To make sense of complex systems, send in the agents

Standard quant models cannot comprehend a radically complex reality, writes Jean-Phillippe Bouchaud

Standard quant models cannot comprehend a radically complex reality, writes Jean-Phillippe Bouchaud

Good science strives to deliver accurate, testable predictions. Economists have tried to conform to this same standard when forecasting everything from GDP growth to inflation and exchange rates. This is no easy task. The economy is a complex system, populated by a large number of heterogeneous, interactive agents of different categories and sizes. In such environments, even qualitative predictions are hard. So, perhaps economists should abandon the pretence of exactitude and adopt a different method, one based on scenario identification and agent-based modelling.   

One of the biggest challenges in economics is modelling the emergent organisation, co-operation and co-ordination of the motley crowd of micro-units that comprise the economy. Most macro models use ‘representative’ agents, effectively treating each unique business or household as a facsimile of all the others. This assumption throws the baby out with the bathwater. But capturing and characterising the emergent properties of micro-units in a system is difficult. And micro-observations do not always translate to the macro level. One well-known example is the Schelling model, which reveals that even when all agents prefer to live in mixed neighbourhoods, myopic dynamics can lead to completely segregated communities.

In this case, Adam Smith’s invisible hand badly fails.

More generally, slight differences in micro rules and parameters can upend macro outcomes. This is the idea of ‘phase transitions’, where a slight change in a parameter leads to sudden discontinuities and crises. Feedback loops, heterogeneities and non-linearities mean these surprises are hard – if not impossible – to imagine or anticipate, even with the aid of the best mathematical apparatus.

This is what I like to call ‘radical complexity’. Simple models can lead to unknowable behaviour. Black swans or ‘unknown unknowns’ can be present even if all the rules of the model are known in detail. In conventional economic models, even probabilities are hard to pin down, and rationality is de facto limited. For example, the probability of rare events can be exponentially sensitive to the model parameters, and therefore unknowable in practice. In these circumstances, precise quantitative predictions are nearly impossible.

But this does not imply the demise of the scientific method. Faced with a radically complex reality, economists should adopt a more qualitative, scenario-based approach that emphasises mechanisms and feedback loops, rather than reaching for precise but misleading forecasts based on unrealistic assumptions. This is actually the path taken, for example, by modern climate change science.

Creating a world and seeing how it unfolds has tremendous pedagogical merits

The first step is to establish a list of possible or plausible scenarios. This can be done with numerical simulations of agent-based models (ABMs). While experimenting on large-scale human systems is still cumbersome (but becoming easier with the use of web-based protocols), running ABM experiments in silico is relatively easy. Such experiments reveal all sorts of unexpected phenomena and can elicit scenarios that would be nearly impossible to imagine, given the feedback loops and non-linearities involved. Think, for example, of the spontaneous synchronisation of fireflies, which took nearly 70 years to explain. Complex endogenous dynamics are pervasive. But they are hard to see without the appropriate tools.

Experimenting with ABMs is rewarding on many counts. One hugely important aspect, in my view, is that it allows students to learn about how complex social and economic systems work in a playful and engaging way. Such simulations foster their intuition and imagination, in the same way lab experiments train physicists to think about the real world, beyond abstract mathematical formalism.

A versatile tool

Creating a world and seeing how it unfolds has tremendous pedagogical merits. It is also an intellectual exercise of genuine value: if we cannot make sense of emergent phenomenon in a world where we set all the rules, how can we expect to comprehend the real world? These experiments can train the mind to grasp these collective phenomena and understand how and why some scenarios can materialise when others do not. The versatility of ABMs can accommodate ingredients that are almost impossible to include in classical economic models, and explore their impact on the dynamics of the systems.

ABMs are often spurned because they are generally hard to calibrate. Therefore, the numbers they spit out cannot – and should not – be taken at face value. They should rather be regarded as all-purpose scenario generators. ABMs can reveal hidden phenomena, uncover different possibilities and reduce the realm of Black Swans. The latter are often the result of a lack of imagination and the simplicity of models, rather than being inherently impossible to foresee. And when viewed through this lens, swans that appear black to the myopic eye may in fact be generically white.

Experimenting with toy-models of economic complexity will create a useful corpus of scenario-based, qualitative macroeconomics. Rather than aiming for precise numerical predictions based on unrealistic assumptions, economists should strive to build models that rely on plausible causal mechanisms and encompass all plausible scenarios, even when these scenarios cannot be fully characterised mathematically. A qualitative approach to the complexity of economics should be high on the research agenda. As Keynes said, it is better to be roughly right than exactly wrong.

Jean-Phillipe Bouchaud is chairman of Capital Fund Management and member of the Académie des Sciences

Only users who have a paid subscription or are part of a corporate subscription are able to print or copy content.

To access these options, along with all other subscription benefits, please contact info@risk.net or view our subscription options here: http://subscriptions.risk.net/subscribe

You are currently unable to copy this content. Please contact info@risk.net to find out more.

Digging deeper into deep hedging

Dynamic techniques and gen-AI simulated data can push the limits of deep hedging even further, as derivatives guru John Hull and colleagues explain

You need to sign in to use this feature. If you don’t have a Risk.net account, please register for a trial.

Sign in
You are currently on corporate access.

To use this feature you will need an individual account. If you have one already please sign in.

Sign in.

Alternatively you can request an individual account here