Exploiting Long Short-term Memory Neural Network for Stock Price Prediction

Abstract

Stock as a high yield, high risk investment has been favored by the public. In order to increase the return on investing in stocks, investors need to predict stock prices. In the past, investors used traditional mathematical methods to make predictions. Now, neural networks are used by investors to predict stocks, which can improve the accuracy of stock forecasting. To further verify the effectiveness of these methods, this work discusses the effects of different network structures and hyperparameters on stock prediction models using short-term memory (LSTM) neural networks. The results show that deeper network layer can get better training effect, but it needs more training time, resulting in a lot of time waste. In addition, this experiment tests the prediction effect under different dropout parameters. The results show that the dropout function should not be too large or too small. Multiple experiments are needed to find an appropriate dropout value.