The Duck of Minerva

ABM in the Social Sciences

11 August 2009

While I am in the throes of designing and implementing an agent-based modeling approach to study how democracies react to extreme external shocks, I wanted to take a brief break from coding and writing to highlight two very interesting pieces in the current issue of Nature that address ABM directly. The first, “Economics: Meltdown modelling,” discusses how advanced agent-based models might be able to help predict future economic crashes—complete with a vignette where a futuristic ABM prevents a collapse. The problem, as the article asserts, is that ABM is often rejected by mainstream economists.

Many [economists] argue that agent-based models haven’t had the same level of testing…agent-based model of a market with many diverse players and a rich structure may contain many variable parameters. So even if its output matches reality, it’s not always clear if this is because of careful tuning of those parameters, or because the model succeeds in capturing realistic system dynamics. That leads many economists and social scientists to wonder whether any such model can be trusted. But agent-based enthusiasts counter that conventional economic models also contain many tunable parameters and are therefore subject to the same criticism.

This aversion to ABM is persistent throughout the social sciences, which creates an odd dynamic where ABM enthusiasts must often spend a great deal of time justifying their use before research can even begin. What’s baffling about this situation, however, is that ABM is just a tool; useful in for some research questions, but ultimately an imperfect device—just as nearly all other research methods in the social sciences are imperfect. This is precisely the sentiment of the authors of the second article, an op-ed entitled, “The economy needs agent-based modelling.” In discussing the current state of the art in analytical economic models the authors note:

The best models they have are of two types, both with fatal flaws. Type one is econometric: empirical statistical models that are fitted to past data. These successfully forecast a few quarters ahead as long as things stay more or less the same, but fail in the face of great change. Type two goes by the name of ‘dynamic stochastic general equilibrium’. These models assume a perfect world, and by their very nature rule out crises of the type we are experiencing now…As a result, economic policy-makers are basing their decisions on common sense, and on anecdotal analogies to previous crises such as Japan’s ‘lost decade’ or the Great Depression. The leaders of the world are flying the economy by the seat of their pants.

Why then, is ABM treated as being particularly fallible? As a user and developer I have pondered this many times. I believe the primary issue for many critics is the notion of “creating a universe for experimentation,” i.e. the belief that an ABM can account for all of the complexity. The easy response to such a critique is simple: no one believes that. My first exposure to ABM were zero intelligence agents, and I was struck by how such simple models could predict the dynamics of real markets (so much so, that I thought I might name a blog after them someday). Quality ABM’s focus on a narrow set of agent attributes, and attempt to glean the maximum insight from these simple mechanics. For a more philosophical response I will paraphrase the great econometrician Neal Beck in saying that, “all of statistics is a sub-field of theology.” That is, with any model we assume to know the “real truth,” but accept the inherent error and still attempt to build knowledge from the analytsis. ABM are no different, however, these models simply leverage a different technology and analytical framework to produce conclusions.

I welcome both critics and supporters of ABM to make the case for and against their use. It should be noted, however, that those railing against new technology often become victims of their own shortsightedness.

Photo: Nature