statsmodels.api

[Josef Perktold]

Oscar Benjamin <oscar.j.benjamin <at> gmail.com> writes:

> 
> On 17 September 2013 14:35, Josef Pktd <josef.pktd <at> gmail.com> wrote:
> >> (As an aside, this is all much simpler if you're using Ubuntu or some
> >> other Linux distro rather than Windows.)
> >
> > scientific python on a stick
> >
> > https://code.google.com/p/winpython/wiki/PackageIndex_33
> 
> Thanks, I've just installed that and I'll try it out later.
> 
> > I haven't seen any problems so far on python 3.3
> > The statsmodels test suite passes without problems on python 3.3 also,
as far as I remember.
> > (and no problems using Windows. just use the right binaries.)
> 
> Well that's exactly my point. On a Linux distro you would have the
> right binaries first time. No need to search through project webpages
> or documentation, weigh up different installers, or download 750MB of
> software that you mostly won't use. Similarly on a Linux distro it's a
> lot easier to get all of the build tools set up to build these things
> from source if you'd prefer.

This might be true for many Linux users.
However, the last time I tried to install something in a virtual Linux OS
that was not in the standard repository, I was completely lost.
(I'm a Windows user.)


> 
> Windows users are often dependent on inconsistent sources of binaries.
> In this case I imagine that the OP installed numpy from sourceforge
> since it has 3.3 binaries but it doesn't have the same for scipy at
> which point googling would easily lead to Cristoph's page.

Another problem with relying on binaries on Windows is when the matching
binaries are not available. For example, the Windows binaries on pypi of
pandas and statsmodels are compiled against the latest numpy release and
will not work with older numpy versions.
(
http://stackoverflow.com/questions/17709641/valueerror-numpy-dtype-has-the-wrong-size-try-recompiling/18369312
)

On the other hand, python-xy comes with MingW, and I never had any problems
compiling pandas and statsmodels for any version combination of python and
numpy that I tested (although 32 bit only so far, I never set up the
Microsoft sdk).

(latest news: moving to MingW-64 for numpy and scipy, and related packages,
might be on the way.)

Josef


> 
> Oscar
>