"Pradyumna S Upadrashta" <[EMAIL PROTECTED]> wrote:
> Are there any dynamical folks who speak statistics well enough that you
> can clarify what is meant by "nonlinear correlation"? are there any
> statisticians who can give me the correct way to think about correlation
> structure in a time-series, and how "nonlinear correlations" could
> affect a process (besides leading to some form of nonstationarity)?

Try exploring around http://www.mpipks-dresden.mpg.de/~tisean

You have to drop the linear-specific notion of 'correlation', and adopt a
more general notion of prediction error (correlation is the reduction in
prediction error using a linear univariate model, where the prediction
error is measured as the percentage of variance explained).

It turns out that you can predict pretty well in some cases by using a
nonlinear method of nearest neighbors. Imagine that you notice that the
time series has been behaving a little like once in the past: you then
predict directly from the matching history. Let me illustrate this with a
text series:

If you saw me say blah, did I say ....
[We notice that "say " was followed by "blah" in the past, so we predict
"blah", and we compute the prediction error.]

Aleks


.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to