Financial Market Trend Forecast Based On Deep Learning And Natural Language Processing

With the continuous development of domestic financial market,the number of investors participating in financial market investment is increasing year by year.The main method for investors to participate in the financial market is to buy and sell stocks.However,as the stock price is highly volatile,if the investors cannot accurately grasp the trend of the stock price,they will not get profits,but will suffer economic losses.In addition,when the market suffers large fluctuations,accurate control of the market trend is also conducive to the authorities to intervene and promulgate rescue policies.In view of the above application scenarios,this thesis proposes a market trend prediction system based on deep learning and natural language processing technology,which can predict the rise and fall of individual stocks with high accuracy according to the text of financial public opinion and stock numerical data.The main work of this thesis is as follows:First,an improved method of word vector quadratic training is proposed.The disadvantages of the traditional word vector model in the field of emotion analysis lead to the difficulty in distinguishing the optimistic words from the pessimistic words in the public opinion.In order to alleviate this situation,this thesis used the method of text emotion classification for reference,and combined with the background of this thesis to improve the method,experiments showed that the improved method adjusted the model added the emotional information contained in the word vector.Secondly,a neural network model is proposed to characterize the text of financial public opinion.Due to the need of direct processing of large-scale text data,this thesis,based on the idea of hierarchical model,proposes a network model suitable for public opinion text data,which is called Bi-LSTM and CNN hierarchical model based on self-attention mechanism.The results of comparison with other models show that the model takes into account the training time and precision.In addition,an improved self-attention mechanism is proposed in this thesis,and experiments show that the accuracy of the improved model is improved.Finally,a mixed model based on public opinion text data and stock numerical data is proposed.This model processes both text data and numerical data at the same time,and adds the stacked recurrent neural network and highway network structure in the final part,which can process data with a multi-day span.This model is called Bi-LSTM and CNN layered hybrid model based on improved self-attention mechanism.The final experimental results show that the model proposed in this thesis can reach 59.2%accuracy in the data of 5 days.