As you might expect, we get a lot of emails.
Most of them are very nice. A few are nasty. After Ben’s “Lord Make Me Chaste…But Not Yet” brief, I think one reader called us “everything that is wrong with America.”
Tough but fair.
The most common emails by far, however, are from readers cluing us into examples of narratives they have observed in the wild. Only the common usage often isn’t the same thing that we mean by the term. When most people say ‘Narrative’, they mean a story that they perceive to be manufactured or artificial. Usually it is a pejorative, referring to something that is intentionally misleading, something which is decided in advance and attaches facts as convenient. They mean spin. This is…not exactly what we mean when we say it. When you read ‘Narrative’ on Epsilon Theory, you should read it as ‘an abstracted and symbolic representation of reality that replaces that reality as the locus of our thinking about a topic.’ A Narrative may be malicious or benign, but it is always most powerful and relevant to our interests when it becomes Common Knowledge, and especially when it leverages or is itself a Meme.
Why is this important to understand? Because if you are always looking for someone telling a story, you’ll miss the far more common types of narrative abstractions: facts, figures and models that stand in as proxies for the mechanic of reality they seek to model. Ben’s piece last year – Cartoons Against Humanity – is probably our clearest explanation of what we mean by this.
Yes, most narratives aren’t twisted, biased interpretations of facts. They aren’t fiat news. They aren’t overtly manipulative. They are numbers and data, unadorned, presented before you to do with what you will. The narrative’s power lies in our presumption of their sufficiently explanatory power.
We’ve talked about labor numbers and other cartoonified numbers. How about one of the most mind-numbingly obvious examples that still manages to suck us into the cartoon? Now, you may not know what I mean when I say, “run, run, pass,” but we have thousands of subscribers and readers in the Pacific Northwest who will know exactly what I mean. It is the clearest description of the Seattle Seahawks’ offensive gameplan for most of the 2018 season, including their most recent unfortunate playoff loss to my Dallas Cowboys. And why not? Any student of the most basic NFL analytics will tell you that the number of rushing attempts in a game is among the most explanatory variables for the team’s victory. No story there. That’s just a fact. If you ran the ball a lot, you probably won the game. Except, you know, if you ran the ball a lot, it’s usually because you were winning. Teams run when they’re ahead. Remember how I said this was mind-numbingly obvious (if you follow the sport, anyway)?
But as a data point, this idea that establishing the run is how you win is the natural progression in logic from the pure data, and it fuels all sorts of other more traditional story-driven narratives and, yes, common knowledge. Well, it sets up the pass. It sets up the play action. It’s necessary as a threat to keep the opposing alignment and defensive schemes honest and closer to the line of scrimmage. And the fact that all these things are a little bit true make it feel like you’re really on to something, when what you have is a garden variety, underdetermined mess of a model for how to win.
In the same way, we investors conveniently tell ourselves that so many of our models don’t work out-of-sample when they worked perfectly well in-sample because the underlying effect got arbed away, and so we move on to finding some new thing that hopefully won’t get arbed away. We convince ourselves that there’s economic intuition behind our variables, a good practice, as far as it goes. Or we convince ourselves that we’re accessing something that can’t be arbed away because it is a fundamental feature of human nature, which is probably true if our investment horizon is infinite (which it isn’t).
But no matter how accurate it may appear to be, a model is always a model. The moment we start believing that it’s as good as reality is usually the moment we get run over by a Mack truck.
What’s to be done?
In the absence of data to continually improve our estimates, we diversify our models. We take in inputs and mechanics from people with different priors. We ruthlessly evaluate the sensitivity of our models to every assumption we can identify, both explicit and implicit. Not just variables, but the structure of our estimation and prediction frameworks. We love the pieces that the team at Newfound Research have put out on this very topic. The last few posts here are focused on it. I highly recommend reading them. As you do, think about how treating your models and data like fact by not diversifying them is, itself, among the most powerful kinds of stories we tell ourselves.
And go Cowboys.