Forums  > Pricing & Modelling  > Agent based models  
     
Page 1 of 1
Display using:  

DrGrumpy


Total Posts: 136
Joined: Nov 2008
 
Posted: 2017-10-12 08:28
I'm interested in looking at using agents to simulate a multi-asset world (more than 2!).

I've seen there are a few links (1, 2, 3, 4) here but everything is rather old.

Does anyone still do this or has it fallen from favour? Can anyone point me to any recent work in the field?

As this is a bit of a "fun in some free time" exercise, rather than mission critical it would be awesome if someone's already done some of the heavy lifting in coding the obvious agents... Wink

A girl goes into a bar and asks for an example of double entendre, so the barman gives her one.

billWalker


Total Posts: 172
Joined: Feb 2005
 
Posted: 2017-10-12 16:00
I think the consensus is that agent-based models can produce behavior that looks like the real world, but that due to extreme sensitivity to initial conditions (itself due to the iterative nature of the simulations), they are no good for empirical inference. That might be why there has not been a lot of work done.

One of the links you provide mentions the Santa Fe Artificial Stock Market, and I think most work on this stuff is still being done by W Brian Arthur at the Santa Fe Institute. He published Complexity and the Economy a few years ago, which is a collection of his papers on the subject, including the original agent-based modelling paper that spawned the Santa Fe Artificial Stock Market. I found nothing in there that would be directly applicable for a PM, though. His interest moved on to the application of agent-based modeling to the spread of innovation in economies.

Complexity Explorer at the Santa Fe Institute has some MOOCs on agent-based modeling.

This is a hobby-level interest of mine, as well, as I believe agents *are* the correct model for price movements. I've come to the conclusion that the problem is probably just too hard, however. I'm not optimistic that deep learning or whatever quantum fairy dust arises will be any better able to solve it, again because of the importance of initial conditions in an iterative model. I think it's more complicated than non-linear classification.

That said, perhaps people in HFT are doing interesting stuff with these.

I've been meaning to read Barabasi's Network Science as my next step in trying to understand all this. Has anybody had any experience with that?

"Plausible regularities may be present but swamped by changes in attendant circumstances." Ole Peters

DrGrumpy


Total Posts: 136
Joined: Nov 2008
 
Posted: 2017-10-13 06:14
Thanks for the thoughts...

I guess your point on the sensitivity to initial conditions is the nail in the coffin. There's not even many papers referencing the Santa Fe work, and only a handful in the last year or so.

FWIW, there's a Swarm-Like Agent Protocol in Python (SLAPP). Might be a useful starting point if someone wants to do some coding.

A girl goes into a bar and asks for an example of double entendre, so the barman gives her one.

rod


Total Posts: 375
Joined: Nov 2006
 
Posted: 2017-10-19 14:29
Jean-Philippe Bouchaud of Capital Fund Management is advocating the use of agent-based modeling by central banks:

The workhorse models of central bankers need a radical change
Previous Thread :: Next Thread 
Page 1 of 1