Prediction of stock price direction with machine learning models

• Main Author: Kant Kaw Khin
• Other Authors: Wong Jia Yiing, Patricia
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Artificial intelligence; Machine learning; Stock price prediction
• Online Access: https://hdl.handle.net/10356/176943
• Institution: Nanyang Technological University

Machine Learning (ML) algorithms drew a great deal of attention in the recent years as promising models in time-series predictions, allowing investors to leverage on these computational abilities to perform stock analysis more efficiently. Stock analysis can be done through Technical Analysis (TA), Fundamental Analysis (FA) and Sentiment Analysis (SA). This project investigates and compares how well ML models can predict the price direction of prominent stocks listed on the SGX based on (1) TA, (2) SA, and (3) a combination of both TA and SA. The ML models used in this paper are Random Forest (RF) Classification, XGBoost Classification, Long Short-Term Memory (LSTM) Classification and LSTM Regression model. The classification models are used to predict price stock direction, while the regression model is used to predict closing prices. RF and XGBoost mostly supported the project’s objective that a model based on TA + SA will perform better than models based on TA and SA individually when predicting stock price direction. Using the stock of Singapore Telecommunications Limited (Z74.SI) as a reference, RF and XGBoost produced accuracy rates of 82% and 78.4% for TA + SA analysis respectively, which is higher than that of models conducting TA and SA individually. However, LSTM Classification model did not perform satisfactorily, with accuracy rate in TA + SA (52.3%) falling behind that of TA only (60.1%). LSTM Regression model was also used to predict closing prices, and its performances was evaluated against a well-known time-series prediction ARIMA model. The results were satisfactory with the LSTM Regression model outperforming for TA + SA as compared to TA and SA individually.