Thanks! Two clarifications: 1) If the procedure is correct, I should always mark LS's that I know are there through domain knowledge, it's not a good thing to leave such information out if I know it independently. Correct? And 2) X-12-ARIMA's issue is with all outliers/interventions, including those marked by me, or is it only with outliers/interventions that X-12-ARIMA identifies automatically, through outlier?

I might add, I really like X-12-ARIMA for ARIMA work. It's slightly awkward to use with R, but the pages of diagnostic output that it gives are rather amazing... an education in itself, really. And there's the cachet of using the same software that governments around the world use. I look forward to X-13 which has been in beta for a while now. (Apparently unavailable to us mere mortals.)

Irrespective of your "ability to explain" the ls should be included because if you didn't include them the error process would be malformed. I don't follow your second comment. X12-ARIMA doesn't delete or augment the model , in other words no tests for necessity or sufficiency are incorporated. No tests are done for constancy of parameters or constancy of varaiance over time ; no tests are done on model errors to determine the optimal power transformation but for some reason transformations are selected on the original data which has no requiremrnts to be normally distributed.

In your answer, you say that "X12-ARIMA makes the mistake of developing the ARIMA first and THEN the empirically identified interventions". I'm interpreting "empirically identified" to mean identified by X12 in its automatic outlier detection procedure, as opposed to LS's that I know about a priori and manually enter. Just wanted to make sure I got that part right.

@wayne Yes you are right ! . I should have made that distinction. One other thing you are entering dates of LS that are known de jure. The true points of intervention may have started before or after thus the de facto dates may differ from the de jure dates. I ran into this when I studied the classic paper reflecting traffic deaths on texas highways as it reflected the date of a law change regarding seat belts.

Thanks! I had to submit my initial report a couple of days ago, and in the model I used I hadn't inserted the two additional LS's that I know of (for a total of four), because doing so caused the ACF of the residuals to show significance at lag 2, so I was skeptical. I'll add those two additional LS's and move forward investigating the ACF issue. (Four level shifts in 3.5 years, not counting little tweaks here and there, with a fifth LS about to hit us this month. That's our electricity provider...)

@Wayne If you can direct me how to do it we can go to a chat room for more detailed discussion. I have no clue as to how to do that.

If you follow the link on the last message you posted, it takes you to here, where you can type in the bottom of the screen and send.

I have to wait until I get home from the day job to actually work with the data, but am looking forward to adding the two LS's. They lower my projection -- which is good in this case -- as long as I can make sure that all the diagnostics look reasonable. Also working on a DLM solution: one uses regressions for indicator variables for level shifts and has a larger CI, and I'm working on one that dramatically increases the allowed variance for the trend after level shifts to allow it to adapt.

ok i am logged in

Wayne, if you have the data and tour model and the x12-arima output I would be glad to look at it.

You can send it to me at dave@autobox.com and I will post back here any findings.

r u there ?

es, had to step away.

I can get the data this evening, unless I'm lucky and put something in Dropbox so I can get to it from work.

Just had a look, and I don't see it here at work. Best I have is a graph of the data (and trend, prediction, Prediction Interval) that I have in a report I emailed to someone.

That's for the ARIMA (0 1 1) (0 1 1)12 model, with input of the actual data (black line up through the end of the red trend), the two level shifts (red trend jumps), refunds (Apr-Jul 2010), and how many degrees more extreme the temperature was each month compared to the mean. I can post more here after I get home.

@Wayne ok