Why does DateOffset.rollback() not work the way I expect it to with days / hours?-CodePudding

I'm trying to move an arbitrary input pd.Timestamp back to the correct start offset unless it is already on offset.

The following code works with BusinessMonthEnd, MonthEnd, MonthStart, and so on:

import pandas as pd
from pandas.tseries.offsets import *


def to_start(t: pd.Timestamp, freq: pd.DateOffset) -> pd.Timestamp:
    return freq.rollback(t)


assert to_start(pd.to_datetime("2021-09-27"), MonthEnd()) == pd.to_datetime("2021-08-31")
assert to_start(pd.to_datetime("2021-09-27"), MonthBegin()) == pd.to_datetime("2021-08-31")
assert to_start(pd.to_datetime("2021-08-27"), BMonthEnd()) == pd.to_datetime("2021-07-30")
assert to_start(pd.to_datetime("2021-08-27"), YearBegin()) == pd.to_datetime("2021-01-01")
assert to_start(pd.to_datetime("2021-08-27"), BYearBegin()) == pd.to_datetime("2021-01-01")

# also it works nicely with holiday calendars
from pandas.tseries.holiday import USFederalHolidayCalendar

us_fed_biz_days = CustomBusinessDay(calendar=USFederalHolidayCalendar())
memorial_day = pd.to_datetime("2021-05-31")
the_friday_before_memorial_day = pd.to_datetime("2021-05-28")
assert to_start(memorial_day, us_fed_biz_days) == the_friday_before_memorial_day

However (and this is driving me crazy) it doesn't seem to work for Day, BusinessDay, Week, Hour etc:

assert to_start(pd.to_datetime("2021-08-27 05:00"), Day()) == pd.to_datetime("2021-08-27")
assert to_start(pd.to_datetime("2021-08-27 05:00"), BDay()) == pd.to_datetime("2021-08-27")
assert to_start(pd.to_datetime("2021-08-27 05:15"), Hour()) == pd.to_datetime("2021-08-27 05:00")
assert to_start(pd.to_datetime("2021-08-26"), pd.tseries.frequencies.to_offset("W-MON")) == pd.to_datetime("2021-08-24")

I have also tried this:

def to_start(t: pd.Timestamp, freq: pd.DateOffset) -> pd.Timestamp:
        return pd.Period(t, freq=freq).start_time

which ironically works with the second set of assertions but not the first?

Is my expectation of the above assertions not failing unreasonable, and if so what am I missing, please?

CodePudding user response：

From your examples, I take it that you want to use fixed frequencies (e.g. an hour) to floor the given date/time while variable frequencies (e.g. business month end) should call .rollback if the check given in .is_on_offset returns True (see source code linked in the questions' comment section).

Ex:

def to_start(t: pd.Timestamp, freq: pd.DateOffset) -> pd.Timestamp:
    try:
        return t.floor(freq) # fixed frequencies should just floor the date/time
    except ValueError: # if freq is variable, we fall into here...
        return freq.rollback(t.floor("D"))

Tests:

# variable offsets that depend on the date
assert to_start(pd.to_datetime("2021-09-27"), pd.tseries.offsets.MonthEnd()) == pd.to_datetime("2021-08-31")
assert to_start(pd.to_datetime("2021-09-27"), pd.tseries.offsets.MonthBegin()) == pd.to_datetime("2021-09-01")
assert to_start(pd.to_datetime("2021-08-27"), pd.tseries.offsets.BMonthEnd()) == pd.to_datetime("2021-07-30")
assert to_start(pd.to_datetime("2021-08-27"), pd.tseries.offsets.YearBegin()) == pd.to_datetime("2021-01-01")
assert to_start(pd.to_datetime("2021-08-27"), pd.tseries.offsets.BYearBegin()) == pd.to_datetime("2021-01-01")
assert to_start(pd.to_datetime("2021-08-26"), pd.tseries.frequencies.to_offset("W-MON")) == pd.to_datetime("2021-08-23")
assert to_start(pd.to_datetime("2021-08-28 05:00"), pd.tseries.offsets.BDay()) == pd.to_datetime("2021-08-27")
# fixed offsets (e.g. an hour is always an hour)
assert to_start(pd.to_datetime("2021-08-27 05:00"), pd.tseries.offsets.Day()) == pd.to_datetime("2021-08-27")
assert to_start(pd.to_datetime("2021-08-27 05:15"), pd.tseries.offsets.Hour()) == pd.to_datetime("2021-08-27 05:00")