Research On Multi-source Heterogeneous Stock Data Analysis Method Based On Deep Learning

The global economic and financial integration has made the role of financial products such as stocks and securities more important,so it is of great economic value to accurately predict the fluctuations and changes of financial market.The fusion of multi-source heterogeneous data composed of different data sources such as stock market,social network and portal website and different information forms such as stock value,text and image provides rich data support for predicting changes in financial market.Numerical data mainly contains the information of stock price fluctuation range,while text data mainly contains the information of stock price fluctuation trend.The effective fusion of heterogeneous data to predict stock trends not only has important academic research value,but also produces important social economic benefits.In recent years,with the development of deep learning and the enhancement of hardware computing power,it provides a powerful tool for using artificial intelligence to predict stock data with timing and multi-source heterogeneity.The current deep learning methods have made great progress in stock trend prediction,but the diffusion of stock text information noise,the irrelevance and noise of stock information,the volatility and uncertainty of stock data are still the key problems facing stock prediction in deep learning.This dissertation focus on the above three key problems and tries to solve the stock trend prediction in a multi-source heterogeneous data environment through deep learning.The main work and contributions are summarized as follows.1.Aiming at the problem of noise diffusion caused by semantic information in tweet context,a hierarchical model of stock movement predictive network(SMPN)with incorporative attention mechanism is proposed to predict stock market.For the noisy data environment of English texts,this thesis proposes a deep learning framework that combines global attention and local attention mechanisms.The main innovations of this method are:combining the contextual information of each layer with local semantics to obtain richer semantic representations of each layer;using local semantics to effectively reduce the spread and diffusion of invalid information(noise)in the text.Experiments show that SMPN can effectively improve the performance of stock prediction.2.Aiming at the issue of irrelevance and noise between Chinese text and target stock information,a Bi-directional gated recurrent unit reinforcement learning model of stock price movement prediction network(SPMPN)is proposed.For the externally crawled Chinese stock text data,this dissertation proposes a deep learning framework that uses characters to segment Chinese semantic information and gated recurrent unit based on reinforcement learning.The main innovations of this method are:local character-level semantic embedding and global context-level semantic embedding are combined to learn more effective news-level representational semantics;while news-level layer noises are automatically filtered based on reinforcement learning at the day-level layer,richer longterm dependencies are captured in the text information.The experimental results verify that reinforcement learning can filter noise more effectively and improve the performance of stock prediction.3.Aiming at the volatility and uncertainty of stock data,a dual-channel stock forecasting model of collaboration network and generative adversarial network(SELFGAN)is proposed.The main innovations of this method are as follows:the dual channel model framework is composed of collaboration network and generative adversarial network.The collaboration network can effectively reduce the volatility of stock data,while the generative adversarial network can effectively reduce the randomness and uncertainty of stock data.Experiments show that the model not only improves the prediction accuracy,but also increases the interpretability of the prediction results,and at the same time improves the generalization ability of the model.In summary,focusing on the problems of low prediction accuracy of single data,complexity of stock external environment and weak generalization performance of general depth model in the stock trend prediction method based on deep learning,this dissertation proposes to introduce external multi-source heterogeneous data into the deep stock prediction model to improve the prediction accuracy of the model.Using the correlation and complementarity between numerical data and text data to realize multi-source heterogeneous data fusion,this dissertation proposes different algorithm for different tasks,and verifies the rationality and effectiveness of the algorithm through a large number of comparative experimental results.