The current behaviour of sign on timedelta64 objects is somewhat surprising:

In [64]: sign(np.timedelta64(3, 'm')) == sign(np.timedelta64(3, 's'))
Out[64]: False

I'm not sure if the return type for the sign of a timedelta64 should even be a timedelta64 or or simply numerical. Of course, normally sign(x) returns a result of the same dtype as x, but for any numerical data we have uint8(1) == int8(1) == float32(1) == .... etc. I'm not sure what the best return value would be for the sign of a timedelta64, but I think it would be desirable if their signs compared equal.