DS reg logreg is producing weak forecasts for down

[andybega]

outcome,from_year,for_years,gwcode,p_up,p_same,p_down
v2x_freexp_altinf,2011,2012 - 2013,2,0.020803108535437917,0.4791968914645621,0.5

Why is it always 0.5?

[andybega]

The lambda values picked are pretty high, so most coefficients are zeroed out.

When the forecast is constant, AUC-ROC and PR fail, btw.

One problem in debugging this is the random variation in CV splits that leads to slightly different results each run.

I've tried running with a different cost measure, that helps a little bit. Maybe using a 0.5 instead of 1 SE rule to pick the lambda value for any given alpha would also be better.