Research On Key Technologies Of Event-driven Stock Market Prediction

Stock predicting is a particularly challenging task.The theoretical basis of this paper is the "effective market hypothesis",namely in the stock market with sound law,good function,high transparency and full competition,all valuable information has been timely,accurately and fully reflected in the trend of the stock price,which is simply "information effective".Therefore,it is reasonable to model the relationship between events and stock price volatility by event study.Recommend appropriate stock buying based on event information,then choose the right time to sell based on morphological matching.This paper mainly includes three research points:event detection,strategic stock selection,and timing selling.Event detection is an important task in information extraction.Previous work mainly consider the sequential representation of sentences.However,long-range dependencies between words in the sentences may hurt the performance of these approaches.We believe that syntactic representations can provide an effective mechanism to directly link words to their informative context in the sentences.In this paper,we propose a novel event detection model based on syntactic dependency trees.In particular,we propose transforming syntactic dependency trees to target-dependent trees where leaf nodes are words and internal nodes are dependency relations,to distinguish the target words.Experimental results show that our approach has excellent effect.Strategic stock selection is an event-driven stock recommendation.The same is the use of deep learning powerful self-learning ability,this paper proposes a hybrid neural network model to learn the implied link between events and stock prices.In order to distinguish the specificity of each stock,a convolutional neural network is used to represent the background information of stock.Considering the digital sensitivity in the financial field,this paper makes a discrete vector representation of the numbers in the text sentence.At the same time,we use the temporal characteristics of long and short term memory network to represent the event sentence,and learn its semantic and structural features.The experimental results show that the model achieves an annualized return over the market.Timing selling is the best selling point selection based on dynamic time warping algorithm.Based on the idea that historical information will be reproduced in the future,we use the historical stock price datas(including the open price,the highest price,the lowest price,the closing price)to build the form base of stock price,and then look for the best matching form of the holding stock based on the dynamic time warping algorithm to get the selling point.The experimental results show that timing selling further improves the annualized return of the model on the basis of strategic stock selection.The heuristic operation strategy proposed in this paper achieves good results,and indirectly illustrates the feasibility of the research.