Diebold Mariano Loss Differential

[nikieb]

A very short, and probably simple question.

I have estimated the RW, ARIMA, VAR and VEC models according to all criteria to forecast inflation with eviews 7.

My next step is to compare the models using the Diebold Mariano test.

I have manually calculated the mean loss differential of 8 periods ahead, and I am interested in t=1 t=4 and t=8.

Now i have to regress a one period differential against a constant to find the t-stat, but i honestly do not know how i can regress only 1 period. E.g. only quarter 1 4 or 8. I calculate everything via the menu, not via a written program. What is the sample range? Should I try sample "3 4" for quarter 4? Filling in quarter 4 4 does not work.

I know the question sounds silly after doing all the prior work, but it just does not work out.

Thank you!

[EViews Glenn]

I suspect the reason there hasn't been an answer to this question is that there's really not enough information on the kind of data that the user has, nor is it clear (at least to me), the exact task to be performed. If you could provide additional information, that might help.

[nikieb]

Hi Glenn,

I am forecasting the inflation rate for Japan. I use four models for this, the VAR VEC Random Walk and ARIMA model. I have done this already. To be able to compare the models in terms of performance, I want to perform the Diebold Mariano test. This test is a regression of the loss differential to a constant, and I have calculated the loss differential (difference between the error terms of 2 models). I have forecasts for 8 subsequent quarters, and q 1 4 and 8 are of interest to me.
My question is: how can i regress one specific quarter? How should I enter this in Eviews? When estimating an equation, should the time frame (sample) for quarter 1 be "0 1" and for q4 and q8 "3 4" and "7 8"?

I only want to use the number associated with one particular quarter for regression, but I dont know how i can only include this single number, that is my problem.

[EViews Glenn]

If I understand correctly, just set the sample of observations to be used prior to performing a specific regression. Something like

Code: Select all

smpl if @quarter=1 equation eq01q1.ls y c x

or

Code: Select all

smpl if @quarter=1 or @quarter=3 equation eq01q1and3.ls y c x

For details, look at the discussion of the smpl command in the docs.

[nikieb]

Dear Glenn,

I think that you indeed understood what i meant.
However, now I receive an error that there are no observations and another error. It really must be one of the simplest things to do in Eviews. I have attached an example of a workfile, it is '"av" that I am trying to regress against a constant. For time 1, 4 and 8.
Could you please help me out?

[nikieb]

It worked for the first quarter just now, but when I reach t=8, i receive the message 'insuffient number of observations'. There are 2 observations required for every regression, so for t=1 is quarter 1 and 2, for t=4 is quarter 4 and 5, and t=8 quarter 8 and 9? If so, I dont have a ninth observation.

[nikieb]

Basically I just need the t-statistics of quarter 1 4 and 8 individually, that is all.

[EViews Gareth]

Your posts are very confusing.

Glenn's answer assumed that you had dated data (monthly or weekly). You do not. Your workfile has 8 observations. I'm assuming they are ordered by the "quarter" variable. Glenn's smpl if @quarter suggestion only works on dated data.

You say you want to run a regression for quarter 1. You only have a single observation when quarter=1. You want to run a regression with a single observation. That makes no mathematical sense. Even if you are just regressing against a constant, all you will do is set the coefficient identically equal to the value of the dependent variable. Equally, you only have one observation for 4 and 8. You cannot run regressions with only a single observation.

[nikieb]

The data included is indeed quarterly data, and it is the average loss differential. The purpose of the Diebold Mariano model is to compare 2 models and to determine which is best in forecasting. I have to find the accompanying t-statistic of the loss differential at those different time frames (t=1,4, at a 5% significance level.

If a regression does not make sense, do you a suggestion how it could be done?