| With large increase in investors and the continuous evolution of financial derivatives,the structure of the financial market has become more complex and changeable,new challenges and opportunities are made for market stability and development.In particular,with the rapid development of the mobile internet,information is rapidly spreading and growing at an unprecedented rate,investors use this information to handle their investment business,which keep affecting the stability and development of the market.As one of the most important source of information,how to make effective use of complicated online news to predict the financial market plays an active role in the stable development of the market and the market participants' business decisions.Text data has many features such as semantic diversity,complex and changeable syntax,metaphorical expression,and so on,which makes forecasting system based on the traditional text mining technology suffer from lots of deficiencies,such as the limited ability of text representation and manual text extraction.Therefore,this paper combines traditional natural language processing techniques and deep learning models to conduct in-depth research and practice of financial market prediction tasks,which base on semantic analysis of text news.Those subtasks includes representation learning,global feature extraction,and temporal data modeling.The specific research content is as follows:1.Forecasting model of S&P 500 Index based on event extraction techniques.The primary task of natural language processing research is how to effectively represent the text semantic information as a machine language.Traditional representation models including Bag-Of-Words(BOW)have some shortcomings such as the inability to characterize contextual semantic information,and the high dimension and high sparseness of feature vectors.This paper uses syntactic parse tree to extract the subject-predicate component of the news headline,uses word embedding techniques to represent the subject-predicate component as a low-dimensional,dense real-valued vector,and combines Tensor neural networks(NTNs)for further feature extraction of event vectors.Finally,Recurrent Neural Networks(RNNs)based on LSTM is used to fit their mapping relations between news events and index fluctuations.2.Forecasting model of S&P 500 Index based on Seq2Seq(Sequence-to-Sequence)models.On the basis of event extraction,this paper further proposes SV(Skip-thought Vector)model to characterize event vectors of news headlines.Large-scale news corpus is used to training the model to convert semantic information into machine language effectively.Information is converted into machine language.At the same time,in order to effectively map all news events on the same day to the only index of the day,this paper uses FV(Fisher Vector)coding to globally characterize and extract event vectors,and designs a bidirectional LSTM-RNN model to fit the sequence relationship between news events and index fluctuations.Based on the above two forecasting models,we use semantic distance,accuracy,MCC,and other indicators to evaluate the model.The two combined forecast models based on news texts achieve 64.06% and 64.55% prediction accuracy respectively on the S&P 500 index forecast task,and according to the experimental analysis,the following conclusions are drawn: 1.The news text can be used as one of the effective characteristic indicators for predicting the future index;2.The news text information has a time-effect on the index effect,which will decrease gradually over time;3.The effective representation of textual information helps to improve the prediction accuracy of the index. |
PDF Full Text Request