Multiple Models for Different Time Length (Model Switching)

For technical questions regarding estimation of single equations, systems, VARs, Factor analysis and State Space Models in EViews. General econometric questions and advice should go in the Econometric Discussions forum.

Moderators: EViews Gareth, EViews Moderator

kszynkar
Posts: 45
Joined: Tue Feb 02, 2016 1:15 am
Location: Switzerland

Multiple Models for Different Time Length (Model Switching)

Postby kszynkar » Thu Apr 07, 2016 1:02 am

Hi,

How would I organize multiple models and switch between them for different forecast horizons?

For example, I may use one model to forecast one variable for one month, and (dynamically) for the second month's forecast. However, I may want to switch to another model for the 3-6 month forecast, then another for the 6+ month forecast.

In my case, I have multiple variables I'm doing this for, so am starting off with a system of equations and an EViews "model" object that I'm "solving" for the 1 model per variable case.

I was wondering if there was a good way to organize these models and switch between them in EViews.

Thanks

trubador
Did you use forum search?
Posts: 1518
Joined: Thu Nov 20, 2008 12:04 pm

Re: Multiple Models for Different Time Length (Model Switchi

Postby trubador » Thu Apr 07, 2016 3:02 am

I am not a fan of "model switching" approach, but you need to write your own program in EViews if that's what you want. Why don't you try "Forecast Averaging" instead? http://www.eviews.com/help/helpintro.ht ... 24.html%23

kszynkar
Posts: 45
Joined: Tue Feb 02, 2016 1:15 am
Location: Switzerland

Re: Multiple Models for Different Time Length (Model Switchi

Postby kszynkar » Thu Apr 07, 2016 4:22 am

The problem with averaging is that the short-term models are not at all accurate over longer time horizons because the error builds up too quickly. I also tried to add the independent variables from the long-term models into the short-run model to make a "supermodel", then the coefficients are not significant or nonsense.

That's why I wanted to switch models, or blend them to at least get some benefit when error from one model vs. another was "borderline" for selecting between the two, and was wondering how best to organize/program this.

kszynkar
Posts: 45
Joined: Tue Feb 02, 2016 1:15 am
Location: Switzerland

Re: Multiple Models for Different Time Length (Model Switchi

Postby kszynkar » Thu Apr 07, 2016 4:47 am

The way I was thinking was:
If I were to have m models and n variables, to produce m*n forecasts, then select between each of the m forecasts for each variable

Or is there a better way?

trubador
Did you use forum search?
Posts: 1518
Joined: Thu Nov 20, 2008 12:04 pm

Re: Multiple Models for Different Time Length (Model Switchi

Postby trubador » Thu Apr 07, 2016 5:22 am

So you have a mixture of short term and long term models at hand? I always find ARDL, VAR and VEC models very useful in both regards, so would stick to forecast averaging. Anyway, in your case, I think you can first compare (and improve if necessary) your models' ability to forecast with a cross-validation exercise. For that purpose you can try tscval add-in. You can also use these results further to choose between models as it implements an exhaustive list of forecast evaluation criteria. I cannot think of any particular selection algorithm at the moment, but it should not be that difficult to program it in EViews as you already have built your models.

kszynkar
Posts: 45
Joined: Tue Feb 02, 2016 1:15 am
Location: Switzerland

Re: Multiple Models for Different Time Length (Model Switchi

Postby kszynkar » Thu Apr 07, 2016 5:36 am

I'll try that out. Thanks for the help

kszynkar
Posts: 45
Joined: Tue Feb 02, 2016 1:15 am
Location: Switzerland

Re: Multiple Models for Different Time Length (Model Switchi

Postby kszynkar » Mon Apr 25, 2016 12:41 am

I ended up using a VEC for what it's worth. Thanks for the help


Return to “Estimation”

Who is online

Users browsing this forum: No registered users and 31 guests