Predicting global ranking based on box office gross: fool’s errands and secret discoveries
Data is horrible*. It struts around like it thinks it’s reality when in fact it’s just a radiant by-product of our analog lives which we comb through American Pickers-style hoping to find new stories to tell that we’re too unsophisticated to find otherwise.
*No, I’m not going to say “Data are horrible”. This is a long article and we’re going to be together for a while. Get used to me.
The problem is that new stories and surprising discoveries can actually happen. Sometimes the data refracts the light of ordinary life across a new spectrum, revealing new sparkles, and new shadows.
But sometimes the data is not kind and we are not its masters. Sometimes the data is less like a “refractor of light” and more like a shit-covered pit viper turning to face you with its jaw distended. Take for example John Perich’s incredible box office meta-analysis at Overthinking It which starts from a very simple premise (find the truest measure of box office success) and finds itself confronted with startling and upsetting revelations. I won’t spoil the punchline; you need to read it for yourself. But suffice to say it calls into question any reasonable understanding of capitalism. Now Mr. Perich, guilty of no crime but curiosity, must answer for the crimes of his data.
Given what happened, you would think that when he open-sourced his data I would have been appropriately cautious about its siren’s song. You would think that I would have the good sense to not open the still-smoldering book held in the charred corpse’s twisted hand.
Nope. I think my exact words were, “HE JUST DIDN’T GO FAR ENOUGH! HIS QUESTIONS WERE TOO SMALL! WHAT IF WE COMBINED HIS RESEARCH WITH THE FLICKCHART GLOBAL RANKINGS!?! MAD, YOU SAY? I’LL SHOW YOU WHO’S MAD!! WOULD A MADMAN DO THIS?! HAHAHAHAHAHAHA!”
In my defense, the premise is almost impossibly alluring. Determining a brand new statistical relationship between financial success and popularity? It would be like a grand unified theory of film criticism. I wasn’t expecting the relationship to be simply linear or anything; I’m not an animal. But some sweet, naive part of me wanted there to be something that I could at least slap a polynomial regression onto.
But no, there was —
Well, here, see for yourself. Here are the top ten box office grossing movies for each of the past thirty years (adjusted for inflation) plotted against their global rankings:
Sure there’s a shape there, but nothing with any self-respecting R2 associated with it. And it makes for shit dataviz besides, especially when it’s properly labelled. So let’s just look at the top one hundred grossers:
Still nothing. It just confirms the first nothing.
Ok, what if we use Perich’s figures which control for U.S. population at the time. After all, that matters right? Maybe it’s simply a matter of “tuning our sieve”…
No. Um, okay, what if I divided total adjusted gross by the number of movie theaters in existence at the time:
Oh God that’s worse. The more rectangular the data gets, the less information we actually have.
Well Perich provides one last data set. What if I combined those last two: adjusted gross per theater controlling for population:
Great. It’s never looked worse.
What was I to make of this? There’s not a regression equation in the world that could make sense of that pile of alley trash.
I mean, what could this mean? Am I to believe that for the vast majority of movies, the vast majority of “successful” movies, there is no relationship between how many of us buy a ticket and how beloved they eventually become? Could the film world be such a chaotic wasteland? I’d expected some pockets of “cult sleepers” and “overrated blockbusters,” but not just a cloud of meaninglessness.
As I pondered these questions, arms wrapped beseeching around my screen, I entered a state I call “datamad,” a depraved, desperate state of mind which clutches and rejects ten thousand strands of datahope per minute, datatears streaming down one’s dataface. It’s the statistician version of Nightmare Alley.
But then through the fog, a pattern emerged. One that, while it may not strictly be “true”, might give us a clue as to the real reason “these” data aren’t playing well with each other.
I’m going to add some lines to that last chart. Now, don’t get mad; these are just some lines, drawn by some idiot blogger who’s desperate for material. Doesn’t mean a thing.
Well would you look at that. It’s not perfect; I can see half a dozen exceptions without even zooming in. But for the most part. . . hm, isn’t that interesting.
While I’m here, with LibreOffice all warmed up, just for fun, let’s take the base-10 logarithm of both sides. Just to crush down some of the exponentials. Totally mathematically valid. Not even really doing anything.
And that’s the last scatter plot I have for you. Things may have shifted around a bit but the affinity in each of the quadrants has solidified. The “female” quadrant’s definition could perhaps be expanded to some broader concept such as “family” or simply “not traditionally American male”.
Ok, I’ll admit we had to look fairly hard for this. And maybe this is just showboating Globetrotter algebra. But to me this offers certain. . . explanations. Not about movies, but about us, the Flickcharters.
And it’s not “just” that we skew male (though that is a priori obvious). It’s that we skew a certain kind of male. Not a Scary Movie male, and not a Clear And Present Danger male. But a Lord of the Rings male (all three films), and a Die Hard male.
If nothing else, this investigation has helped to remind me of the fickleness of legacy, and about the complexity of the human artistic experience. There is still worthwhile work (and metawork) to be done on these topics, but to believe that there is any kind of universally applicable truth in the data which can be summed up neatly in a thousand words, is to fundamentally misunderstand the subjectivity and nuance that makes movies so important.
As it stands, the data could be said to be telling us a story about how the culture’s past collective purchasing decisions have been lensed through the personalities of a certain self-selected group of Netizens. And ultimately this has more to say about the lens than the light. Who are we? And how are we different from everyone else?
Our charts know, just like they know everything else.