Early in Chapter 4 of his marginalism monograph, “Why Marginalism Will Dwindle, and What Will Replace It?,” Cowen reviews what happened between the late 19th century and now. He then makes his way to a striking example from something called Arbitrage Pricing Theory (APT) (pp. 99-100):
There is a recent working paper which is perhaps more striking yet, by Antoine Didisheim, Shikun (Barry) Ke, Bryan T. Kelly, and Semyon Malamud. They pick up from Arbitrage Pricing Theory (APT), a well-established idea from financial economics. APT typically looks for “factors” in the data which predict excess returns, and a traditional APT model might have found five or six such factors. Are “inflation” or perhaps “the term structure of interest rates” useful factors? Well, that can be debated, but if so, those results sound pretty intuitive. But those intuitions seem to be disappearing. In a paper by these authors, they apply machine learning methods to look for more factors. As we know, machine learning is very good at finding non-obvious relationships in the data. The largest model they built has 360,000 (!) factors, and it reduces pricing errors by 54.8 percent relative to the classic six-factor model from Fama and French. Bravo to the authors, but what kinds of intuitions do you think possibly can be supported by those 360,000 factors?
That sent me through a loop. I thought about it, consulted Claude, and ended up writing a rather long working paper, On Method: Computational Compressibility in Complex Natural and Cultural Phenomena. I thought about it some more and then went to ChatGPT this time for another long dialog (14K words). I then asked ChatGPT to summarize the dialog. Here’s the prompt I gave it:
Would you summarize the preceding discussion in terms suitable for a blog post? I want it to start with some kind of quick introductory overview and then go on through asset pricing, De Vany on movies, and Jockers on novels in that order. Two things to bridge between asset pricing and movies: 1) Despite the different objects, movies vs. stocks, the objects are both ‘thick’ and the underlying structures are similar to a first approximation though the analytic methods are quite different. 2) Galloway’s remark about stocks as brands connects movies with stocks as thick objects.
I’ve appended that summary below. The title of this post is the one ChatGPT gave to its summary.
* * * * *
We have been circling around a question raised by Tyler Cowen’s recent reflections on economics and artificial intelligence: what happens when the old economic intuitions no longer seem adequate to the objects economists are trying to understand? Cowen’s worry is that marginalist reasoning, once the pride of economics, may have helped create a world too complex for marginalist intuition to master.
That is a real worry, but perhaps not a hopeless one. If we look across asset pricing, movie dynamics, and literary history, a different picture emerges. Machine learning and high-dimensional modeling are not simply replacing human understanding. They are revealing new conceptual objects. Just as telescopes and microscopes disclosed new physical objects in the early modern world, computational models now disclose new relational objects: pricing kernels, heavy-tailed outcome regimes, temporal diagonals in literary space.
The old intuitions are not enough. But the new models may help us build new intuitions.
Asset Pricing: From Marginalism to High-Dimensional Valuation
Classical asset-pricing theory begins with a powerful marginalist intuition: investors require compensation for bearing risk. An asset’s expected return should depend on its exposure to systematic risks. CAPM gave us beta; later factor models added size, value, momentum, profitability, investment, and other variables.
These models are attractive because they are low-dimensional and intuitively graspable. A few named factors are supposed to explain many asset returns. That is the dream: a compact causal vocabulary.
But recent high-dimensional models suggest that the true pricing structure may not be compressible into five or six factors. The AIPT work we discussed begins with roughly 130 stock characteristics and then generates hundreds of thousands of nonlinear factors. These are used to approximate the market’s pricing kernel, or stochastic discount factor: the hidden valuation structure through which future payoffs are priced under uncertainty.
This is no longer a simple causal story in which one named factor explains one outcome. It is more algorithmic. The model works by probing a vast feature space and finding structure there. The individual factors may not be interpretable as ordinary causes. But the model still has an economic frame: markets price future payoffs, and that pricing process appears to be high-dimensional.
This is where Cowen’s story becomes interesting. Marginalism did not merely explain markets. It helped construct modern finance. Pricing theory made derivatives, securitization, and the decomposition of risk into tradable claims possible. But once those instruments proliferated, they helped create a financial world too complex for the original marginalist intuitions to command.
In short: marginalism may have helped build the world that now requires machine learning to map.
Thick Objects: Stocks, Brands, and Movies
At first glance, financial assets seem very different from movies or novels. A stock is an economic claim; a movie is a cultural artifact. But the distinction begins to blur once we treat both as “thick objects.”
A thick object is not exhausted by one or two measurable properties. It has many dimensions. It is embedded in social interpretation. Its value depends on history, expectation, reputation, performance, and future promise.
Scott Galloway’s remark captures this beautifully: “Stocks are like brands and that is they’re part promise and part performance.” [YouTube, Pivot, May 29, 2026.]
That is a sophisticated observation. A stock is not merely a claim on current earnings. It is a socially circulating judgment about a company’s future. The performance side includes revenue, earnings, margins, growth, debt, liquidity, volatility, and so forth. The promise side includes brand power, technological imagination, managerial credibility, founder charisma, regulatory risk, and collective belief about what the company may become.
That makes stocks resemble brands. A brand is not just a name, logo, or handle. It is a socially stabilized promise. It condenses past performance and future expectation into a recognizable entity.
And that is the bridge to movies. A movie before release is also part promise and part performance. The performance side includes director, stars, genre, budget, studio, distribution, trailers, and reviews. The promise side is what audiences imagine the movie will deliver: spectacle, prestige, emotional satisfaction, social participation, novelty, nostalgia.
So although movies and stocks are different objects, both are thick. Both circulate through populations. Both are interpreted under uncertainty. Both depend on the conversion of signals into expectation and expectation into value.
The analytic methods differ. Asset-pricing models estimate valuation structure. Movie models track social diffusion and outcome distributions. But to a first approximation, the underlying situation is structurally similar: complex objects move through fields of collective judgment.
De Vany: Nobody Knows Anything
Arthur De Vany’s work on Hollywood economics gives us a powerful model of movie dynamics. His basic lesson is that nobody knows anything, at least not in advance. Stars, directors, budgets, genre, publicity, and opening strategy matter, but they do not determine success.
Why? Because movies succeed or fail through audience interaction. Word of mouth cannot be commanded from above. The studio can launch the film, but the audience field decides what happens next.
That produces a distinctive statistical structure. Most movies fail or merely break even. Some become profitable. A few become blockbusters. The distribution is heavy-tailed, not normal. Extreme outcomes dominate the system.
De Vany’s model is therefore both causal and algorithmic. The causal mechanism is audience interaction: people talk, recommend, avoid, imitate, and amplify. The statistical result is a heavy-tailed distribution in which individual outcomes are difficult to predict but population-level structure can be modeled.
That is important for the comparison with asset pricing. In both cases, the visible attributes of the object do not fully determine its fate. The movie is released into an interpretive market. The stock trades in one continuously. Both are subject to cascades, feedback, expectations, and nonlinear amplification.
A movie’s box office trajectory and a stock’s price trajectory are not the same kind of thing. But both are records of collective valuation under uncertainty.
Jockers: The Direction of the Nineteenth-Century Novel
Matthew Jockers’ work on the nineteenth-century Anglophone novel gives us a third case. Here the object is not a stock or a film, but a novel. Jockers represented thousands of novels in a high-dimensional feature space, using hundreds of variables capturing style, theme, and other textual features. He then measured similarity among the novels.
The striking result is that chronology appears in the similarity structure, even though date is not built into the model as an explanatory variable. The novels arrange themselves along a historical diagonal. Literary history appears as direction in feature space.
This is algorithmic discovery. The model does not begin with a causal theory of literary evolution. It computes textual similarity and reveals a pattern. The causal interpretation comes afterward.
That is where cultural evolution enters. Each generation of writers grows up reading the novels produced by earlier generations. Readers, writers, publishers, genres, and conventions form a cultural matrix. A new generation internalizes that matrix, modifies it, and produces the next one. Over time, the system moves.
So Jockers gives us the phenomenon: temporal direction in literary design space. Cultural-evolutionary theory supplies the likely mechanism: a generational ratchet in which each cohort is formed by the artifact ecology inherited from its predecessors.
Algorithmic Models and Causal Models
Across these three domains, the same distinction keeps returning. Algorithmic models reveal structure. Causal models explain why the structure exists.
Classical asset pricing has causal ambition: a few underlying risks explain expected returns. AIPT is more algorithmic: high-dimensional nonlinear factors approximate the pricing kernel better than older models, but the individual factors are not necessarily interpretable causes.
De Vany has a clear causal mechanism, word-of-mouth diffusion, but he gives up on precise ex ante prediction. The model explains the distributional shape of movie outcomes, not the fate of each film in advance.
Jockers is algorithmic at the point of discovery. The high-dimensional model reveals historical direction. The causal explanation must be supplied by a theory of literary and cultural evolution.
This suggests a general research program:
First, use computational models to reveal structures we could not otherwise see.
Second, build causal accounts of those structures.
That is the optimistic answer to Cowen. The failure of low-dimensional intuition does not mean the end of understanding. It means that our inherited vocabulary is no longer adequate to the objects before us. Machine learning discloses new objects; theory then has to catch up.
A Brave New Analytic World
The most surprising thing in this comparison is not that novels and movies can be brought into a common framework. They are both cultural artifacts. They share genres, narrative structures, audiences, markets, and institutions. Many movies are based on novels; successful films can even generate novelizations.
The surprising thing is that financial assets also fit the framework.
Stocks, like films and novels, are thick objects. They are historically produced, socially interpreted, and high-dimensional. They have measurable performance and imagined promise. They circulate through populations. Their value depends on collective judgment.
The old intuitions will not guide us all the way through this world. But that does not mean we are blind. We have new instruments. High-dimensional models are doing for conceptual and cultural space what telescopes and microscopes once did for physical space: revealing objects we did not know how to see.
The adventure now is to learn how to think about them.
No comments:
Post a Comment