Forums  > Trading  > Multi Agent Research  
     
Page 1 of 1
Display using:  

opmtrader
Founding Member

Total Posts: 1333
Joined: Mar 2004
 
Posted: 2005-02-26 04:23
Researchers and traders are using multiple simulated traders to model, study, and forecast the market. This stuff is fascinating to me primarily because I think these individuals are honing into the heart of the market dynamic. These dynamics may manifest themselves in the linear behaviors people such as myself may be able to pick up on and exploit, but unlike us rigid stat based traders, these guys may be able to pick up on the more fluid dynamics going on and predict when all those nice stats may be ready to be thrown out the window.

Here are some interesting links:

http://www.lincoln.ox.ac.uk/fellows/johnson/
http://www.trnmag.com/Stories/2001/082201/Unusual_calms_tell_of_coming_storms_082201.html

And some interesting quotes I picked up along the way:

"Complex systems are epitomised by the occurrence of large changes, or so-called 'extreme events': such extreme events arise far more often than would be expected if the individual agents acted independently - hence they represent moments of global correlation, or cooperativity, in the system. From an academic point of view any 'theory' of complex systems must incorporate an understanding of such extreme events. From a practical point of view, such extreme events are the most important feature of such a system since they tend to dominate the future behaviour of the system in question - they also give rise to the 'fat tails' which are ubiquitous in empirical studies of complex systems. In short, extreme events are rare, but their effects can be catastrophic and long-lasting. The major challenge facing researchers in such Complex Systems is the real-world validation of any theoretical ideas or concepts developed. Our desire to establish an empirical test-bed for our theoretical ideas, led to my group's involvement in setting up a unique Complex Systems 'lab' at Oxford University focusing on financial markets (see www.occf.ox.ac.uk). Why financial markets? Because they are highly data-rich and continuously evolving, and are widely considered to be excellent examples of real-world Complex Systems with traders representing the agents who repeatedly compete for a limited resource in this global 'many-body game'. The underlying rationale is that the fluctuations observed in financial time-series should, at some level, reflect the interactions, feedback, frustration and adaptation of the markets' many (and diverse) participants: i.e. financial markets can be viewed as populations of agents, with limited information, who are repeatedly competing for a limited resource. In short, the financial markets are just one big 'game'."

"The researchers found that the level of predictability goes up just before a large change. "Strategy selection amongst agents seems to converge -- it's kind of a crowd effect," said Neil Johnson, a lecturer in physics at Oxford University. "Without communicating, or even knowing the existence of each other, agents begin to lock into a particular behavior" just before a large change, he said."

Predictability also has to do with how stable a system is, said Johnson. "Take, for example, commuter traffic. Usually everyone drives differently, at slightly different speeds and with slightly different strategies -- this gives the traffic flow some kind of stability. If everyone is driving at the same speed, in the same way, with the same distance between, then this is unstable," he said. This is because if something happens to one car, it is more likely to cause a chain reaction. "[If] something appears on the road, it will cause a massive disruption," he said.

"When the forces become unbalanced, an avalanche-like effect happens "whereby small glimpses of a pattern momentarily emerge and, by chance, become amplified," said Johnson. It is during this amplification period that the predictability of the system grows. "The agents start taking up definite positions -- the two opposing forces are now momentarily out of balance. After about 10 hours the imbalance begins to show itself as a definite trend" in market prices, Johnson said."

opmtrader
Founding Member

Total Posts: 1333
Joined: Mar 2004
 
Posted: 2005-02-26 04:35
So to facilitate discussion I have a few questions.

1) I have not read much of the primary research in this field, however I have some idea of how it works. Besides the dynamics of systems at extremes, what other insights has it generally uncovered?

2) Are these models decent for predicting shorter term moves? In other words are these applicable for continuous trading or are they currently crash alert applications?

3) Has anyone here seen, used, or developed a system such as this? Perhaps some of Oxford's finest will comment on the research lab in those articles. I will fully understand if you are not interested in speaking about it.

YukaRedux
Now with added evil

Total Posts: 650
Joined: Dec 2004
 
Posted: 2005-02-26 04:42


3) Grinham in Australia are (or certainly were) using agent-based systems to trade - and very successfully by all accounts.



9 x 9 = 82

opmtrader
Founding Member

Total Posts: 1333
Joined: Mar 2004
 
Posted: 2005-02-26 05:23
Thanks YukaRedux. Anywhere I can look at anything about their organization? A website perhaps. That name is new to me.

YukaRedux
Now with added evil

Total Posts: 650
Joined: Dec 2004
 
Posted: 2005-02-26 05:38

Hmm, their website www.gmf.com.au doesn't seem to be around any more....

I'll see what I can dig up.



9 x 9 = 82

opmtrader
Founding Member

Total Posts: 1333
Joined: Mar 2004
 
Posted: 2005-02-26 06:12
I used my time machine and got this:

http://web.archive.org/web/20030408053642/http://www.gmf.com.au/

Maybe they are like RenTech and took their website down to increase their privacy.

silverside


Total Posts: 1411
Joined: Jun 2004
 
Posted: 2005-02-26 10:58
This is also very interesting

Sornette's latest paper (or just google for sornette)

it's a lot more concrete than his books - less of the voodoo physics. Basically he reckons you can find 'prediction days' in the NASDAQ based on agent models, he gives a few details, but not all. The main thing i'd like to know if he can go from these days to a trading strategy, backtest it and evaluate it's risk/reward.

I'd love to be able to play around with these models as they seem reasonable, and not *that* complicated.


JamesH83


Total Posts: 697
Joined: Aug 2004
 
Posted: 2005-02-26 12:40
Hi OPM,

I have been interested in this stuff for a while. (Wanted to write my MPhil thesis on it). I read most of Sornettes stuff. There are ofcourse various good papers from the Sante Fe Institute. I was quite interested in the econometrics side of things. Pesaran has done a bit of work on chaotic attractor modelling in econometrics.

This book although a little old is excellent.

While this is a facinating subject, it seems to me that there are easier ways to make money. Yuka, I'd be interested to hear more about Grinham or any other success stories.

James

¦(X)=(Nh)-1K{h-1(X-x1) + ... + h-1(X-xT)},

MoreLiver


Total Posts: 481
Joined: Dec 2004
 
Posted: 2005-02-26 12:41
complex is short like lorenz equation or hurst or luanopov have been proof earlier than we had diapers

Hard for many to understand inflation, but not for me.

opmtrader
Founding Member

Total Posts: 1333
Joined: Mar 2004
 
Posted: 2005-02-26 18:03
silverside, I nearly mentioned Sornette in my first post as the idea that "agents begin to lock into a particular behavior just before a large change" seems to be supported by his work. Now I must read this new paper you've so kindly provided.

JamesH83, I think you are definitely right that there are easier ways to make money. My idea is that these mysterious complex behaviors inevitably throw off data which can be analyzed by linear methods. For example I imagine what Sornette is calling a "prediction day" is what us troglodytic statisticians would call a "trend day".

Holmes


Total Posts: 185
Joined: Dec 2004
 
Posted: 2005-02-26 19:48
i was lucky enought to sit johnsons course

i doubt he or anyone in that group is making money from the thing though. i get the impression they are alot more interested in doing research basically because of the vast amounts of cheap data markets provide: he talked about robots worcking on mars in the same sentences as traders acting as agents

email him! im pretty sure he'd be interested! he is a nice guy


his group: i think i read somewhere they have a model they want to sell or something

also he seemed to believe quite strongly that chaos was a load of nonsense. i THINK that lot of the 'agents' strategies are quite probabilistic

johnson et al have written a book on this, OUP i think
'financial markey complexity', it is basically his lecture notes.

i enjoyed the course, and the book is quite well written

hope that helps

opmtrader
Founding Member

Total Posts: 1333
Joined: Mar 2004
 
Posted: 2005-02-26 20:13
Wow, I just read that paper. As always I wish I understood more of the mechanics of it. I used to swap emails with Didier back in the day but I decided to let him off the hook. He's too busy to be trying to teach me the nuances of his work. He's a good guy though.

In any case, from what I've understood here, is that these agents essentially make follow/fade decisions based on the directions of the past. Would that be correct? Are these past directions based on the global direction or simply by the direction of the individual agents choices? What is the three stage count all about? He shows 001,010,111, etc. Is each time step 1 day? If each agent is rule based, how are they able to all independently choose a similar direction for the next day more often than randomness would suggest? I can see how that would be possible if the choices were predicated on the global state but not on static individual rules.

As you can see I am a bit hazy on the specifics. I will probably need to reread the paper a few more times. Its just more fun to get a discussion going about it.

In any case I'd think Mr. Sornette is doing well for himself these days. If his model can predict the direction of the next day 67% of the time on 39 occurances out of a stretch of 61 days then I'd say he's doing all right.

opmtrader
Founding Member

Total Posts: 1333
Joined: Mar 2004
 
Posted: 2005-02-26 20:16
Thank you Holmes. I may have to buy that book. Maybe I'll even email him after reading it. As you can see from the post below, a lot of this is beyond me right now but it can be very useful to conceptualize how the market works in advanced ways even if one can't exploit it currently.

JamesH83


Total Posts: 697
Joined: Aug 2004
 
Posted: 2005-02-27 04:01
I swapped emails with Lo for a while....about Didier plagarising his work.....

¦(X)=(Nh)-1K{h-1(X-x1) + ... + h-1(X-xT)},

Holmes


Total Posts: 185
Joined: Dec 2004
 
Posted: 2005-03-03 22:43
opmtrader,

this thread got me thinking back a couple of years before i started having nightmares about quarks

have you been to:

http://www.unifr.ch/econophysics/ ?

opmtrader
Founding Member

Total Posts: 1333
Joined: Mar 2004
 
Posted: 2005-03-04 03:27
Thank you Holmes. I have not been to that site. Seems like a wealth of good papers there which I should read sometime.

I just want to emphasize that for me this is just a little brain exercise. I likely don't have the time or the wherewithal to create and implement these models.

To quote a very wise NP member "Tis better to make a living on something simple to support more complex projects than to work on complex things and starve". That NP member can claim credit for his insight here if he wishes.
Previous Thread :: Next Thread 
Page 1 of 1