J. Doyne Farmer, Director of Complexity Economics at INET Oxford, believes that creating economic models that can effectively incorporate behavioural realism to make useful predictions may be the most important problem in economics today.
The simulations used in complexity economics are called ‘Agent-Based Models’ (ABMs). These models substantially differ from the sets of equations used in the traditional economic approach, which assume forward-looking agents acting rationally with predictable behaviour.
Complexity economists on the other hand assume from the outset that agents are boundedly rational, meaning that they have imperfect information and a limited ability to reason. As the name emphasises, the individual building blocks of ABMs (such as consumers, workers, households, firms or governments) have agency, in other words they make decisions that impact not only their individual path, but other agents and the system itself.
Algorithms describing how each agent makes decisions are programmed into the computer (often drawing on research about how agents make decisions in the real world), the agents gather information and make their decisions, their actions impact on the state of the economy, which in turn generates new information, causing the agents to make new decisions, and so on. These feedbacks create dynamics, and unlike traditional models, ABMs are not assumed to settle into a steady state equilibrium, but instead evolve over time.
There is growing evidence that such models can more realistically describe economic phenomena than traditional models, make better predictions, and be a more useful tool for policymakers. Our recently published research demonstrates how ABMs are now 'coming of age' and are finding practical use across three areas: central banking, economics and finance, and AI.
Central banking: Agent-based models poised to play greater role
A new INET Oxford working paper, published by the Bank of England and Bank of Spain, charts the growing use of ABMs by central banks.
The functions of central banks have expanded dramatically since the global financial crisis of 2007-2009 to include new challenges such as cybersecurity, climate change, cryptocurrencies and rising economic inequality. This paper explains how integrating ABMs into the analytical frameworks of central banks can help address these challenges - and provide more effective policy responses. The authors point to several advantages that ABMs offer:
ABMs offer a high level of heterogeneity across multiple dimensions, allowing complex interactions between heterogeneous agents;
They generate non-linear dynamics similar to those observed in the real world, such as boom and bust;
ABMs provide a flexible framework for analysing the impact of policy - not only on aggregate measures such as GDP or unemployment, but also the distributional effects of these policies; This flexibility allows for capturing both the particularities of a given country and the details of real-world policies, as well as for promptly adapting the model for changing economic circumstances.
Co-author of the paper Jagoda Kaszowska-Mojsa said as central banks face mounting challenges—from financial instability to climate risks and digital currencies—their analytical tools must evolve accordingly.
“This paper not only highlights the growing role of agent-based models (ABMs) in policy work but also provides a valuable list of key studies that central bankers can build upon to refine their own analytical frameworks.”
“By leveraging these insights, central banks worldwide can better adapt to an increasingly dynamic and uncertain economic environment and enhance their ability to address emerging risks."
This paper is a collaborative effort with co-authors from 5 different central banks: Bank of England (Marc Hinterschweiger, Arzu Uluc); Banco de España (Adrian Carro); Banca d'Italia (Aldo Glielmo); Narodowy Bank Polski (Jagoda Kaszowska-Mojsa); and Magyar Nemzeti Bank (András Borsos). In line with this collaborative nature, the paper has been published in a coordinated manner by both the Bank of England and the Banco de España, and it will soon be published as part of the Santa Fe Institute's upcoming volume "The Economy as a Complex Evolving System, Part IV".
Agent-Based Modeling in Economics and Finance: Past, Present, and Future
Axtell and Farmer's comprehensive paper recently published in the Journal of Economic Literature, charts the rise of ABMs from their antecedents in the 1950s, through the computational revolution, to the recent pandemic, where models managed to successfully predict the economic impact of Covid-19 on the UK economy.
Pioneers in the field, Axtell and Farmer look ahead to the future trajectory of ABMs, charting a course to modelling that can navigate the big challenges we face like climate change, inequality and financial stability.
Farmer said: "This is a propitious time for ABM due to a confluence of factors. Continued advances in computing hardware, largely driven by Moore’s law, makes bigger models ever more feasible. When combined with the growing quantity and availability of high-quality data from both administrative and commercial sources, large-scale, empirically grounded ABMs are becoming possible and have recently begun to appear."
Commenting on the significance of the work, François Lafond, Deputy Director of Complexity Economics at the Institute for New Economic Thinking at the Oxford Martin School said:
“Agent-based models have been around for a long time – first as a proof of concept for emergent phenomena, later as proper economic models reproducing many stylized facts at once. Today, many technical barriers have been overcome: we have greater computational capabilities, powerful software libraries, methods to calibrate parameters, and data to initialise micro states. As a result, agent-based models are now competing head-to-head with other approaches, and we expect the field to boom in the near future. The Axtell-Farmer paper provides a superb, encyclopaedic review which will serve as a first port of entry for everyone joining this field in the years to come.”
John Muellbauer, Co-Director of the Macroeconomics and Finance Programme at INET Oxford said:
“This is a masterly survey by the two leading exponents of agent-based modelling, making a convincing case for the relevance of the approach in all branches of economics, and social science more generally. Computer simulations for populations of agents following specified behavioural rules and interacting with each other generate dynamic systems behaviour. Such behaviour is often more realistic than that implied by standard economic models, which for tractability make unreasonable assumptions about the information processing capability and rationality of agents.
"For general readers the paper bubbles with fascinating insights into many phenomena in all branches of economics, micro and macro, whether in financial, labour, goods or credit markets, demonstrating the power of moving beyond conventional paradigms. The paper is formidably erudite in setting the development of ABM in its historical context, not just in economics, but in physical and computer sciences, probably setting a new record for the JEL in the length and breadth of citations."
Features of Agent Based Models (from Axtell & Farmer, 2025).
Present in ABMs |
Not present in ABMs |
|
|
|
|
|
|
|
|
|
|
|
AI: Award-winning work on robust policy design
In their latest paper, Akash Agrawal, Joel Dyer, Aldo Glielmo and Michael Wooldridge focus on the challenge of ensuring policies optimised in ABMs perform reliably in real-world environments. Addressing the 'simulation-to-reality-gap', this research helps pave the way for more accurate forecasts that will open up ABM use to more policy-makers.
The work, 'Robust Policy Design in Agent-Based Simulators using Adversarial Reinforcement Learning', achieved the award for Best Paper at the Multi-Agent AI in the Real World Workshop at The 39th Annual AAAI Conference on Artificial Intelligence on 3 March 2025.
Co-author and Postdoctoral Researcher Joel Dyer said:
"The agent-based modelling (ABMing) paradigm offers an avenue for building realistic synthetic representations of complex socioeconomic systems. For this reason, they have the potential to be extremely useful tools for policymakers: by using ABMs to simulate the potential consequences of policy interventions, policies can be tested and optimised – for example, through the use of reinforcement learning methods – to verify that they would improve social welfare if implemented in the real world.
"Unfortunately, given the complexity of the real world, any model – agent-based or otherwise – will imperfectly represent some aspect of reality or another. Care must therefore be taken to ensure that policies designed with the assistance of ABMs are robust to any mismatch that exists between the environment in which they were developed and the real-world system in which they will take effect.
"Our work is a first step towards developing the mathematical and computational tools for ABMs that will help guard against this danger. By investigating methods for improving the robustness of policies developed through computational agent-based simulations, we aim to ensure that ABMs can be effectively used to support good real-world decision-making."
