on Diebold Mariano automatic procedure

[lukeroom]

Suppose I enter two time series *equations* to do a horserace in forecasting in the forecasting window (the window where you can do the Diebold-Mariano test). I let eviews do the forecast and do the DM test. BUT! the forecasts that Eviews does are a sequence of one-step ahead, two-step ahead, .... How can you compute the DM test, where one needs to sum a sequence of, say, one-step ahead forecasts. It seems to me that Eviews computes the DM test wrongly, by summing up apples (only one one-step ahead) and bananas (only one two-step ahead) and pears (one three-steps ahead etc...)
Am I reading correctly what eviews is doing or am I missing something?

Thank you for your help
Luca

[EViews Gareth]

Can you explain exactly how you are performing the DM test?

[lukeroom]

I am just using the automatic procedure in the "forecast evaluation" window.
From that, a DM test is generated (reading the help, for a one step ahead forecast).
But when I look at the figure generated by that forecast, I see that it is a sequence of one-step ahead, two step ahead, three step ahead....

The question is what forecasts are eviews using to compute the DM statistic?
Thank you
Luca

[EViews Gareth]

Are you entering equation objects, or series objects?

[lukeroom]

Are you entering equation objects, or series objects?

equation objects

[EViews Gareth]

Then, yes, you are correct, if that equation is dynamic.

You'd need to compute the one-step ahead forecasts, save them into a series and then enter the series as the forecast comparisons, not the equations.

[random_user]

I agree with this. I think essentially what he/she is saying is that the forecasts generated when you run the forecast evaluation are dynamic (rather than static) forecasts. This is problematic for equations that use lags of the dependent variable (which most forecasting equations do).

For example, let's say you're evaluating two forecast models over the period 2012Q1 to 20222Q4.

In the first period (2012Q1), the lagged dependent variable is observed (it's the value at 2011Q4). But in the second period (2012Q1), the lagged dependent variable is instead replaced with the forecast value rather than from the historical data.

What's the problem? Well, yes that makes the initial forecast for 2012Q1 a 1-step ahead forecast, and the subsequent forecast for 2012Q2 a 2-step ahead forecast. And indeed, none of the diagnostics and forecast evaluation tools make sense anymore, because they're no longer apples-to-apples comparisons.

In general, for forecast evaluations, it's better if you code them yourself, because no matter what EViews does on its own, chances are it won't work for your needs.

That said, there's huge amount of room for improvement for EViews here. It would be nice to have options to do out-of-sample vs. in-sample, for example. Another is to be able to test different forecast horizons, or to be able to at least switch to static forecasts.

[lukeroom]

Ok, thank you.
I suggest for the moment to change the help to clarify this issue.

Luca