Implement tests for evaluation of forecasting models

[Jelena]

Hello,

I am trying to compare the performance of a univariate to a model that nests the univariate model and is extended by a macroeconomic variable.
The t-test is not appropriate for such nested models. Thus, I would like to use the Clark and West (2007) test and the Giacomini and White (2006) test.
Does somebody has an idea how I could programme this in Eviews?

I am happy about every answer and hint!
Thanks in advance!

Jelena

[Jelena]

Here is an explanation (by Paye) of the procedure of Clark and West (2007)

[trubador]

I do not think there is anything further to add or say on the explanation you have posted other than writing the code for you. It is pretty clear, simple and straightforward...

[Jelena]

For you it might be obvious, for me there is nothing in eviews that is "pretty clear, simple and straightforward..."
I just started using eviews 3 weeks ago and besides writing my paper have to come along with it by myself.
Besides the user eviews command reference there is nothing really helpful for writing a program. All books available just illustrate how to use the eviews user interface.

[trubador]

That's the thing. You do not actually need a program. Only a few commands would suffice. Assuming you have the forecasted values, all you need is a few descriptive statistics, which are quite user-friendly and easy to find and run in EViews (see the Help files). It is not that I am reluctant to or refrain from helping you. If you "really" want to learn EViews, you should become capable of doing such basic stuff on your own. And believe me, having someone doing it for you is not the best way to learn. Note that EViews has excellent learning resources other than its built-in Help system.

[Jelena]

Thank you for the response!
I tried two ways, what I think could be correct.

First:

!window = 180
smpl oos
genr MSPE1 = @sumsq(fcast1 - y)/!window
genr MSPE2 = @sumsq(fcast2 - y)/!window
genr MSPE3= @sumsq(fcast1-fcast2)/!window

genr CW= (MSPE1-MSPE2+MSPE3)*1000

This gives me a series with the same value for every month in my sample. Although I wanted to compute a single value... However, Is that a correct way I can do it?

After I found the follwing hint in Payes paper in a foodnote: "Computationally, the test can be executed by regressing the adjusted out-of-sample squared forecast error differences on a constant and examining the associated t-statistic." I tried to do this in the following way:

smpl oos
genr MSPE1a = ((fcast1 - y)^2)/!window

genr MSPE2a = ((fcast2 - y)^2)/!window

genr MSPE3a = ((fcast1-fcast2)^2)/!window

genr CWa= (MSPE1a-MSPE2a+MSPE3a)*1000

freeze(ls_dya) eq1.ls cwa c

I am not sure whether both methods give me the same level of significance.

I will think tomorrow about what the difference in econometric terms is. But if someone has an idea, which one is right I would be happy about every hint!

Thanks in advance