Tom, there is some version of the Giacomini-White test in RATS?.
Best regards.
W
Giacomini-White test
Re: Giacomini-White test
If I'm reading the paper correctly, the calculations are basically the same as Diebold-Mariano (they use a chi-squared distribution rather than Normal to allow for multiple targets). What is different is the preparation and the asymptotic theory. In order to apply the asymptotics, you need the forecasts to be computed with rolling estimates using a fixed width estimation window. The fixed width window keeps the estimates for a correct model from collapsing to a point asymptotically. Seems more like it should have been in JOE than Econometrica.
Re: Giacomini-White test
Tom, thank you very much for the answer, I think it happens only under some circumstances that GW seems DM, but the main point of the test is that it only works with a fixed window, that make it not useful . In my case I have recursive pseudo real-time forecasts in a VAR's compared against some ARIMA's, under these circumstances I think the test that I should use is Clark and McCracken, am I right?.
It´s possible to have a test version of Clark and McCracken in which the forecasts are introduced directly. In the original version (I think posted by the authors) one should do the regressions and forecasts within the routine, is it possible to separate from each other as in the version of DM that is in RATS?.
Best regards
W
It´s possible to have a test version of Clark and McCracken in which the forecasts are introduced directly. In the original version (I think posted by the authors) one should do the regressions and forecasts within the routine, is it possible to separate from each other as in the version of DM that is in RATS?.
Best regards
W
Re: Giacomini-White test
I don't know if you've seen it, but Todd Clark has a fairly long post in the thread
http://www.estima.com/forum/viewtopic.php?f=7&t=742
which describes the difference between Clark-McCracken and Giacomini-White.
First of all, a VAR and an ARIMA model (with MA's) don't nest, so you can use D-M for those. I'm not 100% sure what the effect is on any of these tests for using pseudo-real time data. However, that would also tend to make the test less subject to the nested model problem since the data isn't settling down the way it would with a typical increasing sample.
http://www.estima.com/forum/viewtopic.php?f=7&t=742
which describes the difference between Clark-McCracken and Giacomini-White.
First of all, a VAR and an ARIMA model (with MA's) don't nest, so you can use D-M for those. I'm not 100% sure what the effect is on any of these tests for using pseudo-real time data. However, that would also tend to make the test less subject to the nested model problem since the data isn't settling down the way it would with a typical increasing sample.