Credibility assessment of financial stock tweets

摘要

Social media plays an important role in facilitating conversations and news dissemination. Specifically, Twitter has recently seen use by investors to facilitate discussions surrounding stock exchange-listed companies. Investors depend on timely, credible information being made available in order to make well-informed investment decisions, with credibility being defined as the believability of information. Much work has been done on assessing credibility on Twitter in domains such as politics and natural disaster events, but the work on assessing the credibility of financial statements is scant within the literature. Investments made on apocryphal information could hamper efforts of social media's aim of providing a transparent arena for sharing news and encouraging discussion of stock market events. This paper presents a novel methodology to assess the credibility of financial stock market tweets, which is evaluated by conducting an experiment using tweets pertaining to companies listed on the London Stock Exchange. Three sets of traditional machine learning classifiers (using three different feature sets) are trained using an annotated dataset. We highlight the importance of considering features specific to the domain in which credibility needs to be assessed for - in the case of this paper, financial features. In total, after discarding non-informative features, 34 general features are combined with over 15 novel financial features for training classifiers. Results show that classifiers trained on both general and financial features can yield improved performance than classifiers trained on general features alone, with Random Forest being the top performer, although the Random Forest model requires more features (37) than that of other classifiers (such as K-Nearest Neighbours 9) to achieve such performance.