TSLS with ARMA

[kiber_master]

Hello!

I have noticed that EViews 6 and EViews 7 give different results on TSLS with ARMA.
Was something changed in E7? Why new results give me worser R2?

Moreover, I've noticed, that no matter what type of initials I choose, E7 takes User. Will it fixed in some new release?

Below I attach my comparison in Excell-file. In this file time series was generated with known coefficients (1.234 for factor and 0.156 for AR(1)). Whith this knowledge E6 results are better than E7 results.

Hope to get answers on all my questions,
Thank you!

[startz]

It mightbe helpful to show exactly the command you used in each version.

[kiber_master]

I use "Quick\Estimate Equation..." and select tsls method. y - explained, x - explanatory, z - instrumental.

Here the workfile I've saved.

[startz]

It's possible that youo're running into either numerical error or multiple local maxima. The extremely high R-squared is at least suspicious. Note also that the result is sensitive to the inclusion of lagged instruments.

[kiber_master]

I've exclude lagged instruments and use only z for 1st step. In Excell you can see that series was generated by formula plus some error. It could explaine such high R-squared. It should be high and coefficients should be close to 1.234 and 0.156 respectively, am I right?

[EViews Gareth]

I just estimated that equation in both EV6 and EV7, and received identical results in each.

You appear to be right about the starting values though - I'll investigate.

[kiber_master]

If the results are identical, could you show them here? So I can check which of my version calculates wrong.

[EViews Gareth]

Code: Select all

X 1.232378691838361 AR(1) 0.1086086136395669

[kiber_master]

Thank you, I will explore, what was wrong with my examples.

[startz]

I get different answers from Gareth (using the current 7 release). The answers vary enormously depending on the starting values.

[kiber_master]

I get different answers from Gareth (using the current 7 release). The answers vary enormously depending on the starting values.

So that's what I'm talking about. I'm afraid, that one day I will get non-stationary process instead of stationary because of calculation error... =(

May be, you can give me an advice how to set initial values to get more adequate results?

[startz]

I get different answers from Gareth (using the current 7 release). The answers vary enormously depending on the starting values.

So that's what I'm talking about. I'm afraid, that one day I will get non-stationary process instead of stationary because of calculation error... =(

May be, you can give me an advice how to set initial values to get more adequate results?

It's not a calculation error really. Any nonlinear process can have multiple local optima. The best you can do, with any software, is to try multiple starting values and see if it makes a difference.

With AR models with real data, it's very unusual to see this problem. But it can happen.

[kiber_master]

Thank you for explanation!
I hope my example was enormous and rare.

[EViews Gareth]

As an aside, I've fixed the starting values from TSLS problem, and it should go out in a patch soon.