@@mdevans43 makes sense. Im writing a master thesis of XGboost for pure premium housing insurances in denmark. I have a problem with modelling the frequency, as a lot of the frequencies will be 0, due to no claims during the year. What would your approach be to model frequency? My thought is to model the frequency first, then the severity. And in the end multiply those for the pure premium.
Thanks for the video. One point should be mentioned: xgboost isn't using gradient descent internally so what you are saying about eta from 10:18 is not accurate.