Thanks to visit codestin.com
Credit goes to www.scribd.com

0% found this document useful (0 votes)
99 views27 pages

Deep Learning For Algorithmic Trading

This systematic review examines the integration of deep learning (DL) in algorithmic trading, focusing on predictive models and optimization strategies. It analyzes various DL architectures like RNN, LSTM, and CNN, evaluates their performance in predicting stock prices and market trends, and discusses challenges such as data noise and overfitting. The findings highlight the transformative potential of DL in enhancing decision-making processes within volatile financial markets and outline future research directions.

Uploaded by

gitology
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
99 views27 pages

Deep Learning For Algorithmic Trading

This systematic review examines the integration of deep learning (DL) in algorithmic trading, focusing on predictive models and optimization strategies. It analyzes various DL architectures like RNN, LSTM, and CNN, evaluates their performance in predicting stock prices and market trends, and discusses challenges such as data noise and overfitting. The findings highlight the transformative potential of DL in enhancing decision-making processes within volatile financial markets and outline future research directions.

Uploaded by

gitology
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 27

Array 26 (2025) 100390

Contents lists available at ScienceDirect

Array
journal homepage: www.elsevier.com/locate/array

Deep learning for algorithmic trading: A systematic review of predictive


models and optimization strategies
MD Shahriar Mahmud Bhuiyan a , MD AL Rafi a , Gourab Nicholas Rodrigues a ,
MD Nazmul Hossain Mir a , Adit Ishraq b ,∗, M.F. Mridha c ,∗, Jungpil Shin d
a Washington University of Science and Technology, 2900 Eisenhower ave, Alexandria, VA, 22314, USA
b Stockholm University, Stockholm, Frescativägen, 114 19, Sweden
c
Department of Computer Science, American International University-Bangladesh, Dhaka, 1229, Bangladesh
d
School of Computer Science and Engineering, The University of Aizu, Aizu-wakamatsu, 965-8580, Japan

ARTICLE INFO ABSTRACT

Keywords: Algorithmic trading has revolutionized financial markets, offering rapid and efficient trade execution. The
Algorithmic trading integration of deep learning (DL) into these systems has further enhanced predictive capabilities, providing
Financial market prediction sophisticated models that capture complex, non-linear market patterns. This systematic literature review
Stock price forecasting
explores recent advancements in the application of DL algorithms to algorithmic trading with a focus on
Predictive modeling
optimizing financial market predictions. We analyze and synthesize the key DL architectures, such as recurrent
Deep learning
neural networks (RNN), long short-term memory (LSTM), convolutional neural networks (CNN), and hybrid
models, to evaluate their performance in predicting stock prices, volatility, and market trends. The review
highlights current challenges, such as data noise, overfitting, and interpretability, while discussing emerging
solutions and future research directions. Our findings provide a comprehensive understanding of how DL
reshapes algorithmic trading and its potential to improve decision-making processes in volatile financial
environments.

1. Introduction analyzing chaotic and highly volatile financial markets [11,12]. DL


architectures like Recurrent Neural Networks (RNNs) [13], Long Short-
Algorithmic trading has transformed the financial markets by au- Term Memory (LSTM) networks [14], Convolutional Neural Networks
tomating the process of executing trades, relying on pre-programmed (CNNs) [15], and hybrid models have shown promise in predicting
instructions and sophisticated mathematical models to achieve speed stock prices, detecting market trends, and managing portfolio risks [16,
and efficiency unattainable by human traders [1,2]. This shift from 17].
manual to algorithmic trading has allowed market participants to cap- Financial market predictions have traditionally been challenging
italize on minute price discrepancies and market inefficiencies with re- due to market volatility, noise, and unpredictability [18,19]. How-
markable precision, especially in high-frequency trading (HFT), where ever, introducing DL algorithms offers an unprecedented opportunity
milliseconds can define success or failure [3,4]. The financial sector’s to improve predictive accuracy by modeling dynamic relationships
increasing reliance on computational power to process vast volumes of that elude traditional statistical methods [20,21]. While conventional
market data in real-time has laid the groundwork for advanced tech- algorithmic trading strategies have been driven primarily by technical
niques like machine learning (ML) and deep learning (DL) to optimize indicators, DL enables incorporating a broader array of inputs, such as
trading strategies and market predictions [5–7]. historical market data, macroeconomic factors, and textual data from
news and social media [22]. This shift presents an opportunity for
Deep learning, a subset of machine learning, has gained significant
algorithmic trading systems to execute trades and actively adapt and
attention in recent years due to its ability to handle vast amounts of
optimize strategies based on predictive insights drawn from complex
unstructured data, such as price time series, social media sentiment,
data sets.
and economic indicators [8–10]. Unlike traditional machine learning
Despite the considerable promise of deep learning in algorithmic
algorithms, deep learning models, particularly neural networks, excel
trading, the literature remains fragmented, with many studies focusing
at identifying complex, non-linear patterns, making them ideal for

∗ Corresponding authors.
E-mail addresses: [email protected] (M.S.M. Bhuiyan), [email protected] (M.A. Rafi), [email protected] (G.N. Rodrigues),
[email protected] (M.N.H. Mir), [email protected] (A. Ishraq), [email protected] (M.F. Mridha), [email protected] (J. Shin).

https://doi.org/10.1016/j.array.2025.100390
Received 10 November 2024; Received in revised form 12 March 2025; Accepted 22 March 2025
Available online 4 April 2025
2590-0056/© 2025 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

on isolated aspects of DL without providing a comprehensive view 2. Existing reviews


of its applications across various trading strategies [23,24]. Previous
reviews have often concentrated either on broader machine learning Numerous reviews have examined the role of ML and DL in financial
applications in finance or on specific DL techniques applied to narrow applications, especially in the domain of algorithmic trading [30–32].
problem sets, such as stock price prediction [25]. As a result, there is a These reviews typically focus on isolated components, such as specific
notable gap in the systematic synthesis of research that encompasses algorithms, data types, or market scenarios [33]. However, few offer a
both the range of DL models used and their practical efficacy in systematic synthesis across diverse DL models and their application to
optimizing algorithmic trading. various financial markets. Table 1 shows several key reviews that have
The financial markets are characterized by increasing complexity, contributed to the understanding of DL in trading while highlighting
driven by globalization, technological advancements, and the prolif- their limitations.
eration of high-frequency trading [26,27]. Traditional trading models
For instance, Biju et al. [34] quantitatively analyze ML, AI, and
struggle to cope with the volume and velocity of data generated, while
DL literature through bibliometric methods, identifying trends and
financial regulations and market structure changes require ever more
research themes such as ESG scoring and geographical insights. They
adaptive and intelligent trading systems [28,29]. The demand for more
also highlight algorithmic bias concerns, although they note the lack
robust predictive models necessitates a review that explores how DL
of empirical research on advanced automated financial technologies
techniques are being leveraged to optimize these systems, overcoming
the limitations of earlier approaches. and issues with traceability and replicability in deep learning systems.
This systematic literature review aims to fill this gap by providing a Similarly, Sonkavde et al. [6] review recent ML and DL techniques for
structured and comprehensive analysis of deep learning’s applications stock price prediction, proposing a generic framework for prediction
in algorithmic trading. This paper seeks to identify trends, challenges, and classification. While their work focuses on ensemble models, they
and emerging opportunities by critically evaluating existing research. acknowledge limitations such as insufficient exploration of ensemble
Given the growing influence of artificial intelligence (AI) in finance, methods and the missed integration of sentiment analysis, which could
this review is crucial for guiding future research and development enhance model predictive power.
in algorithmic trading systems that are more predictive, adaptive, Salehpour et al. [32] provide a holistic view of cutting-edge ma-
and ultimately more profitable. This paper makes the following key chine learning applications in algorithmic trading, emphasizing real-
contributions to the intersection of deep learning (DL) and algorithmic world performance and the risks of overfitting and interpretability.
trading: They stress that their review may not encompass all emerging tech-
niques in this rapidly evolving field. Ayitey et al. [35] focus on machine
• The paper examines how deep learning algorithms are currently learning algorithms applied to forex market forecasting from 2010 to
being applied to various financial markets. It explores a broad 2021. Their meta-analysis evaluates model performance but is limited
spectrum of DL architectures, including RNN, LSTM, CNN, and
by the quality of included studies and a lack of standardization in model
hybrid models, used across different financial instruments such as
evaluation metrics.
stocks, commodities, and foreign exchange. The review identifies
Joiner et al. [38] offer insights into financial asset price forecasting
how these models are employed for diverse prediction tasks,
models, highlighting research gaps in data warehousing integration and
including price trends and market volatility forecasting.
algorithmic trading’s impact on market liquidity. However, they face
• This study critically evaluates the challenges and limitations of
limitations in comparability due to inconsistent reporting metrics and
using deep learning in optimizing financial market predictions.
the reliance on specific datasets. Cohen et al. [37] integrate human
It highlights key issues such as the difficulty of managing noisy
and incomplete data, the risk of overfitting in complex models, behavior and social media sentiment with traditional data for algo-
and the lack of interpretability in DL models due to their black- rithmic trading, enhancing predictions of price movements. However,
box nature. Additionally, the review explores the computational they emphasize the risks associated with algorithmic misinterpretation
intensity required for large-scale financial datasets, providing during extreme price movements and the challenges of data quality.
insights into practical and theoretical barriers to adoption. Lastly, Singh et al. [31] review learning methods such as DL, RL, and
• The paper outlines emerging opportunities in the use of DL for DRL in finance, presenting a framework for applying these techniques
algorithmic trading. It identifies promising areas for further in- to financial problems. They identify limitations such as their narrow
vestigation, such as the development of new DL architectures, exploration of single-agent settings and the neglect of data quality
the integration of hybrid models that combine deep learning with issues.
reinforcement learning or traditional methods, and the use of These existing reviews provide a foundation but highlight the frag-
advanced data integration techniques like attention mechanisms mented nature of research in the application of deep learning to al-
and transformers. These opportunities point to ways in which gorithmic trading. Our systematic review seeks to fill these gaps by
DL can enhance predictive power and operational efficiency in providing a comprehensive overview of the diverse DL models, their
financial markets. applications, challenges, and future research directions.

By synthesizing the current state of research, this paper provides


3. Review methodology
a consolidated understanding of deep learning’s role in algorithmic
trading and lays the groundwork for future advancements. These contri-
This section outlines the methodology used to conduct a systematic
butions are intended to drive both academic inquiry and practical appli-
review of DL applications in algorithmic trading. By following a struc-
cations in financial technology, offering novel insights into optimizing
tured and rigorous approach, we aim to ensure that the review is both
trading strategies through advanced predictive models.
The paper is structured as follows: Section 2 provides an overview comprehensive and replicable, allowing future researchers to build on
of existing reviews in the field, summarizing key studies and method- our findings.
ologies. Section 3 outlines the review methodology used to collect and
analyze relevant literature. Section 4 presents the theoretical back- 3.1. Systematic review protocol
ground, covering foundational concepts and models. Section 5 discusses
the results and findings from the reviewed studies. Section 6 identifies This study follows a systematic literature review (SLR) methodol-
challenges in the field, such as data quality and model complexity. ogy, guided by the Preferred Reporting Items for Systematic Reviews
Section 7 offers a discussion of these challenges and suggests future re- and Meta-Analyses (PRISMA) protocol [39,40]. The PRISMA approach
search directions. Finally, Section 8 concludes the paper, summarizing is designed to ensure transparency, rigor, and replicability in conduct-
key insights and recommendations. ing systematic reviews [41]. By adhering to this protocol, we structured

2
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Table 1
Summary of key reviews on DL in algorithmic trading, their contributions, and limitations.
Ref. Year Systematic Dataset Technological Results Challenges & Contribution Limitations
review description advancements analysis future work
[36] 2025 ✓ × × × × This paper integrates AI and ML This paper highlights the lack of
techniques to enhance stock market standardized data frameworks,
prediction accuracy, addresses limited real-time adaptability, the
research gaps in emerging data high computational complexity of
sources, connects predictive models reinforcement learning techniques,
to economic sustainability, and and the potential oversight of
employs a systematic review indirect competitors, which may
methodology for comprehensive impact the accuracy and practicality
analysis. of predictive models.
[34] 2024 ✓ × ✓ × × This paper contributes to finance The paper lacks empirical research
research by quantitatively analyzing on advanced automated financial
ML, AI, and DL literature through technologies, highlights algorithmic
bibliometric methods, identifying biases in critical areas, and discusses
trends, research themes like ESG issues with traceability and
scoring, and geographical insights. It replicability in deep learning
also highlights algorithmic bias systems. Additionally, it identifies an
concerns and advocates for future absence of governance and ambiguity
research directions addressing in the effectiveness of AI/ML over
disruptive innovations in finance. traditional analytical tools.
[6] 2023 ✓ × ✓ × ✓ This paper reviews recent machine The paper’s limitations include
learning and deep learning insufficient exploration of ensemble
techniques for stock price prediction, methods, lack of detailed
implements various models, and hyperparameter tuning strategies,
conducts a comparative analysis with inadequate adaptation framework for
a focus on ensemble models. It also the evolving stock market, focus on
proposes a generic framework for specific stocks limiting
prediction and classification, generalizability, and missed
discussing the implications and integration of sentiment analysis,
limitations of current methodologies, which could enhance model
and guiding future research. predictive power.
[32] 2023 ✓ × ✓ × × This paper reviews cutting-edge The paper highlights several
machine learning applications in limitations, including the risks of
algorithmic trading, validating overfitting, challenges with
previous advancements, evaluating interpretability, reliance on large
real-world performance, and datasets, instability in reinforcement
identifying limitations for future learning, and increased complexity in
research. It offers a holistic view of hybrid methods. Additionally, the
the field, providing insights for review may not encompass all
practitioners and policymakers on emerging techniques in the
responsible innovation at the fast-evolving field of machine
intersection of technology and learning in trading.
finance.
[35] 2023 ✓ × ✓ × ✓ This paper provides a comprehensive The meta-analysis is limited by the
review of machine learning quality of included studies, with poor
algorithms applied to forex market designs and biases affecting overall
forecasting (2010–2021). It evaluates findings. A lack of standardization in
the design of evaluation techniques, model evaluation metrics complicates
conducts a meta-analysis of model comparisons. The paper focuses
performance using metrics like MAE solely on machine learning methods,
and RMSE, identifies trends (LSTM, potentially overlooking other
ANN), and highlights future research forecasting techniques. Dataset
directions to improve prediction division methodology is also
accuracy. questioned.
[37] 2022 × × ✓ × × The paper integrates human behavior The paper highlights several
and social media sentiment with limitations, including the risk of
traditional data for algorithmic algorithmic misinterpretation during
trading, enhancing predictions of extreme price movements,
price movements. It combines dependence on data quality,
technical, fundamental, and challenges in forecasting due to
sentiment analysis, offers practical market complexity, poor algorithm
guidance for traders, and reviews performance under ‘‘Black Swan’’
advanced techniques for improved conditions, and the significant
trading outcomes, emphasizing transaction costs associated with
adaptability and pattern recognition. high-frequency trading systems.

(continued on next page)

3
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Table 1 (continued).
Ref. Year Systematic Dataset Technological Results Challenges & Contribution Limitations
review description advancements analysis future work
Ours – ✓ ✓ ✓ ✓ ✓ This review comprehensively This review has several limitations,
evaluates the application of deep including potential publication bias
learning in algorithmic trading, from underreported negative results
highlighting advancements, results and a focus on English-language
analysis, and future directions. It also studies, which may exclude relevant
identifies critical research gaps and research. Additionally, varying study
discusses the implications of deep quality affects transparency in
learning technologies in finance. methodologies, and the review is
limited to the literature available up
to 2018–2024.

Table 2
Key research questions (RQs) for deep leaning based algorithmic trading review.
Research question Description
RQ1: How are deep learning algorithms currently being Sections 4 & 5
applied in algorithmic trading across different financial
markets? This question examines the types of DL models,
their application in various financial instruments (e.g.,
stocks, commodities, foreign exchange), and the nature of
the predictions they produce (e.g., price trends, volatility).
RQ2: What are the key challenges and limitations associated Section 6
with using DL in optimizing financial market predictions?
This includes both practical challenges such as data quality,
computational demands, and overfitting, as well as
theoretical limitations, such as the interpretability of models
and their generalizability across markets.
RQ3: What are the future opportunities for improving Section 7
algorithmic trading performance using DL techniques? This
question explores emerging trends in DL research, including
new architectures, hybrid models, and the integration of
advanced data sources that may offer better predictive power
and efficiency in financial markets.

Table 3
Search strategy for systematic review.
Component Description
Fig. 1. PRISMA flow diagram detailing the stages of the systematic literature review • IEEE Xplore
process, including study identification, screening, eligibility assessment, and final Databases • SpringerLink
inclusion. Searched • Google Scholar
• arXiv
Search Keywords • ‘‘Deep Learning’’
our review process into several key stages: defining research questions, • ‘‘Algorithmic Trading’’
developing a search strategy, selecting relevant studies, assessing their • ‘‘Financial Market Prediction’’
• ‘‘Stock Price Forecasting’’
quality, and synthesizing the findings. The overall process is visually
• ‘‘Reinforcement Learning’’
represented in Fig. 1, which outlines the key steps taken during the • ‘‘Neural Networks in Trading’’
systematic review, ensuring clarity and reproducibility. • ‘‘Market Volatility Prediction’’
• ‘‘Convolutional Neural Networks in Finance’’
• ‘‘Recurrent Neural Networks in Trading’’
3.2. Research questions (RQs)
• ‘‘High-Frequency Trading using AI’’
• ‘‘Long Short-Term Memory in Financial Markets’’
The review was structured to address the following key research • ‘‘AI-driven Algorithmic Trading Strategies’’
questions (RQs) presented in Table 2: • ‘‘Financial Time-Series Forecasting’’
• ‘‘Sentiment Analysis for Trading’’
• ‘‘Machine Learning in Portfolio Optimization’’
3.3. Search strategy
Search Combination of keywords and Boolean operators such as
Techniques AND, OR, and NOT to refine search results and identify
To ensure comprehensive coverage of relevant literature, we devel- relevant studies.
oped a structured search strategy, outlined in Table 3.
Timeframe for 2020–2024 (to capture the most recent advancements in DL
Publication applications in algorithmic trading).
3.4. Study selection criteria
Manual Search Manual searches in reference lists of key papers to ensure no
relevant studies were missed.
The inclusion and exclusion criteria for study selection were defined
as presented in Table 4.

3.5. Quality assessment • Relevance: The study’s relevance to the research questions and
its contribution to understanding DL applications in algorithmic
To assess the methodological rigor of the selected studies, we ap- trading.
plied a quality assessment framework that evaluated the following • Rigor: The robustness of the study’s experimental design, includ-
criteria: ing the evaluation metrics and comparison to baseline models.

4
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Table 4 4. Theoretical background


Inclusion and exclusion criteria for study selection.
Criteria Inclusion Exclusion Understanding the theoretical underpinnings of algorithmic trad-
Studies applying deep learning Studies focusing solely on ing and the application of deep learning is essential for grasping
Focus models to algorithmic trading traditional machine learning their roles in modern financial markets [42,43]. This section outlines
or financial market prediction. or statistical models without
incorporating deep learning.
the foundational concepts and methodologies employed in algorithmic
trading, followed by a discussion on how deep learning techniques have
Type of Empirical studies, case studies, Studies that do not provide
Studies or review papers reporting sufficient details on methods,
revolutionized financial predictions and trading strategies.
performance metrics. model configurations, or
performance metrics. 4.1. Algorithmic trading
Publication Studies published between Studies published before 2010.
Year 2010 and 2023. Algorithmic trading involves using computer algorithms to auto-
Source Peer-reviewed journals or Non-peer-reviewed articles, matically execute the buying and selling of financial assets according
high-quality conference opinion pieces, or grey to specific, pre-established criteria [44,45]. These algorithms can ex-
proceedings. literature (e.g., blog posts, ecute trades at speeds and frequencies far beyond the capability of
non-academic reports).
human traders. Algorithmic trading strategies are primarily designed to
Language Articles published in English. Articles not written in English. optimize execution efficiency, minimize trading costs, and exploit mar-
Dataset Studies that describe the Studies that do not disclose ket inefficiencies [46,47]. The key strategies employed in algorithmic
Transparency datasets used, including the datasets or fail to provide trading include:
source, size, and features. sufficient dataset details.
Model Repro- Studies providing detailed Studies lacking reproducibility, • High-Frequency Trading (HFT): This involves executing a large
ducibility model parameters, code, or including insufficient details number of orders at extremely high speeds, usually in fractions
steps for replication. for replicating the deep
of a second. The goal is to exploit minute price differentials
learning model.
across markets or within the same asset by rapidly buying and
Financial Studies focusing on diverse Studies limited to very niche
selling large volumes. HFT firms often employ statistical arbitrage
Instruments financial instruments, or non-mainstream financial
including stocks, commodities, instruments without broader and market-making strategies to capture profits from tiny market
and foreign exchange. applicability. inefficiencies.
Comparison Studies that compare deep Studies that do not provide • Arbitrage: Arbitrage trading strategies seek to profit from price
to Baselines learning models to baseline comparative analysis or discrepancies between related financial instruments, such as
models (e.g., traditional ML performance benchmarking. stocks, bonds, commodities, or derivatives. These strategies in-
models or statistical clude statistical arbitrage and index arbitrage, where algorithms
approaches).
identify and exploit mispricings by simultaneously buying and
selling the assets involved.
• Trend-Following: These algorithms detect patterns in the move-
• Replicability: The transparency of the study’s methodology, in- ment of asset prices, such as upward or downward trends, and
cluding the availability of datasets, model parameters, and repro- execute trades accordingly. Trend-following strategies, such as
ducible code. momentum trading, typically rely on technical indicators like
• Publication Quality: Preference was given to studies published moving averages, relative strength indices (RSI), and Bollinger
in reputable journals and conferences with a high citation index. Bands.
• Mean Reversion: Mean reversion strategies are based on the
Each study was scored on a scale of 1 (low) to 5 (high) for each idea that asset prices tend to return to their historical average
criterion, and only studies with an overall score of 3 or higher were over time. Algorithms using this approach detect when an asset
included in the final analysis. is either overvalued or undervalued and place trades, anticipating
that the price will move back toward the average.
3.6. Data extraction and synthesis • Market Making: This approach entails consistently offering buy
and sell prices for financial assets. Market makers earn profits
from the bid–ask spread and provide liquidity to markets. Al-
For each selected study, the following data were extracted:
gorithms automate this process, ensuring orders are executed
efficiently and at competitive prices.
• Year of publication: To track the temporal distribution of studies
• Pairs Trading: This is a type of statistical arbitrage where two
and trends over time.
correlated assets are traded simultaneously. The strategy assumes
• DL models used: The specific deep learning architectures em-
that when the price of one asset diverges significantly from its
ployed (e.g., CNN, RNN, LSTM, hybrid models).
paired asset, the prices will eventually converge. An algorithm
• Financial markets examined: The type of financial instruments
buys the underperforming asset and sells the outperforming one,
studied (e.g., stocks, commodities, foreign exchange) and their
profiting from their convergence.
respective markets.
• Evaluation metrics: Performance measures such as accuracy, Table 5 compares key algorithmic trading strategies based on their
precision, recall, Sharpe ratio, return on investment (ROI), etc. description, key features, and use cases.
• Challenges and limitations: Key challenges identified in ap- Fig. 2 provides a comprehensive overview of AI techniques used in
plying DL to algorithmic trading (e.g., data quality, overfitting, algorithmic trading. The diagram categorizes these techniques into four
model interpretability). main groups: Statistical Time-Series Methods, Machine Learning, Deep
Learning, and Reinforcement Learning (RL). In this literature, we focus
The results were synthesized using a narrative synthesis approach, exclusively on deep learning models, providing a detailed description
categorizing studies based on DL model types, prediction tasks, and of each model shown in Fig. 2 in the following Section 4.2. Addition-
performance outcomes. Where appropriate, comparisons were made ally, we review previous studies that have applied these models in
between the effectiveness of different models and their applications algorithmic trading, highlighting their effectiveness and applications in
across various financial markets. financial markets.

5
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Fig. 2. Overview of different AI models for Algorithmic Trading.

Table 5
Comparison of algorithmic trading strategies.
Trading Description Key features and use cases
strategy
High- Executes a large number of Utilized by institutional
Frequency orders at extremely high traders; exploits inefficiencies
Trading speeds, often in milliseconds, in highly liquid markets
(HFT) to exploit small price through rapid trading.
differentials.
Arbitrage Exploits price discrepancies Includes statistical arbitrage
between related financial and index arbitrage;
instruments or markets by commonly applied across
simultaneously buying and different asset classes (e.g., Fig. 3. Architecture of RNN.
selling to lock in profits. stocks, bonds, commodities).
Trend- Detects price trends (upward Popular in momentum trading;
Following or downward) and executes uses indicators such as moving
trades in the direction of the averages, RSI, and Bollinger 4.2.1. Recurrent Neural Networks (RNNs)
trend. Relies on technical Bands; suitable for medium- to RNNs are particularly adept at time-series forecasting in financial
indicators to confirm trends. long-term strategies. markets due to their inherent ability to capture temporal dependen-
Mean Assumes that prices will revert Based on the statistical cies [50,51]. An RNN maintains a hidden state ℎ𝑡 that updates with
Reversion to their historical mean over concept of reversion to the
each new time step 𝑡, based on the previous hidden state ℎ𝑡−1 and the
time. Trades are made when mean; often used in markets
assets are deemed overbought with cyclical behavior. current input 𝑥𝑡 :
or oversold.
ℎ𝑡 = 𝜎(𝑊ℎ ℎ𝑡−1 + 𝑊𝑥 𝑥𝑡 + 𝑏ℎ ) (1)
Market It refers to the ongoing Provides liquidity to the
Making process of offering both buy market; algorithms automate In this equation, 𝑊ℎ and 𝑊𝑥 represent the weight matrices, 𝑏ℎ is a
and sell prices for a financial order execution to reduce
bias term, and 𝜎 is a non-linear activation function, typically chosen
asset, generating profit from latency and ensure
the difference between the bid competitive pricing. as tanh or ReLU. While RNNs are powerful, they often struggle with
and ask prices. long-term dependencies due to the vanishing gradient problem.
Pairs Trading A statistical arbitrage strategy Buys the underperforming Fig. 3 illustrates the structure of an RNN, where the hidden state ℎ𝑡
where two correlated assets asset and sells the captures information from previous time steps, allowing the model to
are traded simultaneously. The outperforming asset; works make predictions based on historical data. RNNs have been successfully
idea is to bet on their prices well with assets that have a
applied in algorithmic trading for predicting stock price movements
converging. historical correlation.
based on historical prices and trading volumes, as well as forecasting
currency exchange rates by analyzing historical currency data [52,53].
A recent contribution by Lu et al. [52] focuses on the efficiency of
4.2. Deep learning in financial predictions prediction results in big data analysis, emphasizing the importance of
data preprocessing methods. The proposed Time-series RNN (TRNN) for
DL has emerged as a pivotal technology in financial markets, en- stock price prediction incorporates trading volume and employs sliding
abling sophisticated modeling of non-linear relationships and effective windows to process time series data. By expanding the price–volume
processing of vast amounts of data [20,48,49]. In algorithmic trading, relationship from one dimension to two, the TRNN model strengthens
various DL models have been developed to predict market trends, the influence of recent trading volumes on current stock prices. The
optimize trading strategies, and enhance decision-making processes. study contrasts the TRNN with conventional RNN and LSTM models,

6
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Fig. 4. Architecture of LSTM.

Fig. 5. Architecture of CNN.

highlighting its increased efficiency and accuracy, while also exploring


its potential applications in other areas.
Zhao et al. [54] presents a stock price trend prediction model Xu et al. [60] presented a hybrid LSTM-GARCH framework that
utilizing RNN, LSTM, and GRU, enhanced by an attention mechanism. integrates deep learning with classical econometric models to improve
The results indicate that the GRU-M and LSTM-M models outperform volatility risk prediction. The hybrid model achieved a lower MSE of
the RNN-M model, with the GRU-M model achieving slightly better 0.604 compared to standalone models, highlighting its utility in finan-
performance. The study concludes that two layers are optimal for cial risk management. However, its performance was sensitive to pa-
performance, as deeper networks may slow the learning rate and un- rameter selection and input data quality, potentially affecting broader
derperform with limited training data. Aldhyani et al. [55] introduce applicability. Lastly, Zhang et al. [53] investigated the influence of
a hybrid CNN-LSTM framework for predicting stock market prices, investor sentiment on stock market volatility using an LSTM model,
achieving high R-squared values of 98.37% for Tesla and 99.48% for finding a significant negative correlation, particularly during crises. The
Apple. However, the study notes the absence of sentiment analysis study demonstrated that sentiment indices outperform historical data
from financial investigations, which could limit prediction robustness. models for short-term volatility predictions but noted limitations due to
The findings also suggest that current AI stock prediction models have the small sample size and dictionary-based sentiment analysis methods.
not yet achieved full accuracy, indicating opportunities for further
enhancement. 4.2.3. Convolutional Neural Networks (CNNs)
CNNs have demonstrated significant potential in financial market
4.2.2. Long Short-Term Memory Networks (LSTMs) predictions by treating time-series data as one-dimensional signals [61,
LSTMs are a specialized type of RNN designed to overcome the 62]. This application enables CNNs to automatically identify significant
vanishing gradient problem [14,56]. LSTMs utilize memory cells that features from historical price data using their convolutional layers. By
regulate the flow of information using three gates: the input gate, the uncovering patterns and trends that are often overlooked in conven-
forget gate, and the output gate [14]. The cell state 𝐶𝑡 and hidden state tional financial analysis, CNNs improve the precision of predictions. Ad-
ℎ𝑡 are updated as follows: ditionally, their ability to efficiently process large datasets makes them
𝑓𝑡 = 𝜎(𝑊𝑓 [ℎ𝑡−1 , 𝑥𝑡 ] + 𝑏𝑓 ) well-suited for high-frequency trading environments, where timely and
precise forecasts are essential for successful trading strategies [63,64].
𝑖𝑡 = 𝜎(𝑊𝑖 [ℎ𝑡−1 , 𝑥𝑡 ] + 𝑏𝑖 )
The convolution operation for a one-dimensional input 𝑥𝑡 with a filter
𝐶̃𝑡 = tanh(𝑊𝐶 [ℎ𝑡−1 , 𝑥𝑡 ] + 𝑏𝐶 ) 𝑤 can be defined as:
(2)
𝐶𝑡 = 𝑓𝑡 ⋅ 𝐶𝑡−1 + 𝑖𝑡 ⋅ 𝐶̃𝑡 ∑
𝑘−1

𝑜𝑡 = 𝜎(𝑊𝑜 [ℎ𝑡−1 , 𝑥𝑡 ] + 𝑏𝑜 ) 𝑦𝑡 = (𝑥 ∗ 𝑤)(𝑡) = 𝑥𝑡+𝑖 𝑤𝑖 (3)


𝑖=0
ℎ𝑡 = 𝑜𝑡 ⋅ tanh(𝐶𝑡 )
In this formulation, 𝑓𝑡 , 𝑖𝑡 , and 𝑜𝑡 correspond to the forget, input, where 𝑘 denotes the kernel size and 𝑤𝑖 represents the filter weights. Fig.
and output gates, respectively. Fig. 4 illustrates the LSTM architecture, 5 depicts the CNN architecture designed for one-dimensional financial
demonstrating the flow of information through the cell state and gates. time series, showcasing how convolutional layers extract significant
Each gate plays a crucial role in regulating data flow and maintaining features that can be used for predicting market behavior.
long-term dependencies. In algorithmic trading, CNNs have been applied to identify patterns
A recent study by Billah et al. [57] compared the Simple Moving in historical price charts, which aids in trend prediction. Hoseinzade
Average (SMA), Exponential Moving Average (EMA), and LSTM models et al. [65] present CNNpred, a CNN-based framework for stock market
for stock price prediction. The study found that while LSTM signifi- prediction that utilizes diverse variables beyond traditional indicators,
cantly outperforms SMA and EMA in short-term forecasting, with an enhancing prediction accuracy. Their framework demonstrates signif-
RMSE of 12.312 and MAPE of 2.06%, SMA and EMA are better suited icant performance improvements on major stock indices and success-
for long-term predictions. The paper also highlighted the limitations fully extracts higher-level features from varied data sources. However,
of LSTM in long-term prediction due to the complexity of the stock the authors acknowledge limitations due to the complexity of variable
market, noting that the findings were based on datasets from only interactions in financial data, suggesting that traditional CNN filters
six companies, which may limit generalizability. Additionally, Zhang may not be optimal and calling for further research to address these
et al. [58] introduced a novel CNN-BiLSTM-Attention model for stock challenges. Similarly, Lu et al. [66] introduce a novel CNN-BiLSTM-
price prediction, which achieved superior accuracy compared to LSTM, AM method for predicting stock closing prices, effectively combining
CNN-LSTM, and CNN-LSTM-Attention models. Despite its robustness CNN for feature extraction, BiLSTM for temporal dependencies, and an
and effectiveness, the study noted limitations regarding external factors attention mechanism to enhance accuracy. Tested over 1000 trading
and the model’s generalizability due to the lack of testing on North days of the Shanghai Composite Index, this approach achieved the high-
American market data. Gulmez et al. [59] further enhanced LSTM est prediction accuracy with a mean absolute error (MSE) of 21.952.
by optimizing it with the Artificial Rabbits Optimization algorithm The study highlights the need for further parameter optimization and
(LSTM-ARO) for stock price prediction. Their model outperformed broader validation in other time series prediction fields.
traditional ANN and GA-optimized LSTM models in terms of MSE, MAE, Chung et al. [67] propose a method to optimize CNN architecture
MAPE, and R-squared, although the study did not address the model’s using a Genetic Algorithm (GA) for stock market prediction, demon-
performance in volatile market conditions. strating the effectiveness of deep learning in financial forecasting.

7
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Fig. 7. Architecture of Variational Autoencoders (VAE).

conditions. Jung et al. [74] introduce a hybrid autoencoder-LSTM


Fig. 6. Architecture of Autoencoders.
model for forecasting foreign exchange volatility, showing improved
prediction accuracy over traditional LSTM models, especially with data
variability and outliers. The findings underscore the model’s sensitivity
The GA-optimized CNN outperforms traditional models, validated by to extreme outliers and its dependency on data quality, which may limit
statistical tests. However, its effectiveness relies on the fitness function’s generalizability to other currencies or timeframes.
definition and faces challenges due to the stock market’s inherent
complexity and noise, underscoring the need for further exploration of
4.2.5. Variational Autoencoders (VAEs)
advanced deep-learning techniques in this domain. In another study,
Mehtab et al. [68] developed deep learning models for predicting Variational Autoencoders (VAEs) are a generative model that learns
NIFTY 50 stock prices, achieving high forecasting accuracy, particularly a probabilistic mapping from data to a latent space and can generate
with the univariate encoder–decoder convolutional LSTM model. They new data samples by sampling from this latent space [75,76]. Unlike
emphasize fast execution, with the univariate CNN model processing in traditional autoencoders, VAEs impose a distributional prior on the
about 11.17 s. However, limitations include the inherent randomness latent variables, making them suitable for tasks such as anomaly de-
in stock prices affecting accuracy, reliance on historical data for predic- tection, denoising, and generation of new instances in financial data
tions, and the complexity of models requiring significant computational analysis [77]. The encoder outputs parameters of the posterior distri-
resources and expertise. bution, while the decoder attempts to reconstruct the data by sampling
from this distribution.
4.2.4. Autoencoders The VAE framework is defined as:
Autoencoders are unsupervised neural networks designed for di-
mensionality reduction and latent feature extraction [69,70]. The 𝑞𝜙 (𝑧|𝑥) =  (𝑧; 𝜇𝜙 (𝑥), 𝜎𝜙2 (𝑥)) (5)
model comprises an encoder that transforms input data 𝑥 into a latent where 𝜇𝜙 (𝑥) and 𝜎𝜙 (𝑥) represent the mean and standard deviation of the
representation 𝑧 and a decoder that reconstructs 𝑥 from 𝑧: variational approximation of the posterior distribution. The decoder is
𝑧 = 𝜎(𝑊𝑒 𝑥 + 𝑏𝑒 ) (Encoder) then trained to maximize the likelihood of data points given the latent
(4) variables sampled from the prior distribution.
𝑥̂ = 𝜎(𝑊𝑑 𝑧 + 𝑏𝑑 ) (Decoder)
Here, 𝑊𝑒 and 𝑊𝑑 are the weight matrices for the encoder and Fig. 7 illustrates a VAE’s structure, highlighting the model’s proba-
decoder, respectively. Fig. 6 shows the autoencoder structure, high- bilistic nature where the encoder learns to approximate the posterior
lighting the encoder–decoder architecture used for reconstructing input distribution of latent variables. VAEs have been applied in financial
data. This capability is beneficial for feature extraction and anomaly market prediction tasks, including generating synthetic stock data and
detection in financial datasets. detecting anomalies in high-frequency trading patterns [78,79].
In algorithmic trading, autoencoders have been employed for Recent work by Xu et al. [78] leverages VAEs for stock market
anomaly detection, which helps identify irregular trading patterns or prediction, demonstrating that VAEs can generate new sequences of
fraudulent activities. A notable contribution by Soleymani et al. [71] stock prices for stress testing and simulate different market conditions.
introduces DeepBreath, a novel deep reinforcement learning frame- Similarly, Li et al. [79] use VAEs to model uncertainty in stock returns,
work for portfolio management. This framework combines a restricted achieving better robustness in their predictions compared to traditional
stacked autoencoder for feature selection with a CNN for investment models. However, the study highlights the challenge of selecting an
policies, demonstrating superior performance against expert strategies appropriate prior distribution and the difficulty of generalizing the
in return on investment while maintaining scalability and efficiency. model to various market conditions.
However, the study acknowledges challenges in addressing settlement
risk and potential difficulties in generalizing to unforeseen market
4.2.6. Graph Neural Networks (GNNs)
conditions. In another study, Zhang et al. [72] explore the use of
GNNs are designed to process data represented as graphs, captur-
autoencoders for stock-index tracking, demonstrating their superiority
ing the relationships between nodes and their neighbors [80,81]. In
over traditional models, particularly with portfolios of fewer than
financial markets, where entities such as stocks, companies, or even
30 stocks. The effectiveness of specific autoencoder architectures in
capturing complex market representations is highlighted, although countries are interconnected through complex relationships, GNNs have
the findings may be limited by dataset specificity and insufficient been applied to capture dependencies between different financial in-
hyperparameter optimization, potentially affecting generalizability. struments [63,82]. The GNN model operates by iteratively aggregating
Bao et al. [73] present a novel deep-learning framework that com- information from neighboring nodes to update the state of each node,
bines wavelet transforms, stacked autoencoders, and LSTM networks making it highly effective for learning from structured data such as
for stock price forecasting, incorporating macroeconomic variables to financial networks [83,84].
enhance predictive accuracy. This model outperforms others in vari- The general message-passing operation in a GNN is given by:
ous financial markets, demonstrating effective predictive performance. ⎛ ⎞
∑ 1
However, it is time-consuming and may require advanced hyperpa- ℎ(𝑘+1) = 𝜎 ⎜𝑊 (𝑘) ⋅ ℎ(𝑘) + ⋅ 𝑊 (𝑘) ⋅ ℎ(𝑘) ⎟ (6)
rameter selection, with its generalizability limited by specific market
𝑖 ⎜ 𝑖 𝑐 𝑗 ⎟
⎝ 𝑗∈ (𝑖) 𝑖𝑗 ⎠

8
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

where ℎ(𝑘)
𝑖 represents the hidden state of node 𝑖 at the 𝑘-th layer,  (𝑖)
denotes the neighbors of node 𝑖, and 𝑐𝑖𝑗 is a normalization constant.
The function 𝜎 is a non-linear activation function, typically ReLU or
sigmoid.
GNNs have been applied in financial prediction tasks such as mod-
eling stock price movements by considering the correlations between
different assets and companies. For instance, Yilmaz et al. [85] inte-
grate graph representations with time series data for financial fore-
casting, employing an ensemble of deep neural networks. Their model
achieves a 23.52% annual return on the DOW30 index, significantly
outperforming heuristic trading strategies. By leveraging both tempo-
ral dependencies and graph-based relationships, the model enhances
predictive accuracy. However, challenges remain regarding the model’s
complexity, its reliance on historical data, and the necessity for broader
market validation to confirm its robustness across varying financial
conditions.
Beyond forecasting, GNNs have also been employed to detect fraud-
ulent financial market activities. Chen et al. [86] propose a multi-modal
GNN framework for identifying market manipulation in high-frequency
trading. The model constructs dynamic trading networks to capture
intricate interactions among traders, achieving an accuracy of 98.7%
with an 8.3 ms latency. This real-time capability is crucial for regula-
tory agencies and financial institutions aiming to mitigate fraudulent
activities. However, the approach faces computational challenges due
to the intensive nature of high-frequency trading data and remains
sensitive to sudden market fluctuations, necessitating further architec-
tural refinements to enhance adaptability. Another key application of
GNNs in finance is anomaly detection in trading behavior. Li et al. [87]
introduce a novel GNN-based architecture designed to detect anoma-
lies in high-frequency trading data. Their model improves real-time Fig. 8. Architecture of Transformer.
processing efficiency and enhances detection accuracy by 15% over
conventional methods. Additionally, it achieves a 15%–20% higher
F1-score, demonstrating its robustness across different market condi-
The self-attention mechanism computes a score for each pair of
tions. Despite these advantages, the model’s effectiveness is constrained
input elements, which determines how much focus to place on each
by its reliance on high-quality data, implementation complexity, and
element in the sequence. Given an input sequence represented as a
potential limitations in generalizing across diverse financial datasets.
matrix 𝑋, the self-attention function can be expressed as follows:
GNNs have also been utilized to rank investors based on trad- ( )
ing behaviors and social network interactions. Baltakys et al. [88] 𝑄𝐾 𝑇
Attention(𝑄, 𝐾, 𝑉 ) = softmax √ 𝑉 (7)
develop a graph-based ranking tool that evaluates investors by their 𝑑𝑘
level of suspicious trading activity, leveraging social connections to In this equation:
enhance predictability. The results indicate a high level of accuracy in
identifying potentially fraudulent investors, offering valuable insights • 𝑄 (queries), 𝐾 (keys), and 𝑉 (values) are derived from the input
for regulatory oversight. However, challenges such as dependency on sequence 𝑋 through learned linear transformations.
specific data sources, difficulty in generalizing findings across different • 𝑑𝑘 is the dimensionality of the keys, serving as a scaling factor to
financial markets, and the complexity of modeling social interactions stabilize gradients during training.
introduce limitations that require careful interpretation in regulatory • The softmax function normalizes the attention scores to produce
contexts. a distribution that indicates how much focus each input should
While GNNs offer significant advancements in financial applica- receive.
tions, they still face key challenges such as scalability, feature selec-
tion, and sensitivity to the choice of graph structure. Future research Transformers in Fig. 8, are built on an encoder–decoder archi-
should explore adaptive GNN architectures that dynamically refine tecture. The encoder transforms the input sequence into continuous
graph representations based on evolving market conditions. Addition- representations, while the decoder uses these representations to gen-
ally, hybrid models that integrate GNNs with other deep learning erate the output sequence. In the context of financial predictions, the
techniques may further enhance predictive power and resilience in encoder can analyze historical price movements and other relevant
financial decision-making. features, while the decoder can generate forecasts or trading signals.
Recent contributions have shown the potential of Transformer mod-
4.2.7. Transformer models els in financial forecasting. Wang et al. [89] introduced the Adaptive
Transformers have recently gained attention in algorithmic trading Long-Short Pattern Transformer (ALSP-TF), which enhances stock price
due to their ability to handle long-range dependencies in sequential forecasting by improving representation across varying context scales.
data without the limitations of recurrent architectures [89,90]. Unlike The model incorporates a learnable self-attention mechanism and graph
traditional RNNs, which process data sequentially, Transformers use a self-supervised regularization, outperforming state-of-the-art methods
mechanism called self-attention, allowing them to analyze input data in on multiple exchange datasets by capturing both short- and long-term
parallel. This enables the model to weigh the importance of different volatility patterns. However, the study highlights challenges such as the
parts of the input sequence, making it particularly effective for financial model’s complexity, dependence on data quality, and limited validation
time-series analysis [91,92]. across different markets.

9
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Bilokon et al. [93] compared Transformer and LSTM-based models Li et al. [99] presented a deep RL-based trading agent that en-
for financial time series prediction using high-frequency limit order hances DQN and A3C algorithms for dynamic financial environments.
book data. Their work introduced a new LSTM-based model (DL- The model leverages Stacked Denoising Autoencoders (SDAEs) and
STM) and a Transformer architecture, showing that while Transformers LSTMs for feature extraction and integrates practical mechanisms like
excel at absolute price prediction, LSTMs are superior for predict- n-step rewards. This agent outperforms baseline models by generating
ing price differences. However, the focus on specific data types and higher profits and robust risk-adjusted returns. However, sensitivity
the exclusion of models like FEDformer and Autoformer, which have to extreme market conditions and partial observability are limitations
limitations with stationary data, pose constraints on broader applica- that require further research, particularly in the context of multi-asset
tions. Lezmi et al. [90] explored Transformer models for time series trading. Huang et al. [100] proposed an efficient deep SARSA algo-
forecasting in finance, particularly in trend-following strategies and rithm for stock trading, combining Deep Reinforcement Learning (DRL)
portfolio optimization. The study highlighted the models’ capability to with a BiLSTM-Attention network for enhanced feature recognition.
capture long-range dependencies but observed performance declines in The model achieved an impressive 582.17% Cumulative Return and a
volatile markets after 2020. The authors also discussed challenges like Sharpe Ratio of 1.86 on the IXIC dataset, surpassing both traditional
the low signal-to-noise ratio in financial data, the difficulty in model and DRL-based strategies. Despite the promising results, the paper
generalization, and the need for continuous recalibration. lacks a discussion on the challenges associated with the real-world
Finally, Tran et al. [91] introduced a novel neural network layer deployment of such models.
that combines bilinear projection and attention mechanisms to im- In summary, Deep Learning models, including RNNs, LSTMs, CNNs,
prove financial forecasting in multivariate time-series data. The model Autoencoders, Transformers, and Reinforcement Learning, have
enhances interpretability by focusing on temporal information and demonstrated significant potential in financial predictions, particularly
achieves state-of-the-art results in predicting mid-price movements in algorithmic trading. These models leverage the strengths of neural
with minimal computational resources. Despite its strengths, the model networks to analyze complex patterns in historical data, leading to
faces limitations, including potential underfitting with large datasets improved predictive performance and enhanced decision-making capa-
and generalizability concerns due to the specific nature of the data bilities [101,102]. The advantages and disadvantages of the discussed
used. algorithms in algorithmic trading are summarized in Table 6.
In summary, Transformer models provide a robust framework for
algorithmic trading by leveraging self-attention mechanisms to capture 4.3. Role of deep learning in financial markets
intricate patterns in financial time-series data [94,95]. Their ability
to process data in parallel, combined with recent advancements in Deep learning has become a pivotal tool in financial markets due to
training and architecture design, positions them as a powerful tool for its ability to model complex, non-linear relationships that traditional
enhancing prediction accuracy and strategy development in dynamic machine learning models struggle to capture [103–105]. Unlike linear
financial markets. regression models or decision trees, DL can handle vast amounts of
unstructured data (e.g., time-series data, text from news articles, and
4.2.8. Reinforcement Learning (RL) social media sentiment) and learn hidden patterns that drive market
RL has gained prominence in algorithmic trading for its ability to behavior [106,107].
learn optimal trading policies through interaction with the trading en- The main strength of DL in finance is its ability to automate fea-
vironment [96,97]. Unlike traditional supervised learning approaches, ture extraction, allowing models to discover intricate relationships
RL focuses on training an agent to make decisions by taking actions 𝑎𝑡 at between variables without the need for extensive manual preprocess-
each time step 𝑡 to maximize cumulative rewards 𝑅𝑡 . The agent receives ing [108,109]. For instance, in the case of stock price prediction, DL
feedback from the environment based on its actions, which informs its models can combine technical indicators, macroeconomic factors, and
future decision-making process. textual data to produce more accurate forecasts. Recent studies, such
The learning process in RL is governed by the Bellman equation, as those by Billah et al. [57] and Long et al. [110], demonstrate
which captures the relationship between the current state 𝑠𝑡 , action 𝑎𝑡 , how advanced architectures like LSTM and hybrid models significantly
and the expected cumulative reward: improve forecasting accuracy compared to traditional approaches.
In comparison to traditional machine learning, DL models are more
𝑄(𝑠𝑡 , 𝑎𝑡 ) = 𝑟𝑡 + 𝛾 max 𝑄(𝑠𝑡+1 , 𝑎′ ) (8) adept at handling high-dimensional and noisy data, which is charac-
𝑎′
In this equation: teristic of financial markets. This has led to improved performance
in a wide range of financial applications, including price forecast-
• 𝑄(𝑠𝑡 , 𝑎𝑡 ) denotes the action-value function, representing the ex- ing, portfolio optimization, risk management, and algorithmic trading
pected utility of taking action 𝑎𝑡 in state 𝑠𝑡 . strategies [31,111]. For example, Xu et al. [60] introduced a hy-
• 𝑟𝑡 is the immediate reward received after taking action 𝑎𝑡 . brid LSTM-GARCH framework that enhances volatility risk prediction,
• 𝛾 is the discount factor (ranging from 0 to 1) that determines the demonstrating the effective integration of deep learning with classical
importance of future rewards. A 𝛾 close to 0 prioritizes immediate econometric models.
rewards, while a 𝛾 close to 1 emphasizes long-term rewards. However, despite their success, DL models face challenges such
• max𝑎′ 𝑄(𝑠𝑡+1 , 𝑎′ ) signifies the best possible future reward obtain- as overfitting, interpretability, and computational complexity. These
able from the subsequent state 𝑠𝑡+1 after taking action 𝑎𝑡 . challenges highlight the need for further research into hybrid models,
model interpretability, and advanced optimization techniques [112,
In algorithmic trading, RL has been applied to develop dynamic 113]. Biju et al. [34]) emphasize the importance of addressing algorith-
trading strategies that adapt to changing market conditions. Recent mic biases and the lack of governance in AI/ML applications in finance,
work by Khan et al. [98] integrates reinforcement learning techniques, advocating for future research directions to enhance the reliability of
such as Deep Q-learning and Proximal Policy Optimization, to im- these models.
prove trading performance. RL models adapt dynamically to market Recent advancements also indicate a growing focus on improv-
complexities and outperform traditional models, though challenges ing model robustness and generalizability. For instance, the work of
like overfitting, data quality issues, and high computational demands Ghani et al. [114] on the GARCH-MIDAS model highlights the need
remain. The study also highlights regulatory and ethical concerns due for a broader understanding of economic policy uncertainty’s impact
to the autonomous nature of RL-based strategies but showcases how RL on market volatility, particularly in emerging economies. This sug-
enhances decision-making in trading environments. gests that while DL models are powerful, their application should

10
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Table 6
Advantages and disadvantages of algorithms in algorithmic trading.
Model Advantages Disadvantages
∙ Captures temporal dependencies effectively ∙ Struggles with long-term dependencies
Recurrent Neural Networks (RNNs) ∙ Suitable for sequential data ∙ Slower training due to sequential nature
∙ Adapts to varying input lengths ∙ Prone to overfitting with noisy data.
∙ Addresses vanishing gradient issue ∙ Computationally intensive
Long Short-Term Memory Networks (LSTMs) ∙ Excels in modeling long-term dependencies ∙ Requires careful tuning of hyperparameters
∙ Can remember past information for future predictions ∙ May still overfit if not regularized properly
∙ Flexible architecture
∙ Effective in extracting features from time-series data ∙ Requires extensive labeled data
Convolutional Neural Networks (CNNs) ∙ Recognizes patterns and trends in price movements ∙ Less effective for temporal dependencies unless adapted
∙ Processes multiple input features simultaneously
∙ Useful for anomaly detection ∙ Limited interpretability
Autoencoders ∙ Capable of unsupervised learning ∙ May not be effective for sequential decision-making tasks
∙ Identifies hidden patterns in trading data ∙ Requires extensive training data for effective feature
extraction
∙ Generates synthetic financial data for modeling ∙ High computational cost due to variational inference
∙ Learns latent representations of market data ∙ May produce unrealistic outputs if poorly trained
Variational Autoencoders (VAEs)
∙ Effective for risk modeling and anomaly detection ∙ Requires large datasets for meaningful feature learning
∙ Reduces noise while preserving crucial patterns
∙ Models relationships between assets in financial networks ∙ High computational cost for large financial graphs
∙ Captures interdependencies between different stocks ∙ Requires careful graph construction for meaningful insights
Graph Neural Networks (GNNs)
∙ Suitable for analyzing stock correlations and portfolio ∙ Sensitive to noisy financial data
optimization
∙ Effective in predicting market movements using relational
data
∙ Handles long-range dependencies without sequential ∙ Requires large datasets and substantial computational
Transformer Models processing resources
∙ Faster training due to parallel processing ∙ Extensive fine-tuning needed
∙ Effective in capturing relationships in complex datasets ∙ Sensitive to noise in the data
∙ Learns optimal trading strategies through market interaction ∙ Sample inefficiency
Reinforcement Learning (RL) ∙ Adapts to changing market conditions ∙ Requires extensive training time
∙ Discovers novel strategies through exploration ∙ May overfit to historical data
∙ Difficult to implement reward functions that align with
trading goals

be complemented with an understanding of economic and financial 5.1.1. Stock price prediction
contexts. The prediction of stock prices has been one of the primary areas
In conclusion, the role of deep learning in financial markets is evolv- where deep learning models have made a significant impact [116,121].
ing, driven by continuous advancements in technology and methodolo- The ability to model the non-linear and temporal dependencies in mar-
gies. The integration of DL models with traditional economic theories ket data makes recurrent architectures such as LSTM and GRU highly
and real-world data remains crucial for maximizing their potential in effective [50,122]. Several hybrid models that combine traditional
addressing the complexities of financial systems. time-series analysis with advanced DL techniques have been developed
to enhance predictive accuracy.
5. Results and findings Table 7 summarizes key papers in stock price prediction, focusing
on the deep learning architectures used, the datasets analyzed, their
contributions, and the limitations identified in their approaches.
This section provides an in-depth analysis of selected studies focused
on deep learning (DL) applications in algorithmic trading [115]. We
5.1.2. Market volatility prediction
categorize the studies based on the specific financial applications they
Market volatility prediction is crucial for risk management and opti-
target, such as stock price prediction, market volatility prediction, port-
mizing trading strategies. Volatility forecasting has seen advancements
folio optimization, sentiment analysis, automated trading strategies,
through the use of DL models that can dynamically adjust to changing
risk management, and anomaly detection [57,60,116–118]. Each sub-
market conditions [127,128]. RNN-based models like LSTM and GRU
section reviews the key contributions and limitations of prominent DL
have been extensively applied, with some hybrid models integrating
models, providing a holistic view of the effectiveness and applicability
econometric approaches to improve performance [53,60].
of these models in trading environments.
Recent studies have emphasized the significance of incorporating
multiple data streams, including macroeconomic indicators, sentiment
5.1. Overview of selected studies based on applications analysis, and high-frequency trading data, to enhance the predictive
accuracy of volatility models [129,130]. The findings suggest that
The integration of DL in algorithmic trading has facilitated the hybrid models outperform single-model approaches, but data noise and
development of models that can handle vast datasets, uncover intri- overfitting remain challenges. As shown in Table 8, a variety of models
cate patterns, and make informed predictions in real-time [119,120]. have been applied to different datasets, each with specific contributions
As the financial landscape evolves, researchers have explored various and limitations.
applications of deep learning techniques to enhance trading strategies
and risk management. These studies offer important insights into how 5.1.3. Portfolio optimization
deep learning enhances decision-making and adjusts to swiftly evolving Portfolio optimization involves determining the best asset allocation
market dynamics, emphasizing both its potential and the challenges to maximize returns for a given level of risk [133,134]. Traditional
encountered in real-world applications. optimization methods have been enhanced by deep learning models,

11
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Table 7
Summary of selected studies on stock price prediction.
Ref. Year Model Contribution Limitation
[57] 2024 LSTM The paper compares SMA, EMA, and LSTM models for The paper highlights LSTM’s weakness in long-term stock
stock price prediction. It demonstrates LSTM’s superiority price prediction, where SMA and EMA perform better.
in short-term forecasting with an RMSE of 12.312 and Additionally, it acknowledges the stock market’s inherent
MAPE of 2.06%, while SMA and EMA perform better in complexity, which affects prediction reliability, and notes
long-term predictions. the limited generalizability of the findings due to the use
of datasets from only six companies.
[123] 2024 Variational Mode The paper introduces a novel two-stage prediction model The two-stage model has high computational complexity
Decomposition (VMD) combining variational mode decomposition (VMD) and due to the decomposition and ensemble learning
ensemble learning methods, outperforming fourteen other processes. Its performance heavily depends on the quality
models in stock price prediction. The model effectively of the VMD algorithm, and suboptimal decomposition can
decomposes time series data and enhances prediction reduce prediction accuracy. Additionally, the model’s
accuracy, demonstrating superior performance in generalizability to other time series domains remains
empirical validation with multiple evaluation metrics. unexplored.
[124] 2024 GARCH-AI The paper introduces a hybrid model integrating GARCH The paper’s limitations include a primary focus on
with AI techniques like LSTM, GRU, and Transformers to volatility dynamics rather than direct stock price
enhance stock price forecasting in African markets. prediction, which may restrict broader financial
Results show that GARCH-LSTM outperforms standalone forecasting applications. Additionally, it uses Airtel stock
models in metrics such as RMSE, MAE, MAPE, and data from a limited time frame, limiting generalizability
R-squared, demonstrating improved prediction accuracy to other markets, financial instruments, or diverse market
and robustness. conditions.
[125] 2024 GRA-WD- BiLSTM The paper introduces a GRA-WD-BiLSTM hybrid model to The study’s limitations include constraints in selecting
predict stock prices using environmental factors like air environmental data due to equipment and data
quality and weather. It achieved high prediction availability, restricting the influencing factors used.
accuracies of 95.93% for SSEC, 93.02% for SZI, and Additionally, it did not account for seasonal variations in
97.07% for HSI, showcasing the model’s effectiveness environmental factors, which may affect the accuracy of
across multiple stock indices. stock price predictions.
[58] 2023 CNN-BiLSTM The paper presents a novel CNN-BiLSTM-Attention model The model’s limitations stem from its reliance on stock
for stock price prediction, achieving superior accuracy trading data, which is affected by various external
compared to LSTM, CNN-LSTM, and CNN-LSTM-Attention factors. Additionally, the authors did not evaluate the
models. Tested on various stock indices, including the CSI model using North American market data, potentially
300, it demonstrates robustness and effectiveness, restricting its generalizability. Future research should
although it has limitations regarding external factors and incorporate multi-source information and test diverse
market data generalizability. datasets to overcome these limitations.
[59] 2023 LSTM The paper presents an optimized deep LSTM network The paper does not address the LSTM-ARO model’s
enhanced by the Artificial Rabbits Optimization algorithm performance in volatile market conditions or its
for stock price prediction. Using DJIA index stocks, the adaptability to diverse stock datasets. Additionally, its
LSTM-ARO model outperforms other models, including reliance on historical data may overlook sudden market
traditional ANN and GA-optimized LSTMs, based on changes and external factors that can significantly impact
metrics like MSE, MAE, MAPE, and R-squared. stock prices, limiting the model’s robustness and
applicability.
[126] 2023 BiLSTM-MTRAN- TCN The paper presents a novel BiLSTM-MTRAN-TCN model The model, while enhancing prediction accuracy, may
for stock price prediction, improving prediction accuracy struggle to capture all market dynamics due to the
and stability over traditional methods. It achieves 𝑅2 unpredictable nature of financial markets. Additionally,
improvements of 0.3% to 15.6% and reduces RMSE by the paper lacks a thorough discussion of the
24.3% to 93.5%, demonstrating strong generalization computational costs and scalability of the proposed
across different stock indices and time periods. method, raising concerns about its applicability in
real-time scenarios.

particularly reinforcement learning approaches that learn optimal port- Several studies have incorporated sentiment analysis with tradi-
folio strategies through interaction with the market environment [117, tional market data to enhance prediction models. For instance, integrat-
135]. ing sentiment scores from Twitter or financial news with stock price
Studies in this domain have applied actor-critic models and deep data can lead to better forecasting, especially during major market
Q-networks (DQN) for dynamic portfolio rebalancing. For instance, events [149,150]. However, the noisy nature of social media data and
Cui et al. [117] utilized a novel Deep RL hyper-heuristic framework, the ambiguity in sentiment classification pose significant challenges, as
demonstrating improved portfolio performance through the integration illustrated in Table 10.
of expert knowledge and low-level trading strategies. However, a key
limitation of RL models is their dependence on accurate reward signals 5.1.5. Risk management
from the market, which can lead to suboptimal performance during ex- Effective risk management in algorithmic trading involves pre-
treme market conditions, such as black swan events or sudden volatility dicting potential losses and optimizing trading strategies to mitigate
spikes. risks [119,151]. Deep learning models help identify risk factors by
A summary of selected studies on portfolio optimization is provided analyzing historical data and learning complex relationships between
in Table 9, highlighting their contributions and limitations. asset returns and market variables [152,153].
Autoencoders, Variational Autoencoders (VAE), and other unsuper-
5.1.4. Sentiment analysis for trading vised learning techniques have been deployed to detect anomalies that
Sentiment analysis has become popular in trading by using news could lead to portfolio losses. These models are valuable in identifying
articles, social media content, and financial reports to assess market outliers, but the lack of interpretability and the potential for false
sentiment [30,141]. DL models, such as CNNs and LSTMs, are fre- positives remain concerns [154,155]. The following Table 11 summa-
quently used to analyze text data and extract sentiment features that rizes key studies in this area, including contributions and limitations of
impact market behavior [142,143]. different deep learning models used for risk management.

12
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Table 8
Summary of selected studies on market volatility prediction.
Ref. Year Model Contribution Limitation
[60] 2024 LSTM-GARCH The paper presents a hybrid LSTM-GARCH framework The hybrid model’s performance is sensitive to parameter
that enhances volatility risk prediction in financial selection and input data quality, potentially impacting its
markets by integrating deep learning with classical predictive capabilities. Additionally, the study primarily
econometric models. The hybrid model significantly uses historical data from the Nasdaq 100 Index, which
outperforms standalone models, achieving an MSE of may limit the findings’ generalizability to other financial
0.604, thus providing a more accurate tool for volatility markets or asset classes, reducing broader applicability.
forecasting in financial risk management.
[114] 2024 GARCH-MIDAS This study enhances understanding of economic policy The study’s focus on a limited number of countries (US,
uncertainty’s impact on stock market volatility in China, UK, and Pakistan) may hinder the generalizability
emerging economies, particularly Pakistan, using the of findings to other emerging markets. Additionally, the
GARCH-MIDAS model. It finds the US EPU index lack of significant predictive information from Pakistan
significantly predicts volatility, while indices from and China’s EPU indices suggests potential data or model
Pakistan and China show limited predictive power, even limitations that require further investigation.
during economic upheavals like COVID-19.
[131] 2022 VU-GARCH-LSTM The paper introduces the VU-GARCH-LSTM hybrid model, The paper acknowledges challenges in GARCH-type
enhancing stock market volatility prediction by 21.03% models regarding their ability to capture complex
in RMSE compared to existing models. It emphasizes the fluctuations and nonlinear correlations in financial
significance of data distribution characteristics and shows time-series data. It suggests that further modifications to
improved performance in predicting abnormal financial the model’s concave function could enhance prediction
events, addressing limitations in capturing complex performance, indicating that there is still potential for
fluctuations in financial time-series data. improvement in the hybrid modeling approach.
[53] 2021 LSTM The paper investigates how investor sentiment from the The paper’s limitations include a relatively small sample
Xueqiu forum impacts stock market volatility in China, size of 1,230 observations, which may restrict the
revealing a significant negative correlation, especially generalizability of the findings. Additionally, the
during crises. It demonstrates that sentiment indices can dictionary-based sentiment analysis method may overlook
effectively predict short-term volatility, outperforming nuanced sentiments, suggesting a need for integrating
historical data models, while suggesting a causal machine learning techniques. Finally, the predictive
relationship where volatility influences sentiment. power diminishes for long-term volatility predictions.
[132] 2021 HAR-based framework This paper highlights the predictive role of speculative The paper notes that the relationship between speculative
sentiment on gold market volatility, demonstrating its sentiment and volatility is complex, potentially
significant impact on forecast accuracy. Results indicate overlooking certain market dynamics, which calls for
that incorporating sentiment improves models, leading to additional research. Furthermore, its focus on
better cumulative returns and Sharpe ratios, thereby high-frequency data may not accurately reflect long-term
enhancing investment strategies. Further research is trends or behaviors in the gold market, limiting the
needed to explore complex market dynamics. generalizability of the findings.
[19] 2019 ARIMA The paper presents an ARIMA model tailored for The paper acknowledges that the ARIMA model, while
predicting stock market movements in the Indian stock effective, cannot guarantee 100% accuracy due to the
market, achieving an average deviation of approximately unpredictability of auto-regressive processes. It also notes
5% mean percentage error. It demonstrates the model’s the necessity for stationary time series data, complicating
effectiveness in real-world applications while emphasizing the modeling process if differencing is required.
the importance of time series analysis and its potential Additionally, external factors affecting stock prices are
across various fields. not explored.

5.1.6. Anomaly detection and fraud detection for operational efficiency and cost reduction [176,177]. Various ma-
Anomaly detection plays a crucial role in safeguarding trading chine learning models have been applied to improve forecasting accu-
systems by identifying irregularities and suspicious activities that may racy by capturing complex temporal patterns and external factors [178,
indicate fraud or market manipulation [162,163]. In trading envi- 179].
ronments, detecting anomalies involves recognizing deviations from Nguyen et al. [180] presented an LSTM-based method for multi-
normal patterns in transaction data or stock price movements. DL variate time series forecasting and anomaly detection using an LSTM
models, such as autoencoders, RNN, and hybrid models, have shown Autoencoder with OCSVM. While improving prediction accuracy, the
model focuses on past anomalies, faces data access challenges, and
promise in capturing these rare events due to their ability to model
produces single-value outputs, limiting its applicability in complex
non-linear patterns and temporal dependencies [164,165].
multivariate scenarios. This highlights the importance of addressing
Autoencoders, particularly, are effective in unsupervised learning
both temporal dependencies and external anomalies in supply chain
tasks like anomaly detection, as they are trained to reconstruct normal
forecasting. Punia et al. [181] proposed a cross-temporal forecasting
patterns in data [166,167]. When the input deviates significantly from framework (CTFF) using LSTM networks to enhance forecast coherency
expected behavior, reconstruction error increases, signaling potential in retail supply chains. It achieved lower MAE, MSE, and MAPE with
fraud. Similarly, RNN-based models, especially LSTM and GRU, are statistically significant improvements, with confidence levels exceeding
employed to model temporal sequences and detect unusual behavior 95%. However, its effectiveness depends on high-quality data, may not
in real-time transactions [154,168]. generalize to other industries, and involves complex implementation,
Table 12 provides a summary of notable studies applying DL models limiting adoption by smaller organizations. This model addresses fore-
for anomaly and fraud detection in trading systems, outlining key cast accuracy in the retail context, contributing to the growing body
contributions and their respective limitations. of work focusing on improving forecast consistency and accuracy in
supply chains.
Belhadi et al. [181] introduced a hybrid RotF-LB ensemble approach
5.1.7. Supply chain forecasting to forecast credit risk in agriculture 4.0 investments. By identifying 22
Supply chain forecasting is essential for predicting product demand, key variables and offering practical guidelines for SMEs and FSPs, the
managing inventory levels, and optimizing logistics, which are critical model improves forecasting accuracy, utilizing data from 216 SMEs,

13
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Table 9
Summary of selected studies on portfolio optimization.
Ref. Year Model Contribution Limitation
[117] 2024 Deep Reinforcement The paper presents a novel Deep RL hyper-heuristic The paper identifies several limitations, including the
Learning framework for multi-period portfolio optimization, complexity of real-world constraints, such as cardinality
integrating expert knowledge for enhanced asset and turnover, which create a nonconvex search space.
allocation. Evaluated on five capital market instances, it Additionally, the approach’s effectiveness may fluctuate
shows significant performance gains over traditional with changing market conditions, and the inherent
methods, with implications for investment strategies and complexity of DRL algorithms presents computational
regulatory frameworks, despite challenges from real-world challenges in larger problem instances, affecting
constraints and market variability. scalability.
[136] 2024 Method of Moving The paper presents a new portfolio optimization method The paper acknowledges challenges with financial data
Asymptotes (MMA) that directly optimizes portfolio weights through complexity, including noise and non-stationarity, which
constrained penalized regression, enhanced by AutoML may hinder accurate asset dependency modeling. Despite
for model selection and hyperparameter tuning. Tested on a penalization term, overfitting remains a concern in
M6 competition data, the approach achieved a 9.5% high-dimensional settings. Additionally, the method’s
return and a 5.045 information ratio, surpassing success is highly dependent on data quality, potentially
traditional strategies. affecting model accuracy.
[137] 2024 Quantile Regression The paper introduces a novel performance measure The paper’s limitations include a focus on the US market,
Neural Network integrating non-Gaussianity and systemic risk into potentially reducing the generalizability of findings to
portfolio optimization, extending the Conditional other regions. Its complex implementation, requiring
Value-at-Risk framework. It demonstrates superior advanced technical expertise and computational
profitability and resilience in backtesting, especially resources, may limit accessibility. Additionally, reliance
during crises, outperforming benchmark strategies, even on specific modeling assumptions may introduce biases if
with transaction costs included, highlighting practical market conditions differ from those assumed.
applicability in market distress.
[138] 2022 CNN-BiLSTM This paper integrates a CNN-BiLSTM model with the The study’s limitations include its focus on Thai stock
Markowitz mean–variance (MV) model to enhance stock data, limiting broader applicability. Hyperparameters
selection and portfolio optimization. Experimental results were manually tuned, and external factors, such as
show improved performance over traditional models, political events, were not considered. Additionally, the
especially in terms of the Sharpe ratio and risk. A model’s time complexity was not addressed, potentially
five-stock portfolio is identified as optimal for individual affecting real-world implementation and computational
investors. efficiency.
[139] 2021 Random Forest This paper integrates traditional time series models with The study’s limitations include the use of only simple
machine learning and deep learning techniques like historical returns as input features, excluding technical
Random Forest, SVR, LSTM, DMLP, and CNN for stock indicators or economic factors that could enhance model
return prediction. It finds that Random Forest, combined accuracy. Additionally, the high turnover rates in the
with portfolio optimization models, yields the best RF-based portfolio models significantly reduce overall
results, though high turnover rates reduce profitability. profitability, indicating a need for further research to
address these challenges.
[140] 2021 XGBoost This paper presents a hybrid model combining machine The paper acknowledges challenges in accurately
learning and mean–variance optimization for portfolio predicting complex input data, which can impact
construction, achieving superior prediction accuracy and prediction accuracy. It also highlights that transaction
improved returns. The results demonstrate enhanced costs were primarily analyzed as unilateral, potentially
performance over traditional methods and benchmark overlooking real-world trading implications. Additionally,
models, emphasizing the importance of stock selection the model’s effectiveness heavily depends on
and the complexity of input data in investment hyperparameter optimization, necessitating robust
decision-making. techniques to avoid suboptimal performance.

195 enterprises, and 104 FSPs. Limitations include sample selection 5.2. Datasets and resources
bias, dimensionality challenges, and a need for comparisons with more
advanced EML techniques. This research emphasizes the need for hy- In deep learning applications for algorithmic trading, the choice of
brid models to address forecasting challenges in niche sectors, such dataset is crucial, as it significantly impacts model performance and
as agriculture. Weng et al. [182] proposed a hybrid LightGBM-LSTM predictive accuracy. Various datasets, each with unique features and
characteristics, can be utilized for different tasks such as stock price
model for supply chain sales forecasting, offering high accuracy, ef-
prediction, volatility forecasting, anomaly detection, and sentiment
ficiency, and interpretability. This model enhances inventory man-
analysis. Table 14 shows ten noteworthy datasets commonly used in
agement and demand planning, outperforming traditional methods in
deep learning research related to algorithmic trading, along with a
prediction speed and providing insights into sales influences. However, summary highlighting their references, names, and descriptions.
its complexity demands more computational resources, relies on data The Astock dataset [184] focuses on the China A-shares market,
quality, and may have limited generalizability across industries. Weng containing 40,963 stock-specific news items annotated with trading
et al.’s work complements other research by demonstrating the effi- actions. It features 24 stock factors and is divided into in-distribution
cacy of hybrid models in balancing accuracy with interpretability and and out-of-distribution splits, supporting stock movement classification
efficiency. and simulated trading. The Stock Market Dataset [185] offers historical
Table 13 summarizes some studies, highlighting the different mod- daily prices of NASDAQ-traded stocks and ETFs, retrieved via Yahoo
Finance, with each ticker’s data stored in CSV format. The NIFTY-50
els used, their contributions to supply chain forecasting, and the limi-
Stock Market Data [186] provides daily stock price history and trading
tations that must be addressed for broader application.
volumes for fifty stocks in India’s NIFTY-50 index from January 1,
These studies demonstrate the effectiveness of deep learning and 2000, to April 30, 2021, including individual CSV files for each stock
hybrid models in supply chain forecasting by capturing temporal de- and metadata updated monthly for ongoing analysis. The StockNet
pendencies, external influences, and multi-task requirements, offering dataset [187] includes two years of stock price movements for 88
scalable solutions for improving operational performance. stocks across nine sectors, along with Twitter data. It features both

14
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Table 10
Summary of selected studies on sentiment analysis for trading.
Ref. Year Model Contribution Limitation
[118] 2024 LLM The paper explores the use of LLMs like OPT, BERT, and The paper highlights challenges in adopting LLMs for
FinBERT for financial sentiment analysis, demonstrating financial analysis, including the need for technical
their superior performance in market prediction. The OPT expertise, computational resources, and funding. It also
model achieved a 74.4% accuracy in predicting stock acknowledges the complexity of interpreting textual data
returns and generated a 355% gain from a long-short and suggests that further research is needed to explore
strategy between 2021–2023. additional models and refine techniques for financial text
analysis.
[144] 2024 CNN, LSTM, and This paper introduces a sentiment analysis framework for The paper highlights potential data collection bias due to
GRU-CNN Turkish financial tweets using DL models like CNN, specific keyword usage, which may skew sentiment
LSTM, and GRU-CNN. The CNN model with pre-trained representation. Additionally, Turkish tweets’ ambiguity
word embedding achieved the highest accuracy for binary complicates pre-processing, impacting model performance.
classification (83.02%) and multi-class classification The research also notes the limited amount of sentiment
(72.73%), highlighting the effectiveness of deep learning analysis work in Turkish compared to English, indicating
for sentiment analysis in resource-limited languages. a gap for future studies.
[145] 2024 JST and TS-LDA This paper addresses a research gap by applying The paper’s limitations include cryptocurrency data
aspect-based sentiment analysis (JST and TS-LDA) to sources providing short comments, limiting analysis
predict Bitcoin’s directional returns, using polarity, depth. Additionally, classifying comments based on parent
subjectivity, and LDA topics as features. It demonstrates posts may lead to incorrect subject assignments, skewing
enhanced model performance through ROC AUC and results when discussions deviate from the original topic,
accuracy, providing interpretable topics and leveraging suggesting the need for better methods or alternative
diverse textual data sources. data sources for accurate analysis.
[146] 2024 Spatial Federated The paper introduces a novel ensemble technique The paper’s limitations include heavy reliance on the
Learning combining spatial federated learning with sentiment quality and quantity of news data for sentiment analysis,
analysis, utilizing news stored on a blockchain for stock making predictions vulnerable to biased or limited
market predictions. Results show improved prediction information. The ensemble and federated learning
accuracy, outperforming traditional models. A 5-bit framework also increases implementation complexity, and
pattern guides trading decisions, validated on Hong Kong its generalizability to other stock markets requires further
Stock Exchange datasets through sentiment transfer validation across different regions and datasets.
learning.
[147] 2023 ChatGPT 3.5 This paper examines the capabilities of large language The paper acknowledges limitations, including the short
models, focusing on ChatGPT 3.5, for financial sentiment duration of the dataset, which may not capture all factors
analysis in the foreign exchange market. ChatGPT influencing financial markets. Additionally, the models
surpassed FinBERT by 35% in sentiment classification only partially aligned with market movements, suggesting
accuracy and showed a 36% stronger correlation with that sentiment analysis alone may not fully explain price
market returns, demonstrating its effectiveness in variations, requiring further studies with extended time
interpreting market trends and informing trading frames for validation.
strategies.
[148] 2022 Machine Learning The paper presents a novel sentiment index that The paper overlooks transaction fees, impacting the
incorporates weighted textual content and financial real-world applicability of trading strategies derived from
anomalies, enhancing stock trend predictions by the sentiment index. Additionally, it highlights challenges
considering day-of-the-week and holiday effects. faced by retail investors in executing short trades, which
Experimental results show improved predictive ability may restrict the practical implementation of the proposed
across various machine learning models, contributing to strategies in certain markets, limiting their overall
better accuracy, risk management, and potential returns effectiveness and accessibility.
for investors.

raw and preprocessed price data from Yahoo Finance and tweet data in major indices and sector performance, including information on oil
from Twitter, enabling stock movement prediction using text and price price fluctuations following geopolitical events and significant stock
signals. The dataset of stock market indices [188] tracks the perfor- movements. The CryptoBubbles dataset [193] encompasses around 404
mance of eight stock market indices (IPC, S&P 500, DAX, DJIA, FTSE, cryptocurrencies from top exchanges, featuring daily price data and
N225, NDX, CAC) from June 2006 to May 2023, including daily stock tweets for bubble detection, allowing analysis of market behaviors and
indices and various economic indicators sourced from Yahoo Finance trends despite inherent challenges.
and OECD. These datasets serve as essential resources for researchers and prac-
The Stock-Market Sentiment Dataset [189] comprises tweets col- titioners in the field of algorithmic trading, enabling the development
lected from multiple Twitter handles, categorizing economic news into and testing of deep learning models that can effectively analyze market
positive and negative sentiments, serving as a valuable resource for behavior and make informed trading decisions.
sentiment analysis in trading strategies. The Stock Tweets for Sentiment
5.3. Performance metrics
Analysis and Prediction dataset [190] contains over 80,000 tweets
related to the top 25 stock tickers on Yahoo Finance, collected be-
In assessing the performance of deep learning (DL) models in algo-
tween September 30, 2021, and September 30, 2022. It includes tweet
rithmic trading, a range of evaluation metrics is employed to capture
texts, stock names, company names, and corresponding stock market different aspects of model effectiveness. These metrics are drawn from
price and volume data, facilitating sentiment analysis and stock price both the fields of machine learning and finance to measure predictive
prediction. accuracy, profitability, and risk-adjusted returns.
The Dataset for Stock Market Prediction [191] consists of historical
stock prices for three petroleum companies—Pakistan State Oil, Hascol, 5.3.1. Accuracy, precision, and recall
and Attock Petroleum Limited—extracted from the Pakistan Stock Ex- For classification tasks, such as predicting whether stock prices will
change website over the last four years, including daily attributes and rise or fall, metrics such as accuracy, precision, and recall are widely
Twitter data for sentiment analysis. The United States Stock Market In- used [194,195]. These metrics evaluate how well the DL models are
dex dataset [192] captures US stock market activity, detailing changes able to predict binary or multi-class outcomes based on historical data.

15
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Table 11
Summary of selected studies on risk management.
Ref. Year Model Contribution Limitation
[156] 2024 – The paper demonstrates how AI-driven predictive The paper highlights limitations such as the dependence
analytics enhances risk assessment, decision-making, and on high-quality data for accurate AI predictions and the
profitability in finance. It emphasizes AI’s role in complexity of integrating AI into existing financial
real-time risk detection, proactive decision-making, and systems. It also neglects possible ethical and regulatory
regulatory compliance. Additionally, dynamic learning issues, such as data privacy and algorithmic bias, which
models continuously improve predictions, helping are vital considerations for the implementation of AI
institutions mitigate risks and meet regulatory technologies in finance.
requirements efficiently.
[157] 2024 – The paper discusses the integration of AI in stock trading, The paper identifies limitations, including the risk of
highlighting its role in improving risk management and overfitting in AI models, which may hinder their ability
decision-making processes. AI technologies, including to generalize to future market conditions. Additionally,
machine learning and neural networks, enhance trading the reliance on historical data can lead to poor
efficiency while addressing regulatory and ethical performance in changing markets. Ethical concerns, such
concerns. It also explores real-world case studies and as potential market manipulation, also require careful
future AI trends in finance. governance.
[158] 2023 LSTM The paper contributes to predictive analysis and risk The paper points out the challenges of applying machine
management in trading by integrating LSTM networks learning to stock market predictions, especially when
and collaborative filtering, enhancing stock market dealing with time series data, as the outcomes may be
predictions. Tested on the Stock Price EOD Dataset, the unreliable. Although deep learning models have shown
model achieves over 95% accuracy, outperforming promise, additional research is necessary to investigate
traditional benchmarks and highlighting the effectiveness more algorithms and datasets to enhance the robustness
of DL in real-world trading scenarios. and applicability of the proposed models.
[159] 2020 LSTM The paper introduces an Automated Trading System The paper highlights overfitting as a key limitation of
(ATS) using LSTM networks for stock price prediction, Artificial Neural Networks (ANN), affecting real-world
integrated with risk management strategies. The performance. It also raises concerns about model
LSTM-RMODV method achieved a 228.94% return on generalizability due to limited evaluation in actual
invested capital. Despite low prediction accuracy, robust market conditions. Additionally, while technical
risk management ensured consistent profitability, knowledge is useful, deeper market insights are necessary
highlighting its significance in trading systems. for improving model performance, especially with
advanced techniques like VWAP.
[160] 2020 Decision Tree & ANN The paper presents the Artificial Intelligent Risk A key limitation of the study is the complexity of
Management System (AIRMS), employing machine calibrating artificial neural networks, which may delay
learning for risk management in finance. It developed their implementation in real-world scenarios.
two systems, AIRMS-DT and AIRMS-ANN, which Additionally, the research emphasizes the need for further
improved currency portfolio returns by 50% and 40%, investigation into diverse strategies and feature extraction
respectively, effectively classifying trading signals and methods to enhance the model’s overall effectiveness and
outperforming original portfolios. performance.
[161] 2020 Neural Network The paper enhances financial risk management by The study’s primary limitation is the complexity in
developing an AI system that integrates neural networks, calibrating artificial neural networks, which may hinder
digitized news fluctuations, and candlestick chart data to their practical implementation in real-world trading.
improve price forecasts for SiU9 US Dollar futures. Additionally, the research emphasizes the need for further
Results show significant accuracy improvements and investigation into diverse strategies and feature extraction
effective risk assessment, minimizing losses in speculative methods to improve the model’s overall effectiveness and
trading through a neural network algorithm. performance in financial risk management.

• Accuracy: Accuracy measures the ratio of correctly predicted The F1 score is especially useful in trading scenarios where false
observations to the total observations: positives and false negatives have different costs.
𝑇 .𝑃 𝑜𝑠𝑖𝑡𝑖𝑣𝑒 + 𝑇 .𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒
Acc = (9)
𝑇 𝑜𝑡𝑎𝑙 5.3.2. Mean squared error (MSE) and root mean squared error (RMSE)
where 𝑇 .𝑃 𝑜𝑠𝑖𝑡𝑖𝑣𝑒 is the number of true positives, 𝑇 .𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒 is For regression tasks, such as predicting future stock prices, the MSE
the number of true negatives, and Total means sum of all data. and RMSE are commonly used [196,197]. These metrics measure the
• Precision: Precision focuses on the proportion of true positives mean squared distance between the predicted and actual values.
among all positive predictions:
1 ∑
𝑁
𝑇 .𝑃 𝑜𝑠𝑖𝑡𝑖𝑣𝑒 MSE = (𝑥 − 𝑥̂ 𝑖 )2 (13)
Prec = (10) 𝑁 𝑖=1 𝑖
𝑇 .𝑃 𝑜𝑠𝑖𝑡𝑖𝑣𝑒 + 𝐹 .𝑃 𝑜𝑠𝑖𝑡𝑖𝑣𝑒
High precision indicates a low false positive rate, which is crucial where 𝑥𝑖 is the actual value, 𝑥̂ 𝑖 is the predicted value, and 𝑁 is the
for minimizing incorrect buy/sell decisions in algorithmic trading. number of observations. MSE penalizes larger errors more than smaller
• Recall: Recall, or sensitivity, measures the proportion of actual ones, making it sensitive to outliers.
positives that were correctly identified: √
𝑇 .𝑃 𝑜𝑠𝑖𝑡𝑖𝑣𝑒 RMSE = 𝑀𝑆𝐸 (14)
Rec = (11)
𝑇 .𝑃 𝑜𝑠𝑖𝑡𝑖𝑣𝑒 + 𝐹 .𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒 RMSE provides a more interpretable metric since it is on the same scale
In trading, a high recall ensures that the model captures as many as the original data. It is widely used to evaluate stock price predictions
profitable trades (true positives) as possible. or forecast error in financial markets.
• F1-Score: The F1-score balances precision and recall, providing a
single measure of a model’s accuracy when the class distribution
5.3.3. Sharpe Ratio (SR)
is imbalanced:
In algorithmic trading, financial metrics such as the SR are essential
Prec × Rec
F1-Score = 2 × (12) for assessing the risk-adjusted performance of a trading strategy [198].
Prec + Rec

16
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Table 12
Summary of selected studies on anomaly detection and fraud detection.
Ref. Year Model Contribution Limitation
[169] 2024 GNN This paper introduces a novel fraud detection framework The main limitation of this paper is its reliance on a
integrating Graph Neural Networks (GNNs) with anomaly specific dataset, which may not capture the full diversity
detection. Applied to the imbalanced Credit Card Fraud of global fraud patterns. Additionally, while the model
Detection dataset, the model achieves a 95% detection shows robustness against certain fraud schemes, its
rate with a 2% false positive rate, outperforming performance on emerging and highly complex fraud
state-of-the-art methods by 10% and demonstrating tactics remains unexplored, requiring further validation.
robustness against complex fraud schemes.
[170] 2024 Transformer-based The paper presents a novel Transformer-based The paper faces challenges with the curse of
autoencoder autoencoder framework for detecting anomalies in limit dimensionality due to multiple hidden states per feature,
order book (LOB) data. It introduces a dissimilarity complicating the learning process. Additionally, the
function and a trade-based manipulation simulation scarcity of relevant machine learning literature for
methodology, achieving high performance on NASDAQ financial market anomalies and concerns about addressing
stock data, surpassing traditional models like LSTM in the sequential nature of trade-based manipulations limit
detecting fraudulent patterns. the model’s generalizability.
[171] 2024 GNN This paper explores the use of GANs in fraud detection, The paper highlights technical challenges in GANs, such
demonstrating their ability to model complex data as pattern collapse, and limiting data diversity. It also
distributions and detect anomalies in transactions. The raises privacy concerns, particularly in sensitive sectors
proposed GAN-based system enhances transaction like finance and healthcare, where synthetic data
security, outperforming traditional methods, and offers a generation could expose private information. Additionally,
foundational framework for future research on GAN ethical issues surrounding the misuse of GANs, such as
applications across various domains. identity forgery, are addressed.
[172] 2024 – The paper presents a transformative framework for fraud The paper highlights limitations including data privacy
detection in fintech, showcasing how advanced AI concerns due to the extensive personal data required for
techniques, including deep learning and natural language AI-driven fraud detection, risks of algorithmic bias from
processing, outperform traditional methods. It highlights historical data, and the challenge of continuous learning.
AI’s enhanced capabilities in real-time detection and These systems must adapt to evolving fraud tactics, which
unstructured data analysis, emphasizing the need for can be resource-intensive and complex to maintain.
scalable, adaptive, and predictive solutions.
[173] 2024 LightGBM The paper investigates anomaly detection (AD) methods The paper’s limitations include the lower fraud detection
for fraud detection in online credit card payments, performance of AD methods compared to LightGBM and
comparing them to supervised learning approaches like their reliance on normal samples, which limits
LightGBM. While LightGBM performs better across adaptability to new anomalies. Additionally, the results
metrics, AD methods show more robustness to are based on a specific dataset, potentially limiting their
distribution shifts. However, AD methods fall short in applicability to other datasets or contexts with different
fraud detection effectiveness compared to LightGBM. characteristics.
[174] 2023 Random Forests and The paper emphasizes the need for real-time machine The paper highlights challenges such as potential bias in
Gradient Boosting learning in fraud detection to improve financial security machine learning models, leading to unfair targeting of
Machines by moving beyond static systems. Results show that demographics. It also addresses the lack of transparency
machine learning models enhance fraud detection by in models, which can undermine trust and complicate
reducing false positives and increasing detection speed, regulatory compliance. Additionally, current systems still
with fine-tuning improving the balance between precision face inherent challenges for effective fraud detection.
and recall.
[175] 2020 SADE The paper introduces the SADE framework, which detects The paper does not explicitly address limitations, but
anomalies in financial transaction networks by focusing potential challenges include the complexity of selecting
on subgraph-level patterns using role-guided embeddings. parameters for role-guided embeddings, which could
Extensive experiments show that SADE outperforms affect performance. Additionally, the framework’s reliance
existing methods in detecting anomalous subgraphs, on the quality of financial transaction data may hinder
demonstrating its effectiveness in fraud detection and risk effectiveness, as noisy or incomplete data could lead to
modeling without requiring prior knowledge. inaccurate anomaly detection results.

The SR measures the excess return (the return above a risk-free rate) 5.3.5. Sharpe ratio vs. ROI
per unit of volatility (risk): While ROI focuses purely on profitability, the Sharpe Ratio also
E[𝐶𝑥 − 𝐶𝑦 ] considers the risk (volatility) taken to achieve that profit. In financial
SR = (15) markets, a high ROI might not always be desirable if it comes with
𝜎𝑥
where 𝐶𝑥 is the return of the portfolio, 𝐶𝑦 is the risk-free rate, and excessive risk, which is why the Sharpe Ratio is often preferred for
𝜎𝑥 is the standard deviation of the portfolio’s return (volatility). A evaluating DL-based trading strategies [200,201].
higher SR indicates a better risk-adjusted return. In DL-based trading,
the Sharpe Ratio is often used to evaluate the profitability of models 5.4. Effectiveness of performance metrics
while accounting for the risks associated with market volatility.

5.3.4. Return on investment (ROI) The effectiveness of these metrics varies depending on the applica-
ROI is another financial metric used to assess the profitability of tion. For example:
DL-based trading models [199]. It is calculated as:
Net Profit • Accuracy, Precision, and Recall are effective in binary classi-
ROI = × 100 (16) fication tasks like predicting price movements (up/down). They
Investment Cost
where the net profit is the total gain from the investment and the invest- help ensure that DL models minimize incorrect buy/sell signals.
ment cost is the amount initially invested. ROI measures the efficiency Higher values are generally better for these metrics, as they
of the investment and is often used to compare the performance of indicate fewer errors in classification, while lower values suggest
different trading strategies implemented using DL models. more misclassifications and less reliable predictions.

17
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Table 13
Summary of selected studies on supply chain forecasting.
Ref. Year Model Contribution Limitation
[180] 2021 LSTM This paper proposes an LSTM-based method for This anomaly detection model primarily identifies
multivariate time series forecasting and an LSTM anomalies from historical data without predicting future
Autoencoder combined with OCSVM for anomaly anomalies. Data access restrictions from companies
detection. The model demonstrates superior performance limited the study’s scope. Additionally, the LSTM-based
in fashion retail and NASA datasets, optimizing method produces a single value for multivariate data,
hyperparameters and integrating data sources to improve potentially constraining its effectiveness in more complex
prediction accuracy. scenarios.
[181] 2021 Ensemble Model This study introduces a hybrid RotF-LB ensemble This study acknowledges potential sample selection bias
approach to forecast credit risk in agriculture 4.0 and the need for larger datasets for validation.
investments. It identifies 22 key variables and provides Dimensionality challenges persist despite the LB
practical guidelines for SMEs and FSPs. The model algorithm, and further exploration of non-linear methods
outperforms others in forecasting accuracy, utilizing data like autoencoders is suggested. Additionally, the study
from 216 SMEs, 195 enterprises, and 104 FSPs. recommends comparing the RotF-LB approach with more
advanced EML techniques.
[183] 2020 LSTM This paper presents a cross-temporal forecasting The CTFF relies heavily on high-quality point-of-sales
framework (CTFF) using LSTM networks to ensure data, limiting performance with poor data. Its
forecast coherency across retail supply chain levels. It applicability to industries beyond large retail supply
demonstrates lower MAE, MSE, and MAPE compared to chains remains uncertain. The computational complexity
cross-sectional methods, with statistically significant of LSTM-based integration may hinder adoption by
improvements validated by t-tests and p-values at smaller organizations, while external market factors and
confidence levels above 95%. disruptions are not comprehensively addressed.
[182] 2020 LightGBM and LSTM This paper presents a hybrid model combining LightGBM This hybrid model’s complexity requires more
and LSTM for supply chain sales forecasting. It offers computational resources and may be harder to
high accuracy, efficiency, and interpretability, improving implement. Its effectiveness depends on data quality, and
inventory management and demand planning. The model poor data can lead to inaccurate forecasts. Additionally,
outperforms traditional methods in prediction speed and its performance may vary across industries, limiting its
provides valuable insights into sales influences. broader applicability beyond the tested datasets.

• MSE and RMSE are valuable for regression tasks, such as predict- Addressing these issues is crucial for improving model performance
ing future stock prices or financial indices. These metrics provide and reliability, ultimately fostering greater trust and adoption among
insights into the precision of the model’s predictions. Lower val- traders and financial institutions. Continued research and innovation
ues are better for MSE and RMSE, as they indicate less deviation in this domain will be essential for overcoming these obstacles and
from the actual values. Higher values indicate greater deviation unlocking the full potential of deep learning in finance.
from the actual values, suggesting less accurate predictions.
• Sharpe Ratio is critical for evaluating the overall performance 6.1. Data quality
of DL-based trading strategies by accounting for both return and
risk. It helps in selecting strategies that provide the best risk-
One of the most frequently cited challenges in the literature is the
adjusted returns. Higher values are better for the Sharpe Ratio,
quality of financial data. Financial markets are notoriously noisy, and
as they indicate higher returns relative to risk, while lower values
this presents significant difficulties for deep learning (DL) models [208,
indicate less favorable returns for the amount of risk taken.
209]. Financial data is often influenced by a wide range of factors such
• ROI is useful for measuring the raw profitability of a trading strat-
egy, though it should be paired with other metrics like the Sharpe as market sentiment, macroeconomic news, and geopolitical events,
Ratio to provide a complete picture of performance. Higher values which introduce a high degree of variability and randomness.
are better for ROI, as they indicate greater profitability, while
lower values suggest less profitability or potential losses. 6.1.1. Noise in financial data
Noise in financial data refers to random price fluctuations that do
5.5. Comparison of some recent studies not represent genuine market signals. Noise is particularly problematic
for DL models that aim to extract patterns from historical data [208,
In recent years, deep learning techniques have gained significant 210]. Models trained on noisy data are more likely to capture irrelevant
attention in the field of algorithmic trading [100,202]. These studies patterns, leading to poor predictive performance [211]. For instance,
explore various methods for improving trading strategies by lever- stock prices are influenced by short-term market fluctuations that may
aging advanced algorithms like neural networks, decision trees, and not follow any predictable pattern, but DL models may incorrectly
reinforcement learning. Researchers have focused on applying deep interpret these fluctuations as meaningful trends.
learning models to predict stock prices, identify trading signals, and DL models tend to fit the noise in financial time-series data, which
optimize trading decisions. While these studies show promising results, can lead to overfitting [212]. The literature highlights that despite the
they also face challenges such as model overfitting, data quality issues, use of advanced regularization techniques such as dropout and weight
and the adaptability of models to different market conditions.
decay, overfitting remains a persistent issue when models are trained
Table 15 provides a comprehensive comparison of several studies re-
on noisy or non-stationary data.
lated to deep learning for algorithmic trading, summarizing key aspects
such as datasets used, models employed, results achieved, contributions
made, and limitations identified. 6.1.2. Missing data
Missing data is another critical issue, especially in high-frequency
6. Challenges identified trading, where gaps in data can arise due to network latency, exchange
outages, or reporting delays [213,214]. DL models often require com-
Despite the considerable advancements in deep learning for fi- plete datasets for accurate predictions, and missing data points can
nancial applications, these challenges underscore the complexities in- significantly degrade model performance. Imputation techniques are
volved in developing robust models for algorithmic trading [206,207]. commonly used to address this issue, but they introduce their own

18
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Table 14 Here, 𝑥̂ 𝑡 represents the imputed value for time 𝑡 based on the mean
Summary of datasets used in deep learning for algorithmic trading. of the previous 𝑁 observations. While simple imputation methods such
Name Description as this can address missing values, they may fail to capture the intricate
The Astock dataset, focused on the China A-shares market, market dynamics and dependencies, leading to inaccurate predictions.
Astock [184] contains 40,963 stock-specific news items (July
2018–November 2021) annotated with trading actions. It
6.2. Overfitting
features 24 stock factors, is divided into in-distribution and
out-of-distribution splits, and supports stock movement
classification and simulated trading, using relevant financial Overfitting happens when a model excels at the training data but
evaluation metrics. struggles to apply what it learned to new, unseen data [216,217].
Stock Market The dataset contains historical daily prices of This is a major challenge in the application of DL to algorithmic
Dataset [185] NASDAQ-traded stocks and ETFs, retrieved via Yahoo trading. The inherent complexity of DL models, such as RNN and LSTM
Finance using the yfinance Python package, up to April 1, networks, often leads to overfitting, particularly in financial markets
2020. Each ticker’s data is stored in CSV format with fields
characterized by high volatility and unpredictability [218,219].
like date, open, high, low, close, adjusted close, and volume,
organized into ETFs and stock folders.
6.2.1. Impact of overfitting in trading
NIFTY-50 Stock The NIFTY-50 dataset contains daily stock price history and
Market Data [186] trading volumes for fifty stocks in India’s NIFTY-50 index Overfitting can result in models that are highly sensitive to historical
from January 1, 2000, to April 30, 2021. It includes market conditions but ineffective in capturing future trends [220,221].
individual CSV files for each stock and metadata with In algorithmic trading, this can manifest as poor real-time performance,
macro-level information, updated monthly for ongoing where models perform exceptionally well on backtested historical data
analysis.
but fail to maintain profitability in live trading. This challenge is
StockNet Dataset The Stocknet dataset includes two years (2014–2016) of intensified by the ever-changing nature of financial markets, where
[187] stock price movements for 88 stocks across 9 sectors, along
with Twitter data. It features both raw and preprocessed
historical performance does not guarantee future results.
price data from Yahoo Finance and tweet data from Twitter, Various methods have been suggested in the literature to address
enabling stock movement prediction using text and price overfitting, including:
signals.
datasets of stock The dataset tracks the performance of eight stock market • Dropout: A regularization technique where random neurons are
market indices indices (IPC, S&P 500, DAX, DJIA, FTSE, N225, NDX, CAC) dropped during training to prevent the model from becoming too
[188] from June 2006 to May 2023. It includes daily stock indices reliant on specific features. The dropout rate 𝛾 is set between 0
and FX rates, quarterly GDP, and monthly CPI, RFR, UR, and 1, controlling the proportion of neurons to drop:
BOP, and GFCF data, sourced from Yahoo Finance and OECD. {
Stock-Market The Stock-Market Sentiment Dataset comprises tweets (𝑙) 0, with probability 𝛾
ℎ𝑖 = (𝑙) (18)
Sentiment Dataset collected from multiple Twitter handles, categorizing ℎ𝑖 , with probability 1 − 𝛾
[189] economic news into positive (1) and negative (−1)
sentiments. It contains 3,685 positive and 2,106 negative where ℎ(𝑙)
𝑖 is the activation of neuron 𝑖 at layer 𝑙.
entries, organized into two columns: Text and Sentiment.
This dataset serves as an insightful resource for sentiment • Cross-Validation: K-fold cross-validation is commonly used to
analysis in stock trading. assess a model’s generalization capabilities. This method involves
Stock Tweets for The dataset comprises over 80,000 tweets related to the top dividing the dataset into 𝑘 subsets, where the model is trained
Sentiment Analysis 25 stock tickers on Yahoo Finance, collected between on 𝑘 − 1 subsets and evaluated on the remaining subset. This
and Prediction September 30, 2021, and September 30, 2022. It includes procedure is repeated 𝑘 times, and the model’s performance is
[190] tweet texts, stock names, company names, and corresponding
averaged across all folds.
stock market price and volume data, facilitating sentiment
1∑
analysis and stock price prediction. 𝑘

Dataset for Stock The dataset comprises historical stock prices for three MSECV = MSE𝑖 (19)
𝑘 𝑖=1
Market Prediction petroleum companies: Pakistan State Oil (PSO), Hascol, and
[191] Attock Petroleum Limited (APL), extracted from the Pakistan
Despite the use of such techniques, overfitting remains a critical
Stock Exchange (PSX) website over the last four years. It
includes daily attributes like total trade volume, high, low,
issue, particularly in high-frequency and short-term trading strategies,
open, and close prices. Additionally, Twitter data for where the noise in the data often dominates the underlying market
sentiment analysis was collected using Twint and Tweepy, patterns.
capturing user-profiles and tweet attributes to calculate a
composite influence score.
6.3. Model interpretability
United States Stock The dataset captures US stock market activity, detailing
Market Index [192] changes in major indices (Dow, S&P 500, Nasdaq) and sector
Another major challenge identified in the literature is the inter-
performance. It includes information on oil price fluctuations
following geopolitical events, US services sector expansion,
pretability of DL models. DL models, particularly DNNs, are often
labor market indicators, and significant stock movements, referred to as ‘‘black boxes’’ due to their complexity and the difficulty in
such as Spirit Airlines’ plunge amid bankruptcy talks. understanding how they arrive at specific predictions [222,223]. This
CryptoBubbles The CryptoBubbles dataset includes around 404 lack of transparency can be problematic in algorithmic trading, where
dataset [193] cryptocurrencies from top exchanges, covering daily price understanding the rationale behind a model’s decisions is important for
data from March 1, 2016, to April 7, 2021. It features gaining trust from traders and regulators [224,225].
essential price metrics and 2.4 million tweets for bubble
detection using the PSY model, allowing analysis of market
behaviors and trends, despite inherent challenges.
6.3.1. Black-box nature of DL models
In traditional ML models such as decision trees or linear regression,
the decision-making process is more transparent because the model
structure can be easily interpreted [226,227]. However, in deep learn-
challenges, particularly in terms of preserving the temporal dynamics ing models, particularly those with multiple layers and a high number
of financial data [215]. of parameters, it is challenging to discern which features are driving
the model’s predictions.
1 ∑
𝑁
The literature proposes several methods for improving interpretabil-
𝑥̂ 𝑡 = 𝑥 (17)
𝑁 𝑖=1 𝑡−𝑖 ity, including:

19
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Table 15
Comparison of recent studies on deep learning for algorithmic trading.
Ref Year Model Dataset Result Contribution Limitation
[144] 2024 CNN Turkish tweet Accuracy This paper presents a sentiment analysis The paper points out the potential bias in data
dataset [203] 83.02% (Binary) framework for Turkish financial tweets utilizing collection caused by the use of specific
& 72.73% deep learning models such as CNN, LSTM, and keywords, which could distort sentiment
(Multi-Class) GRU-CNN. The CNN model, enhanced with representation. It also highlights the challenges
pre-trained word embeddings, achieved the posed by the ambiguity of Turkish tweets,
best performance, with an accuracy of 83.02% which complicates preprocessing and affects
for binary classification and 72.73% for model performance. Furthermore, the research
multi-class classification, demonstrating the mentions the limited amount of sentiment
effectiveness of deep learning in sentiment analysis conducted in Turkish compared to
analysis for resource-constrained languages. English, suggesting this is an area for future
exploration.
[158] 2023 LSTM Stock Price Accuracy 95% This paper enhances predictive analysis and This paper highlights the difficulties of using
Dataset EOD risk management in trading by combining machine learning for stock market predictions,
[204] LSTM networks with collaborative filtering, particularly when working with time series
improving stock market forecasts. Evaluated on data, as the results may lack reliability. While
the Stock Price EOD Dataset, the model deep learning models have shown potential,
achieves an accuracy of over 95%, surpassing further research is needed to explore additional
traditional benchmarks and demonstrating the algorithms and datasets to improve the
effectiveness of deep learning in practical robustness and applicability of the proposed
trading applications. models.
[57] 2024 LSTM BEXIMCO RMSE 12.312 & This paper evaluates the performance of SMA, The paper emphasizes LSTM’s limitations in
Pharmaceuti- MAPE 2.06% EMA, and LSTM models for stock price long-term stock price prediction, where SMA
cal [205] prediction. It shows that LSTM outperforms the and EMA yield better results. It also
other models in short-term forecasting, acknowledges the intrinsic complexity of the
achieving an RMSE of 12.312 and a MAPE of stock market, which impacts the reliability of
2.06%, whereas SMA and EMA provide better predictions, and highlights the limited
results for long-term predictions. applicability of the findings due to the use of
datasets from only six companies.
[125] 2024 GRA-WD- Own Dataset Accuracy This paper presents a GRA-WD-BiLSTM hybrid This study’s limitations include restrictions in
BiLSTM 95.93% (SSEC), model for forecasting stock prices based on selecting environmental data due to equipment
93.02% (SZI) & environmental factors such as air quality and and data availability, which limited the range
97.07% (HSI) weather. The model demonstrated strong of influencing factors considered. Furthermore,
prediction accuracies of 95.93% for the SSEC, it did not incorporate seasonal variations in
93.02% for the SZI, and 97.07% for the HSI, environmental factors, which could impact the
highlighting its effectiveness across different accuracy of stock price predictions.
stock indices.
[118] 2024 LLM Own Dataset Accuracy 74.4% This paper investigates the application of large This paper discusses the difficulties in using
language models (LLMs) such as OPT, BERT, LLMs for financial analysis, emphasizing the
and FinBERT for financial sentiment analysis, necessity for technical skills, computational
showing their enhanced effectiveness in power, and funding. It also recognizes the
predicting market trends. The OPT model, in challenges in interpreting textual data and
particular, reached a 74.4% accuracy in proposes that further research is required to
forecasting stock returns and yielded a 355% investigate alternative models and improve
return using a long-short strategy between methods for analyzing financial texts.
2021 and 2023.
[126] 2023 BiLSTM- Shanghai and MAE 0.087, MSE This paper introduces an innovative Although the model improves prediction
MTRAN- TCN Shenzhen 0.014, RMSE BiLSTM-MTRAN-TCN model for stock price accuracy, it may face difficulties in fully
stock markets 0.118, & 𝑅2 forecasting, enhancing prediction accuracy and capturing the complexities of market dynamics
Dataset 0.986 stability compared to conventional approaches. due to the inherent unpredictability of financial
It achieves an increase in 𝑅2 ranging from markets. Furthermore, the paper does not
0.3% to 15.6% and a reduction in RMSE provide an in-depth analysis of the
between 24.3% and 93.5%, showing excellent computational costs and scalability of the
generalization across various stock indices and proposed approach, which raises concerns
time frames. about its practicality in real-time applications.

• Feature Importance: Techniques such as SHAP (SHapley Addi- highlight the importance of different time steps in contributing to
tive exPlanations) [228] values and LIME (Local Interpretable a prediction:
Model-agnostic Explanations) [229] can be used to assign impor-
tance scores to features, providing insight into how the model exp(𝑒𝑡 )
𝛼𝑡 = ∑𝑇 (21)
uses input data to make predictions. SHAP values are computed 𝑡=1 exp(𝑒𝑡 )
as:
∑ |𝑆|!(|𝑁| − |𝑆| − 1)! where 𝛼𝑡 is the attention weight assigned to time step 𝑡, and 𝑒𝑡
𝜙𝑖 = [𝑓 (𝑆 ∪ {𝑖}) − 𝑓 (𝑆)] (20) is the attention score. These mechanisms help to identify which
𝑆⊆𝑁⧵{𝑖}
|𝑁|!
market events or time periods are most relevant to the model’s
where 𝜙𝑖 is the Shapley value for feature 𝑖, 𝑆 is a subset of predictions.
features, and 𝑁 is the total set of features. This approach offers
a way to decompose the model’s output into contributions from
However, despite advances in interpretability techniques, the chal-
each feature.
• Attention Mechanisms: Attention mechanisms, commonly used lenge of making complex DL models transparent to traders and regula-
in NLP tasks, have been adapted for use in time-series financial tors remains a significant barrier to their broader adoption in algorith-
data to improve model interpretability [230]. Attention weights mic trading.

20
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

6.4. Computational complexity 6.5.2. High dimensionality of financial data


Financial markets generate massive amounts of data, including price
data, order book information, economic indicators, and social media
Deep learning models are computationally intensive, particularly
sentiment. The high dimensionality of this data poses a challenge for
when applied to high-frequency trading or large-scale financial datasets
DL models, which must learn to filter out irrelevant features while
[231,232]. The training process for deep neural networks can take
capturing the most important ones [242,243]. Dimensionality reduc-
considerable time and resources, especially for architectures such as tion techniques, such as autoencoders and principal component analysis
LSTMs and CNNs, which are often used in financial markets. (PCA), are commonly used to address this issue, but they can also result
in the loss of important information [244,245].
6.4.1. Training time and resources 𝑋̂ = 𝑊 ⋅ 𝑋 (23)
Training deep models on financial data often requires substan-
In the equation above, 𝑋̂ represents the lower-dimensional repre-
tial computational resources, including high-performance GPUs or dis-
sentation of the original data 𝑋, and 𝑊 is the transformation matrix
tributed computing environments [233,234]. This can be a limiting
used for dimensionality reduction. While this approach helps to sim-
factor for smaller institutions or individual traders who may not have
plify the model and reduce computational complexity, it also introduces
access to such infrastructure. In addition to the cost, long training times the risk of discarding valuable information that could improve the
can also delay the deployment of models in live trading environments, model’s performance.
where the ability to quickly adapt to market conditions is critical.
𝑛×𝑑 7. Discussion and future research directions
𝑇train = (22)
𝑟
This review highlights the transformative potential of deep learning
In this equation, 𝑇train represents the total training time, 𝑛 is the models in algorithmic trading and sets the stage for future exploration
number of samples in the dataset, 𝑑 is the model complexity (e.g., the in this dynamic field. As researchers and practitioners continue to
number of layers and parameters), and 𝑟 is the computational resources refine these methodologies, ongoing collaboration between academia
available (e.g., GPU speed). As 𝑛 and 𝑑 increase, the training time and industry will be essential to address practical challenges and drive
grows, requiring more powerful computational setups. innovation.

7.1. Interpretation of findings


6.4.2. Real-time trading and latency
In high-frequency trading, real-time performance is crucial, and This review has synthesized the use of DL models in algorithmic
even small delays in executing trades can lead to missed opportunities trading, revealing several important trends and gaps. First, DL models
and financial losses. The computational complexity of DL models, such as LSTM, CNN, and RL agents dominate the landscape of finan-
particularly those that involve deep architectures or ensemble meth- cial predictions due to their ability to handle non-linear patterns and
ods, can introduce latency in decision-making [235]. This latency is high-dimensional data. LSTMs, in particular, have demonstrated strong
problematic in environments where trades need to be executed within capabilities in time-series forecasting, a crucial task in financial mar-
milliseconds. kets. Reinforcement Learning has also gained traction, particularly in
autonomous trading strategies, where its ability to adapt and optimize
Efforts to reduce latency often involve simplifying the model archi-
in dynamic environments has shown promise [246].
tecture or using techniques such as model pruning and quantization
However, this synthesis also highlights some significant gaps in
to reduce the computational demands of DL models. However, there the literature. One key issue is the over-reliance on historical data,
is often a trade-off between model complexity (and accuracy) and the which may not always capture future market dynamics due to the non-
speed of execution [236]. stationary nature of financial markets [247]. The lack of attention to
market complexity, including the influence of geopolitical events, social
6.5. Market complexity sentiment, and macroeconomic indicators, suggests that current models
are often too narrowly focused [248]. Furthermore, while many studies
report impressive results in back-testing, real-world performance often
Financial markets are shaped by a wide range of factors, such as falls short due to challenges such as data noise, overfitting, and model
macroeconomic trends, geopolitical developments, and market senti- interpretability.
ment, which contribute to their complexity and dynamism. Capturing
these interactions in DL models is challenging, as markets are not 7.2. Theoretical and practical implications
governed by fixed rules but by a multitude of interacting variables,
many of which are difficult to quantify [237,238]. Theoretically, the findings of this review contribute to a deeper
understanding of the potential of DL models in financial markets.
DL architectures like LSTMs and CNNs offer substantial theoretical
6.5.1. Non-stationarity of financial data advantages over traditional machine learning (ML) models by capturing
The main challenge in modeling financial markets is the non- complex dependencies and providing more accurate predictions [48].
stationarity of financial data. Unlike static datasets, financial markets However, the black-box nature of these models raises significant con-
evolve over time, with changes in market structure, trading behavior, cerns regarding their interpretability, particularly in regulated financial
and external factors [239,240]. This makes it difficult for DL models environments where transparency is critical [249].
trained on historical data to generalize to future market conditions. From a practical standpoint, the integration of DL in algorithmic
Traditional statistical models often assume stationarity, but this as- trading offers traders and financial institutions powerful tools to en-
hance decision-making, optimize trading strategies, and reduce human
sumption is frequently violated in financial markets, leading to poor
error [31]. Traders and developers of trading algorithms can benefit
model performance [241].
from DL models’ ability to automate complex trading strategies and
To address this issue, some studies have proposed adaptive learning continuously learn from new data. However, the practical adoption of
methods, where the model parameters are updated as new market data these models is limited by challenges such as high computational costs,
becomes available. Reinforcement learning, in particular, has shown data quality issues, and the need for model explainability. Moreover,
promise in this area, as it allows models to continuously learn from financial institutions must also navigate regulatory challenges when
new market data and adjust their strategies in real-time [165]. deploying such advanced models in real-world trading environments.

21
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

7.3. Limitations of the study explain the decisions of complex DL models in a manner that is acces-
sible to traders, regulators, and other stakeholders. The development
Several limitations in this review must be acknowledged. First, of interpretable DL architectures, such as explainable reinforcement
while an extensive search strategy was employed, the review may learning models, could also help address regulatory concerns about
still suffer from publication bias, as studies with negative results are black-box models.
often underreported. Furthermore, this review mainly concentrated on Finally, addressing the computational demands of deep learning in
studies published in English, which may have led to the exclusion financial markets will be essential for broader adoption. Research into
of important research published in other languages. Another limita- model compression techniques, such as pruning and quantization, could
tion is the varying quality of the selected studies; while most studies reduce the computational cost of deploying DL models in real-time
applied rigorous methodologies, some lacked transparency regarding trading systems. Additionally, advances in distributed computing and
their model training and evaluation processes. Finally, this review was cloud-based infrastructure could provide more accessible resources for
constrained by the available literature between 2018–2024, which may training and deploying DL models at scale.
omit other advancements in deep learning and algorithmic trading.
8. Conclusion
7.4. Emerging trends in DL for financial prediction
This review has highlighted the transformative potential of deep
Recent advancements in deep learning are opening new avenues
learning in algorithmic trading, where models such as LSTM, CNN,
for improving financial predictions. One of the most promising devel-
and Reinforcement Learning have shown substantial improvements in
opments is the rise of attention mechanisms and transformer archi-
predicting financial markets and optimizing trading strategies. How-
tectures, which have shown remarkable success in NLP and are now
ever, significant challenges remain, particularly related to data quality,
being adapted for financial time-series forecasting [250]. Attention
overfitting, and the interpretability of complex DL models. Financial
mechanisms allow models to focus on the most relevant parts of the
markets are noisy, volatile, and influenced by a multitude of factors,
input data, enabling them to better capture important market trends
and events. Transformers, with their self-attention capabilities, offer making it difficult for models to generalize well. Additionally, the
significant advantages in modeling long-term dependencies without the black-box nature of DL models raises concerns for traders and regu-
need for sequential processing, as seen in recurrent architectures [63, lators who require transparency in decision-making. Emerging trends
251]. such as attention mechanisms, transformer architectures, and hybrid
Another emerging trend is the use of hybrid models that com- models offer promising solutions to these challenges, alongside inte-
bine multiple DL techniques. For example, hybrid models that in- grating alternative data sources like social media sentiment and news.
tegrate LSTMs with CNNs or Reinforcement Learning have demon- Future research must focus on improving model robustness, developing
strated superior performance by leveraging the strengths of each ar- explainable AI techniques, and addressing computational efficiency to
chitecture [178]. These models can capture both short-term and long- unlock the full potential of DL in real-world trading environments. By
term market trends, providing more robust predictions and trading overcoming these hurdles, DL can significantly enhance the accuracy
strategies. and effectiveness of algorithmic trading, providing traders with more
Moreover, there is a growing interest in integrating external data powerful tools for navigating complex financial markets.
sources such as social media sentiment, news articles, and macroeco-
nomic indicators into DL models. This approach, often referred to as CRediT authorship contribution statement
‘‘alternative data’’, has the potential to improve prediction accuracy
by providing a more holistic view of market dynamics [252]. These MD Shahriar Mahmud Bhuiyan: Writing – original draft, Method-
data sources can be processed using NLP techniques such as BERT ology, Formal analysis, Conceptualization. MD AL Rafi: Writing – orig-
(Bidirectional Encoder Representations from Transformers) [253] to inal draft, Methodology, Formal analysis, Conceptualization. Gourab
gauge market sentiment and anticipate the impact of news on asset Nicholas Rodrigues: Writing – original draft, Investigation, Formal
prices. analysis, Conceptualization. MD Nazmul Hossain Mir: Visualization,
Validation, Resources, Methodology, Formal analysis, Conceptualiza-
7.5. Addressing current challenges tion. Adit Ishraq: Writing – review & editing, Visualization, Validation,
Methodology, Investigation. M.F. Mridha: Writing – review & editing,
Several challenges discussed in this review require attention in Supervision, Methodology. Jungpil Shin: Writing – review & editing,
future research. One of the most pressing issues is the need for im- Visualization, Validation, Methodology.
proved data preprocessing techniques to handle noise and missing
data more effectively [254]. Advanced data imputation methods, such
Declaration of competing interest
as generative models like Variational Autoencoders (VAEs), can help
in reconstructing missing data without introducing bias. Additionally,
The authors declare the following financial interests/personal rela-
techniques like data augmentation and denoising can be further ex-
tionships which may be considered as potential competing interests:
plored to improve the robustness of DL models against noisy financial
M. F. Mridha reports was provided by American International Univer-
data [255,256].
sity Bangladesh. M. F. Mridha reports a relationship with American
Overfitting remains a persistent challenge in DL-based algorithmic
International University Bangladesh that includes: employment. If there
trading, particularly when models are trained on limited or historical
data [30]. Future research should explore the use of more advanced are other authors, they declare that they have no known competing
regularization techniques, such as adversarial training and ensemble financial interests or personal relationships that could have appeared
methods, to improve model generalization. Moreover, developing adap- to influence the work reported in this paper.
tive learning methods, where models continuously update based on
new market data, could help address the problem of non-stationarity Acknowledgment
in financial markets.
Model interpretability is another critical area for future research. The authors sincerely thank the Advanced Machine Intelligence
While techniques like SHAP [228] values and LIME [229] have been Research Lab (AMIR Lab) for its continuous support and instructions
proposed, there is still a need for more sophisticated tools that can in fulfilling their goals.

22
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

Data availability [28] Rahmani AM, Rezazadeh B, Haghparast M, Chang W-C, Ting SG. Applications
of artificial intelligence in the economy, including applications in stock trading,
market analysis, and risk management. IEEE Access 2023.
No data was used for the research described in the article.
[29] Pathak D, Aishwarya APS, Dhama P, Singh S. Technological paradigms reshap-
ing. In: Role of emerging technologies in social science. Cambridge Scholars
Publishing; 2024, p. 186.
References [30] Ozbayoglu AM, Gudelek MU, Sezer OB. Deep learning for financial applications:
A survey. Appl Soft Comput 2020;93:106384.
[1] Addy WA, Ajayi-Nifise AO, Bello BG, Tula ST, Odeyem O, Falaiye T. Algorithmic [31] Singh V, Chen S-S, Singhania M, Nanavati B, Gupta A, et al. How are
trading and AI: A review of strategies and market impact. World J Adv Eng reinforcement learning and deep learning algorithms used for big data based
Technol Sci 2024;11(1):258–67. decision making in financial industries–A review and research agenda. Int J Inf
[2] Pasqual M. High Frequency Trading: an analysis of the phenomenon and the Manag Data Insights 2022;2(2):100094.
effect of the human component. Università Ca’Foscari Venezia; 2020. [32] Salehpour A, Samadzamini K. Machine learning applications in algorith-
[3] Zaharudin KZ, Young MR, Hsu W-H. High-frequency trading: Definition, mic trading: a comprehensive systematic review. Int J Educ Manag Eng
implications, and controversies. J Econ Surv 2022;36(1):75–107. 2023;13(6):41.
[4] Verbist J, Petitjean M. High Frequency Trading guillemotleft What do we know [33] Sarker IH. Machine learning: Algorithms, real-world applications and research
and where do we need to go?. directions. SN Comput Sci 2021;2(3):160.
[5] Najem R, Bahnasse A, Talea M. Toward an enhanced stock market forecast- [34] Biju AKVN, Thomas AS, Thasneem J. Examining the research taxonomy
ing with machine learning and deep learning models. Procedia Comput Sci of artificial intelligence, deep learning & machine learning in the financial
2024;241:97–103. sphere—A bibliometric analysis. Qual Quant 2024;58(1):849–78.
[6] Sonkavde G, Dharrao DS, Bongale AM, Deokate ST, Doreswamy D, Bhat SK. [35] Ayitey Junior M, Appiahene P, Appiah O, Bombie CN. Forex market forecasting
Forecasting stock market prices using machine learning and deep learning mod- using machine learning: Systematic literature review and meta-analysis. J Big
els: A systematic review, performance analysis and discussion of implications. Data 2023;10(1):9.
Int J Financ Stud 2023;11(3):94. [36] Rezaei A, Abdellatif I, Umar A. Towards economic sustainability: A com-
[7] Khairy M, Mahmoud TM, Omar A, Abd El-Hafeez T. Comparative performance prehensive review of artificial intelligence and machine learning techniques
of ensemble machine learning for Arabic cyberbullying and offensive language in improving the accuracy of stock market movements. Int J Financ Stud
detection. Lang Resour Eval 2024;58(2):695–712. 2025;13(1):28.
[8] Sharifani K, Amini M. Machine learning and deep learning: A review of methods [37] Cohen G. Algorithmic trading and financial forecasting using advanced artificial
and applications. World Inf Technol Eng J 2023;10(07):3897–904. intelligence methodologies. Mathematics 2022;10(18):3302.
[9] Sengupta S, Basak S, Saikia P, Paul S, Tsalavoutis V, Atiah F, et al. A review of [38] Joiner D, Vezeau A, Wong A, Hains G, Khmelevsky Y. Algorithmic trading and
deep learning with special emphasis on architectures, applications and recent short-term forecast for financial time series with machine learning models; state
trends. Knowl-Based Syst 2020;194:105596. of the art and perspectives. In: 2022 IEEE international conference on recent
[10] Zhang D, Yin C, Zeng J, Yuan X, Zhang P. Combining structured and unstruc- advances in systems science and engineering. IEEE; 2022, p. 1–9.
tured data for predictive models: a deep learning approach. BMC Med Inform [39] Ortiz AFH, Camacho EC, Rojas JC, Camacho TC, Guevara SZ, Cuenca NTR, et al.
Decis Mak 2020;20:1–11. A practical guide to perform a systematic literature review and meta-analysis.
[11] Jain S. A comparative study of stock market prediction models: Deep learning Princ Pr Clin Res 2021;7(4):47–57.
approach and machine learning approach. In: Deep learning tools for predicting [40] García-Holgado A, Marcos-Pablos S, García-Peñalvo F. Guidelines for performing
stock market movements. Wiley Online Library; 2024, p. 249–69. systematic research projects reviews. Int J Interact Multim Artif Intell 2020.
[12] Singh T, Kalra R, Mishra S, Satakshi, Kumar M. An efficient real-time [41] Rethlefsen ML, Kirtley S, Waffenschmidt S, Ayala AP, Moher D, Page MJ, et
stock prediction exploiting incremental learning and deep learning. Evol Syst al. PRISMA-S: an extension to the PRISMA statement for reporting literature
2023;14(6):919–37. searches in systematic reviews. Syst Rev 2021;10:1–19.
[13] Bisong E, Bisong E. Recurrent neural networks (RNNs). In: Building machine [42] Lotti L. Fundamentals of algorithmic markets: liquidity, contingency, and the
learning and deep learning models on google cloud platform: A comprehensive incomputability of exchange. Philos Technol 2018;31:43–58.
guide for beginners. Springer; 2019, p. 443–73. [43] Hansen KB. The virtue of simplicity: On machine learning models in algorithmic
[14] Hochreiter S. Long short-term memory. In: Neural computation. MIT-Press; trading. Big Data & Soc 2020;7(1):2053951720926558.
1997. [44] Golub A, Glattfelder JB, Olsen RB. The alpha engine: Designing an automated
[15] Ketkar N, Moolayil J, Ketkar N, Moolayil J. Convolutional neural networks. In: trading algorithm. In: High-performance computing in finance. Chapman and
Deep learning with Python: Learn best practices of deep learning models with Hall/CRC; 2018, p. 49–76.
PyTorch. Springer; 2021, p. 197–242. [45] Pothumsetty R. Application of artificial intelligence in algorithmic trading. Int
[16] Thakkar A, Chaudhari K. A comprehensive survey on portfolio optimization, J Eng Appl Sci Technol 2020;4(12):140–9.
stock price and trend prediction using particle swarm optimization. Arch [46] Velu R. Algorithmic trading and quantitative strategies. Chapman and
Comput Methods Eng 2021;28(4):2133–64. Hall/CRC; 2020.
[17] Thakkar A, Chaudhari K. Fusion in stock market prediction: a decade survey on [47] Syamala SR, Wadhwa K. Trading performance and market efficiency: Evidence
the necessity, recent developments, and potential future directions. Inf Fusion from algorithmic trading. Res Int Bus Financ 2020;54:101283.
2021;65:95–107. [48] Ahmed SF, Alam MSB, Hassan M, Rozbu MR, Ishtiak T, Rafa N, et al. Deep
[18] Atkins A, Niranjan M, Gerding E. Financial news predicts stock market volatility learning modelling techniques: current progress, applications, advantages, and
better than close price. J Financ Data Sci 2018;4(2):120–37. challenges. Artif Intell Rev 2023;56(11):13521–617.
[19] Idrees SM, Alam MA, Agarwal P. A prediction approach for stock market [49] Abdullayeva FJ. Convolutional neural network-based automatic diagnostic
volatility based on time series data. IEEE Access 2019;7:17287–98. system for AL-DDoS attacks detection. Int J Cyber Warf Terror (IJCWT)
[20] Razavi S. Deep learning, explained: Fundamentals, explainability, and bridge- 2022;12(1):1–15.
ability to process-based modelling. Environ Model Softw 2021;144:105159. [50] Zulqarnain M, Ghazali R, Ghouse MG, Hassim YMM, Javid I. Predicting financial
[21] Omar A, Mahmoud TM, Abd-El-Hafeez T. Comparative performance of machine prices of stock market using recurrent convolutional neural networks. Int J Intell
learning and deep learning algorithms for arabic hate speech detection in osns. Syst Appl 2020;13(6):21.
In: Proceedings of the international conference on artificial intelligence and [51] Bhardwaj A. Time series forecasting with recurrent neural networks: An
computer vision. Springer; 2020, p. 247–57. in-depth analysis and comparative study. Perform Eval 2023;2(4).
[22] Nia HM. Stock market prediction by using machine learning and deep learning [52] Lu M, Xu X. TRNN: An efficient time-series recurrent neural network for stock
approaches. 2022. price prediction. Inform Sci 2024;657:119951.
[23] Li S. Algorithmic financial regulation: Limits of computing complex adaptive [53] Zhang W, Gong X, Wang C, Ye X. Predicting stock market volatility based on
systems. Am Univ Bus Law Rev 2023;13(2). textual sentiment: A nonlinear analysis. J Forecast 2021;40(8):1479–500.
[24] Chinthapalli UR, Bommisetti RK, Kondamudi BR, Bagale G, Satyanarayana R. [54] Zhao J, Zeng D, Liang S, Kang H, Liu Q. Prediction model for stock price
Isolated stakeholders’ behavior towards fintech assisted by artificial intelligence trend based on recurrent neural network. J Ambient Intell Humaniz Comput
technology. Ann Oper Res 2021;1–27. 2021;12:745–53.
[25] Kumbure MM, Lohrmann C, Luukka P, Porras J. Machine learning techniques [55] Aldhyani TH, Alzahrani A. Framework for predicting and modeling stock market
and data for stock market forecasting: A literature review. Expert Syst Appl prices based on deep learning algorithms. Electronics 2022;11(19):3149.
2022;197:116659. [56] Mishra A, Tripathi K, Gupta L, Singh KP. Long short-term memory recurrent
[26] Li W. Analysis of financial market efficiency. Int J Glob Econ Manag neural network architectures for melody generation. In: Soft computing for
2024;2(2):244–53. problem solving: SocProS 2017, volume 2. Springer; 2018, p. 41–55.
[27] Frömmel M. International financial markets in the digital era. In: Digitalization [57] Billah MM, Sultana A, Bhuiyan F, Kaosar MG. Stock price prediction: compari-
and the future of financial services: innovation and impact of digital finance. son of different moving average techniques using deep learning model. Neural
Springer; 2022, p. 85–101. Comput Appl 2024;36(11):5861–71.

23
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

[58] Zhang J, Ye L, Lai Y. Stock price prediction using CNN-BiLSTM-Attention model. [88] Baltakys K, Baltakienė M, Heidari N, Iosifidis A, Kanniainen J. Predicting
Mathematics 2023;11(9):1985. the trading behavior of socially connected investors: Graph neural net-
[59] Gülmez B. Stock price prediction with optimized deep LSTM network with work approach with implications to market surveillance. Expert Syst Appl
artificial rabbits optimization algorithm. Expert Syst Appl 2023;227:120346. 2023;228:120285.
[60] Xu K, Wu Y, Jiang M, Sun W, Yang Z. Hybrid LSTM-GARCH framework [89] Wang H, Wang T, Li S, Zheng J, Guan S, Chen W. Adaptive long-short pattern
for financial market volatility risk prediction. J Comput Sci Softw Appl transformer for stock investment selection. In: IJCAI. 2022, p. 3970–7.
2024;4(5):22–9. [90] Lezmi E, Xu J. Time series forecasting with transformer models and application
[61] Kim J, Oh S, Kim H, Choi W. Tutorial on time series prediction using 1D-CNN to asset management. 2023, Available at SSRN 4375798.
and BiLSTM: A case example of peak electricity demand and system marginal [91] Tran DT, Iosifidis A, Kanniainen J, Gabbouj M. Temporal attention-augmented
price prediction. Eng Appl Artif Intell 2023;126:106817. bilinear network for financial time-series data analysis. IEEE Trans Neural
[62] Lee SW, Kim HY. Stock market forecasting with super-high dimensional Networks Learn Syst 2018;30(5):1407–18.
time-series data using ConvLSTM, trend sampling, and specialized data [92] Huang S, Wang D, Wu X, Tang A. Dsanet: Dual self-attention network for mul-
augmentation. Expert Syst Appl 2020;161:113704. tivariate time series forecasting. In: Proceedings of the 28th ACM international
[63] Bao W, Cao Y, Yang Y, Che H, Huang J, Wen S. Data-driven stock forecasting conference on information and knowledge management. 2019, p. 2129–32.
models based on neural networks: A review. Inf Fusion 2024;102616. [93] Bilokon P, Qiu Y. Transformers versus LSTMs for electronic trading. 2023, arXiv
[64] Palaniappan V, Ishak I, Ibrahim H, Sidi F, Zukarnain ZA. A review on high preprint arXiv:2309.11400.
frequency trading forecasting methods: Opportunity and challenges for quantum [94] Zhang Q, Qin C, Zhang Y, Bao F, Zhang C, Liu P. Transformer-based attention
based method. IEEE Access 2024. network for stock movement prediction. Expert Syst Appl 2022;202:117239.
[65] Hoseinzade E, Haratizadeh S. CNNpred: CNN-based stock market prediction [95] Tang Z, Huang J, Rinprasertmeechai D. Period-aggregated transformer for
using a diverse set of variables. Expert Syst Appl 2019;129:273–85. learning latent seasonalities in long-horizon financial time series. PloS One
[66] Lu W, Li J, Wang J, Qin L. A CNN-BiLSTM-AM method for stock price 2024;19(8):e0308488.
prediction. Neural Comput Appl 2021;33(10):4741–53. [96] Felizardo LK, Paiva FCL, de Vita Graves C, Matsumoto EY, Costa AHR, Del-
[67] Chung H, Shin K-s. Genetic algorithm-optimized multi-channel convolu- Moral-Hernandez E, et al. Outperforming algorithmic trading reinforcement
tional neural network for stock market prediction. Neural Comput Appl learning systems: A supervised approach to the cryptocurrency market. Expert
2020;32(12):7897–914. Syst Appl 2022;202:117259.
[68] Mehtab S, Sen J. Stock price prediction using CNN and LSTM-based deep [97] Hirchoua B, Ouhbi B, Frikh B. Deep reinforcement learning based trading
learning models. In: 2020 international conference on decision aid sciences and agents: Risk curiosity driven learning for financial rules-based policy. Expert
application. IEEE; 2020, p. 447–53. Syst Appl 2021;170:114553.
[69] Pinaya WHL, Vieira S, Garcia-Dias R, Mechelli A. Autoencoders. In: Machine [98] Khan A, Farooq O. Algorithmic trading strategies: Enhancing performance with
learning. Elsevier; 2020, p. 193–208. reinforcement learning techniques. Balt Multidiscip Res Lett J 2024;1(1):35–41.
[70] Wickramasinghe CS, Marino DL, Manic M. ResNet autoencoders for unsuper- [99] Li Y, Zheng W, Zheng Z. Deep robust reinforcement learning for practical
vised feature learning from high-dimensional data: Deep models resistant to algorithmic trading. IEEE Access 2019;7:108014–22.
performance degradation. IEEE Access 2021;9:40511–20. [100] Huang Y, Wan X, Zhang L, Lu X. A novel deep reinforcement learning
[71] Soleymani F, Paquet E. Financial portfolio optimization with online deep framework with BiLSTM-attention networks for algorithmic trading. Expert Syst
reinforcement learning and restricted stacked autoencoder—DeepBreath. Expert Appl 2024;240:122581.
Syst Appl 2020;156:113456. [101] Taherdoost H. Deep learning and neural networks: Decision-making implica-
[72] Zhang C, Liang S, Lyu F, Fang L. Stock-index tracking optimization using tions. Symmetry 2023;15(9):1723.
auto-encoders. Front Phys 2020;8:388. [102] Goel A, Goel AK, Kumar A. The role of artificial neural network and machine
[73] Bao W, Yue J, Rao Y. A deep learning framework for financial time se- learning in utilizing spatial information. Spat Inf Res 2023;31(3):275–85.
ries using stacked autoencoders and long-short term memory. PloS One [103] Chowdhury MS, Nabi N, Rana MNU, Shaima M, Esa H, Mitra A, et al. Deep
2017;12(7):e0180944. learning models for stock market forecasting: A comprehensive comparative
[74] Jung G, Choi S-Y. Forecasting foreign exchange volatility using deep learning analysis. J Bus Manag Stud 2024;6(2):95–9.
autoencoder-LSTM techniques. Complexity 2021;2021(1):6647534. [104] Khalil F, Pipa G. Is deep-learning and natural language processing transcending
[75] Pinheiro Cinelli L, Araújo Marins M, Barros da Silva EA, Lima Netto S. the financial forecasting? Investigation through lens of news analytic process.
Variational autoencoder. In: Variational methods for machine learning with Comput Econ 2022;60(1):147–71.
applications to deep networks. Springer; 2021, p. 111–49. [105] Chopra R, Sharma GD. Application of artificial intelligence in stock market
[76] Ehrhardt J, Wilms M. Autoencoders and variational autoencoders in medical forecasting: a critique, review, and research agenda. J Risk Financ Manag
image analysis. In: Biomedical image synthesis and simulation. Elsevier; 2022, 2021;14(11):526.
p. 129–62. [106] Chaudhary PS, Khurana MR, Ayalasomayajula M. Real-world applications of
[77] Neloy AA, Turgeon M. A comprehensive study of auto-encoders for anomaly data analytics, big data, and machine learning. In: Data analytics and machine
detection: Efficiency and trade-offs. Mach Learn Appl 2024;100572. learning: navigating the big data landscape. Springer; 2024, p. 237–63.
[78] Dogariu M, Ştefan L-D, Boteanu BA, Lamba C, Kim B, Ionescu B. Generation of [107] Shrestha YR, Krishna V, von Krogh G. Augmenting organizational decision-
realistic synthetic financial time-series. ACM Trans Multimed Comput Commun making with deep learning algorithms: Principles, promises, and challenges.
Appl (TOMM) 2022;18(4):1–27. J Bus Res 2021;123:588–603.
[79] Hosseini SA, Hosseini SA, Houshmand M. Variational autoencoder-based di- [108] Khalil M, McGough AS, Pourmirza Z, Pazhoohesh M, Walker S. Machine
mension reduction of ichimoku features for improved financial market analysis. learning, deep learning and statistical analysis for forecasting building energy
Frankl Open 2024;8:100135. consumption—A systematic review. Eng Appl Artif Intell 2022;115:105287.
[80] Liu Z, Nguyen T-K, Fang Y. Tail-gnn: Tail-node graph neural networks. In: [109] Kotios D, Makridis G, Fatouros G, Kyriazis D. Deep learning enhancing banking
Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & services: a hybrid transaction classification and cash flow prediction approach.
data mining. 2021, p. 1109–19. J Big Data 2022;9(1):100.
[81] Wu Z, Pan S, Chen F, Long G, Zhang C, Philip SY. A comprehensive [110] Long W, Gao J, Bai K, Lu Z. A hybrid model for stock price prediction based
survey on graph neural networks. IEEE Trans Neural Networks Learn Syst on multi-view heterogeneous data. Financ Innov 2024;10(1):48.
2020;32(1):4–24. [111] Wang Y, Yan G. Survey on the application of deep learning in algorithmic
[82] Li J, Chang Y, Wang Y, Zhu X. Tracking down financial statement fraud trading. Data Sci Financ Econ 2021;1(4):345–61.
by analyzing the supplier-customer relationship network. Comput Ind Eng [112] Abdullah TA, Zahid MSM, Ali W. A review of interpretable ML in health-
2023;178:109118. care: taxonomy, applications, challenges, and future directions. Symmetry
[83] Badawy A, Fisteus JA, Mahmoud TM, Abd El-Hafeez T. Topic extrac- 2021;13(12):2439.
tion and interactive knowledge graphs for learning resources. Sustainability [113] Papadimitroulas P, Brocki L, Chung NC, Marchadour W, Vermet F, Gaubert L, et
2021;14(1):226. al. Artificial intelligence: Deep learning in oncological radiomics and challenges
[84] Mamdouh Farghaly H, Abd El-Hafeez T. A high-quality feature selection method of interpretability and data harmonization. Phys Medica 2021;83:108–21.
based on frequent and correlated items for text classification. Soft Comput [114] Ghani M, Ghani U. Economic policy uncertainty and emerging stock market
2023;27(16):11259–74. volatility. Asia-Pac Financ Mark 2024;31(1):165–81.
[85] Yilmaz M, Keskin MM, Ozbayoglu AM. Algorithmic stock trading based on [115] Eliwa EHI, El Koshiry AM, Abd El-Hafeez T, Omar A. Optimal gasoline
ensemble deep neural networks trained with time graph. Appl Soft Comput price predictions: Leveraging the ANFIS regression model. Int J Intell Syst
2024;163:111847. 2024;2024(1):8462056.
[86] Chen Y, Li M, Shu M, Bi W, Xia S. Multi-modal market manipulation detection [116] Rezaei H, Faaljou H, Mansourfar G. Stock price prediction using deep learning
in high-frequency trading using graph neural networks. J Ind Eng Appl Sci and frequency decomposition. Expert Syst Appl 2021;169:114332.
2024;2(6):111–20. [117] Cui T, Du N, Yang X, Ding S. Multi-period portfolio optimization using a deep
[87] Li M, Shu M, Lu T. Anomaly pattern detection in high-frequency trading using reinforcement learning hyper-heuristic approach. Technol Forecast Soc Change
graph neural networks. J Ind Eng Appl Sci 2024;2(6):77–85. 2024;198:122944.

24
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

[118] Kirtac K, Germano G. Sentiment trading with large language models. Financ [149] Li X, Wu P, Wang W. Incorporating stock prices and news sentiments
Res Lett 2024;62:105227. for stock market prediction: A case of Hong Kong. Inf Process Manage
[119] Devan M, Thirunavukkarasu K, Shanmugam L. Algorithmic trading strategies: 2020;57(5):102212.
Real-time data analytics with machine learning. J Knowl Learn Sci Technol [150] Thormann M-L, Farchmin J, Weisser C, Kruse R-M, Säfken B, Silbersdorff A.
ISSN: 2959- 6386 (Online) 2023;2(3):522–46. Stock price predictions with LSTM neural networks and twitter sentiment. Stat
[120] Olorunnimbe K, Viktor H. Deep learning in the stock market—A sys- Optim Inf Comput 2021;9(2):268–87.
tematic survey of practice, backtesting, and applications. Artif Intell Rev [151] Pathak S, Pawar A, Taware S, Kulkarni S, Akkalkot A. A survey on machine
2023;56(3):2057–109. learning algorithms for risk-controlled algorithmic trading. Int J Sci Res Sci
[121] Singh H, Shukla AK. Analysis of stock market prediction models using deep Technol 2023;1069–89.
learning. Biosci Biotechnol Res Commun 2021;14(9):74–80. [152] Chong E, Han C, Park FC. Deep learning networks for stock market analysis
[122] Das JD, Thulasiram RK, Henry C, Thavaneswaran A. Encoder–decoder based and prediction: Methodology, data representations, and case studies. Expert Syst
LSTM and GRU architectures for stocks and cryptocurrency prediction. J Risk Appl 2017;83:187–205.
Financ Manag 2024;17(5):200. [153] Kou G, Chao X, Peng Y, Alsaadi FE, Herrera Viedma E, et al. Machine learning
[123] Zhang J, Chen X. A two-stage model for stock price prediction based on methods for systemic risk analysis in financial sectors. Vilnius Gediminas
variational mode decomposition and ensemble machine learning method. Soft Technical University Press; 2019.
Comput 2024;28(3):2385–408. [154] Choi K, Yi J, Park C, Yoon S. Deep learning for anomaly detection in time-series
[124] Mutinda JK, Langat AK. Stock price prediction using combined GARCH-AI
data: Review, analysis, and guidelines. IEEE Access 2021;9:120043–65.
models. Sci Afr 2024;e02374.
[155] Basora L, Olive X, Dubot T. Recent advances in anomaly detection methods
[125] Liu B, Pei J, Yu Z. Stock price prediction through GRA-WD-BiLSTM model with
applied to aviation. Aerospace 2019;6(11):117.
air quality and weather factors. Int J Mach Learn Cybern 2024;15(5):1967–84.
[156] Javaid HA. Ai-driven predictive analytics in finance: Transforming risk
[126] Wang S. A stock price prediction method based on BiLSTM and improved
assessment and decision-making. Adv Comput Sci 2024;7(1).
transformer. IEEE Access 2023.
[157] Goel A. Artificial Intelligence-Based Risk Assessment And Management In Stock
[127] Boudri I, El Bouhadi A. Modeling and forecasting historical volatility using
Trading.
econometric and deep learning approaches: Evidence from the Moroccan and
[158] Kavin KV. Applications of machine learning in predictive analysis and risk
Bahraini stock markets. J Risk Financ Manag 2024;17(7):300.
[128] Ge W, Lalbakhsh P, Isai L, Lenskiy A, Suominen H. Neural network– management in trading. Int J Innov Res Comput Sci Technol 2023;11(6):18–25.
based financial volatility forecasting: A systematic review. ACM Comput Surv [159] Silva TR, Li AW, Pamplona EO. Automated trading system for stock index
2022;55(1):1–30. using LSTM neural networks and risk management. In: 2020 international joint
[129] Yang R, Yu L, Zhao Y, Yu H, Xu G, Wu Y, et al. Big data analytics for financial conference on neural networks. IEEE; 2020, p. 1–8.
market volatility forecast based on support vector machine. Int J Inf Manage [160] Chandrinos SK, Sakkas G, Lagaros ND. AIRMS: A risk management tool using
2020;50:452–62. machine learning. Expert Syst Appl 2018;105:34–48.
[130] Abdissa BG. How can high-frequency data and automatic feature selection be [161] Lomakin N, Popov I, Shohneh A, Maramygin M, Gorbunova A. AI-system
used to improve forecasts of volatility and value-at-risk for brent crude oil of stock exchange trading robot for financial risk hedging. In: International
futures? Inland Norway University; 2024. scientific conference ‘‘far east con’’. Atlantis Press; 2020, p. 3273–82.
[131] Koo E, Kim G. A hybrid prediction model integrating garch models with a [162] Tatineni S, Mustyala A. Enhancing financial security: Data science’s role in risk
distribution manipulation strategy based on lstm networks for stock market management and fraud detection. ESP Int J Adv Comput Technology (ESP-
volatility. IEEE Access 2022;10:34743–54. IJACT) 2024;2(2):94–105.
[132] Niu Z, Demirer R, Suleman MT, Zhang H. Speculation, cross-market sen- [163] Hassan MU, Rehmani MH, Chen J. Anomaly detection in blockchain networks:
timent and the predictability of gold market volatility. J Behav Financ A comprehensive survey. IEEE Commun Surv & Tutorials 2022;25(1):289–318.
2024;25(3):278–95. [164] Eren Y, Küçükdemiral İ. A comprehensive review on deep learning approaches
[133] Gunjan A, Bhattacharyya S. A brief review of portfolio optimization techniques. for short-term load forecasting. Renew Sustain Energy Rev 2024;189:114031.
Artif Intell Rev 2023;56(5):3847–86. [165] Hatcher WG, Yu W. A survey of deep learning: Platforms, applications and
[134] Oliinyk V, Kozmenko O. Optimization of investment portfolio management. emerging research trends. IEEE Access 2018;6:24411–32.
Serbian J Manag 2019;14(2):373–87. [166] Fan C, Xiao F, Zhao Y, Wang J. Analytical investigation of autoencoder-based
[135] Niu H, Li S, Li J. MetaTrader: An reinforcement learning approach integrating methods for unsupervised anomaly detection in building energy data. Appl
diverse policies for portfolio optimization. In: Proceedings of the 31st ACM Energy 2018;211:1123–35.
international conference on information & knowledge management. 2022, p. [167] Gong D, Liu L, Le V, Saha B, Mansour MR, Venkatesh S, et al. Memorizing
1573–83. normality to detect anomaly: Memory-augmented deep autoencoder for un-
[136] Huang X, Newton DP, Platanakis E, Sutcliffe C. Single-stage portfolio supervised anomaly detection. In: Proceedings of the IEEE/CVF international
optimization with automated machine learning for M6. Int J Forecast 2024. conference on computer vision. 2019, p. 1705–14.
[137] Lin W, Taamouti A. Portfolio selection under non-gaussianity and sys- [168] Karn AL, Ateeq K, Sengan S, Gandhi I, Ravi L, Sharma DK, Subra-
temic risk: A machine learning based forecasting approach. Int J Forecast maniyaswamy V. B-lstm-Nb based composite sequence learning model for
2024;40(3):1179–88. detecting fraudulent financial activities. Malays J Comput Sci 2022;30–49.
[138] Chaweewanchon A, Chaysiri R. Markowitz mean-variance portfolio optimization
[169] Thilagavathi M, Saranyadevi R, Vijayakumar N, Selvi K, Anitha L, Sudharson K.
with predictive stock selection using machine learning. Int J Financ Stud
AI-driven fraud detection in financial transactions with graph neural networks
2022;10(3):64.
and anomaly detection. In: 2024 international conference on science technology
[139] Ma Y, Han R, Wang W. Portfolio optimization with return prediction using deep
engineering and management. IEEE; 2024, p. 1–6.
learning and machine learning. Expert Syst Appl 2021;165:113973.
[170] Poutré C, Chételat D, Morales M. Deep unsupervised anomaly detection in
[140] Chen W, Zhang H, Mehlawat MK, Jia L. Mean–variance portfolio optimiza-
high-frequency markets. J Financ Data Sci 2024;10:100129.
tion using machine learning-based stock price prediction. Appl Soft Comput
[171] Zhu M, Gong Y, Xiang Y, Yu H, Huo S. Utilizing GANs for fraud detection:
2021;100:106943.
model training with synthetic transaction data. In: International conference on
[141] Nyakurukwa K, Seetharam Y. Sentimental showdown: News media vs. social
image, signal processing, and pattern recognition, vol. 13180. SPIE; 2024, p.
media in stock markets. Heliyon 2024;10(9).
887–94.
[142] Yadav A, Vishwakarma DK. Sentiment analysis using deep learning
[172] Shoetan PO, Familoni BT. Transforming fintech fraud detection with advanced
architectures: a review. Artif Intell Rev 2020;53(6):4335–85.
[143] Jain PK, Saravanan V, Pamula R. A hybrid CNN-LSTM: A deep learning artificial intelligence algorithms. Financ Account Res J 2024;6(4):602–25.
approach for consumer sentiment analysis using qualitative user-generated [173] Thimonier H, Popineau F, Rimmel A, Doan B-L, Daniel F. Comparative
contents. Trans Asian Low- Resour Lang Inf Process 2021;20(5):1–15. evaluation of anomaly detection methods for fraud detection in online credit
[144] Memiş E, Akarkamçı H, Yeniad M, Rahebi J, Lopez-Guede JM. Comparative card payments. In: International congress on information and communication
study for sentiment analysis of financial tweets with deep learning methods. technology. Springer Nature Singapore Singapore; 2024, p. 37–50.
Appl Sci 2024;14(2):588. [174] Potla RT. AI in fraud detection: Leveraging real-time machine learning for
[145] Loginova E, Tsang WK, van Heijningen G, Kerkhove L-P, Benoit DF. Forecasting financial security. J Artif Intell Res Appl 2023;3(2):534–49.
directional bitcoin price returns using aspect-based sentiment analysis on online [175] Pei Y, Lyu F, Van Ipenburg W, Pechenizkiy M. Subgraph anomaly detection in
text data. Mach Learn 2024;113(7):4761–84. financial transaction networks. In: Proceedings of the first ACM international
[146] Sakhare NN, Shaik IS. Spatial federated learning approach for the sentiment conference on AI in finance. 2020, p. 1–8.
analysis of stock news stored on blockchain. Spat Inf Res 2024;32(1):13–27. [176] Adeniran IA, Efunniyi CP, Osundare OS, Abhulimen A. Optimizing logistics and
[147] Fatouros G, Soldatos J, Kouroumali K, Makridis G, Kyriazis D. Transforming supply chain management through advanced analytics: Insights from industries.
sentiment analysis in the financial domain with ChatGPT. Mach Learn Appl Eng Sci Technol J 2024;5(8).
2023;14:100508. [177] Pasupuleti V, Thuraka B, Kodete CS, Malisetty S. Enhancing supply chain
[148] Qiu Y, Song Z, Chen Z. Short-term stock trends prediction based on sentiment agility and sustainability through machine learning: Optimization techniques
analysis and machine learning. Soft Comput 2022;26(5):2209–24. for logistics and inventory management. Logistics 2024;8(3):73.

25
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

[178] Mahmoud A, Mohammed A. Leveraging hybrid deep learning models [207] El Hajj M, Hammoud J. Unveiling the influence of artificial intelligence
for enhanced multivariate time series forecasting. Neural Process Lett and machine learning on financial markets: A comprehensive analysis of AI
2024;56(5):1–25. applications in trading, risk management, and financial operations. J Risk
[179] Hossain MS, Khan A, Das P, Haque MSU, Kamruzzaman F, Akter S, et al. Financ Manag 2023;16(10):434.
Enhanced market trend forecasting using machine learning models: a study with [208] Fischer T, Krauss C. Deep learning with long short-term memory networks for
external factor integration. Int Interdiscip Bus Econ Adv J 2025;6(01):5–12. financial market predictions. European J Oper Res 2018;270(2):654–69.
[180] Nguyen HD, Tran KP, Thomassey S, Hamad M. Forecasting and anomaly [209] Nosratabadi S, Mosavi A, Duan P, Ghamisi P, Filip F, Band SS, et al. Data
detection approaches using LSTM and LSTM autoencoder techniques with the science in economics: comprehensive review of advanced machine learning and
applications in supply chain management. Int J Inf Manage 2021;57:102282. deep learning methods. Mathematics 2020;8(10):1799.
[181] Belhadi A, Kamble SS, Mani V, Benkhati I, Touriki FE. An ensemble machine [210] Karimi D, Dou H, Warfield SK, Gholipour A. Deep learning with noisy labels:
learning approach for forecasting credit risk of agricultural SMEs’ investments Exploring techniques and remedies in medical image analysis. Med Image Anal
in agriculture 4.0 through supply chain finance. Ann Oper Res 2021;1–29. 2020;65:101759.
[182] Weng T, Liu W, Xiao J. Supply chain sales forecasting based on lightGBM and [211] Kim B, Kim H, Kim K, Kim S, Kim J. Learning not to learn: Training deep
LSTM combination model. Ind Manag & Data Syst 2020;120(2):265–79. neural networks with biased data. In: Proceedings of the IEEE/CVF conference
[183] Punia S, Singh SP, Madaan JK. A cross-temporal hierarchical framework and on computer vision and pattern recognition. 2019, p. 9012–20.
deep learning for supply chain forecasting. Comput Ind Eng 2020;149:106796. [212] Song D, Baek AMC, Kim N. Forecasting stock market indices using padding-
[184] Zou J, Cao H, Liu L, Lin Y, Abbasnejad E, Shi JQ. Astock: A new dataset and based fourier transform denoising and time series deep learning models. IEEE
automated stock trading based on stock-specific news analyzing model. 2022, Access 2021;9:83786–96.
arXiv preprint arXiv:2206.06606. [213] Magris M. Volatility modeling and limit-order book analytics with high-
[185] Onyshchak O. Stock market dataset. 2020, http://dx.doi.org/10.34740/ frequency data. Tampere University; 2019.
KAGGLE/DSV/1054465, Kaggle URL https://www.kaggle.com/dsv/1054465. [214] Aldridge I, Krawciw S. Real-time risk: What investors should know about
[186] Vopani. NIFTY-50 stock market data (2000 - 2021). 2021, Kaggle URL https: FinTech, high-frequency trading, and flash crashes. John Wiley & Sons; 2017.
//www.kaggle.com/datasets/rohanrao/nifty50-stock-market-data. [215] Adhikari D, Jiang W, Zhan J, He Z, Rawat DB, Aickelin U, et al. A comprehen-
[187] Xu Y, Cohen SB. Stock movement prediction from tweets and historical prices. sive survey on imputation of missing data in internet of things. ACM Comput
In: Proceedings of the 56th annual meeting of the association for computational Surv 2022;55(7):1–38.
linguistics (volume 1: long papers). 2018, p. 1970–9. [216] Ying X. An overview of overfitting and its solutions. In: Journal of physics:
[188] González-Núñez E, Trejo LA. Datasets of stock market indices. IEEE Dataport; conference series, vol. 1168. IOP Publishing; 2019, 022022.
2024, http://dx.doi.org/10.21227/yvfx-n484. [217] Montesinos López OA, Montesinos López A, Crossa J. Overfitting, model tuning,
and evaluation of prediction performance. In: Multivariate statistical machine
[189] Chaudhary Y. Stock-market sentiment dataset. 2020, http://dx.doi.org/
learning methods for genomic prediction. Springer; 2022, p. 109–39.
10.34740/KAGGLE/DSV/1217821, Kaggle URL https://www.kaggle.com/dsv/
[218] Petrozziello A, Troiano L, Serra A, Jordanov I, Storti G, Tagliaferri R, et al.
1217821.
Deep learning for volatility forecasting in asset management. Soft Comput
[190] Yukhymenko H. Stock tweets for sentiment analysis and prediction. 2022, Kag-
2022;26(17):8553–74.
gle URL https://www.kaggle.com/datasets/equinxx/stock-tweets-for-sentiment-
[219] Yan B, Aasma M, et al. A novel deep learning framework: Prediction and
analysis-and-prediction.
analysis of financial time series using CEEMD and LSTM. Expert Syst Appl
[191] Irfan R. Dataset for stock market prediction. IEEE Dataport; 2022, http://dx.
2020;159:113609.
doi.org/10.21227/e29h-5c78.
[220] Peng Y, de Moraes Souza JG. Chaos, overfitting and equilibrium: To what
[192] US Govt. Stock market dataset. Trading Economics; 2023, URL https://
extent can machine learning beat the financial market? Int Rev Financ Anal
tradingeconomics.com/united-states/stock-market.
2024;95:103474.
[193] Sawhney R, Agarwal S, Mittal V, Rosso P, Nanda V, Chava S. Cryptocurrency
[221] Chatzis SP, Siakoulis V, Petropoulos A, Stavroulakis E, Vlachogiannakis N.
bubble detection: a new stock market dataset, financial task & hyperbolic
Forecasting stock market crisis events using deep and statistical machine
models. 2022, arXiv preprint arXiv:2206.06320.
learning techniques. Expert Syst Appl 2018;112:353–71.
[194] Basak S, Kar S, Saha S, Khaidem L, Dey SR. Predicting the direction of
[222] Buhrmester V, Münch D, Arens M. Analysis of explainers of black box deep
stock market prices using tree-based classifiers. North Am J Econ Financ
neural networks for computer vision: A survey. Mach Learn Knowl Extr
2019;47:552–67.
2021;3(4):966–89.
[195] Duarte JJ, Montenegro González S, Cruz Jr JC. Predicting stock price
[223] Liang Y, Li S, Yan C, Li M, Jiang C. Explaining the black-box model: A survey
falls using news data: Evidence from the Brazilian market. Comput Econ
of local interpretation methods for deep neural networks. Neurocomputing
2021;57(1):311–40.
2021;419:168–82.
[196] Henrique BM, Sobreiro VA, Kimura H. Stock price prediction using support [224] Soundararajan R, Shenbagaraman V. Enhancing financial decision-making
vector regression on daily and up to the minute prices. J Financ Data Sci through explainable AI and blockchain integration: improving trans-
2018;4(3):183–201. parency and trust in predictive models. Educ Administration: Theory Pr
[197] Weng B, Lu L, Wang X, Megahed FM, Martinez W. Predicting short-term stock 2024;30(4):9341–51.
prices using ensemble methods and online data sources. Expert Syst Appl [225] Abdullayeva FJ. Detection of cyberattacks in cloud computing service de-
2018;112:258–73. livery models using correlation based feature selection. In: 2021 IEEE 15th
[198] Tudor C, Sova R. Enhancing trading decision in financial markets: An algo- international conference on application of information and communication
rithmic trading framework with continual mean-variance optimization, window technologies. IEEE; 2021, p. 1–4.
presetting, and controlled early-stopping. IEEE Access 2024. [226] Dumitrescu E, Hué S, Hurlin C, Tokpavi S. Machine learning for credit scoring:
[199] Khattak BHA, Shafi I, Khan AS, Flores ES, Lara RG, Samad MA, et al. A Improving logistic regression with non-linear decision-tree effects. European J
systematic survey of AI models in financial market forecasting for profitability Oper Res 2022;297(3):1178–92.
analysis. IEEE Access 2023. [227] Rudin C. Stop explaining black box machine learning models for high
[200] Kim A, Yang Y, Lessmann S, Ma T, Sung M-C, Johnson JE. Can deep learning stakes decisions and use interpretable models instead. Nat Mach Intell
predict risky retail investors? A case study in financial risk behavior forecasting. 2019;1(5):206–15.
European J Oper Res 2020;283(1):217–34. [228] Scott M, Su-In L, et al. A unified approach to interpreting model predictions.
[201] Sanz-Cruzado J, Mccreadie R, Droukas N, Macdonald C, Ounis I. On transaction- Adv Neural Inf Process Syst 2017;30:4765–74.
based metrics as a proxy for profitability of financial asset recommendations. [229] Ribeiro MT, Singh S, Guestrin C. ‘‘Why should i trust you?’’ explaining
In: Proceedings of the 3rd international workshop on personalization & the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD
recommender systems in financial services. 2022. international conference on knowledge discovery and data mining. 2016, p.
[202] Mostafa G, Mahmoud H, Abd El-Hafeez T, E. ElAraby M. The power of deep 1135–44.
learning in simplifying feature selection for hepatocellular carcinoma: a review. [230] Gao Y, Ruan Y. Interpretable deep learning model for building en-
BMC Med Inform Decis Mak 2024;24(1):287. ergy consumption prediction based on attention mechanism. Energy Build
[203] Guven A. Turkish tweets dataset. 2021, Kaggle URL https://www.kaggle.com/ 2021;252:111379.
datasets/anil1055/turkish-tweet-dataset. [231] Dixon MF, Polson NG, Sokolov VO. Deep learning for spatio-temporal modeling:
[204] Kusawa S. Stock price dataset EOD. 2021, Kaggle URL http://kaggle.com/ dynamic traffic flows and high frequency trading. Appl Stoch Models Bus Ind
datasets/sunnykusawa/stock-price-dataset-eod. 2019;35(3):788–807.
[205] Yahoo. Beximco Pharmaceuticals Limited (R2WA.F) stock historical prices & [232] Sirignano J, Cont R. Universal features of price formation in financial markets:
data. 2018, Yahoo! Finance URL https://finance.yahoo.com/quote/R2WA.F/ perspectives from deep learning. In: Machine learning and AI in finance.
history?p=R2WA.F. Routledge; 2021, p. 5–15.
[206] Horobet A, Boubaker S, Belascu L, Negreanu CC, Dinca Z. Technology-driven [233] Wang S, Zheng H, Wen X, Fu S. Distributed high-performance computing
advancements: Mapping the landscape of algorithmic trading literature. Technol methods for accelerating deep learning training. J Knowl Learn Sci Technol
Forecast Soc Change 2024;209:123746. ISSN: 2959- 6386 (Online) 2024;3(3):108–26.

26
M.S.M. Bhuiyan et al. Array 26 (2025) 100390

[234] Lattuada M, Gianniti E, Ardagna D, Zhang L. Performance prediction of deep [245] Hasan BMS, Abdulazeez AM. A review of principal component anal-
learning applications training in GPU as a service systems. Clust Comput ysis algorithm for dimensionality reduction. J Soft Comput Data Min
2022;25(2):1279–302. 2021;2(1):20–30.
[235] Jauro F, Chiroma H, Gital AY, Almutairi M, Shafi’i MA, Abawajy JH. [246] Schrettenbrunner MB. Artificial-intelligence-driven management: Autonomous
Deep learning architectures in emerging cloud computing architectures: Re- real-time trading and testing of portfolio or inventory strategies. IEEE Eng
cent development, challenges and next research trend. Appl Soft Comput Manag Rev 2023;51(3):65–76.
2020;96:106582. [247] Jamaani F, Roca E. Are the regional gulf stock markets weak-form efficient
[236] Liu D, Kong H, Luo X, Liu W, Subramaniam R. Bringing AI to edge: From deep as single stock markets and as a regional stock market? Res Int Bus Financ
learning’s perspective. Neurocomputing 2022;485:297–320. 2015;33:221–46.
[237] Eissa MA, Al Refai H, Chortareas G. Heterogeneous impacts of geopolitical risk [248] Balli F, Balli HO, Hasan M, Gregory-Allen R. Geopolitical risk spillovers and its
factors on stock markets in the Middle East: A quantile regression analysis determinants. Ann Reg Sci 2022;68(2):463–500.
across four emerging economies. J Econ Asymmetries 2024;30:e00374. [249] Hassija V, Chamola V, Mahapatra A, Singal A, Goel D, Huang K, et al.
[238] Belhoula MM, Mensi W, Naoui K. Impacts of investor’s sentiment, uncertainty Interpreting black-box models: a review on explainable artificial intelligence.
indexes, and macroeconomic factors on the dynamic efficiency of G7 stock Cogn Comput 2024;16(1):45–74.
markets. Qual Quant 2024;58(3):2855–86. [250] Li W, Law KE. Deep learning models for time series forecasting: a review. IEEE
[239] Alhnaity B, Abbod M. A new hybrid financial time series prediction model. Eng Access 2024.
Appl Artif Intell 2020;95:103873. [251] Katharopoulos A, Vyas A, Pappas N, Fleuret F. Transformers are rnns: Fast
[240] Cavalcante RC, Brasileiro RC, Souza VL, Nobrega JP, Oliveira AL. Computa- autoregressive transformers with linear attention. In: International conference
tional intelligence and financial markets: A survey and future directions. Expert on machine learning. PMLR; 2020, p. 5156–65.
Syst Appl 2016;55:194–211. [252] Xu J, Zhou J, Jia Y, Li J, Hui X. An adaptive master-slave regularized model for
[241] Schnaubelt M. A comparison of machine learning model validation schemes unexpected revenue prediction enhanced with alternative data. In: 2020 IEEE
for non-stationary time series data. Technical report, FAU Discussion Papers in 36th international conference on data engineering. IEEE; 2020, p. 601–12.
Economics; 2019. [253] Devlin J. Bert: Pre-training of deep bidirectional transformers for language
[242] Georgiou T, Liu Y, Chen W, Lew M. A survey of traditional and deep learning- understanding. 2018, arXiv preprint arXiv:1810.04805.
based feature descriptors for high dimensional data in computer vision. Int J [254] García S, Ramírez-Gallego S, Luengo J, Benítez JM, Herrera F. Big data
Multimed Inf Retr 2020;9:135–70. preprocessing: methods and prospects. Big Data Anal 2016;1:1–22.
[243] Hua J, Tembe WD, Dougherty ER. Performance of feature-selection methods in [255] Shen M, Zhang H, Cao Y, Yang F, Wen Y. Missing data imputation for solar yield
the classification of high-dimension data. Pattern Recognit 2009;42(3):409–24. prediction using temporal multi-modal variational auto-encoder. In: Proceedings
[244] Alkhayrat M, Aljnidi M, Aljoumaa K. A comparative dimensionality reduction of the 29th ACM international conference on multimedia. 2021, p. 2558–66.
study in telecom customer segmentation using deep learning and PCA. J Big [256] Abayomi-Alli OO, Damaševičius R, Qazi A, Adedoyin-Olowe M, Misra S. Data
Data 2020;7(1):9. augmentation and deep learning methods in sound classification: A systematic
review. Electronics 2022;11(22):3795.

27

You might also like