0
$\begingroup$

I have a few time series (20+) to be jointly forecasted. They are minute-level data with an obvious weekly seasonality. Therefore the seasonality is 7*24*60=10080.

Typical time series model can not handle such a long seasonality, since it requires a huge number of seasonality parameters (I tried DLM and some similar models).

Most time series models require deseasonality before feeding, is there any way to deal with this long seasonality?

$\endgroup$
  • 1
    $\begingroup$ Here is Rob J. Hyndman's blog post specifically on this topic. Search also for related threads on Cross Validated, something like these. $\endgroup$ – Richard Hardy Jun 10 at 17:06
  • $\begingroup$ Thanks @RichardHardy! That's definitely a good start point. $\endgroup$ – kaixu Jun 10 at 20:50
  • $\begingroup$ I'd be thinking about looking at trigonometric seasonality not dummy seasonality. You can probably get an adequate fit with only a few components. Do you seem to have stationarity across weeks? $\endgroup$ – Glen_b Jun 11 at 3:27
  • $\begingroup$ @Glen_b Yes a fourier series seems to be able to capture the seasonality pretty well. $\endgroup$ – kaixu Jun 11 at 18:19
1
$\begingroup$

I would consider possibly using deterministic structure such as 6 dummies for day-of-the-week , 23 dummies for hour-of-the-day and arma structure for short term memory. How to build ARIMA model from my time series? might be of interest

$\endgroup$

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.