I have several groups of data for time series analysis, and I need some automatic way to do the analysis. I try to find a way in SAS to automatically do the job, but I can't find one corresponding to auto.arima() in R.

I have two questions regarding this:

  1. The closest one is to use identify with scan and esacf to find tentative orders for arima model. However, that doesn't test unitroot or whether the data is just a whitenoise. Also, I still need to manually figure out the orders. Does anyone know if there is any procedure in SAS that can do the analysis automatically?

  2. I tried to write a macro to do the job automatically. I need to fist run identify to check unitroot, whitenoise, orders etc. and based on the results, I need to run estimate accordingly with corresponding orders. However, when I write two separate proc arima procedures, SAS requires me to run identify again in the second procedure, otherwise the estimate can't be executed. For example:

    PROC ARIMA data = xxx;
         IDENTIFY VAR = yy scan esacf;
    /* other code */
    PROC ARIMA data = xxx;
         ESTIMATE p=1 q=1; <---------------- ERROR: must run IDENTIFY first

Is there a way to use the results of the identify in the first ARIMA for the estimate in the second ARIMA procedure? Otherwise, it's just a waste of time to run the identify twice.


You are fortunate to ask this question onn this site because IrishStat has been automating ARIMA models for over 30 years (sorry to give away your age Dave). Also Rob Hyndman wrote the auto.arima procedure in R. I have a connection as I took my first time series course in a short course by Box and Tiao at Carnegie - Mellon University in 1974 (giving away my age now). Also when I was the Chief of Statistical Research at Risk Data Corporation (in the early 1990s) I hired Terry Woodfield who authored the ETS software at the SAS Institute just before we were able to draw him away. I am sure PROC ARIMA has gone through many changes but i am sure that if you make contact with Terry he could probably help you.

Personally the way I learned it from Box, Tiao and Pack ARIMA modeling is an iterative process that should be gone through manually in stages with the user making decisions at various stages. That is not to say that good results cannot be obtained by automated procedures. In fact I think that Dave Reilly (IrishStat) along with his son Tom have so much experience doing this that they will contend that they could produce a better model with their algorithm than I can do manually and they may be right. But my point is that for a time series specialist to take that approach takes away some of the steps that help him really get to understand the characteristics of thee seris very well.

One thing that always troubled me in the early years was that the Box-Jenkins methodology was revered a little too much. Estimation is by conditional least squares and so the normality of the residuals is important and often overlooked (a buried secret). In the late 1970s i work on the problem of outliers in time series and Darryl Downing and I published a paper on the topic in JASA in 1982.

Since then other like Doug Martin, George Tiao and Ruey Tsay have made much bigger contributions. IrishStat is aware of that literature and has incorporated their ideas in his software. That is why he emphasizes checking for level shifts and outliers before fixating on an ARIMA model. That aspect of his software makes it somewhat unique. It is different from auto.arima and SAS/ETS. So keep that in mind in your search for other automated procedures using SAS.

I hope you appreciate this as an answer even though it does not directly answer questions 1 or 2. I am sure you can find Terry Woodfield on the internet or go directly to the SAS Institute with your questions which are very specific to SAS and really require someone with intimate knowledge of the SAS algorithms. I don't think you will find anyone on this site who could give you better help.

  • $\begingroup$ Thank you very much for your detailed response! It's really helpful. I agree that an automatic approach may not be good for someone who want to understand the data. On the other hand, I guess the decisions at different stages of the analysis really depend on personal taste, and maybe that's why it's not easy to build an automatic process for everyone. Actually that's one of the reason why I don't like time series or even statistics in general. A lot of times, you just make the decision by experience or just with your gut, and there seems no scientifically standard criterion. $\endgroup$ – EXP0 Jul 27 '12 at 13:57
  • $\begingroup$ I agree with everything you said up until the last sentence. Much of statistics is well grounded in mathematical theory both in probablity and in statistical inference. $\endgroup$ – Michael R. Chernick Jul 27 '12 at 14:27
  • $\begingroup$ Yes, I agree that there are a whole system of statistical inference behind the theory. Maybe I should rephrase my last sentence. Actually, what I try to say is that a lot of statistical analysis are based on personal choice. For example, normally we think p-value less than 0.05 means significant. However, why 0.05? As far as I know, there's no scientific way to show that whether we should use 0.05 or 0.03 or 0.01, they are mainly based on personal experience and choice. Not sure if that's good or bad though. :-) $\endgroup$ – EXP0 Jul 27 '12 at 22:54
  • $\begingroup$ The 0.05 was chosen by Fisher but he never intended it to be cut and dry. You are absolutely right that to always use 0.05 to test for significance is too rigid. Maybe what you are uncomfortable with is making decisions under uncertainty. This is something we do every day in informal ways with most of the decisions in life and almost nothing is certain (especially the future).There is value in formalizing this process. A lot of what statistics is about and time series especially is finding the signal in the noise. it is there but is not always easy to pull out with the naked eye alone. $\endgroup$ – Michael R. Chernick Jul 27 '12 at 23:41

I don't think there is BUT even it were there you need to consider the impact of outliers, level shifts, seasonal pulses and local time trends in the model identification stage. You might want to look at http://www.autobox.com (a piece of software that I helped write ) as a viable alternative. It automatically delivers a final model based upon checking statistical significance and the Gaussian Assumptions.

  • $\begingroup$ thank you for your response! It looks like a great product! $\endgroup$ – EXP0 Jul 27 '12 at 14:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.