Consider a series holding timedelta64[ns] that measures at the time difference between two events A and B:

> time_deltas499900   -1 days +23:45:13
499916   -1 days +23:50:57
499917            00:03:27
499919            00:17:45
499920            00:16:56
499921   -1 days +23:59:26
499922            00:16:34
499923            00:15:46
499928            00:12:56
499929            00:05:54...       
499970            00:00:48
499971   -1 days +23:58:32dtype: timedelta64[ns]

How can I identify negative deltas? (e.g. A happening before B).

This does not work:

> time_deltas[time_deltas<0]
TypeError: invalid type comparison

Also consider the following:

# Negative  time delta example:
> time_deltas.iloc[-1]
Timedelta('-1 days +23:58:32')# Values seem to have integer representation in ns
> time_deltas.iloc[-1].value
-88000000000# Positive time delta example:
> time_deltas.iloc[-2]
Timedelta('0 days 00:00:48')# Again, values seem to have integer representation in ns
> time_deltas.iloc[-1].value
48000000000

But then:

# Trying to use the internal representation fails
> time_deltas.apply(lambda x: x.value>0)
AttributeError: 'numpy.timedelta64' object has no attribute 'value'# Same with
> time_deltas.apply(lambda x: x['value']>0)
IndexError: invalid index to scalar variable.

Compare it against pd.Timedelta(0):

In [60]: time_deltas = pd.to_timedelta(np.random.randint(-10**6, 10**6, size=10))In [61]: time_deltas
Out[61]: 
TimedeltaIndex([         '00:00:00.000809', '-1 days +23:59:59.999034','-1 days +23:59:59.999456', '-1 days +23:59:59.999156','-1 days +23:59:59.999053', '-1 days +23:59:59.999723','-1 days +23:59:59.999523',          '00:00:00.000349','00:00:00.000051',          '00:00:00.000774'],dtype='timedelta64[ns]', freq=None)In [62]: time_deltas < pd.Timedelta(0)
Out[62]: array([False,  True,  True,  True,  True,  True,  True, False, False, False], dtype=bool)