So, I have been reading a lot of stuff about complexity, emergence, agent-based –modeling (ABM), and the like – getting by with a little help from my friends, HT: Roger Koppl. It’s all very intriguing. My underlying agenda is policy relevance, particularly in this crisis environment.
In my last post I raise the question: Can we be certain, as a logical matter, that if certain conditions obtain, certain definite types of outcomes will result from the free-market process? As I understand it, much of effort in ABM-type investigations is into this type of issue. Not so much into the “logic” as we usually understand it, as into the necessary outcomes of certain kinds of processes, such processes being modeled by the behavior of the virtual agents of the model. By investigating these process-outcome variations one hopes to get traction onto the kind of economic-policy one feels appropriate – for example, supporting less intervention, support for property-rights, etc. – or the opposite.
There are some things I find very attractive about ABM (understanding that this may sound naïve, being the reactions of an outsider to the methodology and someone arriving late to the party). In particular, I like two things:
One, the ability to handle large numbers of interactions (variables) without the need to for closure (equilibrium), save for sufficient information to compute the next iteration. The system is open-ended, in disequilibrium.
Second, it makes me think about the uncertainty conundrum/paradox – the problem of how to model uncertainty that is shared by the agent and the modeler. I think it leverages the distinction between real/radical uncertainty (which is of a categorical, open-ended, inarticulabe nature – things will happen that could not have been foreseen or even imagined) and the effects of this uncertainty, which are arguably limited to a set of known categories. The AB model models the effects of the uncertainty – not the uncertainty itself, right?
But I have some concerns that I am struggling to think through. The exercise seems to start from the conclusion. We observe a particular reality. So we wonder what kind of world might be producing this outcome and we try and put the essentials into a computer model. How do we know we are right? We see what comes out. If it is close, we are happy. If it is not, we tweak it until it is close, then we are happy. Where does this really get us? There are probably a very large number of agent-institution-profiles that will produce outcomes close enough to make us happy.
Of course, this is not so different, maybe not different at all, from what really goes on with business-as-usual empirical work in economics, based on classical probability theory and the estimation of structural parameters. There is the pretense of randomness, but, in reality, the researchers select for desirable results, which they report, and discard the undesirable ones. The upshot is a competition of models – often based on the same, or very similar data, and certainly on the same reality. Model choice then becomes a matter of credibility, prior inclination or something else; the “facts” never speak for themselves, and they don’t speak with one voice. In this regard ABM is really no different. The model choice here will depend on the plausibility of the model setup – the realism of the agent-behavior and institutional-framework modeled. I can see some potential for progress, as long as the number of outcomes from a particular setup-type is not too varied (as it is with the standard macro-empirical story). I guess time will tell.
In my next post I want to consider the question of action in disequilibrium and plausible Keynesian stories (ABMs and others) built around that. My punchline is that it all comes down to the burden of proof.