EViews.com

Hello,

When trying to replicate both a dynamic and static forecast and the associated standard errors from an existing equation, Eviews has not been reproducing the same standard errors, even though the forecast is the same as it was before. In fact, the newly calculated standard error for each forecast (both dynamic and static) is constant for all periods.

The equation I am using to forecast has two lagged independent variables and an MA(1) term. This strange result for standard errors only occurs when I forecast the LOG of the dependent variable, but forecasting the dependent variable itself doesn't give anything immediately strange (unfortunately, there is no corresponding earlier forecast of the non-logged series, so I do cannot determine whether the results I am getting now for that are strange or not).

There are other similar equations in the same workfile with similar strange results.

I am currently running the most up-to-date Eviews 7.2 (including 4/4/2012 patch). The workfile (and equations of interest) was initially created in 7.0, I believe with no updates applied. I asked a colleague running 7.0 to try to replicate these forecasts, and she successfully did so. She then upgraded her installation to the 7.2 (as of the 4/4/2012) and experienced the same issue that I described above (with the exact same strange results, at least).

I've attached a screenshot of the issue. The Workfile window has the relevant forecasts series highlighted to show that the command is the same with just a different output series name. The prefix "new_" was appended to the replications. The Equation window to the right is the equation from which I am trying to forecast. The two Group windows below (dynamic on the left, static on the right) have the following layout from left to right: original forecast, original SEs, new forecast, new SEs. The forecasts match while the SEs (which I've highlighted for clarity) do not.

Did we overlook something that may cause this type of result with the same forecasts with different standard errors? Please let me know whether I can provide more information or a clarification.

Many thanks!
Mickey

If I'm reading correctly, you're saying that EViews 7.0 gives different forecast standard errors for ARMA, from 7.2?

If so, that's because there was a bug fix:
http://eviews.com/download/ev7download.html

Hi Gareth,

Well, I just ran a few variations of the example equations above, and it seems that when the only ARMA term is an MA(n) term (I tried MA(2) and MA(3), in addition to the MA(1) above), I get a constant SE for all periods in the static forecast (which is sampled from 2004Q1 to 2010Q4), and that same value for the SE series for the first n periods of the dynamic forecast. At n+1 and on, there is a different constant value for the SEs through the end of the forecast sample (which is 2011Q1 through 2015Q4).

It's the constant values for the SEs that concerns us. Can you please confirm whether this is reasonable to expect, given the bug fix you mentioned?

Thanks,
Mickey

Yes, it should go to a constant value.

Note there is currently a bug, currently being fixed, that prevents coefficient uncertainty from being added to forecast SEs. This shouldn't affect them by much though.

Hi,

I think I have stumbled on the same problem.

I am editing an econometrics course which has upgraded from EViews 6 to EViews 7, and I am running Student Edition 7.2 June 13 2012 build.

I am forecasting with AR and MA models, using dynamic forecasting. Using the exact same data and estimated models, the forecast values are the same for EViews 6 and EViews 7, but the forecast standard errors are not the same. In Eviews 6 the forecast standard errors converge to a constant, but in EV7 they get larger over time.

If I turn off Coef uncertainty in SE calc in EV7, then the forecast SEs do converge to a constant, but not the same value as in EViews 6 (including Coef uncertainy in SE calc).
I am not sure how to interpret such results or how to explain to students what to do in the circumstances.

Is there anything I can do in Eviews 7 to get the same results I got in EViews 6? I think I may just have to skip the forecast standard errors for these models.

Thanks and best wishes.

Jon

UPDATE

Hi again,
I have answered my own post – always a good result!
I looked a little deeper into the formula for the forecast standard error for AR and MA models (which are not included in the EViews help files).
If I untick the box Coef uncertainty in SE calcs when doing the dynamic forecast from an AR1 model, then the SE for the forecast one-period ahead is the SE of the estimated model, as it should be. And the se for forecasts further ahead is equal to the value I get by hand using the formula.
So I can explain and interpret that.
If you could answer the query about the bug in EV7 in the calculation of forecast standard errors in dynamic forecasting, that would be helpful, please, and why it should be different to EV6.
Regards,
Jon

The bug in EV7 was fixed. The bug still exists in EV6.