EViews.com

Dear Eviews Team

I have a monthly series for one of the component of the aggregate CPI and I simply took the log difference *100 in order to get the inflation for the component and called it "cinf" as in the workfile attached to this question. I used Gareth's automatic ARMA selection code to find the best model and I settled with ARMA(8,(8,12)) which yields uncorrelated residuals. This selection yields the lowest AIC but the coefficients of AR(4), AR(5), AR(6) and MA(4), MA(5), and MA(6) are insignificant. When I removed these terms from the estimation (but left the rest untouched), the coefficients of the AR and MA's (except AR(1), MA(1), MA(12)) all become insignificant. When I remove these terms, the remaining ARMA(1,(1,12)) yield significant coefficients, but the residuals are correlated and R-squared is 23%. I am quite not sure how to interpret these changes and find out the best model.

My overall goal is to retrieve the residuals from the best ARMA model and use them as the unexpected inflation. I have been dealing with this issue for long time and I am really stuck. I attached my workfile and a simple program and I would appreciate any help significantly.

Thanks in advance

I just wanted to add a comment that this is my third attempt posting this question. I posted it each time in Estimation and then later it appeared in Econometric Discussion by itself (a built-in filter system maybe?). I removed the posts from the Econometric Discussion section and now it only appears in Estimation.

Thanks

This post belongs in the Econometric Discussions forum.

This post belongs in the Econometric Discussions forum.

I apologise for moving back to the estimation then. I was very puzzled by why it kept coming back to estimation. Do you mind sharing your thoughts on my original inquiry? I am sorry agan for posting in the wrong forum.

We don't give out econometric advice

We don't give out econometric advice

Gareth,
I still appreciated your taking time to read my post and to move it to the appropriate section. This forum has been a great source of knowledge for me and I thank you and other Eviews moderators for your hardwork.

Here are a few thoughts.

1. How do you define best?
2. Be sure that if you're doing significance tests, that you're doing joint tests rather than tests on individual coefficients.
...

3. Unfortunately this may not be a good way of measuring unanticipated inflation, as it depends on the econometrician having the same information set as whoever is doing the anticipating.
...

4. You might try a few different methods and save the residuals from each. If the residuals all look about the same, then it doesn't matter too much which method you use.

Thanks startz for your feedback. I was simply looking at the out-of-sample forecast to determine which model would fit the best. ARMA(1,(1,12)) has a very poor fit with correlated residuals and I was trying to come up with a better model.

I was actually advised to use the Kalman filter to extract the unexpected inflation from the inflation series, but I am not completely certain about how to use the Kalman filter (I already started exploring it in Eviews though). Do you also think that the Kalman filter is the most appropriate way to extract the unexpected inflation?

Thank you in advance for your help.

The issue isn't really whether you use a Kalman filter or an ARMA model, the most important issue is what predictor variables you include.