`Period(weekend_date, 'B')` returns Monday's Period

[max-sixty]

The asof docs state that the method should return the most recent label up to and including the passed label. For period indexes, this doesn't seem to be happening - if I supply a period of a weekend date to a weekday series, it returns the Monday, rather than the Friday.

If I do the same on a DatetimeIndex, it behaves as expected

In [50]:
dates

Out[50]:
<class 'pandas.tseries.period.PeriodIndex'>
[1990-05-07, ..., 2015-07-13]
Length: 6568, Freq: B

In [51]:
dates_ts=dates.to_timestamp()
dates_ts

Out[51]:
<class 'pandas.tseries.index.DatetimeIndex'>
[1990-05-07, ..., 2015-07-13]
Length: 6568, Freq: None, Timezone: None

In [52]:
dates.asof('2015-07-11')

Out[52]:
Period('2015-07-13', 'B')

In [53]:
dates_ts.asof('2015-07-11')

Out[53]:
Timestamp('2015-07-10 00:00:00')

[sinhrks]

Thanks for the report!

Internally it uses get_loc, and these also return different results. Because Period creates next valid date when passed date and freq is invalid, needs special handling in get_loc. PR is appreciated:)

# DatetimeIndex
tidx = pd.date_range('2015-07-01', freq='B', periods=30)
tidx.get_loc('2015-07-19', method='pad')
# 12

# PeriodIndex
pidx = pd.period_range('2015-07-01', freq='B', periods=30)
pidx.get_loc('2015-07-19', method='pad')
# 13

pd.Period('2015-07-19', freq='B')
# Period('2015-07-20', 'B')

[max-sixty]

Is there a reason why it creates the next valid date rather than raising? While we could add a check in get_loc, the behavior we're correcting for seems unexpected.

[sinhrks]

I'm not sure whether there is a background of current Period behavior. It looks get_period_ordinal in period_helper.c doesn't care freqs with gaps.

@jreback @bjonen @cancan101 What do you think, pd.Period('2015-07-18', 'B') (2015-07-18 is Saturday) should raise ValueError?

[max-sixty]

@jreback @bjonen @cancan101, gentle follow up to @sinhrks's comment: do you have any thoughts here?

[max-sixty]

This hit us again today - @jreback do you have a view here?
At worst "this isn't important" or "don't change anything" would be helpful!

[jreback]

Well, you have to have a policy for what to happend when you pass a non-supported freq to your frequency based Period. I don't think rolling forward is that surprising. An alternative is what I gave in the other issue #10978 , e.g. move to a business day, then apply the op. to_period() just does this.

So I think .asof should as well, e.g. coerce the input to the appropriate frequency first.

[max-sixty]

OK. I agree with you for converting frequencies.

The concept of business days is a bit different - they have gaps. (I can't think of any other examples beyond business hours and some fanciful ones such as 'summer months'.)

So I don't think there's a concept of a business day representation of Saturday, while there is a month representation of '2010-01-01', or day representation of '2010-01'.

Although it may be that I'm alone in this view! If anyone else has thoughts + there's any consensus, happy to do a PR.