Cats vs Eviews

A forum for questions and answers regarding the CATS cointegration procedure

Cats vs Eviews

Unread postby randerson » Mon Sep 14, 2015 4:26 pm

I have a co-author who has been estimating VECM/CI models in Eviews. For forecasting, I prefer Rats. I find that VAR estimates are identical in the two packages but VECM estimates differ. Coefficient estimates differ in the first digit, although *most* coefficients that are significant in Eviews also are significant in Rats. But some coef that are significant in Eviews are not in Rats. I talked with staff at Eviews who claim their estimation is the Johansen MLE. Does anyone have experience in these packages?
randerson
 
Posts: 14
Joined: Wed Dec 01, 2010 12:24 pm

Re: Cats vs Eviews

Unread postby TomDoan » Mon Sep 14, 2015 5:56 pm

The one thing that I've run into is that EViews includes an unrestricted constant in a VECM where the cointegrating vector is estimated with a restricted constant (DET=RC in CATS).
TomDoan
 
Posts: 7156
Joined: Wed Nov 01, 2006 5:36 pm

Re: Cats vs Eviews

Unread postby randerson » Mon Sep 14, 2015 9:48 pm

Tom, thanks for the reply. But both Eviews and Rats allow models with no intercept (and/or a variety of other deterministic components). Further experiments, however, reveal a simple (embarrassing) answer: the lag length in Rats refers to the "k" in the VAR equation, while the lag length in Eviews refers to the "k-1" in the ECM form. Specifying the lags correctly obtains coef estimates, in the two packages, extremely close to each other. Standard errors still diff slightly, but not enough to affect inference regarding coefficient significance. The Eviews manual says its estimator for the var-cov matrix uses T-k. I assume the Cats code follows, for example Juselius (2006), in using T as the divisor. This requires further exploration.
randerson
 
Posts: 14
Joined: Wed Dec 01, 2010 12:24 pm

Re: Cats vs Eviews -- further experiments

Unread postby randerson » Wed Sep 16, 2015 4:36 pm

The Eviews manual notes that their variant of the Johansen MLE forces residuals in the CI regression always equal to zero.

I find that VECM/CI models estimated in Eviews and Cats can be quite different. Researchers likely are not aware of the difference.

In experiments with a VAR(4), I noticed these results:

1) No intercepts/trends: With no exog or dummy variables, EViews and Cats provide the identical parameter estimates. When exog variables are included in Eviews and dummy variables are included in Cats, the parameter estimates from Eviews and Cas are completely different. Not even close. (Odd since the exog/dummy variables enter only the "VAR section" in both packages.)

2) Intercepts in CIV and VAR (no trends): With and without exog/dummy variables, Eviews and Cats provide similar but not identical parameter estimates (differ in second sig digit). They differ enough that "results shopping" seems feasible.

3) Intercepts plus trends are included: Some coeffs are "close" between the packages, some are quite different.

No action item here, just FYI for forum members.
randerson
 
Posts: 14
Joined: Wed Dec 01, 2010 12:24 pm

Re: Cats vs Eviews

Unread postby TomDoan » Thu Sep 17, 2015 9:51 am

Can you send an EViews file with the data and models that you're running? (Either attach it here or send to support@estima.com)?
TomDoan
 
Posts: 7156
Joined: Wed Nov 01, 2006 5:36 pm

Re: Cats vs Eviews

Unread postby randerson » Thu Sep 17, 2015 3:22 pm

yes, will do, but it will be a week or so before I can return to it.
randerson
 
Posts: 14
Joined: Wed Dec 01, 2010 12:24 pm


Return to CATS Questions

Who is online

Users browsing this forum: No registered users and 1 guest

cron