LQ45 STOCK INDEX PREDICTION USING MACHINE LEARNING (BIDIRECTIONAL LONG-SHORT TERM MEMORY) METHOD WITH HYPERPARAMETER TUNNING

• Main Author: Timotius Oei, Michael
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/76158
• Institution: Institut Teknologi Bandung

One of the methods that can be applied to predict stock index values is Bidirectional Long-Short Term Memory (BLSTM). In the BLSTM method, information flows forward and backward through the network, allowing for better modeling to anticipate trend patterns and recognize complex patterns based on historical stock data. Furthermore, hyperparameter tuning is applied to find the best combination of BLSTM hyperparameters (epochs, batch size, learning rate, and the number of neurons) to obtain optimal output. As a comparative method, Long-Short Term Memory (LSTM) and Gaussian Process Regression (GPR) are also used. This study aims to analyze the performance of BLSTM and the influence of hyperparameter tuning on stock index prediction results, which tend to experience significant index changes due to the pandemic. The data used consists of LQ45 stock data from March 2017 to July 2023. The data is divided into training data, testing data, and prediction data. The training data covers the period from March 2017 to April 2022. Meanwhile, the testing and prediction data respectively cover time ranges from April 2021 to July 2023 and August 2023 to October 2023. Based on the analysis and processing of LQ45 data, the prediction accuracy is obtained as 93.222% for the BLSTM method with hyperparameter tuning, 93.086% for the LSTM method with hyperparameter tuning, and 93.212% for the GPR method. Additionally, future stock index graphs and buy & sell graphs based on the application of Moving Average (MA) are also obtained, which are expected to be considered in investment decisions.