On 5th November 2008 the Queen opened a new building at the London School of Economics. The visit included a short briefing on the credit crunch from Luis Garicano, the department’s director of research, who had a chart to go with it.1 When he finished, she asked him one question, which he later recalled as “Why did nobody notice it?”2

The crisis was driven by a series of feedback loops. Rising house prices made mortgages look safe to lend on, and the extra lending pushed prices up again. Banks were also lending to each other, so trouble at one of them could spread to the rest. When prices turned, the same process worked in reverse.

Before 2008, most central banks were using models built around a single representative household, with disturbances arriving from outside the economy.3 A process involving feedback between borrowers, lenders and banks was difficult to capture in those models. Whether a different kind of model would have picked up the warning signs is harder to say.

Two Ways of Modelling the Economy

The standard approach looks for equilibrium: a position where nobody wants to change what they are doing, given what everyone else is doing. Complexity economics takes a different starting point. It treats the economy as a system in which people and firms keep adjusting to one another, often creating patterns and changes that nobody intended.4

The idea has older roots than the label. The authors of one Santa Fe Institute volume trace it back to Adam Smith.4 The institute gave researchers a place to develop it. In September 1987 it held a conference organised by the physicists Philip Anderson and David Pines and the economist Kenneth Arrow. This led to its first research programme, “The Economy as an Evolving Complex System”, which Brian Arthur ran.4 Arthur remembers that Arrow invited ten economists and Anderson ten physicists and other scientists, among them Doyne Farmer, Stuart Kauffman and John Holland.5

Complexity economics was not originally intended to replace the conventional approach. A 1997 Santa Fe volume described it as complementary to equilibrium economics. It is also worth remembering that mainstream economics is not one single school of thought. When referring to “standard models” here, the term means DSGE models (dynamic stochastic general equilibrium), the family of models that central banks used heavily before 2008.

What the Standard Models Missed

There were three features of those models that made a crisis like 2008 difficult to represent. Stiglitz (2017) argues that their microfoundations were wrong and that relying on a representative agent left them badly placed to explain deep downturns.3 Romer (2016) objects to the way they produce fluctuations: they rely on random shocks without explaining where those shocks come from, while the decisions made by people within the model receive too little attention.6 There is also a practical problem with using one household to represent everybody. It leaves no clear distinction between borrowers and lenders, and therefore no obvious way for a default to spread from one to the other. This final point is an inference from the model's structure, rather than a claim made directly by the sources.

That criticism fits the basic versions of these models best. Models with financial frictions existed before the crisis. Bernanke, Gertler and Gilchrist (1999) developed a “financial accelerator”, in which a borrower whose finances weaken faces higher borrowing costs, making the downturn worse.7 Kiyotaki and Moore (1997) modelled a related credit cycle.8 After 2008 these approaches were extended, and some models began to include banks explicitly. One estimate found that the financial accelerator helps account for the fall in US output in 2008.9 Even so, a survey cautions that models of financial frictions are not necessarily models of financial crises, and that only some newer non-linear models build crises directly.10

The criticism, then, needs to be more specific than saying that equilibrium models blame everything on outside forces. They can now show how a shock is amplified through the financial system. The more difficult question is whether they can explain a crisis that develops within the system itself, as weakness at one institution spreads through a network.

What an Agent-Based Model Does

An agent-based model (ABM) simulates a large number of households, firms and banks. Each follows its own rules and reacts to the behaviour of the others. The modeller can then see what happens across the system as a whole.

A simple housing-market example helps explain how this works. This is an illustration, not the result of any one study. House prices rise, making lenders more willing to offer mortgages. More lending then pushes prices up further. Rising prices also increase the value of the property used as collateral for existing loans, giving banks another reason to lend. Borrowers and lenders become more exposed as the boom continues. If prices then fall, some borrowers may be forced to sell, and those sales can push prices down further.

Some research points in this direction. Geanakoplos et al. (2012) built an agent-based model of the housing market using individual data from the Washington, DC area. Their preliminary analysis suggests that the 1997 to 2007 boom and bust came mostly from changes in leverage rather than from interest rates.11 The Bank of England also worked with Oxford researchers on a model of the UK housing market. It included first-time buyers, owners, buy-to-let investors and renters, interacting with banks and a central bank. The researchers found that a bigger buy-to-let sector may amplify house price cycles, while a cap on loans relative to income dampens them.12 Haldane (2016) argues that models of this kind can produce booms and busts without requiring an outside shock.13

Banks can be modelled as a network too. Haldane and May (2011) drew on ideas from ecology and the study of infectious disease. If banks adopt similar strategies, each may appear safer on its own. But the system becomes less diverse and more vulnerable to a chain of failures. The authors were careful to point out that financial systems and ecosystems are not the same.14 The comparison is useful for identifying possible risks, but it does not tell us exactly how a financial system will behave.

Why They Are Hard to Validate

Building an agent-based model is one thing; showing that it works is another. There are several reasons why this can be difficult.

The rules governing agents are flexible. There are many plausible ways to describe how a bank or household makes decisions, and a modeller's choices may reflect personal preferences that another researcher would disagree with.15 The results can change depending on those choices.

There is also a difference between reproducing something that has already happened and identifying it before it happens. Platt (2019) reports that validation is mostly qualitative, based on whether a model can reproduce patterns already seen in the data. Even the better calibration methods, he argues, need further work before large models can be pinned down properly.16 A model that reproduces past booms and busts has shown that it can describe those events. That does not necessarily mean it would have identified the conditions leading up to them. Finally, severe crises are rare, which leaves relatively little historical data with which to test a model under the conditions where it matters most. This is a further limitation arising from the scarcity of crisis events.

Conventional models have an advantage here. Their relationships are comparatively easy to fit to data.15 Axtell and Farmer (2025), in their own review of the field, also identify a number of problems that agent-based modelling still needs to address.17

How Far Central Banks Have Gone

It is worth being careful about claims that central banks have adopted these models. There are different levels of use, from research projects to models that form part of a bank's regular toolkit. A Bank of England working paper surveying the field describes agent-based models as complementary tools for central banks.18 At the Bank itself, one agent-based model appeared in 2008, followed by a gap of several years. Three more were developed between 2016 and 2018.19 The UK housing model discussed above was among the first attempts to use an agent-based model to support central bank policy analysis, and was published as a research working paper.20

Canada appears to have gone further. The University of Amsterdam reports that the Bank of Canada has added an agent-based model, CANVAS, to its official model portfolio. It describes this as the first large-scale macroeconomic agent-based model actively used by a central bank.21 That account comes from a university rather than the Bank of Canada itself, so it would be worth checking against the Bank's own documents. Most of the other examples found are research projects or exploratory exercises.

Was the Queen’s Question Answerable?

No model can reliably tell us the date of a crash. A model of systemic risk can, however, show which conditions make a financial system fragile. These are different tasks, and it is important not to confuse them.

For identifying fragility, the complex-systems approach has a stronger case. A model with linked balance sheets, high leverage and feedback between institutions can show how a system might become vulnerable before anything has actually gone wrong. A model built around one average household has much less scope to represent those relationships. The economists who wrote to the Queen in 2009 made a similar point, arguing that the crisis reflected a failure to imagine the risks to the system as a whole.1 Garicano offered a related explanation at the time. People had followed their incentives, but in many cases those incentives rewarded behaviour that was damaging to society.22

The question of whether these models could have given an early warning is harder to answer. No test has been identified in which an agent-based model, using only data available at the time, identified an emerging weakness that a conventional model missed. Until there is evidence of that kind, the case for agent-based models rests mainly on their ability to represent and explain crises, rather than to predict them in advance. Financial-friction models and agent-based models are also beginning to cover some of the same ground. The important question may eventually be less about choosing between them and more about understanding which problems each approach is best suited to solving.

Footnotes

  1. Palgrave Macmillan, Queen Elizabeth II and the Economists (opens in a new tab), 2013. 2

  2. London School of Economics, FinFacts: Global Financial Crisis: Warnings of danger from 2001 and a Fed all bark but no bite - with monumental consequences (opens in a new tab), 25th November 2008.

  3. National Bureau of Economic Research, Where Modern Macroeconomics Went Wrong (opens in a new tab), September 2017. 2

  4. Santa Fe Institute Press, Complexity Economics: An Introduction (opens in a new tab), 2020. 2 3

  5. Network Law Review, Some Background to Complexity Economics (opens in a new tab), 25th July 2023.

  6. New York University, Stern School of Business, The Trouble With Macroeconomics (opens in a new tab), 14th September 2016.

  7. Elsevier, The financial accelerator in a quantitative business cycle framework (opens in a new tab), 1999.

  8. University of Chicago Press, Credit Cycles (opens in a new tab), April 1997.

  9. CEPREMAP, The role of financial frictions during the crisis: an estimated DSGE model (opens in a new tab), January 2014.

  10. University of Kent, Financial Frictions in Macroeconomic Models (opens in a new tab), December 2017.

  11. American Economic Association, Getting at Systemic Risk via an Agent-Based Model of the Housing Market (opens in a new tab), May 2012.

  12. Bank of England, Macroprudential policy in an agent-based model of the UK housing market (opens in a new tab), October 2016.

  13. Bank of England, The Dappled World (opens in a new tab), 10th November 2016.

  14. Nature, Systemic risk in banking ecosystems (opens in a new tab), 19th January 2011.

  15. arXiv, Physics and Financial Economics (1776-2014): Puzzles, Ising and Agent-Based models (opens in a new tab), 1st April 2014. 2

  16. arXiv, A Comparison of Economic Agent-Based Model Calibration Methods (opens in a new tab), 15th February 2019.

  17. American Economic Association, Agent-Based Modeling in Economics and Finance: Past, Present, and Future (opens in a new tab), March 2025.

  18. Bank of England, Agent-based modeling at central banks: recent developments and new challenges (opens in a new tab), February 2025.

  19. American Economic Association, Making a Breach: The Incorporation of Agent-Based Models into the Bank of England’s Toolkit (opens in a new tab), 2021.

  20. INET Oxford, INET Oxford and the Bank of England develop an agent-based model of the UK housing market (opens in a new tab), 13th October 2016.

  21. University of Amsterdam, New economic models give central banks better tools in times of crisis (opens in a new tab), 28th November 2025.

  22. Associated Press, Sorry Ma’am — we just didn’t see it coming (opens in a new tab), 26th July 2009.