2024 8th International Conference on Computing, Communication, Control and Automation (ICCUBEA)

Pimpri Chinchwad College of Engineering (PCCOE), Pune, India. Aug 23-24, 2024

Comparative Analysis of Stock Price Prediction

using Time Series Models
2024 8th International Conference on Computing, Communication, Control and Automation (ICCUBEA) | 979-8-3503-9177-0/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICCUBEA61740.2024.10775112

Harsh Notaria Shriya Shah Devarshee Thopte

Department of Computer Engineering Department of Computer Engineering Department of Computer Engineering
Dwarkadas J. Sanghvi College of Dwarkadas J. Sanghvi College of Dwarkadas J. Sanghvi College of
Engineering Engineering Engineering
Mumbai, India Mumbai, India Mumbai, India
[email protected] [email protected] [email protected]

Hemang Soneji Prof. Pranit Bari Prof. Khushali Deulkar

Department of Computer Engineering Department of Computer Engineering Department of Computer Engineering
Dwarkadas J. Sanghvi College of Dwarkadas J. Sanghvi College of Dwarkadas J. Sanghvi College of
Engineering Engineering Engineering
Mumbai, India Mumbai, India Mumbai, India
[email protected] [email protected] [email protected]

Abstract— The inconsistency of the stock market is The objective of this project is to explore different
ungovernable, making it prone to imminent financial risks. Our prediction techniques to predict future stock returns using
research indicates that the system displays remarkable precision historical returns and numerical news indicators. The aim is to
in producing prognostics based on historical data derived from construct a diversified portfolio of multiple stocks to minimize
Yahoo Finance in the form of graphs and values. Resorting to risk. To accomplish this, we analyze the seemingly
our model, financial analysts and investors can alleviate risk and unpredictable market data and utilize various prediction
capitalize on market opportunities, facilitating informed algorithms and conventional statistical models for forecasting
decision-making. Our study’s primary focus is evaluating the
stock prices.
following four models for predicting stock market prices: Long
Short-Term Memory, Auto-Regressive Integrated Moving II. LITERATURE SURVEY
Average, Seasonal Auto-Regressive Integrated Moving Average
using eXogenous factors, and Auto-Regressive Moving Average. [1] conducts an analysis of strategies using predictive
ARIMA utilizes differencing to stabilize non-stationary time algorithms and Machine Learning architectures. This provides
series, while ARMA amalgamates autoregressive and moving a model-independent approach to stock price prediction.
average components to record linear dependencies. SARIMAX Based on the LSTM architecture of RNN, their model is
strengthens ARIMA by integrating seasonal fluctuations, capable of classifying and predicting time series data. The
suitable for modeling seasonal trends in financial data. Deep RMSE is estimated by this methodology using rigorous
learning techniques like LSTM excel at recognising complex mathematical research, which increases forecast accuracy.
temporal correlations. These findings showcase the importance
of utilizing advanced forecasting techniques to navigate the An experiment is carried out by [2] regarding the closing
ever-changing financial landscape, offering valuable insights for stock prices of 25 businesses that are formally registered on
professional and academic purposes. the Indonesia Stock Exchange (IDX). They put forth an
approach to trading: when to purchase or when to sell, by
Keywords— LSTM, RNN, ARMA, ARIMA, SARIMAX, Time utilizing the LSTM and Extreme Gradient Boosting
Series Analysis. (XGBoost) and defining two thresholds. Their 99%
performance in XGBoost was the best they could muster.
I. INTRODUCTION
With the use of LSTM, [3] has reached a maximum
It is said that the stock market resembles a device that accuracy of 99%. It forecasts stock market prices to help
transfers money from the impatient to the patient. In today's investors make better informed and accurate choices.
fast-paced world, patience being a virtue, is often overlooked
while choosing the right stock. Our approach addresses this by In [4], an LSTM model is fitted to data using Python tools
empowering individuals to make judicious, long-term that automatically download previous market data in order to
decisions while navigating the market. predict future stock prices.
The research aims to analyse the predictive competence of [5] shows how to combine recurrent neural networks to
the four models, covering various financial datasets. develop an algorithm that forecasts changes in the price of
Conventional models such as ARMA and ARIMA do a stocks. About 3.58% more is expected for them than for the
satisfactory job of capturing linear dependencies and actual values.
stabilizing trends, with SARIMAX further improving by
giving consideration to seasonal variations. On the other hand, The LSTM RNN is employed in [6] to analyze the trend
LSTM stands out in recognising non-linear correlations in of stock price in Build Your Dreams company, which is the
sequential data. Through panoramic evaluations and studies, leading energy vehicle manufacturer in China. With general
we aim to project the pros and cons of the respective models recursive neural networks, it resolves the lengthy sequence
in stock price prediction. Ultimately, our goal is to provide dependence issue. The input, output, and forgetting gates
financial analysts and investors with valuable insights to make increase the financial time series prediction's accuracy.
informed decisions amidst volatile market conditions.

979-8-3503-9177-0/24/$31.00 ©2024 IEEE 1

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on January 17,2025 at 09:59:18 UTC from IEEE Xplore. Restrictions apply.
 [7] selected Apple Inc. adjusted closing prices for the III. DATASET
period spanning 2018 to 2019. Next, using the Autoregressive We have taken the data from Yahoo Finance which is a
Moving Average model predicted the closing prices of Apple popular financial platform that provides comprehensive
Inc. for the upcoming five days. The main difference made in financial information, including stock market data, news, and
the original data was making the sequence stationary. This analysis. Yahoo Finance provided historical stock price data
model is appropriate for short-term price prediction when the for the last 20 years that was used in this study. The report
anticipated time series is compared with the actual value and contained daily stock prices for well-known companies
reveals that the error rates of the data are minimal. including Apple, Google, Microsoft, Facebook, and others. To
Concurrently, it demonstrates how the time series model is a help readers analyze this time series data, Yahoo Finance
useful stimulant for finance research. provides us with information and a variety of tools, including
[8] outlines a hybrid approach to stock price forecasting technical indicators, charting tools, news, and expert
that combines the Markov model, BPNN (back propagation commentary from the financial industry.
Neural Network), and ARMA. The nonlinear and linear
constituents of the stock price series are solved by ARMA and IV. METHODOLOGY
Back Propagation Neural Network respectively. The outcome Proposed pipeline for Comparative Analysis of Stock
might be enhanced by the Markov Model. Price Prediction using Time Series Models
A unique stock market signal is examined and evaluated
in the paper [9] utilizing an ARMA (Autoregressive Moving
Average) model with a distinct number of poles and zeros. The
model forecasts the price for the very following day.
[10] shows the design of the ARIMA model and the
support-vector machine-aided XGBoost to estimate the price
of the Shanghai 300 index and Shenzhen inventory. SVM
(Support Vector Machine) yields an accuracy of 0.98 and a
recall rate of 0.95 for the XGBoost model.
The results obtained in [11] showed that the ARIMA
model has an increased latency for short-term estimation and
can compete approvingly. The model was applied to stock
data from the New York Stock Exchange (NYSE) along with
the Nigerian Stock Exchange (NSE).
[12] employs a seasonal ARIMA model of (0,1,1) (0,1,2)
12 which forecasts index values of NIFTY 50 by using a
historical dataset. Auto-correlation functions (ACF & PACF)
have been used to test the train testing strategy of ML and AIC
on eight years' worth of monthly data.
Ultimately, it is determined that [13] is not able to predict
the real-time stock market value of the Apple corporation
based on research done using SARIMA models with AIC Fig. 1. Pipeline
scores of 580.165, 451.591, and 114.612. The research also
found a MAPE score of 36.05%. The machine learning models implemented in this
[14] is based on SARIMA and XGBoost and employs a research were LSTM, ARMA, ARIMA and SARIMAX.
ten-year stock data set from Yahoo Finance to create A. Long Short Term Memory
predictions. XGBoost is used to implement gradient boosted
decision trees that are designed for execution and speed. It LSTM is an RNN (Recurrent Neural Network) which can
considers the factors such as the open and adjusted end of day be abbreviated for long term relationships in sequential data,
price, peak of the day, low of the day, and the totality in which makes forecasting time series more convenient. This
volume. There were three results: an accuracy level of specific model architecture is sometimes appropriate for stock
89.48%, a 15.612% mean-absolute error (MAE), and a mean- price prediction and delivers more accurate forecasts, due to
absolute percentage error came out as 10.52%. its ability to recognise complex patterns within former price
sequences.
A website that displays the anticipated LSTM graph of the
stock that a user wants to search for was made available to all The dataset used in the LSTM model is scaled and then
types of users [15]. It has a news section with updates and divided into sets of training and testing. A collection of closing
information about the stock market as well as the variables prices and their target values is designed using this data.
influencing stock market developments. However, because the LSTM model accepts only NumPy
arrays, the datasets are transformed to the concerned format
In [16], SARIMAX was used to forecast previously and restructured.
specified elements from 2021 to 2050 in Saudi Arabia. The
training data was utilized from the previous 40 years (1980 Two 128-unit and 64-unit LSTM layers substantiated by
2020). It proved the consideration and categorization of two Dense Layers from the Keras framework are incorporated
SARIMAX performed better than ARIMA. into the LSTM model. The Adam optimiser and the MSE loss

2
Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on January 17,2025 at 09:59:18 UTC from IEEE Xplore. Restrictions apply.
function is used to institute the model. Later, the model is using Matplotlib the forecasted and actual stock price patterns
trained for one epoch on the training set. are plotted.
B. Autoregressive Integrated Moving Average D. Seasonal Autoregressive Integrated Moving Average
The analysis of data patterns and temporal relationships is with eXogenous factors
enhanced by the largely used ARIMA time series forecasting SARIMA is an amplified variation of ARIMA that
model. Due to its combination of moving average (MA) and predicts time series by maneuvering both seasonal and non-
autoregressive (AR) components, it is used for predicting seasonal changes. SARIMAX, built upon the SARIMA
stock prices and helps to find the underlying patterns and model, is used for recording compound time series with the
fluctuations that are usual for stock price data. To know the help of seasonal patterns and exogenous variables. Exogenous
temporal trends and patterns that enable highlights of future variables in SARIMAX are external factors which increase
price movements, the model analyzes previous stock price accuracy of the prediction algorithm.
data to forecast future stock values. By using previous stock
prices, the model gives investors helpful information to make The SARIMAX general formula is given below:
decisions about buying, selling, and holding stocks. SARIMAX(p,d,q)(P,D,Q)m:
To predict stock prices, the ARIMA model has various Yt = c + ࢥ1 + Yt-1 + ࢥ2 + Yt-2 + … + ࢥp + Yt-p + ࣅt –
components like differencing, autoregressive components, ș1ࣅt-1 – ș2ࣅt-2 –… – șqࣅt-q įXt + ȖZt + ࢥ1Yt-m + ࢥ2Yt-
and moving averages. The differencing component of 2m + … + ࢥPYt-Pm – ș1ࣅt-m – ș2ࣅt-2m – … – șQࣅt-Qm (2)
ARIMA promises to be constant by removing trends, whereas
the AR component shows the connection between an Where:
observation and lag observations. The moving average x t – time,
component of the model tells us the link between the residual x Zt – other exogenous variables, if present,
errors and an observation of the model. x Yt – value of time series at t,
x Xt – exogenous variable at t,
C. Autoregressive Moving Average
x c – constant term,
Separating the autoregressive and moving average x q – number of MA terms,
components from the differencing component is what
x p – number of AR terms,
condensed variants of ARIMA, like ARMA, incorporate.
Although ARMA is not as complex as ARIMA, it can x m – seasonal period,
nevertheless identify some of the time series patterns present x D – degree of seasonal differencing,
in stock market price data, thus making it a good choice for x șș«șT– MA parameters,
predicting jobs. x ࢥࢥ«ࢥS– AR parameters,
x P – number of seasonal AR terms,
Historical stock price data retrieved from Yahoo Finance
is divided into testing and training sets, the ARMA (Auto x Q – number of seasonal MA terms,
Regressive Moving Average) model approach repeatedly x į– coefficient for Xt,
finds the ideal parameters (p and q) by minimizing the mean x Ȗ– coefficient for Zt,
squared error (MSE) using grid search. The equation is given x ࣅt – error in terms of white noise at t.
as:
The implemented SARIMAX model consists of the
following parameters:
Yt = c + ࢥ1 + Yt-1 + ࢥ2 + Yt-12 + … + ࢥp + Yt-p + ࣅt - ș1ࣅt-1 - ș2ࣅt-
2- … - șqࣅt-q (1) x Order: (1,1,0) are the non-seasonal parameters of
ARIMA. Here, the model is an ARIMA with p = 1, d =
Where: 1, q = 0.
x t – time, x Trend = 'n,' suggesting that there is no trend component
x Yt – time series value at t, in our model.
x c – constant, x Temporal Sequence: (3,0,3,4)
x ARIMA seasonal order: depicts seasonal ARIMA
x q – count of MA terms,
values. This is a SARIMAX (3,0,3) model with P = 3, D
x p – count of AR terms,
= 0, Q = 3, and m = 4 (the time steps count in each
x ࣅt – error in terms of white noise at t,
seasonal period is denoted by m). It specifies three
x șș«șT– moving-average parameters, moving average terms, three autoregressive terms, and a
x ࢥࢥ«ࢥS– autoregressive parameters. seasonal period of 4.
Where the white noise error term is depicted by ࣅtșLDQG The model is first set up with appropriate parameters to
ࢥi are moving average coefficients, autoregressive record both seasonal as well as non-seasonal constituents of
coefficients, while c is a constant. At time t, X is the observed the time series data ('ts_decompose'). These parameters
value. Forecasts for the testing period are then produced by consist of a seasonal order of (3,0,3,4) which shows the
applying the best-fit model to the training set. Performance seasonal autoregressive term, the seasonal differences term,
indicators namely mean absolute percentage error (MAPE), and the seasonal moving average term along with a seasonal
root mean-squared error (RMSE), and mean-squared error period of 4, with no sign of trend ({trend='n'}), and an ARIMA
(MSE) are applied. The prediction accuracy of the model is order of (1,1,0) for moving average, differencing, and
evaluated using MAPE and accuracy percentage. Finally, autoregressive components.

3
Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on January 17,2025 at 09:59:18 UTC from IEEE Xplore. Restrictions apply.
 After fitting the model to the data (ts_decompose),
parameters are estimated to produce a prediction structure that
is specific to the features of the time series. On the basis of
this model, projections are then produced that provide insights
into potential future patterns in stock prices.
Additionally, the algorithm quantifies forecast accuracy
over real stock prices by assessing the SARIMAX algorithm's
efficacy using the Root-Mean-Squared Error (RMSE) and
mean absolute percent error (MAPE) metrics.
V. RESULTS
The ideal model is the model that yields the most accurate
results for better investment suggestions and profits for the
user’s finances. The metrics that we have used to evaluate the
accuracy for each model are MAPE, MSE, and RMSE. We Fig. 4. Stock price prediction graph for Apple dataset using SARIMAX
model
have also plotted the stock price prediction graphs, helping the
user to find the robustness and reliability of the models. The
metrics that we have used will help any investor in
understanding the predictive trends to invest better and get
better returns. As shown below, the accuracy of the
SARIMAX model emerged as the highest out of the four
models when tried on the data of different companies like
Apple, Google, Microsoft, and Amazon.
In the above model comparative study SARIMAX
displays the highest accuracy of 98.54%, proving to be the
prime model out of the four. In contrast to LSTM, ARMA, and
ARIMA, SARIMAX showed its ability to handle the
intricacies of financial time series data, including seasonal
components and exogenous regressors, by demonstrating
increased prediction accuracy. SARIMAX is an appealing
alternative for investors and financial analysts seeking
credible forecasts to aid decision-making due to its
outstanding accuracy, which displays its reliability and
robustness in projecting stock prices. The results of our study Fig. 5. Accuracy comparison of the time series models we tested.
show that SARIMAX is the most effective model for stock
price prediction tasks and offers a strong tool for reliably and
confidently negotiating the complexities of financial markets.

Fig. 2. Stock price prediction graph for Apple dataset using LSTM model.

Fig. 6. Accuracy comparison of our model (SARIMAX) with the existing

models. comparing each time series model with their output
parameters.

TABLE I. COMPARING EACH TIME SERIES MODEL WITH THEIR

OUTPUT PARAMETERS

LSTM ARMA ARIMA SARIMAX

MSE 17.13 929.05 130.38 3.28
MAPE 3.6% 13.5% 7.5% 1.46%
RMSE 4.14 30.48 11.42 1.81
ACCURACY 96.4% 86.5% 93.67% 98.54%
Fig. 3. Stock price prediction graph for Apple dataset using ARIMA
model.

4
Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on January 17,2025 at 09:59:18 UTC from IEEE Xplore. Restrictions apply.
 TABLE II. COMPARING THE ACCURACY OF OUR MODEL WITH enabling them to make wise investment choices, investors
EXISTING MODELS
may be informed about current events to improve the aspect
SARIMAX (OUR SARIMA + SARIMA of financial security. The NEWS API may be integrated to do
MODEL) XGBOOST
this. Hybrid model integration is a potential aspect that can be
ACCURACY 98.54% 89.48% 63.95%
performed to increase the prediction accuracy. Another future
TABLE III. COMPARING SARIMAX ACCURACY OBTAINED FOR
gap can be bridged by adding a portfolio optimization feature
DIFFERENT COMPANIES which helps by investing in the appropriate assets that offer
significant returns. This enhances the investment strategies
COMPANY APPLE GOOGLE MICROSOFT AMAZON
NAME allowing maximum returns and minimizing the risks. Trading
ACCURACY 98.54% 98.17% 98.54% 97.75% bots for the stock market can be built on to automate the
(USING process conceding smooth trading.
SARIMAX)

REFERENCES
[1] K. Sutradhar, S. Sutradhar, I. Ahmed Jhimel, S. Kumar Gupta and M.
Monirujjaman Khan, "Stock Market Prediction using Recurrent Neural
Network’s LSTM Architecture," 2021 IEEE 12th Annual Ubiquitous
Computing, Electronics & Mobile Communication Conference
(UEMCON), New York, NY, USA, 2021, pp. 0541-0547, doi:
10.1109/UEMCON53757.2021.9666562.
[2] A. B. Gumelar et al., "Boosting the Accuracy of Stock Market
Prediction using XGBoost and Long Short-Term Memory," 2020
International Seminar on Application for Technology of Information
and Communication (iSemantic), Semarang, Indonesia, 2020, pp. 609-
613, doi: 10.1109/iSemantic50169.2020.9234256.
[3] K. J, H. E, M. S. Jacob and D. R, "Stock Price Prediction Based on
LSTM Deep Learning Model," 2021 International Conference on
System, Computation, Automation and Networking (ICSCAN),
Puducherry, India, 2021, pp. 1-4, doi:
10.1109/ICSCAN53069.2021.9526491.
[4] P. Srivastava and P. K. Mishra, "Stock Market Prediction Using RNN
LSTM," 2021 2nd Global Conference for Advancement in Technology
(GCAT), Bangalore, India, 2021, pp. 1-5, doi:
10.1109/GCAT52182.2021.9587540.
[5] A. K. S. C. Chua, E. D. Pua, E. P. Dadios and R. K. C. Billones, "AI
To Predict Price Movements in the Stock Market," 2021 IEEE 13th
Fig. 7. Graphical representation of the SARIMAX accuracy comparison International Conference on Humanoid, Nanotechnology, Information
for different companies. Technology, Communication and Control, Environment, and
Management (HNICEM), Manila, Philippines, 2021, pp. 1-6, doi:
10.1109/HNICEM54116.2021.9731955.
VI. CONCLUSION [6] Y. Zhang and G. M. Tumibay, "Stock Market Prediction Model Based
The comparative analysis between the time series on LSTM Deep Learning: The Case of Top Corporate Company in
prediction models LSTM, ARMA, ARIMA, and SARIMAX China," 2022 4th International Conference on Artificial Intelligence
and Advanced Manufacturing (AIAM), Hamburg, Germany, 2022, pp.
for stock price forecasting calls attention to the significance of 451-455, doi: 10.1109/AIAM57466.2022.00092.
choosing a pertinent model to handle financial data. Every [7] H. Tang, "Stock Prices Prediction Based on ARMA Model," 2021
model shown has the potential to encapsulate specific aspects International Conference on Computer, Blockchain and Financial
of the temporal dynamics observed in stock market pricing. Development (CBFD), Nanjing, China, 2021, pp. i-iv, doi:
Nevertheless, our illations reveal that SARIMAX is the 10.1109/CBFD52659.2021.00046.
optimal choice, offering the highest accuracy. SARIMAX's [8] Shuzhen Shi, Wenlong Liu and Minglu Jin, "Stock price forecasting
ability to account for both autoregressive and seasonal using a hybrid ARMA and BP neural network and Markov model,"
2012 IEEE 14th International Conference on Communication
components, along with its capability to handle non-stationary Technology, Chengdu, 2012, pp. 981-985, doi:
data, proved its worth in achieving superior predictive 10.1109/ICCT.2012.6511341.
performance. [9] M. F. Anaghi and Y. Norouzi, "A model for stock price forecasting
based on ARMA systems," 2012 2nd International Conference on
However, the capability of LSTM to identify long term Advances in Computational Tools for Engineering Applications
dependencies and nonlinear interactions opens up interesting (ACTEA), Beirut, Lebanon, 2012, pp. 265-268, doi:
areas for further study. SARIMAX has emerged as one of the 10.1109/ICTEA.2012.6462880.
few leaders in the field of prediction model development, and [10] Y. Wang and Z. Wang, "Stock Price Volatility Prediction in Financial
as stock markets continue to change, the quest for strong and Big Data on XGBoost and ARIMA Models," 2023 International
trustworthy models will always be ongoing. SARIMAX is Conference on Ambient Intelligence, Knowledge Informatics and
Industrial Electronics (AIKIIE), Ballari, India, 2023, pp. 1-5, doi:
more flexible for forecasting jobs including both seasonal 10.1109/AIKIIE60097.2023.10390131.
patterns and external impacts than SARIMA, which is limited [11] A. A. Ariyo, A. O. Adewumi and C. K. Ayo, "Stock Price Prediction
to modeling seasonal patterns within a time series. SARIMAX Using the ARIMA Model," 2014 UKSim-AMSS 16th International
also includes the option to incorporate external variables that Conference on Computer Modelling and Simulation, Cambridge, UK,
may impact the target variable. 2014, pp. 106-112, doi: 10.1109/UKSim.2014.67.
[12] K. Pant, R. Rawat and A. Pant, "Forecasting of Nifty 50 Index
VII. FUTURE SCOPE Volatility Using Seasonal ARIMA Model," 2023 International
Conference on Advancement in Computation & Computer
With 98.54% correct forecasts, the SARIMAX-based Technologies (InCACCT), Gharuan, India, 2023, pp. 271-274, doi:
stock price prediction algorithm that has been provided is 10.1109/InCACCT57535.2023.10141726.
quite dependable and ensures safe investments. In addition to

5
Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on January 17,2025 at 09:59:18 UTC from IEEE Xplore. Restrictions apply.
[13] Daryl, A. Winata, S. Kumara and D. Suhartono, "Predicting Stock
Market Prices using Time Series SARIMA," 2021 1st International
Conference on Computer Science and Artificial Intelligence (ICCSAI),
Jakarta, Indonesia, 2021, pp. 92-99, doi:
10.1109/ICCSAI53272.2021.9609720.
[14] D. S. Kumar, T. B. C, S. N. R. C, V. A, A. S. Devi and D. Kavitha,
"Analysis and Prediction of Stock Price Using Hybridization of
SARIMA and XGBoost," 2022 International Conference on
Communication, Computing and Internet of Things (IC3IoT), Chennai,
India, 2022, pp. 1-4, doi: 10.1109/IC3IOT53935.2022.9767868.
[15] R. Gandhi, G. Anand, V. Vekaria, T. Ahuja and L. D’Mello, "Stock
Market Graph Prediction using LSTM," 2023 14th International
Conference on Computing Communication and Networking
Technologies (ICCCNT), Delhi, India, 2023, pp. 1-4, doi:
10.1109/ICCCNT56998.2023.10306549.
[16] Alharbi, Fahad & Csala, Dénes. (2022). A Seasonal Autoregressive
Integrated Moving Average with Exogenous Factors (SARIMAX)
Forecasting Model-Based Time Series Approach. Inventions. 7. 94.
10.3390/inventions7040094.

6
Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on January 17,2025 at 09:59:18 UTC from IEEE Xplore. Restrictions apply.