Am 13.11.2018 um 18:55 schrieb Allin
Cottrell:
On Mon, 12
Nov 2018, Sven Schreiber wrote:
I may have "misspoken" in my original message on this in 2013.
What we actually did at that time, or not long thereafter, was to
replace use of the t-statistic as the (only) test-down criterion
with a choice of AIC, BIC or t-statistic, with AIC as the default.
However, the Modified AIC and BIC are specific to the ADF-GLS
method. So, for example, if you specify --test-down=BIC then MBIC
is used with GLS, plain BIC otherwise (and similarly for AIC).
OK
That's now rectified in git. I've also added a verbose switch.
...
OK, done in git.
Excellent, thanks!
- adf and coint2 have a --seasonals
option, but 'coint' doesn't, according to the doc and the
missing GUI item. Any particular reason why?
I guess not. But if we enabled it for "coint" would we want to
include the seasonals in the final regression (ADF test on uhat
from the cointegrating regression)? Or only in the initial ADF
tests?
Well, I haven't talked about the initial ADF tests, because in my
view they are not part of the Engle-Granger test which presupposes
I(1) variables I'd say. (BTW Jack, take note how there is an almost
officially endorsed pre-testing problem here, when you wanted to get
rid of something much weaker in the BoxCoxFuncForm package ;-)
What I meant was in the same place where all other deterministics
appear, in the cointegrating regression. (And by implication then
also in the previous ADF tests, if selected, I guess.) Definitely
not also in the final residual test regression.
As an imaginary example, suppose you wanted to test a Fisher-type
relationship for cointegration. So you have a nominal interest rate
which is not seasonal, and then you have inflation which is
seasonal. (And for the various good reasons out there you don't want
to use seasonally adjusted data.) If you ran the Engle-Granger 1st
stage regression without seasonal dummies, your residuals inherit
the seasonal pattern. If you run the 2nd stage on them, chances are
that that pattern is mistaken for mean reversion -> potentially
spurious cointegration. (Yes, with enough lags the pattern could
implicitly be modelled as seasonal quasi-unit roots. But it would
still be misspecified.)
thanks,
sven