The Research On Event-driven Stock Prediction

Stock prediction is mainly based on stock-market data.It is focused due to its considerable return.However,comparing with the prediction task in other fields,this task is also more difficult because of the high randomness in the stock market.In recent years,researchers improve the accuracy of stock prediction with news,social media,and price series.Stock trends are related to the capital direction and hot events,and the events have a driving effect on the stock market.As the U.S.stock market has a long history,a large trading volume and a wide influence in the world,this paper researches the impact of various data on the U.S.stock market from the perspective of event driving and comprehensively analyzes the role of news,social media and price series in the stock prediction.(1)To solve the quantification of event influence in the stock market,we propose a multielement hierarchical attention capsule network.The model quantifies the influence of important information from the news and social media through the multi-element hierarchical attention mechanism and retains more context information through the capsule network.At the same time,a combined data set was constructed to maintain the complementarity of social media and news,and finally,we achieved the state-of-the-art results.(2)To solve the effective encoding of text from the stock market,we propose a capsule network based on the Transformer encoder to extract the deep semantic features of social media and capture the structural features of the text.In this work,we utilize different baselines for comparison and verification,and the results show that the model achieves the text representation more effectively and finally improve the prediction accuracy.(3)To research the delay effect on stock prediction and solve the problem of incomplete information coverage from news and social media,we adopt the U.S.stock price data sets and propose a Long Short-Term Memory network based on auto-encoder.We increase the implicit variables by auto-encoder,make the randomness of the stock market relatively controllable,and improve the generalization ability of the model.Through the Long Short-Term Memory network based on attention mechanism,we redistribute the time series weights and improve the prediction accuracy.Meanwhile,to analyze the delay impact in the market,we compare the performance of several models by delay prediction.Experimental results show that there is a delay effect on the market and the prediction accuracy can be improved by delay.