EViews.com

This thread is about the tscval add-in.

The add-in performs time-series cross validation (rolling out-of-sample accuracy assessment). It is intended to be used to evaluate competing specifications for equation and VAR objects. The add-in returns table and vector objects with error estimates at different horizons.

The add-in can be downloaded via the EViews Add-in manager or directly from GitHub at https://github.com/jameslamb/ML4EVIEWS/raw/master/Installers/tscval.aipz. Please see the docs https://github.com/jameslamb/ML4EVIEWS/blob/master/tscval/Docs/tscval.pdf for support. If those do not answer your question, post here and we will try to help you.

Happy coding,

James & Rita

A very nice and an extremely useful add-in. Thanks to you both...

Two minor updates in the last week:

v1.0.1 (April 4, 2016):

[1] Fixed a bug in the VAR proc that was causing tscval to try to set the sample outside the workfile range for users performing cross-validation up to the edge of the workfile range.

v1.0.2 (April 8, 2016)

[2] Fixed a bug in the VAR and EQ procs that was causing tscval to fail if called in GUI from untitled objects.


Thanks to CharlieEVIEWS for catching both of these!

An update this week:

v1.0.3 (May 31, 2016):

[1] Added median squared error (medSE) to the list of error options
[2] Fixed a small bug. A bit of code was referencing the nonexistent @med() function instead of @median().

Another small update:

v1.0.4 (June 1, 2016):

[1] Added median squared percentage error (medSPE) to the list of error options.

v1.0.5 (June 20, 2016):

[1] Added "mean forecast error" (MFE) and "median forecast error" (medFE) to the list of error options.

v1.0.6 (July 19, 2016):

[1] Added "keep_mats" option to the equation proc to allow users to view matrices of raw forecasts and errors.

v1.1.1 (July 26, 2016):

[1] Added tscv_score, a proc for scoring the error vector generated from tscval() (move to v1.1.* reflects this addition)
[2] Added documentation for tscv_score in tscval.pdf
[3] Fixed minor bug referencing an empty string in assigning a temporary name to the matrix of forecasts in tscval
[4] Fixed minor bug that was causing vector objects from the VAR proc to be overwritten if tscval had already been run in the wf
[5] Added keep_mats option to the VAR proc

v1.1.2 (Aug 10, 2016):

[1] Added "VAR_FORM" option for the equation proc, to allow users the option to see error stats for dependent variable
specified by expression (default in tscval is to use dependent variable with transformation removed)

Thanks to Yakup Kutsal Koca for suggesting this feature.

v1.1.3 (Nov 23, 2016):

[1] Fixed bug in which users trying to run tscval on legacy EViews version (pre EV9) were getting a cryptic error message.

Thanks to Eric Lascelles for reporting this.

Hi

how can I get/save the out of sample forecasts for a given variable for each period?

Thanks

Hello afonsosm, thanks for using tscval!

tscval does not currently allow you to retrieve all of the out of sample forecasts, but you can get matrices of all of the errors by specifying option . I say matrices (plural) because you'll get one for each type of error. If you choose an error type that preserves the sign and scale of the data (such as MAE), then you'll be able to access the raw differences between the target variable and each forecast.

For more documentation, see https://github.com/jameslamb/ML4EVIEWS/ ... tscval.pdf

Thanks @jaylamb20

For the exercise I'm performing, I really need to save the forecasts. (I'm doing bridge equation models to nowcast some variables, so I need the forecasts of the higher frequency variables).
If you join that possibility to the add-in, let us know!

Thank you very much!

@afonsosm unfortunately, tscval is not under active maintenance. I've recorded this feature request (https://github.com/jameslamb/ML4EVIEWS/issues/4) but I don't know when I'll be able to get to it.

You are welcome to contribute it, or to add any details to that liked issue to explain the desired functionality in more detail.