Re: [Gretl-users] Variable selection
Am 16.06.2018 um 19:11 schrieb Riccardo (Jack) Lucchetti:
On Sat, 16 Jun 2018, Henrique Andrade wrote:
> I would like to make some speed tests using the approaches suggested
> by me
> and Jack before we decide the best way (in terms of speed) to use the
> powerset function inside Gretl. I think we could consider using C
> code too
> (the powerset is really a bottleneck when we have more than 15
> variables to
> combine).
I guess that the computational cost of the combinatorics is probably
comparable to the actual CPU usage for estimation. For example:
I was surprised by this statement, but for all variables (except the LHS
WW) in the dataset from mroz87.gdt I get this:
t0 = 0.32339221 # producing all combinations
t1 = 31.887871 # estimating them
So I think it's not necessary to implement the powerset thing in C (but
wouldn't hurt of course). Instead the natural speedup would be
parallelization of the estimations I think.
cheers,
sven