Elastic net fits are pretty powerful and it'd be nice to cite back my other paper in JASSS.
The problem is that they do not produce very good standard errors because LASSO just isn't consistent and as such there is no kosher way of generating prediction intervals either.
What I think I will do is use the trick from "learning form data" book which argues that if consistency is important we can re-run the standard OLS with the parameters selected by LASSO.
The problem here is that elastic net also have a ridge component that would get lost. So you potentially end up predicting quite far (and I worry about crazy results where the estimated point is not even in the confidence interval). However on the other hand OLS bootstrap prediction intervals are of very high quality in all simulations so that it will probably still be quite effective.
Elastic net fits are pretty powerful and it'd be nice to cite back my other paper in JASSS.
The problem is that they do not produce very good standard errors because LASSO just isn't consistent and as such there is no kosher way of generating prediction intervals either.
What I think I will do is use the trick from "learning form data" book which argues that if consistency is important we can re-run the standard OLS with the parameters selected by LASSO.
The problem here is that elastic net also have a ridge component that would get lost. So you potentially end up predicting quite far (and I worry about crazy results where the estimated point is not even in the confidence interval). However on the other hand OLS bootstrap prediction intervals are of very high quality in all simulations so that it will probably still be quite effective.