Hi Matthias,

I think the log is a nice trick if there's non-constant variance. I usually test its impact with cross-validation