|
PureBytes Links
Trading Reference Links
|
Thank you to the list for Neural Networks answers.
It seems that neural network are not the holy grail.
I post some remarks of Ward System about their system NeuroShell trader.
>1)What I have seen is that most people I could talk to using neural nets
told
>me it took about 3 to 6 months of intensive work prior to getting results.
>It was more when I started looking seriously into technical analysis.
We pretty much agree with this. However, the biggest mistake people make is
that they look hard for ways to predict their favorite issue, which may not
be very predictable at all. We usually recommend looking for issues that
will predict well and trade them instead. If you do that you can often
reduce the time to be successful.
>2) Has Neuroshell trader a function to scan a lot of stocks :
>currently I have 4000 stocks data in several metastock folders classified
>by industry groups and I scan for specific patterns like indicators
>combinations and cross overs on a lot of stocks.
It is not really built for scanning, since it is building some pretty
sophisticated models. You can add several issues to a chart, and it will
apply the model in the chart to the issues you add. Then as you add data
each day, you can look at the new buy/sell signals (the Professional gives
you alerts). The number of issues you can have in a chart depends on how
much RAM you have since the complex models are held in memory until saved;
generally 5 to 30 issues reasonable. Of course you can have as many charts
as you want.
By the way, the Trader and Trader Professional are not currently intraday
systems. They process end of day data only. Later we will have an intraday
system.
>3) I wanted to know also if Neuroshell trader neural networks could be
>trained to detect head and shoulders, cup and handle, diamonds and other
>patterns like these (like Kasanjian Research's PatternSmasher say it does).
Many of these indicators are already built-in. If they aren't, you can
program them yourself in a DLL that the Trader Professional can call. The
neural net can then try to predict the value of an indicator you build. But
the neural net is usually used to predict the future value of an issue using
indicators as inputs.
>4) One function I am looking for is multiple time frames, i.e. detect a
>pattern on a weekly unit basis and detect another complementary in a daily
>basis, all in the same chart or exploration. With Metastock, I had to
>rewrite all indicators I wish to use on a weekly basis using a daily chart?
Our charts are either daily, weekly, or monthly. A given indicator will work
either in either a daily or weekly chart. However, if you have a daily
chart, the indicators cannot be told to work weekly unless you can rebuild
them yourself.
>5) I start reading some books on neural nets by Mark Jurik or Zirilli and I
>found that generally speaking, neural nets parameters preprocessing should
>be rather different than just input values of indicators in classic
>technical analysis, sya like rather use delta values or log or other
>preprocessed parameters.
We haven't found this to be true. Usually, standard indicators work best for
us and our customers.
>The complexity of using neural nets seems to be there whoosing the correct
number
>of parameters (not too much) and giving the correct preprocessed
information.
> What is necessary with Neuroshell trader ?
> Do you have a user forum, online suggestions for neural nets
models implementation? Do you have suggested readings on neural nets?
We agree with this. We don't have a forum yet, but we are building one. But
we have many online suggestions and good examples. The problem with reading
is that 1) If they were making money, they probably wouldn't be telling you
their indicators and 2) they may be trying to sell you their indicators. We
don't suggest trying to find out what someone else did, because they could
be successful today and their models don't work two months from now, and new
ones have to be built. There are no Holy Grail inputs.
>6) Do you carry a trial package?
We have a thirty day money back return period. We will refund the price of
the software (but not shipping costs) if you return it within 30 days. We do
get a few back, but they are mostly from people who are looking for a get
rich quick scheme, and who do not have the sophistication to build their own
systems. The Trader is best learned from the videos we supply, which we will
be mailing you. However, you sound very sophisticated; for you the best
thing is to order it, use it for 30 days, and feel free to return it if you
don't like it for any reason. But we think you will, judging from your
experience.
-----Original Message-----
From: owner-metastock@xxxxxxxxxxxxx
[mailto:owner-metastock@xxxxxxxxxxxxx]On Behalf Of Laurent GITTLER
Sent: Monday, July 12, 1999 5:38 PM
To: 'Metastock@xxxxxxxxxxxxx'
Subject: Neural networks
I read a book by Zirrilli about neural networks and I found it great. Also
Mark Jurik says some customers were having great returns using this
technology. So far I don't know whether it can be be profitable or not.
I was wondering if anybody has tested Neuroshell trader which combines
Neural networks with Charting software. They claim to have 700 indicators.
I have ssen a lot of people saying it is overe optimized for the past and
that the market is changing, but as I see the thread about optimal f, I am
not so sure it is not a good way to go into the market using portofolio
management rules.
I don't know if any software does this, but normally neural networks should
be great tools to detect Head and Shoulders, cup and handles and so on,
neural networks act mainly as a fuzzy optimized filter. It has great
performance on OCR (Optical Character Recognition), when you tell neural
networks whether it is the good pattern or not as you give character images
on the neural network.
Same thing for Virus, I know that neural networks gave great results
detecting virus, with an accuracy of more than 99%, including virus which
were not detected yet, by analogy. The problem was to give the proper input
to them, and not the virus code.
So I was also wondering if someone has tried to use and train neural
networks this way.
Laurent Gittler
|