Add cdfplot

[TomAugspurger]

WIP for now. Closes #2669

I'm using statsmodels' KDE implementation right now; it has a cdf method, but scipy's gaussian_kde doesn't. I need to check the math, but I think something like doing a cumsum on the density and normalizing by the sum should be the same.

I'm also going to add kwargs for things like the inverse cdf.

[jtratner]

TODO (soonish): we should make this into a dict instead, would be simpler:

PLOT_KINDS = {"line": LinePlot,
 "bar": BarPlot,
 "barh": BarPlot,
 "kde": KdePlot,
 "cdf": CDFPlot}
# ... other stuff
if kind not in PLOT_KINDS:
   raise ValueError("Invalid chart type given: %s" % kind)
klass = PLOT_KINDS[kind]

[jtratner]

@TomAugspurger not saying the dict thing is required to get this merged, but if you wanted to add a second commit that changed this all to two dicts (one for Series, one for DataFrame), rather than if/elif statements, that'd be a complementary addition to this PR.

[jseabold]

FWIW, in statsmodels (and places in scipy.stats), we compute the cdf from the pdf using scipy.integrate.quad.

https://github.com/statsmodels/statsmodels/blob/master/statsmodels/nonparametric/kde.py#L161

[TomAugspurger]

Okay this one can sit until 0.14 has started and I can add a release notes entry. I got on a bit of DRYness run, so I ended up combining the KdePlot and CDFPlot into a DistributionPlot class.

I've only used statsmodels for the actual fitting of the cdf. Scipy doesn't provide a cdf on their kde object, so we'd have to duplicate the code that @jseabold linked to above. May as well just use what they've done.

And of course if this has fallen out of scope for pandas then close without merging!

[ghost]

We can't have a dependency on statsmodels, except for testing.
I remember kde plots were a bit suspect in terms of scope too,
expanding further is therefore suspect too.

Let's call this out of scope.