The Naked Future (13 page)

In a 2010 paper titled “Green-lighting Movie Scripts: Revenue Forecasting and Risk Management,” Eliashberg puts forward a statistical method for forecasting a movie's box-office success on the basis of whether it's a good example of its genre type. He calls his system a Bayesian Additive Regression Tree for Quasi-Linear (BART-QL) model. Rather than give a single-point forecast, his system provides a range of forecasts that change when new information is added (and naturally it also provides a confidence interval).

To create the model, Eliashberg and his coauthors Sam K. Hui and John Zhang amassed a data set of two hundred movie scripts for films that were released during a ten-year time frame (1995 to 2006). For each script, they recorded the genre, how much the movie cost to make, and how much the investment returned (domestic box-office revenue). They then had a group of readers score the movie on the basis of twenty-three “content” parameters that came straight from the critics. Did the hero have a strong nemesis? Did the characters evolve as the movie progressed? Was the ending surprising? Did tension build? Was the premise believable, important, or both?

They also looked at how elements in the script interacted by analyzing how the action was described according to a list of thirty weighted key words. “Ship,” “sword,” and “chamber” received scores of -.25, -.35, and -.29, respectively; “girl,” “mom,” and “office” got .24, .24, and .16. Eliashberg terms these
“correlation coefficients,” meaning that words with similar weights are more likely to show up in the presence of one another if the movie is written in a way that's genre consistent. A negative score for many of the key words does not in any way mean that your movie will tank. You just want your script populated with lots of key words that have similar values. You want swords and chambers together, not swords and offices. But there's room for nuance. The BART-QL method can then determine which
combination of variables were most likely to produce success and by how much.

The result is an algorithm that weights the wisdom of the critics and can tell you how well your script will perform at the box office. The primary variable, not surprisingly, is initial spend. Movie production budget and box-office take have a 70 percent correlation, which suggests that movie studios are already fairly good at allocating resources to the right script, and resource allocation in the movie business is
somewhat
of a self-fulfilling prophecy. But the effect diminishes after a certain budget point. In those movies in which the correlation coefficients are tightly packed together, where the men on ships with their swords stay away from the moms in the offices—movies, in other words, that exemplify their genre—the likelihood for success is greater.

The BART-QL can tell you how much a movie will make with a degree of accuracy measured by a mean-square error of .8698, versus a mean-square error of .9406 for a linear regression (average) of movie budgets alone. That may not sound like much but it works out to millions of dollars per movie.

More important, once you tinker with the script, the model can tell you if you've made the movie more profitable or less, by how much, and again with what degree of confidence. In this way it gives you a naked future view of how a potential audience will respond to changes that you make. The approach doesn't impose strict taboos. It won't tell you what sort of script to write. Rather, it provides a means to quantify
the
cost
of certain decisions in plot or character over others. If you really need that scene where the pirate (brandishing his sword) gets sucked into a time warp and finds himself in the middle of a busy insurance office, you can keep it, but that decision really limits the other choices you can make. It's a high-cost jump that will tax the audience's patience and use up goodwill.

“Everything is conditional,” says Eliashberg. “I can't tell you that ‘you want to maximize the number of exterior scenes' or ‘try
to set the dialogue such that every character has the same amount of air time.' It all depends on many things. This is why I'm not in a position to recommend an optimal story line.”

In the BART-QL method, the wisdom of the critics, and the great Aristotelian tradition they represent, has been codified and systematized. Whereas the human critic sitting in the darkened theater allows the experience of film to wash over him and then recalls bits and pieces later to construct a review, the formula recalls the most essential ingredients and then calculates how they interact with one another.

There's nothing in this that's beyond the capability of David Denby or A. O. Scott. The difference is a question of effort and time spent. Just as it takes an extremely gifted and ambitious computer scientist to undertake the sort of personal record keeping and self-quantification displayed by Stephen Wolfram and Ray Kurzweil in the 1980s, only the most eccentric critic would try to review a film by recording every plot point, the number of times particular words appear, the length of external scenes, the distribution of speaking time across the players, et cetera, and then pronounce the movie a success or a failure. But the same price-depreciation phenomenon that's enabling more people to collect and use personal information with less effort is playing out in the way the Eliashberg method serves as a critic with a perfect memory. Instead of giving consumers advice on which movies to see, it tells studio heads which scripts will earn how much and how to make scripts better. In so doing, the method looks toward a different series of hypothetical futures. One day it may be available in the form of an app that screenwriters keep on their phones.

Studios develop anywhere from twelve to twenty-five movies a year, but they keep one hundred to four hundred film scripts in development. That's a lot of money and energy going into making films that don't make back what it cost to produce or market them. The main problem with the studio system, says Eliashberg, is that big choices are made in a relative vacuum of information.

“They [studio executives] sit down in a green-lighting meeting,
look at a script, and come up with a statement like: this is
Spider-Man
meets
Superman
. Then they look up
Spider-Man
and they look up
Superman
and they see how they performed in the box office. They take an average. That's how they decide the fate of a new script.”

Eliashberg calls this “high-level comping” because this process examines the script from a height that omits important detail and compares it with another script from the same thousand-feet-up vantage point. It does nothing to solve the question of why so many people like
Spider-Man
enough to see multiple movies about him but so few people like
Green Lantern
,
Punisher
, or any of the other recent superhero super flops.

Although Eliashberg's formula could save studios millions a year, it hasn't been widely adopted. Like any true talent, he feels unfairly slighted by the Hollywood establishment and all the big egos therein. “You say, ‘we developed a methodology to help you more efficiently green-light scripts.' They say, ‘I've been in this business forty years . . .'” Still, he's optimistic. More and more, Hollywood is bankrolled by hedge funds and more and more hedge funds are run by Wharton students of the sort that show up in his class every semester.

But even if you can solve the question of which movie to make for maximum box-office spend, you may be working out the wrong equations; box office is not going to be the determinant of success that it once was, as movie-theater attendance has been falling for more than a decade.
⁴
This shouldn't surprise you. Today's moviegoing experience—modern though it appears and feels—is an artifact of the industrial age. Millions of people pouring into cities to work in factories provided the necessary market for a new form of entertainment, one that was based around what was a very new technology. The device revolution has given us the movie theater you carry in your pocket. The allure of the velvet-seated film house is on the decline. The greater mystery is why it survived for so long. The rise of such streaming services as Netflix suggests a future of more personalized movie viewing and a
relative
waning of blockbusters in the future.

Herein lies my major disagreement with Netflix's Amatriain. Netflix, at the moment at least, seems willing to perpetuate the studio system's worst practices. In fact, high-level comping is exactly how Netflix recommends movies to you, on the basis of genres and actors in which you've expressed interest and how your interests correlate with others. The problem is that these genre tags and
House of Cards
tricks are only an approximation of the story itself. Netflix is firmly stuck in the big data present.

Bayesian systems such as the BART-QL could be most useful in extrapolating from
sparse
data sets, such as the sorts of movies one viewer has seen and liked. “Let's say you have multiple consumers, and for each consumer you have very few observations, like what movies he or she has seen, or liked, or whatever. Bayesian analysis allows you to aggregate across different consumers and make individual level predictions for each
consumer, in instances where for each consumer you have very few data points,” says Eliashberg.

There is no perfect movie but, in the naked future, there may be a statistically perfect movie for
you
. Locked away in Netflix's servers is, perhaps, the data to answer the question of why you like the stories that you do.

What would this look like? The perfectly personalized film probably has a lot in common with a very modern, extremely story-intensive video game in that the viewer (or player) would be able to make numerous decisions about the characters populating the film and even provide input into the story. That level of influence could be customized for different users. Because more and more video games have an online component, they provide a unique opportunity for telemetric data collection and thus prediction. In 2011 a group of researchers from the University of North Carolina showed that they could accurately predict how a player would respond to a key challenge or threat in a massively multiplayer online role-playing game, data that game developers could use to make their games better, both in the design stage and later, during game play.
⁵

Imagine that same data collection applied to the act of selecting a movie to watch from a streaming service such as Netflix. Let's
say you like thrillers that end on a down note but where the protagonist lives. That system should be able to steer you toward movies that meet that plot criteria. What about surprise twists? You probably have a tolerance for certain types of surprises over others (there are surely some that you hate). At the very least, a system looking to anticipate your absolutely perfect Friday 9
P.M.
movie should be able to keep you from having to see the love interest die, or the charming nerd character become a tough guy, or Kevin Costner's bare ass if that's the sort of thing that a system with enough data on you can predict will ruin your night.

Video games, by definition, are an active experience. We think of movies as primarily passive, the same way we think of reading. But in the naked future there is no such thing as a purely passive entertainment experience. For instance, as millions of readers move from print to e-readers, both book retailers and publishers are discovering new insights about what makes a good book.

In July 2012,
Wall Street Journal
writer Alexandra Alter broke the story that Amazon, through its digital-book sales, had already begun telemetrically analyzing reader behavior to better predict which books will sell better overall, and which books will sell better to individuals. The rise of e-books is creating digital, data-driven dialogues between readers and publishers, and one day, maybe even between readers and authors, transforming, in Alter's words, “the activity [of reading] into something measurable and quasi public.”

So far, Amazon has discovered that people who read popular series books like Harry Potter tend to read them rapidly and completely, that serious nonfiction gets read in spasms of interest, and that literary types, who you might think would be the most patient and determined readers, are the most likely to jump from book to book, passage to passage, like small-winged birds from branch to branch. If you have the Kindle app installed on your Mac, as I do, you can actually see which passages of books others have highlighted and how the world is responding to the text in real time. (By the way, please highlight that. Thanks.)

“The bigger trend we're trying to unearth is where are those
drop-offs in certain kinds of books, and what can we do with publishers to prevent that?” Jim Hilt, Barnes & Noble's vice president of e-books, told Alter, “If we can help authors create even better books than they create today, it's a win for everybody.”
⁶

Not every consumer is comfortable with the notion of their e-reader's reading them, or with Netflix's analyzing not just personal movie choices but also movie-watching behavior in order to deliver more personalized recommendations. True, back in the 1990s (hardly a golden age for personal privacy) people felt the same way about Amazon features that are common today, such as the notification appearing at the bottom of the Amazon screen that says “Customers Who Bought This Item Also Bought
 . . .
”
“People were freaked out about that in 1998,” Russ Grandinetti, vice president for Kindle Content, told a group in San Francisco. Today, says Grandinetti, “We've seen higher rates across the industry of people willing to share.” That's good news whether you want Amazon tracking your behavior or not. Even if some people opt out, they'll still enjoy better recommendations because millions of other people let Netflix watch them watch movies.

Even if the entertainment products themselves don't seem to be changing, best-practice guidelines for creating entertainment are in some surprising ways. In 2010 psychologists James E. Cutting, Jordan E. DeLong, and Christine E. Nothelfer published a paper showing that, over the course of the last seventy years, the length of shots in particular movies was coming ever closer to resembling the attention patterns of the human brain, as measured via a mathematical wave analysis technique called the Fourier analysis. Movie shots, overall, are getting shorter. The average shot in the 2008 film
Quantum of Solace
lasts 1.7 seconds. Compare that with Alfred Hitchcock's experimental 1948 film
Rope
, which includes several shots that last nearly 10 minutes.
⁷
But shot length won't continue to collapse until it reaches nothing. No one wants to watch a movie composed of millions of millisecond flashes of story. So what's the optimal shot length under what circumstances? By bringing the current trend in line with a larger theory of human attention, Cutting's
research provides a framework to predict what movies of the future will look like: they'll resemble, at least in feel and rhythm, the chaotic world that made us the animals we are today.