基于分步降维框架的股吧短文本数据关键信息抽取与个股收益预测

Social media is an important source for stock market investors to obtain financial information, where the emotional and other relevant signals contained in users’ posts are closely related to the stock prices. However, since these posts are mostly freely expressed, highly colloquial, short in length and extremely low in semantic density, the traditional text analysis method based on sentiment dictionary faces with the risk of losing valuable information. The recent multi-step dimension reduction framework proposed by Fan et al. (2021) attempts to improve the preciseness of substantial information extraction from text data by making full use of the semantic features within the text in a data-driven way. This paper extends this framework to a social media scenario and systematically explores whether the short texts on Eastmoney Guba provide effective leading information for individual stock prices. Specifically, the principal component analysis method is used to extract common factors in the text, and then variable screening is performed on the residual matrix to further filter features of words in the text. Then Lasso regression is used to build a prediction model, by which the unique semantics for individual stocks contained in the text are extracted. The results show that the framework can indeed extract the useful information from the short texts in Guba for individual stock returns prediction. In addition, the identified vocabulary sets with predictive power also reflect the characteristics of social media short texts, which are different from not only other financial texts but also the traditional financial sentiment dictionaries. Therefore, this multi-step dimension reduction framework provides a new path for leveraging social media short-text data in various domains.