Fix test gmm poisson #1550

Merged
merged 2 commits into from Apr 4, 2014

Projects

None yet

2 participants

@josef-pkt
Member

fixes the test failures on scipy >= 0.13 in #1420

use log-linear OLS as starting values for GMM-Poisson

@jseabold
Member
jseabold commented Apr 4, 2014

Ah, good. That was easy. Odd that starting values matter so much between scipy versions only?

@jseabold jseabold merged commit 051c9ce into statsmodels:master Apr 4, 2014

1 check passed

continuous-integration/travis-ci The Travis CI build passed
Details
@josef-pkt
Member

Odd that starting values matter so much between scipy versions only?

I was going through the stepsize code in scipy optimize again, with pdb this time.
Because of the exp the gradient is large when we are too far away from the optimum, and the exp overflows pretty fast. But I didn't find an obvious way to limit the stepsize in scipy.
stepsize is the only code that changed in scipy 0.13 AFAICS.
inf and nan handling might have changed in scipy 0.10, IIRC.

BTW: Now that Optimizer is merged, I can switch to using it in GMM. I wanted to see if lbfgs is better, but it's not connected to GMM.

BTW: Given that I was reading about Poisson versus log-linear OLS, I realized that log-linear OLS makes pretty good starting values. TODO for discrete.

@josef-pkt josef-pkt added the PR label Apr 14, 2014
@josef-pkt josef-pkt deleted the josef-pkt:fix_test_gmm_poisson branch Jul 27, 2015
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment