Addition of two DateOffsets

[1pakch]

It is convenient to be able to add one DateOffset to another one:

from pandas.tseries.offsets import DateOffset as DO
DO(hour=2) + DO(hour=3) # raises ValueError

Currently it is not supported. Note that DateOffsets basically replicates the functionality of relativedeltas from dateutil module which support addition

from dateutil.relativedelta import relativedelta as RD
RD(hour=2) + RD(hour=3) # yields relativedelta(hour=3)

[jreback]

In [17]: Timestamp('20130101') + DO(hour=2) + DO(hour=3)
Out[17]: Timestamp('2013-01-01 03:00:00')

These do not form composite objects. I suppose they could, but not sure if it has a useful case as these can operate directly on Series/Timestamps.

@aickley pls show the utility here.

[1pakch]

This would be useful when 1) offsets are defined at a different place from where there are used and 2) when offsets are defined relative to each other. For instance, currently I have some code conceptually similar to

# Parameters of a study
predictors_window_bdays = 3
predictors_window_end = DateOffset(...) # offset from an event
predictors_window_start = predictors_window_start - DateOffset(bdays=predictors_window_bdays)
# response_window_{bdays,start,end} defined similarly

# The code where the real work is done
# events is a DatetimeIndex()
x = predictors.ix[events+predictors_window_end] - predictors.ix[events+predictors_window_start]
y = response.ix[events+response_window_end]  - response.ix[events+response_window_start]
# analyze the relationship between x and y

Now it is not possible to specify predictor_window_end corresponding to, say, 12:00 of a business day preceding the day of the event (in my view the most intuitive syntax for that would be DateOffset(bdays=-1, hour=12) as suggested in #10903). Additionally, it is not possible to define predictor_window_start relative to predictor_window_end using an intuitive syntax described above.

As a workaround, I am currently storing lists of DateOffsets that are being applied to the index one after the other.

[jreback]

This is currently not allowed because adding anything to a BusinessDays (except other business days) is not defined without an absolute reference point. E.g. if you had Timestamp('20130101 11:59:59) + BDay(1) + Second(1) it is clear what to do.

But if you have BDay(1) + Second() then these have to be kept (and applied) independently.

So aside from BusinessHours which I think could be added to a BuisinessDay I don't think this is really possible.

I'll leave it open for you to provide a reasonable proof-of-concept.

[jreback]

For non-business days I suppose these could be combinable as Timedeltas are.

e.g.

Day() + Hour() -> DateOffset(days=1,hours=1)

[1pakch]

Ok, I understand that BDays is not a time unit and hence it's very different from Hour, Minute etc. That being said, in principle, DateOffset can have a bdays offset independent from the other components. When such an extended DateOffset is applied to a date the only ambiguity is whether offsetting by business days or usual time units is done first.

Apart from that, combining BDay with absolute time of the day (BDay() + DateOffset(hour=12, minute=1)) is a transparent operation which is very useful when looking at intraday financial time series.

BTW, surprisingly enough, now one can combine BDay and Second

In [33]: delta = BDay() + Second()
In [34]: (pandas.DatetimeIndex(['2015-08-24']) - delta)[0]
Out[34]: Timestamp('2015-08-21 00:00:01')

and the result is different from the direct application

In [35]: (pandas.DatetimeIndex(['2015-08-24']) - BDay() - Second())[0]
Out[35]: Timestamp('2015-08-20 23:59:59')

which is clearly a bug. The same applies to Hour (I checked) and, I guess, all the other units.

[chris-b1]

@aickley It's not so clear to me that's a bug, because offsets are not necessarily associative.

The way addition of a delta on BDay works right now is that the offset remains a BusinessDay - the additional delta is applied at the last step without regard to sign/count on the offset.

E.g:

In [5]: delta = pd.offsets.BDay() + pd.offsets.Hour()
In [6]: delta
Out[6]: <BusinessDay: offset=Timedelta('0 days 01:00:00')>

In [9]: pd.Timestamp('2015-1-1') + delta
Out[9]: Timestamp('2015-01-02 01:00:00')

In [10]: pd.Timestamp('2015-1-1') + 2 * delta
Out[10]: Timestamp('2015-01-05 01:00:00')

In [11]: pd.Timestamp('2015-1-1') - 2 * delta
Out[11]: Timestamp('2014-12-30 01:00:00')

This definition actually makes sense if you think of the delta as an 'anchored' offset - that is, 1:00 on each business day. But it may be less surprising if the sign/count propagated to the additional delta and that maybe could be changed. It is a bit ambiguous / surprising.

In any case, I'm not seeing why you can't just apply multiple offsets to your data? That would avoid any ambiguity.

[mroeschke]

Yeah I think this has been disallowed specifically as it raises a TypeError now and due to #10902 (comment) probably not easy to support so closing