STOCK PRICE PREDICTION USING LONG SHORT-TERM MEMORY NETWORK

• Main Author: Pradjanata, Roselina
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39797
• Institution: Institut Teknologi Bandung

Building stock price prediction model generally requires considerable time and data, because each stock has a different price movement characteristic. Even though the prediction results in general will be more accurate, computation time and cost to make model for each stock are too high. This problem can be overcome by creating a model using a stock index that is already a combined price of certain stocks. This research focuses on modeling the prices of Indonesian stocks included in the JKSE by using the LQ45 index which consists of 45 stocks with the largest capitalization. Modeling will use artificial neural networks called long short-term memory (LSTM) which is very suitable for time-series data types and can remember long-term relationships from data. The experiment was carried out to get the model with the best hyperparameter. The hyperparameter tuning in the experiment is batch size, number of epochs, and number of neurons in the hidden layer. Due to the constantly increasing stock price data, a stock price prediction system will be developed that makes the model always updated with the latest data and the best hyperparameters. The results of the experiment are the LQ45 index model which can be used to predict all stocks included in the JKSE. Prediction performance will be measured using the mean absolute percentage error (MAPE) metric so that it can be compared between individual stocks. From the test results, predictive performance of the stocks that have high volatility tends to be lower than stocks that are fairly stable and have low volatility. Meanwhile, whether of these stocks included in the LQ45 index has little effect on predictive performance.