Research On Price Forecasting Model Of US Stock Index Based On LSTM Neural Network

Since the birth of the stock market, people have been constantly using various data models, machine learning and data mining tools to predict future movements of stock prices in order to gain huge profits. Among those tools, the neural network is widely used,which is due to its high degree of self-learning, stability and modeling capabilities. Compared to statistics and econometrics mathematical model, neural network has advantages.In this paper, based on analysis of the problems faced by the stock price short-term forecasting, the paper compared a variety of stock price forecasting methods,and explored the BP neural network, RNN neural network and LSTM neural network to predict the short-term stock price and make appropriate comparison among them. Based on the LSTM,the paper had a discuss and improved the algrithm and structure of the model. According to Comparative validation, as well as theoretical studies, LSTM neural network model can learn the past data very well, and identify the impact of the relationship between time series, and can take advantage of selective memory of advanced machine learning capabilities, mine the inherent law of change, in order to predict changes in next short period. In this paper, the main research work are:The paper analyzed against conventional BP neural network model for predicting stock price, and interpreted the irrationality of using it to predict time series model theoretically, then made evidence. Meanwhile,the paper introduced the concept of RNN neural network which is proper for prediction of time series, and compared it with BP neural network. Then, based on the RNN model,the paper introduced LSTM neural network, and used it in the experiment to verify the feasibility of this model for more accurate prediction. Finally, discussed the theory of LSTM, and improved model, and used it in experiments.As for the factors that will influence stock price data,the paper chose the most critical factors including closing price, open price, highest price, lowest price and trading volume as the independent variable in cells of each layer neural networks. Meanwhile, the experiment chose Shanghai Composite Index and the S & P 500 index and the Dow Jones Industrial Average to study,those index are representative in China and the US stock market. The study of US stock index were divided into two parts which were the indices with minimum and maximum daily trading volume in order to compare the two indices to test model accuracy, while, according to the different characteristics of China and the US stock market,the paper verified the usefulness and accuracy of the model.