TSA bug with frequency inference

[jseabold]

Somehow the frequency is not being set properly here and when properly set it gets set to None somewhere.

http://stackoverflow.com/questions/15515019/statsmodel-using-arma

[jseabold]

The problem is that these dates do not have a discernible frequency. So we need to be able to not return the wrapped results here or to just assign an index of integers rather.

[jseabold]

What should we return for the dates index for prediction in cases of non-consecutive days frequencies, for instance, business days? Should we also return business days dates, or just fall back on an integer index for the series in this case too?

[josef-pkt]

Do you have enough information to create dates, future dates that are in the corresponding business day frequency/sequence?
If you can handle it, we could just pretend it's a different calender where non-business days don't exist, one-step is one business day independent of actual/real time.

The conflicting interpretation (which we don't support yet) is to assume subordination of a continuous time process (for example). The underlying process is in "real" time but we observe it only at irregularly spaced time intervals.
(or think of AR in space with irregular distances instead of time). In this case, one step ahead would be one real time unit, not observed calender time unit.

[jseabold]

Yeah I do have enough info. It's no extra work, we get it for free from pandas, I just don't know if it makes sense. I'll go ahead and let business day through.

The thing to fix is just when there's no frequency, it's irregularly spaced (or there's bank holidays interspersed). Then we'll just return an integer index.

[jseabold]

There are a few corner cases here and this is a bit more work than I thought, so I'm going to lay out what needs to be done if I can't finish this shortly, so I can return to it. Some of it is already done in a local branch. When there is no discernible frequency to the dates, we need to

1. Raise an exception when start date is not on the dates index. If dates are 12-30-09, 1-3-10, 1-4-10, we don't know what to do with a start date of 1-1-10. In the future, we will intelligently roll forward onto 1-3-10, but I'm not going to code this until we drop support for pandas 0.7.x, so it's easier and cleaner to do.
2. If end is a date, it must be on the index in-sample. If not, raise an exception. If it is, then predict_dates should just be a slice of the in-sample date index.
3. If you want to predict out of sample, end must be an integer. We won't know what predict from 1-1-2010 1-8-2010 means if the last in-sample time is, for example, 1-2-2010 without any frequency information.
4. Don't forget to also test the forecast method.