update rolling doc string

[jhamman]

minor update of rolling doc string. Missed this update after @shoyer's last review.

xref: #668

[shoyer]

Does pandas follow this same convention for handling missing values? It's probably worth checking...

[jhamman]

I think bottleneck and pandas differ on how they handle the min_periods argument.

Bottleneck:

min_count: {int, None}, optional :

If the number of non-NaN values in a window is less than min_count, then a value of NaN is assigned to the window. By default min_count is None, which is equivalent to setting min_count equal to window.

Pandas doesn't say in its doc string:

min_periods : int, default None

Minimum number of observations in window required to have a value (otherwise result is NA).

So, comparing their behavior, we see they both set min_periods to the size of the window.

In [1]: import pandas as pd
In [2]: s = pd.Series(range(8))    
In [3]: pd.rolling_mean(s, 3)
Out[3]: 
0   NaN
1   NaN
2     1
3     2
4     3
5     4
6     5
7     6
dtype: float64
In [4]: import bottleneck as bn
In [6]: bn.move_mean(s, 3)
Out[6]: array([ nan,  nan,   1.,   2.,   3.,   4.,   5.,   6.])

[shoyer]

Something seems to be out of sync for NaN handling, though:

In [27]: d = xr.DataArray([0, np.nan, 1, 2, np.nan, 3, 4, 5, np.nan, 6, 7], dims='x')

In [28]: d.rolling(x=2).mean()
Out[28]:
<xarray.DataArray (x: 11)>
array([ nan,  0. ,  1. ,  1.5,  2. ,  3. ,  3.5,  4.5,  5. ,  6. ,  6.5])
Coordinates:
  * x        (x) int64 0 1 2 3 4 5 6 7 8 9 10

# using the pandas RC for v0.18
In [29]: d.to_series().rolling(2).mean().to_xarray()
Out[29]:
<xarray.DataArray (x: 11)>
array([ nan,  nan,  nan,  1.5,  nan,  nan,  3.5,  4.5,  nan,  nan,  6.5])
Coordinates:
  * x        (x) int64 0 1 2 3 4 5 6 7 8 9 10

[jhamman]

hmmm, you must not have bottleneck in your environment because I get:

In [6]: d.rolling(x=2).mean()
Out[6]: 
<xarray.DataArray (x: 11)>
array([ nan,  nan,  nan,  1.5,  nan,  nan,  3.5,  4.5,  nan,  nan,  6.5])
Coordinates:
  * x        (x) int64 0 1 2 3 4 5 6 7 8 9 10

I suppose we should open an issue on this. I guess we need to use the non-nan safe numpy methods in Rolling.reduce to get the same behavior. We'll have to come up with a solution to get this to work in a vectorized manor.

[shoyer]

Yep, no bottleneck on my work machine.

[shoyer]

made a new issue :#776