ENH: support .groupby().ewm() directly

[jreback]

would be nice to support .groupby().ewm() directly.

In [2]: df = pd.DataFrame({'tid': np.random.randint(0, 10, size=100), 'value': np.random.randn(100)})

In [3]: df.groupby('tid').rolling(3).value.mean()
Out[3]: 
tid    
0    3          NaN
     6          NaN
     10   -0.596818
     16   -0.619943
     23   -0.688623
     24   -0.167816
             ...   
9    74    0.575155
     80    1.046240
     81    0.279683
     85    0.296154
     86   -0.267774
     92    0.320419
Name: value, dtype: float64

In [4]: df.groupby('tid').apply(lambda x: x.ewm(5).value.mean())
Out[4]: 
tid    
0    3    -1.182125
     6    -0.118510
     10   -0.616044
     16   -0.820601
     23   -0.435397
     24   -0.279619
             ...   
9    74    0.505610
     80    0.657754
     81    0.451367
     85    0.372517
     86    0.226676
     92    0.429832
Name: value, dtype: float64

[vallahaye]

Hello everyone!

Is this issue still planned for a future major release ?
I'm currently building a library over Pandas at my job and having the possibility to call "ewm" on a SeriesGroupBy object would reduce the complexity of my code.
Also on a performance standpoint, I've recently come across issues when grouping my data by more than one columns (which leaded to the creation of around two million groups). Would having the "ewm" function accessible directly on the SeriesGroupBy object be beneficial for the execution speed ? I imagine it would take some kind of "fast path" in the execution process by not having to call a lambda function (grouped.transform(lambda x: x.ewm...)) for each group.

Thanks in advance

[TomAugspurger]

I don't think anyone is working on this right now.

[vallahaye]

Yeah I have to feeling it's been forgotten over the years. How hard is it to implement such a feature ? Am I right thinking the performances would be better ? Maybe I could have a look at it someday.