I am new to ARIMA modelling. I understand most of the basic concepts, and I've read a lot of topics about ARIMA on this site.

At present I am pretty comfortable with analysing ACF and PACF graphs, checking for seasonality and creating model from my data.

Next step I want to deal with outliers and periodic pulses in my data. So, how do I do that using R? Should I just create some dummy variable and code periodic pulses with, for example 1, other outliers with 2 and everything else with 0, and then pass this variable to "xreg" option in arima command ?

I am also afraid of mixing things up by not understanding terminology correctly. There is a lot of talk about pulses, level-shifts, seasonal pulses, local time trands, etc. Is there any guide or paper discussing all this stuff and how to deal with it ?


If you want to empirically identify Pulses, Seasonal Pulses, Level(Step) Shifts and/or Local Time Trends you might want to look at How do I detect shifts in sales mix? or Detect changes in time series or Outlier detection for generic time series . Some commercial packages offer Intervention Analysis which does not include Intervention Detection which is what you are pursuing.

  • $\begingroup$ thank you for your answer, and links. But, for more clarity, let's take some example. For instance currency intraday trading. There is always huge gap between last price on Friday and first on Monday. I know it's always there, and I whant my arima model to deal with it. How should I treat such case? $\endgroup$ – GrayR May 8 '12 at 13:36
  • 1
    $\begingroup$ Dave Reilly is the expert on this. He is being modest by not mentioning that his product autobox does this automatically for you. it will save you a lot of time by not having to reinvent the wheel in R (assuming R doesn't already have a package) but you do have to license it. $\endgroup$ – Michael Chernick May 8 '12 at 15:33
  • 1
    $\begingroup$ @GrayR If there is exceptional movement from Friday to Monday this ( in the absence of a "cause series" ) could be considered a Pulse or a one-time outlier/inlier. Intervention Detection is based upon an in-model evaluation as compared to an out-of-model study of the current model residuals. The advantage of the in-model approach is that the "unusual value to be detected" doesn't adversely impact the standard deviation that is being used to test significance. I would be glad to talk to you about this further. $\endgroup$ – IrishStat May 9 '12 at 0:09
  • 2
    $\begingroup$ I should add if this activity i.e. a positive change between Friday and a Monday is persistent then this would identified as a seasonal pulse and therefore a predictable activity. $\endgroup$ – IrishStat May 9 '12 at 0:17

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.