2

I have 4 correlated time series, and I want to predict one of them, from the other 3. There is a clear seasonal effect in the 4 time series, so my first thought was to fit a multivariate ARIMA model, but I can not seem to find an R-function for this.

chl
  • 50,972
  • 18
  • 205
  • 364
utdiscant
  • 1,430
  • 4
  • 19
  • 22

3 Answers3

7

As @IrishStat says, you want a transfer function or ARMAX model. This can be fitted using the TSA package in R.

Rob Hyndman
  • 51,928
  • 23
  • 126
  • 178
  • I think I would need some help working out how to use the function. Should I create new Question for that purpose, and close this thread, or should I ask it here? – utdiscant Nov 14 '11 at 18:38
  • I added a follow up question here: http://stats.stackexchange.com/questions/18375/how-to-fit-an-arimax-model-with-r – utdiscant Nov 15 '11 at 15:00
3

It appears that you want to model 1 of them given the other 3. This is called a Transfer Function and also sometimes an ARMAX model. You will be interested in capturing not only contemporaneous effects but lag effects.The unexplained component ( the current error term) might be further partitioned into some autoregressive structure (ARIMA) and/or sOme deterministic structure ( Pulses, Level Shifts , Seasonal Pulses, Local Time Trends. I am not an R expert but I don't believe that functionality currently exists. You might want to use the internet http://www.google.com/search?sourceid=navclient&ie=UTF-8&rlz=1T4SUNA_enUS287US288&q=multivariate+box-jenkins and to search for "MULTIVARIATE BOX-JENKINS".

IrishStat
  • 27,906
  • 5
  • 29
  • 55
  • After reading your answer, I am not completely sure what I want exactly. I have 4 time series, and I want to be able to predict only one of them. In that sense, I do not actually have the other 3 given (I know that is what I wrote before). So I guess I want to build a model using the four different time series, to predict one of them. – utdiscant Nov 13 '11 at 15:56
  • :utdiscant Let me clarify what I mean by "given". If you assume that you have 1 endogenous series and 3 exogenous series then you can build an ARMAX model. To forecast this one endogenous series you might need future values of the exogenous series depending on the evidented lag structure and the length of the required forecast. The future values of the exogenous series would have to be delivered either as user-specified or the software might self-forecast these series using the appropriate ARIMA structure.If this is still not clear please either arrange for a chat room or contact me via email. – IrishStat Nov 13 '11 at 17:05
1

I am not sure if you are still looking for a solution or not but I think windowing in Rapidminer might work in here. you can give this a horizon value, using that value, you will build your model. for example you will take yesterday values of other 3 variables to predict today's fourth independent variable.

Rio
  • 141
  • 5