EViews.com

I've been using:

freeze(mode=overwrite,{%series}_stats) {%series}.autoarma(tform=auto,diff=2,select=aic,periods=12,kpsssig=5,eqname={%series}_eq,maxar=4,maxma=3,maxsar=1,maxsma=1,seed=131071,forclen={!fcst_length},etable,agraph,atable,fgraph) forecast

without issue.

However, when choosing the transform as Box-Cox (tform=bc, bc=2) I get errors stating "Attempt to raise a negative number to a non integer power". What is going on here? I have written a program that will find "optimal" lambdas and would like to use it, but even with the standard options I get this error.

Any ideas? Thanks

Is your copy of EViews up to date?


Seems to work for me:

We are using the Enterprise Edition of 10+ Nov 30 2017 Build. Check for update says we are good. Rebooted machine and still the error. Error does not occur when using the GUI. The error only occurs when the tform option changes from tform=auto to tform=bc, bc=2. I plan to pass the lambda to bc with the optimal one found.

Note that if I use bc=1, 0.5, 0, -0.5, -1 are ok.

if lambda is 2 or -2 or higher/lower, EViews fails. With the positive 2 I get the error "Attempt to raise a negative number to a non integer power"
if I se lambda of -2 I get error "^ function given an out of range argument"

Lambdas of 2 and -2 are common, so I am not sure why EViews is bombing at the command line and not the GUI.

The manual says bc=int Set the power of the Box-Cox transformation. Only applicable if the tform=bc option is used.

Again, the GUI handles everything fine.

Did you try the code:

Well I am totally messed up here...your code worked. I don't see where I am going wrong.

Here is my code. I still get the errors



Don't know why this does not work.

In the EViews database I am using, there are a total of 5 variables that exhibit this behaviour. That is, they cause crashes of EViews form within the program code, but work perfectly within the GUI. I can calculate Box-Cox transforms on my own and they work fine. This is a weird problem, which may be with EViews. Any other insight is great.

Would it be possible for you to provide the data?

Yes I can. Workfile is attached (hopefully)

Can confirm I receive the same error. We'll investigate.

Thanks...let me know when you can.

So what's going on is that for the (2,2,0,0) model, the forecast of the transform of the variable goes negative. Then EViews tries to un-do the transformation (which involves taking the square root), and fails, giving the error message.

You could try creating a series that is equal to the Box-Cox transformation of the variable yourself, doing the autoarma on it, selecting the correct model and then undoing the transformation manually. Yes this is a bit of a pain.

Thanks. I am guessing that autoarma is differencing (this is auto) the Box-Coxed variable and forecasting a drop (negative value) and transforming that negative drop back ("un Box-Cox"). This is problematic with a lambda of 2 (square root of negative number). I guess EViews tries to transform back before making a level forecast? If so, then perhaps it could be altered to take the forecasted change (which is what a difference is), add it to the previous Box-Cox value, then transform the whole thing back.

So my options, if wanting to use a Box-Cox transform are to:

1. Make the transform manually
2. Run autoarma to get the level forecasts (from differences if applicable)
3. Manually un-Box-Cox the forecast back.

This seem ok?

Thanks for your checking.

Right.