I'm not sure at what stage the development of Eviews 10 is but here are some suggestions re: VECM (i.e. the Cointegrated VAR). Those stem from my fairly advanced knowledge and practice of this class of models. Allow me to put in my comments along each step of the standard procedure for estimating CVARs.
To start, there are three important considerations: (1) choice of the lag length, (2) choice of the sample and (3) choice of deterministics (dummies).
1. VAR lag length
In practice the usual information criteria (AIC, HQ, SC...) are used to select the lag length. The lag length is an important choice, yet in practical applications the lag length chosen by the minimization of alternative information criteria is often different. This is of course expected. I'm no expert in this area but I'm sure papers have been written on VAR lag length selection... A forthcoming book by Kilian and Lutkepohl (2017) discusses this at length (pun intended).
Suggestion#1: Still, I suggest that more information criteria be introduced in order to choose a lag length more confidently. Maybe introduce "modified" versions of those criteria?
Suggestion#2: Introduce systems reduction tests --is a VAR(4) adding significant information compared to a VAR(3)? Compared to a VAR(2), etc...
2. VAR sample choice
Quite often we forget that all inference (including lag length selection) is based on a correctly specified model. Here I mean that the lag length should be treated as a parameter. Using the whole sample just because we have data is nonsensical unless the estimated parameters are stable. Therefore, system-wide stability tests for VAR models are required. Eviews does not offer those, to the best of my knowledge; it's a shame.
Suggestion#3: Introduce system-wide VAR stability tests.
3. VAR deterministics (dummies)
Dummies should capture large, exogenous events (Juselius 2006). "Exogenous" is in the eye of the beholder to some extent, but "large is not". For instance one can decide that a residual is large if its value is greater than 3 times the residual series standard error.
Suggestion#4: Code some kind of procedure for finding large residuals.
4. Model check
Once a base VAR model has been estimated it should be checked for residual normality. Eviews provides all the tools for that at the moment (I don't have any suggestion for improvement in that regard).
Next comes cointegration testing.
5. Johansen cointegration
Here there are two areas where Eviews could improve: (1) small sample correction factor --because in empirical work everything is a small sample and (2) much improved handling of deterministics and (weakly) exogenous I(0) or I(1) variables.
Suggestion#5: Introduce a small-sample Bartlett correction factor (Johansen has written on this)
Suggestion#6: Bootstrap everything instead of relying on asymptotic inference.
Next comes the identification of the long run and short run structures. I'm fine with Eviews powers in this respect.
Those are my thoughts for now. I may jump in and provide some further toughts.
For making suggestions and/or requests for new features you'd like added to EViews.
1 post • Page 1 of 1
Who is online
Users browsing this forum: No registered users and 1 guest