Where’s the “narrative” in “narrative economics”?

Bob Shiller defines “narrative economics” as the study of “how narrative contagion affects economic events”. This research program focuses on two things: “(1) the word-of-mouth contagion of ideas in the form of stories and (2) the efforts that people make to generate new contagious stories or to make stories more contagious.” In other words, if you could somehow get people to stop telling each other tall tales, then stock prices, GDP, interest rates, housing sales, etc would be different.

I’m a huge fan of this work. And even the most ardent critics of narrative economics still appreciate the power of a good narrative. For example, like all economics papers, the Campbell-Cochrane habit-formation paper has an introduction. The introduction gives an intuitive explanation of how the paper works using evocative language. The paper is called By Force of Habit for god’s sake. If that isn’t an effort to make the paper’s story more contagious, I don’t know what is.

But many researchers feel that narrative economics (in its current form) is less than scientific. And I think there’s something to these claims. Narrative economics says that economic outcomes are different because of the fact that people tell each other stories and embellish these stories in the process. But there are no specific parameters corresponding to these two tendencies in any economic model. Put differently, There’s no “narrative” in models of “narrative economics”. Whether or not a paper gets classified as a narrative-based model often comes down to how it’s written. This makes it hard to analyze how tall tales affect economic outcomes. What would the counterfactual world without the contagious narratives look like?

Compare and contrast this state of affairs how other economic forces get modeled. For example, the Campbell-Cochrane habit-formation paper allows risk aversion to vary over time. We know this because we can directly point to this parameter in the model. And, as a result, we can imagine a world where risk aversion is no longer allowed to vary over time. It’s not clear how to do the same thing in narrative-based models. There’s no parameter that corresponds to the story-telling instinct.

To place narrative economics on firm footing, we need some way of toggling on and off peoples’ tendency to tell tall tales in our models. There needs to be a narrative module in our models. That way, we can analyze the effect of this module on things like stock prices, GDP, interest rates, housing sales, etc. This post highlights two problems with current narrative-based models that make it difficult to accomplish this goal. Then, I conclude by suggesting a way to put the “narrative” into models of narrative economics.

Narratives aren’t the only kind of epidemic

When Shiller talks and writes about narrative economics, he plays up the importance of how stories go “viral”. The reason why narratives have the power to affect economic outcomes is that they can spread contagiously from person to person via word of mouth like a meme or viruse. Shiller has been a strong proponent of using epidemiological models to study financial markets. I’m a huge fan of this idea! I’ve even got a paper which takes this exact approach. Epidemiological models can tell us a lot about how trader interactions affect market outcomes, leading to things like booms and busts.

That being said, it’s important to emphasize that there is no “narrative” in epidemiological models. In these models, something is exchanged when two agents interact with one another. That something could be a virus, a story, or an Egg McMuffin recipe. Anything that these models say about narratives must also apply to a virus or a delicious new way to start your day. Epidemiological models are models of interacting agents not models of what happens when agents interact. Social finance and narrative economics are distinct fields.

Suppose there are people, of which, are currently sick. The remaining people are healthy. Each instant, a sick person encounters a healthy person with probability . And, when that happens, the healthy person becomes sick at a rate of per interaction. Otherwise, sick people recover with probability each instant. This implies that the total population of sick people will evolve as

(1)  

Flipping the logic around then says that the population of healthy people must evolve according to:

(2)  

Notice that there is no biology in this model. There’s nothing specific to structure of viruses or the life cycle of bacteria. There are just two interacting populations. These could be sick and healthy people. Or they could be excited speculators and rational investors. Epidemiological models can capture how an economic narrative spreads through a population. But they can tell us nothing about what an economic narrative actually is. This is a non-starter. We want to be able to plug the narrative module (whatever that happens to be) into an epidemiological model. But a narrative is different from how it spreads.

The tell-tale signs of people telling tall tales

Right now, economists mainly think in terms of models where investors solve forward-looking constrained optimization problems. Narrative economics argues that it’s valuable to think about narratives and models rather than just models. If this is true, then the story-telling instinct must be able to explain phenomenon that the existing paradigm can’t. Narratives must be more than just bad explanations. They must be something that is fundamentally outside the currently modeling paradigm. Otherwise, it won’t be possible to distinguish the implications of the narrative from the implications of a fine-tuned economic model.

Here’s what I mean. There are several recent papers (e.g., see here and here) that study narratives by incorporating ideas from the causal-inference literature. These papers model narratives using directed acyclical graphs (DAGs). If you have an underlying structural-equation model for the economy, then you can represent this model’s causal implications using a DAG in a way that is largely independent of the magnitudes of the parameter values involved. All that matters is the zero vs nonzero distinction.

Causal relationships without nitty-gritty parameter estimates… this might at first seem like a promising way of modeling narratives. But here’s the thing: anything that can be captured by a DAG can also be written down as a standard economic model. So, if you model narratives using DAGs, it can never be clear which is the real driver—the narrative or the underlying model.

To illustrate, suppose that the Campbell-Cochrane habit-formation model were strictly true. Suppose that the model in that paper was actually the data-generating process for observed asset-prices in the real world. In this fictitious scenario, further suppose that whenever you ask traders, they talk about the world in exactly the way that Campbell and Cochrane do in their introduction. In this set up, traders’ would have a clear narrative about what was going on in the market, and this narrative would fit perfectly into a DAG. But the narrative would not be responsible for the observed market data. If you didn’t ask traders about what they were doing, all economic outcomes would be the exact same.

DAGs capture the component of narratives that could be incorporated into well-posed model. If it’s important to consider narratives in addition to standard economic models, then their contribution must come from something that cannot be captured by simply adding a new variable to a DAG. Narrative economics must represent more than just throwing away some of the information in a model.

Narratives determine how people construe events

Narrative economics says that economic outcomes are different because people tell each other stories that get exaggerated with each retelling. To test this claim, we want some way of adding a narrative to an existing economic model. Then, we can flip the story-telling instinct on and off in the model and examine the consequences.

Whatever this narrative module looks like, it won’t be tied to epidemiological models. These models aren’t specific to narratives. They’re called “epidemiological models” for a reason. What’s more, if we want to distinguish narrative economics from the existing model-based paradigm, then a narrative must be more than just a new variable or parameter. Otherwise, it would be easy to achieve the same results using a standard model. If all you do is show that prices tend to go up with there are more positive words in the Wall Street Journal, it’s easy to gin up alternative stories that don’t involve investors tell each other good stories.

We can’t define a narrative by studying epidemiological models. And we can’t associate a narrative with a single new variable or parameter? Where does that leave us?

In his 2013 Nobel Prize lecture, Bob Shiller urged economists to incorporate more ideas from psychology, sociology, and other fields. And I think this is exactly the right way to go. But, rather than turning to epidemiology, let’s look at the subfield of psychology that studies the interface between language and the mind—namely, cognitive linguistics. We want to identify the economic effects of conveying information between people via the medium of story. It stands to reason that stories might be stored differently by the brain relative to statistics, song, interpretive dance, divine proclamation, etc.

Cognitive linguistics tells us that stories affect how people “construe events” being related to one another. For example, the force dynamics paradigm says that letting something be is not the same as pushing on it with zero force even though these two situations are identical according to every physics textbook. The narrative module we are looking for should tell us when to construe the same events in different ways.

Here’s another example. Suppose there are people with a deadly disease and doctors are asked to choose between two treatments. Treatment A results in deaths. Under treatment B there is a chance that no one will die but a chance that all people will die. The narrative should explain whether doctors frame this famous choice as “(treatment A saves and treatment B kills everyone of the time)” or as “(treatment A kills people and treatment B saves everyone of the time)”.

There is direct evidence that human brains reason about stories using something like the force dynamics paradigm (e.g., see studies like this one). So we are not talking about layering on a “narrative interpretation” to a model as is the case with epidemiological models. And, if a story can change the *relationships among entire collections of variables*, it’s not easy to account for its predictions using a single existing model. Because its effects are non-local, you would need an entirely new model for each construal of events. Turning off the narrative module would be akin to steadfastly adhering to only one model. A narrative module should account for the way that people pick and choose which model to apply at different points in time. That’s my guess about how to model the “narrative” in “narrative economics”.