Research On Multi-feature Stock Trend Forecast Based On LSTM

With the rapid development of China’s economy and the increase of residents’ disposable income,the demand for investment has risen rapidly.With its relatively higher yield,Stock market has become one of the main ways of investment.The stock market has been the focus of academia for its social functions of promoting economic development and optimizing the allocation of resources.There are many types of algorithm models applied to stock forecasting,and the range of data sources used by these models is constantly expanding.As a complex non-linear system,the stock market is influenced by many kinds of information sources,and the influence of various information is ultimately reflected in the changes of stock prices.A high-performance stock forecasting system should guarantee the diversity of information sources and integrate different information sources.In the existing research,due to the difficulty of different types of information fusion,stock forecasting models are often just built based on one single information source,which limits the performance of the models.Aiming at this problem,in order to explore the application of multi-information sources in stock forecasting,this paper chooses data sources and constructs multi-features on the premise of considering hierarchy,coverage and representativeness.Using LSTM networks to build a multi-feature stock trend forecasting model to predict the daily rise and fall trend of individual stock transactions.This paper analyzes stock-related data sources and determine feature construction methods.According to the model,this paper selects three kinds of information sources: stock history data,financial news,stock social sensation.And determine the construction methods of features out of the basic data and financial text.Establish a financial text processing model.According to the characteristics of financial text class construction,based on the different structures of LSTM network,a sentiment text sentiment analysis model and a financial news text representation model are established: the former obtains the emotional tendency category of social sentiment,and the latter extracts the hidden text characteristics of financial news.Establish a stock trend forecasting model with multiple features.Based on the results of text processing models,the characteristics of financial texts are constructed.Therepresentative stock technical indicators are selected as the technical indicators features.The EMD algorithm and fuzzy entropy algorithm are used to decompose and reconstruct the stock closing price sequence to obtain the hidden signal features.According to the characteristics of the constructed features,they are input into a model in which the LSTM network is integrated with the fully connected neural network.Based this structure to construct a stock trend forecasting model and obtain the trend of individual stocks.This paper empirically analyzes the above research based on the Python environment.Build web crawlers to download financial texts,and verify the effectiveness of financial news and social public opinion information sources through experimental analysis of financial text processing models.The stock trend forecasting model is used to predict the rise and fall of individual stocks in the SSE 50 Index.The experimental results show that the multi-feature model constructed in this paper is superior to the comparison model in F1-value.And in the comparison experiment,it is found that in most cases,as the data source increases,the prediction accuracy of the same model is improved,which proves the rationality of multiple data sources.