[BUG] NumPy timedelta64 order of magnitude not changed during date operations

[dmvianna]

It seems that whatever the unit of time used in timedelta64 (day, week, hour, year), when operating with a datetime64[ns] value, the timedelta integer is operated as nanosecond. Hence, if we intend to add or subtract one week (or day, or hour, or year), only one nanosecond is subtracted instead.

In [7]: import numpy as np

In [8]: import pandas as pd

In [9]: from pandas import Series

In [10]: a = np.timedelta64(1, 'W')

In [11]: a
Out[11]: numpy.timedelta64(1,'W')

In [12]: b = Series(pd.to_datetime('1/1/2010'))

In [13]: b
Out[13]:
0   2010-01-01 00:00:00
dtype: datetime64[ns]

In [14]: b - a
Out[14]:
0   2009-12-31 23:59:59.999999999
dtype: datetime64[ns]

In [15]: pd.__version__, np.__version__
Out[15]: ('0.12.0', '1.7.1')

[hayd]

I think this will be fixed in 0.13 (as it appears to be working in current master):

In [7]: b - a
Out[7]:
0   2009-12-25 00:00:00
dtype: datetime64[ns]

[jreback]

not the exact fix #4534, but timedelta support is much improved in master/0.13