|
PureBytes Links
Trading Reference Links
|
I think it depends on how weights are adjusted during training.
For example, often it is better to randomly pre-set the weights prior to
training runs. This helps to kick the network out of local minima. If you
have very different results from retraining, you may have too much
dimensionality (I hope that is the right term) in your network solution, or
maybe you are just grinding your network deeper into different local minima.
But from what I remember, retraining a network should only be done to
determine these types of things, not for creating the final product. It felt
like an art; touch it too much in the wrong way, and it would get worse.
Just some thoughts.
----- Original Message -----
From: <TSOKAKIS@xxxxxxxxx>
To: <amibroker@xxxxxxxxxxxxxxx>
Sent: Monday, February 17, 2003 2:13 AM
Subject: [amibroker] Re: Training the past
> Fred,
> the market will not change in one night. Nobody can erase 3 bearish
> years at once.
> It is not the same as 99. Many stocks were over-overbought at the end
> of this period.
> We do not see [yet] over-oversold stocks. Many p/e are still high.
> As for the NN thesis, you may read it and then talk some points, if
> it is interesting for the list.[BTW, the training period was in the
> bullish era...]
> I noticed some usual mistakes in the text. One of them is the total
> absense of multiple training.
> Since you wrote that you have some experience from NN, you should
> know this usual NN trap: When you re-train NN you do not get the same
> results !!
> I have asked some NN vendors to explain this fact, but,
> unfortunately, no reply yet [perhaps they... train their replies too]
> If I do not have some replies to this basic NN question, I will not
> pay closer attention to this [black box] method.
> Dimitris Tsokakis
> --- In amibroker@xxxxxxxxxxxxxxx, "Fred <fctonetti@xxxx>"
> <fctonetti@xxxx> wrote:
> > Given that as described it was written for and tested on some
> segment
> > of the bear market would one really expect it to continue to work
> > when bear becomes bull ? What happens if the bear morphs into bull
> > tomorrow ? In order to have a clue I suggest you test it on an out
> > of sample for a couple of years prior to 2000. The "real world" as
> > you referred to it clearly takes in more then the last 3 years.
> This
> > bear won't last forever.
> >
> > --- In amibroker@xxxxxxxxxxxxxxx, "Dimitris Tsokakis"
> <TSOKAKIS@xxxx>
> > wrote:
> > > An interesting ref was at
> > > http://groups.yahoo.com/group/amibroker/message/33830
> > > thanks to jnk1997 research.
> > > Lieutenant Jason Kutsurelis thesis gives many details about the 5-
> > year training period.
> > > He does not mention how long was the tested period AFTER the NN
> > training.
> > > If NN is to approximate the past, then there are simpler methods.
> > > I have an excellent oscillator since Jan2000.
> > > It touches gently the edges of its horizontal channel, when the
> > main N100 changes occur.
> > > Sometimes anticipating and, in general, with insignificant lag.
> > > It gives nice 6/6/0 results with a simple up/down level system.
> > > Does any body know the future behavior of this fine oscillator ?
> > > Certainly no !!
> > > We may only hope to see a repetition, statistics is on our side.
> > > 2 years later, another statistical project will explain, AFTER
> THE
> > REAL FACTS, if the oscillator was good or not.
> > > We are in the wild world of real trading, not in a university
> > classroom.
> > > Dimitris Tsokakis
>
>
> Post AmiQuote-related messages ONLY to: amiquote@xxxxxxxxxxxxxxx
> (Web page: http://groups.yahoo.com/group/amiquote/messages/)
>
> Check group FAQ at:
http://groups.yahoo.com/group/amibroker/files/groupfaq.html
>
> Your use of Yahoo! Groups is subject to http://docs.yahoo.com/info/terms/
>
>
>
Post AmiQuote-related messages ONLY to: amiquote@xxxxxxxxxxxxxxx
(Web page: http://groups.yahoo.com/group/amiquote/messages/)
Check group FAQ at: http://groups.yahoo.com/group/amibroker/files/groupfaq.html
Your use of Yahoo! Groups is subject to http://docs.yahoo.com/info/terms/
|