I am working on a project where I am to do the intervention analysis and forecasting based on the time series. The problem is something like:

I have a normal time series entries but in between them some known event like natural calamities (storm, tornado) happens. I have the data for that and it affects the normal time series. Now my objective is to forecast the value of time series both in normal mode and also when I have a prediction of storm coming.

I have been reading Forecasting with dynamic regression chapter 7 about intervention analysis. I am also reading about the transfer function modeling. Can you please help me as in which model is good for this kind of time series analysis? Or may be some link which can guide me as how to do it? I will appreciate a link with some example in R or some examples.

EDIT: I guess I was not correct in description but I know the exact time information of all the previous storm events and I sort of want to find out the effect of storm intervention on the time series and I can forecast more closely if I know that there is a storm happening right now.

  • 3
    $\begingroup$ R does not have transfer function modeling capability, see this, this examples. $\endgroup$ – forecaster Jun 9 '15 at 19:36
  • 1
    $\begingroup$ Doesn't the TSA package have the transfer function modeling capability? The arimax function of cran.r-project.org/web/packages/TSA/TSA.pdf seems to show that they do internally transfer function modeling. $\endgroup$ – Naman Jun 9 '15 at 19:56
  • 2
    $\begingroup$ Ok, try it out. According to this website from Rob Hyndman "Cryer and Chan also ignore the CRAN pack­ages, and pro­vide their own R pack­age (TSA), but it is poorly writ­ten and I would not rec­om­mend using it." and also it is important to note that TSA doesnt allow you to forecast, it simply fits a model to the data. $\endgroup$ – forecaster Jun 9 '15 at 20:15
  • $\begingroup$ @forecaster Ohh I didn't know about that. What model and package do you recommend for this kind of analysis? I truly appreciate your help and time. I am fairly new in this field and I am not even completely sure that transfer models are the best thing to do this kind of stuff. It's just I didn't want to go in completely wrong path that's why I asked this question. All your thoughts and ideas are very much welcome. $\endgroup$ – Naman Jun 9 '15 at 20:56
  • $\begingroup$ As far as I am aware, in SAS/SPSS you can build custom transfer function models. In SAs you can easily visualize the transfer function and the notation is very simple. See this SAS conf. paper that was published in 1987! has all the transfer function examples you need. $\endgroup$ – forecaster Jun 9 '15 at 21:07

It is one thing to do intervention modelling , it is another thing to do intervention detection before you do intervention modelling. Intervention detection is discussed nearly everywhere ( search this site ) and is neatly summarized by Tsay http://ri.search.yahoo.com/_ylt=A0LEVjJuP3dVXdEA7uYPxQt.;_ylu=X3oDMTEzNDgyZmhhBGNvbG8DYmYxBHBvcwMxBHZ0aWQDRkZHRUMwXzEEc2VjA3Ny/RV=2/RE=1433907183/RO=10/RU=http%3a%2f%2fwww.unc.edu%2f~jbhill%2ftsay.pdf/RK=0/RS=0cgymNUizBmJRyvOEdid7Pb4dQU-. Detecting interventions in causal models is a little bit tricky requiring specialized commercial softare as the form of the causal model structure also needs to be simultaneously identified.

  • $\begingroup$ I am sorry but I am unable to understand as why is it intervention detection? I guess my problem details are not clear. I know the time of all storm events in the past. So I wanna fit that with my time series and then with the new model I want to forecast if I have a storm coming in near future. Am I on wrong path? I appreciate your help a lot. $\endgroup$ – Naman Jun 9 '15 at 19:53
  • $\begingroup$ If you know the points of the intervention then there is no need to identify them. If however you don't know BUT wish to find data points that have been interfered with then one need to identify those points ( if they exist ). Intervention variables may be pulses (in your case) or level shifts or seasonal pulses or local time trends. $\endgroup$ – IrishStat Jun 9 '15 at 22:49
  • $\begingroup$ Exactly, I don't want to find out the point of interventions. I sort of want to create a model that has point of intervention as it's predictors. So my model should know the effects of these interventions. And so when I forecast and I know that a storm is happening right now then the model can give better forecast then the basic model. Can you please point as what are the best models or packages for this sort of work? Thanks a lot for your help. $\endgroup$ – Naman Jun 9 '15 at 22:57
  • $\begingroup$ Forming a model with a deterministic intervention predictor might involve not only using the contemporary effect but there may be lead and lag effects around the known date of the intervention. Additionally there may be auto-regressive memory needed (ARIMA structure) and of course there may be non-constant error variance over time. I have helped develop a piece of commercial software which can be useful in automatically ( or non-automatically) forming a model.It is called AUTOBOX and there are 30 day demo versions . If you post your data I will use it to analyze your data and report results. $\endgroup$ – IrishStat Jun 10 '15 at 0:32
  • $\begingroup$ That's exactly what I have been researching about past some days and till now I learnt that I should be fitting ARIMA model first to data and then use some transfer function to correct for interventions. I would get access to data in couple of days and then I can definitely post some sample, although I think it is rather unlikely that we can use other commercial software than R or SAS. I would still greatly appreciate all your inputs, since I am in researching phase of the project currently. $\endgroup$ – Naman Jun 10 '15 at 14:03

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.