[Gretl-devel] Re: ACF and documentation of Bartlett standard errors

I think the choice of Bartlett standard errors for the correlogram in the GUI is relatively recent (not sure), while the corresponding --bartlett option for the corrgm command has been there for a long time. In any case, I was made aware of the special null hypothesis underlying the Bartlett variant: namely that for the interval at lag k the null hypothesis is an MA(k-1). This is not a problem (of course), but the standard confidence intervals come from white noise as the null hypothesis. So the choice of Bartlett standard errors in this case isn't just some robustification or small-sample thing (as opposed to a Bartlett correction in other contexts, for example), but involves a different view. So far so good, but I haven't found any mention of what the Bartlett errors actually mean. I have checked the GUI help text (for "Correlogram"), the documentation of the corrgm command, and the user guide. Actually, I was a bit surprised that the user guide and specifically the univariate time series chapter does not mention the ACF/PACF/correlogram at all. (Searching for "corrgm" or "ACF" doesn't show anything, and "correlogram" only shows up in a different context as cross-correlogram.) This is not a criticism, but just an observation -- maybe there are places I have overlooked. I think wherever these Bartlett standard errors are mentioned as an option, the different null hypothesis should be made explicit. I also think it would be good to put the word "Bartlett" into the ACF plot if it's chosen. I'm posting this here for discussion, maybe I'm misunderstanding the intention or assumption for those Bartlett standard errors, since I haven't checked the source code (yet).