Research On Technologies Of Stock Market Prediction Based On Quantity-price And Sentiment Analysis

As a very difficult and meaningful task,stock price prediction has been concerned by researchers from all walks of life.Under the premise of the efficient market hypothesis,it is difficult to predict the price when stock market is fair and has information that is transparent and unconstrained price that can fluctuate with the disclosure of information.However,in practice,it is impossible for all investors to get the same information at the same time.Especially,in today's rapid development of the Internet,network information is huge and chaotic.Different investors have different ability to deal with information.Faced with such a large amount of information,they will have different views on price trends.With this point,we hope to get as much information as possible through the analysis of existing technology,so as to predict the price trend as accurately as possible.At present,there are mainly structured data(quantitative data)and unstructured data(textual data)in the stock market.Therefore,this paper mainly embarks from the following three directions: quantitative prediction,sentiment analysis,joint prediction.Quantitative prediction is the most basic but also the most important task in the whole stock price prediction task.Its essence is time series analysis.Most of the existing researches are based on rules or statistics to model the price series.We need to know lots of knowledge in the financial domain.However,the changing rules of the stock market and the complexity of the stock market itself make this kind of analysis very difficult and unsatisfactory.Therefore,this paper proposes a time series prediction model based on stacked autoencoder(SAE)combined with long short-term memory network(LSTM),which extracts features automatically by SAE,and performs time series analysis through LSTM.The results show that the model has achieved good returns.Based on the analysis of news texts,sentiment information is further introduced to improve the prediction effect.In this paper,hierarchical structure is used to model news text information,and multi-document level representation is obtained.Then,by using the query-driven attention mechanism,sentiment information in news texts can be introduced from different perspectives according to needs.Sentiment information can be represented at multi-document level which includes sentiment information,thus the implicit relationship between news text and price trend can be modeled.The results show that the introduction of sentiment information can further improve the prediction effect based on news text.Joint prediction is the integration of the above two parts.Considering the quantitative data,it also needs to introduce the sentiment information of news text.This part first encodes these two parts of data to get the combined representation,then obtains the implicit feature vectors through variational inference,and finally uses the feature vectors which contain multi-source information to predict the stock price trend.The experimental results show that the introduction of multiple sources of information can better predict the stock price trend,but also reflects the feasibility of this study.