REAL TIME STOCK
PRICE PREDICTION
Jeevitha.P
Harini.G SRM INSTITUTE OF SCIENCE
Department of CSE AND TECHNOLOGY
SRM INSTITUTE OF SCIENCE Trichy,India
AND TECHNOLOGY jp2364@[Link]
Trichy,India
hg4525@[Link]
completely graphical in nature and
provide you with real time
ABSTRACT - In this project, we information on live or virtual stock
demonstrate a stock price prediction market trends that lets you compare
system based on a hybrid deep
actual share price with prediction to
learning model, combining with
show how close it comes. Using these
CNN, LSTM and transformer. Each
part plays a role: CNN to get the cutting edge techniques in
short-term pattern in stock, LSTM conjunction with each other, the
identifies the long trend of market, hybrid model provides more accurate
and Transformer applies attention predictions and explanations for
mechanism enhancing prediction market performance and serves as a
accuracy. The system has a
bridge between sophisticated deep
thoroughly modern design that offers
learning methodologies and tangible
an interactive web based interface
for the public users and touch screen finance domains.
based kiosk frontend for the
operators, both types of scrutiny are INTRODUCTION:
It is challenging to predict stock prices
because of the inherent uncertain nature
of financial markets. Conventional
approaches have difficulty in dealing
with these fast dynamic movements,
but deep learning is a promising
solution.
In this project, a novel real-time stock Built a sleek web app serving dynamic
price prediction system is proposed by comparisons of real and predicted stock
and large for the first time with CNN– prices via interactive charts, facilitating
LSTM–Transformer hybrid model. The user engagement and comprehension.
CNN has a small receptive field so that it 3. Improved Forecast
Accuracy:
can effectively capture short-term
Obtained improved prediction accuracy
patterns, and the LSTM captures long-
through the complementation of CNN,
term trends of time series data; LSTM and Transformer layers.
meanwhile the Transformer pays 4. Practical Integration of
attention to key features. Research:
A web-based interactive platform, which Brought deep learning research to real-
doesn’t distinguish between real and world financial forecasting, a user-
artificial stock data, displays this friendly data-driven platform for
investors and researchers.
comparison interactively with typical life
updating price charts. By integrating
BACKGROUND AND
these advanced deep learning algorithms,
MOTIVATION :
the model achieves higher forecast A. Overview of stock price :
accuracy and can capture useful Stock prices are difficult to
information associated with market predict due to the market
motions. volatility; and the complexity in
the data patterns. Such variance
[Link] is not always well-represented in
We summarise our main contributions : traditional models . Time-series
This work provides several important forecasting has been enhanced
contributions for the task of learn in recent years by
financial models form data using deep advancements of deep learning,
learning: particularly the CNN, LSTM
1. Hybrid Model Design: and Transformer models. In this
work, these models are
Proposed a new hybrid deep learning
combined in a hybrid
model which combines three kinds of
architecture to improve
neural networks including CNN, LSTM
accuracy and robustness of
and Transformer for learning the short-
stock prediction. Snellier aims to
term patterns, long-term dependences
and attention-based relations in stock create a web-based system that
data. would analyze deep learning
2. Interactive Web Application: research and link it to actionable
financial insights, all in real For Stock Prediction ARIMA/LR was
time. only on-demand during early times but
B. Challenges in stock price they were helpless in front of non-
prediction : In the process of linear and noisy nature of the financial
prediction of the stock price, data.
volatility in the market, non- B. Machine Learning
linear trends and lots of Approaches
affecting factors make it a very Models like SVM, Random Forest,
challenging event. XGBoost have been searched to
achieve good results but were not able
RELATED WORK to capture the proper trends with
A. Traditional methods : varying time based differences.
Traditional models have limited C. Deep Learning Models
capabilities to handle these RNN & LSTM tried to overcome this
difficult patterns but tend to not by modeling sequential data which
find out. So, to tackle above successful dealt with the short
mentioned challenges sequences but again struggles with
B. Hybrid and transformer global data and no attention
models : CNN–LSTM– mechanism.
Transformer hybrid method is in D. Hybrid/Transformer Models
view to capture both short term C-Car roll & LSTM were able to
& long term dependencies with extract both local & temporal features
accuracy enhanced through efficiently, where as Transformer gives
attention. a better understanding through self
C. Research gap : The aim is to attentions to keep long range
provide real-time integrated dependencies well.
validation removing the E . Research Gap
dependency on the User itself Different from the existing
from the system that has enough
potential to need given the Model type strengths limita
situation or make impulsive tion
buying decisions on.
ARIM Statisti simple, fails
A cal interpreta with
ble non
DATASETS AND linear
PROCESSING data
SVM/ Machin Handless No
A. Conventional Methods RF e nonlinear time
learnin data seque Feature extraction is an important
g nce aspect in improving the accuracy of
learni stock price prediction. This project will
ng perform feature extraction using a
hybrid deep learning model integrating
LSTM deep captures slow
learnin long term , CNN, LSTM, and Transformer
g trends vanis components.
hing
gradi [Link]: Convolutional Neural
ents Network
CNN deep detects misse These CNN layers extract local
learnin local s temporal patterns from the stock data,
g patterns long for instance, short-term price
term movements and fluctuations. The key
relati trends are highlighted as the
on convolutional filters also reduce noise
transfor deep focuses needs and provide a refined feature map for
mer learnin on key large further processing.
g time steps data ,
heav [Link]: LONG SHORT-TERM
y MEMORY
comp
utatio The LSTM layers extract long-term
n dependencies and sequential
relationships in the time-series data,
hybrid deep combines slight helping the model understand how past
CNN - learnin pattern , ly trends influence future prices.
LSTM g memory highe
Transfo and r Transformer (Attention Mechanism)
rmer attention comp The Transformer component
for high utatio highlights, through self-attention, the
accuracy n most relevant features across the
cost sequence to ensure that the model
focuses on the critical patterns in
changing stock prices.
MODELING OF FEATURE
EXTRACTION AND MATHEMATICS USED IN THIS
MATHEMATICS PROJECT
1. Data Normalization
Before training, data is scaled using \text{MAE} = \frac{1}{n} \
Min–Max Normalization to bring all sum_{i=1}^{n} |y_i - \hat{y}_i|
features within a similar range (0 to 1):
X' = X - Xmin/Xmax - Xmin PROPOSED MULTI-MODAL
This not only ensures faster SYSTEM
convergence but also avoids feature
dominance from larger values. In this paper, we develop a hybrid deep
learning system that combines the
CNN, LSTM and Transformer models
2. CNN – Convolution Operation for effective neural stock price
In the CNN layer, mathematical prediction in real time. Each of the
convolution is used to extract local factors handles some limitation
patterns. observed in previous methods, and
S(t) = (X * W)(t) = \sum_{i} X(i) \cdot hence contributes to better
W(t - i) characterizing the intricate behaviors of
financial markets.
3. LSTM – Gating Mechanisms CNN Module - Local Feature
LSTM uses mathematical gates to Extractor
control the flow of information.
The CNN module learns the short-term
f_t = \sigma(W_f \cdot [h_{t-1}, x_t] +
temporal patterns and local
b_f) dependencies from a sequence of stock
i_t = \sigma(W_i \cdot [h_{t-1}, x_t] + prices.
b_i)
C_t = f_t * C_{t-1} + i_t * \\
tanh(W_C \\cdot [h_{t-1}, x_t] + b_C)
4. Transformer – The Attention
Mechanism
Self-attention in the Transformer is
Attention(Q,K,V)=softmax〖(QK
defined as:
Fig.1 hybrid CNN LSTM Transformer architecture
^T/√d_k)V〗 diagram for stock market prediction
It helps the model focus on the most
relevant time steps in the sequence.
MODEL ARCHITECTURE
5. Evaluation Metrics AND TRAINING
Performance can be quantified using
mathematical error functions such as: Model ArchitectureThe new model is a
\text{RMSE} = \sqrt{\frac{1}{n}\ fusion of the merits from CNN, LSTM
sum_{i=1}^{n}(y_i - \hat{y}_i)^2} and Transformer for the successful
analysis and prediction of stock price Transformer model overcomes the
movements. disadvantages of low prediction
performance on time-series data in existing
Input Layer:Takes preprocessed stock models, and further provides interpretation
data, which includes into which features contribute to stock
Open,High,Low,Close and Volume as price movements.
features.
Model Explainability : The attention
CNN Layer:It extracts local temporal mechanism of the Transformer further
patterns and diminishes noise by improves the explainability of the model
convolutions to detect the short-term price by detecting which time steps and features
change. (like price trends, volume changes or
volatility) have most influenced
LSTM Layer:Is used to encode the predictions.
features extracted to capture long-range
dependence and sequential relationship
between time series data.
Transformer Layer:Uses
EXPERIMENTAL
RESULTS :
Stock RMSE MAE R^2
symbol Score MODEL COMPONENT
AAPL 1.24 0.86 0.94 EVALUATION AND INPUT
TSLA 2.51 1.79 0.91 SENSITIVITY TESTING
MSFT 1.08 0.73 0.95
● Model Component
The experimental results show that the Evaluation : An ablation study
hybrid CNN–LSTM–Transformer model was performed to analyze the effect
performs consistently well across multiple of each part in hybrid architecture.
stocks. Low RMSE and MAE values, Each model (CNN, LSTM, and
along with high R² scores, indicate the Transformer) was first trained
model’s strong accuracy and stability in independently after which they
predicting stock prices. were combined in different ways
(CNN–LSTM, LSTM–
EXPLAINABILITY AND FEATURE Transformer, CNN–Transformer)
INSIGHTS that led to the final hybrid model.
Explainability and Feature Findings: Effect of different methods
Insights : Interpreting why a deep on long-term trend The CNN network can
learning model makes predictions is catch the short-term ups and downs well,
important to enhance trust and but it can’t keep its good performance in
interpretability in financial markets longer periods. Although the LSTM
forecasting. The hybrid CNN–LSTM–
model was able to capture sequential System Architecture : Tiers of the
dependencies, it failed with Deployment *The deployment is
organized in three tiers:
Frontend: Developed using React. js
user interface, which is quick, modern and
responsive. Interactive charts (via Chart. js
or Plotly. js) display
LIMITATIONS AND
THREADS OF VALIDITY
The study has some limitations and threats
to validity. The dataset considered is
CASE STUDIES somewhat circumscribed in terms of time
frame and market scenario, which might
To validate and appraise the performance affect the results’ generalizability. Simple
of the proposed hybrid CNN–LSTM– data quality, feature selection and modest
Transformer model in reality, three case hyperparameter tuning may impact model
studies were conducted with real-time and performance as well.
historical stock prices of Apple (AAPL),
Tesla (TSLA), as well as Microsoft Risks to the internal validity stem from
(MSFT). These firms were selected possible data leakages, or over fitting
because they exhibit high trading volume during training; external ones are hitched
and active stock market behavior. with the fact that a model might rely on
outer circumstances or specific datasets.
Case 1: Apple Inc. (AAPL) Construct
Goal: Test the model’s capability to
capture stable market trends in a non-
FUTURE WORK
volatile stock. Such a dataset can now act as a baseline in
further studies that extend to multi-
Findings: The hybrid model accurately
markets, long-term periods or multilingual
forecasted daily Apple
sentiment sources to improve model
generalization. Advanced feature
DEPLOYMENT engineering and automatic hyperparameter
tuning can be employed to increase
Real-time access to stock price predictions
prediction accuracy.
was provided by the deployed hybrid
CNN–LSTM–Transformer model in a Furthermore, combining explainable AI
web application that can be used approaches will assist in interpreting
interactively. The deployment comprises feature contributions and mitigating bias.
connecting the model that has been Likewise, participants can consider a real-
trained with a front-end and back-end time deployment mechanism with online
framework to predict in real time or model retraining to keep pace with
simulate projections on an interactive dynamic market trends and changing
dashboard. investor behaviors.
RESULTS AND limited generalization, and feature bias are
DISCUSSION a few of the threats to validity that were
identified through the study.
The proposed stock prediction model Generally, the research puts forward the
showed promising performance in potential of hybrid predictive models in
predicting short-term price movements.
Experimental results showed that the financial forecasting and lays the foundation
integration of technical indicators with for future work in scalability, explainability,
sentiment features improves overall and real-time adaptability within dynamic
prediction accuracy and reduces mean
squared error compared to the baseline market conditions.
models.
ACKNOWLEDGEMENT
Feature importance analysis suggested that
The authors thank SRM Institute of Science
features related to sentiment and moving
averages contributed most to performance. and Technology for support and the
maintainers of publicly available datasets.
However, variations in accuracy across
time periods demonstrate the model's
sensitivity to market volatility and REFERENCES
changing sentiment trends. This calls for [1] Y. LeCun, Y. Bengio, and G. Hinton,
adaptive retraining mechanisms and “Deep Learning,” Nature, vol. 521, pp.
continuous data updates in order to 436–444, 2015.
maintain reliability in real-world
applications. → (Fundamental paper introducing deep
learning foundations used in CNNs and
The results confirm that the approach of LSTMs.)
combining quantitative and qualitative
data is effective, but also highlight the
challenges relating to generalization across [2] S. Hochreiter and J. Schmidhuber,
diverse market environments. “Long Short-Term Memory,” Neural
Computation, vol. 9, no. 8, pp. 1735–1780,
CONCLUSION 1997.
This paper presents a machine learning- → (Original paper introducing LSTM
based methodology for stock price networks for sequence learning.)
predictions in real time, which merges
historical market data with sentiment
[3] A. Vaswani et al., “Attention Is All
analysis. The results indicate that the
You Need,” in Proc. Advances in Neural
integration of technical indicators with
Information Processing Systems
textual features of sentiment can lead to
(NeurIPS), pp. 5998–6008, 2017.
better forecasting accuracy and deeper
→ (Introduced the Transformer model
insight into market behavior.
While the performance achieved was and self-attention mechanism.)
competitive, the model is heavily
dependent on the quality of data, time [4] M. F. Dixon, D. Klabjan, and J. H.
span, and market conditions. Data leakage, Bang, “Classification-based Financial
Markets Prediction using Deep Neural → (Describes integration of attention
Networks,” Algorithmic Finance, vol. 6, and hybrid models for market
no. 3–4, pp. 67–77, 2017.
forecasting.)
→ (Shows application of deep learning
models for financial prediction.) [9] Yahoo Finance API Documentation.
[Link]
[5] X. Chen, L. Zhou, and Y. Dai, “Stock → (Used for real-time stock market data
Price Prediction Using CNN and LSTM
retrieval.)
Based on Attention Mechanism,” IEEE
Access, vol. 8, pp. 23227–23236, 2020.
[10] Alpha Vantage API
→ (Demonstrates hybrid CNN–LSTM
Documentation.
models with attention for time-series
[Link] (Used
forecasting.)
for historical and live financial data
[6] J. Selvin, R. Vinayakumar, E. A. access.
Gopalakrishnan, V. K. Menon, and K. P.
Soman, “Stock Price Prediction Using
LSTM, RNN, and CNN-Sliding Window
Model,” International Conference on
Advances in Computing, Communications
and Informatics (ICACCI), pp. 1643–
1647, 2017.
→ (Comparative study of LSTM, CNN,
and hybrid models for stock prediction.)
[7] S. Li, Y. Jin, and Q. Li, “Hybrid Deep
Learning Framework for Financial Time
Series Forecasting,” Expert Systems with
Applications, vol. 184, 115455, 2021.
→ (Proposes a hybrid architecture for
improving stock prediction
performance.)
[8] M. Zhang, H. Zhang, and W. Xu,
“Stock Market Prediction with Multi-Head
Attention and Hybrid Deep Learning
Models,” Applied Intelligence, vol. 52, no.
8, pp. 8733–8750, 2022.