Adding Timedelta parameter to time-based .rolling()

[chrisaycock]

Starting with a simple DataFrame:

n = 10
df = pd.DataFrame({'value': np.arange(n)},
                  index=pd.date_range('2015-12-24', periods=n, freq="D"))

I can invoke a time-based .rolling() with a string for the offset:

df.rolling('3d').sum()

I would like to invoke it with a Timedelta since it is easier to arithmetically change the window length:

df.rolling(pd.Timedelta(days=3)).sum()

[jreback]

hmm, it allows offsets already (in fact internally that's all its doing), so this must be just an oversight

In [7]: df.rolling(pd.tseries.frequencies.to_offset(pd.Timedelta(days=3))).sum()
Out[7]: 
            value
2015-12-24    0.0
2015-12-25    1.0
2015-12-26    3.0
2015-12-27    6.0
2015-12-28    9.0
2015-12-29   12.0
2015-12-30   15.0
2015-12-31   18.0
2016-01-01   21.0
2016-01-02   24.0

[chrisaycock]

Ah, I hadn't known about to_offset()! That's neat.

[fjanoos]

Thanks ! it would be great though if Timedelta were directly supported !

[jreback]

@fjanoos we already use to_offset internally

In [1]: from pandas.tseries.frequencies import to_offset

In [2]: to_offset(pd.offsets.Day(1)).nanos
Out[2]: 86400000000000

In [4]: to_offset(pd.Timedelta('1 day')).nanos
Out[4]: 86400000000000

this is just a matter of fixing the checking.

pull-requests are welcomed!

[fjanoos]

Is this change in the master branch now ?

[jreback]

yep closed via #15443