On 24 Sep 2003 at 16:23, Rich Ulrich wrote:

Just one comment:, see below.

> I find the question itself to be puzzling -- so I have split 
> it up and numbered the pieces.
> 
> On 19 Sep 2003 21:35:22 -0700, [EMAIL PROTECTED] (Anwar) wrote:
> 
> "Does anyone know of regression methods that 
> satisfy the following three criteria: 
>  1) nonlinear, 
>  2) variable coefficients (i.e. updating of coefficient values), AND 
>  3) minimize out-of-sample, one-step ahead, forecasting error 
> rather than in-sample fit? "
> ====== end of quote from the post.
> 
> On (1).  What sort of 'nonlinear' ? 
> Engineers and psychologists often include simple polynomials 
> when they say 'nonlinear', since these can draw curved lines;
> whereas statisticians are concerned with nonlinear-in-the-
> coefficients, which are more complicated.  
> 
> So, do you mean the easy nonlinear, or the tough one? 

That was the state some time ago non-linear -in-coefficients models
are not more any more difficult than linear models, there is a lot of 
high-quality optimization software around now! If your statistics 
package continues to make it difficult, or you does'n need it, or you 
change package. It is time to start to use the more natural 
terminology you say engeeneers and psychologists use. Some well-known 
statisticians have also said so publically, like Jim Lindsey, which 
uses the terminology that way in his books and on-line documents. 
Algorithmic difficulties should not determine terminology any more. 

> 
> 
> On (2).  Frankly, 'variable coefficients'  practically 
> freezes my brain:  when the phrase is re-interpreted, 

In time-series situations it is rather natural to have coefficients 
with change with time, and such models can be estimated with the 
Kalman filter. 

Kjetil Halvorsen

> so that it is not referring to "coefficients of variables"  but
> rather "updating of coefficient values."  
> 
> Okay.  Is this a matter of adapting robust lines to 
> windows of certain widths?  Or, could it be a request for 
> adaptive-learning?  Something else?  For either of the
> first two, I think that calling it a computer implementation
> of <  ... >   would be better that calling it a 'regression
> method'   but I'm prejudiced because I don't know what
> would be a single acceptable example, ignoring entirely
> the precepts of (1) and (3).
> 
> 
> On (3).  Huh?  It seems to me, this says nothing more
> than,  "The model does seem to comprise a time-series."
> And someone wants prediction.    Plus,  they use a 
> different vocabulary than what   I would use.  
> 
> This isn't one more prospectus for 
> outwitting the stock market,  is it?
> 
> 
> -- 
> Rich Ulrich, [EMAIL PROTECTED]
> http://www.pitt.edu/~wpilib/index.html
> "Taxes are the price we pay for civilization." 
> .
> .
> =================================================================
> Instructions for joining and leaving this list, remarks about the
> problem of INAPPROPRIATE MESSAGES, and archives are available at:
> .                  http://jse.stat.ncsu.edu/                    .
> =================================================================


.
.
=================================================================
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