A STOCK PRICE PREDICTION MODEL BASED ON

INVESTOR SENTIMENT AND OPTIMIZED DEEP LEARNING

KAVURI VENKATA GOPI1, PG Student - MOHAMMAD ISMAIL KHAN2,

MCA, Department of MCA, Chirala Assistant Professor, Department of
Engineering College, Chirala. CSE ,Chirala Engineering College, Chirala,
[email protected] [email protected]

Abstract: The research presents the MS-SSA-LSTM overall predictive performance. To facilitate user
model, which integrates multi-source data, sentiment interaction and testing, a user-friendly Flask
analysis, swarm intelligence algorithms, and deep framework with SQLite support was developed,
learning techniques to enhance stock price streamlining signup, signin, and model evaluation
predictions. This model incorporates sentiment processes.
analysis from East Money forum posts, creating a
Index terms - Deep learning, LSTM model, stock
unique sentiment dictionary and calculating a
price prediction, sentiment analysis, sentiment
sentiment index. This offers valuable insights into
dictionary, sparrow search algorithm.
market sentiment's influence on stock prices. The
Sparrow Search Algorithm (SSA) is used to fine-tune
1. INTRODUCTION
LSTM hyperparameters, optimizing prediction
accuracy. • Experimental results showcase the With the maturity of China’s stock market and the
MS-SSA-LSTM model's superior performance. It's a rapid growth of Internet finance, many people realize
valuable tool for accurate stock price predictions. the importance of investment and choose to enter the
Tailored for China's volatile financial market, the financial market. However, the stock market is
model excels in short-term stock price predictions, characterized by massive data and enormous
offering insights for dynamic decision-making by volatility. Many retail investors need more data-
investors. And also, a hybrid LSTM+GRU model mining skills to make money. Therefore, accurate
was introduced for stock sentiment classification. stock price prediction can reduce investment risks
Additionally, a robust ensemble strategy was and improve investment returns for investors and
adopted, incorporating a Voting Classifier (AdaBoost enterprises.
+ RandomForest) for sentiment analysis and a Voting
Early scholars used statistical methods to construct a
Regressor (LinearRegression +
linear model to fit the stock price time series trend.
RandomForestRegressor + KNeighborsRegressor) for
The traditional methods contain ARMA, ARIMA,
stock price prediction. These ensembles seamlessly
GARCH, etc. The ARMA is established to conduct a
integrated with existing models (MLP, CNN, LSTM,
time series stock analysis [1]. The ARIMA model is
MS-LSTM, MS-SSA-LSTM), collectively enhancing
developed based on the ARMA and predicts the trend experimental data, ANN has the benefits of quick
of stock price changes [2]. The ARIMA model can convergence and high accuracy [8], [9], [10].
also introduce wavelet analysis to improve the fitting Moghaddam and Esfandyari [11] evaluated the effect
accuracy of the Shanghai Composite Index [3]. The of several feedforward artificial neural networks on
GARCH model provides innovative ideas for stock the market stock price forecast through experiments.
time series prediction through a time window [4]. At Liu and Hou [12] improved the BP (Back
the same time, some scholars have combined ARMA Propagation) neural network using the Bayesian
and GARCH to build a new prediction model, which regularization method. Nevertheless, the traditional
provided theoretical support for the volumetric price neural network method has the following areas for
analysis of multivariate stocks [5]. Generally improvement. Generalization ability is not strong,
speaking, these classical methods only capture quickly leads to overfitting, and falls into local
regular and structured data. However, traditional optimization. Since many samples need to be trained,
forecasting methods require assumptions that are better models must be found to solve these problems.
uncommon in real life. Therefore, It is challenging to
A new model for predicting stock prices is proposed
describe nonlinear financial data using statistical
in this paper (MS-SSA-LSTM), which matches the
methods.
characteristics of multi-source data with LSTM
Subsequently, many researchers attempt to anticipate neural networks and uses the Sparrow Search
stock prices using machine learning approaches such Algorithm. The MS-SSA-LSTM stock price forecast
as Support Vector Machines (SVM) and Neural model can forecast the stock price in advance and
Networks. Machine learning’s core idea is to use help investors and traders make more informed
algorithms to parse data, learn from it, and make investment decisions. Investors and traders obtain the
predictions about new data. Because the SVM shows data of individual stocks they want to invest in,
unique benefits in dealing with limited samples, high- including historical transaction data and comment
dimensional data, and nonlinear situations, many information of stock market shareholders, and input
scholars use it in stock forecasting. Hossain and them into the MS-SSA-LSTM model. The model
Nasser [6] found that the SVM method is superior to automatically outputs a stock price trend chart and
the statistical ones in stock prediction accuracy. Chai forecasts the stock price for the next day
et al. [7] suggested a hybrid SVM model to anticipate
2. LITERATURE SURVEY
the HS300 index’s ups and downs and found that the
least squares SVM combined with the Genetic
The presence and changes in, long memory features
Algorithm (GA) performed better. However, when
in the returns and volatility dynamics of S&P 500 and
the SVM applies to large-scale training samples,
London Stock Exchange using ARMA model [1].
much memory and computing time will be consumed,
Recently, multifractal analysis has been evolved as an
which may limit its development space in predicting
important way to explain the complexity of financial
a large amount of stock data. Then, Artificial Neural
markets which can hardly be described by linear
Networks (ANN) and multi-layer ANN address
methods of efficient market theory. In financial
financial time series issues. According to the
markets, the weak form of the efficient market minimize their risk by portfolio diversification. Gold
hypothesis implies that price returns are serially earlier was only purchased at the time of marriage or
uncorrelated sequences. In other words, prices should other rituals in India but now it has gained
follow a random walk behavior. The random walk importance in the eyes of investors also, so it has
hypothesis is evaluated against alternatives become necessary to predict the price of Gold with
accommodating either unifractality or multifractality. suitable method.
Several studies find that the return volatility of stocks
The GARCH model and its numerous variants have
tends to exhibit long-range dependence, heavy tails,
been applied widely both in the financial literature
and clustering. Because stochastic processes with
and in practice. For purposes of quasi maximum
self-similarity possess long-range dependence and
likelihood estimation, innovations to GARCH
heavy tails, it has been suggested that self-similar
processes are typically assumed to be identically and
processes be employed to capture these
independently distributed, with mean zero and unit
characteristics in return volatility modeling. The
variance (strong GARCH) [4]. Under less restrictive
present study applies monthly and yearly forecasting
assumptions (the absence of unconditional
of Time Series Stock Returns in S&P 500 and
correlation, weak GARCH), higher order dependency
London Stock Exchange using ARMA model. [1]
patterns might be exploited for the ex ante forecasting
The statistical analysis of S&P 500 shows that the
of GARCH innovations, and hence, stock returns. In
ARMA model for S&P 500 outperforms the London
this paper, rolling windows of empirical stock returns
stock exchange and it is capable for predicting
are used to test the independence of consecutive
medium or long horizons using real known values.
GARCH innovations. Rolling -values from
The statistical analysis in London Stock Exchange
independence testing reflect the time variation of
shows that the ARMA model for monthly stock
serial dependence, and provide useful information for
returns outperforms the yearly. A comparison
signaling one-step-ahead directions of stock price
between S&P 500 and London Stock Exchange
changes. Ex ante forecasting gains are documented
shows that both markets are efficient and have
for nonparametric innovation predictions, especially
Financial Stability during periods of boom and bust.
if the sign of the innovation predictors is combined
The study gives an inside view of the application of with independence diagnostics ( -values) and/or the
ARIMA time series model to forecast the future Gold sign of linear return forecasts.
price in Indian browser based on past data from
In the recent years, the use of GARCH type
November 2003 to January 2014 to mitigate the risk
(especially, ARMA-GARCH) models and
in purchases of gold. Hence, to give guideline for the
computational-intelligence-based techniques—
investor when to buy or sell the yellow metal. [2]This
Support Vector Machine (SVM) and Relevance
financial instrument has gained a lot of momentum in
Vector Machine (RVM) have been successfully used
recent past as Indian economy is curbed with factors
for financial forecasting. [2,6]This paper deals with
like changing political scenario, global clues & high
the application of ARMA-GARCH, recurrent SVM
inflation etc, so researcher, investors and speculators
(RSVM) and recurrent RVM (RRVM) in volatility
are in search of different financial instrument to
forecasting. Based on RSVM and RRVM, two sentiment analysis, and swarm intelligence
GARCH methods are used and are compared with algorithms. [14,15,16,30] By optimizing LSTM
parametric GARCHs (Pure and ARMA-GARCH) in hyperparameters with the Sparrow Search Algorithm,
terms of their ability to forecast multi-periodically. the system excels in forecasting stock prices with
These models are evaluated on four performance exceptional accuracy. Experimental results affirm its
metrics: MSE, MAE, DS, and linear regression R superiority over other models, underlining its
squared. The real data in this study uses two Asian universal applicability and potential to enhance
stock market composite indices of BSE SENSEX and predictive performance. This model is compared with
NIKKEI225. This paper also examines the effects of MLP, CNN, LSTM, MS-LSTM. And also, a hybrid
outliers on modeling and forecasting volatility. Our LSTM+GRU model was introduced for stock
experiment shows that both the RSVM and RRVM sentiment classification. Additionally, a robust
perform almost equally, but better than the GARCH ensemble strategy was adopted, incorporating a
type models in forecasting. The ARMA-GARCH Voting Classifier (AdaBoost + RandomForest) for
model is superior to the pure GARCH and only the sentiment analysis and a Voting Regressor
RRVM with RSVM hold the robustness properties in (LinearRegression + RandomForestRegressor +
forecasting. KNeighborsRegressor) for stock price prediction.
These ensembles seamlessly integrated with existing
This paper proposes an EMD-LSSVM (empirical
models (MLP, CNN, LSTM, MS-LSTM, MS-SSA-
mode decomposition least squares support vector
LSTM), collectively enhancing overall predictive
machine) model to analyze the CSI 300 index. A
performance. To facilitate user interaction and
WD-LSSVM (wavelet denoising least squares
testing, a user-friendly Flask framework with SQLite
support machine) is also proposed as a benchmark to
support was developed, streamlining signup, signin,
compare with the performance of EMD-LSSVM [7].
and model evaluation processes.
Since parameters selection is vital to the performance
of the model, different optimization methods are ii) System Architecture:
used, including simplex, GS (grid search), PSO
(particle swarm optimization), and GA (genetic The initial step is to import datasets, including the
algorithm). Experimental results show that the EMD- Stock Tweets Dataset, Single Stock Data, and Multi-
LSSVM model with GS algorithm outperforms other Source Data. These datasets serve as the foundation
methods in predicting stock market movement for both sentiment analysis and stock price
direction. prediction. Text data from the Stock Tweets Dataset
undergoes cleaning, which includes removing
3. METHODOLOGY punctuations, HTML tags, URLs, and emojis. This
step ensures the text is ready for sentiment analysis.
i) Proposed Work:
The Single Stock Data and Multi-Source Data are

The project introduces the MS-SSA-LSTM model, a processed to handle null values, remove duplicates,

cutting-edge system for stock price prediction. This and scale the data. This prepares the financial data for

model seamlessly integrates multi-source data, stock price prediction. Several models, including
MLP, CNN, LSTM, MS-LSTM, MS-SSA-LSTM, how social media affects stock prices and market
extensions- Voting Classifier, and LSTM + GRU, are trends to help investors and traders.
trained for sentiment classification. They analyze the So, these are the top 5 rows of the dataset
cleaned tweet data to determine market sentiment.
Another set of models, including MLP, CNN, LSTM,
MS-LSTM, MS-SSA-LSTM, and extension- Voting
Regression, are trained for stock price prediction.
They utilize processed financial data to forecast stock
prices. After the models are trained, they are used to
make predictions. In the case of sentiment analysis,
Fig 2 Stock tweets dataset
predictions provide insights into market sentiment.
For stock price prediction, the models forecast future ALL STOCK DATASET
stock prices. The predictions from sentiment analysis The "All Stock Dataset" is a comprehensive
and stock price models play a crucial role in aiding collection of financial data from various sources. It
investors and traders in making informed decisions. provides a wealth of information for in-depth stock
The combined results help users navigate the market research. In our project, we used this dataset
complex landscape of the stock market, reduce risks, to enhance our stock price prediction model. We
and optimize investment returns. aimed to improve the accuracy of stock price
forecasts by leveraging diverse data sources,
ultimately benefitting investors and businesses.
THIS IS THE SAMPLE DATASET

Fig 3 All stock dataset

Fig 1 Proposed architecture
iv) Data Processing:
iii) Dataset collection:
Data processing involves transforming raw data into
STOCK TWEETS DATASET valuable information for businesses. Generally, data
scientists process data, which includes collecting,
The "Stock Tweets" dataset includes posts about
organizing, cleaning, verifying, analyzing, and
stocks and financial markets on social media. We
converting it into readable formats such as graphs or
used it to understand people's feelings and reactions
documents. Data processing can be done using three
to market news [1,4,7,8]. This helped us create tools
methods i.e., manual, mechanical, and electronic. The
for stock trading and investments. We wanted to see
aim is to increase the value of information and
facilitate decision-making. This enables businesses to
improve their operations and make timely strategic adjusted during training to optimize the network's
decisions. Automated data processing solutions, such ability to learn complex patterns in data. The final
as computer software programming, play a output layer generates predictions or classifications.
significant role in this. It can help turn large amounts MLPs are used in a wide range of applications, from
of data, including big data, into meaningful insights image recognition to financial forecasting, owing to
for quality management and decision-making. their capacity to model intricate relationships in data.

v) Feature selection:

Feature selection is the process of isolating the most

consistent, non-redundant, and relevant features to
use in model construction. Methodically reducing the
size of datasets is important as the size and variety of
datasets continue to grow. The main goal of feature
Fig 4 MLP
selection is to improve the performance of a
predictive model and reduce the computational cost A Convolutional Neural Network (CNN) is a type
of modeling. of deep learning model suitable for various data
beyond images. It processes data through layers that
Feature selection, one of the main components of
apply convolutions and pooling operations, enabling
feature engineering, is the process of selecting the
the network to automatically learn relevant patterns
most important features to input in machine learning
or features within the data. This makes CNNs
algorithms. Feature selection techniques are
valuable for tasks involving sequential data or grids,
employed to reduce the number of input variables by
such as time series analysis or structured data
eliminating redundant or irrelevant features and
processing. They excel at capturing intricate
narrowing down the set of features to those most
relationships and hierarchies, contributing to their
relevant to the machine learning model. The main
versatility in different domains, including natural
benefits of performing feature selection in advance,
language processing and financial predictions.
rather than letting the machine learning model figure
out which features are most important.

vi) Algorithms:

A Multilayer Perceptron (MLP) operates by

processing data through a series of layers. It begins
with an input layer that receives data and proceeds to
hidden layers, where each neuron calculates a
weighted sum of inputs, applies an activation Fig 5 CNN
function for non-linearity, and passes the result to the
next layer. These weights between neurons are
A Long Short-Term Memory (LSTM) is a type of system in scenarios where diverse data sources play a
recurrent neural network (RNN) designed for critical role.
sequential data analysis. Unlike traditional RNNs,
LSTMs are adept at capturing and preserving
dependencies over long sequences, making them
ideal for tasks where data points have complex,
distant relationships. LSTMs utilize specialized
memory cells and gates that enable them to
remember, update, or forget information, facilitating
precise modeling of sequential patterns. This has
found applications in various fields, including natural
language processing, speech recognition, and
financial time series analysis, where understanding Fig 7 MS-LSTM

historical context and predicting future trends are

The MS-SSA-LSTM model, or Multi-Source
crucial.
Sparrow Search Algorithm Long Short-Term
Memory, represents a sophisticated approach to stock
price prediction. It combines multi-source data from
various origins, employs sentiment analysis, and
optimizes the Long Short-Term Memory (LSTM)
network using the Sparrow Search Algorithm (SSA).
This advanced model effectively addresses the
challenges of financial forecasting by offering a more
accurate and robust way to predict stock prices. It
Fig 6 LSTM outperforms conventional models and holds high
universal applicability, making it a valuable tool for
The Multi-Source Long Short-Term Memory investors and enterprises operating in dynamic
(MS-LSTM) is an extended variant of the traditional financial markets.
LSTM neural network designed to process data from
various sources simultaneously. It excels at handling
comprehensive information by integrating data inputs
from multiple origins, making it particularly valuable
for complex tasks such as stock price prediction.
[30,32] MS-LSTM enhances the model's capacity to
capture and analyze intricate dependencies and
patterns by leveraging a broad range of data, thus
improving the overall predictive capabilities of the

Fig 8 MS-SSA-LSTM
The Voting Regressor is an ensemble machine
learning technique that combines the predictions of
multiple regression algorithms to improve predictive
performance. In this case, it incorporates three
Fig 10 LSTM + GRU
diverse regressors: Linear Regression, Random
Forest Regressor, and k-Neighbors Regressor. By The Voting Classifier is a key component for
aggregating their individual predictions, it aims to sentiment classification in this project, combining the
create a more accurate and robust model for strengths of AdaBoost and Random Forest (RF)
regression tasks. This approach leverages the [18,39]. It harnesses AdaBoost's boosting
strengths of each base regressor, such as the linearity capabilities, where multiple weak learners are
of Linear Regression, the adaptability of Random combined to form a strong classifier, and RF's
Forest, and the proximity-based learning of k- ensemble learning approach, which aggregates
Neighbors Regression, to enhance overall predictive predictions from multiple decision trees. By
capabilities. integrating these two techniques, the Voting
Classifier enhances the accuracy and robustness of
sentiment classification, making it a powerful tool for
analyzing market sentiment in our research.

Fig 11 Voting classifier

Fig 9 Voting Regressor

4. EXPERIMENTAL RESULTS
The LSTM+GRU is an advanced recurrent neural
network (RNN) architecture that combines the
Precision: Precision evaluates the fraction of
capabilities of Long Short-Term Memory (LSTM)
correctly classified instances or samples among the
and Gated Recurrent Unit (GRU) cells. It enhances
ones classified as positives. Thus, the formula to
the model's ability to capture sequential patterns in
calculate the precision is given by:
data by leveraging both LSTM's memory retention
and GRU's computational efficiency. This Precision = True positives/ (True positives + False
combination is particularly effective for tasks positives) = TP/(TP + FP)
involving time series data, natural language
processing, and sequential pattern recognition, as it
addresses the limitations of each cell type
individually, resulting in improved performance and
training efficiency.
Fig 12 Precision comparison graph

Recall: Recall is a metric in machine learning that

measures the ability of a model to identify all
relevant instances of a particular class. It is the ratio Fig 14 Accuracy graph
of correctly predicted positive observations to the
total actual positives, providing insights into a F1 Score: The F1 Score is the harmonic mean of
model's completeness in capturing instances of a precision and recall, offering a balanced measure that
given class. considers both false positives and false negatives,
making it suitable for imbalanced datasets.

Fig 15 F1Score
Fig 13 Recall comparison graph

Accuracy: Accuracy is the proportion of correct

predictions in a classification task, measuring the
overall correctness of a model's predictions.
Fig 16 Performance Evaluation

Fig 17 Home page

Fig 19 Login page

Fig 20 User input

Fig 18 Signin page Fig 21 Result

Fig 22 Graphs
 interface. The MS-SSA-LSTM model and its
extensions offer valuable insights, reducing
investment risks, and enhancing decision-making in
the dynamic landscape of the Chinese financial
market.

6. FUTURE SCOPE
Fig 23 Graphs
Expanding the model's capabilities to handle real-
5. CONCLUSION
time data feeds can enable investors to make even
more timely decisions. Integrating data sources that
The project aimed to enhance stock market
provide up-to-the-minute information could be a
predictions, with a focus on the MS-SSA-LSTM
valuable addition. [34] Further refining the sentiment
model. The research explored various models,
analysis component by incorporating natural
emphasizing the significance of sentiment analysis
language processing (NLP) techniques and
and innovative algorithms for optimized forecasting
sentiment-specific machine learning models can
[26]. The MS-SSA-LSTM model stood out for its
provide a more nuanced understanding of market
dual proficiency in stock price prediction and
sentiment. Exploring and integrating data from
sentiment classification. Leveraging diverse data
diverse sources, such as social media, news feeds,
sources and advanced techniques, it offered a
and macroeconomic indicators, can offer a
comprehensive approach to risk reduction and
comprehensive view of the market and potentially
improved returns. Existing models (MLP, CNN,
improve predictive accuracy. Developing tools or
LSTM, MS-LSTM) demonstrated competence, while
features that offer explanations for the model's
the MS-SSA-LSTM model showcased superiority,
predictions can make it more transparent and user-
particularly in short-term predictions for China's
friendly. Investors may benefit from understanding
dynamic market. Ensemble models (Voting
the reasons behind specific forecasts. Extending the
Classifier, LSTM+GRU, Voting Regressor)
model's capabilities to include risk assessment and
introduced in the extension phase expanded the
portfolio optimization can provide investors with a
predictive toolkit. LSTM+GRU excelled in sentiment
holistic approach to managing their investments. This
classification, and the Voting Regressor
could involve considering the diversification of assets
outperformed in stock price prediction, contributing
and risk-adjusted returns.
reliable alternatives. The Flask extension facilitated
user-friendly interaction, allowing input of ticker
REFERENCES
symbols for accurate predictions. LSTM+GRU for
sentiment and Voting Regressor for stock price [1] M. M. Rounaghi and F. N. Zadeh, ‘‘Investigation
predictions were seamlessly deployed, enhancing of market efficiency and financial stability between
accessibility for users and investors. Investors, S&P 500 and London stock exchange: Monthly and
traders, and businesses stand to benefit from the yearly forecasting of time series stock returns using
project's robust predictive models and user-friendly ARMA model,’’ Phys. A, Stat. Mech. Appl., vol.
456, pp. 10–21, Aug. 2016, doi: Comput. Appl., vol. 124, no. 12, pp. 11–15, Aug.
10.1016/j.physa.2016.03.006. 2015, doi: 10.5120/ijca2015905681.

[2] G. Bandyopadhyay, ‘‘Gold price forecasting [9] D. Banjade, ‘‘Forecasting Bitcoin price using
using ARIMA model,’’ J. Adv. Manage. Sci., vol. 4, artificial neural network,’’ Jan. 2020, doi:
no. 2, pp. 117–121, 2016, doi: 10.2139/ssrn.3515702.
10.12720/joams.4.2.117-121.
[10] J. Zahedi and M. M. Rounaghi, ‘‘Application of
[3] H. Shi, Z. You, and Z. Chen, ‘‘Analysis and artificial neural network models and principal
prediction of Shanghai composite index by ARIMA component analysis method in predicting stock prices
model based on wavelet analysis,’’ J. Math. Pract. on Tehran stock exchange,’’ Phys. A, Stat. Mech.
Theory, vol. 44, no. 23, pp. 66–72, 2014. Appl., vol. 438, pp. 178–187, Nov. 2015, doi:
10.1016/j.physa.2015.06.033.
[4] H. Herwartz, ‘‘Stock return prediction under
GARCH—An empirical assessment,’’ Int. J. [11] A. H. Moghaddam, M. H. Moghaddam, and M.
Forecasting, vol. 33, no. 3, pp. 569–580, Jul. 2017, Esfandyari, ‘‘Stock market index prediction using
doi: 10.1016/j.ijforecast.2017.01.002. artificial neural network,’’ J. Econ., Finance
Administ. Sci., vol. 21, no. 41, pp. 89–93, Dec. 2016,
[5] H. Mohammadi and L. Su, ‘‘International
doi: 10.1016/j.jefas.2016.07.002.
evidence on crude oil price dynamics: Applications
of ARIMA-GARCH models,’’ Energy Econ., vol. 32, [12] H. Liu and Y. Hou, ‘‘Application of Bayesian
no. 5, pp. 1001–1008, Sep. 2010, doi: neural network in prediction of stock time series,’’
10.1016/j.eneco.2010.04.009. Comput. Eng. Appl., vol. 55, no. 12, pp. 225–229,
2019.
[6] A. Hossain and M. Nasser, ‘‘Recurrent support
and relevance vector machines based model with [13] A. M. Rather, A. Agarwal, and V. N. Sastry,
application to forecasting volatility of financial ‘‘Recurrent neural network and a hybrid model for
returns,’’ J. Intell. Learn. Syst. Appl., vol. 3, no. 4, prediction of stock returns,’’ Expert Syst. Appl., vol.
pp. 230–241, 2011, doi: 10.4236/jilsa.2011.34026. 42, no. 6, pp. 3234–3241, Apr. 2015, doi:
10.1016/j.eswa.2014.12.003.
[7] J. Chai, J. Du, K. K. Lai, and Y. P. Lee, ‘‘A
hybrid least square support vector machine model [14] A. Sherstinsky, ‘‘Fundamentals of recurrent
with parameters optimization for stock forecasting,’’ neural network (RNN) and long short-term memory
Math. Problems Eng., vol. 2015, pp. 1–7, Jan. 2015, (LSTM) network,’’ Phys. D, Nonlinear Phenomena,
doi: 10.1155/2015/231394. vol. 404, Mar. 2020, Art. no. 132306, doi:
10.1016/j.physd.2019.132306.
[8] A. Murkute and T. Sarode, ‘‘Forecasting market
price of stock using artificial neural network,’’ Int. J. [15] G. Ding and L. Qin, ‘‘Study on the prediction of
stock price based on the associated network model of
LSTM,’’ Int. J. Mach. Learn. Cybern., vol. 11, no. 6, [22] J. Borade, ‘‘Stock prediction and simulation of
pp. 1307–1317, Nov. 2019, doi: 10.1007/s13042-019- trade using support vector regression,’’ Int. J. Res.
01041-1. Eng. Technol., vol. 7, no. 4, pp. 52–57, Apr. 2018,
doi: 10.15623/ijret.2018.0704009.
[16] X. Yan, W. Weihan, and M. Chang, ‘‘Research
on financial assets transaction prediction model based [23] X. Li and P. Tang, ‘‘Stock price prediction based
on LSTM neural network,’’ Neural Comput. Appl., on technical analysis, fundamental analysis and deep
vol. 33, no. 1, pp. 257–270, May 2020, doi: learning,’’ Stat. Decis., vol. 38, no. 2, pp. 146–150,
10.1007/s00521-020- 04992-7. 2022, doi: 10.13546/j.cnki.tjyjc.2022.02.029.

[17] M. Nabipour, P. Nayyeri, H. Jabani, A. Mosavi, [24] J. Heo and J. Y. Yang, ‘‘Stock price prediction
E. Salwana, and S. Shahab, ‘‘Deep learning for stock based on financial statements using SVM,’’ Int. J.
market prediction,’’ Entropy, vol. 22, no. 8, p. 840, Hybrid Inf. Technol., vol. 9, no. 2, pp. 57–66, Feb.
Jul. 2020, doi: 10.3390/e22080840. 2016, doi: 10.14257/ijhit.2016.9.2.05.

[18] Z. D. Aksehir and E. Kiliç, ‘‘How to handle data [25] J. B. De Long, A. Shleifer, L. H. Summers, and
imbalance and feature selection problems in CNN- R. J. Waldmann, ‘‘Noise trader risk in financial
based stock price forecasting,’’ IEEE Access, vol. 10, markets,’’ J. Political Economy, vol. 98, no. 4, pp.
pp. 31297–31305, 2022, doi: 10.1109/ACCESS.2022. 703–738, Aug. 1990, doi: 10.1086/261703.
3160797.
[26] H. Cui and Y. Zhang, ‘‘Does investor sentiment
[19] Y. Ji, A. W. Liew, and L. Yang, ‘‘A novel affect stock price crash risk?’’ Appl. Econ. Lett., vol.
improved particle swarm optimization with long- 27, no. 7, pp. 564–568, Jul. 2019, doi:
short term memory hybrid model for stock indices 10.1080/13504851.2019.1643448.
forecast,’’ IEEE Access, vol. 9, pp. 23660–23671,
[27] R. P. Schumaker, Y. Zhang, C.-N. Huang, and
2021, doi: 10.1109/ACCESS.2021.3056713.
H. Chen, ‘‘Evaluating sentiment in financial news
[20] X. Zeng, J. Cai, C. Liang, and C. Yuan, ‘‘A articles,’’ Decis. Support Syst., vol. 53, no. 3, pp.
hybrid model integrating long short-term memory 458–464, Jun. 2012, doi: 10.1016/j.dss.2012.03.001.
with adaptive genetic algorithm based on individual
[28] M. Nofer and O. Hinz, ‘‘Using Twitter to predict
ranking for stock index prediction,’’ PLoS ONE, vol.
the stock market,’’ Bus. Inf. Syst. Eng., vol. 57, no. 4,
17, no. 8, Aug. 2022, Art. no. e0272637, doi:
pp. 229–242, Jun. 2015, doi: 10.1007/s12599- 015-
10.1371/journal.pone.0272637.
0390-4.
[21] J. Xue and B. Shen, ‘‘A novel swarm
[29] P. Fan, Y. Yang, Z. Zhang, and M. Chen, ‘‘The
intelligence optimization approach: Sparrow search
relationship between individual stock investor
algorithm,’’ Syst. Sci. Control Eng., vol. 8, no. 1, pp.
sentiment and stock yield-based on the perspective of
22–34, Jan. 2020, doi:
10.1080/21642583.2019.1708830.
stock evaluation information,’’ Math. Pract. Theory, analysis of online news,’’ World Wide Web, vol. 17,
vol. 51, no. 16, pp. 305–320, 2021. no. 4, pp. 723–742, Jun. 2013, doi: 10.1007/s11280-
013-0221-9.
[30] Z. Jin, Y. Yang, and Y. Liu, ‘‘Stock closing
price prediction based on sentiment analysis and [37] L. Yang and T. Zhai, ‘‘Research on sentiment
LSTM,’’ Neural Comput. Appl., vol. 32, no. 13, pp. tendency analysis of video reviews based on
9713–9729, Sep. 2019, doi: 10.1007/s00521-019- sentiment dictionary,’’ Netw. Secur. Technol. Appl.,
04504-2. vol. 255, no. 3, pp. 53–56, 2022.

[31] X. Xu and K. Tian, ‘‘A novel financial text [38] A. Fathy, T. M. Alanazi, H. Rezk, and D.
sentiment analysis-based approach for stock index Yousri, ‘‘Optimal energy management of micro-grid
prediction,’’ J. Quantum Technol. Econ., vol. 38, no. using sparrow search algorithm,’’ Energy Rep., vol.
12, pp. 124–145, 2021, doi: 8, pp. 758–773, Nov. 2022, doi: 10.1016/j.egyr.2021.
10.13653/j.cnki.jqte.2021.12.009. 12.022.

[32] C.-R. Ko and H.-T. Chang, ‘‘LSTM-based [39] Y. Chen, Z. Liu, C. Xu, X. Zhao, L. Pang, K. Li,
sentiment analysis for stock price forecast,’’ PeerJ and Y. Shi, ‘‘Heavy metal content prediction based
Comput. Sci., vol. 7, p. e408, Mar. 2021, doi: on random forest and sparrow search algorithm,’’ J.
10.7717/peerj-cs.408. Chemometrics, vol. 36, no. 10, Sep. 2022, Art. no.
e3445, doi: 10.1002/cem.3445.
[33] Y. Li and Y. Pan, ‘‘A novel ensemble deep
learning model for stock prediction based on stock [40] S. Hochreiter and J. Schmidhuber, ‘‘Long short-
prices and news,’’ Int. J. Data Sci. Anal., vol. 13, no. term memory,’’ Neural Comput., vol. 9, no. 8, pp.
2, pp. 139–149, Sep. 2021, doi: 10.1007/s41060-021- 1735–1780, Nov. 1997, doi:
00279-9. 10.1162/neco.1997.9.8.1735.

[34] C. Kearney and S. Liu, ‘‘Textual sentiment in

finance: A survey of methods and models,’’ Int. Rev.
Financial Anal., vol. 33, pp. 171–185, May 2014, doi:
10.1016/j.irfa.2014.02.006.

[35] T. Wang and Z. Zhang, ‘‘Research on the

construction method of emotional lexicon for movie
review,’’ Comput. Digit. Eng., vol. 50, no. 4, pp.
843–848, 2022, doi: 10.3969/j.issn.1672-
9722.2022.04.031.

[36] Y. Rao, J. Lei, L. Wenyin, Q. Li, and M. Chen,

‘‘Building emotional dictionary for sentiment