0
$\begingroup$

I have a series with double seasonality, one daily and one weekly. I modelled the daily seasonality with SARIMA, and I want to include weekly dummy variables.

I only included weekday variables for the times when spikes occur in my data, and dropped the rest. That's how a long term forecast looks with this model:

enter image description here

Now my problem: it isn't supposed to have an upward trend like that.
If I don't include the weekday dummies, it stays on the same level (which I want), but with weekday dummies, the above happens and I have an upward trend (in the spikes).

I don't know why that happens. Any clue why that happens/how I can fix that?

edit:

Regardless of the exact SARIMA specification, there is always no trend without dummies, and always a trend when i include "variablewochentag" (German for weekday variable). There are 2 spikes per day in my data, and my data is Mo-Fr, so i have 10 weekday variables. There is also a seasonal difference (65 = daily seasonality) in my estimation, and i have no constant. I have 2080 observations which i used for estimation, added weekday dummies for future observations and then have run the forecast which results in the above picture.

Estimation Output: enter image description here

edit2:

http://www.filedropper.com/data_10 Here is my data. Note: the first 2080 data points are my actual observations, the rest are just zeros + the relevant variables to do the long term forecast. Seasonality is 65 for daily, 325 for weekly.

|cite|improve this question
$\endgroup$
  • $\begingroup$ Could you include the estimated model coefficients? $\endgroup$ – Richard Hardy Dec 5 '15 at 18:51
  • 1
    $\begingroup$ please provide the data ...... and all meaningful output . You probably have a differenced model with a steady state positive constant $\endgroup$ – IrishStat Dec 5 '15 at 18:54
  • $\begingroup$ @Richardhardy i included the estimation output. $\endgroup$ – Jariel Dec 5 '15 at 23:05
  • $\begingroup$ @IrishStat i included the estimation output. The model is differenced (seasonally), but it doesn't have a constant. The trend doesn't exist if i don't include my dummy variables. $\endgroup$ – Jariel Dec 5 '15 at 23:06
  • $\begingroup$ also added my data. $\endgroup$ – Jariel Dec 6 '15 at 1:31
1
$\begingroup$

Your model is way over-specified (kitchen-sink modelling) probably because of your reliance on a simple statistic like AIC/BIC which premises no outliers/level shifts/time trends/constant parameters/constant error variance and many other things. If you have redundant/cancelling parameters like AR(65) and MA(65) as you do all bets are off with the normal expectation of stable (i.e.non-trending) forecasts. Furthermore if your coefficients are non-invertible or nearly non-invertible like your AR(1) and AR(2) this can lead to "explosive forecasts" . Consider an AR(2) model with coefficients .6 and .5 ....which is non-invertible because the coefficients sum to more than 1.0 , you get explosive forecasts . KISS is the order of the day. Model identification is important.

|cite|improve this answer
$\endgroup$
  • $\begingroup$ even if i only include Ma(1) and SMA(65), i still get an "explosive forecast" if i also include my dummy variables. $\endgroup$ – Jariel Dec 7 '15 at 9:17
  • $\begingroup$ If i include a constant in my model (together with my seasonal differencing), i also get a trend. Why is described here: robjhyndman.com/hyndsight/arimaconstants Could there be the same problem with dummies? $\endgroup$ – Jariel Dec 7 '15 at 10:34
  • $\begingroup$ I think i know now why i get a trend. I need to difference my dummy variables which i haven't done. $\endgroup$ – Jariel Dec 7 '15 at 10:55
  • $\begingroup$ That is possible .... $\endgroup$ – IrishStat Dec 7 '15 at 13:24
  • $\begingroup$ Any luck ... I doubt it .. $\endgroup$ – IrishStat Dec 8 '15 at 9:48

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.