Predicting the Future With Social Media

Sitaram Asur Bernardo A. Huberman

Social Computing Lab Social Computing Lab
HP Labs HP Labs
Palo Alto, California Palo Alto, California
Email: sitaram.asur@hp.com Email: bernardo.huberman@hp.com

Abstract—In recent years, social media has become ubiquitous This paper reports on such a study. Specifically we consider
and important for social networking and content sharing. And the task of predicting box-office revenues for movies using
yet, the content that is generated from these websites remains the chatter from Twitter, one of the fastest growing social
largely untapped. In this paper, we demonstrate how social media
content can be used to predict real-world outcomes. In particular, networks in the Internet. Twitter 1 , a micro-blogging network,
we use the chatter from Twitter.com to forecast box-office has experienced a burst of popularity in recent months leading
revenues for movies. We show that a simple model built from to a huge user-base, consisting of several tens of millions of
the rate at which tweets are created about particular topics can users who actively participate in the creation and propagation
outperform market-based predictors. We further demonstrate of content.
how sentiments extracted from Twitter can be further utilized to
improve the forecasting power of social media. We have focused on movies in this study for two main
reasons.
I. I NTRODUCTION • The topic of movies is of considerable interest among
the social media user community, characterized both by
Social media has exploded as a category of online discourse large number of users discussing movies, as well as a
where people create content, share it, bookmark it and network substantial variance in their opinions.
at a prodigious rate. Examples include Facebook, MySpace, • The real-world outcomes can be easily observed from
Digg, Twitter and JISC listservs on the academic side. Because box-office revenue for movies.
of its ease of use, speed and reach, social media is fast Our goals in this paper are as follows. First, we assess how
changing the public discourse in society and setting trends buzz and attention is created for different movies and how that
and agendas in topics that range from the environment and changes over time. Movie producers spend a lot of effort and
politics to technology and the entertainment industry. money in publicizing their movies, and have also embraced
Since social media can also be construed as a form of the Twitter medium for this purpose. We then focus on the
collective wisdom, we decided to investigate its power at mechanism of viral marketing and pre-release hype on Twitter,
predicting real-world outcomes. Surprisingly, we discovered and the role that attention plays in forecasting real-world box-
that the chatter of a community can indeed be used to make office performance. Our hypothesis is that movies that are well
quantitative predictions that outperform those of artificial talked about will be well-watched.
markets. These information markets generally involve the Next, we study how sentiments are created, how positive and
trading of state-contingent securities, and if large enough and negative opinions propagate and how they influence people.
properly designed, they are usually more accurate than other For a bad movie, the initial reviews might be enough to
techniques for extracting diffuse information, such as surveys discourage others from watching it, while on the other hand, it
and opinions polls. Specifically, the prices in these markets is possible for interest to be generated by positive reviews and
have been shown to have strong correlations with observed opinions over time. For this purpose, we perform sentiment
outcome frequencies, and thus are good indicators of future analysis on the data, using text classifiers to distinguish
outcomes [4], [5]. positively oriented tweets from negative.
In the case of social media, the enormity and high vari- Our chief conclusions are as follows:
ance of the information that propagates through large user • We show that social media feeds can be effective indica-
communities presents an interesting opportunity for harnessing tors of real-world performance.
that data into a form that allows for specific predictions • We discovered that the rate at which movie tweets
about particular outcomes, without having to institute market are generated can be used to build a powerful model
mechanisms. One can also build models to aggregate the for predicting movie box-office revenue. Moreover our
opinions of the collective population and gain useful insights predictions are consistently better than those produced
into their behavior, while predicting future trends. Moreover, by an information market such as the Hollywood Stock
gathering information on how people converse regarding par- Exchange, the gold standard in the industry [4].
ticular products can be helpful when designing marketing and
advertising campaigns [1], [3]. 1 http://www.twitter.com
 • Our analysis of the sentiment content in the tweets shows in Jan 3 ). It can be considered a directed social network, where
that they can improve box-office revenue predictions each user has a set of subscribers known as followers. Each
based on tweet rates only after the movies are released. user submits periodic status updates, known as tweets, that
This paper is organized as follows. Next, we survey recent consist of short messages of maximum size 140 characters.
related work. We then provide a short introduction to Twitter These updates typically consist of personal information about
and the dataset that we collected. In Section 5, we study how the users, news or links to content such as images, video
attention and popularity are created and how they evolve. and articles. The posts made by a user are displayed on the
We then discuss our study on using tweets from Twitter user’s profile page, as well as shown to his/her followers. It is
for predicting movie performance. In Section 6, we present also possible to send a direct message to another user. Such
our analysis on sentiments and their effects. We conclude messages are preceded by @userid indicating the intended
in Section 7. We describe our prediction model in a general destination.
context in the Appendix. A retweet is a post originally made by one user that is
forwarded by another user. These retweets are a popular means
II. R ELATED W ORK of propagating interesting posts and links through the Twitter
community.
Although Twitter has been very popular as a web service, Twitter has attracted lots of attention from corporations
there has not been considerable published research on it. for the immense potential it provides for viral marketing.
Huberman and others [2] studied the social interactions on Due to its huge reach, Twitter is increasingly used by news
Twitter to reveal that the driving process for usage is a sparse organizations to filter news updates through the community.
hidden network underlying the friends and followers, while A number of businesses and organizations are using Twitter
most of the links represent meaningless interactions. Java et or similar micro-blogging services to advertise products and
al [7] investigated community structure and isolated different disseminate information to stakeholders.
types of user intentions on Twitter. Jansen and others [3]
have examined Twitter as a mechanism for word-of-mouth IV. DATASET C HARACTERISTICS
advertising, and considered particular brands and products The dataset that we used was obtained by crawling hourly
while examining the structure of the postings and the change in feed data from Twitter.com. To ensure that we obtained all
sentiments. However the authors do not perform any analysis tweets referring to a movie, we used keywords present in the
on the predictive aspect of Twitter. movie title as search arguments. We extracted tweets over
There has been some prior work on analyzing the correlation frequent intervals using the Twitter Search Api 4 , thereby
between blog and review mentions and performance. Gruhl ensuring we had the timestamp, author and tweet text for
and others [9] showed how to generate automated queries our analysis. We extracted 2.89 million tweets referring to 24
for mining blogs in order to predict spikes in book sales. different movies released over a period of three months.
And while there has been research on predicting movie sales, Movies are typically released on Fridays, with the exception
almost all of them have used meta-data information on the of a few which are released on Wednesday. Since an average of
movies themselves to perform the forecasting, such as the 2 new movies are released each week, we collected data over
movies genre, MPAA rating, running time, release date, the a time period of 3 months from November to February to have
number of screens on which the movie debuted, and the sufficient data to measure predictive behavior. For consistency,
presence of particular actors or actresses in the cast. Joshi we only considered the movies released on a Friday and only
and others [10] use linear regression from text and metadata those in wide release. For movies that were initially in limited
features to predict earnings for movies. Sharda and Delen [8] release, we began collecting data from the time it became
have treated the prediction problem as a classification problem wide. For each movie, we define the critical period as the
and used neural networks to classify movies into categories time from the week before it is released, when the promotional
ranging from ’flop’ to ’blockbuster’. Apart from the fact campaigns are in full swing, to two weeks after release, when
that they are predicting ranges over actual numbers, the best its initial popularity fades and opinions from people have been
accuracy that their model can achieve is fairly low. Zhang disseminated.
and Skiena [6] have used a news aggregation model along Some details on the movies chosen and their release dates
with IMDB data to predict movie box-office numbers. We are provided in Table 1. Note that, some movies that were
have shown how our model can generate better results when released during the period considered were not used in this
compared to their method. study, simply because it was difficult to correctly identify
tweets that were relevant to those movies. For instance,
III. T WITTER for the movie 2012, it was impractical to segregate tweets
Launched on July 13, 2006, Twitter 2 is an extremely talking about the movie, from those referring to the year. We
popular online microblogging service. It has a very large user have taken care to ensure that the data we have used was
base, consisting of several millions of users (23M unique users 3 http://blog.compete.com/2010/02/24/compete-ranks-top-sites-for-january-
2010/
2 http://www.twitter.com 4 http://search.twitter.com/api/
 Movie Release Date
2
Armored 2009-12-04
Release weekend
Avatar 2009-12-18 1.9

The Blind Side 2009-11-20 1.8

The Book of Eli 2010-01-15 1.7

Daybreakers 2010-01-08

Tweets per authors

1.6
Dear John 2010-02-05
Did You Hear About The Morgans 2009-12-18 1.5

Edge Of Darkness 2010-01-29 1.4

Extraordinary Measures 2010-01-22 1.3

From Paris With Love 2010-02-05

1.2
The Imaginarium of Dr Parnassus 2010-01-08
Invictus 2009-12-11 1.1

Leap Year 2010-01-08 1

2 4 6 8 10 12 14 16 18 20

Legion 2010-01-22 Days

Twilight : New Moon 2009-11-20

Pirate Radio 2009-11-13
Princess And The Frog 2009-12-11 Fig. 2. Number of tweets per unique authors for different movies
Sherlock Holmes 2009-12-25
Spy Next Door 2010-01-15
The Crazies 2010-02-26 14

Tooth Fairy 2010-01-22

Transylmania 2009-12-04 12

When In Rome 2010-01-29

10
Youth In Revolt 2010-01-08

log(frequency)
8

TABLE I
NAMES AND RELEASE DATES FOR THE MOVIES WE CONSIDERED IN OUR 6

ANALYSIS .
4

disambiguated and clean by choosing appropriate keywords 0

0 1 2 3 4 5 6 7 8
log(tweets)
and performing sanity checks.

Fig. 3. Log distribution of authors and tweets.

4500

release weekend weekend 2

4000

3500

3000
authors over the critical period. The X-axis shows the number
2500
of tweets in the log scale, while the Y-axis represents the
corresponding frequency of authors in the log scale. We can
2000

observe that it is close to a Zipfian distribution, with a few

1500
authors generating a large number of tweets. This is consistent
1000
with observed behavior from other networks [12]. Next, we
500
examine the distribution of authors over different movies. Fig 4
2 4 6 8 10 12 14 16 18 20 shows the distribution of authors and the number of movies
they comment on. Once again we find a power-law curve, with
a majority of the authors talking about only a few movies.
Fig. 1. Time-series of tweets over the critical period for different movies.
V. ATTENTION AND P OPULARITY
The total data over the critical period for the 24 movies We are interested in studying how attention and popularity
we considered includes 2.89 million tweets from 1.2 million are generated for movies on Twitter, and the effects of this
users. attention on the real-world performance of the movies consid-
Fig 1 shows the timeseries trend in the number of tweets ered.
for movies over the critical period. We can observe that the
busiest time for a movie is around the time it is released, A. Pre-release Attention:
following which the chatter invariably fades. The box-office Prior to the release of a movie, media companies and and
revenue follows a similar trend with the opening weekend producers generate promotional information in the form of
generally providing the most revenue for a movie. trailer videos, news, blogs and photos. We expect the tweets
Fig 2 shows how the number of tweets per unique author for movies before the time of their release to consist primarily
changes over time. We find that this ratio remains fairly of such promotional campaigns, geared to promote word-of-
consistent with a value between 1 and 1.5 across the critical mouth cascades. On Twitter, this can be characterized by
period. Fig 3 displays the distribution of tweets by different tweets referring to particular urls (photos, trailers and other
 5
x 10
Features Adjusted R2 p-value
10
Avg Tweet-rate 0.80 3.65e-09
9
Tweet-rate timeseries 0.93 5.279e-09
8 Tweet-rate timeseries + thcnt 0.973 9.14e-12
7
HSX timeseries + thcnt 0.965 1.030e-10
6

TABLE IV
Authors

5
C OEFFICIENT OF D ETERMINATION (R2 ) VALUES USING DIFFERENT
4 PREDICTORS FOR MOVIE BOX - OFFICE REVENUE FOR THE FIRST WEEKEND .
3

0
there is a greater percentage of tweets containing urls in the
2 4 6 8 10 12 14 16 18 20 22 24
Number of Movies
week prior to release than afterwards. This is consistent with
our expectation. In the case of retweets, we find the values to
Fig. 4. Distribution of total authors and the movies they comment on. be similar across the 3 weeks considered. In all, we found the
retweets to be a significant minority of the tweets on movies.
Features Week 0 Week 1 Week 2 One reason for this could be that people tend to describe their
url 39.5 25.5 22.5 own expectations and experiences, which are not necessarily
retweet 12.1 12.1 11.66 propaganda.
We want to determine whether movies that have greater
TABLE II
U RL AND RETWEET PERCENTAGES FOR CRITICAL WEEK publicity, in terms of linked urls on Twitter, perform better in
the box office. When we examined the correlation between the
urls and retweets with the box-office performance, we found
the correlation to be moderately positive, as shown in Table
promotional material) as well as retweets, which involve users 3. However, the adjusted R2 value is quite low in both cases,
forwarding tweet posts to everyone in their friend-list. Both indicating that these features are not very predictive of the
these forms of tweets are important to disseminate information relative performance of movies. This result is quite surprising
regarding movies being released. since we would expect promotional material to contribute
First, we examine the distribution of such tweets for dif- significantly to a movie’s box-office income.
ferent movies, following which we examine their correlation
B. Prediction of first weekend Box-office revenues
with the performance of the movies.
Next, we investigate the power of social media in predicting
real-world outcomes. Our goal is to observe if the knowledge
0.7
Week 0
Week 1
that can be extracted from the tweets can lead to reasonably
Week 2
0.6 accurate prediction of future outcomes in the real world.
The problem that we wish to tackle can be framed as
Tweets with urls (percentage)

0.5

follows. Using the tweets referring to movies prior to their

0.4
release, can we accurately predict the box-office revenue
0.3 generated by the movie in its opening weekend?
0.2
7
x 10
15

Tweet−rate
0.1 HSX

0
2 4 6 8 10 12 14 16 18 20 22 24
Movies
10
Actual revenue

Fig. 5. Percentages of urls in tweets for different movies.

Table 2 shows the percentages of urls and retweets in the

tweets over the critical period for movies. We can observe that
0
0 2 4 6 8 10 12 14 16
7
x 10
Predicted Box−office Revenue
Features Correlation R2
url 0.64 0.39
retweet 0.5 0.20 Fig. 6. Predicted vs Actual box office scores using tweet-rate and HSX
predictors
TABLE III
C ORRELATION AND R2 VALUES FOR URLS AND RETWEETS BEFORE
RELEASE . To use a quantifiable measure on the tweets, we define the
tweet-rate, as the number of tweets referring to a particular
 Predictor AMAPE Score
movie per hour. Regnobudget +nReg1w 3.82 96.81
Avg Tweet-rate + thcnt 1.22 98.77
|tweets(mov)|
T weet − rate(mov) = (1) Tweet-rate Timeseries + thcnt 0.56 99.43
|T ime (in hours)|
TABLE V
Our initial analysis of the correlation of the average tweet- AMAPE AND S CORE VALUE COMPARISON WITH EARLIER WORK .
rate with the box-office gross for the 24 movies considered
showed a strong positive correlation, with a correlation coeffi-
cient value of 0.90. This suggests a strong linear relationship
among the variables considered. Accordingly, we constructed From the results in Table 4, it can be seen that our
a linear regression model using least squares of the average regression model built from social media provides an
of all tweets for the 24 movies considered over the week accurate prediction of movie performances at the box
prior to their release. We obtained an adjusted R2 value office. Furthermore, the model built using the tweet rate
of 0.80 with a p-value of 3.65e − 09 ∗ ∗∗, where the ’***’ timeseries outperforms the HSX-based model. The graph
shows significance at 0.001, indicating a very strong predictive outlining the predicted and actual values of this model is also
relationship. Notice that this performance was achieved using shown in Fig 6, outlining the utility of harvesting social media.
only one variable (the average tweet rate). To evaluate our
predictions, we employed real box-office revenue information, Comparison with News-based Prediction:
extracted from the Box Office Mojo website 5 . In earlier work, Zhang and others [6] have developed a
The movie T ransylmania that opened on Dec 4th had news-based model for predicting movie revenue. The best-
easily the lowest tweet-rates of all movies considered. For performing method in the aforementioned work is the com-
the week prior to its release, it received on an average 2.75 bined model obtained by using predictors from IMDB and
tweets per hour. As a result of this lack of attention, the news. The corresponding R2 value for this combined model
movie captured the record for the lowest-grossing opening for is 0.788, which is far lower than the ones obtained by
a movie playing at over 1,000 sites, making only $263,941 our predictors. We computed the AMAPE (Adjusted Mean
in its opening weekend, and was subsequently pulled from Absolute Percentage/Relative Error) measure, that the authors
theaters at the end of the second week. On the other end use, for our data. The comparative values are shown in Table
of the spectrum, two movies that made big splashes in their 5. We can observe that our values are far better than the ones
opening weekends, Twilight:New Moon (making 142M) and reported in the earlier work. Note however, that since historical
Avatar(making 77M) had, for their pre-release week, averages information on tweets are not available, we were able to use
of 1365.8 and 1212.8 tweets per hour respectively. This once data on only the movies we have collected, while the authors
again illustrates the importance of attention in social media. in the earlier paper have used a larger database of movies for
Next, we performed a linear regression of the time series their analysis.
values of the tweet-rate for the 7 days before the release.
We used 7 variables each corresponding to the tweet-rate C. Predicting HSX prices
for a particular day. An additional variable we used was the Given that social media can accurately predict box office
number of theaters the movies were released in, thcnt. The results, we also tested their efficacy at forecasting the stock
results of the regression experiments are shown in Table 4. prices of the HSX index. At the end of the first weekend,
Note that, in all cases, we are using only data available prior the Hollywood stock exchange adjusts the price for a movie
to the release to predict box-office for the opening weekend. stock to reflect the actual box office gross. If the movie does
not perform well, the price goes down and vice versa. We
Comparison with HSX: conducted an experiment to see if we could predict the price
To compare with our tweet-based model, we used the Hol- of the HSX movie stock at the end of the opening weekend
lywood Stock Exchange index. The fact that artificial online for the movies we have considered. We used the historical
markets such as the Foresight Exchange and the Hollywood HSX prices as well as the tweet-rates, individually, for the
Stock Exchange are good indicators of future outcomes has week prior to the release as predictive variables. The response
been shown previously [4], [5]. The prices in these markets variable was the adjusted price of the stock. We also used
have been shown to have strong correlations with observed the theater count as a predictor in both cases, as before. The
outcome frequencies. In the case of movies, the Hollywood results are summarized in Table 6. As is apparent, the tweet-
Stock Exchange (http://www.hsx.com/), is a popular play- rate proves to be significantly better at predicting the actual
money market, where the prices for movie stocks can ac- values than the historical HSX prices. This again illustrates
curately predict real box office results. Hence, to compare the power of the buzz from social media.
with our tweet-rate predictor, we considered regression on
D. Predicting revenues for all movies for a given weekend
the movie stock prices from the Hollywood Stock Exchange,
which can be considered the gold standard [4]. Until now, we have considered the problem of predicting
opening weekend revenue for movies. Given the success of
5 http://boxofficemojo.com the regression model, we now attempt to predict revenue for
 Predictor Adjusted R2 p − value
HSX timeseries + thcnt 0.95 4.495e-10
Tweet-rate timeseries + thnt 0.97 2.379e-11

TABLE VI
P REDICTION OF HSX END OF OPENING WEEKEND PRICE .

Predictor Adjusted R2 p − value

Avg Tweet-rate 0.79 8.39e-09
Avg Tweet-rate + thcnt 0.83 7.93e-09
Avg Tweet-rate + PNratio 0.92 4.31e-12
Tweet-rate timeseries 0.84 4.18e-06
Tweet-rate timeseries + thcnt 0.863 3.64e-06
Tweet-rate timeseries + PNratio 0.94 1.84e-08

TABLE VIII
P REDICTION OF SECOND WEEKEND BOX - OFFICE GROSS

Weekend Adjusted R2
Jan 15-17 0.92
Sentiment analysis is a well-studied problem in linguistics
Jan 22-24 0.97 and machine learning, with different classifiers and language
Jan 29-31 0.92 models employed in earlier work [13], [14]. It is common
Feb 05-07 0.95 to express this as a classification problem where a given
text needs to be labeled as P ositive, N egative or N eutral.
TABLE VII
C OEFFICIENT OF D ETERMINATION (R2 ) VALUES USING TWEET- RATE Here, we constructed a sentiment analysis classifier using the
TIMESERIES FOR DIFFERENT WEEKENDS LingPipe linguistic analysis package 6 which provides a set
of open-source java libraries for natural language processing
tasks. We used the DynamicLMClassifier which is a language
model classifier that accepts training events of categorized
all movies over a particular weekend. The Hollywood Stock character sequences. Training is based on a multivariate es-
Exchange de-lists movie stocks after 4 weeks of release, which timator for the category distribution and dynamic language
means that there is no timeseries available for movies after models for the per-category character sequence estimators.
4 weeks. In the case of tweets, people continue to discuss To obtain labeled training data for the classifier, we utilized
movies long after they are released. Hence, we attempt to use workers from the Amazon Mechanical Turk 7 . It has been
the timeseries of tweet-rate, over 7 days before the weekend, shown that manual labeling from Amazon Turk can correlate
to predict the box-office revenue for that particular weekend. well with experts [11]. We used thousands of workers to assign
Table 7 shows the results for 3 weekends in January and sentiments for a large random sample of tweets, ensuring that
1 in February. Note, that there were movies that were two each tweet was labeled by three different people. We used
months old in consideration for this experiment. Apart from only samples for which the vote was unanimous as training
the time series, we used two additional variables - the theater data. The samples were initially preprocessed in the following
count and the number of weeks the movie has been released. ways:
We used the coefficient of determination (adjusted R2 ) to
evaluate the regression models. From Table 7, we find that • Elimination of stop-words
the tweets continue to be good predictors even in this case, • Elimination of all special characters except exclamation
with an adjusted R2 consistently greater than 0.90. The results marks which were replaced by < EX > and question
have shown that the buzz from social media can be accurate marks (< QM >)
indicators of future outcomes. The fact that a simple linear • Removal of urls and user-ids
regression model considering only the rate of tweets on movies • Replacing the movie title with < M OV >
can perform better than artificial money markets, illustrates the We used the pre-processed samples to train the classifier using
power of social media. an n-gram model. We chose n to be 8 in our experiments.
The classifier was trained to predict three classes - Positive,
VI. S ENTIMENT A NALYSIS Negative and Neutral. When we tested on the training-set with
Next, we would like to investigate the importance of sen- cross-validation, we obtained an accuracy of 98%. We then
timents in predicting future outcomes. We have seen how used the trained classifier to predict the sentiments for all the
efficient the attention can be in predicting opening weekend tweets in the critical period for all the movies considered.
box-office values for movies. Hence we consider the problem
of utilizing the sentiments prevalent in the discussion for 6 http://www.alias-i.com/lingpipe

forecasting. 7 https://www.mturk.com/
 Movie Subjectivity Variable p − value
1.6
(Intercept) 0.542
1.4 Avg Tweet-rate 2.05e-11 (***)
PNRatio 9.43e-06 (***)
1.2

1 TABLE IX
R EGRESSION USING THE AVERAGE TWEET- RATE AND THE POLARITY
0.8
(PNR ATIO ). T HE SIGNIFICANCE LEVEL (*:0.05, **: 0.01, ***: 0.001) IS
ALSO SHOWN .
0.6

0.4

0.2

0
positive than negative tweets is likely to be successful.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Week 0 Week 1 Week 2

|T weets with P ositive Sentiment|
P N ratio = (3)
Fig. 7. Movie Subjectivity values |T weets with N egative Sentiment|
Fig 8 shows the polarity values for the movies considered
Movie Polarity in the critical period. We find that there are more positive
12
sentiments than negative in the tweets for almost all the
movies. The movie with the enormous increase in positive
10

sentiment after release is The Blind Side (5.02 to 9.65). The

8
movie had a lukewarm opening weekend sales (34M) but then
6 boomed in the next week (40.1M), owing largely to positive
4
sentiment. The movie New Moon had the opposite effect. It
released in the same weekend as Blind Side and had a great
2
first weekend but its polarity reduced (6.29 to 5), as did its
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
box-office revenue (142M to 42M) in the following week.
Considering that the polarity measure captured some vari-
Week 0 Week 1 Week 2
ance in the revenues, we examine the utility of the sentiments
in predicting box-office sales. In this case, we considered
Fig. 8. Movie Polarity values the second weekend revenue, since we have seen subjectivity
increasing after release. We use linear regression on the
revenue as before, using the tweet-rate and the PNratio as an
additional variable. The results of our regression experiments
A. Subjectivity
are shown in Table 8. We find that the sentiments do provide
Our expectation is that there would be more value for improvements, although they are not as important as the rate
sentiments after the movie has released, than before. We of tweets themselves. The tweet-rate has close to the same
expect tweets prior to the release to be mostly anticipatory predictive power in the second week as the first. Adding the
and stronger positive/negative tweets to be disseminated later sentiments, as an additional variable, to the regression equation
following the release. Positive sentiments following the release improved the prediction to 0.92 while used with the average
can be considered as recommendations by people who have tweet-rate, and 0.94 with the tweet-rate timeseries. Table 9
seen the movie, and are likely to influence others from shows the regression p-values using the average tweet rate
watching the same movie. To capture the subjectivity, we and the sentiments. We can observe that the coefficients are
defined a measure as follows. highly significant in both cases.
|P ositive and N egative T weets|
Subjectivity = (2) VII. C ONCLUSION
|N eutral T weets|
In this article, we have shown how social media can be
When we computed the subjectivity values for all the movies, utilized to forecast future outcomes. Specifically, using the
we observed that our hypothesis was true. There were more rate of chatter from almost 3 million tweets from the popular
sentiments discovered in tweets for the weeks after release, site Twitter, we constructed a linear regression model for
than in the pre-release week. Fig 7 shows the ratio of subjec- predicting box-office revenues of movies in advance of their
tive to objective tweets for all the movies over the three weeks. release. We then showed that the results outperformed in
We can observe that for most of the movies, the subjectivity accuracy those of the Hollywood Stock Exchange and that
increases after release. there is a strong correlation between the amount of attention
a given topic has (in this case a forthcoming movie) and
B. Polarity its ranking in the future. We also analyzed the sentiments
To quantify the sentiments for a movie, we measured the present in tweets and demonstrated their efficacy at improving
ratio of positive to negative tweets. A movie that has far more predictions after a movie has released.
 While in this study we focused on the problem of predicting [4] D. M. Pennock, S. Lawrence, C. L. Giles, and F. Å. Nielsen. The real
box office revenues of movies for the sake of having a clear power of artificial markets. Science, 291(5506):987–988, Jan 2001.
[5] Kay-Yut Chen, Leslie R. Fine and Bernardo A. Huberman. Predicting
metric of comparison with other methods, this method can be the Future. Information Systems Frontiers, 5(1):47–61, 2003.
extended to a large panoply of topics, ranging from the future [6] W. Zhang and S. Skiena. Improving movie gross prediction through news
rating of products to agenda setting and election outcomes. At analysis. In Web Intelligence, pages 301304, 2009.
[7] Akshay Java, Xiaodan Song, Tim Finin and Belle Tseng. Why we twitter:
a deeper level, this work shows how social media expresses a understanding microblogging usage and communities. Proceedings of the
collective wisdom which, when properly tapped, can yield an 9th WebKDD and 1st SNA-KDD 2007 workshop on Web mining and social
extremely powerful and accurate indicator of future outcomes. network analysis, pages 56–65, 2007.
[8] Ramesh Sharda and Dursun Delen. Predicting box-office success of
motion pictures with neural networks. Expert Systems with Applications,
VIII. A PPENDIX : G ENERAL P REDICTION M ODEL FOR vol 30, pp 243–254, 2006.
S OCIAL M EDIA [9] Daniel Gruhl, R. Guha, Ravi Kumar, Jasmine Novak and Andrew
Tomkins. The predictive power of online chatter. SIGKDD Conference
Although we focused on movie revenue prediction in this on Knowledge Discovery and Data Mining, 2005.
paper, the method that we advocate can be extended to other [10] Mahesh Joshi, Dipanjan Das, Kevin Gimpel and Noah A. Smith. Movie
Reviews and Revenues: An Experiment in Text Regression NAACL-HLT,
products of consumer interest. 2010.
We can generalize our model for predicting the revenue [11] Rion Snow, Brendan O’Connor, Daniel Jurafsky and Andrew Y. Ng.
of a product using social media as follows. We begin with Cheap and Fast - But is it Good? Evaluating Non-Expert Annotations for
Natural Language Tasks. Proceedings of EMNLP, 2008.
data collected regarding the product over time, in the form [12] Fang Wu, Dennis Wilkinson and Bernardo A. Huberman. Feeback Loops
of reviews, user comments and blogs. Collecting the data of Attention in Peer Production. Proceedings of SocialCom-09: The 2009
over time is important as it can measure the rate of chatter International Conference on Social Computing, 2009.
[13] Bo Pang and Lillian Lee. Opinion Mining and Sentiment Analysis
effectively. The data can then be used to fit a linear regression Foundations and Trends in Information Retrieval, 2(1-2), pp. 1135, 2008.
model using least squares. The parameters of the model [14] Namrata Godbole, Manjunath Srinivasaiah and Steven Skiena. Large-
include: Scale Sentiment Analysis for News and Blogs. Proc. Int. Conf. Weblogs
and Social Media (ICWSM), 2007.
• A : rate of attention seeking
• P : polarity of sentiments and reviews
• D : distribution parameter

Let y denote the revenue to be predicted and  the error. The

linear regression model can be expressed as :
y = βa ∗ A + βp ∗ P + βd ∗ D +  (4)
where the β values correspond to the regression coefficients.
The attention parameter captures the buzz around the product
in social media. In this article, we showed how the rate of
tweets on Twitter can capture attention on movies accurately.
We found this coefficient to be the most significant in our
experiments. The polarity parameter relates to the opinions
and views that are disseminated in social media. We observed
that this gains importance after the movie has been released
and adds to the accuracy of the predictions. In the case of
movies, the distribution parameter is the number of theaters a
particular movie is released in. In the case of other products,
it can reflect their availability in the market.

IX. ACKNOWLEDGEMENT
This material is based upon work supported by the National
Science Foundation under Grant # 0937060 to the Computing
Research Association for the CIFellows Project.

R EFERENCES
[1] Jure Leskovec, Lada A. Adamic and Bernardo A. Huberman. The
dynamics of viral marketing. In Proceedings of the 7th ACM Conference
on Electronic Commerce, 2006.
[2] Bernardo A. Huberman, Daniel M. Romero, and Fang Wu. Social
networks that matter: Twitter under the microscope. First Monday, 14(1),
Jan 2009.
[3] B. Jansen, M. Zhang, K. Sobel, and A. Chowdury. Twitter power:
Tweets as electronic word of mouth. Journal of the American Society
for Information Science and Technology, 2009.