As most readers will know, gretl offers a --test-down option for the
ADF test. To date (until recent CVS), the only form of "testing
down" that we offered was that proposed by Ng and Perron in the
mid-1990s. Namely, we start at the maximum lag order, and delete the
last lag (of the difference of the variable to be tested) so long as
it is not "significant" at the 10 percent level on a two-tailed
The problem with this procedure is that (as later work by Ng and
Perron showed), it tends to over-parameterize the ADF regression,
hence reducing its power. A 2001 Econometrica article by these
authors suggested using a modified version of the AIC criterion to
optimize the lag order.
That's now implemented in gretl, via --test-down=MAIC (and for good
measure, we also offer the correspondingly modified Schwarz Bayesian
criterion, via --test-down=MBIC).
I think it would be a good idea to make the MAIC criterion the
default for "test-down". This would be backward incompatible, in the
sense that the same script, using the --test-down option with the
"adf" command, and with no additional parameter, would (in general)
produce different results under gretl >= 1.9.13 versus gretl <=
1.9.12. But you could still get the old behaviour under new gretl
So my question to you: Does this sound OK? Any objections?