Analysis of the impact of investor sentiment on stock price using the latent dirichlet allocation topic model

Investor sentiment has always been an active research topic in finance. In recent years, text mining, machine learning and sentiment analysis have been very fruitful, and researchers can extract valuable information from social platforms more promptly and accurately. Unsupervised learning avoids the subjective human selection of data while reducing the workload. This paper uses the primary model for the unsupervised learning total probability generative model LDA (Latent Dirichlet Allocation). Natural language processing and word-splitting tools empirically analyze text data from a well-known financial and stock information website. An attempt is made to explore the correlation with stock excess return. The significant findings are as follows. First, investor sentiment classified by theme is positively correlated with excess return. Second, different themes have different degrees of influence, with "broad market sentiment" affecting the short term, corporate development involving a long time, and "corporate dividends" affecting both. Third, there is an asymmetric effect of investor sentiment on excess return.