Stock Market Prediction Based on LSTM Neural Networks

Purpose – This study aims to more accurately and effectively predict trends in portfolio prices by building a model using LSTM neural networks, and investigating the risk and profit prediction of investment portfolios. Design/Methodology/Approach – To obtain a return on stocks, this study used 60 monthly transaction data from major countries, including the United States and Korea, for five ETFs, BNDX, BND, VXUS, VTI, and 122630.KS, for five years from January 2016 to December of 2021. In addition, a related portfolio was constructed using modern portfolio theory. Through Min-Max normalization, five ETFs and closing data from April 20 to July 20, 2022 were normalized. The input data were classified into two characteristic dimensions, and an LSTM time series model was constructed with the number of nodes in six hidden layers. Findings – By establishing a portfolio and making regression predictions, it was possible to effectively reduce situations in which prediction accuracy was lowered due to large fluctuations in index-based stocks. Research Implications – The predicted results were tested using OLS regression analysis. The relationship between the risk of building a tangential portfolio with the same composition with different weights, the accuracy of stock price prediction by effectively reducing the low prediction accuracy of highly volatile stocks in the portfolio, and changing the set risk-free interest rate were examined.