ردیابی رسانه های اجتماعی متعدد برای پیش بینی رویدادها در بازار سهام Tracking Multiple Social Media for Stock Market Event Prediction

Predictions concerning financial markets are complicated by their inherent volatility. Capturing signals of this volatility and providing proper estimates about ‘market flips’ is of prime interest to economists. This problem has attracted great interest from researchers in diverse disciplines such as economics, statistics and data science. Consequently this has led to a wide variety of methods aimed at modeling stock markets [11, 16, 17, 20, 23]. In most of the traditional approaches, researchers characterize the stock market by the historical records of prices and try to find signatures that indicate rising or falling prices based on this historical time series. However, such financial time series methods are generally incognizant of human indicators, such as public reaction, and have frequently been found wanting in their accuracy at predicting sudden, large changes in market value [4]. Recently, with the pervasive growth of social media [6, 14] which allow individuals to readily express their sentiments [21], views and concerns, real-time mining of such factors has become possible. Furthermore, different aspects of public sentiment can now be extracted by analyzing multiple social networks. In this paper, we collect and analyze global search trend data from Google, archived news articles from Bloomberg News and relevant tweets from Twitter. Using unsupervised methods, we extract features from these publicly available data sources. Using these features, we design a set of experiments to investigate the correlations between human behavior and market fluctuations in South American markets. With this analysis, we propose models that predict large changes (events) in market value using the most informative extracted factors. To be specific, given these three data sources at day d and historical stock prices for a market, our proposed models attempt to predict the stock market value of at least day d + 1. The key contributions of this paper are: – We propose a systematic analysis of Google Search Trends, Bloomberg News and Twitter to gather information about market trends and quantify these social media trends. – We identify burst features from Twitter and further group burst features into burst events. We also investigate and find the correlations of these burst events with market trends. – We present Delta Naive Bayes Model to predict finance market fluctuations by fusing multiple social media sources. Though there have been earlier attempts at investigating combinations of sources for finance related applications [12, 19], most of the work has focused on surveying datasets. In that respect, to the best our knowledge, our work is the first not only to compare sources but also to propose predictive models that leverage multiple sources. – Finally, we present our findings about the underlying cross-correlation among market indicators extracted from multiple social media sources and extensively analyze the information from each data source.