N-Step ahead Out-of-Sample-Forecasts

[chrisw]

Good evening everybody,

I have got a problem regarding n-step ahead out-of-sample-forecasting. I would like to perform a (static) n-step ahead forecast without including actuals. Hence, the n-th step value should be derived by using the previously (n-1) forecasted values (no actual values) and so forth. Then, I would like to continue with another n-step ahead forecast after re-estimating the model and including one more observation.

How can I implement this in eViews5 and eViews6?

Thanks everybody in advance!

[EViews Gareth]

Isn't what you are describing just a dynamic forecast?

[chrisw]

Possibly it is.

But how can I perform a dynamic forecast without using state space equations (since I am not familiar with it)? If I apply the forecast command there is no option to chose a value for n. Hence, I just perform a one-step ahead forecast before including the next observation, reestimating the model and calculating the next one-step ahead forecast.

But for my application it is crucial to calculate a n-step forecast before reestimating the model.

Thanks for your help and patience!

[startz]

Possibly it is.

But how can I perform a dynamic forecast without using state space equations (since I am not familiar with it)? If I apply the forecast command there is no option to chose a value for n. Hence, I just perform a one-step ahead forecast before including the next observation, reestimating the model and calculating the next one-step ahead forecast.

But for my application it is crucial to calculate a n-step forecast before reestimating the model.

Thanks for your help and patience!

The forecast command lets you specify a SMPL for the period to be forecast. Perhaps that will help.

[chrisw]

Thanks a lot for this advice! Actually, I could solve the problem by iteratively adjusting the forecast sample period.

[kanonathena]

Possibly it is.

But how can I perform a dynamic forecast without using state space equations (since I am not familiar with it)? If I apply the forecast command there is no option to chose a value for n. Hence, I just perform a one-step ahead forecast before including the next observation, reestimating the model and calculating the next one-step ahead forecast.

But for my application it is crucial to calculate a n-step forecast before reestimating the model.

Thanks for your help and patience!

The forecast command lets you specify a SMPL for the period to be forecast. Perhaps that will help.

What is SMPL? I can't find such an option in forecast window.

I exactly the same problem as chrisw. I want to forecast inflation 8 quarters head using lags of forcasted inflation and actual value of other regressors (that is assuming you do not know future inflation but have perfect foresight of other regressors), adding one observation at a time.

[startz]

If you'r using an equation object, look at the field in the lower left of the Forecast window labelled "Forecast sample."

[kanonathena]

Thanks for the quick reply. I see what you mean. What should I do if I want the estimation sample to end at the 4th quarter of each year, so it is not adding one observation at a time but four(quarters)? And also my forecast horizon is 8 quarters head, how do I specify n-step ahead forecast without having to manually adjusting the sample every time?

[pablotillan]

I have the same question, I need to perfrom a 2,4 and 8 step ahead forecast from a VAR,AR and Random Walk models. I am using inflation, interest rate and unemployment data. The VAR has been performed with 9 lags. Can someone help me in do the forecast?
thanks!

[KrilleJ]

The RW forecast is simply the last value in the estimation sample.

For the equation object and the VAR object you can set the sample for which you want to do the forecast and then use the forecast command.

For example, with a VAR object (with series x and y and two lags) called var01 you can do

Code: Select all

smpl 1990Q1 2013Q4
var var01.ls 1 2 x y
smpl 2014Q1 2015Q4
var01.forecast(f=na) f

This will produce dynamic forecasts for the two series.

A corresponding methodology goes for the equation object.

If you want to do several forecasts, you can loop over the sample periods.

/Krille