SECTOR STOCK PREDICTION MODEL USING RECURRENT NEURAL NETWORK FOR TIME SERIES FORECASTING

• Main Author: Yulia Ariyanti, Dwiani
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/66644
• Institution: Institut Teknologi Bandung

Stock analysis conducts analysis to find out estimated future stock prices by using news as one of the analytical materials. This final project aims to obtain a model that can predict sectoral stock prices with this information and determine the effect of sectoral conditions on the company's stock prices. Since the model must be able to handle time series data to capture stock prices patterns and store sequential information for a long time, RNN (Recurrent Neural Network) is used by utilizing the LSTM (Long Short-Term Memory) layer to overcome vanishing gradient problem for stock prices data which is sequential data with a long time span. The model is used to predict stock prices from sectoral point of view and company's stock prices with adding feature of sectoral information. To reduce error of the prediction result, data manipulation is implemented by combining the value of news sentiment in the last few days and adjusting the time steps range as the input for model training. As the result, the RNN model with input stock price information and news sentiment has MAPE (Mean Average Percentage Error) score of 1.13%. It performs a better error score compared to the model that was trained without using news sentiment with MAPE score of 2.28%. In general, the prediction of the company's stock price with adding sectoral information has a better performance (lower MAPE scores) compared to model that is trained using only the company's stock prices. In particular, the best performance was obtained for BBCA's stock prediction, with a MAPE score of 1.63%.