Abstract: With the rapid development of economy and technology,the stock market has become an important part of the current financial market.Traditional machine learning methods have limitations in processing stock prediction problems with nonlinearization,high-noise or strong volatility.In recent years,the rise of deep neural networks has provided new solutions to stock trend forecasting problems.In this paper,longshort-term memory network (LSTM) is used to deal with long-distance stock temporal problems,and a multi-category feature system is constructed as the input for long-term and short-term memory networks for training,including common technical indicators,multiple key features,and real event information for individual stocks.Meanwhile,the experimental part comprehensively analyzes the effectiveness of various characteristics for stock trend prediction,and the comparison results show that the multi-category characteristic system performs well in the prediction,and can reach a short-term forecast accuracy of 68.77%.In addition,LSTM is compared with other models such as convolutional neural network (CNN),recurrent neural network (RNN) and multilayer perceptron (MLP).Experimental results show that LSTM is superior to other models in solving this problem.

Key words: Deep learning, Feature construction, Long short-term memory, Stock trend forecasting