autoarma in batch run

[LindaR]

Hi,
I am using quarterly data with autoarma to generate 5 year out forecasts in EViews 9.5. When I run the code on any single series it works fine but when I try to do all the series (441) the program crashes - I get a message saying that EViews9 has stopped working and I need to close it - but the number of series it manages to process before crashing varies. Any suggestions?

[EViews Gareth]

Can you provide the data too?

[LindaR]

Sure thing, here it is, thanks.

[EViews Gareth]

Is your copy of EViews up to date (Help->About EViews).

[LindaR]

Yes, I think so - EViews 9 ((.5 as far as I am aware).

[LindaR]

Sorry, didn't read all your post - it is the Enterprise Edition May 1 2015 build, thanks.

[EViews Gareth]

Ah, that's very out of date - please update to the latest version.

[LindaR]

I finally got out IT to update EViews and it is working fine but the model estimates I get do not seem to be stable - if I run the program twice I get different models selected for some series, sometimes even if they have converged within the 500 iterations - is this normal and if so why ? I attach the data and program.

I have some additional questions about autoarma:
1. If you get a dependent variable like D(serxx, 2) does the , 2 mean it has been differenced twice?
2. Is it possible to retrieve how the starting values have been set?
3. What is the data-based algorithm used to determine starting values in the first instance?
4. Is it possible to increase the number of iterations?

Thanks for your assistance,

Linda

[EViews Gareth]

For arma models that do not converge with the standard starting value algorithms, EViews resorts to random starting values, which may mean that running it twice will give different results.

1. Yes.
2. No.
3. Least squares regression of residuals against lagged residuals (for AR terms) and innovations (for MA terms), where innovations are obtained by first regressing residuals against many lags of residuals. The number of lags to use for MA is determined by:
whiten_lags = floor(_min((nobs-maxar)/3.0,12.0) * pow(nobs/100.0, 0.25));
4. It uses the global option (Options->General Options->Estimation)

[LindaR]

Thanks Gareth,
This is really helpful. A lot to digest - I will have to take some time to think about it.
Linda