Sina Seyfi

Can AI Discover the Next Big Economic Theory? Behind the Scenes of AutoTheory

As economists, we traditionally view the development of theory as a quintessentially human endeavor—a creative spark fueled by intuition and years of domain expertise. But my colleagues and I are exploring a provocative question: can we formalize and automate the generation of economic theories using Large Language Models (LLMs)?

Our answer is AutoTheory, a system we build to turn theory discovery into a scalable, “generate–test–retain” process.

The Philosophy: Creativity as Variation and Selection

Creativity (and scientific discovery) requires a large number of attempts. Researchers try many different models and “some” of them turn out to be good. After all, the best way to have a good idea is to have a lot of idea. We model our system after the “blind variation and selective retention” (BVSR) framework.

The LLM acts as our variation engine, proposing diverse mechanisms. However, the real heavy lifting happens in our selection pipeline, where candidate theories face rigorous mathematical audits, code execution, and data calibration. A theory either reproduces the empirical data or it does not—regardless of how persuasively the LLM describes it.

How the Engine Works

To ensure we do not just get “average” ideas, we implement a few key strategies:

• Expert Personas: Each run draws from 20 methodological perspectives—like “Bayesian,” “Physicist,” or “Minimalist”—to force the AI to look at puzzles from unconventional angles.
• Theory Evolution: We use evolutionary algorithms where the AI mutates successful ideas based on “referee reports” or crosses over two different models to find a hybrid solution.
• The Math Audit: LLMs are notorious for algebraic slips. Every proposal in our system undergoes an independent mathematical audit to verify market-clearing conditions and first-order derivations before any code is written.

Putting It to the Test: Solving the Multiplier Puzzle

We apply AutoTheory to the price-multiplier scaling puzzle—a significant gap between how much asset prices should move versus how they actually move in the data.

Good news! AutoTheory discovers multiple distinct mechanisms that match the data perfectly, such as the Bayesian-Breadth-Cost CARA model and Dual-Agent Limited-Diversification frameworks.

The “Parsimony Trap”

We also challenge the system with the Equity Term Structure puzzle, which requires matching ten different moment conditions. This proves much harder. We identify a “parsimony trap“: while the AI can find complex models to fit the data, they often require too many parameters to be considered “good” theory. This teaches us that for truly difficult puzzles, the bottleneck is not just generating ideas—it is finding the simplest mechanism that still works.

The Future of the “Dismal Science”

What does this mean for our field? We believe this shifts the human role from “producer” to “curator“. Instead of spending years hand-crafting a single mechanism, researchers can now use tools like AutoTheory to systematically explore the landscape of possible models in a matter of days.

The next great economic insight might not start at a whiteboard, but in a massively parallel search through the space of ideas. I am excited to see where this automated “Aha!” moment takes us next.