Broken roundrip DatetimeIndex -> CategoricalIndex -> DatetimeIndex

[toobaz]

Code Sample, a copy-pastable example if possible

In [2]: pd.DatetimeIndex(pd.CategoricalIndex(pd.DatetimeIndex(['2015-10-10'], tz='US/Eastern')))
Out[2]: DatetimeIndex(['2015-10-10 04:00:00'], dtype='datetime64[ns]', freq=None)

Problem description

Out[2] has no timezone information. Related to this comment

Notice that it makes sense to loose freq, as this is a property of a specific collection of dates, such as a DatetimeIndex is. Instead, tz is a property of each date, and it should be hence kept.

This is (I think) the reason why

In [2]: pd.core.dtypes.cast.maybe_cast_to_datetime(pd.Timestamp('2015-10-10'), None)
Out[2]: array(['2015-10-10T00:00:00.000000000'], dtype='datetime64[ns]')

but

In [3]: pd.core.dtypes.cast.maybe_cast_to_datetime(pd.Timestamp('2015-10-10', tz='US/Eastern'), None)
Out[3]: DatetimeIndex(['2015-10-10 00:00:00-04:00'], dtype='datetime64[ns, US/Eastern]', freq=None)

Certainly related to #13783, #14052, #13238 and others, but probably requires a separate fix (as long as we don't have a real tz-aware dtype).

Expected Output

Out[2]: DatetimeIndex(['2015-10-10 00:00:00-04:00'], dtype='datetime64[ns, US/Eastern]', freq=None)

Output of pd.show_versions()

INSTALLED VERSIONS

commit: fdba133
python: 3.5.3.final.0
python-bits: 64
OS: Linux
OS-release: 4.9.0-4-amd64
machine: x86_64
processor:
byteorder: little
LC_ALL: None
LANG: it_IT.UTF-8
LOCALE: it_IT.UTF-8

pandas: 0.22.0.dev0+301.gfdba13333
pytest: 3.2.3
pip: 9.0.1
setuptools: 36.7.0
Cython: 0.25.2
numpy: 1.12.1
scipy: 0.19.0
pyarrow: None
xarray: None
IPython: 6.2.1
sphinx: 1.5.6
patsy: 0.4.1
dateutil: 2.6.1
pytz: 2017.2
blosc: None
bottleneck: 1.2.0dev
tables: 3.3.0
numexpr: 2.6.1
feather: 0.3.1
matplotlib: 2.0.0
openpyxl: 2.3.0
xlrd: 1.0.0
xlwt: 1.3.0
xlsxwriter: 0.9.6
lxml: 4.1.1
bs4: 4.5.3
html5lib: 0.999999999
sqlalchemy: 1.0.15
pymysql: None
psycopg2: None
jinja2: 2.10
s3fs: None
fastparquet: None
pandas_gbq: None
pandas_datareader: 0.2.1

[jbrockmendel]

@TomAugspurger any idea what's going on here? Seems like something to address as long as we're all focused on the DTA/TDA/PA constructors

[TomAugspurger]

Seems like the DTI constructor goes CategoricalIndex -> ndarray[datetime64[ns]], which loses the tzinfo.

In [2]: a = pd.CategoricalIndex(pd.DatetimeIndex(['2015-01-01'], tz='US/Eastern'))
In [3]: pd.DatetimeIndex(a)
> /Users/taugspurger/sandbox/pandas/pandas/core/indexes/datetimes.py(244)__new__()
-> if isinstance(data, Index):
(Pdb) data
CategoricalIndex(['2015-01-01 00:00:00-05:00'], categories=[2015-01-01 00:00:00-05:00], ordered=False, dtype='category')
(Pdb) c
> /Users/taugspurger/sandbox/pandas/pandas/core/indexes/datetimes.py(244)__new__()
-> if isinstance(data, Index):
(Pdb) data
array(['2015-01-01T05:00:00.000000000'], dtype='datetime64[ns]')

We would ideally follow this check down the datetimetz path, but is_dateimtetz(CategoricalIndex[datetime64[ns, tz]]) is false.

279  ->         if not (is_datetime64_dtype(data) or is_datetimetz(data) or
280                     is_integer_dtype(data) or lib.infer_dtype(data) == 'integer'):

(Pdb) data
CategoricalIndex(['2015-01-01 00:00:00-05:00'], categories=[2015-01-01 00:00:00-05:00], ordered=False, dtype='category')
(Pdb) is_datetimetz(data)
False

To the extent possible, I would recommend an array from the Series / Index as early as possible. Or we could maybe update is_datetimetz to look into index classes.

[TomAugspurger]

This is also fishy:

(Pdb) tools.to_datetime(data)
DatetimeIndex(['2015-01-01 05:00:00'], dtype='datetime64[ns]', freq=None)
(Pdb) tools.to_datetime(data.categories)
DatetimeIndex(['2015-01-01 00:00:00-05:00'], dtype='datetime64[ns, US/Eastern]', freq=None)

Those should probably both be datetime64[ns, US/Eastern]. We'd want to fix that for user code which may hit it, but again would be solved by unboxing arrays early in the index constructor (which maybe has to wait till we have lossless arrays for everything).