[Gretl-devel] Re: nonlinearity tests, etc. (2)
Am 28.02.2021 um 17:07 schrieb Allin Cottrell:
On Sun, 28 Feb 2021, Sven Schreiber wrote:
I can see this could be useful if you just have a residual series
you
want to test (e.g. after ARMA). So I've put it into git on a trial
basis. There's a new C-plugin, bdstest, which supports (for now) a
"hidden" function, _bdstest(). It's based on LeBaron's "fast"
code.
Whoa, even though there was LeBaron's code, that was super quick!
if we pursue this it should probably be expressed as a fraction of
the
standard deviation of x (a rule of thumb seems to be eps = 0.5*sd(x)).
That Kanzler (1999) paper also investigated something like that. He
seems to recommend either eps closer to 2*sd(x), or for other cases to
calibrate it such that the dim-1 correlation integral becomes around 0.8.
1 near the top of the file. In that mode the program does Monte
Carlo,
which reveals that the test over-rejects quite substantially for series
of typical econometric length -- so if we pursue this we may want to
look into MBDS.
Yes; but even in the original form where the test is liberal it could be
useful I think. At least you know that there's probably nothing you need
to do after, say, ARMA if it cannot reject. Refinements could be added
later I guess - Eviews offers a lot of tweaks and variants, but it seems
to be the only package to do so.
thanks
sven