In the 1980s and ’90s, SFI played laboratory to a promising new method known as agent-based modeling. Instead of relying on averages and unrealistic mathematical assumptions, these models were built from the ground up to account for individual differences and complex interactions.
In ensuing decades, agent-based models (ABMs) proliferated across many fields. In economics, they advanced from little computer dots roving on a sugary landscape to complex models of housing bubbles, national supply chains, and the economic impact of COVID-19.
But mainstream economics largely sidelined ABMs, which until recently were typically highly stylized (studying a generic economy at a generic point in time, rather than a specific economy at the present moment), and thus limited in efficacy.
In the background, economic microdata has exploded — not benefiting traditional economic models built for averages, yet imbuing ABMs with potent predictive ability. Add in the exponential growth of computing power, and agent-based models may soon transform the future of economic policy.
To supercharge wider use of ABMs in economics, SFI hosted a month-long working group — the longest ever held at the Institute. The new Gurley Forum, opened in April, made the experimental format possible. “Economic Agent-Based Models: Crossing Over the Tipping Point” brought together economists, central bankers, computer scientists, social scientists, physicists, and other researchers from August 4–29, supported by the Zegar Family Foundation and the Omidyar Network.
“Mainstream economic models have not been able to solve some big problems: What causes inflation? What determines interest rates? What are the main factors behind recessions? What forces drive inequality?” says SFI External Professor J. Doyne Farmer (University of Oxford). “At the SFI working group, we developed an approach to building an agent-based macromodel that could finally resolve these paradoxes.” Farmer co-organized the working group to fill in gaps and expand the real-world impact of agent-based models.
For example, participants demonstrated why agent-based models could help economists better understand vexing issues, including supply-chain dynamics, job vacancies, unemployment, and the green-energy transition.
“As a result of the SFI working group, we can see our way to creating a model of a whole country’s economy in software,” says SFI External Faculty Fellow Rob Axtell (George Mason University).
Leading mainstream economists from institutions like Columbia, Johns Hopkins, Cambridge, and L’Observatoire Français des Conjonctures Économiques participated in the meeting, addressing a major gap in the use of agent-based models. Representatives from the central banks of Italy, Canada, Hungary, and Chile also attended, sharing why they have adopted agent-based models for certain purposes.
“We saw some of the first examples of agent-based modeling competing directly with the conventional approach,” says Axtell. “In the last year or two, several countries have built high-quality agent-based models for central banks and policymakers to use because they’re so good at ingesting microdata, instead of statistical averages.”
The month-long meeting format allowed for intensive collaboration between established experts and the next generation of interdisciplinary and social-science researchers.
Participant Jordan Kemp recently completed a Ph.D. working with SFI External Professor and Science Board Member Luís Bettencourt (University of Chicago) and now is an Oxford postdoctoral fellow. At the meeting, he appreciated conversations on how to navigate the intersection between complexity science and economics.
“It’s not so often that economists talk to people who study ABMs. We are all asking the same questions in very different ways, and at the working group, we actually made progress towards finding compatibilities between complexity and classical economics,” Kemp says.
Finally, participants explored machine-learning techniques (such as large language models, or LLMs) that can vastly improve the predictions of agent-based models.
“We can now encode agents in ABMs with the rich priors of LLMs, which allows agent behavior to adapt fluidly to context,” says participant Valentina Semenova, an International Monetary Fund economist. “At the working group, we talked about early successes, like multi-agent LLM systems that study political polarization. We are only scratching the surface of what these cognitively enhanced agents can reveal about the real world.”
Speakers presented examples such as LLMs that simulate individual responses to impending environmental disasters, and “cocktail parties” of LLM agents bumping into each other and exchanging ideas.
“In the 1990s and early 2000s, ABMs played an important role in helping understand the underlying generative mechanisms that animate economic systems, but they struggled to produce tactical insights on specific interventions,” says SFI Vice President for Applied Complexity Will Tracy. “Recent advances in AI, data, and compute have fueled new ABMs capable of guiding policymakers. This extraordinary working group is doing the intellectual work needed to make modern agent-based modeling available to a larger cohort of decision-makers.”
Organizers launched several new work streams, with plans for publications next year. “Because agent-based models can include so much detail, they address the heterogeneity of our society much better than traditional models,” says Farmer. “After four decades of progress, we now have the computing power and most of the data we need. The time is right for agent-based models to play a role in economic policy.”