Abstract:   As a complex nonlinear system, stock market is affected by many kinds of information sources, and its trend adjustment does not follow the uniform time process. According to the characteristics of the stock market, and combined with the characteristics of LSTM (long short term memory) neural network fitting complex time series data, a multi-feature stock trend prediction model based on LSTM is proposed by using information from different levels of sources such as news platform, social platform, historical data, etc.. Firstly, empirical mode decomposition (EMD) is used to denoise the basic stock data to obtain the law of stock price change; secondly, the financial text data is pushed to the improved LSTM network model with attention mechanism and automatic encoder for training and testing, so as to reflect the investors’ psychology; then, the combined prediction of stock trend is carried out by LSTM neural network. The results show that the time series financial text feature can effectively improve the prediction performance of the model.

Key words:  multi-source data, multi-feature, LSTM, stock trend prediction