Tweets better than Google Trends at forecasting TV program ratings

How well does the emotional and instantaneous content in tweets perform relative to the more deliberate searches recorded in Google Trends in forecasting future TV ratings? In a massive big data analysis using data from Twitter, Google Trends and other widely used websites for entertainment information, a forthcoming article in the INFORMS journal Marketing Science finds that mining Twitter content is significantly more effective than Google Trends in its ability to predict future TV ratings.

In a business world where the number of episodes and advertising rates are determined by the expected viewership of shows, the science of forecasting views is more important than ever. The research conducted by Professor Xiao Liu of New York University and Professors Param Vir Singh and Kannan Srinivasan of Carnegie Mellon University compared the content of more than 1.8 billion tweets versus 113 million Google searches of about 30 US.

TV shows and regular season NFL games during the 2008-2012 seasons. The authors found that tweets significantly outperform Google searches at predicting future views. Online reviews at IMDB, as well as page views of Wikipedia or Huffington Post also did not efficiently forecast views.

The authors identified some interesting differences in audience interactions with the different websites. Tweets peaked during the show, Google searches peaked after the show, and views of Wikipedia and Huffington Post gradually increased for several days after the show. These results suggest that Twitter captures more of the emotions generated during the show, while the other sites capture more of the information about the show. Professor Kannan Srinivasan conjectured, "It is perhaps the detailed emotional content in the tweets that we are able to extract through our algorithms that plausibly makes Twitter mining such fertile ground for viewership forecasting."

The authors noted that their analysis used state-of-the-art machine learning techniques to mine the text content in tweets. "Merely classifying tweet sentiment as positive and negative is not enough for good forecasting," said Professor Liu. "We developed an approach where the machine learning algorithm would automatically analyze the content of the data to identify relevant features that would be useful for predicting ratings and then automatically classify tweets based on these features."

On the scope of 1.8 billion tweets, Professor Singh explained, "Our automated algorithm ensured the discovery of all relevant tweet characteristics, without human preconceptions, and we exploited the latest in scalable and distributed computing technologies for our analysis."

The authors cautioned that their conclusion about Twitter's superiority in forecasting TV ratings should not be generalized to other settings. Professor Liu noted that "while the emotions captured in Twitter may be more useful for predicting TV ratings, Google Search might be better at forecasting the flu, as information search may be more important here than recording emotion."