Emergence is not a metaphor

Most quantitative models of an economy, a market or an institution begin with an act of compression: replace a population with its average. It is a reasonable simplification and it works well over the range where behaviour is roughly linear and the population is roughly homogeneous. It fails precisely where the questions get expensive.

The averaging problem

Consider a portfolio of institutions with different leverage. An aggregate model holds average leverage and asks what a shock does to it. But a forced sale is not triggered by average leverage — it is triggered by an individual institution crossing an individual threshold. The distribution is not decoration around the mean; in this case it is the entire mechanism. The average institution never sells. The tail does, and then the price it sells into becomes everyone else's mark.

This is the general shape of the failure. Wherever a system contains thresholds, constraints, or feedback from individual actions to a shared price or a shared resource, aggregate dynamics are not a summary of individual dynamics. They are a different thing.

What agent-based modelling actually changes

An agent-based model does not aggregate first. It represents the participants — heterogeneous, boundedly rational, adapting — lets them interact under explicit rules, and observes what the aggregate does as a consequence. The macro pattern is an output, not an input.

• Heterogeneity is preserved rather than assumed away
• Interaction is local, so structure and network position matter
• Behaviour can be adaptive, so responses change as conditions do
• Aggregate outcomes are generated, so they can be traced back to the interactions that produced them

That last property is the one that matters for anyone who has to defend a result. When an aggregate model produces a surprising number, you can inspect the equation. When a simulation produces one, you can inspect the run: which agents acted, in what order, and what each of them was responding to.

The cost, stated honestly

Agent-based models are harder to specify, harder to calibrate and harder to validate than the equation-based alternative. They have more degrees of freedom, which means more opportunity to fit the past and more responsibility to prove you have not merely done so. Anyone selling simulation without acknowledging this is selling something else.

The choice is not between a simple model and a complex one. It is between a model whose simplifications are visible and one whose simplifications are structural.

The reasonable position is selective. Use aggregate models where behaviour is smooth and the population is well described by its mean, which is most of the time. Reach for simulation when the question is about a transition, a threshold, a distributional effect, or an adaptation — the cases where the average is not the mechanism.