Movie Analytics and the Future of Film Finance. Are Oscars and Box Office Revenue Predictable?

Abstract

This chapter discusses to which extent modern analytics techniques can help us understand the success of movies, as measured by their box office or attributed Oscars. Interesting lessons emerge from our analyses. Predicting box office revenue on the basis of data available before the release of the movie remains difficult, even with state-of-the-art techniques. Prediction markets are a remarkably powerful tool at predicting success at Oscars. A moderate amount of controversy, as measured by the number of underlying themes raised by movie reviewers, may prove to be helpful in obtaining an Academy Award for Best Picture .

1 Introduction: Movie Analytics and Film Finance

This chapter examines to which extent modern analytics techniques help us understand the success of movies. We will describe essential analytics techniques as needed here and discuss them in the context of the prediction of box office revenue and Oscar attribution. The work in this chapter lies in the context of the broader issue of film financing . A series of papers published by slated.com (Brown, 2015a, 2015b) makes the case that the industry has now become an attractive investment domain, and the success of the industry in terms of both raw revenue and revenue growth implies that investors must have at least some sense of how to control the risk of investments in that industry. It, however, remains true that predicting the success of a film, even with modern data mining techniques at hand, is a difficult task, as will be detailed further in the chapter (see for example El Assady et al. 2013).

This chapter addresses in detail the problem of predicting box office revenue on the basis of data available before the movie is released (Sect. 2). The methodology for data collection and analysis based on state-of-the-art data mining models is described. In all, this discussion draws a sobering lesson: state-of-the-art methods can identify those variables which are important for predicting box office revenue, such as the budget, whether the movie is part of a series, and the “star power ” of the distributor, actors, and producer, but it remains difficult to actually predict box office revenue with decent accuracy because of the presence of very strong outliers in the dataset. The chapter then turns to a discussion of the role of “prediction markets ,” that is, exchange-traded markets created for the purpose of trading the outcome of specific events, in predicting Oscar wins (Sect. 3). The performance of the Intrade market (now taken off-line for reasons explained below) in predicting the 2013 Best Picture Award (attributed to the movie Argo) is described, on the basis of data extracted from Intrade before it was taken off-line. Finally, the chapter discusses the role of “controversy,” as identified by a text mining of movie reviews, in Oscar attribution (Sect. 4). The analysis suggests that too many themes underlying the reviews may be too complex for a voting audience to rally on. On the other hand, too few may be too simple. The movie Argo, with six underlying themes, may very well have reached a happy medium. Hope may very well lie, both in terms of predicting box office revenue and awards such as Oscars, in preproduction analyses of scripts. However, such analyses, as performed, for example, by slated.com, involve human judgment. Parts of this chapter are based on Haughton et al. (2015).

2 Box Office Revenue Prediction

Strong financial stakes are linked to a good box office prediction for a film, both for producers of this film and for cinema managers. The literature in this area is large, but datasets used are often unintentionally biased by their authors who exclude films with a low budget or low success (observed ex post) because information about such films is less reliable.

Managers of cinemas are investors, and as such, minimizing the risk of commercial failure of the films they select is a major issue. To achieve this, several methods are at their disposal. They may, for example, broadcast a film with an actor or a director who is recognized and appreciated by the public, a film adapted from a popular book, or a sequel to a successful film; or they might count on an effective advertising campaign. However, these simple strategies do not guarantee the commercial success of a film. To address this problem, predictive models have been developed for several years to identify factors that help make a movie a success.

Extensive literature exists on the subject with early attempts dating back to the 1990s and new developments each year. In general, models display relatively good predictive rates (with R-squares of more than 70% and relatively low error rates).

Most of these studies, however, have a selection bias inherent in the availability and quality of the information used to construct the models: they base their prediction only on a subsample of films with either a high production budget or, worse as far as bias is concerned, a relatively high box office revenue. The problem is that such information is not available “ex ante” for a cinema wishing to decide whether or not to accept a given film. Typical samples in the literature eliminate a considerable number of “outliers” (up to 46% of the films with total box office revenues under US$1 million). We also note that this 46% is probably an underestimate of the proportion of film with revenues below US$1 million since information related to certain failures was probably not included on the site www.boxofficemojo.com used in this chapter. This explains in large part the good “predictions” in the aforementioned studies (see, e.g., McKenzie (2012) for a review article). These models are however not practical for a film director wishing to identify a potentially successful movie. Determinants of success among films with a relatively high box office revenue are not necessarily the same as those for films “ignored” because of their low visibility.

In this section, we explore to what extent it is possible to predict the financial success of a film before its release, with the goal of potentially helping a cinema manager decide which films to accept to maximize profit. We consider all categories of films such as those that can be found on the US site www.boxofficemojo.com, a website that tracks box office revenue in a systematic, algorithmic way. We will focus on box office revenue for the first week after a film’s release, since most profits are realized during this first week. We compare the results of estimates using different statistical methods, more or less recent and more or less complex. Our main finding is that a precise prediction of box office receipts on the basis of prerelease information only is much more difficult than when the selection of the sample is biased by the use of “ex post” information. We will therefore try to predict box office revenue from the available “ex ante” information for the manager whatever the film. More precisely, we will only use information available before the shooting of the movie begins: a good prediction with this information only would allow the manager to position himself/herself earlier on promising films. We will test several forecasting models.

2.1 Methodology

Our forecasts will focus on the box office revenue obtained during the first week after the release of the movie. This is highly relevant to cinema managers since, as illustrated in Fig. 1, the bulk of the revenue from a film is obtained during the first week (40% on average), reflecting the delay in other cinemas deciding to show the movie.

Fig. 1

Average weekly box office performance and average number of movie theaters showing a movie. Source: Computed by the authors

Concerning our predictions, we want to put ourselves in a “real situation,” faced by a realistic cinema manager, with the information available to him/her. We will therefore predict the revenues of the films released after the end of 2011 (2166 films), from models built on data available “ex ante”: on the basis of films released before January 2012. More specifically, we consider all those films whose first week of broadcasting was over by January 1, 2012. Different time series learning samples will be tested to construct our models:

• The 5874 films released between 2000 and 2012

• The 3392 films released between 2006 and 2012

• The 2193 films released between 2008 and 2012

• The 1090 films released between 2010 and 2012

• The 566 films released in 2012

This temporal distinction offers us the possibility of testing what is most appropriate for predicting box office revenues and of guiding the choice of the most suitable reference sample. This helps control for an unobserved economic conjuncture. We will refer to this process as the choice of an “optimal temporal horizon.” We will construct the following predictive models:

• Linear regression

• Decision tree

• Random forest

• Conditional forest

• Gradient boosting

We will also use the so-called stacking method to optimize our results. These different models will finally be applied to the ex ante data on the films released after January 2012, estimated as if we were at the end of 2012. Therefore, it is clear that, in contrast to previous studies, we will not exclude any film. We, finally, compare our predictions with the actual revenues realized by films released in 2013, 2014, and 2015.

In summary, our objective is to identify the optimal time span a manager must observe as a test period to construct his/her estimation model and the modeling technique with the highest predictive power.

2.2 Data

2.2.1 Data Collection: Web-Scraping

The database used in this chapter contains more or less detailed information (total, weekly or even daily) on the box office revenue of 15,459 films in the United States. It also contains “classical” details about the genre of the film, its production budget, distributor, cast members, etc. This information was gathered via web-scraping on the American reference site: www.boxofficemojo.com.

The data collection method (fully implemented via R) proceeded as follows:

• Collection of all the links to the pages referencing the films

• Creation of a “scraping function” to apply to these links

This function must be general enough to be able to retrieve the information available on any type of page, which implies extensive investigative work on the site (discovery of patterns in the coding of pages) and management/anticipation of potential errors. Among other things, the function must be able to determine the granularity of the available information (daily, weekly, etc.). To speed up data collection, and to avoid overloading the server of the query site, we aim for the function to visit the smallest number of links possible. This function returns the results of each page as a frame, with one row for each “period” of box office revenue (e.g., 1 week). Finally, this function is applied to all links gathered in the first stage. Once data collection was completed, a dataset of 107,760 weekly box office revenues was obtained for 11,544 different films. We recall that we focus on first-week revenues in this study.

2.3 Data Processing: Creation of Variables

We now possess a very rich database, but several data transformations were necessary in order to be able to make use of it.

2.3.1 Box Office Revenue Deflation

We choose to measure box office revenues in financial terms rather than in terms of audience (number of viewers) since we are focusing on the financial return of the films. To make the intertemporal comparison possible, we deflated box office revenue by the monthly CPI of the weeks in which they were launched. When that information was not available, we simply deflated the global box office revenue by the CPI of their launch year. Table 1 displays the ten highest revenue films, in both nominal and real terms.

Table 1 Ten films with the highest box office revenue: nominal and actual (basis, January 2010)

2.3.2 Production Budget

The film production budget is only available for a little less than a quarter of the movies. To retain this 25% of available information, we will split the production budget into two “subvariables”:

• A first binary variable indicating whether or not the information is available.

• A second variable of interaction between the binary availability variable and the production budget. This variable is therefore zero when the information is not available and is equal to the production budget otherwise.

This “dichotomy” is the best way to maintain this variable which is positively correlated with the box office, hereby avoiding a non-negligible bias.

2.3.3 Experience, Quality, and Star Power

The decision to “consume” a film is special because the good is consumed once (with a few exceptions, since a few individuals will see the same film multiple times in a cinema). The individual therefore bases his/her choice of viewing on an implicit estimate of the quality of the film. Watching trailers, reading opinions by experts, or observing the enthusiasm of other better-informed individuals (measured on social networks, e.g., see Mestyán, Yasseri, & Kertész, 2013) can influence this estimate. The observation of these different variables makes it possible to estimate in a very fine way the box office revenue and even its dynamic evolution over time. Without any variables related to the viewing of the film (“after-launch” variables), information on casting (actors, producers, directors) are our only means of obtaining an indication of quality, for three main reasons:

• 1. The reputation of the actors, their “star power,” attracts consumers because it sends, a priori, a positive signal on the quality of the film: an actor who has acted “well” in good films in the past will have a good reputation and will attract more people than less well-known actors, all things being equal.

• 2. The experience of the director is another indicator of ex ante quality for spectators: the more experienced a director is, the more he/she is likely to attract viewers.

• 3. The “quality” (judged by specialists) of past performances of actors suggests that quality will be present in the new film.

For example, an Oscar can be perceived as a guarantee of quality for an actor. It is therefore necessary to create a variable summarizing the star power and the experience of the actors, etc. involved in a film. We use the following approach: for a given film, the star power of each actor in the film is estimated by the sum of the box office revenues of the films in which he/she played a major role in the past. The total star power of the actors of the film is then obtained by summing these individual star powers. The star power of the directors, the producers, and the distributors are computed in the same way.

For a given film, the experience of each actor is simply estimated by the number of films in which he/she played a major role in the past. We obtain the total experience of the actors of a film by summing up the individual experiences. The same applies to directors, producers, and studio distributors. For a given film, the recognized reward-based quality of an actor or a director is estimated by the number of nominations he/she had for an Oscar in the past. For a producer or distributor, this quality is captured by the number of films in which he/she has intervened and that has been nominated for a Best Film Oscar. The total reward-based quality of a given new film is simply estimated by the sum of the individual reward-based qualities of the actors, distributors, producers, and directors.¹

2.4 Descriptive Statistics

2.4.1 Box Office Revenue

As illustrated in Fig. 2, the distribution of the logarithm of box office revenue in the first week follows a bimodal distribution, with a main mode representing movies with a relatively low (<US$100,000) box office revenue in the first week. The second mode represents movies with an average to high box office revenue. It can be seen here that the box office revenue is clearly not normally distributed, and it is in this bimodality that the problem lies. The usual literature which implicitly excludes low-performing films, actually the majority of films, concentrates only on the “second mode” of the distribution.

Fig. 2

Distribution of operating revenue in the first week. Source: Computed by the authors

Of course, if we do not have an ex ante method of knowing which mode the box office revenue will belong to, working only on higher-revenue films does not inform cinema managers if the model for higher-revenue movies does not apply to lower-revenue movies. In this chapter we will work directly with this bimodal dependent variable.

2.4.2 Production Budget

Table 2 gives summary statistics for the availability and the deflated level of budgets.

Table 2 Description: production budget

2.4.3 Star Power: Actors, Directors, Producers, and Distributors (Table 3)

Table 3 Description: star power

2.4.4 Other Explanatory Variables and Controls

The other variables that the manager has ex ante to make his/her predictions are genre and MPAA rating . We also know if the film is a remake, if it is a book adaptation, if it belongs to a series of films, if it is a prequel, if it is broadcast in a foreign language (non-English), and if it received a golden palm at Cannes before arriving on US territory. We also add a dummy indicating whether the film has had a limited release before its wide release and a variable with the duration of this limited release. One may also consider the day of the week when the movie was launched. Finally, seasonal effects are captured by the month of release of the film.

All of these variables are described in Tables 4, 5, 6, and 7.

Table 4 Description: genres of films (a movie may have several genres)

Table 5 MPAA rating: proportions

Table 6 Description of other controls

Table 7 Control variables: seasonality and day of release

2.5 Models

We now describe the models we will use to predict box office revenue in terms of ex ante available predictors.

2.5.1 Classification and Regression Tree

A classification and regression tree (CART) decision tree is an improved nonlinear and entirely nonparametric statistical learning technique introduced by Breiman, Friedman, Stone, and Olshen (1984) which allows to classify or predict a dependent variable from independent variables. A decision tree is built intuitively and is easily interpretable, thanks to its graphic appearance. The construction of the tree proceeds as follows: at each node, the algorithm splits the dataset into two subsets, using any possible predictor and any cutoff point for continuous predictors, in such a way that the two subsets are as homogeneous as possible with respect to the dependent variable. This technique has the advantage of being nonparametric, thus not postulating any a priori assumption on the distribution of the data, being robust to outliers, and supporting all types of variables. In addition, the CART algorithm handles missing values in an effective manner. When the learning sample is large as is the case here for most reference periods, the CART algorithm has properties which are similar to the nearest neighbor algorithm. On the other hand, limitations include the inability to detect combinations of variables as effective predictors and the need for a large sample (which may be problematic for the periods 2010–2012 and 2011–2012).

2.6 Random Forests

Random forest is a powerful statistical learning technique (often considered as the most powerful predictor available) developed by Breiman in 2001 (Breiman, 2001) that adapts decision trees for bootstrap aggregating (bagging). Bagging is a technique used to reduce the variance of an estimated prediction function while maintaining a relatively low bias. Here this technique is particularly well suited since the variance of the box office revenue variable is very large. It is therefore expected that this method will be more efficient than decision trees. On the other hand, as for all models built by aggregation, there is no direct interpretation. The random forest algorithm proceeds with a double random selection of both predictors and data (via a bootstrap of the learning sample), and majority vote on the resulting CART trees (hence the name of random forests).

2.7 Conditional Forests

The conditional forest algorithm developed by Hothorn, Hornik, and Zeileis (2006) makes it possible to remedy the problems faced by random forests such as selection bias or overfitting. It is therefore expected that this technique will perform at least as well as random forests. One of the main disadvantages of this method is that the underlying algorithm takes much longer to run than random forests since it performs tests to select the variables.

2.8 Gradient Boosting

The gradient boosting algorithm introduced by Freund and Schapire (1996) is a prediction method that minimizes several types of loss function with respect to a prediction function. This method can adapt to any type of data even when the number of variables exceeds the number of observations and gives very good results.

2.9 Results

We now present results obtained by each of these methods for predicting box office revenue (Table 8).

Table 8 Estimation results for box office revenue in the first week

In general, whatever the estimation method used, the R-square tends to increase with the number of observations serving as reference. The more films we have to construct our model, the better we can explain the variance. The results are similar when using movies released between 2000 and 2012 or those released between 2006 and 2012 (the best explained variance of 79% is for this latter period). Thus, it would appear that our optimal time range of films to be considered for estimation is 6 years (with a preference for the 2006–2012 range), if we want to maximize the explained variance. Random forests, conditional forests, and gradient boosting seem to be the three methods giving marginally better results. This makes sense given their complexity. However, the difference in performance between the best models and classical linear regression remains marginal.

To obtain higher coefficients of determination, we would need other variables that measure expectations of the quality of the film itself, which we do not have (and which are not so easily obtainable by cinema managers). As far as the root mean square error (RMSE) is concerned, it can be seen that it is very large, at US$10 million (January 2010 basis), and the average error rate is also very high (from 596.99% maximum to 22.30% minimum). This is due to the presence of extreme values and is the reason why we are looking at the root median square error and the median error rate to remedy this issue.

These indicators of predictive power are much more favorable than the previous ones since the root median square error fluctuates between 110,000 and US$2,000,000 regardless of the technique used. These indicators clearly identify random forests and gradient boosting as the best models.

Gradient boosting seems to dominate when it has the longest range of data available (2000–2012 or 2006–2012) with a median error rate ranging between 100 and 110%. These rates, however, remain very large. This implies that even the most advanced predictive techniques, with ex ante information only, fail to correctly discriminate between films that will achieve a low box office revenue and those that will achieve such a medium or large revenue. Stronger predictive models would need data from social networks, for example, quite a bit more difficult to extract than the data utilized in this chapter. We can however identify which variables are more important than others in predicting box office revenue, as shown in Fig. 3.

Fig. 3

Importance of variables. Source: Computed by the authors

From the graph of the importance of variables in our “best” prediction model (random, based on the period 2006–2012), we find that the production budget is the most important variable. We also find a few binary variables that are of importance, notably belonging to a series of films or being adapted from a book. Seasonality is also an important variable.

3 Predicting Oscars from “Prediction Markets”

The Intrade market (Intrade.com) was an online predictive betting exchange operated by Intrade The Prediction Market Limited. It allowed members to purchase or sell contracts on whether a future event will occur. Popular topics included upcoming elections, movie and music awards, and financial predictions of stock market indexes. Intrade did not participate in the buying or selling of contracts directly but instead had a flat monthly fee structure for members regardless of the participation level of that member. Trading was done on a per-unit basis with each unit paying US$10.00 if the event occurs and US$0 if the event does not occur. The contracts traded on a 100-point scale with 100 points representing the full US$10.00 value. For example, a contract might have stated “Mitt Romney will win the U.S. presidential election in 2012,” and the contract might have traded at 25 points. Therefore, a member would purchase this contract for the value of US$2.50, and if Mitt Romney was elected, then the member would receive US$10.00. If Mitt Romney was not elected, then the member would lose the US$2.50 to the person who sold the contract.

Intrade received significant media exposure during the 2012 presidential elections with the accurate prediction of nearly all US state electoral contests, but the exposure was overshadowed later in the year with the filing of a civil suit on unregulated trading by the US Commodity Futures Trading Commission. On December 23, 2012, Intrade ceased allowing US members from participating, resulting in a significant drop in overall participation, and on March 10, 2013, Intrade ceased all trading. The prediction market Paddy Power (www.paddypower.com) typically hosts best picture bets in Academy Award competitions, as does Bet Victor (www.betvictor.com). But unlike for Intrade, there is no convenient way to get historical pricing information out of Paddy Power or Bet Victor, and rapid changes in pricing may be difficult to track without employing some form of screen scraping. Researchers in the United States may also be shut out from even looking at international betting sites (such as www.WilliamHill.com).

Scholars have used Intrade data to investigate issues such as participation in the Euro currency (Shambaugh, 2012), the probability of a US recession (Leamer, 2008), elections (Saxon, 2010; Rothschild & Wolfers, 2008; Erikson & Wlezien, 2008), and entertainment awards such as the Grammy Awards and Oscars (Gold, McClarren, & Gaughan, 2013). Prediction market estimates of the probability of a win are considered to be very accurate, at least for events such as Oscar wins. Prediction markets were reportedly successful again in predicting the 2015 Academy Awards (Leonhardt, 2015). Figure 4 displays the price per contract for each of the nominees winning the Oscar for the 2013 Best Picture Award. We can see that up until December, there was no clear front-runner. Then beginning in December, the film Lincoln emerged as the clear favorite. However, in late January, the film Argo began to gain on Lincoln, surpassing Lincoln and holding onto that position until the end.

Fig. 4

Close Intrade contract prices for each nominated movie to win the 2013 Best Picture Award. Source: Extracted by the authors

In analyzing the average contract price for each movie, we see that the five serious contenders for the Best Picture Award were in alphabetical order Argo, Les Misérables, Lincoln, Silver Linings Playbook, and Zero Dark Thirty. What happened over this time period that could have contributed to the perceptions of which film would win the Award? Specifically, what occurred around the time frame of late January that caused such a dramatic change? On January 26, 2013, a Los Angeles Times headline read “The Gold Standard; now for real insight into Oscars – by the guilds.” The article reported that the guilds’ awards, beginning with PGAs, had been fairly reliable predictors.

The “PGAs” denote the Producer Guild of America Awards (PGAs) which were announced that evening. The headline coming out of the PGAs that evening was that Argo won the top prize of the night, the Zanuck Award for Outstanding Producer of Theatrical Motion Pictures.

This awards ceremony was followed the next evening with the 19th Annual Screen Actors Guild (SAG) awards . Their top award is the Outstanding Performance by a Cast in a Motion Picture which was awarded to Argo. So can we simply use those two awards ceremonies to predict the Oscar’s Best Picture Award? Although we focused only on the 2012 movies, we can take a quick look at the winners over the past decade. Over those 10 years, the PGA and SAG have awarded the same picture five times, and four of those times, the Oscars have followed suit and awarded the same film. In the other 5 years where the PGA and SAG have awarded different films, the Oscars selected one of the two 4 of those 5 years. Only in 2004 did the PGA, the SAG, and the Oscars each give the top award to a different film (Table 9).

Table 9 PGA awards, SAG awards, and Oscars

4 Predicting Oscars from Movie Review Data

In this section, we focus attention on whether text reviews of movies which are nominated for a Best Picture Award carry any sign of the likelihood of a movie winning the Award. We suggest that a measure of how controversial the movie is perceived to be, the value of which could be extracted by a text analysis of the reviews, is a potential predictor of a win, aside from other predictors identified in the past work.

4.1 IMDb Review Data

4.1.1 IMDb Review

In terms of text mining the opinions of movie watchers, IMDb user reviews have several advantages compared to tweets. First, most user reviews on IMDb are much longer than tweets (which are constrained to a maximum of 140 characters). Therefore, a review can contain richer and more complex thoughts than a tweet. Second, some review writers on IMDb are prolific authors, while the quality of tweets is not guaranteed at all. One can filter out reviews by non-prolific authors by choosing the “Prolific Author” filter on the IMDb review page (Fig. 5). Third, IMDb review readers can vote up or down to a review, as in “2 out of 12 people found the following review useful” in the middle part of Fig. 5. We can use it to measure the quality of a review. However, IMDb does not provide any API or structured database for downloading movie reviews. Therefore, we need to crawl the raw HTML webpage to extract review data.

Fig. 5

“Argo” IMDb reviews including prolific authors only. Source: Extracted by the authors

4.1.2 XPath and R XML Library

In this chapter, XPath is used to mechanically navigate through elements and attributes in an XML document, such as all IMDb reviews on one webpage. It is easy to read and easy to reuse and is supported by most programming languages and software packages such as Python or R. XPath expressions such as in Table 10 are not difficult to learn. Handy XPath tutorials are available at http://www.w3schools.com/xpath/.

Table 10 Path expressions for XPath (http://www.w3schools.com/xpath/xpath_syntax.asp)

We can get the XPath of the part of a webpage we are interested in by simply using the Google Chrome web browser. First, we open an IMDb review webpage in Chrome, choose the part we want to crawl, right-click on it, and from the pop-up menu select “Inspect Element.” Once that option is selected, we see two windows on the bottom side of the browser, and the part chosen earlier is highlighted in the left-down side HTML code area. We right-click on the highlighted area, and select “Copy XPath” from the pop-up menu and then obtain the raw XPath expression for the IMDb review, such as “//*[@id="tn15content"]/p[1]/text()” (Fig. 6).

Fig. 6

Obtaining the XPath of review text using Google Chrome. Source: Extracted by the authors

In this chapter, we use R to handle the process of crawling, transforming, and loading IMDb reviews. To handle the XPath in R, we need to first install the “XML” package. After installation, we can run the R code in Fig. 7 to crawl and parse movie reviews. The result looks like that in Fig. 8.

Fig. 7

R code to extract IMDb reviews. Source: Extracted by the authors

Fig. 8

Extracted IMDb reviews. Source: Extracted by the authors

4.1.3 Text Mining Using SAS Enterprise Miner

In the next step, we handle the textual dataset using SAS Text Miner, which is a plug-in for the SAS Enterprise Miner environment. The Enterprise Miner interface is displayed in Fig. 9, after we have created a New Project and New Diagram. We can then create a SAS dataset from the IMDb review documents, using the Text Import node or File Import node . The Text Parsing node in SAS Text Miner decomposes the documents into detailed terms or phrases, and the Text Filter node automatically detects misspelling in the data and transforms the quantitative representation into a compact and informative format. The Text Cluster node clusters documents into disjoint sets of documents, and the Text Topic node creates topics for each document, where one document can be associated with more than one topic (Fig. 10).

Fig. 9

SAS Enterprise Miner diagram. Source: Extracted by the authors

Fig. 10

SAS Enterprise Miner diagram for movie tweets. Source: Extracted by the authors

4.1.4 Review Themes and Oscar Chances

In this section, we discuss how a text mining of the IMDb pre-Oscars reviews gives an idea of the numbers of different themes which are perceived by reviewers for each movie and potentially yields a preliminary measure of perceived controversy. The question is then of how much “controversy” is optimal for Oscar winning purposes.

Measures of controversy and how they are used in marketing are discussed in Zhang and Li (2010). A quote from this article is very pertinent to our discussion. “From a persuasion point of view, our belief is that a convincing argument is not necessarily a mono-color picture, but instead a meaningful “bag” of positive and negative reflections” (p. 2).

This point of view may very well apply to movie chances for Oscars and other measures of success such as profit. Zhang and Li (2010) state that one possible quantitative measure of controversy is simply the standard deviation of the numerical ratings, and we adopt that point of view here as well. To extract themes from movie reviews , we use the text mining algorithms proposed by SAS Text Miner within the Enterprise Miner platform. Details of the algorithm are published elsewhere, but the algorithms work essentially as follows. Each review is defined to be a document, and a very large but sparse matrix is constructed with documents as rows and all possible terms (words in documents and their grammatical relatives, such as begin, began, beginning, etc.) as columns. Singular value decomposition (SVD) techniques are used to reduce the matrix without losing too much information, and a cluster analysis is applied to the reduced matrix, yielding for each set of reviews, a set of clusters of documents. The list of most common terms in these documents is then obtained and gives an idea of the main themes in that cluster.

As an illustration, Table 11 displays the results of this clustering exercise for Argo. For example, the main theme in cluster 3 is clearly related to perceived Oscar chances for the movie, director, and leading actor (Ben Affleck), and the main theme in cluster 4 is about the thrilling aspects of the movie.

Table 11 Clusters and main terms for Argo reviews

In the case of Argo, the text analysis yielded six clusters. Reviews of Amour, a movie by a controversial director (Haneke) on a very complex theme, related to death and euthanasia, yielded 23 clusters. Aside from whether these extracted themes are expressing positive or negative sentiments (and the approach to measuring perceived controversy in Zhang and Li (2010) does use the number of positive and negative sentiments), it is reasonable to surmise that the number of issues such a complex movie rises may be simply too large for a group to rally on.

Our study, based on just nine movies, is still preliminary; we suggest that a very interesting future direction would involve looking at measures of complexity for a much larger number of movies and investigating how correlated such measure would be to, for example, the standard deviation of ratings.

Figure 11 displays a scatter plot of the standard deviation of ratings against the number of clusters extracted by the text analysis for the nine movies.

Fig. 11

Standard deviation of ratings and number of clusters for each nominated movie to win the 2013 Best Picture Award . Source: Computed by the authors

With a few caveats, first that the standard deviation of ratings is fairly small for all nine movies, and that Zero and Amour act as outliers, we can see that the standard deviation of ratings shows a propensity to increase with the number of clusters. It is interesting to note that the five serious contenders for the Best Picture Award as perceived by the Intrade market (Argo, Les Misérables, Lincoln, Silver Linings Playbook, and Zero Dark Thirty) tend to yield a moderate number of clusters, in other words tend to raise a number of issues which a group can potentially rally on (see Fig. 11 and Table 12).

Table 12 Number of extracted themes and statistics for the nine movies nominated for a Best Picture Award

Further investigations of controversy indices in movie reviews and their role on measures of success would be very interesting. Controversy is highly correlated with word-of-mouth (WOM) activity and WOM marketing. WOM is the process of information exchange, involving in particular recommendations about products and services, between two people in an informal way (O’Leary & Sheehan, 2008). WOM communication could have a strong influence on consumer short-term and long-term purchasing behavior, influencing both short-term and long-term judgments (Bone, 1995). Another advantage of WOM is that cost of WOM marketing is low, for both online and off-line channels.

WOM could be positive or negative. The question is do customers’ negative opinions always fall on the bad side of the coin, or is there any advocacy to brand coming from negative WOM or mixed WOM (so-called controversy )? Some research indicates the possibility that controversy arising from consumers’ opinion might have a positive impact. Liu (2011) finds that movie box office revenue is correlated with the volume of WOM activity but not correlated with the percentage of negative critical reviews. Zhang and Li (2010) observe that controversy can attract market attention and potentially yield strong sales. On the other hand, consider the role of controversy in the 2014 Best Animated Feature Award, as pertains to Hayao Miyazaki’s The Wind Rises. The Wind Rises lost the award in large part because of controversy surrounding the theme in the movie. We suggest that getting a better handle on what constitutes controversy, relying on progress in text mining techniques, is likely to illuminate problems which are to date difficult to apprehend.

To conclude, this chapter has discussed a number of approaches to predicting box office revenue using variables available before the release of the movie and has also presented a number of correlates of Oscar awards. It is clear that data analysis, coupled with strong human judgment, is likely to be the key combination to investor’s risk control. In that respect, investment in the entertainment industry, for all the passion it may entail, shares many common features with those in other areas of business activity.