Exploring the Impact of Financial News Sentiment

on Stock Price Forecasting: A Comparative Deep

Learning Approach
1st Abhishek Rajhans 2nd Trinanjan Das
Computer Science and Engineering Computer Science and Engineering
Indian Institute of Information Technology Ranchi, India Indian Institute of Information Technology Ranchi, India
[email protected] [email protected]

3rd Anish Kumar 4th Bam Bahadur Sinha

Computer Science and Engineering Computer Science and Engineering
Indian Institute of Information Technology Ranchi, India National Institute of Technology Sikkim, India
[email protected] [email protected]

Abstract—The application of Artificial Intelligence (AI) in This forecasting challenge can be categorized into three
stock price prediction has demonstrated significant advance- distinct timeframes:
ments, with Machine Learning and Deep Learning techniques
proving highly efficient in this domain. Two widely adopted a. Short-term forecasting, encompasses predictions over brief
architectures for stock price prediction are the Long Short Term periods such as seconds, minutes, hours, few days, weeks,
Memory (LSTM) and Gated Recurrent Unit (GRU) models. or months.
Recognizing the potential impact of financial news sentiment on b. Medium-term forecasting, extending the prediction horizon
forecasting, this study investigates whether incorporating such to one to two years.
sentiment yields superior results compared to relying solely on
historical stock prices. Accurate predictions in the financial c. Long-term forecasting, involving predictions over periods
market are challenging due to its inherent high volatility and exceeding two years.
non-linear nature. To contribute insights into the effectiveness Numerous models and techniques have been developed and
of various deep learning architectures, a comparative analysis
utilized for stock price prediction. Our focus lies within time
was conducted on Long Short-Term Memory (LSTM), Gated
Recurrent Unit (GRU), Convolutional Neural Network (CNN), series forecasting, broadly categorized into two classes:
and the Prophet model. The evaluation focused on three com- a. Linear Models: This encompasses ARIMA [1] and its
panies listed on the National Stock Exchange (NSE). The study variations, including SARIMA [2]. These models utilize
employs the Mean Squared Error (MSE) metric to rigorously
assess and compare the models’ performance. The results of the
predefined equations to fit mathematical models to univari-
comparative analysis provide valuable implications for precision ate time series data. Our paper will delve into the ARIMA
in stock price prediction within the context of market volatility model, discussed in a subsequent section.
and non-linearity. The study adds to the growing body of research b. Non-Linear Models: This category includes deep learning
on the application of AI in financial forecasting, emphasizing algorithms, GARCH [3], and others. Deep learning algo-
the importance of considering sentiment analysis alongside tra-
ditional approaches for enhanced predictive accuracy.
rithms, renowned for capturing non-linear patterns, are of
Index Terms—LSTM, GRU, Prophet, Time Series Analysis, particular interest. In our research, we employ Convolu-
National Stock Exchange (NSE) tional Neural Network (CNN) [4] and Long Short-Term
Memory (LSTM) models. LSTMs, integral to Recurrent
I. I NTRODUCTION Neural Networks (RNNs), possess the ability to retain input
Stock price prediction is a compelling domain of interest, information over extended periods, a crucial aspect in stock
attracting researchers seeking to enhance forecasting and anal- price prediction for enhanced precision and accuracy.
ysis within the financial landscape. Time series forecasting, a Given the vast and highly non-linear nature of stock market
prevalent method applied to variables evolving, is particularly data, the preference for deep learning models is evident.
relevant in tracking the dynamic nature of stock prices. The However, working with time series data demands careful
overarching objective of this research is to enhance the pre- consideration of certain aspects:
dictive performance compared to existing models designed for
a. Stationarity: This involves assessing statistical properties
stock price projection. The intricate nature of stock price pre-
such as mean, variance, covariance, and standard de-
diction arises from its inherent volatility and unpredictability.
viation, ensuring they remain constant over time for a
Identify applicable funding agency here. If none, delete this. stationary time series.
b. Seasonality: Recognizing periodic fluctuations or patterns In a separate investigation outlined in [10] explored stock
within the time series, denoted as seasonality, is crucial. price prediction using Seasonal Autoregressive Integrated
Any predictable oscillation repeating over time falls within Moving Average (SARIMA) and Prophet Prediction Model.
this category. Their models achieved a root mean square error (RMSE)
c. Autocorrelation: This metric measures the correlation be- difference of 44.9545.
tween a variable’s current value and its past values, indi- As a whole, these investigations emphasise the possible
cating the degree of correlation across different successive effectiveness of machine learning and deep learning method-
time intervals. ologies in forecasting stock prices. The application of these
In summary, the immense and non-linear nature of stock mar- methodologies to forecast the prices of individual companies
ket data necessitates sophisticated models, and our research or stock index movements utilising daily closing prices shows
aims to contribute to this domain by exploring and enhancing potential for improving the accuracy of predictions. Further-
the performance of both linear and non-linear time series more, it has been suggested in the literature that specific
forecasting models. models demonstrate noteworthy efficacy, including the Hybrid
The remainder of the paper follows a structured organiza- LSTM-GRU model and the Recurrent Neural Network (RNN)
tion. Section 2 provides a comprehensive review of related with Long Short-Term Memory (LSTM). It has been observed
work, delving into existing literature in the field. Section that convolutional neural network (CNN) and ARIMA models
3 outlines the research methodology employed, elucidating perform admirably in scenarios requiring short-term predic-
the approach taken in conducting the study. Subsequently, in tions. It is acknowledged, however, that model performance
Section 4, the obtained results are discussed, shedding light on may vary depending on the specific dataset and characteristics
the outcomes of the research. Finally, Section 5 serves as the of the population under study.
conclusion of the paper, summarizing key findings, and also
exploring the future scope of the study. III. P ROPOSED F LOW
Figure 1 illustrates the proposed flow of the research.
II. R ELATED BACKGROUND
An expanding corpus of literature addresses the utilization
of different machine-learning techniques as well as deep-
learning techniques for stock price prediction. Several studies
have substantiated the effectiveness of these methodologies
across various datasets and techniques in forecasting stock
prices.
A notable contribution to the field of stock price predic-
tion is presented in a study [6], wherein they employed the
Autoregressive Integrated Moving Average (ARIMA) model
[5]. The research, titled ”Stock Price Prediction Using the
ARIMA Model,” meticulously details the process of construct-
ing a predictive model for stock prices. Utilizing stock data
published in the Nigeria Stock Exchange (NSE) and New
York Stock Exchange (NYSE) the authors developed stock Fig. 1. Proposed Flow
price predictive models. Their findings highlight the ARIMA
model’s significant potential for short-term predictions.
Subsequently, the same researchers, along with Otokiti A. Data Preparation
Sunday O, extended their work in [8]. This time, they explored Historical stock price datasets for TCS, ICICI, and Pow-
stock price prediction using Artificial Neural Networks (ANN) ergrid were procured from the National Stock Exchange
[7] and introduced a hybridized approach. This approach (NSE) website, spanning a one-year timeframe. These datasets
combines variables from technical and fundamental analyses encompass diverse attributes, including but not limited to
of stock market indicators to enhance existing prediction ’OPEN’, ’Date’, ’LOW’, ’HIGH,’, ’Series, ’PREV. CLOSE’,
methodologies. ’Close’ ’VWAP’,’LTP’,’52W L’,’52W H’,’VALUE’, ’VOL-
Another study [9] delves into stock price prediction utilizing UME’ and ’# Trades.’ These attributes collectively offer a
various deep learning models, specifically Long Short-Term comprehensive overview of the daily trading activities and
Memory (LSTM), Recurrent Neural Network (RNN), and performance metrics associated with the respective stocks.
Convolutional Neural Network (CNN) with a Sliding Window Each dataset, representing TCS, ICICI, and Powergrid,
Model. The primary emphasis is on making price forecasts underwent meticulous analysis utilizing five distinct ma-
for individual companies or anticipating the movements of chine learning models. The selected models comprised Long
stock indexes using daily closing prices. The method under Short-Term Memory (LSTM), Convolutional Neural Net-
consideration employs a model-independent strategy. work (CNN), AutoRegressive Integrated Moving Average
(ARIMA), a Hybrid Model integrating LSTM and Gated Re-
current Unit (GRU), and Facebook’s Prophet. This methodical
approach was employed to systematically explore and evaluate
the predictive capabilities inherent in each model concern-
ing the unique characteristics of the individual datasets. By
adopting such a comprehensive strategy, the research sought
to enhance the depth and breadth of its findings, providing a
nuanced understanding of the performance of each model in
the context of the specific stock datasets under consideration.
B. Exploratory Data Analysis
The target variable selected for forecasting is ’close,’
with its preceding timestamp being utilized for analysis.
Consequently, all the analyses outlined below will be
executed with a specific emphasis on this feature.

Rolling Mean and Rolling Std Analysis

The analyses involve the computation of the moving average
(Rolling Mean) and the moving standard deviation (Rolling
Std) of a time series. These techniques are employed to
mitigate short-term fluctuations, accentuate trends, and Fig. 3. Seasonal Decomposition Analysis - Powergrid dataset
pinpoint potential periods of volatility. Often utilized as
a pre-processing step, they contribute to a more profound
comprehension of the underlying patterns in the data. Figure null hypothesis positing data stationarity is rejected.
2 demonstrates the rolling mean and rolling std analysis for Figure 4 illustrates the Augmented Dickey-Fuller Test
Airtel stock prices. on Powergrid dataset.

Fig. 4. ADF Test - Powergrid dataset

iii. ACF (Autocorrelation Function) and PACF (Partial

Fig. 2. Airtel Stock Price with Rolling Mean & Rolling Std Autocorrelation: ACF and PACF plots are graphical
tools employed for the analysis of autocorrelation and
partial autocorrelation in a time series, respectively. The
C. Time Series Analysis
utilization of these plots aids in the identification of
i. Seasonal Decomposition Analysis: Seasonal decompo- patterns and dependencies among observations at various
sition entails the dissection of a time series into its lags.
constituent components: trend, seasonality, and residuals.
The objective of this analysis is to discern recurring D. Deep-Learning Models
patterns and trends within the data. This step is deemed i. LSTM: Long Short-Term Memory (LSTM) networks,
crucial in comprehending the cyclic nature inherent in classified as a type of recurrent neural network (RNN),
the time series. Figure 3 illustrates the seasonal decom- are well-suited for modeling sequences and time-
position analysis of the Powergrid dataset. dependent data. These networks are characterized by
ii. ADF Test (Augmented Dickey-Fuller Test): The ADF memory cells capable of storing and retrieving informa-
test serves as a statistical tool for evaluating the sta- tion over extended sequences. An LSTM cell comprises
tionarity of a time series, aiding in the determination of key components, namely the input gate, forget gate, cell
whether differencing is requisite to achieve stationarity. state, and output gate.
Stationarity, a critical assumption for various time series The cell state, represented as Dt , undergoes updates
models, is assessed through this test. In this instance, based on both the input data and the preceding cell
as the p-value exceeds the 5% significance level, the state. The control of this update process is governed
 flow within the GRU, facilitating the decision-making
process regarding which information to update and what
to discard.
The update and reset gates collectively contribute to the
modification of the hidden state of the GRU, denoted as
zt . The mathematical representation of GRU is discussed
via Equation 3.

zt = σ(Xz · [zt−1 , yt ] + az )
rt = σ(Xr · [zt−1 , yt ] + ar )
(3)
z˜t = tanh(Xh · [rt ⊙ zt−1 , yt ] + ah )
zt = (1 − zt ) ⊙ zt−1 + zt ⊙ z˜t
Fig. 5. Autocorrelation - Powergrid dataset
Here, σ is the sigmoid activation function, ⊙ denotes
element-wise multiplication, and Xz , Xr , Xh are weight
matrices.
iii. Hybrid Model of GRU and LSTM: The hybrid model,
integrating both Long Short-Term Memory (LSTM) and
Gated Recurrent Unit (GRU), capitalizes on the respec-
tive strengths of each architecture to enhance predictive
performance. Formulated as a weighted combination
of individual LSTM and GRU predictions, the hybrid
prediction incorporates an optimized weight parameter,
denoted as σ, to achieve optimal predictive accuracy.
Equation 4 demonstrates the mathematical representation
of the hybrid prediction model.
Fig. 6. Partial Autocorrelation - Powergrid dataset Hybrid = σ × LST M + (1 − σ) × GRU (4)

The parameter σ represents a tunable parameter gov-

by the input gate (jt ), determining the extent to which erning the contribution of each model to the final pre-
new information is incorporated into the cell state. The diction. The identification of the optimal value for σ
mathematical representation for cell state and input gate involves an iterative process. A for loop is employed to
is discussed via Equation 1. systematically vary σ across a specified range. During
each iteration, the mean squared error (MSE) score is
Dt = gt ⊙ Dt−1 + jt ⊙ D̃t computed by comparing the hybrid predictions against
D̃t = tanh(Xc · [zt−1 , yt ] + ac ) (1) the actual values. The σ value corresponding to the
minimum MSE score is then chosen as the optimal
jt = σ(Xi · [zt−1 , yt ] + ai )
weighting factor.
Here, gt is the forget gate output, ⊙ denotes element-
wise multiplication, σ is the sigmoid activation function, E. Model Assessment
and Xc , Xi are weight matrices. The Mean Squared Error (MSE) stands as a widely utilized
The output of the LSTM cell, zt , is determined by metric for evaluating prediction accuracy. It quantifies the av-
the output gate ot and the updated cell state Dt . The erage squared difference between predicted and actual values,
mathematical representation for computing the output of offering a quantitative assessment of the prediction error. The
LSTM cell is discussed via Equation 2. formulation of MSE is expressed via equation 5.
N
ot = σ(Wo · [zt−1 , yt ] + ao ) 1 X
(2) M SE = (ad − pd )2 (5)
zt = ot ⊙ tanh(Dt ) N
d=1

Here, Wo is another weight matrix, and σ is the sigmoid In this expression, ’N’ signifies the total number of data points,
activation function. ad denotes the actual value on day ’d’, and pd represents
ii. GRU: The Gated Recurrent Unit (GRU), classified as an- the corresponding predicted value. The MSE is computed by
other variant of recurrent neural network (RNN), features squaring the difference between each actual and predicted
a cell with two primary gates: the update gate (zt ) and value, summing these squared differences, and subsequently
the reset gate (rt ). These gates govern the information dividing by the total number of data points.
 Fig. 7. LSTM-GRU & PROPHET - TCS data Fig. 9. CNN model - TCS data

Fig. 8. LSTM model - TCS data Fig. 10. LSTM-GRU & PROPHET - Powergrid data

IV. R ESULTS & D ISCUSSION Figure 7, 10 and 13 green line indicates the ’Hybrid’ and the
This section provides a concise and lucid overview of orange indicates the ’Prophet’ model. In Figure 8, 11 and 14
the results, accompanied by an analysis of their significance orange line indicates training prediction and the green line
within the context of the study topic. The historical prices for indicates test prediction.
TCS, ICICI Bank, and Powergrid spanned a one-year duration The calculation of the Root Mean Squared Error (RMSE)
and were directly sourced from the National Stock Exchange score for the aforementioned models has also been undertaken.
(NSE). In this study, the training of our model exclusively It can be asserted that, in terms of the RMSE score, the
utilized the ’close’ column. LSTM-GRU Hybrid model demonstrates superior performance
in forecasting the TCS stock price.
A. TCS The calculation of the Root Mean Squared Error (RMSE)
The implementation of the LSTM, LSTM-GRU hybrid
model, CNN, and Prophet model was carried out on the TCS
dataset. The prediction graphs for the aforementioned models
are illustrated via Figure 7, 8 and 9.
B. POWERGRID
The implementation of the LSTM, LSTM-GRU hybrid
model, CNN, and Prophet model has been executed on the
Powergrid dataset. The prediction graphs for the aforemen-
tioned models are illustrated via Figure 10, 11 and 12.
C. ICICI BANK
The implementation of the LSTM, LSTM-GRU hybrid
model, CNN, and Prophet model has been carried out on the
ICICI Bank dataset. The prediction graphs for the aforemen-
tioned models are illustrated via Figure 13, 14 and 15. In Fig. 11. LSTM model - Powergrid data
 Fig. 12. CNN model - Powergrid data Fig. 15. CNN model - ICICI BANK data

Fig. 13. LSTM-GRU & PROPHET - ICICI BANK data

Fig. 16. RMSE Score of models - ICICI, Powergrid and TCS data

score for the aforementioned models has also been conducted.

It can be asserted from figure 14 that, in terms of the V. C ONCLUSION AND F UTURE S COPE
RMSE score, the CNN model exhibits superior performance
in forecasting the Powergrid stock price.
The calculation of the Root Mean Squared Error (RMSE) In conclusion, this research paper delved into the application
score for the aforementioned models has also been undertaken. of various deep learning-based methods for predicting stock
It can be concluded from figure 16 that, in terms of the RMSE prices, utilizing one-year historical stock data for three distinct
score, the CNN model demonstrates superior performance in stocks. The algorithms considered in this study encompassed
forecasting the ICICI Bank stock price. LSTM, Prophet, Convolutional Neural Network (CNN), and
the LSTM-GRU Hybrid model. LSTM has traditionally been a
preferred algorithm for sequential data analysis. However, this
study revealed that the CNN model outperformed all other
algorithms, achieving the lowest Root Mean Squared Error
(RMSE) for both the Powergrid and ICICI Bank datasets. In-
terestingly, the LSTM-GRU Hybrid model exhibited superior
performance only in the TCS dataset. These findings suggest
that the CNN model may yield a more accurate and robust
predictive model.
Potential avenues for future research in this domain may
involve exploring the integration of Reinforcement Learning
with LSTM or other hybrid approaches. Additionally, different
feature selection methods could be investigated to enhance
model performance. Lastly, evaluating the models on a larger
dataset or in a real-time scenario could provide insights into
Fig. 14. LSTM model - ICICI BANK data their generalizability and practical applicability.
 R EFERENCES
[1] Shumway, R. H., Stoffer, D. S., Shumway, R. H., & Stoffer, D. S.
(2017). ARIMA models. Time series analysis and its applications: with
R examples, 75-163.
[2] Cheng, J., Tiwari, S., Khaled, D., Mahendru, M., & Shahzad, U. (2024).
Forecasting Bitcoin prices using artificial intelligence: Combination of
ML, SARIMA, and Facebook Prophet models. Technological Forecast-
ing and Social Change, 198, 122938.
[3] Otto, P., & Schmid, W. (2023). A general framework for spatial GARCH
models. Statistical Papers, 64(5), 1721-1747.
[4] Dhanalakshmi, R., K, B., Sinha, B. B., & Gopalakrishnan, R. (2023).
Tomato leaf disease identification by modified inception based sequential
convolution neural networks. The Imaging Science Journal, 71(5), 408-
424.
[5] Kontopoulou, V. I., Panagopoulos, A. D., Kakkos, I., & Matsopoulos,
G. K. (2023). A review of ARIMA vs. machine learning approaches for
time series forecasting in data driven networks. Future Internet, 15(8),
255.
[6] Ariyo, A. A., Adewumi, A. O., & Ayo, C. K. (2014, March). Stock
price prediction using the ARIMA model. In 2014 UKSim-AMSS 16th
International Conference on computer modelling and simulation (pp.
106-112). IEEE.
[7] Kurani, A., Doshi, P., Vakharia, A., & Shah, M. (2023). A comprehensive
comparative study of artificial neural network (ANN) and support vector
machines (SVM) on stock forecasting. Annals of Data Science, 10(1),
183-208.
[8] Adebiyi, A. A., Ayo, C. K., Adebiyi, M., & Otokiti, S. O. (2012). Stock
price prediction using neural network with hybridized market indicators.
Journal of Emerging Trends in Computing and Information Sciences,
3(1).
[9] Zhang, J., Ye, L., & Lai, Y. (2023). Stock Price Prediction Using CNN-
BiLSTM-Attention Model. Mathematics, 11(9), 1985.
[10] Vishwakarma, A., Singh, A., Mahadik, A., & Pradhan, R. (2020).
Stock price prediction using Sarima and Prophet machine learning
model. Journal of Advanced Research in Science, Communication and
Technology (IJARSCT), 9(1).