Trend Extraction from High Dimensional Stock News based on Markov Chain

Abstract

High dimensional data causes a challenge for the researchers to extract useful features. Information Extraction achieves a better interpretability and retains physical properties of the original features on these data. Our research focuses on extracting relevant features from high dimensional stock news to support stock price status or trend prediction. In our previous research, trend related features have been extracted from sentence-level contents in news using Named Entities and their relations. This paper extracts stock trends from the trend related information in one or more news based on Markov Chain model. Based on these features in the model, there are three states: Decrease (D), Stable (S) and Increase (I). The trend prediction results are based on the probability of trend fluctuations from news by using transition probability matrices and initial state vectors. The accuracy of the prediction results are calculated by comparing the different transition types and different data periods.