Wednesday, December 08, 2010

Building a Computational Model of the Crisis

A team of four researchers affiliated with the Santa Fe Institute has secured a grant from the Institute for New Economic Thinking to fund the development of an agent-based computational model of the financial crisis. The model will explicitly consider "housing and mortgage markets, banks and other financial institutions, securitization processes and hedge fund investors, manufacturing and service firms, and regulatory agencies," with the goal of discovering "the essential elements needed to reproduce the crisis, while investigating alternative policies that may have reduced its intensity and strategies for recovery."

It's an interesting and multidisciplinary group, composed of Doyne Farmer, John Geanakoplos, Peter Howitt and Robert Axtell. Genakoplos and Howitt are two of the most creative economists around, and I have discussed the work of the former on leverage and the latter on learning in earlier posts. Axtell is the co-author (with Joshua Epstein) of a fascinating book called Growing Artificial Societies, in which they develop an elaborate computational model of the interaction between a renewable resource base and the human population that depends on it. The model reproduces spatial patterns of resource depletion and recovery as well as population growth, migration and decline. Farmer is a physicist by training but has been working on finance for as long as I can remember. I discussed some of his work in an earlier post making a case for greater methodological pluralism in economics in general, and agent-based modeling in particular.

The team is looking for a graduate student or postdoctoral fellow to join them for a couple of years. For a young researcher interested in finance, the microfoundations of macroeconomics, and the agent-based computational methodology, this could be a fantastic opportunity.


  1. The Internet Age!
    It is entirely possible to get a broad sample of folks to give an online diary of what they did, week by week in 2007 or 2008. Run their actual activities in a simulation, the actual agent, not some model.

  2. Matt, the agents have to be programmed with behavioral rules that are flexible enough to respond to a broad range of situations, not just what they faced in 2007-8. Otherwise you can't explore counterfactuals. Diaries might give you clues about the rules people were plausibly following but they are just clues. The point is to see how these rules, coupled with the interaction structure and the regulatory environment, determine economic outcomes. Then one can explore the implications of alternative policy interventions or regulatory regimes.

  3. As long as the agents postulated are not averse to collusive fraud, this could work.