BUG: indexing with string in multi-index for Period gives KeyError

[jorisvandenbossche]

From SO:

import pandas as pd
import itertools
a = pd.period_range('2013Q1','2013Q4', freq='Q')
i = (1111, 2222, 3333)
idx = pd.MultiIndex.from_tuples(list(itertools.product(a, i)),
                            names=('Periode', 'CVR'))
df = pd.DataFrame(index=idx,
              columns=('OMS', 'OMK','RES','DRIFT_IND','OEVRIG_IND','FIN_IND','VARE_UD','LOEN_UD','FIN_UD'))

Trying to select with a string gives a ValueError:

In [15]: df.loc[('2013Q1',1111),'OMS']
---------------------------------------------------------------------------
KeyError: ('2013Q1', 1111)

while it works with an explicit Period:

In [16]: df.loc[(pd.Period('2013Q1'),1111),'OMS']
Out[16]: nan

and it also works when not using multi-index notation (only accessing the first level): df.loc['2013Q1', 'OMS']

and this did work in 0.14.1.

[jorisvandenbossche]

Further, the string syntax also does not work for assigning (but this was also the case in 0.14.2). While for a DatetimeIndex both accessing and assigning using a string indexer works.

[jreback]

this is essentially a dupe of #3462 as the multi-index delegation is to the PeriodIndex which doesn't support this ATM.

[jorisvandenbossche]

@jreback Possibly, but this did work in 0.14.1, while the issue you linked to was already broken (or not implemented) then.

But of course possible the underlying issue is the same

[jreback]

hmm, ok, Period does need a bit of attention

[mroeschke]

This works again in 0.19.1:

In [31]: df.loc[('2013Q1',1111),'OMS']
Out[31]: 
Periode  CVR 
2013Q1   1111    NaN
Name: OMS, dtype: object

However the result isn't the same as:

In [32]: df.loc[(pd.Period('2013Q1'),1111),'OMS']
Out[32]: nan

Should these return the same result?

[jbrockmendel]

@mroeschke was there discussion somewhere else about the expected result here? i think the scalar np.nan makes more sense than a Series