Research On Classification Of Breaking Financial News Based On Expectation Deviation

Along with the economics development of China,financial markets and people's lives are more and more tightly correlated together.Existing research works have demonstrated that breaking financial news will strongly affect the financial market immediately.Moreover,with the rapid development of Internet technology and social networks,such impact will be greatly enlarged.Usually,for good news,stock prices will go up,and fall down for the bad news.Unfortunately,it is observed from recent data that quite a number of good news still drive down the stock prices,and this greatly challenges the conventional financial data mining techniques.Most of existing financial data mining techniques,specifically news classification method,fail to classify well in the observed case.The reason is that the traditional classification methods are usually proposed for term features,i.e.,use the characteristics of the news as the feature space,and then predict their class label.However,it is not enough for this observed case.Therefore,after a carefully investigation,we argue that there exist some anchor events which play a key role in affecting stock prices.After this anchor event,if the so called good news or bad news is not consistent with the anchor event,then that news cannot affect the corresponding stocks even if it is a good one.Accordingly,we propose a novel classification method which is based on the proposed expectation deviation.The expectation deviation well illustrates the relationship between the coming news and the anchor event.We extract the expectation by using the topic model,and then propose the classification method.First,we use topic model to find the topics of the news and find whether they can match with each other or not.Then,we use the descriptive dictionary to predict the deviation of the news.Finally,we use our model to predict the label of the news.Our works can be summarized as follows.In order to extract effective news from a large amount of news,we adopt strong vibration filtering method and apply K-means clustering method to screen out trivial news which is loose connected with each other.After calculating expectation deviation,we propose a novel classification method based on it.We propose to judge the consistency by checking whether the topics of anchor event and coming news can overlap or not.If the extent of overlapping exceeds a predefined threshold,then the news is considered to be consistent.The overlap of news is calculated by extracting topics first and then calculating the topic similarity.Only if the topics are similar,we can then quantify the deviation based on dictionary based method.Then,we combine the consistency results calculated by LDA and the quantified deviation together to construct the classifiers for the expectation deviation based news classification task.At last,we performed the careful experiments,and the promising experimental results have demonstrated that,when the financial market is abnormal,the classification method proposed by us can classify the news more accurately than the traditional classification methods.