Research On Key Technologies Of Stock Forecasting With Multi-Scale Characteristics Fusion

The price fluctuation of the stock market is closely related to the national economy.Studying the price trend of the stock market has always been a hot topic that relevant researchers have paid attention to.Traditional stock forecasts are generally characterized by technical indicators to build traditional time sequence models to predict future stock prices.With the continuous development of natural language processing technology and machine learning algorithms,it has become more diverse in the data of stock forecasts.Text data such as financial news and stock reviews began to be used for stock forecasting.By extracting the emotion in text data as the input of the stock to the stock Price predicting,traditional time sequence-based technical index stock forecast models are not enough to face complex and changeable financial stock markets.Considering the low emotional value of the current text,it is impossible to express text emotions accurately.Therefore,based on the characteristics of financial news,this article proposes a model prediction model of deep learning stocks with mult-scale characteristics.Empirical research is conducted with the three stocks of the liquor industry and the real estate industry as an empirical object.The research content is as follows:Firstly,financial news feature extraction based on machine learning.This article builds CNN based financial news event classification models,CNN based news emotional classification models,and LDA based news theme models to extract financial news.In the classification of news events,the average accuracy of CNN in the test set reached 0.9062,which was 2 to 5percentage points higher than SVM,decision tree,XGBoost and other models.The average accuracy of the emotional analysis model based on semantic matching is 0.9339;in the news theme word model,the theme word probability distribution of highly expressing news information is obtained.Based on the above experiments and its results,the characteristics of financial news are extracted,and one-dimensional emotional characteristics,multi-dimensional emotional characteristics,and the characteristics of financial news.Secondly,stock forecast based on multi-scale characteristics fusion.This article will build LSTM based technology index stock prediction models as the benchmark model.The average accuracy rate of 6 stocks is 58.71%,which is greater than 56%.It proves that the model has certain prediction capabilities.On the basis of the benchmark model,the one-dimensional emotional characteristics are integrated to obtain a stock prediction model based on LSTM based emotional characteristics fusion technology indicators.The average accuracy of the model of this model reaches 63.26%.Ability to predict.On the basis of one-dimensional emotional characteristics,the classification of news events is introduced to obtain multi-dimensional emotional characteristics classified by categories.The LSTM based multi-dimensional emotional characteristics fusion technology indicators of the stock predictive model.The predictive ability of simple emotions has increased by 3.79%,indicating that multi dimensional emotional characteristics can better express the emotions of financial news and obtain higher prediction accuracy.In order to explore the role of highly summarized text on stock forecasting,based on the benchmark model,add the characteristics of the news theme,and obtain the LSTM based financial news theme characteristic fusion technology indicator.60.53%,slightly higher than the benchmark model,indicating that the highly summarized text also helps improve the accuracy of the forecast,but compared with the other two emotional characteristics extraction methods,the effect is slightly worse.