Prev: Plot Error Bar graph
Next: Logarithmic Bins for Histogram
From: Bruno on 24 May 2010 23:01
Hi,

I'm trying to predict the next point in a time series using wavelet decomposition and neural networks. But I'm having problems choosing the data to feed the NN after the signal decomposition. Can anyone clarify which data should train my network?

Thanks,
Bruno
From: Greg Heath on 25 May 2010 08:11
On May 24, 11:01 pm, "Bruno " <pale...(a)pobox.com> wrote:
> Hi,
>
> I'm trying to predict the next point in a time series using wavelet
decomposition and neural networks. But I'm having problems
choosing the data to feed the NN after the signal decomposition.
> Can anyone clarify which data should train my network?

Forget about NNs for a moment.

I'm not familiar with wavelet decomposition. However, if you
used the fft for spectral decomposition, then it cannot be done.
The reason is that evaluating an ifft beyond the original time
interval just yields a periodic extension.

Does evaluating a wavelet expression beyond the original
time interval yield a reasonable extrapolation?

Given the above, I don't see how a NN fits into the picture.

Please explain.

Hope this helps.

Greg
From: Bruno on 25 May 2010 14:23
Greg,

What i trying to do is decompose the input signal into different levels using wavelet. Since wavelet preserve time information it can be used as a filter to predict trend movements.
So I'm trying to figure out how to feed this decomposition into a NN to predict the next point.

Greg Heath <heath(a)alumni.brown.edu> wrote in message <047d52f9-470e-43f3-9784-307c45d4142f(a)e21g2000vbl.googlegroups.com>...
> On May 24, 11:01 pm, "Bruno " <pale...(a)pobox.com> wrote:
> > Hi,
> >
> > I'm trying to predict the next point in a time series using wavelet
> decomposition and neural networks. But I'm having problems
> choosing the data to feed the NN after the signal decomposition.
> > Can anyone clarify which data should train my network?
>
> Forget about NNs for a moment.
>
> I'm not familiar with wavelet decomposition. However, if you
> used the fft for spectral decomposition, then it cannot be done.
> The reason is that evaluating an ifft beyond the original time
> interval just yields a periodic extension.
>
> Does evaluating a wavelet expression beyond the original
> time interval yield a reasonable extrapolation?
>
> Given the above, I don't see how a NN fits into the picture.
>
> Please explain.
>
> Hope this helps.
>
> Greg
From: TideMan on 25 May 2010 16:53
On May 26, 6:23 am, "Bruno " <pale...(a)pobox.com> wrote:
> Greg,
>
> What i trying to do is decompose the input signal into different levels using wavelet. Since wavelet preserve time information it can be used as a filter to predict trend movements.
> So I'm trying to figure out how to feed this decomposition into a NN to predict the next point.
>
> Greg Heath <he...(a)alumni.brown.edu> wrote in message <047d52f9-470e-43f3-9784-307c45d41...(a)e21g2000vbl.googlegroups.com>...
> > On May 24, 11:01 pm, "Bruno " <pale...(a)pobox.com> wrote:
> > > Hi,
>
> > > I'm trying to predict the next point in a time series using wavelet
> > decomposition and neural networks. But I'm having problems
> > choosing the data to feed the NN after the signal decomposition.
> > > Can anyone clarify which data should train my network?
>
> > Forget about NNs for a moment.
>
> > I'm not familiar with wavelet decomposition. However, if you
> > used the fft for spectral decomposition, then it cannot be done.
> > The reason is that evaluating an ifft beyond the original time
> > interval just yields a periodic extension.
>
> > Does evaluating a wavelet expression beyond the original
> > time interval yield a reasonable extrapolation?
>
> > Given the above, I don't see how a NN fits into the picture.
>
> > Please explain.
>
> > Hope this helps.
>
> > Greg

Well, I know nothing about NN, but I know a little bit about
orthogonal wavelet decomposition.
Greg's point about Fourier decomposition also applies to wavelets.
You need to forecast the wavelet coefficients using NN or whatever,
then recompose the time series using the scaling function (i.e., the
Matlab function wrcoef).

 | 
Pages: 1
Prev: Plot Error Bar graph
Next: Logarithmic Bins for Histogram