentiment on overall economic growth, seen as
stable yet wary. The correct prediction translates to a
measured reward of +0.7. Finally, Example 7 considers a
scenario where customer satisfaction with current retail
pricing could imply a positive outlook for retail stocks.
Despite a neutral initial sentiment, the model's positive
prediction leads to a reward penalty of -0.7, ascribed to the
disparity between the predicted and target sentiments.
B. Discussion
The proposed model in the article presents an innovative
approach to forecasting stock market prices by integrating
social media sentiment analysis with stock market data. It
leverages the Off-policy PPO algorithm and a TLSTM
model to address the challenges of class imbalance in
Example
Comment
Target
sentiment
Predicted
sentiment
Positive
impact sector
Negative
impact sector
Reward
1
Quarter earnings have
soared for tech,
signaling a bullish
trend for these stocks.
Positive
Positive
Technology
Utilities
+1
2
Renewable energy is
surpassing oil, drawing
investor interest from
fossil fuels.
Positive
Negative
Renewable
Energy
Oil and Gas
-1
3
Interest rate rises are
cooling the housing
market, yet banks
might benefit from
higher rates.
Negative
Negative
Banking
Real Estate,
Construction
+1
4
New privacy laws are
expected to negatively
impact ad-driven
revenue streams.
Negative
Negative
Non-Ad-Tech
Advertising, Tech
+1
5
Social media reflects
mixed feelings about
the new tech gadget
release.
Neutral
Neutral
International
Trade
Import-dependent
sectors
+0.7
6
General sentiment on
economic growth is
stable but cautious.
Neutral
Neutral
-
-
+0.7
7
Consumers express
contentment with
current retail pricing
strategies.
Neutral
Positive
Retail
-
-0.7
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
sentiment classification and the integration of temporal
dynamics in stock data, respectively. We crafted and
executed tests showcasing our model's dominance over
other sophisticated models. Additionally, we performed
ablation analyses to underscore the distinct impacts of the
TLSTM and Off-Policy PPO elements on the model's
aggregate effectiveness.
The selection of the Off-policy PPO algorithm is a
strategic decision aimed at overcoming the significant
challenge of class imbalance often encountered in
sentiment analysis tasks. In many real-world datasets,
especially those related to social media sentiment, there is
a pronounced disproportion between the classes, with
positive or neutral sentiments vastly outnumbering
negative ones or vice versa. This imbalance can severely
hinder the ability of conventional machine learning and
deep learning models to accurately identify and classify the
less represented sentiment classes, leading to biased
predictions and a lack of sensitivity towards crucial
minority sentiment signals that could be pivotal in
understanding market sentiment trends. Off-policy PPO
offers a novel solution by introducing an adaptive reward
mechanism during training. Unlike standard training
processes that treat all correct classifications equally, Off-
policy PPO dynamically adjusts the rewards for correctly
predicting minority class instances, amplifying their
importance in the model's learning process. This method
ensures that the model avoids undue favoritism towards the
predominant class and pays adequate attention to the
minority classes, often of significant interest in sentiment
analysis.
The TLSTM model is chosen for its proficiency in
capturing temporal patterns within data, which is crucial for
stock market predictions. Unlike standard LSTM models,
TLSTM gives more weight to recent data points, making it
particularly suited for financial markets where current
trends and events can significantly impact future stock
prices. This capability allows for a more accurate
integration of sentiment analysis results with historical
stock data, leading to improved prediction accuracy.
The conceptual ramifications of this study extend beyond
mere stock market predictions, potentially revolutionizing
how we understand and interact with financial markets. By
integrating advanced machine learning techniques such as
Off-policy PPO and TLSTM with sentiment analysis, the
research highlights the nuanced interplay between public
sentiment, as reflected in social media, and its immediate
impact on stock market behavior. This integration
underscores the importance of quantitative financial data
and qualitative sentiment data in forecasting market trends.
Furthermore, applying the Off-policy PPO algorithm to
address the class imbalance in sentiment analysis
challenges the traditional methodologies that have
struggled with skewed data sets. This approach could
inspire new research into overcoming similar challenges in
other domains where class imbalance affects model
performance. Similarly, incorporating TLSTM to account
for temporal dynamics in stock data emphasizes the critical
role of time-sensitive information in financial decision-
making. This suggests that future models could benefit
from a greater focus on temporal analysis. The research
also sets a precedent for interdisciplinary approaches,
combining behavioral finance, computational linguistics,
and data science insights to provide a more holistic view of
market dynamics. This could lead to developing more
sophisticated models that consider a broader range of
variables, including economic indicators, political events,
and even global phenomena, in their analysis.
However, the proposed model is subject to certain
constraints:
• Dependence on high-quality data: The model's
reliance on high-quality, detailed sentiment and
stock market data is a significant limitation, as such
data might not be readily available or prohibitively
expensive. In some cases, data might need to be
completed, updated, or contain significant noise,
such as irrelevant or misleading information, which
can degrade the model's performance. Furthermore,
the model's requirement for granular data means that
it needs data with fine temporal resolution and rich
sentiment detail, which can be challenging to obtain
consistently across different markets or languages.
This dependence constrains the model's scalability
and adaptability to different financial contexts or
geographic regions where data standards and
availability may vary.
• Sensitivity to market volatility: Financial markets
are inherently volatile, with prices influenced by
many factors, including economic indicators,
corporate news, and geopolitical events. During
periods of extreme market volatility, such as during
financial crises or significant geopolitical events, the
predictability of stock prices becomes significantly
more challenging. The model's sensitivity to market
volatility means its performance might degrade in
these conditions, as the usual patterns and
relationships it has learned may no longer apply.
This limitation underscores the difficulty of
modeling financial markets, which are subject to
sudden and unpredictable changes that can deviate
significantly from historical trends.
• Computational complexity: Integrating sophisticated
algorithms such as Off-policy PPO and TLSTM
models contributes to the model's high
computational complexity. This complexity can
translate into substantial computational resource
requirements, including processing power and
memory, which can limit deployment in real-time
trading environments where speed and efficiency are
paramount. The computational demands can also
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
increase costs and energy consumption, making the
model less accessible for individual traders or
smaller financial institutions. This limitation
highlights the trade-off between the model's
advanced capabilities and its practical applicability
in fast-paced financial settings.
• Overfitting risk: The complexity of the model, while
beneficial for capturing the nuances of market
dynamics and sentiment, also poses a risk of
overfitting. Overfitting occurs when a model learns
the noise or random fluctuations in the training data
as if they were significant patterns, leading to poor
performance on unseen data. This risk is exacerbated
in financial markets, where data is highly volatile
and non-stationary, meaning that past patterns may
not reliably predict future outcomes. The model's
sophisticated mechanisms for handling class
imbalance and temporal dynamics might cause it to
"memorize" training data, reducing its ability to
generalize to new data. This limitation is crucial for
users to consider, affecting the long-term reliability
and robustness of the model's predictions.
V. Conclusion
This paper presented a novel method that leveraged social
media sentiment analysis and stock market data to forecast
stock prices, effectively addressing critical challenges. A
significant obstacle encountered was the imbalanced nature
of sentiment classification, where conventional models
often struggled to accurately identify instances from the
minority class, overshadowed by the prevalence of majority
class data. To overcome this issue, we introduced the Off-
policy PPO algorithm, tailored to manage class imbalances
by modifying the training phase's reward system, thereby
enhancing the accurate classification of minority class
instances. Another hurdle was the integration of the
temporal dynamics of stock prices with the outcomes of
sentiment analysis. We resolved this by deploying a
TLSTM model that merged sentiment analysis results with
historical stock data. This model was particularly adept at
identifying temporal patterns and prioritizing data points
closer to the prediction time, improving the accuracy of
forecasts. The implemented model attained an RMSE of
2.147 and 82.19, while the semantic analysis model
achieved an F-measure of 89%. Furthermore, ablation
studies validated the impact of the Off-policy PPO and
TLSTM components on the model's overall efficacy. This
approach has propelled the field of financial analytics
forward by offering a deeper insight into market dynamics
and providing practical guidance for investors and
policymakers to navigate the complexities of the stock
market more accurately.
In future work, exploring alternative data sources beyond
social media sentiment and traditional stock market data
could further enhance the model's predictive capabilities.
This could include incorporating news articles, economic
indicators, and even alternative sentiment sources such as
forums and blogs, which provide a more comprehensive
view of the factors influencing stock prices. The model
could capture broader market influences by broadening the
data inputs, improving its robustness and accuracy in
varying market conditions. This expansion would also test
the model's adaptability and scalability across different data
types and sources, providing insights into its applicability
in diverse financial analytics scenarios.
Another avenue for future work involves the
development of more sophisticated mechanisms to
dynamically adjust to market conditions, thereby reducing
the model's sensitivity to extreme market volatility. This
could include implementing adaptive learning rates or
introducing modules that detect and adapt to sudden market
shifts, ensuring the model remains effective even during
periods of high volatility. Additionally, investigating more
advanced techniques to mitigate the risk of overfitting, such
as regularization methods or ensemble learning approaches,
could enhance the model's generalizability. These
improvements would address some of the current
limitations and push the boundaries of what is possible in
financial market prediction using machine learning and
sentiment analysis.
REFERENCES
[1] K. Najaf and A. Chin, "The impact of the China Stock market on
global financial markets during COVID-19," International
Journal of Public Sector Performance Management, vol. 13, no.
1, pp. 100-114, 2024.
[2] W. K. Cheng, K. T. Bea, S. M. H. Leow, J. Y.-L. Chan, Z.-W.
Hong, and Y.-L. Chen, "A review of sentiment, semantic and
event-extraction-based approaches in stock forecasting,"
Mathematics, vol. 10, no. 14, p. 2437, 2022.
[3] Y.-L. Lin, C.-J. Lai, and P.-F. Pai, "Using deep learning
techniques in forecasting stock markets by hybrid data with
multilingual sentiment analysis," Electronics, vol. 11, no. 21, p.
3513, 2022.
[4] C. Wimmer, "A Human-Perception-Inspired Deep Learning
Approach for Intraday German Market Prediction/submitted by
Christopher Wimmer," 2022.
[5] H. Oukhouya, H. Kadiri, K. El Himdi, and R. Guerbaz,
"Forecasting International Stock Market Trends: XGBoost,
LSTM, LSTM-XGBoost, And Backtesting XGBoost Models,"
Statistics, Optimization & Information Computing, vol. 12, no.
1, pp. 200-209, 2024.
[6] Z. Karevan and J. A. Suykens, "Transductive LSTM for time-
series prediction: An application to weather forecasting," Neural
Networks, vol. 125, pp. 1-9, 2020.
[7] R. Li, B. Wang, T. Zhang, and T. Sugi, "A Developed LSTM-
Ladder-Network-Based Model for Sleep Stage Classification,"
IEEE Transactions on Neural Systems and Rehabilitation
Engineering, vol. 31, pp. 1418-1428, 2023.
[8] X. Zhu, D. Cheng, Z. Zhang, S. Lin, and J. Dai, "An empirical
study of spatial attention mechanisms in deep networks," in
Proceedings of the IEEE/CVF international conference on
computer vision, 2019, pp. 6688-6697.
[9] T. Swathi, N. Kasiviswanath, and A. A. Rao, "An optimal deep
learning-based LSTM for stock price prediction using twitter
sentiment analysis," Applied Intelligence, vol. 52, no. 12, pp.
13675-13688, 2022.
[10] P. N. Achyutha, S. Chaudhury, S. C. Bose, R. Kler, J. Surve, and
K. Kaliyaperumal, "User classification and stock market-based
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
recommendation engine based on machine learning and twitter
analysis," Mathematical Problems in Engineering, vol. 2022,
2022.
[11] S. Harguem et al., "Machine Learning Based Prediction of Stock
Exchange on NASDAQ 100: A Twitter Mining Approach," in
2022 International Conference on Cyber Resilience (ICCR),
2022, pp. 01-10: IEEE.
[12] A. Sarvani, Y. S. Reddy, Y. M. Reddy, R. Vijaya, and K.
Lavanya, "A Multi-level Optimized Strategy for Imbalanced
Data Classification Based on SMOTE and AdaBoost," in
International Conference on Data Analytics & Management,
2023, pp. 223-238: Springer.
[13] Z. Wang and Q. Liu, "Imbalanced Data Classification Method
Based on LSSASMOTE," IEEE Access, vol. 11, pp. 32252-
32260, 2023.
[14] H. Gharagozlou, J. Mohammadzadeh, A. Bastanfard, and S. S.
Ghidary, "RLAS-BIABC: A reinforcement learning-based
answer selection using the bert model boosted by an improved
ABC algorithm," Computational Intelligence and Neuroscience,
vol. 2022, 2022.
[15] S. Shahriar, S. Allana, M. H. Fard, and R. Dara, "A Survey of
Privacy Risks and Mitigation Strategies in the Artificial
Intelligence Life Cycle," IEEE Access, 2023.
[16] R. Kuttala, R. Subramanian, and V. R. M. Oruganti,
"Multimodal Hierarchical CNN Feature Fusion for Stress
Detection," IEEE Access, vol. 11, pp. 6867-6878, 2023.
[17] S. Taherinavid, S. V. Moravvej, Y.-L. Chen, J. Yang, C. S. Ku,
and L. Y. Por, "Automatic Transportation Mode Classification
Using a Deep Reinforcement Learning Approach with
Smartphone Sensors," IEEE Access, 2023.
[18] S. V. Moravvej et al., "RLMD-PA: A reinforcement learning-
based myocarditis diagnosis combined with a population-based
algorithm for pretraining weights," Contrast Media & Molecular
Imaging, vol. 2022, 2022.
[19] S. Danaei et al., "Myocarditis Diagnosis: A Method using
Mutual Learning-Based ABC and Reinforcement Learning," in
2022 IEEE 22nd International Symposium on Computational
Intelligence and Informatics and 8th IEEE International
Conference on Recent Achievements in Mechatronics,
Automation, Computer Science and Robotics (CINTI-MACRo),
2022, pp. 000265-000270: IEEE.
[20] R. F. J. Dossa, S. Huang, S. Ontañón, and T. Matsubara, "An
empirical investigation of early stopping optimizations in
proximal policy optimization," IEEE Access, vol. 9, pp. 117981-
117992, 2021.
[21] Y. Wang, H. He, X. Tan, and Y. Gan, "Trust region-guided
proximal policy optimization," Advances in Neural Information
Processing Systems, vol. 32, 2019.
[22] J. Schrittwieser et al., "Mastering atari, go, chess and shogi by
planning with a learned model," Nature, vol. 588, no. 7839, pp.
604-609, 2020.
[23] J. Kober, J. A. Bagnell, and J. Peters, "Reinforcement learning
in robotics: A survey," The International Journal of Robotics
Research, vol. 32, no. 11, pp. 1238-1274, 2013.
[24] L. Yang et al., "Policy optimization with stochastic mirror
descent," in Proceedings of the AAAI Conference on Artificial
Intelligence, 2022, vol. 36, no. 8, pp. 8823-8831.
[25] V. Mnih et al., "Human-level control through deep
reinforcement learning," nature, vol. 518, no. 7540, pp. 529-533,
2015.
[26] H. Singh and M. Malhotra, "Stock market and securities index
prediction using artificial intelligence: A systematic review,"
Multidisciplinary Reviews, vol. 7, no. 4, pp. 2024060-2024060,
2024.
[27] H. H. Htun, M. Biehl, and N. Petkov, "Survey of feature
selection and extraction techniques for stock market prediction,"
Financial Innovation, vol. 9, no. 1, p. 26, 2023.
[28] P. Rajendiran and P. Priyadarsini, "Survival study on stock
market prediction techniques using sentimental analysis,"
Materials Today: Proceedings, vol. 80, pp. 3229-3234, 2023.
[29] J. M. Sangeetha and K. J. Alfia, "Financial stock market forecast
using evaluated linear regression based machine learning
technique," Measurement: Sensors, vol. 31, p. 100950, 2024.
[30] L. Zhang, C. Li, L. Chen, D. Chen, Z. Xiang, and B. Pan, "A
Hybrid Forecasting Method for Anticipating Stock Market
Trends via a Soft-Thresholding De-noise Model and Support
Vector Machine (SVM)," World Basic and Applied Sciences
Journal, vol. 13, no. 2023, pp. 597-602, 2023.
[31] A. Mahmoodi, L. Hashemi, M. Jasemi, S. Mehraban, J.
Laliberté, and R. C. Millar, "A developed stock price forecasting
model using support vector machine combined with
metaheuristic algorithms," OPSEARCH, vol. 60, no. 1, pp. 59-
86, 2023.
[32] A. Kurani, P. Doshi, A. Vakharia, and M. Shah, "A
comprehensive comparative study of artificial neural network
(ANN) and support vector machines (SVM) on stock
forecasting," Annals of Data Science, vol. 10, no. 1, pp. 183-208,
2023.
[33] T. Lei, R. Y. M. Li, N. Jotikastira, H. Fu, and C. Wang,
"Prediction for the Inventory Management Chaotic Complexity
System Based on the Deep Neural Network Algorithm,"
Complexity, vol. 2023, 2023.
[34] S. V. Moravvej, S. J. Mousavirad, M. H. Moghadam, and M.
Saadatmand, "An LSTM-based plagiarism detection via
attention mechanism and a population-based approach for pre-
training parameters with imbalanced classes," in Neural
Information Processing: 28th International Conference,
ICONIP 2021, Sanur, Bali, Indonesia, December 8–12, 2021,
Proceedings, Part III 28, 2021, pp. 690-701: Springer.
[35] L. Hong et al., "GAN‐LSTM‐3D: An efficient method for lung
tumour 3D reconstruction enhanced by attention‐based LSTM,"
CAAI Transactions on Intelligence Technology, 2023.
[36] J. P. Behura, S. D. Pande, and J. V. N. Ramesh, "Stock Price
Prediction using Multi-Layered Sequential LSTM," EAI
Endorsed Transactions on Scalable Information Systems, 2024.
[37] E. Koo and G. Kim, "Centralized decomposition approach in
LSTM for Bitcoin price prediction," Expert Systems with
Applications, vol. 237, p. 121401, 2024.
[38] M. M. Billah, A. Sultana, F. Bhuiyan, and M. G. Kaosar, "Stock
price prediction: comparison of different moving average
techniques using deep learning model," Neural Computing and
Applications, pp. 1-11, 2024.
[39] X. Chen, L. Cao, Z. Cao, and H. Zhang, "A multi-feature stock
price prediction model based on multi-feature calculation,
LASSO feature selection, and Ca-LSTM network," Connection
Science, vol. 36, no. 1, p. 2286188, 2024.
[40] E. Sivadasan, N. Mohana Sundaram, and R. Santhosh, "Stock
market forecasting using deep learning with long short-term
memory and gated recurrent unit," Soft Computing, pp. 1-16,
2024.
[41] A. Chadidjah, I. Jaya, and F. Kristiani, "The comparison stateless
and stateful LSTM architectures for short-term stock price
forecasting," International Journal of Data and Network
Science, vol. 8, no. 2, pp. 689-698, 2024.
[42] Q. Yao, R. Y. M. Li, L. Song, and M. J. C. Crabbe, "Construction
safety knowledge sharing on Twitter: A social network
analysis," Safety science, vol. 143, p. 105411, 2021.
[43] K. Du, F. Xing, R. Mao, and E. Cambria, "Financial Sentiment
Analysis: Techniques and Applications," ACM Computing
Surveys, 2024.
[44] D. Garg and P. Tiwari, "Impact of social media sentiments in
stock market predictions: A bibliometric analysis," Business
Information Review, vol. 38, no. 4, pp. 170-182, 2021.
[45] D. Rousidis, P. Koukaras, and C. Tjortjis, "Social media
prediction: a literature review," Multimedia Tools and
Applications, vol. 79, no. 9-10, pp. 6279-6311, 2020.
[46] M. S. Amin et al., "Harmonizing Macro-Financial Factors and
Twitter Sentiment Analysis in Forecasting Stock Market
Trends," Journal of Computer Science and Technology Studies,
vol. 6, no. 1, pp. 58-67, 2024.
[47] S. Verma, S. P. Sahu, and T. P. Sahu, "Stock Market Forecasting
Using Additive Ratio Assessment-Based Ensemble Learning,"
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
in International Conference On Innovative Computing And
Communication, 2023, pp. 325-335: Springer.
[48] B. A. Abdelfattah, S. M. Darwish, and S. M. Elkaffas,
"Enhancing the Prediction of Stock Market Movement Using
Neutrosophic-Logic-Based Sentiment Analysis," Journal of
Theoretical and Applied Electronic Commerce Research, vol.
19, no. 1, pp. 116-134, 2024.
[49] S. BL and S. BR, "Combined deep learning classifiers for stock
market prediction: integrating stock price and news sentiments,"
Kybernetes, vol. 52, no. 3, pp. 748-773, 2023.
[50] S. Wu, Y. Liu, Z. Zou, and T.-H. Weng, "S_I_LSTM: stock price
prediction based on multiple data sources and sentiment
analysis," Connection Science, vol. 34, no. 1, pp. 44-62, 2022.
[51] Z. Ye, Y. Wu, H. Chen, Y. Pan, and Q. Jiang, "A stacking
ensemble deep learning model for bitcoin price prediction using
Twitter comments on bitcoin," Mathematics, vol. 10, no. 8, p.
1307, 2022.
[52] I. Gupta, T. K. Madan, S. Singh, and A. K. Singh, "HiSA-
SMFM: historical and sentiment analysis based stock market
forecasting model," arXiv preprint arXiv:2203.08143, 2022.
[53] P. Mehta, S. Pandya, and K. Kotecha, "Harvesting social media
sentiment analysis to enhance stock market prediction using
deep learning," PeerJ Computer Science, vol. 7, p. e476, 2021.
[54] P. Yu and X. Yan, "Stock price prediction based on deep neural
networks," Neural Computing and Applications, vol. 32, pp.
1609-1628, 2020.
[55] W. Meng, Q. Zheng, Y. Shi, and G. Pan, "An off-policy trust
region policy optimization method with monotonic improvement
guarantee for deep reinforcement learning," IEEE Transactions
on Neural Networks and Learning Systems, vol. 33, no. 5, pp.
2223-2235, 2021.
[56] Z. D. Akşehir and E. Kılıç, "A new denoising approach based on
mode decomposition applied to the stock market time series:
2LE-CEEMDAN," PeerJ Computer Science, vol. 10, p. e1852,
2024.
[57] S. Hochreiter and J. Schmidhuber, "Long short-term memory,"
Neural computation, vol. 9, no. 8, pp. 1735-1780, 1997.
[58] A. Graves, "Generating sequences with recurrent neural
networks," arXiv preprint arXiv:1308.0850, 2013.
[59] M. S. Akhtar, T. Garg, and A. Ekbal, "Multi-task learning for
aspect term extraction and aspect sentiment classification,"
Neurocomputing, vol. 398, pp. 247-256, 2020.
[60] X. Xin, A. Wumaier, Z. Kadeer, and J. He, "SSEMGAT:
Syntactic and Semantic Enhanced Multi-Layer Graph Attention
Network for Aspect-Level Sentiment Analysis," Applied
Sciences, vol. 13, no. 8, p. 5085, 2023.
[61] X. Zhu, Z. Kuang, and L. Zhang, "A prompt model with
combined semantic refinement for aspect sentiment analysis,"
Information Processing & Management, vol. 60, no. 5, p.
103462, 2023.
[62] L. Mingzheng, H. Zelin, L. Jiadong, and L. Wei, "Aspect-level
sentiment analysis model fused with GPT and multi-layer
attention," in Third International Conference on Artificial
Intelligence and Computer Engineering (ICAICE 2022), 2023,
vol. 12610, pp. 279-284: SPIE.
[63] Y. He, X. Huang, S. Zou, and C. Zhang, "PSAN: Prompt
Semantic Augmented Network for aspect-based sentiment
analysis," Expert Systems with Applications, vol. 238, p. 121632,
2024.
[64] Ö. Özdemir and E. B. Sönmez, "Weighted cross-entropy for
unbalanced data with application on covid x-ray images," in
2020 Innovations in Intelligent Systems and Applications
Conference (ASYU), 2020, pp. 1-6: IEEE.
[65] F. Huang, J. Li, and X. Zhu, "Balanced Symmetric Cross
Entropy for Large Scale Imbalanced and Noisy Data," arXiv
preprint arXiv:2007.01618, 2020.
[66] X. Li, X. Sun, Y. Meng, J. Liang, F. Wu, and J. Li, "Dice loss
for data-imbalanced NLP tasks," arXiv preprint
arXiv:1911.02855, 2019.
[67] S. S. M. Salehi, D. Erdogmus, and A. Gholipour, "Tversky loss
function for image segmentation using 3D fully convolutional
deep networks," in Machine Learning in Medical Imaging: 8th
International Workshop, MLMI 2017, Held in Conjunction with
MICCAI 2017, Quebec City, QC, Canada, September 10, 2017,
Proceedings 8, 2017, pp. 379-387: Springer.
[68] S. A. Taghanaki et al., "Combo loss: Handling input and output
imbalance in multi-organ segmentation," Computerized Medical
Imaging and Graphics, vol. 75, pp. 24-33, 2019.
Ali Peivandizadeh is a
distinguished graduate of the
University of Houston,
specializing in machine learning
and pioneering advancements in
predictive analytics. His expertise
has notably impacted stock
market prediction and social
media analysis, enhancing data-driven decision-
making processes.
Sima Hatami is an esteemed
scholar at Islamic Azad
University, Qazvin, Iran, known
for her extensive expertise in
Business Management. She is
deeply involved in cutting-edge
research, particularly in the areas
of Machine Learning, Stock
Market Prediction, and Social
Media Analysis.
Amirhossein Nakhjavani is a
dedicated medical student at
Mashhad University of Medical
Sciences, who merges his medical
training with a fervent interest in
IT. His research, enriched by his
passion for machine learning, is
focused on groundbreaking
applications in stock market prediction and social
media analysis.
Lida Khoshsima is currently
R&D Specialist and Human
Resource Specialist with 10 years
of experience in human resources
and HR software development.
Honored with the title of
exemplary employee, Shiva Co.,
Tehran, Iran, 2018.she got M.Sc.
degree in Economics, Raja University, Qazvin, Iran.
Her research interests include Business cycle analysis,
Machine Learning, Macroeconomic forecasting, and
Stock market.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8 VOLUME XX, 2017
Mohammad Reza Chalak
Qazani received the Bachelor of
Engineering in manufacturing and
production from University of
Tabriz, Tabriz, Iran in 2010 and
the master's degree in robotic and
mechanical engineering from the
Tarbiat Modares University,
Tehran, Iran, in 2013. He received the Ph.D. degree in
modelling and simulation of a motion cueing algorithm
using prediction and computational intelligence
techniques from the Institute for Intelligent Systems
Research and Innovation (IISRI), Deakin University,
Australia, in 2021. He was an Alfred Deakin
Postdoctoral Research Fellow in the IISRI for two
years working in the areas of model predictive control,
motion cueing algorithm and soft computing
controllers. He is currently an Assistant Professor in the
Faculty of Computing and Information Technology
(FoCIT), Sohar University, Sohar, Sohar 311, Oman.
His teaching areas are data structure and algorithm,
enterprise resource planning modelling and
implementation, modelling and visualization,
computer architecture, introduction to artificial
intelligence, and advance machine learning.
Muhammad Haleem is currently
an Assistant Professor with the
Department of Computer Science,
Faculty of Engineering, Kardan
University, Kabul, Afghanistan.
He holds an MSCS degree from
the Department of Computer
Science, Abdul Wali Khan
University, Mardan, Pakistan. His research interests
include the Internet of Things, machine learning, cloud
computing, particle swarm optimization, power
systems, and data analytics.
Roohallah Alizadehsani obtained
a Bachelor of Science degree in
Computer Engineering-Software
from Sharif University of
Technology and then received a
Master of Science in Computer
Engineering-Software from the
same institution. Roohallah's
research interests include mainly in the areas of Data
Mining, Machine Learning, Bioinformatics, Heart
Disease, Skin Disease, Diabetes Disease, Hepatitis
Disease, and Cancer Disease. He is currently a
Research Fellow at Deakin University of Australia.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3399548
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
