#399 closed enhancement (wontfix)
Optional auto-scaling of variables as estimate enhancement
Reported by: | jimarin | Owned by: | Víctor de Buen Remiro |
---|---|---|---|
Priority: | normal | Milestone: | |
Component: | Math | Version: | head |
Severity: | minor | Keywords: | |
Cc: |
Description
Hi there,
I think it would be of value to let Estimate function to auto-scale input series
so as not to suffer much from numerical problem at inner computations.
This feature could be optional so uninterested users could turn it off.
Thanks !
jima
Change History (3)
comment:1 Changed 19 years ago by
Status: | new → assigned |
---|
comment:2 Changed 17 years ago by
Resolution: | → wontfix |
---|---|
Status: | assigned → closed |
Note: See
TracTickets for help on using
tickets.
It is peculiar that mathematicians are in fact people to whom numbers do not
interest much.
I think it's a better idea to set the scale of each variable according to its
own nature. Human mind, even statistical analysts, cannot handle with very large
o small numbers. Ad-hoc scaling forces the analyst to know the units of each
variable making more intelligible models and avoiding interpretation errors.
Even so, optional internal auto-scaling could be a good feature to be used in
massive automated modelling processes, in order to avoid finite arithmetic
derived troubles. But automatic scaling methods are non trivial because standard
normalization trends to cumulate all eigenvalues very close to 1, and different
numerical problems could appear.
We will try it
Thanks to report this proposal.