Could Algorithms Be Better at Picking the Next Big Blockbuster Than Studio Execs? (wired.com)

In a world where artificial intelligence is no longer just a Spielberg-Kubrick collaboration, could algorithms be better at picking the next big blockbuster than studio execs? From a report: "Filmmakers are getting closer to understanding what moviegoers go to theaters to see thanks to neural networks fed off of data from previous box office hits," says Landon Starr, the head of data science at Clearlink, which uses machine learning to help companies understand consumer behavior. "Although this technology isn't spot-on quite yet, AI-powered predictions are likely stronger than the human calculations used in the past." And they're advancing quickly.

Vault, an Israeli startup founded in 2015, is developing a neural-network algorithm based on 30 years of box office data, nearly 400,000 story features found in scripts, and data like film budgets and audience demographics to estimate a movie's opening weekend. The company is only a couple years in, but founder David Stiff recently said that roughly 75 percent of Vault's predictions "come 'pretty close'" to films' actual opening grosses.

Scriptbook takes a similar approach, using its own AI platform to predict a movie's success based on the screenplay only. The Antwerp startup's AI analyzed 62 movies from 2015 and 2016, and claims it was able to successfully predict the box office failure or success of 52 of them, judging 30 movies correctly as profitable and 22 movies correctly as not profitable.

Comments

By the rule of headlines

[H3lldr0p]

The answer is of course "No".

The reason being is that it's mostly random. Kinda like the stock market. You can't make predictions based on past performance. Something or someone may fall out of favor in the public's eye. Or something or someone may be suddenly popular.

Re:By the rule of headlines

[Actually, I do RTFA]

I was going to highlight this as a headline example where the obvious answer is "yes". Studio executives seem like they can be replaced by the AI of a magic eight ball with no loss in efficiency. (At least on the script picking side. I'm sure there's a lot of work that goes into Hollywood accounting.)

Re:By the rule of headlines

[JoeDuncan]

I'm sure there's a lot of work that goes into Hollywood accounting.)

It IS really hard to make a movie like Infinity War LOSE money on paper...

Re:By the rule of headlines

[Anonymous Coward]

We would assume it would be hard to fuck up star wars. Yet here we are.

I see THREE SHELLS!

[Thud457]

Like where in San Angeles, all restaurants are Taco Bell,
in TEH FUTAR, all movies will be Pixar Marvel Wars.
It is an inescapable conclusion.

No.

[Anonymous Coward]

The goal is not to create good movies.

The goal is to create movies that sell.

That means pandering to the lowest common denominator of the movie going public.

Movies are not made to entertain people of above-average intelligence.

Yeah, but it's not enough profit

[rsilvergun]

the question is what made Avengers clear 1.5 billion+ and Batman v Superman only do half that? Can you quantify that so that every movie performs to the limits of the market (e.g. every movie goer sees it)?

I"m inclined to say yes. These aren't high concept art movies, they're popcorn flix. Like a pop song they follow a formula. Eventually that formula can be understood. Kinda like "PsychoHistory" from the Foundation novels. Eventually the math will be understood.

Re:Yeah, but it's not enough profit

[alvinrod]

The question is what made Avengers clear 1.5 billion+ and Batman v Superman only do half that?

People are pretty good at finding patterns. In this particular case the pattern is that most Marvel universe movies are pretty good and will be enjoyable and that if a DC universe film is good, it was probably an accident. There isn't really a formula here, just a case of a shitty reputation. You can probably look for the underlying causes of that if you want to avoid such things in the future, or try to correct them now, but on the surface it's not difficult to reason out why one film did vastly more business.

Re:Yeah, but it's not enough profit

[Anonymous Coward]

I do research in machine learning, literally designing the sort of predictors you're talking about, and I'd say no. Even given a lot of data (and even data on every movie in history wouldn't get near to what you'd call "big" data) you're still dealing with noisy samples - whether that noise comes from hidden variables or just dumb luck - and the system is fundamentally time-varying in either hard or impossible to predict ways (last years blockbuster could be this years flop as tastes evolve. If you don't believe me take a look at some of the unbelievably long and boring hits from the 40s and 50s).

Futurama

[sexconker]

Futurama did it with the Execubots.

Executive Alpha, programmed to like things it has seen before.
Executive Beta, programmed to roll dice to determine the fall schedule.
Executive Gamma, programmed to underestimate middle America.

Personal Experience

[nuckfuts]

A number of years ago I was excited to try a movie recommendation website. The premise was that you would rate a bunch of movies, and their algorithm would learn which ones you liked and which ones you didn't. It would then compare your results with other users. Suppose someone else rated a bunch of movies the same way you rated them, but in addition rated some movie very highly that you hadn't seen yet. That movie should make a great recommendation for you!

I started by spending a fair chunk of time rating a whole bunch of movies. I figured that the more I trained the algorithm on my tastes, the better results I would get. Finally, I decided to try for a recommendation. Lo and behold - up came some movie I had never even heard of, that was rated very highly by people who shared my taste.

Well, not only did I not enjoy the movie, I absolutely HATED it. I have distrusted recommendation systems ever since.

Don't doubt algorithms could improve the median.

[hey!]

Every person in charge of acquiring new material for a big media company is always on the lookout for the same thing, only different. That's because the financial backers have two, largely mutually exclusive goals: guaranteed audiences and a runaway hit. In economic terms they're looking for an investment with higher than normal returns for its risk.

That's why pop culture is so clogged with retreads. It's only a matter of time before we see Star Trek: With Tits.

Now I happen to know more about publishing than movie making, so I'll focus on that for a moment. New authors submit their manuscripts on spec to agents and publisher acquisition editors. These agents and editors are usually pretty sharp, but that makes their time valuable. So someone like an intern has to wade through the "slush pile". It's a horrible job because 99.9% of the slushpile is pure rubbish.

What the slushpile reader does for hours on end is skim the first page, and toss, skim the first page, and toss. The first page is about ten lines of text in standard manuscript format. But if an algorithm could make the first cut, it would be able to examine entire manuscripts for the desired combination of (a) resemblance to past hits and (b) differences from recently published books, winnowing hundreds or thousands of manuscripts down to a couple dozen candidates fit for human eyes.

The exact same process could be used for movie or TV spec scripts. American broadcast TV shows often have a problem ginning up enough story ideas to fill an entire season, but accepting spec scripts means someone has to deal with the slushpile. So there's usually a couple of writers-don't-have-any-ideas episodes each season. If you could process a couple of thousand spec scripts and pick a dozen candidates that fit the show, you might find an idea you could use.