Hello,

NN is a fine tool for extracting various pattern. Of course that
requires the existence of a pattern. But what to do if patterns are
rarely found or if there is a very blurred pattern structure.

I think this applies to data whith a considerably random-like origin
(for example: stock markets, results of sportevents).

In this context: Wich kind of neural Network (MLP, RBF,  ...) should
be used and what should be avoided?

Alea

On Sep 24, 4:05=A0pm, p...@couptreffer.de wrote:
> Hello,
>
> NN is a fine tool for extracting various pattern. Of course that
> requires the existence of a pattern. But what to do if patterns are
> rarely found or if there is a very blurred pattern structure.
>
> I think this applies to data whith a considerably random-like origin
> (for example: stock markets, results of sportevents).
>
> In this context: Wich kind of neural Network (MLP, RBF, =A0...) should
> be used and what should be avoided?

NNs are effective methods for approximating an underlying
I/O relationship given enough diverse training samples with
a sufficient SNR.
Both MLP and  RBF are universal approximators. The easiest
to implement depends on the problem.

The most successful nets use a subset of well chosen
and preprocessed inputs.

Black Box approaches with little aforethought typically
fail as the I/O relationships become more complicated.

A thorugh data analysis which may include linear models,
PCA and/or clustering may be the best way to make
decisions about the most appropriate inputs and type of
NN to use.

Hope this helps.

Greg

On Sep 24, 9:05=A0pm, p...@couptreffer.de wrote:
> Hello,
>
> NN is a fine tool for extracting various pattern. Of course that
> requires the existence of a pattern. But what to do if patterns are
> rarely found or if there is a very blurred pattern structure.
>
> I think this applies to data whith a considerably random-like origin
> (for example: stock markets, results of sportevents).
>
> In this context: Wich kind of neural Network (MLP, RBF, =A0...) should
> be used and what should be avoided?
>
> Alea

With random data the normal approach is to place all the stock prices
into a vector and construct a matrix with the stock market prices
against time. The next thing you do is to construct a correlation
matrix of :-

S(i)T(j-k) where s(i) is stock i and j-k represents a time difference.
You then invert thir resultant matrix.

You have a maximum time difference of j-k and you can condition this
data by taking the first ten time differences and then combining times
for longer time differences. This is quite computationally intensive,
but it is basically what pure numerical stock prediction is all about.


  - Ian Parker

Hello,
thank you both for your replies.

@ Greg
I understand that the right choise of the input is most important.
But if i choose the input with the most information gain (WEKA is a
good tool for that) the result of the NN-run is pretty trivial:
Always bet on the favourite.

@ Ian P.
Thank you for the insight of computerized stock picking.
But if that strategy is initially successfull and known to the public,
it will have an impact on the stock price. This will tremendously
reduce the advantage.

Alea

On Sep 28, 8:49=A0pm, Alea <p...@couptreffer.de> wrote:

> @ Ian P.
> Thank you for the insight of computerized stock picking.
> But if that strategy is initially successfull and known to the public,
> it will have an impact on the stock price. This will tremendously
> reduce the advantage.
>
This is very much a $64,000 (or should I say $64 trillion) question.
When everyone is using computers to predict stocks we are then in a
situation where we can analyse stability. We all know what happened
with the sub-prime market. Confidence was lost and this was
infectious. Obama BTW has made what might be termed a "killing" on the
government purchases of bank stocks.

We can look at the matrices that are coming out and derive definite
stability criteria. A system is STABLE when all the eigenvalues are
stable (definition).

I believe along with other citizens (I am British BTW) that regulation
is required to prevent a repeat of the present crisis. All discussion
though has been concentrated on greed and the size of banker's
bonuses. There has been no discussion of AI or intrinsic stability. In
the future this will be of vital importance.

BTW - What is the value of bankers? Should AI have the Porshe or the
Ferrari? There is research on automatic driving going on?


  - Ian Parker

post@couptreffer.de wrote:
> Hello,
> 
> NN is a fine tool for extracting various pattern. Of course that
> requires the existence of a pattern. But what to do if patterns are
> rarely found or if there is a very blurred pattern structure.
> 
> I think this applies to data whith a considerably random-like origin
> (for example: stock markets, results of sportevents).
> 
> In this context: Wich kind of neural Network (MLP, RBF,  ...) should
> be used and what should be avoided?
> 
> Alea

Patterns in mostly random data. Interesting problem.

There is plenty of literature on noise reduction and pattern recognition.

The nub of it is that you have to know the kinds of patterns your looking
for. This is a subjective choice.

Ignore that fundamental requirement, and you'll find that everything is 
a pattern.

Reference: Satosi Watanabe's Theorem of the Ugly Duckling.

Tom.