Analyzing the value of ‘reddit’ sentiment in predicting the success of a movie release, and the corresponding stock price movement of the risk bearing firm

(1)

Analyzing the Value of ‘Reddit’ Sentiment in Predicting the Success of a Movie

Release, and the Corresponding Stock Price Movement of the Risk Bearing Firm.

BSc Thesis Economics

University of Amsterdam

July 2015

Author: Max M. van Rossem

Student number: 10071059

Supervisor:

Ben A. Loerakker

(2)

Statement of Originality

This document is written by Student Max van Rossem who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the content.

(3)

Abstract

The objective of thesis is to determine whether the success of a movie can be predicted using Reddit sentiment.

It is posed that the successful release of a film is related to the stock price movement of the risk bearing firm. The ‘Reddit’ sentiment is proposed to reflect the overall awareness (volume) and polarization (‘valence’) of consumers about a film. We test whether sentiment holds predictive qualities by

constructing 5 variables from the sentiment scores: (i) Negative sentiment, (ii) positive sentiment, (iii) difference in negative and positive sentiment scores, (iv) a ratio of negative to positive sentiment and (v) the volume of sentiment activity. Sentiment data is additionally adjusted for (i) movie budget and (ii) budget to market capitalization ratio. Through these adaptations we determine whether sentiment has a statistically significant relationship that can be discerned in predicting the change in stock price (after the opening weekend).

The sample set was specified to 21 movies released in North America with minimal bias errors. Given the sample, tests showed no statistically significant relationship was discernable in predicting stock price movement using sentiment. However, in line with previous studies we have found that box office

performance is related to the volume of online consumer activity. Additionally we find that the sentiment valence (ratio of negative to positive comments) is possibly linked to the post-release public rating of movies.

(4)

1. Introduction

Online communication has come to play a critical role in the formation and expression of current views and opinions in society. Approximately 88% of the North-American population has internet access (World statistics, 2014). Societies are now able to share information or use data in a way that was not possible only decades ago. It is now possible to collect views and opinions from the internet to gauge market trends. Relevant relationships that can be drawn out from quantified sentiment is a subject on the frontier of economic analysis.

Field studies of sentiment analysis already performed by Baker and Wurgler showed capabilities in enhancing forecast models (2006). A popular source of ‘chatter’ is the application Twitter. Researchers compiled Twitter-chatter data and demonstrated its ability to reduce stock price forecast errors (Bollen, Mao and Zeng, 2010). The sentiment measurements used are based on the popular social news platform known as ‘Reddit’. Sentiment statistics are obtained with the proprietary analysis tool Keenalytics. We focus on the relevance of sentiment specifically within the movie industry. Economic relevance of the movie industry has been bolstered since its status of a multi-billion dollar status (Eliashberg, Elberse, & Leenders, 2006). The movie industry is a high risk environment (Liu, 2006, p.75). No previous Reddit sentiment related research has yet been conducted in relation to release success. Distribution companies stand to benefit from being able to predict the quantity of viewers a particular film will have to optimize its strategy (Liu, 2006).

It is not the purpose of this study to develop a full forecasting system in which the Reddit sentiment data is a complete predictor for share price. Rather we intend to illustrate how economic parameters such as box office results and share price might be correlated to actual word of mouth (hereafter referred to as WOM) information. We narrow the question down specifically to Reddit pre-launch sentiment and its predictive quality in analyzing the weekend release success of films.

Conducting a literature review we look at the theory behind stock market efficiencies and existing research on the effect of consumer reviews. We evaluate Reddit as a source of data, after which we review how the sentiment algorithm provided by Keenalytics functions. Following the literature review two general models are proposed and the constituent sentiment variable configurations are explained. The sample set is defined, the model regression results are discussed followed by conclusions that can be drawn.

(6)

2.Literature review

In the following section the relevant literature and economic background is provided. The significance of Reddit and its representative qualities are discussed. Subsequently we look at the significance of

consumer comments from a literary perspective to substantiate the model.

The increased use of the internet brings the consumer and producer in more frequent contact but also consumers amongst themselves (from which most of the sentiment data originates). The internet has resulted in resources being reallocated to their most productive uses (Freund and Weinhold, 2002). Freedom of information is theorized to bring the market closer to a perfect market (Litan and Rivlin, 2001). As markets become more transparent, true product values are realized. Stock price changes reflect disparity through the principles of information asymmetry and the efficient capital market (ECM) hypothesis (Fama, 1970). There exists uncertainty in the financial impact of a film, when true product value is realized, investor expectations align.

For the relationship between sentiment and product success there are many industries to consider. For the purpose of this paper the movie industry is chosen for a number of reasons:

 Films are widely commented upon by the general public providing a large amount of comment data.

 Films are experience products; the WOM phenomenon in these industries plays a crucial role (Liu, 2006).

 New films launch on a frequent basis. The product life span is short, the bulk of the revenue is generated within a few months (Joshi & Hanssens, 2009). As our data stretches over a limited time span, starting in August 2014, the frequent product releases increase available data.

 Publicly listed studios (e.g. Disney) can be financially impacted by the performance of a single film (Joshi & Hanssens, 2009)

 “The movie industry is known for its high risk” (Liu, 2006). High risk gives rise to volatility, films are heterogeneous goods. Being able to find insight in a high risk environment with heterogeneous goods is useful.

(7)

2.1 The effect of critics:

The theory of information asymmetry applies as newly released films do not possess unbiased product evaluation (Wenjing, Bin & Whinston, 2008). Consumers investigate a service or product before making a decision on whether to buy it. The WOM phenomenon provides reviews besides the likely, biased advertisements from its producers. Films are experience goods, researchers have found that such goods are influenced by critics. In a study by Basuroy, Chatterjee & Ravid (2003) a dual relationship was found for the effect of movie reviews on movie success. Their research established that critics could

simultaneously “influence and predict box office revenue”. Their research studied the effect of criticism delivered by professional critics whereas the Reddit sentiment data looks at the discussion of the American general public. Americans actively seek the advice of film critics. It is not claimed that pre-release Reddit chatter is a direct reflection of collective professional critics’ consensus, rather it is a sample measurement of the current attitude held by the consumer population. A study by Asur and Huberman (2010) has shown crowd chatter to be a reliable indicator of movie performance. Using twitter they found a “strong correlation between the amounts of attention a given topic has (a forthcoming

movie) and it’s ranking in the future”. We will determine whether this is also the case for Reddit.

2.2 The effect of marketing

The Efficient Capital Markets (ECM) hypothesis refers to the incorporation of all expected future cash flows into the stock price. Thus, in the case of motion pictures, a commercial hit should increase stock prices, and a flop should cause them to fall. Through ECM the stock price is assumed to incorporate all information released by the producing corporation such as script, director, release date, cast and budget. In a study by Joshi and Hanssens (2009) on the effects of Movie advertising (proportional to movie budget) and the stock market valuation, they found that pre-launch advertising played a large role in resulting stock returns. It was found that large advertising budgets pushed for a “profit expectations-laden

environment” inducing raised stock prices. In general 90% of advertising expenditure is utilized within

the weeks before the actual launch of a movie (Elberse and Anand, 2007). We will relate the advertising budget to the volume of comments that are found to exist for a particular film. Investors have received open access to large amounts of information regarding individual movies. “If markets are efficient, all this

information would have been incorporated into the studio’s pre-launch stock price, without bias” (Joshi

and Hanssens, 2009). They go further to say that any excess stock return immediately after the movie’s launch is related to the “actual movie performance relative to its pre-launch prediction”.

(8)

This is where the utility of Reddit sentiment is conceivably present, as an indicator to actual future performance. The Reddit sentiment can serve as a proxy for measuring the effect of advertising activities, star power, director popularity and all other such factors that determine the inherent consumer interest. The utility of Reddit sentiment is held within two aspects, namely its ‘valence’ and ‘volume’. As described in the research performed by Chintagunta, Gopinath and Venkataraman these two parameters within sentiment measure the total ‘public momentum’ when regarding a film’s expected performance (2010). Throughout our investigation we attempt to split valence and volume attributed effects of sentiment. Economic reasoning would support that a validation of a high quality (of a film) would result in a successful release. Since movie ticket prices are stable across time and for each film (Davis,

2005)(Digital Marketing Stats, 2015),we can expect high quality would result in more sales relative to lower quality movies at the same price. We create a variable for ‘sentiment valence’ to represent a proxy of quality. The second aspect of volume indicates relevance as a measurement of total consumer interest. Without public awareness of an available high quality good, fewer sales and thus reduced success can logically be expected.

2.3 Reddit as a source

Reddit.com, sometimes called ‘the front page of the internet’ is a platform through which users interact with one and other by posting content, commenting on content and voting on content worth. The site gets around 5.2 billion monthly page views. Reddit is amongst the top 10 most visited websites in the USA and is world-wide the 29th most visited website, around 60% of the users on Reddit are from the US (Alexa, 2015). The website works on the principle that interesting or agreeable posts are voted ‘up’ by its users. The more ‘up votes’ a post has the more visible it becomes for other users. The up-vote and down-vote functionality in Reddit can be compared to consumer interaction on preferences. The way that Reddit is constructed gives a form of 'weight' to the more relevant comments. This ‘weight’ measured by the number of ‘up’ votes relative to ‘down’ votes is used by the Keenalytics algorithm that constructs the sentiment score.

The site is constructed in such a way that users can specify their areas of interest (interests are divided into different subjects, known as 'sub-Reddits'). The result of this mechanism is that individual users receive a larger proportion of posts on their own personalized ‘front page’ according to their selected interests. There is wide array of more than 850,000 ‘sub-Reddits’ that correspond to areas of interest

(9)

subscribers (redditlist, 2015). The sentiment data is however collected from the whole site, not specifically the movie 'sub-Reddit'.

There are no barriers for people to join Reddit once they are online, only an email address is necessary. For viewing the posts that are made by users, an account is not needed. In table 1 an overview of the demographics of Reddit and the film industry have been collected.

There are a number of relevant demographics between the film going public and Reddit users that correspond (table 1). The high number of Reddit users allows for a broad user base, permitting for a fairer representation of the American population. A large part of the movie industry’s revenue comes from people between the ages of 18 and 39 years (43%), while 84% of Reddit users are between the ages of 18 and 34. The ethnicity distribution of North-American moviegoers and Reddit users is almost identical. Caucasians are the largest group, followed by Hispanics and black people respectively. In the income distribution broad similarities can be observed between the Reddit user and filmgoer demographics. We argue Reddit is a useful representation of the North-American film consumer population.

Reddit is a suitable platform from which to derive sentiment data for a number of practical reasons:

 Limited editorial influence.

 Voting by users facilitates quantification of popularity of both posts and comments.

 English language as the standard language.

(10)

3.Sentiment extraction

A meaningful relationship regarding the sentiment cannot be claimed without a degree of understanding of how the sentiment scores are determined. Hence this section outlines the approach taken to measure sentiment. There are a number of ways that programmers logically disseminate large quantities of text into useful statistics. Keenalytics relies on producing word vectors as well as machine learning (ML) to analyze sentiment.

3.1 Word vector representations

In order to analyze the texts posted by Reddit users, an implementation of GloVe (Global Vectors for Word Representation) is used (Pennington, Socher & Manning, 2014). This method provides an efficient application to convert words into vectors of real numbers that can be used in computation. These representations (i.e. vectors) can subsequently be used in natural language processing applications. The reason why this conversion is essential to the sentiment quantification is the following:

A piece of string text cannot be interpreted by a computer in the way that humans interpret words. For example, the word 'nice' - for the computer, are four characters 'n', 'i', 'c' and 'e' arranged in a specific order. There is no pre-programed meaning attached to this sequence of letters resulting in the word 'nice'. GloVe attaches representations to words so a computer can derive meaning from them. Specifically, GloVe assigns each word a unique vector of numbers using an unsupervised learning algorithm. The values of these numbers gives each word some conceptual meaning. For example the vectors attached to the word 'good' and 'nice' hold similar values.

The GloVe method has the power to detect conceptual similarities: "a is to b as c is to d". For example Amsterdam is to the Netherlands as Paris is to _____”. Using only word vectors; “Amsterdam to Netherlands, “Paris” would output the word vector “France”. Likewise "fly is to flying as dance is to _____” would output “dancing” using only the three initial vectors as input.

These GloVe vectors are used to conceptualize longer strings of text (i.e. comments on Reddit), which in turn hold some concept of sentiment about a certain topic. In order to judge the sentiment of these strings of text, Machine Learning is used.

(11)

3.2 Machine Learning

Machine learning (ML) algorithms are algorithms that learn from data. Word vector representations can construct sentence vector representations - i.e. vectors that represent some 'conceptual' meaning of a sentence. From Reddit, there are data sets of positive and negative reviews relating to variety of movie products from a variety of sub-Reddits. A review vector is constructed (similar to the word vector described above) so the data set then becomes a combination of review vectors and review scores (low or high). In order to further use this data, a ML algorithm is used to train the computer to recognize what goes into a positive review and a negative review about a certain subject. This “trained algorithm” has been applied to comments on Reddit.

For example: A comment on Reddit is as follows: "The movie X is bad, I was really disappointed". Using pre-constructed word vectors, a comment vector is constructed. This runs through the ML algorithm, it outputs this with a probability of being a positive sentiment comment. The ML algorithm is not only counting negative words. The algorithm is constructed through feeding a wealth of reviews and a pre-assigned associated quantitative score. The ML algorithm detects the patterns by itself related to positive or negative comments and as such operates independently. It is through ML that the algorithm constructed a process used in relating comments to a score. More information can be found at Keenalytics.com.

No program to date is able to perfectly interpret human comments. Keenalytics has limitations in its capability to interpret and quantify comments, sarcasm amongst other nuances in writing are difficult for the algorithm to take into account.

Sentiment data is obtained on a daily basis. Figure 1 provides a fragment of the data for one movie. The total data base covers a total of 31 movies. ‘Activity’ and ‘Num_com’ are experimental values that Keenalytics generated that have not been used.

(12)

The scores used throughout the analysis are Sent_pos and Sent_neg, which respectively represent the measurements of daily positive and negative sentiment estimated by the Keenalytics algorithm. These sentiment scores are later subdivided into two parts of interest, namely ‘valence’ and volume. As Pang and Lee (2008) describe; “work centered within economics and marketing sectors study whether the effect

of polarity (often referred to as “valence”) and/or volume of reviews have a measurable, significant influence on actual consumer purchasing”. In the next section we quantitatively define and model these

two aspects, valence and volume, in a manner that allows for their separate as well as combined strength in estimating the direction of stock price movement.

(13)

4.Defining the model and variables

To find out whether Reddit sentiment holds predictive capacities we construct a model that uses sentiment to predict a future movement in stock price. The dependent variable is the change in (adjusted) closing stock price of the concerned company. Adjusted closing stock prices have been adjusted for dividends and splits. The historical stock price data was retrieved from Yahoofinance.com.

We use OLS regressions to determine the validity of our models. There are a total of 34 observations (films) that were usable (table 4), however as will be explained the sample is specified down to a 21 film sample set to reduce sample errors. We estimate the relationship between created sentiment variables 𝑆𝑣 and the percentage change in stock price Y. We model a single variable model as well as a duel variable model shown by equations (1) and (2):

𝑌 = 𝑎 + 𝛽𝑆𝑣+ 𝜀 (1)

𝑌 = 𝑎 + 𝛽𝑆𝑣𝑜𝑙𝑢𝑚𝑒+ 𝜋𝑆𝑣𝑎𝑙𝑒𝑛𝑐𝑒+ 𝜀 (2) ε signifies the error term.

𝑌 measures change in the stock price shortly after a movie’s release. An important aspect of the analysis is which horizon to use in terms of the measured effect of the stock price reaction. Investor uncertainty of the performance of a film is greatest in the opening days. Information asymmetry and the misalignment of investor expectations of the performance of a film is largest at the point before release (Eliashberg, Elberse & Leenders, 2005). Investor’s re-evaluate expectations given the new information gained during the performance of the newly released film. Stock prices of distribution companies are affected by many sources after the release of a new film, as such the time span is kept close to the release day.

The first relevant stock price movement is the change in stock price (SP) on the day of the release. The second relevant SP, and the SP change that has been used for our regressions, is the second day after the release of a film. Note that day 2 is defined not as the second release day but rather the second day that the stock market was open of the release period for a film (usually a Monday). As films are usually released on a Friday, the change in SP after the release weekend would reflect a change in investor expectations due to information gained from the opening weekend. The opening weekend is a particularly important indicator for movies that subsequently become blockbusters (Sawhney & Eliashberg, 1996). News bulletins on film attendance and box office results during the opening weekend are frequently

(14)

posted on finance information feeds which change anticipated success. A example of this is the article by Mendelson (2014), titled: 'Focus' Nabs Mediocre $19M Weekend, 'Fifty Shades' Nears $500M. Such articles are widespread during the opening weekends of important movie releases. We hence assume stock prices on the second day that the stock market is open after the release of a film are the most relevant in reflecting the adjusted investor expectations.

The change in SP of day 2 is calculated using the closing SP the day of the release (day 1) as a reference point. This is the last point at which we assume investors did not yet have a clear indication on the financial impact due to the performance of a film. The percentage change in SP for day 2 is then calculated as follows:

Y = [𝐶𝑙𝑜𝑠𝑖𝑛𝑔 𝑆𝑃 𝑟𝑒𝑙𝑒𝑎𝑠𝑒 𝑑𝑎𝑦 2− 𝐶𝑙𝑜𝑠𝑖𝑛𝑔 𝑆𝑃 𝑑𝑎𝑦 1 ]

𝐶𝑙𝑜𝑠𝑖𝑛𝑔 𝑆𝑃 𝑑𝑎𝑦 1 ∗ 100%

The percentage SP change due to the performance of a film is independent for the release of each firm. There are however industry wide shocks which can put pressure on the stock price. Correcting for market movement the Dow-Jones United States Broadcasting (DJUSBC) index was integrated. The percentage stock price movement was adjusted for the movement of the index DJUSBC.

If the expectations of success are not aligned with the actual results of the individual films the stock price varies accordingly. The ECM hypothesis, assumes that stock prices completely and immediately incorporate all information that may affect the future cash flow of the company. Investors’ expectations can undervalue, overvalue or align itself with the true result. A commercial success should increase stock prices, and a box office failure should cause the SP to fall. This assumes the effect of unexpected commercial success is not masked by other factors that may affect the stock price.

Direct financial effect of a complete box office failure of a movie is the budget itself. In terms of stock price movement this effect could be insignificant for companies with a large market capitalization. However a box office failure might indicate to investors a company’s lacking in being able to produce winning movies. The change in stock price can result in a larger change in market capitalization than the total budget of a movie.

(15)

Reasoning for using pre-weekend sentiment is that buyer intent for the movie is being quantified. Trailers, advertisements, star visits to talk shows and journalism all provide input to the expectation and sentiment around a certain movie. While weekend sentiment may be more defined by the overall experience of filmgoers.

From the sentiment data we have extracted the following 5 sentiment variables (general notation: 𝑆𝑣):

1. Positive sentiment 𝑆𝑝𝑜𝑠

2. Negative sentiment 𝑆𝑛𝑒𝑔

3. Negative sentiment subtracted by positive sentiment 𝑆𝑁𝑠𝑢𝑏𝑃

4. Ratio of negative to positive sentiment 𝑆𝑁𝑑𝑖𝑣𝑃

5. Volume of mentions 𝑆𝑣𝑜𝑙

Model format (1) will be regressed using each of these variables. Model format (2) will be regressed using only 𝑆𝑁𝑑𝑖𝑣𝑃 and 𝑆𝑣𝑜𝑙.

The reasons for this is that variables 𝑆𝑝𝑜𝑠, 𝑆𝑛𝑒𝑔 and 𝑆𝑁𝑠𝑢𝑏𝑃 each hold both qualities of valence and volume. Variables 𝑆𝑁𝑑𝑖𝑣𝑃 and 𝑆𝑣𝑜𝑙 only hold one of these qualities, respectively valence and volume. Arguably we can hence expect less multicollinearity between 𝑆𝑁𝑑𝑖𝑣𝑃 and 𝑆𝑣𝑜𝑙.

The sentiment data has been used in these variables has been adjusted in 2 main ways:

1. Sentiment accounted for budget 𝑟𝑆𝑣

2. Sentiment adjusted for budget relative to market capitalization 𝑏𝑆𝑣

Each of these adjustments and the reason for applying them are discussed separately in this section. Throughout the following section references are made toward parameter definitions, a summation of which is found in table 7. As variable 𝑆𝑁𝑑𝑖𝑣𝑃 is a ratio of negative to positive sentiment, these adjustments have no effect.

(16)

4.1 Time span of aggregated sentiment

The total sentiment score is dependent on the number of days that data had been gathered before its release. The time span prior to release is different for all movies investigated. The ‘total sentiment’ accumulation ranges from 6 days to 223 days of sentiment data. This is not based on choice but is caused by the fact that the Keenalytics program is a recent development, the oldest data stems from August 2014 and continues until April 2015.

As mentioned the majority of marketing expenditures occur in the weeks leading up to the release. As posts on Reddit about specific movies occur somewhat sporadically sentiment values are preferable to be congregated over a longer period. The three week sentiment accumulation period is deemed most

appropriate as increasing the accumulation time reduces the sample size. This is due to the release date of films within the data set occurring close to the start of Keenalytics data gathering period. Using sentiment averages for differing available time spans will likely contain averaging errors making a fixed period for all films more appropriate.

4.2 Estimators for volume and valence of sentiment data

From the data it can be observed that the majority of accumulations of sentiment are greater for negative than for positive sentiment. A brief data summarization is provided in Table 3. For the 7822 days of sentiment data, 81.51 % sentiment values saw a higher negative than positive score (not including 0 sentiment days). For all films whether successful or not, this negative sentiment bias was present. There are two proposed causes.

Firstly it is known there exists a culture of inherent pessimism or skepticism of online commenters (Baumeister, Bratslavsky, Finkenauer & Vohs, 2001).This inherent pessimism, or probability skew toward negative comments puts bias toward a negative sentiment score. Baumeister et. al. additionally argue that seasonal effects may influence the general level of pessimistic comments (2001). Due to limitations of the data set seasonality affects are not modeled. It is assumed inherent pessimism does not change for the duration of the data gathering period.

The second and more likely cause is that the negative skew is due to the Keenalytics algorithm itself. The details as to how the algorithm accrues a negative bias is beyond the scope of this analysis. Instead the assumptions is made that the level of ‘inherent pessimism bias’ or ‘Keenalytics negativity bias’ is constant, making the error systematic in nature.

(17)

It clear that positive and negative sentiment scores are strongly correlated. Across used time span of 3 weeks, the correlation value between 𝑆𝑝𝑜𝑠 and 𝑆𝑛𝑒𝑔 is 0.976. The reason for this is that the sentiment scores incorporate the volume of mentions of the movie. So the positive and negative score for a given movie will be similar based on how often it is mentioned.

Regressing the three week cumulative positive sentiment for the corrected sample set (𝑆𝑝𝑜𝑠) against the three week cumulative negative sentiment for the sample set (𝑆𝑛𝑒𝑔) we get a coefficient of 1.29 (S.E.: 0.065) and an R square of 0.9801.

Variance Inflation Factor (VIF) can be used as an indicator of multi-collinearity (Stock & Watson, 2011). It can be used to measure the severity of multicollinearity in OLS regressions. Using the R-squared value from Stata output 2 we calculate it as follows:

𝑉𝐼𝐹 = 1

1−𝑅_{𝑆𝑝𝑆𝑛}2 = 1

1−0.98012 = 25.38

A Rule of thumb is that if the VIF > 10 there is likely some form of multicollinearity (Stock & Watson, 2011), we can conclude there is strong evidence of multi-collinearity. We can expect inflated standard errors if the effect of multi-collinearity is not addressed. On way to deal with multicollinearity is to construct a new variable to mitigate its effect. By creating a ratio of negative to positive sentiment we transform the multi-collinear variables in a logical way to mitigate its effect. Two variables are created. In one we subtract the positive sentiment value from the negative value (𝑆𝑁𝑠𝑢𝑏𝑃). In the other a ratio is made by dividing the negative sentiment by the positive sentiment (𝑆𝑁𝑑𝑖𝑣𝑃). Following is an example in

calculating the 𝑆𝑁𝑑𝑖𝑣𝑃 ratio for the film ‘50 shades of grey’ for a 3 week aggregated sentiment.

𝑆𝑁𝑑𝑖𝑣𝑃 = (𝑆𝑛𝑒𝑔)/(𝑆𝑝𝑜𝑠) = 40342

20472= 1.971

In some cases the sentiment of both negative and positive scores were zero, in this case the ratio has been given a value of one. The argument follows that as there was no discussion found on the topic to

determine a value of sentiment, it indicates there was neither a negative or positive preference. This corresponds to a ratio of 1. There were no cases in which there was a zero sentiment score for either positive or negative sentiment while the other had a non-zero sentiment score.

𝑆𝑁𝑑𝑖𝑣𝑃 is primarily a measurement of valence of sentiment. We can test whether our valence variable holds a foundation in reality by comparing it with real world data on movie quality. This would provide

(18)

We use post release public ranking scores to test for a relationship to support the quality of valence within sentiment. Two well-known sites for consumer and professional ranking are Rotten Tomatoes abbreviated as RT and the International Movie Database abbreviated IMBD. Both websites are in the top 200 most visited websites in the US (Alexa, 2015). These scores were retrieved on May 4th 2015 from their respective websites and constructed into a new average score. Regressing the average score ([RT score + IMDB score] / 2) against the ratio 𝑆𝑁𝑑𝑖𝑣𝑃 we obtain a significant negative coefficient -0.95 (S.E.: 0.37) (overview found in table 8).

We can conclude with a 95% confidence level that there is a negative relationship between 𝑆𝑁𝑑𝑖𝑣𝑃 and average consumer rankings. The coefficient between the RT_IMDB average ranking and the ratio of negative scores relative to positive scores is negative, as we would expect. The more negative comments there are relative to positive comments the lower the average the RT _IMDB score. This gives an indication of Keenalytic’s capacity in successfully incorporating valence into its sentiment score. An estimator of volume is created by adding the negative and positive sentiment:

𝑆𝑣𝑜𝑙= 𝑆𝑛𝑒𝑔+ 𝑆𝑝𝑜𝑠

This variable does not give any indication to the preferences but measures the amount of activity. Given the data set we see that volume is positively correlated to weekend box office revenue with a p-value below 0.01 (table 8). As we expect a higher level of activity is indicative of consumer awareness and hence movie box office revenue.

4.3 Difference in negative and positive sentiment

The parameter “𝑆𝑁𝑠𝑢𝑏𝑃” is created to measure the difference between negative and positive sentiment scores. Negative scores are predominantly greater than positive scores. This takes into account the volume of difference while the ‘𝑆_{𝑁𝑑𝑖𝑣𝑃}’ ratio measures a proportionary difference. Volume of communication is an indicator of consumer awareness, the difference in negative to positive activity creates a new

parameter with attributes of both qualities of sentiment valence and volume. Table 8 shows the 𝑆𝑁𝑠𝑢𝑏𝑃 holds a statistically significant (p<0.01) positive relationship with the weekend box office revenue. No significant relationship is found with post release review scores.

(19)

4.4 Sentiment accounted for budget

On average around 15 % of the movie budget goes toward marketing (Joshi and Hanssens, 2006). Hence movies with relatively larger budgets gain more attention of the public. In general larger budgets have resulted in larger box office results (Prag and Casavant, 1994). Regardless of what direction the sentiment is biased toward, the sentiment score (volume) increases with budget. We can confirm this by regressing budget (retrieved from IMDB.com) against negative sentiment. Preforming this regression using the corrected sample set, we see a highly significant positive coefficient for both negative sentiment

(coefficient = 934.66, Std. Err. = 299.45, p-value= 0.005) and positive sentiment (coefficient = 1182.17, Std. Err. = 403.06, p-value= 0.008). The full results of these regressions are given in Stata output 1 and 2. An indication of a successful launch is being able to effectively create ‘buzz’ or social media activity (Asur and Huberman, 2010). Substantial funds are spent on marketing to ensure successful opening weekends and ensure that the movie theatre owners keep the movie in theatres.A successful advertising campaign is characterized by high values (volume) in the sentiment data, representing the degree of public awareness. Controlling for the budget of the films gives an indicator of the success of the

marketing campaign. Put differently we measure the ‘activity per dollar’ of a movie’s budget. This creates a form of sentiment that is more effective in measuring ‘unexpected’ activity, possibly relatable to

unexpected movie success.

4.5 Sentiment accounted for budget to market capitalization ratio

Parent companies of the producing studios have different levels of market capitalization (see Table 2), resulting in films having varying potential to impact financial gain or loss. The movies which are used are produced and/or distributed by 7 major companies. A small overview of these companies and their differing market capitalizations is provided in table 2.

The market capitalizations (retrieved from wolframalpha.com) on the starting and end date of the data gathering period show two important things:

1. The total market capitalization differs substantially between comparing firms, ranging from $4.9 billion for Lionsgate to $146.7 billion for Comcast (April 3rd 2015).

2. The market capitalization changes over the data collection period for each firm. Notably Sony’s market capitalization increased by more than 69% during our data span.

(20)

As mentioned stock price fluctuations are problematic to attribute to one single cause. We aim to measure SP variations attributed to the performance of film investments. Although the general direction of a stock movement might be in line with movie success, the magnitude of a SP movement can be out of proportion (Einav & Ravid 2009). Market capitalization of a company can change by more than the total budget of a film. A reason for this is that investors may change their perception on future capabilities of a firm due to the performance on one of its ventures.

Taking the market capitalization the day before the release of the film of the relevant companies in combination with the budget of each individual film, a ratio is made. The ratio is an estimate as to what extent a company is financially at risk with regard to a particular film. The method used to create the new sentiment variables (r𝑆𝑣) is as followed:

𝑟𝑆𝑝𝑜𝑠 = (

𝐵𝑢𝑑𝑔𝑒𝑡 𝑜𝑓 𝑓𝑖𝑙𝑚

𝑀𝑎𝑟𝑘𝑒𝑡 𝑐𝑎𝑝𝑖𝑡𝑎𝑙𝑖𝑧𝑎𝑡𝑖𝑜𝑛 𝑜𝑓 𝑝𝑎𝑟𝑒𝑛𝑡 𝑐𝑜𝑚𝑝𝑎𝑛𝑦) ∗ 𝑆𝑝𝑜𝑠 (similar for 𝑟𝑆𝑛𝑒𝑔) 𝑟𝑆𝑁𝑠𝑢𝑏𝑃 = 𝑟𝑆𝑛𝑒𝑔 - 𝑟𝑆𝑝𝑜𝑠

𝑟𝑆𝑣𝑜𝑙 = 𝑟𝑆𝑛𝑒𝑔 + 𝑟𝑆𝑝𝑜𝑠

(21)

5. Sample set specification

In this section we outline the steps taken in determining the sample set. We attempt to minimize the extent of sample induced bias while simultaneously assuring we do not tamper with required randomness. All movies used for the analysis are first released in North America. Table 4 provides an overview of all movies part of the initial sample population. The data eliminations are given in the following section, there are 3 eliminations steps which have been categorized, these are:

1. Satisfying preconditions for analysis (step 1) 2. Selecting an appropriate ‘string’ (step 2) 3. Generic film title elimination (step 3) Table 7 summarized films removed at each step.

5.1 Step 1: Necessary preconditions of the sample set

In step 1 the sample population of films had to account for the following properties: 1. Initial movie release in North America.

2. Release took place within the time span August 17th 2014 to April 3rd 2015. 3. North American distributor is a publicly listed company.

The initial North American release assures that there is no preliminary grand scale public sentiment available. The North American movie market is significant with over $10.9 Billion in box office revenue in 2013 which is more than 30% of the global revenue (Motion Picture Association of America, 2013). Financial impact is arguably significant when a film fails in the North American movie market.

The release timing had to correspond to the sentiment data set available. Some films fell within the time range but were not included due to their release falling shortly after august 17 (for example ‘If I stay’ was released 22nd of august).

The film producers of the investigated films are also the main distributor. There are however a few cases in which the relevant company is solely the distributor (table 6). For example the film ‘Focus’ is distributed by Warner bros. (Time Warner) but it is not a producer of this film. There was no public

(22)

extent to which we can measure Warner Bros. exposure. Distributers are at risk when selecting films to distribute as there are opportunity costs as well as marketing costs which must be recuperated. The distributors are “dependent on the success of their potential blockbusters, which puts enormous pressure

on the industry, and forces studios to find creative ways to turn the release of a movie into an 'event'”

(Eliashberg et. al., 2005). Most companies we use in this analysis have stakes invested both as producer and distributer. Companies that are classified as being part of “the majors or the mini-majors” have high risks involved when investing in a film to distribute (Eliashberg et. al, 2005). Films for which one of the ‘major’ or ‘mini-major’ companies was the main North-American distributor were hence included.

There are a various factors which affect movie revenues, TV-advertisements, timing of release, quality factors such as star power, director, budget, type of film and at what studio the film was produced (Elliott & Simmons, 2011). Prag and Casavant (1994) further add parental guidance rating as another indicator. These factors are proposed to affect the success of individual films. However these properties stay constant and are public information. Assuming perfect capital markets, expectations are aligned with their state and would not induce variations in stock price during the opening. However there are reasons for eliminating certain films from the sample set as they generate bias.

5.2 Step 2: String of most significance

The ‘string’ refers to the title that the Keenalytics algorithm uses to search for related sentiment data. For example a film such as ‘The Hobbit: Battle of the Five Armies’ there are alternative strings which can be used to define this topic. If the string is defined as the whole movie title ‘The Hobbit: Battle of the Five

Armies’ there will be less sentiment data as few comments incorporate the whole title of the film when

referring to the movie. Instead the string can be defined as ‘The Hobbit’ in which case there is a larger chance of incorporating related comments as well as un-related comments. The assumption is that in the weeks approaching the release the comments will have a high probability of being related to the film. The corrected sample set is expressed by the following changes in ‘string’ definitions.

‘Hungergames: Mockingjay part 1’ changed to ‘Hungergames’

(23)

Continuing on the example of ‘The Hobbit’, for the string ‘The Hobbit: Battle of the Five Armies’ the negative sentiment score on the day of the release was 2, while using the sting ‘The Hobbit’ the negative sentiment score was 12,701.

5.3 Step 3: Eliminating generic titles

The Keenalytics algorithm cannot perfectly distinguish between film related comments and other comments using the same words. Some of the films were hence found to have an unusual level of activity. As an example the movie ‘Focus’ is considered. The word focus is used in contexts which are not related to sentiment regarding the actual film. Apart from cognitive reasoning for removing the most apparent movie titles that are susceptible to this issue we attempt to create a simple indicator.

Looking across the time span we can compare ‘activity segments’. Three segments were created: 1. The release period – 20 days before and after the release day

2. Preceding-Interest period – 80 days before release 3. Succeeding-Interest period – 80 days after release

The preceding and succeeding interest periods are periods in which it assumed that no or negligible sentiment data specified to the movie is generated as they are over 80 days away from the release day. Table 5 shows the average number of negative comments in Reddit on the search terms for the movies.

To quantify the overuse table 5 shows the ratio of expected minimal activity (preceding and succeeding interest) in comparison to the release period. Given the different release dates and the short data span some values that were unattained are left blank. Determining what films suffer from the generic string problem is not purely based on the ratio’s in table 5.

(24)

Primarily we use common sense to determine the following list of films suspected to have a generic title:

Focus Unbroken

The Interview Get hard

Exists Into the storm

If I stay The good lie

Addicted The Judge

Annabelle Insurgent

Some of the films suspected to have a generic title are confirmed by their activity ratio’s shown in table 5. The sentiment averages become smaller for less popular releases, the ratios become less relevant as they are affected by sporadic comments. For example ‘The Maze Runner’ is not suspected of a generic string even though its ratio of irrelevant activity is 13.5.

A negative sentiment moving average is illustrated in Figure 2, which show the 7 day moving average for the film Focus. It can be seen in figure 2that the film ‘Focus’ has a large degree of preceding release activity (shown in red) as compared to the release period (shown in green). Notably the trailer release has little impact on the overall activity pattern.

(25)

We will determine whether a variable is significant at a significance level of 95 %. Using model 1 & 2 we test the explained variables for relevance.

𝑌 = 𝑎 + 𝛽𝑆𝑣+ 𝜀 (1) 𝑌 = 𝑎 + 𝛽𝑆𝑣𝑜𝑙+ 𝜋𝑆𝑁𝑑𝑖𝑣𝑃+ 𝜀 (2)

Given the explanations the expected coefficients for each is summarized below:

Model format Variable (𝑆𝑣) Hypothesis: Coefficient

(1)

𝑆𝑝𝑜𝑠

Positive

(1)

𝑆𝑛𝑒𝑔

Negative

(1)

𝑆𝑁𝑠𝑢𝑏𝑃

Negative

(1)

𝑆𝑁𝑑𝑖𝑣𝑃

Negative

(1)

𝑆𝑣𝑜𝑙

Positive

(2)

𝑆𝑁𝑠𝑢𝑏𝑃

Negative

(2)

𝑆𝑣𝑜𝑙

Positive

(26)

6. Evaluation

As discussed there are two primary attributes to the sentiment data, namely volume and valance. The variables created from the sentiment data, express these two attributes in different manners. The variables 𝑆𝑝𝑜𝑠, 𝑆𝑛𝑒𝑔 and 𝑆𝑁𝑠𝑢𝑏𝑃 each represent a combination of valence and volume. 𝑆𝑁𝑑𝑖𝑣𝑃 is arguably purely a measurement of valence while 𝑆𝑣𝑜𝑙 solely represents the activity measured in the three week period approaching the release.

Each variable is inserted into model equation (1) and is regressed with the adaptations of unadjusted sentiment( 𝑆𝑣 ), sentiment adjusted for budget (𝑏𝑆𝑣 )and sentiment adjusted for market capitalization (𝑟𝑆𝑣 ). The variables 𝑆𝑣𝑜𝑙 and 𝑆𝑁𝑑𝑖𝑣𝑃 are regressed together in the format given by model (2). The final results of these regressions are presented in table 8 and 9.

Regressing model format (1), for each variable adjusted for market capitalization (𝑟𝑆𝑣) the significance level increases. For visibility we have included an indicator at p<0.10. Although this is not a level at which conclusions can be drawn it still poses a question as to why the significance increased. We propose that the market capitalization which is used to adjust the sentiment data is itself related to the stock price change. This could explain the increase in significance as the adjustment incorporates some degree of correlation. With the dependent variable the result of market capitalization regressed against the stock price change is found in Stata output 3. With a p-value of 0.083 we find a coefficient of 0.0072 (S.E. 0.0039). The relation is possibly strong enough to affect the result of the regressions. The increase in significance might then be attributed to the leverage effect. Smaller companies can be expected to be more volatile. Higher volatility stocks more often exhibit larger negative movements than equal positive movements (Christie, 1982)( Figlewski & Wang, 2000).

Continuing to look at model (1) we see that 𝑆𝑝𝑜𝑠 which should presumably be positively correlated with stock price movement, resulted in a negative coefficient. Counterintuitively the 𝑆𝑝𝑜𝑠 coefficient is notably more negative compared to the coefficient for 𝑆𝑛𝑒𝑔. The results for positive sentiment are counter to the expected coefficients. The correlation between positive and negative sentiment was found to be 0.976 with a VIF of 25.38 , this multi-collinearity is driven by the fact that the high volume of comments drives both positive and negative scores.

(27)

The negative sentiment coefficients are all negative as was hypothesized, however none of the coefficients are statistically relevant.

The coefficient is more negative for the 𝑆𝑁𝑠𝑢𝑏𝑃 model as would have been expected, however as both models for 𝑆𝑝𝑜𝑠 and 𝑆𝑛𝑒𝑔 did not result in relevant coefficients it follows that for the 𝑆𝑁𝑠𝑢𝑏𝑃 model we also find no statistically relevant results. So too is the case for the single variable 𝑆𝑣𝑜𝑙 model, with a negative coefficient that was not hypothesized.

It was supposed that the most relevant variable would be the 𝑆𝑁𝑑𝑖𝑣𝑃 as its relation to the final review score (the average of RT and IMDB) was found. The 𝑆𝑁𝑑𝑖𝑣𝑃variable was regressed against the Average_RT_IMBD_D4 (table 8) with a coefficient of -0.95 at p < 0.05. This implied the ability of 𝑆𝑁𝑑𝑖𝑣𝑃 to relate to quality it was supposed this would provide relevance when predicting release success. However as we can see from the final results the coefficient is not negative or statistically relevant. For Model 2 the final results in table 9 show no statistically relevant relationships. It was expected the independent variable 𝑆𝑣𝑜𝑙 would be have a positive coefficient while 𝑆𝑁𝑑𝑖𝑣𝑃 would have a negative coefficient. The model corresponds to both the attributed qualities of sentiment, volume and valance. With a high volume the inherent interest of the public would allegedly be at a high level which would result in in widespread awareness. Combined with the parameter for quality of 𝑆𝑁𝑑𝑖𝑣𝑃 the consumer factor for product success was proposed to be modeled.

To gain some insight in whether the regression model is well specified we provide two residual plots. The

(28)

There is a relatively random distribution across the residual plot 1. More of the data points are situated on the lower end of the sentiment scale. This implies an asymmetric representation of data over the

regression coefficient.

Residual plot 2 might indicate some form of heteroscedasticity because of the seemingly higher

variability at low end of NdivP ratio scores than at the high end. Using the same model for residual plot 2 we can test for heteroscedasticity using the Breusch-Pagan test (Stock and Watson, 2011). We cannot conclude there is heteroscedasticity ( Prob >chi2 = 0.286).

The models that were proposed were built on the following steps:

Step I: Excessive positive (negative) sentiment translates into higher (lower) than expected sales. Step II: Higher (lower) than expected sales result in stock price appreciation (depreciation).

For causes as to the incapability of extracting a significant relationship, we look at the assumptions that had to be made in constructing the model. The reality of some of these assumptions may be quite different to the theory.

 Efficient Capital Markets theorem holds in the movie industry stock market.

Each film has a different process through which it finally reaches the market. Films are heterogeneous goods which are difficult to analyze. Factors such as star power and director popularity may not hold a unanimous meaning for investors. With heterogeneous expectations it is difficult to find patterns in stock price movement.

 ‘Skepticism’ and ‘Keenalytics measurement bias’ are systematic errors.

The negative sentiment score is higher for the majority of pre-release data points, contrary to Liu (2006) who found that the “average percentage of positive word of mouth is greater than that of negative WOM

in the early time periods, especially during the pre-release”. There is little insight that can be given into

the actual nature of the error. As described the word-to-vec process within the algorithm uses lists of negative and positive words to class sentiment. A film title can be related to discussions that could polarize sentiment in a direction irrespective of the perceived film quality or interest.

(29)

 Companies advertise in proportion to the budget of the film.

The exact data on marketing expenditure for each film is not publically available. Adjusting the sentiment for budget may hence have arbitrary effects. Advertising spending has a stronger impact on the success of blockbusters than low budget films (Holbrook and Addis, 2007).

 The performance of a film affects the perceived future profitability of the relevant firm.

This assumption was made to give each film relevance in the determination of the stock price during its release. However the reality may be that some films do little to affect stock price, and are hence disruptive within the sample set. Low budget films released by large firms such as Comcast may have little to no effect on the firm’s stock price..

 The distributor owns the risk of a film’s financial performance.

Although the firms mentioned are the primary distributors of the films, there is very little information available on the exact deals that the distributor has with its co-producers in terms of profit sharing or revenue sharing. This makes it difficult to determine to what extent a firm is liable to a films box office performance.

(30)

7. Conclusion

There are many variables that play a role in determining the price of stock. The release of a movie, although sometimes an influence, may not always affect the stock price. Given our data set and

significance level, the models did not prove to be statistically significant when attempting to explain the movement of stock prices after the opening weekend

There were some inherent qualities of the movie releases that the sentiment data could predict. Similar to Liu (2006) we found that the volume of word of mouth holds explanatory power in predicting box office revenue. As we see in table 8, variables that had the sentiment ‘trait’ of volume incorporated into them (𝑆𝑝𝑜𝑠, 𝑆𝑛𝑒𝑔, 𝑆𝑁𝑠𝑢𝑏𝑃 and 𝑆𝑣𝑜𝑙), showed to be significant in determining box office revenue. All reached a level of significance at p<0.01.

The other quality of sentiment, valence, was interpreted by 𝑆𝑁𝑑𝑖𝑣𝑃 to a certain extent. Using the RT and IMDB user ranked scores as a measure for quality, 𝑆𝑁𝑑𝑖𝑣𝑃 was significant to p < 0.05. We are tentative however in proposing predictive power. As seen in residual plot 1 most of the data is situated on the lower end of the sentiment scale. When discussing the relevance of activity ratios in table 5, we saw that ratios created with small amounts of sentiment data become more unreliable. As for any empirical relationship developed in the region in which the observed data falls, the relationship found might break down were it extrapolated to a region in which less data has been observed.

In summation we found:

1. There is no clear cut correlation between the release weekend stock price change and the sentiment data of the movies investigated.

2. The box office revenue of a movie is likely correlated with volume of comments related to it.

3. The Keenalytics algorithm in combination with Reddit extracted comments, shows a degree of capability in predicting future movie ratings, however we are tentative to claim forecasting capability

(31)

Stock price movement in relation to movie release success was difficult to be fully realized, it would be interesting to carry out the same test for only for the less capital intensive distributors. Movie releases would then possibly be more relatable to the stock price movements as the risk involved with each film is higher.

Instead of a continuous stock price change, it would be interesting to create a dummy variable to classify success within stock price movements purely as an increase, or decrease in stock price. Furthermore in determining the predictive power of the model the closing day stock price is used, a more specific time may have been more effective. Weekend results may have been more notable within the first hours of opening the stock exchange.

The sentiment data itself could also be adjusted for a type of index of general activity levels. Similarly to how stock price change was corrected for industry wide shocks, sentiment volume could be corrected for website wide (regardless of topic) shocks or slumps in activity.

(32)

8. Data outputs and tables

Stata output 1

Regressing budget against negative sentiment

Stata output 2

Regressing budget against positive sentiment

Stata output 3

(33)

Table 1

Demographic summary

Description moviegoer statistics

Theatrical Market Statistics of North

America (in %) Description Reddit statistics Reddit statistics ( %)

Gender M/F 52 / 48 Gender M/F 78 / 21

Age Frequent Moviegoers 18-24 20 Age of Reddit users 18-24 49 Age Frequent Moviegoers 25-39 23 Age of Reddit users 25-34 35 Age Frequent Moviegoers 40-49 9 Age of Reddit users 35-44 7 Age Frequent Moviegoers 50-59 11 Age of Reddit users 45-54 2 Age Frequent Moviegoers 60+ 10 Age of Reddit users 55+ 1

Caucasian 60 Caucasian Reddit users 69

Hispanic 20 Hispanic Reddit users 22

Black 11

Black and other Reddit

users 9

Other 9 N/A N/A

Income <$30,000 24 Income <$30,000 24 Income $30,000 –$49,999 20 Income $30,000 – $49,999 24 Income $50,000 – $74,999 21 Income $50,000 – $74,999 28 Income >$75,000 35 Income >$75,000 24 1. Reddit Demographics, (2014) 2. Nielsen Cooperation, (2008)

(34)

Table 2

Corporations within the data set (Market capitalizations, films released and budget range)

Company Market Cap.

Aug 17, 2014 (Billion $) Market Cap. Apr 3, 2015 (Billion $) Number of films released during data period Range of film budgets smallest to largest (million $) FOX 78.54 70.56 4 34 to 127 Time Warner 65.39 70.21 14 6.5 to 250 Comcast (Universal Pictures) 142.40 146.70 3 4 to 65 Sony 19.21 32.55 2 44 to 49 Lions Gate 4.49 4.90 4 5 to 125 Disney 153.30 180.20 4 28 to 165 Viacom 34.47 26.84 1 165 Wolframalpha.com Table 3

Descriptive statistics for study variables

Variable Mean Maximum Minimum Standard Deviation Budget ($ million) 69 250 4 63 Market Capitalization ($ billion) 78 181 4.5 46 Negative sentiment 4506 323,059 0 18,669 Positive sentiment 3113 219,334 0 13,284 Weekend Box Office

revenue ($ million)

(35)

Table 4

Films that fulfilled the initial preconditions (sample set: step 1) TITLE

50 Shades of Grey Interstellar

Addicted Into the Woods

Alexander and the Terrible Into the storm American Sniper Jessabelle

Annabelle Jupiter Ascending

Big Hero 6 Kingsman: The Secret Service Chappie Hungergames: Mockingjay part 1 Cinderella Night at the Museum

Dolphin tale 2 Run All Night

Exists Taken 3

Focus The Boy Next Door

Get Hard The Hobbit: Battle of the Five Armies Horrible Bosses 2 The Interview

If I stay The Judge

Inherent Vice The Maze Runner

Insurgent The good lie

Unbroken This is where I leave you

Table 5

Generic strings: ratios of relevant to irrelevant activity Film Preceding interest average Release period activity average Succeeding interest average Ratio before:during Ratio After:during Focus 67,113 102,212 0.66 The interview 5,284 26,511 12,278 0.20 0.46 Exists 20,903 29,583 1.42 The Judge 10,323 13,047 1.26 The Hobbit 2,853 9,055 3,225 0.32 0.36 Addicted 6,815 8,534 1.25 Interstellar 5,054 3,010 0.60 American Sniper 177 4,462 0.04 Hunger Games 401 3,008 1,120 0.13 0.37

(36)

Film Preceding interest average Release period activity average Succeeding interest average Ratio before:during Ratio After:during Unbroken 724 2,785 1,422 0.26 0.51

Into the woods 1,341 2,735 889 0.49 0.33

Cinderella 403 2,712 0.15 50 shades of grey 459 2,475 0.19 Insurgent 2,336 1,814 1.29 Jupiter Ascending 1 1,089 0.00 Chappie 111 696 0.16 Taken 3 98 537 16 0.18 0.03 Get Hard 2,001 453 4.42 Mockingjay 20 398 139 0.05 0.35 Kingsman 2 370 0.01 Inherent vice 11 169 132 0.07 0.78 If I stay 147 513 3.50 Big hero 6 144 315 2.19 Horrible Bosses 2 6 54 11 0.11 0.20

The boy next door 5 37 0.13

Night at the Museum 15 29 4 0.52 0.14

This is where I leave you 23 7 0.31

Kingsman: The Secret

Service 0 22 0.00

Mockingjay part 1 2 19 29 0.09 1.51

Annabelle 10 80 7.73

Into the storm 10 138 13.63

Alexander and the terrible 8 37 4.68

Run all Night 6 4 1.50

Hobbit: Battle of 5 Armies 0 4 1 0.00 0.25

The Maze Runner 2 27 13.50

Dolphin tale 2 0 2

Jessabelle 0 0

The Good Lie 0 0 0

(37)

Table 6

Company information: Distributor & Producer

TITLE Distributor Producer

50 Shades of Grey CMCSA

Addicted Lionsgate Lionsgate

Alexander and the Terrible Walt Disney Walt Disney

American Sniper Warner Bros Warner Bros

Annabelle Warner Bros New Line Cinema

Big Hero 6 Walt Disney Walt Disney

Chappie Sony Columbia Pictures

Cinderella Walt Disney Walt Disney

Dolphin tale 2 Warner Bros

Focus Warner Bros

Get Hard Warner Bros Warner Bros

Horrible Bosses 2 Warner Bros

Inherent Vice Warner Bros Warner Bros

Insurgent Lionsgate

Interstellar VIA Warner Bros & VIA

Into the Woods Walt Disney Walt Disney

Jessabelle Lionsgate Lionsgate

Jupiter Ascending Warner Bros Warner Bros

kingsman 20 Century Fox 20th Century Fox

hunger games Lionsgate Lionsgate

Night at the Museum 20 Century Fox 20th Century Fox

Run All Night Warner Bros

Taken 3 20th Century Fox

The Boy Next Door CMCSA UPI

The Hobbit Warner Bros MGM

The Interview Columbia Pictures Columbia Pictures

The Judge Warner Bros Warner Bros

The Maze Runner 20 Century Fox 20th Century Fox

This is where I leave you Warner Bros Warner Bros

(38)

Table 7

Descriptor reference table (Descriptor variable format: z𝑽𝒙)

z Adjustment to sentiment

𝑉𝑥 Unadjusted sentiment

b 𝑉𝑥 Sentiment adjusted for budget: cumulative sentiment / budget r 𝑉𝑥 Sentiment adjusted for market capitalization ratio:

Cumulative sentiment * [budget / company market cap]

𝒙 Sentiment variable

Pos Cumulative positive sentiment Neg Cumulative negative sentiment

NsubP Cumulative negative sentiment – cumulative positive sentiment NdivP Cumulative negative sentiment / cumulative positive sentiment Vol Cumulative negative sentiment + cumulative positive sentiment description Other relevant data

Boxrev_ Box office revenue

revDivscr_ Box office revenue divided by number of opening screens

RT_IMDB Average score of user ratings: (Rotten tomatoes score + IMDB score) / 2 MarketCap Market Capitalization of relevant firm at the time of release

Step Removal of films from sample set

1

Films without clear publically listed owner removed or data span less than 1 week:

Exists

Into the storm The good lie If I stay Jessabelle

2

String of most significance included, removed:

‘Hungergames: Mockingjay part 1 changed to ‘Hungergames’ ‘Kingsman: The Secret Service’ changed to ‘Kingsman’ The Hobbit: Battle of the Five Armies’ changed to ‘The Hobbit’

3

Film titles susceptible to overuse removed:

If I stay Addicted Focus The Interview The Judge Unbroken Get hard

(39)

Table 8

Variable strength in predicting review scores and box office results

Independent sentiment variable 𝑆𝑣 Dependent variable: RT_IMDB . . Coefficient (Standard Error) Dependent variable: Weekend Box office

. . Coefficient (Standard Error) 𝑆𝑝𝑜𝑠 .136* (.068) 343.15a (117.21) 𝑆𝑛𝑒𝑔 .083* (.053) 287.90a 84.45 𝑆𝑁𝑠𝑢𝑏𝑃 .052* (.185) 966.79a (272.36) 𝑆𝑁𝑑𝑖𝑣𝑃 -0.950b (.374) -3,686,077 (8,158,783) 𝑆𝑣𝑜𝑙 .053c* (.030) 158.74a (49.29) a p<0.01, bp <0.05, cP<0.10 , scaling adjustment* ( 𝑆𝑣 / 10,000)

Box office revenue data: boxofficemojo.com IMDB data: imdb.com

(40)

Table 9

Results model regressions Model 1: 𝑌 = 𝑎 + 𝛽𝑆𝑣+ 𝜀

Unadjusted Adjusted for budget

Adjusted for market capitalization Independent variable (𝑆𝑣) Coefficient (Standard Error) Coefficient (Standard Error) Coefficient (Standard Error) 𝑆𝑝𝑜𝑠 -0.029* (0.067) -8.816* (5.706) -10.287* c (5.688) 𝑆𝑛𝑒𝑔 -0.031* (-.144) -6.483* (4.033) -6.942* c (3.768) 𝑆𝑁𝑠𝑢𝑏𝑃 -0.152* (0.164) -10.688* (35.199) -276.744* c (114.776) 𝑆𝑁𝑑𝑖𝑣𝑃 0.007 (0.391) - - 𝑆𝑣𝑜𝑙 -0.016* (0.029) -3.767* (2.371) -4.147* c (2.267) Model 2: 𝑌 = 𝑎 + 𝛽𝑆𝑣𝑜𝑙+ 𝜋𝑆𝑁𝑑𝑖𝑣𝑃+ 𝜀

clean Adjusted for budget

Adjusted for market capitalization 𝑆𝑉𝑜𝑙 -0.016* (.030) -3.767* ( 2.43629) 4.889* (18.34739 ) 𝑆𝑁𝑑𝑖𝑣𝑃 -0.008 (0.399) 0.002 (0.377) -87.702 (176.641) a p<0.01, bp <0.05, cP<0.10 , scaling adjustment* ( 𝑆𝑣 / 10,000)

(41)

9. Bibliography

Alexa Internet (2015). How popular is reddit.com? Retrieved from: http://www.alexa.com

Asur, S. and Huberman, B. A. (2010). Predicting the Future with Social Media. WI-IATW

(1) 492 – 499.

Baker, M. and Wurgler, J. (2006). Investor Sentiment and the Cross-Section of Stock Returns. The Journal of Finance 61 (4), 1645-1680.

Baumeister, R. F., Bratslavsky, E., Finkenauer, C. and Vohs, K. D. (2001). Bad is stronger than good.

Review of general Psychology, 5 (4), 323-370

Basuroy, S., Chatterjee, S. and Ravid, S.A. (2003). How Critical Are Critical Reviews?

The Box Office Effects of Film Critics, Star Power, and Budgets. Journal of Marketing, 67, 103– 117.

Bollen, J., Mao, H. and Zeng, X. (2010). Twitter mood predicts the stock market. Journal of Computational Science 2 (1), 1-8

Box Office Mojo (n.d.). Retrieved from: http://www.boxofficemojo.com

Chintagunta, P. K., Gopinath, S. and Venkataraman, S. (2010). The effects of online user reviews on movie box-office performance: Accounting for sequential rollout and aggregation across local markets. Marketing science, September 2010.

Christie, A.A. (1982). The stochastic Behavior of Common Stock Variances. Value, Leverage and Interest Rate Effects. Journal of Financial Economics 10, 410

Davis, P. (2005). The Effect of Local Competition on Admission Prices in the U.S. Motion Picture Exhibition Market. Journal of Law and Economics 48 (2), 677-707

Digital Marketing Stats (2015). By the Numbers: 50+ Amazing Reddit Statistics. Retrieved from: http://expandedramblings.com/index.php/reddit-stats/

Einav, L. and Ravid, S. A. (2009). Stock market response to changes in movies’ opening dates

Journal of Cultural Economics. 33, 311–319.

Elberse, A. and Anand, B. (2007). The effectiveness of the pre-release advertising for the motion pictures: An empirical investigation using a simulated market. Information Economics and Policy 19 (3-4), 319 – 343.

Analyzing the value of ‘reddit’ sentiment in predicting the success of a movie release, and the corresponding stock price movement of the risk bearing firm