News Sentiment and Network Effects

Abstract: Timely understanding of public perception is of vital importance for policymakers, as this measure influences how they respond to public issues. I use the wealth of data available from Twitter to measure public opinion that is traditionally captured with time consuming and expensive surveys. In particular, this study focuses on improving forecasting performance of social and economic activities, such as daily presidential approval ratings, through machine learning techniques. It measures media sentiment shocks and evaluates the network effects via both active (retweets) and passive (favorites) network activities. Additionally, I build a neural network to account for nonlinearity in emotion-based predictions. I train the sentiment index measured from news and its network response on measures available from polls, and find that using social media data can establish a presidential approval index in real-time and with more precision. I also find that shocks in media sentiments have a more immediate effect, while network predictors have a longer lasting impact. Granger analysis shows that the past values of the media sentiment and its network effects are important beyond past values of the approval index alone. Results indicate that the model can successfully classify the public’s opinions and emotions. In the second part of the paper I evaluate the effects of “self-promotion”. Results suggest that signals found in the sentiment and the network response to the President’s own tweets are predecessors to the same effects in media, and consequently to the President’s approval index.   

Markets and Customer Confidence: Analysis of Economic and Financial News and Social Network Signals

Abstract: Through the opinion mining approach, I test whether a data-driven model of opinion dynamics is able to accurately forecast public sentiment from active online participants. I construct daily economic and market sentiment indices to evaluate if network signals have predictive capabilities of economic indicators. Results indicate that volatility index is best predicted with short term rolling averages of three days. For the consumer confidence index, a longer time frame is needed, up to four weeks, as customers are much more dependent on macroeconomic factors than short term market fluctuations. Hence, findings suggest that we can have interim snapshots, by using social media, as long as we adjust the windows and properly select keywords.