BUG: to_numpy on categorical data with tz aware datetime categories returns datetime64

[jorisvandenbossche]

In [1]: s = pd.Series(pd.Categorical(pd.date_range("2012", periods=3, tz='Europe/Brussels')))

In [2]: s.to_numpy() 
Out[2]: 
array(['2011-12-31T23:00:00.000000000', '2012-01-01T23:00:00.000000000',
       '2012-01-02T23:00:00.000000000'], dtype='datetime64[ns]')

In [3]: s = pd.Series(pd.date_range("2012", periods=3, tz='Europe/Brussels'))

In [4]: s.to_numpy() 
Out[4]: 
array([Timestamp('2012-01-01 00:00:00+0100', tz='Europe/Brussels', freq='D'),
       Timestamp('2012-01-02 00:00:00+0100', tz='Europe/Brussels', freq='D'),
       Timestamp('2012-01-03 00:00:00+0100', tz='Europe/Brussels', freq='D')],
      dtype=object)

We opted for the default behaviour of to_numpy to be as much preserving the values, so for normal datetimetz series, this is object dtype. So probably we should do the same for Categorical?

@TomAugspurger since to_numpy is still new, I suppose we can see this is as a bug fix and not change in behaviour (to not go through deprecation cycle)?

[TomAugspurger]

we can see this is as a bug fix

Agreed. Tagging for 0.25.0 though, so it doesn't sit for too long.