BUG: Inconsistent datetime comparison with Tz

[sinhrks]

Related to #8306. On current master, Timestamp comparison results in TypeError if its timezones are different. However, Index and Series implicitly converts tz to GMT

pd.Timestamp('2016-01-01 12:00', tz='US/Eastern') > pd.Timestamp('2016-01-01 08:00')
# TypeError: Cannot compare tz-naive and tz-aware timestamps

# same result as idx.tz_convert(None) > pd.Timestamp('2016-01-01 08:00')
idx = pd.date_range('2016-01-01 12:00', periods=10, freq='H', tz='Asia/Tokyo')
idx > pd.Timestamp('2016-01-01 08:00')
# array([False, False, False, False, False, False,  True,  True,  True,  True], dtype=bool)

Numeric ops raises TypeError as expected.

idx - pd.Timestamp('2016-01-01 08:00')
# TypeError: Timestamp subtraction must have the same timezones or no timezones

[gliptak]

I opened numpy/numpy#7390 Does this belong to pandas instead? Thanks

[jreback]

this is an invalid dtype for numpy and not defined there
further what you are doing doesn't make any sense

[gliptak]

@jreback I'm validating that the ts column has datetime64 with timezone (just comparing it to datetime64 fails ...). How would this need to be coded?

[jreback]

use .select_dtypes or an com.is_datetimelike or com.is_datetime64tz_dtype

numpy doesn't know about/respect this (its really a bug in the dtype definition and i don't know when/if ever will be fixed/allowed).

[jreback]

Here is also the method to coerce. EDT is not a timezone, and what dateutil is doing is wrong and doesn't give you anything useful.

In [34]: df = pd.DataFrame(["Mar 10, 2016 11:20 PM EDT"], columns=['ts'])

In [35]: pd.to_datetime(df['ts']).astype('datetime64[us, US/Eastern]')
Out[35]: 
0   2016-03-10 23:20:00-05:00
Name: ts, dtype: datetime64[ns, US/Eastern]

[gliptak]

Thank you for the pointers.

In [4]: df = pd.DataFrame([parse("Mar 10, 2016 11:20 PM EDT")], columns=['ts'])
In [16]: df['ts'] = pd.to_datetime(df['ts']).astype('datetime64[us, US/Eastern]')
In [19]: df.dtypes['ts'] == np.dtype('datetime64[ns]')
Out[19]: False

So how am I to compare? Thanks

[jreback]

what are you trying to do? why do you need to compare? what are you comparing? most ops will simply work, you rarely actually need to compare things, if you need to sub-select use .select_dtypes(...) as I indicated.

[gliptak]

Sorry, I didn't offer context. I came across this working unit tests for pydata/pandas-datareader#188

dtypes = [np.dtype(x) for x in ['float64', 'float64', 'datetime64[ns]']]
tm.assert_series_equal(df.dtypes, pd.Series(dtypes, index=exp_columns))

I had to force no timezone for the compare above to succeed ...
Could you show how to rewrite df.dtypes['ts'] == np.dtype('datetime64[ns]') with .select_dtypes(...)?
Thanks

[jreback]

not really sure what you are doing

http://pandas.pydata.org/pandas-docs/stable/generated/pandas.DataFrame.select_dtypes.html?highlight=select_dtypes#pandas.DataFrame.select_dtypes

[sinhrks]

@gliptak I don't quite understood also. pydata/pandas-datareader#188 is merged and the test has been passed. Pls update pydata/pandas-datareader#188 if there is any problem. I assume this issue is unrelated to yours.