Short-term Forecast Of Stock Market Based On Text Analysis On Social Network

The arrival of internet age marked the great changes of our way of life. We can get varies kinds of information we want through the internet. And internet has also been becoming an important distributing center of financial information, especially with the development of web, technologies that can provide interaction have been constantly emerging, such as BBS, blog, etc. In recent years, stock investors have been tending to publish personal view of the current stock market on BBS, which generates a lot of web texts of great research value, for these texts often contain the comments on the stock market and investment plan information of stock investors, it is an effective way to understand investors’behavior through these comments.So far, attempts on predicting short-term stock markets have been made by researchers trying to analyze social networks. Researchers abroad mainly focused on the relatively mature European and American markets, so their methods is not adaptive to the immature Chinese stock market. While most of domestic existing works are exploratory, lacking of systemic and quantifiable forecasting work. Therefore, this paper aims at the designing of stock market prediction model by combing the extraction of text source relating stock market with the sentiment analysis. Below illustrates the main research work and contributions in this paper:Firstly, for comments of stock on the internet are likely to reflect the current stock market, they can be used to forecast the stock market quotation in future. In this paper, a modeling method for stock comment text is proposed based on vector space model and word vector model. The obtained word vectors are clustered into k classes by k-means algorithm. Then the mapping rules from text to word set is put forward, through which short texts can be mapped to K-dimension vectors, and finally modeling for the text finished. Experimental results show that the accuracy 68% obtained by the word vector based approaches is significantly higher than that 63.8% got from the vector space model, and both of them are higher than the results shown in literatures.Secondly, owing to the fact that the approaches based on textual features only consider the characteristics of the surface layer of texts and suffer the limit of descript deep information, we propose a sentiment analysis based stock prediction method, In this method, a small amount of words that were labeled with emotional polarity are selected as seeds in advance to calculate the correlation of unknown emotional vocabulary words and seeds, then we got the emotional polarity of unknown words to eventually generate the stock emotional lexicon, with which we can make feature selection for text. Experimental results show that the fusion method performs better than simple text characteristic based methods.