1
$\begingroup$

My issue is really simple: I need to compute a seasonal arima model on traffic data (5 min frequency). The data exhibits daily seasonality (288 observations).

This is causing me issues in computing the model using R. (SO question: https://stackoverflow.com/questions/30804281/r-arima-method-blocks-when-adding-seasonality)

I know that seasonal arima models are not particulary suited for long seasonality periods, however I read tons of articles regarding traffic forecasting that exploits them in order to make predictions.

I will appreciate any suggestion regarding software tools (in addition to R) that would do the trick. Thank you.

| cite | improve this question | |
$\endgroup$
  • $\begingroup$ Have you tried using option method="CSS" in the arima function? Perhaps that could reduce the computational burden. $\endgroup$ – Richard Hardy Jun 17 '15 at 14:11
  • $\begingroup$ @RichardHardy yes, I have already tried it without success. Someone also suggested to add the "optim.method" option to the method, same negative result. $\endgroup$ – riccamini Jun 17 '15 at 14:20
  • $\begingroup$ I tried a SARIMA(2,0,2)(1,0,1) with 288 period on a 10,000-long series, and that was quite memory-consuming, unfortunately... $\endgroup$ – Richard Hardy Jun 17 '15 at 14:20
  • $\begingroup$ Yes I know. For this reason I am working on a 16Gb machine and I have managed to avoid swapping. However the lack of memory doesn't seem to be the issue, as I get a "optim error" from R $\endgroup$ – riccamini Jun 17 '15 at 14:30
  • $\begingroup$ Have you tried using the Fourier series as recommended? If so, further try the lower case arima command and see if you can successfully choose an order for the arima sequence parametrically. $\endgroup$ – RegressForward Jun 17 '15 at 18:14
0
$\begingroup$

Care should be taken when dealing with daily data to properly/possibly include both deterministic (daily effects) and auto-projective structure (memory) while incorporating pulse/level shift and local time trends AND any necessary transformation required to ensure a homogenous error variance. In my opinion (somewhat biased based upon my subject/domain knowledge and partial developer) would be to use AUTOBOX http://www.autobox.com/cms/ . If you wish you can post your data and I will try and help.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ First let me thank you for the suggestion, I will look at it. For the data, since I am not the owner and it's not public, I think I cannot post it. $\endgroup$ – riccamini Jun 17 '15 at 14:22
  • $\begingroup$ You might doubly scale/transform the data before you post it. Simply subtract/add a constant and then divide/multiply by yet another constant to effectively mask the data.. If you don't want to post it publicly I would be glad to help you as I am always looking for data sets that can teach me or that i can learn from.. $\endgroup$ – IrishStat Jun 17 '15 at 14:30
0
$\begingroup$

In terms of decomposing timeseries the tool that comes to mind is X-13 ARIMA-SEATS Seasonal Adjustment Program by the US Census Bureau, which is available on multiple platforms. Take a look:

https://www.census.gov/srd/www/x13as/

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ I've just read the documentation, it has limits on both the maximum seasonal period (12), and series length (780) that my case exceeds. Thank you. $\endgroup$ – riccamini Jun 17 '15 at 20:07

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.