|
PureBytes Links
Trading Reference Links
|
Hi Fred,
I'm looking forward to following along in Howard's forthcoming book
using AmiBroker. I really like his security selection chapter that
he previews and I'm excited to learn how to do that with AmiBroker.
> One other comment about NN's ... Although they may have a clue what
> they are doing, the algorithms that they develop based on the data
> you feed them are for the most part hidden ... This makes
utilization
> of them roughly equivalent to using a black box ... imho, not a
great
> idea, but to each his/her own ...
As you know but others may not, in most uses neural networks are
a form of fancy non-linear regression. The "equation" is hidden
because it is a large series of multiplications, summations and non-
linear functions that map input indicators to predicted output
signals. This complexity effectively results in a black box. You
are quite right. They, like everything, are merely a tool with their
own strengths and weaknesses.
> > Neural Nets ... Possibly but I have yet to see anyone come out
with
> a
> > product that generates a system that does well out of sample ...
This is because generally a NN is regressed on the past, "curve
fitted" as some call it (this is analogous to optimizing any trading
system on past data), validated on recent past, then the model is
most often used as-is going forward, but since the markets are non-
stationary, and if your input indicators are not robust to changing
markets, the NN equity curve climbs, eventually flattens, then
declines and one needs to remodel. Alternatively, one could walk-
forward train the model moving forward.
> > Dakota may be the exception
Dakota presently does not have any neural networks on-board, though
one could incorporate them.
> but because of things I've seen that
> > particular organization do in the past it is not likely I'd ever
> use
> > one of their products.
In the late '90s, early 2000's, I posted some NN based models on the
web and changed the models from time to time, somewhat based on the
effects above, somewhat to show the effects of different modeling
strategies, somewhat due to changes in available data (some symbols
that feed the models change or disappear on you altogether, the DM
via Equis with the advent of the Euro being one example) and I think
there was a misunderstanding about what I was showing. Probably my
error in presentation.
Tomasz has created a great product and I look forward to learning how
to use it to follow along with Howard's forthcoming book.
P.S. You've done some amazing work creating "IO". I read over your
docs real quick... impressive.
Thanks,
Carl
Content-Description: "AVG certification"
No virus found in this incoming message.
Checked by AVG Free Edition.
Version: 7.1.409 / Virus Database: 268.13.27/517 - Release Date: 11/3/2006
|