0
$\begingroup$

Currently, I am working on forecasting. I try various models. One of the models I tried is an AR model. As I have monthly data, I use the 12th period back in time. So the model is like Y(t)=Y(t-12)+e.

As Eviews9 contains helpfull tools to estimate models automatically, I requested a demo. Today, I could make use of eviews9. However, my syntax code do not result in the same output compared to eviews8.

So my questions are the following:

  1. Why is the output different (displayed below), while my input was exactly the same?
  2. How can I interpret the SIGMASQ?
  3. As you can see, the number of included observations differs between eviews8 and eviews9. However, in both programs, I restricted the sample by using this command:smpl 2002m1 2011m12. Why the samples are not the same in the output, as the restriction should be the same in both cases?

The syntax I used in both programs was: LS(DERIV=AA) num10994 ar(12)

Outputs:

Output Eviews8

Output Eviews9

$\endgroup$
  • 1
    $\begingroup$ the default estimation approach seems to have changed from least squares to ML $\endgroup$ – Christoph Hanck Sep 9 '15 at 15:47
  • 1
    $\begingroup$ Your question about interpretation (2) is on topic here, but questions about how eviews works are off topic. Check the documentation or contact their tech support. $\endgroup$ – gung - Reinstate Monica Sep 9 '15 at 16:08
  • $\begingroup$ another thing that differs is the sample used - the second already starts in 2002 already. I am not sure whether that is because of the estimation approach, but another sample may of course yield other estimates. You could set the samples to match and see if differences persist. $\endgroup$ – Christoph Hanck Sep 10 '15 at 6:11
0
$\begingroup$

It appears to me that the ar coefficient (1.102) in "8" is not invertable as it exceeds 1.0 . It was estimated using conditional least squares. If you use maximimum liklehood (as they did in version "9" ) you can control the sample space ( i.e. valid range for the coefficient ) thus .8223 . Notice the reduction in R_square and the proportionate increase in error variance.

Version 8's coefficient suggests that things are growing by 10.2% over last yeaar at this point in time. Such models are formally outside the range of valid Box-Jenkins models but I have my doubts as "growth models" are commonplace.

$\endgroup$

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.