Research On Graph Representation Learning Based Stock Trend Forecasting

Stocks are an important part of the financial industry.With the rapid development of the global economy,the stock market has also developed vigorously,and stocks have become a very important investment channel in our daily life.The goal of stock trend forecasting is to predict the future trend of stock prices,and stock trend forecasting is essential for stock investors to obtain higher returns from the stock market.Therefore,tock trend forecasting has very high academic research and practical application value.With the rapid development of deep learning technology,remarkable results have been achieved in the fields of computer vision and natural language processing.Some work has also begun to apply deep learning to the stock trend forecasting problem.For example,some event-driven methods use neural networks to mine information about events that affect stock price changes from information such as news,social media,and discussion boards,and use these event information to predict stock prices.Other methods of technical analysis feed the stock’s historical price and trading volume data into a neural network to predict the stock’s future price trend.Nevertheless,most of the existing deep learning-based stock forecasting work treats each stock as a separate individual,ignoring the internal connections between stocks.In recent years,graph representation learning techniques such as graph neural networks have become a popular direction of graph data mining,and have shown outstanding advantages in fully mining graph structure data and semantic information on graphs.Therefore,this thesis studies the stock prediction technology based on graph representation learning,and improves the effect of stock prediction by introducing the graph representation learning technology to mine the rich internal relationship between stocks and stocks.This thesis firstly conducts basic research on the graph representation learning,and studies two graph representation learning techniques,knowledge graph representation learning and heterogeneous graph representation learning.Then,for the stock trend forecasting technology based on graph representation learning,in the event-driven method,this thesis studies the relational event-driven stock trend forecasting method;in the technical analysis,this thesis studies the stock-concept heterogeneous graph representation learning based stock trend forecasting,and the instance graph representation learning based multivariate time series(stock)forecasting method.In summary,the main contributions and innovations of this thesis are as follows:First,basic research on graph representation learning.This thesis first designs a segmented knowledge graph representation learning framework to solve the problem that existing knowledge graph representation learning methods cannot balance model complexity and model expressiveness.Our approach achieves an increase in model expressivity without increasing model complexity by promoting sufficient feature interactions and symmetry and anti-symmetry of modeled relationships.Then,this thesis proposes a heterogeneous graph representation learning method that fuses structural information,by encoding node centrality,which measures the importance of nodes,and incorporating graph structural information into aggregation modules on meta-paths.Therefore,structural information is fused into heterogeneous graph representation learning.Second,for event-driven stock trend forecasting,this thesis proposes a relational event-driven stock trend forecasting method.Existing event-driven methods ignore the impact of event information distinguished by the properties of the stock itself and the impact of event information of other related stocks.Our proposed method models stock contexts and learns the impact of event information on stocks in different contexts.At the same time,this work constructs a stock graph and design a new propagation layer to propagate the impact of event information from other stocks.Third,stock trend forecasting of technical analysis.This thesis first proposes a stock forecasting method based on stock-concept heterogeneous graph representation learning,which fully mines concept-oriented shared information from predefined concepts and hidden concepts,and then utilizes the shared information of stocks to improve the stock trend forecasting,so as to solve the problem of existing methods ignoring the dynamic correlation between stocks and concepts and ignoring the valuable shared information contained in hidden concepts.Then,in response to the problem of existing methods ignoring the internal connections between different stocks at different time stamps,this thesis propose a multivariate time series(stock)forecasting method based on instance graph representation learning,which mines the interdependence between different stocks at different time stamps to improve the stock trend forecasting.