[SciPy-User] scikits.timeseries: How to define frequency of15minutes

Georges Schutz georges.schutz@internet...
Fri Apr 16 16:17:06 CDT 2010


OK I got that, and I thank you for your clear words stating 
the current situation. I have not yet analyzed the 
consequences of working with missing data. But I also think 
that 93% of missing data is some how waste of memory 
especially with large datasets which is often the case using  
time-series.

I suppose I have to little background in C programming and my 
experiences in C are too long ago to be able to propose an 
useful C-based patch for that issue.

I will experience with the current code and will follow 
closely the evolution of scikits.timeseries.

Keep doing what you are doing it is really helpful.

Georges Schutz 

Pierre GM wrote:

> On Apr 16, 2010, at 2:07 PM, Matt Knox wrote:
>> Georges Schutz <georges.schutz <at> internet.lu> writes:
>> 
>>> 
>>> I know that the underlying code is in C and that this is 
also the reason
>>> of the fast runtime of this python modules what I 
appreciate of cause.
>>> My initial question was if there exist a method to realize 
custom
>>> frequencies which seems not being the case. Know the 
question is is
>>> nobody else needing such a feature or is it somewhere on a 
todo list or
>>> even someone already implemented it in a one extension.
>> 
>> This isn't on my to-do list, or Pierre's either so far as I 
know.
>> Although I certainly agree it is a useful feature. As far 
as major new
>> features, I am kind of waiting to see what transpires with 
the numpy
>> datetime data type when it matures a bit and then maybe see 
how we might
>> integrate that into the time series package, at which point 
it would also
>> be useful to revisit this feature.
> 
> 
> I agree with Matt: the new numpy dtypes may be a good lead, 
but it's still
> far too early to tell how easy it'll be to merge it w/ 
scikits.timeseries.
> Keep in mind that the package was initially designed to 
provide some
> convenience while analyzing data at frequencies of a day or 
lower, because
> that was what Matt and I had to deal with. However, like 
Matt, I recognize
> that having the possibility to define specific frequencies 
would be great.
> Unfortunately, we are at this point stuck with the existing 
frequencies.
> There are a couple of ways to handle 15-min frequencies, 
though. As
> suggested before, you could use a minutely frequency. 
However, you'd end
> up with 93% of missing data, which can be a bit wasteful. A 
second
> possibility is to use a hour frequency with a (n,4) series, 
with n the
> number of hours. In turn, this might be a problem if you 
want to convert
> the series to another frequency (nD series are not 
recognized by
> .convert), but that could get you started. Of course, 
that'll depend on
> what you want to do. Let us know how it goes P.




More information about the SciPy-User mailing list