| Stock researchers and investors have been working hard to make reliable forecasts on the stock market and maximize investment returns.Stock volatility is affected by many factors,such as historical stock price data,social media public opinion,investor sentiment,etc.Stock price and stock text fusion is an acceptable stock forecasting method.However,problems remain,such as poor time dependence on historical price data,low stock text availability,and insufficient fused features’ effectiveness.Noisy,low-quality,and incomplete abnormal data in the existing stock data lead to inaccurate learned stock features and poor model prediction performance.In addition,most current models improve the availability of stock data sets by changing the network structure of stock forecasting and lack in-depth research on the uncertainty factors of stock data.Therefore,this paper starts with improving the availability of stock input,entirely selecting stock-related features,and selecting an efficient forecasting model to improve the performance of the stock forecasting model.The critical research contents are as follows:(1)Considering the complex and diverse influencing factors of the stock market and the low availability of stock-related data,this paper proposes a stock price prediction model Credible Net based on heterogeneous data and multi-layer attention mechanisms.The core innovation lies in the historical price of stock dependency relationship captured,and the lowquality text in the stock is processed using the two-layer attention mechanisms of the word layer and the sentence layer.Then,the two heterogeneous data of stock text and stock price are fused,and the time-level attention mechanism is used to extract valuable information in the stock fusion features to obtain efficient fusion data information.The results show that attention mechanisms can effectively solve the problems of low quality,low credibility,and low information availability in stock data.In addition,this paper uses a Gaussian mixture model to model the output of the neural network and then quantifies the uncertainty of the stock forecast task and analyzes the uncertainty of the stock forecast.Experimental results thoroughly verify that the proposed model performs well on the tweet and stock price datasets.(2)Aiming at the problem of unsatisfactory prediction effect and low prediction model accuracy caused by abnormal data in stock data,this paper uses a highly randomized tree model to select features of historical stock price transaction data.Then it quantifies uncertainty.The idea is introduced into the stock forecasting model to improve the accuracy of the model.On this basis,this paper proposes a stock price prediction network Log Net based on feature selection and uncertainty quantification,uses a Gaussian mixture model to model the logit of the neural network’s output,and selects stock data with lower uncertainty for stock prediction.When selecting a prediction model,this paper uses the SDENet model to predict the future trend of stocks.This model enables us to extract information from low-uncertainty stock characteristics,thereby improving the model’s prediction performance.The results of comparative experiments show that Log Net has significantly improved performance compared to other stock forecasting neural network models.(3)Investor sentiment has an essential impact on stock price fluctuations,but most existing research rarely considers the integration of sentiment values and other heterogeneous data.To solve this problem,this paper proposes a stock price prediction model Sen Trans Net based on sentiment analysis and Transformer.The model first constructs sentiment scores on unstructured data obtained from Twitter and then fuses the sentiment scores with historical price indicators for stocks.In selecting the stock prediction model,this paper uses the Transformer model to realize the practical analysis of the stock fusion data,thereby improving the stock price prediction performance.To effectively evaluate the effectiveness and stability of the Sen Trans Net model,this paper selects four different stock stocks in four other industries as data sets.The results show that the proposed Sen Trans Net stock prediction model has better prediction performance and robustness for stock prediction tasks. |
