Does Twitter chatter matter? Online reviews and box office revenues

ABSTRACT

With a rapid rise of text-based social media and online Word-of-Mouth (WOM) activity, millions of people express their thoughts and opinions on a variety of topics. Considering that nowadays WOM is a most influential source of information when guiding consumers’ choice and purchase decisions, in this paper we look at the relationship between Twitter messages (tweets) and cinema box office revenues. Using static and dynamic panel data regression approaches, we show that the frequency, sentiment and timing of tweets posted about a film are correlated to different extent with the movie’s box office revenues, with negative tweets being particularly damaging to the box office revenues. From a managerial perspective, this is important to know, such that film production companies and distributors can adjust their strategy accordingly.

KEYWORDS:

• Social media (Twitter)
• cinema box-office sales
• panel data models

JEL CLASSIFICATION:

• M30
• C33
• C53
• C55

Acknowledgments

We would like to thank Felix Ming Fai Wong for sharing the data from the paper of Wong, Sen, and Chiang (2012).

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ Kaplan and Haenlein (2010) classified social media into six categories by characteristic: (1) Social networks (e.g. Facebook) allow people to reunite or form new communities and attract millions of users on a daily basis; (2) Collaborative projects (e.g. Wikipedia) put a twist on the traditional encyclopaedias and provide the community with easy access to knowledge; (3) Blogs and microblogs (e.g. Twitter) permit people to share their opinions with the world; (4) Content communities (e.g. YouTube) allow various kinds of video sharing for the purpose of entertainment and truth telling; (5) Virtual social worlds (e.g. Second Life) allow users to experience various aspects of face-to-face interaction in a virtual world; (6) Virtual game worlds (e.g. World of Warcraft) are similar to virtual social worlds but with less self-disclosure as user behaviour is limited by strict guidelines.

² O’Connor et al. (2010) connect measures of public opinion measured from polls with sentiment measured from Twitter text messages and demonstrate that relevant sentiments observed from tweets correlate well with polling data on consumer confidence and political opinion, and can also predict future movements in the polls.

³ According to Statista (2016), the number of active Twitter users in 2016 at about 300 million was comparable to Instagram (400 million), Skype (300 million) and Viber (249 million), but it is lagging behind Facebook (1,590 million) and WhatsApp (1,000 million).

⁴ Source: http://www.the-numbers.com/market/ .

⁵ Nielsen ‘Trust in Advertising Report’ (http://www.nielsen.com/us/en/insights/news/2013/under-the-influence-consumer-trust-in-advertising.html) shows that in 2013, 84% of respondents around the globe say that WOM recommendations from friends and family, often referred to as earned advertising, are the most influential, while 68% trust consumer opinions posted online when purchasing goods and services.

⁶ Pentheny (2015) further differs between the source of the review – consumer or critic, and takes into account the identity of the consumer. Koschat (2012) also stresses that the impact of the third-party endorsements on a consumer depends on the consumer’s type and his/her media portfolio, while Chatterjee (2001) specifies that better-informed consumers are less likely affected by the negative reviews. In their paper on the relationship between expert reviews and the demand for wine, Friberg and Grönqvist (2012) argue that identity of a reviewer is also important.

⁷ In order to forecast cinema box office sales using regression or stochastic models, previous research uses information about the movies themselves, such as the movies genre, Motion Picture Association of America (MPAA) rating, running time, release date, the number of screens on which the movie debuted, and the presence of particular actors or actresses in the cast (see for example, Simonoff and Sparrow 2000; De Vany and Walls 1999).

⁸ Based on the information from www.boxofficemojo.com, The Iron Lady was released on the 30 December 2011, The Grey was released on the 27 January 2012, both Chronicle and The Woman in Black were released on the 3 February 2012, while This Means War was released on the 17 February 2012.

⁹ Tweets used for this paper are part of the final dataset by Wong, Sen, and Chiang (2012), collected in the period from 0:00 h on the 3 February 2012 to 0:00 h on the 8 March 2012. They used the Twitter Streaming API (https://dev.twitter.com/streaming/overview – a Twitter tool that enables real time extraction of tweets) to gather a total of 12 million worldwide tweets by tracking keywords in the titles of their 34 selected movies. The posting language was limited to English due to the high level of difficulty to filter for keywords in other languages. After filtering out the spurious tweets with keywords in the wrong order, and accounting for variation in spacing and punctuation, they obtained their final dataset of 1.77 million tweets. Out of these 1.77 million tweets, 51% were classified as irrelevant, and the authors focus their analysis on the remainder of 49% of relevant tweets.

¹⁰ www.boxofficemojo.com.

¹¹ Meryl Streep won the Oscar for the Best Performance by an Actress in a Leading Role, while Mark Coulier and J. Roy Helland won the Oscar for the Best Achievement in Makeup for this movie.

¹² Random effects estimator, although (asymptotically) most efficient estimator, often suffers from the problem of endogenous regressors. Hence, we have opted for the fixed effects estimator here, since we explicitly deal with the potential endogeneity of regressors when estimating the dynamic panel regression results.

¹³ Generally, one can experiment with a second or deeper lags to find a good instrument, but using longer lags reduces sample size. If the number of movies is large enough, one may use all available lags (second and further lags), as well as variables in levels (system GMM) as instruments.

¹⁴ Due to the small number of movies in the sample, a large number of instruments generated by the difference GMM estimator causes the Sargan test of overidentifying restrictions to be weak, hence it is not reported.

¹⁵ To test the robustness of the estimated time effects, we have re-specified the models, including a linear time trend or week dummies instead of a month dummy. We have further tried specification with the inclusion of separate day-of-the-week dummies instead of a weekend dummy. The results for these models led to the same empirical conclusions as the ones presented and are available upon request. We have opted for the reported specifications because they are parsimonious.

¹⁶ For example, Marta Kagan’s work on social media shows that the Millennials have already outnumbered the Baby Boomers, and that 96% of them have already joined the social networks. They further care more what they friends think, that they trust commercial advertising. For further information, see: http://www.slideshare.net/mzkagan/what-the-fk-social-media .