Concat Series[datetimetz] & Series[category] loses timezone

[TomAugspurger]

In [8]: a = pd.Series(pd.date_range('20170101', periods=4, tz='US/Pacific'))

In [9]: b = pd.Series(['a', 'b'], dtype='category')

In [10]: pd.concat([a, b], ignore_index=True)[0]
Out[10]: Timestamp('2017-01-01 08:00:00')

Out[10] should have a tz like the following

In [11]: pd.concat([a, b.astype(object)], ignore_index=True)[0]
Out[11]: Timestamp('2017-01-01 00:00:00-0800', tz='US/Pacific', freq='D')

[evangelinehl]

Hi,

Can my team and I contribute to this issue?

Thanks.

[TomAugspurger]

Sure, you can give it a shot.

I'm not sure where the best place is for the fix yet. Somewhere in core/dtypes/concat.py, but it'll be a bit tricky to find the best time to convert to objects.

[TomAugspurger]

That's the offset from the timezone ("US/Pacific").

Looking more closely, I think the issue is at

pandas/pandas/core/dtypes/concat.py

Line 194 in dc8d35a

def _concat_asobject(to_concat):

.

We call np.asarray(x) on non-categorical, like our datetimes. I wonder if we could rewrite that as

to_concat = [np.asarray(x.astype(object)) for x in to_concat]

We know we need object dtype anyway, so there's no avoiding that. We need the asarray(...) for things like SparseArray, since SparseArray.astype(object) is still sparse.

Maybe give that change a shot and see if the tests still pass.

[evangelinehl]

Okay I'll take a look at the suggested change. Thanks!

[evangelinehl]

This should be closed

[jreback]

@evangelineliu things are closed with a merged pull request