|
PureBytes Links
Trading Reference Links
|
Jeff and others,
On Wed, 7 Feb 2001 10:12:29 -0500 (EST), you wrote:
>You do make an important distinction about making "one's targets
>clear" and give the example of predicting turning points. I
>do know of some success with predicting turning points with NNs.
>
>I should have made the distinction that I don't know of any
>NN-based models that have done much better at price forecasting
>than you could obtain by a random walk model. I have
>a bit of firsthand knowledge of this, I worked for a quant
>firm at the CBOT in the late 80s... this is at a time when
>NNs were all the rage... and continued that way into the 90s.
>Folks were treating NNs as blackboxes that might instantly
>give them an edge. The rage gradually faded...
Any serious modeling task should have a clear target, but imo there is
more to consider:
NN is only a _modeling technique_ (one among many). Therefore, we
first have to have an adequate _model setup_ (e.g. in terms of input /
output parameters) for any problem to be modeled successfully.
GA is only an _optimization technique_ (one among many). So it can be
used to optimize e.g. the parameters of a neural model (like
structure, size, transfer functions, convergency criteria, etc).
Additionally, we have to turn our attention to the _generalization
aspect_, because we want only to pick-up features and patterns from
the past, that will also be valid for the future, at least to some
extent. Of course, here again are some helpful procedures, like
cross-validation, etc.
And only, if we use a valid _model setup_, employing an appropriate
_modeling technique_, carefully used in context with a powerful
_optimization technique_, and all done under consequent consideration
of _generalization aspects_, we possibly may be successful.
Imo, as a consequence, it does not make much sense to stick to
something like:" NNs are bad", or "GAs are powerful". It all depends
from the context, where and how these tools & techniques are used.
>My intent here is not to be overly-pessimistic or confrontational,
>but rather I want to point out that with NNs in particular
>(and I presume also with GAs) there is a lot of preprocessing
>of data and understanding of the underlying phenomenon necessary
>in order to get decent results.
Imo, its not necessary _to really understand_ the underlying
phenomenons, if we use "open-minded" modeling techniques, like e.g.
NNs. It is sufficient to have a valid model set-up (i.e. parameter
set) _and_ enough(!) valid examples from the past. Working this line,
we even can "discover" new knowledge and increase our understanding.
What imo could be much more important, is to do some basic tests on
the modeling problem to solve, to get some impression, if it might be
possible to be successful. The test procedures from nonlinear dynamic
modeling theory can be _very_ helpful here, especially if we work on
time series (e.g. prices). This way we can avoid to spend too much
work on modeling problems, which can be a-priori identified as truly
random, whereas other problems, which are said to be random often
offer nice chances for appropriate modeling.
mfg rudolf stricker
| Disclaimer: The views of this user are strictly his own.
|