forecast - confidence interval

[sfawver]

so i fit and forecast this equation and output the standard errors.

my question is why doesn't the forecasted value +/- 2 se equal the confidence intervals shown when i click the eq_1 object and click forecast?

[EViews Gareth]

The manual has a good description of this. Chapter 22 of User Guide II, Section "Forecasting from Equations with Expressions", sub-section "Plotted Standard Errors" and "Saved Forecast Standard Errors".

[sfawver]

took a look.

however, my quetsion has not been answered. the se's should match. why doesn't output +/-2 se match ci?

looks like bug to me.

[EViews Gareth]

Perhaps you could provide details of why you think they don't match? You don't actually perform any calculations in the program you posted, so it is hard to tell what (if anything) you've done incorrectly.

[sfawver]

did my own calc in excel using the se's.

the eviews se's are incorrect. they're only correct when ex: dep=dep(-1) and forecasting dep.

incorrect when ex: log(dep)=log(dep(-1)) and forecasting dep.

ch22 is bs. smells of bug to me.

[EViews Gareth]

Again, perhaps you could provide details of your calculations? It is hard for us to fix a bug without knowing exactly what the bug is.

[sfawver]

ok. insert this and run it. compare the results to what you see in the eq_1 forecast graph. they don't match.

series ci_up=(cdq_pct_se*2)+cdq_pct_saf
series ci_down=cdq_pct_saf-(cdq_pct_se*2)

[EViews Gareth]

Nor should they. As described in the manual, that isn't the correct way to do it. Following the exact description given in the manual, try this:

Code: Select all

eq_1.forecast(d) log_mean log_se freeze(forcgr) eq_1.forecast(g) mean se group g exp(log_mean) exp(log_mean+2*log_se) exp(log_mean-2*log_se) freeze(mangr) g.line show forcgr mangr

[sfawver]

why doesn't eviews do this automatically? what's the point of displaying the confidence intervals if they're not correct?

[sfawver]

what i meant is why are the standard errors not automatically adjusted? what is being calculated by default?

[EViews Gareth]

How could they be automatically adjusted? How would EViews know how many standard errors you were interested in? Indeed, how would EViews know what you were going to use the standard errors for?

[EViews Glenn]

Just to provide some context here for those who might be trying to follow along or those wondering whether there is indeed an issue...

The user is constructing a model specified with log(DEP) as the dependent variable and telling EViews to forecast DEP. The EViews forecast command constructs a forecast graph using the exponential of the fitted log(DEP) and the exponential of log(DEP)+/-2*SE, where SE is the forecast error of log(DEP). This provides the asymmetric confidence intervals associated with the appropriate nonlinear transformation of the estimation equation.

The forecast command also produces a series containing the forecasts computed from the exponential of the fitted log(DEP) and a series containing *approximate* forecast standard errors for exp(log(DEP)). Note the *approximation* here is important since the actual confidence bounds for exp(log(DEP)) are asymmetric. Since we can only put one thing into a single series we provide the standard errors associated with a linearization around the forecast (akin to doing a delta method in standard errors). As a consequence, taking the saved forecast series and adding plus and minus the saved forecast error will not match the forecast graph since the latter is a linear approximation to the correct interval.

Gareth politely pointed out that our manual has a detailed discussion of this issue, along with code examples of how to compute the equivalent of our forecast graph. I believe that should be sufficient.