| With the continuous development and improvement of financial markets,an increasing number of investors are participating in the financial market,which drives the application of new scientific technologies in stock market prediction.Accurate prediction can help investors reduce investment risks and improve investment returns.However,the high volatility and randomness of stock data make the research on stock trend forecasting a huge challenge.Among the numerous stock prediction studies,stock correlation research is one of the important directions.However,the correlation between stocks is time-varying,which makes the stocks in the set of statically correlated stocks constructed in advance according to some correlation rules less correlated in reality.In addition,a variety of raw data related to stocks,such as closing price and trading volume,represent various information in the stock market that affects stock price trends from different perspectives.There is a certain correlation between the data from different perspectives,and the existing prediction models often ignore this correlation information.This paper proposes a stock prediction model combining multi-view stock data features with dynamic market correlation information(MDF-DMC).The model extracts the internal and external effective features that affect the stock trend by modeling the internal factors that affect stock fluctuations(multi-view data of the stock itself)and external factors(market stock data),and combines the two features to improve the predictive ability of the model.MDF-DMC uses a mixed multi-layer perceptron(MLP-Mixer)to interact with multiview data,and uses a gate recurrent unit(GRU)to enhance information.MLP-Mixer can interact with data from two dimensions of time and space,so the model can learn the correlation between data from different perspectives.Then through GRU,the correlation features extracted by MLP-Mixer are enhanced from the perspective of timing,in order to extract internal multiview data features with more representational capabilities;MDF-DMC also uses an improved Transformer encoder to model stock market data,and dynamically extracts external market correlation features that affect stock fluctuations in the form of sliding windows,which solves the problem of poor timeliness of pre-built correlation stock sets.Finally,the model integrates internal multi-view features and external market correlation features,and uses fully connected layers to complete stock trend prediction.Experimental results on a large amount of stock data in different countries show that MDF-DMC improves both predictive performance and generalization ability compared with recent state-of-the-art models.Aiming at the drawbacks in the previous research on stock prediction,this paper provides a new way of thinking:1)Using MLP-Mixer to extract the correlation between multi-view data of individual stocks,and using GRU to enhance the extracted correlation features,so as to make full use of the complementarity of multi-view data of stocks and effectively improve the prediction performance of the model.2)The multi-head attention mechanism is used to establish the dynamic correlation between the stock to be predicted and the stock market,without any prior knowledge,to obtain the influence of other stocks in the market on the predicted stock in real time.3)Comprehensively consider the internal and external factors that affect the stock,and improve the model’s prediction performance and generalization ability.Finally,based on the above research,this paper implements a stock prediction system to provide users with stock market visualization,trend prediction and risk control. |
