Research On Stock Prediction Based On Multi-head Self-attention Mechanism And Media Features

Trend prediction and rule mining based on stock price and other financial time series data in the stock market can help investors to formulate more reasonable investment strategies,help promote the smooth operation of the stock market and other capital markets,and effectively promote the technical development and theoretical innovation of time series data prediction in finance and other fields.Financial time series data such as stock price are characterized by easy change,high noise and low stability,and their change rules are interfered by market noise data as well as multiple internal and external influencing factors.It is a challenging problem to grasp their rules and achieve accurate prediction.Prediction methods of statistical analysis and machine learning has generalization ability is insufficient,feature extraction,such problems as insufficient ability and nonlinear mapping ability,and the prediction model of neural network can overcome the above problems and insufficient has long-range dependence and memory ability and gradient disappeared and the problems such as explosion,these factors make prediction effect is influenced by the larger shares.Based on the traditional deep learning model,this thesis further improves the performance of the stock prediction model by optimizing the structure of the neural network,expanding the data features and influence factors,and integrating the multi-head self-attention mechanism to extract depth features.First,a stock price prediction model based on weighted BiGRU and the media features is constructed.On the basis of deep learning and neural network model,the network structure is further optimized,the weighted BiGRU network structure is adopted as the main model,and the media features is introduced as the information source of the model.The weighted BiGRU network can effectively solve the problems of the traditional network such as long range dependence,gradient disappearance and explosion.By considering the dual influence of the historical direction and the future direction simultaneously,it can not only increase the memory capacity but also expand the influence factor direction.The text vector of media features obtained by Doc2 vec training can extract the media features through BiGRU classifier,which can reflect the positive or negative sentiment of relevant enterprises or industries in the stock market,and expand the influence factor of forecast point data.Then,a prediction model of stock price trend is constructed which integrates the multihead self-attention mechanism and the deep media features.On the basis of the basic media features obtained by the above prediction model,the model introduces multi-head selfattention mechanism,which is used to calculate the attention distribution value(importance weight)of each media feature,and then extracts the deep media features.Compared with the basic features,the deep feature can distinguish the different influence degrees of different time series and topic texts on the predicted point data,and the prominent influence of key texts should be considered.The deep media features combined with the stock historical volume and price index was fused into the weighted BiGRU model,which fully optimized and expanded the data source to further improve the prediction accuracy.Finally,the effect and performance of the stock prediction method are demonstrated and analyzed through a large number of simulation iterative experiments,and the performance indicators are compared with the traditional prediction model.Experimental results show that compared with the traditional prediction model,the proposed prediction method can further reduce the error and improve the prediction ability.