“Predicting Stock Price Using Machine Learning”

Amanpreet Kaur Kasavath Darshan Rahul Chauhan

Computer Science and Engineering Computer Science and Engineering Computer Science and Engineering
Department Department Department
Chandigarh University Chandigarh University Chandigarh University
Punjab, India Punjab, India Punjab, India
[email protected] [email protected] [email protected]

Kakarapalli Manikanta Tej Deepanshu Raghu Arnav Mehta

Computer Science and Engineering Computer Science and Engineering Computer Science and Engineering
Department Department Department
Chandigarh University Chandigarh University Chandigarh University
Punjab, India Punjab, India Punjab, India
[email protected] [email protected] [email protected]

Abstract — In today's volatile stock market, accurate price machine learning techniques, notably DL algorithms such as
forecasting is crucial. Investors seek high returns, Long Short-Term Memory. This Research paper delves into the
necessitating reliable predictive models. Traditional time convergence of traditional time series models and cutting-edge
series analysis contrasts with machine learning, especially deep learning approaches, with a specific focus on LSTM neural
Support Vector Machines (SVMs) and deep learning like RNN networks, to unravel the complexities of stock price prediction.
and LSTM networks, adept at non-linear data and long-term By leveraging a dataset comprising 5,000 observations of the
dependencies. This study compares ARIMA and GARCH
widely tracked S&P500 index, the study aims to explore whether
with LSTMs using 5,000 S&P500 data points, highlighting
LSTM's superior predictive abilities, handling non-linearity, deep learning methodologies, LSTM networks, can significantly
and encoding long- term memory for better stock price enhance the accuracy of stock price forecasts compared to
estimation. Traditional statistical models like Autoregressive conventional time series models. The motivation behind this
Integrated Moving Average and (GARCH) Generalized research lies in the urgent need to explore innovative techniques
Autoregressive Conditional Heteroskedasticity against that of capable of capturing intricate patterns within stock market data,
LSTM networks. Through comparative analysis, it including non- linear relationships and long-term dependencies
underscores the superior predictive capabilities of LSTM that pose challenges for traditional models.
networks, attributed to their ability to process non-linear data
and encode essential information in long- term memory, In this paper, we offer a comprehensive theoretical overview of
thereby offering enhanced stock price estimation. both time series models and LSTM neural networks,
supplemented by practical insights derived from rigorous
Keywords → Stock market prediction; Machine learning; stock empirical analysis. Through comparing predictions using
market price prediction; sentiment-analysis; enhanced learning- metrics like root mean square error, we aim to assess the
based method; time series data prediction, Recurrent Neural effectiveness of deep learning techniques, particularly LSTM
Networks (RNN), Long Short-Term Memory. neural networks, in capturing the subtleties of stock price
movements. Our structured approach includes detailed
exploration of methodologies, datasets, experimental setups, and
I. INTRODUCTION implications of findings. By addressing this pivotal question, we
contribute valuable insights to the ongoing discourse on
In the fast-paced and volatile realm of financial markets, predictive analytics in financial markets, facilitating more
accurately forecasting stock prices is crucial for investors, informed investment strategies and regulatory policies. As
financial institutions, and regulatory bodies. With negative investors increasingly rely on accurate predictions to inform
returns and heightened risk factors, the need for reliable strategic decision-making, understanding the potential of deep
prediction models has never been more pressing. learning methodologies assumes paramount importance. By
Traditionally, mathematical models, particularly those bridging the gap between traditional and advanced techniques,
based on time series analysis, have been instrumental in this research aims to provide valuable insights into the efficacy
predicting stock prices. However, the emergence of of LSTM networks in improving accuracy.

1
 II. RELATED WORK III. DATASETS USED

The pursuit of advanced computational techniques for stock These include utilizing application programming interfaces
price forecasting has become a central focus in financial (APIs) provided by relevant companies, purchasing data from
research, given its direct impact on investment decisions specialized data firms, and accessing datasets from open-source
and market stability. While traditional methods like communities. For this study, two essential types of datasets are
autoregressive models and moving averages have long been required.
relied upon for insights into market trends, their limitations
in capturing non-linear patterns and long- term Firstly, stock market price data is indispensable. Secondly, news
dependencies have spurred exploration into more articles sourced from mainstream media are crucial for
sophisticated approaches. Recent literature underscores a comprehensive analysis. These datasets, obtained through
growing interest in employing machine learning various means, form the foundation for conducting in- depth
algorithms, particularly deep learning models, for stock research on stock price forecasting and its correlation with news
price prediction. Notably, Recurrent Neural Networks coverage.
(RNNs), particularly Long Short-Term Memory (LSTM) The datasets utilized in this research encompass a wide array of
networks, have gained traction due to their adeptness at financial data pertinent to stock price forecasting, with a primary
capturing sequential patterns and inherent temporal focus on the widely tracked S&P500 index. These datasets span
dependencies within financial data. multiple years, ensuring a diverse range of market conditions to
facilitate robust model training and testing. Key components of
the datasets include daily opening, closing, high, and low prices
Pioneering work by Hochreiter and Schmid Huber
of S&P500 stocks, alongside corresponding trading volumes.
introduced LSTM networks, showcasing their prowess in
Additionally, supplementary datasets may include economic
learning and retaining long-term dependencies.
indicators, global market trends, and relevant news sentiment
Subsequently, LSTM networks have found successful
data, enriching the analysis and providing a comprehensive
applications in diverse fields, including finance. For
perspective on market dynamics.
instance, Zhang et al. utilized an LSTM network to predict
stock prices within the Chinese stock market, achieving
IV. METHODOLOGY
superior accuracy compared to traditional models.
Furthermore, the adaptability of LSTM networks has been
evidenced in various financial scenarios. Gu et al. extended LSTM-Based Stock Price Forecasting:
LSTM networks to predict stock price movements using
text news data, highlighting the model's capability in Data Collection: The initial step involves gathering historical
handling unstructured information sources. Recent stock price data for the S&P500 index, spanning multiple years.
research has also explored ensemble methods by This comprehensive dataset includes daily opening, closing,
combining LSTM networks with other algorithms to high, and low prices, as well as trading volumes, ensuring varied
enhance prediction accuracy. Brown and Johnson, as well market conditions for effective LSTM model training and
as Chen and Wang, delved into LSTM networks' testing.
effectiveness in capturing complex market patterns. Brown
and Johnson compared LSTM networks with traditional Data Preprocessing: Before inputting the data into the LSTM
time series models, while Wang investigated hybrid network, preprocessing steps are applied to enhance model
models merging deep learning techniques with statistical efficiency. This includes normalization to scale the data to a
approaches. suitable range for the activation function. Normalization ensures
a balanced influence during training, preventing dominance by
certain features due to scale differences.
However, a gap exists in the literature concerning the
systematic evaluation of LSTM networks against time
Sequence Generation: Processed data is transformed into
series models like ARIMA and GARCH using the S&P500
sequences to train the LSTM model effectively. Each input
dataset. This research endeavors to offer a nuanced
sequence contains a fixed number of historical data points, such
understanding of LSTM networks' capabilities and
as prices and volumes for the past N days. The corresponding
limitations in this context. LSTM, short for Long Short-
output sequence predicts the stock price for the next day,
Term Memory, is a type of recurrent neural network (RNN)
enabling the LSTM network to capture temporal patterns in stock
architecture designed to overcome the limitations of
prices.
traditional RNNs in capturing long-term dependencies in
sequential data.
\\
 LSTM Architecture: The architecture of the LSTM neural V. RESULT AND DISCUSSION
network is pivotal. We design a deep LSTM model
comprising multiple layers of LSTM cells. Our research findings on utilizing the LSTM method for stock
price forecasting exhibit promising results, showcasing its efficacy
Model Training: The LSTM model is trained using pre- in capturing intricate market patterns. Through rigorous
processed and sequenced data. During training, the model experimentation and evaluation, the LSTM model consistently
learns to capture patterns within historical stock prices. surpassed traditional time series models like ARIMA and GARCH
in predicting the S&P500 index. Lower values of accuracy metrics
Model Evaluation: Following LSTM model training, such as MAE, MSE, and RMSE were consistently observed for
evaluation is conducted using a distinct test dataset. Metrics LSTM predictions compared to traditional models. Furthermore,
such as Mean Absolute Error (MAE), Mean Squared Error the LSTM model demonstrated a notable capability to capture
(MSE), and Root Mean Squared Error gauge prediction long-term dependencies and adapt to varying market conditions.
accuracy against actual stock prices. Visualizations, like Even during periods of high volatility, where conventional models
line charts comparing predicted and actual prices, offer often struggle, the LSTM model exhibited remarkable stability and
qualitative insights into model performance. accuracy. These qualities, along with its adaptability to different
economic contexts and robust performance in real-time
Hyperparameter Tuning and Optimization: Key LSTM predictions, underscore the model's potential for practical
model parameters, including learning rate, LSTM layers, applications in financial markets.
units, and sequence length, significantly impact
performance. Systematic experimentation and validation LSTM, short for Long Short-Term Memory, is a type of recurrent
fine-tune these hyperparameters to optimize forecasting neural network (RNN) architecture designed to overcome the
accuracy limitations of traditional RNNs in capturing long-term
dependencies in sequential data. It is particularly well-suited for
Comparative Analysis: The LSTM model's efficacy is tasks involving time-series data, natural language processing,
verified through comparison with traditional time series and speech recognition
models like ARIMA and GARCH. Superiority in capturing
intricate S&P500 stock price patterns is demonstrated using The success of the LSTM method in stock price forecasting is
diverse evaluation metrics. credited to its recurrent neural network architecture, enabling it
to effectively learn and remember intricate patterns within
Sensitivity Analysis: Sensitivity analysis evaluates the sequential data. Unlike traditional models, LSTM networks
LSTM model's robustness under varied market conditions. excel in capturing nonlinear relationships and temporal
Performance across different timeframes, economic events, dependencies, making them well-suited for the dynamic nature
and market volatility levels is assessed, ensuring the model's of financial markets.
adaptability and reliability in providing predictions.

\\
 Moreover, the implications of this research extend to timely predictions for financial markets. The implications of this
algorithmic trading strategies. The LSTM model's real- research are substantial for investors, traders, and financial
time prediction accuracy paves the way for automated institutions, offering valuable guidance for strategic decision-
trading systems capable of executing trades based on making. LSTM-based forecasting, stakeholders in the financial
reliable forecasts, enhancing efficiency and potentially sector can optimize investment portfolios, mitigate risks, and
maximizing returns. capitalize on emerging market trends. Furthermore, this research
lays the groundwork for the integration of advanced machine
While the LSTM method demonstrates remarkable learning models into real-time trading systems, augmenting the
performance, it's crucial to acknowledge its limitations. efficiency and efficacy of algorithmic trading strategies.
Predictions rely on historical data, and unforeseen external
events like economic crises or geopolitical developments VII. ACKNOWLEDGMENT
may influence market behavior in ways historical data
cannot anticipate. Additionally, the computational We extend our sincere gratitude to our advisors for their
complexity of deep learning models, including training invaluable guidance. Special thanks to the participants and
time and resource requirements, warrants consideration in organizations for providing essential data. We appreciate the
practical implementations. support from colleagues and friends, enhancing the research
quality. Our families' unwavering encourage men was
This research underscores the potential of the LSTM indispensable. Thank you all for making this research possible.
method in stock price forecasting, emphasizing its ability Furthermore, I am thankful to the participants who generously
to capture market dynamics and adapt to changing shared their insights and feedback, enriching the project's depth
conditions. Embracing advanced techniques like LSTM- and accuracy. This project signifies the convergence of
based forecasting can foster more informed and strategic innovation and persistence, and it wouldn't have been possible
decision-making in the global finance landscape, without the unwavering support of my peers and the invaluable
benefiting investors, financial institutions, and resources provided by our institution. As we celebrate this
policymakers alike. milestone, I am filled with a profound sense of gratitude for the
opportunity to contribute to the advancements in predictive
analytics within the realm of financial markets.

VIII. REFERENCES

[1] Hiba Sadia, Aditya Sharma, Adarsh Paul, Sarmistha Padhi, Saurav Sanyal-
“Stock Market Prediction Using Machine Learning Algorithms”, IJEAT,
2019.
[2] Raut Sushrut Deepak, Shinde Isha Uday, Dr D. Malathi, “Machine Learning
Approach In Stock Market Prediction”, IJPAM 2017.
[3] M. Roondiwala, H Patel and S. Varma, "Predicting stock prices using
LSTM," International Journal of Science and Research (IJSR), vol. 6, no. 4,
pp. 17541756, 2017.
[4] T. Kim and H. Y. Kim, "Forecasting stock prices with a feature fusion
LSTM-CNN model using different representations of the same data," one,
vol. 14, no. 2, p. e0212320, April 2019.
[5] S. Selvin, R. Vinaya Kumar, E. A. Gopalakrishnan, V. K. Menon and K. P.
Soman, "Stock price prediction using LSTM, RNN, and CNN-sliding
window model," in International Conference on Communications and
Informatics, 2017.
VI. CONCLUSION AND FUTURE WORK [6] Wang, S., and Y. Luo. 2012. “Signal Processing: The Rise of the Machines.”
Deutsche Bank Quantitative Strategy (5 June).M. Young, The Technical
Writer’s Handbook. Mill Valley, CA: University Science, 1989.
The effectiveness of LSTM-based stock price forecasting,
[7] Moritz, B., & Zimmermann, T. (2016). Tree-based conditional portfolio
particularly within the context of the widely followed sorts: The relation between past and future stock returns. Available at SSRN
S&P500 index, has been convincingly demonstrated in this 2740751. [5] Olah, C. (2015). Understanding lstm networks–colah’s blog.
research paper. Through meticulous experimentation and Colah. github. io.
comparative analysis, the LSTM model has showcased [8] Emerson, S., Kennedy, R., O'Shea, L., & O'Brien, J. (2019, May). Trends
superior accuracy, resilience to market volatility, and and Applications of Machine Learning in Quantitative Finance. In 8th
International Conference on Economics and Finance Research (ICEFR
adeptness in capturing long-term dependencies, surpassing 2019).
conventional time series models like ARIMA and GARCH. [9] Heaton, J. B., Polson, N. G., & Witte, J. H. (2017). Deep learning for
These findings underscore the potential of deep learning finance: deep portfolios. Applied Stochastic Models in Business and
techniques, especially LSTM networks, in furnishing Industry, 33(1), 3-12.
dependable and
\\