Friday, August 05, 2011


I went along to some talks at ICEC 2011 here at Liverpool, including an interesting invited talk by Robert May on Stability and Complexity in Banking Ecosystems, along with some others in the general area of agent-based computational economics (ACE). A lot of this research is experimental: you set up a virtual marketplace along with traders who compete in it, then “The modeler then steps back to observe the development of the system over time without further intervention” (to quote the wikipedia page).

Of course, current economic problems serve to motivate this topic a great deal and perhaps will cause it to attract a bigger share of research funds, over the next few years. Robert May's talk reported on some experiments involving models of interbank lending aimed at predicting the spread of bank failures, along with some basic (by the standards of most theory people, I would guess) calculations of how the probability of systemic failure is affected by the extent to which banks diversify their assets. Another talk that I went to, compared two markets in which traders bought and sold a single good: one version was a centralised market in which a single price was maintained, while the other version worked by bringing together pairs of traders at random, and a trade would go ahead if the trader with higher value, could afford to buy from the other one. The latter version does not perform “price discovery” so effectively, since trades can take place at the same time at different prices, but there’s apparently a sense in which it’s more stable. Also, incompetant traders get “weeded out” in the decentralised version since they get ripped off, while in the centralised version they can free-ride on the wisdom of the crowd, due to the single current price. You may or may not consider that to be a good thing. The trading strategies were produced by some kind of evolutionary algorithm in which the poorly-performing ones get eliminated while the good ones prevail (and maybe get mutated, I can’t remember).

I would conjecture that there is scope for the theory community to contribute to this research area, somewhat by analogy to Ingo Wegener’s work on genetic programming: genetic and evolutionary algorithms was purely experimental, and he and co-workers contributed the first proofs of performance guarantees for some of these. Such a contribution could help to guide experimental design in computational economics, or address concerns about the realism of the models used in simulations.

Here is a video of an earlier similar talk by Robert May.

No comments: