date_range should take timedelta as freq

[hayd]

http://stackoverflow.com/questions/21643407/define-a-pandas-tseries-index-datetimeindex-using-2-datetimes-dt-start-and-dt-s/21651263#21651263

[c0indev3l]

I think passing a timedelta as freq can only make sense for timeframe below 1 month but it would still be very useful. Maybe an idea could be to provide a way to build a DateOffset from a timedelta

[hayd]

Agree, re making an offset. A from_timedelta method would make sense.

[jreback]

I think u could directly interpret a time delta as a passed freq
like to do a PR

[hayd]

I changed the pd.tseries.frequencies.to_offset to accept timedeltas in the PR.

It lazily forces it to Nano/Micro, could be more clever...

[jreback]

hmm these should try to upcast these (eg return seconds, minutes, days etc if possible
eg maybe extracts days, seconds , microseconds from timedeltas and then convert as appropriate ( their might be some methods in tseries.timedeltas to help with this - don't remember )

[jbrockmendel]

This works in 0.21.1

>>> td = pd.Timedelta(days=3, hours=4)
pd.date_range('2016-01-01', periods=4, freq=td)
DatetimeIndex(['2016-01-01 00:00:00', '2016-01-04 04:00:00',
               '2016-01-07 08:00:00', '2016-01-10 12:00:00'],
              dtype='datetime64[ns]', freq='76H')

This can probably be closed.

[jreback]

yes this is well tested

@jschendel did a PR. awhile back on this iirc