Weekly anchored frequency offsets don't behave as expected

[ijmcf]

For example:

period_index = pandas.period_range('2012/6/3', periods=10, freq='W-FRI')
<class 'pandas.tseries.period.PeriodIndex'>
freq: W-FRI
[02-Jun-2012/08-Jun-2012, ..., 04-Aug-2012/10-Aug-2012]
length: 10

Although the frequency is "W-FRI", the start dates listed (02-Jun-2012, ..., 04-Aug-2012) are Saturdays.

Converting the frequency of the PeriodIndex to daily using "start", you get Saturdays:

period_index.asfreq(freq='D', how='start')
<class 'pandas.tseries.period.PeriodIndex'>
freq: D
[02-Jun-2012, ..., 04-Aug-2012]
length: 10

And if you convert the PeriodIndex to a DatetimeIndex using to_timestamp(), you get a frequency that is explicitly "W-SAT":

period_index.to_timestamp(freq='D', how='s')
<class 'pandas.tseries.index.DatetimeIndex'>
[2012-06-02 00:00:00, ..., 2012-08-04 00:00:00]
Length: 10, Freq: W-SAT, Timezone: None

[changhiskhan]

The week anchoring designates end of week as opposed to beginning of week. We have the option to either change the API or make the docs more explicit.

@wesm do you have any opinion on the API here? I personally think beginning of week anchoring is slightly more intuitive, but I'm afraid of an API change like this that's likely to cause a good deal of silent breakages.

[ijmcf]

It would never occur to me that an anchored offset called "W-THU" (Weekly - Thursday) would start on Friday and end on Thursday. I can live with it, but explaining it to the user base here is going to be interesting.

If it's going to stay as it is now, then converting a PeriodIndex to DatetimeIndex should show a consistent behavior - currently the frequency changes, as shown in my example.

[changhiskhan]

I think anchoring at end of week was inherited from scikits.timeseries behavior. It's also more consistent with other anchored offsets (e.g., A-FEB) where the anchor is the end.

And with regards to consistency between DatetimeIndex and PeriodIndex, since anchoring is end of week, period_index.to_timestamp(freq='D', how='e') gives you the week ending timestamps with the same freq:

In [9]: prng
Out[9]:
<class 'pandas.tseries.period.PeriodIndex'>
freq: W-SUN
[2012-05-28/2012-06-03, ..., 2012-06-11/2012-06-17]
length: 3

In [10]: prng.to_timestamp(freq='D', how='e')
Out[10]:
<class 'pandas.tseries.index.DatetimeIndex'>
[2012-06-03 00:00:00, ..., 2012-06-17 00:00:00]
Length: 3, Freq: W-SUN, Timezone: None

[changhiskhan]

Closing

(can reopen if anyone feels strongly against end-of-week anchoring)