Re: Trade Software

Brad Wallet KF4UYM (bwallet@osf1.gmu.edu)
Fri, 16 Jan 1998 22:12:31 +0100 (MET)

On Fri, 19 Dec 1997, Jerry Whitnell wrote:

> In article <Pine.OSF.3.95q.971218104012.9153A-100000@osf1.gmu.edu>, Brad
> Wallet KF4UYM <bwallet@osf1.gmu.edu> wrote:
>
> > On Wed, 17 Dec 1997, Marc Champigny wrote:
> >
> > > Hello everyone,
> > >=20
> > > I=92m a graduate student in developmental biology who is interested in
> > > acquiring neural network/ genetic algorithm/ fuzzy logic software for
> > > the purpose of futures trading.
> >
> > You should probably begin by learning a bit about economics. In
> > particular, I would research the Theory of Market Efficiency. Then,
> > get a hold of a statistics text and learn about random walks. After that,
> > you should probably consider a new project.=20
>
> "Listen to the experts. They'll tell you why it can't be done. Then do it."
>
> Heinlien (probably misquoted).

If you can't explain what you are doing in a theoretical sense, you should
be uncomfortable. If what you are trying to do perpendicular to the
theory, you should be very afraid. That's as simple a matter of life as
you can get.

The fact of the matter is that many people doing data mining/evolutionary
computation/neural nets/etc feel they can simply skip learning about the
problem area and just throw their method au jour at the problem without
understanding the theory, the assumptions, or the risks. Such individuals
clearly play the role of fool. Yes, these are very flexible tools, and
they allow you to attack a broad range of problems, but if you ignore the
problem domain, you ask the wrong questions, solve the wrong problems, and
generally do boneheaded things.

For example, I feel like crying everytime I see someone attempting to do
discriminant analysis in high dimensional space. Yes, there are good
reasons why Bellman's "Curse of Dimensionality" may not be relevant, and
I've seen some pretty convincing articles that make one think things
aren't as we had thought. But, one cannot simply ignore the theoretical
problems involved in the sparseness of high dimensional space---one must
understand why it may be problematic.

While I am increasingly convinced that discriminant analysis in high
dimensional space is not a lost cause, I am increasingly weary of
dimensionality reduction methods. Specifically, "mining" of high
dimensional data is done at great peril. I've done a number of
calculation, and I think we need to be very, very worried about the
generally ignored beast called selection bias. It is out there, it is
hard to calculate, it is almost uniformly ignored, and it is very
significant. People doing subset selection for regression in the
statistics community have known about it for year (Fogel, et al, I am
sorry that I failed to respond to earlier queries on comments by me on
this matter. Soon, I promise.). Such statistician may not have been able
to calculate such things. In fact, Diehr and Hoflin (1974), called the
distribution "hopelessly complex." Still, they were aware of it, and they
tried. Because of this, some progress was made including such things as
ridge regression. At the very least, they were able to characterize the
problem using simulation (Rencher and Pun, 1990). Our results are invalid
if we ignore the spector of selection bias.

> I've been studying this same problem and have several references to
> studies that have used neural nets and genetic algorithem to trade the
> markets. One includes a review of market efficiency theory and concludes
> that it is not true for all markets at all times (email me for reference,

Several references, and only one mentions the theory of market efficiency?
I think we should all be very troubled by this. And, yes, it does not
hold for all markets at all times, but we need to be aware of what
assumptions it makes. Publishing your method makes the theory's
assumptions valid.

Brad