USING GENETIC ALGORITHM AND ARTIFICIAL NEURAL NETWORK FOR STOCK PRICES FORECASTING

• Main Author: GALUH SEKAR WARDANI (NIM 234 03 028); Pembimbing : Ir. Imam Istiyanto,MBA. , UDRIANA
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/20588
• Institution: Institut Teknologi Bandung

Artificial Neural Network (ANN) is commonly used in financial domain. In this study, ANN was used to forecast stock prices data because of the ANN ability to model complex problems. ANN can map historical values and future values of time series data through learning process that mimic process in human brain. However, one of the main difficulties using ANN is design ANN architecture that appropriate to different data. Genetic Algorithm (GA) is used to overcome that problems, where GA act as a tool to select parameter combination which consist of number of processing element in hidden layer, learning rate and momentum to find appropriate architecture. Forecasting is applied to daily data for the LQ45 index, HMSP and ISAT stock prices. <br /> <br /> <br /> This study was made to identify the effect of using GA, different length time series data and set data distribution to ANN prediction. The empirical findings of this study show that GA increase generalization of ANN. The use of GA was significantly in the NMSE testing value. The performance value of set data testing can be use to indicate how well ANN perform the new values. Different distribution data set show the affect to the ANN prediction. The best prediction for forecasting 1 day ahead for each data results: index LQ45, NMSE 4,3549:0,0210, correlation 95,89%:96,50%, accuracy 44,38%:48,52% ; HMSP stock, NMSE 0,0097:0,0078, correlation 99,61%:99,64%, accuracy 32,61%:37,33% ; ISAT stock, NMSE 0,0305:0,0246, correlation 95,87%:95,91%, accuracy 47,34%:46,15%. Other empirical findings in this study that longer time series data did not always result better forecasting accuracy, and for each set distribution data there are no indication that one of them always bring to best performance. This evidences show that the ANN learning processes need different information to get the best forecasting performance.