Article

Comparative Analysis of Deep Learning Models for Stock Price

Prediction in the Indian Market
Moumita Barua 1, *, Teerath Kumar 1,2, *, Kislay Raj 2 and Arunabha M. Roy 3

1 Department of Business Analytics, Dublin Business School, 13/14 Aungier St, D02 WC04 Dublin, Ireland
2 School of Computing, Dublin City University, D09 V209 Dublin, Ireland; kislay.raj2@mail.dcu.ie
3 Aerospace Engineering Department, University of Michigan, Ann Arbor, MI 48109, USA;
arunabhr.umich@gmail.com
* Correspondence: moumita.barua22@gmail.com (M.B.); teerath.menghwar2@mail.dcu.ie or
teerath.kumar@dbs.ie (T.K.)

Abstract: This research presents a comparative analysis of various deep learning models—including
Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), Convolutional Neural Net-
works (CNN), Gated Recurrent Units (GRU), and Attention LSTM—in predicting stock prices of
major companies in the Indian stock market, specifically HDFC, TCS, ICICI, Reliance, and Nifty.
The study evaluates model performance using key regression metrics such as Mean Absolute Er-
ror (MAE), Mean Squared Error (MSE), and R-Squared (R²). The results indicate that CNN and
GRU models generally outperform the others, depending on the specific stock, and demonstrate
superior capabilities in forecasting stock price movements. This investigation provides insights into
the strengths and limitations of each model while highlighting potential avenues for improvement
through feature engineering and hyperparameter optimization.

Keywords: stock prediction; deep learning; recurrent neural networks; long short-term memory;
convolutional neural networks; Indian stock market

JEL Classification: C80

Citation: Barua, M.; Kumar, T.; Raj, K.;

Roy, A.M. Comparative Analysis of
1. Introduction
Deep Learning Models for Stock Price
Prediction in the Indian Market. In the finance sector, stock price prediction remains a critical challenge for organiza-
FinTech 2024, 3, 551–568. https:// tions as it offers key insights that investors, traders, and financial institutions need to make
doi.org/10.3390/fintech3040029 informed decisions, manage risks, and enhance their profitability. The inherent complexity,
dynamism, and non-linearity of financial markets make it particularly difficult to develop
Academic Editor: David
accurate forecasting models [1,2]. For decades, traditional statistical techniques such as
(Yulong) Liu
Autoregressive Integrated Moving Average (ARIMA) and Generalised Autoregressive
Received: 13 October 2024 Conditional Heteroskedasticity (GARCH) have been the backbone of financial time-series
Revised: 15 November 2024 forecasting. However, these models typically assume linearity and often struggle to capture
Accepted: 20 November 2024 the intricate non-linear dependencies present in stock price data.
Published: 28 November 2024 Efforts to improve prediction accuracy through machine learning techniques have
seen moderate success. Support Vector Machines (SVM) and Random Forests, for instance,
have been applied to stock price forecasting, yielding promising results [3–6]. Nevertheless,
these methods often fall short of capturing long-term dependencies, which are crucial for
Copyright: © 2024 by the authors.
accurate predictions in the financial domain.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
In contrast, deep learning models, particularly Recurrent Neural Networks (RNNs),
distributed under the terms and
Long Short-Term Memory (LSTM) networks, and Gated Recurrent Units (GRUs), have
conditions of the Creative Commons emerged as powerful tools for time-series analysis within the financial sector [7,8]. RNNs
Attribution (CC BY) license (https:// are designed to handle temporal dependencies, but their performance is frequently hin-
creativecommons.org/licenses/by/ dered by the vanishing gradient problem, which limits their ability to model long-term
4.0/).

FinTech 2024, 3, 551–568. https://doi.org/10.3390/fintech3040029 https://www.mdpi.com/journal/fintech

FinTech 2024, 3 552

dependencies. LSTM and GRU architectures, with their memory cells and gating mech-
anisms, effectively address these limitations, enabling them to discern both short-term
fluctuations and long-term trends in time-series data.
Additionally, Convolutional Neural Networks (CNNs), which have primarily been
utilized for image processing, have demonstrated potential in financial forecasting by
identifying local temporal patterns in stock prices [7]. Hybrid models that integrate LSTM
networks with attention mechanisms further enhance predictive accuracy by focusing on
essential components of the input sequence.
Despite the growing body of literature on the application of deep learning models for
stock price prediction, there remains a paucity of comprehensive comparisons regarding
the efficacy of these models across various stocks, particularly within the Indian market.
This study aims to bridge this gap by conducting a thorough comparative analysis of
RNN, LSTM, CNN, GRU, and Attention LSTM models in predicting the stock prices of
five major Indian companies: HDFC, TCS, ICICI, Reliance, and the Nifty 50 index [9].
These companies are pivotal to key sectors of the Indian economy, making them suitable
candidates for evaluating the effectiveness of different deep-learning strategies [10].
Furthermore, this study will not only elucidate the performance of various models in
stock price prediction but will also discuss the implications of these findings for investors
and practitioners in the financial sector. The following sections will provide a comprehen-
sive literature review of existing research on stock price prediction using deep learning
methodologies, detailing the processes of data acquisition and model deployment. This
will be followed by an in-depth comparison of the results and a discussion of the broader
implications of the study’s findings.

2. Related Work and Inspiration for Methodology

There have been several problems solved by deep learning in various domains, in-
cluding the image domain [11–19], audio domain [20–24], text domain [18,25–27], and
many others [26–31]. Among these applications, predicting stock prices has long been
regarded as a formidable challenge due to the inherent volatility and unpredictability of
financial markets. Traditional approaches to financial forecasting have primarily relied
on time series models, such as Autoregressive Integrated Moving Average (ARIMA) and
Generalized Autoregressive Conditional Heteroskedasticity (GARCH). While these models
are beneficial for specific forecasting tasks, they have limitations, particularly in relying
on linearity, which is problematic given the non-linear nature of stock market data [32]. In
recent years, researchers have increasingly turned to machine learning models for stock
price prediction. A pivotal study by Patel et al. examined various traditional machine
learning approaches, including Support Vector Machines (SVM), Random Forest (RF), and
Artificial Neural Networks (ANN), and found that machine learning models outperformed
statistical methods in handling complex relationships and processing large datasets [32].
However, these models also faced challenges in addressing the sequential and temporal
dynamics inherent to stock market data.
Recurrent Neural Networks (RNNs) were among the first deep learning architectures
applied to time series forecasting, designed to capture temporal dependencies. However,
they encounter limitations due to the vanishing gradient problem, which hinders their
capacity to represent long-range relationships in time-series data [33]. To overcome these
challenges, Long Short-Term Memory (LSTM) networks, introduced by Hochreiter and
Schmidhuber in 1997, incorporate memory cells and gating mechanisms to enhance predic-
tive accuracy for financial time series. Numerous studies have demonstrated the efficacy
of LSTMs in stock price forecasting. For instance, Shen and Shafiq [34] investigated the
application of LSTMs for short-term trend predictions in the Chinese stock market, finding
that LSTMs provided superior accuracy and stability compared to traditional machine
learning models. Similarly, research by Fathali et al. [4], which compared LSTM, RNN, and
CNN models for forecasting the Nifty 50 index, confirmed that LSTMs generally exhibited
lower mean squared error (MSE) and higher prediction accuracy.
FinTech 2024, 3 553

Initially developed for image processing tasks, Convolutional Neural Networks

(CNNs) have been adapted for time series forecasting, including stock price predictions.
CNNs excel at capturing local temporal patterns, making them particularly effective for
stocks exhibiting stable and repetitive behaviors. For example, Sen et al. [35] explored a
hybrid CNN-LSTM model for stock price prediction in the Indian markets, showing that
this combined approach outperformed models utilizing LSTMs or CNNs in isolation.
Gated Recurrent Units (GRUs) simplify the architecture of LSTMs by merging the
forget and input gates into a single gate, enhancing computational efficiency while retain-
ing the ability to model long-term dependencies effectively. Studies have indicated that
GRUs perform well in predicting stock prices, particularly for volatile assets. Singh [3]
examined the effectiveness of GRU models in predicting stock prices for prominent Indian
companies, concluding that GRUs offer similar predictive accuracy to LSTMs while being
more computationally efficient.
Recent research has focused on hybrid models that combine various deep-learning
techniques to improve prediction accuracy. The Attention LSTM model, which integrates
attention mechanisms with LSTM architecture, is one such approach. This model enables
the network to focus on critical time steps or features relevant to the prediction task,
enhancing interpretability and performance [36]. Hernández et al. [36] further noted that
attention mechanisms can refine a model’s focus on essential temporal trends, thereby
improving the precision of time-series forecasting.
Despite these advancements, the current literature lacks comprehensive analyses of
various methodologies applied across different market conditions, particularly in the Indian
context. This study contributes to the field by the following:
• Conducting a thorough analysis of RNN, LSTM, CNN, GRU, and Attention LSTM
models for stock price prediction.
• Evaluating these models using a selection of Indian stocks, including HDFC, TCS,
ICICI, Reliance, and the Nifty 50 index.
• Assessing the strengths and weaknesses of each model, with a particular focus on
their ability to manage both volatile and stable stocks.
The following sections will detail the processes of data collection, preprocessing, and
the experimental framework designed to evaluate the performance of these models.

3. Materials and Methods

In this section of the analysis, we examine the dataset that underpins predicting
stock prices, describe the preprocessing activities carried out, showcase the deep learning
models implemented, detail the training procedure, and present the evaluation criteria.
Each component of the methodology aims to secure the most precise and efficient evaluation
of different deep-learning models for making stock price forecasts within the Indian stock
market. Figure 1 depicts the methodology flow of the experiment.

Figure 1. Methodology Diagram.

FinTech 2024, 3 554

3.1. Data Collection

The stock price data for this study were sourced from the National Stock Exchange
(NSE) of India, focusing on five major companies: HDFC, TCS, ICICI, Reliance, and the
Nifty 50 index. The selection of these companies was deliberate, as they represent various
sectors of the Indian economy, including finance, technology, energy, and the broader stock
market itself through the Nifty 50. The dataset encompasses a timeframe from January
2016 to December 2021, providing over five years of daily stock price data. The collected
data include the following key metrics:
• Open Price: The price at which the stock begins trading when the market opens.
• Close Price: The price at which the stock concludes trading when the market closes.
• High Price: The highest price at which the stock was traded during the day.
• Low Price: The lowest price at which the stock was traded during the day.
• Volume: The total number of shares traded during the day.
Utilizinga five-year dataset is crucial to ensure that the models are exposed to a
diverse range of market conditions, including bull and bear markets, as well as periods of
heightened volatility that are typical in stock trading. While several previous studies have
employed similar datasets for stock price forecasting, they have predominantly focused on
shorter timeframes.

3.2. Data Preprocessing

Stock market data are often noisy and incomplete due to factors such as non-trading
days and market fluctuations, necessitating thorough preprocessing before the application
of deep learning models. The following steps were implemented to ensure data quality:

3.2.1. Handling Missing Data

In stock market time series, missing data primarily arise due to market closures on
non-trading days (e.g., weekends and holidays) or during market outages. Although it
is expected that no trading occurs on these days, the time series models often require a
continuous, uniform time series for effective training and prediction. Gaps due to non-
trading days can lead to inconsistencies in the model’s temporal structure, particularly
for models that rely on a fixed time-step format. To address this, a linear interpolation
technique was employed, estimating values for non-trading days based on data from
preceding and subsequent trading days [37]. This method maintains continuity in the
dataset, allowing models to process sequential data without the distortions that may arise
from abrupt temporal gaps or random imputation. By preserving the temporal consistency,
linear interpolation ensures that the model’s learning remains focused on true market
patterns rather than on handling uneven data intervals.

3.2.2. Normalization
With the diverse nature of stock prices and trading volumes, we normalized all features
using Min-Max scaling, resulting in a range of [0,1]. Achieving faster convergence speed
in deep learning models while ensuring none overpower the learning process depends
significantly on this normalization technique [38,39]. Many financial forecasting techniques
find MinMax normalization helpful, as it improves performance with extensive datasets
featuring different scales. The normalization formula is as follows:
x − xmin
xnorm = (1)
xmax − xmin
where x is the original feature value, xmin is the minimum value of the feature, and xmax is
the maximum value of the feature.

3.2.3. Train-Test Split

To accurately evaluate the performance of deep learning models on unseen data, the
dataset was divided chronologically into training and testing sets. Specifically, the training
FinTech 2024, 3 555

data comprised 80% of the total dataset, covering the period from January 2016 through
December 2020, while the testing set included the remaining 20%, spanning January to
December 2021.
This chronological split is crucial for time-series data, as it prevents information
leakage from future data points into the training process, thereby preserving the model’s
predictive validity [39,40]. By leveraging historical data for training and unseen recent
data for validation, the split simulates realistic forecasting conditions. Additionally, this
partitioning exposed the model to diverse market phases, including growth cycles, market
corrections, and periods of heightened volatility, thereby enriching its learning experience
under varied conditions.

3.3. Deep Learning Architectures

This study explores five different deep learning architectures: A Recurrent Neural
Network (RNN), a Long Short-Term Memory (LSTM), a Convolutional Neural Network
(CNN), a Gated Recurrent Unit (GRU), and an Attention LSTM are all part of the discussion.
The selection of each architecture was due to its relevance for time-series prediction tasks,
especially within financial markets [1,41–45].

3.3.1. Vanilla RNN

The sequential data processing of neural network models includes RNNs among the
earliest in existence. The skill to preserve inconspicuous states that amass information
from earlier time steps makes them perfect for time-series data, in which each data point is
linked to preceding data [33,38,40]. Nonetheless, RNNs suffer from the vanishing gradient
challenge, making them ineffective for the representation of extended relationships in time
series. In this work, we trained a basic RNN of one hidden layer and 100 units using the
Adam optimizer.

3.3.2. LSTM
The creation of LSTM networks was meant to address the deficiencies of regular RNNs
through the deployment of memory cells that can hold information for prolonged lengths of
time [46]. The recognition capability of memory cells in LSTMs for immediate and historical
dependencies makes them extremely valuable for stock price prediction, because trends
usually follow long-term styles [37,40,47]. The LSTM model as part of this research includes
two LSTM layers, each composed of 64 units, leading to a dense layer for final predictions.
An addition of a dropout layer helped alleviate overfitting [48].

3.3.3. CNN
Generally speaking, CNNs are employed in image processing, but they have also been
modified for time-series data owing to their strength in local pattern recognition [39,49]. By
sliding a filter over the time series, CNNs in stock price forecasting can detect short-term
trends, including rapid price changes and minor corrections [35,39,50]. The CNN model
implemented in this study consists of one convolutional layer, which has 64 filters and a
kernel that is 3 in size. Once these layers are complete, the maximum pooling layers are
in place to contract feature maps and lower the computational weight. In the end, the
model features a fully connected dense layer to deliver the predicted stock prices. CNNs
have shown a remarkable ability to uncover short-term trends while successfully adding to
financial forecasting models [39].

3.3.4. GRU
GRUs exhibit a prototype of LSTMs that simplify computational complexity by fusing
the forget gate and the input gate into an update gate [33,40,44,47]. In conditions where
computational resources are limited, the fast training capabilities of these GRUs together
with their potential to analyze long-term dependencies make them suitable for stock price
prediction [47]. The GRU model employed in this study has two layers of GRU with
FinTech 2024, 3 556

64 units each, followed by an output layer that is dense. Showings reveal that GRUs can
perform better than LSTMs in specific time-series forecasting jobs thanks to their simplified
designs and faster convergence.

3.3.5. LSTM with Attention Mechanism

Due to their ability to target important parts of the input sequence, attention mecha-
nisms have recently become a popular choice in time-series forecasting [36]. The model
can measure the significance of particular time steps by applying attention mechanisms to
stock price prediction, which provides both improved interpretability and better prediction
accuracy [36]. The attention layer of the Attention LSTM model featured in this study
operates after an LSTM layer, assigning weight to each time step based on its importance
for the prediction task. Successful implementation of attention mechanisms has occurred
in many areas, such as stock market prediction, where they facilitate better results by
permitting the model to focus on key market happenings.

3.4. Model Training

All models used the Adam optimizer, a widely adopted choice in deep learning for
its capability to modify learning rates during the training phase [51]. At the outset, the
learning rate was 0.001, and the application of early stopping helped to ward off overfitting.
After 10 unchanging epochs of validation loss, triggering early stopping is the subsequent
action taken. The training for the models took place over 50 epochs, with a batch size of
64, a frequent selection for time series data. Within training, the models were gauged on a
validation set to keep track of their performance and make hyperparameter corrections.
We carried out the training using an NVIDIA Tesla GPU that greatly decreased the training
time when compared to training utilizing CPUs. Thanks to GPU acceleration, we were able
to efficiently train large models like LSTM and Attention LSTM, which need additional
computational power because of their elaborated architectures. Earlier research has pointed
out that GPU acceleration plays a key role in the training of deep learning models with
large amounts of data, which includes stock market data.

3.5. Evaluation Metrics

To evaluate the performance of the deep learning models, we used several standard
regression metrics commonly applied in financial forecasting:
Mean Absolute Error (MAE): MAE measures the average magnitude of prediction
errors, providing a straightforward assessment of accuracy by treating both large and small
errors equally [52]. The MAE is given by the following:
n
1
MAE =
n ∑ |yi − ŷi | (2)
i =1

where yi is the actual value, ŷi is the predicted value, and n is the total number of
data points.
Mean Squared Error (MSE): MSE places greater emphasis on larger errors, making it
more sensitive to outliers, which can be significant in stock price forecasting [52].
n
1
MSE =
n ∑ (yi − ŷi )2 (3)
i =1

Root Mean Squared Error (RMSE): RMSE, the square root of MSE, provides an error
metric in the original scale of stock prices, often preferred for its interpretability in financial
applications [52]. s
n
1
RMSE =
n ∑ (yi − ŷi )2 (4)
i =1
FinTech 2024, 3 557

Mean Absolute Percentage Error (MAPE): MAPE represents the prediction error as a
percentage, offering an intuitive perspective on forecasting accuracy, which is particularly
useful for comparing errors across datasets with different scales.
n
1 yi − ŷi
MAPE =
n ∑ yi
× 100 (5)
i =1

R-Squared (R²): R-Squared measures how closely predicted values align with actual
values, assessing the model’s ability to explain variance in stock prices [52].

∑in=1 (yi − ŷi )2

R2 = 1 − (6)
∑in=1 (yi − ȳ)2

where ȳ is the mean of actual values.

These metrics enable comprehensive evaluation, analyzing both typical accuracy
(MAE, MAPE) and the model’s sensitivity to extreme values (MSE, RMSE), along with the
model’s explanatory power (R²).

4. Results and Discussion

This section presents the results of the comparative analysis of five deep learning
models—RNN, LSTM, CNN, GRU, and Attention LSTM—used for stock price prediction on
five major Indian stocks: HDFC, TCS, ICICI, Reliance, along with Nifty 50. The evaluation
of the models considered Mean Absolute Error (MAE), Mean Squared Error (MSE), Root
Mean Squared Error (RMSE), and R-Square (R²). The discussion pertains to each stock in
regard to data analysis and model performance.

4.1. HDFC Stock

The performance of the models for HDFC stock prediction is summarized in Table 1.
The RNN model demonstrated the lowest performance, with an R-Square of −0.014790,
indicating a poor correlation between predicted and actual stock prices. The LSTM model
slightly improved upon this with an R-Square of 0.015231. However, the CNN model
performed the best for HDFC, achieving an MAE of 0.201941, an MSE of 0.074096, and an R-
Square of 0.007445, indicating better prediction accuracy. GRU and Attention LSTM models
performed similarly, though slightly worse than CNN. Note the CNN model achieved
the lowest Mean Absolute Error (MAE) among the models, indicating it had the smallest
average prediction error. However, other models, such as the GRU, obtained slightly
better Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and R² values, which
may be preferable depending on the model’s sensitivity to larger errors or the importance
of trend-following ability. Therefore, in our assessment, CNN offers an optimal balance
between simplicity and general predictive accuracy, while alternative models may provide
advantages in specific scenarios.

Table 1. Model performances on the HDFC Stock Data in terms of the evaluation metrics.

Model MAE MSE RMSE R-Square MAPE

RNN 0.211425 0.075756 0.275238 −0.014790 0.362591
LSTM 0.205168 0.073515 0.271136 0.015231 0.073838
CNN 0.201941 0.074096 0.272206 0.007445 0.073531
GRU 0.204980 0.073715 0.271506 0.012543 0.075203
Attention LSTM 0.205012 0.074433 0.272824 0.002933 0.074322

The prediction for the models specifically for the HDFC Stock is shown in Figure 2 below.
FinTech 2024, 3 558

(a) RNN prediction (b) LSTM prediction

(c) CNN prediction (d) GRU prediction

(e) LSTM with Attention prediction

Figure 2. Model predictions for the HDFC Stock.

As for the models, the basic RNN performs worst in predicting the actual stock price
path with considerable distortions from the real trends (Figure 2a). Although RNNs are
basic for sequential data analysis, they are often replaced by more effective models, such
as LSTM and GRU in cases when patterns are more comprehensive or when the analysis
involves a larger memory of dependencies. The LSTM model validates the potential
of approximating the variation in HDFC’s stock price accurately in all the directions
(Figure 2b).
Specifically, the general behavior of the CNN model is to stay close to the trendline
of the underlying stock price data though, at specific moments, it considerably differs
especially during highly volatile moments (Figure 2c). It is very good at capturing bigger
trends but not so great at capturing specific highs and lows.
The resulting GRU model makes for a decent performance which plots the general
trajectory of the stock prices as follows (Figure 2d). However, it also has a weakness; it does
not capture the abrupt changes in the market such as the CNN’s ability to demonstrate a
steep rise and a steep fall. GRUs perform adequately for slight deviations and are highly
reliable but fail to capture the overall extrema concerning price fluctuations.
As seen in the graphs above, the Attention LSTM model is highly accurate in predicting
the trends and fluctuations in HDFC stock prices (Figure 2e). It also predicts actual results
well, especially in areas where the stock has made sharp movements.

4.2. TCS Stock

The model performance for TCS stock prediction is presented in Table 2. The RNN
model exhibited the worst performance with an R-Square of −1.792222, reflecting a poor
FinTech 2024, 3 559

fit to the data. The LSTM and CNN models showed better predictive performance, with
the LSTM model achieving an MAE of 0.286867 and an R-Square of −0.226954. The CNN
model performed similarly, with an MSE of 0.119834. The Attention LSTM model had
the best performance for TCS, with an MAE of 0.275316 and an R-Square of −0.051711,
indicating an improvement over the other models.

Table 2. Model performances on the TCS stock data in terms of the evaluation metrics.

Model MAE MSE RMSE R-Square MAPE

RNN 0.433488 0.292760 0.541073 −1.792222 0.117696
LSTM 0.286867 0.128644 0.358670 −0.226954 0.211829
CNN 0.288108 0.119834 0.346170 −0.142925 0.119429
GRU 0.291298 0.121207 0.348147 −0.156017 0.117161
Attention LSTM 0.275316 0.110270 0.332070 −0.051711 0.110265

The prediction for the models specifically for the TCS stock is shown in Figure 3 below.

(a) RNN prediction (b) LSTM prediction

(c) CNN prediction (d) GRU prediction

(e) LSTM with Attention prediction

Figure 3. Model predictions for the TCS stock.

As seen from the first graph of the RNN model (Figure 3a), there is quite a major
problem with fitting the low and high intensities in predicting the TCS stock prices as they
prominently under and overfit. This behavior shows another problem in the RNN, it does
not work well with patterns in stock data. Because of their structure, although they are
considered a staple for sequential data processing, the problems associated with them, for
FinTech 2024, 3 560

instance, long dependency sequences and vanishing gradients, impact the performance of
the RNNs in dynamic settings like the stock market where precise data are required.
The second graph shows the impact of the LSTM model (Figure 3b) in forecasting the
movements of TCS stock price movements better than RNN. Since the LSTM can retain
information over long periods, owing to its complex gating mechanism, the volatility of
the stock prices is effectively managed by the model. It quickly adapts to sharp density
fluctuations of the price level and provides a closer match to the cyclic patterns inherent in
the stock exchange.
The third graph depicts that the CNN (Figure 3c) has a fairly good performance in
mimicking the overall trend of TCS stock prices but lacks in capturing both the high and low
fluctuations of the prices. CNNs are employed to recognize spatial patterns over the sliding
time windows within time series analysis; these are well-known for that purpose. These
results imply that while moderate and calm patterns are captured correctly by this model,
it does not compare as well with the sudden and sharp patterns that are characteristic of
financial markets.
The fourth graph shows the GRU model (Figure 3d) as it has a similar performance to
the LSTM but is more sensitive to price changes. Like LSTMs, GRUs have a great capability
the handle temporal dependencies, but they have a simpler structure. This attribute makes
the GRU most effective for tick data such as stock prices, as they require the model to
pay attention to relevant information especially when it is new, without the increased
computational requirement of other advanced gating mechanisms.
Interestingly, the Attention LSTM model (Figure 3e) for the investment of the stocks
does not reveal the best alignment with actual stock prices. It fails to capture the minute
details of the change in the data of TCS, along with the broader trends. The addition of
an attention mechanism, which should make the model pay more attention to specific
segments of the input data, does not seem to account for the increase in the model’s accuracy
in this case.

4.3. ICICI Stock

The model performance for ICICI stock prediction is summarized in Table 3. Both
the RNN and LSTM models performed similarly, with an R-Square of 0.138293, showing
moderate predictive accuracy. The CNN model had a slightly lower R-Square of 0.039685,
while the GRU model showed better results than CNN with an R-Square of 0.054675.
Attention LSTM, however, had the weakest performance with an R-Square of −0.026750,
making it the least effective model for ICICI stock prediction.

Table 3. Model performances on the ICICI stock data in terms of the evaluation metrics.

Model MAE MSE RMSE R-Square MAPE

RNN 0.202412 0.064328 0.253630 0.138293 0.362591
LSTM 0.202412 0.064328 0.253630 0.138293 0.062717
CNN 0.205847 0.071689 0.267748 0.039685 0.070235
GRU 0.205716 0.070570 0.265651 0.054675 0.069123
Attention LSTM 0.210881 0.076649 0.276855 −0.026750 0.070589

Figure 4 shows the prediction of the models for the last 10 samples of the ICICI
stock data.
The first graph (Figure 4a) shows the forecasted values with the help of the RNN for
incremental stock prices of ICICI. While it indicates some similarity to actual equity flows,
it also presents important discrepancies including the inability to show more fluctuation.
As a basic recurrent neural network, the given architecture fails to deal with the intrica-
cies and sharp fluctuations that are characteristic of the stock data, which highlights the
inherent weakness of the approach caused by the cardinality of long sequences with the
corresponding problems, such as vanishing gradients.
FinTech 2024, 3 561

(a) RNN prediction (b) CNN prediction

(c) GRU prediction (d) LSTM with attention prediction

Figure 4. Model predictions for the ICICI stock.

As Figure 4b shows, the CNN model is also different from the real ICICI stock prices;
it mimics the general trends, though it may not capture the peaks and troughs of the model
accurately. CNNs, which are remarkable in finding spatial patterns in diverse fields, use
their pattern recognition feature over moving windows in time series. This caused the
CNN to pick up more general trends, although its performance drops when faced with fast,
highly unpredictable oscillations suitable to financial contexts.
The third graph (Figure 4c) below reveals the GRU’s forecast whereby it displays a
better fit of the real stock prices compared to the RNN. Using GRU, the complex form
of LSTM is reduced; the temporal dependencies can still not be ignored as in the case of
applied financial time series data. Its performance here suggests that the network can easily
update information relevant to the market, hence providing a good base for efficient and
effective responses to the market.
The fourth graph (Figure 4d) depicting the Attention LSTM offers an indecisive result.
Although it tries to mimic the general course of the stock value, one can clearly see that it
fails to be precise toward the last data points, producing jagged values that are too high
when approaching an accumulation peak or too low when approaching a trough.

4.4. Reliance Stock

Table 4 summarizes the performance of the models for predicting Reliance stock prices.
Both RNN and LSTM performed poorly, with an R-Square of −1.251171. The CNN model
demonstrated better performance with an MAE of 0.328005 and an R-Square of 0.008383,
while GRU showed slightly lower performance. The best results were achieved by the
Attention LSTM model, with an MAE of 0.317918 and an R-Square of 0.077228, indicating
its effectiveness in predicting Reliance stock prices.

Table 4. Model performances on the Reliance stock data in terms of the evaluation metrics.

Model MAE MSE RMSE R-Square MAPE

RNN 0.406110 0.296743 0.544741 −1.251171 0.296743
LSTM 0.406110 0.296743 0.544741 −1.251171 0.118239
CNN 0.328005 0.130712 0.361541 0.008383 0.131148
GRU 0.333059 0.136330 0.369229 −0.034238 0.135297
Attention LSTM 0.317918 0.121637 0.348765 0.077228 0.132587
FinTech 2024, 3 562

Figure 5 below depicts the performances of the implemented models on the Reliance stock.

(a) RNN prediction (b) LSTM prediction

(c) CNN prediction (d) GRU prediction

Figure 5. Model predictions for the Reliance stock.

Based on the pattern of the graph herein showing the performance of the RNN
(Figure 5a), it is seen that the model correlates well with the stock prices of Reliance but
also shows inconsistencies and underscores the failure of the RNN to depict other moun-
tains and valleys and to undertake a more accurate workout of the dramatic movements
of the Reliance stock prices. Such behavior is inherent to basic RNNs; they are capable of
processing sequential data but suffer from problems of long memory dependencies, or van-
ishing gradients. These drawbacks can significantly reduce their capacity to appropriately
simulate the high level of variation in stock market data.
The performance graph of the CNN (Figure 5c) revealed it as moderately capable of
tracing fluctuations in Reliance’s stock prices. It gives the overall trend with reasonable
efficiency but fails to provide precise estimates for the larger oscillations. CNNs are well-
recognized spatial pattern discriminative models and are used here in a similar fashion to
identify patterns over sliding time windows to detect trends. This performance implies
that although CNNs can work for broader movements, their efficiency is reduced when it
comes to faster less predictable market movements, which are vital in the volatile world of
stock trading.
It can be noticed that the graph of the GRU model resembles real stock prices more
than the RNN and CNN models, specifically in capturing both the overall movement of the
prices and other fluctuations (Figure 5d). Compared to LSTMs, GRUs make some changes
in the architecture but still hold a strong ability to handle temporal dependencies. This
attribute makes them especially useful for financial datasets in which a timely and efficient
ability to respond to new information is a valuable asset. It is perceptible from the graph
that GRUs are capable of capturing the quantitative volatility of stock prices and are thus
promising in representing predictive tasks in finance.
Lastly, the LSTM model graph shows efficiency, synchronizing with the actual Reliance
stock price movements (Figure 5b). It tracks both contraction and expansionary movements
in the markets well and displays high sensitivity to market fluctuations. When enhanced
for temporal sequence learning, LSTMs are good alternatives to RNNs because they have
accomplished the challenge of learning long-term dependencies with their elaborate gating
mechanisms. This capability enables them to perform well when used in places like the
FinTech 2024, 3 563

prediction of stock prices where information on trends can significantly improve upon
the prediction.

4.5. Nifty 50
The performance of models for Nifty 50 stock is summarized in Table 5. Both RNN and
LSTM models showed poor predictive accuracy, with negative R-Square values. The CNN
and GRU models outperformed the others, with CNN achieving an MAE of 0.232070 and
an R-Square of −0.049331. However, the Attention LSTM model performed the worst, with
an R-Square of −1.937318, indicating poor predictive capability for Nifty 50 stock.

Table 5. Model performances on the Nifty50 Index in terms of the evaluation metrics.

Model MAE MSE RMSE R-Square MAPE

RNN 0.252424 0.105110 0.324207 −0.203869 0.9
LSTM 0.246831 0.101502 0.318594 −0.162540 0.8
CNN 0.232070 0.091618 0.302684 −0.049331 0.7
GRU 0.230114 0.090895 0.301488 −0.041058 0.75
Attention LSTM 0.446610 0.256458 0.506417 −1.937318 0.95

The performance of the models for the Nifty50 index data is depicted in Figure 6.

(a) RNN prediction (b) LSTM prediction

(c) CNN prediction (d) GRU prediction

(e) LSTM with attention prediction

Figure 6. Model predictions for the Nifty50 Index.

The above-indicated RNN graph also establishes the model’s general performance in
following the changes in the Nifty50 Index but has weak performances during fluctuations
FinTech 2024, 3 564

(Figure 6a). The forecasts are sometimes too high or too low in relation to concrete changes,
and usually do not capture important highs and lows. They also exposed the basic problems
of RNNs like long-term dependencies and vanishing gradients that hamper its fitness for
use in markets that require precise timely predictions.
It is evidently clear from the graph of the CNN model that the overall trends of the
Nifty50 Index can be tracked with modest accuracy (Figure 6c). Even though it describes
the overall trend, it fails to predict small changes, especially the sharp changes in the market.
CNNs are particularly good at identifying spatial patterns, and when used in time series,
they apply pattern detection across the sliding windows. From this transition method, it
emerged that although CNNs are proficient in identifying trends, they are lacking in agility
in turning market swifts around.
The graph for the GRU model is more consistent with real market data than the RNN
and CNN and corresponds to most of the change in the index (Figure 6d). GRUs make
LSTM structures less complicated and preserve the force of temporal dependency control,
which makes them perfect for application to numerical financial information since timely
changes to new data are essential. What this analysis shows about the GRU’s performance
is that it is valuable in tracking the complex patterns of stock index prices accurately.
Among the proposed LSTM models, the LSTM model graph demonstrates the highest
performance of tracking the actual movement of the Nifty50 Index (Figure 6b). It correctly
aligns with the markets and changes to address their needs appropriately which proves it
has better ways of handling long-term dependencies. This characteristic is particularly rele-
vant in stock market conditions where information patterns in the past are often extremely
powerful for shaping future trends and hence the LSTM is a most promising tool in the
field of financial prediction.
The Attention LSTM’s prediction graph of the Nifty 50 Index offers a clear insight
into how this more complex model goes about forecasting the trends of the stock market
(Figure 6e). Comparing the predictions with the last 50 values of the Nifty50 Index as
depicted in the figure, it can easily be seen that apart from a few fluctuations, the Attention
LSTM commonly predicts the overall directionality of the index with weak accuracy.
The model showed some disparities especially when forecasting the stock’s future trend.

4.6. Model Comparison

Both the CNN and GRU models captured favorable performance for different eq-
uity commodities including the HDFC and the Nifty 50 as a stable stock. The Attention
LSTM model performed brilliantly in the prediction of prices for volatile stocks including
Reliance which evidences that it is capable of comprehending the complex mechanisms
of stock prices in such a scenario. The RNN and LSTM models presented less precise
patterns, whereas LSTM presented, in general, better performance than RNN. CNN has
been developed as the best model for predicting the stock prices and it has been able to
give a compromise of higher accuracy and less time for computations.

5. Conclusions
This study conducted a comprehensive comparative analysis of five deep learning
models—Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Convo-
lutional Neural Network (CNN), Gated Recurrent Unit (GRU), and Attention LSTM—for
predicting the stock prices of five major Indian companies: HDFC, TCS, ICICI, Reliance,
and the Nifty 50 group of stocks. The regression metrics employed for model evaluation
included MAE, MSE, RMSE, and R². The outcomes indicated that model performance var-
ied based on the individual stock being analyzed. CNN and GRU architectures effectively
outperformed RNN and LSTM models, particularly for stable stocks like HDFC and Nifty
50, with the CNN model achieving lower error rates and more robust R² values in most
implementations. The Attention LSTM model proved to be significantly more efficient than
its counterparts for volatile stocks, including Reliance.
FinTech 2024, 3 565

Despite the overall effectiveness of the models, the study revealed notable challenges.
The consistently poor performance of the RNN model highlighted its inability to cap-
ture long-term dependencies, while the LSTM model showed improvements in specific
instances. The Attention LSTM model, although effective for some stocks, performed
poorly with Nifty 50, suggesting that its complexity may not be advantageous for all stock
types. This emphasizes the need for careful consideration of model selection based on the
characteristics of the stock being forecasted.
Although the MAE values across models exceeded 0.2, indicating a significant mean
forecasting error, these findings underscore the complexities inherent to stock price predic-
tions. Future work should explore methods to improve accuracy, including incorporating
external data and refining model hyperparameters. Additionally, addressing potential
biases in model results is crucial; future studies should investigate the impact of biases such
as data selection bias and model complexity bias on prediction accuracy and explore ways
to mitigate these effects. Overall, this study contributes valuable insights to the expanding
research on applying deep learning in forecasting financial time series, offering critical
implications for both academic research and financial practitioners.

Future Work
Several pathways for future research and development arise from this study. First,
there is a strong opportunity to enhance model performance by integrating additional
external data, such as macroeconomic indicators, market news, and social media sentiment
analysis. These external attributes could provide crucial context to stock price movements
and improve models’ ability to detect market dynamics affecting prices.
Second, investigating advanced hyperparameter optimization techniques, including
Bayesian optimization and grid search, could further increase model accuracy. While stan-
dard hyperparameters were applied in this study, fine-tuning them specifically for each
stock may yield better results, especially for more volatile stocks.
Additionally, exploring advanced deep learning models, particularly Transformer
networks, warrants consideration for stock price prediction. The capability of transformer
architectures to utilize self-attention mechanisms and manage extensive data sequences
could enable the better identification of long-term dependencies and trends in stock prices
compared to traditional RNN models. Furthermore, hybrid models that integrate deep
learning with traditional financial analysis methods may enhance both interpretability and
robustness. For example, combining deep learning with econometric models could bridge
the gap between opaque predictions and established financial theories, providing more
actionable insights for investors and traders.
Moreover, comparing our findings from the Indian market with studies conducted in
other capital markets will help identify whether our results align with or differ significantly
from those in other contexts. Future studies could delve deeper into these comparisons,
exploring factors that may influence discrepancies or similarities across different mar-
kets. By focusing on these aspects, we aim to enhance the robustness and applicability
of our models in stock price forecasting and contribute to a broader understanding of
market behavior.
Finally, extending this research to include a broader array of stocks and longer forecast-
ing horizons would provide a more comprehensive understanding of model performance
across different sectors and time periods. Additional investigation into the real-time appli-
cation of these models for stock price predictions and their potential integration with auto-
mated trading systems is essential for assessing their practical utility in financial markets.

Author Contributions: Conceptualization, M.B. and T.K.; methodology, M.B., T.K. and K.R.; software,
T.K. and K.R.; validation, M.B., T.K. and K.R.; formal analysis, M.B. and T.K.; investigation, M.B.,
T.K. and K.R.; resources, M.B. and K.R.; data curation, M.B., T.K. and K.R.; writing—original draft
preparation, M.B. and T.K.; writing—review and editing, M.B., T.K., K.R. and A.M.R.; visualization,
M.B., T.K. and K.R.; supervision, A.M.R.; project administration, M.B. and T.K.; funding acquisition,
A.M.R. All authors have read and agreed to the published version of the manuscript.
FinTech 2024, 3 566

Funding: This research received no external funding.

Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: National Stock Exchange (NSE) of India: www.nseindia.com.
Conflicts of Interest: The authors declare no conflicts of interest.

Abbreviations
The following abbreviations are used in this manuscript:

RNN Recurrent Neural Networks

LSTM Long Short-Term Memory
CNN Convolutional Neural Netoworks
GRU Gated Recurrent Unit
MAE Mean Absolute Error
MSE Mean Squared Error
RMSE Root Mean Squared Error
R2 R-Squared (Coefficient of Determination)
SVM Support Vector Machine
RF Random Forest
ANN Artificial Neural Network
ARIMA Autoregressive Integrated Moving Average
GARCH Generalized Autoregressive Conditional Heteroskedasticity
NSE National Stock Exchange (India)

References
1. Sachdeva, A.; Jethwani, G.; Manjunath, C.; Balamurugan, M.; Krishna, A.V.N. An Effective Time Series Analysis for Equity
Market Prediction Using Deep Learning Model. In Proceedings of the 2019 International Conference on Data Science and
Communication (IconDSC), Bangalore, India, 1–2 March 2019; pp. 1–5. [CrossRef]
2. Selvamuthu, D.; Kumar, V.; Mishra, A. Indian stock market prediction using artificial neural networks on tick data. Financ. Innov.
2019, 5, 16. [CrossRef]
3. Singh, J. Indian Stock Markets Data Analysis and Prediction using Machine Learning. Int. J. Res. Appl. Sci. Eng. Technol. 2020,
8, 631–636. [CrossRef]
4. Fathali, Z.; Kodia, Z.; Said, L.B. Stock Market Prediction of NIFTY 50 Index Applying Machine Learning Techniques. Appl. Artif.
Intell. 2022, 36, 2111134. [CrossRef]
5. Piyush; Amarjeet; Sharma, A.; Kumar, S.; Ansari, N.N. Stock Price Prediction Using Machine Learning. In Proceedings of the
International Conference on Recent Trends in Computing, Jammu, India, 26–27 October 2023; Springer: Berlin/Heidelberg,
Germany, 2023; pp. 79–87. [CrossRef]
6. Weng, B.; Lu, L.; Wang, X.; Megahed, F.M.; Martinez, W. Predicting short-term stock prices using ensemble methods and online
data sources. Expert Syst. Appl. 2018, 112, 258–273. [CrossRef]
7. Kanwal, A.; Lau, M.F.; Ng, S.P.; Sim, K.Y.; Chandrasekaran, S. BiCuDNNLSTM-1dCNN—A hybrid deep learning-based predictive
model for stock price prediction. Expert Syst. Appl. 2022, 202, 117123. [CrossRef]
8. Parray, I.R.; Khurana, S.S.; Kumar, M.; Altalbe, A.A. Time series data analysis of stock price movement using machine learning
techniques. Soft Comput. 2020, 24, 16509–16517. [CrossRef]
9. Gupta, D.; Parikh, A.; Datta, T.K. A multi-criteria decision-making approach to rank the sectoral stock indices of national stock
exchange of India based on their performances. Natl. Account. Rev. 2021, 3, 272–292. [CrossRef]
10. Ab.Khalil, M.; Bakar, A.A. A Comparative Study of Deep Learning Algorithms in Univariate and Multivariate Forecasting of the
Malaysian Stock Market. Sains Malays. 2023, 52, 993–1009. [CrossRef]
11. Kumar, T.; Park, J.; Ali, M.S.; Uddin, A.S.; Ko, J.H.; Bae, S.H. Binary-classifiers-enabled filters for semi-supervised learning.
IEEE Access 2021, 9, 167663–167673. [CrossRef]
12. Chandio, A.; Gui, G.; Kumar, T.; Ullah, I.; Ranjbarzadeh, R.; Roy, A.M.; Hussain, A.; Shen, Y. Precise single-stage detector. arXiv
2022, arXiv:2210.04252.
13. Roy, A.M.; Bhaduri, J.; Kumar, T.; Raj, K. WilDect-YOLO: An efficient and robust computer vision-based accurate object
localization model for automated endangered wildlife detection. Ecol. Inform. 2023, 75, 101919. [CrossRef]
14. Kumar, T.; Mileo, A.; Bendechache, M. KeepOriginalAugment: Single Image-based Better Information-Preserving Data Augmen-
tation Approach. In Proceedings of the 20th International Conference on Artificial Intelligence Applications and Innovations,
Corfu, Greece, 27–30 June 2024.
FinTech 2024, 3 567

15. Vavekanand, R.; Sam, K.; Kumar, S.; Kumar, T. CardiacNet: A Neural Networks Based Heartbeat Classifications using ECG Signals.
Stud. Med. Health Sci. 2024, 1, 1–17. [CrossRef]
16. Raj, K.; Kumar, T.; Mileo, A.; Bendechache, M. OxML Challenge 2023: Carcinoma classification using data augmentation. arXiv
2024, arXiv:2409.10544.
17. Baea, S.H. Class Specific Autoencoders Enhance Sample Diversity. J. Broadcast Eng. 2021, 26, 844–854.
18. Kumar, T.; Mileo, A.; Brennan, R.; Bendechache, M. Advanced Data Augmentation Approaches: A Comprehensive Survey and
Future directions. arXiv 2023, arXiv:2301.02830.
19. Kumar, T.; Singh, A.; Raj, K.; Roy, A.M. Efficient Paddy Grains Quality Assessment Approach Utilizing Affordable Sensors. AI
2024, 5, 686–703. [CrossRef]
20. Park, J.; Kumar, T.; Bae, S.H. Search of an optimal sound augmentation policy for environmental sound classification with deep
neural networks. In Proceedings of the Korean Society of Broadcast Engineers Conference; The Korean Institute of Broadcast and Media
Engineers: Seoul, Republic of Korea, 2020; pp. 18–21.
21. Kumar, T.; Park, J.; Bae, S.H. Intra-Class Random Erasing (ICRE) augmentation for audio classification. In Proceedings of the
Korean Society of Broadcast Engineers Conference; The Korean Institute of Broadcast and Media Engineers: Seoul, Republic of Korea
2020; pp. 244–247.
22. Khan, W.; Zhang, C.; Luo, B.; Kumar, T.; Ahmed, E. Robust Partitioning Scheme for Accelerating SQL Database. In Proceedings
of the 2021 IEEE International Conference on Emergency Science and Information Technology (ICESIT), Chongqing, China, 22–24
November 2021; pp. 369–376.
23. Kumar, T.; Turab, M.; Mileo, A.; Bendechache, M.; Saber, T. AudRandAug: Random Image Augmentations for Audio Classification.
arXiv 2023, arXiv:2309.04762.
24. Turab, M.; Kumar, T.; Bendechache, M.; Saber, T. Investigating multi-feature selection and ensembling for audio classification.
arXiv 2022, arXiv:2206.07511. [CrossRef]
25. Kumar, T.; Turab, M.; Talpur, S.; Brennan, R.; Bendechache, M. Forged Character Detection Datasets: Passports, Driving Licences
and Visa Stickers. Int. J. Artif. Intell. Appl. 2022, 13, 2. [CrossRef]
26. Raj, K.; Singh, A.; Mandal, A.; Kumar, T.; Roy, A.M. Understanding EEG signals for subject-wise definition of armoni activities.
arXiv 2023, arXiv:2301.00948.
27. Khan, W.; Kumar, T.; Cheng, Z.; Raj, K.; Roy, A.; Luo, B. SQL and NoSQL Databases Software architectures performance analysis
and assessments—A Systematic Literature review. Big Data Cogn. Comput. 2023, 7, 97. [CrossRef]
28. Aleem, S.; Kumar, T.; Little, S.; Bendechache, M.; Brennan, R.; McGuinness, K. Random data augmentation based enhancement:
A generalized enhancement approach for medical datasets. In Proceedings of the 24th Irish Machine Vision and Image Processing
(IMVIP) Conference, Belfast, UK, 31 August–2 September 2022.
29. Roy, A.M.; Bhaduri, J.; Kumar, T.; Raj, K. A computer vision-based object localization model for endangered wildlife detection.
Ecol. Econ. Forthcom. 2022. Available online: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4315295 (accessed on 12
October 2024).
30. Singh, A.; Raj, K.; Kumar, T.; Verma, S.; Roy, A.M. Deep learning-based cost-effective and responsive robot for autism treatment.
Drones 2023, 7, 81. [CrossRef]
31. Ranjbarzadeh, R.; Jafarzadeh Ghoushchi, S.; Tataei Sarshar, N.; Tirkolaee, E.B.; Ali, S.S.; Kumar, T.; Bendechache, M. ME-CCNN:
Multi-encoded images and a cascade convolutional neural network for breast tumor segmentation and recognition. Artif. Intell.
Rev. 2023, 56, 10099–10136. [CrossRef]
32. Patel, J.; Shah, S.; Thakkar, P.; Kotecha, K. Predicting stock and stock price index movement using Trend Deterministic Data
Preparation and machine learning techniques. Expert Syst. Appl. 2015, 42, 259–268. [CrossRef]
33. Salehinejad, H.; Sankar, S.; Barfett, J.; Colak, E.; Valaee, S. Recent Advances in Recurrent Neural Networks. arXiv 2018,
arXiv:1801.01078.
34. Shen, J.; Shafiq, M.O. Short-term stock market price trend prediction using a comprehensive deep learning system. J. Big Data
2020, 7, 66. [CrossRef]
35. Sen, J.; Mehtab, S. A Robust Predictive Model for Stock Price Prediction Using Deep Learning and Natural Language Processing.
arXiv 2021, arXiv:1912.07700. [CrossRef]
36. Hernández, A.; Amigó, J.M. Attention Mechanisms and Their Applications to Complex Systems. Entropy 2021, 23, 283. [CrossRef]
37. Sunday, D.O. Application of Long Short-Term Memory (LSTM) in Stock Price Prediction. Int. J. Dev. Econ. Sustain. 2024, 12, 36–45.
[CrossRef]
38. Pipin, S.J.; Purba, R.; Kurniawan, H. Prediksi Saham Menggunakan Recurrent Neural Network (RNN-LSTM) dengan Optimasi
Adaptive Moment Estimation. J. Comput. Syst. Inform. (JoSYC) 2023, 4, 806–815. [CrossRef]
39. Vijh, S.; Pandey, A.K.; Vijh, G.; Kumar, S. Stock Forecasting for Time Series Data using Convolutional Neural Network.
In Proceedings of the 2021 11th International Conference on Cloud Computing, Data Science & Engineering (Confluence), Noida,
India, 28–29 January 2021; pp. 866–870. [CrossRef]
40. Satria, D. Predicting Banking Stock Prices Using Rnn, Lstm, and Gru Approach. Appl. Comput. Sci. 2023, 19, 82–94. [CrossRef]
41. Chaudhary, P. A Deep Learning-based Approach for Stock Price Prediction. Turk. J. Comput. Math. Educ. (TURCOMAT) 2020,
11, 1983–1989. [CrossRef]
FinTech 2024, 3 568

42. Long, W.; Lu, Z.; Cui, L. Deep learning-based feature engineering for stock price movement prediction. Knowl. Based Syst. 2019,
164, 163–173. [CrossRef]
43. Paul, M.K.; Das, P. A Comparative Study of Deep Learning Algorithms for Forecasting Indian Stock Market Trends. Int. J. Adv.
Comput. Sci. Appl. 2023, 14. [CrossRef]
44. Hemajothi, S.; Babu, S.S.; Prathiksha, V.; Sangeetha, B.; Sivani, E.; Soundari, K. Prediction of Stock Price Model through the
Implementation of Hybrid BiLSTM-GRU. In Proceedings of the 2024 International Conference on Recent Advances in Electrical,
Electronics, Ubiquitous Communication, and Computational Intelligence (RAEEUCCI), Chennai, India, 17–18 April 2024; pp. 1–6.
[CrossRef]
45. Sudriyanto, S.; Faid, M.; Malik, K.; Supriadi, A. Evaluasi Model Jaringan Saraf Tiruan Berbasis LSTM dalam Memprediksi
Fluktuasi Harga Bitcoin. J. Adv. Res. Inform. 2024, 2, 15–22. [CrossRef]
46. Jafar, S.H.; Akhtar, S.; El-Chaarani, H.; Khan, P.A.; Binsaddig, R. Forecasting of NIFTY 50 Index Price by Using Backward
Elimination with an LSTM Model. J. Risk Financ. Manag. 2023, 16, 423. [CrossRef]
47. Polepally, V.; Reddy, N.S.N.; Sindhuja, M.; Anjali, N.; Reddy, K.J. A Deep Learning Approach for Prediction of Stock Price
Based on Neural Network Models: LSTM and GRU. In Proceedings of the 2021 12th International Conference on Computing
Communication and Networking Technologies (ICCCNT), Kharagpur, India, 6–8 July 2021; pp. 1–4. [CrossRef]
48. Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; Salakhutdinov, R. Dropout: A Simple Way to Prevent Neural Networks
from Overfitting. J. Mach. Learn. Res. 2014, 15, 1929–1958.
49. Mehtab, S.; Sen, J. Stock Price Prediction Using Convolutional Neural Networks on a Multivariate Time Series. arXiv 2021,
arXiv:2001.09769.
50. Sen, J.; Mehtab, S.; Nath, G. Stock Price Prediction Using Deep Learning Models. TechRxiv 2021. Available online: https:
//www.techrxiv.org/doi/full/10.36227/techrxiv.16640197.v1 (accessed on 12 October 2024).
51. Ahmed, S.F.; Alam, M.S.B.; Hassan, M.; Rozbu, M.R.; Ishtiak, T.; Rafa, N.; Mofijur, M.; Ali, A.B.M.S.; Gandomi, A.H. Deep learning
modelling techniques: Current progress, applications, advantages, and challenges. Artif. Intell. Rev. 2023, 56, 13521–13617.
[CrossRef]
52. Spuler, M.; Sarasola-Sanz, A.; Birbaumer, N.; Rosenstiel, W.; Ramos-Murguialday, A. Comparing metrics to evaluate performance
of regression methods for decoding of neural signals. In Proceedings of the 2015 37th Annual International Conference of the
IEEE Engineering in Medicine and Biology Society (EMBC), Milan, Italy, 25–29 August 2015; pp. 1083–1086. [CrossRef]

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