When I use this approach I'm seeing a Gaussian distribution for my severity predictions, even though the actual severity data clearly looks Gamma-distributed. I'm using log link. Hmm...