Tuesday, January 11, 2011

Agent based modelling - sensible when modelling machines?

Agent based modelling defines simple agents and lets them act and interact in a simulation. A common complaint is that this is stupid because people are more sophisticated than these agents are. However (cfr. last post), when the major share of trading is done by algorithmic software, then this objection loses force. Surely, software should be able to mimic software. In principle, it should even be possible to collect actual trading software and pit these against each other in a virtual and purely simulated market. In practice, I doubt firms would let their "proprietary" software out of their sight.

BTW - I'm not saying the guy mentioned below (LeBaron) has found true results or that this method is splendid. I don't know enough about it to make any such claims. But the method sounds interesting: Let agents with different hypotheses about their environment, different ways of using past info to distill patterns and predictions, compete in an evolving ecosystem where growth is related to past success - and see what aggregate outcomes, persistent regularities and interesting properties we get out of it.

Amplify’d from www.nature.com

'stability' is a word few would use to describe the chaotic markets of the past few years, when complex, nonlinear feedbacks fuelled the boom and bust of the dot-com and housing bubbles, and when banks took extreme risks in pursuit of ever higher profits.

In an effort to deal with such messy realities, a few economists — often working with physicists and others outside the economic mainstream — have spent the past decade or so exploring 'agent-based' models that make only minimal assumptions about human behaviour or inherent market stability (see page 685). The idea is to build a virtual market in a computer and populate it with artificially intelligent bits of software — 'agents' — that interact with one another much as people do in a real market. The computer then lets the overall behaviour of the market emerge from the actions of the individual agents, without presupposing the result.

Agent-based models have roots dating back to the 1940s and the first 'cellular automata', which were essentially just simulated grids of on–off switches that interacted with their nearest neighbours. But they didn't spark much interest beyond the physical-science community until the 1990s, when advances in computer power began to make realistic social simulations more feasible. Since then they have found increasing use in problems such as traffic flow and the spread of infectious diseases (see page 687). Indeed, points out Helbing, agent-based models are the social-science analogue of the computational simulations now routinely used elsewhere in science to explore complex nonlinear processes such as the global climate.

LeBaron has spent the past decade and a half working with colleagues, including a number of physicists, to develop an agent-based model of the stock market. In this model, several hundred agents attempt to profit by buying and selling stock, basing their decisions on patterns they perceive in past stock movements. Because the agents can learn from and respond to emerging market behaviour, they often shift their strategies, leading other agents to change their behaviour in turn. As a result, prices don't settle down into a stable equilibrium, as standard economic theory predicts. Much as in the real stock market, the prices keep bouncing up and down erratically, driven by an ever-shifting ecology of strategies and behaviours.

Nor is the resemblance just qualitative, says LeBaron. Detailed analyses of the agent-based model show that it reproduces the statistical features of real markets, especially their susceptibility to sudden, large price movements. "Traditional models do not go very far in explaining these features," LeBaron says.

Read more at www.nature.com